xref: /linux/tools/perf/tests/perf-record.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 #include <errno.h>
2 #include <inttypes.h>
3 /* For the CLR_() macros */
4 #include <pthread.h>
5 
6 #include <sched.h>
7 #include "evlist.h"
8 #include "evsel.h"
9 #include "perf.h"
10 #include "debug.h"
11 #include "tests.h"
12 
13 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
14 {
15 	int i, cpu = -1, nrcpus = 1024;
16 realloc:
17 	CPU_ZERO(maskp);
18 
19 	if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
20 		if (errno == EINVAL && nrcpus < (1024 << 8)) {
21 			nrcpus = nrcpus << 2;
22 			goto realloc;
23 		}
24 		perror("sched_getaffinity");
25 			return -1;
26 	}
27 
28 	for (i = 0; i < nrcpus; i++) {
29 		if (CPU_ISSET(i, maskp)) {
30 			if (cpu == -1)
31 				cpu = i;
32 			else
33 				CPU_CLR(i, maskp);
34 		}
35 	}
36 
37 	return cpu;
38 }
39 
40 int test__PERF_RECORD(int subtest __maybe_unused)
41 {
42 	struct record_opts opts = {
43 		.target = {
44 			.uid = UINT_MAX,
45 			.uses_mmap = true,
46 		},
47 		.no_buffering = true,
48 		.mmap_pages   = 256,
49 	};
50 	cpu_set_t cpu_mask;
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;
58 	u64 prev_time = 0;
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;
64 	u32 cpu;
65 	int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
66 	char sbuf[STRERR_BUFSIZE];
67 
68 	if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
69 		evlist = perf_evlist__new_default();
70 
71 	if (evlist == NULL) {
72 		pr_debug("Not enough memory to create evlist\n");
73 		goto out;
74 	}
75 
76 	/*
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.
81 	 */
82 	err = perf_evlist__create_maps(evlist, &opts.target);
83 	if (err < 0) {
84 		pr_debug("Not enough memory to create thread/cpu maps\n");
85 		goto out_delete_evlist;
86 	}
87 
88 	/*
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.
93 	 */
94 	err = perf_evlist__prepare_workload(evlist, &opts.target, argv, false, NULL);
95 	if (err < 0) {
96 		pr_debug("Couldn't run the workload!\n");
97 		goto out_delete_evlist;
98 	}
99 
100 	/*
101 	 * Config the evsels, setting attr->comm on the first one, etc.
102 	 */
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);
108 
109 	err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
110 	if (err < 0) {
111 		pr_debug("sched__get_first_possible_cpu: %s\n",
112 			 str_error_r(errno, sbuf, sizeof(sbuf)));
113 		goto out_delete_evlist;
114 	}
115 
116 	cpu = err;
117 
118 	/*
119 	 * So that we can check perf_sample.cpu on all the samples.
120 	 */
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;
125 	}
126 
127 	/*
128 	 * Call sys_perf_event_open on all the fds on all the evsels,
129 	 * grouping them if asked to.
130 	 */
131 	err = perf_evlist__open(evlist);
132 	if (err < 0) {
133 		pr_debug("perf_evlist__open: %s\n",
134 			 str_error_r(errno, sbuf, sizeof(sbuf)));
135 		goto out_delete_evlist;
136 	}
137 
138 	/*
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)).
142 	 */
143 	err = perf_evlist__mmap(evlist, opts.mmap_pages, false);
144 	if (err < 0) {
145 		pr_debug("perf_evlist__mmap: %s\n",
146 			 str_error_r(errno, sbuf, sizeof(sbuf)));
147 		goto out_delete_evlist;
148 	}
149 
150 	/*
151 	 * Now that all is properly set up, enable the events, they will
152 	 * count just on workload.pid, which will start...
153 	 */
154 	perf_evlist__enable(evlist);
155 
156 	/*
157 	 * Now!
158 	 */
159 	perf_evlist__start_workload(evlist);
160 
161 	while (1) {
162 		int before = total_events;
163 
164 		for (i = 0; i < evlist->nr_mmaps; i++) {
165 			union perf_event *event;
166 
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);
170 
171 				++total_events;
172 				if (type < PERF_RECORD_MAX)
173 					nr_events[type]++;
174 
175 				err = perf_evlist__parse_sample(evlist, event, &sample);
176 				if (err < 0) {
177 					if (verbose > 0)
178 						perf_event__fprintf(event, stderr);
179 					pr_debug("Couldn't parse sample\n");
180 					goto out_delete_evlist;
181 				}
182 
183 				if (verbose > 0) {
184 					pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
185 					perf_event__fprintf(event, stderr);
186 				}
187 
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);
191 					++errs;
192 				}
193 
194 				prev_time = sample.time;
195 
196 				if (sample.cpu != cpu) {
197 					pr_debug("%s with unexpected cpu, expected %d, got %d\n",
198 						 name, cpu, sample.cpu);
199 					++errs;
200 				}
201 
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);
205 					++errs;
206 				}
207 
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);
211 					++errs;
212 				}
213 
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);
221 					++errs;
222 				}
223 
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);
229 					++errs;
230 				}
231 
232 				switch (type) {
233 				case PERF_RECORD_COMM:
234 					if (strcmp(event->comm.comm, cmd)) {
235 						pr_debug("%s with unexpected comm!\n", name);
236 						++errs;
237 					}
238 					break;
239 				case PERF_RECORD_EXIT:
240 					goto found_exit;
241 				case PERF_RECORD_MMAP:
242 					mmap_filename = event->mmap.filename;
243 					goto check_bname;
244 				case PERF_RECORD_MMAP2:
245 					mmap_filename = event->mmap2.filename;
246 				check_bname:
247 					bname = strrchr(mmap_filename, '/');
248 					if (bname != NULL) {
249 						if (!found_cmd_mmap)
250 							found_cmd_mmap = !strcmp(bname + 1, cmd);
251 						if (!found_libc_mmap)
252 							found_libc_mmap = !strncmp(bname + 1, "libc", 4);
253 						if (!found_ld_mmap)
254 							found_ld_mmap = !strncmp(bname + 1, "ld", 2);
255 					} else if (!found_vdso_mmap)
256 						found_vdso_mmap = !strcmp(mmap_filename, "[vdso]");
257 					break;
258 
259 				case PERF_RECORD_SAMPLE:
260 					/* Just ignore samples for now */
261 					break;
262 				default:
263 					pr_debug("Unexpected perf_event->header.type %d!\n",
264 						 type);
265 					++errs;
266 				}
267 
268 				perf_evlist__mmap_consume(evlist, i);
269 			}
270 		}
271 
272 		/*
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.
276 		 */
277 		if (total_events == before && false)
278 			perf_evlist__poll(evlist, -1);
279 
280 		sleep(1);
281 		if (++wakeups > 5) {
282 			pr_debug("No PERF_RECORD_EXIT event!\n");
283 			break;
284 		}
285 	}
286 
287 found_exit:
288 	if (nr_events[PERF_RECORD_COMM] > 1) {
289 		pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
290 		++errs;
291 	}
292 
293 	if (nr_events[PERF_RECORD_COMM] == 0) {
294 		pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
295 		++errs;
296 	}
297 
298 	if (!found_cmd_mmap) {
299 		pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
300 		++errs;
301 	}
302 
303 	if (!found_libc_mmap) {
304 		pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
305 		++errs;
306 	}
307 
308 	if (!found_ld_mmap) {
309 		pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
310 		++errs;
311 	}
312 
313 	if (!found_vdso_mmap) {
314 		pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
315 		++errs;
316 	}
317 out_delete_evlist:
318 	perf_evlist__delete(evlist);
319 out:
320 	return (err < 0 || errs > 0) ? -1 : 0;
321 }
322