libperf-sampling.txt (6665B)
1libperf-sampling(7) 2=================== 3 4NAME 5---- 6libperf-sampling - sampling interface 7 8 9DESCRIPTION 10----------- 11The sampling interface provides API to measure and get count for specific perf events. 12 13The following test tries to explain count on `sampling.c` example. 14 15It is by no means complete guide to sampling, but shows libperf basic API for sampling. 16 17The `sampling.c` comes with libperf package and can be compiled and run like: 18 19[source,bash] 20-- 21$ gcc -o sampling sampling.c -lperf 22$ sudo ./sampling 23cpu 0, pid 0, tid 0, ip ffffffffad06c4e6, period 1 24cpu 0, pid 4465, tid 4469, ip ffffffffad118748, period 18322959 25cpu 0, pid 0, tid 0, ip ffffffffad115722, period 33544846 26cpu 0, pid 4465, tid 4470, ip 7f84fe0cdad6, period 23687474 27cpu 0, pid 0, tid 0, ip ffffffffad9e0349, period 34255790 28cpu 0, pid 4465, tid 4469, ip ffffffffad136581, period 38664069 29cpu 0, pid 0, tid 0, ip ffffffffad9e55e2, period 21922384 30cpu 0, pid 4465, tid 4470, ip 7f84fe0ebebf, period 17655175 31... 32-- 33 34It requires root access, because it uses hardware cycles event. 35 36The `sampling.c` example profiles/samples all CPUs with hardware cycles, in a 37nutshell it: 38 39- creates events 40- adds them to the event list 41- opens and enables events through the event list 42- sleeps for 3 seconds 43- disables events 44- reads and displays recorded samples 45- destroys the event list 46 47The first thing you need to do before using libperf is to call init function: 48 49[source,c] 50-- 51 12 static int libperf_print(enum libperf_print_level level, 52 13 const char *fmt, va_list ap) 53 14 { 54 15 return vfprintf(stderr, fmt, ap); 55 16 } 56 57 23 int main(int argc, char **argv) 58 24 { 59 ... 60 40 libperf_init(libperf_print); 61-- 62 63It will setup the library and sets function for debug output from library. 64 65The `libperf_print` callback will receive any message with its debug level, 66defined as: 67 68[source,c] 69-- 70enum libperf_print_level { 71 LIBPERF_ERR, 72 LIBPERF_WARN, 73 LIBPERF_INFO, 74 LIBPERF_DEBUG, 75 LIBPERF_DEBUG2, 76 LIBPERF_DEBUG3, 77}; 78-- 79 80Once the setup is complete we start by defining cycles event using the `struct perf_event_attr`: 81 82[source,c] 83-- 84 29 struct perf_event_attr attr = { 85 30 .type = PERF_TYPE_HARDWARE, 86 31 .config = PERF_COUNT_HW_CPU_CYCLES, 87 32 .disabled = 1, 88 33 .freq = 1, 89 34 .sample_freq = 10, 90 35 .sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_CPU|PERF_SAMPLE_PERIOD, 91 36 }; 92-- 93 94Next step is to prepare CPUs map. 95 96In this case we will monitor all the available CPUs: 97 98[source,c] 99-- 100 42 cpus = perf_cpu_map__new(NULL); 101 43 if (!cpus) { 102 44 fprintf(stderr, "failed to create cpus\n"); 103 45 return -1; 104 46 } 105-- 106 107Now we create libperf's event list, which will serve as holder for the cycles event: 108 109[source,c] 110-- 111 48 evlist = perf_evlist__new(); 112 49 if (!evlist) { 113 50 fprintf(stderr, "failed to create evlist\n"); 114 51 goto out_cpus; 115 52 } 116-- 117 118We create libperf's event for the cycles attribute we defined earlier and add it to the list: 119 120[source,c] 121-- 122 54 evsel = perf_evsel__new(&attr); 123 55 if (!evsel) { 124 56 fprintf(stderr, "failed to create cycles\n"); 125 57 goto out_cpus; 126 58 } 127 59 128 60 perf_evlist__add(evlist, evsel); 129-- 130 131Configure event list with the cpus map and open event: 132 133[source,c] 134-- 135 62 perf_evlist__set_maps(evlist, cpus, NULL); 136 63 137 64 err = perf_evlist__open(evlist); 138 65 if (err) { 139 66 fprintf(stderr, "failed to open evlist\n"); 140 67 goto out_evlist; 141 68 } 142-- 143 144Once the events list is open, we can create memory maps AKA perf ring buffers: 145 146[source,c] 147-- 148 70 err = perf_evlist__mmap(evlist, 4); 149 71 if (err) { 150 72 fprintf(stderr, "failed to mmap evlist\n"); 151 73 goto out_evlist; 152 74 } 153-- 154 155The event is created as disabled (note the `disabled = 1` assignment above), 156so we need to enable the events list explicitly. 157 158From this moment the cycles event is sampling. 159 160We will sleep for 3 seconds while the ring buffers get data from all CPUs, then we disable the events list. 161 162[source,c] 163-- 164 76 perf_evlist__enable(evlist); 165 77 sleep(3); 166 78 perf_evlist__disable(evlist); 167-- 168 169Following code walks through the ring buffers and reads stored events/samples: 170 171[source,c] 172-- 173 80 perf_evlist__for_each_mmap(evlist, map, false) { 174 81 if (perf_mmap__read_init(map) < 0) 175 82 continue; 176 83 177 84 while ((event = perf_mmap__read_event(map)) != NULL) { 178 179 /* process event */ 180 181108 perf_mmap__consume(map); 182109 } 183110 perf_mmap__read_done(map); 184111 } 185 186-- 187 188Each sample needs to get parsed: 189 190[source,c] 191-- 192 85 int cpu, pid, tid; 193 86 __u64 ip, period, *array; 194 87 union u64_swap u; 195 88 196 89 array = event->sample.array; 197 90 198 91 ip = *array; 199 92 array++; 200 93 201 94 u.val64 = *array; 202 95 pid = u.val32[0]; 203 96 tid = u.val32[1]; 204 97 array++; 205 98 206 99 u.val64 = *array; 207100 cpu = u.val32[0]; 208101 array++; 209102 210103 period = *array; 211104 212105 fprintf(stdout, "cpu %3d, pid %6d, tid %6d, ip %20llx, period %20llu\n", 213106 cpu, pid, tid, ip, period); 214-- 215 216And finally cleanup. 217 218We close the whole events list (both events) and remove it together with the threads map: 219 220[source,c] 221-- 222113 out_evlist: 223114 perf_evlist__delete(evlist); 224115 out_cpus: 225116 perf_cpu_map__put(cpus); 226117 return err; 227118 } 228-- 229 230REPORTING BUGS 231-------------- 232Report bugs to <linux-perf-users@vger.kernel.org>. 233 234LICENSE 235------- 236libperf is Free Software licensed under the GNU LGPL 2.1 237 238RESOURCES 239--------- 240https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git 241 242SEE ALSO 243-------- 244libperf(3), libperf-counting(7)