1 // SPDX-License-Identifier: GPL-2.0
2 #include <errno.h>
3 #include <inttypes.h>
4 #include <linux/string.h>
5 /* For the CLR_() macros */
6 #include <pthread.h>
7
8 #include <sched.h>
9 #include "evlist.h"
10 #include "evsel.h"
11 #include "debug.h"
12 #include "record.h"
13 #include "tests.h"
14 #include "util/mmap.h"
15
sched__get_first_possible_cpu(pid_t pid,cpu_set_t * maskp)16 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
17 {
18 int i, cpu = -1, nrcpus = 1024;
19 realloc:
20 CPU_ZERO(maskp);
21
22 if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
23 if (errno == EINVAL && nrcpus < (1024 << 8)) {
24 nrcpus = nrcpus << 2;
25 goto realloc;
26 }
27 perror("sched_getaffinity");
28 return -1;
29 }
30
31 for (i = 0; i < nrcpus; i++) {
32 if (CPU_ISSET(i, maskp)) {
33 if (cpu == -1)
34 cpu = i;
35 else
36 CPU_CLR(i, maskp);
37 }
38 }
39
40 return cpu;
41 }
42
test__PERF_RECORD(struct test * test __maybe_unused,int subtest __maybe_unused)43 int test__PERF_RECORD(struct test *test __maybe_unused, int subtest __maybe_unused)
44 {
45 struct record_opts opts = {
46 .target = {
47 .uid = UINT_MAX,
48 .uses_mmap = true,
49 },
50 .no_buffering = true,
51 .mmap_pages = 256,
52 };
53 cpu_set_t cpu_mask;
54 size_t cpu_mask_size = sizeof(cpu_mask);
55 struct evlist *evlist = perf_evlist__new_dummy();
56 struct evsel *evsel;
57 struct perf_sample sample;
58 const char *cmd = "sleep";
59 const char *argv[] = { cmd, "1", NULL, };
60 char *bname, *mmap_filename;
61 u64 prev_time = 0;
62 bool found_cmd_mmap = false,
63 found_coreutils_mmap = false,
64 found_libc_mmap = false,
65 found_vdso_mmap = false,
66 found_ld_mmap = false;
67 int err = -1, errs = 0, i, wakeups = 0;
68 u32 cpu;
69 int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
70 char sbuf[STRERR_BUFSIZE];
71
72 if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
73 evlist = perf_evlist__new_default();
74
75 if (evlist == NULL) {
76 pr_debug("Not enough memory to create evlist\n");
77 goto out;
78 }
79
80 /*
81 * Create maps of threads and cpus to monitor. In this case
82 * we start with all threads and cpus (-1, -1) but then in
83 * perf_evlist__prepare_workload we'll fill in the only thread
84 * we're monitoring, the one forked there.
85 */
86 err = perf_evlist__create_maps(evlist, &opts.target);
87 if (err < 0) {
88 pr_debug("Not enough memory to create thread/cpu maps\n");
89 goto out_delete_evlist;
90 }
91
92 /*
93 * Prepare the workload in argv[] to run, it'll fork it, and then wait
94 * for perf_evlist__start_workload() to exec it. This is done this way
95 * so that we have time to open the evlist (calling sys_perf_event_open
96 * on all the fds) and then mmap them.
97 */
98 err = perf_evlist__prepare_workload(evlist, &opts.target, argv, false, NULL);
99 if (err < 0) {
100 pr_debug("Couldn't run the workload!\n");
101 goto out_delete_evlist;
102 }
103
104 /*
105 * Config the evsels, setting attr->comm on the first one, etc.
106 */
107 evsel = evlist__first(evlist);
108 perf_evsel__set_sample_bit(evsel, CPU);
109 perf_evsel__set_sample_bit(evsel, TID);
110 perf_evsel__set_sample_bit(evsel, TIME);
111 perf_evlist__config(evlist, &opts, NULL);
112
113 err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
114 if (err < 0) {
115 pr_debug("sched__get_first_possible_cpu: %s\n",
116 str_error_r(errno, sbuf, sizeof(sbuf)));
117 goto out_delete_evlist;
118 }
119
120 cpu = err;
121
122 /*
123 * So that we can check perf_sample.cpu on all the samples.
124 */
125 if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) {
126 pr_debug("sched_setaffinity: %s\n",
127 str_error_r(errno, sbuf, sizeof(sbuf)));
128 goto out_delete_evlist;
129 }
130
131 /*
132 * Call sys_perf_event_open on all the fds on all the evsels,
133 * grouping them if asked to.
134 */
135 err = evlist__open(evlist);
136 if (err < 0) {
137 pr_debug("perf_evlist__open: %s\n",
138 str_error_r(errno, sbuf, sizeof(sbuf)));
139 goto out_delete_evlist;
140 }
141
142 /*
143 * mmap the first fd on a given CPU and ask for events for the other
144 * fds in the same CPU to be injected in the same mmap ring buffer
145 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
146 */
147 err = evlist__mmap(evlist, opts.mmap_pages);
148 if (err < 0) {
149 pr_debug("evlist__mmap: %s\n",
150 str_error_r(errno, sbuf, sizeof(sbuf)));
151 goto out_delete_evlist;
152 }
153
154 /*
155 * Now that all is properly set up, enable the events, they will
156 * count just on workload.pid, which will start...
157 */
158 evlist__enable(evlist);
159
160 /*
161 * Now!
162 */
163 perf_evlist__start_workload(evlist);
164
165 while (1) {
166 int before = total_events;
167
168 for (i = 0; i < evlist->core.nr_mmaps; i++) {
169 union perf_event *event;
170 struct mmap *md;
171
172 md = &evlist->mmap[i];
173 if (perf_mmap__read_init(md) < 0)
174 continue;
175
176 while ((event = perf_mmap__read_event(md)) != NULL) {
177 const u32 type = event->header.type;
178 const char *name = perf_event__name(type);
179
180 ++total_events;
181 if (type < PERF_RECORD_MAX)
182 nr_events[type]++;
183
184 err = perf_evlist__parse_sample(evlist, event, &sample);
185 if (err < 0) {
186 if (verbose > 0)
187 perf_event__fprintf(event, stderr);
188 pr_debug("Couldn't parse sample\n");
189 goto out_delete_evlist;
190 }
191
192 if (verbose > 0) {
193 pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
194 perf_event__fprintf(event, stderr);
195 }
196
197 if (prev_time > sample.time) {
198 pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
199 name, prev_time, sample.time);
200 ++errs;
201 }
202
203 prev_time = sample.time;
204
205 if (sample.cpu != cpu) {
206 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
207 name, cpu, sample.cpu);
208 ++errs;
209 }
210
211 if ((pid_t)sample.pid != evlist->workload.pid) {
212 pr_debug("%s with unexpected pid, expected %d, got %d\n",
213 name, evlist->workload.pid, sample.pid);
214 ++errs;
215 }
216
217 if ((pid_t)sample.tid != evlist->workload.pid) {
218 pr_debug("%s with unexpected tid, expected %d, got %d\n",
219 name, evlist->workload.pid, sample.tid);
220 ++errs;
221 }
222
223 if ((type == PERF_RECORD_COMM ||
224 type == PERF_RECORD_MMAP ||
225 type == PERF_RECORD_MMAP2 ||
226 type == PERF_RECORD_FORK ||
227 type == PERF_RECORD_EXIT) &&
228 (pid_t)event->comm.pid != evlist->workload.pid) {
229 pr_debug("%s with unexpected pid/tid\n", name);
230 ++errs;
231 }
232
233 if ((type == PERF_RECORD_COMM ||
234 type == PERF_RECORD_MMAP ||
235 type == PERF_RECORD_MMAP2) &&
236 event->comm.pid != event->comm.tid) {
237 pr_debug("%s with different pid/tid!\n", name);
238 ++errs;
239 }
240
241 switch (type) {
242 case PERF_RECORD_COMM:
243 if (strcmp(event->comm.comm, cmd)) {
244 pr_debug("%s with unexpected comm!\n", name);
245 ++errs;
246 }
247 break;
248 case PERF_RECORD_EXIT:
249 goto found_exit;
250 case PERF_RECORD_MMAP:
251 mmap_filename = event->mmap.filename;
252 goto check_bname;
253 case PERF_RECORD_MMAP2:
254 mmap_filename = event->mmap2.filename;
255 check_bname:
256 bname = strrchr(mmap_filename, '/');
257 if (bname != NULL) {
258 if (!found_cmd_mmap)
259 found_cmd_mmap = !strcmp(bname + 1, cmd);
260 if (!found_coreutils_mmap)
261 found_coreutils_mmap = !strcmp(bname + 1, "coreutils");
262 if (!found_libc_mmap)
263 found_libc_mmap = !strncmp(bname + 1, "libc", 4);
264 if (!found_ld_mmap)
265 found_ld_mmap = !strncmp(bname + 1, "ld", 2);
266 } else if (!found_vdso_mmap)
267 found_vdso_mmap = !strcmp(mmap_filename, "[vdso]");
268 break;
269
270 case PERF_RECORD_SAMPLE:
271 /* Just ignore samples for now */
272 break;
273 default:
274 pr_debug("Unexpected perf_event->header.type %d!\n",
275 type);
276 ++errs;
277 }
278
279 perf_mmap__consume(md);
280 }
281 perf_mmap__read_done(md);
282 }
283
284 /*
285 * We don't use poll here because at least at 3.1 times the
286 * PERF_RECORD_{!SAMPLE} events don't honour
287 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
288 */
289 if (total_events == before && false)
290 evlist__poll(evlist, -1);
291
292 sleep(1);
293 if (++wakeups > 5) {
294 pr_debug("No PERF_RECORD_EXIT event!\n");
295 break;
296 }
297 }
298
299 found_exit:
300 if (nr_events[PERF_RECORD_COMM] > 1 + !!found_coreutils_mmap) {
301 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
302 ++errs;
303 }
304
305 if (nr_events[PERF_RECORD_COMM] == 0) {
306 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
307 ++errs;
308 }
309
310 if (!found_cmd_mmap && !found_coreutils_mmap) {
311 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
312 ++errs;
313 }
314
315 if (!found_libc_mmap) {
316 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
317 ++errs;
318 }
319
320 if (!found_ld_mmap) {
321 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
322 ++errs;
323 }
324
325 if (!found_vdso_mmap) {
326 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
327 ++errs;
328 }
329 out_delete_evlist:
330 evlist__delete(evlist);
331 out:
332 return (err < 0 || errs > 0) ? -1 : 0;
333 }
334