Re: [PATCH v2 3/3] xen/sched: fix latent races accessing vcpu->dirty_cpu

[Jan Beulich <jbeulich@suse.com>]

On 14.05.2020 10:50, Jürgen Groß wrote:
> On 14.05.20 09:10, Jan Beulich wrote:
>> On 11.05.2020 13:28, Juergen Gross wrote:
>>> @@ -1956,13 +1958,17 @@ void sync_local_execstate(void)
>>>     void sync_vcpu_execstate(struct vcpu *v)
>>>   {
>>> -    if ( v->dirty_cpu == smp_processor_id() )
>>> +    unsigned int dirty_cpu = read_atomic(&v->dirty_cpu);
>>> +
>>> +    if ( dirty_cpu == smp_processor_id() )
>>>           sync_local_execstate();
>>> -    else if ( vcpu_cpu_dirty(v) )
>>> +    else if ( is_vcpu_dirty_cpu(dirty_cpu) )
>>>       {
>>>           /* Remote CPU calls __sync_local_execstate() from flush IPI handler. */
>>> -        flush_mask(cpumask_of(v->dirty_cpu), FLUSH_VCPU_STATE);
>>> +        flush_mask(cpumask_of(dirty_cpu), FLUSH_VCPU_STATE);
>>>       }
>>> +    ASSERT(!is_vcpu_dirty_cpu(dirty_cpu) ||
>>> +           read_atomic(&v->dirty_cpu) != dirty_cpu);
>>
>> Repeating my v1.1 comments:
>>
>> "However, having stared at it for a while now - is this race
>>   free? I can see this being fine in the (initial) case of
>>   dirty_cpu == smp_processor_id(), but if this is for a foreign
>>   CPU, can't the vCPU have gone back to that same CPU again in
>>   the meantime?"
>>
>> and later
>>
>> "There is a time window from late in flush_mask() to the assertion
>>   you add. All sorts of things can happen during this window on
>>   other CPUs. IOW what guarantees the vCPU not getting unpaused or
>>   its affinity getting changed yet another time?"
>>
>> You did reply that by what is now patch 2 this race can be
>> eliminated, but I have to admit I don't see why this would be.
>> Hence at the very least I'd expect justification in either the
>> description or a code comment as to why there's no race left
>> (and also no race to be expected to be re-introduced by code
>> changes elsewhere - very unlikely races are, by their nature,
>> unlikely to be hit during code development and the associated
>> testing, hence I'd like there to be sufficiently close to a
>> guarantee here).
>>
>> My reservations here may in part be due to not following the
>> reasoning for patch 2, which therefore I'll have to rely on the
>> scheduler maintainers to judge on.
> 
> sync_vcpu_execstate() isn't called for a running or runnable vcpu any
> longer. I can add an ASSERT() and a comment explaining it if you like
> that better.

This would help (hopefully people adding new uses of the function
would run into this assertion/comment), but for example the uses
in mapcache_current_vcpu() or do_tasklet_work() look to be pretty
hard to prove they can't happen for a runnable vCPU.

Jan