Re: [Xen-devel] [PATCH v3 05/10] x86/MSI: track host and guest masking separately

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 01.04.16 at 09:40, <liang.z.li@xxxxxxxxx> wrote:
> A couple of weeks ago, Jianzhong reported an issue, the SRIOV NICs (Intel 
> 82599, 82571 ) can't work correctly in Windows guest.
> By debugging, we found your this patch, the commit ID 
> ad28e42bd1d28d746988ed71654e8aa670629753,  caused the regression.

That was obvious.

> Could you help to take a look which part of your change lead to the 
> regression?
> The SRIOV NICs works well in Linux guest.

As you may have seen, I've already found and fixed one aspect.
I'm in the process of fixing another because, no, SRIOV NICs do not
work well in Linux guests, at least not in recent ones (first noticed in
4.4, but last time I had tried before was with 4.0 I think) with
CONFIG_XEN off.

Beyond that I guess I could use some help in at least diagnosing the
issue, e.g. by logging the actions done by Xen, so we can understand
why guest_mask_msi_irq() doesn't get called to unmask the MSI-X
vector(s). I've found the reason why this isn't happening on Linux,
and I am testing a possible fix. However, from the logs I've seen so
far for Windows (with older Xen, and with the above mentioned fix
backported) it is already clear that this change won't help there
(unless of course the issue you see is Windows version dependent).

And just to be clear - the change previously proposed by Jianzhong
is still not going to be an option: We need to understand and address
the root cause of the issue, not paper over it by guessing when to
unmask interrupts.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel