X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Falloc_background.c;h=9aa0b42b26b671f483dd1969ae12c10835143721;hb=7fd6c3ffe45b3b42c0bc8a8c5d1387a5e3316a54;hp=83caa05e6b1fa93ff7f16e99010a3b7e36ed4362;hpb=82c7e8ff35ac0c501290cd4779e8ab2c553aa09d;p=bcachefs-tools-debian diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c index 83caa05..113273b 100644 --- a/libbcachefs/alloc_background.c +++ b/libbcachefs/alloc_background.c @@ -2,17 +2,24 @@ #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 #include @@ -21,58 +28,31 @@ #include #include #include -#include -static const char * const bch2_alloc_field_names[] = { -#define x(name, bytes) #name, - BCH_ALLOC_FIELDS() +/* Persistent alloc info: */ + +static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = { +#define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8, + BCH_ALLOC_FIELDS_V1() #undef x - NULL }; -static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int); - -/* Ratelimiting/PD controllers */ - -static void pd_controllers_update(struct work_struct *work) -{ - struct bch_fs *c = container_of(to_delayed_work(work), - struct bch_fs, - pd_controllers_update); - struct bch_dev *ca; - unsigned i; - - for_each_member_device(ca, c, i) { - struct bch_dev_usage stats = bch2_dev_usage_read(c, ca); - - u64 free = bucket_to_sector(ca, - __dev_buckets_free(ca, stats)) << 9; - /* - * Bytes of internal fragmentation, which can be - * reclaimed by copy GC - */ - s64 fragmented = (bucket_to_sector(ca, - stats.buckets[BCH_DATA_USER] + - stats.buckets[BCH_DATA_CACHED]) - - (stats.sectors[BCH_DATA_USER] + - stats.sectors[BCH_DATA_CACHED])) << 9; - - fragmented = max(0LL, fragmented); - - bch2_pd_controller_update(&ca->copygc_pd, - free, fragmented, -1); - } - - schedule_delayed_work(&c->pd_controllers_update, - c->pd_controllers_update_seconds * HZ); -} - -/* 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 get_alloc_field(const struct bch_alloc *a, - const void **p, unsigned field) +static inline u64 alloc_field_v1_get(const struct bch_alloc *a, + const void **p, unsigned field) { - unsigned bytes = BCH_ALLOC_FIELD_BYTES[field]; + unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field]; u64 v; if (!(a->fields & (1 << field))) @@ -99,1098 +79,1930 @@ static inline u64 get_alloc_field(const struct bch_alloc *a, return v; } -static inline void put_alloc_field(struct bkey_i_alloc *a, void **p, - unsigned field, u64 v) +static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out, + struct bkey_s_c k) { - unsigned bytes = BCH_ALLOC_FIELD_BYTES[field]; + const struct bch_alloc *in = bkey_s_c_to_alloc(k).v; + const void *d = in->data; + unsigned idx = 0; - if (!v) - return; + out->gen = in->gen; - a->v.fields |= 1 << field; +#define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++); + BCH_ALLOC_FIELDS_V1() +#undef x +} - switch (bytes) { - case 1: - *((u8 *) *p) = v; - break; - case 2: - *((__le16 *) *p) = cpu_to_le16(v); +static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out, + struct bkey_s_c k) +{ + struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k); + const u8 *in = a.v->data; + const u8 *end = bkey_val_end(a); + unsigned fieldnr = 0; + int ret; + u64 v; + + out->gen = a.v->gen; + out->oldest_gen = a.v->oldest_gen; + out->data_type = a.v->data_type; + +#define x(_name, _bits) \ + if (fieldnr < a.v->nr_fields) { \ + ret = bch2_varint_decode_fast(in, end, &v); \ + if (ret < 0) \ + return ret; \ + in += ret; \ + } else { \ + v = 0; \ + } \ + out->_name = v; \ + if (v != out->_name) \ + return -1; \ + fieldnr++; + + BCH_ALLOC_FIELDS_V2() +#undef x + return 0; +} + +static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out, + struct bkey_s_c k) +{ + struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k); + const u8 *in = a.v->data; + const u8 *end = bkey_val_end(a); + unsigned fieldnr = 0; + int ret; + u64 v; + + 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) \ + if (fieldnr < a.v->nr_fields) { \ + ret = bch2_varint_decode_fast(in, end, &v); \ + if (ret < 0) \ + return ret; \ + in += ret; \ + } else { \ + v = 0; \ + } \ + out->_name = v; \ + if (v != out->_name) \ + return -1; \ + fieldnr++; + + BCH_ALLOC_FIELDS_V2() +#undef x + return 0; +} + +static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k) +{ + struct bkey_alloc_unpacked ret = { .gen = 0 }; + + switch (k.k->type) { + case KEY_TYPE_alloc: + bch2_alloc_unpack_v1(&ret, k); break; - case 4: - *((__le32 *) *p) = cpu_to_le32(v); + case KEY_TYPE_alloc_v2: + bch2_alloc_unpack_v2(&ret, k); break; - case 8: - *((__le64 *) *p) = cpu_to_le64(v); + case KEY_TYPE_alloc_v3: + bch2_alloc_unpack_v3(&ret, k); break; - default: - BUG(); } - *p += bytes; + return ret; } -struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k) +static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a) { - struct bkey_alloc_unpacked ret = { .gen = 0 }; + unsigned i, bytes = offsetof(struct bch_alloc, data); - if (k.k->type == KEY_TYPE_alloc) { - const struct bch_alloc *a = bkey_s_c_to_alloc(k).v; - const void *d = a->data; - unsigned idx = 0; + for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++) + if (a->fields & (1 << i)) + bytes += BCH_ALLOC_V1_FIELD_BYTES[i]; - ret.gen = a->gen; + return DIV_ROUND_UP(bytes, sizeof(u64)); +} -#define x(_name, _bits) ret._name = get_alloc_field(a, &d, idx++); - BCH_ALLOC_FIELDS() -#undef x - } +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); + int ret = 0; + + /* allow for unknown fields */ + 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; } -void bch2_alloc_pack(struct bkey_i_alloc *dst, - const struct bkey_alloc_unpacked src) +int bch2_alloc_v2_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, + struct printbuf *err) { - unsigned idx = 0; - void *d = dst->v.data; - unsigned bytes; + struct bkey_alloc_unpacked u; + int ret = 0; - dst->v.fields = 0; - dst->v.gen = src.gen; + bkey_fsck_err_on(bch2_alloc_unpack_v2(&u, k), c, err, + alloc_v2_unpack_error, + "unpack error"); +fsck_err: + return ret; +} -#define x(_name, _bits) put_alloc_field(dst, &d, idx++, src._name); - BCH_ALLOC_FIELDS() -#undef x +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; - bytes = (void *) d - (void *) &dst->v; - set_bkey_val_bytes(&dst->k, bytes); - memset_u64s_tail(&dst->v, 0, bytes); + bkey_fsck_err_on(bch2_alloc_unpack_v3(&u, k), c, err, + alloc_v2_unpack_error, + "unpack error"); +fsck_err: + return ret; } -static unsigned bch_alloc_val_u64s(const struct bch_alloc *a) +int bch2_alloc_v4_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, struct printbuf *err) { - unsigned i, bytes = offsetof(struct bch_alloc, data); + struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k); + int ret = 0; - for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_FIELD_BYTES); i++) - if (a->fields & (1 << i)) - bytes += BCH_ALLOC_FIELD_BYTES[i]; + 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(a.v->dirty_sectors || + a.v->cached_sectors || + 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(!a.v->dirty_sectors, 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 || + a.v->dirty_sectors || + 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; +} - return DIV_ROUND_UP(bytes, sizeof(u64)); +void bch2_alloc_v4_swab(struct bkey_s k) +{ + 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); + } } -const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k) +void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k) { - struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k); + struct bch_alloc_v4 _a; + const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a); + unsigned i; - if (k.k->p.inode >= c->sb.nr_devices || - !c->devs[k.k->p.inode]) - return "invalid device"; + 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]); + } - /* allow for unknown fields */ - if (bkey_val_u64s(a.k) < bch_alloc_val_u64s(a.v)) - return "incorrect value size"; + printbuf_indent_sub(out, 2); + } - return NULL; + printbuf_indent_sub(out, 2); } -void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, - struct bkey_s_c k) +void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out) { - struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k); - const void *d = a.v->data; - unsigned i; + if (k.k->type == KEY_TYPE_alloc_v4) { + void *src, *dst; + + *out = *bkey_s_c_to_alloc_v4(k).v; + + src = alloc_v4_backpointers(out); + SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s); + dst = alloc_v4_backpointers(out); - pr_buf(out, "gen %u", a.v->gen); + if (src < dst) + memset(src, 0, dst - src); - for (i = 0; i < BCH_ALLOC_FIELD_NR; i++) - if (a.v->fields & (1 << i)) - pr_buf(out, " %s %llu", - bch2_alloc_field_names[i], - get_alloc_field(a.v, &d, i)); + 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); + } } -int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys) +static noinline struct bkey_i_alloc_v4 * +__bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k) { - struct btree_trans trans; - struct btree_iter *iter; - struct bkey_s_c k; - struct bch_dev *ca; - struct journal_key *j; - unsigned i; - int ret; + struct bkey_i_alloc_v4 *ret; - bch2_trans_init(&trans, c, 0, 0); + ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4))); + if (IS_ERR(ret)) + return ret; - for_each_btree_key(&trans, iter, BTREE_ID_ALLOC, POS_MIN, 0, k, ret) - bch2_mark_key(c, k, 0, 0, NULL, 0, - BCH_BUCKET_MARK_ALLOC_READ| - BCH_BUCKET_MARK_NOATOMIC); + if (k.k->type == KEY_TYPE_alloc_v4) { + void *src, *dst; - ret = bch2_trans_exit(&trans) ?: ret; - if (ret) { - bch_err(c, "error reading alloc info: %i", ret); - return ret; - } + bkey_reassemble(&ret->k_i, k); - for_each_journal_key(*journal_keys, j) - if (j->btree_id == BTREE_ID_ALLOC) - bch2_mark_key(c, bkey_i_to_s_c(j->k), - 0, 0, NULL, 0, - BCH_BUCKET_MARK_ALLOC_READ| - BCH_BUCKET_MARK_NOATOMIC); + src = alloc_v4_backpointers(&ret->v); + SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s); + dst = alloc_v4_backpointers(&ret->v); - percpu_down_write(&c->mark_lock); - bch2_dev_usage_from_buckets(c); - percpu_up_write(&c->mark_lock); + if (src < dst) + memset(src, 0, dst - src); - mutex_lock(&c->bucket_clock[READ].lock); - for_each_member_device(ca, c, i) { - down_read(&ca->bucket_lock); - bch2_recalc_oldest_io(c, ca, READ); - up_read(&ca->bucket_lock); + 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); } - mutex_unlock(&c->bucket_clock[READ].lock); + return ret; +} - mutex_lock(&c->bucket_clock[WRITE].lock); - for_each_member_device(ca, c, i) { - down_read(&ca->bucket_lock); - bch2_recalc_oldest_io(c, ca, WRITE); - up_read(&ca->bucket_lock); - } - mutex_unlock(&c->bucket_clock[WRITE].lock); +static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k) +{ + struct bkey_s_c_alloc_v4 a; - return 0; + 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); + + return __bch2_alloc_to_v4_mut(trans, k); } -enum alloc_write_ret { - ALLOC_WROTE, - ALLOC_NOWROTE, - ALLOC_END, -}; +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); +} -static int bch2_alloc_write_key(struct btree_trans *trans, - struct btree_iter *iter, - unsigned flags) +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 bkey_s_c k; - struct bch_dev *ca; - struct bucket_array *ba; - struct bucket *g; - struct bucket_mark m; - struct bkey_alloc_unpacked old_u, new_u; - __BKEY_PADDED(k, 8) alloc_key; /* hack: */ - struct bkey_i_alloc *a; + struct bkey_i_alloc_v4 *a; int ret; -retry: - k = bch2_btree_iter_peek_slot(iter); - ret = bkey_err(k); - if (ret) - goto err; - old_u = bch2_alloc_unpack(k); + k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos, + BTREE_ITER_WITH_UPDATES| + BTREE_ITER_CACHED| + BTREE_ITER_INTENT); + ret = bkey_err(k); + if (unlikely(ret)) + return ERR_PTR(ret); - if (iter->pos.inode >= c->sb.nr_devices || - !c->devs[iter->pos.inode]) - return ALLOC_END; + a = bch2_alloc_to_v4_mut_inlined(trans, k); + ret = PTR_ERR_OR_ZERO(a); + if (unlikely(ret)) + goto err; + return a; +err: + bch2_trans_iter_exit(trans, iter); + return ERR_PTR(ret); +} - percpu_down_read(&c->mark_lock); - ca = bch_dev_bkey_exists(c, iter->pos.inode); - ba = bucket_array(ca); +static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset) +{ + *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK; - if (iter->pos.offset >= ba->nbuckets) { - percpu_up_read(&c->mark_lock); - return ALLOC_END; - } + pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS; + return pos; +} - g = &ba->b[iter->pos.offset]; - m = READ_ONCE(g->mark); - new_u = alloc_mem_to_key(g, m); - percpu_up_read(&c->mark_lock); +static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset) +{ + pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS; + pos.offset += offset; + return pos; +} - if (!bkey_alloc_unpacked_cmp(old_u, new_u)) - return ALLOC_NOWROTE; +static unsigned alloc_gen(struct bkey_s_c k, unsigned offset) +{ + return k.k->type == KEY_TYPE_bucket_gens + ? bkey_s_c_to_bucket_gens(k).v->gens[offset] + : 0; +} - a = bkey_alloc_init(&alloc_key.k); - a->k.p = iter->pos; - bch2_alloc_pack(a, new_u); +int bch2_bucket_gens_invalid(struct bch_fs *c, struct bkey_s_c k, + enum bkey_invalid_flags flags, + struct printbuf *err) +{ + int ret = 0; - bch2_trans_update(trans, iter, &a->k_i); - ret = bch2_trans_commit(trans, NULL, NULL, - BTREE_INSERT_NOFAIL| - BTREE_INSERT_NOMARK| - flags); -err: - if (ret == -EINTR) - goto retry; + 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; } -int bch2_alloc_write(struct bch_fs *c, unsigned flags, bool *wrote) +void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k) { - struct btree_trans trans; - struct btree_iter *iter; - struct bch_dev *ca; + struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k); unsigned i; - int ret = 0; - - BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8); - - bch2_trans_init(&trans, c, 0, 0); - iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, POS_MIN, - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); + for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) { + if (i) + prt_char(out, ' '); + prt_printf(out, "%u", g.v->gens[i]); + } +} - for_each_rw_member(ca, c, i) { - unsigned first_bucket; +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; - percpu_down_read(&c->mark_lock); - first_bucket = bucket_array(ca)->first_bucket; - percpu_up_read(&c->mark_lock); + for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN, + BTREE_ITER_PREFETCH, k, ret) { + /* + * 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; - bch2_btree_iter_set_pos(iter, POS(i, first_bucket)); + gen = bch2_alloc_to_v4(k, &a)->gen; + pos = alloc_gens_pos(iter.pos, &offset); - while (1) { - ret = bch2_alloc_write_key(&trans, iter, flags); - if (ret < 0 || ret == ALLOC_END) + if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) { + ret = commit_do(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc| + BCH_TRANS_COMMIT_lazy_rw, + bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0)); + if (ret) break; - if (ret == ALLOC_WROTE) - *wrote = true; - bch2_btree_iter_next_slot(iter); + have_bucket_gens_key = false; } - if (ret < 0) { - percpu_ref_put(&ca->io_ref); - break; + if (!have_bucket_gens_key) { + bkey_bucket_gens_init(&g.k_i); + g.k.p = pos; + have_bucket_gens_key = true; } + + g.v.gens[offset] = gen; } + bch2_trans_iter_exit(trans, &iter); + + if (have_bucket_gens_key && !ret) + ret = commit_do(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc| + BCH_TRANS_COMMIT_lazy_rw, + bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0)); - bch2_trans_exit(&trans); + bch2_trans_put(trans); - return ret < 0 ? ret : 0; + if (ret) + bch_err_fn(c, ret); + return ret; } -int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k) +int bch2_alloc_read(struct bch_fs *c) { - struct btree_trans trans; - struct btree_iter *iter; + struct btree_trans *trans = bch2_trans_get(c); + struct btree_iter iter; + struct bkey_s_c k; + struct bch_dev *ca; int ret; - bch2_trans_init(&trans, c, 0, 0); + down_read(&c->gc_lock); - iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, k->k.p, - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); + if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_bucket_gens) { + const struct bch_bucket_gens *g; + u64 b; - ret = bch2_alloc_write_key(&trans, iter, - BTREE_INSERT_NOFAIL| - BTREE_INSERT_LAZY_RW| - BTREE_INSERT_JOURNAL_REPLAY| - BTREE_INSERT_NOMARK); - bch2_trans_exit(&trans); - return ret < 0 ? ret : 0; -} + for_each_btree_key(trans, iter, BTREE_ID_bucket_gens, POS_MIN, + BTREE_ITER_PREFETCH, k, ret) { + 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; -/* Bucket IO clocks: */ + if (k.k->type != KEY_TYPE_bucket_gens) + continue; -static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw) -{ - struct bucket_clock *clock = &c->bucket_clock[rw]; - struct bucket_array *buckets = bucket_array(ca); - struct bucket *g; - u16 max_last_io = 0; - unsigned i; + g = bkey_s_c_to_bucket_gens(k).v; + + /* + * 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; - lockdep_assert_held(&c->bucket_clock[rw].lock); + ca = bch_dev_bkey_exists(c, k.k->p.inode); - /* Recalculate max_last_io for this device: */ - for_each_bucket(g, buckets) - max_last_io = max(max_last_io, bucket_last_io(c, g, rw)); + 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]; + } + bch2_trans_iter_exit(trans, &iter); + } else { + struct bch_alloc_v4 a; + + for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN, + BTREE_ITER_PREFETCH, k, ret) { + /* + * 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); - ca->max_last_bucket_io[rw] = max_last_io; + *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen; + } + bch2_trans_iter_exit(trans, &iter); + } - /* Recalculate global max_last_io: */ - max_last_io = 0; + bch2_trans_put(trans); + up_read(&c->gc_lock); - for_each_member_device(ca, c, i) - max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]); + if (ret) + bch_err_fn(c, ret); - clock->max_last_io = max_last_io; + return ret; } -static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw) +/* 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) { - struct bucket_clock *clock = &c->bucket_clock[rw]; - struct bucket_array *buckets; - struct bch_dev *ca; - struct bucket *g; - unsigned i; + 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; - trace_rescale_prios(c); + if (a->data_type != BCH_DATA_free && + a->data_type != BCH_DATA_need_discard) + return 0; - for_each_member_device(ca, c, i) { - down_read(&ca->bucket_lock); - buckets = bucket_array(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; - for_each_bucket(g, buckets) - g->io_time[rw] = clock->hand - - bucket_last_io(c, g, rw) / 2; + 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_DATA_need_discard: + btree = BTREE_ID_need_discard; + k->k.p = alloc_k.k->p; + break; + default: + return 0; + } - bch2_recalc_oldest_io(c, ca, rw); + old = bch2_bkey_get_iter(trans, &iter, btree, + bkey_start_pos(&k->k), + BTREE_ITER_INTENT); + ret = bkey_err(old); + if (ret) + return ret; - up_read(&ca->bucket_lock); + 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 inline u64 bucket_clock_freq(u64 capacity) +static noinline int bch2_bucket_gen_update(struct btree_trans *trans, + struct bpos bucket, u8 gen) { - return max(capacity >> 10, 2028ULL); + 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; + + g = bch2_trans_kmalloc(trans, sizeof(*g)); + ret = PTR_ERR_OR_ZERO(g); + if (ret) + return ret; + + 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) + return ret; + + 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 void bch2_inc_clock_hand(struct io_timer *timer) +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_clock *clock = container_of(timer, - struct bucket_clock, rescale); - struct bch_fs *c = container_of(clock, - struct bch_fs, bucket_clock[clock->rw]); - struct bch_dev *ca; - u64 capacity; - unsigned i; + 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; + + /* + * Deletion only happens in the device removal path, with + * BTREE_TRIGGER_NORUN: + */ + BUG_ON(new->k.type != KEY_TYPE_alloc_v4); + + old_a = bch2_alloc_to_v4(old, &old_a_convert); + new_a = &bkey_i_to_alloc_v4(new)->v; - mutex_lock(&clock->lock); + new_a->data_type = alloc_data_type(*new_a, new_a->data_type); - /* if clock cannot be advanced more, rescale prio */ - if (clock->max_last_io >= U16_MAX - 2) - bch2_rescale_bucket_io_times(c, clock->rw); + if (new_a->dirty_sectors > old_a->dirty_sectors || + new_a->cached_sectors > old_a->cached_sectors) { + 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); + } - BUG_ON(clock->max_last_io >= U16_MAX - 2); + 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); + } - for_each_member_device(ca, c, i) - ca->max_last_bucket_io[clock->rw]++; - clock->max_last_io++; - clock->hand++; + 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; + } - mutex_unlock(&clock->lock); + 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)); - capacity = READ_ONCE(c->capacity); + old_lru = alloc_lru_idx_read(*old_a); + new_lru = alloc_lru_idx_read(*new_a); - if (!capacity) - return; + 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; + } - /* - * we only increment when 0.1% of the filesystem capacity has been read - * or written too, this determines if it's time - * - * XXX: we shouldn't really be going off of the capacity of devices in - * RW mode (that will be 0 when we're RO, yet we can still service - * reads) - */ - timer->expire += bucket_clock_freq(capacity); + new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a, + bch_dev_bkey_exists(c, new->k.p.inode)); - bch2_io_timer_add(&c->io_clock[clock->rw], timer); -} + 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; + } -static void bch2_bucket_clock_init(struct bch_fs *c, int rw) -{ - struct bucket_clock *clock = &c->bucket_clock[rw]; + if (old_a->gen != new_a->gen) { + ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen); + if (ret) + return ret; + } - clock->hand = 1; - clock->rw = rw; - clock->rescale.fn = bch2_inc_clock_hand; - clock->rescale.expire = bucket_clock_freq(c->capacity); - mutex_init(&clock->lock); + return 0; } -/* 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. + * 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); -#define BUCKET_GC_GEN_MAX 96U + if (bkey_err(k)) + return k; -/** - * wait_buckets_available - wait on reclaimable buckets - * - * If there aren't enough available buckets to fill up free_inc, wait until - * there are. - */ -static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca) -{ - unsigned long gc_count = c->gc_count; - int ret = 0; + if (k.k->type) { + return k; + } else { + struct btree_iter iter2; + struct bpos next; - ca->allocator_state = ALLOCATOR_BLOCKED; - closure_wake_up(&c->freelist_wait); + bch2_trans_copy_iter(&iter2, iter); - while (1) { - set_current_state(TASK_INTERRUPTIBLE); - if (kthread_should_stop()) { - ret = 1; - break; - } + if (!bpos_eq(iter->path->l[0].b->key.k.p, SPOS_MAX)) + end = bkey_min(end, bpos_nosnap_successor(iter->path->l[0].b->key.k.p)); - if (gc_count != c->gc_count) - ca->inc_gen_really_needs_gc = 0; + end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1)); - if ((ssize_t) (dev_buckets_available(c, ca) - - ca->inc_gen_really_needs_gc) >= - (ssize_t) fifo_free(&ca->free_inc)) - break; + /* + * btree node min/max is a closed interval, upto takes a half + * open interval: + */ + k = bch2_btree_iter_peek_upto(&iter2, end); + next = iter2.pos; + bch2_trans_iter_exit(iter->trans, &iter2); - up_read(&c->gc_lock); - schedule(); - try_to_freeze(); - down_read(&c->gc_lock); - } + BUG_ON(next.offset >= iter->pos.offset + U32_MAX); - __set_current_state(TASK_RUNNING); - ca->allocator_state = ALLOCATOR_RUNNING; - closure_wake_up(&c->freelist_wait); + if (bkey_err(k)) + return k; - return ret; + 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 bch2_can_invalidate_bucket(struct bch_dev *ca, - size_t bucket, - struct bucket_mark mark) +static bool next_bucket(struct bch_fs *c, struct bpos *bucket) { - u8 gc_gen; + struct bch_dev *ca; + unsigned iter; - if (!is_available_bucket(mark)) - return false; + if (bch2_dev_bucket_exists(c, *bucket)) + return true; - if (ca->buckets_nouse && - test_bit(bucket, ca->buckets_nouse)) - return false; + if (bch2_dev_exists2(c, bucket->inode)) { + ca = bch_dev_bkey_exists(c, bucket->inode); - gc_gen = bucket_gc_gen(ca, bucket); + if (bucket->offset < ca->mi.first_bucket) { + bucket->offset = ca->mi.first_bucket; + return true; + } - if (gc_gen >= BUCKET_GC_GEN_MAX / 2) - ca->inc_gen_needs_gc++; + bucket->inode++; + bucket->offset = 0; + } - if (gc_gen >= BUCKET_GC_GEN_MAX) - ca->inc_gen_really_needs_gc++; + 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 gc_gen < BUCKET_GC_GEN_MAX; + return ca != NULL; } -/* - * Determines what order we're going to reuse buckets, smallest bucket_key() - * first. - * - * - * - We take into account the read prio of the bucket, which gives us an - * indication of how hot the data is -- we scale the prio so that the prio - * farthest from the clock is worth 1/8th of the closest. - * - * - The number of sectors of cached data in the bucket, which gives us an - * indication of the cost in cache misses this eviction will cause. - * - * - If hotness * sectors used compares equal, we pick the bucket with the - * smallest bucket_gc_gen() - since incrementing the same bucket's generation - * number repeatedly forces us to run mark and sweep gc to avoid generation - * number wraparound. - */ - -static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca, - size_t b, struct bucket_mark m) +static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole) { - unsigned last_io = bucket_last_io(c, bucket(ca, b), READ); - unsigned max_last_io = ca->max_last_bucket_io[READ]; + 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; - /* - * Time since last read, scaled to [0, 8) where larger value indicates - * more recently read data: - */ - unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io; + if (!k.k->type) { + struct bpos bucket = bkey_start_pos(k.k); - /* How much we want to keep the data in this bucket: */ - unsigned long data_wantness = - (hotness + 1) * bucket_sectors_used(m); + if (!bch2_dev_bucket_exists(c, bucket)) { + if (!next_bucket(c, &bucket)) + return bkey_s_c_null; - unsigned long needs_journal_commit = - bucket_needs_journal_commit(m, c->journal.last_seq_ondisk); + bch2_btree_iter_set_pos(iter, bucket); + goto again; + } - return (data_wantness << 9) | - (needs_journal_commit << 8) | - (bucket_gc_gen(ca, b) / 16); -} + if (!bch2_dev_bucket_exists(c, k.k->p)) { + struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode); -static inline int bucket_alloc_cmp(alloc_heap *h, - struct alloc_heap_entry l, - struct alloc_heap_entry r) -{ - return cmp_int(l.key, r.key) ?: - cmp_int(r.nr, l.nr) ?: - cmp_int(l.bucket, r.bucket); + bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset); + } + } + + return k; } -static inline int bucket_idx_cmp(const void *_l, const void *_r) +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) { - const struct alloc_heap_entry *l = _l, *r = _r; + 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; - return cmp_int(l->bucket, r->bucket); -} + 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); -static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca) -{ - struct bucket_array *buckets; - struct alloc_heap_entry e = { 0 }; - size_t b, i, nr = 0; + ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode); + if (!ca->mi.freespace_initialized) + return 0; - ca->alloc_heap.used = 0; + a = bch2_alloc_to_v4(alloc_k, &a_convert); - mutex_lock(&c->bucket_clock[READ].lock); - down_read(&ca->bucket_lock); + 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; - buckets = bucket_array(ca); + 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; - bch2_recalc_oldest_io(c, ca, READ); + bkey_init(&update->k); + update->k.type = discard_key_type; + update->k.p = discard_iter->pos; - /* - * 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_mark m = READ_ONCE(buckets->b[b].mark); - unsigned long key = bucket_sort_key(c, ca, b, m); + ret = bch2_trans_update(trans, discard_iter, update, 0); + if (ret) + goto err; + } - if (!bch2_can_invalidate_bucket(ca, b, m)) - continue; + 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 (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 (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); - cond_resched(); + ret = bch2_trans_update(trans, freespace_iter, update, 0); + if (ret) + goto err; } - if (e.nr) - heap_add_or_replace(&ca->alloc_heap, e, - -bucket_alloc_cmp, NULL); + 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; - for (i = 0; i < ca->alloc_heap.used; i++) - nr += ca->alloc_heap.data[i].nr; + 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); + } - 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); - } + g->v.gens[gens_offset] = a->gen; - up_read(&ca->bucket_lock); - mutex_unlock(&c->bucket_clock[READ].lock); + 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 void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca) +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 bucket_array *buckets = bucket_array(ca); - struct bucket_mark m; - size_t b, start; + struct bch_fs *c = trans->c; + struct bch_dev *ca; + struct bkey_s_c k; + struct printbuf buf = PRINTBUF; + int ret; - if (ca->fifo_last_bucket < ca->mi.first_bucket || - ca->fifo_last_bucket >= ca->mi.nbuckets) - ca->fifo_last_bucket = ca->mi.first_bucket; + ca = bch_dev_bkey_exists(c, start.inode); + if (!ca->mi.freespace_initialized) + return 0; - start = ca->fifo_last_bucket; + bch2_btree_iter_set_pos(freespace_iter, start); - do { - ca->fifo_last_bucket++; - if (ca->fifo_last_bucket == ca->mi.nbuckets) - ca->fifo_last_bucket = ca->mi.first_bucket; + k = bch2_btree_iter_peek_slot(freespace_iter); + ret = bkey_err(k); + if (ret) + goto err; - b = ca->fifo_last_bucket; - m = READ_ONCE(buckets->b[b].mark); + *end = bkey_min(k.k->p, *end); - if (bch2_can_invalidate_bucket(ca, b, m)) { - struct alloc_heap_entry e = { .bucket = b, .nr = 1, }; + 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)); - heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); - if (heap_full(&ca->alloc_heap)) - break; - } + ret = PTR_ERR_OR_ZERO(update); + if (ret) + goto err; - cond_resched(); - } while (ca->fifo_last_bucket != start); + 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 void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca) +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 bucket_array *buckets = bucket_array(ca); - struct bucket_mark m; - size_t checked, i; + struct bch_fs *c = trans->c; + struct bkey_s_c k; + struct printbuf buf = PRINTBUF; + unsigned i, gens_offset, gens_end_offset; + int ret; - 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 (c->sb.version < bcachefs_metadata_version_bucket_gens) + return 0; - m = READ_ONCE(buckets->b[b].mark); + bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset)); - if (bch2_can_invalidate_bucket(ca, b, m)) { - struct alloc_heap_entry e = { .bucket = b, .nr = 1, }; + k = bch2_btree_iter_peek_slot(bucket_gens_iter); + ret = bkey_err(k); + if (ret) + goto err; - heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); - if (heap_full(&ca->alloc_heap)) - break; + 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; + } } - cond_resched(); - } + if (need_update) { + struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g)); - sort(ca->alloc_heap.data, - ca->alloc_heap.used, - sizeof(ca->alloc_heap.data[0]), - bucket_idx_cmp, NULL); + ret = PTR_ERR_OR_ZERO(u); + if (ret) + goto err; - /* 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; + 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 size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca) +static noinline_for_stack int __bch2_check_discard_freespace_key(struct btree_trans *trans, + struct btree_iter *iter) { - size_t i, nr = 0; + 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; - ca->inc_gen_needs_gc = 0; + pos = iter->pos; + pos.offset &= ~(~0ULL << 56); + genbits = iter->pos.offset & (~0ULL << 56); - switch (ca->mi.replacement) { - case CACHE_REPLACEMENT_LRU: - find_reclaimable_buckets_lru(c, ca); - break; - case CACHE_REPLACEMENT_FIFO: - find_reclaimable_buckets_fifo(c, ca); - break; - case CACHE_REPLACEMENT_RANDOM: - find_reclaimable_buckets_random(c, ca); - break; - } + alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0); + ret = bkey_err(alloc_k); + if (ret) + return ret; - heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL); + 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|BCH_TRANS_COMMIT_lazy_rw); + goto out; +} + +static int bch2_check_discard_freespace_key(struct btree_trans *trans, + struct btree_iter *iter, + struct bpos end) +{ + if (!btree_id_is_extents(iter->btree_id)) { + return __bch2_check_discard_freespace_key(trans, iter); + } else { + int ret = 0; - for (i = 0; i < ca->alloc_heap.used; i++) - nr += ca->alloc_heap.data[i].nr; + while (!bkey_eq(iter->pos, end) && + !(ret = btree_trans_too_many_iters(trans) ?: + __bch2_check_discard_freespace_key(trans, iter))) + bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos)); - return nr; + return ret; + } } -static inline long next_alloc_bucket(struct bch_dev *ca) +/* + * 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 noinline_for_stack +int bch2_check_bucket_gens_key(struct btree_trans *trans, + struct btree_iter *iter, + struct bkey_s_c k) { - struct alloc_heap_entry e, *top = ca->alloc_heap.data; - - while (ca->alloc_heap.used) { - if (top->nr) { - size_t b = top->bucket; + 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; - top->bucket++; - top->nr--; - return b; + 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; + } - heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); + 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; } - return -1; + 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; } -/* - * 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) +int bch2_check_alloc_info(struct bch_fs *c) { - if (m.journal_seq_valid) { - u64 journal_seq = atomic64_read(&c->journal.seq); - u64 bucket_seq = journal_seq; + 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; - bucket_seq &= ~((u64) U16_MAX); - bucket_seq |= m.journal_seq; + 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); - if (bucket_seq > journal_seq) - bucket_seq -= 1 << 16; + while (1) { + struct bpos next; - return bucket_seq; - } else { - return 0; + 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_trans_commit(trans, NULL, NULL, + BCH_TRANS_COMMIT_no_enospc| + BCH_TRANS_COMMIT_lazy_rw); + 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 (ret < 0) + goto err; + + ret = for_each_btree_key2(trans, iter, + BTREE_ID_need_discard, POS_MIN, + BTREE_ITER_PREFETCH, k, + bch2_check_discard_freespace_key(trans, &iter, k.k->p)) ?: + for_each_btree_key2(trans, iter, + BTREE_ID_freespace, POS_MIN, + BTREE_ITER_PREFETCH, k, + bch2_check_discard_freespace_key(trans, &iter, k.k->p)) ?: + for_each_btree_key_commit(trans, iter, + BTREE_ID_bucket_gens, POS_MIN, + BTREE_ITER_PREFETCH, k, + NULL, NULL, BCH_TRANS_COMMIT_no_enospc|BCH_TRANS_COMMIT_lazy_rw, + bch2_check_bucket_gens_key(trans, &iter, k)); +err: + bch2_trans_put(trans); + if (ret) + bch_err_fn(c, ret); + return ret; } -static int bch2_invalidate_one_bucket2(struct btree_trans *trans, - struct bch_dev *ca, - struct btree_iter *iter, - u64 *journal_seq, unsigned flags) +static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans, + struct btree_iter *alloc_iter) { -#if 0 - __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key; -#else - /* hack: */ - __BKEY_PADDED(k, 8) alloc_key; -#endif struct bch_fs *c = trans->c; - struct bkey_i_alloc *a; - struct bkey_alloc_unpacked u; - struct bucket *g; - struct bucket_mark m; - struct bkey_s_c k; - bool invalidating_cached_data; - size_t b; + 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; - BUG_ON(!ca->alloc_heap.used || - !ca->alloc_heap.data[0].nr); - b = ca->alloc_heap.data[0].bucket; + alloc_k = bch2_btree_iter_peek(alloc_iter); + if (!alloc_k.k) + return 0; - /* first, put on free_inc and mark as owned by allocator: */ - percpu_down_read(&c->mark_lock); - spin_lock(&c->freelist_lock); + ret = bkey_err(alloc_k); + if (ret) + return ret; - verify_not_on_freelist(c, ca, b); + a = bch2_alloc_to_v4(alloc_k, &a_convert); - BUG_ON(!fifo_push(&ca->free_inc, b)); + if (a->data_type != BCH_DATA_cached) + return 0; - bch2_mark_alloc_bucket(c, ca, b, true, gc_pos_alloc(c, NULL), 0); + 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; - spin_unlock(&c->freelist_lock); - percpu_up_read(&c->mark_lock); + 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)) || + 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))) { + u64 read_time = a->io_time[READ] ?: + atomic64_read(&c->io_clock[READ].now); + + ret = bch2_lru_set(trans, + alloc_k.k->p.inode, + bucket_to_u64(alloc_k.k->p), + read_time); + if (ret) + goto err; + + if (a->io_time[READ] != read_time) { + 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] = read_time; + ret = bch2_trans_update(trans, alloc_iter, + &a_mut->k_i, BTREE_TRIGGER_NORUN); + if (ret) + goto err; + } + } +err: +fsck_err: + bch2_trans_iter_exit(trans, &lru_iter); + printbuf_exit(&buf); + return ret; +} - BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8); +int bch2_check_alloc_to_lru_refs(struct bch_fs *c) +{ + struct btree_iter iter; + struct bkey_s_c k; + int ret = 0; - bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b)); -retry: - k = bch2_btree_iter_peek_slot(iter); - ret = bkey_err(k); + 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|BCH_TRANS_COMMIT_lazy_rw, + bch2_check_alloc_to_lru_ref(trans, &iter))); if (ret) - return ret; + bch_err_fn(c, ret); + return ret; +} - /* - * The allocator has to start before journal replay is finished - thus, - * we have to trust the in memory bucket @m, not the version in the - * btree: - */ - percpu_down_read(&c->mark_lock); - g = bucket(ca, b); - m = READ_ONCE(g->mark); - u = alloc_mem_to_key(g, m); - percpu_up_read(&c->mark_lock); +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; - invalidating_cached_data = m.cached_sectors != 0; + 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; + } - u.gen++; - u.data_type = 0; - u.dirty_sectors = 0; - u.cached_sectors = 0; - u.read_time = c->bucket_clock[READ].hand; - u.write_time = c->bucket_clock[WRITE].hand; + if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) { + (*open)++; + goto out; + } - a = bkey_alloc_init(&alloc_key.k); - a->k.p = iter->pos; - bch2_alloc_pack(a, u); + 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; + } - bch2_trans_update(trans, iter, &a->k_i); + 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; - /* - * XXX: - * when using deferred btree updates, we have journal reclaim doing - * btree updates and thus requiring the allocator to make forward - * progress, and here the allocator is requiring space in the journal - - * so we need a journal pre-reservation: - */ - ret = bch2_trans_commit(trans, NULL, - invalidating_cached_data ? journal_seq : NULL, - BTREE_INSERT_NOUNLOCK| - BTREE_INSERT_NOCHECK_RW| - BTREE_INSERT_NOFAIL| - BTREE_INSERT_USE_RESERVE| - BTREE_INSERT_USE_ALLOC_RESERVE| - BTREE_INSERT_BUCKET_INVALIDATE| - flags); - if (ret == -EINTR) - goto retry; - - if (!ret) { - /* remove from alloc_heap: */ - struct alloc_heap_entry e, *top = ca->alloc_heap.data; - - top->bucket++; - top->nr--; - - if (!top->nr) - heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); + a = bch2_alloc_to_v4_mut(trans, k); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + goto out; - /* - * Make sure we flush the last journal entry that updated this - * bucket (i.e. deleting the last reference) before writing to - * this bucket again: - */ - *journal_seq = max(*journal_seq, bucket_journal_seq(c, m)); - } else { - size_t b2; + if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) { + a->v.gen++; + SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false); + goto write; + } - /* remove from free_inc: */ - percpu_down_read(&c->mark_lock); - spin_lock(&c->freelist_lock); + 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; + } - bch2_mark_alloc_bucket(c, ca, b, false, - gc_pos_alloc(c, NULL), 0); + 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; + } - BUG_ON(!fifo_pop_back(&ca->free_inc, b2)); - BUG_ON(b != b2); + goto out; + } - spin_unlock(&c->freelist_lock); - percpu_up_read(&c->mark_lock); + 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; + + this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]); + (*discarded)++; +out: + (*seen)++; + bch2_trans_iter_exit(trans, &iter); + percpu_ref_put(&ca->io_ref); + printbuf_exit(&buf); return ret; } -static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca, - size_t bucket, u64 *flush_seq) +static void bch2_do_discards_work(struct work_struct *work) { - struct bucket_mark m; - - percpu_down_read(&c->mark_lock); - spin_lock(&c->freelist_lock); + 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; - bch2_invalidate_bucket(c, ca, bucket, &m); + /* + * 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_key2(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)); +} - verify_not_on_freelist(c, ca, bucket); - BUG_ON(!fifo_push(&ca->free_inc, bucket)); +void bch2_do_discards(struct bch_fs *c) +{ + 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); +} - spin_unlock(&c->freelist_lock); +static int invalidate_one_bucket(struct btree_trans *trans, + struct btree_iter *lru_iter, + struct bkey_s_c lru_k, + s64 *nr_to_invalidate) +{ + 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; - bucket_io_clock_reset(c, ca, bucket, READ); - bucket_io_clock_reset(c, ca, bucket, WRITE); + if (*nr_to_invalidate <= 0) + return 1; - percpu_up_read(&c->mark_lock); + if (!bch2_dev_bucket_exists(c, bucket)) { + prt_str(&buf, "lru entry points to invalid bucket"); + goto err; + } - *flush_seq = max(*flush_seq, bucket_journal_seq(c, m)); + if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset)) + return 0; - return m.cached_sectors != 0; -} + a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + goto out; -/* - * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc: - */ -static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca) -{ - struct btree_trans trans; - struct btree_iter *iter; - u64 journal_seq = 0; - int ret = 0; + /* 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; - bch2_trans_init(&trans, c, 0, 0); + BUG_ON(a->v.data_type != BCH_DATA_cached); - iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, - POS(ca->dev_idx, 0), - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); + if (!a->v.cached_sectors) + bch_err(c, "invalidating empty bucket, confused"); - /* Only use nowait if we've already invalidated at least one bucket: */ - while (!ret && - !fifo_full(&ca->free_inc) && - ca->alloc_heap.used) - ret = bch2_invalidate_one_bucket2(&trans, ca, iter, &journal_seq, - BTREE_INSERT_GC_LOCK_HELD| - (!fifo_empty(&ca->free_inc) - ? BTREE_INSERT_NOWAIT : 0)); + cached_sectors = a->v.cached_sectors; - bch2_trans_exit(&trans); + 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); - /* If we used NOWAIT, don't return the error: */ - if (!fifo_empty(&ca->free_inc)) - ret = 0; - if (ret) { - bch_err(ca, "error invalidating buckets: %i", ret); - return ret; - } + 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; - if (journal_seq) - ret = bch2_journal_flush_seq(&c->journal, journal_seq); - if (ret) { - bch_err(ca, "journal error: %i", ret); - return ret; + 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; } - return 0; + goto out; } -static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket) +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; - while (1) { - set_current_state(TASK_INTERRUPTIBLE); - - spin_lock(&c->freelist_lock); - for (i = 0; i < RESERVE_NR; i++) - if (fifo_push(&ca->free[i], bucket)) { - fifo_pop(&ca->free_inc, bucket); - - closure_wake_up(&c->freelist_wait); - ca->allocator_state = ALLOCATOR_RUNNING; - - spin_unlock(&c->freelist_lock); - goto out; - } + ret = bch2_btree_write_buffer_flush(trans); + if (ret) + goto err; - if (ca->allocator_state != ALLOCATOR_BLOCKED_FULL) { - ca->allocator_state = ALLOCATOR_BLOCKED_FULL; - closure_wake_up(&c->freelist_wait); - } + for_each_member_device(ca, c, i) { + s64 nr_to_invalidate = + should_invalidate_buckets(ca, bch2_dev_usage_read(ca)); - spin_unlock(&c->freelist_lock); + 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 ((current->flags & PF_KTHREAD) && - kthread_should_stop()) { - ret = 1; + if (ret < 0) { + percpu_ref_put(&ca->ref); break; } - - schedule(); - try_to_freeze(); } -out: - __set_current_state(TASK_RUNNING); - return ret; +err: + bch2_trans_put(trans); + bch2_write_ref_put(c, BCH_WRITE_REF_invalidate); } -/* - * Pulls buckets off free_inc, discards them (if enabled), then adds them to - * freelists, waiting until there's room if necessary: - */ -static int discard_invalidated_buckets(struct bch_fs *c, struct bch_dev *ca) +void bch2_do_invalidates(struct bch_fs *c) { - while (!fifo_empty(&ca->free_inc)) { - size_t bucket = fifo_peek(&ca->free_inc); - - 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, bucket), - ca->mi.bucket_size, GFP_NOIO, 0); - - if (push_invalidated_bucket(c, ca, bucket)) - return 1; - } - - return 0; + 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); } -/** - * 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) +int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca, + u64 bucket_start, u64 bucket_end) { - struct bch_dev *ca = arg; - struct bch_fs *c = ca->fs; - size_t nr; + 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(); - ca->allocator_state = ALLOCATOR_RUNNING; + 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) { - cond_resched(); + 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); - pr_debug("discarding %zu invalidated buckets", - fifo_used(&ca->free_inc)); + if (bkey_ge(iter.pos, end)) { + ret = 0; + break; + } - ret = discard_invalidated_buckets(c, ca); + k = bch2_get_key_or_hole(&iter, end, &hole); + ret = bkey_err(k); if (ret) - goto stop; + goto bkey_err; - down_read(&c->gc_lock); + 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_invalidate_buckets(c, ca); - if (ret) { - up_read(&c->gc_lock); - goto stop; - } + ret = bch2_bucket_do_index(trans, k, a, true) ?: + bch2_trans_commit(trans, NULL, NULL, + BCH_TRANS_COMMIT_lazy_rw| + BCH_TRANS_COMMIT_no_enospc); + if (ret) + goto bkey_err; - if (!fifo_empty(&ca->free_inc)) { - up_read(&c->gc_lock); - continue; + bch2_btree_iter_advance(&iter); + } else { + struct bkey_i *freespace; + + freespace = bch2_trans_kmalloc(trans, sizeof(*freespace)); + ret = PTR_ERR_OR_ZERO(freespace); + if (ret) + goto bkey_err; + + bkey_init(&freespace->k); + freespace->k.type = KEY_TYPE_set; + freespace->k.p = k.k->p; + freespace->k.size = k.k->size; + + ret = bch2_btree_insert_trans(trans, BTREE_ID_freespace, freespace, 0) ?: + bch2_trans_commit(trans, NULL, NULL, + BCH_TRANS_COMMIT_lazy_rw| + BCH_TRANS_COMMIT_no_enospc); + if (ret) + goto bkey_err; + + bch2_btree_iter_set_pos(&iter, k.k->p); } +bkey_err: + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + continue; + if (ret) + break; + } - pr_debug("free_inc now empty"); - - do { - /* - * Find some buckets that we can invalidate, either - * they're completely unused, or only contain clean data - * that's been written back to the backing device or - * another cache tier - */ + bch2_trans_iter_exit(trans, &iter); + bch2_trans_put(trans); - pr_debug("scanning for reclaimable buckets"); + if (ret < 0) { + bch_err_msg(ca, ret, "initializing free space"); + return ret; + } - nr = find_reclaimable_buckets(c, ca); + 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); - pr_debug("found %zu buckets", nr); + return 0; +} - trace_alloc_batch(ca, nr, ca->alloc_heap.size); +int bch2_fs_freespace_init(struct bch_fs *c) +{ + struct bch_dev *ca; + unsigned i; + int ret = 0; + bool doing_init = false; - 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); - } + /* + * We can crash during the device add path, so we need to check this on + * every mount: + */ - /* - * If we found any buckets, we have to invalidate them - * before we scan for more - but if we didn't find very - * many we may want to wait on more buckets being - * available so we don't spin: - */ - if (!nr || - (nr < ALLOC_SCAN_BATCH(ca) && - !fifo_empty(&ca->free[RESERVE_NONE]))) { - ret = wait_buckets_available(c, ca); - if (ret) { - up_read(&c->gc_lock); - goto stop; - } - } - } while (!nr); + for_each_member_device(ca, c, i) { + if (ca->mi.freespace_initialized) + continue; - up_read(&c->gc_lock); + if (!doing_init) { + bch_info(c, "initializing freespace"); + doing_init = true; + } - pr_debug("%zu buckets to invalidate", nr); + 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; + } + } - /* - * alloc_heap is now full of newly-invalidated buckets: next, - * write out the new bucket gens: - */ + if (doing_init) { + mutex_lock(&c->sb_lock); + bch2_write_super(c); + mutex_unlock(&c->sb_lock); + bch_verbose(c, "done initializing freespace"); } -stop: - pr_debug("alloc thread stopping (ret %i)", ret); - ca->allocator_state = ALLOCATOR_STOPPED; - closure_wake_up(&c->freelist_wait); 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) @@ -1199,12 +2011,12 @@ 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); for_each_online_member(ca, c, i) { - struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi; + struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi; ra_pages += bdi->ra_pages; } @@ -1230,8 +2042,9 @@ 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 */ @@ -1239,8 +2052,6 @@ void bch2_recalc_capacity(struct bch_fs *c) dev_reserve *= ca->mi.bucket_size; - ca->copygc_threshold = dev_reserve; - capacity += bucket_to_sector(ca, ca->mi.nbuckets - ca->mi.first_bucket); @@ -1262,22 +2073,21 @@ void bch2_recalc_capacity(struct bch_fs *c) c->bucket_size_max = bucket_size_max; - if (c->capacity) { - bch2_io_timer_add(&c->io_clock[READ], - &c->bucket_clock[READ].rescale); - bch2_io_timer_add(&c->io_clock[WRITE], - &c->bucket_clock[WRITE].rescale); - } else { - bch2_io_timer_del(&c->io_clock[READ], - &c->bucket_clock[READ].rescale); - bch2_io_timer_del(&c->io_clock[WRITE], - &c->bucket_clock[WRITE].rescale); - } - /* Wake up case someone was waiting for buckets */ 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; @@ -1288,7 +2098,7 @@ 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); } @@ -1301,8 +2111,6 @@ 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++) @@ -1313,40 +2121,7 @@ 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, &ca->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 @@ -1376,279 +2151,9 @@ 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)) - return PTR_ERR(p); - - get_task_struct(p); - rcu_assign_pointer(ca->alloc_thread, p); - wake_up_process(p); - return 0; -} - -static bool flush_held_btree_writes(struct bch_fs *c) -{ - struct bucket_table *tbl; - struct rhash_head *pos; - struct btree *b; - bool nodes_unwritten; - size_t i; -again: - cond_resched(); - nodes_unwritten = false; - - if (bch2_journal_error(&c->journal)) - return true; - - rcu_read_lock(); - for_each_cached_btree(b, c, tbl, i, pos) - if (btree_node_need_write(b)) { - if (btree_node_may_write(b)) { - rcu_read_unlock(); - btree_node_lock_type(c, b, SIX_LOCK_read); - bch2_btree_node_write(c, b, SIX_LOCK_read); - six_unlock_read(&b->lock); - goto again; - } else { - nodes_unwritten = true; - } - } - rcu_read_unlock(); - - if (c->btree_roots_dirty) { - bch2_journal_meta(&c->journal); - goto again; - } - - return !nodes_unwritten && - !bch2_btree_interior_updates_nr_pending(c); -} - -static void allocator_start_issue_discards(struct bch_fs *c) -{ - struct bch_dev *ca; - unsigned dev_iter; - size_t bu; - - for_each_rw_member(ca, c, dev_iter) - while (fifo_pop(&ca->free_inc, bu)) - blkdev_issue_discard(ca->disk_sb.bdev, - bucket_to_sector(ca, bu), - ca->mi.bucket_size, GFP_NOIO, 0); -} - -static int resize_free_inc(struct bch_dev *ca) -{ - alloc_fifo free_inc; - - if (!fifo_full(&ca->free_inc)) - return 0; - - if (!init_fifo(&free_inc, - ca->free_inc.size * 2, - GFP_KERNEL)) - return -ENOMEM; - - fifo_move(&free_inc, &ca->free_inc); - swap(free_inc, ca->free_inc); - free_fifo(&free_inc); - return 0; -} - -static bool bch2_fs_allocator_start_fast(struct bch_fs *c) -{ - struct bch_dev *ca; - unsigned dev_iter; - bool ret = true; - - if (test_alloc_startup(c)) - return false; - - down_read(&c->gc_lock); - - /* Scan for buckets that are already invalidated: */ - for_each_rw_member(ca, c, dev_iter) { - struct bucket_array *buckets; - struct bucket_mark m; - long bu; - - down_read(&ca->bucket_lock); - buckets = bucket_array(ca); - - for (bu = buckets->first_bucket; - bu < buckets->nbuckets; bu++) { - m = READ_ONCE(buckets->b[bu].mark); - - if (!buckets->b[bu].gen_valid || - !is_available_bucket(m) || - m.cached_sectors || - (ca->buckets_nouse && - test_bit(bu, ca->buckets_nouse))) - continue; - - percpu_down_read(&c->mark_lock); - bch2_mark_alloc_bucket(c, ca, bu, true, - gc_pos_alloc(c, NULL), 0); - percpu_up_read(&c->mark_lock); - - fifo_push(&ca->free_inc, bu); - - discard_invalidated_buckets(c, ca); - - if (fifo_full(&ca->free[RESERVE_BTREE])) - break; - } - up_read(&ca->bucket_lock); - } - - up_read(&c->gc_lock); - - /* did we find enough buckets? */ - for_each_rw_member(ca, c, dev_iter) - if (!fifo_full(&ca->free[RESERVE_BTREE])) - ret = false; - - return ret; -} - -int bch2_fs_allocator_start(struct bch_fs *c) -{ - struct bch_dev *ca; - unsigned dev_iter; - u64 journal_seq = 0; - bool wrote; - long bu; - int ret = 0; - - if (!test_alloc_startup(c) && - bch2_fs_allocator_start_fast(c)) - return 0; - - pr_debug("not enough empty buckets; scanning for reclaimable buckets"); - - /* - * We're moving buckets to freelists _before_ they've been marked as - * invalidated on disk - we have to so that we can allocate new btree - * nodes to mark them as invalidated on disk. - * - * However, we can't _write_ to any of these buckets yet - they might - * have cached data in them, which is live until they're marked as - * invalidated on disk: - */ - set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags); - - down_read(&c->gc_lock); - do { - wrote = false; - - for_each_rw_member(ca, c, dev_iter) { - find_reclaimable_buckets(c, ca); - - while (!fifo_full(&ca->free[RESERVE_BTREE]) && - (bu = next_alloc_bucket(ca)) >= 0) { - ret = resize_free_inc(ca); - if (ret) { - percpu_ref_put(&ca->io_ref); - up_read(&c->gc_lock); - goto err; - } - - bch2_invalidate_one_bucket(c, ca, bu, - &journal_seq); - - fifo_push(&ca->free[RESERVE_BTREE], bu); - } - } - - pr_debug("done scanning for reclaimable buckets"); - - /* - * XXX: it's possible for this to deadlock waiting on journal reclaim, - * since we're holding btree writes. What then? - */ - ret = bch2_alloc_write(c, - BTREE_INSERT_NOCHECK_RW| - BTREE_INSERT_USE_ALLOC_RESERVE| - BTREE_INSERT_NOWAIT, &wrote); - - /* - * If bch2_alloc_write() did anything, it may have used some - * buckets, and we need the RESERVE_BTREE freelist full - so we - * need to loop and scan again. - * And if it errored, it may have been because there weren't - * enough buckets, so just scan and loop again as long as it - * made some progress: - */ - } while (wrote); - up_read(&c->gc_lock); - - if (ret) - goto err; - - pr_debug("flushing journal"); - - ret = bch2_journal_flush(&c->journal); - if (ret) - goto err; - - pr_debug("issuing discards"); - allocator_start_issue_discards(c); -err: - clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags); - closure_wait_event(&c->btree_interior_update_wait, - flush_held_btree_writes(c)); - - return ret; -} - void bch2_fs_allocator_background_init(struct bch_fs *c) { spin_lock_init(&c->freelist_lock); - bch2_bucket_clock_init(c, READ); - bch2_bucket_clock_init(c, WRITE); - - c->pd_controllers_update_seconds = 5; - INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update); + INIT_WORK(&c->discard_work, bch2_do_discards_work); + INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work); }