perf test: Skip sigtrap test on old kernels
Namhyung Kim
maintainerPATCH 1/1 perf test: Skip sigtrap test on old kernels
If it runs on an old kernel, perf_event_open would fail because of the new fields sigtrap and sig_data. Just skip the test if it failed.
--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/tests/sigtrap.c +1 -0
@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ; goto out_restore_sigaction; }
Signature · 1 line
2.37.2.789.g6183377224-goog
Quoted 15 lines · Namhyung Kim
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0
tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;
Wouldn't we be interested if perf_event_open() fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.
That's my naive thinking at least - what do other perf tests usually do in this case?
Thanks,
-- Marco
Arnaldo Carvalho de Melo
maintainerEm Sat, Sep 03, 2022 at 08:52:01AM +0200, Marco Elver escreveu:
Quoted 25 lines · Namhyung Kim
Quoted 15 lines
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
Yeah, I was going to try and check if this is the only way that, with the given arguments, perf_event_open would fail, but its better to at least check errno against -EINVAL or something?
- Arnaldo
Namhyung Kim
maintainerQuoted 27 lines · Arnaldo Carvalho de Melo
Em Sat, Sep 03, 2022 at 08:52:01AM +0200, Marco Elver escreveu:
Quoted 25 lines · Namhyung Kim
Quoted 15 lines
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.
Quoted 6 lines
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
Yeah, I was going to try and check if this is the only way that, with the given arguments,
perf_event_openwould fail, but its better to at least check errno against-EINVALor something?
EINVAL would be too generic and the kernel returns it in many places. I really wish we could have a better error reporting mechanism.
Maybe we could retry perf_event_open with sigtrap and sig_data cleared. If it succeeded, then we can skip the test. If it still failed, then report the error. But it still couldn't find a bug in the sigtrap code. What do you think?
Thanks,
Namhyung
Quoted 36 lines · Namhyung Kim
Quoted 27 lines · Arnaldo Carvalho de Melo
Em Sat, Sep 03, 2022 at 08:52:01AM +0200, Marco Elver escreveu:
Quoted 25 lines · Namhyung Kim
Quoted 15 lines
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.Quoted 6 lines
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
Yeah, I was going to try and check if this is the only way that, with the given arguments,
perf_event_openwould fail, but its better to at least check errno against-EINVALor something?
Quoted 7 lines
EINVAL would be too generic and the kernel returns it in many places. I really wish we could have a better error reporting mechanism.
Maybe we could retry
perf_event_openwith sigtrap andsig_datacleared. If it succeeded, then we can skip the test. If it still failed, then report the error. But it still couldn't find a bug in the sigtrap code. What do you think?
Yes, that's what I meant, that it could point out an issue with sigtrap perf_event_open().
If there's no clear way to determine if it's just not supported or a bug, it'd be better to leave it as-is.
Some other options:
1. Provide a way to disable certain tests, if it's known they will fail for otherwise benign reasons i.e. no support.
2. Provide a command line option to skip instead of fail tests where perf_event_open() returns some particular errnos. The default will be fail, but you can then choose to trust that failure of perf_event_open() means no support, and pass the option.
Quoted 45 lines · Marco Elver
Quoted 36 lines · Namhyung Kim
Quoted 27 lines · Arnaldo Carvalho de Melo
Em Sat, Sep 03, 2022 at 08:52:01AM +0200, Marco Elver escreveu:
Quoted 25 lines · Namhyung Kim
Quoted 15 lines
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.Quoted 6 lines
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
Yeah, I was going to try and check if this is the only way that, with the given arguments,
perf_event_openwould fail, but its better to at least check errno against-EINVALor something?Quoted 7 lines
EINVAL would be too generic and the kernel returns it in many places. I really wish we could have a better error reporting mechanism.
Maybe we could retry
perf_event_openwith sigtrap andsig_datacleared. If it succeeded, then we can skip the test. If it still failed, then report the error. But it still couldn't find a bug in the sigtrap code. What do you think?
Quoted 5 lines
Yes, that's what I meant, that it could point out an issue with sigtrap
perf_event_open().If there's no clear way to determine if it's just not supported or a bug, it'd be better to leave it as-is.
perf could go fancy and try to add a probe using a source file+line where older kernels would fail 8-)
Anyway, perf already does all sorts of kernel capability checks, perhaps this is one of can for sure detect it's something available :-/
One new way could be to look at BTF?
Quoted 10 lines
Some other options:
1. Provide a way to disable certain tests, if it's known they will fail for otherwise benign reasons i.e. no support.
2. Provide a command line option to skip instead of fail tests where
perf_event_open()returns some particular errnos. The default will be fail, but you can then choose to trust that failure ofperf_event_open()means no support, and pass the option.
Namhyung Kim
maintainerQuoted 52 lines · Arnaldo Carvalho de Melo
Quoted 48 lines · Marco Elver
Quoted 36 lines · Namhyung Kim
Quoted 27 lines · Arnaldo Carvalho de Melo
Em Sat, Sep 03, 2022 at 08:52:01AM +0200, Marco Elver escreveu:
Quoted 25 lines · Namhyung Kim
Quoted 15 lines
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.Quoted 6 lines
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
Yeah, I was going to try and check if this is the only way that, with the given arguments,
perf_event_openwould fail, but its better to at least check errno against-EINVALor something?Quoted 7 lines
EINVAL would be too generic and the kernel returns it in many places. I really wish we could have a better error reporting mechanism.
Maybe we could retry
perf_event_openwith sigtrap andsig_datacleared. If it succeeded, then we can skip the test. If it still failed, then report the error. But it still couldn't find a bug in the sigtrap code. What do you think?Yes, that's what I meant, that it could point out an issue with sigtrap
perf_event_open().
Quoted 8 lines
Quoted 2 lines
If there's no clear way to determine if it's just not supported or a bug, it'd be better to leave it as-is.
perf could go fancy and try to add a probe using a source file+line where older kernels would fail 8-)
Anyway, perf already does all sorts of kernel capability checks, perhaps this is one of can for sure detect it's something available :-/
One new way could be to look at BTF?
Yeah, we could check BTF if it had the attr.sigtrap field and skip if not. Let me see how I can do that. :)
Thanks,
Namhyung
Quoted 25 lines · Marco Elver
Quoted 15 lines · Namhyung Kim
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
I missed this discussion but I just submitted a patch with a similar issue [1]. To me, it doesn't make sense to have the tests pass on older kernels if this lowers the value of the tests by accepting possibly invalid values. If you want to test older kernels then just use older tests, but maybe there is some use case that I'm not aware of.
[1]: "[PATCH 0/1] perf test: Fix attr tests for PERF_FORMAT_LOST"
Quoted 3 lines
Thanks,
-- Marco
Namhyung Kim
maintainerHello,
Quoted 29 lines · James Clark
Quoted 25 lines · Marco Elver
Quoted 15 lines · Namhyung Kim
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.
Quoted 8 lines
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
I missed this discussion but I just submitted a patch with a similar issue [1]. To me, it doesn't make sense to have the tests pass on older kernels if this lowers the value of the tests by accepting possibly invalid values. If you want to test older kernels then just use older tests, but maybe there is some use case that I'm not aware of.
Thanks for your opinion. But my test environment is running the tests on random machines which may run some old kernel. I agree that it should not skip the real problems but I think we can find a good way to detect old, unsupported kernels reliably like using BTF.
Thanks,
Namhyung
Quoted 41 lines · Namhyung Kim
Hello,
Quoted 29 lines · James Clark
Quoted 25 lines · Marco Elver
Quoted 15 lines · Namhyung Kim
If it runs on an old kernel,
perf_event_openwould fail because of the new fields sigtrap andsig_data. Just skip the test if it failed.Cc Marco Elver <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/sigtrap.c | 1 + 1 file changed, 1 insertion(+)tools/perf/tests/sigtrap.c +0
-0tools/perf/tests/sigtrap.c static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m · line 143
Quoted 4 lines
tools/perf/tests/sigtrap.c +1
-0@@ -140,6 +140,7 @@ static int test__sigtrap(struct test_suite *test __maybe_unused, int subtest __m fd =sys_perf_event_open (&attr,0 ,-1 ,-1 ,perf_event_open_cloexec_flag ());if (fd <0 ) {pr_debug ("FAILED sys_perf_event_open(): %s\n" ,str_error_r (errno, sbuf,sizeof (sbuf))); + ret =TEST_SKIP ;Wouldn't we be interested if
perf_event_open()fails because it could actually be a bug? By skipping we'll be more likely to miss the fact there's a real problem.Quoted 8 lines
Quoted 2 lines
That's my naive thinking at least - what do other perf tests usually do in this case?
I missed this discussion but I just submitted a patch with a similar issue [1]. To me, it doesn't make sense to have the tests pass on older kernels if this lowers the value of the tests by accepting possibly invalid values. If you want to test older kernels then just use older tests, but maybe there is some use case that I'm not aware of.
Quoted 4 lines
Thanks for your opinion. But my test environment is running the tests on random machines which may run some old kernel. I agree that it should not skip the real problems but I think we can find a good way to detect old, unsupported kernels reliably like using BTF.
I'm currently adding a test for the new VG user register [1]. I'd like to test that it's available when the system has SVE but not when it hasn't so that the ABI is locked down. But that test will also depend on the kernel version, because it will never be available on older ones.
So for new tests should we always add two versions of the test, one for newer and one for older kernels? That seems like that has not been done so far in the tests. I'd still question the value of running the new tests on old kernels because they are mainly to validate that new code doesn't cause regressions. It's a lot of work to keep them backwards compatible for the case of running them on old kernels when old tests can be used instead.
If I am do add a version check would just looking at the uname be enough? Or how would the BTF version of that look?
Quoted 3 lines
Thanks,
Namhyung