+// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
#include "alloc_background.h"
#include "alloc_foreground.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 "debug.h"
#include "ec.h"
#include "error.h"
-#include "journal_io.h"
+#include "recovery.h"
#include <linux/kthread.h>
#include <linux/math64.h>
struct bch_fs,
pd_controllers_update);
struct bch_dev *ca;
+ s64 free = 0, fragmented = 0;
unsigned i;
for_each_member_device(ca, c, i) {
- struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
+ struct bch_dev_usage stats = bch2_dev_usage_read(ca);
- u64 free = bucket_to_sector(ca,
+ 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);
+ fragmented += max_t(s64, 0, (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);
}
+ bch2_pd_controller_update(&c->copygc_pd, free, fragmented, -1);
schedule_delayed_work(&c->pd_controllers_update,
c->pd_controllers_update_seconds * HZ);
}
*p += bytes;
}
+struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
+{
+ struct bkey_alloc_unpacked ret = { .gen = 0 };
+
+ 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;
+
+ ret.gen = a->gen;
+
+#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);
get_alloc_field(a.v, &d, i));
}
-static void __alloc_read_key(struct bucket *g, const struct bch_alloc *a)
-{
- const void *d = a->data;
- unsigned idx = 0;
-
- g->_mark.gen = a->gen;
- g->gen_valid = 1;
- g->io_time[READ] = get_alloc_field(a, &d, idx++);
- g->io_time[WRITE] = get_alloc_field(a, &d, idx++);
- g->_mark.data_type = get_alloc_field(a, &d, idx++);
- g->_mark.dirty_sectors = get_alloc_field(a, &d, idx++);
- g->_mark.cached_sectors = get_alloc_field(a, &d, idx++);
-}
-
-static void __alloc_write_key(struct bkey_i_alloc *a, struct bucket *g,
- struct bucket_mark m)
-{
- unsigned idx = 0;
- void *d = a->v.data;
-
- a->v.fields = 0;
- a->v.gen = m.gen;
-
- d = a->v.data;
- put_alloc_field(a, &d, idx++, g->io_time[READ]);
- put_alloc_field(a, &d, idx++, g->io_time[WRITE]);
- put_alloc_field(a, &d, idx++, m.data_type);
- put_alloc_field(a, &d, idx++, m.dirty_sectors);
- put_alloc_field(a, &d, idx++, m.cached_sectors);
-
- set_bkey_val_bytes(&a->k, (void *) d - (void *) &a->v);
-}
-
-static void bch2_alloc_read_key(struct bch_fs *c, struct bkey_s_c k)
+static int bch2_alloc_read_fn(struct bch_fs *c, enum btree_id id,
+ unsigned level, struct bkey_s_c k)
{
struct bch_dev *ca;
- struct bkey_s_c_alloc a;
-
- if (k.k->type != KEY_TYPE_alloc)
- return;
+ struct bucket *g;
+ struct bkey_alloc_unpacked u;
- a = bkey_s_c_to_alloc(k);
- ca = bch_dev_bkey_exists(c, a.k->p.inode);
+ if (level || k.k->type != KEY_TYPE_alloc)
+ return 0;
- if (a.k->p.offset >= ca->mi.nbuckets)
- return;
+ ca = bch_dev_bkey_exists(c, k.k->p.inode);
+ g = __bucket(ca, k.k->p.offset, 0);
+ u = bch2_alloc_unpack(k);
+
+ g->_mark.gen = u.gen;
+ g->_mark.data_type = u.data_type;
+ g->_mark.dirty_sectors = u.dirty_sectors;
+ g->_mark.cached_sectors = u.cached_sectors;
+ g->io_time[READ] = u.read_time;
+ g->io_time[WRITE] = u.write_time;
+ g->oldest_gen = u.oldest_gen;
+ g->gen_valid = 1;
- percpu_down_read_preempt_disable(&c->mark_lock);
- __alloc_read_key(bucket(ca, a.k->p.offset), a.v);
- percpu_up_read_preempt_enable(&c->mark_lock);
+ return 0;
}
-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 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);
- }
+ down_read(&c->gc_lock);
+ ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_ALLOC,
+ NULL, bch2_alloc_read_fn);
+ up_read(&c->gc_lock);
- ret = bch2_btree_iter_unlock(&iter);
- if (ret)
+ if (ret) {
+ bch_err(c, "error reading alloc info: %i", ret);
return ret;
-
- list_for_each_entry(r, journal_replay_list, list) {
- struct bkey_i *k, *n;
- struct jset_entry *entry;
-
- 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));
}
- for_each_member_device(ca, c, i)
- bch2_dev_usage_from_buckets(c, ca);
+ 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) {
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)
+static int bch2_alloc_write_key(struct btree_trans *trans,
+ struct btree_iter *iter,
+ unsigned flags)
{
-#if 0
- __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key;
-#else
- /* hack: */
- __BKEY_PADDED(k, 8) alloc_key;
-#endif
- struct bkey_i_alloc *a = bkey_alloc_init(&alloc_key.k);
+ 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, new;
+ struct bucket_mark m;
+ struct bkey_alloc_unpacked old_u, new_u;
+ __BKEY_PADDED(k, 8) alloc_key; /* hack: */
+ struct bkey_i_alloc *a;
int ret;
+retry:
+ bch2_trans_begin(trans);
- BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
-
- a->k.p = POS(ca->dev_idx, b);
-
- bch2_btree_iter_set_pos(iter, a->k.p);
-
- ret = bch2_btree_iter_traverse(iter);
+ ret = bch2_btree_key_cache_flush(trans,
+ BTREE_ID_ALLOC, iter->pos);
if (ret)
- return ret;
-
- percpu_down_read_preempt_disable(&c->mark_lock);
- g = bucket(ca, b);
- m = READ_ONCE(g->mark);
+ goto err;
- if (!m.dirty) {
- percpu_up_read_preempt_enable(&c->mark_lock);
- return 0;
- }
+ k = bch2_btree_iter_peek_slot(iter);
+ ret = bkey_err(k);
+ if (ret)
+ goto err;
- __alloc_write_key(a, g, m);
- percpu_up_read_preempt_enable(&c->mark_lock);
+ old_u = bch2_alloc_unpack(k);
- bch2_btree_iter_cond_resched(iter);
+ percpu_down_read(&c->mark_lock);
+ ca = bch_dev_bkey_exists(c, iter->pos.inode);
+ ba = bucket_array(ca);
- ret = bch2_btree_insert_at(c, NULL, journal_seq,
- BTREE_INSERT_NOCHECK_RW|
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_USE_RESERVE|
- BTREE_INSERT_USE_ALLOC_RESERVE|
- flags,
- BTREE_INSERT_ENTRY(iter, &a->k_i));
- if (ret)
- return ret;
+ 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);
- new = m;
- new.dirty = false;
- atomic64_cmpxchg(&g->_mark.v, m.v.counter, new.v.counter);
+ if (!bkey_alloc_unpacked_cmp(old_u, new_u))
+ return 0;
- if (ca->buckets_written)
- set_bit(b, ca->buckets_written);
+ a = bkey_alloc_init(&alloc_key.k);
+ a->k.p = iter->pos;
+ bch2_alloc_pack(a, new_u);
- return 0;
+ 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_replay_key(struct bch_fs *c, struct bkey_i *k)
+int bch2_dev_alloc_write(struct bch_fs *c, struct bch_dev *ca, unsigned flags)
{
- struct bch_dev *ca;
- struct btree_iter iter;
- int ret;
+ struct btree_trans trans;
+ struct btree_iter *iter;
+ u64 first_bucket, nbuckets;
+ int ret = 0;
- if (k->k.p.inode >= c->sb.nr_devices ||
- !c->devs[k->k.p.inode])
- return 0;
+ percpu_down_read(&c->mark_lock);
+ first_bucket = bucket_array(ca)->first_bucket;
+ nbuckets = bucket_array(ca)->nbuckets;
+ percpu_up_read(&c->mark_lock);
- ca = bch_dev_bkey_exists(c, k->k.p.inode);
+ BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
- if (k->k.p.offset >= ca->mi.nbuckets)
- return 0;
+ bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
- bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, k->k.p,
- BTREE_ITER_INTENT);
+ iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC,
+ POS(ca->dev_idx, first_bucket),
+ BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
- ret = bch2_btree_iter_traverse(&iter);
- if (ret)
- goto err;
+ while (iter->pos.offset < nbuckets) {
+ bch2_trans_cond_resched(&trans);
- /* check buckets_written with btree node locked: */
+ ret = bch2_alloc_write_key(&trans, iter, flags);
+ if (ret)
+ break;
+ bch2_btree_iter_next_slot(iter);
+ }
+
+ bch2_trans_exit(&trans);
- ret = test_bit(k->k.p.offset, ca->buckets_written)
- ? 0
- : bch2_btree_insert_at(c, NULL, NULL,
- BTREE_INSERT_NOFAIL|
- BTREE_INSERT_JOURNAL_REPLAY,
- BTREE_INSERT_ENTRY(&iter, k));
-err:
- bch2_btree_iter_unlock(&iter);
return ret;
}
-int bch2_alloc_write(struct bch_fs *c, bool nowait, bool *wrote)
+int bch2_alloc_write(struct bch_fs *c, unsigned flags)
{
struct bch_dev *ca;
unsigned i;
int ret = 0;
- *wrote = false;
-
for_each_rw_member(ca, c, i) {
- struct btree_iter iter;
- struct bucket_array *buckets;
- size_t b;
-
- bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
- BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
-
- down_read(&ca->bucket_lock);
- buckets = bucket_array(ca);
-
- for (b = buckets->first_bucket;
- b < buckets->nbuckets;
- b++) {
- if (!buckets->b[b].mark.dirty)
- continue;
-
- ret = __bch2_alloc_write_key(c, ca, b, &iter, NULL,
- nowait
- ? BTREE_INSERT_NOWAIT
- : 0);
- if (ret)
- break;
-
- *wrote = true;
- }
- up_read(&ca->bucket_lock);
- bch2_btree_iter_unlock(&iter);
-
+ bch2_dev_alloc_write(c, ca, flags);
if (ret) {
percpu_ref_put(&ca->io_ref);
break;
mutex_init(&clock->lock);
}
+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 bch_dev *ca = bch_dev_bkey_exists(c, dev);
+ struct btree_iter *iter;
+ struct bucket *g;
+ struct bkey_i_alloc *a;
+ struct bkey_alloc_unpacked u;
+ u16 *time;
+ int ret = 0;
+
+ iter = bch2_trans_get_iter(trans, BTREE_ID_ALLOC, POS(dev, bucket_nr),
+ BTREE_ITER_CACHED|
+ BTREE_ITER_CACHED_NOFILL|
+ BTREE_ITER_INTENT);
+ if (IS_ERR(iter))
+ return PTR_ERR(iter);
+
+ a = bch2_trans_kmalloc(trans, BKEY_ALLOC_U64s_MAX * 8);
+ ret = PTR_ERR_OR_ZERO(a);
+ if (ret)
+ goto out;
+
+ percpu_down_read(&c->mark_lock);
+ g = bucket(ca, bucket_nr);
+ u = alloc_mem_to_key(g, READ_ONCE(g->mark));
+ percpu_up_read(&c->mark_lock);
+
+ bkey_alloc_init(&a->k_i);
+ a->k.p = iter->pos;
+
+ time = rw == READ ? &u.read_time : &u.write_time;
+ if (*time == c->bucket_clock[rw].hand)
+ goto out;
+
+ *time = c->bucket_clock[rw].hand;
+
+ bch2_alloc_pack(a, u);
+
+ ret = bch2_trans_update(trans, iter, &a->k_i, 0) ?:
+ bch2_trans_commit(trans, NULL, NULL, 0);
+out:
+ bch2_trans_iter_put(trans, iter);
+ return ret;
+}
+
/* Background allocator thread: */
/*
* commands to the newly free buckets, then puts them on the various freelists.
*/
-#define BUCKET_GC_GEN_MAX 96U
-
/**
* wait_buckets_available - wait on reclaimable buckets
*
static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
{
unsigned long gc_count = c->gc_count;
+ u64 available;
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()) {
if (gc_count != c->gc_count)
ca->inc_gen_really_needs_gc = 0;
- if ((ssize_t) (dev_buckets_available(c, ca) -
- ca->inc_gen_really_needs_gc) >=
- (ssize_t) fifo_free(&ca->free_inc))
+ available = max_t(s64, 0, dev_buckets_available(ca) -
+ ca->inc_gen_really_needs_gc);
+
+ if (available > fifo_free(&ca->free_inc) ||
+ (available &&
+ (!fifo_full(&ca->free[RESERVE_BTREE]) ||
+ !fifo_full(&ca->free[RESERVE_MOVINGGC]))))
break;
up_read(&c->gc_lock);
}
__set_current_state(TASK_RUNNING);
+ ca->allocator_state = ALLOCATOR_RUNNING;
+ closure_wake_up(&c->freelist_wait);
+
return ret;
}
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;
}
-static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
- size_t bucket, u64 *flush_seq)
+/*
+ * 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;
+ bool invalidating_cached_data;
+ size_t b;
+ int ret = 0;
+
+ BUG_ON(!ca->alloc_heap.used ||
+ !ca->alloc_heap.data[0].nr);
+ b = ca->alloc_heap.data[0].bucket;
- percpu_down_read_preempt_disable(&c->mark_lock);
+ /* first, put on free_inc and mark as owned by allocator: */
+ percpu_down_read(&c->mark_lock);
spin_lock(&c->freelist_lock);
- bch2_invalidate_bucket(c, ca, bucket, &m);
+ verify_not_on_freelist(c, ca, b);
- verify_not_on_freelist(c, ca, bucket);
- BUG_ON(!fifo_push(&ca->free_inc, bucket));
+ BUG_ON(!fifo_push(&ca->free_inc, b));
+
+ g = bucket(ca, b);
+ m = READ_ONCE(g->mark);
+
+ invalidating_cached_data = m.cached_sectors != 0;
+
+ /*
+ * If we're not invalidating cached data, we only increment the bucket
+ * gen in memory here, the incremented gen will be updated in the btree
+ * by bch2_trans_mark_pointer():
+ */
+
+ if (!invalidating_cached_data)
+ bch2_invalidate_bucket(c, ca, b, &m);
+ else
+ 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);
- bucket_io_clock_reset(c, ca, bucket, READ);
- bucket_io_clock_reset(c, ca, bucket, WRITE);
+ if (!invalidating_cached_data)
+ goto out;
- percpu_up_read_preempt_enable(&c->mark_lock);
+ /*
+ * If the read-only path is trying to shut down, we can't be generating
+ * new btree updates:
+ */
+ if (test_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags)) {
+ ret = 1;
+ goto out;
+ }
- if (m.journal_seq_valid) {
- u64 journal_seq = atomic64_read(&c->journal.seq);
- u64 bucket_seq = journal_seq;
+ BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
- bucket_seq &= ~((u64) U16_MAX);
- bucket_seq |= m.journal_seq;
+ bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
+retry:
+ ret = bch2_btree_iter_traverse(iter);
+ if (ret)
+ return ret;
- if (bucket_seq > journal_seq)
- bucket_seq -= 1 << 16;
+ percpu_down_read(&c->mark_lock);
+ g = bucket(ca, iter->pos.offset);
+ m = READ_ONCE(g->mark);
+ u = alloc_mem_to_key(g, m);
+
+ percpu_up_read(&c->mark_lock);
+
+ invalidating_cached_data = u.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;
- *flush_seq = max(*flush_seq, bucket_seq);
+ 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;
+out:
+ 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 m.cached_sectors != 0;
+ return ret < 0 ? ret : 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_CACHED|
+ BTREE_ITER_CACHED_NOFILL|
+ 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_blocked_full = false;
+ ca->allocator_state = ALLOCATOR_RUNNING;
spin_unlock(&c->freelist_lock);
goto out;
}
+ }
- if (!ca->allocator_blocked_full) {
- ca->allocator_blocked_full = true;
+ if (ca->allocator_state != ALLOCATOR_BLOCKED_FULL) {
+ ca->allocator_state = ALLOCATOR_BLOCKED_FULL;
closure_wake_up(&c->freelist_wait);
}
int ret;
set_freezable();
+ ca->allocator_state = ALLOCATOR_RUNNING;
while (1) {
cond_resched();
+ if (kthread_should_stop())
+ break;
pr_debug("discarding %zu invalidated buckets",
fifo_used(&ca->free_inc));
*/
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;
}
void bch2_recalc_capacity(struct bch_fs *c)
{
struct bch_dev *ca;
- u64 capacity = 0, reserved_sectors = 0, gc_reserve;
+ u64 capacity = 0, reserved_sectors = 0, gc_reserve, copygc_threshold = 0;
unsigned bucket_size_max = 0;
unsigned long ra_pages = 0;
unsigned i, j;
dev_reserve *= ca->mi.bucket_size;
- ca->copygc_threshold = dev_reserve;
+ copygc_threshold += dev_reserve;
capacity += bucket_to_sector(ca, ca->mi.nbuckets -
ca->mi.first_bucket);
reserved_sectors = min(reserved_sectors, capacity);
+ c->copygc_threshold = copygc_threshold;
c->capacity = capacity - reserved_sectors;
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);
}
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->copygc_write_point);
bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
bch2_writepoint_stop(c, ca, &c->btree_write_point);
void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
{
- closure_wait_event(&c->freelist_wait, ca->allocator_blocked_full);
+ if (ca->alloc_thread)
+ closure_wait_event(&c->freelist_wait,
+ ca->allocator_state != ALLOCATOR_RUNNING);
}
/* stop allocator thread: */
return 0;
}
-static void flush_held_btree_writes(struct bch_fs *c)
-{
- struct bucket_table *tbl;
- struct rhash_head *pos;
- struct btree *b;
- bool nodes_blocked;
- size_t i;
- struct closure cl;
-
- closure_init_stack(&cl);
-
- clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
-again:
- pr_debug("flushing dirty btree nodes");
- cond_resched();
- closure_wait(&c->btree_interior_update_wait, &cl);
-
- nodes_blocked = false;
-
- 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_blocked = true;
- }
- }
- rcu_read_unlock();
-
- if (c->btree_roots_dirty)
- bch2_journal_meta(&c->journal);
-
- if (nodes_blocked) {
- closure_sync(&cl);
- goto again;
- }
-
- closure_wake_up(&c->btree_interior_update_wait);
- closure_sync(&cl);
-
- closure_wait_event(&c->btree_interior_update_wait,
- !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 int __bch2_fs_allocator_start(struct bch_fs *c)
-{
- struct bch_dev *ca;
- unsigned dev_iter;
- u64 journal_seq = 0;
- long bu;
- int ret = 0;
-
- if (test_alloc_startup(c))
- goto not_enough;
-
- /* Scan for buckets that are already invalidated: */
- for_each_rw_member(ca, c, dev_iter) {
- struct bucket_array *buckets;
- struct bucket_mark m;
-
- down_read(&ca->bucket_lock);
- percpu_down_read_preempt_disable(&c->mark_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 ||
- !test_bit(bu, ca->buckets_nouse) ||
- !is_available_bucket(m) ||
- m.cached_sectors)
- continue;
-
- bch2_mark_alloc_bucket(c, ca, bu, true,
- gc_pos_alloc(c, NULL), 0);
-
- fifo_push(&ca->free_inc, bu);
-
- discard_invalidated_buckets(c, ca);
-
- if (fifo_full(&ca->free[RESERVE_BTREE]))
- break;
- }
- percpu_up_read_preempt_enable(&c->mark_lock);
- up_read(&ca->bucket_lock);
- }
-
- /* did we find enough buckets? */
- for_each_rw_member(ca, c, dev_iter)
- if (!fifo_full(&ca->free[RESERVE_BTREE])) {
- percpu_ref_put(&ca->io_ref);
- goto not_enough;
- }
-
- return 0;
-not_enough:
- 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);
-
- while (1) {
- bool 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);
- return ret;
- }
-
- bch2_invalidate_one_bucket(c, ca, bu,
- &journal_seq);
-
- fifo_push(&ca->free[RESERVE_BTREE], bu);
- bucket_set_dirty(ca, 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, true, &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:
- */
- if (!wrote && ret)
- return ret;
- if (!wrote && !ret)
- break;
- }
-
- pr_debug("flushing journal");
-
- ret = bch2_journal_flush(&c->journal);
- if (ret)
- return ret;
-
- pr_debug("issuing discards");
- allocator_start_issue_discards(c);
-
- set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
-
- /* now flush dirty btree nodes: */
- flush_held_btree_writes(c);
-
- return 0;
-}
-
-int bch2_fs_allocator_start(struct bch_fs *c)
-{
- struct bch_dev *ca;
- unsigned i;
- bool wrote;
- int ret;
-
- down_read(&c->gc_lock);
- ret = __bch2_fs_allocator_start(c);
- up_read(&c->gc_lock);
-
- if (ret)
- return ret;
-
- for_each_rw_member(ca, c, i) {
- ret = bch2_dev_allocator_start(ca);
- if (ret) {
- percpu_ref_put(&ca->io_ref);
- return ret;
- }
- }
-
- return bch2_alloc_write(c, false, &wrote);
-}
-
void bch2_fs_allocator_background_init(struct bch_fs *c)
{
spin_lock_init(&c->freelist_lock);
projects / bcachefs-tools-debian / blobdiff