Linux Perf
perf-record.c
Go to the documentation of this file.
1 // SPDX-License-Identifier: GPL-2.0
2 #include <errno.h>
3 #include <inttypes.h>
4 /* For the CLR_() macros */
5 #include <pthread.h>
6 
7 #include <sched.h>
8 #include "evlist.h"
9 #include "evsel.h"
10 #include "perf.h"
11 #include "debug.h"
12 #include "tests.h"
13 
14 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
15 {
16  int i, cpu = -1, nrcpus = 1024;
17 realloc:
18  CPU_ZERO(maskp);
19 
20  if (sched_getaffinity(pid, sizeof(*maskp), maskp) == -1) {
21  if (errno == EINVAL && nrcpus < (1024 << 8)) {
22  nrcpus = nrcpus << 2;
23  goto realloc;
24  }
25  perror("sched_getaffinity");
26  return -1;
27  }
28 
29  for (i = 0; i < nrcpus; i++) {
30  if (CPU_ISSET(i, maskp)) {
31  if (cpu == -1)
32  cpu = i;
33  else
34  CPU_CLR(i, maskp);
35  }
36  }
37 
38  return cpu;
39 }
40 
41 int test__PERF_RECORD(struct test *test __maybe_unused, int subtest __maybe_unused)
42 {
43  struct record_opts opts = {
44  .target = {
45  .uid = UINT_MAX,
46  .uses_mmap = true,
47  },
48  .no_buffering = true,
49  .mmap_pages = 256,
50  };
51  cpu_set_t cpu_mask;
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;
59  u64 prev_time = 0;
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;
65  u32 cpu;
66  int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
67  char sbuf[STRERR_BUFSIZE];
68 
69  if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */
70  evlist = perf_evlist__new_default();
71 
72  if (evlist == NULL) {
73  pr_debug("Not enough memory to create evlist\n");
74  goto out;
75  }
76 
77  /*
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.
82  */
83  err = perf_evlist__create_maps(evlist, &opts.target);
84  if (err < 0) {
85  pr_debug("Not enough memory to create thread/cpu maps\n");
86  goto out_delete_evlist;
87  }
88 
89  /*
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.
94  */
95  err = perf_evlist__prepare_workload(evlist, &opts.target, argv, false, NULL);
96  if (err < 0) {
97  pr_debug("Couldn't run the workload!\n");
98  goto out_delete_evlist;
99  }
100 
101  /*
102  * Config the evsels, setting attr->comm on the first one, etc.
103  */
104  evsel = perf_evlist__first(evlist);
106  perf_evsel__set_sample_bit(evsel, TID);
107  perf_evsel__set_sample_bit(evsel, TIME);
108  perf_evlist__config(evlist, &opts, NULL);
109 
110  err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask);
111  if (err < 0) {
112  pr_debug("sched__get_first_possible_cpu: %s\n",
113  str_error_r(errno, sbuf, sizeof(sbuf)));
114  goto out_delete_evlist;
115  }
116 
117  cpu = err;
118 
119  /*
120  * So that we can check perf_sample.cpu on all the samples.
121  */
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;
126  }
127 
128  /*
129  * Call sys_perf_event_open on all the fds on all the evsels,
130  * grouping them if asked to.
131  */
132  err = perf_evlist__open(evlist);
133  if (err < 0) {
134  pr_debug("perf_evlist__open: %s\n",
135  str_error_r(errno, sbuf, sizeof(sbuf)));
136  goto out_delete_evlist;
137  }
138 
139  /*
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)).
143  */
144  err = perf_evlist__mmap(evlist, opts.mmap_pages);
145  if (err < 0) {
146  pr_debug("perf_evlist__mmap: %s\n",
147  str_error_r(errno, sbuf, sizeof(sbuf)));
148  goto out_delete_evlist;
149  }
150 
151  /*
152  * Now that all is properly set up, enable the events, they will
153  * count just on workload.pid, which will start...
154  */
155  perf_evlist__enable(evlist);
156 
157  /*
158  * Now!
159  */
161 
162  while (1) {
163  int before = total_events;
164 
165  for (i = 0; i < evlist->nr_mmaps; i++) {
166  union perf_event *event;
167  struct perf_mmap *md;
168 
169  md = &evlist->mmap[i];
170  if (perf_mmap__read_init(md) < 0)
171  continue;
172 
173  while ((event = perf_mmap__read_event(md)) != NULL) {
174  const u32 type = event->header.type;
175  const char *name = perf_event__name(type);
176 
177  ++total_events;
178  if (type < PERF_RECORD_MAX)
179  nr_events[type]++;
180 
181  err = perf_evlist__parse_sample(evlist, event, &sample);
182  if (err < 0) {
183  if (verbose > 0)
184  perf_event__fprintf(event, stderr);
185  pr_debug("Couldn't parse sample\n");
186  goto out_delete_evlist;
187  }
188 
189  if (verbose > 0) {
190  pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
191  perf_event__fprintf(event, stderr);
192  }
193 
194  if (prev_time > sample.time) {
195  pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
196  name, prev_time, sample.time);
197  ++errs;
198  }
199 
200  prev_time = sample.time;
201 
202  if (sample.cpu != cpu) {
203  pr_debug("%s with unexpected cpu, expected %d, got %d\n",
204  name, cpu, sample.cpu);
205  ++errs;
206  }
207 
208  if ((pid_t)sample.pid != evlist->workload.pid) {
209  pr_debug("%s with unexpected pid, expected %d, got %d\n",
210  name, evlist->workload.pid, sample.pid);
211  ++errs;
212  }
213 
214  if ((pid_t)sample.tid != evlist->workload.pid) {
215  pr_debug("%s with unexpected tid, expected %d, got %d\n",
216  name, evlist->workload.pid, sample.tid);
217  ++errs;
218  }
219 
220  if ((type == PERF_RECORD_COMM ||
221  type == PERF_RECORD_MMAP ||
222  type == PERF_RECORD_MMAP2 ||
223  type == PERF_RECORD_FORK ||
224  type == PERF_RECORD_EXIT) &&
225  (pid_t)event->comm.pid != evlist->workload.pid) {
226  pr_debug("%s with unexpected pid/tid\n", name);
227  ++errs;
228  }
229 
230  if ((type == PERF_RECORD_COMM ||
231  type == PERF_RECORD_MMAP ||
232  type == PERF_RECORD_MMAP2) &&
233  event->comm.pid != event->comm.tid) {
234  pr_debug("%s with different pid/tid!\n", name);
235  ++errs;
236  }
237 
238  switch (type) {
239  case PERF_RECORD_COMM:
240  if (strcmp(event->comm.comm, cmd)) {
241  pr_debug("%s with unexpected comm!\n", name);
242  ++errs;
243  }
244  break;
245  case PERF_RECORD_EXIT:
246  goto found_exit;
247  case PERF_RECORD_MMAP:
248  mmap_filename = event->mmap.filename;
249  goto check_bname;
250  case PERF_RECORD_MMAP2:
251  mmap_filename = event->mmap2.filename;
252  check_bname:
253  bname = strrchr(mmap_filename, '/');
254  if (bname != NULL) {
255  if (!found_cmd_mmap)
256  found_cmd_mmap = !strcmp(bname + 1, cmd);
257  if (!found_libc_mmap)
258  found_libc_mmap = !strncmp(bname + 1, "libc", 4);
259  if (!found_ld_mmap)
260  found_ld_mmap = !strncmp(bname + 1, "ld", 2);
261  } else if (!found_vdso_mmap)
262  found_vdso_mmap = !strcmp(mmap_filename, "[vdso]");
263  break;
264 
265  case PERF_RECORD_SAMPLE:
266  /* Just ignore samples for now */
267  break;
268  default:
269  pr_debug("Unexpected perf_event->header.type %d!\n",
270  type);
271  ++errs;
272  }
273 
274  perf_mmap__consume(md);
275  }
277  }
278 
279  /*
280  * We don't use poll here because at least at 3.1 times the
281  * PERF_RECORD_{!SAMPLE} events don't honour
282  * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
283  */
284  if (total_events == before && false)
285  perf_evlist__poll(evlist, -1);
286 
287  sleep(1);
288  if (++wakeups > 5) {
289  pr_debug("No PERF_RECORD_EXIT event!\n");
290  break;
291  }
292  }
293 
294 found_exit:
295  if (nr_events[PERF_RECORD_COMM] > 1) {
296  pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
297  ++errs;
298  }
299 
300  if (nr_events[PERF_RECORD_COMM] == 0) {
301  pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
302  ++errs;
303  }
304 
305  if (!found_cmd_mmap) {
306  pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
307  ++errs;
308  }
309 
310  if (!found_libc_mmap) {
311  pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
312  ++errs;
313  }
314 
315  if (!found_ld_mmap) {
316  pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
317  ++errs;
318  }
319 
320  if (!found_vdso_mmap) {
321  pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
322  ++errs;
323  }
324 out_delete_evlist:
325  perf_evlist__delete(evlist);
326 out:
327  return (err < 0 || errs > 0) ? -1 : 0;
328 }
int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target, const char *argv[], bool pipe_output, void(*exec_error)(int signo, siginfo_t *info, void *ucontext))
Definition: evlist.c:1428
void perf_mmap__consume(struct perf_mmap *map)
Definition: mmap.c:118
struct perf_evlist::@110 workload
int test__PERF_RECORD(struct test *test __maybe_unused, int subtest __maybe_unused)
Definition: perf-record.c:41
int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
Definition: evlist.c:1066
struct perf_mmap * mmap
Definition: evlist.h:45
struct perf_evlist * perf_evlist__new_default(void)
Definition: evlist.c:64
u32 tid
Definition: event.h:39
void perf_evlist__enable(struct perf_evlist *evlist)
Definition: evlist.c:369
int int err
Definition: 5sec.c:44
void perf_evlist__delete(struct perf_evlist *evlist)
Definition: evlist.c:133
#define perf_evsel__set_sample_bit(evsel, bit)
Definition: evsel.h:260
int perf_evlist__start_workload(struct perf_evlist *evlist)
Definition: evlist.c:1542
Definition: mmap.h:17
struct perf_evlist * perf_evlist__new_dummy(void)
Definition: evlist.c:76
struct target target
Definition: perf.h:38
int perf_mmap__read_init(struct perf_mmap *map)
Definition: mmap.c:271
const char * name
int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
Definition: evlist.c:491
#define pr_debug(fmt,...)
Definition: json.h:27
pid_t pid
Definition: evlist.h:42
int perf_evlist__open(struct perf_evlist *evlist)
Definition: evlist.c:1398
int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event, struct perf_sample *sample)
Definition: evlist.c:1561
u32 pid
Definition: event.h:193
u32 tid
Definition: event.h:193
unsigned int mmap_pages
Definition: perf.h:67
#define CPU(he)
#define event
void perf_evlist__config(struct perf_evlist *evlist, struct record_opts *opts, struct callchain_param *callchain)
Definition: record.c:135
u32 cpu
Definition: event.h:201
const char * perf_event__name(unsigned int id)
Definition: event.c:75
int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
Definition: evlist.c:1061
int nr_mmaps
Definition: evlist.h:32
void perf_mmap__read_done(struct perf_mmap *map)
Definition: mmap.c:329
u32 pid
Definition: hists_common.c:15
uid_t uid
Definition: target.h:13
static struct perf_evsel * perf_evlist__first(struct perf_evlist *evlist)
Definition: evlist.h:215
union perf_event * perf_mmap__read_event(struct perf_mmap *map)
Definition: mmap.c:78
Definition: tests.h:30
u64 time
Definition: event.h:194
size_t perf_event__fprintf(union perf_event *event, FILE *fp)
Definition: event.c:1461
#define pr_info(fmt,...)
Definition: json.h:24
struct comm_event comm
Definition: event.h:627
#define STRERR_BUFSIZE
Definition: debug.h:43
char comm[16]
Definition: event.h:40
int verbose
Definition: jevents.c:53
u32 pid
Definition: event.h:39
static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp)
Definition: perf-record.c:14