Skip to contentSkip to conversation

Various 32-bit and test fixes

PATCH Applied Ian Rogers 2 participants 8 messages Sep 4, 2024 perf events

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

v2 cover Various 32-bit and test fixes

Running perf test as an i386 executable yielded a number of failures, some of which are addressed here.

The first 2 are straightforward use strtoull issues when parsing a 64-bit quantity in 32-bit land.

The 3rd patch just avoids a fail when perf probe isn't compiled in (in my case as LIBELF wasn't present).

The 4th and 5th cases fix the breakpoint length, on i386 so the sizeof(long) used matches the kernel's sizeof(long). On aarch64 the value is change to 4 instead of sizeof(long), ie 8, as future kernels may make 8 an invalid argument.

The final change addresses i386 watchpoint support not supporting 8-byte values.

v2: Fix a signed comparison build issue with gcc.

Ian Rogers (6):
perf pmus: Fix name comparisons on 32-bit systems perf time-utils: Fix 32-bit nsec parsing
perf test: Skip uprobe test if probe command isn't present perf parse-events: Add default_breakpoint_len helper perf parse-events: Vary default_breakpoint_len on i386 and arm64 perf test: Make watchpoint data 32-bits on i386

 tools/perf/tests/bp_account.c                 |  4 +++-
 tools/perf/tests/bp_signal.c                  |  3 ++-
 tools/perf/tests/bp_signal_overflow.c         |  3 ++-
 tools/perf/tests/parse-events.c               |  3 ++-
 .../shell/test_uprobe_from_different_cu.sh    |  7 ++++++
 tools/perf/tests/wp.c                         |  5 ++++
 tools/perf/util/parse-events.c                | 23 ++++++++++++++++++-
 tools/perf/util/parse-events.h                |  2 ++
 tools/perf/util/pmus.c                        |  6 ++---
 tools/perf/util/time-utils.c                  |  4 ++--
 10 files changed, 50 insertions(+), 10 deletions(-)
Signature · 1 line
2.46.0.469.g59c65b2a67-goog
author reviewer
#1 lore raw

PATCH v2 1/6 perf pmus: Fix name comparisons on 32-bit systems

The hex PMU suffix maybe 64-bit but the comparisons were "unsigned long" or 32-bit on 32-bit systems. This was causing the "PMU name comparison" test to fail in a 32-bit build.

---
 tools/perf/util/pmus.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
tools/perf/util/pmus.c +3 -3
index 3fcabfd8fca1..769b920d9250 100644
@@ -69,7 +69,7 @@ size_t pmu_name_len_no_suffix(const char *str)
 
 int pmu_name_cmp(const char *lhs_pmu_name, const char *rhs_pmu_name)
 {
-	unsigned long lhs_num = 0, rhs_num = 0;
+	unsigned long long lhs_num = 0, rhs_num = 0;
 	size_t lhs_pmu_name_len = pmu_name_len_no_suffix(lhs_pmu_name);
 	size_t rhs_pmu_name_len = pmu_name_len_no_suffix(rhs_pmu_name);
 	int ret = strncmp(lhs_pmu_name, rhs_pmu_name,
@@ -79,9 +79,9 @@ int pmu_name_cmp(const char *lhs_pmu_name, const char *rhs_pmu_name)
 		return ret;
 
 	if (lhs_pmu_name_len + 1 < strlen(lhs_pmu_name))
-		lhs_num = strtoul(&lhs_pmu_name[lhs_pmu_name_len + 1], NULL, 16);
+		lhs_num = strtoull(&lhs_pmu_name[lhs_pmu_name_len + 1], NULL, 16);
 	if (rhs_pmu_name_len + 1 < strlen(rhs_pmu_name))
-		rhs_num = strtoul(&rhs_pmu_name[rhs_pmu_name_len + 1], NULL, 16);
+		rhs_num = strtoull(&rhs_pmu_name[rhs_pmu_name_len + 1], NULL, 16);
 
 	return lhs_num < rhs_num ? -1 : (lhs_num > rhs_num ? 1 : 0);
 }
Signature · 1 line
2.46.0.469.g59c65b2a67-goog
author reviewer
#2 lore raw

PATCH v2 2/6 perf time-utils: Fix 32-bit nsec parsing

In reply to Ian Rogers

The "time utils" test fails in 32-bit builds:

```
...
parse_nsec_time("18446744073.709551615")
Failed. ptime 4294967295709551615 expected 18446744073709551615
...
```

Switch strtoul to strtoull as an unsigned long in 32-bit build isn't 64-bits.

Fixes c284d669a20d ("perf tools: Move parse_nsec_time to time-utils.c")
---
 tools/perf/util/time-utils.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/time-utils.c +2 -2
index 302443921681..1b91ccd4d523 100644
@@ -20,7 +20,7 @@ int parse_nsec_time(const char *str, u64 *ptime)
 	u64 time_sec, time_nsec;
 	char *end;
 
-	time_sec = strtoul(str, &end, 10);
+	time_sec = strtoull(str, &end, 10);
 	if (*end != '.' && *end != '\0')
 		return -1;
 
@@ -38,7 +38,7 @@ int parse_nsec_time(const char *str, u64 *ptime)
 		for (i = strlen(nsec_buf); i < 9; i++)
 			nsec_buf[i] = '0';
 
-		time_nsec = strtoul(nsec_buf, &end, 10);
+		time_nsec = strtoull(nsec_buf, &end, 10);
 		if (*end != '\0')
 			return -1;
 	} else
Signature · 1 line
2.46.0.469.g59c65b2a67-goog
author reviewer
#3 lore raw

PATCH v2 3/6 perf test: Skip uprobe test if probe command isn't present

In reply to Ian Rogers

The probe command is dependent on libelf. Skip the test if the required probe command isn't present.

---
 tools/perf/tests/shell/test_uprobe_from_different_cu.sh | 7 +++++++
 1 file changed, 7 insertions(+)
tools/perf/tests/shell/test_uprobe_from_different_cu.sh +7 -0
index 82bc774a078a..33387c329f92 100755
@@ -4,6 +4,13 @@
 
 set -e
 
+# Skip if there's no probe command.
+if ! perf | grep probe
+then
+        echo "Skip: probe command isn't present"
+        exit 2
+fi
+
 # skip if there's no gcc
 if ! [ -x "$(command -v gcc)" ]; then
         echo "failed: no gcc compiler"
Signature · 1 line
2.46.0.469.g59c65b2a67-goog
author reviewer
#4 lore raw

PATCH v2 4/6 perf parse-events: Add default_breakpoint_len helper

In reply to Ian Rogers

The default breakpoint length is "sizeof(long)" however this is incorrect on platforms like Aarch64 where sizeof(long) is 8 but the breakpoint length is 4. Add a helper function that can be used to determine the correct breakpoint length, in this change it just returns the existing default sizeof(long) value.

Use the helper in the bp_account test so that, when modifying the event from a watchpoint to a breakpoint, the breakpoint length is appropriate for the architecture and not just sizeof(long).

---
 tools/perf/tests/bp_account.c         | 4 +++-
 tools/perf/tests/bp_signal.c          | 3 ++-
 tools/perf/tests/bp_signal_overflow.c | 3 ++-
 tools/perf/tests/parse-events.c       | 3 ++-
 tools/perf/util/parse-events.c        | 7 ++++++-
 tools/perf/util/parse-events.h        | 2 ++
 6 files changed, 17 insertions(+), 5 deletions(-)
tools/perf/tests/bp_account.c +3 -1
index 6f921db33cf9..4cb7d486b5c1 100644
@@ -16,6 +16,7 @@
 #include "tests.h"
 #include "debug.h"
 #include "event.h"
+#include "parse-events.h"
 #include "../perf-sys.h"
 #include "cloexec.h"
 
@@ -50,7 +51,7 @@ static int __event(bool is_x, void *addr, struct perf_event_attr *attr)
 	attr->config = 0;
 	attr->bp_type = is_x ? HW_BREAKPOINT_X : HW_BREAKPOINT_W;
 	attr->bp_addr = (unsigned long) addr;
-	attr->bp_len = sizeof(long);
+	attr->bp_len = is_x ? default_breakpoint_len() : sizeof(long);
 
 	attr->sample_period = 1;
 	attr->sample_type = PERF_SAMPLE_IP;
@@ -92,6 +93,7 @@ static int bp_accounting(int wp_cnt, int share)
 	attr_mod = attr;
 	attr_mod.bp_type = HW_BREAKPOINT_X;
 	attr_mod.bp_addr = (unsigned long) test_function;
+	attr_mod.bp_len = default_breakpoint_len();
 
 	ret = ioctl(fd[0], PERF_EVENT_IOC_MODIFY_ATTRIBUTES, &attr_mod);
 	TEST_ASSERT_VAL("failed to modify wp\n", ret == 0);
tools/perf/tests/bp_signal.c +2 -1
index 1f2908f02389..3faeb5b6fe0b 100644
@@ -26,6 +26,7 @@
 #include "tests.h"
 #include "debug.h"
 #include "event.h"
+#include "parse-events.h"
 #include "perf-sys.h"
 #include "cloexec.h"
 
@@ -111,7 +112,7 @@ static int __event(bool is_x, void *addr, int sig)
 	pe.config = 0;
 	pe.bp_type = is_x ? HW_BREAKPOINT_X : HW_BREAKPOINT_W;
 	pe.bp_addr = (unsigned long) addr;
-	pe.bp_len = sizeof(long);
+	pe.bp_len = is_x ? default_breakpoint_len() : sizeof(long);
 
 	pe.sample_period = 1;
 	pe.sample_type = PERF_SAMPLE_IP;
tools/perf/tests/bp_signal_overflow.c +2 -1
index 4e897c2cf26b..ee560e156be6 100644
@@ -25,6 +25,7 @@
 #include "tests.h"
 #include "debug.h"
 #include "event.h"
+#include "parse-events.h"
 #include "../perf-sys.h"
 #include "cloexec.h"
 
@@ -88,7 +89,7 @@ static int test__bp_signal_overflow(struct test_suite *test __maybe_unused, int
 	pe.config = 0;
 	pe.bp_type = HW_BREAKPOINT_X;
 	pe.bp_addr = (unsigned long) test_function;
-	pe.bp_len = sizeof(long);
+	pe.bp_len = default_breakpoint_len();
 
 	pe.sample_period = THRESHOLD;
 	pe.sample_type = PERF_SAMPLE_IP;
tools/perf/tests/parse-events.c +2 -1
index edc2adcf1bae..639e65a9bf61 100644
@@ -5,6 +5,7 @@
 #include <api/fs/fs.h>
 #include "tests.h"
 #include "debug.h"
+#include "parse-events.h"
 #include "pmu.h"
 #include "pmus.h"
 #include <dirent.h>
@@ -262,7 +263,7 @@ static int test__checkevent_breakpoint_x(struct evlist *evlist)
 	TEST_ASSERT_VAL("wrong config", test_config(evsel, 0));
 	TEST_ASSERT_VAL("wrong bp_type",
 			HW_BREAKPOINT_X == evsel->core.attr.bp_type);
-	TEST_ASSERT_VAL("wrong bp_len", sizeof(long) == evsel->core.attr.bp_len);
+	TEST_ASSERT_VAL("wrong bp_len", default_breakpoint_len() == evsel->core.attr.bp_len);
 	return TEST_OK;
 }
 
tools/perf/util/parse-events.c +6 -1
index fab01ba54e34..792e5ed94870 100644
@@ -670,6 +670,11 @@ static int add_tracepoint_multi_sys(struct parse_events_state *parse_state,
 }
 #endif /* HAVE_LIBTRACEEVENT */
 
+size_t default_breakpoint_len(void)
+{
+	return sizeof(long);
+}
+
 static int
 parse_breakpoint_type(const char *type, struct perf_event_attr *attr)
 {
@@ -728,7 +733,7 @@ int parse_events_add_breakpoint(struct parse_events_state *parse_state,
 	/* Provide some defaults if len is not specified */
 	if (!len) {
 		if (attr.bp_type == HW_BREAKPOINT_X)
-			len = sizeof(long);
+			len = default_breakpoint_len();
 		else
 			len = HW_BREAKPOINT_LEN_4;
 	}
tools/perf/util/parse-events.h +2 -0
index b735cd9e0acf..68bfea9ffa70 100644
@@ -286,4 +286,6 @@ static inline bool is_sdt_event(char *str __maybe_unused)
 }
 #endif /* HAVE_LIBELF_SUPPORT */
 
+size_t default_breakpoint_len(void);
+
 #endif /* __PERF_PARSE_EVENTS_H */
Signature · 1 line
2.46.0.469.g59c65b2a67-goog
author reviewer
#5 lore raw

PATCH v2 5/6 perf parse-events: Vary default_breakpoint_len on i386 and arm64

In reply to Ian Rogers

On arm64 the breakpoint length should be 4-bytes but 8-bytes is tolerated as perf passes that as sizeof(long). Just pass the correct value.

On i386 the sizeof(long) check in the kernel needs to match the kernel's long size. Check using an environment (uname checks) whether 4 or 8 bytes needs to be passed. Cache the value in a static.

---
 tools/perf/util/parse-events.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
tools/perf/util/parse-events.c +16 -0
index 792e5ed94870..fa71f7f1d8d3 100644
@@ -8,6 +8,7 @@
 #include <sys/ioctl.h>
 #include <sys/param.h>
 #include "term.h"
+#include "env.h"
 #include "evlist.h"
 #include "evsel.h"
 #include <subcmd/parse-options.h>
@@ -672,7 +673,22 @@ static int add_tracepoint_multi_sys(struct parse_events_state *parse_state,
 
 size_t default_breakpoint_len(void)
 {
+#if defined(__i386__)
+	static int len;
+
+	if (len == 0) {
+		struct perf_env env = {};
+
+		perf_env__init(&env);
+		len = perf_env__kernel_is_64_bit(&env) ? sizeof(u64) : sizeof(long);
+		perf_env__exit(&env);
+	}
+	return len;
+#elif defined(__aarch64__)
+	return 4;
+#else
 	return sizeof(long);
+#endif
 }
 
 static int
Signature · 1 line
2.46.0.469.g59c65b2a67-goog
author reviewer
#6 lore raw

PATCH v2 6/6 perf test: Make watchpoint data 32-bits on i386

In reply to Ian Rogers

i386 only supports watchpoints up to size 4, 8 bytes causes extra counts and test failures.

---
 tools/perf/tests/wp.c | 5 +++++
 1 file changed, 5 insertions(+)
tools/perf/tests/wp.c +5 -0
index cc8719609b19..6c178985e37f 100644
@@ -20,7 +20,12 @@ do {                                            \
 	TEST_ASSERT_VAL(text, count == val);    \
 } while (0)
 
+#ifdef __i386__
+/* Only breakpoint length less-than 8 has hardware support on i386. */
+volatile u32 data1;
+#else
 volatile u64 data1;
+#endif
 volatile u8 data2[3];
 
 #ifndef __s390x__
Signature · 1 line
2.46.0.469.g59c65b2a67-goog
maintainer Applied
#7 lore raw
In reply to Ian Rogers
Quoted 4 lines · Ian Rogers

The 4th and 5th cases fix the breakpoint length, on i386 so the sizeof(long) used matches the kernel's sizeof(long). On aarch64 the value is change to 4 instead of sizeof(long), ie 8, as future kernels may make 8 an invalid argument.

Applied these, the others were already in tmp.perf-tools-next, thanks.

- Arnaldo