| Message ID | 20190419160509.66298-17-colyli@suse.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | bcache: fix journal no-space deadlock | expand |
On 4/19/19 6:05 PM, Coly Li wrote: > In btree_flush_write(), iterating all cached btree nodes and adding them > into ordered heap c->flush_btree takes quite long time. In order to > protect ordered heap c->flush_btree, spin lock c->journal.lock is held > for all the iteration and heap ordering. When journal space is fully > occupied, btree_flush_write() might be called frequently, if the cached > btree node iteration takes too much time, kenrel will complain that > normal journal kworkers are blocked too long. Of cause write performance > drops at this moment. > > This patch introduces a new spin lock member in struct journal, named > flush_write_lock. This lock is only used in btree_flush_write() and > protect the ordered heap c->flush_btree during all the cached btree node > iteration. Then there won't be lock contention on c->journal.lock. > > After this fix, when journal space is fully occupied, it is very rare to > observe the journal kworker blocking timeout warning. > > Signed-off-by: Coly Li <colyli@suse.de> > --- > drivers/md/bcache/journal.c | 5 +++-- > drivers/md/bcache/journal.h | 1 + > 2 files changed, 4 insertions(+), 2 deletions(-) > Reviewed-by: Hannes Reinecke <hare@suse.com> Cheers, Hannes
diff --git a/drivers/md/bcache/journal.c b/drivers/md/bcache/journal.c index 8536e76fcac9..6e38470f6924 100644 --- a/drivers/md/bcache/journal.c +++ b/drivers/md/bcache/journal.c @@ -519,7 +519,7 @@ static void btree_flush_write(struct cache_set *c) atomic_long_inc(&c->flush_write); retry: - spin_lock(&c->journal.lock); + spin_lock(&c->journal.flush_write_lock); if (heap_empty(&c->flush_btree)) { for_each_cached_btree(b, c, i) if (btree_current_write(b)->journal) { @@ -540,7 +540,7 @@ static void btree_flush_write(struct cache_set *c) b = NULL; heap_pop(&c->flush_btree, b, journal_min_cmp); - spin_unlock(&c->journal.lock); + spin_unlock(&c->journal.flush_write_lock); if (b) { mutex_lock(&b->write_lock); @@ -1099,6 +1099,7 @@ int bch_journal_alloc(struct cache_set *c) struct journal *j = &c->journal; spin_lock_init(&j->lock); + spin_lock_init(&j->flush_write_lock); INIT_DELAYED_WORK(&j->work, journal_write_work); c->journal_delay_ms = 100; diff --git a/drivers/md/bcache/journal.h b/drivers/md/bcache/journal.h index a8be14c6f6d9..d8ad99f6191b 100644 --- a/drivers/md/bcache/journal.h +++ b/drivers/md/bcache/journal.h @@ -103,6 +103,7 @@ struct journal_write { /* Embedded in struct cache_set */ struct journal { spinlock_t lock; + spinlock_t flush_write_lock; /* used when waiting because the journal was full */ struct closure_waitlist wait; struct closure io;
In btree_flush_write(), iterating all cached btree nodes and adding them into ordered heap c->flush_btree takes quite long time. In order to protect ordered heap c->flush_btree, spin lock c->journal.lock is held for all the iteration and heap ordering. When journal space is fully occupied, btree_flush_write() might be called frequently, if the cached btree node iteration takes too much time, kenrel will complain that normal journal kworkers are blocked too long. Of cause write performance drops at this moment. This patch introduces a new spin lock member in struct journal, named flush_write_lock. This lock is only used in btree_flush_write() and protect the ordered heap c->flush_btree during all the cached btree node iteration. Then there won't be lock contention on c->journal.lock. After this fix, when journal space is fully occupied, it is very rare to observe the journal kworker blocking timeout warning. Signed-off-by: Coly Li <colyli@suse.de> --- drivers/md/bcache/journal.c | 5 +++-- drivers/md/bcache/journal.h | 1 + 2 files changed, 4 insertions(+), 2 deletions(-)