On Wed, Jul 13, 2016 at 10:34:07AM +0300, Ilya Maximets wrote:

On 12.07.2016 08:53, Ilya Maximets wrote:

quoted

On 12.07.2016 05:43, Yuanhan Liu wrote:

quoted

On Mon, Jul 11, 2016 at 02:47:56PM +0300, Ilya Maximets wrote:

quoted

On 11.07.2016 14:05, Yuanhan Liu wrote:

quoted

On Mon, Jul 11, 2016 at 12:50:24PM +0300, Ilya Maximets wrote:

quoted

On 11.07.2016 11:38, Yuanhan Liu wrote:

quoted

On Sun, Jul 10, 2016 at 09:17:31PM +0800, Yuanhan Liu wrote:

quoted

On Fri, Jul 08, 2016 at 02:48:56PM +0300, Ilya Maximets wrote:

quoted

Another point is that crash constantly happens on queue_id=3 (second RX queue) in
my scenario. It is newly allocated virtqueue while reconfiguration from rxq=1 to
rxq=2.

That's a valuable message: what's your DPDK HEAD commit while triggering
this issue?

fbfd99551ca3 ("mbuf: add raw allocation function")

quoted

I guess I have understood what goes wrong in you case.

I would guess that your vhost has 2 queues (here I mean queue-pairs,
including one Tx and Rx queue; below usage is the same) configured,
so does to your QEMU. However, you just enabled 1 queue while starting
testpmd inside the guest, and you want to enable 2 queues by running
following testpmd commands:

    stop
    port stop all
    port config all rxq 2
    port config all txq 2
    port start all

Badly, that won't work for current virtio PMD implementation, and what's
worse, it triggers a vhost crash, the one you saw.

Here is how it comes. Since you just enabled 1 queue while starting
testpmd, it will setup 1 queue only, meaning only one queue's **valid**
information will be sent to vhost. You might see SET_VRING_ADDR
(and related vhost messages) for the other queue as well, but they
are just the dummy messages: they don't include any valid/real
information about the 2nd queue: the driver don't setup it after all.

So far, so good. It became broken when you run above commands. Those
commands do setup for the 2nd queue, however, they failed to trigger
the QEMU virtio device to start the vhost-user negotiation, meaning
no SET_VRING_ADDR will be sent for the 2nd queue, leaving vhost
untold and not updated.

What's worse, above commands trigger the QEMU to send SET_VRING_ENABLE
messages, to enable all the vrings. And since the vrings for the 2nd
queue are not properly configured, the crash happens.

Hmm, why 2nd queue works properly with my fix to vhost in this case?

Hmm, really? You are sure that data flows in your 2nd queue after those
commands? From what I know is that your patch just avoid a crash, but
does not fix it.

Oh, sorry. Yes, it doesn't work. With my patch applied I have a QEMU hang.

The crash actually could be avoided by commit 0823c1cb0a73 ("vhost:
workaround stale vring base"), accidentally. That's why I asked you
above the HEAD commit you were using.

Thanks for pointing this. I'll check it.

I've checked my DPDK HEAD with above commit backported. Yes, it helps to
avoid vhost crash in my scenario. As expected, after reconfiguration new
virtqueue doesn't work, QEMU hangs sometimes.

quoted

quoted

quoted

quoted

quoted

quoted

So maybe we should do virtio reset on port start?

I guess it was removed by this patch:
a85786dc816f ("virtio: fix states handling during initialization").

Seems yes.

Actually, we should not do that: do reset on port start. The right fix
should be allocating MAX queues virtio device supports (2 here). This
would allow us changing the queue number dynamically.

Yes, I agree that this is the right way to fix this issue.
 

quoted

But this doesn't sound a simple fix; it involves many code changes, due
to it was not designed this way before. Therefore, we will not fix it
in this release, due to it's too late. Let's fix it in the next release
instead. For the crash issue, it will not happen with the latest HEAD.
Though it's accident fix, I think we are fine here.

This scenario fixed somehow, I agree. But this patch still needed to protect
vhost from untrusted VM, from malicious or buggy virtio application.
Maybe we could change the commit-message and resend this patch as a
security enhancement? What do you think?

Indeed, but I'm a bit concerned about the performance regression found
by Rich, yet I am not quite sure why it happens, though Rich claimed
that it seems to be a problem related to compiler.

	--yliu