bcache: Fix a journalling performance bug
authorKent Overstreet <kmo@daterainc.com>
Tue, 8 Oct 2013 22:50:46 +0000 (15:50 -0700)
committerKent Overstreet <kmo@daterainc.com>
Mon, 11 Nov 2013 05:55:53 +0000 (21:55 -0800)
drivers/md/bcache/journal.c
drivers/md/bcache/journal.h

index 8435f81e5d858012e8aca6be8204e923a34b1d01..7c9e6bf6aababae1c94abd3971f49af2b375c0c7 100644 (file)
@@ -554,32 +554,26 @@ static void journal_write_endio(struct bio *bio, int error)
        struct journal_write *w = bio->bi_private;
 
        cache_set_err_on(error, w->c, "journal io error");
-       closure_put(&w->c->journal.io.cl);
+       closure_put(&w->c->journal.io);
 }
 
 static void journal_write(struct closure *);
 
 static void journal_write_done(struct closure *cl)
 {
-       struct journal *j = container_of(cl, struct journal, io.cl);
-       struct cache_set *c = container_of(j, struct cache_set, journal);
-
+       struct journal *j = container_of(cl, struct journal, io);
        struct journal_write *w = (j->cur == j->w)
                ? &j->w[1]
                : &j->w[0];
 
        __closure_wake_up(&w->wait);
-
-       if (c->journal_delay_ms)
-               closure_delay(&j->io, msecs_to_jiffies(c->journal_delay_ms));
-
-       continue_at(cl, journal_write, system_wq);
+       continue_at_nobarrier(cl, journal_write, system_wq);
 }
 
 static void journal_write_unlocked(struct closure *cl)
        __releases(c->journal.lock)
 {
-       struct cache_set *c = container_of(cl, struct cache_set, journal.io.cl);
+       struct cache_set *c = container_of(cl, struct cache_set, journal.io);
        struct cache *ca;
        struct journal_write *w = c->journal.cur;
        struct bkey *k = &c->journal.key;
@@ -660,7 +654,7 @@ static void journal_write_unlocked(struct closure *cl)
 
 static void journal_write(struct closure *cl)
 {
-       struct cache_set *c = container_of(cl, struct cache_set, journal.io.cl);
+       struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 
        spin_lock(&c->journal.lock);
        journal_write_unlocked(cl);
@@ -669,7 +663,10 @@ static void journal_write(struct closure *cl)
 static void __journal_try_write(struct cache_set *c, bool noflush)
        __releases(c->journal.lock)
 {
-       struct closure *cl = &c->journal.io.cl;
+       struct closure *cl = &c->journal.io;
+       struct journal_write *w = c->journal.cur;
+
+       w->need_write = true;
 
        if (!closure_trylock(cl, &c->cl))
                spin_unlock(&c->journal.lock);
@@ -688,18 +685,24 @@ void bch_journal_meta(struct cache_set *c, struct closure *cl)
 
        if (CACHE_SYNC(&c->sb)) {
                spin_lock(&c->journal.lock);
-
                w = c->journal.cur;
-               w->need_write = true;
 
                if (cl)
                        BUG_ON(!closure_wait(&w->wait, cl));
 
-               closure_flush(&c->journal.io);
                __journal_try_write(c, true);
        }
 }
 
+static void journal_write_work(struct work_struct *work)
+{
+       struct cache_set *c = container_of(to_delayed_work(work),
+                                          struct cache_set,
+                                          journal.work);
+       spin_lock(&c->journal.lock);
+       journal_try_write(c);
+}
+
 /*
  * Entry point to the journalling code - bio_insert() and btree_invalidate()
  * pass bch_journal() a list of keys to be journalled, and then
@@ -739,7 +742,6 @@ void bch_journal(struct closure *cl)
        }
 
        w = c->journal.cur;
-       w->need_write = true;
        b = __set_blocks(w->data, w->data->keys + n, c);
 
        if (b * c->sb.block_size > PAGE_SECTORS << JSET_BITS ||
@@ -755,8 +757,6 @@ void bch_journal(struct closure *cl)
 
                BUG_ON(!closure_wait(&w->wait, cl));
 
-               closure_flush(&c->journal.io);
-
                journal_try_write(c);
                continue_at(cl, bch_journal, bcache_wq);
        }
@@ -768,11 +768,15 @@ void bch_journal(struct closure *cl)
        atomic_inc(op->journal);
 
        if (op->flush_journal) {
-               closure_flush(&c->journal.io);
                closure_wait(&w->wait, cl->parent);
+               journal_try_write(c);
+       } else if (!w->need_write) {
+               schedule_delayed_work(&c->journal.work,
+                                     msecs_to_jiffies(c->journal_delay_ms));
+               spin_unlock(&c->journal.lock);
+       } else {
+               spin_unlock(&c->journal.lock);
        }
-
-       journal_try_write(c);
 out:
        bch_btree_insert_async(cl);
 }
@@ -790,6 +794,7 @@ int bch_journal_alloc(struct cache_set *c)
 
        closure_init_unlocked(&j->io);
        spin_lock_init(&j->lock);
+       INIT_DELAYED_WORK(&j->work, journal_write_work);
 
        c->journal_delay_ms = 100;
 
index 3d7851274b0413d28d89dec4b82693b760982e39..3ca93d3d566cea572e9fd8caf462a807033cc013 100644 (file)
@@ -140,7 +140,8 @@ struct journal {
        spinlock_t              lock;
        /* used when waiting because the journal was full */
        struct closure_waitlist wait;
-       struct closure_with_timer io;
+       struct closure          io;
+       struct delayed_work     work;
 
        /* Number of blocks free in the bucket(s) we're currently writing to */
        unsigned                blocks_free;