[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] BUG: NOW() seems to (sometimes) go backwards!



>>> On 09.06.16 at 13:25, <JBeulich@xxxxxxxx> wrote:
>>>> On 09.06.16 at 12:24, <joao.m.martins@xxxxxxxxxx> wrote:
>>>>>> Yet when the scaling values get set only once at boot, monotonic
>>>>>> (cross-CPU) TSC means monotonic (cross-CPU) returns from NOW().
>>>>>>
>>>>> Yep. And at this point, this is what needs to be verified, I guess...
>>>> I think get_s_time_fixed doesn't guarantee monotonicity across CPUs being 
>>>> it
>>>> different socket or (SMT) siblings. local_tsc_stamp is seeded differently 
>>>> on 
> 
>> 
>>>> each CPU
>>>> i.e. rdtsc() right after doing the platform time read on the calibration 
>>>> rendezvous.
>>>> Plus stime_local_stamp is seeded with values taken from platform timer 
>>>> (HPET, ACPI,
>>>> PIT) on local_time_calibration which means that get_s_time isn't solely 
>>>> based on TSC
>>>> and that there will always be a gap between stime_local_stamp and 
>>>> local_tsc_stamp.
>>>> TSC is indeed monotonic on a TSC invariant box, but the delta that is 
>>>> computed
>>>> differs from cpu to cpu according to when time calibration happens on each 
>>>> CPU - thus
>>>> not guaranteeing the desired monotonicity property. Having 
>>>> stime_local_stamp 
> 
>> 
>>>> be based
>>>> on the same timestamp that of the local_tsc_stamp plus having a single
>>>> local_tsc_stamp as reference would address this behaviour - see also below.
>>> 
>>> The quality of get_s_time_fixed() output indeed heavily depends on
>>> t->local_tsc_stamp and t->stime_local_stamp being a properly
>>> matched pair. Yet in local_time_calibration()'s constant-TSC case,
>>> they're a direct copy from the respective cpu_calibration fields.
>>> The
>>> main issue I could see here is that on SMT siblings the hardware
>>> switching between the two may introduce arbitrary delays between
>>> them. And with CPU frequency changes, the latency between the
>>> rdtsc() and the execution of get_s_time() finishing could also be
>>> pretty variable. I wonder whether c->local_tsc_stamp wouldn't
>>> better be written with the TSC value used by get_s_time() (or,
>>> which should amount to the same effect, whether we shouldn't
>>> simply call get_s_time_fixed() here with the just sampled TSC value).
>> Indeed, but notice that in this copy for the constant-TSC case:
>> t->stime_local_stamp is written with c->stime_master_stamp - ending
>> up the former being discarded. So even changing that pair to correctly
>> match it wouldn't change the result. At point of which I wonder if copying
>> rendezvous c->stime_master_stamp to t->stime_local_stamp on
>> local_time_calibration is correct?
> 
> Yeah, I stumbled across that too, and I am in the process of
> trying whether this in fact was just a copy-and-paste mistake
> many years ago.

And indeed changing that makes all of the dozens-of-microseconds
backwards moves (compared cross-CPU) go away.

The largest ones left are now around 3.7us, and apparently always
between CPU0 and CPU1 (which I suspect may be due to the special
role CPU0 plays here).

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.