Re: error when pass through device to guest with qemu-xen-dir-remote

[Anthony PERARD <anthony.perard@citrix.com>]

On 09/18/2012 01:46 AM, Zhang, Yang Z wrote:
> Zhang, Yang Z wrote on 2012-09-11:
>>   wrote on August 09, 2012 2:49 PM To: Ian Campbell; Stefano Stabellini:
>>> I rebuild the upstream QEMU according to the wiki, but device static
>>> assignment doesn't work, no lspci output in guest. However hotplug &
>>> unplug works fine.
>>>
>>> -----Original Message----- From: xen-devel-bounces@lists.xen.org
>>> [mailto:xen-devel-bounces@lists.xen.org] On Behalf Of Ian Campbell Sent:
>>> Friday, August 03, 2012 6:36 PM To: Stefano Stabellini Cc: Zhang, Yang
>>> Z; Anthony Perard; xen-devel Subject: Re: [Xen-devel] error when pass
>>> through device to guest with qemu-xen-dir-remote
>>>
>>> On Fri, 2012-08-03 at 11:29 +0100, Stefano Stabellini wrote:
>>>> On Fri, 3 Aug 2012, Zhang, Yang Z wrote:
>>>>> When create guest with device assigned, it shows the error and the
>>>>> device wasn't able to work inside guest: libxl: error:
>>>>> libxl_qmp.c:288:qmp_handle_error_response: received an error message
>>>>> from QMP server: Parameter 'driver' expects a driver name
>>>>>
>>>>> It only fails with qemu-xen-dir-remote(Is this tree more close to
>>>>> upstream qemu?). I don't see the error with the traditional Qemu. I
>>>>> also tried qemu-upstream, but it fails when I try to enable pci
>>>>> pass-through
>>> for xen. I think Anthony's patch to add pci pass-through support for Xen is
>>> accepted by qemu-upstream, am I right?
>>>>
>>>> Yes, it was accepted, but it is present only in upstream QEMU (from
>>>> git://git.qemu.org/qemu.git), not the tree we are currently using in
>>>> xen-unstable for development
>>>> (git://xenbits.xensource.com/qemu-upstream-unstable.git).
>>>> Make sure you are using the right tree!
>>>
>>> http://wiki.xen.org/wiki/QEMU_Upstream has some notes on how to use the
>>> upstream qemu tree instead of our stable branch of upstream.
>>>
>>>>
>>>> Anthony is currently on vacation and is going to be back in about a
>>>> week.
>>>>
>>>>> Another question:
>>>>> Now I am trying to add some features (relevant to pass through device) to
>>> Qemu, which tree should I use? Since traditional qemu is great
>>> different from qemu-upstream, it is too old to develop patch base on
>>> it. But besides the old one, I cannot find a working qemu.
>>>>
>>>> You should use upstream QEMU, I am going to rebase our tree on that
>>>> early on in the 4.3 release cycle.
>>
>> Hi Anthony
>>
>> I found the issue is caused by PV driver. Since RHEL6 integrated PV driver by
>> default, when booting rhel6 as guest, it will load the PV driver and write the xen
>> platform device to notify qemu to unplug all NICs including the pass through
>> device. This definitely is wrong. We only need to unplug the emulator device, and
>> leave pass through device.
>
> Hi Anthony,
>
> Any comments for this?

Hi,

So, we'll have to add a check in the unplug code of qemu to ignore 
passthrough devices.

Here is a patches, I did not test it.

diff --git a/hw/xen_platform.c b/hw/xen_platform.c
index 0d6c2ff..2d3978e 100644
--- a/hw/xen_platform.c
+++ b/hw/xen_platform.c
@@ -85,8 +85,10 @@ static void log_writeb ...

  static void unplug_nic(PCIBus *b, PCIDevice *d, void *o)
  {
+    /* We have to ignore passthrough devices */
      if (pci_get_word(d->config + PCI_CLASS_DEVICE) ==
-            PCI_CLASS_NETWORK_ETHERNET) {
+            PCI_CLASS_NETWORK_ETHERNET
+        && strcmp(d->name, "xen-pci-passthrough") != 0) {
          qdev_free(&d->qdev);
      }
  }


Thanks,

-- 
Anthony PERARD