Skip to contentSkip to conversation

perf vendor events amd: add L3 cache events for Family 17h

PATCH In review Kim Phillips 3 participants 8 messages Sep 19, 2019 – Sep 23, 2019 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/5 perf vendor events amd: add L3 cache events for Family 17h

Allow users to symbolically specify L3 events for Family 17h processors using the existing AMD Uncore driver.

Source of events descriptions are from section 2.1.15.4.1 "L3 Cache PMC Events" of the latest Family 17h PPR, available here:

https://www.amd.com/system/files/TechDocs/55570-B1_PUB.zip

Only BriefDescriptions added, since they show with and without the -v and --details flags.

Tested with:

 # perf stat -e l3_request_g1.caching_l3_cache_accesses,amd_l3/event=0x01,umask=0x80/,l3_comb_clstr_state.request_miss,amd_l3/event=0x06,umask=0x01/ perf bench mem memcpy -s 4mb -l 100 -f default
...
         7,006,831      l3_request_g1.caching_l3_cache_accesses
         7,006,830      amd_l3/event=0x01,umask=0x80/
           366,530      l3_comb_clstr_state.request_miss
           366,568      amd_l3/event=0x06,umask=0x01/
Cc Janakarajan Natarajan <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexander Shishkin <[email protected]>
Cc Andi Kleen <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Martin Liska <[email protected]>
Cc Luke Mujica <[email protected]>
Cc Jin Yao <[email protected]>
Cc Kan Liang <[email protected]>
---
 .../pmu-events/arch/x86/amdfam17h/cache.json  | 42 +++++++++++++++++++
 tools/perf/pmu-events/jevents.c               |  1 +
 2 files changed, 43 insertions(+)
tools/perf/pmu-events/arch/x86/amdfam17h/cache.json +42 -0
index fad4af9142cb..6221a840fcea 100644
@@ -283,5 +283,47 @@
     "BriefDescription": "Total cycles spent with one or more fill requests in flight from L2.",
     "PublicDescription": "Total cycles spent with one or more fill requests in flight from L2.",
     "UMask": "0x1"
+  },
+  {
+    "EventName": "l3_request_g1.caching_l3_cache_accesses",
+    "EventCode": "0x01",
+    "BriefDescription": "Caching: L3 cache accesses",
+    "UMask": "0x80",
+    "Unit": "L3PMC"
+  },
+  {
+    "EventName": "l3_lookup_state.all_l3_req_typs",
+    "EventCode": "0x04",
+    "BriefDescription": "All L3 Request Types",
+    "UMask": "0xff",
+    "Unit": "L3PMC"
+  },
+  {
+    "EventName": "l3_comb_clstr_state.other_l3_miss_typs",
+    "EventCode": "0x06",
+    "BriefDescription": "Other L3 Miss Request Types",
+    "UMask": "0xfe",
+    "Unit": "L3PMC"
+  },
+  {
+    "EventName": "l3_comb_clstr_state.request_miss",
+    "EventCode": "0x06",
+    "BriefDescription": "L3 cache misses",
+    "UMask": "0x01",
+    "Unit": "L3PMC"
+  },
+  {
+    "EventName": "xi_sys_fill_latency",
+    "EventCode": "0x90",
+    "BriefDescription": "L3 Cache Miss Latency. Total cycles for all transactions divided by 16. Ignores SliceMask and ThreadMask.",
+    "UMask": "0x00",
+    "Unit": "L3PMC"
+  },
+  {
+    "EventName": "xi_ccx_sdp_req1.all_l3_miss_req_typs",
+    "EventCode": "0x9a",
+    "BriefDescription": "All L3 Miss Request Types. Ignores SliceMask and ThreadMask.",
+    "UMask": "0x3f",
+    "Unit": "L3PMC"
   }
 ]
tools/perf/pmu-events/jevents.c +1 -0
index d413761621b0..9e37287da924 100644
@@ -239,6 +239,7 @@ static struct map {
 	{ "hisi_sccl,ddrc", "hisi_sccl,ddrc" },
 	{ "hisi_sccl,hha", "hisi_sccl,hha" },
 	{ "hisi_sccl,l3c", "hisi_sccl,l3c" },
+	{ "L3PMC", "amd_l3" },
 	{}
 };
 
Signature · 1 line
2.23.0
#1 lore raw

PATCH 2/5 perf vendor events amd: remove redundant '['

Remove the redundant '['.

perf list output before:
ex_ret_brn
[[Retired Branch Instructions]

perf list output after:
ex_ret_brn
[Retired Branch Instructions]

Cc Janakarajan Natarajan <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexander Shishkin <[email protected]>
Cc Andi Kleen <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Martin Liska <[email protected]>
Cc Luke Mujica <[email protected]>
Cc Jin Yao <[email protected]>
Cc Kan Liang <[email protected]>
Fixes 98c07a8f74f8 ("perf vendor events amd: perf PMU events for AMD Family 17h")
---
 tools/perf/pmu-events/arch/x86/amdfam17h/core.json | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/pmu-events/arch/x86/amdfam17h/core.json +1 -1
index 7b285b0a7f35..1079544eeed5 100644
@@ -13,7 +13,7 @@
   {
     "EventName": "ex_ret_brn",
     "EventCode": "0xc2",
-    "BriefDescription": "[Retired Branch Instructions.",
+    "BriefDescription": "Retired Branch Instructions.",
     "PublicDescription": "The number of branch instructions retired. This includes all types of architectural control flow changes, including exceptions and interrupts."
   },
   {
Signature · 1 line
2.23.0
#2 lore raw

PATCH 3/5 perf vendor events: minor fixes to the README

In reply to Kim Phillips

Some grammatical fixes, and updates to some path references that have since changed.

Cc Janakarajan Natarajan <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexander Shishkin <[email protected]>
Cc Andi Kleen <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Martin Liska <[email protected]>
Cc Luke Mujica <[email protected]>
Cc Jin Yao <[email protected]>
Cc Kan Liang <[email protected]>
---
 tools/perf/pmu-events/README | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)
tools/perf/pmu-events/README +11 -11
index e62b09b6a844..de7efa2cebd1 100644
@@ -30,9 +30,9 @@ the topic. Eg: "Floating-point.json".
 All the topic JSON files for a CPU model/family should be in a separate
 sub directory. Thus for the Silvermont X86 CPU:
 
-	$ ls tools/perf/pmu-events/arch/x86/Silvermont_core
-	Cache.json 	Memory.json 	Virtual-Memory.json
-	Frontend.json 	Pipeline.json
+	$ ls tools/perf/pmu-events/arch/x86/silvermont
+	cache.json     memory.json    virtual-memory.json
+	frontend.json  pipeline.json
 
 The JSONs folder for a CPU model/family may be placed in the root arch
 folder, or may be placed in a vendor sub-folder under the arch folder
@@ -94,7 +94,7 @@ users to specify events by their name:
 
 where 'pm_1plus_ppc_cmpl' is a Power8 PMU event.
 
-However some errors in processing may cause the perf build to fail.
+However some errors in processing may cause the alias build to fail.
 
 Mapfile format
 ===============
@@ -119,7 +119,7 @@ where:
 
 	Header line
 		The header line is the first line in the file, which is
-		always _IGNORED_. It can empty.
+		always _IGNORED_. It can be empty.
 
 	CPUID:
 		CPUID is an arch-specific char string, that can be used
@@ -138,15 +138,15 @@ where:
 		files, relative to the directory containing the mapfile.csv
 
 	Type:
-		indicates whether the events or "core" or "uncore" events.
+		indicates whether the events are "core" or "uncore" events.
 
 
 	Eg:
 
-	$ grep Silvermont tools/perf/pmu-events/arch/x86/mapfile.csv
-	GenuineIntel-6-37,V13,Silvermont_core,core
-	GenuineIntel-6-4D,V13,Silvermont_core,core
-	GenuineIntel-6-4C,V13,Silvermont_core,core
+	$ grep silvermont tools/perf/pmu-events/arch/x86/mapfile.csv
+	GenuineIntel-6-37,v13,silvermont,core
+	GenuineIntel-6-4D,v13,silvermont,core
+	GenuineIntel-6-4C,v13,silvermont,core
 
 	i.e the three CPU models use the JSON files (i.e PMU events) listed
-	in the directory 'tools/perf/pmu-events/arch/x86/Silvermont_core'.
+	in the directory 'tools/perf/pmu-events/arch/x86/silvermont'.
Signature · 1 line
2.23.0
#4 lore raw

Re: [PATCH 3/5] perf vendor events: minor fixes to the README

Em Thu, Sep 19, 2019 at 03:09:28PM -0700, Andi Kleen escreveu:

Quoted 3 lines

For all the patches except the last

Thanks, applying all but the last till we get that further discussed.

Signature · 1 line
- Arnaldo
#5 lore raw

PATCH 4/5 perf list: allow plurals for metric, metricgroup

In reply to Kim Phillips

Enhance usability by allowing the same plurality used in the output title, for the command line parameter.

BEFORE, perf deceitfully acts as if there are no metrics to be had:

  $ perf list metrics

List of pre-defined events (to be used in -e):

Metric Groups:

$

But singular 'metric' shows a list of metrics:

$ perf list metric

List of pre-defined events (to be used in -e):

Metrics:

IPC
[Instructions Per Cycle (per logical thread)]
UPI
[Uops Per Instruction]

AFTER, when asking for 'metrics', we actually see the metrics get listed:

$ perf list metrics

List of pre-defined events (to be used in -e):

Metrics:

IPC
[Instructions Per Cycle (per logical thread)]
UPI
[Uops Per Instruction]

Cc Janakarajan Natarajan <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexander Shishkin <[email protected]>
Cc Andi Kleen <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Martin Liska <[email protected]>
Cc Luke Mujica <[email protected]>
Cc Jin Yao <[email protected]>
Cc Kan Liang <[email protected]>
Fixes 71b0acce78d1 ("perf list: Add metric groups to perf list")
---
 tools/perf/builtin-list.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/builtin-list.c +2 -2
index e290f6b348d8..08e62ae9d37e 100644
@@ -81,9 +81,9 @@ int cmd_list(int argc, const char **argv)
 						long_desc_flag, details_flag);
 		else if (strcmp(argv[i], "sdt") == 0)
 			print_sdt_events(NULL, NULL, raw_dump);
-		else if (strcmp(argv[i], "metric") == 0)
+		else if (strcmp(argv[i], "metric") == 0 || strcmp(argv[i], "metrics") == 0)
 			metricgroup__print(true, false, NULL, raw_dump, details_flag);
-		else if (strcmp(argv[i], "metricgroup") == 0)
+		else if (strcmp(argv[i], "metricgroup") == 0 || strcmp(argv[i], "metricgroups") == 0)
 			metricgroup__print(false, true, NULL, raw_dump, details_flag);
 		else if ((sep = strchr(argv[i], ':')) != NULL) {
 			int sep_idx;
Signature · 1 line
2.23.0
#6 lore raw

PATCH 5/5 perf list: specify metrics are to be used with -M

In reply to Kim Phillips

Output of 'perf list metrics' before:

  $ perf list metrics

List of pre-defined events (to be used in -e):

Metrics:

C2_Pkg_Residency
[C2 residency percent per package]
...

This misleads the uninitiated user to try:

  $ perf stat -e C2_Pkg_Residency

which gets:

event syntax error: 'C2_Pkg_Residency'
\___ parser error
Run 'perf list' for a list of valid events

Explicitly clarify that metrics and metricgroups are meant to be used with -M, and correct the grammar for the -e equivalent statement.

Output of 'perf list metrics' after:

  $ perf list metrics

List of pre-defined events (to be used with -e):

Metrics (to be used with -M):

C2_Pkg_Residency
[C2 residency percent per package]
...

Cc Janakarajan Natarajan <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Alexander Shishkin <[email protected]>
Cc Andi Kleen <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Martin Liska <[email protected]>
Cc Luke Mujica <[email protected]>
Cc Jin Yao <[email protected]>
Cc Kan Liang <[email protected]>
---
 tools/perf/builtin-list.c     | 2 +-
 tools/perf/util/metricgroup.c | 5 +++--
 2 files changed, 4 insertions(+), 3 deletions(-)
tools/perf/builtin-list.c +1 -1
index 08e62ae9d37e..be8e878aa556 100644
@@ -51,7 +51,7 @@ int cmd_list(int argc, const char **argv)
 	setup_pager();
 
 	if (!raw_dump && pager_in_use())
-		printf("\nList of pre-defined events (to be used in -e):\n\n");
+		printf("\nList of pre-defined events (to be used with -e):\n\n");
 
 	if (argc == 0) {
 		print_events(NULL, raw_dump, !desc_flag, long_desc_flag,
tools/perf/util/metricgroup.c +3 -2
index a7c0424dbda3..f116848be9f7 100644
@@ -377,9 +377,10 @@ void metricgroup__print(bool metrics, bool metricgroups, char *filter,
 	}
 
 	if (metricgroups && !raw)
-		printf("\nMetric Groups:\n\n");
+		printf("\nMetric Groups");
 	else if (metrics && !raw)
-		printf("\nMetrics:\n\n");
+		printf("\nMetrics");
+	printf(" (to be used with -M):\n\n");
 
 	for (node = rb_first_cached(&groups.entries); node; node = next) {
 		struct mep *me = container_of(node, struct mep, nd);
Signature · 1 line
2.23.0
#7 lore raw

Re: [PATCH 5/5] perf list: specify metrics are to be used with -M

Quoted 3 lines

This misleads the uninitiated user to try:

  $ perf stat -e C2_Pkg_Residency

Actually I guess we could just fix -e to support metrics too. Would probably not be that difficult.

-Andi