[0/4] perf cs_etm: Basic support for virtual/kernel timestamps

Message

German Gomez May 4, 2022, 3:02 p.m. UTC

Hi,

This change has a soft dependency on [1], but assuming the name/location
of the new sysfs interface (ts_source) doesn't change, it should be safe
to apply.

The new 'ts_source' interface allows perf to detect if the timestamps in
the trace correspond to the value of CNTVCT_EL0, which we can convert to
a perf timestamp and store it in the instruction and branch samples.

Due to the way the trace is compressed and decoded by OpenCSD, we only
know the precise time of the first instruction in a range, but I think
for now this is better than not having timestamps at all...

Thanks,
German

[1] https://lore.kernel.org/all/[e-mail]/

German Gomez (4):
  perf pmu: Add function to check if a pmu file exists
  perf cs_etm: Keep separate symbols for ETMv4 and ETE parameters
  perf cs_etm: Record ts_source in AUXTRACE_INFO for ETMv4 and ETE
  perf cs_etm: Set the time field in the synthetic samples

 tools/perf/arch/arm/util/cs-etm.c |  89 +++++++++++++++++++--
 tools/perf/util/cs-etm.c          | 126 +++++++++++++++++++++++++-----
 tools/perf/util/cs-etm.h          |  13 ++-
 tools/perf/util/pmu.c             |  17 ++++
 tools/perf/util/pmu.h             |   2 +
 5 files changed, 221 insertions(+), 26 deletions(-)

Comments

Mike Leach May 5, 2022, 9:56 a.m. UTC | #1

Hi German,

On Wed, 4 May 2022 at 16:02, German Gomez <[e-mail]> wrote:
>
> Hi,
>
> This change has a soft dependency on [1], but assuming the name/location
> of the new sysfs interface (ts_source) doesn't change, it should be safe
> to apply.
>
> The new 'ts_source' interface allows perf to detect if the timestamps in
> the trace correspond to the value of CNTVCT_EL0, which we can convert to
> a perf timestamp and store it in the instruction and branch samples.
>
> Due to the way the trace is compressed and decoded by OpenCSD, we only
> know the precise time of the first instruction in a range, but I think
> for now this is better than not having timestamps at all...
>

The output of OpenCSD is reflective of the architecture of the trace hardware.

Timestamps in the trace are always associated with waypoint elements -
primarily branches.
Hardware trace compression results in only these elements being
output, i.e. E / N atoms representing branches taken or not taken.
Instructions between branches have no explicit element appearing in
the trace.
The decode process implies all the instructions between these elements
to form a range of executed instructions, hence the timestamp being
associated with the first instruction in a range.
Moreover, even though we may request a high timestamp rate, output of
other trace is prioritised over timestamp trace, so there is never any
guarantee that we can force timestamps to appear at the start of every
range.

Any attempt to increase number of timestamps in a trace range would
have to be done by some software interpolation mechanism

Regards

Mike


> Thanks,
> German
>
> [1] https://lore.kernel.org/all/[e-mail]/
>
> German Gomez (4):
>   perf pmu: Add function to check if a pmu file exists
>   perf cs_etm: Keep separate symbols for ETMv4 and ETE parameters
>   perf cs_etm: Record ts_source in AUXTRACE_INFO for ETMv4 and ETE
>   perf cs_etm: Set the time field in the synthetic samples
>
>  tools/perf/arch/arm/util/cs-etm.c |  89 +++++++++++++++++++--
>  tools/perf/util/cs-etm.c          | 126 +++++++++++++++++++++++++-----
>  tools/perf/util/cs-etm.h          |  13 ++-
>  tools/perf/util/pmu.c             |  17 ++++
>  tools/perf/util/pmu.h             |   2 +
>  5 files changed, 221 insertions(+), 26 deletions(-)
>
> --
> 2.25.1
>


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