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

[Liang, Kan]

On 2023-01-24 4:10 a.m., Stephane Eranian wrote:
> On Tue, Jan 24, 2023 at 12:52 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>>
>> On Mon, Jan 23 2023 at 10:27, kan liang wrote:
>>> From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>>>
>>> It's useful to provide a NMI safe function to convert a given time to
>>> monotonic. For example, the perf_event subsystem wants to convert a TSC
>>> of a PEBS record to a monotonic clock in a NMI handler.
>>
>> Why? That's a postprocessing problem, really.
>>

Besides the reason Stephane mentioned, the current perf tool assumes
that kernel always gives the time it asked. For example, without the
patch, the kernel uses ktime_get_mono_fast_ns() to get the MONOTONIC
time and dump it to user space.

If we want to break the assumption, I think we have to modify the
current interfaces and dump more extra data to indicate the clock
source. That makes the existing interface more complex. Preparing and
dumping the extra data also brings overhead. We also have to force the
user to update their tools.


Thanks,
Kan

> Because you want to correlate samples captured by PEBS with samples
> from applications timestamped with a user available clock such as
> CLOCK_MONOTONIC, for instance.
> When I create a perf_event event and I stipulate that
> event_attr.clockid=MONOTONIC, I expect all the samples from
> that event to be timestamped using the same clock source, regardless
> of PEBS or IBS.
> 
> 
> 
>> Thanks,
>>
>>         tglx