Re: [RFT v3 1/4] perf cs-etm: Generate branch sample for missed packets

From: Mathieu Poirier
Date: Tue May 29 2018 - 12:05:01 EST

On 28 May 2018 at 18:25, Leo Yan <leo.yan@xxxxxxxxxx> wrote:
> Hi Mathieu,
>
> On Mon, May 28, 2018 at 04:13:47PM -0600, Mathieu Poirier wrote:
>> Leo and/or Robert,
>>
>> On Mon, May 28, 2018 at 04:45:00PM +0800, Leo Yan wrote:
>> > Commit e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight
>> > traces") reworks the samples generation flow from CoreSight trace to
>> > match the correct format so Perf report tool can display the samples
>> > properly.
>> >
>> > But the change has side effect for branch packet handling, it only
>> > generate branch samples by checking previous packet flag
>> > 'last_instr_taken_branch' is true, this results in below three kinds
>> > packets are missed to generate branch samples:
>> >
>> > - The start tracing packet at the beginning of tracing data;
>> > - The exception handling packet;
>> > - If one CS_ETM_TRACE_ON packet is inserted, we also miss to handle it
>> > for branch samples. CS_ETM_TRACE_ON packet itself can give the info
>> > that there have a discontinuity in the trace, on the other hand we
>> > also miss to generate proper branch sample for packets before and
>> > after CS_ETM_TRACE_ON packet.
>> >
>> > This patch is to add branch sample handling for up three kinds packets:
>> >
>> > - In function cs_etm__sample(), check if 'prev_packet->sample_type' is
>> > zero and in this case it generates branch sample for the start tracing
>> > packet; furthermore, we also need to handle the condition for
>> > prev_packet::end_addr is zero in the cs_etm__last_executed_instr();
>> >
>> > - In function cs_etm__sample(), check if 'prev_packet->exc' is true and
>> > generate branch sample for exception handling packet;
>> >
>> > - If there has one CS_ETM_TRACE_ON packet is coming, we firstly generate
>> > branch sample in the function cs_etm__flush(), this can save complete
>> > info for the previous CS_ETM_RANGE packet just before CS_ETM_TRACE_ON
>> > packet. We also generate branch sample for the new CS_ETM_RANGE
>> > packet after CS_ETM_TRACE_ON packet, this have two purposes, the
>> > first one purpose is to save the info for the new CS_ETM_RANGE packet,
>> > the second purpose is to save CS_ETM_TRACE_ON packet info so we can
>> > have hint for a discontinuity in the trace.
>> >
>> > For CS_ETM_TRACE_ON packet, its fields 'packet->start_addr' and
>> > 'packet->end_addr' equal to 0xdeadbeefdeadbeefUL which are emitted in
>> > the decoder layer as dummy value. This patch is to convert these
>> > values to zeros for more readable; this is accomplished by functions
>> > cs_etm__last_executed_instr() and cs_etm__first_executed_instr(). The
>> > later one is a new function introduced by this patch.
>> >
>> > Reviewed-by: Robert Walker <robert.walker@xxxxxxx>
>> > Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces")
>> > Signed-off-by: Leo Yan <leo.yan@xxxxxxxxxx>
>> > ---
>> > tools/perf/util/cs-etm.c | 93 +++++++++++++++++++++++++++++++++++++-----------
>> > 1 file changed, 73 insertions(+), 20 deletions(-)
>> >
>> > diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c
>> > index 822ba91..8418173 100644
>> > --- a/tools/perf/util/cs-etm.c
>> > +++ b/tools/perf/util/cs-etm.c
>> > @@ -495,6 +495,20 @@ static inline void cs_etm__reset_last_branch_rb(struct cs_etm_queue *etmq)
>> > static inline u64 cs_etm__last_executed_instr(struct cs_etm_packet *packet)
>> > {
>> > /*
>> > + * The packet is the start tracing packet if the end_addr is zero,
>> > + * returns 0 for this case.
>> > + */
>> > + if (!packet->end_addr)
>> > + return 0;
>>
>> What is considered to be the "start tracing packet"? Right now the only two
>> kind of packets inserted in the decoder packet buffer queue are INST_RANGE and
>> TRACE_ON. How can we hit a condition where packet->end-addr == 0?
>
> When the first CS_ETM_RANGE packet is coming, etmq->prev_packet is
> initialized by the function cs_etm__alloc_queue(), so
> etmq->prev_packet->end_addr is zero:
>
> etmq->prev_packet = zalloc(szp);
>
> As you mentioned, we should only have two kind of packets for
> CS_ETM_RANGE and CS_ETM_TRACE_ON. Currently we skip to handle the
> first CS_ETM_TRACE_ON packet in function cs_etm__flush(), we also can
> refine the function cs_etm__flush() to handle the first coming
> CS_ETM_TRACE_ON packet, after that all packets will be CS_ETM_RANGE
> and CS_ETM_TRACE_ON and have no chance to hit 'packet->end_addr = 0'.
>
> Does this make sense for you?

That is the right way to handle this condition and it gives us a
reliable state machine.

>
> --- Packet dumping when first packet coming ---
> cs_etm__flush: prev_packet: sample_type=0 exc=0 exc_ret=0 cpu=0 start_addr=0x0 end_addr=0x0 last_instr_taken_branch=0
> cs_etm__flush: packet: sample_type=2 exc=0 exc_ret=0 cpu=1 start_addr=0xdeadbeefdeadbeef end_addr=0xdeadbeefdeadbeef last_instr_taken_branch=0
>
>> > +
>> > + /*
>> > + * The packet is the CS_ETM_TRACE_ON packet if the end_addr is
>> > + * magic number 0xdeadbeefdeadbeefUL, returns 0 for this case.
>> > + */
>> > + if (packet->end_addr == 0xdeadbeefdeadbeefUL)
>> > + return 0;
>>
>> As it is with the above, I find triggering on addresses to be brittle and hard
>> to maintain on the long run. Packets all have a sample_type field that should
>> be used in cases like this one. That way we know exactly the condition that is
>> targeted.
>
> Will do this.
>
>> While working on this set, please spin-off another patch that defines
>> CS_ETM_INVAL_ADDR 0xdeadbeefdeadbeefUL and replace all the cases where the
>> numeral is used. That way we stop using the hard coded value.
>
> Will do this.

Much appreciated.

>
> As now this patch is big with more complex logic, so I consider to
> split it into small patches:
>
> - Define CS_ETM_INVAL_ADDR;
> - Fix for CS_ETM_TRACE_ON packet;
> - Fix for exception packet;
>
> Does this make sense for you? I have concern that this patch is a
> fixing patch, so not sure after spliting patches will introduce
> trouble for applying them for other stable kernels ...

Reverse the order:

- Fix for CS_ETM_TRACE_ON packet;
- Fix for exception packet;
- Define CS_ETM_INVAL_ADDR;

But you may not need to - see next comment.

>
>> > +
>> > + /*
>> > * The packet records the execution range with an exclusive end address
>> > *
>> > * A64 instructions are constant size, so the last executed
>> > @@ -505,6 +519,18 @@ static inline u64 cs_etm__last_executed_instr(struct cs_etm_packet *packet)
>> > return packet->end_addr - A64_INSTR_SIZE;
>> > }
>> >
>> > +static inline u64 cs_etm__first_executed_instr(struct cs_etm_packet *packet)
>> > +{
>> > + /*
>> > + * The packet is the CS_ETM_TRACE_ON packet if the start_addr is
>> > + * magic number 0xdeadbeefdeadbeefUL, returns 0 for this case.
>> > + */
>> > + if (packet->start_addr == 0xdeadbeefdeadbeefUL)
>> > + return 0;
>>
>> Same comment as above.
>
> Will do this.
>
>> > +
>> > + return packet->start_addr;
>> > +}
>> > +
>> > static inline u64 cs_etm__instr_count(const struct cs_etm_packet *packet)
>> > {
>> > /*
>> > @@ -546,7 +572,7 @@ static void cs_etm__update_last_branch_rb(struct cs_etm_queue *etmq)
>> >
>> > be = &bs->entries[etmq->last_branch_pos];
>> > be->from = cs_etm__last_executed_instr(etmq->prev_packet);
>> > - be->to = etmq->packet->start_addr;
>> > + be->to = cs_etm__first_executed_instr(etmq->packet);
>> > /* No support for mispredict */
>> > be->flags.mispred = 0;
>> > be->flags.predicted = 1;
>> > @@ -701,7 +727,7 @@ static int cs_etm__synth_branch_sample(struct cs_etm_queue *etmq)
>> > sample.ip = cs_etm__last_executed_instr(etmq->prev_packet);
>> > sample.pid = etmq->pid;
>> > sample.tid = etmq->tid;
>> > - sample.addr = etmq->packet->start_addr;
>> > + sample.addr = cs_etm__first_executed_instr(etmq->packet);
>> > sample.id = etmq->etm->branches_id;
>> > sample.stream_id = etmq->etm->branches_id;
>> > sample.period = 1;
>> > @@ -897,13 +923,28 @@ static int cs_etm__sample(struct cs_etm_queue *etmq)
>> > etmq->period_instructions = instrs_over;
>> > }
>> >
>> > - if (etm->sample_branches &&
>> > - etmq->prev_packet &&
>> > - etmq->prev_packet->sample_type == CS_ETM_RANGE &&
>> > - etmq->prev_packet->last_instr_taken_branch) {
>> > - ret = cs_etm__synth_branch_sample(etmq);
>> > - if (ret)
>> > - return ret;
>> > + if (etm->sample_branches && etmq->prev_packet) {
>> > + bool generate_sample = false;
>> > +
>> > + /* Generate sample for start tracing packet */
>> > + if (etmq->prev_packet->sample_type == 0 ||
>>
>> What kind of packet is sample_type == 0 ?
>
> Just as explained above, sample_type == 0 is the packet which
> initialized in the function cs_etm__alloc_queue().
>
>> > + etmq->prev_packet->sample_type == CS_ETM_TRACE_ON)
>> > + generate_sample = true;
>> > +
>> > + /* Generate sample for exception packet */
>> > + if (etmq->prev_packet->exc == true)
>> > + generate_sample = true;
>>
>> Please don't do that. Exception packets have a type of their own and can be
>> added to the decoder packet queue the same way INST_RANGE and TRACE_ON packets
>> are. Moreover exception packet containt an address that, if I'm reading the
>> documenation properly, can be used to keep track of instructions that were
>> executed between the last address of the previous range packet and the address
>> executed just before the exception occurred. Mike and Rob will have to confirm
>> this as the decoder may be doing all that hard work for us.
>
> Sure, will wait for Rob and Mike to confirm for this.
>
> At my side, I dump the packet, the exception packet isn't passed to
> cs-etm.c layer, the decoder layer only sets the flag
> 'packet->exc = true' when exception packet is coming [1].

That's because we didn't need the information. Now that we do a
function that will insert a packet in the decoder packet queue and
deal with the new packet type in the main decoder loop [2]. At that
point your work may not be eligible for stable anymore and I think it
is fine. Robert's work was an enhancement over mine and yours is an
enhancement over his.

[2]. https://elixir.bootlin.com/linux/v4.17-rc7/source/tools/perf/util/cs-etm.c#L999

>
> [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c#n364
>
>> > +
>> > + /* Generate sample for normal branch packet */
>> > + if (etmq->prev_packet->sample_type == CS_ETM_RANGE &&
>> > + etmq->prev_packet->last_instr_taken_branch)
>> > + generate_sample = true;
>> > +
>> > + if (generate_sample) {
>> > + ret = cs_etm__synth_branch_sample(etmq);
>> > + if (ret)
>> > + return ret;
>> > + }
>> > }
>> >
>> > if (etm->sample_branches || etm->synth_opts.last_branch) {
>> > @@ -922,11 +963,16 @@ static int cs_etm__sample(struct cs_etm_queue *etmq)
>> > static int cs_etm__flush(struct cs_etm_queue *etmq)
>> > {
>> > int err = 0;
>> > + struct cs_etm_auxtrace *etm = etmq->etm;
>> > struct cs_etm_packet *tmp;
>> >
>> > - if (etmq->etm->synth_opts.last_branch &&
>> > - etmq->prev_packet &&
>> > - etmq->prev_packet->sample_type == CS_ETM_RANGE) {
>> > + if (!etmq->prev_packet)
>> > + return 0;
>> > +
>> > + if (etmq->prev_packet->sample_type != CS_ETM_RANGE)
>> > + return 0;
>> > +
>> > + if (etmq->etm->synth_opts.last_branch) {
>>
>> If you add:
>>
>> if (!etmq->etm->synth_opts.last_branch)
>> return 0;
>>
>> You can avoid indenting the whole block.
>
> No, here we cannot do like this. Except we need to handle the
> condition for 'etmq->etm->synth_opts.last_branch', we also need to
> handle 'etm->sample_branches'. These two conditions are saperate and
> decide by different command parameters from 'perf script'.

Pardon me - I didn't see the addition of the new '}' just below.

>
>> > /*
>> > * Generate a last branch event for the branches left in the
>> > * circular buffer at the end of the trace.
>> > @@ -939,18 +985,25 @@ static int cs_etm__flush(struct cs_etm_queue *etmq)
>> > err = cs_etm__synth_instruction_sample(
>> > etmq, addr,
>> > etmq->period_instructions);
>> > + if (err)
>> > + return err;
>> > etmq->period_instructions = 0;
>> > + }
>> >
>> > - /*
>> > - * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for
>> > - * the next incoming packet.
>> > - */
>> > - tmp = etmq->packet;
>> > - etmq->packet = etmq->prev_packet;
>> > - etmq->prev_packet = tmp;
>> > + if (etm->sample_branches) {
>> > + err = cs_etm__synth_branch_sample(etmq);
>> > + if (err)
>> > + return err;
>> > }
>> >
>> > - return err;
>> > + /*
>> > + * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for
>> > + * the next incoming packet.
>> > + */
>> > + tmp = etmq->packet;
>> > + etmq->packet = etmq->prev_packet;
>> > + etmq->prev_packet = tmp;
>>
>> Robert, I remember noticing that when you first submitted the code but forgot to
>> go back to it. What is the point of swapping the packets? I understand
>>
>> etmq->prev_packet = etmq->packet;
>>
>> But not
>>
>> etmq->packet = tmp;
>>
>> After all etmq->packet will be clobbered as soon as cs_etm_decoder__get_packet()
>> is called, which is alwasy right after either cs_etm__sample() or
>> cs_etm__flush().
>
> Yeah, I have the same question for this :)
>
> Thanks for suggestions and reviewing.
>
>> Thanks,
>> Mathieu
>>
>>
>>
>> > + return 0;
>> > }
>> >
>> > static int cs_etm__run_decoder(struct cs_etm_queue *etmq)
>> > --
>> > 2.7.4
>> >