Re: [Qemu-devel] [PATCH] cpus: reset throttle_thread_scheduled after sleep

[Felipe Franciosi]

> On 1 Jun 2017, at 15:36, Dr. David Alan Gilbert <address@hidden> wrote:
> 
> * Jason J. Herne (address@hidden) wrote:
>> On 05/19/2017 05:29 PM, Felipe Franciosi wrote:
>>> Currently, the throttle_thread_scheduled flag is reset back to 0 before
>>> sleeping (as part of the throttling logic). Given that throttle_timer
>>> (well, any timer) may tick with a slight delay, it so happens that under
>>> heavy throttling (ie. close or on CPU_THROTTLE_PCT_MAX) the tick may
>>> schedule a further cpu_throttle_thread() work item after the flag reset,
>>> but before the previous sleep completed. This results on the vCPU thread
>>> sleeping continuously for potentially several seconds in a row.
>>> 
>>> The chances of that happening can be drastically minimised by resetting
>>> the flag after the sleep.
>>> 
>>> Signed-off-by: Felipe Franciosi <address@hidden>
>>> Signed-off-by: Malcolm Crossley <address@hidden>
>>> ---
>>> cpus.c | 2 +-
>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>> 
>>> diff --git a/cpus.c b/cpus.c
>>> index 516e5cb..f42eebd 100644
>>> --- a/cpus.c
>>> +++ b/cpus.c
>>> @@ -677,9 +677,9 @@ static void cpu_throttle_thread(CPUState *cpu, 
>>> run_on_cpu_data opaque)
>>>     sleeptime_ns = (long)(throttle_ratio * CPU_THROTTLE_TIMESLICE_NS);
>>> 
>>>     qemu_mutex_unlock_iothread();
>>> -    atomic_set(&cpu->throttle_thread_scheduled, 0);
>>>     g_usleep(sleeptime_ns / 1000); /* Convert ns to us for usleep call */
>>>     qemu_mutex_lock_iothread();
>>> +    atomic_set(&cpu->throttle_thread_scheduled, 0);
>>> }
>>> 
>>> static void cpu_throttle_timer_tick(void *opaque)
>>> 
>> 
>> This seems to make sense to me.
>> 
>> Acked-by: Jason J. Herne <address@hidden>
>> 
>> I'm CC'ing Juan, Amit and David as they are all active in the migration area
>> and may have
>> opinions on this. Juan and David were also reviewers for the original
>> series.
> 
> The description is interesting and sounds reasonable; it'll be
> interesting to see what difference it makes to the autoconverge
> behaviour for those workloads that need this level of throttle.

To get some hard data, we wrote a little application that:
1) spawns multiple threads (one per vCPU)
2) each thread mmap()s+mlock()s a certain workset (eg. 30GB/#threads for a 32GB 
VM)
3) each thread writes a word to the beginning of every page in a tight loop
4) the parent thread periodically reports the number of dirtied pages

Even on a dedicated 10G link, that is pretty much guaranteed to require 99% 
throttle to converge.

Before the patch, Qemu migrates the VM (depicted above) fairly quickly (~40s) 
after reaching 99% throttle. The application reported a few seconds at a time 
with lockups which we initially thought was just that thread not running 
between Qemu-induced vCPU sleeps (and later attributed it to the reported bug).

Then we used a 1G link. This time, the migration had to run for a lot longer 
even at 99%. That made the bug more likely to happen and we observed soft 
lockups (reported by the guest's kernel on the console) of 70+ seconds.

Using the patch, and back on a 10G link, the migration completes after a few 
more iterations than before (took just under 2mins after reaching 99%). If you 
want further validation of the bug, instrumenting 
cpus-common.c:process_queued_cpu_work() could be done to show that 
cpu_throttle_thread() is running back-to-back under these cases.

In summary we believe this patch is immediately required to prevent the 
lockups. A more elaborate throttling solution should be considered as future 
work. Perhaps a per-vCPU timer which throttles more precisely or a new 
convergence design altogether.

Thanks,
Felipe

> 
> Dave
> 
>> -- 
>> -- Jason J. Herne (address@hidden)
>> 
> --
> Dr. David Alan Gilbert / address@hidden / Manchester, UK