Fix a regression caused by an incorrect assumption about auxtrace indices. Then add a test that would have caught it.
--- Changes in v2: - Fix cover letter and subject - Link to v1: https://lore.kernel.org/r/20260119-james-perf-coresight-cpu-map-segfault-v1-...
--- James Clark (2): perf cs-etm: Fix decoding for sparse CPU maps perf cs-etm: Test sparse CPU maps
tools/perf/tests/shell/test_arm_coresight.sh | 54 ++++++++++++++++++++++++++++ tools/perf/util/cs-etm.c | 3 +- 2 files changed, 56 insertions(+), 1 deletion(-) --- base-commit: 571d29baa07e83e637075239f379f91353c24ec9 change-id: 20260115-james-perf-coresight-cpu-map-segfault-e250af5aa778
Best regards,
The ETM decoder incorrectly assumed that auxtrace queue indices were equivalent to CPU number. This assumption is used for inserting records into the queue, and for fetching queues when given a CPU number. This assumption held when Perf always opened a dummy event on every CPU, even if the user provided a subset of CPUs on the commandline, resulting in the indices aligning.
For example:
# event : name = cs_etm//u, , id = { 2451, 2452 }, type = 11 (cs_etm), size = 136, config = 0x4010, { sample_period, samp> # event : name = dummy:u, , id = { 2453, 2454, 2455, 2456 }, type = 1 (PERF_TYPE_SOFTWARE), size = 136, config = 0x9 (PER>
0 0 0x200 [0xd0]: PERF_RECORD_ID_INDEX nr: 6 ... id: 2451 idx: 2 cpu: 2 tid: -1 ... id: 2452 idx: 3 cpu: 3 tid: -1 ... id: 2453 idx: 0 cpu: 0 tid: -1 ... id: 2454 idx: 1 cpu: 1 tid: -1 ... id: 2455 idx: 2 cpu: 2 tid: -1 ... id: 2456 idx: 3 cpu: 3 tid: -1
Since commit 811082e4b668 ("perf parse-events: Support user CPUs mixed with threads/processes") the dummy event no longer behaves in this way, making the ETM event indices start from 0 on the first CPU recorded regardless of its ID:
# event : name = cs_etm//u, , id = { 771, 772 }, type = 11 (cs_etm), size = 144, config = 0x4010, { sample_period, sample> # event : name = dummy:u, , id = { 773, 774 }, type = 1 (PERF_TYPE_SOFTWARE), size = 144, config = 0x9 (PERF_COUNT_SW_DUM>
0 0 0x200 [0x90]: PERF_RECORD_ID_INDEX nr: 4 ... id: 771 idx: 0 cpu: 2 tid: -1 ... id: 772 idx: 1 cpu: 3 tid: -1 ... id: 773 idx: 0 cpu: 2 tid: -1 ... id: 774 idx: 1 cpu: 3 tid: -1
This causes the following segfault when decoding:
$ perf record -e cs_etm//u -C 2,3 -- true $ perf report
perf: Segmentation fault -------- backtrace -------- #0 0xaaaabf9fd020 in ui__signal_backtrace setup.c:110 #1 0xffffab5c7930 in __kernel_rt_sigreturn [vdso][930] #2 0xaaaabfb68d30 in cs_etm_decoder__reset cs-etm-decoder.c:85 #3 0xaaaabfb65930 in cs_etm__get_data_block cs-etm.c:2032 #4 0xaaaabfb666fc in cs_etm__run_per_cpu_timeless_decoder cs-etm.c:2551 #5 0xaaaabfb6692c in (cs_etm__process_timeless_queues cs-etm.c:2612 #6 0xaaaabfb63390 in cs_etm__flush_events cs-etm.c:921 #7 0xaaaabfb324c0 in auxtrace__flush_events auxtrace.c:2915 #8 0xaaaabfaac378 in __perf_session__process_events session.c:2285 #9 0xaaaabfaacc9c in perf_session__process_events session.c:2442 #10 0xaaaabf8d3d90 in __cmd_report builtin-report.c:1085 #11 0xaaaabf8d6944 in cmd_report builtin-report.c:1866 #12 0xaaaabf95ebfc in run_builtin perf.c:351 #13 0xaaaabf95eeb0 in handle_internal_command perf.c:404 #14 0xaaaabf95f068 in run_argv perf.c:451 #15 0xaaaabf95f390 in main perf.c:558 #16 0xffffaab97400 in __libc_start_call_main libc_start_call_main.h:74 #17 0xffffaab974d8 in __libc_start_main@@GLIBC_2.34 libc-start.c:128 #18 0xaaaabf8aa8f0 in _start perf[7a8f0]
Fix it by inserting into the queues based on CPU number, rather than using the index.
Fixes: 811082e4b668 ("perf parse-events: Support user CPUs mixed with threads/processes") Signed-off-by: James Clark james.clark@linaro.org --- tools/perf/util/cs-etm.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index 25d56e0f1c07..12b55c2bc2ca 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -3086,7 +3086,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o
if (aux_offset >= auxtrace_event->offset && aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) { - struct cs_etm_queue *etmq = etm->queues.queue_array[auxtrace_event->idx].priv; + struct cs_etm_queue *etmq = cs_etm__get_queue(etm, auxtrace_event->cpu);
/* * If this AUX event was inside this buffer somewhere, create a new auxtrace event @@ -3095,6 +3095,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o auxtrace_fragment.auxtrace = *auxtrace_event; auxtrace_fragment.auxtrace.size = aux_size; auxtrace_fragment.auxtrace.offset = aux_offset; + auxtrace_fragment.auxtrace.idx = etmq->queue_nr; file_offset += aux_offset - auxtrace_event->offset + auxtrace_event->header.size;
pr_debug3("CS ETM: Queue buffer size: %#"PRI_lx64" offset: %#"PRI_lx64
On Mon, Jan 19, 2026 at 10:18:35AM +0000, Coresight ML wrote:
The ETM decoder incorrectly assumed that auxtrace queue indices were equivalent to CPU number. This assumption is used for inserting records into the queue, and for fetching queues when given a CPU number. This assumption held when Perf always opened a dummy event on every CPU, even if the user provided a subset of CPUs on the commandline, resulting in the indices aligning.
For example:
# event : name = cs_etm//u, , id = { 2451, 2452 }, type = 11 (cs_etm), size = 136, config = 0x4010, { sample_period, samp> # event : name = dummy:u, , id = { 2453, 2454, 2455, 2456 }, type = 1 (PERF_TYPE_SOFTWARE), size = 136, config = 0x9 (PER>
0 0 0x200 [0xd0]: PERF_RECORD_ID_INDEX nr: 6 ... id: 2451 idx: 2 cpu: 2 tid: -1 ... id: 2452 idx: 3 cpu: 3 tid: -1 ... id: 2453 idx: 0 cpu: 0 tid: -1 ... id: 2454 idx: 1 cpu: 1 tid: -1 ... id: 2455 idx: 2 cpu: 2 tid: -1 ... id: 2456 idx: 3 cpu: 3 tid: -1
Since commit 811082e4b668 ("perf parse-events: Support user CPUs mixed with threads/processes") the dummy event no longer behaves in this way, making the ETM event indices start from 0 on the first CPU recorded regardless of its ID:
# event : name = cs_etm//u, , id = { 771, 772 }, type = 11 (cs_etm), size = 144, config = 0x4010, { sample_period, sample> # event : name = dummy:u, , id = { 773, 774 }, type = 1 (PERF_TYPE_SOFTWARE), size = 144, config = 0x9 (PERF_COUNT_SW_DUM>
0 0 0x200 [0x90]: PERF_RECORD_ID_INDEX nr: 4 ... id: 771 idx: 0 cpu: 2 tid: -1 ... id: 772 idx: 1 cpu: 3 tid: -1 ... id: 773 idx: 0 cpu: 2 tid: -1 ... id: 774 idx: 1 cpu: 3 tid: -1
Seems to me that this patch works around the issue by using the CPU ID instead, but event->auxtrace.idx is broken.
Should we store the correct index in event->auxtrace.idx (e.g., in the __perf_event__synthesize_id_index()) ?
Thanks, Leo
This causes the following segfault when decoding:
$ perf record -e cs_etm//u -C 2,3 -- true $ perf report
perf: Segmentation fault -------- backtrace -------- #0 0xaaaabf9fd020 in ui__signal_backtrace setup.c:110 #1 0xffffab5c7930 in __kernel_rt_sigreturn [vdso][930] #2 0xaaaabfb68d30 in cs_etm_decoder__reset cs-etm-decoder.c:85 #3 0xaaaabfb65930 in cs_etm__get_data_block cs-etm.c:2032 #4 0xaaaabfb666fc in cs_etm__run_per_cpu_timeless_decoder cs-etm.c:2551 #5 0xaaaabfb6692c in (cs_etm__process_timeless_queues cs-etm.c:2612 #6 0xaaaabfb63390 in cs_etm__flush_events cs-etm.c:921 #7 0xaaaabfb324c0 in auxtrace__flush_events auxtrace.c:2915 #8 0xaaaabfaac378 in __perf_session__process_events session.c:2285 #9 0xaaaabfaacc9c in perf_session__process_events session.c:2442 #10 0xaaaabf8d3d90 in __cmd_report builtin-report.c:1085 #11 0xaaaabf8d6944 in cmd_report builtin-report.c:1866 #12 0xaaaabf95ebfc in run_builtin perf.c:351 #13 0xaaaabf95eeb0 in handle_internal_command perf.c:404 #14 0xaaaabf95f068 in run_argv perf.c:451 #15 0xaaaabf95f390 in main perf.c:558 #16 0xffffaab97400 in __libc_start_call_main libc_start_call_main.h:74 #17 0xffffaab974d8 in __libc_start_main@@GLIBC_2.34 libc-start.c:128 #18 0xaaaabf8aa8f0 in _start perf[7a8f0]
Fix it by inserting into the queues based on CPU number, rather than using the index.
Fixes: 811082e4b668 ("perf parse-events: Support user CPUs mixed with threads/processes") Signed-off-by: James Clark james.clark@linaro.org
tools/perf/util/cs-etm.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index 25d56e0f1c07..12b55c2bc2ca 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -3086,7 +3086,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o if (aux_offset >= auxtrace_event->offset && aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) {
struct cs_etm_queue *etmq = etm->queues.queue_array[auxtrace_event->idx].priv;
struct cs_etm_queue *etmq = cs_etm__get_queue(etm, auxtrace_event->cpu);/* * If this AUX event was inside this buffer somewhere, create a new auxtrace event @@ -3095,6 +3095,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o auxtrace_fragment.auxtrace = *auxtrace_event; auxtrace_fragment.auxtrace.size = aux_size; auxtrace_fragment.auxtrace.offset = aux_offset;
auxtrace_fragment.auxtrace.idx = etmq->queue_nr;pr_debug3("CS ETM: Queue buffer size: %#"PRI_lx64" offset: %#"PRI_lx64
-- 2.34.1
CoreSight mailing list -- coresight@lists.linaro.org To unsubscribe send an email to coresight-leave@lists.linaro.org
On 19/01/2026 11:15 am, Leo Yan wrote:
On Mon, Jan 19, 2026 at 10:18:35AM +0000, Coresight ML wrote:
The ETM decoder incorrectly assumed that auxtrace queue indices were equivalent to CPU number. This assumption is used for inserting records into the queue, and for fetching queues when given a CPU number. This assumption held when Perf always opened a dummy event on every CPU, even if the user provided a subset of CPUs on the commandline, resulting in the indices aligning.
For example:
# event : name = cs_etm//u, , id = { 2451, 2452 }, type = 11 (cs_etm), size = 136, config = 0x4010, { sample_period, samp> # event : name = dummy:u, , id = { 2453, 2454, 2455, 2456 }, type = 1 (PERF_TYPE_SOFTWARE), size = 136, config = 0x9 (PER>
0 0 0x200 [0xd0]: PERF_RECORD_ID_INDEX nr: 6 ... id: 2451 idx: 2 cpu: 2 tid: -1 ... id: 2452 idx: 3 cpu: 3 tid: -1 ... id: 2453 idx: 0 cpu: 0 tid: -1 ... id: 2454 idx: 1 cpu: 1 tid: -1 ... id: 2455 idx: 2 cpu: 2 tid: -1 ... id: 2456 idx: 3 cpu: 3 tid: -1
Since commit 811082e4b668 ("perf parse-events: Support user CPUs mixed with threads/processes") the dummy event no longer behaves in this way, making the ETM event indices start from 0 on the first CPU recorded regardless of its ID:
# event : name = cs_etm//u, , id = { 771, 772 }, type = 11 (cs_etm), size = 144, config = 0x4010, { sample_period, sample> # event : name = dummy:u, , id = { 773, 774 }, type = 1 (PERF_TYPE_SOFTWARE), size = 144, config = 0x9 (PERF_COUNT_SW_DUM>
0 0 0x200 [0x90]: PERF_RECORD_ID_INDEX nr: 4 ... id: 771 idx: 0 cpu: 2 tid: -1 ... id: 772 idx: 1 cpu: 3 tid: -1 ... id: 773 idx: 0 cpu: 2 tid: -1 ... id: 774 idx: 1 cpu: 3 tid: -1
Seems to me that this patch works around the issue by using the CPU ID instead, but event->auxtrace.idx is broken.
Should we store the correct index in event->auxtrace.idx (e.g., in the __perf_event__synthesize_id_index()) ?
Thanks, Leo
I don't think it's a workaround, I think it should have been written this way in the first place.
Idx is just what it says it is, it's the position in the array. If there are only two events then they are idx 0 and 1, but CPU can be anything. If ETM didn't open a dummy event at the same time then I think it always would have behaved like that.
This causes the following segfault when decoding:
$ perf record -e cs_etm//u -C 2,3 -- true $ perf report
perf: Segmentation fault -------- backtrace -------- #0 0xaaaabf9fd020 in ui__signal_backtrace setup.c:110 #1 0xffffab5c7930 in __kernel_rt_sigreturn [vdso][930] #2 0xaaaabfb68d30 in cs_etm_decoder__reset cs-etm-decoder.c:85 #3 0xaaaabfb65930 in cs_etm__get_data_block cs-etm.c:2032 #4 0xaaaabfb666fc in cs_etm__run_per_cpu_timeless_decoder cs-etm.c:2551 #5 0xaaaabfb6692c in (cs_etm__process_timeless_queues cs-etm.c:2612 #6 0xaaaabfb63390 in cs_etm__flush_events cs-etm.c:921 #7 0xaaaabfb324c0 in auxtrace__flush_events auxtrace.c:2915 #8 0xaaaabfaac378 in __perf_session__process_events session.c:2285 #9 0xaaaabfaacc9c in perf_session__process_events session.c:2442 #10 0xaaaabf8d3d90 in __cmd_report builtin-report.c:1085 #11 0xaaaabf8d6944 in cmd_report builtin-report.c:1866 #12 0xaaaabf95ebfc in run_builtin perf.c:351 #13 0xaaaabf95eeb0 in handle_internal_command perf.c:404 #14 0xaaaabf95f068 in run_argv perf.c:451 #15 0xaaaabf95f390 in main perf.c:558 #16 0xffffaab97400 in __libc_start_call_main libc_start_call_main.h:74 #17 0xffffaab974d8 in __libc_start_main@@GLIBC_2.34 libc-start.c:128 #18 0xaaaabf8aa8f0 in _start perf[7a8f0]
Fix it by inserting into the queues based on CPU number, rather than using the index.
Fixes: 811082e4b668 ("perf parse-events: Support user CPUs mixed with threads/processes") Signed-off-by: James Clark james.clark@linaro.org
tools/perf/util/cs-etm.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index 25d56e0f1c07..12b55c2bc2ca 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -3086,7 +3086,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o if (aux_offset >= auxtrace_event->offset && aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) {
struct cs_etm_queue *etmq = etm->queues.queue_array[auxtrace_event->idx].priv;
struct cs_etm_queue *etmq = cs_etm__get_queue(etm, auxtrace_event->cpu);/* * If this AUX event was inside this buffer somewhere, create a new auxtrace event @@ -3095,6 +3095,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o auxtrace_fragment.auxtrace = *auxtrace_event; auxtrace_fragment.auxtrace.size = aux_size; auxtrace_fragment.auxtrace.offset = aux_offset;
auxtrace_fragment.auxtrace.idx = etmq->queue_nr;pr_debug3("CS ETM: Queue buffer size: %#"PRI_lx64" offset: %#"PRI_lx64
-- 2.34.1
CoreSight mailing list -- coresight@lists.linaro.org To unsubscribe send an email to coresight-leave@lists.linaro.org
On Mon, Jan 19, 2026 at 11:55:53AM +0000, James Clark wrote:
[...]
0 0 0x200 [0x90]: PERF_RECORD_ID_INDEX nr: 4 ... id: 771 idx: 0 cpu: 2 tid: -1 ... id: 772 idx: 1 cpu: 3 tid: -1 ... id: 773 idx: 0 cpu: 2 tid: -1 ... id: 774 idx: 1 cpu: 3 tid: -1
Seems to me that this patch works around the issue by using the CPU ID instead, but event->auxtrace.idx is broken.
I don't think it's a workaround, I think it should have been written this way in the first place.
Agree. Sorry for confusion in my rely.
[...]
@@ -3086,7 +3086,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o if (aux_offset >= auxtrace_event->offset && aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) {
struct cs_etm_queue *etmq = etm->queues.queue_array[auxtrace_event->idx].priv;
struct cs_etm_queue *etmq = cs_etm__get_queue(etm, auxtrace_event->cpu);
During decoding, cs_etm allocates its own AUX queues and needs to convert the AUX records into a cs_etm-specific format.
For example:
event_id=405 idx=0 cpu=0 tid=6673 event_id=406 idx=1 cpu=1 tid=6673 event_id=407 idx=2 cpu=3 tid=6673 event_id=408 idx=3 cpu=4 tid=6673 event_id=409 idx=4 cpu=5 tid=6673 ` ` `> CPU ID `> Event ID `> Generic buffer IDs
In this case, I deliberately hotplugged off CPU2. As a result, the buffer IDs are consecutive from 0 to 4, while the CPU ID for CPU2 is missing. When adding AUX records to the cs_etm queues, we need to convert from the generic buffer index to the corresponding CPU ID.
So the above change makes sense to me.
/* * If this AUX event was inside this buffer somewhere, create a new auxtrace event@@ -3095,6 +3095,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o auxtrace_fragment.auxtrace = *auxtrace_event; auxtrace_fragment.auxtrace.size = aux_size; auxtrace_fragment.auxtrace.offset = aux_offset;
auxtrace_fragment.auxtrace.idx = etmq->queue_nr;
I am still confused about this. Because above auxtrace will be queued on cs_etm queues, should we convert from the generic buffer index to the cs_etm specific one? E.g.,
auxtrace_fragment.auxtrace.idx = auxtrace_event->cpu;
BTW, if my understanding above is valid, it is good to go through the cs_etm.c file for the "idx <-> CPU ID" conversion.
Thanks, Leo
file_offset += aux_offset - auxtrace_event->offset + auxtrace_event->header.size; pr_debug3("CS ETM: Queue buffer size: %#"PRI_lx64" offset: %#"PRI_lx64-- 2.34.1
CoreSight mailing list -- coresight@lists.linaro.org To unsubscribe send an email to coresight-leave@lists.linaro.org
On 19/01/2026 2:53 pm, Leo Yan wrote:
On Mon, Jan 19, 2026 at 11:55:53AM +0000, James Clark wrote:
[...]
0 0 0x200 [0x90]: PERF_RECORD_ID_INDEX nr: 4 ... id: 771 idx: 0 cpu: 2 tid: -1 ... id: 772 idx: 1 cpu: 3 tid: -1 ... id: 773 idx: 0 cpu: 2 tid: -1 ... id: 774 idx: 1 cpu: 3 tid: -1Seems to me that this patch works around the issue by using the CPU ID instead, but event->auxtrace.idx is broken.
I don't think it's a workaround, I think it should have been written this way in the first place.
Agree. Sorry for confusion in my rely.
[...]
@@ -3086,7 +3086,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o if (aux_offset >= auxtrace_event->offset && aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) {
struct cs_etm_queue *etmq = etm->queues.queue_array[auxtrace_event->idx].priv;
struct cs_etm_queue *etmq = cs_etm__get_queue(etm, auxtrace_event->cpu);During decoding, cs_etm allocates its own AUX queues and needs to convert the AUX records into a cs_etm-specific format.
For example:
event_id=405 idx=0 cpu=0 tid=6673 event_id=406 idx=1 cpu=1 tid=6673 event_id=407 idx=2 cpu=3 tid=6673 event_id=408 idx=3 cpu=4 tid=6673 event_id=409 idx=4 cpu=5 tid=6673 ` ` `> CPU ID `> Event ID `> Generic buffer IDs
In this case, I deliberately hotplugged off CPU2. As a result, the buffer IDs are consecutive from 0 to 4, while the CPU ID for CPU2 is missing. When adding AUX records to the cs_etm queues, we need to convert from the generic buffer index to the corresponding CPU ID.
So the above change makes sense to me.
/* * If this AUX event was inside this buffer somewhere, create a new auxtrace event@@ -3095,6 +3095,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o auxtrace_fragment.auxtrace = *auxtrace_event; auxtrace_fragment.auxtrace.size = aux_size; auxtrace_fragment.auxtrace.offset = aux_offset;
auxtrace_fragment.auxtrace.idx = etmq->queue_nr;I am still confused about this. Because above auxtrace will be queued on cs_etm queues, should we convert from the generic buffer index to the cs_etm specific one? E.g.,
auxtrace_fragment.auxtrace.idx = auxtrace_event->cpu;
This is basically what the change is already doing. etmq->queue_nr == CPU because cs_etm__setup_queue() is called for every CPU up to max_cpu, not only for ones that were recorded so it has a 1:1 mapping. Using etmq->queue_nr also works in per-thread mode where queue 0 is used, which your suggestion doesn't handle. Otherwise they would be equivalent.
To be honest the whole decoder has become hacks on top of hacks. I think we might want to do a full from scratch re-write when we come to do the proper timestamp to MMAP event correlation or the change to not keep resetting the decoder for TRBE wrap mode.
BTW, if my understanding above is valid, it is good to go through the cs_etm.c file for the "idx <-> CPU ID" conversion.
Do you mean there are other mistakes? I did check the rest and did some testing and didn't see any issues.
Thanks, Leo
file_offset += aux_offset - auxtrace_event->offset + auxtrace_event->header.size; pr_debug3("CS ETM: Queue buffer size: %#"PRI_lx64" offset: %#"PRI_lx64-- 2.34.1
CoreSight mailing list -- coresight@lists.linaro.org To unsubscribe send an email to coresight-leave@lists.linaro.org
On Mon, Jan 19, 2026 at 03:16:01PM +0000, James Clark wrote:
[...]
/* * If this AUX event was inside this buffer somewhere, create a new auxtrace event@@ -3095,6 +3095,7 @@ static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_o auxtrace_fragment.auxtrace = *auxtrace_event; auxtrace_fragment.auxtrace.size = aux_size; auxtrace_fragment.auxtrace.offset = aux_offset;
auxtrace_fragment.auxtrace.idx = etmq->queue_nr;I am still confused about this. Because above auxtrace will be queued on cs_etm queues, should we convert from the generic buffer index to the cs_etm specific one? E.g.,
auxtrace_fragment.auxtrace.idx = auxtrace_event->cpu;
This is basically what the change is already doing. etmq->queue_nr == CPU because cs_etm__setup_queue() is called for every CPU up to max_cpu, not only for ones that were recorded so it has a 1:1 mapping. Using etmq->queue_nr also works in per-thread mode where queue 0 is used, which your suggestion doesn't handle. Otherwise they would be equivalent.
Indeed, per-thread mode needs to be handled separately.
On the Juno board (6 CPUs), I see "etmq->queue_nr=6". It seems a bit odd that this value is used as the index. Even in per-thread mode, the index is still set to 6. For correctness, should we instead set the index to the CPU ID, and use 0 for per-thread mode?
To be honest the whole decoder has become hacks on top of hacks. I think we might want to do a full from scratch re-write when we come to do the proper timestamp to MMAP event correlation or the change to not keep resetting the decoder for TRBE wrap mode.
BTW, if my understanding above is valid, it is good to go through the cs_etm.c file for the "idx <-> CPU ID" conversion.
Do you mean there are other mistakes?
Yes.
I did check the rest and did some testing and didn't see any issues.
Thanks a lot for the checking.
Leo
On Mon, Jan 19, 2026 at 03:43:18PM +0000, Coresight ML wrote:
[...]
auxtrace_fragment.auxtrace.idx = etmq->queue_nr;
Indeed, per-thread mode needs to be handled separately.
On the Juno board (6 CPUs), I see "etmq->queue_nr=6". It seems a bit odd that this value is used as the index. Even in per-thread mode, the index is still set to 6. For correctness, should we instead set the index to the CPU ID, and use 0 for per-thread mode?
Sorry I made a mistake.
I messed up "etm->queues.nr_queues" and "etmq->queue_nr". I can confirm that "etmq->queue_nr" stores the CPU ID and 0 for per-thread mode. I verified the patch:
Tested-by: Leo Yan leo.yan@arm.com
We only currently test with default (all CPUs) or --per-thread mode. Different permutations of the "-C" option can affect decoding so add tests for some of them.
Signed-off-by: James Clark james.clark@linaro.org --- tools/perf/tests/shell/test_arm_coresight.sh | 54 ++++++++++++++++++++++++++++ 1 file changed, 54 insertions(+)
diff --git a/tools/perf/tests/shell/test_arm_coresight.sh b/tools/perf/tests/shell/test_arm_coresight.sh index 1c750b67d141..bbf89e944e7b 100755 --- a/tools/perf/tests/shell/test_arm_coresight.sh +++ b/tools/perf/tests/shell/test_arm_coresight.sh @@ -198,6 +198,58 @@ arm_cs_etm_basic_test() { arm_cs_report "CoreSight basic testing with '$*'" $err }
+arm_cs_etm_test_cpu_list() { + echo "Testing sparse CPU list: $1" + perf record -o ${perfdata} -e cs_etm//u -C $1 \ + -- taskset --cpu-list $1 true > /dev/null 2>&1 + perf_script_branch_samples true + err=$? + arm_cs_report "CoreSight sparse CPUs with '$*'" $err +} + +arm_cs_etm_sparse_cpus_test() { + # Iterate for every ETM device + cpus=() + for dev in /sys/bus/event_source/devices/cs_etm/cpu*; do + # Canonicalize the path + dev=`readlink -f $dev` + + # Find the ETM device belonging to which CPU + cpus+=("$(cat $dev/cpu)") + done + + mapfile -t cpus < <(printf '%s\n' "${cpus[@]}" | sort -n) + total=${#cpus[@]} + + # Need more than 1 to test + if [ $total -le 1 ]; then + return 0 + fi + + half=$((total / 2)) + + # First half + first_half=$(IFS=,; echo "${cpus[*]:0:$half}") + arm_cs_etm_test_cpu_list $first_half + + # Second half + second_half=$(IFS=,; echo "${cpus[*]:$half}") + arm_cs_etm_test_cpu_list $second_half + + # Odd list is the same as halves unless >= 4 CPUs + if [ $total -lt 4 ]; then + return 0 + fi + + # Odd indices + odd_cpus=() + for ((i=1; i<total; i+=2)); do + odd_cpus+=("${cpus[$i]}") + done + odd_list=$(IFS=,; echo "${odd_cpus[*]}") + arm_cs_etm_test_cpu_list $odd_list +} + arm_cs_etm_traverse_path_test arm_cs_etm_system_wide_test arm_cs_etm_snapshot_test @@ -211,4 +263,6 @@ arm_cs_etm_basic_test -e cs_etm/timestamp=1/ -a arm_cs_etm_basic_test -e cs_etm/timestamp=0/ arm_cs_etm_basic_test -e cs_etm/timestamp=1/
+arm_cs_etm_sparse_cpus_test + exit $glb_err
On Mon, Jan 19, 2026 at 10:18:36AM +0000, Coresight ML wrote:
We only currently test with default (all CPUs) or --per-thread mode. Different permutations of the "-C" option can affect decoding so add tests for some of them.
Signed-off-by: James Clark james.clark@linaro.org
LGTM and verified on Juno board:
Tested-by: Leo Yan leo.yan@arm.com
On Tue, Jan 20, 2026 at 11:22:38AM +0000, Leo Yan wrote:
On Mon, Jan 19, 2026 at 10:18:36AM +0000, Coresight ML wrote:
We only currently test with default (all CPUs) or --per-thread mode. Different permutations of the "-C" option can affect decoding so add tests for some of them.
Signed-off-by: James Clark james.clark@linaro.org
LGTM and verified on Juno board:
Tested-by: Leo Yan leo.yan@arm.com
Thanks, applied both patches to perf-tools-next,
- Arnaldo
-
Arnaldo Carvalho de Melo -
James Clark -
Leo Yan