Re: Testing the virtio-vhost-user QEMU patch

[Alyssa Ross]

Alyssa Ross <hi@alyssa.is> writes:

> Stefan Hajnoczi <stefanha@redhat.com> writes:
>
>> On Tue, Jul 21, 2020 at 07:14:38AM +0000, Alyssa Ross wrote:
>>> Hi -- I hope it's okay me reaching out like this.
>>> 
>>> I've been trying to test out the virtio-vhost-user implementation that's
>>> been posted to this list a couple of times, but have been unable to get
>>> it to boot a kernel following the steps listed either on
>>> <https://wiki.qemu.org/Features/VirtioVhostUser> or
>>> <https://ndragazis.github.io/dpdk-vhost-vvu-demo.html>.
>>> 
>>> Specifically, the kernel appears to be unable to write to the
>>> virtio-vhost-user device's PCI registers.  I've included the full panic
>>> output from the kernel at the end of this message.  The panic is
>>> reproducible with two different kernels I tried (with different configs
>>> and versions).  I tried both versions of the virtio-vhost-user I was
>>> able to find[1][2], and both exhibited the same behaviour.
>>> 
>>> Is this a known issue?  Am I doing something wrong?
>>
>> Hi,
>> Unfortunately I'm not sure what the issue is. This is an early
>> virtio-pci register access before a driver for any specific device type
>> (net, blk, vhost-user, etc) comes into play.
>
> Small update here: I tried on another computer, and it worked.  Made
> sure that it was exactly the same QEMU binary, command line, and VM
> disk/initrd/kernel, so I think I can fairly confidently say the panic
> depends on what hardware QEMU is running on.  I set -cpu value to the
> same on both as well (SandyBridge).
>
> I also discovered that it works on my primary computer (the one it
> panicked on before) with KVM disabled.
>
> Note that I've only got so far as finding that it boots on the other
> machine -- I haven't verified yet that it actually works.
>
> Bad host CPU:  Intel(R) Core(TM) i5-2520M CPU @ 2.50GHz
> Good host CPU: AMD EPYC 7401P 24-Core Processor
>
> May I ask what host CPUs other people have tested this on?  Having more
> data would probably be useful.  Could it be an AMD vs. Intel thing?

I think I've figured it out!

Sandy Bridge and Ivy Bridge hosts encounter this panic because the
"additional resources" bar size is too big, at 1 << 36.  If I change
this to 1 << 35, no more kernel panic.

Skylake and later are fine with 1 << 36.  In between Ivy Bridge and
Skylake were Haswell and Broadwell, but I couldn't find anybody who was
able to help me test on either of those, so I don't know what they do.

Perhaps related, the hosts that produce panics all seem to have a
physical address size of 36 bits, while the hosts that work have larger
physical address sizes, as reported by lscpu.