1 /* For the CLR_() macros */
11 static int sched__get_first_possible_cpu(pid_t pid
, cpu_set_t
*maskp
)
13 int i
, cpu
= -1, nrcpus
= 1024;
17 if (sched_getaffinity(pid
, sizeof(*maskp
), maskp
) == -1) {
18 if (errno
== EINVAL
&& nrcpus
< (1024 << 8)) {
22 perror("sched_getaffinity");
26 for (i
= 0; i
< nrcpus
; i
++) {
27 if (CPU_ISSET(i
, maskp
)) {
38 int test__PERF_RECORD(int subtest __maybe_unused
)
40 struct record_opts opts
= {
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
;
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;
63 int total_events
= 0, nr_events
[PERF_RECORD_MAX
] = { 0, };
64 char sbuf
[STRERR_BUFSIZE
];
66 if (evlist
== NULL
) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
67 evlist
= perf_evlist__new_default();
69 if (evlist
== NULL
|| argv
== NULL
) {
70 pr_debug("Not enough memory to create evlist\n");
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.
80 err
= perf_evlist__create_maps(evlist
, &opts
.target
);
82 pr_debug("Not enough memory to create thread/cpu maps\n");
83 goto out_delete_evlist
;
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.
92 err
= perf_evlist__prepare_workload(evlist
, &opts
.target
, argv
, false, NULL
);
94 pr_debug("Couldn't run the workload!\n");
95 goto out_delete_evlist
;
99 * Config the evsels, setting attr->comm on the first one, etc.
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
);
107 err
= sched__get_first_possible_cpu(evlist
->workload
.pid
, &cpu_mask
);
109 pr_debug("sched__get_first_possible_cpu: %s\n",
110 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
111 goto out_delete_evlist
;
117 * So that we can check perf_sample.cpu on all the samples.
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
;
126 * Call sys_perf_event_open on all the fds on all the evsels,
127 * grouping them if asked to.
129 err
= perf_evlist__open(evlist
);
131 pr_debug("perf_evlist__open: %s\n",
132 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
133 goto out_delete_evlist
;
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)).
141 err
= perf_evlist__mmap(evlist
, opts
.mmap_pages
, false);
143 pr_debug("perf_evlist__mmap: %s\n",
144 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
145 goto out_delete_evlist
;
149 * Now that all is properly set up, enable the events, they will
150 * count just on workload.pid, which will start...
152 perf_evlist__enable(evlist
);
157 perf_evlist__start_workload(evlist
);
160 int before
= total_events
;
162 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
163 union perf_event
*event
;
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
);
170 if (type
< PERF_RECORD_MAX
)
173 err
= perf_evlist__parse_sample(evlist
, event
, &sample
);
176 perf_event__fprintf(event
, stderr
);
177 pr_debug("Couldn't parse sample\n");
178 goto out_delete_evlist
;
182 pr_info("%" PRIu64
" %d ", sample
.time
, sample
.cpu
);
183 perf_event__fprintf(event
, stderr
);
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
);
192 prev_time
= sample
.time
;
194 if (sample
.cpu
!= cpu
) {
195 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
196 name
, cpu
, sample
.cpu
);
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
);
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
);
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
);
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
);
231 case PERF_RECORD_COMM
:
232 if (strcmp(event
->comm
.comm
, cmd
)) {
233 pr_debug("%s with unexpected comm!\n", name
);
237 case PERF_RECORD_EXIT
:
239 case PERF_RECORD_MMAP
:
240 mmap_filename
= event
->mmap
.filename
;
242 case PERF_RECORD_MMAP2
:
243 mmap_filename
= event
->mmap2
.filename
;
245 bname
= strrchr(mmap_filename
, '/');
248 found_cmd_mmap
= !strcmp(bname
+ 1, cmd
);
249 if (!found_libc_mmap
)
250 found_libc_mmap
= !strncmp(bname
+ 1, "libc", 4);
252 found_ld_mmap
= !strncmp(bname
+ 1, "ld", 2);
253 } else if (!found_vdso_mmap
)
254 found_vdso_mmap
= !strcmp(mmap_filename
, "[vdso]");
257 case PERF_RECORD_SAMPLE
:
258 /* Just ignore samples for now */
261 pr_debug("Unexpected perf_event->header.type %d!\n",
266 perf_evlist__mmap_consume(evlist
, i
);
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.
275 if (total_events
== before
&& false)
276 perf_evlist__poll(evlist
, -1);
280 pr_debug("No PERF_RECORD_EXIT event!\n");
286 if (nr_events
[PERF_RECORD_COMM
] > 1) {
287 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
291 if (nr_events
[PERF_RECORD_COMM
] == 0) {
292 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd
);
296 if (!found_cmd_mmap
) {
297 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd
);
301 if (!found_libc_mmap
) {
302 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
306 if (!found_ld_mmap
) {
307 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
311 if (!found_vdso_mmap
) {
312 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
316 perf_evlist__delete(evlist
);
318 return (err
< 0 || errs
> 0) ? -1 : 0;