Re: [PATCH v2 7/7] Documentation: coresight: Add PID tracing description

[Mike Leach]

Hi,

On Thu, 4 Feb 2021 at 11:08, Suzuki K Poulose <suzuki.poulose@xxxxxxx> wrote:
>
> On 2/4/21 4:09 AM, Leo Yan wrote:
> > Hi Mike,
> >
> > On Wed, Feb 03, 2021 at 05:39:54PM +0000, Mike Leach wrote:
> >
> > [...]
> >
> >>> +2.2) Tracing PID
> >>> +
> >>> +When the kernel is running at EL2 with Virtualization Host Extensions (VHE),
> >>> +perf records CONTEXTIDR_EL2 in the trace data and can be used as PID when
> >>> +decoding; and if the kernel is running at EL1 with nVHE, CONTEXTIDR_EL1 is
> >>> +traced for PID.
> >>> +
> >>
> >> Would this introductory paragraph be better if is explained where the
> >> kernel stores the PID for the different levels, then we logically move
> >> on to how to trace this in perf.
> >>
> >> e.g:-
> >>
> >> "The lernel can be built to write the PID value into the PE ContextID registers.
> >> For a kernel running at EL1, the PID is stored in CONTEXTIDR_EL1.
> >> A PE may implement ARM Virtualisation Host Extensions (VHE), were the
> >> kernel can run at EL2 as a virtualisation host.
> >> In this case the PID value is stored in CONTEXTIDR_EL2.
> >> perf provides PMU options which program the ETM to insert these values
> >> into the trace data."
> >
> > Will in next spin; thanks a lot for writing up!
> >
> >>> +To support tracing PID for the kernel runs at different exception levels,
> >>> +the PMU formats are defined as follow:
> >>> +
> >>> +  "contextid1": Available on both EL1 kernel and EL2 kernel.  When the
> >>> +                kernel is running at EL1, "contextid1" enables the PID
> >>> +                tracing; when the kernel is running at EL2, this enables
> >>> +                tracing the PID of guest applications.
> >>> +
> >>> +  "contextid2": Only usable when the kernel is running at EL2.  When
> >>> +                selected, enables PID tracing on EL2 kernel.
> >>> +
> >>> +  "contextid":  Will be an alias for the option that enables PID
> >>> +                tracing.  I.e,
> >>> +                contextid == contextid1, on EL1 kernel.
> >>> +                contextid == contextid2, on EL2 kernel.
> >>> +
> >>> +The perf tool automatically sets corresponding bit for the "contextid" config,
> >>> +therefore, the user doesn't have to bother which EL the kernel is running.
> >>> +
> >>> +  i.e, perf record -e cs_etm/contextid/u -- uname
> >>> +    or perf record -e cs_etm//u -- uname
> >>> +
> >>> +will always do the "PID" tracing, independent of the kernel EL.
> >>> +
> >>
> >> This is telling me that both cs_etm// and cs_etm/contextid/ have the
> >> same effect - trace PID. Is this correct?
> >
>
> Just to make this clear, this is not a side effect of the patch.

Which is fine - but the documentation should accurately reflect what
is happening on the system.
This is a new paragraph about the PID tracing or otherwise, Even if
some of the effects pre-date this patch, they have to be accurately
communicated.
I am also reading the new paragraph in the context of the rest of the
coresight.rst document - which is a user level document explaining the
basic operation of the coresight system and tools.
This document mentions no other perf command line parameters relevant
to coresight other than the @sink option.It actually calls out to the
OpenCSD docs to provide further information.

> The perf
> tool driver automatically adds the "contextid" tracing and timestamp for
> "system wide" and process bound events, as they traces get mixed into
> the single sink. So these options are added implicitly by the perf tool
> to make the decoding easier.
>

That's fine - I have no problem with contextID trace enabled by
default. Context ID is relatively low overhead - and only emitted at
start of trace  / context changes.
But the explanation of the parameters currently reads as though they
always have an effect - and not putting them in there will omit the
effect - unless you spot the very subtle line at the end.

The user does not need to know about parameters that have no effect!

Perhaps a better approach would be to explain the above - an explicit
statement that "perf will always enable PID/ contextID tracing at the
relevant EL - but for EL2 it is possible to make specific adjustments
using parameters......."

Cheers

Mike


> > Correct.
> >
> >> If so, then contextid, contextid1 and contextid2 have no effect except
> >> in specific EL2 circumstances.
>
> These are required when perf tool may not automatically request them.
> With this series the EL2 is on par with the EL1, where we get the PID
> automatcially in the trace.
>
> And as you rightly said, contextid1, contextid2 are for EL2 specific
> usage.
>
> Cheers
> Suzuki
>
> >
> > Yes, exactly.
> >
> > Thanks,
> > Leo
> >
> >>> +When the kernel is running at EL2 with VHE, if user wants to trace both the
> >>> +PIDs for both host and guest, the two configs "contextid1" and "contextid2"
> >>> +can be set at the same time:
> >>> +
> >>> +  perf record -e cs_etm/contextid1,contextid2/u -- uname
> >>> +
> >>>
> >>
> >>
> >> Regards
> >>
> >> Mike
> >>
> >>
> >>>   Generating coverage files for Feedback Directed Optimization: AutoFDO
> >>>   ---------------------------------------------------------------------
> >>> --
> >>> 2.25.1
> >>>
> >>
> >>
> >> --
> >> Mike Leach
> >> Principal Engineer, ARM Ltd.
> >> Manchester Design Centre. UK
>


-- 
Mike Leach
Principal Engineer, ARM Ltd.
Manchester Design Centre. UK