From mboxrd@z Thu Jan  1 00:00:00 1970
From: Marcelo Tosatti <mtosatti@redhat.com>
Subject: Re: [RFC] Multiple QEMU AIO implementations
Date: Wed, 16 Apr 2008 20:52:20 -0300
Message-ID: <20080416235220.GB4869@dmt>
References: <4806716A.70504@us.ibm.com>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit
Cc: Marcelo Tosatti <marcelo@kvack.org>,
	kvm-devel <kvm-devel@lists.sourceforge.net>,
	Avi Kivity <avi@qumranet.com>
To: Anthony Liguori <aliguori@us.ibm.com>
Return-path: <kvm-devel-bounces@lists.sourceforge.net>
Content-Disposition: inline
In-Reply-To: <4806716A.70504@us.ibm.com>
List-Unsubscribe: <https://lists.sourceforge.net/lists/listinfo/kvm-devel>,
	<mailto:kvm-devel-request@lists.sourceforge.net?subject=unsubscribe>
List-Archive: <http://sourceforge.net/mailarchive/forum.php?forum_name=kvm-devel>
List-Post: <mailto:kvm-devel@lists.sourceforge.net>
List-Help: <mailto:kvm-devel-request@lists.sourceforge.net?subject=help>
List-Subscribe: <https://lists.sourceforge.net/lists/listinfo/kvm-devel>,
	<mailto:kvm-devel-request@lists.sourceforge.net?subject=subscribe>
Sender: kvm-devel-bounces@lists.sourceforge.net
Errors-To: kvm-devel-bounces@lists.sourceforge.net
List-Id: kvm.vger.kernel.org

On Wed, Apr 16, 2008 at 04:36:42PM -0500, Anthony Liguori wrote:
> This isn't fully cooked yet, but pretty close.  The basic idea is to 
> make the aio usage in block-raw go to a set of function pointers and 
> allow multiple simultaneous AIO implementations.

The AIO API looks great. Looking forward to a linux-aio backend.

> I converted the posix-aio support to this, and also introduced a "unix" 
> aio which just uses O_NONBLOCK and select().  The later only supports 1 
> simultaneous request per-fd but currently posix-aio is limited to that 
> too.  At least with my QEMU testing, the unix aio implementation 
> outperforms posix-aio by a factor of 2.

Unfortunately O_NONBLOCK is not honoured for block backed storage,
so its a no-go.

> And it uses no signals...
> 
> I'm inclined to suggest that we use signalfd with posix-aio, and for 
> older guests, just fall back to unix aio.  We can also introduce a 
> linux-aio and use that when possible.

I think Avi's suggestion to emulate signalfd with a separate thread
which does sigtimedwait+write-to-a-pipe is sensate.


-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone