#include "disk_groups.h"
#include "ec.h"
#include "error.h"
-#include "io.h"
+#include "io_write.h"
#include "journal.h"
#include "movinggc.h"
#include "nocow_locking.h"
void bch2_reset_alloc_cursors(struct bch_fs *c)
{
- struct bch_dev *ca;
- unsigned i;
-
rcu_read_lock();
- for_each_member_device_rcu(ca, c, i, NULL)
+ for_each_member_device_rcu(c, ca, NULL)
ca->alloc_cursor = 0;
rcu_read_unlock();
}
if (cl)
closure_wait(&c->open_buckets_wait, cl);
- if (!c->blocked_allocate_open_bucket)
- c->blocked_allocate_open_bucket = local_clock();
-
+ track_event_change(&c->times[BCH_TIME_blocked_allocate_open_bucket],
+ &c->blocked_allocate_open_bucket, true);
spin_unlock(&c->freelist_lock);
return ERR_PTR(-BCH_ERR_open_buckets_empty);
}
ca->nr_open_buckets++;
bch2_open_bucket_hash_add(c, ob);
- if (c->blocked_allocate_open_bucket) {
- bch2_time_stats_update(
- &c->times[BCH_TIME_blocked_allocate_open_bucket],
- c->blocked_allocate_open_bucket);
- c->blocked_allocate_open_bucket = 0;
- }
+ track_event_change(&c->times[BCH_TIME_blocked_allocate_open_bucket],
+ &c->blocked_allocate_open_bucket, false);
- if (c->blocked_allocate) {
- bch2_time_stats_update(
- &c->times[BCH_TIME_blocked_allocate],
- c->blocked_allocate);
- c->blocked_allocate = 0;
- }
+ track_event_change(&c->times[BCH_TIME_blocked_allocate],
+ &c->blocked_allocate, false);
spin_unlock(&c->freelist_lock);
return ob;
ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
if (!ob)
- iter.path->preserve = false;
+ set_btree_iter_dontneed(&iter);
err:
- if (iter.trans && iter.path)
+ if (iter.path)
set_btree_iter_dontneed(&iter);
bch2_trans_iter_exit(trans, &iter);
printbuf_exit(&buf);
struct bucket_alloc_state *s,
struct closure *cl)
{
- struct btree_iter iter;
- struct bkey_s_c k;
+ struct btree_iter iter, citer;
+ struct bkey_s_c k, ck;
struct open_bucket *ob = NULL;
- u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
- u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
+ u64 first_bucket = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
+ u64 alloc_start = max(first_bucket, READ_ONCE(ca->alloc_cursor));
+ u64 alloc_cursor = alloc_start;
int ret;
+
+ /*
+ * Scan with an uncached iterator to avoid polluting the key cache. An
+ * uncached iter will return a cached key if one exists, but if not
+ * there is no other underlying protection for the associated key cache
+ * slot. To avoid racing bucket allocations, look up the cached key slot
+ * of any likely allocation candidate before attempting to proceed with
+ * the allocation. This provides proper exclusion on the associated
+ * bucket.
+ */
again:
for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
BTREE_ITER_SLOTS, k, ret) {
continue;
a = bch2_alloc_to_v4(k, &a_convert);
-
if (a->data_type != BCH_DATA_free)
continue;
+ /* now check the cached key to serialize concurrent allocs of the bucket */
+ ck = bch2_bkey_get_iter(trans, &citer, BTREE_ID_alloc, k.k->p, BTREE_ITER_CACHED);
+ ret = bkey_err(ck);
+ if (ret)
+ break;
+
+ a = bch2_alloc_to_v4(ck, &a_convert);
+ if (a->data_type != BCH_DATA_free)
+ goto next;
+
s->buckets_seen++;
ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
+next:
+ set_btree_iter_dontneed(&citer);
+ bch2_trans_iter_exit(trans, &citer);
if (ob)
break;
}
bch2_trans_iter_exit(trans, &iter);
+ alloc_cursor = iter.pos.offset;
ca->alloc_cursor = alloc_cursor;
if (!ob && ret)
ob = ERR_PTR(ret);
- if (!ob && alloc_cursor > alloc_start) {
- alloc_cursor = alloc_start;
+ if (!ob && alloc_start > first_bucket) {
+ alloc_cursor = alloc_start = first_bucket;
goto again;
}
ob = try_alloc_bucket(trans, ca, watermark,
alloc_cursor, s, k, cl);
if (ob) {
- iter.path->preserve = false;
+ set_btree_iter_dontneed(&iter);
break;
}
}
}
/**
- * bch_bucket_alloc - allocate a single bucket from a specific device
+ * bch2_bucket_alloc_trans - allocate a single bucket from a specific device
+ * @trans: transaction object
+ * @ca: device to allocate from
+ * @watermark: how important is this allocation?
+ * @cl: if not NULL, closure to be used to wait if buckets not available
+ * @usage: for secondarily also returning the current device usage
*
- * Returns index of bucket on success, 0 on failure
+ * Returns: an open_bucket on success, or an ERR_PTR() on failure.
*/
static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
struct bch_dev *ca,
goto again;
}
- if (!c->blocked_allocate)
- c->blocked_allocate = local_clock();
+ track_event_change(&c->times[BCH_TIME_blocked_allocate],
+ &c->blocked_allocate, true);
ob = ERR_PTR(-BCH_ERR_freelist_empty);
goto err;
struct open_bucket *ob;
bch2_trans_do(c, NULL, NULL, 0,
- PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, watermark,
+ PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(trans, ca, watermark,
cl, &usage)));
return ob;
}
bch_dev_bkey_exists(c, ob->dev)->mi.durability;
BUG_ON(*nr_effective >= nr_replicas);
- BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
__clear_bit(ob->dev, devs_may_alloc->d);
- *nr_effective += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
- ? durability : 1;
+ *nr_effective += durability;
*have_cache |= !durability;
ob_push(c, ptrs, ob);
struct dev_alloc_list devs_sorted;
struct ec_stripe_head *h;
struct open_bucket *ob;
- struct bch_dev *ca;
unsigned i, ec_idx;
int ret = 0;
}
goto out_put_head;
got_bucket:
- ca = bch_dev_bkey_exists(c, ob->dev);
-
ob->ec_idx = ec_idx;
ob->ec = h->s;
ec_stripe_new_get(h->s, STRIPE_REF_io);
devs = target_rw_devs(c, wp->data_type, target);
/* Don't allocate from devices we already have pointers to: */
- for (i = 0; i < devs_have->nr; i++)
- __clear_bit(devs_have->devs[i], devs.d);
+ darray_for_each(*devs_have, i)
+ __clear_bit(*i, devs.d);
open_bucket_for_each(c, ptrs, ob, i)
__clear_bit(ob->dev, devs.d);
/**
* should_drop_bucket - check if this is open_bucket should go away
+ * @ob: open_bucket to predicate on
+ * @c: filesystem handle
* @ca: if set, we're killing buckets for a particular device
* @ec: if true, we're shutting down erasure coding and killing all ec
* open_buckets
* otherwise, return true
+ * Returns: true if we should kill this open_bucket
*
* We're killing open_buckets because we're shutting down a device, erasure
* coding, or the entire filesystem - check if this open_bucket matches:
return wp;
}
+static noinline void
+deallocate_extra_replicas(struct bch_fs *c,
+ struct open_buckets *ptrs,
+ struct open_buckets *ptrs_no_use,
+ unsigned extra_replicas)
+{
+ struct open_buckets ptrs2 = { 0 };
+ struct open_bucket *ob;
+ unsigned i;
+
+ open_bucket_for_each(c, ptrs, ob, i) {
+ unsigned d = bch_dev_bkey_exists(c, ob->dev)->mi.durability;
+
+ if (d && d <= extra_replicas) {
+ extra_replicas -= d;
+ ob_push(c, ptrs_no_use, ob);
+ } else {
+ ob_push(c, &ptrs2, ob);
+ }
+ }
+
+ *ptrs = ptrs2;
+}
+
/*
* Get us an open_bucket we can allocate from, return with it locked:
*/
int ret;
int i;
+ if (!IS_ENABLED(CONFIG_BCACHEFS_ERASURE_CODING))
+ erasure_code = false;
+
BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
BUG_ON(!nr_replicas || !nr_replicas_required);
goto alloc_done;
/* Don't retry from all devices if we're out of open buckets: */
- if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
- goto allocate_blocking;
+ if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty)) {
+ int ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
+ target, erasure_code,
+ nr_replicas, &nr_effective,
+ &have_cache, watermark,
+ flags, cl);
+ if (!ret ||
+ bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
+ bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
+ goto alloc_done;
+ }
/*
* Only try to allocate cache (durability = 0 devices) from the
&have_cache, watermark,
flags, cl);
} else {
-allocate_blocking:
ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
target, erasure_code,
nr_replicas, &nr_effective,
if (ret)
goto err;
+ if (nr_effective > nr_replicas)
+ deallocate_extra_replicas(c, &ptrs, &wp->ptrs, nr_effective - nr_replicas);
+
/* Free buckets we didn't use: */
open_bucket_for_each(c, &wp->ptrs, ob, i)
open_bucket_free_unused(c, ob);
projects / bcachefs-tools-debian / blobdiff