1 // SPDX-License-Identifier: GPL-2.0
4 #include <linux/string.h>
5 /* For the CLR_() macros */
15 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
17 int i, cpu = -1, nrcpus = 1024;
21 if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
22 if (errno == EINVAL && nrcpus < (1024 << 8)) {
26 perror("sched_getaffinity");
30 for (i = 0; i < nrcpus; i++) {
31 if (CPU_ISSET(i, maskp)) {
42 int test__PERF_RECORD(struct test *test __maybe_unused, int subtest __maybe_unused)
44 struct record_opts opts = {
53 size_t cpu_mask_size = sizeof(cpu_mask);
54 struct evlist *evlist = perf_evlist__new_dummy();
56 struct perf_sample sample;
57 const char *cmd = "sleep";
58 const char *argv[] = { cmd, "1", NULL, };
59 char *bname, *mmap_filename;
61 bool found_cmd_mmap = false,
62 found_coreutils_mmap = false,
63 found_libc_mmap = false,
64 found_vdso_mmap = false,
65 found_ld_mmap = false;
66 int err = -1, errs = 0, i, wakeups = 0;
68 int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
69 char sbuf[STRERR_BUFSIZE];
71 if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
72 evlist = perf_evlist__new_default();
75 pr_debug("Not enough memory to create evlist\n");
80 * Create maps of threads and cpus to monitor. In this case
81 * we start with all threads and cpus (-1, -1) but then in
82 * perf_evlist__prepare_workload we'll fill in the only thread
83 * we're monitoring, the one forked there.
85 err = perf_evlist__create_maps(evlist, &opts.target);
87 pr_debug("Not enough memory to create thread/cpu maps\n");
88 goto out_delete_evlist;
92 * Prepare the workload in argv[] to run, it'll fork it, and then wait
93 * for perf_evlist__start_workload() to exec it. This is done this way
94 * so that we have time to open the evlist (calling sys_perf_event_open
95 * on all the fds) and then mmap them.
97 err = perf_evlist__prepare_workload(evlist, &opts.target, argv, false, NULL);
99 pr_debug("Couldn't run the workload!\n");
100 goto out_delete_evlist;
104 * Config the evsels, setting attr->comm on the first one, etc.
106 evsel = perf_evlist__first(evlist);
107 perf_evsel__set_sample_bit(evsel, CPU);
108 perf_evsel__set_sample_bit(evsel, TID);
109 perf_evsel__set_sample_bit(evsel, TIME);
110 perf_evlist__config(evlist, &opts, NULL);
112 err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
114 pr_debug("sched__get_first_possible_cpu: %s\n",
115 str_error_r(errno, sbuf, sizeof(sbuf)));
116 goto out_delete_evlist;
122 * So that we can check perf_sample.cpu on all the samples.
124 if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) {
125 pr_debug("sched_setaffinity: %s\n",
126 str_error_r(errno, sbuf, sizeof(sbuf)));
127 goto out_delete_evlist;
131 * Call sys_perf_event_open on all the fds on all the evsels,
132 * grouping them if asked to.
134 err = evlist__open(evlist);
136 pr_debug("perf_evlist__open: %s\n",
137 str_error_r(errno, sbuf, sizeof(sbuf)));
138 goto out_delete_evlist;
142 * mmap the first fd on a given CPU and ask for events for the other
143 * fds in the same CPU to be injected in the same mmap ring buffer
144 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
146 err = perf_evlist__mmap(evlist, opts.mmap_pages);
148 pr_debug("perf_evlist__mmap: %s\n",
149 str_error_r(errno, sbuf, sizeof(sbuf)));
150 goto out_delete_evlist;
154 * Now that all is properly set up, enable the events, they will
155 * count just on workload.pid, which will start...
157 evlist__enable(evlist);
162 perf_evlist__start_workload(evlist);
165 int before = total_events;
167 for (i = 0; i < evlist->nr_mmaps; i++) {
168 union perf_event *event;
171 md = &evlist->mmap[i];
172 if (perf_mmap__read_init(md) < 0)
175 while ((event = perf_mmap__read_event(md)) != NULL) {
176 const u32 type = event->header.type;
177 const char *name = perf_event__name(type);
180 if (type < PERF_RECORD_MAX)
183 err = perf_evlist__parse_sample(evlist, event, &sample);
186 perf_event__fprintf(event, stderr);
187 pr_debug("Couldn't parse sample\n");
188 goto out_delete_evlist;
192 pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
193 perf_event__fprintf(event, stderr);
196 if (prev_time > sample.time) {
197 pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
198 name, prev_time, sample.time);
202 prev_time = sample.time;
204 if (sample.cpu != cpu) {
205 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
206 name, cpu, sample.cpu);
210 if ((pid_t)sample.pid != evlist->workload.pid) {
211 pr_debug("%s with unexpected pid, expected %d, got %d\n",
212 name, evlist->workload.pid, sample.pid);
216 if ((pid_t)sample.tid != evlist->workload.pid) {
217 pr_debug("%s with unexpected tid, expected %d, got %d\n",
218 name, evlist->workload.pid, sample.tid);
222 if ((type == PERF_RECORD_COMM ||
223 type == PERF_RECORD_MMAP ||
224 type == PERF_RECORD_MMAP2 ||
225 type == PERF_RECORD_FORK ||
226 type == PERF_RECORD_EXIT) &&
227 (pid_t)event->comm.pid != evlist->workload.pid) {
228 pr_debug("%s with unexpected pid/tid\n", name);
232 if ((type == PERF_RECORD_COMM ||
233 type == PERF_RECORD_MMAP ||
234 type == PERF_RECORD_MMAP2) &&
235 event->comm.pid != event->comm.tid) {
236 pr_debug("%s with different pid/tid!\n", name);
241 case PERF_RECORD_COMM:
242 if (strcmp(event->comm.comm, cmd)) {
243 pr_debug("%s with unexpected comm!\n", name);
247 case PERF_RECORD_EXIT:
249 case PERF_RECORD_MMAP:
250 mmap_filename = event->mmap.filename;
252 case PERF_RECORD_MMAP2:
253 mmap_filename = event->mmap2.filename;
255 bname = strrchr(mmap_filename, '/');
258 found_cmd_mmap = !strcmp(bname + 1, cmd);
259 if (!found_coreutils_mmap)
260 found_coreutils_mmap = !strcmp(bname + 1, "coreutils");
261 if (!found_libc_mmap)
262 found_libc_mmap = !strncmp(bname + 1, "libc", 4);
264 found_ld_mmap = !strncmp(bname + 1, "ld", 2);
265 } else if (!found_vdso_mmap)
266 found_vdso_mmap = !strcmp(mmap_filename, "[vdso]");
269 case PERF_RECORD_SAMPLE:
270 /* Just ignore samples for now */
273 pr_debug("Unexpected perf_event->header.type %d!\n",
278 perf_mmap__consume(md);
280 perf_mmap__read_done(md);
284 * We don't use poll here because at least at 3.1 times the
285 * PERF_RECORD_{!SAMPLE} events don't honour
286 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
288 if (total_events == before && false)
289 perf_evlist__poll(evlist, -1);
293 pr_debug("No PERF_RECORD_EXIT event!\n");
299 if (nr_events[PERF_RECORD_COMM] > 1 + !!found_coreutils_mmap) {
300 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
304 if (nr_events[PERF_RECORD_COMM] == 0) {
305 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
309 if (!found_cmd_mmap && !found_coreutils_mmap) {
310 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
314 if (!found_libc_mmap) {
315 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
319 if (!found_ld_mmap) {
320 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
324 if (!found_vdso_mmap) {
325 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
329 evlist__delete(evlist);
331 return (err < 0 || errs > 0) ? -1 : 0;