Re: [PATCH v2 2/2] perf cs-etm: Add support sample flags

[leo . yan]

On Tue, Nov 20, 2018 at 09:53:41AM -0700, Mathieu Poirier wrote:

[...]

> > > +static void cs_etm__fixup_flags(struct cs_etm_queue *etmq)
> > > +{
> > > +     /*
> > > +      * Decoding stream might insert one TRACE_OFF packet in the
> > > +      * middle of instruction packets, this means it doesn't
> > > +      * contain the pair packets with TRACE_OFF and TRACE_ON.
> > > +      * For this case, the instruction packet follows with
> > > +      * TRACE_OFF packet so we need to fixup prev_packet with flag
> > > +      * PERF_IP_FLAG_TRACE_BEGIN, this flag finally is used by the
> > > +      * instruction packet to generate samples.
> > > +      */
> > > +     if (etmq->prev_packet->sample_type == CS_ETM_TRACE_OFF &&
> > > +         etmq->packet->sample_type == CS_ETM_RANGE)
> > > +             etmq->prev_packet->flags = PERF_IP_FLAG_BRANCH |
> > > +                                        PERF_IP_FLAG_TRACE_BEGIN;
> > > +
> > > +     if (etmq->prev_packet->sample_type == CS_ETM_RANGE) {
> > > +             /*
> > > +              * When the exception packet is inserted, update flags
> > > +              * so tell perf it is exception related branches.
> > > +              */
> > > +             if (etmq->packet->sample_type == CS_ETM_EXCEPTION ||
> > > +                 etmq->packet->sample_type == CS_ETM_EXCEPTION_RET)
> > > +                     etmq->prev_packet->flags = etmq->packet->flags;
> > > +
> > > +             /*
> > > +              * The trace is discontinuous, weather this is caused by
> > > +              * TRACE_ON packet or TRACE_OFF packet is coming, if the
> > > +              * previous packet is instruction packet, simply set flag
> > > +              * PERF_IP_FLAG_TRACE_END for previous packet.
> > > +              */
> > > +             if (etmq->packet->sample_type == CS_ETM_TRACE_ON ||
> > > +                 etmq->packet->sample_type == CS_ETM_TRACE_OFF)
> > > +                     etmq->prev_packet->flags |= PERF_IP_FLAG_TRACE_END;
> > > +     }
> > > +}
> > > +
> >
> > I think it would be better to keep all the flag related processing in
> > cs-etm-decoder.c so that things in cs-etm.c are only concered with dealing with
> > perf.
> >
> > Look at function cs_etm__alloc_queue(), there you'll find "d_params.data = etmq".
> >
> > In function cs_etm_decoder__new(), decoder->data = d_params->data;
> >
> > This means that anywhere you have a decoder, decoder->data is an etmq.  I've
> > used this profusely in my work on CPU-wide trace scenarios.  Because you're
> > getting there ahead of me you'll need to fix the declaration of struct
> > cs_etm_queue but that's easy.
> 
> I've been thinking further about this and manipulating the etmq packet
> and prev_packet from the cs-etm-decoder.c won't work because all we
> have at that time is the decoder's packet queue.  My goal is to
> manipulate the flags in only one place - either in cs-etm.c or
> cs-etm-decoder.c but not in both.  It might be worth trying to do the
> implementation in cs-etm.c since there is already a lot of packet flow
> intelligence happening there.

Agree.  cs-etm.c has more context info than cs-etm-decoder.c, will
try to refactor in single place in cs-etm.c.

[...]

Thanks,
Leo Yan