Message-ID: <20201203063941.GE28939@leoy-ThinkPad-X240s>
Date:   Thu, 3 Dec 2020 14:39:41 +0800
From:   Leo Yan <leo.yan@...aro.org>
To:     Will Deacon <will@...nel.org>
Cc:     James Clark <james.clark@....com>,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
        linux-perf-users@...r.kernel.org,
        Mark Rutland <mark.rutland@....com>,
        Al Grant <al.grant@....com>,
        John Garry <john.garry@...wei.com>,
        Suzuki K Poulose <suzuki.poulose@....com>
Subject: Re: [PATCH] drivers/perf: Enable PID_IN_CONTEXTIDR with SPE

Hi Will,

[ + Mathieu ]

On Tue, Dec 01, 2020 at 11:09:36PM +0000, Will Deacon wrote:
> On Tue, Dec 01, 2020 at 12:10:40PM +0800, Leo Yan wrote:
> > On Mon, Nov 30, 2020 at 04:46:51PM +0000, Will Deacon wrote:
> > > On Mon, Nov 30, 2020 at 06:24:54PM +0200, James Clark wrote:
> > > > Enable PID_IN_CONTEXTIDR by default when Arm SPE is enabled.
> > > > This flag is required to get PID data in the SPE trace. Without
> > > > it the perf tool will report 0 for PID which isn't very useful,
> > > > especially when doing system wide profiling or profiling
> > > > applications that fork.
> > > 
> > > Can perf not figure out the pid some other way? (e.g. by tracing context
> > > switches and correlating that with the SPE data?).
> > 
> > For perf 'per-thread' mode, we can use context switch trace event as
> > assisted info to select thread context.  But for "system wide" mode and
> > "snapshot" mode in perf tool, since the trace data is continuous, I
> > think we cannot use context switch trace event to correlate the SPE
> > trace data.
> 
> Is there no way to correlate them with something like CNTVCT?

Good point.  Yes, we can convert CNTVCT to system time; I read the
code in the perf's intel-pt.c and found the timestamp is used to
correlate the auxtrace heap.  I think it's better to dig more for
detailed implementation.

> > > Also, how does this work with pid namespaces?
> > 
> > Here we are studying the implemetation of Intel-PT and Arm CoreSight.
> > 
> > The context ID is stored into the hardware trace data when record;
> > afterwards when perf tool decodes the trace data and detects the
> > packet for context ID, it will select the machine's thread context in
> > perf [1].  Since the perf tool gathers all the threads infomation in
> > perf data file, based on the context ID, it can find the corresponding
> > thread pointer with function machine__find_thread() [2].
> > 
> > Since your question is for "pid namespace", to be honest, I don't know
> > how perf tool to handle any confliction for differrent processes share
> > the same PID, and I am not sure if you are asking CGroup related stuff
> > or not.  If this cannot answer your question, please let me know.
> 
> My point was that the pid value written to CONTEXTIDR is a global pid
> and does not take namespacing into account. If perf is run inside a pid
> namespace, it will therefore not work.

Understand now.

The perf events PERF_RECORD_ITRACE_START/PERF_RECORD_SWITCH/
PERF_RECORD_SWITCH_CPU_WIDE can be used to set pid/tid in perf.  So this
would be a safe way for perf tool running in pid namespace.

Loop in Mathieu, this is a common issue for both Arm SPE and CoreSight
(IIRC, though CoreSight's timestamp is not strictly attaching to Arm arch
timer counter, the trend is to unify this for using arch timer
counter).

I think James could continue to upstream a new patch by following your
suggestion for enabling PID_IN_CONTEXTIDR, eventually, it's a feature
for Arm SPE to record CONTEXTIDR in its packet.

Your questions inspired me, thanks!

Leo

Powered by blists - more mailing lists