On Sun, Nov 11, 2018 at 12:59:40PM +0800, Leo Yan wrote: > At the end of trace buffer handling, function cs_etm__flush() is invoked > to flush any remaining branch stack entries. As a side effect, it also > generates branch sample, because the 'etmq->packet' doesn't contains any > new coming packet but point to one stale packet after packets swapping, > so it wrongly makes synthesize branch samples with stale packet info. > > We could review below detailed flow which causes issue: > > Packet1: start_addr=0xffff000008b1fbf0 end_addr=0xffff000008b1fbfc > Packet2: start_addr=0xffff000008b1fb5c end_addr=0xffff000008b1fb6c > > step 1: cs_etm__sample(): > sample: ip=(0xffff000008b1fbfc-4) addr=0xffff000008b1fb5c > > step 2: flush packet in cs_etm__run_decoder(): > cs_etm__run_decoder() > `-> err = cs_etm__flush(etmq, false); > sample: ip=(0xffff000008b1fb6c-4) addr=0xffff000008b1fbf0 > > Packet1 and packet2 are two continuous packets, when packet2 is the new > coming packet, cs_etm__sample() generates branch sample for these two > packets and use [packet1::end_addr - 4 => packet2::start_addr] as branch > jump flow, thus we can see the first generated branch sample in step 1. > At the end of cs_etm__sample() it swaps packets so 'etm->prev_packet'= > packet2 and 'etm->packet'=packet1, so far it's okay for branch sample. > > If packet2 is the last one packet in trace buffer, even there have no > any new coming packet, cs_etm__run_decoder() invokes cs_etm__flush() to > flush branch stack entries as expected, but it also generates branch > samples by taking 'etm->packet' as a new coming packet, thus the branch > jump flow is as [packet2::end_addr - 4 => packet1::start_addr]; this > is the second sample which is generated in step 2. So actually the > second sample is a stale sample and we should not generate it. > > This patch is to add new argument 'new_packet' for cs_etm__flush(), we > can pass 'true' for this argument if there have a new packet, otherwise > it will pass 'false' for the purpose of only flushing branch stack > entries and avoid to generate sample for stale packet.
Very good explanation, thanks for taking the time to write this. > > Signed-off-by: Leo Yan <[email protected]> > --- > tools/perf/util/cs-etm.c | 20 +++++++++++++++++--- > 1 file changed, 17 insertions(+), 3 deletions(-) > > diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c > index fe18d7b..f4fa877 100644 > --- a/tools/perf/util/cs-etm.c > +++ b/tools/perf/util/cs-etm.c > @@ -955,7 +955,7 @@ static int cs_etm__sample(struct cs_etm_queue *etmq) > return 0; > } > > -static int cs_etm__flush(struct cs_etm_queue *etmq) > +static int cs_etm__flush(struct cs_etm_queue *etmq, bool new_packet) > { > int err = 0; > struct cs_etm_auxtrace *etm = etmq->etm; > @@ -989,6 +989,20 @@ static int cs_etm__flush(struct cs_etm_queue *etmq) > > } > > + /* > + * If 'new_packet' is false, this time call has no a new packet > + * coming and 'etmq->packet' contains the stale packet which is > + * set at the previous time with packets swapping. In this case > + * this function is invoked only for flushing branch stack at > + * the end of buffer handling. > + * > + * Simply to say, branch samples should be generated when every > + * time receive one new packet; otherwise, directly bail out to > + * avoid generate branch sample with stale packet. > + */ > + if (!new_packet) > + return 0; > + > if (etm->sample_branches && > etmq->prev_packet->sample_type == CS_ETM_RANGE) { > err = cs_etm__synth_branch_sample(etmq); > @@ -1075,7 +1089,7 @@ static int cs_etm__run_decoder(struct cs_etm_queue > *etmq) > * Discontinuity in trace, flush > * previous branch stack > */ > - cs_etm__flush(etmq); > + cs_etm__flush(etmq, true); > break; > case CS_ETM_EMPTY: > /* > @@ -1092,7 +1106,7 @@ static int cs_etm__run_decoder(struct cs_etm_queue > *etmq) > > if (err == 0) > /* Flush any remaining branch stack entries */ > - err = cs_etm__flush(etmq); > + err = cs_etm__flush(etmq, false); I understand what you're doing and it will yield the correct results. What I'm not sure about is if we wouldn't be better off splitting cs_etm__flush() in order to reduce the complexity of the main decoding loop. That is rename cs_etm__flush() to something like cs_etm__trace_on() and spin off a new cs_etm__end_block(). It does introduce a little bit of code duplication but I think we'd win in terms of readability and flexibility. Thanks, Mathieu > } > > return err; > -- > 2.7.4 >
