Misleading hint to select CONFIG_PERF_EVENTS if driver sets PERF_PMU_CAP_NO_INTERRUPT

I have a question about user-space perf handling error numbers. The problem is that PMU interrupts are not supported in arc700 architecture and it is impossible to evaluate perf record command. In our perf implementation we set PERF_PMU_CAP_NO_INTERRUPT flag so core perf infrastructure knows we don't have interrupts.

Kernel sys_perf_event_open handler checks if PMU interrupts are supported and returns ENOTSUPP (524) error code.
I'd expect that perf implementation checks the return value of syscalls and gives the user understandable error message.
But now I see:
--------------------------------->8-----------------------------------

# perf record ls
The sys_perf_event_open() syscall returned with 524 (Unknown error 524)
for event (cycles:ppp).
/bin/dmesg may provide additional information.
No CONFIG_PERF_EVENTS=y kernel support configured?
--------------------------------->8-----------------------------------

As you can see the root cause of this error message is not obvious. CONFIG_PERF_EVENTS is selected but still there's a problem while existing suggestion barely makes any sense.
So probably there could be a way to determine if CONFIG_PERF_EVENTS was selected or not.

I am not sure about the correct way of solving this problem. Maybe I should add some checks of syscalls return values and give user a warning when not PMU interrupts are available.?
Any suggestions are appreciated.

Regards,
Lada Trimasova.

I wonder if there are any thoughts about this question. Please treat this as a polite reminder to take a look to this problem.

Regards,
Lada

On Fri, 2016-04-22@16:25 +0300, Lada Trimasova wrote:

I have a question about user-space perf handling error numbers. The problem is that PMU interrupts are not supported in arc700 architecture and it is impossible to evaluate perf record command. In our perf implementation we set PERF_PMU_CAP_NO_INTERRUPT flag so core perf infrastructure knows we don't have interrupts.

Kernel sys_perf_event_open handler checks if PMU interrupts are supported and returns ENOTSUPP (524) error code.
I'd expect that perf implementation checks the return value of syscalls and gives the user understandable error message.
But now I see:
--------------------------------->8-----------------------------------

# perf record ls
The sys_perf_event_open() syscall returned with 524 (Unknown error 524)
for event (cycles:ppp).
/bin/dmesg may provide additional information.
No CONFIG_PERF_EVENTS=y kernel support configured?
--------------------------------->8-----------------------------------

As you can see the root cause of this error message is not obvious. CONFIG_PERF_EVENTS is selected but still there's a problem while existing suggestion barely makes any sense.
So probably there could be a way to determine if CONFIG_PERF_EVENTS was selected or not.

I am not sure about the correct way of solving this problem. Maybe I should add some checks of syscalls return values and give user a warning when not PMU interrupts are available.
Any suggestions are appreciated.

Regards,
Lada Trimasova.em

I think what we have now is sufficient - but u seem to want a prettier failure output.

Anyhow, this print is coming from util/evsel.c: perf_evsel__open_strerror() At the very least you want another entry in switch case for ENOTSUPP and then check if event was sampling one (
evsel->attr.sample_period) - use that as a hint for saying sampling events not supported.

But this will print the same even if CONFIG_PERF_EVENTS=n. To really fix this you would want to change the error code returned by SYSCALL_DEFINE5(perf_event_open for PERF_PMU_CAP_NO_INTERRUPT to say -EOPNOTSUPP and use the sample_period to say this was for samplign events! However this is an ABI change and might not be acceptable as some existing scripts etc might break.

On Tue, 2016-04-26@12:42 +0000, Vineet Gupta wrote:

On Friday 22 April 2016 06:56 PM, Lada Trimasova wrote:
I think what we have now is sufficient - but u seem to want a prettier failure output.

Anyhow, this print is coming from util/evsel.c: perf_evsel__open_strerror() At the very least you want another entry in switch case for ENOTSUPP and then check if event was sampling one (
evsel->attr.sample_period) - use that as a hint for saying sampling events not supported.

ENOTSUPP is not visible for user programs. So it's impossible to add this entry

to mentioned switch.

I think that there is no good way to make error message more understandable

without breaking existing api.

But this will print the same even if CONFIG_PERF_EVENTS=n. To really fix this you would want to change the error code returned by SYSCALL_DEFINE5(perf_event_open for PERF_PMU_CAP_NO_INTERRUPT to say -EOPNOTSUPP and use the sample_period to say this was for samplign events! However this is an ABI change and might not be acceptable as some existing scripts etc might break.

/bin/dmesg may provide additional information.
No CONFIG_PERF_EVENTS=y kernel support configured? --------------------------------->8-----------------------------------

As you can see the root cause of this error message is not obvious. CONFIG_PERF_EVENTS is selected but still there's a problem while existing suggestion barely makes any sense.
So probably there could be a way to determine if CONFIG_PERF_EVENTS was selected or not.

I am not sure about the correct way of solving this problem. Maybe I should add some checks of syscalls return values and give user a warning when not PMU interrupts are available.
Any suggestions are appreciated.

Regards,
Lada Trimasova.
_______________________________________________
linux-snps-arc mailing list
linux-snps-arc at lists.infradead.org<mailto:linux-snps-arc at lists.infradead.org>
http://lists.infradead.org/mailman/listinfo/linux-snps-arc

Regards,

Lada.

I meant EOPNOTSUPP. Any errno value has to be visible to user programs otherwise there is no point defining it.

Guys, is the ABI change for returning -EOPNOTSUPP vs. -ENOTSUPP for is_sampling_event() && (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) acceptable ? We could use this in userspace to print pretty rather than half cooked error msg above !

-Vineet