X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Falloc_background.c;h=769c37f9e90710c1db2531d6b318c9e819249d71;hb=b6740e5392a0e0a5fddc5ad4ffac0567078e114a;hp=bf3611e76912e8df9b3efd3fa9ebd541905ebb4f;hpb=99b12cd3a4cdd19985624b79a8c54716cad649bd;p=bcachefs-tools-debian diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c index bf3611e..769c37f 100644 --- a/libbcachefs/alloc_background.c +++ b/libbcachefs/alloc_background.c @@ -2,18 +2,23 @@ #include "bcachefs.h" #include "alloc_background.h" #include "alloc_foreground.h" +#include "backpointers.h" #include "btree_cache.h" #include "btree_io.h" #include "btree_key_cache.h" #include "btree_update.h" #include "btree_update_interior.h" #include "btree_gc.h" +#include "btree_write_buffer.h" #include "buckets.h" +#include "buckets_waiting_for_journal.h" #include "clock.h" #include "debug.h" #include "ec.h" #include "error.h" +#include "lru.h" #include "recovery.h" +#include "trace.h" #include "varint.h" #include @@ -23,14 +28,8 @@ #include #include #include -#include -const char * const bch2_allocator_states[] = { -#define x(n) #n, - ALLOC_THREAD_STATES() -#undef x - NULL -}; +/* Persistent alloc info: */ static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = { #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8, @@ -38,7 +37,17 @@ static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = { #undef x }; -/* Persistent alloc info: */ +struct bkey_alloc_unpacked { + u64 journal_seq; + u8 gen; + u8 oldest_gen; + u8 data_type; + bool need_discard:1; + bool need_inc_gen:1; +#define x(_name, _bits) u##_bits _name; + BCH_ALLOC_FIELDS_V2() +#undef x +}; static inline u64 alloc_field_v1_get(const struct bch_alloc *a, const void **p, unsigned field) @@ -70,36 +79,6 @@ static inline u64 alloc_field_v1_get(const struct bch_alloc *a, return v; } -static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p, - unsigned field, u64 v) -{ - unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field]; - - if (!v) - return; - - a->v.fields |= 1 << field; - - switch (bytes) { - case 1: - *((u8 *) *p) = v; - break; - case 2: - *((__le16 *) *p) = cpu_to_le16(v); - break; - case 4: - *((__le32 *) *p) = cpu_to_le32(v); - break; - case 8: - *((__le64 *) *p) = cpu_to_le64(v); - break; - default: - BUG(); - } - - *p += bytes; -} - static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out, struct bkey_s_c k) { @@ -160,6 +139,8 @@ static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out, out->gen = a.v->gen; out->oldest_gen = a.v->oldest_gen; out->data_type = a.v->data_type; + out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v); + out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v); out->journal_seq = le64_to_cpu(a.v->journal_seq); #define x(_name, _bits) \ @@ -181,53 +162,9 @@ static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out, return 0; } -static void bch2_alloc_pack_v3(struct bkey_alloc_buf *dst, - const struct bkey_alloc_unpacked src) -{ - struct bkey_i_alloc_v3 *a = bkey_alloc_v3_init(&dst->k); - unsigned nr_fields = 0, last_nonzero_fieldnr = 0; - u8 *out = a->v.data; - u8 *end = (void *) &dst[1]; - u8 *last_nonzero_field = out; - unsigned bytes; - - a->k.p = POS(src.dev, src.bucket); - a->v.gen = src.gen; - a->v.oldest_gen = src.oldest_gen; - a->v.data_type = src.data_type; - a->v.journal_seq = cpu_to_le64(src.journal_seq); - -#define x(_name, _bits) \ - nr_fields++; \ - \ - if (src._name) { \ - out += bch2_varint_encode_fast(out, src._name); \ - \ - last_nonzero_field = out; \ - last_nonzero_fieldnr = nr_fields; \ - } else { \ - *out++ = 0; \ - } - - BCH_ALLOC_FIELDS_V2() -#undef x - BUG_ON(out > end); - - out = last_nonzero_field; - a->v.nr_fields = last_nonzero_fieldnr; - - bytes = (u8 *) out - (u8 *) &a->v; - set_bkey_val_bytes(&a->k, bytes); - memset_u64s_tail(&a->v, 0, bytes); -} - -struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k) +static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k) { - struct bkey_alloc_unpacked ret = { - .dev = k.k->p.inode, - .bucket = k.k->p.offset, - .gen = 0, - }; + struct bkey_alloc_unpacked ret = { .gen = 0 }; switch (k.k->type) { case KEY_TYPE_alloc: @@ -244,13 +181,6 @@ struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k) return ret; } -void bch2_alloc_pack(struct bch_fs *c, - struct bkey_alloc_buf *dst, - const struct bkey_alloc_unpacked src) -{ - bch2_alloc_pack_v3(dst, src); -} - static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a) { unsigned i, bytes = offsetof(struct bch_alloc, data); @@ -262,808 +192,1825 @@ static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a) return DIV_ROUND_UP(bytes, sizeof(u64)); } -const char *bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k) +int bch2_alloc_v1_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, + struct printbuf *err) { struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k); - - if (k.k->p.inode >= c->sb.nr_devices || - !c->devs[k.k->p.inode]) - return "invalid device"; + int ret = 0; /* allow for unknown fields */ - if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) - return "incorrect value size"; - - return NULL; + bkey_fsck_err_on(bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v), c, err, + alloc_v1_val_size_bad, + "incorrect value size (%zu < %u)", + bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v)); +fsck_err: + return ret; } -const char *bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k) +int bch2_alloc_v2_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, + struct printbuf *err) { struct bkey_alloc_unpacked u; + int ret = 0; - if (k.k->p.inode >= c->sb.nr_devices || - !c->devs[k.k->p.inode]) - return "invalid device"; - - if (bch2_alloc_unpack_v2(&u, k)) - return "unpack error"; - - return NULL; + bkey_fsck_err_on(bch2_alloc_unpack_v2(&u, k), c, err, + alloc_v2_unpack_error, + "unpack error"); +fsck_err: + return ret; } -const char *bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k) +int bch2_alloc_v3_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, + struct printbuf *err) { struct bkey_alloc_unpacked u; + int ret = 0; - if (k.k->p.inode >= c->sb.nr_devices || - !c->devs[k.k->p.inode]) - return "invalid device"; + bkey_fsck_err_on(bch2_alloc_unpack_v3(&u, k), c, err, + alloc_v2_unpack_error, + "unpack error"); +fsck_err: + return ret; +} - if (bch2_alloc_unpack_v3(&u, k)) - return "unpack error"; +int bch2_alloc_v4_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, struct printbuf *err) +{ + struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k); + int ret = 0; - return NULL; + bkey_fsck_err_on(alloc_v4_u64s(a.v) > bkey_val_u64s(k.k), c, err, + alloc_v4_val_size_bad, + "bad val size (%u > %zu)", + alloc_v4_u64s(a.v), bkey_val_u64s(k.k)); + + bkey_fsck_err_on(!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) && + BCH_ALLOC_V4_NR_BACKPOINTERS(a.v), c, err, + alloc_v4_backpointers_start_bad, + "invalid backpointers_start"); + + bkey_fsck_err_on(alloc_data_type(*a.v, a.v->data_type) != a.v->data_type, c, err, + alloc_key_data_type_bad, + "invalid data type (got %u should be %u)", + a.v->data_type, alloc_data_type(*a.v, a.v->data_type)); + + switch (a.v->data_type) { + case BCH_DATA_free: + case BCH_DATA_need_gc_gens: + case BCH_DATA_need_discard: + bkey_fsck_err_on(bch2_bucket_sectors(*a.v) || a.v->stripe, + c, err, alloc_key_empty_but_have_data, + "empty data type free but have data"); + break; + case BCH_DATA_sb: + case BCH_DATA_journal: + case BCH_DATA_btree: + case BCH_DATA_user: + case BCH_DATA_parity: + bkey_fsck_err_on(!bch2_bucket_sectors_dirty(*a.v), + c, err, alloc_key_dirty_sectors_0, + "data_type %s but dirty_sectors==0", + bch2_data_types[a.v->data_type]); + break; + case BCH_DATA_cached: + bkey_fsck_err_on(!a.v->cached_sectors || + bch2_bucket_sectors_dirty(*a.v) || + a.v->stripe, + c, err, alloc_key_cached_inconsistency, + "data type inconsistency"); + + bkey_fsck_err_on(!a.v->io_time[READ] && + c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_to_lru_refs, + c, err, alloc_key_cached_but_read_time_zero, + "cached bucket with read_time == 0"); + break; + case BCH_DATA_stripe: + break; + } +fsck_err: + return ret; } -void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, - struct bkey_s_c k) +void bch2_alloc_v4_swab(struct bkey_s k) { - struct bkey_alloc_unpacked u = bch2_alloc_unpack(k); - - pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu", - u.gen, u.oldest_gen, bch2_data_types[u.data_type], - u.journal_seq); -#define x(_name, ...) pr_buf(out, " " #_name " %llu", (u64) u._name); - BCH_ALLOC_FIELDS_V2() -#undef x + struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v; + struct bch_backpointer *bp, *bps; + + a->journal_seq = swab64(a->journal_seq); + a->flags = swab32(a->flags); + a->dirty_sectors = swab32(a->dirty_sectors); + a->cached_sectors = swab32(a->cached_sectors); + a->io_time[0] = swab64(a->io_time[0]); + a->io_time[1] = swab64(a->io_time[1]); + a->stripe = swab32(a->stripe); + a->nr_external_backpointers = swab32(a->nr_external_backpointers); + a->fragmentation_lru = swab64(a->fragmentation_lru); + + bps = alloc_v4_backpointers(a); + for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) { + bp->bucket_offset = swab40(bp->bucket_offset); + bp->bucket_len = swab32(bp->bucket_len); + bch2_bpos_swab(&bp->pos); + } } -static int bch2_alloc_read_fn(struct btree_trans *trans, struct bkey_s_c k) +void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k) { - struct bch_fs *c = trans->c; - struct bch_dev *ca; - struct bucket *g; - struct bkey_alloc_unpacked u; + struct bch_alloc_v4 _a; + const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a); + unsigned i; - if (!bkey_is_alloc(k.k)) - return 0; + prt_newline(out); + printbuf_indent_add(out, 2); + + prt_printf(out, "gen %u oldest_gen %u data_type %s", + a->gen, a->oldest_gen, + a->data_type < BCH_DATA_NR + ? bch2_data_types[a->data_type] + : "(invalid data type)"); + prt_newline(out); + prt_printf(out, "journal_seq %llu", a->journal_seq); + prt_newline(out); + prt_printf(out, "need_discard %llu", BCH_ALLOC_V4_NEED_DISCARD(a)); + prt_newline(out); + prt_printf(out, "need_inc_gen %llu", BCH_ALLOC_V4_NEED_INC_GEN(a)); + prt_newline(out); + prt_printf(out, "dirty_sectors %u", a->dirty_sectors); + prt_newline(out); + prt_printf(out, "cached_sectors %u", a->cached_sectors); + prt_newline(out); + prt_printf(out, "stripe %u", a->stripe); + prt_newline(out); + prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy); + prt_newline(out); + prt_printf(out, "io_time[READ] %llu", a->io_time[READ]); + prt_newline(out); + prt_printf(out, "io_time[WRITE] %llu", a->io_time[WRITE]); + prt_newline(out); + prt_printf(out, "fragmentation %llu", a->fragmentation_lru); + prt_newline(out); + prt_printf(out, "bp_start %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a)); + prt_newline(out); + + if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) { + struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k); + const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v); + + prt_printf(out, "backpointers: %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v)); + printbuf_indent_add(out, 2); + + for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) { + prt_newline(out); + bch2_backpointer_to_text(out, &bps[i]); + } - ca = bch_dev_bkey_exists(c, k.k->p.inode); - g = bucket(ca, k.k->p.offset); - u = bch2_alloc_unpack(k); - - g->_mark.gen = u.gen; - g->_mark.data_type = u.data_type; - g->_mark.dirty_sectors = u.dirty_sectors; - g->_mark.cached_sectors = u.cached_sectors; - g->_mark.stripe = u.stripe != 0; - g->stripe = u.stripe; - g->stripe_redundancy = u.stripe_redundancy; - g->io_time[READ] = u.read_time; - g->io_time[WRITE] = u.write_time; - g->oldest_gen = u.oldest_gen; - g->gen_valid = 1; + printbuf_indent_sub(out, 2); + } - return 0; + printbuf_indent_sub(out, 2); } -int bch2_alloc_read(struct bch_fs *c) +void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out) { - struct btree_trans trans; - int ret; + if (k.k->type == KEY_TYPE_alloc_v4) { + void *src, *dst; - bch2_trans_init(&trans, c, 0, 0); - down_read(&c->gc_lock); - ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_alloc, bch2_alloc_read_fn); - up_read(&c->gc_lock); - bch2_trans_exit(&trans); - if (ret) { - bch_err(c, "error reading alloc info: %i", ret); - return ret; - } + *out = *bkey_s_c_to_alloc_v4(k).v; - return 0; + src = alloc_v4_backpointers(out); + SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s); + dst = alloc_v4_backpointers(out); + + if (src < dst) + memset(src, 0, dst - src); + + SET_BCH_ALLOC_V4_NR_BACKPOINTERS(out, 0); + } else { + struct bkey_alloc_unpacked u = bch2_alloc_unpack(k); + + *out = (struct bch_alloc_v4) { + .journal_seq = u.journal_seq, + .flags = u.need_discard, + .gen = u.gen, + .oldest_gen = u.oldest_gen, + .data_type = u.data_type, + .stripe_redundancy = u.stripe_redundancy, + .dirty_sectors = u.dirty_sectors, + .cached_sectors = u.cached_sectors, + .io_time[READ] = u.read_time, + .io_time[WRITE] = u.write_time, + .stripe = u.stripe, + }; + + SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s); + } } -static int bch2_alloc_write_key(struct btree_trans *trans, - struct btree_iter *iter, - unsigned flags) +static noinline struct bkey_i_alloc_v4 * +__bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k) { - struct bch_fs *c = trans->c; - struct bkey_s_c k; - struct bch_dev *ca; - struct bucket *g; - struct bucket_mark m; - struct bkey_alloc_unpacked old_u, new_u; - struct bkey_alloc_buf a; - int ret; -retry: - bch2_trans_begin(trans); + struct bkey_i_alloc_v4 *ret; - ret = bch2_btree_key_cache_flush(trans, - BTREE_ID_alloc, iter->pos); - if (ret) - goto err; + ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4))); + if (IS_ERR(ret)) + return ret; - k = bch2_btree_iter_peek_slot(iter); - ret = bkey_err(k); - if (ret) - goto err; + if (k.k->type == KEY_TYPE_alloc_v4) { + void *src, *dst; - old_u = bch2_alloc_unpack(k); + bkey_reassemble(&ret->k_i, k); - percpu_down_read(&c->mark_lock); - ca = bch_dev_bkey_exists(c, iter->pos.inode); - g = bucket(ca, iter->pos.offset); - m = READ_ONCE(g->mark); - new_u = alloc_mem_to_key(iter, g, m); - percpu_up_read(&c->mark_lock); + src = alloc_v4_backpointers(&ret->v); + SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s); + dst = alloc_v4_backpointers(&ret->v); - if (!bkey_alloc_unpacked_cmp(old_u, new_u)) - return 0; + if (src < dst) + memset(src, 0, dst - src); - bch2_alloc_pack(c, &a, new_u); - ret = bch2_trans_update(trans, iter, &a.k, - BTREE_TRIGGER_NORUN) ?: - bch2_trans_commit(trans, NULL, NULL, - BTREE_INSERT_NOFAIL|flags); -err: - if (ret == -EINTR) - goto retry; + SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v, 0); + set_alloc_v4_u64s(ret); + } else { + bkey_alloc_v4_init(&ret->k_i); + ret->k.p = k.k->p; + bch2_alloc_to_v4(k, &ret->v); + } return ret; } -int bch2_alloc_write(struct bch_fs *c, unsigned flags) +static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k) { - struct btree_trans trans; - struct btree_iter iter; - struct bch_dev *ca; - unsigned i; - int ret = 0; - - bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0); - bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN, - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); + struct bkey_s_c_alloc_v4 a; - for_each_member_device(ca, c, i) { - bch2_btree_iter_set_pos(&iter, - POS(ca->dev_idx, ca->mi.first_bucket)); + if (likely(k.k->type == KEY_TYPE_alloc_v4) && + ((a = bkey_s_c_to_alloc_v4(k), true) && + BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) == 0)) + return bch2_bkey_make_mut_noupdate_typed(trans, k, alloc_v4); - while (iter.pos.offset < ca->mi.nbuckets) { - ret = bch2_alloc_write_key(&trans, &iter, flags); - if (ret) { - percpu_ref_put(&ca->ref); - goto err; - } - bch2_btree_iter_advance(&iter); - } - } -err: - bch2_trans_iter_exit(&trans, &iter); - bch2_trans_exit(&trans); - return ret; + return __bch2_alloc_to_v4_mut(trans, k); } -/* Bucket IO clocks: */ +struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k) +{ + return bch2_alloc_to_v4_mut_inlined(trans, k); +} -int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev, - size_t bucket_nr, int rw) +struct bkey_i_alloc_v4 * +bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter, + struct bpos pos) { - struct bch_fs *c = trans->c; - struct bch_dev *ca = bch_dev_bkey_exists(c, dev); - struct btree_iter iter; - struct bucket *g; - struct bkey_alloc_buf *a; - struct bkey_alloc_unpacked u; - u64 *time, now; - int ret = 0; + struct bkey_s_c k; + struct bkey_i_alloc_v4 *a; + int ret; - bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(dev, bucket_nr), + k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos, + BTREE_ITER_WITH_UPDATES| BTREE_ITER_CACHED| - BTREE_ITER_CACHED_NOFILL| BTREE_ITER_INTENT); - ret = bch2_btree_iter_traverse(&iter); - if (ret) - goto out; + ret = bkey_err(k); + if (unlikely(ret)) + return ERR_PTR(ret); - a = bch2_trans_kmalloc(trans, sizeof(struct bkey_alloc_buf)); + a = bch2_alloc_to_v4_mut_inlined(trans, k); ret = PTR_ERR_OR_ZERO(a); - if (ret) - goto out; - - percpu_down_read(&c->mark_lock); - g = bucket(ca, bucket_nr); - u = alloc_mem_to_key(&iter, g, READ_ONCE(g->mark)); - percpu_up_read(&c->mark_lock); - - time = rw == READ ? &u.read_time : &u.write_time; - now = atomic64_read(&c->io_clock[rw].now); - if (*time == now) - goto out; - - *time = now; - - bch2_alloc_pack(c, a, u); - ret = bch2_trans_update(trans, &iter, &a->k, 0) ?: - bch2_trans_commit(trans, NULL, NULL, 0); -out: - bch2_trans_iter_exit(trans, &iter); - return ret; + if (unlikely(ret)) + goto err; + return a; +err: + bch2_trans_iter_exit(trans, iter); + return ERR_PTR(ret); } -/* Background allocator thread: */ - -/* - * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens - * (marking them as invalidated on disk), then optionally issues discard - * commands to the newly free buckets, then puts them on the various freelists. - */ - -static bool bch2_can_invalidate_bucket(struct bch_dev *ca, size_t b, - struct bucket_mark m) +static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset) { - u8 gc_gen; - - if (!is_available_bucket(m)) - return false; - - if (m.owned_by_allocator) - return false; + *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK; - if (ca->buckets_nouse && - test_bit(b, ca->buckets_nouse)) - return false; - - gc_gen = bucket_gc_gen(bucket(ca, b)); - - ca->inc_gen_needs_gc += gc_gen >= BUCKET_GC_GEN_MAX / 2; - ca->inc_gen_really_needs_gc += gc_gen >= BUCKET_GC_GEN_MAX; - - return gc_gen < BUCKET_GC_GEN_MAX; + pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS; + return pos; } -/* - * Determines what order we're going to reuse buckets, smallest bucket_key() - * first. - */ - -static unsigned bucket_sort_key(struct bucket *g, struct bucket_mark m, - u64 now, u64 last_seq_ondisk) +static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset) { - unsigned used = bucket_sectors_used(m); - - if (used) { - /* - * Prefer to keep buckets that have been read more recently, and - * buckets that have more data in them: - */ - u64 last_read = max_t(s64, 0, now - g->io_time[READ]); - u32 last_read_scaled = max_t(u64, U32_MAX, div_u64(last_read, used)); - - return -last_read_scaled; - } else { - /* - * Prefer to use buckets with smaller gc_gen so that we don't - * have to walk the btree and recalculate oldest_gen - but shift - * off the low bits so that buckets will still have equal sort - * keys when there's only a small difference, so that we can - * keep sequential buckets together: - */ - return (bucket_needs_journal_commit(m, last_seq_ondisk) << 4)| - (bucket_gc_gen(g) >> 4); - } + pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS; + pos.offset += offset; + return pos; } -static inline int bucket_alloc_cmp(alloc_heap *h, - struct alloc_heap_entry l, - struct alloc_heap_entry r) +static unsigned alloc_gen(struct bkey_s_c k, unsigned offset) { - return cmp_int(l.key, r.key) ?: - cmp_int(r.nr, l.nr) ?: - cmp_int(l.bucket, r.bucket); + return k.k->type == KEY_TYPE_bucket_gens + ? bkey_s_c_to_bucket_gens(k).v->gens[offset] + : 0; } -static inline int bucket_idx_cmp(const void *_l, const void *_r) +int bch2_bucket_gens_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, + struct printbuf *err) { - const struct alloc_heap_entry *l = _l, *r = _r; + int ret = 0; - return cmp_int(l->bucket, r->bucket); + bkey_fsck_err_on(bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens), c, err, + bucket_gens_val_size_bad, + "bad val size (%zu != %zu)", + bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens)); +fsck_err: + return ret; } -static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca) +void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k) { - struct bucket_array *buckets; - struct alloc_heap_entry e = { 0 }; - u64 now, last_seq_ondisk; - size_t b, i, nr = 0; + struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k); + unsigned i; - down_read(&ca->bucket_lock); + for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) { + if (i) + prt_char(out, ' '); + prt_printf(out, "%u", g.v->gens[i]); + } +} - buckets = bucket_array(ca); - ca->alloc_heap.used = 0; - now = atomic64_read(&c->io_clock[READ].now); - last_seq_ondisk = c->journal.last_seq_ondisk; +int bch2_bucket_gens_init(struct bch_fs *c) +{ + struct btree_trans *trans = bch2_trans_get(c); + struct btree_iter iter; + struct bkey_s_c k; + struct bch_alloc_v4 a; + struct bkey_i_bucket_gens g; + bool have_bucket_gens_key = false; + unsigned offset; + struct bpos pos; + u8 gen; + int ret; - /* - * Find buckets with lowest read priority, by building a maxheap sorted - * by read priority and repeatedly replacing the maximum element until - * all buckets have been visited. - */ - for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) { - struct bucket *g = &buckets->b[b]; - struct bucket_mark m = READ_ONCE(g->mark); - unsigned key = bucket_sort_key(g, m, now, last_seq_ondisk); + ret = for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN, + BTREE_ITER_PREFETCH, k, ({ + /* + * Not a fsck error because this is checked/repaired by + * bch2_check_alloc_key() which runs later: + */ + if (!bch2_dev_bucket_exists(c, k.k->p)) + continue; - cond_resched(); + gen = bch2_alloc_to_v4(k, &a)->gen; + pos = alloc_gens_pos(iter.pos, &offset); - if (!bch2_can_invalidate_bucket(ca, b, m)) - continue; + if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) { + ret = commit_do(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc, + bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0)); + if (ret) + break; + have_bucket_gens_key = false; + } - if (e.nr && e.bucket + e.nr == b && e.key == key) { - e.nr++; - } else { - if (e.nr) - heap_add_or_replace(&ca->alloc_heap, e, - -bucket_alloc_cmp, NULL); - - e = (struct alloc_heap_entry) { - .bucket = b, - .nr = 1, - .key = key, - }; + if (!have_bucket_gens_key) { + bkey_bucket_gens_init(&g.k_i); + g.k.p = pos; + have_bucket_gens_key = true; } - } - if (e.nr) - heap_add_or_replace(&ca->alloc_heap, e, - -bucket_alloc_cmp, NULL); + g.v.gens[offset] = gen; + 0; + })); - for (i = 0; i < ca->alloc_heap.used; i++) - nr += ca->alloc_heap.data[i].nr; + if (have_bucket_gens_key && !ret) + ret = commit_do(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc, + bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0)); - while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) { - nr -= ca->alloc_heap.data[0].nr; - heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL); - } + bch2_trans_put(trans); - up_read(&ca->bucket_lock); + bch_err_fn(c, ret); + return ret; } -static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca) +int bch2_alloc_read(struct bch_fs *c) { - struct bucket_array *buckets = bucket_array(ca); - struct bucket_mark m; - size_t b, start; - - if (ca->fifo_last_bucket < ca->mi.first_bucket || - ca->fifo_last_bucket >= ca->mi.nbuckets) - ca->fifo_last_bucket = ca->mi.first_bucket; - - start = ca->fifo_last_bucket; - - do { - ca->fifo_last_bucket++; - if (ca->fifo_last_bucket == ca->mi.nbuckets) - ca->fifo_last_bucket = ca->mi.first_bucket; - - b = ca->fifo_last_bucket; - m = READ_ONCE(buckets->b[b].mark); - - if (bch2_can_invalidate_bucket(ca, b, m)) { - struct alloc_heap_entry e = { .bucket = b, .nr = 1, }; + struct btree_trans *trans = bch2_trans_get(c); + struct btree_iter iter; + struct bkey_s_c k; + struct bch_dev *ca; + int ret; - heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); - if (heap_full(&ca->alloc_heap)) - break; - } + down_read(&c->gc_lock); - cond_resched(); - } while (ca->fifo_last_bucket != start); -} + if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_bucket_gens) { + const struct bch_bucket_gens *g; + u64 b; -static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca) -{ - struct bucket_array *buckets = bucket_array(ca); - struct bucket_mark m; - size_t checked, i; + ret = for_each_btree_key(trans, iter, BTREE_ID_bucket_gens, POS_MIN, + BTREE_ITER_PREFETCH, k, ({ + u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset; + u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset; - for (checked = 0; - checked < ca->mi.nbuckets / 2; - checked++) { - size_t b = bch2_rand_range(ca->mi.nbuckets - - ca->mi.first_bucket) + - ca->mi.first_bucket; + if (k.k->type != KEY_TYPE_bucket_gens) + continue; - m = READ_ONCE(buckets->b[b].mark); + g = bkey_s_c_to_bucket_gens(k).v; - if (bch2_can_invalidate_bucket(ca, b, m)) { - struct alloc_heap_entry e = { .bucket = b, .nr = 1, }; + /* + * Not a fsck error because this is checked/repaired by + * bch2_check_alloc_key() which runs later: + */ + if (!bch2_dev_exists2(c, k.k->p.inode)) + continue; - heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); - if (heap_full(&ca->alloc_heap)) - break; - } + ca = bch_dev_bkey_exists(c, k.k->p.inode); - cond_resched(); + for (b = max_t(u64, ca->mi.first_bucket, start); + b < min_t(u64, ca->mi.nbuckets, end); + b++) + *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK]; + 0; + })); + } else { + struct bch_alloc_v4 a; + + ret = for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN, + BTREE_ITER_PREFETCH, k, ({ + /* + * Not a fsck error because this is checked/repaired by + * bch2_check_alloc_key() which runs later: + */ + if (!bch2_dev_bucket_exists(c, k.k->p)) + continue; + + ca = bch_dev_bkey_exists(c, k.k->p.inode); + + *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen; + 0; + })); } - sort(ca->alloc_heap.data, - ca->alloc_heap.used, - sizeof(ca->alloc_heap.data[0]), - bucket_idx_cmp, NULL); + bch2_trans_put(trans); + up_read(&c->gc_lock); - /* remove duplicates: */ - for (i = 0; i + 1 < ca->alloc_heap.used; i++) - if (ca->alloc_heap.data[i].bucket == - ca->alloc_heap.data[i + 1].bucket) - ca->alloc_heap.data[i].nr = 0; + bch_err_fn(c, ret); + return ret; } -static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca) +/* Free space/discard btree: */ + +static int bch2_bucket_do_index(struct btree_trans *trans, + struct bkey_s_c alloc_k, + const struct bch_alloc_v4 *a, + bool set) { - size_t i, nr = 0; + struct bch_fs *c = trans->c; + struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode); + struct btree_iter iter; + struct bkey_s_c old; + struct bkey_i *k; + enum btree_id btree; + enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted; + enum bch_bkey_type new_type = set ? KEY_TYPE_set : KEY_TYPE_deleted; + struct printbuf buf = PRINTBUF; + int ret; - ca->inc_gen_needs_gc = 0; - ca->inc_gen_really_needs_gc = 0; + if (a->data_type != BCH_DATA_free && + a->data_type != BCH_DATA_need_discard) + return 0; - switch (ca->mi.replacement) { - case BCH_CACHE_REPLACEMENT_lru: - find_reclaimable_buckets_lru(c, ca); - break; - case BCH_CACHE_REPLACEMENT_fifo: - find_reclaimable_buckets_fifo(c, ca); + k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k)); + if (IS_ERR(k)) + return PTR_ERR(k); + + bkey_init(&k->k); + k->k.type = new_type; + + switch (a->data_type) { + case BCH_DATA_free: + btree = BTREE_ID_freespace; + k->k.p = alloc_freespace_pos(alloc_k.k->p, *a); + bch2_key_resize(&k->k, 1); break; - case BCH_CACHE_REPLACEMENT_random: - find_reclaimable_buckets_random(c, ca); + case BCH_DATA_need_discard: + btree = BTREE_ID_need_discard; + k->k.p = alloc_k.k->p; break; + default: + return 0; } - heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL); - - for (i = 0; i < ca->alloc_heap.used; i++) - nr += ca->alloc_heap.data[i].nr; - - return nr; -} - -/* - * returns sequence number of most recent journal entry that updated this - * bucket: - */ -static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m) -{ - if (m.journal_seq_valid) { - u64 journal_seq = atomic64_read(&c->journal.seq); - u64 bucket_seq = journal_seq; - - bucket_seq &= ~((u64) U16_MAX); - bucket_seq |= m.journal_seq; - - if (bucket_seq > journal_seq) - bucket_seq -= 1 << 16; + old = bch2_bkey_get_iter(trans, &iter, btree, + bkey_start_pos(&k->k), + BTREE_ITER_INTENT); + ret = bkey_err(old); + if (ret) + return ret; - return bucket_seq; - } else { - return 0; + if (ca->mi.freespace_initialized && + c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info && + bch2_trans_inconsistent_on(old.k->type != old_type, trans, + "incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n" + " for %s", + set ? "setting" : "clearing", + bch2_btree_id_str(btree), + iter.pos.inode, + iter.pos.offset, + bch2_bkey_types[old.k->type], + bch2_bkey_types[old_type], + (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) { + ret = -EIO; + goto err; } + + ret = bch2_trans_update(trans, &iter, k, 0); +err: + bch2_trans_iter_exit(trans, &iter); + printbuf_exit(&buf); + return ret; } -static int bucket_invalidate_btree(struct btree_trans *trans, - struct bch_dev *ca, u64 b) +static noinline int bch2_bucket_gen_update(struct btree_trans *trans, + struct bpos bucket, u8 gen) { - struct bch_fs *c = trans->c; - struct bkey_alloc_buf *a; - struct bkey_alloc_unpacked u; - struct bucket *g; - struct bucket_mark m; struct btree_iter iter; + unsigned offset; + struct bpos pos = alloc_gens_pos(bucket, &offset); + struct bkey_i_bucket_gens *g; + struct bkey_s_c k; int ret; - bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, - POS(ca->dev_idx, b), - BTREE_ITER_CACHED| - BTREE_ITER_CACHED_NOFILL| - BTREE_ITER_INTENT); - - a = bch2_trans_kmalloc(trans, sizeof(*a)); - ret = PTR_ERR_OR_ZERO(a); + g = bch2_trans_kmalloc(trans, sizeof(*g)); + ret = PTR_ERR_OR_ZERO(g); if (ret) - goto err; + return ret; - ret = bch2_btree_iter_traverse(&iter); + k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos, + BTREE_ITER_INTENT| + BTREE_ITER_WITH_UPDATES); + ret = bkey_err(k); if (ret) - goto err; + return ret; - percpu_down_read(&c->mark_lock); - g = bucket(ca, b); - m = READ_ONCE(g->mark); - u = alloc_mem_to_key(&iter, g, m); - percpu_up_read(&c->mark_lock); - - u.gen++; - u.data_type = 0; - u.dirty_sectors = 0; - u.cached_sectors = 0; - u.read_time = atomic64_read(&c->io_clock[READ].now); - u.write_time = atomic64_read(&c->io_clock[WRITE].now); - - bch2_alloc_pack(c, a, u); - ret = bch2_trans_update(trans, &iter, &a->k, - BTREE_TRIGGER_BUCKET_INVALIDATE); -err: + if (k.k->type != KEY_TYPE_bucket_gens) { + bkey_bucket_gens_init(&g->k_i); + g->k.p = iter.pos; + } else { + bkey_reassemble(&g->k_i, k); + } + + g->v.gens[offset] = gen; + + ret = bch2_trans_update(trans, &iter, &g->k_i, 0); bch2_trans_iter_exit(trans, &iter); return ret; } -static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca, - u64 *journal_seq, unsigned flags) +int bch2_trans_mark_alloc(struct btree_trans *trans, + enum btree_id btree_id, unsigned level, + struct bkey_s_c old, struct bkey_i *new, + unsigned flags) { - struct bucket *g; - struct bucket_mark m; - size_t b; + struct bch_fs *c = trans->c; + struct bch_alloc_v4 old_a_convert, *new_a; + const struct bch_alloc_v4 *old_a; + u64 old_lru, new_lru; int ret = 0; - BUG_ON(!ca->alloc_heap.used || - !ca->alloc_heap.data[0].nr); - b = ca->alloc_heap.data[0].bucket; + /* + * Deletion only happens in the device removal path, with + * BTREE_TRIGGER_NORUN: + */ + BUG_ON(new->k.type != KEY_TYPE_alloc_v4); - /* first, put on free_inc and mark as owned by allocator: */ - percpu_down_read(&c->mark_lock); - g = bucket(ca, b); - m = READ_ONCE(g->mark); + old_a = bch2_alloc_to_v4(old, &old_a_convert); + new_a = &bkey_i_to_alloc_v4(new)->v; - BUG_ON(m.dirty_sectors); + new_a->data_type = alloc_data_type(*new_a, new_a->data_type); - bch2_mark_alloc_bucket(c, ca, b, true); + if (bch2_bucket_sectors(*new_a) > bch2_bucket_sectors(*old_a)) { + new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now)); + new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now)); + SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true); + SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true); + } - spin_lock(&c->freelist_lock); - verify_not_on_freelist(c, ca, b); - BUG_ON(!fifo_push(&ca->free_inc, b)); - spin_unlock(&c->freelist_lock); + if (data_type_is_empty(new_a->data_type) && + BCH_ALLOC_V4_NEED_INC_GEN(new_a) && + !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) { + new_a->gen++; + SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false); + } - /* - * If we're not invalidating cached data, we only increment the bucket - * gen in memory here, the incremented gen will be updated in the btree - * by bch2_trans_mark_pointer(): - */ - if (!m.cached_sectors && - !bucket_needs_journal_commit(m, c->journal.last_seq_ondisk)) { - BUG_ON(m.data_type); - bucket_cmpxchg(g, m, m.gen++); - percpu_up_read(&c->mark_lock); - goto out; + if (old_a->data_type != new_a->data_type || + (new_a->data_type == BCH_DATA_free && + alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) { + ret = bch2_bucket_do_index(trans, old, old_a, false) ?: + bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true); + if (ret) + return ret; + } + + if (new_a->data_type == BCH_DATA_cached && + !new_a->io_time[READ]) + new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now)); + + old_lru = alloc_lru_idx_read(*old_a); + new_lru = alloc_lru_idx_read(*new_a); + + if (old_lru != new_lru) { + ret = bch2_lru_change(trans, new->k.p.inode, + bucket_to_u64(new->k.p), + old_lru, new_lru); + if (ret) + return ret; + } + + new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a, + bch_dev_bkey_exists(c, new->k.p.inode)); + + if (old_a->fragmentation_lru != new_a->fragmentation_lru) { + ret = bch2_lru_change(trans, + BCH_LRU_FRAGMENTATION_START, + bucket_to_u64(new->k.p), + old_a->fragmentation_lru, new_a->fragmentation_lru); + if (ret) + return ret; } - percpu_up_read(&c->mark_lock); + if (old_a->gen != new_a->gen) { + ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen); + if (ret) + return ret; + } /* - * If the read-only path is trying to shut down, we can't be generating - * new btree updates: + * need to know if we're getting called from the invalidate path or + * not: */ - if (test_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags)) { - ret = 1; - goto out; + + if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) && + old_a->cached_sectors) { + ret = bch2_update_cached_sectors_list(trans, new->k.p.inode, + -((s64) old_a->cached_sectors)); + if (ret) + return ret; } - ret = bch2_trans_do(c, NULL, journal_seq, - BTREE_INSERT_NOCHECK_RW| - BTREE_INSERT_NOFAIL| - BTREE_INSERT_JOURNAL_RESERVED| - flags, - bucket_invalidate_btree(&trans, ca, b)); -out: - if (!ret) { - /* remove from alloc_heap: */ - struct alloc_heap_entry e, *top = ca->alloc_heap.data; + return 0; +} - top->bucket++; - top->nr--; +/* + * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for + * extents style btrees, but works on non-extents btrees: + */ +static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole) +{ + struct bkey_s_c k = bch2_btree_iter_peek_slot(iter); - if (!top->nr) - heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); + if (bkey_err(k)) + return k; + + if (k.k->type) { + return k; + } else { + struct btree_iter iter2; + struct bpos next; + + bch2_trans_copy_iter(&iter2, iter); + + struct btree_path *path = btree_iter_path(iter->trans, iter); + if (!bpos_eq(path->l[0].b->key.k.p, SPOS_MAX)) + end = bkey_min(end, bpos_nosnap_successor(path->l[0].b->key.k.p)); + + end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1)); /* - * Make sure we flush the last journal entry that updated this - * bucket (i.e. deleting the last reference) before writing to - * this bucket again: + * btree node min/max is a closed interval, upto takes a half + * open interval: */ - *journal_seq = max(*journal_seq, bucket_journal_seq(c, m)); - } else { - size_t b2; + k = bch2_btree_iter_peek_upto(&iter2, end); + next = iter2.pos; + bch2_trans_iter_exit(iter->trans, &iter2); + + BUG_ON(next.offset >= iter->pos.offset + U32_MAX); + + if (bkey_err(k)) + return k; + + bkey_init(hole); + hole->p = iter->pos; + + bch2_key_resize(hole, next.offset - iter->pos.offset); + return (struct bkey_s_c) { hole, NULL }; + } +} + +static bool next_bucket(struct bch_fs *c, struct bpos *bucket) +{ + struct bch_dev *ca; + unsigned iter; + + if (bch2_dev_bucket_exists(c, *bucket)) + return true; + + if (bch2_dev_exists2(c, bucket->inode)) { + ca = bch_dev_bkey_exists(c, bucket->inode); + + if (bucket->offset < ca->mi.first_bucket) { + bucket->offset = ca->mi.first_bucket; + return true; + } + + bucket->inode++; + bucket->offset = 0; + } + + rcu_read_lock(); + iter = bucket->inode; + ca = __bch2_next_dev(c, &iter, NULL); + if (ca) + *bucket = POS(ca->dev_idx, ca->mi.first_bucket); + rcu_read_unlock(); + + return ca != NULL; +} + +static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole) +{ + struct bch_fs *c = iter->trans->c; + struct bkey_s_c k; +again: + k = bch2_get_key_or_hole(iter, POS_MAX, hole); + if (bkey_err(k)) + return k; + + if (!k.k->type) { + struct bpos bucket = bkey_start_pos(k.k); + + if (!bch2_dev_bucket_exists(c, bucket)) { + if (!next_bucket(c, &bucket)) + return bkey_s_c_null; + + bch2_btree_iter_set_pos(iter, bucket); + goto again; + } + + if (!bch2_dev_bucket_exists(c, k.k->p)) { + struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode); + + bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset); + } + } + + return k; +} - /* remove from free_inc: */ - percpu_down_read(&c->mark_lock); - spin_lock(&c->freelist_lock); +static noinline_for_stack +int bch2_check_alloc_key(struct btree_trans *trans, + struct bkey_s_c alloc_k, + struct btree_iter *alloc_iter, + struct btree_iter *discard_iter, + struct btree_iter *freespace_iter, + struct btree_iter *bucket_gens_iter) +{ + struct bch_fs *c = trans->c; + struct bch_dev *ca; + struct bch_alloc_v4 a_convert; + const struct bch_alloc_v4 *a; + unsigned discard_key_type, freespace_key_type; + unsigned gens_offset; + struct bkey_s_c k; + struct printbuf buf = PRINTBUF; + int ret; + + if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c, + alloc_key_to_missing_dev_bucket, + "alloc key for invalid device:bucket %llu:%llu", + alloc_k.k->p.inode, alloc_k.k->p.offset)) + return bch2_btree_delete_at(trans, alloc_iter, 0); + + ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode); + if (!ca->mi.freespace_initialized) + return 0; - bch2_mark_alloc_bucket(c, ca, b, false); + a = bch2_alloc_to_v4(alloc_k, &a_convert); - BUG_ON(!fifo_pop_back(&ca->free_inc, b2)); - BUG_ON(b != b2); + discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0; + bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p); + k = bch2_btree_iter_peek_slot(discard_iter); + ret = bkey_err(k); + if (ret) + goto err; + + if (k.k->type != discard_key_type && + (c->opts.reconstruct_alloc || + fsck_err(c, need_discard_key_wrong, + "incorrect key in need_discard btree (got %s should be %s)\n" + " %s", + bch2_bkey_types[k.k->type], + bch2_bkey_types[discard_key_type], + (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) { + struct bkey_i *update = + bch2_trans_kmalloc(trans, sizeof(*update)); + + ret = PTR_ERR_OR_ZERO(update); + if (ret) + goto err; + + bkey_init(&update->k); + update->k.type = discard_key_type; + update->k.p = discard_iter->pos; - spin_unlock(&c->freelist_lock); - percpu_up_read(&c->mark_lock); + ret = bch2_trans_update(trans, discard_iter, update, 0); + if (ret) + goto err; } - return ret < 0 ? ret : 0; + freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0; + bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a)); + k = bch2_btree_iter_peek_slot(freespace_iter); + ret = bkey_err(k); + if (ret) + goto err; + + if (k.k->type != freespace_key_type && + (c->opts.reconstruct_alloc || + fsck_err(c, freespace_key_wrong, + "incorrect key in freespace btree (got %s should be %s)\n" + " %s", + bch2_bkey_types[k.k->type], + bch2_bkey_types[freespace_key_type], + (printbuf_reset(&buf), + bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) { + struct bkey_i *update = + bch2_trans_kmalloc(trans, sizeof(*update)); + + ret = PTR_ERR_OR_ZERO(update); + if (ret) + goto err; + + bkey_init(&update->k); + update->k.type = freespace_key_type; + update->k.p = freespace_iter->pos; + bch2_key_resize(&update->k, 1); + + ret = bch2_trans_update(trans, freespace_iter, update, 0); + if (ret) + goto err; + } + + bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset)); + k = bch2_btree_iter_peek_slot(bucket_gens_iter); + ret = bkey_err(k); + if (ret) + goto err; + + if (a->gen != alloc_gen(k, gens_offset) && + (c->opts.reconstruct_alloc || + fsck_err(c, bucket_gens_key_wrong, + "incorrect gen in bucket_gens btree (got %u should be %u)\n" + " %s", + alloc_gen(k, gens_offset), a->gen, + (printbuf_reset(&buf), + bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) { + struct bkey_i_bucket_gens *g = + bch2_trans_kmalloc(trans, sizeof(*g)); + + ret = PTR_ERR_OR_ZERO(g); + if (ret) + goto err; + + if (k.k->type == KEY_TYPE_bucket_gens) { + bkey_reassemble(&g->k_i, k); + } else { + bkey_bucket_gens_init(&g->k_i); + g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset); + } + + g->v.gens[gens_offset] = a->gen; + + ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0); + if (ret) + goto err; + } +err: +fsck_err: + printbuf_exit(&buf); + return ret; +} + +static noinline_for_stack +int bch2_check_alloc_hole_freespace(struct btree_trans *trans, + struct bpos start, + struct bpos *end, + struct btree_iter *freespace_iter) +{ + struct bch_fs *c = trans->c; + struct bch_dev *ca; + struct bkey_s_c k; + struct printbuf buf = PRINTBUF; + int ret; + + ca = bch_dev_bkey_exists(c, start.inode); + if (!ca->mi.freespace_initialized) + return 0; + + bch2_btree_iter_set_pos(freespace_iter, start); + + k = bch2_btree_iter_peek_slot(freespace_iter); + ret = bkey_err(k); + if (ret) + goto err; + + *end = bkey_min(k.k->p, *end); + + if (k.k->type != KEY_TYPE_set && + (c->opts.reconstruct_alloc || + fsck_err(c, freespace_hole_missing, + "hole in alloc btree missing in freespace btree\n" + " device %llu buckets %llu-%llu", + freespace_iter->pos.inode, + freespace_iter->pos.offset, + end->offset))) { + struct bkey_i *update = + bch2_trans_kmalloc(trans, sizeof(*update)); + + ret = PTR_ERR_OR_ZERO(update); + if (ret) + goto err; + + bkey_init(&update->k); + update->k.type = KEY_TYPE_set; + update->k.p = freespace_iter->pos; + bch2_key_resize(&update->k, + min_t(u64, U32_MAX, end->offset - + freespace_iter->pos.offset)); + + ret = bch2_trans_update(trans, freespace_iter, update, 0); + if (ret) + goto err; + } +err: +fsck_err: + printbuf_exit(&buf); + return ret; +} + +static noinline_for_stack +int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans, + struct bpos start, + struct bpos *end, + struct btree_iter *bucket_gens_iter) +{ + struct bch_fs *c = trans->c; + struct bkey_s_c k; + struct printbuf buf = PRINTBUF; + unsigned i, gens_offset, gens_end_offset; + int ret; + + if (c->sb.version < bcachefs_metadata_version_bucket_gens) + return 0; + + bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset)); + + k = bch2_btree_iter_peek_slot(bucket_gens_iter); + ret = bkey_err(k); + if (ret) + goto err; + + if (bkey_cmp(alloc_gens_pos(start, &gens_offset), + alloc_gens_pos(*end, &gens_end_offset))) + gens_end_offset = KEY_TYPE_BUCKET_GENS_NR; + + if (k.k->type == KEY_TYPE_bucket_gens) { + struct bkey_i_bucket_gens g; + bool need_update = false; + + bkey_reassemble(&g.k_i, k); + + for (i = gens_offset; i < gens_end_offset; i++) { + if (fsck_err_on(g.v.gens[i], c, + bucket_gens_hole_wrong, + "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)", + bucket_gens_pos_to_alloc(k.k->p, i).inode, + bucket_gens_pos_to_alloc(k.k->p, i).offset, + g.v.gens[i])) { + g.v.gens[i] = 0; + need_update = true; + } + } + + if (need_update) { + struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g)); + + ret = PTR_ERR_OR_ZERO(u); + if (ret) + goto err; + + memcpy(u, &g, sizeof(g)); + + ret = bch2_trans_update(trans, bucket_gens_iter, u, 0); + if (ret) + goto err; + } + } + + *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0)); +err: +fsck_err: + printbuf_exit(&buf); + return ret; +} + +static noinline_for_stack int bch2_check_discard_freespace_key(struct btree_trans *trans, + struct btree_iter *iter) +{ + struct bch_fs *c = trans->c; + struct btree_iter alloc_iter; + struct bkey_s_c alloc_k; + struct bch_alloc_v4 a_convert; + const struct bch_alloc_v4 *a; + u64 genbits; + struct bpos pos; + enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard + ? BCH_DATA_need_discard + : BCH_DATA_free; + struct printbuf buf = PRINTBUF; + int ret; + + pos = iter->pos; + pos.offset &= ~(~0ULL << 56); + genbits = iter->pos.offset & (~0ULL << 56); + + alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0); + ret = bkey_err(alloc_k); + if (ret) + return ret; + + if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c, + need_discard_freespace_key_to_invalid_dev_bucket, + "entry in %s btree for nonexistant dev:bucket %llu:%llu", + bch2_btree_id_str(iter->btree_id), pos.inode, pos.offset)) + goto delete; + + a = bch2_alloc_to_v4(alloc_k, &a_convert); + + if (fsck_err_on(a->data_type != state || + (state == BCH_DATA_free && + genbits != alloc_freespace_genbits(*a)), c, + need_discard_freespace_key_bad, + "%s\n incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)", + (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf), + bch2_btree_id_str(iter->btree_id), + iter->pos.inode, + iter->pos.offset, + a->data_type == state, + genbits >> 56, alloc_freespace_genbits(*a) >> 56)) + goto delete; +out: +fsck_err: + set_btree_iter_dontneed(&alloc_iter); + bch2_trans_iter_exit(trans, &alloc_iter); + printbuf_exit(&buf); + return ret; +delete: + ret = bch2_btree_delete_extent_at(trans, iter, + iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?: + bch2_trans_commit(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc); + goto out; } /* - * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc: + * We've already checked that generation numbers in the bucket_gens btree are + * valid for buckets that exist; this just checks for keys for nonexistent + * buckets. */ -static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca) +static noinline_for_stack +int bch2_check_bucket_gens_key(struct btree_trans *trans, + struct btree_iter *iter, + struct bkey_s_c k) { - u64 journal_seq = 0; + struct bch_fs *c = trans->c; + struct bkey_i_bucket_gens g; + struct bch_dev *ca; + u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset; + u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset; + u64 b; + bool need_update = false, dev_exists; + struct printbuf buf = PRINTBUF; int ret = 0; - /* Only use nowait if we've already invalidated at least one bucket: */ - while (!ret && - !fifo_full(&ca->free_inc) && - ca->alloc_heap.used) { - if (kthread_should_stop()) { - ret = 1; + BUG_ON(k.k->type != KEY_TYPE_bucket_gens); + bkey_reassemble(&g.k_i, k); + + /* if no bch_dev, skip out whether we repair or not */ + dev_exists = bch2_dev_exists2(c, k.k->p.inode); + if (!dev_exists) { + if (fsck_err_on(!dev_exists, c, + bucket_gens_to_invalid_dev, + "bucket_gens key for invalid device:\n %s", + (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) { + ret = bch2_btree_delete_at(trans, iter, 0); + } + goto out; + } + + ca = bch_dev_bkey_exists(c, k.k->p.inode); + if (fsck_err_on(end <= ca->mi.first_bucket || + start >= ca->mi.nbuckets, c, + bucket_gens_to_invalid_buckets, + "bucket_gens key for invalid buckets:\n %s", + (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) { + ret = bch2_btree_delete_at(trans, iter, 0); + goto out; + } + + for (b = start; b < ca->mi.first_bucket; b++) + if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c, + bucket_gens_nonzero_for_invalid_buckets, + "bucket_gens key has nonzero gen for invalid bucket")) { + g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0; + need_update = true; + } + + for (b = ca->mi.nbuckets; b < end; b++) + if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c, + bucket_gens_nonzero_for_invalid_buckets, + "bucket_gens key has nonzero gen for invalid bucket")) { + g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0; + need_update = true; + } + + if (need_update) { + struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g)); + + ret = PTR_ERR_OR_ZERO(u); + if (ret) + goto out; + + memcpy(u, &g, sizeof(g)); + ret = bch2_trans_update(trans, iter, u, 0); + } +out: +fsck_err: + printbuf_exit(&buf); + return ret; +} + +int bch2_check_alloc_info(struct bch_fs *c) +{ + struct btree_trans *trans = bch2_trans_get(c); + struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter; + struct bkey hole; + struct bkey_s_c k; + int ret = 0; + + bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS_MIN, + BTREE_ITER_PREFETCH); + bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard, POS_MIN, + BTREE_ITER_PREFETCH); + bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace, POS_MIN, + BTREE_ITER_PREFETCH); + bch2_trans_iter_init(trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN, + BTREE_ITER_PREFETCH); + + while (1) { + struct bpos next; + + bch2_trans_begin(trans); + + k = bch2_get_key_or_real_bucket_hole(&iter, &hole); + ret = bkey_err(k); + if (ret) + goto bkey_err; + + if (!k.k) break; + + if (k.k->type) { + next = bpos_nosnap_successor(k.k->p); + + ret = bch2_check_alloc_key(trans, + k, &iter, + &discard_iter, + &freespace_iter, + &bucket_gens_iter); + if (ret) + goto bkey_err; + } else { + next = k.k->p; + + ret = bch2_check_alloc_hole_freespace(trans, + bkey_start_pos(k.k), + &next, + &freespace_iter) ?: + bch2_check_alloc_hole_bucket_gens(trans, + bkey_start_pos(k.k), + &next, + &bucket_gens_iter); + if (ret) + goto bkey_err; } - ret = bch2_invalidate_one_bucket(c, ca, &journal_seq, - (!fifo_empty(&ca->free_inc) - ? BTREE_INSERT_NOWAIT : 0)); - /* - * We only want to batch up invalidates when they're going to - * require flushing the journal: - */ - if (!journal_seq) + ret = bch2_trans_commit(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc); + if (ret) + goto bkey_err; + + bch2_btree_iter_set_pos(&iter, next); +bkey_err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + continue; + if (ret) break; } + bch2_trans_iter_exit(trans, &bucket_gens_iter); + bch2_trans_iter_exit(trans, &freespace_iter); + bch2_trans_iter_exit(trans, &discard_iter); + bch2_trans_iter_exit(trans, &iter); - /* If we used NOWAIT, don't return the error: */ - if (!fifo_empty(&ca->free_inc)) - ret = 0; if (ret < 0) - bch_err(ca, "error invalidating buckets: %i", ret); + goto err; + + ret = for_each_btree_key(trans, iter, + BTREE_ID_need_discard, POS_MIN, + BTREE_ITER_PREFETCH, k, + bch2_check_discard_freespace_key(trans, &iter)); if (ret) - return ret; + goto err; - if (journal_seq) - ret = bch2_journal_flush_seq(&c->journal, journal_seq); - if (ret) { - bch_err(ca, "journal error: %i", ret); - return ret; + bch2_trans_iter_init(trans, &iter, BTREE_ID_freespace, POS_MIN, + BTREE_ITER_PREFETCH); + while (1) { + bch2_trans_begin(trans); + k = bch2_btree_iter_peek(&iter); + if (!k.k) + break; + + ret = bkey_err(k) ?: + bch2_check_discard_freespace_key(trans, &iter); + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) { + ret = 0; + continue; + } + if (ret) { + struct printbuf buf = PRINTBUF; + bch2_bkey_val_to_text(&buf, c, k); + + bch_err(c, "while checking %s", buf.buf); + printbuf_exit(&buf); + break; + } + + bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos)); } + bch2_trans_iter_exit(trans, &iter); + if (ret) + goto err; - return 0; + ret = for_each_btree_key_commit(trans, iter, + BTREE_ID_bucket_gens, POS_MIN, + BTREE_ITER_PREFETCH, k, + NULL, NULL, BCH_TRANS_COMMIT_no_enospc, + bch2_check_bucket_gens_key(trans, &iter, k)); +err: + bch2_trans_put(trans); + if (ret) + bch_err_fn(c, ret); + return ret; } -static void alloc_thread_set_state(struct bch_dev *ca, unsigned new_state) +static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans, + struct btree_iter *alloc_iter) { - if (ca->allocator_state != new_state) { - ca->allocator_state = new_state; - closure_wake_up(&ca->fs->freelist_wait); + struct bch_fs *c = trans->c; + struct btree_iter lru_iter; + struct bch_alloc_v4 a_convert; + const struct bch_alloc_v4 *a; + struct bkey_s_c alloc_k, lru_k; + struct printbuf buf = PRINTBUF; + int ret; + + alloc_k = bch2_btree_iter_peek(alloc_iter); + if (!alloc_k.k) + return 0; + + ret = bkey_err(alloc_k); + if (ret) + return ret; + + a = bch2_alloc_to_v4(alloc_k, &a_convert); + + if (a->data_type != BCH_DATA_cached) + return 0; + + if (fsck_err_on(!a->io_time[READ], c, + alloc_key_cached_but_read_time_zero, + "cached bucket with read_time 0\n" + " %s", + (printbuf_reset(&buf), + bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) { + struct bkey_i_alloc_v4 *a_mut = + bch2_alloc_to_v4_mut(trans, alloc_k); + ret = PTR_ERR_OR_ZERO(a_mut); + if (ret) + goto err; + + a_mut->v.io_time[READ] = atomic64_read(&c->io_clock[READ].now); + ret = bch2_trans_update(trans, alloc_iter, + &a_mut->k_i, BTREE_TRIGGER_NORUN); + if (ret) + goto err; + + a = &a_mut->v; + } + + lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru, + lru_pos(alloc_k.k->p.inode, + bucket_to_u64(alloc_k.k->p), + a->io_time[READ]), 0); + ret = bkey_err(lru_k); + if (ret) + return ret; + + if (fsck_err_on(lru_k.k->type != KEY_TYPE_set, c, + alloc_key_to_missing_lru_entry, + "missing lru entry\n" + " %s", + (printbuf_reset(&buf), + bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) { + ret = bch2_lru_set(trans, + alloc_k.k->p.inode, + bucket_to_u64(alloc_k.k->p), + a->io_time[READ]); + if (ret) + goto err; } +err: +fsck_err: + bch2_trans_iter_exit(trans, &lru_iter); + printbuf_exit(&buf); + return ret; } -static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b) +int bch2_check_alloc_to_lru_refs(struct bch_fs *c) { - unsigned i; + struct btree_iter iter; + struct bkey_s_c k; + + int ret = bch2_trans_run(c, + for_each_btree_key_commit(trans, iter, BTREE_ID_alloc, + POS_MIN, BTREE_ITER_PREFETCH, k, + NULL, NULL, BCH_TRANS_COMMIT_no_enospc, + bch2_check_alloc_to_lru_ref(trans, &iter))); + bch_err_fn(c, ret); + return ret; +} + +static int bch2_discard_one_bucket(struct btree_trans *trans, + struct btree_iter *need_discard_iter, + struct bpos *discard_pos_done, + u64 *seen, + u64 *open, + u64 *need_journal_commit, + u64 *discarded) +{ + struct bch_fs *c = trans->c; + struct bpos pos = need_discard_iter->pos; + struct btree_iter iter = { NULL }; + struct bkey_s_c k; + struct bch_dev *ca; + struct bkey_i_alloc_v4 *a; + struct printbuf buf = PRINTBUF; int ret = 0; - spin_lock(&c->freelist_lock); - for (i = 0; i < RESERVE_NR; i++) { - /* - * Don't strand buckets on the copygc freelist until - * after recovery is finished: - */ - if (i == RESERVE_MOVINGGC && - !test_bit(BCH_FS_STARTED, &c->flags)) - continue; + ca = bch_dev_bkey_exists(c, pos.inode); + if (!percpu_ref_tryget(&ca->io_ref)) { + bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0)); + return 0; + } - if (fifo_push(&ca->free[i], b)) { - fifo_pop(&ca->free_inc, b); - ret = 1; - break; + if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) { + (*open)++; + goto out; + } + + if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal, + c->journal.flushed_seq_ondisk, + pos.inode, pos.offset)) { + (*need_journal_commit)++; + goto out; + } + + k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc, + need_discard_iter->pos, + BTREE_ITER_CACHED); + ret = bkey_err(k); + if (ret) + goto out; + + a = bch2_alloc_to_v4_mut(trans, k); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + goto out; + + if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) { + a->v.gen++; + SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false); + goto write; + } + + if (a->v.journal_seq > c->journal.flushed_seq_ondisk) { + if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) { + bch2_trans_inconsistent(trans, + "clearing need_discard but journal_seq %llu > flushed_seq %llu\n" + "%s", + a->v.journal_seq, + c->journal.flushed_seq_ondisk, + (bch2_bkey_val_to_text(&buf, c, k), buf.buf)); + ret = -EIO; } + goto out; } - spin_unlock(&c->freelist_lock); - ca->allocator_state = ret - ? ALLOCATOR_running - : ALLOCATOR_blocked_full; - closure_wake_up(&c->freelist_wait); + if (a->v.data_type != BCH_DATA_need_discard) { + if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) { + bch2_trans_inconsistent(trans, + "bucket incorrectly set in need_discard btree\n" + "%s", + (bch2_bkey_val_to_text(&buf, c, k), buf.buf)); + ret = -EIO; + } + + goto out; + } + + if (!bkey_eq(*discard_pos_done, iter.pos) && + ca->mi.discard && !c->opts.nochanges) { + /* + * This works without any other locks because this is the only + * thread that removes items from the need_discard tree + */ + bch2_trans_unlock(trans); + blkdev_issue_discard(ca->disk_sb.bdev, + k.k->p.offset * ca->mi.bucket_size, + ca->mi.bucket_size, + GFP_KERNEL); + *discard_pos_done = iter.pos; + + ret = bch2_trans_relock_notrace(trans); + if (ret) + goto out; + } + + SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false); + a->v.data_type = alloc_data_type(a->v, a->v.data_type); +write: + ret = bch2_trans_update(trans, &iter, &a->k_i, 0) ?: + bch2_trans_commit(trans, NULL, NULL, + BCH_WATERMARK_btree| + BCH_TRANS_COMMIT_no_enospc); + if (ret) + goto out; + + count_event(c, bucket_discard); + (*discarded)++; +out: + (*seen)++; + bch2_trans_iter_exit(trans, &iter); + percpu_ref_put(&ca->io_ref); + printbuf_exit(&buf); return ret; } -static void discard_one_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b) +static void bch2_do_discards_work(struct work_struct *work) { - if (ca->mi.discard && - blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev))) - blkdev_issue_discard(ca->disk_sb.bdev, bucket_to_sector(ca, b), - ca->mi.bucket_size, GFP_NOFS, 0); + struct bch_fs *c = container_of(work, struct bch_fs, discard_work); + struct btree_iter iter; + struct bkey_s_c k; + u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0; + struct bpos discard_pos_done = POS_MAX; + int ret; + + /* + * We're doing the commit in bch2_discard_one_bucket instead of using + * for_each_btree_key_commit() so that we can increment counters after + * successful commit: + */ + ret = bch2_trans_run(c, + for_each_btree_key(trans, iter, + BTREE_ID_need_discard, POS_MIN, 0, k, + bch2_discard_one_bucket(trans, &iter, &discard_pos_done, + &seen, + &open, + &need_journal_commit, + &discarded))); + + if (need_journal_commit * 2 > seen) + bch2_journal_flush_async(&c->journal, NULL); + + bch2_write_ref_put(c, BCH_WRITE_REF_discard); + + trace_discard_buckets(c, seen, open, need_journal_commit, discarded, + bch2_err_str(ret)); } -static bool allocator_thread_running(struct bch_dev *ca) +void bch2_do_discards(struct bch_fs *c) { - unsigned state = ca->mi.state == BCH_MEMBER_STATE_rw && - test_bit(BCH_FS_ALLOCATOR_RUNNING, &ca->fs->flags) - ? ALLOCATOR_running - : ALLOCATOR_stopped; - alloc_thread_set_state(ca, state); - return state == ALLOCATOR_running; + if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) && + !queue_work(c->write_ref_wq, &c->discard_work)) + bch2_write_ref_put(c, BCH_WRITE_REF_discard); } -static int buckets_available(struct bch_dev *ca, unsigned long gc_count) +static int invalidate_one_bucket(struct btree_trans *trans, + struct btree_iter *lru_iter, + struct bkey_s_c lru_k, + s64 *nr_to_invalidate) { - s64 available = dev_buckets_reclaimable(ca) - - (gc_count == ca->fs->gc_count ? ca->inc_gen_really_needs_gc : 0); - bool ret = available > 0; + struct bch_fs *c = trans->c; + struct btree_iter alloc_iter = { NULL }; + struct bkey_i_alloc_v4 *a = NULL; + struct printbuf buf = PRINTBUF; + struct bpos bucket = u64_to_bucket(lru_k.k->p.offset); + unsigned cached_sectors; + int ret = 0; + + if (*nr_to_invalidate <= 0) + return 1; + + if (!bch2_dev_bucket_exists(c, bucket)) { + prt_str(&buf, "lru entry points to invalid bucket"); + goto err; + } + + if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset)) + return 0; - alloc_thread_set_state(ca, ret - ? ALLOCATOR_running - : ALLOCATOR_blocked); + a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + goto out; + + /* We expect harmless races here due to the btree write buffer: */ + if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v)) + goto out; + + BUG_ON(a->v.data_type != BCH_DATA_cached); + + if (!a->v.cached_sectors) + bch_err(c, "invalidating empty bucket, confused"); + + cached_sectors = a->v.cached_sectors; + + SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false); + a->v.gen++; + a->v.data_type = 0; + a->v.dirty_sectors = 0; + a->v.cached_sectors = 0; + a->v.io_time[READ] = atomic64_read(&c->io_clock[READ].now); + a->v.io_time[WRITE] = atomic64_read(&c->io_clock[WRITE].now); + + ret = bch2_trans_update(trans, &alloc_iter, &a->k_i, + BTREE_TRIGGER_BUCKET_INVALIDATE) ?: + bch2_trans_commit(trans, NULL, NULL, + BCH_WATERMARK_btree| + BCH_TRANS_COMMIT_no_enospc); + if (ret) + goto out; + + trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors); + --*nr_to_invalidate; +out: + bch2_trans_iter_exit(trans, &alloc_iter); + printbuf_exit(&buf); return ret; +err: + prt_str(&buf, "\n lru key: "); + bch2_bkey_val_to_text(&buf, c, lru_k); + + prt_str(&buf, "\n lru entry: "); + bch2_lru_pos_to_text(&buf, lru_iter->pos); + + prt_str(&buf, "\n alloc key: "); + if (!a) + bch2_bpos_to_text(&buf, bucket); + else + bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i)); + + bch_err(c, "%s", buf.buf); + if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) { + bch2_inconsistent_error(c); + ret = -EINVAL; + } + + goto out; } -/** - * bch_allocator_thread - move buckets from free_inc to reserves - * - * The free_inc FIFO is populated by find_reclaimable_buckets(), and - * the reserves are depleted by bucket allocation. When we run out - * of free_inc, try to invalidate some buckets and write out - * prios and gens. - */ -static int bch2_allocator_thread(void *arg) +static void bch2_do_invalidates_work(struct work_struct *work) +{ + struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work); + struct bch_dev *ca; + struct btree_trans *trans = bch2_trans_get(c); + struct btree_iter iter; + struct bkey_s_c k; + unsigned i; + int ret = 0; + + ret = bch2_btree_write_buffer_tryflush(trans); + if (ret) + goto err; + + for_each_member_device(ca, c, i) { + s64 nr_to_invalidate = + should_invalidate_buckets(ca, bch2_dev_usage_read(ca)); + + ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru, + lru_pos(ca->dev_idx, 0, 0), + lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX), + BTREE_ITER_INTENT, k, + invalidate_one_bucket(trans, &iter, k, &nr_to_invalidate)); + + if (ret < 0) { + percpu_ref_put(&ca->ref); + break; + } + } +err: + bch2_trans_put(trans); + bch2_write_ref_put(c, BCH_WRITE_REF_invalidate); +} + +void bch2_do_invalidates(struct bch_fs *c) { - struct bch_dev *ca = arg; - struct bch_fs *c = ca->fs; - unsigned long gc_count = c->gc_count; - size_t nr; + if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) && + !queue_work(c->write_ref_wq, &c->invalidate_work)) + bch2_write_ref_put(c, BCH_WRITE_REF_invalidate); +} + +int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca, + u64 bucket_start, u64 bucket_end) +{ + struct btree_trans *trans = bch2_trans_get(c); + struct btree_iter iter; + struct bkey_s_c k; + struct bkey hole; + struct bpos end = POS(ca->dev_idx, bucket_end); + struct bch_member *m; + unsigned long last_updated = jiffies; int ret; - set_freezable(); + BUG_ON(bucket_start > bucket_end); + BUG_ON(bucket_end > ca->mi.nbuckets); + bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, + POS(ca->dev_idx, max_t(u64, ca->mi.first_bucket, bucket_start)), + BTREE_ITER_PREFETCH); + /* + * Scan the alloc btree for every bucket on @ca, and add buckets to the + * freespace/need_discard/need_gc_gens btrees as needed: + */ while (1) { - ret = kthread_wait_freezable(allocator_thread_running(ca)); + if (last_updated + HZ * 10 < jiffies) { + bch_info(ca, "%s: currently at %llu/%llu", + __func__, iter.pos.offset, ca->mi.nbuckets); + last_updated = jiffies; + } + + bch2_trans_begin(trans); + + if (bkey_ge(iter.pos, end)) { + ret = 0; + break; + } + + k = bch2_get_key_or_hole(&iter, end, &hole); + ret = bkey_err(k); if (ret) - goto stop; + goto bkey_err; + + if (k.k->type) { + /* + * We process live keys in the alloc btree one at a + * time: + */ + struct bch_alloc_v4 a_convert; + const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert); + + ret = bch2_bucket_do_index(trans, k, a, true) ?: + bch2_trans_commit(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc); + if (ret) + goto bkey_err; - while (!ca->alloc_heap.used) { - cond_resched(); + bch2_btree_iter_advance(&iter); + } else { + struct bkey_i *freespace; - ret = kthread_wait_freezable(buckets_available(ca, gc_count)); + freespace = bch2_trans_kmalloc(trans, sizeof(*freespace)); + ret = PTR_ERR_OR_ZERO(freespace); if (ret) - goto stop; + goto bkey_err; - gc_count = c->gc_count; - nr = find_reclaimable_buckets(c, ca); + bkey_init(&freespace->k); + freespace->k.type = KEY_TYPE_set; + freespace->k.p = k.k->p; + freespace->k.size = k.k->size; - trace_alloc_scan(ca, nr, ca->inc_gen_needs_gc, - ca->inc_gen_really_needs_gc); + ret = bch2_btree_insert_trans(trans, BTREE_ID_freespace, freespace, 0) ?: + bch2_trans_commit(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc); + if (ret) + goto bkey_err; - if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) || - ca->inc_gen_really_needs_gc) && - c->gc_thread) { - atomic_inc(&c->kick_gc); - wake_up_process(c->gc_thread); - } + bch2_btree_iter_set_pos(&iter, k.k->p); } - - ret = bch2_invalidate_buckets(c, ca); +bkey_err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + continue; if (ret) - goto stop; + break; + } - while (!fifo_empty(&ca->free_inc)) { - u64 b = fifo_peek(&ca->free_inc); + bch2_trans_iter_exit(trans, &iter); + bch2_trans_put(trans); - discard_one_bucket(c, ca, b); + if (ret < 0) { + bch_err_msg(ca, ret, "initializing free space"); + return ret; + } - ret = kthread_wait_freezable(push_invalidated_bucket(c, ca, b)); - if (ret) - goto stop; + mutex_lock(&c->sb_lock); + m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx); + SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true); + mutex_unlock(&c->sb_lock); + + return 0; +} + +int bch2_fs_freespace_init(struct bch_fs *c) +{ + struct bch_dev *ca; + unsigned i; + int ret = 0; + bool doing_init = false; + + /* + * We can crash during the device add path, so we need to check this on + * every mount: + */ + + for_each_member_device(ca, c, i) { + if (ca->mi.freespace_initialized) + continue; + + if (!doing_init) { + bch_info(c, "initializing freespace"); + doing_init = true; } + + ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets); + if (ret) { + percpu_ref_put(&ca->ref); + bch_err_fn(c, ret); + return ret; + } + } + + if (doing_init) { + mutex_lock(&c->sb_lock); + bch2_write_super(c); + mutex_unlock(&c->sb_lock); + bch_verbose(c, "done initializing freespace"); } -stop: - alloc_thread_set_state(ca, ALLOCATOR_stopped); + return 0; } +/* Bucket IO clocks: */ + +int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev, + size_t bucket_nr, int rw) +{ + struct bch_fs *c = trans->c; + struct btree_iter iter; + struct bkey_i_alloc_v4 *a; + u64 now; + int ret = 0; + + a = bch2_trans_start_alloc_update(trans, &iter, POS(dev, bucket_nr)); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + return ret; + + now = atomic64_read(&c->io_clock[rw].now); + if (a->v.io_time[rw] == now) + goto out; + + a->v.io_time[rw] = now; + + ret = bch2_trans_update(trans, &iter, &a->k_i, 0) ?: + bch2_trans_commit(trans, NULL, NULL, 0); +out: + bch2_trans_iter_exit(trans, &iter); + return ret; +} + /* Startup/shutdown (ro/rw): */ void bch2_recalc_capacity(struct bch_fs *c) @@ -1072,7 +2019,7 @@ void bch2_recalc_capacity(struct bch_fs *c) u64 capacity = 0, reserved_sectors = 0, gc_reserve; unsigned bucket_size_max = 0; unsigned long ra_pages = 0; - unsigned i, j; + unsigned i; lockdep_assert_held(&c->state_lock); @@ -1103,8 +2050,9 @@ void bch2_recalc_capacity(struct bch_fs *c) * allocations for foreground writes must wait - * not -ENOSPC calculations. */ - for (j = 0; j < RESERVE_NONE; j++) - dev_reserve += ca->free[j].size; + + dev_reserve += ca->nr_btree_reserve * 2; + dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */ dev_reserve += 1; /* btree write point */ dev_reserve += 1; /* copygc write point */ @@ -1137,6 +2085,17 @@ void bch2_recalc_capacity(struct bch_fs *c) closure_wake_up(&c->freelist_wait); } +u64 bch2_min_rw_member_capacity(struct bch_fs *c) +{ + struct bch_dev *ca; + unsigned i; + u64 ret = U64_MAX; + + for_each_rw_member(ca, c, i) + ret = min(ret, ca->mi.nbuckets * ca->mi.bucket_size); + return ret; +} + static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca) { struct open_bucket *ob; @@ -1147,7 +2106,7 @@ static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca) ob++) { spin_lock(&ob->lock); if (ob->valid && !ob->on_partial_list && - ob->ptr.dev == ca->dev_idx) + ob->dev == ca->dev_idx) ret = true; spin_unlock(&ob->lock); } @@ -1160,8 +2119,6 @@ void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca) { unsigned i; - BUG_ON(ca->alloc_thread); - /* First, remove device from allocation groups: */ for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++) @@ -1172,40 +2129,7 @@ void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca) */ bch2_recalc_capacity(c); - /* Next, close write points that point to this device... */ - for (i = 0; i < ARRAY_SIZE(c->write_points); i++) - bch2_writepoint_stop(c, ca, &c->write_points[i]); - - bch2_writepoint_stop(c, ca, &c->copygc_write_point); - bch2_writepoint_stop(c, ca, &c->rebalance_write_point); - bch2_writepoint_stop(c, ca, &c->btree_write_point); - - mutex_lock(&c->btree_reserve_cache_lock); - while (c->btree_reserve_cache_nr) { - struct btree_alloc *a = - &c->btree_reserve_cache[--c->btree_reserve_cache_nr]; - - bch2_open_buckets_put(c, &a->ob); - } - mutex_unlock(&c->btree_reserve_cache_lock); - - while (1) { - struct open_bucket *ob; - - spin_lock(&c->freelist_lock); - if (!ca->open_buckets_partial_nr) { - spin_unlock(&c->freelist_lock); - break; - } - ob = c->open_buckets + - ca->open_buckets_partial[--ca->open_buckets_partial_nr]; - ob->on_partial_list = false; - spin_unlock(&c->freelist_lock); - - bch2_open_bucket_put(c, ob); - } - - bch2_ec_stop_dev(c, ca); + bch2_open_buckets_stop(c, ca, false); /* * Wake up threads that were blocked on allocation, so they can notice @@ -1235,81 +2159,9 @@ void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca) set_bit(ca->dev_idx, c->rw_devs[i].d); } -void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca) -{ - if (ca->alloc_thread) - closure_wait_event(&c->freelist_wait, - ca->allocator_state != ALLOCATOR_running); -} - -/* stop allocator thread: */ -void bch2_dev_allocator_stop(struct bch_dev *ca) -{ - struct task_struct *p; - - p = rcu_dereference_protected(ca->alloc_thread, 1); - ca->alloc_thread = NULL; - - /* - * We need an rcu barrier between setting ca->alloc_thread = NULL and - * the thread shutting down to avoid bch2_wake_allocator() racing: - * - * XXX: it would be better to have the rcu barrier be asynchronous - * instead of blocking us here - */ - synchronize_rcu(); - - if (p) { - kthread_stop(p); - put_task_struct(p); - } -} - -/* start allocator thread: */ -int bch2_dev_allocator_start(struct bch_dev *ca) -{ - struct task_struct *p; - - /* - * allocator thread already started? - */ - if (ca->alloc_thread) - return 0; - - p = kthread_create(bch2_allocator_thread, ca, - "bch-alloc/%s", ca->name); - if (IS_ERR(p)) { - bch_err(ca->fs, "error creating allocator thread: %li", - PTR_ERR(p)); - return PTR_ERR(p); - } - - get_task_struct(p); - rcu_assign_pointer(ca->alloc_thread, p); - wake_up_process(p); - return 0; -} - void bch2_fs_allocator_background_init(struct bch_fs *c) { spin_lock_init(&c->freelist_lock); -} - -void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c) -{ - struct open_bucket *ob; - - for (ob = c->open_buckets; - ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); - ob++) { - spin_lock(&ob->lock); - if (ob->valid && !ob->on_partial_list) { - pr_buf(out, "%zu ref %u type %s\n", - ob - c->open_buckets, - atomic_read(&ob->pin), - bch2_data_types[ob->type]); - } - spin_unlock(&ob->lock); - } - + INIT_WORK(&c->discard_work, bch2_do_discards_work); + INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work); }