Re: [PATCHES v3 0/2] perf tools: Add cached probe type detection for evsel

[Arnaldo Melo <arnaldo.melo@gmail.com>]

On June 17, 2026 7:48:15 PM GMT-03:00, Namhyung Kim <namhyung@kernel.org> wrote:
>On Tue, Jun 16, 2026 at 03:25:44PM -0300, Arnaldo Carvalho de Melo wrote:
>> Hi,
>> 
>> Checking whether an evsel is a kprobe or uprobe currently requires
>> walking the PMU list via evsel__find_pmu() on every call.  This is
>> wasteful when the same evsel is checked repeatedly in hot paths like
>> trace__fprintf_tp_fields().
>> 
>> Patch 1 adds evsel__is_kprobe(), evsel__is_uprobe(), and
>> evsel__is_probe() helpers that resolve the probe type on first call
>> and cache the result in a 3-bit field that fits in existing struct
>> padding.
>> 
>> Detection covers both PMU-based probes (kprobe/uprobe PMU, detected
>> by PMU name) and ftrace-based dynamic probes (kprobes/uprobes/fprobes
>> created via tracefs, which report as PMU "tracepoint").  Ftrace probes
>> are detected by the __probe_ip field that the kernel adds to all
>> dynamic probe formats — this is authoritative regardless of the
>> group/system name the user chose.
>> 
>> Patch 2 is the first user: it guards the __probe_ip field name
>> comparison in perf trace with evsel__is_probe(), so the strcmp is
>> skipped entirely for the common case of non-probe tracepoint events.
>> 
>> Patch 3 adds inline stubs for evsel__field() and evsel__common_field()
>> when libtraceevent is absent.  These were previously declared
>> unconditionally but defined only inside #ifdef HAVE_LIBTRACEEVENT,
>> which was harmless until patch 1 added a call from always-compiled
>> code in evsel.c.
>> 
>> Build-tested with gcc and clang.
>> 
>> Changes in v3:
>>   - Add patch 3: no-libtraceevent stubs for evsel__field() and
>>     evsel__common_field().  Without this, NO_LIBTRACEEVENT=1 builds
>>     fail to link because evsel__resolve_probe_type() calls
>>     evsel__field() unconditionally (sashiko-bot).
>> 
>> Changes in v2 (patch 1 only):
>>   - Detect ftrace-based dynamic probes (kprobes/uprobes/fprobes created
>>     via tracefs) that report PMU "tracepoint" instead of "kprobe"/
>>     "uprobe".  Check for the __probe_ip field presence in the tracepoint
>>     format.  Without this, evsel__is_probe() returned false for ftrace
>>     probes, breaking __probe_ip suppression in perf trace for probes
>>     created via 'perf probe' (sashiko-bot).
>>   - Widen probe_type bitfield from 2 to 3 bits for PROBE__FTRACE.
>> 
>> Arnaldo Carvalho de Melo (3):
>>   perf evsel: Add lazy-initialized probe type detection helpers
>>   perf trace: Guard __probe_ip suppression with evsel__is_probe()
>>   perf evsel: Add no-libtraceevent stubs for evsel__field() and evsel__common_field()
>
>It seems you squashed patch 3 into 1. :)


I noticed it but accepted it at the end, I guess part of the process of using AI to speed up the process :-/

- Arnaldo


>
>Thanks,
>Namhyung
>
>> 
>>  tools/perf/builtin-trace.c |  2 +-
>>  tools/perf/util/evsel.c    | 53 ++++++++++++++++++++++++++++++++++++++++++++++
>>  tools/perf/util/evsel.h    | 19 +++++++++++++++++
>>  3 files changed, 73 insertions(+), 1 deletion(-)
>> 
>> Cc: Aaron Tomlin <atomlin@atomlin.com>
>> Cc: Ian Rogers <irogers@google.com>
>> Cc: Namhyung Kim <namhyung@kernel.org>
>> 
>> Developed with AI assistance (Claude/sashiko), tagged in commits.

- Arnaldo