3 /* For the CLR_() macros */
13 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
15 int i, cpu = -1, nrcpus = 1024;
19 if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
20 if (errno == EINVAL && nrcpus < (1024 << 8)) {
24 perror("sched_getaffinity");
28 for (i = 0; i < nrcpus; i++) {
29 if (CPU_ISSET(i, maskp)) {
40 int test__PERF_RECORD(int subtest __maybe_unused)
42 struct record_opts opts = {
51 size_t cpu_mask_size = sizeof(cpu_mask);
52 struct perf_evlist *evlist = perf_evlist__new_dummy();
53 struct perf_evsel *evsel;
54 struct perf_sample sample;
55 const char *cmd = "sleep";
56 const char *argv[] = { cmd, "1", NULL, };
57 char *bname, *mmap_filename;
59 bool found_cmd_mmap = false,
60 found_libc_mmap = false,
61 found_vdso_mmap = false,
62 found_ld_mmap = false;
63 int err = -1, errs = 0, i, wakeups = 0;
65 int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
66 char sbuf[STRERR_BUFSIZE];
68 if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
69 evlist = perf_evlist__new_default();
72 pr_debug("Not enough memory to create evlist\n");
77 * Create maps of threads and cpus to monitor. In this case
78 * we start with all threads and cpus (-1, -1) but then in
79 * perf_evlist__prepare_workload we'll fill in the only thread
80 * we're monitoring, the one forked there.
82 err = perf_evlist__create_maps(evlist, &opts.target);
84 pr_debug("Not enough memory to create thread/cpu maps\n");
85 goto out_delete_evlist;
89 * Prepare the workload in argv[] to run, it'll fork it, and then wait
90 * for perf_evlist__start_workload() to exec it. This is done this way
91 * so that we have time to open the evlist (calling sys_perf_event_open
92 * on all the fds) and then mmap them.
94 err = perf_evlist__prepare_workload(evlist, &opts.target, argv, false, NULL);
96 pr_debug("Couldn't run the workload!\n");
97 goto out_delete_evlist;
101 * Config the evsels, setting attr->comm on the first one, etc.
103 evsel = perf_evlist__first(evlist);
104 perf_evsel__set_sample_bit(evsel, CPU);
105 perf_evsel__set_sample_bit(evsel, TID);
106 perf_evsel__set_sample_bit(evsel, TIME);
107 perf_evlist__config(evlist, &opts, NULL);
109 err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
111 pr_debug("sched__get_first_possible_cpu: %s\n",
112 str_error_r(errno, sbuf, sizeof(sbuf)));
113 goto out_delete_evlist;
119 * So that we can check perf_sample.cpu on all the samples.
121 if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) {
122 pr_debug("sched_setaffinity: %s\n",
123 str_error_r(errno, sbuf, sizeof(sbuf)));
124 goto out_delete_evlist;
128 * Call sys_perf_event_open on all the fds on all the evsels,
129 * grouping them if asked to.
131 err = perf_evlist__open(evlist);
133 pr_debug("perf_evlist__open: %s\n",
134 str_error_r(errno, sbuf, sizeof(sbuf)));
135 goto out_delete_evlist;
139 * mmap the first fd on a given CPU and ask for events for the other
140 * fds in the same CPU to be injected in the same mmap ring buffer
141 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
143 err = perf_evlist__mmap(evlist, opts.mmap_pages, false);
145 pr_debug("perf_evlist__mmap: %s\n",
146 str_error_r(errno, sbuf, sizeof(sbuf)));
147 goto out_delete_evlist;
151 * Now that all is properly set up, enable the events, they will
152 * count just on workload.pid, which will start...
154 perf_evlist__enable(evlist);
159 perf_evlist__start_workload(evlist);
162 int before = total_events;
164 for (i = 0; i < evlist->nr_mmaps; i++) {
165 union perf_event *event;
167 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
168 const u32 type = event->header.type;
169 const char *name = perf_event__name(type);
172 if (type < PERF_RECORD_MAX)
175 err = perf_evlist__parse_sample(evlist, event, &sample);
178 perf_event__fprintf(event, stderr);
179 pr_debug("Couldn't parse sample\n");
180 goto out_delete_evlist;
184 pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
185 perf_event__fprintf(event, stderr);
188 if (prev_time > sample.time) {
189 pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
190 name, prev_time, sample.time);
194 prev_time = sample.time;
196 if (sample.cpu != cpu) {
197 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
198 name, cpu, sample.cpu);
202 if ((pid_t)sample.pid != evlist->workload.pid) {
203 pr_debug("%s with unexpected pid, expected %d, got %d\n",
204 name, evlist->workload.pid, sample.pid);
208 if ((pid_t)sample.tid != evlist->workload.pid) {
209 pr_debug("%s with unexpected tid, expected %d, got %d\n",
210 name, evlist->workload.pid, sample.tid);
214 if ((type == PERF_RECORD_COMM ||
215 type == PERF_RECORD_MMAP ||
216 type == PERF_RECORD_MMAP2 ||
217 type == PERF_RECORD_FORK ||
218 type == PERF_RECORD_EXIT) &&
219 (pid_t)event->comm.pid != evlist->workload.pid) {
220 pr_debug("%s with unexpected pid/tid\n", name);
224 if ((type == PERF_RECORD_COMM ||
225 type == PERF_RECORD_MMAP ||
226 type == PERF_RECORD_MMAP2) &&
227 event->comm.pid != event->comm.tid) {
228 pr_debug("%s with different pid/tid!\n", name);
233 case PERF_RECORD_COMM:
234 if (strcmp(event->comm.comm, cmd)) {
235 pr_debug("%s with unexpected comm!\n", name);
239 case PERF_RECORD_EXIT:
241 case PERF_RECORD_MMAP:
242 mmap_filename = event->mmap.filename;
244 case PERF_RECORD_MMAP2:
245 mmap_filename = event->mmap2.filename;
247 bname = strrchr(mmap_filename, '/');
250 found_cmd_mmap = !strcmp(bname + 1, cmd);
251 if (!found_libc_mmap)
252 found_libc_mmap = !strncmp(bname + 1, "libc", 4);
254 found_ld_mmap = !strncmp(bname + 1, "ld", 2);
255 } else if (!found_vdso_mmap)
256 found_vdso_mmap = !strcmp(mmap_filename, "[vdso]");
259 case PERF_RECORD_SAMPLE:
260 /* Just ignore samples for now */
263 pr_debug("Unexpected perf_event->header.type %d!\n",
268 perf_evlist__mmap_consume(evlist, i);
273 * We don't use poll here because at least at 3.1 times the
274 * PERF_RECORD_{!SAMPLE} events don't honour
275 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
277 if (total_events == before && false)
278 perf_evlist__poll(evlist, -1);
282 pr_debug("No PERF_RECORD_EXIT event!\n");
288 if (nr_events[PERF_RECORD_COMM] > 1) {
289 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
293 if (nr_events[PERF_RECORD_COMM] == 0) {
294 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
298 if (!found_cmd_mmap) {
299 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
303 if (!found_libc_mmap) {
304 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
308 if (!found_ld_mmap) {
309 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
313 if (!found_vdso_mmap) {
314 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
318 perf_evlist__delete(evlist);
320 return (err < 0 || errs > 0) ? -1 : 0;