Re: [PATCH 1/3] timekeeping: NMI safe converter from a given time to monotonic

[Liang, Kan]

On 2023-01-24 2:01 a.m., John Stultz wrote:
> On Mon, Jan 23, 2023 at 10:27 AM <kan.liang@xxxxxxxxxxxxxxx> wrote:
>> +int notrace get_mono_fast_from_given_time(int (*get_time_fn)
>> +                                               (struct system_counterval_t *sys_counterval,
>> +                                               void *ctx),
>> +                                         void *ctx,
>> +                                         u64 *mono_ns)
>> +{
>> +       struct system_counterval_t system_counterval;
>> +       struct tk_fast *tkf = &tk_fast_mono;
>> +       u64 cycles, now, interval_start;
>> +       struct tk_read_base *tkr;
>> +       unsigned int seq;
>> +       int ret;
>> +
>> +       do {
>> +               seq = raw_read_seqcount_latch(&tkf->seq);
>> +               tkr = tkf->base + (seq & 0x01);
>> +
>> +               ret = get_time_fn(&system_counterval, ctx);
>> +               if (ret)
>> +                       return ret;
>> +
>> +               /*
>> +                * Verify that the clocksource associated with the given
>> +                * timestamp is the same as the currently installed
>> +                * timekeeper clocksource
>> +                */
>> +               if (tkr->clock != system_counterval.cs)
>> +                       return -EOPNOTSUPP;
>> +               cycles = system_counterval.cycles;
>> +
>> +               /*
>> +                * Check whether the given timestamp is on the current
>> +                * timekeeping interval.
>> +                */
>> +               now = tk_clock_read(tkr);
>> +               interval_start = tkr->cycle_last;
>> +               if (!cycle_between(interval_start, cycles, now))
>> +                       return -EOPNOTSUPP;
> 
> So. I've not fully thought this out, but it seems like it would be
> quite likely that you'd run into the case where the cycle_last value
> is updated and your earlier TSC timestamp isn't valid for the current
> interval. The get_device_system_crosststamp() logic has a big chunk of
> complex code to try to handle this case by interpolating the cycle
> value back in time. How well does just failing in this case work out?
> 

For the case, perf fallback to the time captured in the NMI handler, via
ktime_get_mono_fast_ns().

The TSC in PEBS is captured by HW when the sample was generated. There
should be a small delta compared with the time captured in the NMI
handler. But I think the delta should be acceptable as a backup solution
for the most analysis cases. Also, I don't think the case (the
cycle_last value is updated during the monitoring) should occur very
often either. So I drop the history support to simplify the function.

Thanks,
Kan