Re: Is: events not being cleared during fast migration over InfiniBand Was: Re: xen 4.3 test report

[George Dunlap <george.dunlap@eu.citrix.com>]

On 05/06/13 19:50, Konrad Rzeszutek Wilk wrote:
> On Tue, Jun 04, 2013 at 04:17:55PM +0400, Vasiliy Tolstov wrote:
>> 2013/6/3 Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>:
>>> The non-debug version tells me it is:
>>>
>>> 289     if ( (port = get_free_port(d)) < 0 )
>>> 290         ERROR_EXIT(port);
>>>
>>> Which gets -EEXIST from get_free_port. But get_free_port only returns
>>> -EINVAL, -ENOMEM, and -ENOSPC in failure modes.
>>>
>>> But we get -EEXIST? Could you re-run git diff and attach output to
>>> this email? I think you tweaked the debug code a bit so I am looking
>>> at something different?
>>
>> Oh sorry. Yes i modify you patch to this version:
> That is OK.
>> -    if ( v->virq_to_evtchn[virq] != 0 )
>> +    if ( v->virq_to_evtchn[virq] != 0 ) {
>> +       gdprintk(XENLOG_WARNING, "d%dv%d [%s:%d], port:%d, rc:%ld\n",
>> d->domain_id,
>> +               vcpu, __func__,__LINE__, v->virq_to_evtchn[virq],
>> (long)-EEXIST);
>>           ERROR_EXIT(-EEXIST);
> OK, so the value was 3 (event channel), and I am not sure what the virq value
> was. But it looks as if somebody did not clear that and we are
> tripping over it.
>
> George, have you seen issues with events not being cleared during migration?

I haven't, no.  Do you know where the virq is supposed to be clear? The 
BUG() is in restore_cpu_virqs(), but at a quick glance I can't find a 
corresponding function to tear down virqs.

  -George