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