libperf-sampling(7) =================== NAME ---- libperf-sampling - sampling interface DESCRIPTION ----------- The sampling interface provides API to measure and get count for specific perf events. The following test tries to explain count on `sampling.c` example. It is by no means complete guide to sampling, but shows libperf basic API for sampling. The `sampling.c` comes with libperf package and can be compiled and run like: [source,bash] -- $ gcc -o sampling sampling.c -lperf $ sudo ./sampling cpu 0, pid 0, tid 0, ip ffffffffad06c4e6, period 1 cpu 0, pid 4465, tid 4469, ip ffffffffad118748, period 18322959 cpu 0, pid 0, tid 0, ip ffffffffad115722, period 33544846 cpu 0, pid 4465, tid 4470, ip 7f84fe0cdad6, period 23687474 cpu 0, pid 0, tid 0, ip ffffffffad9e0349, period 34255790 cpu 0, pid 4465, tid 4469, ip ffffffffad136581, period 38664069 cpu 0, pid 0, tid 0, ip ffffffffad9e55e2, period 21922384 cpu 0, pid 4465, tid 4470, ip 7f84fe0ebebf, period 17655175 ... -- It requires root access, because it uses hardware cycles event. The `sampling.c` example profiles/samples all CPUs with hardware cycles, in a nutshell it: - creates events - adds them to the event list - opens and enables events through the event list - sleeps for 3 seconds - disables events - reads and displays recorded samples - destroys the event list The first thing you need to do before using libperf is to call init function: [source,c] -- 12 static int libperf_print(enum libperf_print_level level, 13 const char *fmt, va_list ap) 14 { 15 return vfprintf(stderr, fmt, ap); 16 } 23 int main(int argc, char **argv) 24 { ... 40 libperf_init(libperf_print); -- It will setup the library and sets function for debug output from library. The `libperf_print` callback will receive any message with its debug level, defined as: [source,c] -- enum libperf_print_level { LIBPERF_ERR, LIBPERF_WARN, LIBPERF_INFO, LIBPERF_DEBUG, LIBPERF_DEBUG2, LIBPERF_DEBUG3, }; -- Once the setup is complete we start by defining cycles event using the `struct perf_event_attr`: [source,c] -- 29 struct perf_event_attr attr = { 30 .type = PERF_TYPE_HARDWARE, 31 .config = PERF_COUNT_HW_CPU_CYCLES, 32 .disabled = 1, 33 .freq = 1, 34 .sample_freq = 10, 35 .sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_CPU|PERF_SAMPLE_PERIOD, 36 }; -- Next step is to prepare CPUs map. In this case we will monitor all the available CPUs: [source,c] -- 42 cpus = perf_cpu_map__new_online_cpus(); 43 if (!cpus) { 44 fprintf(stderr, "failed to create cpus\n"); 45 return -1; 46 } -- Now we create libperf's event list, which will serve as holder for the cycles event: [source,c] -- 48 evlist = perf_evlist__new(); 49 if (!evlist) { 50 fprintf(stderr, "failed to create evlist\n"); 51 goto out_cpus; 52 } -- We create libperf's event for the cycles attribute we defined earlier and add it to the list: [source,c] -- 54 evsel = perf_evsel__new(&attr); 55 if (!evsel) { 56 fprintf(stderr, "failed to create cycles\n"); 57 goto out_cpus; 58 } 59 60 perf_evlist__add(evlist, evsel); -- Configure event list with the cpus map and open event: [source,c] -- 62 perf_evlist__set_maps(evlist, cpus, NULL); 63 64 err = perf_evlist__open(evlist); 65 if (err) { 66 fprintf(stderr, "failed to open evlist\n"); 67 goto out_evlist; 68 } -- Once the events list is open, we can create memory maps AKA perf ring buffers: [source,c] -- 70 err = perf_evlist__mmap(evlist, 4); 71 if (err) { 72 fprintf(stderr, "failed to mmap evlist\n"); 73 goto out_evlist; 74 } -- The event is created as disabled (note the `disabled = 1` assignment above), so we need to enable the events list explicitly. From this moment the cycles event is sampling. We will sleep for 3 seconds while the ring buffers get data from all CPUs, then we disable the events list. [source,c] -- 76 perf_evlist__enable(evlist); 77 sleep(3); 78 perf_evlist__disable(evlist); -- Following code walks through the ring buffers and reads stored events/samples: [source,c] -- 80 perf_evlist__for_each_mmap(evlist, map, false) { 81 if (perf_mmap__read_init(map) < 0) 82 continue; 83 84 while ((event = perf_mmap__read_event(map)) != NULL) { /* process event */ 108 perf_mmap__consume(map); 109 } 110 perf_mmap__read_done(map); 111 } -- Each sample needs to get parsed: [source,c] -- 85 int cpu, pid, tid; 86 __u64 ip, period, *array; 87 union u64_swap u; 88 89 array = event->sample.array; 90 91 ip = *array; 92 array++; 93 94 u.val64 = *array; 95 pid = u.val32[0]; 96 tid = u.val32[1]; 97 array++; 98 99 u.val64 = *array; 100 cpu = u.val32[0]; 101 array++; 102 103 period = *array; 104 105 fprintf(stdout, "cpu %3d, pid %6d, tid %6d, ip %20llx, period %20llu\n", 106 cpu, pid, tid, ip, period); -- And finally cleanup. We close the whole events list (both events) and remove it together with the threads map: [source,c] -- 113 out_evlist: 114 perf_evlist__delete(evlist); 115 out_cpus: 116 perf_cpu_map__put(cpus); 117 return err; 118 } -- REPORTING BUGS -------------- Report bugs to . LICENSE ------- libperf is Free Software licensed under the GNU LGPL 2.1 RESOURCES --------- https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git SEE ALSO -------- libperf(3), libperf-counting(7)