void read_cpu(int c) { perf_event_desc_t *fds; uint64_t val, delta; double ratio; int i, j, n, ret; fds = all_fds[c]; if (fds[0].fd == -1) { printf("CPU%d not monitored\n", c); return; } for(i=0, j = 0; i < options.num_groups; i++) { for(n = 0; n < options.nevents[i]; n++, j++) { ret = read(fds[j].fd, fds[j].values, sizeof(fds[j].values)); if (ret != sizeof(fds[j].values)) { if (ret == -1) err(1, "cannot read event %s : %d", fds[j].name, ret); else { warnx("CPU%d G%-2d could not read event %s, read=%d", c, i, fds[j].name, ret); continue; } } /* * scaling because we may be sharing the PMU and * thus may be multiplexed */ delta = perf_scale_delta(fds[j].values, fds[j].prev_values); val = perf_scale(fds[j].values); ratio = perf_scale_ratio(fds[j].values); printf("CPU%-3d G%-2d %'20"PRIu64" %'20"PRIu64" %s (scaling %.2f%%, ena=%'"PRIu64", run=%'"PRIu64") %s\n", c, i, val, delta, fds[j].name, (1.0-ratio)*100, fds[j].values[1], fds[j].values[2], options.cgroup_name ? options.cgroup_name : ""); fds[j].prev_values[0] = fds[j].values[0]; fds[j].prev_values[1] = fds[j].values[1]; fds[j].prev_values[2] = fds[j].values[2]; if (fds[j].values[2] > fds[j].values[1]) errx(1, "WARNING: time_running > time_enabled %"PRIu64"\n", fds[j].values[2] - fds[j].values[1]); } } }
/* Same as stop above, but without the parameter for faster calling into. */ uint64_t *pfm_pause_counters() { int ret, i; uint64_t values[3]; ret = prctl(PR_TASK_PERF_EVENTS_DISABLE); if (ret) fprintf(stderr, "prctl(disable) failed\n"); /* * now read the results. We use pfp_event_count because * libpfm guarantees that counters for the events always * come first. */ memset(values, 0, sizeof(values)); for (i=0; i < num_fds; i++) { uint64_t val; double ratio; ret = read(fds[i].fd, values, sizeof(values)); if (ret < sizeof(values)) { if (ret == -1) err(1, "cannot read results: %s", strerror(errno)); else warnx("could not read event%d", i); } /* * scaling is systematic because we may be sharing the PMU and * thus may be multiplexed */ val = perf_scale(values); ratio = perf_scale_ratio(values); counter_values[i] = val; #ifdef VERBOSE printf("%'20"PRIu64" %s (%.2f%% scaling, ena=%'"PRIu64", run=%'"PRIu64")\n", val, fds[i].name, (1.0-ratio)*100.0, values[1], values[2]); #endif } return counter_values; }
void perfevent_read() { int i, ret; uint64_t values[3]; /* * now read the results. We use pfp_event_count because * libpfm guarantees that counters for the events always * come first. */ memset(values, 0, sizeof(values)); fprintf(stderr, "============================ Tabulate Statistics ============================\n"); for (i=0; i < perfevent_num_fds; i++) { fprintf(stderr, "%s ", perfevent_fds[i].name); } fprintf(stderr, "\n"); for (i=0; i < perfevent_num_fds; i++) { uint64_t val; double ratio; ret = read(perfevent_fds[i].fd, values, sizeof(values)); if (ret < sizeof(values)) { if (ret == -1) err(1, "cannot read results: %s", strerror(errno)); else warnx("could not read event%d", i); } /* * scaling is systematic because we may be sharing the PMU and * thus may be multiplexed */ val = perf_scale(values); ratio = perf_scale_ratio(values); if (ratio == 1.0) fprintf(stderr, "%lld ", (long long int) val); else if (ratio == 0.0) fprintf(stderr, "NO_VALUE "); else fprintf(stderr, "%lld-SCALED-%.2f%% ", (long long int) val, ratio*100.0); } fprintf(stderr, "\n=============================================================================\n"); }
static void print_counts(perf_event_desc_t *fds, int num, int do_delta) { uint64_t values[3]; ssize_t ret; int i; /* * now simply read the results. */ for(i=0; i < num; i++) { uint64_t val; double ratio; ret = read(fds[i].fd, values, sizeof(values)); if (ret < (ssize_t)sizeof(values)) { if (ret == -1) err(1, "cannot read values event %s", fds[i].name); else warnx("could not read event%d", i); } /* * scaling because we may be sharing the PMU and * thus may be multiplexed */ fds[i].prev_value = fds[i].value; fds[i].value = val = perf_scale(values); ratio = perf_scale_ratio(values); val = do_delta ? (val - fds[i].prev_value): val; if (ratio == 1.0) printf("%20"PRIu64" %s\n", val, fds[i].name); else if (ratio == 0.0) printf("%20"PRIu64" %s (did not run: incompatible events, too many events in a group, competing session)\n", val, fds[i].name); else printf("%20"PRIu64" %s (scaled from %.2f%% of time)\n", val, fds[i].name, ratio*100.0); } }
static void print_counts(perf_event_desc_t *fds, int num) { double ratio; uint64_t val, delta; int i; read_groups(fds, num); for(i=0; i < num; i++) { val = perf_scale(fds[i].values); delta = perf_scale_delta(fds[i].values, fds[i].prev_values); ratio = perf_scale_ratio(fds[i].values); /* separate groups */ if (perf_is_group_leader(fds, i)) putchar('\n'); if (options.print) printf("%'20"PRIu64" %'20"PRIu64" %s (%.2f%% scaling, ena=%'"PRIu64", run=%'"PRIu64")\n", val, delta, fds[i].name, (1.0-ratio)*100.0, fds[i].values[1], fds[i].values[2]); else printf("%'20"PRIu64" %s (%.2f%% scaling, ena=%'"PRIu64", run=%'"PRIu64")\n", val, fds[i].name, (1.0-ratio)*100.0, fds[i].values[1], fds[i].values[2]); fds[i].prev_values[0] = fds[i].values[0]; fds[i].prev_values[1] = fds[i].values[1]; fds[i].prev_values[2] = fds[i].values[2]; } }
static void print_counts(perf_event_desc_t *fds, int num, int do_delta) { ssize_t ret; int i; /* * now simply read the results. */ for(i=0; i < num; i++) { uint64_t val; double ratio; ret = read(fds[i].fd, fds[i].values, sizeof(fds[i].values)); if (ret < (ssize_t)sizeof(fds[i].values)) { if (ret == -1) err(1, "cannot read values event %s", fds[i].name); else warnx("could not read event%d", i); } val = perf_scale(fds[i].values); ratio = perf_scale_ratio(fds[i].values); val = do_delta ? perf_scale_delta(fds[i].values, fds[i].prev_values) : val; fds[i].prev_values[0] = fds[i].values[0]; fds[i].prev_values[1] = fds[i].values[1]; fds[i].prev_values[2] = fds[i].values[2]; if (ratio == 1.0) printf("%20"PRIu64" %s\n", val, fds[i].name); else if (ratio == 0.0) printf("%20"PRIu64" %s (did not run: incompatible events, too many events in a group, competing session)\n", val, fds[i].name); else printf("%20"PRIu64" %s (scaled from %.2f%% of time)\n", val, fds[i].name, ratio*100.0); } }
static void measure(void) { perf_event_desc_t *fds = NULL; int num_fds = 0; uint64_t values[3]; ssize_t n; int i, ret; int pr[2], pw[2]; pid_t pid; char cc = '0'; ret = pfm_initialize(); if (ret != PFM_SUCCESS) err(1, "cannot initialize libpfm"); if (options.cpu == -1) { srandom(getpid()); options.cpu = random() % sysconf(_SC_NPROCESSORS_ONLN); } ret = pipe(pr); if (ret) err(1, "cannot create read pipe"); ret = pipe(pw); if (ret) err(1, "cannot create write pipe"); ret = perf_setup_list_events(options.events, &fds, &num_fds); if (ret || !num_fds) exit(1); for(i=0; i < num_fds; i++) { fds[i].hw.disabled = 1; fds[i].hw.read_format = PERF_FORMAT_SCALE; fds[i].fd = perf_event_open(&fds[i].hw, 0, -1, -1, 0); if (fds[i].fd == -1) err(1, "cannot open event %d", i); } /* * Pin to CPU0, inherited by child process. That will enforce * the ping-pionging and thus stress the PMU context switch * which is what we want */ ret = pin_cpu(getpid(), options.cpu); if (ret) err(1, "cannot pin to CPU%d", options.cpu); printf("Both processes pinned to CPU%d, running for %d seconds\n", options.cpu, options.delay); /* * create second process which is not monitoring at the moment */ switch(pid=fork()) { case -1: err(1, "cannot create child\n"); case 0: /* do not inherit session fd */ for(i=0; i < num_fds; i++) close(fds[i].fd); /* pr[]: write master, read child */ /* pw[]: read master, write child */ close(pr[1]); close(pw[0]); do_child(pr[0], pw[1]); exit(1); } close(pr[0]); close(pw[1]); /* * Let's roll now */ prctl(PR_TASK_PERF_EVENTS_ENABLE); signal(SIGALRM, sig_handler); alarm(options.delay); /* * ping pong loop */ while(!quit) { n = write(pr[1], "c", 1); if (n < 1) err(1, "write failed"); n = read(pw[0], &cc, 1); if (n < 1) err(1, "read failed"); } prctl(PR_TASK_PERF_EVENTS_DISABLE); for(i=0; i < num_fds; i++) { uint64_t val; double ratio; ret = read(fds[i].fd, values, sizeof(values)); if (ret == -1) err(1,"pfm_read error"); if (ret != sizeof(values)) errx(1, "did not read correct amount %d", ret); val = perf_scale(values); ratio = perf_scale_ratio(values); if (ratio == 1.0) printf("%20"PRIu64" %s\n", val, fds[i].name); else if (ratio == 0.0) printf("%20"PRIu64" %s (did not run: competing session)\n", val, fds[i].name); else printf("%20"PRIu64" %s (scaled from %.2f%% of time)\n", val, fds[i].name, ratio*100.0); } /* * kill child process */ kill(SIGKILL, pid); /* * close pipes */ close(pr[1]); close(pw[0]); /* * and destroy our session */ for(i=0; i < num_fds; i++) close(fds[i].fd); perf_free_fds(fds, num_fds); /* free libpfm resources cleanly */ pfm_terminate(); }
int perf_display_sample(perf_event_desc_t *fds, int num_fds, int idx, struct perf_event_header *ehdr, FILE *fp) { perf_event_desc_t *hw; struct { uint32_t pid, tid; } pid; struct { uint64_t value, id; } grp; uint64_t time_enabled, time_running; size_t sz; uint64_t type, fmt; uint64_t val64; const char *str; int ret, e; if (!fds || !fp || !ehdr || num_fds < 0 || idx < 0 || idx >= num_fds) return -1; sz = ehdr->size - sizeof(*ehdr); hw = fds+idx; type = hw->hw.sample_type; fmt = hw->hw.read_format; /* * the sample_type information is laid down * based on the PERF_RECORD_SAMPLE format specified * in the perf_event.h header file. * That order is different from the enum perf_event_sample_format */ if (type & PERF_SAMPLE_IP) { const char *xtra = " "; ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx("cannot read IP"); return -1; } /* * MISC_EXACT_IP indicates that kernel is returning * th IIP of an instruction which caused the event, i.e., * no skid */ if (hw->hw.precise_ip && (ehdr->misc & PERF_RECORD_MISC_EXACT_IP)) xtra = " (exact) "; fprintf(fp, "IIP:%#016"PRIx64"%s", val64, xtra); sz -= sizeof(val64); } if (type & PERF_SAMPLE_TID) { ret = perf_read_buffer(hw, &pid, sizeof(pid)); if (ret) { warnx( "cannot read PID"); return -1; } fprintf(fp, "PID:%d TID:%d ", pid.pid, pid.tid); sz -= sizeof(pid); } if (type & PERF_SAMPLE_TIME) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read time"); return -1; } fprintf(fp, "TIME:%'"PRIu64" ", val64); sz -= sizeof(val64); } if (type & PERF_SAMPLE_ADDR) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read addr"); return -1; } fprintf(fp, "ADDR:%#016"PRIx64" ", val64); sz -= sizeof(val64); } if (type & PERF_SAMPLE_ID) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read id"); return -1; } fprintf(fp, "ID:%"PRIu64" ", val64); sz -= sizeof(val64); } if (type & PERF_SAMPLE_STREAM_ID) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read stream_id"); return -1; } fprintf(fp, "STREAM_ID:%"PRIu64" ", val64); sz -= sizeof(val64); } if (type & PERF_SAMPLE_CPU) { struct { uint32_t cpu, reserved; } cpu; ret = perf_read_buffer(hw, &cpu, sizeof(cpu)); if (ret) { warnx( "cannot read cpu"); return -1; } fprintf(fp, "CPU:%u ", cpu.cpu); sz -= sizeof(cpu); } if (type & PERF_SAMPLE_PERIOD) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read period"); return -1; } fprintf(fp, "PERIOD:%'"PRIu64" ", val64); sz -= sizeof(val64); } /* struct read_format { * { u64 value; * { u64 time_enabled; } && PERF_FORMAT_ENABLED * { u64 time_running; } && PERF_FORMAT_RUNNING * { u64 id; } && PERF_FORMAT_ID * } && !PERF_FORMAT_GROUP * * { u64 nr; * { u64 time_enabled; } && PERF_FORMAT_ENABLED * { u64 time_running; } && PERF_FORMAT_RUNNING * { u64 value; * { u64 id; } && PERF_FORMAT_ID * } cntr[nr]; * } && PERF_FORMAT_GROUP * }; */ if (type & PERF_SAMPLE_READ) { uint64_t values[3]; uint64_t nr; if (fmt & PERF_FORMAT_GROUP) { ret = perf_read_buffer_64(hw, &nr); if (ret) { warnx( "cannot read nr"); return -1; } sz -= sizeof(nr); time_enabled = time_running = 1; if (fmt & PERF_FORMAT_TOTAL_TIME_ENABLED) { ret = perf_read_buffer_64(hw, &time_enabled); if (ret) { warnx( "cannot read timing info"); return -1; } sz -= sizeof(time_enabled); } if (fmt & PERF_FORMAT_TOTAL_TIME_RUNNING) { ret = perf_read_buffer_64(hw, &time_running); if (ret) { warnx( "cannot read timing info"); return -1; } sz -= sizeof(time_running); } fprintf(fp, "ENA=%'"PRIu64" RUN=%'"PRIu64" NR=%"PRIu64"\n", time_enabled, time_running, nr); values[1] = time_enabled; values[2] = time_running; while(nr--) { grp.id = -1; ret = perf_read_buffer_64(hw, &grp.value); if (ret) { warnx( "cannot read group value"); return -1; } sz -= sizeof(grp.value); if (fmt & PERF_FORMAT_ID) { ret = perf_read_buffer_64(hw, &grp.id); if (ret) { warnx( "cannot read leader id"); return -1; } sz -= sizeof(grp.id); } e = perf_id2event(fds, num_fds, grp.id); if (e == -1) str = "unknown sample event"; else str = fds[e].name; values[0] = grp.value; grp.value = perf_scale(values); fprintf(fp, "\t%'"PRIu64" %s (%"PRIu64"%s)\n", grp.value, str, grp.id, time_running != time_enabled ? ", scaled":""); } } else { time_enabled = time_running = 0; /* * this program does not use FORMAT_GROUP when there is only one event */ ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read value"); return -1; } sz -= sizeof(val64); if (fmt & PERF_FORMAT_TOTAL_TIME_ENABLED) { ret = perf_read_buffer_64(hw, &time_enabled); if (ret) { warnx( "cannot read timing info"); return -1; } sz -= sizeof(time_enabled); } if (fmt & PERF_FORMAT_TOTAL_TIME_RUNNING) { ret = perf_read_buffer_64(hw, &time_running); if (ret) { warnx( "cannot read timing info"); return -1; } sz -= sizeof(time_running); } if (fmt & PERF_FORMAT_ID) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read leader id"); return -1; } sz -= sizeof(val64); } fprintf(fp, "ENA=%'"PRIu64" RUN=%'"PRIu64"\n", time_enabled, time_running); values[0] = val64; values[1] = time_enabled; values[2] = time_running; val64 = perf_scale(values); fprintf(fp, "\t%'"PRIu64" %s %s\n", val64, fds[0].name, time_running != time_enabled ? ", scaled":""); } } if (type & PERF_SAMPLE_CALLCHAIN) { uint64_t nr, ip; ret = perf_read_buffer_64(hw, &nr); if (ret) { warnx( "cannot read callchain nr"); return -1; } sz -= sizeof(nr); while(nr--) { ret = perf_read_buffer_64(hw, &ip); if (ret) { warnx( "cannot read ip"); return -1; } sz -= sizeof(ip); fprintf(fp, "\t0x%"PRIx64"\n", ip); } } if (type & PERF_SAMPLE_RAW) { ret = __perf_handle_raw(hw); if (ret == -1) return -1; sz -= ret; } if (type & PERF_SAMPLE_BRANCH_STACK) { ret = perf_display_branch_stack(hw, fp); sz -= ret; } if (type & PERF_SAMPLE_REGS_USER) { ret = perf_display_regs_user(hw, fp); sz -= ret; } if (type & PERF_SAMPLE_STACK_USER) { ret = perf_display_stack_user(hw, fp); sz -= ret; } if (type & PERF_SAMPLE_WEIGHT) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read weight"); return -1; } fprintf(fp, "WEIGHT:%'"PRIu64" ", val64); sz -= sizeof(val64); } if (type & PERF_SAMPLE_DATA_SRC) { ret = perf_read_buffer_64(hw, &val64); if (ret) { warnx( "cannot read data src"); return -1; } fprintf(fp, "DATA_SRC:%'"PRIu64" ", val64); sz -= sizeof(val64); } /* * if we have some data left, it is because there is more * than what we know about. In fact, it is more complicated * because we may have the right size but wrong layout. But * that's the best we can do. */ if (sz) { warnx("did not correctly parse sample leftover=%zu", sz); perf_skip_buffer(hw, sz); } fputc('\n',fp); return 0; }
static void read_groups(perf_event_desc_t *fds, int num) { uint64_t *values = NULL; size_t new_sz, sz = 0; int i, evt, ret; /* * { u64 nr; * { u64 time_enabled; } && PERF_FORMAT_ENABLED * { u64 time_running; } && PERF_FORMAT_RUNNING * { u64 value; * { u64 id; } && PERF_FORMAT_ID * } cntr[nr]; * } && PERF_FORMAT_GROUP"%'20"PRIu64" * * we do not use FORMAT_ID in this program */ for (evt = 0; evt < num; ) { int num_evts_to_read; if (options.format_group) { num_evts_to_read = perf_get_group_nevents(fds, num, evt); new_sz = sizeof(uint64_t) * (3 + num_evts_to_read); } else { num_evts_to_read = 1; new_sz = sizeof(uint64_t) * 3; } if (new_sz > sz) { sz = new_sz; values = realloc(values, sz); } if (!values) err(1, "cannot allocate memory for values\n"); ret = read(fds[evt].fd, values, new_sz); if (ret != new_sz) { /* unsigned */ if (ret == -1) err(1, "cannot read values event %s", fds[evt].name); /* likely pinned and could not be loaded */ warnx("could not read event %d, tried to read %zu bytes, but got %d", evt, new_sz, ret); } /* * propagate to save area */ for (i = evt; i < (evt + num_evts_to_read); i++) { if (options.format_group) values[0] = values[3 + (i - evt)]; /* * scaling because we may be sharing the PMU and * thus may be multiplexed */ fds[i].prev_value = fds[i].value; fds[i].value = perf_scale(values); fds[i].enabled = values[1]; fds[i].running = values[2]; } evt += num_evts_to_read; } if (values) free(values); }