tools/perf/builtin-test.c

   1 /*
   2  * builtin-test.c
   3  *
   4  * Builtin regression testing command: ever growing number of sanity tests
   5  */
   6 #include "builtin.h"
   7
   8 #include "util/cache.h"
   9 #include "util/debug.h"
  10 #include "util/debugfs.h"
  11 #include "util/evlist.h"
  12 #include "util/parse-options.h"
  13 #include "util/parse-events.h"
  14 #include "util/symbol.h"
  15 #include "util/thread_map.h"
  16 #include "util/pmu.h"
  17 #include "../../include/linux/hw_breakpoint.h"
  18
  19 #include <sys/mman.h>
  20
  21 static int vmlinux_matches_kallsyms_filter(struct map *map __used, struct symbol *sym)
  22 {
  23         bool *visited = symbol__priv(sym);
  24         *visited = true;
  25         return 0;
  26 }
  27
  28 static int test__vmlinux_matches_kallsyms(void)
  29 {
  30         int err = -1;
  31         struct rb_node *nd;
  32         struct symbol *sym;
  33         struct map *kallsyms_map, *vmlinux_map;
  34         struct machine kallsyms, vmlinux;
  35         enum map_type type = MAP__FUNCTION;
  36         long page_size = sysconf(_SC_PAGE_SIZE);
  37         struct ref_reloc_sym ref_reloc_sym = { .name = "_stext", };
  38
  39         /*
  40          * Step 1:
  41          *
  42          * Init the machines that will hold kernel, modules obtained from
  43          * both vmlinux + .ko files and from /proc/kallsyms split by modules.
  44          */
  45         machine__init(&kallsyms, "", HOST_KERNEL_ID);
  46         machine__init(&vmlinux, "", HOST_KERNEL_ID);
  47
  48         /*
  49          * Step 2:
  50          *
  51          * Create the kernel maps for kallsyms and the DSO where we will then
  52          * load /proc/kallsyms. Also create the modules maps from /proc/modules
  53          * and find the .ko files that match them in /lib/modules/`uname -r`/.
  54          */
  55         if (machine__create_kernel_maps(&kallsyms) < 0) {
  56                 pr_debug("machine__create_kernel_maps ");
  57                 return -1;
  58         }
  59
  60         /*
  61          * Step 3:
  62          *
  63          * Load and split /proc/kallsyms into multiple maps, one per module.
  64          */
  65         if (machine__load_kallsyms(&kallsyms, "/proc/kallsyms", type, NULL) <= 0) {
  66                 pr_debug("dso__load_kallsyms ");
  67                 goto out;
  68         }
  69
  70         /*
  71          * Step 4:
  72          *
  73          * kallsyms will be internally on demand sorted by name so that we can
  74          * find the reference relocation * symbol, i.e. the symbol we will use
  75          * to see if the running kernel was relocated by checking if it has the
  76          * same value in the vmlinux file we load.
  77          */
  78         kallsyms_map = machine__kernel_map(&kallsyms, type);
  79
  80         sym = map__find_symbol_by_name(kallsyms_map, ref_reloc_sym.name, NULL);
  81         if (sym == NULL) {
  82                 pr_debug("dso__find_symbol_by_name ");
  83                 goto out;
  84         }
  85
  86         ref_reloc_sym.addr = sym->start;
  87
  88         /*
  89          * Step 5:
  90          *
  91          * Now repeat step 2, this time for the vmlinux file we'll auto-locate.
  92          */
  93         if (machine__create_kernel_maps(&vmlinux) < 0) {
  94                 pr_debug("machine__create_kernel_maps ");
  95                 goto out;
  96         }
  97
  98         vmlinux_map = machine__kernel_map(&vmlinux, type);
  99         map__kmap(vmlinux_map)->ref_reloc_sym = &ref_reloc_sym;
 100
 101         /*
 102          * Step 6:
 103          *
 104          * Locate a vmlinux file in the vmlinux path that has a buildid that
 105          * matches the one of the running kernel.
 106          *
 107          * While doing that look if we find the ref reloc symbol, if we find it
 108          * we'll have its ref_reloc_symbol.unrelocated_addr and then
 109          * maps__reloc_vmlinux will notice and set proper ->[un]map_ip routines
 110          * to fixup the symbols.
 111          */
 112         if (machine__load_vmlinux_path(&vmlinux, type,
 113                                        vmlinux_matches_kallsyms_filter) <= 0) {
 114                 pr_debug("machine__load_vmlinux_path ");
 115                 goto out;
 116         }
 117
 118         err = 0;
 119         /*
 120          * Step 7:
 121          *
 122          * Now look at the symbols in the vmlinux DSO and check if we find all of them
 123          * in the kallsyms dso. For the ones that are in both, check its names and
 124          * end addresses too.
 125          */
 126         for (nd = rb_first(&vmlinux_map->dso->symbols[type]); nd; nd = rb_next(nd)) {
 127                 struct symbol *pair, *first_pair;
 128                 bool backwards = true;
 129
 130                 sym  = rb_entry(nd, struct symbol, rb_node);
 131
 132                 if (sym->start == sym->end)
 133                         continue;
 134
 135                 first_pair = machine__find_kernel_symbol(&kallsyms, type, sym->start, NULL, NULL);
 136                 pair = first_pair;
 137
 138                 if (pair && pair->start == sym->start) {
 139 next_pair:
 140                         if (strcmp(sym->name, pair->name) == 0) {
 141                                 /*
 142                                  * kallsyms don't have the symbol end, so we
 143                                  * set that by using the next symbol start - 1,
 144                                  * in some cases we get this up to a page
 145                                  * wrong, trace_kmalloc when I was developing
 146                                  * this code was one such example, 2106 bytes
 147                                  * off the real size. More than that and we
 148                                  * _really_ have a problem.
 149                                  */
 150                                 s64 skew = sym->end - pair->end;
 151                                 if (llabs(skew) < page_size)
 152                                         continue;
 153
 154                                 pr_debug("%#" PRIx64 ": diff end addr for %s v: %#" PRIx64 " k: %#" PRIx64 "\n",
 155                                          sym->start, sym->name, sym->end, pair->end);
 156                         } else {
 157                                 struct rb_node *nnd;
 158 detour:
 159                                 nnd = backwards ? rb_prev(&pair->rb_node) :
 160                                                   rb_next(&pair->rb_node);
 161                                 if (nnd) {
 162                                         struct symbol *next = rb_entry(nnd, struct symbol, rb_node);
 163
 164                                         if (next->start == sym->start) {
 165                                                 pair = next;
 166                                                 goto next_pair;
 167                                         }
 168                                 }
 169
 170                                 if (backwards) {
 171                                         backwards = false;
 172                                         pair = first_pair;
 173                                         goto detour;
 174                                 }
 175
 176                                 pr_debug("%#" PRIx64 ": diff name v: %s k: %s\n",
 177                                          sym->start, sym->name, pair->name);
 178                         }
 179                 } else
 180                         pr_debug("%#" PRIx64 ": %s not on kallsyms\n", sym->start, sym->name);
 181
 182                 err = -1;
 183         }
 184
 185         if (!verbose)
 186                 goto out;
 187
 188         pr_info("Maps only in vmlinux:\n");
 189
 190         for (nd = rb_first(&vmlinux.kmaps.maps[type]); nd; nd = rb_next(nd)) {
 191                 struct map *pos = rb_entry(nd, struct map, rb_node), *pair;
 192                 /*
 193                  * If it is the kernel, kallsyms is always "[kernel.kallsyms]", while
 194                  * the kernel will have the path for the vmlinux file being used,
 195                  * so use the short name, less descriptive but the same ("[kernel]" in
 196                  * both cases.
 197                  */
 198                 pair = map_groups__find_by_name(&kallsyms.kmaps, type,
 199                                                 (pos->dso->kernel ?
 200                                                         pos->dso->short_name :
 201                                                         pos->dso->name));
 202                 if (pair)
 203                         pair->priv = 1;
 204                 else
 205                         map__fprintf(pos, stderr);
 206         }
 207
 208         pr_info("Maps in vmlinux with a different name in kallsyms:\n");
 209
 210         for (nd = rb_first(&vmlinux.kmaps.maps[type]); nd; nd = rb_next(nd)) {
 211                 struct map *pos = rb_entry(nd, struct map, rb_node), *pair;
 212
 213                 pair = map_groups__find(&kallsyms.kmaps, type, pos->start);
 214                 if (pair == NULL || pair->priv)
 215                         continue;
 216
 217                 if (pair->start == pos->start) {
 218                         pair->priv = 1;
 219                         pr_info(" %" PRIx64 "-%" PRIx64 " %" PRIx64 " %s in kallsyms as",
 220                                 pos->start, pos->end, pos->pgoff, pos->dso->name);
 221                         if (pos->pgoff != pair->pgoff || pos->end != pair->end)
 222                                 pr_info(": \n*%" PRIx64 "-%" PRIx64 " %" PRIx64 "",
 223                                         pair->start, pair->end, pair->pgoff);
 224                         pr_info(" %s\n", pair->dso->name);
 225                         pair->priv = 1;
 226                 }
 227         }
 228
 229         pr_info("Maps only in kallsyms:\n");
 230
 231         for (nd = rb_first(&kallsyms.kmaps.maps[type]);
 232              nd; nd = rb_next(nd)) {
 233                 struct map *pos = rb_entry(nd, struct map, rb_node);
 234
 235                 if (!pos->priv)
 236                         map__fprintf(pos, stderr);
 237         }
 238 out:
 239         return err;
 240 }
 241
 242 #include "util/cpumap.h"
 243 #include "util/evsel.h"
 244 #include <sys/types.h>
 245
 246 static int trace_event__id(const char *evname)
 247 {
 248         char *filename;
 249         int err = -1, fd;
 250
 251         if (asprintf(&filename,
 252                      "%s/syscalls/%s/id",
 253                      tracing_events_path, evname) < 0)
 254                 return -1;
 255
 256         fd = open(filename, O_RDONLY);
 257         if (fd >= 0) {
 258                 char id[16];
 259                 if (read(fd, id, sizeof(id)) > 0)
 260                         err = atoi(id);
 261                 close(fd);
 262         }
 263
 264         free(filename);
 265         return err;
 266 }
 267
 268 static int test__open_syscall_event(void)
 269 {
 270         int err = -1, fd;
 271         struct thread_map *threads;
 272         struct perf_evsel *evsel;
 273         struct perf_event_attr attr;
 274         unsigned int nr_open_calls = 111, i;
 275         int id = trace_event__id("sys_enter_open");
 276
 277         if (id < 0) {
 278                 pr_debug("is debugfs mounted on /sys/kernel/debug?\n");
 279                 return -1;
 280         }
 281
 282         threads = thread_map__new(-1, getpid(), UINT_MAX);
 283         if (threads == NULL) {
 284                 pr_debug("thread_map__new\n");
 285                 return -1;
 286         }
 287
 288         memset(&attr, 0, sizeof(attr));
 289         attr.type = PERF_TYPE_TRACEPOINT;
 290         attr.config = id;
 291         evsel = perf_evsel__new(&attr, 0);
 292         if (evsel == NULL) {
 293                 pr_debug("perf_evsel__new\n");
 294                 goto out_thread_map_delete;
 295         }
 296
 297         if (perf_evsel__open_per_thread(evsel, threads, false, NULL) < 0) {
 298                 pr_debug("failed to open counter: %s, "
 299                          "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 300                          strerror(errno));
 301                 goto out_evsel_delete;
 302         }
 303
 304         for (i = 0; i < nr_open_calls; ++i) {
 305                 fd = open("/etc/passwd", O_RDONLY);
 306                 close(fd);
 307         }
 308
 309         if (perf_evsel__read_on_cpu(evsel, 0, 0) < 0) {
 310                 pr_debug("perf_evsel__read_on_cpu\n");
 311                 goto out_close_fd;
 312         }
 313
 314         if (evsel->counts->cpu[0].val != nr_open_calls) {
 315                 pr_debug("perf_evsel__read_on_cpu: expected to intercept %d calls, got %" PRIu64 "\n",
 316                          nr_open_calls, evsel->counts->cpu[0].val);
 317                 goto out_close_fd;
 318         }
 319
 320         err = 0;
 321 out_close_fd:
 322         perf_evsel__close_fd(evsel, 1, threads->nr);
 323 out_evsel_delete:
 324         perf_evsel__delete(evsel);
 325 out_thread_map_delete:
 326         thread_map__delete(threads);
 327         return err;
 328 }
 329
 330 #include <sched.h>
 331
 332 static int test__open_syscall_event_on_all_cpus(void)
 333 {
 334         int err = -1, fd, cpu;
 335         struct thread_map *threads;
 336         struct cpu_map *cpus;
 337         struct perf_evsel *evsel;
 338         struct perf_event_attr attr;
 339         unsigned int nr_open_calls = 111, i;
 340         cpu_set_t cpu_set;
 341         int id = trace_event__id("sys_enter_open");
 342
 343         if (id < 0) {
 344                 pr_debug("is debugfs mounted on /sys/kernel/debug?\n");
 345                 return -1;
 346         }
 347
 348         threads = thread_map__new(-1, getpid(), UINT_MAX);
 349         if (threads == NULL) {
 350                 pr_debug("thread_map__new\n");
 351                 return -1;
 352         }
 353
 354         cpus = cpu_map__new(NULL);
 355         if (cpus == NULL) {
 356                 pr_debug("cpu_map__new\n");
 357                 goto out_thread_map_delete;
 358         }
 359
 360
 361         CPU_ZERO(&cpu_set);
 362
 363         memset(&attr, 0, sizeof(attr));
 364         attr.type = PERF_TYPE_TRACEPOINT;
 365         attr.config = id;
 366         evsel = perf_evsel__new(&attr, 0);
 367         if (evsel == NULL) {
 368                 pr_debug("perf_evsel__new\n");
 369                 goto out_thread_map_delete;
 370         }
 371
 372         if (perf_evsel__open(evsel, cpus, threads, false, NULL) < 0) {
 373                 pr_debug("failed to open counter: %s, "
 374                          "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 375                          strerror(errno));
 376                 goto out_evsel_delete;
 377         }
 378
 379         for (cpu = 0; cpu < cpus->nr; ++cpu) {
 380                 unsigned int ncalls = nr_open_calls + cpu;
 381                 /*
 382                  * XXX eventually lift this restriction in a way that
 383                  * keeps perf building on older glibc installations
 384                  * without CPU_ALLOC. 1024 cpus in 2010 still seems
 385                  * a reasonable upper limit tho :-)
 386                  */
 387                 if (cpus->map[cpu] >= CPU_SETSIZE) {
 388                         pr_debug("Ignoring CPU %d\n", cpus->map[cpu]);
 389                         continue;
 390                 }
 391
 392                 CPU_SET(cpus->map[cpu], &cpu_set);
 393                 if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) {
 394                         pr_debug("sched_setaffinity() failed on CPU %d: %s ",
 395                                  cpus->map[cpu],
 396                                  strerror(errno));
 397                         goto out_close_fd;
 398                 }
 399                 for (i = 0; i < ncalls; ++i) {
 400                         fd = open("/etc/passwd", O_RDONLY);
 401                         close(fd);
 402                 }
 403                 CPU_CLR(cpus->map[cpu], &cpu_set);
 404         }
 405
 406         /*
 407          * Here we need to explicitely preallocate the counts, as if
 408          * we use the auto allocation it will allocate just for 1 cpu,
 409          * as we start by cpu 0.
 410          */
 411         if (perf_evsel__alloc_counts(evsel, cpus->nr) < 0) {
 412                 pr_debug("perf_evsel__alloc_counts(ncpus=%d)\n", cpus->nr);
 413                 goto out_close_fd;
 414         }
 415
 416         err = 0;
 417
 418         for (cpu = 0; cpu < cpus->nr; ++cpu) {
 419                 unsigned int expected;
 420
 421                 if (cpus->map[cpu] >= CPU_SETSIZE)
 422                         continue;
 423
 424                 if (perf_evsel__read_on_cpu(evsel, cpu, 0) < 0) {
 425                         pr_debug("perf_evsel__read_on_cpu\n");
 426                         err = -1;
 427                         break;
 428                 }
 429
 430                 expected = nr_open_calls + cpu;
 431                 if (evsel->counts->cpu[cpu].val != expected) {
 432                         pr_debug("perf_evsel__read_on_cpu: expected to intercept %d calls on cpu %d, got %" PRIu64 "\n",
 433                                  expected, cpus->map[cpu], evsel->counts->cpu[cpu].val);
 434                         err = -1;
 435                 }
 436         }
 437
 438 out_close_fd:
 439         perf_evsel__close_fd(evsel, 1, threads->nr);
 440 out_evsel_delete:
 441         perf_evsel__delete(evsel);
 442 out_thread_map_delete:
 443         thread_map__delete(threads);
 444         return err;
 445 }
 446
 447 /*
 448  * This test will generate random numbers of calls to some getpid syscalls,
 449  * then establish an mmap for a group of events that are created to monitor
 450  * the syscalls.
 451  *
 452  * It will receive the events, using mmap, use its PERF_SAMPLE_ID generated
 453  * sample.id field to map back to its respective perf_evsel instance.
 454  *
 455  * Then it checks if the number of syscalls reported as perf events by
 456  * the kernel corresponds to the number of syscalls made.
 457  */
 458 static int test__basic_mmap(void)
 459 {
 460         int err = -1;
 461         union perf_event *event;
 462         struct thread_map *threads;
 463         struct cpu_map *cpus;
 464         struct perf_evlist *evlist;
 465         struct perf_event_attr attr = {
 466                 .type           = PERF_TYPE_TRACEPOINT,
 467                 .read_format    = PERF_FORMAT_ID,
 468                 .sample_type    = PERF_SAMPLE_ID,
 469                 .watermark      = 0,
 470         };
 471         cpu_set_t cpu_set;
 472         const char *syscall_names[] = { "getsid", "getppid", "getpgrp",
 473                                         "getpgid", };
 474         pid_t (*syscalls[])(void) = { (void *)getsid, getppid, getpgrp,
 475                                       (void*)getpgid };
 476 #define nsyscalls ARRAY_SIZE(syscall_names)
 477         int ids[nsyscalls];
 478         unsigned int nr_events[nsyscalls],
 479                      expected_nr_events[nsyscalls], i, j;
 480         struct perf_evsel *evsels[nsyscalls], *evsel;
 481         int sample_size = __perf_evsel__sample_size(attr.sample_type);
 482
 483         for (i = 0; i < nsyscalls; ++i) {
 484                 char name[64];
 485
 486                 snprintf(name, sizeof(name), "sys_enter_%s", syscall_names[i]);
 487                 ids[i] = trace_event__id(name);
 488                 if (ids[i] < 0) {
 489                         pr_debug("Is debugfs mounted on /sys/kernel/debug?\n");
 490                         return -1;
 491                 }
 492                 nr_events[i] = 0;
 493                 expected_nr_events[i] = random() % 257;
 494         }
 495
 496         threads = thread_map__new(-1, getpid(), UINT_MAX);
 497         if (threads == NULL) {
 498                 pr_debug("thread_map__new\n");
 499                 return -1;
 500         }
 501
 502         cpus = cpu_map__new(NULL);
 503         if (cpus == NULL) {
 504                 pr_debug("cpu_map__new\n");
 505                 goto out_free_threads;
 506         }
 507
 508         CPU_ZERO(&cpu_set);
 509         CPU_SET(cpus->map[0], &cpu_set);
 510         sched_setaffinity(0, sizeof(cpu_set), &cpu_set);
 511         if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) {
 512                 pr_debug("sched_setaffinity() failed on CPU %d: %s ",
 513                          cpus->map[0], strerror(errno));
 514                 goto out_free_cpus;
 515         }
 516
 517         evlist = perf_evlist__new(cpus, threads);
 518         if (evlist == NULL) {
 519                 pr_debug("perf_evlist__new\n");
 520                 goto out_free_cpus;
 521         }
 522
 523         /* anonymous union fields, can't be initialized above */
 524         attr.wakeup_events = 1;
 525         attr.sample_period = 1;
 526
 527         for (i = 0; i < nsyscalls; ++i) {
 528                 attr.config = ids[i];
 529                 evsels[i] = perf_evsel__new(&attr, i);
 530                 if (evsels[i] == NULL) {
 531                         pr_debug("perf_evsel__new\n");
 532                         goto out_free_evlist;
 533                 }
 534
 535                 perf_evlist__add(evlist, evsels[i]);
 536
 537                 if (perf_evsel__open(evsels[i], cpus, threads, false, NULL) < 0) {
 538                         pr_debug("failed to open counter: %s, "
 539                                  "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 540                                  strerror(errno));
 541                         goto out_close_fd;
 542                 }
 543         }
 544
 545         if (perf_evlist__mmap(evlist, 128, true) < 0) {
 546                 pr_debug("failed to mmap events: %d (%s)\n", errno,
 547                          strerror(errno));
 548                 goto out_close_fd;
 549         }
 550
 551         for (i = 0; i < nsyscalls; ++i)
 552                 for (j = 0; j < expected_nr_events[i]; ++j) {
 553                         int foo = syscalls[i]();
 554                         ++foo;
 555                 }
 556
 557         while ((event = perf_evlist__mmap_read(evlist, 0)) != NULL) {
 558                 struct perf_sample sample;
 559
 560                 if (event->header.type != PERF_RECORD_SAMPLE) {
 561                         pr_debug("unexpected %s event\n",
 562                                  perf_event__name(event->header.type));
 563                         goto out_munmap;
 564                 }
 565
 566                 err = perf_event__parse_sample(event, attr.sample_type, sample_size,
 567                                                false, &sample, false);
 568                 if (err) {
 569                         pr_err("Can't parse sample, err = %d\n", err);
 570                         goto out_munmap;
 571                 }
 572
 573                 evsel = perf_evlist__id2evsel(evlist, sample.id);
 574                 if (evsel == NULL) {
 575                         pr_debug("event with id %" PRIu64
 576                                  " doesn't map to an evsel\n", sample.id);
 577                         goto out_munmap;
 578                 }
 579                 nr_events[evsel->idx]++;
 580         }
 581
 582         list_for_each_entry(evsel, &evlist->entries, node) {
 583                 if (nr_events[evsel->idx] != expected_nr_events[evsel->idx]) {
 584                         pr_debug("expected %d %s events, got %d\n",
 585                                  expected_nr_events[evsel->idx],
 586                                  perf_evsel__name(evsel), nr_events[evsel->idx]);
 587                         goto out_munmap;
 588                 }
 589         }
 590
 591         err = 0;
 592 out_munmap:
 593         perf_evlist__munmap(evlist);
 594 out_close_fd:
 595         for (i = 0; i < nsyscalls; ++i)
 596                 perf_evsel__close_fd(evsels[i], 1, threads->nr);
 597 out_free_evlist:
 598         perf_evlist__delete(evlist);
 599 out_free_cpus:
 600         cpu_map__delete(cpus);
 601 out_free_threads:
 602         thread_map__delete(threads);
 603         return err;
 604 #undef nsyscalls
 605 }
 606
 607 static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t **maskp,
 608                                          size_t *sizep)
 609 {
 610         cpu_set_t *mask;
 611         size_t size;
 612         int i, cpu = -1, nrcpus = 1024;
 613 realloc:
 614         mask = CPU_ALLOC(nrcpus);
 615         size = CPU_ALLOC_SIZE(nrcpus);
 616         CPU_ZERO_S(size, mask);
 617
 618         if (sched_getaffinity(pid, size, mask) == -1) {
 619                 CPU_FREE(mask);
 620                 if (errno == EINVAL && nrcpus < (1024 << 8)) {
 621                         nrcpus = nrcpus << 2;
 622                         goto realloc;
 623                 }
 624                 perror("sched_getaffinity");
 625                         return -1;
 626         }
 627
 628         for (i = 0; i < nrcpus; i++) {
 629                 if (CPU_ISSET_S(i, size, mask)) {
 630                         if (cpu == -1) {
 631                                 cpu = i;
 632                                 *maskp = mask;
 633                                 *sizep = size;
 634                         } else
 635                                 CPU_CLR_S(i, size, mask);
 636                 }
 637         }
 638
 639         if (cpu == -1)
 640                 CPU_FREE(mask);
 641
 642         return cpu;
 643 }
 644
 645 static int test__PERF_RECORD(void)
 646 {
 647         struct perf_record_opts opts = {
 648                 .target = {
 649                         .uid = UINT_MAX,
 650                         .uses_mmap = true,
 651                 },
 652                 .no_delay   = true,
 653                 .freq       = 10,
 654                 .mmap_pages = 256,
 655         };
 656         cpu_set_t *cpu_mask = NULL;
 657         size_t cpu_mask_size = 0;
 658         struct perf_evlist *evlist = perf_evlist__new(NULL, NULL);
 659         struct perf_evsel *evsel;
 660         struct perf_sample sample;
 661         const char *cmd = "sleep";
 662         const char *argv[] = { cmd, "1", NULL, };
 663         char *bname;
 664         u64 sample_type, prev_time = 0;
 665         bool found_cmd_mmap = false,
 666              found_libc_mmap = false,
 667              found_vdso_mmap = false,
 668              found_ld_mmap = false;
 669         int err = -1, errs = 0, i, wakeups = 0, sample_size;
 670         u32 cpu;
 671         int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
 672
 673         if (evlist == NULL || argv == NULL) {
 674                 pr_debug("Not enough memory to create evlist\n");
 675                 goto out;
 676         }
 677
 678         /*
 679          * We need at least one evsel in the evlist, use the default
 680          * one: "cycles".
 681          */
 682         err = perf_evlist__add_default(evlist);
 683         if (err < 0) {
 684                 pr_debug("Not enough memory to create evsel\n");
 685                 goto out_delete_evlist;
 686         }
 687
 688         /*
 689          * Create maps of threads and cpus to monitor. In this case
 690          * we start with all threads and cpus (-1, -1) but then in
 691          * perf_evlist__prepare_workload we'll fill in the only thread
 692          * we're monitoring, the one forked there.
 693          */
 694         err = perf_evlist__create_maps(evlist, &opts.target);
 695         if (err < 0) {
 696                 pr_debug("Not enough memory to create thread/cpu maps\n");
 697                 goto out_delete_evlist;
 698         }
 699
 700         /*
 701          * Prepare the workload in argv[] to run, it'll fork it, and then wait
 702          * for perf_evlist__start_workload() to exec it. This is done this way
 703          * so that we have time to open the evlist (calling sys_perf_event_open
 704          * on all the fds) and then mmap them.
 705          */
 706         err = perf_evlist__prepare_workload(evlist, &opts, argv);
 707         if (err < 0) {
 708                 pr_debug("Couldn't run the workload!\n");
 709                 goto out_delete_evlist;
 710         }
 711
 712         /*
 713          * Config the evsels, setting attr->comm on the first one, etc.
 714          */
 715         evsel = list_entry(evlist->entries.next, struct perf_evsel, node);
 716         evsel->attr.sample_type |= PERF_SAMPLE_CPU;
 717         evsel->attr.sample_type |= PERF_SAMPLE_TID;
 718         evsel->attr.sample_type |= PERF_SAMPLE_TIME;
 719         perf_evlist__config_attrs(evlist, &opts);
 720
 721         err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask,
 722                                             &cpu_mask_size);
 723         if (err < 0) {
 724                 pr_debug("sched__get_first_possible_cpu: %s\n", strerror(errno));
 725                 goto out_delete_evlist;
 726         }
 727
 728         cpu = err;
 729
 730         /*
 731          * So that we can check perf_sample.cpu on all the samples.
 732          */
 733         if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, cpu_mask) < 0) {
 734                 pr_debug("sched_setaffinity: %s\n", strerror(errno));
 735                 goto out_free_cpu_mask;
 736         }
 737
 738         /*
 739          * Call sys_perf_event_open on all the fds on all the evsels,
 740          * grouping them if asked to.
 741          */
 742         err = perf_evlist__open(evlist, opts.group);
 743         if (err < 0) {
 744                 pr_debug("perf_evlist__open: %s\n", strerror(errno));
 745                 goto out_delete_evlist;
 746         }
 747
 748         /*
 749          * mmap the first fd on a given CPU and ask for events for the other
 750          * fds in the same CPU to be injected in the same mmap ring buffer
 751          * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
 752          */
 753         err = perf_evlist__mmap(evlist, opts.mmap_pages, false);
 754         if (err < 0) {
 755                 pr_debug("perf_evlist__mmap: %s\n", strerror(errno));
 756                 goto out_delete_evlist;
 757         }
 758
 759         /*
 760          * We'll need these two to parse the PERF_SAMPLE_* fields in each
 761          * event.
 762          */
 763         sample_type = perf_evlist__sample_type(evlist);
 764         sample_size = __perf_evsel__sample_size(sample_type);
 765
 766         /*
 767          * Now that all is properly set up, enable the events, they will
 768          * count just on workload.pid, which will start...
 769          */
 770         perf_evlist__enable(evlist);
 771
 772         /*
 773          * Now!
 774          */
 775         perf_evlist__start_workload(evlist);
 776
 777         while (1) {
 778                 int before = total_events;
 779
 780                 for (i = 0; i < evlist->nr_mmaps; i++) {
 781                         union perf_event *event;
 782
 783                         while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 784                                 const u32 type = event->header.type;
 785                                 const char *name = perf_event__name(type);
 786
 787                                 ++total_events;
 788                                 if (type < PERF_RECORD_MAX)
 789                                         nr_events[type]++;
 790
 791                                 err = perf_event__parse_sample(event, sample_type,
 792                                                                sample_size, true,
 793                                                                &sample, false);
 794                                 if (err < 0) {
 795                                         if (verbose)
 796                                                 perf_event__fprintf(event, stderr);
 797                                         pr_debug("Couldn't parse sample\n");
 798                                         goto out_err;
 799                                 }
 800
 801                                 if (verbose) {
 802                                         pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
 803                                         perf_event__fprintf(event, stderr);
 804                                 }
 805
 806                                 if (prev_time > sample.time) {
 807                                         pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
 808                                                  name, prev_time, sample.time);
 809                                         ++errs;
 810                                 }
 811
 812                                 prev_time = sample.time;
 813
 814                                 if (sample.cpu != cpu) {
 815                                         pr_debug("%s with unexpected cpu, expected %d, got %d\n",
 816                                                  name, cpu, sample.cpu);
 817                                         ++errs;
 818                                 }
 819
 820                                 if ((pid_t)sample.pid != evlist->workload.pid) {
 821                                         pr_debug("%s with unexpected pid, expected %d, got %d\n",
 822                                                  name, evlist->workload.pid, sample.pid);
 823                                         ++errs;
 824                                 }
 825
 826                                 if ((pid_t)sample.tid != evlist->workload.pid) {
 827                                         pr_debug("%s with unexpected tid, expected %d, got %d\n",
 828                                                  name, evlist->workload.pid, sample.tid);
 829                                         ++errs;
 830                                 }
 831
 832                                 if ((type == PERF_RECORD_COMM ||
 833                                      type == PERF_RECORD_MMAP ||
 834                                      type == PERF_RECORD_FORK ||
 835                                      type == PERF_RECORD_EXIT) &&
 836                                      (pid_t)event->comm.pid != evlist->workload.pid) {
 837                                         pr_debug("%s with unexpected pid/tid\n", name);
 838                                         ++errs;
 839                                 }
 840
 841                                 if ((type == PERF_RECORD_COMM ||
 842                                      type == PERF_RECORD_MMAP) &&
 843                                      event->comm.pid != event->comm.tid) {
 844                                         pr_debug("%s with different pid/tid!\n", name);
 845                                         ++errs;
 846                                 }
 847
 848                                 switch (type) {
 849                                 case PERF_RECORD_COMM:
 850                                         if (strcmp(event->comm.comm, cmd)) {
 851                                                 pr_debug("%s with unexpected comm!\n", name);
 852                                                 ++errs;
 853                                         }
 854                                         break;
 855                                 case PERF_RECORD_EXIT:
 856                                         goto found_exit;
 857                                 case PERF_RECORD_MMAP:
 858                                         bname = strrchr(event->mmap.filename, '/');
 859                                         if (bname != NULL) {
 860                                                 if (!found_cmd_mmap)
 861                                                         found_cmd_mmap = !strcmp(bname + 1, cmd);
 862                                                 if (!found_libc_mmap)
 863                                                         found_libc_mmap = !strncmp(bname + 1, "libc", 4);
 864                                                 if (!found_ld_mmap)
 865                                                         found_ld_mmap = !strncmp(bname + 1, "ld", 2);
 866                                         } else if (!found_vdso_mmap)
 867                                                 found_vdso_mmap = !strcmp(event->mmap.filename, "[vdso]");
 868                                         break;
 869
 870                                 case PERF_RECORD_SAMPLE:
 871                                         /* Just ignore samples for now */
 872                                         break;
 873                                 default:
 874                                         pr_debug("Unexpected perf_event->header.type %d!\n",
 875                                                  type);
 876                                         ++errs;
 877                                 }
 878                         }
 879                 }
 880
 881                 /*
 882                  * We don't use poll here because at least at 3.1 times the
 883                  * PERF_RECORD_{!SAMPLE} events don't honour
 884                  * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
 885                  */
 886                 if (total_events == before && false)
 887                         poll(evlist->pollfd, evlist->nr_fds, -1);
 888
 889                 sleep(1);
 890                 if (++wakeups > 5) {
 891                         pr_debug("No PERF_RECORD_EXIT event!\n");
 892                         break;
 893                 }
 894         }
 895
 896 found_exit:
 897         if (nr_events[PERF_RECORD_COMM] > 1) {
 898                 pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
 899                 ++errs;
 900         }
 901
 902         if (nr_events[PERF_RECORD_COMM] == 0) {
 903                 pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
 904                 ++errs;
 905         }
 906
 907         if (!found_cmd_mmap) {
 908                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
 909                 ++errs;
 910         }
 911
 912         if (!found_libc_mmap) {
 913                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
 914                 ++errs;
 915         }
 916
 917         if (!found_ld_mmap) {
 918                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
 919                 ++errs;
 920         }
 921
 922         if (!found_vdso_mmap) {
 923                 pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
 924                 ++errs;
 925         }
 926 out_err:
 927         perf_evlist__munmap(evlist);
 928 out_free_cpu_mask:
 929         CPU_FREE(cpu_mask);
 930 out_delete_evlist:
 931         perf_evlist__delete(evlist);
 932 out:
 933         return (err < 0 || errs > 0) ? -1 : 0;
 934 }
 935
 936
 937 #if defined(__x86_64__) || defined(__i386__)
 938
 939 #define barrier() asm volatile("" ::: "memory")
 940
 941 static u64 rdpmc(unsigned int counter)
 942 {
 943         unsigned int low, high;
 944
 945         asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter));
 946
 947         return low | ((u64)high) << 32;
 948 }
 949
 950 static u64 rdtsc(void)
 951 {
 952         unsigned int low, high;
 953
 954         asm volatile("rdtsc" : "=a" (low), "=d" (high));
 955
 956         return low | ((u64)high) << 32;
 957 }
 958
 959 static u64 mmap_read_self(void *addr)
 960 {
 961         struct perf_event_mmap_page *pc = addr;
 962         u32 seq, idx, time_mult = 0, time_shift = 0;
 963         u64 count, cyc = 0, time_offset = 0, enabled, running, delta;
 964
 965         do {
 966                 seq = pc->lock;
 967                 barrier();
 968
 969                 enabled = pc->time_enabled;
 970                 running = pc->time_running;
 971
 972                 if (enabled != running) {
 973                         cyc = rdtsc();
 974                         time_mult = pc->time_mult;
 975                         time_shift = pc->time_shift;
 976                         time_offset = pc->time_offset;
 977                 }
 978
 979                 idx = pc->index;
 980                 count = pc->offset;
 981                 if (idx)
 982                         count += rdpmc(idx - 1);
 983
 984                 barrier();
 985         } while (pc->lock != seq);
 986
 987         if (enabled != running) {
 988                 u64 quot, rem;
 989
 990                 quot = (cyc >> time_shift);
 991                 rem = cyc & ((1 << time_shift) - 1);
 992                 delta = time_offset + quot * time_mult +
 993                         ((rem * time_mult) >> time_shift);
 994
 995                 enabled += delta;
 996                 if (idx)
 997                         running += delta;
 998
 999                 quot = count / running;
1000                 rem = count % running;
1001                 count = quot * enabled + (rem * enabled) / running;
1002         }
1003
1004         return count;
1005 }
1006
1007 /*
1008  * If the RDPMC instruction faults then signal this back to the test parent task:
1009  */
1010 static void segfault_handler(int sig __used, siginfo_t *info __used, void *uc __used)
1011 {
1012         exit(-1);
1013 }
1014
1015 static int __test__rdpmc(void)
1016 {
1017         long page_size = sysconf(_SC_PAGE_SIZE);
1018         volatile int tmp = 0;
1019         u64 i, loops = 1000;
1020         int n;
1021         int fd;
1022         void *addr;
1023         struct perf_event_attr attr = {
1024                 .type = PERF_TYPE_HARDWARE,
1025                 .config = PERF_COUNT_HW_INSTRUCTIONS,
1026                 .exclude_kernel = 1,
1027         };
1028         u64 delta_sum = 0;
1029         struct sigaction sa;
1030
1031         sigfillset(&sa.sa_mask);
1032         sa.sa_sigaction = segfault_handler;
1033         sigaction(SIGSEGV, &sa, NULL);
1034
1035         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
1036         if (fd < 0) {
1037                 die("Error: sys_perf_event_open() syscall returned "
1038                     "with %d (%s)\n", fd, strerror(errno));
1039         }
1040
1041         addr = mmap(NULL, page_size, PROT_READ, MAP_SHARED, fd, 0);
1042         if (addr == (void *)(-1)) {
1043                 die("Error: mmap() syscall returned "
1044                     "with (%s)\n", strerror(errno));
1045         }
1046
1047         for (n = 0; n < 6; n++) {
1048                 u64 stamp, now, delta;
1049
1050                 stamp = mmap_read_self(addr);
1051
1052                 for (i = 0; i < loops; i++)
1053                         tmp++;
1054
1055                 now = mmap_read_self(addr);
1056                 loops *= 10;
1057
1058                 delta = now - stamp;
1059                 pr_debug("%14d: %14Lu\n", n, (long long)delta);
1060
1061                 delta_sum += delta;
1062         }
1063
1064         munmap(addr, page_size);
1065         close(fd);
1066
1067         pr_debug("   ");
1068
1069         if (!delta_sum)
1070                 return -1;
1071
1072         return 0;
1073 }
1074
1075 static int test__rdpmc(void)
1076 {
1077         int status = 0;
1078         int wret = 0;
1079         int ret;
1080         int pid;
1081
1082         pid = fork();
1083         if (pid < 0)
1084                 return -1;
1085
1086         if (!pid) {
1087                 ret = __test__rdpmc();
1088
1089                 exit(ret);
1090         }
1091
1092         wret = waitpid(pid, &status, 0);
1093         if (wret < 0 || status)
1094                 return -1;
1095
1096         return 0;
1097 }
1098
1099 #endif
1100
1101 static int test__perf_pmu(void)
1102 {
1103         return perf_pmu__test();
1104 }
1105
1106 static struct test {
1107         const char *desc;
1108         int (*func)(void);
1109 } tests[] = {
1110         {
1111                 .desc = "vmlinux symtab matches kallsyms",
1112                 .func = test__vmlinux_matches_kallsyms,
1113         },
1114         {
1115                 .desc = "detect open syscall event",
1116                 .func = test__open_syscall_event,
1117         },
1118         {
1119                 .desc = "detect open syscall event on all cpus",
1120                 .func = test__open_syscall_event_on_all_cpus,
1121         },
1122         {
1123                 .desc = "read samples using the mmap interface",
1124                 .func = test__basic_mmap,
1125         },
1126         {
1127                 .desc = "parse events tests",
1128                 .func = parse_events__test,
1129         },
1130 #if defined(__x86_64__) || defined(__i386__)
1131         {
1132                 .desc = "x86 rdpmc test",
1133                 .func = test__rdpmc,
1134         },
1135 #endif
1136         {
1137                 .desc = "Validate PERF_RECORD_* events & perf_sample fields",
1138                 .func = test__PERF_RECORD,
1139         },
1140         {
1141                 .desc = "Test perf pmu format parsing",
1142                 .func = test__perf_pmu,
1143         },
1144         {
1145                 .func = NULL,
1146         },
1147 };
1148
1149 static bool perf_test__matches(int curr, int argc, const char *argv[])
1150 {
1151         int i;
1152
1153         if (argc == 0)
1154                 return true;
1155
1156         for (i = 0; i < argc; ++i) {
1157                 char *end;
1158                 long nr = strtoul(argv[i], &end, 10);
1159
1160                 if (*end == '\0') {
1161                         if (nr == curr + 1)
1162                                 return true;
1163                         continue;
1164                 }
1165
1166                 if (strstr(tests[curr].desc, argv[i]))
1167                         return true;
1168         }
1169
1170         return false;
1171 }
1172
1173 static int __cmd_test(int argc, const char *argv[])
1174 {
1175         int i = 0;
1176
1177         while (tests[i].func) {
1178                 int curr = i++, err;
1179
1180                 if (!perf_test__matches(curr, argc, argv))
1181                         continue;
1182
1183                 pr_info("%2d: %s:", i, tests[curr].desc);
1184                 pr_debug("\n--- start ---\n");
1185                 err = tests[curr].func();
1186                 pr_debug("---- end ----\n%s:", tests[curr].desc);
1187                 pr_info(" %s\n", err ? "FAILED!\n" : "Ok");
1188         }
1189
1190         return 0;
1191 }
1192
1193 static int perf_test__list(int argc, const char **argv)
1194 {
1195         int i = 0;
1196
1197         while (tests[i].func) {
1198                 int curr = i++;
1199
1200                 if (argc > 1 && !strstr(tests[curr].desc, argv[1]))
1201                         continue;
1202
1203                 pr_info("%2d: %s\n", i, tests[curr].desc);
1204         }
1205
1206         return 0;
1207 }
1208
1209 int cmd_test(int argc, const char **argv, const char *prefix __used)
1210 {
1211         const char * const test_usage[] = {
1212         "perf test [<options>] [{list <test-name-fragment>|[<test-name-fragments>|<test-numbers>]}]",
1213         NULL,
1214         };
1215         const struct option test_options[] = {
1216         OPT_INCR('v', "verbose", &verbose,
1217                     "be more verbose (show symbol address, etc)"),
1218         OPT_END()
1219         };
1220
1221         argc = parse_options(argc, argv, test_options, test_usage, 0);
1222         if (argc >= 1 && !strcmp(argv[0], "list"))
1223                 return perf_test__list(argc, argv);
1224
1225         symbol_conf.priv_size = sizeof(int);
1226         symbol_conf.sort_by_name = true;
1227         symbol_conf.try_vmlinux_path = true;
1228
1229         if (symbol__init() < 0)
1230                 return -1;
1231
1232         return __cmd_test(argc, argv);
1233 }