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

Message ID 1527497103-3593-2-git-send-email-leo.yan@linaro.org
State New
Headers show
Series
  • Perf script: Add python script for CoreSight trace disassembler
Related show

Commit Message

Leo Yan May 28, 2018, 8:45 a.m.
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@arm.com>

Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces")
Signed-off-by: Leo Yan <leo.yan@linaro.org>

---
 tools/perf/util/cs-etm.c | 93 +++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 73 insertions(+), 20 deletions(-)

-- 
2.7.4

Comments

Mathieu Poirier May 28, 2018, 10:13 p.m. | #1
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@arm.com>

> Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces")

> Signed-off-by: Leo Yan <leo.yan@linaro.org>

> ---

>  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?


> +

> +	/*

> +	 * 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.

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.

> +

> +	/*

>  	 * 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.

> +

> +	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 ?

> +		    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.

> +

> +		/* 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.

>  		/*

>  		 * 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().

Thanks,
Mathieu



> +	return 0;

>  }

>  

>  static int cs_etm__run_decoder(struct cs_etm_queue *etmq)

> -- 

> 2.7.4

>
Leo Yan May 29, 2018, 12:25 a.m. | #2
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@arm.com>

> > Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces")

> > Signed-off-by: Leo Yan <leo.yan@linaro.org>

> > ---

> >  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?

--- 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.

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 ...

> > +

> > +	/*

> >  	 * 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].

[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'.

> >  		/*

> >  		 * 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

> >
Mathieu Poirier May 29, 2018, 4:04 p.m. | #3
On 28 May 2018 at 18:25, Leo Yan <leo.yan@linaro.org> 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@arm.com>

>> > Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces")

>> > Signed-off-by: Leo Yan <leo.yan@linaro.org>

>> > ---

>> >  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

>> >
Leo Yan May 30, 2018, 12:28 a.m. | #4
Hi Mathieu,

On Tue, May 29, 2018 at 10:04:49AM -0600, Mathieu Poirier wrote:

[...]

> > 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.


From the discussion context, I think here 'you may not need to' is
referring to my concern for applying patches on stable kernel, so I
should take this patch series as an enhancement and don't need to
consider much for stable kernel.

On the other hand, your suggestion is possible to mean 'not need
to' split into small patches (though I guess this is misunderstanding
for your meaning).

Could you clarify which is your meaning?

> >> > +

> >> > +   /*

> >> >      * 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


Agree, will look into for exception packet and try to add new packet
type for this.

[...]

Thanks,
Leo Yan
Robert Walker May 30, 2018, 9:45 a.m. | #5
On 28/05/18 23:13, 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@arm.com>

>> Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces")

>> Signed-off-by: Leo Yan <leo.yan@linaro.org>

>> ---

>>   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?

> 

> 

>> +

>> +	/*

>> +	 * 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.

> 

> 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.

> 

>> +

>> +	/*

>>   	 * 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.

> 

>> +

>> +	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 ?

> 

>> +		    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.

> 

>> +

>> +		/* 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.

> 

>>   		/*

>>   		 * 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().

>


This is code I inherited from the original versions of these patches, 
but it works because:
- etmq->packet and etmq->prev_packet are pointers to struct 
cs_etm_packet allocated by zalloc() in cs_etm__alloc_queue()
- cs_etm_decoder__get_packet() takes a pointer to struct cs_etm_packet 
and copies the contents of the first packet from the queue into the 
passed location with:
    *packet = decoder->packet_buffer[decoder->head]

So the swap code is only swapping the pointers over, not the contents of 
the packets.

Regards

Rob


> Thanks,

> Mathieu

>

> 

> 

>> +	return 0;

>>   }

>>   

>>   static int cs_etm__run_decoder(struct cs_etm_queue *etmq)

>> -- 

>> 2.7.4

>>
Mathieu Poirier May 30, 2018, 2:45 p.m. | #6
On 29 May 2018 at 18:28, Leo Yan <leo.yan@linaro.org> wrote:
> Hi Mathieu,

>

> On Tue, May 29, 2018 at 10:04:49AM -0600, Mathieu Poirier wrote:

>

> [...]

>

>> > 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.

>

> From the discussion context, I think here 'you may not need to' is

> referring to my concern for applying patches on stable kernel, so I

> should take this patch series as an enhancement and don't need to

> consider much for stable kernel.


Yes, that is what I meant.

>

> On the other hand, your suggestion is possible to mean 'not need

> to' split into small patches (though I guess this is misunderstanding

> for your meaning).

>

> Could you clarify which is your meaning?

>

>> >> > +

>> >> > +   /*

>> >> >      * 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

>

> Agree, will look into for exception packet and try to add new packet

> type for this.

>

> [...]

>

> Thanks,

> Leo Yan
Leo Yan May 30, 2018, 2:49 p.m. | #7
On Wed, May 30, 2018 at 08:45:46AM -0600, Mathieu Poirier wrote:
> On 29 May 2018 at 18:28, Leo Yan <leo.yan@linaro.org> wrote:

> > Hi Mathieu,

> >

> > On Tue, May 29, 2018 at 10:04:49AM -0600, Mathieu Poirier wrote:

> >

> > [...]

> >

> >> > 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.

> >

> > From the discussion context, I think here 'you may not need to' is

> > referring to my concern for applying patches on stable kernel, so I

> > should take this patch series as an enhancement and don't need to

> > consider much for stable kernel.

> 

> Yes, that is what I meant.


Thanks for confirmation, will send new patch series according to the
discussion.

[...]

Thanks,
Leo Yan
Mike Leach May 30, 2018, 3:04 p.m. | #8
Hi Leo,



On 30 May 2018 at 10:45, Robert Walker <robert.walker@arm.com> wrote:
>

>

> On 28/05/18 23:13, 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@arm.com>

>>> Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight

>>> traces")

>>> Signed-off-by: Leo Yan <leo.yan@linaro.org>

>>> ---

>>>   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?

>>

>>

>>> +

>>> +       /*

>>> +        * 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.

>>

>> 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.

>>

>>> +

>>> +       /*

>>>          * 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.

>>

>>> +

>>> +       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 ?

>>

>>> +                   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.

>>


clarification on the exception packets....

The Opencsd output exception packet gives you the exception number,
and optionally the preferred return address. If this address is
present does depend a lot on the underlying protocol - will normally
be there with ETMv4.
Exceptions are marked differently in the underlying protocol - the
OCSD packets abstract away these differences.

consider the code:

0x1000: <some instructions>
0x1100: BR 0x2000
....
0x2000: <some instructions>
0x2020  BZ r4

Without an exception this would result in the packets

OCSD_RANGE(0x1000,0x1104, Last instr type=Br, taken)   // recall that
range packets have start addr inclusive, end addr exclusive.
OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -
depends on condition>

Now consider an exception occurring before the BR 0x2000

this will result in:-
OCSD_RANGE(0x1000, 0x1100, Last instr type=Other)
OCSD_EXECEPTION(IRQ, ret-addr 0x1100)
OCSD_RANGE(IRQ_START, IRQ_END+4, Last instr type = BR, taken)    //
this is more likely to have multiple ranges / branches before any
return, but simplified here.
OCSD_EXCEPTION_RETURN()  // present if exception returns are
explicitly marked in underlying trace - may not always be depending on
circumstances.
OCSD_RANGE(0x1100,0x1104, Last=BR, taken) // continue on with short
range - just the branch
OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -
depends on condition>

Now consider the exception occurring after the BR, but before any
other instructions are executed.

OCSD_RANGE(0x1000,0x1104, Last instr type=Br, taken)   // recall that
range packets have start addr inclusive, end addr exclusive.
OCSD_EXECEPTION(IRQ, ret-addr 0x2000)     // here the preferred return
address is actually the target of the branch.
OCSD_RANGE(IRQ_START, IRQ_END+4, Last instr type = BR, taken)    //
this is more likely to have multiple ranges / branches before any
return, but simplified here.
OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -
depends on condition>

So in general it is possible to arrive in the IRQ_START range with the
previous packet having been either a taken branch, a not taken branch,
or not a branch.
Care must be taken - whether AutoFDO or normal trace disassembly not
to assume that having the last range packet as a taken branch means
that the next range packet is the target, if there is an intervening
exception packet.

Regards

Mike

>>> +

>>> +               /* 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.

>>

>>>                 /*

>>>                  * 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().

>>

>

> This is code I inherited from the original versions of these patches, but it

> works because:

> - etmq->packet and etmq->prev_packet are pointers to struct cs_etm_packet

> allocated by zalloc() in cs_etm__alloc_queue()

> - cs_etm_decoder__get_packet() takes a pointer to struct cs_etm_packet and

> copies the contents of the first packet from the queue into the passed

> location with:

>    *packet = decoder->packet_buffer[decoder->head]

>

> So the swap code is only swapping the pointers over, not the contents of the

> packets.

>

> Regards

>

> Rob

>

>

>

>> Thanks,

>> Mathieu

>>

>>

>>

>>> +       return 0;

>>>   }

>>>     static int cs_etm__run_decoder(struct cs_etm_queue *etmq)

>>> --

>>> 2.7.4

>>>

>




-- 
Mike Leach
Principal Engineer, ARM Ltd.
Manchester Design Centre. UK
Leo Yan May 30, 2018, 3:39 p.m. | #9
Hi Mike,

On Wed, May 30, 2018 at 04:04:34PM +0100, Mike Leach wrote:

[...]

> >>> +               /* 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.

> >>

> 

> clarification on the exception packets....

> 

> The Opencsd output exception packet gives you the exception number,

> and optionally the preferred return address. If this address is

> present does depend a lot on the underlying protocol - will normally

> be there with ETMv4.

> Exceptions are marked differently in the underlying protocol - the

> OCSD packets abstract away these differences.

> 

> consider the code:

> 

> 0x1000: <some instructions>

> 0x1100: BR 0x2000

> ....

> 0x2000: <some instructions>

> 0x2020  BZ r4

> 

> Without an exception this would result in the packets

> 

> OCSD_RANGE(0x1000,0x1104, Last instr type=Br, taken)   // recall that

> range packets have start addr inclusive, end addr exclusive.

> OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -

> depends on condition>

> 

> Now consider an exception occurring before the BR 0x2000

> 

> this will result in:-

> OCSD_RANGE(0x1000, 0x1100, Last instr type=Other)

> OCSD_EXECEPTION(IRQ, ret-addr 0x1100)

> OCSD_RANGE(IRQ_START, IRQ_END+4, Last instr type = BR, taken)    //

> this is more likely to have multiple ranges / branches before any

> return, but simplified here.

> OCSD_EXCEPTION_RETURN()  // present if exception returns are

> explicitly marked in underlying trace - may not always be depending on

> circumstances.

> OCSD_RANGE(0x1100,0x1104, Last=BR, taken) // continue on with short

> range - just the branch

> OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -

> depends on condition>

> 

> Now consider the exception occurring after the BR, but before any

> other instructions are executed.

> 

> OCSD_RANGE(0x1000,0x1104, Last instr type=Br, taken)   // recall that

> range packets have start addr inclusive, end addr exclusive.

> OCSD_EXECEPTION(IRQ, ret-addr 0x2000)     // here the preferred return

> address is actually the target of the branch.

> OCSD_RANGE(IRQ_START, IRQ_END+4, Last instr type = BR, taken)    //

> this is more likely to have multiple ranges / branches before any

> return, but simplified here.

> OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -

> depends on condition>

> 

> So in general it is possible to arrive in the IRQ_START range with the

> previous packet having been either a taken branch, a not taken branch,

> or not a branch.

> Care must be taken - whether AutoFDO or normal trace disassembly not

> to assume that having the last range packet as a taken branch means

> that the next range packet is the target, if there is an intervening

> exception packet.


Thanks a lot for detailed explaination.

IIUC, AutoFDO will not have such issue due every range packet will be
handled for it.  On the other hand, as you remind, the branch samples
(and its consumer trace disassembler) is very dependent on the flag
'last_instr_taken_branch'.

According to your explaination, I think we consider the branch is
taken for below situations:

- The new coming packet is exception packet (both for exception entry
  and exit packets);
- The previous packet is expcetion packet;
- The previous packet is normal range packet with
  'last_instr_taken_branch' = true;

So I'd like to use below function to demonstrate my understanding for
exception packets handling.  I also will send out one new patch for
support exception packet for reviewing.

If you have concern or I miss anything, please let me know.

static bool cs_etm__is_taken_branch(struct cs_etm_packet *prev_packet,
                                    struct cs_etm_packet *packet,)
{
        /* The branch is taken for normal range packet with taken branch flag */
        if (prev_packet->sample_type == CS_ETM_RANGE &&
            prev_packet->last_instr_taken_branch)
                return true;

        /* The branch is taken if previous packet is exception packet */
        if (prev_packet->sample_type == CS_ETM_EXCEPTION ||
            prev_packet->sample_type == CS_ETM_EXCEPTION_RET)
                return true;

        /* The branch is taken for an intervening exception packet */
        if (packet->sample_type == CS_ETM_EXCEPTION ||
            packet->sample_type == CS_ETM_EXCEPTION_RET)
                return true;

        return false;
}

[...]

Thanks,
Leo Yan
Leo Yan May 30, 2018, 3:49 p.m. | #10
On Wed, May 30, 2018 at 11:39:00PM +0800, Leo Yan wrote:
> Hi Mike,

> 

> On Wed, May 30, 2018 at 04:04:34PM +0100, Mike Leach wrote:

> 

> [...]

> 

> > >>> +               /* 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.

> > >>

> > 

> > clarification on the exception packets....

> > 

> > The Opencsd output exception packet gives you the exception number,

> > and optionally the preferred return address. If this address is

> > present does depend a lot on the underlying protocol - will normally

> > be there with ETMv4.

> > Exceptions are marked differently in the underlying protocol - the

> > OCSD packets abstract away these differences.

> > 

> > consider the code:

> > 

> > 0x1000: <some instructions>

> > 0x1100: BR 0x2000

> > ....

> > 0x2000: <some instructions>

> > 0x2020  BZ r4

> > 

> > Without an exception this would result in the packets

> > 

> > OCSD_RANGE(0x1000,0x1104, Last instr type=Br, taken)   // recall that

> > range packets have start addr inclusive, end addr exclusive.

> > OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -

> > depends on condition>

> > 

> > Now consider an exception occurring before the BR 0x2000

> > 

> > this will result in:-

> > OCSD_RANGE(0x1000, 0x1100, Last instr type=Other)

> > OCSD_EXECEPTION(IRQ, ret-addr 0x1100)

> > OCSD_RANGE(IRQ_START, IRQ_END+4, Last instr type = BR, taken)    //

> > this is more likely to have multiple ranges / branches before any

> > return, but simplified here.

> > OCSD_EXCEPTION_RETURN()  // present if exception returns are

> > explicitly marked in underlying trace - may not always be depending on

> > circumstances.

> > OCSD_RANGE(0x1100,0x1104, Last=BR, taken) // continue on with short

> > range - just the branch

> > OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -

> > depends on condition>

> > 

> > Now consider the exception occurring after the BR, but before any

> > other instructions are executed.

> > 

> > OCSD_RANGE(0x1000,0x1104, Last instr type=Br, taken)   // recall that

> > range packets have start addr inclusive, end addr exclusive.

> > OCSD_EXECEPTION(IRQ, ret-addr 0x2000)     // here the preferred return

> > address is actually the target of the branch.

> > OCSD_RANGE(IRQ_START, IRQ_END+4, Last instr type = BR, taken)    //

> > this is more likely to have multiple ranges / branches before any

> > return, but simplified here.

> > OCSD_RANGE(0x2000,0x2024, Last instr type=Br, <taken / not taken -

> > depends on condition>

> > 

> > So in general it is possible to arrive in the IRQ_START range with the

> > previous packet having been either a taken branch, a not taken branch,

> > or not a branch.

> > Care must be taken - whether AutoFDO or normal trace disassembly not

> > to assume that having the last range packet as a taken branch means

> > that the next range packet is the target, if there is an intervening

> > exception packet.

> 

> Thanks a lot for detailed explaination.

> 

> IIUC, AutoFDO will not have such issue due every range packet will be

> handled for it.  On the other hand, as you remind, the branch samples

> (and its consumer trace disassembler) is very dependent on the flag

> 'last_instr_taken_branch'.

> 

> According to your explaination, I think we consider the branch is

> taken for below situations:

> 

> - The new coming packet is exception packet (both for exception entry

>   and exit packets);

> - The previous packet is expcetion packet;

> - The previous packet is normal range packet with

>   'last_instr_taken_branch' = true;

> 

> So I'd like to use below function to demonstrate my understanding for

> exception packets handling.  I also will send out one new patch for

> support exception packet for reviewing.

> 

> If you have concern or I miss anything, please let me know.

> 

> static bool cs_etm__is_taken_branch(struct cs_etm_packet *prev_packet,

>                                     struct cs_etm_packet *packet,)

> {

>         /* The branch is taken for normal range packet with taken branch flag */

>         if (prev_packet->sample_type == CS_ETM_RANGE &&

>             prev_packet->last_instr_taken_branch)

>                 return true;

> 

>         /* The branch is taken if previous packet is exception packet */

>         if (prev_packet->sample_type == CS_ETM_EXCEPTION ||

>             prev_packet->sample_type == CS_ETM_EXCEPTION_RET)

>                 return true;

> 

>         /* The branch is taken for an intervening exception packet */

>         if (packet->sample_type == CS_ETM_EXCEPTION ||

>             packet->sample_type == CS_ETM_EXCEPTION_RET)

>                 return true;

> 

>         return false;

> }


Just clarify, I missed to mention I introduce two extra sample types:
CS_ETM_EXCEPTION and CS_ETM_EXCEPTION_RET, one is for exception
entry packet and another is for exception exit packet.   If this is
hard for understanding, you could hold on for reveiwing new patch.

Thanks,
Leo Yan

Patch

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;
+
+	/*
+	 * 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;
+
+	/*
 	 * 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;
+
+	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 ||
+		    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;
+
+		/* 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) {
 		/*
 		 * 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;
+	return 0;
 }
 
 static int cs_etm__run_decoder(struct cs_etm_queue *etmq)