Re: [PATCH] x86/vpmu: fix race-condition in vpmu_load

[Jan Beulich <jbeulich@suse.com>]

On 19.09.2022 16:11, Tamas K Lengyel wrote:
> On Mon, Sep 19, 2022 at 9:58 AM Jan Beulich <jbeulich@suse.com> wrote:
> 
>> On 19.09.2022 15:24, Tamas K Lengyel wrote:
>>> On Mon, Sep 19, 2022 at 9:21 AM Jan Beulich <jbeulich@suse.com> wrote:
>>>>
>>>> On 19.09.2022 14:25, Tamas K Lengyel wrote:
>>>>> On Mon, Sep 19, 2022 at 5:28 AM Jan Beulich <jbeulich@suse.com> wrote:
>>>>>>
>>>>>> On 16.09.2022 23:35, Boris Ostrovsky wrote:
>>>>>>>
>>>>>>> On 9/16/22 8:52 AM, Jan Beulich wrote:
>>>>>>>> On 15.09.2022 16:01, Tamas K Lengyel wrote:
>>>>>>>>> While experimenting with the vPMU subsystem an ASSERT failure was
>>>>>>>>> observed in vmx_find_msr because the vcpu_runnable state was true.
>>>>>>>>>
>>>>>>>>> The root cause of the bug appears to be the fact that the vPMU
>> subsystem
>>>>>>>>> doesn't save its state on context_switch.
>>>>
>>>> For the further reply below - is this actually true? What is the
>>>> vpmu_switch_from() (resolving to vpmu_save()) doing then early
>>>> in context_switch()?
>>>>
>>>>>>>>> The vpmu_load function will attempt
>>>>>>>>> to gather the PMU state if its still loaded two different ways:
>>>>>>>>>      1. if the current pcpu is not where the vcpu ran before doing
>> a remote save
>>>>>>>>>      2. if the current pcpu had another vcpu active before doing a
>> local save
>>>>>>>>>
>>>>>>>>> However, in case the prev vcpu is being rescheduled on another
>> pcpu its state
>>>>>>>>> has already changed and vcpu_runnable is returning true, thus #2
>> will trip the
>>>>>>>>> ASSERT. The only way to avoid this race condition is to make sure
>> the
>>>>>>>>> prev vcpu is paused while being checked and its context saved.
>> Once the prev
>>>>>>>>> vcpu is resumed and does #1 it will find its state already saved.
>>>>>>>> While I consider this explanation plausible, I'm worried:
>>>>>>>>
>>>>>>>>> --- a/xen/arch/x86/cpu/vpmu.c
>>>>>>>>> +++ b/xen/arch/x86/cpu/vpmu.c
>>>>>>>>> @@ -419,8 +419,10 @@ int vpmu_load(struct vcpu *v, bool_t
>> from_guest)
>>>>>>>>>           vpmu = vcpu_vpmu(prev);
>>>>>>>>>
>>>>>>>>>           /* Someone ran here before us */
>>>>>>>>> +        vcpu_pause(prev);
>>>>>>>>>           vpmu_save_force(prev);
>>>>>>>>>           vpmu_reset(vpmu, VPMU_CONTEXT_LOADED);
>>>>>>>>> +        vcpu_unpause(prev);
>>>>>>>>>
>>>>>>>>>           vpmu = vcpu_vpmu(v);
>>>>>>>>>       }
>>>>>>>> We're running with IRQs off here, yet vcpu_pause() waits for the
>> vcpu
>>>>>>>> to actually be de-scheduled. Even with IRQs on this is already a
>>>>>>>> relatively heavy operation (also including its impact on the remote
>>>>>>>> side). Additionally the function is called from context_switch(),
>> and
>>>>>>>> I'm unsure of the usability of vcpu_pause() on such a path. In
>>>>>>>> particular: Is there a risk of two CPUs doing this mutually to one
>>>>>>>> another? If so, is deadlocking excluded?
>>>>>>>>
>>>>>>>> Hence at the very least I think the description wants extending, to
>>>>>>>> discuss the safety of the change.
>>>>>>>>
>>>>>>>> Boris - any chance you could comment here? Iirc that's code you did
>>>>>>>> introduce.
>>>>>>>
>>>>>>>
>>>>>>> Is the assertion in vmx_find_msr() really needs to be for runnable
>> vcpu or can it be a check on whether vcpu is actually running (e.g.
>> RUNSTATE_running)?
>>>>>>
>>>>>> You cannot safely check for "running", as "runnable" may transition
>>>>>> to/from "running" behind your back.
>>>>>
>>>>> The more I look at this the more I think the only sensible solution is
>>>>> to have the vPMU state be saved on vmexit for all vCPUs.
>>>>
>>>> Do you really mean vmexit? It would suffice if state was reliably
>>>> saved during context-switch-out, wouldn't it? At that point the
>>>> vCPU can't be resumed on another pCPU, yet.
>>>>
>>>>> That way all
>>>>> this having to figure out where and when a context needs saving during
>>>>> scheduling goes away. Yes, it adds a bit of overhead for cases where
>>>>> the vCPU will resume on the same pCPU and that context saved could
>>>>> have been skipped,
>>>>
>>>> If you really mean vmexit, then I'm inclined to say that's more
>>>> than just "a bit of overhead". I'd agree if you really meant
>>>> context-switch-out, but as said further up it looks to me as if
>>>> that was already occurring. Apparently I'm overlooking something
>>>> crucial ...
>>>
>>> Yes, the current setup is doing exactly that, saving the vPMU context
>>> on context-switch-out, and that's where the ASSERT failure occurs
>>> because the vCPU it needs to save the context for is already runnable
>>> on another pCPU.
>>
>> Well, if that's the scenario (sorry for not understanding it that
>> way earlier on), then the assertion is too strict: While in context
>> switch, the vCPU may be runnable, but certainly won't actually run
>> anywhere. Therefore I'd be inclined to suggest to relax the
>> condition just enough to cover this case, without actually going to
>> checking for "running".
>>
> 
> What ensures the vCPU won't actually run anywhere if it's in the runnable
> state?

The fact that the vCPU is the subject of context_switch().

> And how do I relax the condition just enough to cover just this case?

That's the more difficult question. The immediate solution, passing a
boolean or flag to vpmu_switch_from(), may not be practical, as it
would likely mean passing this through many layers.

Utilizing that I have Jürgen sitting next to me, I've discussed this
with him, and he suggested to simply check for v == current. And
indeed - set_current() in context_switch() happens a few lines after
vpmu_switch_from().

However, going back to vmx_find_msr() I find that the v == current
case is already included there. Which makes me wonder again - what
exactly is the scenario that you're observing the assertion
triggering? Would you mind spelling out the call chain, perhaps by
way of the call stack from the assertion?

Jan