Re: [PATCH] perf python: Set up callchain parameters for sessions

From: Samuel Wu

Date: Mon Nov 09 2026 - 14:03:09 EST


On Sun, Oct 18, 2026 at 1:14 AM Lalit Maganti <lalitmaganti@xxxxxxxxx> wrote:
>
> perf.session never calls callchain_param_setup(), so callchains are
> resolved with callchain_param.record_mode left as CALLCHAIN_NONE rather
> than set from the sample type as in perf script and perf report.
>
> On arm64 this loses frames: when a sample hits a function before it has
> pushed its frame record, its caller is only in the link register, and
> recovering it needs record_mode CALLCHAIN_FP. Comparing the callchain of
> every sample with perf script, on a frame pointer recording of a
> two-thread workload on arm64, 4 of 4494 samples miss a frame, e.g.:
>
> perf script: clock_gettime@plt <- now <- spin <- other_path
> perf.session: clock_gettime@plt <- spin <- other_path
>
> Call callchain_param_setup() once the session is created and, as perf
> script does, again as attrs are processed, for pipe mode. With that,
> all callchains match perf script on arm64, for frame pointer and DWARF
> recordings, from a file or a pipe. On x86 they matched before and still
> do.
>
> Assisted-by: LLM
> Signed-off-by: Lalit Maganti <lalitmaganti@xxxxxxxxx>
> ---
> tools/perf/util/python.c | 22 +++++++++++++++++++++-
> 1 file changed, 21 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/util/python.c b/tools/perf/util/python.c
> index 60dbe2dad5eb..92d18e24049d 100644
> --- a/tools/perf/util/python.c
> +++ b/tools/perf/util/python.c
> @@ -4842,6 +4842,24 @@ static int __pyrf_session_tool__stat_round(const struct perf_tool *tool,
> */
> static struct pyrf_session *pyrf_session__active;
>
> +/*
> + * In pipe mode attrs are only known once processed, so set up callchains again,
> + * as perf script does.
> + */
> +static int pyrf_session__process_attr(const struct perf_tool *tool,
> + union perf_event *event,
> + struct evlist **pevlist)
> +{
> + int err = perf_event__process_attr(tool, event, pevlist);
> +
> + if (err)
> + return err;
> + callchain_param_setup(evlist__combined_sample_type(*pevlist),
> + perf_session__e_machine(evsel__session(evlist__last(*pevlist)),
> + /*e_flags=*/NULL));
> + return 0;
> +}
> +
> static bool pyrf_session__check_not_busy(struct pyrf_session *psession)
> {
> if (pyrf_session__active == psession &&
> @@ -5061,7 +5079,7 @@ static PyObject *pyrf_session__new(PyTypeObject *type, PyObject *args, PyObject
> psession->tool.ksymbol = perf_event__process_ksymbol;
> psession->tool.text_poke = perf_event__process_text_poke;
> psession->tool.build_id = perf_event__process_build_id;
> - psession->tool.attr = perf_event__process_attr;
> + psession->tool.attr = pyrf_session__process_attr;
> psession->tool.event_update = perf_event__process_event_update;
> psession->tool.feature = perf_event__process_feature;
> #ifdef HAVE_LIBTRACEEVENT
> @@ -5139,6 +5157,8 @@ static PyObject *pyrf_session__new(PyTypeObject *type, PyObject *args, PyObject
> PyErr_SetString(PyExc_OSError, "perf: symbol__init failed");
> goto err_out;
> }
> + callchain_param_setup(evlist__combined_sample_type(session->evlist),
> + perf_session__e_machine(session, /*e_flags=*/NULL));
> if (psession->symfs && !symbol_conf.symfs[0])
> zfree(&psession->symfs);
>
>
> base-commit: 1c90f0a3c8a1c4e38f9edd97af9e20848805a7fb
> --
> 2.56.0
>

I think we can add:
Fixes: 88439191ad5e ("perf python: Add callchain support")

But otherwise, looks good to me. I did a quick test on next-20261109
with an ARM64 Pixel device using a workload whose hot path has a leaf
function without a frame record, and got the expected results.

Reviewed-by: Samuel Wu <wusamuel@xxxxxxxxxx>
Tested-by: Samuel Wu <wusamuel@xxxxxxxxxx>