From mboxrd@z Thu Jan  1 00:00:00 1970
From: Avi Kivity <avi@qumranet.com>
Subject: Re: direct mmio for passthrough - kernel part
Date: Fri, 18 Apr 2008 19:02:14 +0300
Message-ID: <4808C606.9040402@qumranet.com>
References: <1208352331-27201-1-git-send-email-benami@il.ibm.com>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit
Cc: kvm-devel@lists.sourceforge.net, andrea@qumranet.com,
	allen.m.kay@intel.com
To: benami@il.ibm.com
Return-path: <kvm-devel-bounces@lists.sourceforge.net>
In-Reply-To: <1208352331-27201-1-git-send-email-benami@il.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

benami@il.ibm.com wrote:
> This patch for PCI passthrough devices enables a guest to access a device's
> memory mapped I/O regions directly, without requiring the host to trap and
> emulate every MMIO access. 
>
> Updated from last version: we create a memory slot for each MMIO region of the
> guest's devices, and then use the /sys/bus/pci/.../resource# mapping to find the
> hfn for that MMIO region. The kernel part and the userspace part of this
> patchset apply to Amit's pv-dma tree.  Tested on a Lenovo M57p with an e1000 NIC
> assigned directly to an FC8 guest.
>
> Comments are appreciated. 
>   

I see no support for cache attributes in the page attributes table or 
mtrr.  I guess for most devices this will work (as they will be set as 
uncachable by the mtrrs), but for display cards we'd need to set the 
vram as write-combining to get reasonable performance.  This requires 
mtrr and pat emulation in kvm so we detect the guest's intentions.

-- 
Any sufficiently difficult bug is indistinguishable from a feature.


-------------------------------------------------------------------------
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