On Thu, Nov 18, 2021 at 10:14:12AM -0700, Mathieu Poirier wrote:
Good morning Leo,
On Sat, Nov 13, 2021 at 10:35:40PM +0800, Leo Yan wrote:
Currently, cs-etm passes the tid value for both tid and pid parameters when calling machine__set_current_tid(), this can lead to confusion for thread handling. E.g. we arbitrarily pass the same value for pid and tid, perf tool will be misled to consider it is a main thread (see thread__main_thread()).
On the other hand, Perf tool only can retrieve tid from Arm CoreSight context packet, and we have no chance to know pid (it maps to kernel's task_struct::tgid) from hardware tracing data. For this reason, this patch passes -1 as pid for function machine__set_current_tid().
Signed-off-by: Leo Yan leo.yan@linaro.org
tools/perf/util/cs-etm.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index f323adb1af85..eed1a5930072 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -1118,7 +1118,7 @@ int cs_etm__etmq_set_tid(struct cs_etm_queue *etmq, if (cs_etm__get_cpu(trace_chan_id, &cpu) < 0) return err;
- err = machine__set_current_tid(etm->machine, cpu, tid, tid);
- err = machine__set_current_tid(etm->machine, cpu, -1, tid);
I remember wondering about what to do with the pid parameter when I wrote this patch...
Some updates after I digged into the pid parameter for machine__set_current_tid().
During the recording phase, the perf tool will capture events PERF_RECORD_COMM and PERF_RECORD_FORK; these events contain pid/tid for profiled program. Below is an example for RECORD_FORK/RECORD_COMM events in perf data file:
0x89f0 [0x40]: event: 7 . . ... raw event: size 64 bytes . 0000: 07 00 00 00 00 20 40 00 59 6d 00 00 59 6d 00 00 ..... @.Ym..Ym.. . 0010: 5a 6d 00 00 59 6d 00 00 00 00 00 00 00 00 00 00 Zm..Ym.......... . 0020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ . 0030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
0 0 0x89f0 [0x40]: PERF_RECORD_FORK(27993:27994):(27993:27993)
0x8a30 [0x38]: event: 3 . . ... raw event: size 56 bytes . 0000: 03 00 00 00 00 00 38 00 59 6d 00 00 5a 6d 00 00 ......8.Ym..Zm.. . 0010: 6d 61 69 6e 00 00 00 00 00 00 00 00 00 00 00 00 main............ . 0020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ . 0030: 00 00 00 00 00 00 00 00 ........
0 0 0x8a30 [0x38]: PERF_RECORD_COMM: main:27993/27994
In the reporting phase, perf tool will setup threads structure based on the RECORD_FORK and RECORD_COMM events. This means perf tool will set the pid/tid for every thread, e.g. in up case, it allocates thread context for 'main' program, and its one child thread is setup to thread->pid_ as '27993' and thread->tid as '27994'.
Afterwards, when perf tool decodes CoreSight trace data and handles context packet, at the end, machine__update_thread_pid() is invoked for updating thread's pid:
machine__update_thread_pid(struct machine *machine, struct thread *th, pid_t pid) { if (pid == th->pid_ || pid == -1 || th->pid_ != -1) return;
... }
Whatever we pass the pid parameter as tid or '-1' from the caller function machine__set_current_tid(), it doesn't change anything for the thread context. Since th->pid_ has been initialized and its value is not '-1', no matter what's the pid value is passed via argument, machine__update_thread_pid() will directly bail out. This is why before we pass 'tid' value rather than '-1' for pid, it doesn't cause any error.
For this reason, this patch doesn't improve anything. After discussed with Mathieu offline, I decided to drop this change. So update the info in case someone is interested in the relevant info.
Thanks, Leo