Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include <errno.h>
3#include <inttypes.h>
4#include <linux/string.h>
5
6#include <sched.h>
7#include <perf/mmap.h>
8#include "event.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#include "util/sample.h"
16
17static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
18{
19 int i, cpu = -1, nrcpus = 1024;
20realloc:
21 CPU_ZERO(maskp);
22
23 if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
24 if (errno == EINVAL && nrcpus < (1024 << 8)) {
25 nrcpus = nrcpus << 2;
26 goto realloc;
27 }
28 perror("sched_getaffinity");
29 return -1;
30 }
31
32 for (i = 0; i < nrcpus; i++) {
33 if (CPU_ISSET(i, maskp)) {
34 if (cpu == -1)
35 cpu = i;
36 else
37 CPU_CLR(i, maskp);
38 }
39 }
40
41 return cpu;
42}
43
44static int test__PERF_RECORD(struct test_suite *test __maybe_unused, int subtest __maybe_unused)
45{
46 struct record_opts opts = {
47 .target = {
48 .uid = UINT_MAX,
49 .uses_mmap = true,
50 },
51 .no_buffering = true,
52 .mmap_pages = 256,
53 };
54 cpu_set_t cpu_mask;
55 size_t cpu_mask_size = sizeof(cpu_mask);
56 struct evlist *evlist = evlist__new_dummy();
57 struct evsel *evsel;
58 struct perf_sample sample;
59 const char *cmd = "sleep";
60 const char *argv[] = { cmd, "1", NULL, };
61 char *bname, *mmap_filename;
62 u64 prev_time = 0;
63 bool found_cmd_mmap = false,
64 found_coreutils_mmap = false,
65 found_libc_mmap = false,
66 found_vdso_mmap = false,
67 found_ld_mmap = false;
68 int err = -1, errs = 0, i, wakeups = 0;
69 u32 cpu;
70 int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
71 char sbuf[STRERR_BUFSIZE];
72
73 if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
74 evlist = evlist__new_default();
75
76 if (evlist == NULL) {
77 pr_debug("Not enough memory to create evlist\n");
78 goto out;
79 }
80
81 /*
82 * Create maps of threads and cpus to monitor. In this case
83 * we start with all threads and cpus (-1, -1) but then in
84 * evlist__prepare_workload we'll fill in the only thread
85 * we're monitoring, the one forked there.
86 */
87 err = evlist__create_maps(evlist, &opts.target);
88 if (err < 0) {
89 pr_debug("Not enough memory to create thread/cpu maps\n");
90 goto out_delete_evlist;
91 }
92
93 /*
94 * Prepare the workload in argv[] to run, it'll fork it, and then wait
95 * for evlist__start_workload() to exec it. This is done this way
96 * so that we have time to open the evlist (calling sys_perf_event_open
97 * on all the fds) and then mmap them.
98 */
99 err = evlist__prepare_workload(evlist, &opts.target, argv, false, NULL);
100 if (err < 0) {
101 pr_debug("Couldn't run the workload!\n");
102 goto out_delete_evlist;
103 }
104
105 /*
106 * Config the evsels, setting attr->comm on the first one, etc.
107 */
108 evsel = evlist__first(evlist);
109 evsel__set_sample_bit(evsel, CPU);
110 evsel__set_sample_bit(evsel, TID);
111 evsel__set_sample_bit(evsel, TIME);
112 evlist__config(evlist, &opts, NULL);
113
114 err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
115 if (err < 0) {
116 pr_debug("sched__get_first_possible_cpu: %s\n",
117 str_error_r(errno, sbuf, sizeof(sbuf)));
118 goto out_delete_evlist;
119 }
120
121 cpu = err;
122
123 /*
124 * So that we can check perf_sample.cpu on all the samples.
125 */
126 if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) {
127 pr_debug("sched_setaffinity: %s\n",
128 str_error_r(errno, sbuf, sizeof(sbuf)));
129 goto out_delete_evlist;
130 }
131
132 /*
133 * Call sys_perf_event_open on all the fds on all the evsels,
134 * grouping them if asked to.
135 */
136 err = evlist__open(evlist);
137 if (err < 0) {
138 pr_debug("perf_evlist__open: %s\n",
139 str_error_r(errno, sbuf, sizeof(sbuf)));
140 goto out_delete_evlist;
141 }
142
143 /*
144 * mmap the first fd on a given CPU and ask for events for the other
145 * fds in the same CPU to be injected in the same mmap ring buffer
146 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
147 */
148 err = evlist__mmap(evlist, opts.mmap_pages);
149 if (err < 0) {
150 pr_debug("evlist__mmap: %s\n",
151 str_error_r(errno, sbuf, sizeof(sbuf)));
152 goto out_delete_evlist;
153 }
154
155 /*
156 * Now that all is properly set up, enable the events, they will
157 * count just on workload.pid, which will start...
158 */
159 evlist__enable(evlist);
160
161 /*
162 * Now!
163 */
164 evlist__start_workload(evlist);
165
166 while (1) {
167 int before = total_events;
168
169 for (i = 0; i < evlist->core.nr_mmaps; i++) {
170 union perf_event *event;
171 struct mmap *md;
172
173 md = &evlist->mmap[i];
174 if (perf_mmap__read_init(&md->core) < 0)
175 continue;
176
177 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
178 const u32 type = event->header.type;
179 const char *name = perf_event__name(type);
180
181 ++total_events;
182 if (type < PERF_RECORD_MAX)
183 nr_events[type]++;
184
185 err = evlist__parse_sample(evlist, event, &sample);
186 if (err < 0) {
187 if (verbose > 0)
188 perf_event__fprintf(event, NULL, stderr);
189 pr_debug("Couldn't parse sample\n");
190 goto out_delete_evlist;
191 }
192
193 if (verbose > 0) {
194 pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
195 perf_event__fprintf(event, NULL, stderr);
196 }
197
198 if (prev_time > sample.time) {
199 pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
200 name, prev_time, sample.time);
201 ++errs;
202 }
203
204 prev_time = sample.time;
205
206 if (sample.cpu != cpu) {
207 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
208 name, cpu, sample.cpu);
209 ++errs;
210 }
211
212 if ((pid_t)sample.pid != evlist->workload.pid) {
213 pr_debug("%s with unexpected pid, expected %d, got %d\n",
214 name, evlist->workload.pid, sample.pid);
215 ++errs;
216 }
217
218 if ((pid_t)sample.tid != evlist->workload.pid) {
219 pr_debug("%s with unexpected tid, expected %d, got %d\n",
220 name, evlist->workload.pid, sample.tid);
221 ++errs;
222 }
223
224 if ((type == PERF_RECORD_COMM ||
225 type == PERF_RECORD_MMAP ||
226 type == PERF_RECORD_MMAP2 ||
227 type == PERF_RECORD_FORK ||
228 type == PERF_RECORD_EXIT) &&
229 (pid_t)event->comm.pid != evlist->workload.pid) {
230 pr_debug("%s with unexpected pid/tid\n", name);
231 ++errs;
232 }
233
234 if ((type == PERF_RECORD_COMM ||
235 type == PERF_RECORD_MMAP ||
236 type == PERF_RECORD_MMAP2) &&
237 event->comm.pid != event->comm.tid) {
238 pr_debug("%s with different pid/tid!\n", name);
239 ++errs;
240 }
241
242 switch (type) {
243 case PERF_RECORD_COMM:
244 if (strcmp(event->comm.comm, cmd)) {
245 pr_debug("%s with unexpected comm!\n", name);
246 ++errs;
247 }
248 break;
249 case PERF_RECORD_EXIT:
250 goto found_exit;
251 case PERF_RECORD_MMAP:
252 mmap_filename = event->mmap.filename;
253 goto check_bname;
254 case PERF_RECORD_MMAP2:
255 mmap_filename = event->mmap2.filename;
256 check_bname:
257 bname = strrchr(mmap_filename, '/');
258 if (bname != NULL) {
259 if (!found_cmd_mmap)
260 found_cmd_mmap = !strcmp(bname + 1, cmd);
261 if (!found_coreutils_mmap)
262 found_coreutils_mmap = !strcmp(bname + 1, "coreutils");
263 if (!found_libc_mmap)
264 found_libc_mmap = !strncmp(bname + 1, "libc", 4);
265 if (!found_ld_mmap)
266 found_ld_mmap = !strncmp(bname + 1, "ld", 2);
267 } else if (!found_vdso_mmap)
268 found_vdso_mmap = !strcmp(mmap_filename, "[vdso]");
269 break;
270
271 case PERF_RECORD_SAMPLE:
272 /* Just ignore samples for now */
273 break;
274 default:
275 pr_debug("Unexpected perf_event->header.type %d!\n",
276 type);
277 ++errs;
278 }
279
280 perf_mmap__consume(&md->core);
281 }
282 perf_mmap__read_done(&md->core);
283 }
284
285 /*
286 * We don't use poll here because at least at 3.1 times the
287 * PERF_RECORD_{!SAMPLE} events don't honour
288 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
289 */
290 if (total_events == before && false)
291 evlist__poll(evlist, -1);
292
293 sleep(1);
294 if (++wakeups > 5) {
295 pr_debug("No PERF_RECORD_EXIT event!\n");
296 break;
297 }
298 }
299
300found_exit:
301 if (nr_events[PERF_RECORD_COMM] > 1 + !!found_coreutils_mmap) {
302 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
303 ++errs;
304 }
305
306 if (nr_events[PERF_RECORD_COMM] == 0) {
307 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
308 ++errs;
309 }
310
311 if (!found_cmd_mmap && !found_coreutils_mmap) {
312 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
313 ++errs;
314 }
315
316 if (!found_libc_mmap) {
317 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
318 ++errs;
319 }
320
321 if (!found_ld_mmap) {
322 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
323 ++errs;
324 }
325
326 if (!found_vdso_mmap) {
327 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
328 ++errs;
329 }
330out_delete_evlist:
331 evlist__delete(evlist);
332out:
333 if (err == -EACCES)
334 return TEST_SKIP;
335 if (err < 0 || errs != 0)
336 return TEST_FAIL;
337 return TEST_OK;
338}
339
340static struct test_case tests__PERF_RECORD[] = {
341 TEST_CASE_REASON("PERF_RECORD_* events & perf_sample fields",
342 PERF_RECORD,
343 "permissions"),
344 { .name = NULL, }
345};
346
347struct test_suite suite__PERF_RECORD = {
348 .desc = "PERF_RECORD_* events & perf_sample fields",
349 .test_cases = tests__PERF_RECORD,
350};
1/* For the CLR_() macros */
2#include <pthread.h>
3
4#include <sched.h>
5#include "evlist.h"
6#include "evsel.h"
7#include "perf.h"
8#include "debug.h"
9#include "tests.h"
10
11static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
12{
13 int i, cpu = -1, nrcpus = 1024;
14realloc:
15 CPU_ZERO(maskp);
16
17 if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
18 if (errno == EINVAL && nrcpus < (1024 << 8)) {
19 nrcpus = nrcpus << 2;
20 goto realloc;
21 }
22 perror("sched_getaffinity");
23 return -1;
24 }
25
26 for (i = 0; i < nrcpus; i++) {
27 if (CPU_ISSET(i, maskp)) {
28 if (cpu == -1)
29 cpu = i;
30 else
31 CPU_CLR(i, maskp);
32 }
33 }
34
35 return cpu;
36}
37
38int test__PERF_RECORD(int subtest __maybe_unused)
39{
40 struct record_opts opts = {
41 .target = {
42 .uid = UINT_MAX,
43 .uses_mmap = true,
44 },
45 .no_buffering = true,
46 .mmap_pages = 256,
47 };
48 cpu_set_t cpu_mask;
49 size_t cpu_mask_size = sizeof(cpu_mask);
50 struct perf_evlist *evlist = perf_evlist__new_dummy();
51 struct perf_evsel *evsel;
52 struct perf_sample sample;
53 const char *cmd = "sleep";
54 const char *argv[] = { cmd, "1", NULL, };
55 char *bname, *mmap_filename;
56 u64 prev_time = 0;
57 bool found_cmd_mmap = false,
58 found_libc_mmap = false,
59 found_vdso_mmap = false,
60 found_ld_mmap = false;
61 int err = -1, errs = 0, i, wakeups = 0;
62 u32 cpu;
63 int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
64 char sbuf[STRERR_BUFSIZE];
65
66 if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
67 evlist = perf_evlist__new_default();
68
69 if (evlist == NULL || argv == NULL) {
70 pr_debug("Not enough memory to create evlist\n");
71 goto out;
72 }
73
74 /*
75 * Create maps of threads and cpus to monitor. In this case
76 * we start with all threads and cpus (-1, -1) but then in
77 * perf_evlist__prepare_workload we'll fill in the only thread
78 * we're monitoring, the one forked there.
79 */
80 err = perf_evlist__create_maps(evlist, &opts.target);
81 if (err < 0) {
82 pr_debug("Not enough memory to create thread/cpu maps\n");
83 goto out_delete_evlist;
84 }
85
86 /*
87 * Prepare the workload in argv[] to run, it'll fork it, and then wait
88 * for perf_evlist__start_workload() to exec it. This is done this way
89 * so that we have time to open the evlist (calling sys_perf_event_open
90 * on all the fds) and then mmap them.
91 */
92 err = perf_evlist__prepare_workload(evlist, &opts.target, argv, false, NULL);
93 if (err < 0) {
94 pr_debug("Couldn't run the workload!\n");
95 goto out_delete_evlist;
96 }
97
98 /*
99 * Config the evsels, setting attr->comm on the first one, etc.
100 */
101 evsel = perf_evlist__first(evlist);
102 perf_evsel__set_sample_bit(evsel, CPU);
103 perf_evsel__set_sample_bit(evsel, TID);
104 perf_evsel__set_sample_bit(evsel, TIME);
105 perf_evlist__config(evlist, &opts, NULL);
106
107 err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
108 if (err < 0) {
109 pr_debug("sched__get_first_possible_cpu: %s\n",
110 str_error_r(errno, sbuf, sizeof(sbuf)));
111 goto out_delete_evlist;
112 }
113
114 cpu = err;
115
116 /*
117 * So that we can check perf_sample.cpu on all the samples.
118 */
119 if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) {
120 pr_debug("sched_setaffinity: %s\n",
121 str_error_r(errno, sbuf, sizeof(sbuf)));
122 goto out_delete_evlist;
123 }
124
125 /*
126 * Call sys_perf_event_open on all the fds on all the evsels,
127 * grouping them if asked to.
128 */
129 err = perf_evlist__open(evlist);
130 if (err < 0) {
131 pr_debug("perf_evlist__open: %s\n",
132 str_error_r(errno, sbuf, sizeof(sbuf)));
133 goto out_delete_evlist;
134 }
135
136 /*
137 * mmap the first fd on a given CPU and ask for events for the other
138 * fds in the same CPU to be injected in the same mmap ring buffer
139 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
140 */
141 err = perf_evlist__mmap(evlist, opts.mmap_pages, false);
142 if (err < 0) {
143 pr_debug("perf_evlist__mmap: %s\n",
144 str_error_r(errno, sbuf, sizeof(sbuf)));
145 goto out_delete_evlist;
146 }
147
148 /*
149 * Now that all is properly set up, enable the events, they will
150 * count just on workload.pid, which will start...
151 */
152 perf_evlist__enable(evlist);
153
154 /*
155 * Now!
156 */
157 perf_evlist__start_workload(evlist);
158
159 while (1) {
160 int before = total_events;
161
162 for (i = 0; i < evlist->nr_mmaps; i++) {
163 union perf_event *event;
164
165 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
166 const u32 type = event->header.type;
167 const char *name = perf_event__name(type);
168
169 ++total_events;
170 if (type < PERF_RECORD_MAX)
171 nr_events[type]++;
172
173 err = perf_evlist__parse_sample(evlist, event, &sample);
174 if (err < 0) {
175 if (verbose)
176 perf_event__fprintf(event, stderr);
177 pr_debug("Couldn't parse sample\n");
178 goto out_delete_evlist;
179 }
180
181 if (verbose) {
182 pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
183 perf_event__fprintf(event, stderr);
184 }
185
186 if (prev_time > sample.time) {
187 pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
188 name, prev_time, sample.time);
189 ++errs;
190 }
191
192 prev_time = sample.time;
193
194 if (sample.cpu != cpu) {
195 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
196 name, cpu, sample.cpu);
197 ++errs;
198 }
199
200 if ((pid_t)sample.pid != evlist->workload.pid) {
201 pr_debug("%s with unexpected pid, expected %d, got %d\n",
202 name, evlist->workload.pid, sample.pid);
203 ++errs;
204 }
205
206 if ((pid_t)sample.tid != evlist->workload.pid) {
207 pr_debug("%s with unexpected tid, expected %d, got %d\n",
208 name, evlist->workload.pid, sample.tid);
209 ++errs;
210 }
211
212 if ((type == PERF_RECORD_COMM ||
213 type == PERF_RECORD_MMAP ||
214 type == PERF_RECORD_MMAP2 ||
215 type == PERF_RECORD_FORK ||
216 type == PERF_RECORD_EXIT) &&
217 (pid_t)event->comm.pid != evlist->workload.pid) {
218 pr_debug("%s with unexpected pid/tid\n", name);
219 ++errs;
220 }
221
222 if ((type == PERF_RECORD_COMM ||
223 type == PERF_RECORD_MMAP ||
224 type == PERF_RECORD_MMAP2) &&
225 event->comm.pid != event->comm.tid) {
226 pr_debug("%s with different pid/tid!\n", name);
227 ++errs;
228 }
229
230 switch (type) {
231 case PERF_RECORD_COMM:
232 if (strcmp(event->comm.comm, cmd)) {
233 pr_debug("%s with unexpected comm!\n", name);
234 ++errs;
235 }
236 break;
237 case PERF_RECORD_EXIT:
238 goto found_exit;
239 case PERF_RECORD_MMAP:
240 mmap_filename = event->mmap.filename;
241 goto check_bname;
242 case PERF_RECORD_MMAP2:
243 mmap_filename = event->mmap2.filename;
244 check_bname:
245 bname = strrchr(mmap_filename, '/');
246 if (bname != NULL) {
247 if (!found_cmd_mmap)
248 found_cmd_mmap = !strcmp(bname + 1, cmd);
249 if (!found_libc_mmap)
250 found_libc_mmap = !strncmp(bname + 1, "libc", 4);
251 if (!found_ld_mmap)
252 found_ld_mmap = !strncmp(bname + 1, "ld", 2);
253 } else if (!found_vdso_mmap)
254 found_vdso_mmap = !strcmp(mmap_filename, "[vdso]");
255 break;
256
257 case PERF_RECORD_SAMPLE:
258 /* Just ignore samples for now */
259 break;
260 default:
261 pr_debug("Unexpected perf_event->header.type %d!\n",
262 type);
263 ++errs;
264 }
265
266 perf_evlist__mmap_consume(evlist, i);
267 }
268 }
269
270 /*
271 * We don't use poll here because at least at 3.1 times the
272 * PERF_RECORD_{!SAMPLE} events don't honour
273 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
274 */
275 if (total_events == before && false)
276 perf_evlist__poll(evlist, -1);
277
278 sleep(1);
279 if (++wakeups > 5) {
280 pr_debug("No PERF_RECORD_EXIT event!\n");
281 break;
282 }
283 }
284
285found_exit:
286 if (nr_events[PERF_RECORD_COMM] > 1) {
287 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
288 ++errs;
289 }
290
291 if (nr_events[PERF_RECORD_COMM] == 0) {
292 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
293 ++errs;
294 }
295
296 if (!found_cmd_mmap) {
297 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
298 ++errs;
299 }
300
301 if (!found_libc_mmap) {
302 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
303 ++errs;
304 }
305
306 if (!found_ld_mmap) {
307 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
308 ++errs;
309 }
310
311 if (!found_vdso_mmap) {
312 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
313 ++errs;
314 }
315out_delete_evlist:
316 perf_evlist__delete(evlist);
317out:
318 return (err < 0 || errs > 0) ? -1 : 0;
319}