+// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
#include "alloc_background.h"
#include "alloc_foreground.h"
#include "buckets.h"
#include "clock.h"
#include "debug.h"
+#include "ec.h"
#include "error.h"
-#include "journal_io.h"
+#include "recovery.h"
#include <linux/kthread.h>
#include <linux/math64.h>
#include <linux/sort.h>
#include <trace/events/bcachefs.h>
+static const char * const bch2_alloc_field_names[] = {
+#define x(name, bytes) #name,
+ BCH_ALLOC_FIELDS()
+#undef x
+ NULL
+};
+
static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int);
/* Ratelimiting/PD controllers */
/* Persistent alloc info: */
-static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
-{
- unsigned bytes = offsetof(struct bch_alloc, data);
-
- if (a->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
- bytes += 2;
- if (a->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
- bytes += 2;
-
- return DIV_ROUND_UP(bytes, sizeof(u64));
-}
-
-const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k)
-{
- if (k.k->p.inode >= c->sb.nr_devices ||
- !c->devs[k.k->p.inode])
- return "invalid device";
-
- switch (k.k->type) {
- case BCH_ALLOC: {
- struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
-
- if (bch_alloc_val_u64s(a.v) != bkey_val_u64s(a.k))
- return "incorrect value size";
- break;
- }
- default:
- return "invalid type";
- }
-
- return NULL;
-}
-
-void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
- struct bkey_s_c k)
+static inline u64 get_alloc_field(const struct bch_alloc *a,
+ const void **p, unsigned field)
{
- switch (k.k->type) {
- case BCH_ALLOC: {
- struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
-
- pr_buf(out, "gen %u", a.v->gen);
- break;
- }
- }
-}
+ unsigned bytes = BCH_ALLOC_FIELD_BYTES[field];
+ u64 v;
-static inline unsigned get_alloc_field(const u8 **p, unsigned bytes)
-{
- unsigned v;
+ if (!(a->fields & (1 << field)))
+ return 0;
switch (bytes) {
case 1:
- v = **p;
+ v = *((const u8 *) *p);
break;
case 2:
- v = le16_to_cpup((void *) *p);
+ v = le16_to_cpup(*p);
break;
case 4:
- v = le32_to_cpup((void *) *p);
+ v = le32_to_cpup(*p);
+ break;
+ case 8:
+ v = le64_to_cpup(*p);
break;
default:
BUG();
return v;
}
-static inline void put_alloc_field(u8 **p, unsigned bytes, unsigned v)
+static inline void put_alloc_field(struct bkey_i_alloc *a, void **p,
+ unsigned field, u64 v)
{
+ unsigned bytes = BCH_ALLOC_FIELD_BYTES[field];
+
+ if (!v)
+ return;
+
+ a->v.fields |= 1 << field;
+
switch (bytes) {
case 1:
- **p = v;
+ *((u8 *) *p) = v;
break;
case 2:
*((__le16 *) *p) = cpu_to_le16(v);
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_read_key(struct bch_fs *c, struct bkey_s_c k)
+struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
{
- struct bch_dev *ca;
- struct bkey_s_c_alloc a;
- struct bucket_mark new;
- struct bucket *g;
- const u8 *d;
+ struct bkey_alloc_unpacked ret = { .gen = 0 };
- if (k.k->type != BCH_ALLOC)
- return;
+ 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;
- a = bkey_s_c_to_alloc(k);
- ca = bch_dev_bkey_exists(c, a.k->p.inode);
+ ret.gen = a->gen;
- if (a.k->p.offset >= ca->mi.nbuckets)
- return;
+#define x(_name, _bits) ret._name = get_alloc_field(a, &d, idx++);
+ BCH_ALLOC_FIELDS()
+#undef x
+ }
+ return ret;
+}
+
+void bch2_alloc_pack(struct bkey_i_alloc *dst,
+ const struct bkey_alloc_unpacked src)
+{
+ unsigned idx = 0;
+ void *d = dst->v.data;
+ unsigned bytes;
+
+ dst->v.fields = 0;
+ dst->v.gen = src.gen;
+
+#define x(_name, _bits) put_alloc_field(dst, &d, idx++, src._name);
+ BCH_ALLOC_FIELDS()
+#undef x
+
+ bytes = (void *) d - (void *) &dst->v;
+ set_bkey_val_bytes(&dst->k, bytes);
+ memset_u64s_tail(&dst->v, 0, bytes);
+}
+
+static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
+{
+ unsigned i, bytes = offsetof(struct bch_alloc, data);
+
+ for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_FIELD_BYTES); i++)
+ if (a->fields & (1 << i))
+ bytes += BCH_ALLOC_FIELD_BYTES[i];
+
+ return DIV_ROUND_UP(bytes, sizeof(u64));
+}
+
+const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k)
+{
+ 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";
- percpu_down_read_preempt_disable(&c->usage_lock);
+ /* allow for unknown fields */
+ if (bkey_val_u64s(a.k) < bch_alloc_val_u64s(a.v))
+ return "incorrect value size";
- g = bucket(ca, a.k->p.offset);
- bucket_cmpxchg(g, new, ({
- new.gen = a.v->gen;
- new.gen_valid = 1;
- }));
+ return NULL;
+}
+
+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);
+ const void *d = a.v->data;
+ unsigned i;
- d = a.v->data;
- if (a.v->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
- g->io_time[READ] = get_alloc_field(&d, 2);
- if (a.v->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
- g->io_time[WRITE] = get_alloc_field(&d, 2);
+ pr_buf(out, "gen %u", a.v->gen);
- percpu_up_read_preempt_enable(&c->usage_lock);
+ 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));
}
-int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
+int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys)
{
- struct journal_replay *r;
- struct btree_iter iter;
+ struct btree_trans trans;
+ struct btree_and_journal_iter iter;
struct bkey_s_c k;
struct bch_dev *ca;
unsigned i;
- int ret;
+ int ret = 0;
- for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS_MIN, 0, k) {
- bch2_alloc_read_key(c, k);
- bch2_btree_iter_cond_resched(&iter);
- }
+ bch2_trans_init(&trans, c, 0, 0);
- ret = bch2_btree_iter_unlock(&iter);
- if (ret)
- return ret;
+ bch2_btree_and_journal_iter_init(&iter, &trans, journal_keys,
+ BTREE_ID_ALLOC, POS_MIN);
- list_for_each_entry(r, journal_replay_list, list) {
- struct bkey_i *k, *n;
- struct jset_entry *entry;
+ while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
+ bch2_mark_key(c, k, 0, 0, NULL, 0,
+ BTREE_TRIGGER_ALLOC_READ|
+ BTREE_TRIGGER_NOATOMIC);
- for_each_jset_key(k, n, entry, &r->j)
- if (entry->btree_id == BTREE_ID_ALLOC)
- bch2_alloc_read_key(c, bkey_i_to_s_c(k));
+ bch2_btree_and_journal_iter_advance(&iter);
}
+ ret = bch2_trans_exit(&trans) ?: ret;
+ if (ret) {
+ bch_err(c, "error reading alloc info: %i", ret);
+ return ret;
+ }
+
+ percpu_down_write(&c->mark_lock);
+ bch2_dev_usage_from_buckets(c);
+ percpu_up_write(&c->mark_lock);
+
mutex_lock(&c->bucket_clock[READ].lock);
for_each_member_device(ca, c, i) {
down_read(&ca->bucket_lock);
return 0;
}
-static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
- size_t b, struct btree_iter *iter,
- u64 *journal_seq, unsigned flags)
+enum alloc_write_ret {
+ ALLOC_WROTE,
+ ALLOC_NOWROTE,
+ ALLOC_END,
+};
+
+static int bch2_alloc_write_key(struct btree_trans *trans,
+ struct btree_iter *iter,
+ unsigned flags)
{
- struct bucket_mark m;
- __BKEY_PADDED(k, DIV_ROUND_UP(sizeof(struct bch_alloc), 8)) alloc_key;
+ 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;
- u8 *d;
-
- percpu_down_read_preempt_disable(&c->usage_lock);
- g = bucket(ca, b);
+ int ret;
+retry:
+ k = bch2_btree_iter_peek_slot(iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
- m = READ_ONCE(g->mark);
- a = bkey_alloc_init(&alloc_key.k);
- a->k.p = POS(ca->dev_idx, b);
- a->v.fields = 0;
- a->v.gen = m.gen;
- set_bkey_val_u64s(&a->k, bch_alloc_val_u64s(&a->v));
-
- d = a->v.data;
- if (a->v.fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
- put_alloc_field(&d, 2, g->io_time[READ]);
- if (a->v.fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
- put_alloc_field(&d, 2, g->io_time[WRITE]);
- percpu_up_read_preempt_enable(&c->usage_lock);
-
- bch2_btree_iter_cond_resched(iter);
-
- bch2_btree_iter_set_pos(iter, a->k.p);
-
- return bch2_btree_insert_at(c, NULL, journal_seq,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|
- BTREE_INSERT_USE_ALLOC_RESERVE|
- flags,
- BTREE_INSERT_ENTRY(iter, &a->k_i));
-}
+ old_u = bch2_alloc_unpack(k);
-int bch2_alloc_replay_key(struct bch_fs *c, struct bpos pos)
-{
- struct bch_dev *ca;
- struct btree_iter iter;
- int ret;
+ if (iter->pos.inode >= c->sb.nr_devices ||
+ !c->devs[iter->pos.inode])
+ return ALLOC_END;
- if (pos.inode >= c->sb.nr_devices || !c->devs[pos.inode])
- return 0;
+ percpu_down_read(&c->mark_lock);
+ ca = bch_dev_bkey_exists(c, iter->pos.inode);
+ ba = bucket_array(ca);
- ca = bch_dev_bkey_exists(c, pos.inode);
+ if (iter->pos.offset >= ba->nbuckets) {
+ percpu_up_read(&c->mark_lock);
+ return ALLOC_END;
+ }
- if (pos.offset >= ca->mi.nbuckets)
- return 0;
+ 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);
- bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
- BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+ if (!bkey_alloc_unpacked_cmp(old_u, new_u))
+ return ALLOC_NOWROTE;
- ret = __bch2_alloc_write_key(c, ca, pos.offset, &iter, NULL, 0);
- bch2_btree_iter_unlock(&iter);
+ a = bkey_alloc_init(&alloc_key.k);
+ a->k.p = iter->pos;
+ bch2_alloc_pack(a, new_u);
+
+ bch2_trans_update(trans, iter, &a->k_i,
+ BTREE_TRIGGER_NORUN);
+ ret = bch2_trans_commit(trans, NULL, NULL,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_USE_RESERVE|
+ flags);
+err:
+ if (ret == -EINTR)
+ goto retry;
return ret;
}
-int bch2_alloc_write(struct bch_fs *c)
+int bch2_alloc_write(struct bch_fs *c, unsigned flags, bool *wrote)
{
+ struct btree_trans trans;
+ struct btree_iter *iter;
struct bch_dev *ca;
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_each_rw_member(ca, c, i) {
- struct btree_iter iter;
- unsigned long bucket;
+ unsigned first_bucket;
- bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
- BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+ percpu_down_read(&c->mark_lock);
+ first_bucket = bucket_array(ca)->first_bucket;
+ percpu_up_read(&c->mark_lock);
- down_read(&ca->bucket_lock);
- for_each_set_bit(bucket, ca->buckets_dirty, ca->mi.nbuckets) {
- ret = __bch2_alloc_write_key(c, ca, bucket,
- &iter, NULL, 0);
- if (ret)
- break;
+ bch2_btree_iter_set_pos(iter, POS(i, first_bucket));
- clear_bit(bucket, ca->buckets_dirty);
+ while (1) {
+ ret = bch2_alloc_write_key(&trans, iter, flags);
+ if (ret < 0 || ret == ALLOC_END)
+ break;
+ if (ret == ALLOC_WROTE)
+ *wrote = true;
+ bch2_btree_iter_next_slot(iter);
}
- up_read(&ca->bucket_lock);
- bch2_btree_iter_unlock(&iter);
- if (ret) {
+ if (ret < 0) {
percpu_ref_put(&ca->io_ref);
break;
}
}
- return ret;
+ bch2_trans_exit(&trans);
+
+ return ret < 0 ? ret : 0;
+}
+
+int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
+{
+ struct btree_trans trans;
+ struct btree_iter *iter;
+ int ret;
+
+ bch2_trans_init(&trans, c, 0, 0);
+
+ iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, k->k.p,
+ BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+
+ ret = bch2_alloc_write_key(&trans, iter,
+ BTREE_INSERT_NOFAIL|
+ BTREE_INSERT_LAZY_RW|
+ BTREE_INSERT_JOURNAL_REPLAY);
+ bch2_trans_exit(&trans);
+ return ret < 0 ? ret : 0;
}
/* Bucket IO clocks: */
unsigned long gc_count = c->gc_count;
int ret = 0;
+ ca->allocator_state = ALLOCATOR_BLOCKED;
+ closure_wake_up(&c->freelist_wait);
+
while (1) {
set_current_state(TASK_INTERRUPTIBLE);
if (kthread_should_stop()) {
}
__set_current_state(TASK_RUNNING);
+ ca->allocator_state = ALLOCATOR_RUNNING;
+ closure_wake_up(&c->freelist_wait);
+
return ret;
}
if (!is_available_bucket(mark))
return false;
+ if (ca->buckets_nouse &&
+ test_bit(bucket, ca->buckets_nouse))
+ return false;
+
gc_gen = bucket_gc_gen(ca, bucket);
if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
struct alloc_heap_entry l,
struct alloc_heap_entry r)
{
- return (l.key > r.key) - (l.key < r.key) ?:
- (l.nr < r.nr) - (l.nr > r.nr) ?:
- (l.bucket > r.bucket) - (l.bucket < r.bucket);
+ return cmp_int(l.key, r.key) ?:
+ cmp_int(r.nr, l.nr) ?:
+ cmp_int(l.bucket, r.bucket);
}
static inline int bucket_idx_cmp(const void *_l, const void *_r)
{
const struct alloc_heap_entry *l = _l, *r = _r;
- return (l->bucket > r->bucket) - (l->bucket < r->bucket);
+ return cmp_int(l->bucket, r->bucket);
}
static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
return -1;
}
+/*
+ * 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;
+
+ return bucket_seq;
+ } else {
+ return 0;
+ }
+}
+
+static int bch2_invalidate_one_bucket2(struct btree_trans *trans,
+ struct bch_dev *ca,
+ struct btree_iter *iter,
+ u64 *journal_seq, unsigned flags)
+{
+#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;
+ int ret;
+
+ BUG_ON(!ca->alloc_heap.used ||
+ !ca->alloc_heap.data[0].nr);
+ b = ca->alloc_heap.data[0].bucket;
+
+ /* first, put on free_inc and mark as owned by allocator: */
+ percpu_down_read(&c->mark_lock);
+ spin_lock(&c->freelist_lock);
+
+ verify_not_on_freelist(c, ca, b);
+
+ BUG_ON(!fifo_push(&ca->free_inc, b));
+
+ bch2_mark_alloc_bucket(c, ca, b, true, gc_pos_alloc(c, NULL), 0);
+
+ spin_unlock(&c->freelist_lock);
+ percpu_up_read(&c->mark_lock);
+
+ BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
+
+ bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
+retry:
+ k = bch2_btree_iter_peek_slot(iter);
+ ret = bkey_err(k);
+ if (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);
+
+ invalidating_cached_data = m.cached_sectors != 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;
+
+ a = bkey_alloc_init(&alloc_key.k);
+ a->k.p = iter->pos;
+ bch2_alloc_pack(a, u);
+
+ bch2_trans_update(trans, iter, &a->k_i,
+ BTREE_TRIGGER_BUCKET_INVALIDATE);
+
+ /*
+ * 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|
+ 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);
+
+ /*
+ * 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;
+
+ /* remove from free_inc: */
+ percpu_down_read(&c->mark_lock);
+ spin_lock(&c->freelist_lock);
+
+ bch2_mark_alloc_bucket(c, ca, b, false,
+ gc_pos_alloc(c, NULL), 0);
+
+ BUG_ON(!fifo_pop_back(&ca->free_inc, b2));
+ BUG_ON(b != b2);
+
+ spin_unlock(&c->freelist_lock);
+ percpu_up_read(&c->mark_lock);
+ }
+
+ return ret;
+}
+
static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
size_t bucket, u64 *flush_seq)
{
struct bucket_mark m;
- percpu_down_read_preempt_disable(&c->usage_lock);
+ percpu_down_read(&c->mark_lock);
spin_lock(&c->freelist_lock);
bch2_invalidate_bucket(c, ca, bucket, &m);
bucket_io_clock_reset(c, ca, bucket, READ);
bucket_io_clock_reset(c, ca, bucket, WRITE);
- percpu_up_read_preempt_enable(&c->usage_lock);
-
- 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;
+ percpu_up_read(&c->mark_lock);
- if (bucket_seq > journal_seq)
- bucket_seq -= 1 << 16;
-
- *flush_seq = max(*flush_seq, bucket_seq);
- }
+ *flush_seq = max(*flush_seq, bucket_journal_seq(c, m));
return m.cached_sectors != 0;
}
*/
static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
{
- struct btree_iter iter;
+ struct btree_trans trans;
+ struct btree_iter *iter;
u64 journal_seq = 0;
int ret = 0;
- long b;
- bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0),
- BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+ bch2_trans_init(&trans, c, 0, 0);
+
+ iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC,
+ POS(ca->dev_idx, 0),
+ BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
/* Only use nowait if we've already invalidated at least one bucket: */
while (!ret &&
!fifo_full(&ca->free_inc) &&
- (b = next_alloc_bucket(ca)) >= 0) {
- bool must_flush =
- bch2_invalidate_one_bucket(c, ca, b, &journal_seq);
-
- ret = __bch2_alloc_write_key(c, ca, b, &iter,
- must_flush ? &journal_seq : NULL,
- !fifo_empty(&ca->free_inc) ? BTREE_INSERT_NOWAIT : 0);
- }
+ 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));
- bch2_btree_iter_unlock(&iter);
+ bch2_trans_exit(&trans);
/* If we used NOWAIT, don't return the error: */
if (!fifo_empty(&ca->free_inc))
set_current_state(TASK_INTERRUPTIBLE);
spin_lock(&c->freelist_lock);
- for (i = 0; i < RESERVE_NR; i++)
+ for (i = 0; i < RESERVE_NR; i++) {
+
+ /*
+ * Don't strand buckets on the copygc freelist until
+ * after recovery is finished:
+ */
+ if (!test_bit(BCH_FS_STARTED, &c->flags) &&
+ i == RESERVE_MOVINGGC)
+ continue;
+
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;
}
+ }
+
+ if (ca->allocator_state != ALLOCATOR_BLOCKED_FULL) {
+ ca->allocator_state = ALLOCATOR_BLOCKED_FULL;
+ closure_wake_up(&c->freelist_wait);
+ }
+
spin_unlock(&c->freelist_lock);
if ((current->flags & PF_KTHREAD) &&
int ret;
set_freezable();
+ ca->allocator_state = ALLOCATOR_RUNNING;
while (1) {
cond_resched();
pr_debug("free_inc now empty");
do {
- if (test_bit(BCH_FS_GC_FAILURE, &c->flags)) {
- up_read(&c->gc_lock);
- bch_err(ca, "gc failure");
- goto stop;
- }
-
/*
* Find some buckets that we can invalidate, either
* they're completely unused, or only contain clean data
*/
if (!nr ||
(nr < ALLOC_SCAN_BATCH(ca) &&
- !fifo_full(&ca->free[RESERVE_MOVINGGC]))) {
- ca->allocator_blocked = true;
- closure_wake_up(&c->freelist_wait);
-
+ !fifo_empty(&ca->free[RESERVE_NONE]))) {
ret = wait_buckets_available(c, ca);
if (ret) {
up_read(&c->gc_lock);
}
} while (!nr);
- ca->allocator_blocked = false;
up_read(&c->gc_lock);
pr_debug("%zu buckets to invalidate", nr);
stop:
pr_debug("alloc thread stopping (ret %i)", ret);
+ ca->allocator_state = ALLOCATOR_STOPPED;
+ closure_wake_up(&c->freelist_wait);
return 0;
}
}
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);
+
/*
* Wake up threads that were blocked on allocation, so they can notice
* the device can no longer be removed and the capacity has changed:
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)
{
return 0;
}
-static void flush_held_btree_writes(struct bch_fs *c)
+static bool flush_held_btree_writes(struct bch_fs *c)
{
struct bucket_table *tbl;
struct rhash_head *pos;
struct btree *b;
- bool flush_updates;
- size_t i, nr_pending_updates;
-
- clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
+ bool nodes_unwritten;
+ size_t i;
again:
- pr_debug("flushing dirty btree nodes");
cond_resched();
+ nodes_unwritten = false;
- flush_updates = false;
- nr_pending_updates = bch2_btree_interior_updates_nr_pending(c);
+ if (bch2_journal_error(&c->journal))
+ return true;
rcu_read_lock();
for_each_cached_btree(b, c, tbl, i, pos)
- if (btree_node_dirty(b) && (!b->written || b->level)) {
+ if (btree_node_need_write(b)) {
if (btree_node_may_write(b)) {
rcu_read_unlock();
btree_node_lock_type(c, b, SIX_LOCK_read);
six_unlock_read(&b->lock);
goto again;
} else {
- flush_updates = true;
+ nodes_unwritten = true;
}
}
rcu_read_unlock();
- if (c->btree_roots_dirty)
+ if (c->btree_roots_dirty) {
bch2_journal_meta(&c->journal);
-
- /*
- * This is ugly, but it's needed to flush btree node writes
- * without spinning...
- */
- if (flush_updates) {
- closure_wait_event(&c->btree_interior_update_wait,
- bch2_btree_interior_updates_nr_pending(c) <
- nr_pending_updates);
goto again;
}
+ return !nodes_unwritten &&
+ !bch2_btree_interior_updates_nr_pending(c);
}
static void allocator_start_issue_discards(struct bch_fs *c)
ca->mi.bucket_size, GFP_NOIO, 0);
}
-static int __bch2_fs_allocator_start(struct bch_fs *c)
+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;
- u64 journal_seq = 0;
- long bu;
- bool invalidating_data = false;
- int ret = 0;
+ bool ret = true;
- if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
- return -1;
+ if (test_alloc_startup(c))
+ return false;
- if (test_alloc_startup(c)) {
- invalidating_data = true;
- goto not_enough;
- }
+ down_read(&c->gc_lock);
/* Scan for buckets that are already invalidated: */
for_each_rw_member(ca, c, dev_iter) {
- struct btree_iter iter;
+ struct bucket_array *buckets;
struct bucket_mark m;
- struct bkey_s_c k;
+ long bu;
- for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0), 0, k) {
- if (k.k->type != BCH_ALLOC)
- continue;
+ down_read(&ca->bucket_lock);
+ buckets = bucket_array(ca);
- bu = k.k->p.offset;
- m = READ_ONCE(bucket(ca, bu)->mark);
+ for (bu = buckets->first_bucket;
+ bu < buckets->nbuckets; bu++) {
+ m = READ_ONCE(buckets->b[bu].mark);
- if (!is_available_bucket(m) || m.cached_sectors)
+ 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_preempt_disable(&c->usage_lock);
+ percpu_down_read(&c->mark_lock);
bch2_mark_alloc_bucket(c, ca, bu, true,
- gc_pos_alloc(c, NULL),
- BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
- BCH_BUCKET_MARK_GC_LOCK_HELD);
- percpu_up_read_preempt_enable(&c->usage_lock);
+ gc_pos_alloc(c, NULL), 0);
+ percpu_up_read(&c->mark_lock);
fifo_push(&ca->free_inc, bu);
- if (fifo_full(&ca->free_inc))
+ discard_invalidated_buckets(c, ca);
+
+ if (fifo_full(&ca->free[RESERVE_BTREE]))
break;
}
- bch2_btree_iter_unlock(&iter);
+ up_read(&ca->bucket_lock);
}
- /* did we find enough buckets? */
- for_each_rw_member(ca, c, dev_iter)
- if (fifo_used(&ca->free_inc) < ca->free[RESERVE_BTREE].size) {
- percpu_ref_put(&ca->io_ref);
- goto not_enough;
- }
-
- return 0;
-not_enough:
- pr_debug("did not find enough empty buckets; issuing discards");
+ up_read(&c->gc_lock);
- /* clear out free_inc, we'll be using it again below: */
+ /* did we find enough buckets? */
for_each_rw_member(ca, c, dev_iter)
- discard_invalidated_buckets(c, ca);
+ if (!fifo_full(&ca->free[RESERVE_BTREE]))
+ ret = false;
- pr_debug("scanning for reclaimable buckets");
-
- for_each_rw_member(ca, c, dev_iter) {
- find_reclaimable_buckets(c, ca);
+ return ret;
+}
- while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
- (bu = next_alloc_bucket(ca)) >= 0) {
- invalidating_data |=
- bch2_invalidate_one_bucket(c, ca, bu, &journal_seq);
+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;
- fifo_push(&ca->free[RESERVE_BTREE], bu);
- set_bit(bu, ca->buckets_dirty);
- }
- }
+ if (!test_alloc_startup(c) &&
+ bch2_fs_allocator_start_fast(c))
+ return 0;
- pr_debug("done scanning for reclaimable buckets");
+ pr_debug("not enough empty buckets; scanning for reclaimable buckets");
/*
* We're moving buckets to freelists _before_ they've been marked as
* have cached data in them, which is live until they're marked as
* invalidated on disk:
*/
- if (invalidating_data) {
- pr_debug("invalidating existing data");
- set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
- } else {
- pr_debug("issuing discards");
- allocator_start_issue_discards(c);
- }
-
- /*
- * 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);
- if (ret)
- return ret;
+ set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
- if (invalidating_data) {
- pr_debug("flushing journal");
+ down_read(&c->gc_lock);
+ do {
+ wrote = false;
- ret = bch2_journal_flush_seq(&c->journal, journal_seq);
- if (ret)
- return ret;
+ for_each_rw_member(ca, c, dev_iter) {
+ find_reclaimable_buckets(c, ca);
- pr_debug("issuing discards");
- allocator_start_issue_discards(c);
- }
+ 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;
+ }
- set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
+ bch2_invalidate_one_bucket(c, ca, bu,
+ &journal_seq);
- /* now flush dirty btree nodes: */
- if (invalidating_data)
- flush_held_btree_writes(c);
+ fifo_push(&ca->free[RESERVE_BTREE], bu);
+ }
+ }
- return 0;
-}
+ pr_debug("done scanning for reclaimable buckets");
-int bch2_fs_allocator_start(struct bch_fs *c)
-{
- struct bch_dev *ca;
- unsigned i;
- int ret;
+ /*
+ * 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);
- down_read(&c->gc_lock);
- ret = __bch2_fs_allocator_start(c);
+ /*
+ * 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)
- return ret;
+ goto err;
- for_each_rw_member(ca, c, i) {
- ret = bch2_dev_allocator_start(ca);
- if (ret) {
- percpu_ref_put(&ca->io_ref);
- return ret;
- }
- }
+ pr_debug("flushing journal");
- return bch2_alloc_write(c);
+ 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)
projects / bcachefs-tools-debian / blobdiff