8 static int sched__get_first_possible_cpu(pid_t pid
, cpu_set_t
*maskp
)
10 int i
, cpu
= -1, nrcpus
= 1024;
14 if (sched_getaffinity(pid
, sizeof(*maskp
), maskp
) == -1) {
15 if (errno
== EINVAL
&& nrcpus
< (1024 << 8)) {
19 perror("sched_getaffinity");
23 for (i
= 0; i
< nrcpus
; i
++) {
24 if (CPU_ISSET(i
, maskp
)) {
35 int test__PERF_RECORD(int subtest __maybe_unused
)
37 struct record_opts opts
= {
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
;
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;
60 int total_events
= 0, nr_events
[PERF_RECORD_MAX
] = { 0, };
61 char sbuf
[STRERR_BUFSIZE
];
63 if (evlist
== NULL
) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
64 evlist
= perf_evlist__new_default();
66 if (evlist
== NULL
|| argv
== NULL
) {
67 pr_debug("Not enough memory to create evlist\n");
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.
77 err
= perf_evlist__create_maps(evlist
, &opts
.target
);
79 pr_debug("Not enough memory to create thread/cpu maps\n");
80 goto out_delete_evlist
;
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.
89 err
= perf_evlist__prepare_workload(evlist
, &opts
.target
, argv
, false, NULL
);
91 pr_debug("Couldn't run the workload!\n");
92 goto out_delete_evlist
;
96 * Config the evsels, setting attr->comm on the first one, etc.
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
);
104 err
= sched__get_first_possible_cpu(evlist
->workload
.pid
, &cpu_mask
);
106 pr_debug("sched__get_first_possible_cpu: %s\n",
107 strerror_r(errno
, sbuf
, sizeof(sbuf
)));
108 goto out_delete_evlist
;
114 * So that we can check perf_sample.cpu on all the samples.
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
;
123 * Call sys_perf_event_open on all the fds on all the evsels,
124 * grouping them if asked to.
126 err
= perf_evlist__open(evlist
);
128 pr_debug("perf_evlist__open: %s\n",
129 strerror_r(errno
, sbuf
, sizeof(sbuf
)));
130 goto out_delete_evlist
;
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)).
138 err
= perf_evlist__mmap(evlist
, opts
.mmap_pages
, false);
140 pr_debug("perf_evlist__mmap: %s\n",
141 strerror_r(errno
, sbuf
, sizeof(sbuf
)));
142 goto out_delete_evlist
;
146 * Now that all is properly set up, enable the events, they will
147 * count just on workload.pid, which will start...
149 perf_evlist__enable(evlist
);
154 perf_evlist__start_workload(evlist
);
157 int before
= total_events
;
159 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
160 union perf_event
*event
;
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
);
167 if (type
< PERF_RECORD_MAX
)
170 err
= perf_evlist__parse_sample(evlist
, event
, &sample
);
173 perf_event__fprintf(event
, stderr
);
174 pr_debug("Couldn't parse sample\n");
175 goto out_delete_evlist
;
179 pr_info("%" PRIu64
" %d ", sample
.time
, sample
.cpu
);
180 perf_event__fprintf(event
, stderr
);
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
);
189 prev_time
= sample
.time
;
191 if (sample
.cpu
!= cpu
) {
192 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
193 name
, cpu
, sample
.cpu
);
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
);
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
);
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
);
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
);
228 case PERF_RECORD_COMM
:
229 if (strcmp(event
->comm
.comm
, cmd
)) {
230 pr_debug("%s with unexpected comm!\n", name
);
234 case PERF_RECORD_EXIT
:
236 case PERF_RECORD_MMAP
:
237 mmap_filename
= event
->mmap
.filename
;
239 case PERF_RECORD_MMAP2
:
240 mmap_filename
= event
->mmap2
.filename
;
242 bname
= strrchr(mmap_filename
, '/');
245 found_cmd_mmap
= !strcmp(bname
+ 1, cmd
);
246 if (!found_libc_mmap
)
247 found_libc_mmap
= !strncmp(bname
+ 1, "libc", 4);
249 found_ld_mmap
= !strncmp(bname
+ 1, "ld", 2);
250 } else if (!found_vdso_mmap
)
251 found_vdso_mmap
= !strcmp(mmap_filename
, "[vdso]");
254 case PERF_RECORD_SAMPLE
:
255 /* Just ignore samples for now */
258 pr_debug("Unexpected perf_event->header.type %d!\n",
263 perf_evlist__mmap_consume(evlist
, i
);
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.
272 if (total_events
== before
&& false)
273 perf_evlist__poll(evlist
, -1);
277 pr_debug("No PERF_RECORD_EXIT event!\n");
283 if (nr_events
[PERF_RECORD_COMM
] > 1) {
284 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
288 if (nr_events
[PERF_RECORD_COMM
] == 0) {
289 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd
);
293 if (!found_cmd_mmap
) {
294 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd
);
298 if (!found_libc_mmap
) {
299 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
303 if (!found_ld_mmap
) {
304 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
308 if (!found_vdso_mmap
) {
309 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
313 perf_evlist__delete(evlist
);
315 return (err
< 0 || errs
> 0) ? -1 : 0;