netdev - Re: TX from KVM guest virtio

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Fri, 11 Mar 2011 16:49:26 +1030
From:	Rusty Russell <rusty@...tcorp.com.au>
To:	Shirley Ma <mashirle@...ibm.com>,
	"Michael S. Tsirkin" <mst@...hat.com>
Cc:	Tom Lendacky <tahm@...ux.vnet.ibm.com>,
	Krishna Kumar2 <krkumar2@...ibm.com>,
	David Miller <davem@...emloft.net>, kvm@...r.kernel.org,
	netdev@...r.kernel.org, steved@...ibm.com
Subject: Re: TX from KVM guest virtio_net to vhost issues

On Wed, 09 Mar 2011 13:46:36 -0800, Shirley Ma <mashirle@...ibm.com> wrote:
> Since we have lots of performance discussions about virtio_net and vhost
> communication. I think it's better to have a common understandings of
> the code first, then we can seek the right directions to improve it. We
> also need to collect more statistics data on both virtio and vhost.
> 
> Let's look at TX first: from virtio_net(guest) to vhost(host), send vq
> is shared between guest virtio_net and host vhost, it uses memory
> barriers to sync the changes.
> 
> In the start:
> 
> Guest virtio_net TX send completion interrupt (for freeing used skbs) is
> disable. Guest virtio_net TX send completion interrupt is enabled only
> when send vq is overrun, guest needs to wait vhost to consume more
> available skbs. 
> 
> Host vhost notification is enabled in the beginning (for consuming
> available skbs); It is disable whenever the send vq is not empty. Once
> the send vq is empty, the notification is enabled by vhost.
> 
> In guest start_xmit(), it first frees used skbs, then send available
> skbs to vhost, ideally guest never enables TX send completion interrupts
> to free used skbs if vhost keeps posting used skbs in send vq.
> 
> In vhost handle_tx(), it wakes up by guest whenever the send vq has a
> skb to send, once the send vq is not empty, vhost exits handle_tx()
> without enabling notification. Ideally if guest keeps xmit skbs in send
> vq, the notification is never enabled.
> 
> I don't see issues on this implementation.
> 
> However, in our TCP_STREAM small message size test, we found that
> somehow guest couldn't see more used skbs to free, which caused
> frequently TX send queue overrun.

So it seems like the guest is outrunning the host?

> In our TCP_RR small message size multiple streams test, we found that
> vhost couldn't see more xmit skbs in send vq, thus it enabled
> notification too often.

And here the host is outrunning the guest?

> What's the possible cause here in xmit? How guest, vhost are being
> scheduled? Whether it's possible, guest virtio_net cooperates with vhost
> for ideal performance: both guest virtio_net and vhost can be in pace
> with send vq without many notifications and exits?

We tend to blame the scheduler for all kinds of things, until we find
the real cause :)  Nailing things to different CPUs might help us
determine if it really is scheduler...

But if one side outruns the other, it does a lot of unnecessary work
notifying/interrupting it over and over again before the host/guest gets
a chance to shut notifications/interrupts off.  Hence the last_used
publishing patch (Xen does this right, I screwed it up).

Long weekend here, and I'm otherwise committed.  But if noone has
cleaned up that patch by early next week, I'll try to do so and see if
we can make a real difference.

Cheers,
Rusty.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html