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_coreutils_mmap = false, 62 found_libc_mmap = false, 63 found_vdso_mmap = false, 64 found_ld_mmap = false; 65 int err = -1, errs = 0, i, wakeups = 0; 66 u32 cpu; 67 int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, }; 68 char sbuf[STRERR_BUFSIZE]; 69 70 if (evlist == NULL) /* Fallback for kernels lacking PERF_COUNT_SW_DUMMY */ 71 evlist = perf_evlist__new_default(); 72 73 if (evlist == NULL) { 74 pr_debug("Not enough memory to create evlist\n"); 75 goto out; 76 } 77 78 /* 79 * Create maps of threads and cpus to monitor. In this case 80 * we start with all threads and cpus (-1, -1) but then in 81 * perf_evlist__prepare_workload we'll fill in the only thread 82 * we're monitoring, the one forked there. 83 */ 84 err = perf_evlist__create_maps(evlist, &opts.target); 85 if (err < 0) { 86 pr_debug("Not enough memory to create thread/cpu maps\n"); 87 goto out_delete_evlist; 88 } 89 90 /* 91 * Prepare the workload in argv[] to run, it'll fork it, and then wait 92 * for perf_evlist__start_workload() to exec it. This is done this way 93 * so that we have time to open the evlist (calling sys_perf_event_open 94 * on all the fds) and then mmap them. 95 */ 96 err = perf_evlist__prepare_workload(evlist, &opts.target, argv, false, NULL); 97 if (err < 0) { 98 pr_debug("Couldn't run the workload!\n"); 99 goto out_delete_evlist; 100 } 101 102 /* 103 * Config the evsels, setting attr->comm on the first one, etc. 104 */ 105 evsel = perf_evlist__first(evlist); 106 perf_evsel__set_sample_bit(evsel, CPU); 107 perf_evsel__set_sample_bit(evsel, TID); 108 perf_evsel__set_sample_bit(evsel, TIME); 109 perf_evlist__config(evlist, &opts, NULL); 110 111 err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask); 112 if (err < 0) { 113 pr_debug("sched__get_first_possible_cpu: %s\n", 114 str_error_r(errno, sbuf, sizeof(sbuf))); 115 goto out_delete_evlist; 116 } 117 118 cpu = err; 119 120 /* 121 * So that we can check perf_sample.cpu on all the samples. 122 */ 123 if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, &cpu_mask) < 0) { 124 pr_debug("sched_setaffinity: %s\n", 125 str_error_r(errno, sbuf, sizeof(sbuf))); 126 goto out_delete_evlist; 127 } 128 129 /* 130 * Call sys_perf_event_open on all the fds on all the evsels, 131 * grouping them if asked to. 132 */ 133 err = perf_evlist__open(evlist); 134 if (err < 0) { 135 pr_debug("perf_evlist__open: %s\n", 136 str_error_r(errno, sbuf, sizeof(sbuf))); 137 goto out_delete_evlist; 138 } 139 140 /* 141 * mmap the first fd on a given CPU and ask for events for the other 142 * fds in the same CPU to be injected in the same mmap ring buffer 143 * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)). 144 */ 145 err = perf_evlist__mmap(evlist, opts.mmap_pages); 146 if (err < 0) { 147 pr_debug("perf_evlist__mmap: %s\n", 148 str_error_r(errno, sbuf, sizeof(sbuf))); 149 goto out_delete_evlist; 150 } 151 152 /* 153 * Now that all is properly set up, enable the events, they will 154 * count just on workload.pid, which will start... 155 */ 156 perf_evlist__enable(evlist); 157 158 /* 159 * Now! 160 */ 161 perf_evlist__start_workload(evlist); 162 163 while (1) { 164 int before = total_events; 165 166 for (i = 0; i < evlist->nr_mmaps; i++) { 167 union perf_event *event; 168 struct perf_mmap *md; 169 170 md = &evlist->mmap[i]; 171 if (perf_mmap__read_init(md) < 0) 172 continue; 173 174 while ((event = perf_mmap__read_event(md)) != NULL) { 175 const u32 type = event->header.type; 176 const char *name = perf_event__name(type); 177 178 ++total_events; 179 if (type < PERF_RECORD_MAX) 180 nr_events[type]++; 181 182 err = perf_evlist__parse_sample(evlist, event, &sample); 183 if (err < 0) { 184 if (verbose > 0) 185 perf_event__fprintf(event, stderr); 186 pr_debug("Couldn't parse sample\n"); 187 goto out_delete_evlist; 188 } 189 190 if (verbose > 0) { 191 pr_info("%" PRIu64" %d ", sample.time, sample.cpu); 192 perf_event__fprintf(event, stderr); 193 } 194 195 if (prev_time > sample.time) { 196 pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n", 197 name, prev_time, sample.time); 198 ++errs; 199 } 200 201 prev_time = sample.time; 202 203 if (sample.cpu != cpu) { 204 pr_debug("%s with unexpected cpu, expected %d, got %d\n", 205 name, cpu, sample.cpu); 206 ++errs; 207 } 208 209 if ((pid_t)sample.pid != evlist->workload.pid) { 210 pr_debug("%s with unexpected pid, expected %d, got %d\n", 211 name, evlist->workload.pid, sample.pid); 212 ++errs; 213 } 214 215 if ((pid_t)sample.tid != evlist->workload.pid) { 216 pr_debug("%s with unexpected tid, expected %d, got %d\n", 217 name, evlist->workload.pid, sample.tid); 218 ++errs; 219 } 220 221 if ((type == PERF_RECORD_COMM || 222 type == PERF_RECORD_MMAP || 223 type == PERF_RECORD_MMAP2 || 224 type == PERF_RECORD_FORK || 225 type == PERF_RECORD_EXIT) && 226 (pid_t)event->comm.pid != evlist->workload.pid) { 227 pr_debug("%s with unexpected pid/tid\n", name); 228 ++errs; 229 } 230 231 if ((type == PERF_RECORD_COMM || 232 type == PERF_RECORD_MMAP || 233 type == PERF_RECORD_MMAP2) && 234 event->comm.pid != event->comm.tid) { 235 pr_debug("%s with different pid/tid!\n", name); 236 ++errs; 237 } 238 239 switch (type) { 240 case PERF_RECORD_COMM: 241 if (strcmp(event->comm.comm, cmd)) { 242 pr_debug("%s with unexpected comm!\n", name); 243 ++errs; 244 } 245 break; 246 case PERF_RECORD_EXIT: 247 goto found_exit; 248 case PERF_RECORD_MMAP: 249 mmap_filename = event->mmap.filename; 250 goto check_bname; 251 case PERF_RECORD_MMAP2: 252 mmap_filename = event->mmap2.filename; 253 check_bname: 254 bname = strrchr(mmap_filename, '/'); 255 if (bname != NULL) { 256 if (!found_cmd_mmap) 257 found_cmd_mmap = !strcmp(bname + 1, cmd); 258 if (!found_coreutils_mmap) 259 found_coreutils_mmap = !strcmp(bname + 1, "coreutils"); 260 if (!found_libc_mmap) 261 found_libc_mmap = !strncmp(bname + 1, "libc", 4); 262 if (!found_ld_mmap) 263 found_ld_mmap = !strncmp(bname + 1, "ld", 2); 264 } else if (!found_vdso_mmap) 265 found_vdso_mmap = !strcmp(mmap_filename, "[vdso]"); 266 break; 267 268 case PERF_RECORD_SAMPLE: 269 /* Just ignore samples for now */ 270 break; 271 default: 272 pr_debug("Unexpected perf_event->header.type %d!\n", 273 type); 274 ++errs; 275 } 276 277 perf_mmap__consume(md); 278 } 279 perf_mmap__read_done(md); 280 } 281 282 /* 283 * We don't use poll here because at least at 3.1 times the 284 * PERF_RECORD_{!SAMPLE} events don't honour 285 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does. 286 */ 287 if (total_events == before && false) 288 perf_evlist__poll(evlist, -1); 289 290 sleep(1); 291 if (++wakeups > 5) { 292 pr_debug("No PERF_RECORD_EXIT event!\n"); 293 break; 294 } 295 } 296 297 found_exit: 298 if (nr_events[PERF_RECORD_COMM] > 1 + !!found_coreutils_mmap) { 299 pr_debug("Excessive number of PERF_RECORD_COMM events!\n"); 300 ++errs; 301 } 302 303 if (nr_events[PERF_RECORD_COMM] == 0) { 304 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd); 305 ++errs; 306 } 307 308 if (!found_cmd_mmap && !found_coreutils_mmap) { 309 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd); 310 ++errs; 311 } 312 313 if (!found_libc_mmap) { 314 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc"); 315 ++errs; 316 } 317 318 if (!found_ld_mmap) { 319 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld"); 320 ++errs; 321 } 322 323 if (!found_vdso_mmap) { 324 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]"); 325 ++errs; 326 } 327 out_delete_evlist: 328 perf_evlist__delete(evlist); 329 out: 330 return (err < 0 || errs > 0) ? -1 : 0; 331 } 332