Re: Shutdown panic in disable_nonboot_cpus after cpupool-numa-split

[Juergen Gross <jgross@suse.com>]

On 07/07/2014 04:08 PM, Stefan Bader wrote:
> On 07.07.2014 15:03, Jürgen Groß wrote:
>> On 07/07/2014 02:49 PM, Stefan Bader wrote:
>>> On 07.07.2014 14:38, Jürgen Groß wrote:
>>>> On 07/07/2014 02:00 PM, Andrew Cooper wrote:
>>>>> On 07/07/14 12:33, Stefan Bader wrote:
>>>>>> I recently noticed that I get a  panic (rebooting the system) on shutdown in
>>>>>> some
>>>>>    > cases. This happened only on my AMD system and also not all the time.
>>>>> Finally
>>>>>    > realized that it is related to the use of using cpupool-numa-split
>>>>>    > (libxl with xen-4.4 maybe, but not 100% sure 4.3 as well).
>>>>>    >
>>>>>    > What happens is that on shutdown the hypervisor runs
>>>>> disable_nonboot_cpus which
>>>>>    > call cpu_down for each online cpu. There is a BUG_ON in the code for
>>>>> the case of
>>>>>    > cpu_down returning -EBUSY. This happens in my case as soon as the
>>>>> first cpu that
>>>>>    > has been moved to pool-1 by cpupool-numa-split is attempted. The error is
>>>>>    > returned by running the notifier_call_chain and I suspect that ends
>>>>> up calling
>>>>>    > cpupool_cpu_remove which always returns EBUSY for cpus not in pool0.
>>>>>    >
>>>>>    > I am not sure which end needs to be fixed but looping over all online
>>>>> cpus in
>>>>>    > disable_nonboot_cpus sounds plausible. So maybe the check for pool-0 in
>>>>>    > cpupool_cpu_remove is wrong...?
>>>>>    >
>>>>>    > -Stefan
>>>>>
>>>>> Hmm yes - this looks completely broken.
>>>>>
>>>>> cpupool_cpu_remove() only has a single caller which is from cpu_down(),
>>>>> and will unconditionally fail for cpus outside of the default pool.
>>>>>
>>>>> It is not obvious at all how this is supposed to work, and the comment
>>>>> beside cpupool_cpu_remove() doesn't help.
>>>>>
>>>>> Can you try the following (only compile tested) patch, which looks
>>>>> plausibly like it might DTRT.  The for_each_cpupool() is a little nastly
>>>>> but there appears to be no cpu_to_cpupool mapping available.
>>>>
>>>> Your patch has the disadvantage to support hot-unplug of the last cpu in
>>>> a cpupool. The following should work, however:
>>>
>>> Disadvantage and support sounded a bit confusing. But I think it means
>>> hot-unplugging the last cpu of a pool is bad and should not be working.
>>
>> Correct.
>>
>>>
>>>>
>>>> diff --git a/xen/common/cpupool.c b/xen/common/cpupool.c
>>>> index 4a0e569..73249d3 100644
>>>> --- a/xen/common/cpupool.c
>>>> +++ b/xen/common/cpupool.c
>>>> @@ -471,12 +471,24 @@ static void cpupool_cpu_add(unsigned int cpu)
>>>>     */
>>>>    static int cpupool_cpu_remove(unsigned int cpu)
>>>>    {
>>>> -    int ret = 0;
>>>> +    int ret = -EBUSY;
>>>> +    struct cpupool **c;
>>>>
>>>>        spin_lock(&cpupool_lock);
>>>> -    if ( !cpumask_test_cpu(cpu, cpupool0->cpu_valid))
>>>> -        ret = -EBUSY;
>>>> +    if ( cpumask_test_cpu(cpu, cpupool0->cpu_valid) )
>>>> +        ret = 0;
>>>>        else
>>>> +    {
>>>> +        for_each_cpupool(c)
>>>> +        {
>>>> +            if ( cpumask_test_cpu(cpu, (*c)->cpu_suspended ) )
>>>
>>> The rest seems to keep the semantics the same as before (though does that mean
>>> unplugging the last cpu of pool-0 is ok?) But why testing for suspended here to
>>> succeed (and not valid)?
>>
>> Testing valid would again enable to remove the last cpu of a cpupool in
>> case of hotplugging. cpu_suspended is set if all cpus are to be removed
>> due to shutdown, suspend to ram/disk, ...
>
> Ah, ok. Thanks for the detail explanation. So I was trying this change in
> parallel and can confirm that it gets rid of the panic on shutdown. But when I
> try to offline any cpu in pool1 (if echoing 0 into /sys/devices/xen_cpu/xen_cpu?
> is the correct method) I always get EBUSY. IOW I cannot hot-unplug any cpu that
> is in a pool other than 0. It does only work after removing it from pool1, then
> add it to pool0 and then echo 0 into online.

That's how it was designed some years ago. I don't want to change the
behavior in the hypervisor. Adding some tool support could make sense,
however.

Juergen