Patchwork [v2,2/3] aio: use g_slice_alloc() for AIOCB pooling

login
register
mail settings
Submitter Kevin Wolf
Date Nov. 12, 2012, 10:34 a.m.
Message ID <50A0D09B.5080202@redhat.com>
Download mbox | patch
Permalink /patch/198357/
State New
Headers show

Comments

Kevin Wolf - Nov. 12, 2012, 10:34 a.m.
Am 31.10.2012 16:34, schrieb Stefan Hajnoczi:
> AIO control blocks are frequently acquired and released because each aio
> request involves at least one AIOCB.  Therefore, we pool them to avoid
> heap allocation overhead.
> 
> The problem with the freelist approach in AIOPool is thread-safety.  If
> we want BlockDriverStates to associate with AioContexts that execute in
> multiple threads, then a global freelist becomes a problem.
> 
> This patch drops the freelist and instead uses g_slice_alloc() which is
> tuned for per-thread fixed-size object pools.  qemu_aio_get() and
> qemu_aio_release() are now thread-safe.
> 
> Note that the change from g_malloc0() to g_slice_alloc() should be safe
> since the freelist reuse case doesn't zero the AIOCB either.

Of course the real reason is that all fields are set anyway. We could
even express this fact more clearly in the code:


If you agree, I'll commit this on top of your series.

Kevin
Paolo Bonzini - Nov. 12, 2012, 10:37 a.m.
Il 12/11/2012 11:34, Kevin Wolf ha scritto:
>> Note that the change from g_malloc0() to g_slice_alloc() should be safe
>> since the freelist reuse case doesn't zero the AIOCB either.
> 
> Of course the real reason is that all fields are set anyway.

This doesn't necessarily apply for "subclasses", though.  (It does for
the reason Stefan mentioned, but it is not possible to make it apparent
in the code).

Paolo

> diff --git a/block.c b/block.c
> index 854ebd6..6f5a0e7 100644
> --- a/block.c
> +++ b/block.c
> @@ -3910,10 +3910,13 @@ void *qemu_aio_get(const AIOCBInfo *aiocb_info,
> BlockDriverState *bs,
>      BlockDriverAIOCB *acb;
> 
>      acb = g_slice_alloc(aiocb_info->aiocb_size);
> -    acb->aiocb_info = aiocb_info;
> -    acb->bs = bs;
> -    acb->cb = cb;
> -    acb->opaque = opaque;
> +    *acb = (BlockDriverAIOCB) {
> +        .aiocb_info = aiocb_info,
> +        .bs         = bs,
> +        .cb         = cb,
> +        .opaque     = opaque,
> +    };
> +
>      return acb;
>  }
> 
> If you agree, I'll commit this on top of your series.
> 
> Kevin
>
Kevin Wolf - Nov. 12, 2012, 10:42 a.m.
Am 12.11.2012 11:37, schrieb Paolo Bonzini:
> Il 12/11/2012 11:34, Kevin Wolf ha scritto:
>>> Note that the change from g_malloc0() to g_slice_alloc() should be safe
>>> since the freelist reuse case doesn't zero the AIOCB either.
>>
>> Of course the real reason is that all fields are set anyway.
> 
> This doesn't necessarily apply for "subclasses", though.  (It does for
> the reason Stefan mentioned, but it is not possible to make it apparent
> in the code).

See? I really shouldn't review code any more. :-)

Kevin

Patch

diff --git a/block.c b/block.c
index 854ebd6..6f5a0e7 100644
--- a/block.c
+++ b/block.c
@@ -3910,10 +3910,13 @@  void *qemu_aio_get(const AIOCBInfo *aiocb_info,
BlockDriverState *bs,
     BlockDriverAIOCB *acb;

     acb = g_slice_alloc(aiocb_info->aiocb_size);
-    acb->aiocb_info = aiocb_info;
-    acb->bs = bs;
-    acb->cb = cb;
-    acb->opaque = opaque;
+    *acb = (BlockDriverAIOCB) {
+        .aiocb_info = aiocb_info,
+        .bs         = bs,
+        .cb         = cb,
+        .opaque     = opaque,
+    };
+
     return acb;
 }