1 // SPDX-License-Identifier: GPL-2.0
3 #include "alloc_background.h"
4 #include "alloc_foreground.h"
5 #include "backpointers.h"
6 #include "btree_cache.h"
8 #include "btree_key_cache.h"
9 #include "btree_update.h"
10 #include "btree_update_interior.h"
12 #include "btree_write_buffer.h"
14 #include "buckets_waiting_for_journal.h"
24 #include <linux/kthread.h>
25 #include <linux/math64.h>
26 #include <linux/random.h>
27 #include <linux/rculist.h>
28 #include <linux/rcupdate.h>
29 #include <linux/sched/task.h>
30 #include <linux/sort.h>
32 /* Persistent alloc info: */
34 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
35 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
40 struct bkey_alloc_unpacked {
47 #define x(_name, _bits) u##_bits _name;
52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
53 const void **p, unsigned field)
55 unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
58 if (!(a->fields & (1 << field)))
63 v = *((const u8 *) *p);
82 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
85 const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
86 const void *d = in->data;
91 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
96 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
99 struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
100 const u8 *in = a.v->data;
101 const u8 *end = bkey_val_end(a);
102 unsigned fieldnr = 0;
107 out->oldest_gen = a.v->oldest_gen;
108 out->data_type = a.v->data_type;
110 #define x(_name, _bits) \
111 if (fieldnr < a.v->nr_fields) { \
112 ret = bch2_varint_decode_fast(in, end, &v); \
120 if (v != out->_name) \
124 BCH_ALLOC_FIELDS_V2()
129 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
132 struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
133 const u8 *in = a.v->data;
134 const u8 *end = bkey_val_end(a);
135 unsigned fieldnr = 0;
140 out->oldest_gen = a.v->oldest_gen;
141 out->data_type = a.v->data_type;
142 out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
143 out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
144 out->journal_seq = le64_to_cpu(a.v->journal_seq);
146 #define x(_name, _bits) \
147 if (fieldnr < a.v->nr_fields) { \
148 ret = bch2_varint_decode_fast(in, end, &v); \
156 if (v != out->_name) \
160 BCH_ALLOC_FIELDS_V2()
165 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
167 struct bkey_alloc_unpacked ret = { .gen = 0 };
171 bch2_alloc_unpack_v1(&ret, k);
173 case KEY_TYPE_alloc_v2:
174 bch2_alloc_unpack_v2(&ret, k);
176 case KEY_TYPE_alloc_v3:
177 bch2_alloc_unpack_v3(&ret, k);
184 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
186 unsigned i, bytes = offsetof(struct bch_alloc, data);
188 for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
189 if (a->fields & (1 << i))
190 bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
192 return DIV_ROUND_UP(bytes, sizeof(u64));
195 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
196 enum bkey_invalid_flags flags,
197 struct printbuf *err)
199 struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
201 /* allow for unknown fields */
202 if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
203 prt_printf(err, "incorrect value size (%zu < %u)",
204 bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
205 return -BCH_ERR_invalid_bkey;
211 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
212 enum bkey_invalid_flags flags,
213 struct printbuf *err)
215 struct bkey_alloc_unpacked u;
217 if (bch2_alloc_unpack_v2(&u, k)) {
218 prt_printf(err, "unpack error");
219 return -BCH_ERR_invalid_bkey;
225 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
226 enum bkey_invalid_flags flags,
227 struct printbuf *err)
229 struct bkey_alloc_unpacked u;
231 if (bch2_alloc_unpack_v3(&u, k)) {
232 prt_printf(err, "unpack error");
233 return -BCH_ERR_invalid_bkey;
239 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
240 enum bkey_invalid_flags flags, struct printbuf *err)
242 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
244 if (alloc_v4_u64s(a.v) > bkey_val_u64s(k.k)) {
245 prt_printf(err, "bad val size (%u > %zu)",
246 alloc_v4_u64s(a.v), bkey_val_u64s(k.k));
247 return -BCH_ERR_invalid_bkey;
250 if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
251 BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
252 prt_printf(err, "invalid backpointers_start");
253 return -BCH_ERR_invalid_bkey;
256 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
257 prt_printf(err, "invalid data type (got %u should be %u)",
258 a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
259 return -BCH_ERR_invalid_bkey;
262 switch (a.v->data_type) {
264 case BCH_DATA_need_gc_gens:
265 case BCH_DATA_need_discard:
266 if (a.v->dirty_sectors ||
267 a.v->cached_sectors ||
269 prt_printf(err, "empty data type free but have data");
270 return -BCH_ERR_invalid_bkey;
274 case BCH_DATA_journal:
277 case BCH_DATA_parity:
278 if (!a.v->dirty_sectors) {
279 prt_printf(err, "data_type %s but dirty_sectors==0",
280 bch2_data_types[a.v->data_type]);
281 return -BCH_ERR_invalid_bkey;
284 case BCH_DATA_cached:
285 if (!a.v->cached_sectors ||
286 a.v->dirty_sectors ||
288 prt_printf(err, "data type inconsistency");
289 return -BCH_ERR_invalid_bkey;
292 if (!a.v->io_time[READ] &&
293 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_to_lru_refs) {
294 prt_printf(err, "cached bucket with read_time == 0");
295 return -BCH_ERR_invalid_bkey;
298 case BCH_DATA_stripe:
305 static inline u64 swab40(u64 x)
307 return (((x & 0x00000000ffULL) << 32)|
308 ((x & 0x000000ff00ULL) << 16)|
309 ((x & 0x0000ff0000ULL) >> 0)|
310 ((x & 0x00ff000000ULL) >> 16)|
311 ((x & 0xff00000000ULL) >> 32));
314 void bch2_alloc_v4_swab(struct bkey_s k)
316 struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
317 struct bch_backpointer *bp, *bps;
319 a->journal_seq = swab64(a->journal_seq);
320 a->flags = swab32(a->flags);
321 a->dirty_sectors = swab32(a->dirty_sectors);
322 a->cached_sectors = swab32(a->cached_sectors);
323 a->io_time[0] = swab64(a->io_time[0]);
324 a->io_time[1] = swab64(a->io_time[1]);
325 a->stripe = swab32(a->stripe);
326 a->nr_external_backpointers = swab32(a->nr_external_backpointers);
328 bps = alloc_v4_backpointers(a);
329 for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
330 bp->bucket_offset = swab40(bp->bucket_offset);
331 bp->bucket_len = swab32(bp->bucket_len);
332 bch2_bpos_swab(&bp->pos);
336 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
338 struct bch_alloc_v4 _a;
339 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
343 printbuf_indent_add(out, 2);
345 prt_printf(out, "gen %u oldest_gen %u data_type %s",
346 a->gen, a->oldest_gen,
347 a->data_type < BCH_DATA_NR
348 ? bch2_data_types[a->data_type]
349 : "(invalid data type)");
351 prt_printf(out, "journal_seq %llu", a->journal_seq);
353 prt_printf(out, "need_discard %llu", BCH_ALLOC_V4_NEED_DISCARD(a));
355 prt_printf(out, "need_inc_gen %llu", BCH_ALLOC_V4_NEED_INC_GEN(a));
357 prt_printf(out, "dirty_sectors %u", a->dirty_sectors);
359 prt_printf(out, "cached_sectors %u", a->cached_sectors);
361 prt_printf(out, "stripe %u", a->stripe);
363 prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
365 prt_printf(out, "io_time[READ] %llu", a->io_time[READ]);
367 prt_printf(out, "io_time[WRITE] %llu", a->io_time[WRITE]);
369 prt_printf(out, "fragmentation %llu", a->fragmentation_lru);
371 prt_printf(out, "bp_start %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
374 if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
375 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
376 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
378 prt_printf(out, "backpointers: %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
379 printbuf_indent_add(out, 2);
381 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
383 bch2_backpointer_to_text(out, &bps[i]);
386 printbuf_indent_sub(out, 2);
389 printbuf_indent_sub(out, 2);
392 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
394 if (k.k->type == KEY_TYPE_alloc_v4) {
397 *out = *bkey_s_c_to_alloc_v4(k).v;
399 src = alloc_v4_backpointers(out);
400 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
401 dst = alloc_v4_backpointers(out);
404 memset(src, 0, dst - src);
406 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(out, 0);
408 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
410 *out = (struct bch_alloc_v4) {
411 .journal_seq = u.journal_seq,
412 .flags = u.need_discard,
414 .oldest_gen = u.oldest_gen,
415 .data_type = u.data_type,
416 .stripe_redundancy = u.stripe_redundancy,
417 .dirty_sectors = u.dirty_sectors,
418 .cached_sectors = u.cached_sectors,
419 .io_time[READ] = u.read_time,
420 .io_time[WRITE] = u.write_time,
424 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
428 static noinline struct bkey_i_alloc_v4 *
429 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
431 struct bkey_i_alloc_v4 *ret;
433 ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4)));
437 if (k.k->type == KEY_TYPE_alloc_v4) {
440 bkey_reassemble(&ret->k_i, k);
442 src = alloc_v4_backpointers(&ret->v);
443 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
444 dst = alloc_v4_backpointers(&ret->v);
447 memset(src, 0, dst - src);
449 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v, 0);
450 set_alloc_v4_u64s(ret);
452 bkey_alloc_v4_init(&ret->k_i);
454 bch2_alloc_to_v4(k, &ret->v);
459 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
461 struct bkey_s_c_alloc_v4 a;
463 if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
464 ((a = bkey_s_c_to_alloc_v4(k), true) &&
465 BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) == 0))
466 return bch2_bkey_make_mut_noupdate_typed(trans, k, alloc_v4);
468 return __bch2_alloc_to_v4_mut(trans, k);
471 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
473 return bch2_alloc_to_v4_mut_inlined(trans, k);
476 struct bkey_i_alloc_v4 *
477 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
481 struct bkey_i_alloc_v4 *a;
484 k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos,
485 BTREE_ITER_WITH_UPDATES|
492 a = bch2_alloc_to_v4_mut_inlined(trans, k);
493 ret = PTR_ERR_OR_ZERO(a);
498 bch2_trans_iter_exit(trans, iter);
502 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
504 *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
506 pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
510 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
512 pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
513 pos.offset += offset;
517 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
519 return k.k->type == KEY_TYPE_bucket_gens
520 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
524 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
525 enum bkey_invalid_flags flags,
526 struct printbuf *err)
528 if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
529 prt_printf(err, "bad val size (%zu != %zu)",
530 bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
531 return -BCH_ERR_invalid_bkey;
537 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
539 struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
542 for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
545 prt_printf(out, "%u", g.v->gens[i]);
549 int bch2_bucket_gens_init(struct bch_fs *c)
551 struct btree_trans *trans = bch2_trans_get(c);
552 struct btree_iter iter;
554 struct bch_alloc_v4 a;
555 struct bkey_i_bucket_gens g;
556 bool have_bucket_gens_key = false;
562 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
563 BTREE_ITER_PREFETCH, k, ret) {
565 * Not a fsck error because this is checked/repaired by
566 * bch2_check_alloc_key() which runs later:
568 if (!bch2_dev_bucket_exists(c, k.k->p))
571 gen = bch2_alloc_to_v4(k, &a)->gen;
572 pos = alloc_gens_pos(iter.pos, &offset);
574 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
575 ret = commit_do(trans, NULL, NULL,
577 BTREE_INSERT_LAZY_RW,
578 bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0));
581 have_bucket_gens_key = false;
584 if (!have_bucket_gens_key) {
585 bkey_bucket_gens_init(&g.k_i);
587 have_bucket_gens_key = true;
590 g.v.gens[offset] = gen;
592 bch2_trans_iter_exit(trans, &iter);
594 if (have_bucket_gens_key && !ret)
595 ret = commit_do(trans, NULL, NULL,
597 BTREE_INSERT_LAZY_RW,
598 bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0));
600 bch2_trans_put(trans);
607 int bch2_alloc_read(struct bch_fs *c)
609 struct btree_trans *trans = bch2_trans_get(c);
610 struct btree_iter iter;
615 down_read(&c->gc_lock);
617 if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_bucket_gens) {
618 const struct bch_bucket_gens *g;
621 for_each_btree_key(trans, iter, BTREE_ID_bucket_gens, POS_MIN,
622 BTREE_ITER_PREFETCH, k, ret) {
623 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
624 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
626 if (k.k->type != KEY_TYPE_bucket_gens)
629 g = bkey_s_c_to_bucket_gens(k).v;
632 * Not a fsck error because this is checked/repaired by
633 * bch2_check_alloc_key() which runs later:
635 if (!bch2_dev_exists2(c, k.k->p.inode))
638 ca = bch_dev_bkey_exists(c, k.k->p.inode);
640 for (b = max_t(u64, ca->mi.first_bucket, start);
641 b < min_t(u64, ca->mi.nbuckets, end);
643 *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
645 bch2_trans_iter_exit(trans, &iter);
647 struct bch_alloc_v4 a;
649 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
650 BTREE_ITER_PREFETCH, k, ret) {
652 * Not a fsck error because this is checked/repaired by
653 * bch2_check_alloc_key() which runs later:
655 if (!bch2_dev_bucket_exists(c, k.k->p))
658 ca = bch_dev_bkey_exists(c, k.k->p.inode);
660 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
662 bch2_trans_iter_exit(trans, &iter);
665 bch2_trans_put(trans);
666 up_read(&c->gc_lock);
674 /* Free space/discard btree: */
676 static int bch2_bucket_do_index(struct btree_trans *trans,
677 struct bkey_s_c alloc_k,
678 const struct bch_alloc_v4 *a,
681 struct bch_fs *c = trans->c;
682 struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
683 struct btree_iter iter;
687 enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
688 enum bch_bkey_type new_type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
689 struct printbuf buf = PRINTBUF;
692 if (a->data_type != BCH_DATA_free &&
693 a->data_type != BCH_DATA_need_discard)
696 k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
701 k->k.type = new_type;
703 switch (a->data_type) {
705 btree = BTREE_ID_freespace;
706 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
707 bch2_key_resize(&k->k, 1);
709 case BCH_DATA_need_discard:
710 btree = BTREE_ID_need_discard;
711 k->k.p = alloc_k.k->p;
717 old = bch2_bkey_get_iter(trans, &iter, btree,
718 bkey_start_pos(&k->k),
724 if (ca->mi.freespace_initialized &&
725 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info &&
726 bch2_trans_inconsistent_on(old.k->type != old_type, trans,
727 "incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n"
729 set ? "setting" : "clearing",
730 bch2_btree_ids[btree],
733 bch2_bkey_types[old.k->type],
734 bch2_bkey_types[old_type],
735 (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
740 ret = bch2_trans_update(trans, &iter, k, 0);
742 bch2_trans_iter_exit(trans, &iter);
747 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
748 struct bpos bucket, u8 gen)
750 struct btree_iter iter;
752 struct bpos pos = alloc_gens_pos(bucket, &offset);
753 struct bkey_i_bucket_gens *g;
757 g = bch2_trans_kmalloc(trans, sizeof(*g));
758 ret = PTR_ERR_OR_ZERO(g);
762 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos,
764 BTREE_ITER_WITH_UPDATES);
769 if (k.k->type != KEY_TYPE_bucket_gens) {
770 bkey_bucket_gens_init(&g->k_i);
773 bkey_reassemble(&g->k_i, k);
776 g->v.gens[offset] = gen;
778 ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
779 bch2_trans_iter_exit(trans, &iter);
783 int bch2_trans_mark_alloc(struct btree_trans *trans,
784 enum btree_id btree_id, unsigned level,
785 struct bkey_s_c old, struct bkey_i *new,
788 struct bch_fs *c = trans->c;
789 struct bch_alloc_v4 old_a_convert, *new_a;
790 const struct bch_alloc_v4 *old_a;
791 u64 old_lru, new_lru;
795 * Deletion only happens in the device removal path, with
796 * BTREE_TRIGGER_NORUN:
798 BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
800 old_a = bch2_alloc_to_v4(old, &old_a_convert);
801 new_a = &bkey_i_to_alloc_v4(new)->v;
803 new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
805 if (new_a->dirty_sectors > old_a->dirty_sectors ||
806 new_a->cached_sectors > old_a->cached_sectors) {
807 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
808 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
809 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
810 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
813 if (data_type_is_empty(new_a->data_type) &&
814 BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
815 !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
817 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
820 if (old_a->data_type != new_a->data_type ||
821 (new_a->data_type == BCH_DATA_free &&
822 alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
823 ret = bch2_bucket_do_index(trans, old, old_a, false) ?:
824 bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
829 if (new_a->data_type == BCH_DATA_cached &&
830 !new_a->io_time[READ])
831 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
833 old_lru = alloc_lru_idx_read(*old_a);
834 new_lru = alloc_lru_idx_read(*new_a);
836 if (old_lru != new_lru) {
837 ret = bch2_lru_change(trans, new->k.p.inode,
838 bucket_to_u64(new->k.p),
844 new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
845 bch_dev_bkey_exists(c, new->k.p.inode));
847 if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
848 ret = bch2_lru_change(trans,
849 BCH_LRU_FRAGMENTATION_START,
850 bucket_to_u64(new->k.p),
851 old_a->fragmentation_lru, new_a->fragmentation_lru);
856 if (old_a->gen != new_a->gen) {
857 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
866 * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
867 * extents style btrees, but works on non-extents btrees:
869 static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
871 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
879 struct btree_iter iter2;
882 bch2_trans_copy_iter(&iter2, iter);
884 if (!bpos_eq(iter->path->l[0].b->key.k.p, SPOS_MAX))
885 end = bkey_min(end, bpos_nosnap_successor(iter->path->l[0].b->key.k.p));
887 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
890 * btree node min/max is a closed interval, upto takes a half
893 k = bch2_btree_iter_peek_upto(&iter2, end);
895 bch2_trans_iter_exit(iter->trans, &iter2);
897 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
905 bch2_key_resize(hole, next.offset - iter->pos.offset);
906 return (struct bkey_s_c) { hole, NULL };
910 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
915 if (bch2_dev_bucket_exists(c, *bucket))
918 if (bch2_dev_exists2(c, bucket->inode)) {
919 ca = bch_dev_bkey_exists(c, bucket->inode);
921 if (bucket->offset < ca->mi.first_bucket) {
922 bucket->offset = ca->mi.first_bucket;
931 iter = bucket->inode;
932 ca = __bch2_next_dev(c, &iter, NULL);
934 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
940 static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
942 struct bch_fs *c = iter->trans->c;
945 k = bch2_get_key_or_hole(iter, POS_MAX, hole);
950 struct bpos bucket = bkey_start_pos(k.k);
952 if (!bch2_dev_bucket_exists(c, bucket)) {
953 if (!next_bucket(c, &bucket))
954 return bkey_s_c_null;
956 bch2_btree_iter_set_pos(iter, bucket);
960 if (!bch2_dev_bucket_exists(c, k.k->p)) {
961 struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
963 bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
970 static noinline_for_stack
971 int bch2_check_alloc_key(struct btree_trans *trans,
972 struct bkey_s_c alloc_k,
973 struct btree_iter *alloc_iter,
974 struct btree_iter *discard_iter,
975 struct btree_iter *freespace_iter,
976 struct btree_iter *bucket_gens_iter)
978 struct bch_fs *c = trans->c;
980 struct bch_alloc_v4 a_convert;
981 const struct bch_alloc_v4 *a;
982 unsigned discard_key_type, freespace_key_type;
983 unsigned gens_offset;
985 struct printbuf buf = PRINTBUF;
988 if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
989 "alloc key for invalid device:bucket %llu:%llu",
990 alloc_k.k->p.inode, alloc_k.k->p.offset))
991 return bch2_btree_delete_at(trans, alloc_iter, 0);
993 ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
994 if (!ca->mi.freespace_initialized)
997 a = bch2_alloc_to_v4(alloc_k, &a_convert);
999 discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1000 bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1001 k = bch2_btree_iter_peek_slot(discard_iter);
1006 if (k.k->type != discard_key_type &&
1007 (c->opts.reconstruct_alloc ||
1008 fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1010 bch2_bkey_types[k.k->type],
1011 bch2_bkey_types[discard_key_type],
1012 (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1013 struct bkey_i *update =
1014 bch2_trans_kmalloc(trans, sizeof(*update));
1016 ret = PTR_ERR_OR_ZERO(update);
1020 bkey_init(&update->k);
1021 update->k.type = discard_key_type;
1022 update->k.p = discard_iter->pos;
1024 ret = bch2_trans_update(trans, discard_iter, update, 0);
1029 freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1030 bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1031 k = bch2_btree_iter_peek_slot(freespace_iter);
1036 if (k.k->type != freespace_key_type &&
1037 (c->opts.reconstruct_alloc ||
1038 fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1040 bch2_bkey_types[k.k->type],
1041 bch2_bkey_types[freespace_key_type],
1042 (printbuf_reset(&buf),
1043 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1044 struct bkey_i *update =
1045 bch2_trans_kmalloc(trans, sizeof(*update));
1047 ret = PTR_ERR_OR_ZERO(update);
1051 bkey_init(&update->k);
1052 update->k.type = freespace_key_type;
1053 update->k.p = freespace_iter->pos;
1054 bch2_key_resize(&update->k, 1);
1056 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1061 bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1062 k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1067 if (a->gen != alloc_gen(k, gens_offset) &&
1068 (c->opts.reconstruct_alloc ||
1069 fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1071 alloc_gen(k, gens_offset), a->gen,
1072 (printbuf_reset(&buf),
1073 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1074 struct bkey_i_bucket_gens *g =
1075 bch2_trans_kmalloc(trans, sizeof(*g));
1077 ret = PTR_ERR_OR_ZERO(g);
1081 if (k.k->type == KEY_TYPE_bucket_gens) {
1082 bkey_reassemble(&g->k_i, k);
1084 bkey_bucket_gens_init(&g->k_i);
1085 g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1088 g->v.gens[gens_offset] = a->gen;
1090 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1096 printbuf_exit(&buf);
1100 static noinline_for_stack
1101 int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1104 struct btree_iter *freespace_iter)
1106 struct bch_fs *c = trans->c;
1109 struct printbuf buf = PRINTBUF;
1112 ca = bch_dev_bkey_exists(c, start.inode);
1113 if (!ca->mi.freespace_initialized)
1116 bch2_btree_iter_set_pos(freespace_iter, start);
1118 k = bch2_btree_iter_peek_slot(freespace_iter);
1123 *end = bkey_min(k.k->p, *end);
1125 if (k.k->type != KEY_TYPE_set &&
1126 (c->opts.reconstruct_alloc ||
1127 fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1128 " device %llu buckets %llu-%llu",
1129 freespace_iter->pos.inode,
1130 freespace_iter->pos.offset,
1132 struct bkey_i *update =
1133 bch2_trans_kmalloc(trans, sizeof(*update));
1135 ret = PTR_ERR_OR_ZERO(update);
1139 bkey_init(&update->k);
1140 update->k.type = KEY_TYPE_set;
1141 update->k.p = freespace_iter->pos;
1142 bch2_key_resize(&update->k,
1143 min_t(u64, U32_MAX, end->offset -
1144 freespace_iter->pos.offset));
1146 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1152 printbuf_exit(&buf);
1156 static noinline_for_stack
1157 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1160 struct btree_iter *bucket_gens_iter)
1162 struct bch_fs *c = trans->c;
1164 struct printbuf buf = PRINTBUF;
1165 unsigned i, gens_offset, gens_end_offset;
1168 if (c->sb.version < bcachefs_metadata_version_bucket_gens)
1171 bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1173 k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1178 if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1179 alloc_gens_pos(*end, &gens_end_offset)))
1180 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1182 if (k.k->type == KEY_TYPE_bucket_gens) {
1183 struct bkey_i_bucket_gens g;
1184 bool need_update = false;
1186 bkey_reassemble(&g.k_i, k);
1188 for (i = gens_offset; i < gens_end_offset; i++) {
1189 if (fsck_err_on(g.v.gens[i], c,
1190 "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1191 bucket_gens_pos_to_alloc(k.k->p, i).inode,
1192 bucket_gens_pos_to_alloc(k.k->p, i).offset,
1200 struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1202 ret = PTR_ERR_OR_ZERO(u);
1206 memcpy(u, &g, sizeof(g));
1208 ret = bch2_trans_update(trans, bucket_gens_iter, u, 0);
1214 *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1217 printbuf_exit(&buf);
1221 static noinline_for_stack int __bch2_check_discard_freespace_key(struct btree_trans *trans,
1222 struct btree_iter *iter)
1224 struct bch_fs *c = trans->c;
1225 struct btree_iter alloc_iter;
1226 struct bkey_s_c alloc_k;
1227 struct bch_alloc_v4 a_convert;
1228 const struct bch_alloc_v4 *a;
1231 enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1232 ? BCH_DATA_need_discard
1234 struct printbuf buf = PRINTBUF;
1238 pos.offset &= ~(~0ULL << 56);
1239 genbits = iter->pos.offset & (~0ULL << 56);
1241 alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1242 ret = bkey_err(alloc_k);
1246 if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1247 "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1248 bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1251 a = bch2_alloc_to_v4(alloc_k, &a_convert);
1253 if (fsck_err_on(a->data_type != state ||
1254 (state == BCH_DATA_free &&
1255 genbits != alloc_freespace_genbits(*a)), c,
1256 "%s\n incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
1257 (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1258 bch2_btree_ids[iter->btree_id],
1261 a->data_type == state,
1262 genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1266 set_btree_iter_dontneed(&alloc_iter);
1267 bch2_trans_iter_exit(trans, &alloc_iter);
1268 printbuf_exit(&buf);
1271 ret = bch2_btree_delete_extent_at(trans, iter,
1272 iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
1273 bch2_trans_commit(trans, NULL, NULL,
1274 BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW);
1278 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1279 struct btree_iter *iter,
1282 if (!btree_id_is_extents(iter->btree_id)) {
1283 return __bch2_check_discard_freespace_key(trans, iter);
1287 while (!bkey_eq(iter->pos, end) &&
1288 !(ret = btree_trans_too_many_iters(trans) ?:
1289 __bch2_check_discard_freespace_key(trans, iter)))
1290 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
1297 * We've already checked that generation numbers in the bucket_gens btree are
1298 * valid for buckets that exist; this just checks for keys for nonexistent
1301 static noinline_for_stack
1302 int bch2_check_bucket_gens_key(struct btree_trans *trans,
1303 struct btree_iter *iter,
1306 struct bch_fs *c = trans->c;
1307 struct bkey_i_bucket_gens g;
1309 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1310 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1312 bool need_update = false, dev_exists;
1313 struct printbuf buf = PRINTBUF;
1316 BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1317 bkey_reassemble(&g.k_i, k);
1319 /* if no bch_dev, skip out whether we repair or not */
1320 dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1322 if (fsck_err_on(!dev_exists, c,
1323 "bucket_gens key for invalid device:\n %s",
1324 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1325 ret = bch2_btree_delete_at(trans, iter, 0);
1330 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1331 if (fsck_err_on(end <= ca->mi.first_bucket ||
1332 start >= ca->mi.nbuckets, c,
1333 "bucket_gens key for invalid buckets:\n %s",
1334 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1335 ret = bch2_btree_delete_at(trans, iter, 0);
1339 for (b = start; b < ca->mi.first_bucket; b++)
1340 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1341 "bucket_gens key has nonzero gen for invalid bucket")) {
1342 g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1346 for (b = ca->mi.nbuckets; b < end; b++)
1347 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1348 "bucket_gens key has nonzero gen for invalid bucket")) {
1349 g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1354 struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1356 ret = PTR_ERR_OR_ZERO(u);
1360 memcpy(u, &g, sizeof(g));
1361 ret = bch2_trans_update(trans, iter, u, 0);
1365 printbuf_exit(&buf);
1369 int bch2_check_alloc_info(struct bch_fs *c)
1371 struct btree_trans *trans = bch2_trans_get(c);
1372 struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1377 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS_MIN,
1378 BTREE_ITER_PREFETCH);
1379 bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1380 BTREE_ITER_PREFETCH);
1381 bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1382 BTREE_ITER_PREFETCH);
1383 bch2_trans_iter_init(trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1384 BTREE_ITER_PREFETCH);
1389 bch2_trans_begin(trans);
1391 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1400 next = bpos_nosnap_successor(k.k->p);
1402 ret = bch2_check_alloc_key(trans,
1412 ret = bch2_check_alloc_hole_freespace(trans,
1413 bkey_start_pos(k.k),
1416 bch2_check_alloc_hole_bucket_gens(trans,
1417 bkey_start_pos(k.k),
1424 ret = bch2_trans_commit(trans, NULL, NULL,
1425 BTREE_INSERT_NOFAIL|
1426 BTREE_INSERT_LAZY_RW);
1430 bch2_btree_iter_set_pos(&iter, next);
1432 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1437 bch2_trans_iter_exit(trans, &bucket_gens_iter);
1438 bch2_trans_iter_exit(trans, &freespace_iter);
1439 bch2_trans_iter_exit(trans, &discard_iter);
1440 bch2_trans_iter_exit(trans, &iter);
1445 ret = for_each_btree_key2(trans, iter,
1446 BTREE_ID_need_discard, POS_MIN,
1447 BTREE_ITER_PREFETCH, k,
1448 bch2_check_discard_freespace_key(trans, &iter, k.k->p)) ?:
1449 for_each_btree_key2(trans, iter,
1450 BTREE_ID_freespace, POS_MIN,
1451 BTREE_ITER_PREFETCH, k,
1452 bch2_check_discard_freespace_key(trans, &iter, k.k->p)) ?:
1453 for_each_btree_key_commit(trans, iter,
1454 BTREE_ID_bucket_gens, POS_MIN,
1455 BTREE_ITER_PREFETCH, k,
1456 NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1457 bch2_check_bucket_gens_key(trans, &iter, k));
1459 bch2_trans_put(trans);
1465 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1466 struct btree_iter *alloc_iter)
1468 struct bch_fs *c = trans->c;
1469 struct btree_iter lru_iter;
1470 struct bch_alloc_v4 a_convert;
1471 const struct bch_alloc_v4 *a;
1472 struct bkey_s_c alloc_k, lru_k;
1473 struct printbuf buf = PRINTBUF;
1476 alloc_k = bch2_btree_iter_peek(alloc_iter);
1480 ret = bkey_err(alloc_k);
1484 a = bch2_alloc_to_v4(alloc_k, &a_convert);
1486 if (a->data_type != BCH_DATA_cached)
1489 lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1490 lru_pos(alloc_k.k->p.inode,
1491 bucket_to_u64(alloc_k.k->p),
1492 a->io_time[READ]), 0);
1493 ret = bkey_err(lru_k);
1497 if (fsck_err_on(!a->io_time[READ], c,
1498 "cached bucket with read_time 0\n"
1500 (printbuf_reset(&buf),
1501 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1502 fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1503 "missing lru entry\n"
1505 (printbuf_reset(&buf),
1506 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1507 u64 read_time = a->io_time[READ] ?:
1508 atomic64_read(&c->io_clock[READ].now);
1510 ret = bch2_lru_set(trans,
1512 bucket_to_u64(alloc_k.k->p),
1517 if (a->io_time[READ] != read_time) {
1518 struct bkey_i_alloc_v4 *a_mut =
1519 bch2_alloc_to_v4_mut(trans, alloc_k);
1520 ret = PTR_ERR_OR_ZERO(a_mut);
1524 a_mut->v.io_time[READ] = read_time;
1525 ret = bch2_trans_update(trans, alloc_iter,
1526 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1533 bch2_trans_iter_exit(trans, &lru_iter);
1534 printbuf_exit(&buf);
1538 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1540 struct btree_iter iter;
1544 ret = bch2_trans_run(c,
1545 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1546 POS_MIN, BTREE_ITER_PREFETCH, k,
1547 NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1548 bch2_check_alloc_to_lru_ref(trans, &iter)));
1554 static int bch2_discard_one_bucket(struct btree_trans *trans,
1555 struct btree_iter *need_discard_iter,
1556 struct bpos *discard_pos_done,
1559 u64 *need_journal_commit,
1562 struct bch_fs *c = trans->c;
1563 struct bpos pos = need_discard_iter->pos;
1564 struct btree_iter iter = { NULL };
1567 struct bkey_i_alloc_v4 *a;
1568 struct printbuf buf = PRINTBUF;
1571 ca = bch_dev_bkey_exists(c, pos.inode);
1572 if (!percpu_ref_tryget(&ca->io_ref)) {
1573 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1577 if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1582 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1583 c->journal.flushed_seq_ondisk,
1584 pos.inode, pos.offset)) {
1585 (*need_journal_commit)++;
1589 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1590 need_discard_iter->pos,
1596 a = bch2_alloc_to_v4_mut(trans, k);
1597 ret = PTR_ERR_OR_ZERO(a);
1601 if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1603 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1607 if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1608 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1609 bch2_trans_inconsistent(trans,
1610 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1613 c->journal.flushed_seq_ondisk,
1614 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1620 if (a->v.data_type != BCH_DATA_need_discard) {
1621 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1622 bch2_trans_inconsistent(trans,
1623 "bucket incorrectly set in need_discard btree\n"
1625 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1632 if (!bkey_eq(*discard_pos_done, iter.pos) &&
1633 ca->mi.discard && !c->opts.nochanges) {
1635 * This works without any other locks because this is the only
1636 * thread that removes items from the need_discard tree
1638 bch2_trans_unlock(trans);
1639 blkdev_issue_discard(ca->disk_sb.bdev,
1640 k.k->p.offset * ca->mi.bucket_size,
1643 *discard_pos_done = iter.pos;
1645 ret = bch2_trans_relock_notrace(trans);
1650 SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1651 a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1653 ret = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1654 bch2_trans_commit(trans, NULL, NULL,
1655 BCH_WATERMARK_btree|
1656 BTREE_INSERT_NOFAIL);
1660 this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1664 bch2_trans_iter_exit(trans, &iter);
1665 percpu_ref_put(&ca->io_ref);
1666 printbuf_exit(&buf);
1670 static void bch2_do_discards_work(struct work_struct *work)
1672 struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1673 struct btree_iter iter;
1675 u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1676 struct bpos discard_pos_done = POS_MAX;
1680 * We're doing the commit in bch2_discard_one_bucket instead of using
1681 * for_each_btree_key_commit() so that we can increment counters after
1682 * successful commit:
1684 ret = bch2_trans_run(c,
1685 for_each_btree_key2(trans, iter,
1686 BTREE_ID_need_discard, POS_MIN, 0, k,
1687 bch2_discard_one_bucket(trans, &iter, &discard_pos_done,
1690 &need_journal_commit,
1693 if (need_journal_commit * 2 > seen)
1694 bch2_journal_flush_async(&c->journal, NULL);
1696 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1698 trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1702 void bch2_do_discards(struct bch_fs *c)
1704 if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1705 !queue_work(c->write_ref_wq, &c->discard_work))
1706 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1709 static int invalidate_one_bucket(struct btree_trans *trans,
1710 struct btree_iter *lru_iter,
1711 struct bkey_s_c lru_k,
1712 s64 *nr_to_invalidate)
1714 struct bch_fs *c = trans->c;
1715 struct btree_iter alloc_iter = { NULL };
1716 struct bkey_i_alloc_v4 *a = NULL;
1717 struct printbuf buf = PRINTBUF;
1718 struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1719 unsigned cached_sectors;
1722 if (*nr_to_invalidate <= 0)
1725 if (!bch2_dev_bucket_exists(c, bucket)) {
1726 prt_str(&buf, "lru entry points to invalid bucket");
1730 if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1733 a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1734 ret = PTR_ERR_OR_ZERO(a);
1738 /* We expect harmless races here due to the btree write buffer: */
1739 if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1742 BUG_ON(a->v.data_type != BCH_DATA_cached);
1744 if (!a->v.cached_sectors)
1745 bch_err(c, "invalidating empty bucket, confused");
1747 cached_sectors = a->v.cached_sectors;
1749 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1752 a->v.dirty_sectors = 0;
1753 a->v.cached_sectors = 0;
1754 a->v.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
1755 a->v.io_time[WRITE] = atomic64_read(&c->io_clock[WRITE].now);
1757 ret = bch2_trans_update(trans, &alloc_iter, &a->k_i,
1758 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1759 bch2_trans_commit(trans, NULL, NULL,
1760 BCH_WATERMARK_btree|
1761 BTREE_INSERT_NOFAIL);
1765 trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1766 --*nr_to_invalidate;
1768 bch2_trans_iter_exit(trans, &alloc_iter);
1769 printbuf_exit(&buf);
1772 prt_str(&buf, "\n lru key: ");
1773 bch2_bkey_val_to_text(&buf, c, lru_k);
1775 prt_str(&buf, "\n lru entry: ");
1776 bch2_lru_pos_to_text(&buf, lru_iter->pos);
1778 prt_str(&buf, "\n alloc key: ");
1780 bch2_bpos_to_text(&buf, bucket);
1782 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1784 bch_err(c, "%s", buf.buf);
1785 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) {
1786 bch2_inconsistent_error(c);
1793 static void bch2_do_invalidates_work(struct work_struct *work)
1795 struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1797 struct btree_trans *trans = bch2_trans_get(c);
1798 struct btree_iter iter;
1803 ret = bch2_btree_write_buffer_flush(trans);
1807 for_each_member_device(ca, c, i) {
1808 s64 nr_to_invalidate =
1809 should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1811 ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
1812 lru_pos(ca->dev_idx, 0, 0),
1813 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1814 BTREE_ITER_INTENT, k,
1815 invalidate_one_bucket(trans, &iter, k, &nr_to_invalidate));
1818 percpu_ref_put(&ca->ref);
1823 bch2_trans_put(trans);
1824 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1827 void bch2_do_invalidates(struct bch_fs *c)
1829 if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1830 !queue_work(c->write_ref_wq, &c->invalidate_work))
1831 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1834 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1835 unsigned long *last_updated)
1837 struct btree_trans *trans = bch2_trans_get(c);
1838 struct btree_iter iter;
1841 struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1842 struct bch_member *m;
1845 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1846 POS(ca->dev_idx, ca->mi.first_bucket),
1847 BTREE_ITER_PREFETCH);
1849 * Scan the alloc btree for every bucket on @ca, and add buckets to the
1850 * freespace/need_discard/need_gc_gens btrees as needed:
1853 if (*last_updated + HZ * 10 < jiffies) {
1854 bch_info(ca, "%s: currently at %llu/%llu",
1855 __func__, iter.pos.offset, ca->mi.nbuckets);
1856 *last_updated = jiffies;
1859 bch2_trans_begin(trans);
1861 if (bkey_ge(iter.pos, end)) {
1866 k = bch2_get_key_or_hole(&iter, end, &hole);
1873 * We process live keys in the alloc btree one at a
1876 struct bch_alloc_v4 a_convert;
1877 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1879 ret = bch2_bucket_do_index(trans, k, a, true) ?:
1880 bch2_trans_commit(trans, NULL, NULL,
1881 BTREE_INSERT_LAZY_RW|
1882 BTREE_INSERT_NOFAIL);
1886 bch2_btree_iter_advance(&iter);
1888 struct bkey_i *freespace;
1890 freespace = bch2_trans_kmalloc(trans, sizeof(*freespace));
1891 ret = PTR_ERR_OR_ZERO(freespace);
1895 bkey_init(&freespace->k);
1896 freespace->k.type = KEY_TYPE_set;
1897 freespace->k.p = k.k->p;
1898 freespace->k.size = k.k->size;
1900 ret = bch2_btree_insert_trans(trans, BTREE_ID_freespace, freespace, 0) ?:
1901 bch2_trans_commit(trans, NULL, NULL,
1902 BTREE_INSERT_LAZY_RW|
1903 BTREE_INSERT_NOFAIL);
1907 bch2_btree_iter_set_pos(&iter, k.k->p);
1910 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1916 bch2_trans_iter_exit(trans, &iter);
1917 bch2_trans_put(trans);
1920 bch_err_msg(ca, ret, "initializing free space");
1924 mutex_lock(&c->sb_lock);
1925 m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1926 SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1927 mutex_unlock(&c->sb_lock);
1932 int bch2_fs_freespace_init(struct bch_fs *c)
1937 bool doing_init = false;
1938 unsigned long last_updated = jiffies;
1941 * We can crash during the device add path, so we need to check this on
1945 for_each_member_device(ca, c, i) {
1946 if (ca->mi.freespace_initialized)
1950 bch_info(c, "initializing freespace");
1954 ret = bch2_dev_freespace_init(c, ca, &last_updated);
1956 percpu_ref_put(&ca->ref);
1963 mutex_lock(&c->sb_lock);
1964 bch2_write_super(c);
1965 mutex_unlock(&c->sb_lock);
1966 bch_verbose(c, "done initializing freespace");
1972 /* Bucket IO clocks: */
1974 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1975 size_t bucket_nr, int rw)
1977 struct bch_fs *c = trans->c;
1978 struct btree_iter iter;
1979 struct bkey_i_alloc_v4 *a;
1983 a = bch2_trans_start_alloc_update(trans, &iter, POS(dev, bucket_nr));
1984 ret = PTR_ERR_OR_ZERO(a);
1988 now = atomic64_read(&c->io_clock[rw].now);
1989 if (a->v.io_time[rw] == now)
1992 a->v.io_time[rw] = now;
1994 ret = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1995 bch2_trans_commit(trans, NULL, NULL, 0);
1997 bch2_trans_iter_exit(trans, &iter);
2001 /* Startup/shutdown (ro/rw): */
2003 void bch2_recalc_capacity(struct bch_fs *c)
2006 u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2007 unsigned bucket_size_max = 0;
2008 unsigned long ra_pages = 0;
2011 lockdep_assert_held(&c->state_lock);
2013 for_each_online_member(ca, c, i) {
2014 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2016 ra_pages += bdi->ra_pages;
2019 bch2_set_ra_pages(c, ra_pages);
2021 for_each_rw_member(ca, c, i) {
2022 u64 dev_reserve = 0;
2025 * We need to reserve buckets (from the number
2026 * of currently available buckets) against
2027 * foreground writes so that mainly copygc can
2028 * make forward progress.
2030 * We need enough to refill the various reserves
2031 * from scratch - copygc will use its entire
2032 * reserve all at once, then run against when
2033 * its reserve is refilled (from the formerly
2034 * available buckets).
2036 * This reserve is just used when considering if
2037 * allocations for foreground writes must wait -
2038 * not -ENOSPC calculations.
2041 dev_reserve += ca->nr_btree_reserve * 2;
2042 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2044 dev_reserve += 1; /* btree write point */
2045 dev_reserve += 1; /* copygc write point */
2046 dev_reserve += 1; /* rebalance write point */
2048 dev_reserve *= ca->mi.bucket_size;
2050 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2051 ca->mi.first_bucket);
2053 reserved_sectors += dev_reserve * 2;
2055 bucket_size_max = max_t(unsigned, bucket_size_max,
2056 ca->mi.bucket_size);
2059 gc_reserve = c->opts.gc_reserve_bytes
2060 ? c->opts.gc_reserve_bytes >> 9
2061 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2063 reserved_sectors = max(gc_reserve, reserved_sectors);
2065 reserved_sectors = min(reserved_sectors, capacity);
2067 c->capacity = capacity - reserved_sectors;
2069 c->bucket_size_max = bucket_size_max;
2071 /* Wake up case someone was waiting for buckets */
2072 closure_wake_up(&c->freelist_wait);
2075 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2077 struct open_bucket *ob;
2080 for (ob = c->open_buckets;
2081 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2083 spin_lock(&ob->lock);
2084 if (ob->valid && !ob->on_partial_list &&
2085 ob->dev == ca->dev_idx)
2087 spin_unlock(&ob->lock);
2093 /* device goes ro: */
2094 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2098 /* First, remove device from allocation groups: */
2100 for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2101 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2104 * Capacity is calculated based off of devices in allocation groups:
2106 bch2_recalc_capacity(c);
2108 bch2_open_buckets_stop(c, ca, false);
2111 * Wake up threads that were blocked on allocation, so they can notice
2112 * the device can no longer be removed and the capacity has changed:
2114 closure_wake_up(&c->freelist_wait);
2117 * journal_res_get() can block waiting for free space in the journal -
2118 * it needs to notice there may not be devices to allocate from anymore:
2120 wake_up(&c->journal.wait);
2122 /* Now wait for any in flight writes: */
2124 closure_wait_event(&c->open_buckets_wait,
2125 !bch2_dev_has_open_write_point(c, ca));
2128 /* device goes rw: */
2129 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2133 for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2134 if (ca->mi.data_allowed & (1 << i))
2135 set_bit(ca->dev_idx, c->rw_devs[i].d);
2138 void bch2_fs_allocator_background_init(struct bch_fs *c)
2140 spin_lock_init(&c->freelist_lock);
2141 INIT_WORK(&c->discard_work, bch2_do_discards_work);
2142 INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);