This discussion has been locked.

You can no longer post new replies to this discussion. If you have a question you can start a new discussion

Perf record

Hi,

I'd like to use perf record to look into some performance counters. I'm using a ThunderX2 system with Centos 8. When i use any of the uncore counters such as uncore_l3c_0/evict_request/, uncore_dmc_1/cnt_cycles/, uncore_dmc_0/cnt_cycles/ I get the same error "The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event ..." Do you guys have any suggestions? see below for more details. Thanks

# rpm -qa |grep perf
python3-perf-4.18.0-240.el8.aarch64
perftest-4.4-3.el8.aarch64
perf-4.18.0-240.el8.aarch64
qperf-0.4.11-1.el8.aarch64
# uname -r
4.18.0-240.el8.aarch64
# perf record -a -e uncore_dmc_0/cnt_cycles/ ls
Lowering default frequency rate to 1600.
Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (uncore_dmc_0/cnt_cycles/).
/bin/dmesg | grep -i perf may provide additional information.
# /bin/dmesg | grep -i perf  |tail -n 3
[194444.313119] perf: interrupt took too long (38453 > 38357), lowering kernel.perf_event_max_sample_rate to 5200
[194444.388648] perf: interrupt took too long (48338 > 48066), lowering kernel.perf_event_max_sample_rate to 4100
[194961.878397] perf: interrupt took too long (60611 > 60422), lowering kernel.perf_event_max_sample_rate to 3200

Top replies

james.clark over 3 years ago +1 verified

Hi, Unfortunately the error feedback for perf event open isn't that great but it looks like this is the -22 (EINVAL) return code that you are getting. Uncore events just don't support sampling (i.e....

+1 james.clark over 3 years ago

Hi,

Unfortunately the error feedback for perf event open isn't that great but it looks like this is the -22 (EINVAL) return code that you are getting. Uncore events just don't support sampling (i.e. perf record) and this is a limitation in the ThunderX2 PMU driver:

	/*
	 * SOC PMU counters are shared across all cores.
	 * Therefore, it does not support per-process mode.
	 * Also, it does not support event sampling mode.
	 */
	if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
		return -EINVAL;

If you want the counter values then perf stat works for me on TX2:

$ sudo ./perf stat -e uncore_dmc_0/cnt_cycles/ ls
...

 Performance counter stats for 'system wide':

        15,638,408      uncore_dmc_0/cnt_cycles/                                    

       0.001504234 seconds time elapsed

Thanks

James

0 YHuerta over 3 years ago in reply to james.clark

Thanks for the tip. I tested it and it worked.

# started on Wed May  5 13:11:18 2021


 Performance counter stats for 'system wide':

       38565488800      uncore_dmc_0/cnt_cycles/                                    
           1783808      uncore_dmc_0/data_transfers/                                   

       3.766226605 seconds time elapsed