Skip to contentSkip to conversation

perf pmu: Create x86 specific perf_pmu__valid_suffix

PATCH In review Jin Yao 2 participants 2 messages Jul 20, 2021 perf events

lore
Share & tools
lore

Jin Yao

author
#0 lore raw

PATCH 1/1 perf pmu: Create x86 specific perf_pmu__valid_suffix

The commit c47a5599eda3 ("perf tools: Fix pattern matching for same substring in different PMU type") breaks arm64 system because it assumes the first token must be followed by a '_', but it is possibly a numeric on arm64.

For example, perf_pmu__valid_suffix("hisi_sccl3_l3c7", "hisi_sccl") fails. "hisi_sccl3_l3c7" is pmu name and "hisi_sccl" is token. "hisi_sccl" is followed by a digit but not followed by a '_' ('3' in this example).

Since the PMU alias format on arm64 has difference than the format on x86. Create a x86 specific perf_pmu__valid_suffix. For other arch, the weak function always returns true to keep original behavior unchanged.

Fixes c47a5599eda3 ("perf tools: Fix pattern matching for same substring in different PMU type")
---
 tools/perf/arch/x86/util/pmu.c | 22 ++++++++++++++++++++++
 tools/perf/util/pmu.c          | 27 ++++++---------------------
 tools/perf/util/pmu.h          |  1 +
 3 files changed, 29 insertions(+), 21 deletions(-)
tools/perf/arch/x86/util/pmu.c +22 -0
index d48d608517fd..519da0811308 100644
@@ -3,6 +3,7 @@
 
 #include <linux/stddef.h>
 #include <linux/perf_event.h>
+#include <linux/ctype.h>
 
 #include "../../../util/intel-pt.h"
 #include "../../../util/intel-bts.h"
@@ -18,3 +19,24 @@ struct perf_event_attr *perf_pmu__get_default_config(struct perf_pmu *pmu __mayb
 #endif
 	return NULL;
 }
+
+bool perf_pmu__valid_suffix(char *pmu_name, char *tok)
+{
+	char *p;
+
+	if (strncmp(pmu_name, tok, strlen(tok)))
+		return false;
+
+	p = pmu_name + strlen(tok);
+	if (*p == 0)
+		return true;
+
+	if (*p != '_')
+		return false;
+
+	++p;
+	if (*p == 0 || !isdigit(*p))
+		return false;
+
+	return true;
+}
tools/perf/util/pmu.c +6 -21
index 44b90d638ad5..a671b16f2d3e 100644
@@ -742,27 +742,6 @@ struct pmu_events_map *__weak pmu_events_map__find(void)
 	return perf_pmu__find_map(NULL);
 }
 
-static bool perf_pmu__valid_suffix(char *pmu_name, char *tok)
-{
-	char *p;
-
-	if (strncmp(pmu_name, tok, strlen(tok)))
-		return false;
-
-	p = pmu_name + strlen(tok);
-	if (*p == 0)
-		return true;
-
-	if (*p != '_')
-		return false;
-
-	++p;
-	if (*p == 0 || !isdigit(*p))
-		return false;
-
-	return true;
-}
-
 bool pmu_uncore_alias_match(const char *pmu_name, const char *name)
 {
 	char *tmp = NULL, *tok, *str;
@@ -1906,3 +1885,9 @@ int perf_pmu__match(char *pattern, char *name, char *tok)
 
 	return 0;
 }
+
+bool __weak perf_pmu__valid_suffix(char *pmu_name __maybe_unused,
+				   char *tok __maybe_unused)
+{
+	return true;
+}
tools/perf/util/pmu.h +1 -0
index 926da483a141..901812987b79 100644
@@ -134,5 +134,6 @@ void perf_pmu__warn_invalid_config(struct perf_pmu *pmu, __u64 config,
 
 bool perf_pmu__has_hybrid(void);
 int perf_pmu__match(char *pattern, char *name, char *tok);
+bool perf_pmu__valid_suffix(char *pmu_name, char *tok);
 
 #endif /* __PMU_H */
Signature · 1 line
2.17.1
#1 lore raw
Quoted 4 lines · Jin Yao

The commit c47a5599eda3 ("perf tools: Fix pattern matching for same substring in different PMU type") breaks arm64 system because it assumes the first token must be followed by a '_', but it is possibly a numeric on arm64.

I wouldn't just say arm64. The core code should support matching multiple tokens, but it just so happens that only arm64 uses it.

Quoted 11 lines

For example, perf_pmu__valid_suffix("hisi_sccl3_l3c7", "hisi_sccl") fails. "hisi_sccl3_l3c7" is pmu name and "hisi_sccl" is token. "hisi_sccl" is followed by a digit but not followed by a '_' ('3' in this example).

Since the PMU alias format on arm64 has difference than the format on x86. Create a x86 specific perf_pmu__valid_suffix. For other arch, the weak function always returns true to keep original behavior unchanged.

Quoted 4 lines
Fixes c47a5599eda3 ("perf tools: Fix pattern matching for same substring in different PMU type")

---

I tend not to agree with this solution. Let's not make it x86 vs non-x86 problem. Indeed, if we continue to support custom PMU alias matching per-arch, then it becomes a maintenance burden.

So this change fixes my initial problem:

-----8<------

From bbc9c7bf65c9cea756f0716dd717e9cfc767ba39 Mon Sep 17 00:00:00 2001
From: John Garry <[email protected]>
Date: Tue, 20 Jul 2021 09:48:22 +0100
Subject: [PATCH] perf pmu: Only check valid suffix for final token

Since commit 730670b1d108 ("perf pmu: Support more complex PMU event aliasing"), matching multiple alias tokens has been supported.

However, commit c47a5599eda32 ("perf tools: Fix pattern matching for same substring in different PMU type"), ignored the possibility of multiple tokens, and assumed that we can only match on valid suffix and a single token.

Fix multiple token support by only checking valid suffix from final token.

tools/perf/util/pmu.c +7 -2
index 44b90d638ad5..6de2275b2ae2 100644
@@ -789,12 +789,17 @@ bool pmu_uncore_alias_match(const char *pmu_name, const char *name)
 	 *	    match "socket" in "socketX_pmunameY" and then "pmuname" in
 	 *	    "pmunameY".
 	 */
-	for (; tok; name += strlen(tok), tok = strtok_r(NULL, ",", &tmp)) {
+	for (; ; ) {
+		char *next_tok = strtok_r(NULL, ",", &tmp);
 		name = strstr(name, tok);
-		if (!name || !perf_pmu__valid_suffix((char *)name, tok)) {
+		if (!name || (!perf_pmu__valid_suffix((char *)name, tok) && !next_tok)) {
 			res = false;
 			goto out;
 		}
+		if (!next_tok)
+			break;
+		name += strlen(tok);
+		tok = next_tok;
 	}
 
 	res = true;
Signature · 10 lines
2.26.2

----->8------

Please check it.

I need to check the other place perf_pmu__valid_suffix() is referenced (perf_pmu__match()) to see if anything is broken there.

Thanks,
John