On 08/21/2014 02:28 PM, Zhangjie (HZ) wrote:
> On 2014/8/21 12:29, Jason Wang wrote:
>> On 08/20/2014 05:23 PM, Zhangjie (HZ) wrote:
>>> On 2014/8/19 12:56, Jason Wang wrote:
>>>> commit a9f98bb5ebe6fb1869321dcc58e72041ae626ad8 vhost: multiqueue
>>> call it before setting
>>>> Zhang Jie, please test this patch to see if it fixes the issue.
>>>> +static void vhost_net_set_vq_index(struct vhost_net *net, int vq_index)
>>>> +{
>>>> + net->dev.vq_index = vq_index;
>>>> +}
>>> int vq_index)
>>>> ...
>>> Because of vhost_net_set_vq_index, VM can be start successfully.
>>> But, after about 80 times of migration under my environment, virtual nic
>>> became unreachable again.
>>> When I use jprobe to notify tap, the virtual nic becomes reachable again.
>>> This shows that interrupts missing causes
>>> the problem.
>> Thanks for the testing. A questions is can you reproduce this when vhost
>> is disabled?
> After migration, vhost is not disabled, virtual nic became unreachable
> because vhost is not awakened.
> By the logical of EVENT_IDX, virtio-net will not kick vhost again if the used
> idx is not updated.
> So, if one interrupts is lost during migration, virtio_net will not kick
> vhost again.
> Then, no skb from virtio-net can be sent to tap.
Yes and I mean to test vhost=off to see if it was the issue of vhost.
>
> Jason's patch reduced the probability of occurrence, from about 1/20 to 1/80.
> It is really effective. I think the patch should be acked.
> May be we can try to solve the problem from another perspective. Do you have
> some methods to sense the migration?
> We can make up a signal from virtio-net after the migration.
You can make a patch like this to debug. If problem disappears, it means
interrupt was really lost anyway.
>
>> Anyway, I will try to reproduce it by myself.
>>
> The test environment is really terrible, I build a environment myself, but it
> problem did not occur.
> The environment I use now is from a colleague Responsible for test work.
> Two hosts, every host has about 20 vms, they send packages(ipv4 and ipv6)
> between each other.
> The VM to be migrated also sens packages itself, and there is a ping(-i
> 0.001) from another host to it.
> The physical nic is 1GE, connected through a internal nework.
Just want to confirm. For the problem did not occur, you mean with my
patch on top?
