1 // SPDX-License-Identifier: GPL-2.0
4 /* For the CLR_() macros */
14 static int sched__get_first_possible_cpu(pid_t pid
, cpu_set_t
*maskp
)
16 int i
, cpu
= -1, nrcpus
= 1024;
20 if (sched_getaffinity(pid
, sizeof(*maskp
), maskp
) == -1) {
21 if (errno
== EINVAL
&& nrcpus
< (1024 << 8)) {
25 perror("sched_getaffinity");
29 for (i
= 0; i
< nrcpus
; i
++) {
30 if (CPU_ISSET(i
, maskp
)) {
41 int test__PERF_RECORD(struct test
*test __maybe_unused
, int subtest __maybe_unused
)
43 struct record_opts opts
= {
52 size_t cpu_mask_size
= sizeof(cpu_mask
);
53 struct perf_evlist
*evlist
= perf_evlist__new_dummy();
54 struct perf_evsel
*evsel
;
55 struct perf_sample sample
;
56 const char *cmd
= "sleep";
57 const char *argv
[] = { cmd
, "1", NULL
, };
58 char *bname
, *mmap_filename
;
60 bool found_cmd_mmap
= false,
61 found_libc_mmap
= false,
62 found_vdso_mmap
= false,
63 found_ld_mmap
= false;
64 int err
= -1, errs
= 0, i
, wakeups
= 0;
66 int total_events
= 0, nr_events
[PERF_RECORD_MAX
] = { 0, };
67 char sbuf
[STRERR_BUFSIZE
];
69 if (evlist
== NULL
) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
70 evlist
= perf_evlist__new_default();
73 pr_debug("Not enough memory to create evlist\n");
78 * Create maps of threads and cpus to monitor. In this case
79 * we start with all threads and cpus (-1, -1) but then in
80 * perf_evlist__prepare_workload we'll fill in the only thread
81 * we're monitoring, the one forked there.
83 err
= perf_evlist__create_maps(evlist
, &opts
.target
);
85 pr_debug("Not enough memory to create thread/cpu maps\n");
86 goto out_delete_evlist
;
90 * Prepare the workload in argv[] to run, it'll fork it, and then wait
91 * for perf_evlist__start_workload() to exec it. This is done this way
92 * so that we have time to open the evlist (calling sys_perf_event_open
93 * on all the fds) and then mmap them.
95 err
= perf_evlist__prepare_workload(evlist
, &opts
.target
, argv
, false, NULL
);
97 pr_debug("Couldn't run the workload!\n");
98 goto out_delete_evlist
;
102 * Config the evsels, setting attr->comm on the first one, etc.
104 evsel
= perf_evlist__first(evlist
);
105 perf_evsel__set_sample_bit(evsel
, CPU
);
106 perf_evsel__set_sample_bit(evsel
, TID
);
107 perf_evsel__set_sample_bit(evsel
, TIME
);
108 perf_evlist__config(evlist
, &opts
, NULL
);
110 err
= sched__get_first_possible_cpu(evlist
->workload
.pid
, &cpu_mask
);
112 pr_debug("sched__get_first_possible_cpu: %s\n",
113 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
114 goto out_delete_evlist
;
120 * So that we can check perf_sample.cpu on all the samples.
122 if (sched_setaffinity(evlist
->workload
.pid
, cpu_mask_size
, &cpu_mask
) < 0) {
123 pr_debug("sched_setaffinity: %s\n",
124 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
125 goto out_delete_evlist
;
129 * Call sys_perf_event_open on all the fds on all the evsels,
130 * grouping them if asked to.
132 err
= perf_evlist__open(evlist
);
134 pr_debug("perf_evlist__open: %s\n",
135 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
136 goto out_delete_evlist
;
140 * mmap the first fd on a given CPU and ask for events for the other
141 * fds in the same CPU to be injected in the same mmap ring buffer
142 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
144 err
= perf_evlist__mmap(evlist
, opts
.mmap_pages
);
146 pr_debug("perf_evlist__mmap: %s\n",
147 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
148 goto out_delete_evlist
;
152 * Now that all is properly set up, enable the events, they will
153 * count just on workload.pid, which will start...
155 perf_evlist__enable(evlist
);
160 perf_evlist__start_workload(evlist
);
163 int before
= total_events
;
165 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
166 union perf_event
*event
;
168 while ((event
= perf_evlist__mmap_read(evlist
, i
)) != NULL
) {
169 const u32 type
= event
->header
.type
;
170 const char *name
= perf_event__name(type
);
173 if (type
< PERF_RECORD_MAX
)
176 err
= perf_evlist__parse_sample(evlist
, event
, &sample
);
179 perf_event__fprintf(event
, stderr
);
180 pr_debug("Couldn't parse sample\n");
181 goto out_delete_evlist
;
185 pr_info("%" PRIu64
" %d ", sample
.time
, sample
.cpu
);
186 perf_event__fprintf(event
, stderr
);
189 if (prev_time
> sample
.time
) {
190 pr_debug("%s going backwards in time, prev=%" PRIu64
", curr=%" PRIu64
"\n",
191 name
, prev_time
, sample
.time
);
195 prev_time
= sample
.time
;
197 if (sample
.cpu
!= cpu
) {
198 pr_debug("%s with unexpected cpu, expected %d, got %d\n",
199 name
, cpu
, sample
.cpu
);
203 if ((pid_t
)sample
.pid
!= evlist
->workload
.pid
) {
204 pr_debug("%s with unexpected pid, expected %d, got %d\n",
205 name
, evlist
->workload
.pid
, sample
.pid
);
209 if ((pid_t
)sample
.tid
!= evlist
->workload
.pid
) {
210 pr_debug("%s with unexpected tid, expected %d, got %d\n",
211 name
, evlist
->workload
.pid
, sample
.tid
);
215 if ((type
== PERF_RECORD_COMM
||
216 type
== PERF_RECORD_MMAP
||
217 type
== PERF_RECORD_MMAP2
||
218 type
== PERF_RECORD_FORK
||
219 type
== PERF_RECORD_EXIT
) &&
220 (pid_t
)event
->comm
.pid
!= evlist
->workload
.pid
) {
221 pr_debug("%s with unexpected pid/tid\n", name
);
225 if ((type
== PERF_RECORD_COMM
||
226 type
== PERF_RECORD_MMAP
||
227 type
== PERF_RECORD_MMAP2
) &&
228 event
->comm
.pid
!= event
->comm
.tid
) {
229 pr_debug("%s with different pid/tid!\n", name
);
234 case PERF_RECORD_COMM
:
235 if (strcmp(event
->comm
.comm
, cmd
)) {
236 pr_debug("%s with unexpected comm!\n", name
);
240 case PERF_RECORD_EXIT
:
242 case PERF_RECORD_MMAP
:
243 mmap_filename
= event
->mmap
.filename
;
245 case PERF_RECORD_MMAP2
:
246 mmap_filename
= event
->mmap2
.filename
;
248 bname
= strrchr(mmap_filename
, '/');
251 found_cmd_mmap
= !strcmp(bname
+ 1, cmd
);
252 if (!found_libc_mmap
)
253 found_libc_mmap
= !strncmp(bname
+ 1, "libc", 4);
255 found_ld_mmap
= !strncmp(bname
+ 1, "ld", 2);
256 } else if (!found_vdso_mmap
)
257 found_vdso_mmap
= !strcmp(mmap_filename
, "[vdso]");
260 case PERF_RECORD_SAMPLE
:
261 /* Just ignore samples for now */
264 pr_debug("Unexpected perf_event->header.type %d!\n",
269 perf_evlist__mmap_consume(evlist
, i
);
274 * We don't use poll here because at least at 3.1 times the
275 * PERF_RECORD_{!SAMPLE} events don't honour
276 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
278 if (total_events
== before
&& false)
279 perf_evlist__poll(evlist
, -1);
283 pr_debug("No PERF_RECORD_EXIT event!\n");
289 if (nr_events
[PERF_RECORD_COMM
] > 1) {
290 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
294 if (nr_events
[PERF_RECORD_COMM
] == 0) {
295 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd
);
299 if (!found_cmd_mmap
) {
300 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd
);
304 if (!found_libc_mmap
) {
305 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
309 if (!found_ld_mmap
) {
310 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
314 if (!found_vdso_mmap
) {
315 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
319 perf_evlist__delete(evlist
);
321 return (err
< 0 || errs
> 0) ? -1 : 0;