mempool: fix first round failure behavior
Tejun Heo [Tue, 10 Jan 2012 23:08:28 +0000 (15:08 -0800)]
mempool modifies gfp_mask so that the backing allocator doesn't try too
hard or trigger warning message when there's pool to fall back on.  In
addition, for the first try, it removes __GFP_WAIT and IO, so that it
doesn't trigger reclaim or wait when allocation can be fulfilled from
pool; however, when that allocation fails and pool is empty too, it waits
for the pool to be replenished before retrying.

Allocation which could have succeeded after a bit of reclaim has to wait
on the reserved items and it's not like mempool doesn't retry with
__GFP_WAIT and IO.  It just does that *after* someone returns an element,
pointlessly delaying things.

Fix it by retrying immediately if the first round of allocation attempts
w/o __GFP_WAIT and IO fails.

[akpm@linux-foundation.org: shorten the lock hold time]
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/mempool.c

index e3a802a..d904981 100644 (file)
@@ -221,14 +221,23 @@ repeat_alloc:
                return element;
        }
 
-       /* We must not sleep in the GFP_ATOMIC case */
+       /*
+        * We use gfp mask w/o __GFP_WAIT or IO for the first round.  If
+        * alloc failed with that and @pool was empty, retry immediately.
+        */
+       if (gfp_temp != gfp_mask) {
+               spin_unlock_irqrestore(&pool->lock, flags);
+               gfp_temp = gfp_mask;
+               goto repeat_alloc;
+       }
+
+       /* We must not sleep if !__GFP_WAIT */
        if (!(gfp_mask & __GFP_WAIT)) {
                spin_unlock_irqrestore(&pool->lock, flags);
                return NULL;
        }
 
        /* Let's wait for someone else to return an element to @pool */
-       gfp_temp = gfp_mask;
        init_wait(&wait);
        prepare_to_wait(&pool->wait, &wait, TASK_UNINTERRUPTIBLE);