From mboxrd@z Thu Jan  1 00:00:00 1970
From: Gleb Natapov <gleb@redhat.com>
Subject: Re: major slowdown in i/o (due to high CPU usage) from 2.6.32 to
 current
Date: Tue, 19 Jul 2011 12:16:29 +0300
Message-ID: <20110719091629.GQ2400@redhat.com>
References: <4E249499.4050502@msgid.tls.msk.ru>
 <4E254AFC.8040405@redhat.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Cc: Michael Tokarev <mjt@tls.msk.ru>, KVM list <kvm@vger.kernel.org>,
	Svante Signell <srs@kth.se>
To: Avi Kivity <avi@redhat.com>
Return-path: <kvm-owner@vger.kernel.org>
Received: from mx1.redhat.com ([209.132.183.28]:62331 "EHLO mx1.redhat.com"
	rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP
	id S1752346Ab1GSJQf (ORCPT <rfc822;kvm@vger.kernel.org>);
	Tue, 19 Jul 2011 05:16:35 -0400
Content-Disposition: inline
In-Reply-To: <4E254AFC.8040405@redhat.com>
Sender: kvm-owner@vger.kernel.org
List-ID: <kvm.vger.kernel.org>

On Tue, Jul 19, 2011 at 12:14:36PM +0300, Avi Kivity wrote:
> On 07/18/2011 11:16 PM, Michael Tokarev wrote:
> >It all started by Svante Signell on IRC, see
> >http://bugs.debian.org/634149 .  The short story:
> >running hurd guest image (on ide drive); when changing
> >only host kernel from 2.6.32 to (in my case) 3.0.0-rc7,
> >the i/o (especially write) speed of the guest reduces
> >dramatically, as is host CPU usage.  I used a cd-1.iso
> >install image from the URL provided in the bugreport,
> >http://people.debian.org/~sthibault/hurd-i386/installer/cdimage/ --
> >even installing packages is very significantly slower
> >in current kernels than it was in 2.6.32.
> >
> >Typical kvm_stats output on 3.0.0 host kernel:
> >
> >kvm-18166 [001] 47471.398406: kvm_emulate_insn:     80000000:160750: f3 66 6f
> >kvm-18166 [001] 47471.398406: kvm_pio:              pio_write at 0x1f0 size 2 count 1
> >kvm-18166 [001] 47471.398407: kvm_userspace_exit:   reason KVM_EXIT_IO (2)
> >kvm-18166 [001] 47471.398408: kvm_entry:            vcpu 0
> >kvm-18166 [001] 47471.398409: kvm_exit:             reason EXIT_IOIO rip 0x160750 info 1f0012c  0
> >
> >
> >Gleb, Avi -- what we're missing here?
> >
> >It does not look like the problem we have with FreeBSD,
> >this is apparently something different...
> 
> Looks like hurd doesn't use DMA.  Instead it uses rep/outsw, which
> indeed has a performance regression - it now exits on every word
> instead of on every sector.
> 
> Gleb?
> 
Yes, we are discussing this right now on IRC with Michael. This is exactly the case :(
Hurd developers says that they support IDE DMA though.

--
			Gleb.