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 8 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t *maskp) 9 { 10 int i, cpu = -1, nrcpus = 1024; 11 realloc: 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 35 int test__PERF_RECORD(void) 36 { 37 struct perf_record_opts opts = { 38 .target = { 39 .uid = UINT_MAX, 40 .uses_mmap = true, 41 }, 42 .no_delay = true, 43 .freq = 10, 44 .mmap_pages = 256, 45 }; 46 cpu_set_t cpu_mask; 47 size_t cpu_mask_size = sizeof(cpu_mask); 48 struct perf_evlist *evlist = perf_evlist__new_default(); 49 struct perf_evsel *evsel; 50 struct perf_sample sample; 51 const char *cmd = "sleep"; 52 const char *argv[] = { cmd, "1", NULL, }; 53 char *bname, *mmap_filename; 54 u64 prev_time = 0; 55 bool found_cmd_mmap = false, 56 found_libc_mmap = false, 57 found_vdso_mmap = false, 58 found_ld_mmap = false; 59 int err = -1, errs = 0, i, wakeups = 0; 60 u32 cpu; 61 int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, }; 62 63 if (evlist == NULL || argv == NULL) { 64 pr_debug("Not enough memory to create evlist\n"); 65 goto out; 66 } 67 68 /* 69 * Create maps of threads and cpus to monitor. In this case 70 * we start with all threads and cpus (-1, -1) but then in 71 * perf_evlist__prepare_workload we'll fill in the only thread 72 * we're monitoring, the one forked there. 73 */ 74 err = perf_evlist__create_maps(evlist, &opts.target); 75 if (err < 0) { 76 pr_debug("Not enough memory to create thread/cpu maps\n"); 77 goto out_delete_evlist; 78 } 79 80 /* 81 * Prepare the workload in argv[] to run, it'll fork it, and then wait 82 * for perf_evlist__start_workload() to exec it. This is done this way 83 * so that we have time to open the evlist (calling sys_perf_event_open 84 * on all the fds) and then mmap them. 85 */ 86 err = perf_evlist__prepare_workload(evlist, &opts.target, argv, 87 false, false); 88 if (err < 0) { 89 pr_debug("Couldn't run the workload!\n"); 90 goto out_delete_maps; 91 } 92 93 /* 94 * Config the evsels, setting attr->comm on the first one, etc. 95 */ 96 evsel = perf_evlist__first(evlist); 97 perf_evsel__set_sample_bit(evsel, CPU); 98 perf_evsel__set_sample_bit(evsel, TID); 99 perf_evsel__set_sample_bit(evsel, TIME); 100 perf_evlist__config(evlist, &opts); 101 102 err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask); 103 if (err < 0) { 104 pr_debug("sched__get_first_possible_cpu: %s\n", strerror(errno)); 105 goto out_delete_maps; 106 } 107 108 cpu = err; 109 110 /* 111 * So that we can check perf_sample.cpu on all the samples. 112 */ 113 if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) { 114 pr_debug("sched_setaffinity: %s\n", strerror(errno)); 115 goto out_delete_maps; 116 } 117 118 /* 119 * Call sys_perf_event_open on all the fds on all the evsels, 120 * grouping them if asked to. 121 */ 122 err = perf_evlist__open(evlist); 123 if (err < 0) { 124 pr_debug("perf_evlist__open: %s\n", strerror(errno)); 125 goto out_delete_maps; 126 } 127 128 /* 129 * mmap the first fd on a given CPU and ask for events for the other 130 * fds in the same CPU to be injected in the same mmap ring buffer 131 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)). 132 */ 133 err = perf_evlist__mmap(evlist, opts.mmap_pages, false); 134 if (err < 0) { 135 pr_debug("perf_evlist__mmap: %s\n", strerror(errno)); 136 goto out_close_evlist; 137 } 138 139 /* 140 * Now that all is properly set up, enable the events, they will 141 * count just on workload.pid, which will start... 142 */ 143 perf_evlist__enable(evlist); 144 145 /* 146 * Now! 147 */ 148 perf_evlist__start_workload(evlist); 149 150 while (1) { 151 int before = total_events; 152 153 for (i = 0; i < evlist->nr_mmaps; i++) { 154 union perf_event *event; 155 156 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) { 157 const u32 type = event->header.type; 158 const char *name = perf_event__name(type); 159 160 ++total_events; 161 if (type < PERF_RECORD_MAX) 162 nr_events[type]++; 163 164 err = perf_evlist__parse_sample(evlist, event, &sample); 165 if (err < 0) { 166 if (verbose) 167 perf_event__fprintf(event, stderr); 168 pr_debug("Couldn't parse sample\n"); 169 goto out_err; 170 } 171 172 if (verbose) { 173 pr_info("%" PRIu64" %d ", sample.time, sample.cpu); 174 perf_event__fprintf(event, stderr); 175 } 176 177 if (prev_time > sample.time) { 178 pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n", 179 name, prev_time, sample.time); 180 ++errs; 181 } 182 183 prev_time = sample.time; 184 185 if (sample.cpu != cpu) { 186 pr_debug("%s with unexpected cpu, expected %d, got %d\n", 187 name, cpu, sample.cpu); 188 ++errs; 189 } 190 191 if ((pid_t)sample.pid != evlist->workload.pid) { 192 pr_debug("%s with unexpected pid, expected %d, got %d\n", 193 name, evlist->workload.pid, sample.pid); 194 ++errs; 195 } 196 197 if ((pid_t)sample.tid != evlist->workload.pid) { 198 pr_debug("%s with unexpected tid, expected %d, got %d\n", 199 name, evlist->workload.pid, sample.tid); 200 ++errs; 201 } 202 203 if ((type == PERF_RECORD_COMM || 204 type == PERF_RECORD_MMAP || 205 type == PERF_RECORD_MMAP2 || 206 type == PERF_RECORD_FORK || 207 type == PERF_RECORD_EXIT) && 208 (pid_t)event->comm.pid != evlist->workload.pid) { 209 pr_debug("%s with unexpected pid/tid\n", name); 210 ++errs; 211 } 212 213 if ((type == PERF_RECORD_COMM || 214 type == PERF_RECORD_MMAP || 215 type == PERF_RECORD_MMAP2) && 216 event->comm.pid != event->comm.tid) { 217 pr_debug("%s with different pid/tid!\n", name); 218 ++errs; 219 } 220 221 switch (type) { 222 case PERF_RECORD_COMM: 223 if (strcmp(event->comm.comm, cmd)) { 224 pr_debug("%s with unexpected comm!\n", name); 225 ++errs; 226 } 227 break; 228 case PERF_RECORD_EXIT: 229 goto found_exit; 230 case PERF_RECORD_MMAP: 231 mmap_filename = event->mmap.filename; 232 goto check_bname; 233 case PERF_RECORD_MMAP2: 234 mmap_filename = event->mmap2.filename; 235 check_bname: 236 bname = strrchr(mmap_filename, '/'); 237 if (bname != NULL) { 238 if (!found_cmd_mmap) 239 found_cmd_mmap = !strcmp(bname + 1, cmd); 240 if (!found_libc_mmap) 241 found_libc_mmap = !strncmp(bname + 1, "libc", 4); 242 if (!found_ld_mmap) 243 found_ld_mmap = !strncmp(bname + 1, "ld", 2); 244 } else if (!found_vdso_mmap) 245 found_vdso_mmap = !strcmp(mmap_filename, "[vdso]"); 246 break; 247 248 case PERF_RECORD_SAMPLE: 249 /* Just ignore samples for now */ 250 break; 251 default: 252 pr_debug("Unexpected perf_event->header.type %d!\n", 253 type); 254 ++errs; 255 } 256 257 perf_evlist__mmap_consume(evlist, i); 258 } 259 } 260 261 /* 262 * We don't use poll here because at least at 3.1 times the 263 * PERF_RECORD_{!SAMPLE} events don't honour 264 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does. 265 */ 266 if (total_events == before && false) 267 poll(evlist->pollfd, evlist->nr_fds, -1); 268 269 sleep(1); 270 if (++wakeups > 5) { 271 pr_debug("No PERF_RECORD_EXIT event!\n"); 272 break; 273 } 274 } 275 276 found_exit: 277 if (nr_events[PERF_RECORD_COMM] > 1) { 278 pr_debug("Excessive number of PERF_RECORD_COMM events!\n"); 279 ++errs; 280 } 281 282 if (nr_events[PERF_RECORD_COMM] == 0) { 283 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd); 284 ++errs; 285 } 286 287 if (!found_cmd_mmap) { 288 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd); 289 ++errs; 290 } 291 292 if (!found_libc_mmap) { 293 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc"); 294 ++errs; 295 } 296 297 if (!found_ld_mmap) { 298 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld"); 299 ++errs; 300 } 301 302 if (!found_vdso_mmap) { 303 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]"); 304 ++errs; 305 } 306 out_err: 307 perf_evlist__munmap(evlist); 308 out_close_evlist: 309 perf_evlist__close(evlist); 310 out_delete_maps: 311 perf_evlist__delete_maps(evlist); 312 out_delete_evlist: 313 perf_evlist__delete(evlist); 314 out: 315 return (err < 0 || errs > 0) ? -1 : 0; 316 } 317