Re: [PATCH] x86/time: use correct (local) time stamp in constant-TSC calibration fast path

[Joao Martins <joao.m.martins@oracle.com>]

[changing Dario address to citrix.com as it was bouncing for me ]

On 06/09/2016 04:52 PM, Jan Beulich wrote:
>>>> On 09.06.16 at 17:00, <joao.m.martins@oracle.com> wrote:
>> On 06/09/2016 01:57 PM, Jan Beulich wrote:
>>>>>> On 09.06.16 at 14:11, <joao.m.martins@oracle.com> wrote:
>>> So in effect for the fast path the patch
>>> changes the situation from c->stime_local_stamp being effectively
>>> unused to c->stime_master_stamp being so. In the former case, if
>>> that really hadn't been a typo, deleting the write of that field from
>>> time_calibration_std_rendezvous() would have made sense, as
>>> get_s_time() certainly is more overhead than the simply memory
>>> read and write needed for keeping c->stime_master_stamp up to
>>> date (despite not being used).
>> I agree, but what I meant previously was more of a concern meaning: CPU 0 is 
>> doing an
>> expensive read_platform_time (e.g. HPET supposedly microseconds order, plus 
>> a
>> non-contented lock) to set stime_master_stamp that doesn't get used at all -
>> effectively not using the clocksource set initially at boot.
> 
> Yeah, there's likely some cleanup potential here, but of course we
> need to be pretty careful about not doing something that may be
> needed by some code paths but not others. But if you think you
> can help the situation without harming maintainability, feel free to
> go ahead.
> 
OK, Makes sense. I'll likely do already so of it on my related series.

>> What if verify_tsc_reliability clears out X86_FEATURE_TSC_RELIABLE when 
>> failing
>> the warp test? The calibration function is set early on right after 
>> interrupts are
>> enabled and the time warp check later on when all CPUs are up. So on 
>> problematic
>> platforms it's possible that std_rendezvous is used with a constant TSC but 
>> still
>> deemed unreliable. We still keep incrementing deltas at roughly about the 
>> same time,
>> but in effect with this change the stime_local_stamp would be TSC-only based 
>> thus
>> leading to warps with an unreliable TSC? And there's also the CPU 
>> hotplug/onlining
>> case that we once discussed.
> 
> I agree that we're likely in trouble in such a case. But for the
> moment I'd be glad if we could get the "normal" case work right.
> 
OK. Apologies for the noise, I was just pointing out things that I tried and some
also discussed here in the PVCLOCK_TSC_STABLE_BIT series, although didn't cross me
that Xen own idea of time could be a little broken. IMO adding another clocksource
for TSC would be more correct if we are only using TSC (and having its associated
limitations made aware/explicit to the user) rather then being on the back of another
clocksource in use. But it wouldn't cover the normal case :( unless set manually

NB: Guests on the other hand aren't affected since they take care of keeping the
latest stamp when different vCPUS slightly diverge.

Joao

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