| Message ID | 1526892748-326-2-git-send-email-leo.yan@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | Perf script: Add python script for CoreSight trace disassembler | expand |
Hi Leo, On 21/05/18 09:52, 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 packet handling, it only > generate samples when 'prev_packet->last_instr_taken_branch' is true, > this results in the start tracing packet and exception packets are > dropped. > > This patch checks extra two conditions for complete samples: > > - If 'prev_packet->sample_type' is zero we can use this condition to > get to know this is the start tracing packet; for this case, the start > packet's end_addr is zero as well so we need to handle it in the > function cs_etm__last_executed_instr(); > I think you also need to add something in to handle discontinuities in trace - for example it is possible to configure the ETM to only trace execution in specific code regions or to trace a few cycles every so often. In these cases, prev_packet->sample_type will not be zero, but whatever the previous packet was. You will get a CS_ETM_TRACE_ON packet in such cases, generated by an I_TRACE_ON element in the trace stream. You also get this on exception return. However, you should also keep the test for prev_packet->sample_type == 0 as you may not see a CS_ETM_TRACE_ON when decoding a buffer that has wrapped. Regards Rob > - If 'prev_packet->exc' is true, we can know the previous packet is > exception handling packet so need to generate sample for exception > flow. > > Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces") > Cc: Mike Leach <mike.leach@arm.com> > Cc: Robert Walker <robert.walker@arm.com> > Cc: Mathieu Poirier <mathieu.poirier@linaro.org> > Signed-off-by: Leo Yan <leo.yan@linaro.org> > --- > tools/perf/util/cs-etm.c | 35 ++++++++++++++++++++++++++++------- > 1 file changed, 28 insertions(+), 7 deletions(-) > > diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c > index 822ba91..378953b 100644 > --- a/tools/perf/util/cs-etm.c > +++ b/tools/perf/util/cs-etm.c > @@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address > * > * A64 instructions are constant size, so the last executed > @@ -897,13 +904,27 @@ 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) > + 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) { >
Hi Rob, On Mon, May 21, 2018 at 12:27:42PM +0100, Robert Walker wrote: > Hi Leo, > > On 21/05/18 09:52, 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 packet handling, it only > >generate samples when 'prev_packet->last_instr_taken_branch' is true, > >this results in the start tracing packet and exception packets are > >dropped. > > > >This patch checks extra two conditions for complete samples: > > > >- If 'prev_packet->sample_type' is zero we can use this condition to > > get to know this is the start tracing packet; for this case, the start > > packet's end_addr is zero as well so we need to handle it in the > > function cs_etm__last_executed_instr(); > > > > I think you also need to add something in to handle discontinuities in > trace - for example it is possible to configure the ETM to only trace > execution in specific code regions or to trace a few cycles every so > often. In these cases, prev_packet->sample_type will not be zero, but > whatever the previous packet was. You will get a CS_ETM_TRACE_ON packet in > such cases, generated by an I_TRACE_ON element in the trace stream. > You also get this on exception return. > > However, you should also keep the test for prev_packet->sample_type == 0 > as you may not see a CS_ETM_TRACE_ON when decoding a buffer that has > wrapped. Thanks for reviewing. Let's dig more detailed into this issue, especially for handling packet CS_ETM_TRACE_ON, I'd like divide into two sub cases. - The first case is for using python script: I use python script to analyze packets with below command: ./perf script --itrace=ril128 -s arm-cs-trace-disasm.py -F cpu,event,ip,addr,sym -- -v -d objdump -k ./vmlinux What I observe is after we pass python script with parameter '-s arm-cs-trace-disasm.py', then instruction tracing options '--itrace=ril128' isn't really used; the perf tool creates another new process for launch python script and re-enter cmd_script() function, but at the second time when invoke cmd_script() for python script execution the option '--itrace=ril128' is dropped and all parameters are only valid defined by the python script. As result, I can the variable 'etmq->etm->synth_opts.last_branch' is always FALSE for running python script. So all CS_ETM_TRACE_ON packets will be ignored in the function cs_etm__flush(). Even the CS_ETM_TRACE_ON packets are missed to handle, the program flow still can work well. The reason is without the interference by CS_ETM_TRACE_ON, the CS_ETM_RANGE packets can smoothly create instruction range by ignore the middle CS_ETM_TRACE_ON packet. Please see below example, in this example there have 3 packets, the first one packet is CS_ETM_RANGE packet which is labelled with 'PACKET_1', the first one packet can properly generate branch sample data with previous packet as expected; the second packet is PACKET_2 which is CS_ETM_TRACE_ON, but 'etmq->etm->synth_opts.last_branch' is false so function cs_etm__flush() doesn't handle it and skip the swap operation "etmq->prev_packet = tmp"; the third packet is PACKET_3, which is CS_ETM_RANGE packet and we can see it's smoontly to create continous instruction range between PACKET_1 and PACKET_3. cs_etm__sample: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f79c end_addr=0xffff000008a5f7bc last_instr_taken_branch=1 PACKET_1: cs_etm__sample: packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 cs_etm__synth_branch_sample: ip=0xffff000008a5f7b8 addr=0xffff000008a5f858 pid=2290 tid=2290 id=1000000021 stream_id=1000000021 period=1 cpu=1 flags=0 cpumode=2 cs_etm__flush: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 PACKET_2: cs_etm__flush: packet: sample_type=2 exc=0 exc_ret=0 cpu=2 start_addr=0xdeadbeefdeadbeef end_addr=0xdeadbeefdeadbeef last_instr_taken_branch=1 cs_etm__sample: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 PACKET_3: cs_etm__sample: packet: sample_type=1 exc=0 exc_ret=0 cpu=2 start_addr=0xffff000008be7528 end_addr=0xffff000008be7538 last_instr_taken_branch=1 cs_etm__synth_branch_sample: ip=0xffff000008a5f860 addr=0xffff000008be7528 pid=2290 tid=2290 id=1000000021 stream_id=1000000021 period=1 cpu=2 flags=0 cpumode=2 So seems to me, the CS_ETM_TRACE_ON packet doesn't introduce trouble for the program flow analysis if we can handle all CS_ETM_RANGE packets and without handling CS_ETM_TRACE_ON packet for branch samples. - The second case is for --itrace option without python script: ./perf script --itrace=ril -F cpu,event,ip,addr,sym -k ./vmlinux In this case, the flag 'etmq->etm->synth_opts.last_branch' is true so CS_ETM_TRACE_ON packet will be handled; but I can observe the CS_ETM_RANGE packet in etmq->prev_packet isn't handled in the function cs_etm__flush() for branch sample, so actually we miss some branch sample for this case. So I think we also need handle CS_ETM_RANGE packet in function cs_etm__flush() to generate branch samples. But this has side effect, we introduce the extra track for CS_ETM_TRACE_ON packet for branch samples, so we will see one branch range like: [ 0xdeadbeefdeadbeef .. 0xdeadbeefdeadbeef ]. Please reivew below change is okay for you? Thanks a lot for suggestions. diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index 822ba91..37d3722 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address * * A64 instructions are constant size, so the last executed @@ -897,13 +904,28 @@ 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) { @@ -921,12 +943,17 @@ static int cs_etm__sample(struct cs_etm_queue *etmq) static int cs_etm__flush(struct cs_etm_queue *etmq) { + struct cs_etm_auxtrace *etm = etmq->etm; int err = 0; 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 +966,25 @@ 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) -- 2.7.4
On Tue, May 22, 2018 at 04:39:20PM +0800, Leo Yan wrote:
[...]
Rather than the patch I posted in my previous email, I think below new
patch is more reasonable for me.
In the below change, 'etmq->prev_packet' is only used to store the
previous CS_ETM_RANGE packet, we don't need to save CS_ETM_TRACE_ON
packet into 'etmq->prev_packet'.
On the other hand, cs_etm__flush() can use 'etmq->period_instructions'
to indicate if need to generate instruction sample or not. If it's
non-zero, then generate instruction sample and
'etmq->period_instructions' will be cleared; so next time if there
have more tracing CS_ETM_TRACE_ON packet, it can skip to generate
instruction sample due 'etmq->period_instructions' is zero.
How about you think for this?
Thanks,
Leo Yan
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c
index 822ba91..dd354ad 100644
--- a/tools/perf/util/cs-etm.c
+++ b/tools/perf/util/cs-etm.c
@@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address
*
* A64 instructions are constant size, so the last executed
@@ -897,13 +904,27 @@ 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)
+ 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 +943,12 @@ 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_packet *tmp;
if (etmq->etm->synth_opts.last_branch &&
etmq->prev_packet &&
- etmq->prev_packet->sample_type == CS_ETM_RANGE) {
+ etmq->prev_packet->sample_type == CS_ETM_RANGE &&
+ etmq->period_instructions) {
+
/*
* Generate a last branch event for the branches left in the
* circular buffer at the end of the trace.
@@ -940,14 +962,6 @@ static int cs_etm__flush(struct cs_etm_queue *etmq)
etmq, addr,
etmq->period_instructions);
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;
}
return err;
--
2.7.4
Hi Leo, On 22/05/18 10:52, Leo Yan wrote: > On Tue, May 22, 2018 at 04:39:20PM +0800, Leo Yan wrote: > > [...] > > Rather than the patch I posted in my previous email, I think below new > patch is more reasonable for me. > > In the below change, 'etmq->prev_packet' is only used to store the > previous CS_ETM_RANGE packet, we don't need to save CS_ETM_TRACE_ON > packet into 'etmq->prev_packet'. > > On the other hand, cs_etm__flush() can use 'etmq->period_instructions' > to indicate if need to generate instruction sample or not. If it's > non-zero, then generate instruction sample and > 'etmq->period_instructions' will be cleared; so next time if there > have more tracing CS_ETM_TRACE_ON packet, it can skip to generate > instruction sample due 'etmq->period_instructions' is zero. > > How about you think for this? > > Thanks, > Leo Yan > I don't think this covers the cases where CS_ETM_TRACE_ON is used to indicate a discontinuity in the trace. For example, there is work in progress to configure the ETM so that it only traces a few thousand cycles with a gap of many thousands of cycles between each chunk of trace - this can be used to sample program execution in the form of instruction events with branch stacks for feedback directed optimization (AutoFDO). In this case, the raw trace is something like: ... I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0x0000007E7B886908; I_ATOM_F3 : Atom format 3.; EEN I_ATOM_F1 : Atom format 1.; E # Trace stops here # Some time passes, and then trace is turned on again I_TRACE_ON : Trace On. I_ADDR_CTXT_L_64IS0 : Address & Context, Long, 64 bit, IS0.; Addr=0x00000057224322F4; Ctxt: AArch64,EL0, NS; I_ATOM_F3 : Atom format 3.; ENN I_ATOM_F5 : Atom format 5.; ENENE ... This results in the following packets from the decoder: CS_ETM_RANGE: [0x7e7b886908-0x7e7b886930] br CS_ETM_RANGE: [0x7e7b88699c-0x7e7b8869a4] br CS_ETM_RANGE: [0x7e7b8869d8-0x7e7b8869f0] CS_ETM_RANGE: [0x7e7b8869f0-0x7e7b8869fc] br CS_ETM_TRACE_ON CS_ETM_RANGE: [0x57224322f4-0x5722432304] br CS_ETM_RANGE: [0x57224320e8-0x57224320ec] CS_ETM_RANGE: [0x57224320ec-0x57224320f8] CS_ETM_RANGE: [0x57224320f8-0x572243212c] br CS_ETM_RANGE: [0x5722439b80-0x5722439bec] CS_ETM_RANGE: [0x5722439bec-0x5722439c14] br CS_ETM_RANGE: [0x5722437c30-0x5722437c6c] CS_ETM_RANGE: [0x5722437c6c-0x5722437c7c] br Without handling the CS_ETM_TRACE_ON, this would be interpreted as a branch from 0x7e7b8869f8 to 0x57224322f4, when there is actually a gap of many thousand instructions between these. I think this patch will break the branch stacks - by removing the prev_packet swap from cs_etm__flush(), the next time a CS_ETM_RANGE packet is handled, cs_etm__sample() will see prev_packet contains the last CS_ETM_RANGE from the previous block of trace, causing an erroneous call to cs_etm__update_last_branch_rb(). In the example above, the branch stack will contain an erroneous branch from 0x7e7b8869f8 to 0x57224322f4. I think what you need to do is add a check for the previous packet being a CS_ETM_TRACE_ON when determining the generate_sample value. Regards Rob > > diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c > index 822ba91..dd354ad 100644 > --- a/tools/perf/util/cs-etm.c > +++ b/tools/perf/util/cs-etm.c > @@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address > * > * A64 instructions are constant size, so the last executed > @@ -897,13 +904,27 @@ 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) > + generate_sample = true; Also check for etmq->prev_packet->sample_type == CS_ETM_TRACE_ON here and set 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 +943,12 @@ 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_packet *tmp; > > if (etmq->etm->synth_opts.last_branch && > etmq->prev_packet && > - etmq->prev_packet->sample_type == CS_ETM_RANGE) { > + etmq->prev_packet->sample_type == CS_ETM_RANGE && > + etmq->period_instructions) { > + I don't think this is needed. > /* > * Generate a last branch event for the branches left in the > * circular buffer at the end of the trace. > @@ -940,14 +962,6 @@ static int cs_etm__flush(struct cs_etm_queue *etmq) > etmq, addr, > etmq->period_instructions); > 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; This should not be changed as discussed above. > } > > return err; >
Hi Rob, On Wed, May 23, 2018 at 12:21:18PM +0100, Robert Walker wrote: > Hi Leo, > > On 22/05/18 10:52, Leo Yan wrote: > >On Tue, May 22, 2018 at 04:39:20PM +0800, Leo Yan wrote: > > > >[...] > > > >Rather than the patch I posted in my previous email, I think below new > >patch is more reasonable for me. > > > >In the below change, 'etmq->prev_packet' is only used to store the > >previous CS_ETM_RANGE packet, we don't need to save CS_ETM_TRACE_ON > >packet into 'etmq->prev_packet'. > > > >On the other hand, cs_etm__flush() can use 'etmq->period_instructions' > >to indicate if need to generate instruction sample or not. If it's > >non-zero, then generate instruction sample and > >'etmq->period_instructions' will be cleared; so next time if there > >have more tracing CS_ETM_TRACE_ON packet, it can skip to generate > >instruction sample due 'etmq->period_instructions' is zero. > > > >How about you think for this? > > > >Thanks, > >Leo Yan > > > > I don't think this covers the cases where CS_ETM_TRACE_ON is used to > indicate a discontinuity in the trace. For example, there is work in > progress to configure the ETM so that it only traces a few thousand cycles > with a gap of many thousands of cycles between each chunk of trace - this > can be used to sample program execution in the form of instruction events > with branch stacks for feedback directed optimization (AutoFDO). > > In this case, the raw trace is something like: > > ... > I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0x0000007E7B886908; > I_ATOM_F3 : Atom format 3.; EEN > I_ATOM_F1 : Atom format 1.; E > # Trace stops here > > # Some time passes, and then trace is turned on again > I_TRACE_ON : Trace On. > I_ADDR_CTXT_L_64IS0 : Address & Context, Long, 64 bit, IS0.; > Addr=0x00000057224322F4; Ctxt: AArch64,EL0, NS; > I_ATOM_F3 : Atom format 3.; ENN > I_ATOM_F5 : Atom format 5.; ENENE > ... > > This results in the following packets from the decoder: > > CS_ETM_RANGE: [0x7e7b886908-0x7e7b886930] br > CS_ETM_RANGE: [0x7e7b88699c-0x7e7b8869a4] br > CS_ETM_RANGE: [0x7e7b8869d8-0x7e7b8869f0] > CS_ETM_RANGE: [0x7e7b8869f0-0x7e7b8869fc] br > CS_ETM_TRACE_ON > CS_ETM_RANGE: [0x57224322f4-0x5722432304] br > CS_ETM_RANGE: [0x57224320e8-0x57224320ec] > CS_ETM_RANGE: [0x57224320ec-0x57224320f8] > CS_ETM_RANGE: [0x57224320f8-0x572243212c] br > CS_ETM_RANGE: [0x5722439b80-0x5722439bec] > CS_ETM_RANGE: [0x5722439bec-0x5722439c14] br > CS_ETM_RANGE: [0x5722437c30-0x5722437c6c] > CS_ETM_RANGE: [0x5722437c6c-0x5722437c7c] br > > Without handling the CS_ETM_TRACE_ON, this would be interpreted as a branch > from 0x7e7b8869f8 to 0x57224322f4, when there is actually a gap of many > thousand instructions between these. > > I think this patch will break the branch stacks - by removing the > prev_packet swap from cs_etm__flush(), the next time a CS_ETM_RANGE packet > is handled, cs_etm__sample() will see prev_packet contains the last > CS_ETM_RANGE from the previous block of trace, causing an erroneous call to > cs_etm__update_last_branch_rb(). In the example above, the branch stack > will contain an erroneous branch from 0x7e7b8869f8 to 0x57224322f4. > > I think what you need to do is add a check for the previous packet being a > CS_ETM_TRACE_ON when determining the generate_sample value. I still can see there have hole for packets handling with your suggestion, let's focus on below three packets: CS_ETM_RANGE: [0x7e7b8869f0-0x7e7b8869fc] br CS_ETM_TRACE_ON: [0xdeadbeefdeadbeef-0xdeadbeefdeadbeef] CS_ETM_RANGE: [0x57224322f4-0x5722432304] br When the CS_ETM_TRACE_ON packet is coming, cs_etm__flush() doesn't handle for 'etmq->prev_packet' to generate branch sample, this results in we miss the info for 0x7e7b8869fc, and with packet swapping 'etmq->prev_packet' is assigned to CS_ETM_TRACE_ON packet. When the last CS_ETM_RANGE packet is coming, cs_etm__sample() will combine the values from CS_ETM_TRACE_ON packet and the last CS_ETM_RANGE packet to generate branch sample packet; at the end we get below sample packets: packet(n): sample::addr=0x7e7b8869f0 packet(n+1): sample::ip=0xdeadbeefdeadbeeb sample::addr=0x57224322f4 So I think we also need to generate branch sample, and we can get below results: packet(n): sample::addr=0x7e7b8869f0 packet(n+1): sample::ip=0x7e7b8869f8 sample::addr=0xdeadbeefdeadbeef packet(n+2): sample::ip=0xdeadbeefdeadbeeb sample::addr=0x57224322f4 So we also can rely on this to get to know there have one address range is [0xdeadbeefdeadbeef..0xdeadbeefdeadbeeb] to indicate there have a discontinuity in the trace. > Regards > > Rob > > > > >diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c > >index 822ba91..dd354ad 100644 > >--- a/tools/perf/util/cs-etm.c > >+++ b/tools/perf/util/cs-etm.c > >@@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address > > * > > * A64 instructions are constant size, so the last executed > >@@ -897,13 +904,27 @@ 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) > >+ generate_sample = true; > > Also check for etmq->prev_packet->sample_type == CS_ETM_TRACE_ON here and > set generate_sample = true. Agree, will add this. > >+ > >+ /* 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 +943,12 @@ 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_packet *tmp; > > if (etmq->etm->synth_opts.last_branch && > > etmq->prev_packet && > >- etmq->prev_packet->sample_type == CS_ETM_RANGE) { > >+ etmq->prev_packet->sample_type == CS_ETM_RANGE && > >+ etmq->period_instructions) { > >+ > > I don't think this is needed. Okay, I will keep this. > > /* > > * Generate a last branch event for the branches left in the > > * circular buffer at the end of the trace. > >@@ -940,14 +962,6 @@ static int cs_etm__flush(struct cs_etm_queue *etmq) > > etmq, addr, > > etmq->period_instructions); > > 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; > > This should not be changed as discussed above. Okay, will keep this. But I suggest we add some change like below: + if (etm->sample_branches) { + err = cs_etm__synth_branch_sample(etmq); + if (err) + return err; + } If so, could you review my posted another patch for this? http://archive.armlinux.org.uk/lurker/message/20180522.083920.184f1f78.en.html Thanks, Leo Yan > > } > > return err; > >
Hi Leo, On 23/05/18 14:22, Leo Yan wrote: > Hi Rob, > > On Wed, May 23, 2018 at 12:21:18PM +0100, Robert Walker wrote: >> Hi Leo, >> >> On 22/05/18 10:52, Leo Yan wrote: >>> On Tue, May 22, 2018 at 04:39:20PM +0800, Leo Yan wrote: >>> >>> [...] >>> >>> Rather than the patch I posted in my previous email, I think below new >>> patch is more reasonable for me. >>> >>> In the below change, 'etmq->prev_packet' is only used to store the >>> previous CS_ETM_RANGE packet, we don't need to save CS_ETM_TRACE_ON >>> packet into 'etmq->prev_packet'. >>> >>> On the other hand, cs_etm__flush() can use 'etmq->period_instructions' >>> to indicate if need to generate instruction sample or not. If it's >>> non-zero, then generate instruction sample and >>> 'etmq->period_instructions' will be cleared; so next time if there >>> have more tracing CS_ETM_TRACE_ON packet, it can skip to generate >>> instruction sample due 'etmq->period_instructions' is zero. >>> >>> How about you think for this? >>> >>> Thanks, >>> Leo Yan >>> >> I don't think this covers the cases where CS_ETM_TRACE_ON is used to >> indicate a discontinuity in the trace. For example, there is work in >> progress to configure the ETM so that it only traces a few thousand cycles >> with a gap of many thousands of cycles between each chunk of trace - this >> can be used to sample program execution in the form of instruction events >> with branch stacks for feedback directed optimization (AutoFDO). >> >> In this case, the raw trace is something like: >> >> ... >> I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0x0000007E7B886908; >> I_ATOM_F3 : Atom format 3.; EEN >> I_ATOM_F1 : Atom format 1.; E >> # Trace stops here >> >> # Some time passes, and then trace is turned on again >> I_TRACE_ON : Trace On. >> I_ADDR_CTXT_L_64IS0 : Address & Context, Long, 64 bit, IS0.; >> Addr=0x00000057224322F4; Ctxt: AArch64,EL0, NS; >> I_ATOM_F3 : Atom format 3.; ENN >> I_ATOM_F5 : Atom format 5.; ENENE >> ... >> >> This results in the following packets from the decoder: >> >> CS_ETM_RANGE: [0x7e7b886908-0x7e7b886930] br >> CS_ETM_RANGE: [0x7e7b88699c-0x7e7b8869a4] br >> CS_ETM_RANGE: [0x7e7b8869d8-0x7e7b8869f0] >> CS_ETM_RANGE: [0x7e7b8869f0-0x7e7b8869fc] br >> CS_ETM_TRACE_ON >> CS_ETM_RANGE: [0x57224322f4-0x5722432304] br >> CS_ETM_RANGE: [0x57224320e8-0x57224320ec] >> CS_ETM_RANGE: [0x57224320ec-0x57224320f8] >> CS_ETM_RANGE: [0x57224320f8-0x572243212c] br >> CS_ETM_RANGE: [0x5722439b80-0x5722439bec] >> CS_ETM_RANGE: [0x5722439bec-0x5722439c14] br >> CS_ETM_RANGE: [0x5722437c30-0x5722437c6c] >> CS_ETM_RANGE: [0x5722437c6c-0x5722437c7c] br >> >> Without handling the CS_ETM_TRACE_ON, this would be interpreted as a branch >> from 0x7e7b8869f8 to 0x57224322f4, when there is actually a gap of many >> thousand instructions between these. >> >> I think this patch will break the branch stacks - by removing the >> prev_packet swap from cs_etm__flush(), the next time a CS_ETM_RANGE packet >> is handled, cs_etm__sample() will see prev_packet contains the last >> CS_ETM_RANGE from the previous block of trace, causing an erroneous call to >> cs_etm__update_last_branch_rb(). In the example above, the branch stack >> will contain an erroneous branch from 0x7e7b8869f8 to 0x57224322f4. >> >> I think what you need to do is add a check for the previous packet being a >> CS_ETM_TRACE_ON when determining the generate_sample value. > I still can see there have hole for packets handling with your > suggestion, let's focus on below three packets: > > CS_ETM_RANGE: [0x7e7b8869f0-0x7e7b8869fc] br > CS_ETM_TRACE_ON: [0xdeadbeefdeadbeef-0xdeadbeefdeadbeef] > CS_ETM_RANGE: [0x57224322f4-0x5722432304] br > > When the CS_ETM_TRACE_ON packet is coming, cs_etm__flush() doesn't > handle for 'etmq->prev_packet' to generate branch sample, this results > in we miss the info for 0x7e7b8869fc, and with packet swapping > 'etmq->prev_packet' is assigned to CS_ETM_TRACE_ON packet. > > When the last CS_ETM_RANGE packet is coming, cs_etm__sample() will > combine the values from CS_ETM_TRACE_ON packet and the last > CS_ETM_RANGE packet to generate branch sample packet; at the end > we get below sample packets: > > packet(n): sample::addr=0x7e7b8869f0 > packet(n+1): sample::ip=0xdeadbeefdeadbeeb sample::addr=0x57224322f4 > > So I think we also need to generate branch sample, and we can get > below results: > > packet(n): sample::addr=0x7e7b8869f0 > packet(n+1): sample::ip=0x7e7b8869f8 sample::addr=0xdeadbeefdeadbeef > packet(n+2): sample::ip=0xdeadbeefdeadbeeb sample::addr=0x57224322f4 > > So we also can rely on this to get to know there have one address > range is [0xdeadbeefdeadbeef..0xdeadbeefdeadbeeb] to indicate there > have a discontinuity in the trace. Yes, I agree you need the extra branch sample from cs_etm__flush(). With a discontinuity in trace, I get output from perf script like this: branches:u: 59ee6e2e08 sqlite3VdbeExec (speedtest1) => 59ee6e2e64 sqlite3VdbeExec (spe branches:u: 59ee6e2e7c sqlite3VdbeExec (speedtest1) => 59ee6e2eec sqlite3VdbeExec (spe branches:u: 59ee6e2efc sqlite3VdbeExec (speedtest1) => 59ee6e2f14 sqlite3VdbeExec (spe branches:u: 59ee6e2f3c sqlite3VdbeExec (speedtest1) => deadbeefdeadbeef [unknown] ([unknown]) branches:u: deadbeefdeadbeeb [unknown] ([unknown]) => 769949daa0 memcpy (/system/lib64/libc.so) branches:u: 769949dacc memcpy (/system/lib64/libc.so) => 59ee6f0664 insertCell (speedtest1) branches:u: 59ee6f0664 insertCell (speedtest1) => 59ee6f0684 insertCell (speedtest1) branches:u: 59ee6f06a4 insertCell (speedtest1) => 59ee6a4d50 memmove@plt (speedtest1) branches:u: 59ee6a4d5c memmove@plt (speedtest1) => 769949ebf8 memmove (/system/lib64/libc.so) Showing there is a break in trace between 59ee6e2f3c and 769949daa0. The deadbeefdeadbeef addresses are a bit ugly - these are just dummy values emitted in the decoder layer - maybe these should be changed to 0. Or you could add a new sample type (i.e. not branch) to indicate the start / end of trace, with only the valid address. With this change, it becomes the same as the patch from your previous mail. Regards Rob >> Regards >> >> Rob >> >>> diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c >>> index 822ba91..dd354ad 100644 >>> --- a/tools/perf/util/cs-etm.c >>> +++ b/tools/perf/util/cs-etm.c >>> @@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address >>> * >>> * A64 instructions are constant size, so the last executed >>> @@ -897,13 +904,27 @@ 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) >>> + generate_sample = true; >> Also check for etmq->prev_packet->sample_type == CS_ETM_TRACE_ON here and >> set generate_sample = true. > Agree, will add this. > >>> + >>> + /* 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 +943,12 @@ 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_packet *tmp; >>> if (etmq->etm->synth_opts.last_branch && >>> etmq->prev_packet && >>> - etmq->prev_packet->sample_type == CS_ETM_RANGE) { >>> + etmq->prev_packet->sample_type == CS_ETM_RANGE && >>> + etmq->period_instructions) { >>> + >> I don't think this is needed. > Okay, I will keep this. > >>> /* >>> * Generate a last branch event for the branches left in the >>> * circular buffer at the end of the trace. >>> @@ -940,14 +962,6 @@ static int cs_etm__flush(struct cs_etm_queue *etmq) >>> etmq, addr, >>> etmq->period_instructions); >>> 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; >> This should not be changed as discussed above. > Okay, will keep this. But I suggest we add some change like below: > > + if (etm->sample_branches) { > + err = cs_etm__synth_branch_sample(etmq); > + if (err) > + return err; > + } > > If so, could you review my posted another patch for this? > http://archive.armlinux.org.uk/lurker/message/20180522.083920.184f1f78.en.html WIll do - with these changes, it is the same as your original patch. > Thanks, > Leo Yan > >>> } >>> return err; >>>
Hi Leo, Following the discussions from your reply to this with a simplified patch, this version of the patch works better as you also need to emit a branch sample when handling a CS_ETM_TRACE_ON packet to indicate the end of a block of trace. This patch does not break the output from perf inject to generate instruction samples for AutoFDO, so I am happy with that. Regards Rob Reviewed-by: Robert Walker <robert.walker@arm.com> On 22/05/18 09:39, Leo Yan wrote: > Hi Rob, > > On Mon, May 21, 2018 at 12:27:42PM +0100, Robert Walker wrote: >> Hi Leo, >> >> On 21/05/18 09:52, 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 packet handling, it only >>> generate samples when 'prev_packet->last_instr_taken_branch' is true, >>> this results in the start tracing packet and exception packets are >>> dropped. >>> >>> This patch checks extra two conditions for complete samples: >>> >>> - If 'prev_packet->sample_type' is zero we can use this condition to >>> get to know this is the start tracing packet; for this case, the start >>> packet's end_addr is zero as well so we need to handle it in the >>> function cs_etm__last_executed_instr(); >>> >> I think you also need to add something in to handle discontinuities in >> trace - for example it is possible to configure the ETM to only trace >> execution in specific code regions or to trace a few cycles every so >> often. In these cases, prev_packet->sample_type will not be zero, but >> whatever the previous packet was. You will get a CS_ETM_TRACE_ON packet in >> such cases, generated by an I_TRACE_ON element in the trace stream. >> You also get this on exception return. >> >> However, you should also keep the test for prev_packet->sample_type == 0 >> as you may not see a CS_ETM_TRACE_ON when decoding a buffer that has >> wrapped. > Thanks for reviewing. Let's dig more detailed into this issue, > especially for handling packet CS_ETM_TRACE_ON, I'd like divide into two > sub cases. > > - The first case is for using python script: > > I use python script to analyze packets with below command: > ./perf script --itrace=ril128 -s arm-cs-trace-disasm.py -F cpu,event,ip,addr,sym -- -v -d objdump -k ./vmlinux > > What I observe is after we pass python script with parameter '-s > arm-cs-trace-disasm.py', then instruction tracing options > '--itrace=ril128' isn't really used; the perf tool creates another > new process for launch python script and re-enter cmd_script() > function, but at the second time when invoke cmd_script() for python > script execution the option '--itrace=ril128' is dropped and all > parameters are only valid defined by the python script. > > As result, I can the variable 'etmq->etm->synth_opts.last_branch' is > always FALSE for running python script. So all CS_ETM_TRACE_ON > packets will be ignored in the function cs_etm__flush(). > > Even the CS_ETM_TRACE_ON packets are missed to handle, the program > flow still can work well. The reason is without the interference by > CS_ETM_TRACE_ON, the CS_ETM_RANGE packets can smoothly create > instruction range by ignore the middle CS_ETM_TRACE_ON packet. > > Please see below example, in this example there have 3 packets, the > first one packet is CS_ETM_RANGE packet which is labelled with > 'PACKET_1', the first one packet can properly generate branch sample > data with previous packet as expected; the second packet is > PACKET_2 which is CS_ETM_TRACE_ON, but > 'etmq->etm->synth_opts.last_branch' is false so function > cs_etm__flush() doesn't handle it and skip the swap operation > "etmq->prev_packet = tmp"; the third packet is PACKET_3, which is > CS_ETM_RANGE packet and we can see it's smoontly to create > continous instruction range between PACKET_1 and PACKET_3. > > cs_etm__sample: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f79c end_addr=0xffff000008a5f7bc last_instr_taken_branch=1 > PACKET_1: cs_etm__sample: packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 > cs_etm__synth_branch_sample: ip=0xffff000008a5f7b8 addr=0xffff000008a5f858 pid=2290 tid=2290 id=1000000021 stream_id=1000000021 period=1 cpu=1 flags=0 cpumode=2 > > cs_etm__flush: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 > PACKET_2: cs_etm__flush: packet: sample_type=2 exc=0 exc_ret=0 cpu=2 start_addr=0xdeadbeefdeadbeef end_addr=0xdeadbeefdeadbeef last_instr_taken_branch=1 > > cs_etm__sample: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 > PACKET_3: cs_etm__sample: packet: sample_type=1 exc=0 exc_ret=0 cpu=2 start_addr=0xffff000008be7528 end_addr=0xffff000008be7538 last_instr_taken_branch=1 > cs_etm__synth_branch_sample: ip=0xffff000008a5f860 addr=0xffff000008be7528 pid=2290 tid=2290 id=1000000021 stream_id=1000000021 period=1 cpu=2 flags=0 cpumode=2 > > So seems to me, the CS_ETM_TRACE_ON packet doesn't introduce trouble > for the program flow analysis if we can handle all CS_ETM_RANGE > packets and without handling CS_ETM_TRACE_ON packet for branch > samples. > > - The second case is for --itrace option without python script: > ./perf script --itrace=ril -F cpu,event,ip,addr,sym -k ./vmlinux > > In this case, the flag 'etmq->etm->synth_opts.last_branch' is true > so CS_ETM_TRACE_ON packet will be handled; but I can observe the > CS_ETM_RANGE packet in etmq->prev_packet isn't handled in the > function cs_etm__flush() for branch sample, so actually we miss some > branch sample for this case. > > So I think we also need handle CS_ETM_RANGE packet in function > cs_etm__flush() to generate branch samples. But this has side > effect, we introduce the extra track for CS_ETM_TRACE_ON packet for > branch samples, so we will see one branch range like: > [ 0xdeadbeefdeadbeef .. 0xdeadbeefdeadbeef ]. > > Please reivew below change is okay for you? Thanks a lot for > suggestions. > > diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c > index 822ba91..37d3722 100644 > --- a/tools/perf/util/cs-etm.c > +++ b/tools/perf/util/cs-etm.c > @@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address > * > * A64 instructions are constant size, so the last executed > @@ -897,13 +904,28 @@ 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) { > @@ -921,12 +943,17 @@ static int cs_etm__sample(struct cs_etm_queue *etmq) > > static int cs_etm__flush(struct cs_etm_queue *etmq) > { > + struct cs_etm_auxtrace *etm = etmq->etm; > int err = 0; > 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 +966,25 @@ 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)
Em Fri, May 25, 2018 at 03:03:47PM +0100, Robert Walker escreveu: > Hi Leo, > > Following the discussions from your reply to this with a simplified patch, > this version of the patch works better as you also need to emit a branch > sample when handling a CS_ETM_TRACE_ON packet to indicate the end of a block > of trace. > > This patch does not break the output from perf inject to generate > instruction samples for AutoFDO, so I am happy with that. > > Regards > > Rob > > Reviewed-by: Robert Walker <robert.walker@arm.com> So, Leo, can you please resubmit, bumping the v2 to v3 (or the latest one, I haven't fully reread this thread) add this "Reviewed-by: Robert" tag and any other that people may have provided, so that I can merge it? Thanks, - Arnaldo > > On 22/05/18 09:39, Leo Yan wrote: > > Hi Rob, > > > > On Mon, May 21, 2018 at 12:27:42PM +0100, Robert Walker wrote: > > > Hi Leo, > > > > > > On 21/05/18 09:52, 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 packet handling, it only > > > > generate samples when 'prev_packet->last_instr_taken_branch' is true, > > > > this results in the start tracing packet and exception packets are > > > > dropped. > > > > > > > > This patch checks extra two conditions for complete samples: > > > > > > > > - If 'prev_packet->sample_type' is zero we can use this condition to > > > > get to know this is the start tracing packet; for this case, the start > > > > packet's end_addr is zero as well so we need to handle it in the > > > > function cs_etm__last_executed_instr(); > > > > > > > I think you also need to add something in to handle discontinuities in > > > trace - for example it is possible to configure the ETM to only trace > > > execution in specific code regions or to trace a few cycles every so > > > often. In these cases, prev_packet->sample_type will not be zero, but > > > whatever the previous packet was. You will get a CS_ETM_TRACE_ON packet in > > > such cases, generated by an I_TRACE_ON element in the trace stream. > > > You also get this on exception return. > > > > > > However, you should also keep the test for prev_packet->sample_type == 0 > > > as you may not see a CS_ETM_TRACE_ON when decoding a buffer that has > > > wrapped. > > Thanks for reviewing. Let's dig more detailed into this issue, > > especially for handling packet CS_ETM_TRACE_ON, I'd like divide into two > > sub cases. > > > > - The first case is for using python script: > > > > I use python script to analyze packets with below command: > > ./perf script --itrace=ril128 -s arm-cs-trace-disasm.py -F cpu,event,ip,addr,sym -- -v -d objdump -k ./vmlinux > > > > What I observe is after we pass python script with parameter '-s > > arm-cs-trace-disasm.py', then instruction tracing options > > '--itrace=ril128' isn't really used; the perf tool creates another > > new process for launch python script and re-enter cmd_script() > > function, but at the second time when invoke cmd_script() for python > > script execution the option '--itrace=ril128' is dropped and all > > parameters are only valid defined by the python script. > > > > As result, I can the variable 'etmq->etm->synth_opts.last_branch' is > > always FALSE for running python script. So all CS_ETM_TRACE_ON > > packets will be ignored in the function cs_etm__flush(). > > > > Even the CS_ETM_TRACE_ON packets are missed to handle, the program > > flow still can work well. The reason is without the interference by > > CS_ETM_TRACE_ON, the CS_ETM_RANGE packets can smoothly create > > instruction range by ignore the middle CS_ETM_TRACE_ON packet. > > > > Please see below example, in this example there have 3 packets, the > > first one packet is CS_ETM_RANGE packet which is labelled with > > 'PACKET_1', the first one packet can properly generate branch sample > > data with previous packet as expected; the second packet is > > PACKET_2 which is CS_ETM_TRACE_ON, but > > 'etmq->etm->synth_opts.last_branch' is false so function > > cs_etm__flush() doesn't handle it and skip the swap operation > > "etmq->prev_packet = tmp"; the third packet is PACKET_3, which is > > CS_ETM_RANGE packet and we can see it's smoontly to create > > continous instruction range between PACKET_1 and PACKET_3. > > > > cs_etm__sample: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f79c end_addr=0xffff000008a5f7bc last_instr_taken_branch=1 > > PACKET_1: cs_etm__sample: packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 > > cs_etm__synth_branch_sample: ip=0xffff000008a5f7b8 addr=0xffff000008a5f858 pid=2290 tid=2290 id=1000000021 stream_id=1000000021 period=1 cpu=1 flags=0 cpumode=2 > > > > cs_etm__flush: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 > > PACKET_2: cs_etm__flush: packet: sample_type=2 exc=0 exc_ret=0 cpu=2 start_addr=0xdeadbeefdeadbeef end_addr=0xdeadbeefdeadbeef last_instr_taken_branch=1 > > > > cs_etm__sample: prev_packet: sample_type=1 exc=0 exc_ret=0 cpu=1 start_addr=0xffff000008a5f858 end_addr=0xffff000008a5f864 last_instr_taken_branch=1 > > PACKET_3: cs_etm__sample: packet: sample_type=1 exc=0 exc_ret=0 cpu=2 start_addr=0xffff000008be7528 end_addr=0xffff000008be7538 last_instr_taken_branch=1 > > cs_etm__synth_branch_sample: ip=0xffff000008a5f860 addr=0xffff000008be7528 pid=2290 tid=2290 id=1000000021 stream_id=1000000021 period=1 cpu=2 flags=0 cpumode=2 > > > > So seems to me, the CS_ETM_TRACE_ON packet doesn't introduce trouble > > for the program flow analysis if we can handle all CS_ETM_RANGE > > packets and without handling CS_ETM_TRACE_ON packet for branch > > samples. > > > > - The second case is for --itrace option without python script: > > ./perf script --itrace=ril -F cpu,event,ip,addr,sym -k ./vmlinux > > > > In this case, the flag 'etmq->etm->synth_opts.last_branch' is true > > so CS_ETM_TRACE_ON packet will be handled; but I can observe the > > CS_ETM_RANGE packet in etmq->prev_packet isn't handled in the > > function cs_etm__flush() for branch sample, so actually we miss some > > branch sample for this case. > > > > So I think we also need handle CS_ETM_RANGE packet in function > > cs_etm__flush() to generate branch samples. But this has side > > effect, we introduce the extra track for CS_ETM_TRACE_ON packet for > > branch samples, so we will see one branch range like: > > [ 0xdeadbeefdeadbeef .. 0xdeadbeefdeadbeef ]. > > > > Please reivew below change is okay for you? Thanks a lot for > > suggestions. > > > > diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c > > index 822ba91..37d3722 100644 > > --- a/tools/perf/util/cs-etm.c > > +++ b/tools/perf/util/cs-etm.c > > @@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address > > * > > * A64 instructions are constant size, so the last executed > > @@ -897,13 +904,28 @@ 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) { > > @@ -921,12 +943,17 @@ static int cs_etm__sample(struct cs_etm_queue *etmq) > > static int cs_etm__flush(struct cs_etm_queue *etmq) > > { > > + struct cs_etm_auxtrace *etm = etmq->etm; > > int err = 0; > > 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 +966,25 @@ 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)
Hi Arnaldo, Rob, On Fri, May 25, 2018 at 12:27:13PM -0300, Arnaldo Carvalho de Melo wrote: > Em Fri, May 25, 2018 at 03:03:47PM +0100, Robert Walker escreveu: > > Hi Leo, > > > > Following the discussions from your reply to this with a simplified patch, > > this version of the patch works better as you also need to emit a branch > > sample when handling a CS_ETM_TRACE_ON packet to indicate the end of a block > > of trace. I also will follow the suggestion as Rob mentioned in another email: "The deadbeefdeadbeef addresses are a bit ugly - these are just dummy values emitted in the decoder layer - maybe these should be changed to 0." > > This patch does not break the output from perf inject to generate > > instruction samples for AutoFDO, so I am happy with that. Thanks for confirmation. > > Regards > > > > Rob > > > > Reviewed-by: Robert Walker <robert.walker@arm.com> > > So, Leo, can you please resubmit, bumping the v2 to v3 (or the latest > one, I haven't fully reread this thread) add this "Reviewed-by: Robert" > tag and any other that people may have provided, so that I can merge it? Sure! I will respin the v3 patch series by following up Rob's suggestion and add Rob's review tag. BTW, I'd like to get ack from Mathieu as well. Mathieu is working on CPU wide tracing, so I talked with Mathieu he will review the patch series if has conflict with CPU wide tracing. [...] Thanks, Leo Yan
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index 822ba91..378953b 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -495,6 +495,13 @@ 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 records the execution range with an exclusive end address * * A64 instructions are constant size, so the last executed @@ -897,13 +904,27 @@ 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) + 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) {
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 packet handling, it only generate samples when 'prev_packet->last_instr_taken_branch' is true, this results in the start tracing packet and exception packets are dropped. This patch checks extra two conditions for complete samples: - If 'prev_packet->sample_type' is zero we can use this condition to get to know this is the start tracing packet; for this case, the start packet's end_addr is zero as well so we need to handle it in the function cs_etm__last_executed_instr(); - If 'prev_packet->exc' is true, we can know the previous packet is exception handling packet so need to generate sample for exception flow. Fixes: e573e978fb12 ("perf cs-etm: Inject capabilitity for CoreSight traces") Cc: Mike Leach <mike.leach@arm.com> Cc: Robert Walker <robert.walker@arm.com> Cc: Mathieu Poirier <mathieu.poirier@linaro.org> Signed-off-by: Leo Yan <leo.yan@linaro.org> --- tools/perf/util/cs-etm.c | 35 ++++++++++++++++++++++++++++------- 1 file changed, 28 insertions(+), 7 deletions(-) -- 2.7.4