lkml.org 
[lkml]   [2017]   [Nov]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v8 6/6] x86/tsc: use tsc early
From
Date
Hi Pavel,

At 11/15/2017 05:46 AM, Pavel Tatashin wrote:
> Hi Dou,
>
> Great comments, my replies below:
>
>>> static inline unsigned long long paravirt_sched_clock(void)
>>> {
>>> - return PVOP_CALL0(unsigned long long, pv_time_ops.sched_clock);
>>> + return PVOP_CALL0(unsigned long long,
>>> pv_time_ops.active_sched_clock);
>>> }
>>>
>>
>> Should in the 5th patch
>
> Actually, it has to be in patch 6, because otherwise patch 5 without
> patch 6 would cause native_sched_clock() to be used even when a
> platform specific clock is set, thus may cause performance regressions
> where it is not optimal to use tsc for clock.
>
Indeed.

>> + tsc_early_disable();
>> + __sched_clock_offset = sched_clock_early() - sched_clock();
>> + pr_info("sched clock early is finished, offset[%lld.%09lds]\n", t, r);
>> +}

s/sched clock early/early sched clock/

>> +#endif /* CONFIG_X86_TSC */
>> Add definitions for the situation of X86_TSC = no :
>>
>> #else /* CONFIG_X86_TSC */
>> static inline void tsc_early_init(unsigned int khz) { }
>> static inline void tsc_early_fini(void) { }
>
> Excellent point, I totally forgot about X86_TSC = no, however, a
> better fix is to simply remove #ifdef CONFIG_X86_TSC from my
> functions. Apparently, even with X86_TSC=no we can use TSC unless

Agree with removing #ifdef CONFIG_X86_TSC

> notsc kernel parameter is passed. This will be in the next patchset.
>
>>
>> According to tsc_early_delay_calibrate(), if (!boot_cpu_has(X86_FEATURE_TSC
>> || !tsc_khz ), tsc_early_init(tsc_khz)
>> will never be called, so here is redundant.
>>
>>> return;
>>> }
>>>
>>> @@ -1302,6 +1385,7 @@ void __init tsc_init(void)
>>> if (!tsc_khz) {
>>> mark_tsc_unstable("could not calculate TSC khz");
>>> setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
>>> + tsc_early_fini();
>>
>>
>> ditto
>
> Right, in both case we still want to call tsc_early_fini(). Because,
> it calls tsc_early_disable() even when tsc_early_init() was never
> called. tsc_early_disable() either sets static branch
> __tsc_early_static to false or changes active_sched_clock to be
> platform specific, depending on CONFIG_PARAVIRT.

Yes, the function names confused me, the actually purpose of
tsc_early_fini() is switching schedule clock to the final one. right?

How about replacing *return;* with *goto final_sched_clock;*

Thanks,
dou.
>
>> BTW, seems you forgot to cc Peter Zijlstra <peterz@infradead.org> in both V7
>> and V8 patchsets.
>
> Thank you for noticing this! I will include Peter when I send out
> patchset version 9.
>
> Thank you,
> Pavel
>
>
>


\
 
 \ /
  Last update: 2017-11-15 04:53    [W:0.067 / U:0.948 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site