Re: [PATCH 2/5] x86/time: implement tsc as clocksource

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

On 03/23/2016 07:28 AM, Jan Beulich wrote:
>>>> On 22.03.16 at 21:40, <joao.m.martins@oracle.com> wrote:
>> On 03/22/2016 04:02 PM, Jan Beulich wrote:
>>>>>> On 22.03.16 at 16:51, <joao.m.martins@oracle.com> wrote:
>>>> On 03/22/2016 12:46 PM, Jan Beulich wrote:
>>>>>>>> On 22.03.16 at 13:41, <joao.m.martins@oracle.com> wrote:
>>>>>>>
>>>>>> I found out one issue in the tsc clocksource initialization path with respect to
>>>>>> the reliability check. This check is running when initializing platform timer,
>>>>>> but not all CPUS are up at that point (init_xen_time()) which means that the
>>>>>> check will always succeed. So for clocksource=tsc I need to defer initialization
>>>>>> to a later point (on verify_tsc_reliability()) and if successful switch to that
>>>>>> clocksource. Unless you disagree, v2 would have this and just requires one
>>>>>> additional preparatory change prior to this patch.
>>>>>
>>>>> Hmm, that's suspicious when thinking about CPU hotplug: What
>>>>> do you intend to do when a CPU comes online later, failing the
>>>>> check?
>>>> Good point, but I am not sure whether that would happen. The initcall
>>>> verify_tsc_reliability seems to be called only for the boot processor. Or are
>>>> you saying that it's case that initcalls are called too when hotplugging cpus
>>>> later on? If that's the case then my suggestion wouldn't work as you point out -
>>>> or rather without having runtime switching support as you point out below.
>>>
>>> Looks like I didn't express myself clearly enough: "failing the check"
>>> wasn't meant to imply the failure would actually occur, but rather
>>> that failure would occur if the check was run on that CPU. IOW the
>>> case of a CPU getting hotplugged which doesn't satisfy the needs
>>> for selecting TSC as the clock source.
>> Ah, I see. I believe this wouldn't be an issue with the current rendezvous
>> mechanism (std_rendezvous), as the delta is computed between local_tsc_stamp
>> taken in that (hotplugged) CPU in the calibration and the rdtsc() in the 
>> guest
>> same CPU, even though having CPU0 TSC (master) as system_time.
>>
>> However it can be a problem with PVCLOCK_TSC_STABLE_BIT as the hotplugged CPU
>> could have its TSC drifted, and since setting this flag relies on
>> synchronization of TSCs we would need to clear the flag enterily.
> 
> Except that we can't, after guests already got started, validly clear
> that flag afaics.
Correct.

> The only option I see here would be to never set
> this flag if CPU hotplug is possible - by looking at the hot pluggable
> CPU count and, if non-zero, perhaps allowing a command line
> override to indicate no hotplug is intended (it may well be that such
> is already implicitly available).
OK, will add this then to allow the flag only if the conditions above are met.
Thanks for the pointer!

>>>>> So far we don't do runtime switching of the clock source,
>>>>> and I'm not sure that's a good idea to do when there are running
>>>>> guests.
>>>> Totally agree, but I would be proposing to be at initialization phase where
>>>> there wouldn't be guests running. We would start with HPET, then only switch 
>>>> to
>>>> TSC if that check has passed (and would happen once).
>>>>
>>>> Simpler alternative would be to call init_xen_time after all CPUs are 
>>>> brought up
>>>> (and would also keep this patch as is), but I am not sure about the
>>>> repercussions of that.
>>>
>>> I don't see how either would help with the hotplug case.
>> This was in response to what I thought you meant with your earlier question
>> (which I misunderstood). But my question is still valid I believe. The 
>> reason
>> for choosing between my suggested approaches is that tsc_check_reliability()
>> requires all CPUs up so that the check is correctly performed.
> 
> Sure - it seems quite obvious that all boot time available CPUs
> should be checked.
Cool, so I will go with moving init_xen_time right after all CPUs are up but
before initcalls are invoked.

Joao

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