Re: [PATCH v2 4/4] perf arm-spe: Support hardware-based PID tracing

[Leo Yan]

On Thu, Nov 11, 2021 at 12:23:08PM +0000, German Gomez wrote:
> On 11/11/2021 08:30, Leo Yan wrote:
> > On Wed, Nov 10, 2021 at 11:59:05PM -0800, Namhyung Kim wrote:

[...]

> >>>>> +static int arm_spe_set_tid(struct arm_spe_queue *speq, pid_t tid)
> >>>>> +{
> >>>>> +       struct arm_spe *spe = speq->spe;
> >>>>> +       int err = machine__set_current_tid(spe->machine, speq->cpu, tid, tid);
> >>>>
> >>>> I think we should pass -1 as pid as we don't know the real pid.
> >>>>
> >>> AFAICT, I observe one case for machine__set_current_tid() returning error
> >>> is 'speq->cpu' is -1 (this is the case for per-thread tracing).  In
> >>> this case, if pass '-1' for pid/tid, it still will return failure.
> >>>
> >>> So here should return the error as it is.  Am I missing anything?
> >>
> >> I'm not saying about the error.  It's about thread status.
> >> In the machine__set_current_tid(), it calls
> >> machine__findnew_thread() with given pid and tid.
> >>
> >> I suspect it can set pid to a wrong value if the thread has
> >> no pid value at the moment.
> >
> > Here we should avoid to write pid '-1' with
> > machine__set_current_tid().
> 
> If the kernel is writing the tids to the contextidr, isn't it wrong to
> assume tid == pid when decoding the context packets here? I haven't
> observed any impact in the built-in commands though, so there must be
> something I'm not seeing.

Okay, let me correct myself :)

I checked Intel-pt's implementation, I understand now that we need to 
distinguish the cases for pid/tid from context switch event and only tid
from SPE context packet.

Since the context switch event contains the correct pid and tid
values, we should set both of them, see Intel-PT's implementation [1].

As Namhyung pointed, we need to set pid as '-1' when we only know the
tid value from SPE context packet; see [2].

So we should use the same with Intel-pt.

Sorry for I didn't really understand well Namhyung's suggestion and
thanks you both pointed out the issue.

Leo

P.s. an offline topic is we should send a patch to fix cs-etm issue
as well [3].

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/util/intel-pt.c#n2920
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/util/intel-pt.c#n2215
[3] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/util/cs-etm.c#n1121