[RFC PATCH] perf cs-etm: Handle valid-but-zero timestamps

Leo Yan leo.yan at linaro.org
Tue May 11 01:26:31 PDT 2021


Hi Denis,

On Tue, May 11, 2021 at 01:06:03AM -0700, Denis Nikitin wrote:
> Hi Leo,
> 
> > Just remind, as Mike has mentioned that if the timestamp is zero, it
> > means the hardware setting for timestamp is not enabled properly.  So
> > for system wide or per CPU mode tracing, it's better to double check
> > what's the reason the timestamp is not enabled properly.
> 
> The bug is confirmed by HW verification.

Yeah.

> > IIUC, this patch breaks the existed rational in the code.  Let's think
> > about there have 4 CPUs, every CPU has its own AUX trace buffer, and
> > when decode the trace data, it will use 4 queues to track the packets
> > and every queue has its timestamp.
> >
> >   CPU0: cs_etm_queue -> ... -> packet_queue->timestamp
> >   CPU1: cs_etm_queue -> ... -> packet_queue->timestamp
> >   CPU2: cs_etm_queue -> ... -> packet_queue->timestamp
> >   CPU3: cs_etm_queue -> ... -> packet_queue->timestamp
> >
> > The issue is if all CPUs' timestamp are zero, it's impossible to find
> > a way to synthesize samples in the right time order.
> 
> Is it really impossible or it just can lead to incorrect decoding?

Thanks for correcting.  Just clarifying: with this change, perf can
decode and synthesize samples, but the sequence of samples is not
time-based ordering.

> I verified the profiles generated with zero timestamps and this patch
> on Trogdor (8 CPU cores) and I don't see any significant difference
> in the quality of the AutoFDO profiles.
> 
> If mixed packets don't cause errors in reconstructing the branches
> but instead mess up with their timeline then it probably won't matter
> for AutoFDO which just collects statistics of the branches.
> What do you think?

Understand.

CoreSight trace data can be used for two purposes: tracing and
profiling.  For AutoFDO profiling, it's okay for with zero timestamps
based on your conclusion; on the other hand, the change can introduce
confusion if any user wants to use CoreSight for tracing and review the
program flow (like use "perf script") command.

The change in this patch is valid for me, but it's better to connect
with a new option (like "--force-aux-ts-zero" mentioned in my another
replying), this can allow users to explictly to force AUX trace
timestamp as zero (or in other word, users can use this option to ignore
AUX timestamp if the timestamp is not reliable).

Thanks,
Leo



More information about the linux-arm-kernel mailing list