Re: [PATCH] perf cs-etm: Add support for coresight trace for any range of CPUs

From: Ganapatrao Kulkarni
Date: Thu Apr 20 2023 - 10:30:59 EST


Hi James,

On 20-04-2023 06:07 pm, Ganapatrao Kulkarni wrote:


On 20-04-2023 06:00 pm, James Clark wrote:


On 20/04/2023 12:47, Ganapatrao Kulkarni wrote:

Hi James,

On 20-04-2023 03:13 pm, James Clark wrote:


On 19/04/2023 18:21, Ganapatrao Kulkarni wrote:
The current implementation supports coresight trace for a range of
CPUs, if the first CPU is CPU0.

Adding changes to enable coresight trace for any range of CPUs by
decoding the first CPU also from the header.
Later, first CPU id is used instead of CPU0 across the decoder
functions.

Signed-off-by: Ganapatrao Kulkarni <gankulkarni@xxxxxxxxxxxxxxxxxxxxxx>
---
   .../perf/util/cs-etm-decoder/cs-etm-decoder.c |  4 +-
   .../perf/util/cs-etm-decoder/cs-etm-decoder.h |  3 +-
   tools/perf/util/cs-etm.c                      | 62 ++++++++++++-------
   3 files changed, 42 insertions(+), 27 deletions(-)

diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c
b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c
index 82a27ab90c8b..41ab299b643b 100644
--- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c
+++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c
@@ -724,7 +724,7 @@ cs_etm_decoder__create_etm_decoder(struct
cs_etm_decoder_params *d_params,
   }
     struct cs_etm_decoder *
-cs_etm_decoder__new(int decoders, struct cs_etm_decoder_params
*d_params,
+cs_etm_decoder__new(int first_decoder, int decoders, struct
cs_etm_decoder_params *d_params,
               struct cs_etm_trace_params t_params[])
   {
       struct cs_etm_decoder *decoder;
@@ -769,7 +769,7 @@ cs_etm_decoder__new(int decoders, struct
cs_etm_decoder_params *d_params,
       /* init raw frame logging if required */
       cs_etm_decoder__init_raw_frame_logging(d_params, decoder);
   -    for (i = 0; i < decoders; i++) {
+    for (i = first_decoder; i < decoders; i++) {
           ret = cs_etm_decoder__create_etm_decoder(d_params,
                                &t_params[i],
                                decoder);
diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h
b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h
index 92a855fbe5b8..b06193fc75b4 100644
--- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h
+++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h
@@ -90,7 +90,8 @@ int cs_etm_decoder__process_data_block(struct
cs_etm_decoder *decoder,
                          size_t len, size_t *consumed);
     struct cs_etm_decoder *
-cs_etm_decoder__new(int num_cpu,
+cs_etm_decoder__new(int first_decoder,
+            int decoders,
               struct cs_etm_decoder_params *d_params,
               struct cs_etm_trace_params t_params[]);
   diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c
index 94e2d02009eb..2619513ae088 100644
--- a/tools/perf/util/cs-etm.c
+++ b/tools/perf/util/cs-etm.c
@@ -55,6 +55,8 @@ struct cs_etm_auxtrace {
       u8 has_virtual_ts; /* Virtual/Kernel timestamps in the trace. */
         int num_cpu;
+    int first_cpu;
+    int last_cpu;
       u64 latest_kernel_timestamp;
       u32 auxtrace_type;
       u64 branches_sample_type;
@@ -638,14 +640,13 @@ static void cs_etm__set_trace_param_ete(struct
cs_etm_trace_params *t_params,
   }
     static int cs_etm__init_trace_params(struct cs_etm_trace_params
*t_params,
-                     struct cs_etm_auxtrace *etm,
-                     int decoders)
+                     struct cs_etm_auxtrace *etm)
   {
       int i;
       u32 etmidr;
       u64 architecture;
   -    for (i = 0; i < decoders; i++) {
+    for (i = etm->first_cpu; i < etm->last_cpu; i++) {
           architecture = etm->metadata[i][CS_ETM_MAGIC];
             switch (architecture) {
@@ -817,7 +818,7 @@ static void cs_etm__free(struct perf_session
*session)
       /* Then the RB tree itself */
       intlist__delete(traceid_list);
   -    for (i = 0; i < aux->num_cpu; i++)
+    for (i = aux->first_cpu; i < aux->last_cpu; i++)
           zfree(&aux->metadata[i]);
         thread__zput(aux->unknown_thread);
@@ -921,7 +922,8 @@ static struct cs_etm_queue
*cs_etm__alloc_queue(struct cs_etm_auxtrace *etm,
        * Each queue can only contain data from one CPU when
unformatted, so only one decoder is
        * needed.
        */
-    int decoders = formatted ? etm->num_cpu : 1;
+    int first_decoder = formatted ? etm->first_cpu : 0;
+    int decoders = first_decoder + (formatted ? etm->num_cpu : 1);
         etmq = zalloc(sizeof(*etmq));
       if (!etmq)
@@ -937,7 +939,7 @@ static struct cs_etm_queue
*cs_etm__alloc_queue(struct cs_etm_auxtrace *etm,
       if (!t_params)
           goto out_free;
   -    if (cs_etm__init_trace_params(t_params, etm, decoders))
+    if (cs_etm__init_trace_params(t_params, etm))
           goto out_free;
         /* Set decoder parameters to decode trace packets */
@@ -947,8 +949,7 @@ static struct cs_etm_queue
*cs_etm__alloc_queue(struct cs_etm_auxtrace *etm,
                       formatted))
           goto out_free;
   -    etmq->decoder = cs_etm_decoder__new(decoders, &d_params,
-                        t_params);
+    etmq->decoder = cs_etm_decoder__new(first_decoder, decoders,
&d_params, t_params);
         if (!etmq->decoder)
           goto out_free;
@@ -2959,11 +2960,11 @@ static int cs_etm__queue_aux_records(struct
perf_session *session)
    * Loop through the ETMs and complain if we find at least one where
ts_source != 1 (virtual
    * timestamps).
    */
-static bool cs_etm__has_virtual_ts(u64 **metadata, int num_cpu)
+static bool cs_etm__has_virtual_ts(u64 **metadata, struct
cs_etm_auxtrace *etm)
   {
       int j;
   -    for (j = 0; j < num_cpu; j++) {
+    for (j = etm->first_cpu; j < etm->last_cpu; j++) {
           switch (metadata[j][CS_ETM_MAGIC]) {
           case __perf_cs_etmv4_magic:
               if (HAS_PARAM(j, ETMV4, TS_SOURCE) ||
metadata[j][CS_ETMV4_TS_SOURCE] != 1)
@@ -2982,13 +2983,14 @@ static bool cs_etm__has_virtual_ts(u64
**metadata, int num_cpu)
   }
     /* map trace ids to correct metadata block, from information in
metadata */
-static int cs_etm__map_trace_ids_metadata(int num_cpu, u64 **metadata)
+static int cs_etm__map_trace_ids_metadata(struct cs_etm_auxtrace *etm)
   {
       u64 cs_etm_magic;
+    u64 **metadata = etm->metadata;
       u8 trace_chan_id;
       int i, err;
   -    for (i = 0; i < num_cpu; i++) {
+    for (i = etm->first_cpu; i < etm->last_cpu; i++) {
           cs_etm_magic = metadata[i][CS_ETM_MAGIC];
           switch (cs_etm_magic) {
           case __perf_cs_etmv3_magic:
@@ -3015,12 +3017,13 @@ static int cs_etm__map_trace_ids_metadata(int
num_cpu, u64 **metadata)
    * If we found AUX_HW_ID packets, then set any metadata marked as
unused to the
    * unused value to reduce the number of unneeded decoders created.
    */
-static int cs_etm__clear_unused_trace_ids_metadata(int num_cpu, u64
**metadata)
+static int cs_etm__clear_unused_trace_ids_metadata(struct
cs_etm_auxtrace *etm)
   {
       u64 cs_etm_magic;
+    u64 **metadata = etm->metadata;
       int i;
   -    for (i = 0; i < num_cpu; i++) {
+    for (i = etm->first_cpu; i < etm->last_cpu; i++) {
           cs_etm_magic = metadata[i][CS_ETM_MAGIC];
           switch (cs_etm_magic) {
           case __perf_cs_etmv3_magic:
@@ -3049,7 +3052,7 @@ int cs_etm__process_auxtrace_info_full(union
perf_event *event,
       int event_header_size = sizeof(struct perf_event_header);
       int total_size = auxtrace_info->header.size;
       int priv_size = 0;
-    int num_cpu;
+    int num_cpu, first_cpu = 0, last_cpu;
       int err = 0;
       int aux_hw_id_found;
       int i, j;
@@ -3068,22 +3071,31 @@ int cs_etm__process_auxtrace_info_full(union
perf_event *event,
       /* First the global part */
       ptr = (u64 *) auxtrace_info->priv;
       num_cpu = ptr[CS_PMU_TYPE_CPUS] & 0xffffffff;
-    metadata = zalloc(sizeof(*metadata) * num_cpu);
+
+    /* Start parsing after the common part of the header */
+    i = CS_HEADER_VERSION_MAX;
+
+    /*Get CPU id of first event */
+    first_cpu = ptr[i + CS_ETM_CPU];
+    last_cpu = first_cpu + num_cpu;
+
+    if (first_cpu > cpu__max_cpu().cpu ||
+            last_cpu > cpu__max_cpu().cpu)
+        return -EINVAL;
+
+    metadata = zalloc(sizeof(*metadata) * last_cpu);

Hi Ganapatrao,

I think I see what the problem is, but I'm wondering if a better fix
would be to further decouple the CPU ID from the index in the array.

With your change it's not clear what happens with sparse recordings, for
example 'perf record -e cs_etm// -C 1,3,5'. And it seems like there is

This patch fixes for any range of CPUs.
Record with sparse list of CPUs will not work with current code still.


Is there a major issue that means sparse can't be done? I'm thinking it
would be best to fix both issues with one change while we understand
this part rather than a half fix that might have do be completely
re-understood and re-done later anyway. Unless there is some big blocker
but I can't see it?

some wastage in the zalloc here for example if only CPU 256 is traced
then we'd still make 256 decoders but 255 of them would be unused?

I tried to test sparse recordings, but your change doesn't apply to the
latest coresight/next branch. I did notice that 'perf report -D' doesn't
work with them on coresight/next (it just quits), but I couldn't see if
that's fixed with your change.

My patch is rebased on 6.3-RC7 codebase with Mike's 3 perf patches
related to dynamic id [1] support(queued for 6.4).

"perf report -D" works for me.

I was referring to sparse CPU lists, which I think you mentioned above
doesn't work even with this patch.


[1] https://www.spinics.net/lists/linux-perf-users/msg27452.html


It should be based on the next branch here:
git://git.kernel.org/pub/scm/linux/kernel/git/coresight/linux.git

OK.


Would a better fix not be to keep the metadata loops from 0-N and
instead save the CPU ID in cs_etm_decoder_params or the decoder. That
way it would support both sparse and not starting from 0 cases? I think

Yep, I though this initially, it got complicated due to for more
for-loops. I will try again and post V2.


I can't imagine it would need any extra for loops off the top of my
head. Just saving the CPU ID in a few structs and using it wherever it's
needed instead of the loop index. I imagine most of the loops would
actually stay the same rather than be changed like you have in V1.

Working on V2, I will post it at the earliest.

My codebase is "6.3-RC7 + Mike's patche[1] + Anshuman/Suzuki patches[3]" and below diff works for any CPU range and sparse list as well.

I will send the V2 ASAP.

[root@B0-213 perf]# git diff
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c
index 94e2d02009eb..bfed0bb8be98 100644
--- a/tools/perf/util/cs-etm.c
+++ b/tools/perf/util/cs-etm.c
@@ -274,6 +274,22 @@ static int cs_etm__metadata_set_trace_id(u8 trace_chan_id, u64 *cpu_metadata)
(typeof(_mask))(((_reg) & (_mask)) >> __bf_shf(_mask)); \
})

+
+static u64 *get_cpu_data(struct cs_etm_auxtrace *etm, int cpu)
+{
+ int i;
+ u64 *metadata = NULL;
+
+ for (i = 0; i < etm->num_cpu; i++) {
+ if (etm->metadata[i][CS_ETM_CPU] == (u64)cpu) {
+ metadata = etm->metadata[i];
+ break;
+ }
+ }
+
+ return metadata;
+}
+
/*
* Handle the PERF_RECORD_AUX_OUTPUT_HW_ID event.
*
@@ -344,7 +360,7 @@ static int cs_etm__process_aux_output_hw_id(struct perf_session *session,
}

/* not one we've seen before - lets map it */
- cpu_data = etm->metadata[cpu];
+ cpu_data = get_cpu_data(etm, cpu);
err = cs_etm__map_trace_id(trace_chan_id, cpu_data);
if (err)
return err;


I tried below commands and works well with this diff.

timeout 8s ./perf record -e cs_etm// -C 0 dd if=/dev/zero of=/dev/null
timeout 8s ./perf record -e cs_etm// -C 1 dd if=/dev/zero of=/dev/null
timeout 8s ./perf record -e cs_etm// -C 1,3,5 dd if=/dev/zero of=/dev/null
timeout 8s ./perf record -e cs_etm// -C 10,3,12 dd if=/dev/zero of=/dev/null
timeout 8s ./perf record -e cs_etm//u -C 0-10 dd if=/dev/zero of=/dev/null
timeout 8s ./perf record -e cs_etm//u -C 10-20 dd if=/dev/zero of=/dev/nu

However I could not be able to apply to next branch of coresight repo
git://git.kernel.org/pub/scm/linux/kernel/git/coresight/linux.git

Then I looked at the code, it does not have Mike's dynamic ID patches[1] which are merged to perf/next.
It seems the code is not the latest or it is next for coresight driver only.

I think either patch should be rebased to most-recent kernel RC1/RCx(which I did) or perf-tools-next of [2] for perf tool.

[1] https://www.spinics.net/lists/linux-perf-users/msg27452.html
[2] git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git
[3] https://www.spinics.net/lists/arm-kernel/msg1059012.html


the code would be better if it's worded like "i < recorded_cpus" rather
than "i < cpu" to make it clear that i isn't actually the CPU ID it's
just an index.

Yes, makes sense to call it "recorded_cpus".


Also a new test for this scenario would probably be a good idea.

Thanks
James

Thanks,
Ganapat

Thanks,
Ganapat

Thanks,
Ganapat