Skip to contentSkip to conversation

perf header: Fix missing PMU caps recording in write_pmu_caps

PATCH In review Wei Li 2 participants 4 messages Aug 15, 2023 – Aug 18, 2023 perf events

lore
Share & tools
lore

Wei Li

author
#0 lore raw

PATCH 1/1 perf header: Fix missing PMU caps recording in write_pmu_caps

An error occurred when executing "perf script syscall-counts" on arm64:
Press control+C to stop and show the summary
0x6a7c [0x14]: failed to process type: 80

After debugging i found that it failed to process HEADER_PMU_CAPS, looking back on write_pmu_caps(), it get all pmus with caps except "cpu" first, then handle the hybrid core pmus, and the uncore pmus finally.

But on arm64 platforms, the core pmu is named "armv8_pmuv3" etc, not "cpu". So it's counted but no caps recorded here, this is broken since commit 178ddf3bad98 ("perf header: Avoid hybrid PMU list in write_pmu_caps"), as the logic for determining hybrid is changed from name matching to "core" attribute judging.

To fix this, change the logic of recording pmu caps to be consistent with the preceding counting, also make the code clear. The compatibility with old perf tool is ignored since it has been already broken.

Fixes 178ddf3bad98 ("perf header: Avoid hybrid PMU list in write_pmu_caps")
---
 tools/perf/util/header.c | 16 ++--------------
 1 file changed, 2 insertions(+), 14 deletions(-)
tools/perf/util/header.c +2 -14
index 52fbf526fe74..7d2bd1fb8ef2 100644
@@ -1618,22 +1618,10 @@ static int write_pmu_caps(struct feat_fd *ff,
 	if (!nr_pmu)
 		return 0;
 
-	/*
-	 * Write hybrid pmu caps first to maintain compatibility with
-	 * older perf tool.
-	 */
-	if (perf_pmus__num_core_pmus() > 1) {
-		pmu = NULL;
-		while ((pmu = perf_pmus__scan_core(pmu))) {
-			ret = __write_pmu_caps(ff, pmu, true);
-			if (ret < 0)
-				return ret;
-		}
-	}
-
 	pmu = NULL;
 	while ((pmu = perf_pmus__scan(pmu))) {
-		if (pmu->is_core || !pmu->nr_caps)
+		if (!pmu->name || !strcmp(pmu->name, "cpu") ||
+		    !pmu->nr_caps)
 			continue;
 
 		ret = __write_pmu_caps(ff, pmu, true);
Signature · 1 line
2.25.1
reviewer
#1 lore raw
Quoted 44 lines · Wei Li

An error occurred when executing "perf script syscall-counts" on arm64:
Press control+C to stop and show the summary
0x6a7c [0x14]: failed to process type: 80

After debugging i found that it failed to process HEADER_PMU_CAPS, looking back on write_pmu_caps(), it get all pmus with caps except "cpu" first, then handle the hybrid core pmus, and the uncore pmus finally.

But on arm64 platforms, the core pmu is named "armv8_pmuv3" etc, not "cpu". So it's counted but no caps recorded here, this is broken since commit 178ddf3bad98 ("perf header: Avoid hybrid PMU list in write_pmu_caps"), as the logic for determining hybrid is changed from name matching to "core" attribute judging.

To fix this, change the logic of recording pmu caps to be consistent with the preceding counting, also make the code clear. The compatibility with old perf tool is ignored since it has been already broken.

Fixes 178ddf3bad98 ("perf header: Avoid hybrid PMU list in write_pmu_caps")
---
 tools/perf/util/header.c | 16 ++--------------
 1 file changed, 2 insertions(+), 14 deletions(-)
tools/perf/util/header.c +0 -11
index 52fbf526fe74..7d2bd1fb8ef2 100644
@@ -1618,22 +1618,10 @@ static int write_pmu_caps(struct feat_fd *ff,
        if (!nr_pmu)
                return 0;
 
-       /*
-        * Write hybrid pmu caps first to maintain compatibility with
-        * older perf tool.
-        */
-       if (perf_pmus__num_core_pmus() > 1) {
-               pmu = NULL;
-               while ((pmu = perf_pmus__scan_core(pmu))) {
-                       ret = __write_pmu_caps(ff, pmu, true);
-                       if (ret < 0)
-                               return ret;
-               }

tools/perf/util/header.c static int write_pmu_caps(struct feat_fd *ff, · line 1627

Quoted 10 lines
tools/perf/util/header.c +2 -3
-       }
-
        pmu = NULL;
        while ((pmu = perf_pmus__scan(pmu))) {
-               if (pmu->is_core || !pmu->nr_caps)
+               if (!pmu->name || !strcmp(pmu->name, "cpu") ||
+                   !pmu->nr_caps)
                        continue;
 
                ret = __write_pmu_caps(ff, pmu, true);

Sorry for the breakage. I think all of the branches are interfering here, I think a simpler solution is:

```
--- a/tools/perf/util/header.c
+++ b/tools/perf/util/header.c
@@ -1605,8 +1605,7 @@ static int write_pmu_caps(struct feat_fd *ff,
       int ret;

       while ((pmu = perf_pmus__scan(pmu))) {
-               if (!pmu->name || !strcmp(pmu->name, "cpu") ||
-                   perf_pmu__caps_parse(pmu) <= 0)
+               if (perf_pmu__caps_parse(pmu) <= 0)
                       continue;
               nr_pmu++;
       }
@@ -1619,23 +1618,13 @@ static int write_pmu_caps(struct feat_fd *ff,
               return 0;

       /*
-        * Write hybrid pmu caps first to maintain compatibility with
-        * older perf tool.
+        * Note older perf tools assume core PMUs come first, this is a property
+        * of perf_pmus__scan.
        */
-       if (perf_pmus__num_core_pmus() > 1) {
-               pmu = NULL;
-               while ((pmu = perf_pmus__scan_core(pmu))) {
-                       ret = __write_pmu_caps(ff, pmu, true);
-                       if (ret < 0)
-                               return ret;
-               }
-       }
-
       pmu = NULL;
       while ((pmu = perf_pmus__scan(pmu))) {
-               if (pmu->is_core || !pmu->nr_caps)
+               if (perf_pmu__caps_parse(pmu) <= 0)
                       continue;
-
               ret = __write_pmu_caps(ff, pmu, true);
               if (ret < 0)
                       return ret;
```

Could you try it out?

Thanks,
Ian

Quoted 3 lines

--
2.25.1

author
#2 lore raw

Hi Ian:

On 2023/8/17 11:31, Ian Rogers wrote:
(SNIP)

Quoted 76 lines
Quoted 20 lines
tools/perf/util/header.c +0 -11
index 52fbf526fe74..7d2bd1fb8ef2 100644
@@ -1618,22 +1618,10 @@ static int write_pmu_caps(struct feat_fd *ff,
        if (!nr_pmu)
                return 0;
 
-       /*
-        * Write hybrid pmu caps first to maintain compatibility with
-        * older perf tool.
-        */
-       if (perf_pmus__num_core_pmus() > 1) {
-               pmu = NULL;
-               while ((pmu = perf_pmus__scan_core(pmu))) {
-                       ret = __write_pmu_caps(ff, pmu, true);
-                       if (ret < 0)
-                               return ret;
-               }

tools/perf/util/header.c static int write_pmu_caps(struct feat_fd *ff, · line 1627

Quoted 10 lines
tools/perf/util/header.c +2 -3
-       }
-
        pmu = NULL;
        while ((pmu = perf_pmus__scan(pmu))) {
-               if (pmu->is_core || !pmu->nr_caps)
+               if (!pmu->name || !strcmp(pmu->name, "cpu") ||
+                   !pmu->nr_caps)
                        continue;
 
                ret = __write_pmu_caps(ff, pmu, true);

Sorry for the breakage. I think all of the branches are interfering here, I think a simpler solution is:

```

tools/perf/util/header.c +4 -15
@@ -1605,8 +1605,7 @@ static int write_pmu_caps(struct feat_fd *ff,
       int ret;
 
       while ((pmu = perf_pmus__scan(pmu))) {
-               if (!pmu->name || !strcmp(pmu->name, "cpu") ||
-                   perf_pmu__caps_parse(pmu) <= 0)
+               if (perf_pmu__caps_parse(pmu) <= 0)
                       continue;
               nr_pmu++;
       }
@@ -1619,23 +1618,13 @@ static int write_pmu_caps(struct feat_fd *ff,
               return 0;
 
       /*
-        * Write hybrid pmu caps first to maintain compatibility with
-        * older perf tool.
+        * Note older perf tools assume core PMUs come first, this is a property
+        * of perf_pmus__scan.
        */
-       if (perf_pmus__num_core_pmus() > 1) {
-               pmu = NULL;
-               while ((pmu = perf_pmus__scan_core(pmu))) {
-                       ret = __write_pmu_caps(ff, pmu, true);
-                       if (ret < 0)
-                               return ret;
-               }
-       }
-
       pmu = NULL;
       while ((pmu = perf_pmus__scan(pmu))) {
-               if (pmu->is_core || !pmu->nr_caps)
+               if (perf_pmu__caps_parse(pmu) <= 0)
                       continue;
-
               ret = __write_pmu_caps(ff, pmu, true);
               if (ret < 0)
                       return ret;

```

Quoted 2 lines

Could you try it out?

I have tested your patch, it does make "perf script" work, but your modification records all the pmus with caps, including "cpu" on x86. This does not match the spec in tools/perf/Documentation/perf.data-file-format.txt:

HEADER_PMU_CAPS = 31,

List of pmu capabilities (except cpu pmu which is already covered by HEADER_CPU_PMU_CAPS). Note that hybrid cpu pmu capabilities are also stored here.

---
The header info of perf.data:

# bpf_prog_info 5809: bpf_prog_6deef7357e7b4530 addr 0xffffffffc11a10e3 size 61
# cpu pmu capabilities: branches=32, max_precise=3, pmu_name=skylake
# cpu pmu capabilities: branches=32, max_precise=3, pmu_name=skylake
# intel_pt pmu capabilities: topa_multiple_entries=1, psb_cyc=1, single_range_output=1, mtc_periods=249, ip_filtering=1, output_subsys=0, cr3_filtering=1, psb_periods=3f, cycle_thresholds=3fff, power_event_t>
# missing features: TRACING_DATA BRANCH_STACK GROUP_DESC AUXTRACE STAT CLOCKID DIR_FORMAT COMPRESSED CLOCK_DATA HYBRID_TOPOLOGY

Thanks,
Wei

reviewer
#3 lore raw
Quoted 104 lines · liwei (GF)

Hi Ian:

On 2023/8/17 11:31, Ian Rogers wrote:
(SNIP)

Quoted 76 lines
Quoted 20 lines
tools/perf/util/header.c +0 -11
index 52fbf526fe74..7d2bd1fb8ef2 100644
@@ -1618,22 +1618,10 @@ static int write_pmu_caps(struct feat_fd *ff,
        if (!nr_pmu)
                return 0;
 
-       /*
-        * Write hybrid pmu caps first to maintain compatibility with
-        * older perf tool.
-        */
-       if (perf_pmus__num_core_pmus() > 1) {
-               pmu = NULL;
-               while ((pmu = perf_pmus__scan_core(pmu))) {
-                       ret = __write_pmu_caps(ff, pmu, true);
-                       if (ret < 0)
-                               return ret;
-               }

tools/perf/util/header.c static int write_pmu_caps(struct feat_fd *ff, · line 1627

Quoted 10 lines
tools/perf/util/header.c +2 -3
-       }
-
        pmu = NULL;
        while ((pmu = perf_pmus__scan(pmu))) {
-               if (pmu->is_core || !pmu->nr_caps)
+               if (!pmu->name || !strcmp(pmu->name, "cpu") ||
+                   !pmu->nr_caps)
                        continue;
 
                ret = __write_pmu_caps(ff, pmu, true);

Sorry for the breakage. I think all of the branches are interfering here, I think a simpler solution is:

```

tools/perf/util/header.c +4 -15
@@ -1605,8 +1605,7 @@ static int write_pmu_caps(struct feat_fd *ff,
       int ret;
 
       while ((pmu = perf_pmus__scan(pmu))) {
-               if (!pmu->name || !strcmp(pmu->name, "cpu") ||
-                   perf_pmu__caps_parse(pmu) <= 0)
+               if (perf_pmu__caps_parse(pmu) <= 0)
                       continue;
               nr_pmu++;
       }
@@ -1619,23 +1618,13 @@ static int write_pmu_caps(struct feat_fd *ff,
               return 0;
 
       /*
-        * Write hybrid pmu caps first to maintain compatibility with
-        * older perf tool.
+        * Note older perf tools assume core PMUs come first, this is a property
+        * of perf_pmus__scan.
        */
-       if (perf_pmus__num_core_pmus() > 1) {
-               pmu = NULL;
-               while ((pmu = perf_pmus__scan_core(pmu))) {
-                       ret = __write_pmu_caps(ff, pmu, true);
-                       if (ret < 0)
-                               return ret;
-               }
-       }
-
       pmu = NULL;
       while ((pmu = perf_pmus__scan(pmu))) {
-               if (pmu->is_core || !pmu->nr_caps)
+               if (perf_pmu__caps_parse(pmu) <= 0)
                       continue;
-
               ret = __write_pmu_caps(ff, pmu, true);
               if (ret < 0)
                       return ret;

```

Quoted 2 lines

Could you try it out?

I have tested your patch, it does make "perf script" work, but your modification records all the pmus with caps, including "cpu" on x86. This does not match the spec in tools/perf/Documentation/perf.data-file-format.txt:

HEADER_PMU_CAPS = 31,

List of pmu capabilities (except cpu pmu which is already covered by HEADER_CPU_PMU_CAPS). Note that hybrid cpu pmu capabilities are also stored here.

---
The header info of perf.data:

# bpf_prog_info 5809: bpf_prog_6deef7357e7b4530 addr 0xffffffffc11a10e3 size 61
# cpu pmu capabilities: branches=32, max_precise=3, pmu_name=skylake
# cpu pmu capabilities: branches=32, max_precise=3, pmu_name=skylake
# intel_pt pmu capabilities: topa_multiple_entries=1, psb_cyc=1, single_range_output=1, mtc_periods=249, ip_filtering=1, output_subsys=0, cr3_filtering=1, psb_periods=3f, cycle_thresholds=3fff, power_event_t>
# missing features: TRACING_DATA BRANCH_STACK GROUP_DESC AUXTRACE STAT CLOCKID DIR_FORMAT COMPRESSED CLOCK_DATA HYBRID_TOPOLOGY
Quoted 2 lines

Thanks,
Wei

Thanks Wei! I'll update the patch I wrote and skip "cpu" with the comment that it is already covered by HEADER_CPU_PMU_CAPS. It feels that caps should really cover hybrid, but I was trying to refactor and cleanup not change the perf.data format.

Thanks,
Ian