linux/tools/perf/builtin-test.c
<<
>>
Prefs
   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
  21static 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
  28static 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) {
 139next_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;
 158detour:
 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        }
 238out:
 239        return err;
 240}
 241
 242#include "util/cpumap.h"
 243#include "util/evsel.h"
 244#include <sys/types.h>
 245
 246static 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
 268static 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;
 321out_close_fd:
 322        perf_evsel__close_fd(evsel, 1, threads->nr);
 323out_evsel_delete:
 324        perf_evsel__delete(evsel);
 325out_thread_map_delete:
 326        thread_map__delete(threads);
 327        return err;
 328}
 329
 330#include <sched.h>
 331
 332static 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
 438out_close_fd:
 439        perf_evsel__close_fd(evsel, 1, threads->nr);
 440out_evsel_delete:
 441        perf_evsel__delete(evsel);
 442out_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 */
 458static 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
 482        for (i = 0; i < nsyscalls; ++i) {
 483                char name[64];
 484
 485                snprintf(name, sizeof(name), "sys_enter_%s", syscall_names[i]);
 486                ids[i] = trace_event__id(name);
 487                if (ids[i] < 0) {
 488                        pr_debug("Is debugfs mounted on /sys/kernel/debug?\n");
 489                        return -1;
 490                }
 491                nr_events[i] = 0;
 492                expected_nr_events[i] = random() % 257;
 493        }
 494
 495        threads = thread_map__new(-1, getpid(), UINT_MAX);
 496        if (threads == NULL) {
 497                pr_debug("thread_map__new\n");
 498                return -1;
 499        }
 500
 501        cpus = cpu_map__new(NULL);
 502        if (cpus == NULL) {
 503                pr_debug("cpu_map__new\n");
 504                goto out_free_threads;
 505        }
 506
 507        CPU_ZERO(&cpu_set);
 508        CPU_SET(cpus->map[0], &cpu_set);
 509        sched_setaffinity(0, sizeof(cpu_set), &cpu_set);
 510        if (sched_setaffinity(0, sizeof(cpu_set), &cpu_set) < 0) {
 511                pr_debug("sched_setaffinity() failed on CPU %d: %s ",
 512                         cpus->map[0], strerror(errno));
 513                goto out_free_cpus;
 514        }
 515
 516        evlist = perf_evlist__new(cpus, threads);
 517        if (evlist == NULL) {
 518                pr_debug("perf_evlist__new\n");
 519                goto out_free_cpus;
 520        }
 521
 522        /* anonymous union fields, can't be initialized above */
 523        attr.wakeup_events = 1;
 524        attr.sample_period = 1;
 525
 526        for (i = 0; i < nsyscalls; ++i) {
 527                attr.config = ids[i];
 528                evsels[i] = perf_evsel__new(&attr, i);
 529                if (evsels[i] == NULL) {
 530                        pr_debug("perf_evsel__new\n");
 531                        goto out_free_evlist;
 532                }
 533
 534                perf_evlist__add(evlist, evsels[i]);
 535
 536                if (perf_evsel__open(evsels[i], cpus, threads, false, NULL) < 0) {
 537                        pr_debug("failed to open counter: %s, "
 538                                 "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 539                                 strerror(errno));
 540                        goto out_close_fd;
 541                }
 542        }
 543
 544        if (perf_evlist__mmap(evlist, 128, true) < 0) {
 545                pr_debug("failed to mmap events: %d (%s)\n", errno,
 546                         strerror(errno));
 547                goto out_close_fd;
 548        }
 549
 550        for (i = 0; i < nsyscalls; ++i)
 551                for (j = 0; j < expected_nr_events[i]; ++j) {
 552                        int foo = syscalls[i]();
 553                        ++foo;
 554                }
 555
 556        while ((event = perf_evlist__mmap_read(evlist, 0)) != NULL) {
 557                struct perf_sample sample;
 558
 559                if (event->header.type != PERF_RECORD_SAMPLE) {
 560                        pr_debug("unexpected %s event\n",
 561                                 perf_event__name(event->header.type));
 562                        goto out_munmap;
 563                }
 564
 565                err = perf_evlist__parse_sample(evlist, event, &sample, false);
 566                if (err) {
 567                        pr_err("Can't parse sample, err = %d\n", err);
 568                        goto out_munmap;
 569                }
 570
 571                evsel = perf_evlist__id2evsel(evlist, sample.id);
 572                if (evsel == NULL) {
 573                        pr_debug("event with id %" PRIu64
 574                                 " doesn't map to an evsel\n", sample.id);
 575                        goto out_munmap;
 576                }
 577                nr_events[evsel->idx]++;
 578        }
 579
 580        list_for_each_entry(evsel, &evlist->entries, node) {
 581                if (nr_events[evsel->idx] != expected_nr_events[evsel->idx]) {
 582                        pr_debug("expected %d %s events, got %d\n",
 583                                 expected_nr_events[evsel->idx],
 584                                 perf_evsel__name(evsel), nr_events[evsel->idx]);
 585                        goto out_munmap;
 586                }
 587        }
 588
 589        err = 0;
 590out_munmap:
 591        perf_evlist__munmap(evlist);
 592out_close_fd:
 593        for (i = 0; i < nsyscalls; ++i)
 594                perf_evsel__close_fd(evsels[i], 1, threads->nr);
 595out_free_evlist:
 596        perf_evlist__delete(evlist);
 597out_free_cpus:
 598        cpu_map__delete(cpus);
 599out_free_threads:
 600        thread_map__delete(threads);
 601        return err;
 602#undef nsyscalls
 603}
 604
 605static int sched__get_first_possible_cpu(pid_t pid, cpu_set_t **maskp,
 606                                         size_t *sizep)
 607{
 608        cpu_set_t *mask;
 609        size_t size;
 610        int i, cpu = -1, nrcpus = 1024;
 611realloc:
 612        mask = CPU_ALLOC(nrcpus);
 613        size = CPU_ALLOC_SIZE(nrcpus);
 614        CPU_ZERO_S(size, mask);
 615
 616        if (sched_getaffinity(pid, size, mask) == -1) {
 617                CPU_FREE(mask);
 618                if (errno == EINVAL && nrcpus < (1024 << 8)) {
 619                        nrcpus = nrcpus << 2;
 620                        goto realloc;
 621                }
 622                perror("sched_getaffinity");
 623                        return -1;
 624        }
 625
 626        for (i = 0; i < nrcpus; i++) {
 627                if (CPU_ISSET_S(i, size, mask)) {
 628                        if (cpu == -1) {
 629                                cpu = i;
 630                                *maskp = mask;
 631                                *sizep = size;
 632                        } else
 633                                CPU_CLR_S(i, size, mask);
 634                }
 635        }
 636
 637        if (cpu == -1)
 638                CPU_FREE(mask);
 639
 640        return cpu;
 641}
 642
 643static int test__PERF_RECORD(void)
 644{
 645        struct perf_record_opts opts = {
 646                .target = {
 647                        .uid = UINT_MAX,
 648                        .uses_mmap = true,
 649                },
 650                .no_delay   = true,
 651                .freq       = 10,
 652                .mmap_pages = 256,
 653        };
 654        cpu_set_t *cpu_mask = NULL;
 655        size_t cpu_mask_size = 0;
 656        struct perf_evlist *evlist = perf_evlist__new(NULL, NULL);
 657        struct perf_evsel *evsel;
 658        struct perf_sample sample;
 659        const char *cmd = "sleep";
 660        const char *argv[] = { cmd, "1", NULL, };
 661        char *bname;
 662        u64 prev_time = 0;
 663        bool found_cmd_mmap = false,
 664             found_libc_mmap = false,
 665             found_vdso_mmap = false,
 666             found_ld_mmap = false;
 667        int err = -1, errs = 0, i, wakeups = 0;
 668        u32 cpu;
 669        int total_events = 0, nr_events[PERF_RECORD_MAX] = { 0, };
 670
 671        if (evlist == NULL || argv == NULL) {
 672                pr_debug("Not enough memory to create evlist\n");
 673                goto out;
 674        }
 675
 676        /*
 677         * We need at least one evsel in the evlist, use the default
 678         * one: "cycles".
 679         */
 680        err = perf_evlist__add_default(evlist);
 681        if (err < 0) {
 682                pr_debug("Not enough memory to create evsel\n");
 683                goto out_delete_evlist;
 684        }
 685
 686        /*
 687         * Create maps of threads and cpus to monitor. In this case
 688         * we start with all threads and cpus (-1, -1) but then in
 689         * perf_evlist__prepare_workload we'll fill in the only thread
 690         * we're monitoring, the one forked there.
 691         */
 692        err = perf_evlist__create_maps(evlist, &opts.target);
 693        if (err < 0) {
 694                pr_debug("Not enough memory to create thread/cpu maps\n");
 695                goto out_delete_evlist;
 696        }
 697
 698        /*
 699         * Prepare the workload in argv[] to run, it'll fork it, and then wait
 700         * for perf_evlist__start_workload() to exec it. This is done this way
 701         * so that we have time to open the evlist (calling sys_perf_event_open
 702         * on all the fds) and then mmap them.
 703         */
 704        err = perf_evlist__prepare_workload(evlist, &opts, argv);
 705        if (err < 0) {
 706                pr_debug("Couldn't run the workload!\n");
 707                goto out_delete_evlist;
 708        }
 709
 710        /*
 711         * Config the evsels, setting attr->comm on the first one, etc.
 712         */
 713        evsel = list_entry(evlist->entries.next, struct perf_evsel, node);
 714        evsel->attr.sample_type |= PERF_SAMPLE_CPU;
 715        evsel->attr.sample_type |= PERF_SAMPLE_TID;
 716        evsel->attr.sample_type |= PERF_SAMPLE_TIME;
 717        perf_evlist__config_attrs(evlist, &opts);
 718
 719        err = sched__get_first_possible_cpu(evlist->workload.pid, &cpu_mask,
 720                                            &cpu_mask_size);
 721        if (err < 0) {
 722                pr_debug("sched__get_first_possible_cpu: %s\n", strerror(errno));
 723                goto out_delete_evlist;
 724        }
 725
 726        cpu = err;
 727
 728        /*
 729         * So that we can check perf_sample.cpu on all the samples.
 730         */
 731        if (sched_setaffinity(evlist->workload.pid, cpu_mask_size, cpu_mask) < 0) {
 732                pr_debug("sched_setaffinity: %s\n", strerror(errno));
 733                goto out_free_cpu_mask;
 734        }
 735
 736        /*
 737         * Call sys_perf_event_open on all the fds on all the evsels,
 738         * grouping them if asked to.
 739         */
 740        err = perf_evlist__open(evlist, opts.group);
 741        if (err < 0) {
 742                pr_debug("perf_evlist__open: %s\n", strerror(errno));
 743                goto out_delete_evlist;
 744        }
 745
 746        /*
 747         * mmap the first fd on a given CPU and ask for events for the other
 748         * fds in the same CPU to be injected in the same mmap ring buffer
 749         * (using ioctl(PERF_EVENT_IOC_SET_OUTPUT)).
 750         */
 751        err = perf_evlist__mmap(evlist, opts.mmap_pages, false);
 752        if (err < 0) {
 753                pr_debug("perf_evlist__mmap: %s\n", strerror(errno));
 754                goto out_delete_evlist;
 755        }
 756
 757        /*
 758         * Now that all is properly set up, enable the events, they will
 759         * count just on workload.pid, which will start...
 760         */
 761        perf_evlist__enable(evlist);
 762
 763        /*
 764         * Now!
 765         */
 766        perf_evlist__start_workload(evlist);
 767
 768        while (1) {
 769                int before = total_events;
 770
 771                for (i = 0; i < evlist->nr_mmaps; i++) {
 772                        union perf_event *event;
 773
 774                        while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 775                                const u32 type = event->header.type;
 776                                const char *name = perf_event__name(type);
 777
 778                                ++total_events;
 779                                if (type < PERF_RECORD_MAX)
 780                                        nr_events[type]++;
 781
 782                                err = perf_evlist__parse_sample(evlist, event, &sample, false);
 783                                if (err < 0) {
 784                                        if (verbose)
 785                                                perf_event__fprintf(event, stderr);
 786                                        pr_debug("Couldn't parse sample\n");
 787                                        goto out_err;
 788                                }
 789
 790                                if (verbose) {
 791                                        pr_info("%" PRIu64" %d ", sample.time, sample.cpu);
 792                                        perf_event__fprintf(event, stderr);
 793                                }
 794
 795                                if (prev_time > sample.time) {
 796                                        pr_debug("%s going backwards in time, prev=%" PRIu64 ", curr=%" PRIu64 "\n",
 797                                                 name, prev_time, sample.time);
 798                                        ++errs;
 799                                }
 800
 801                                prev_time = sample.time;
 802
 803                                if (sample.cpu != cpu) {
 804                                        pr_debug("%s with unexpected cpu, expected %d, got %d\n",
 805                                                 name, cpu, sample.cpu);
 806                                        ++errs;
 807                                }
 808
 809                                if ((pid_t)sample.pid != evlist->workload.pid) {
 810                                        pr_debug("%s with unexpected pid, expected %d, got %d\n",
 811                                                 name, evlist->workload.pid, sample.pid);
 812                                        ++errs;
 813                                }
 814
 815                                if ((pid_t)sample.tid != evlist->workload.pid) {
 816                                        pr_debug("%s with unexpected tid, expected %d, got %d\n",
 817                                                 name, evlist->workload.pid, sample.tid);
 818                                        ++errs;
 819                                }
 820
 821                                if ((type == PERF_RECORD_COMM ||
 822                                     type == PERF_RECORD_MMAP ||
 823                                     type == PERF_RECORD_FORK ||
 824                                     type == PERF_RECORD_EXIT) &&
 825                                     (pid_t)event->comm.pid != evlist->workload.pid) {
 826                                        pr_debug("%s with unexpected pid/tid\n", name);
 827                                        ++errs;
 828                                }
 829
 830                                if ((type == PERF_RECORD_COMM ||
 831                                     type == PERF_RECORD_MMAP) &&
 832                                     event->comm.pid != event->comm.tid) {
 833                                        pr_debug("%s with different pid/tid!\n", name);
 834                                        ++errs;
 835                                }
 836
 837                                switch (type) {
 838                                case PERF_RECORD_COMM:
 839                                        if (strcmp(event->comm.comm, cmd)) {
 840                                                pr_debug("%s with unexpected comm!\n", name);
 841                                                ++errs;
 842                                        }
 843                                        break;
 844                                case PERF_RECORD_EXIT:
 845                                        goto found_exit;
 846                                case PERF_RECORD_MMAP:
 847                                        bname = strrchr(event->mmap.filename, '/');
 848                                        if (bname != NULL) {
 849                                                if (!found_cmd_mmap)
 850                                                        found_cmd_mmap = !strcmp(bname + 1, cmd);
 851                                                if (!found_libc_mmap)
 852                                                        found_libc_mmap = !strncmp(bname + 1, "libc", 4);
 853                                                if (!found_ld_mmap)
 854                                                        found_ld_mmap = !strncmp(bname + 1, "ld", 2);
 855                                        } else if (!found_vdso_mmap)
 856                                                found_vdso_mmap = !strcmp(event->mmap.filename, "[vdso]");
 857                                        break;
 858
 859                                case PERF_RECORD_SAMPLE:
 860                                        /* Just ignore samples for now */
 861                                        break;
 862                                default:
 863                                        pr_debug("Unexpected perf_event->header.type %d!\n",
 864                                                 type);
 865                                        ++errs;
 866                                }
 867                        }
 868                }
 869
 870                /*
 871                 * We don't use poll here because at least at 3.1 times the
 872                 * PERF_RECORD_{!SAMPLE} events don't honour
 873                 * perf_event_attr.wakeup_events, just PERF_EVENT_SAMPLE does.
 874                 */
 875                if (total_events == before && false)
 876                        poll(evlist->pollfd, evlist->nr_fds, -1);
 877
 878                sleep(1);
 879                if (++wakeups > 5) {
 880                        pr_debug("No PERF_RECORD_EXIT event!\n");
 881                        break;
 882                }
 883        }
 884
 885found_exit:
 886        if (nr_events[PERF_RECORD_COMM] > 1) {
 887                pr_debug("Excessive number of PERF_RECORD_COMM events!\n");
 888                ++errs;
 889        }
 890
 891        if (nr_events[PERF_RECORD_COMM] == 0) {
 892                pr_debug("Missing PERF_RECORD_COMM for %s!\n", cmd);
 893                ++errs;
 894        }
 895
 896        if (!found_cmd_mmap) {
 897                pr_debug("PERF_RECORD_MMAP for %s missing!\n", cmd);
 898                ++errs;
 899        }
 900
 901        if (!found_libc_mmap) {
 902                pr_debug("PERF_RECORD_MMAP for %s missing!\n", "libc");
 903                ++errs;
 904        }
 905
 906        if (!found_ld_mmap) {
 907                pr_debug("PERF_RECORD_MMAP for %s missing!\n", "ld");
 908                ++errs;
 909        }
 910
 911        if (!found_vdso_mmap) {
 912                pr_debug("PERF_RECORD_MMAP for %s missing!\n", "[vdso]");
 913                ++errs;
 914        }
 915out_err:
 916        perf_evlist__munmap(evlist);
 917out_free_cpu_mask:
 918        CPU_FREE(cpu_mask);
 919out_delete_evlist:
 920        perf_evlist__delete(evlist);
 921out:
 922        return (err < 0 || errs > 0) ? -1 : 0;
 923}
 924
 925
 926#if defined(__x86_64__) || defined(__i386__)
 927
 928#define barrier() asm volatile("" ::: "memory")
 929
 930static u64 rdpmc(unsigned int counter)
 931{
 932        unsigned int low, high;
 933
 934        asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter));
 935
 936        return low | ((u64)high) << 32;
 937}
 938
 939static u64 rdtsc(void)
 940{
 941        unsigned int low, high;
 942
 943        asm volatile("rdtsc" : "=a" (low), "=d" (high));
 944
 945        return low | ((u64)high) << 32;
 946}
 947
 948static u64 mmap_read_self(void *addr)
 949{
 950        struct perf_event_mmap_page *pc = addr;
 951        u32 seq, idx, time_mult = 0, time_shift = 0;
 952        u64 count, cyc = 0, time_offset = 0, enabled, running, delta;
 953
 954        do {
 955                seq = pc->lock;
 956                barrier();
 957
 958                enabled = pc->time_enabled;
 959                running = pc->time_running;
 960
 961                if (enabled != running) {
 962                        cyc = rdtsc();
 963                        time_mult = pc->time_mult;
 964                        time_shift = pc->time_shift;
 965                        time_offset = pc->time_offset;
 966                }
 967
 968                idx = pc->index;
 969                count = pc->offset;
 970                if (idx)
 971                        count += rdpmc(idx - 1);
 972
 973                barrier();
 974        } while (pc->lock != seq);
 975
 976        if (enabled != running) {
 977                u64 quot, rem;
 978
 979                quot = (cyc >> time_shift);
 980                rem = cyc & ((1 << time_shift) - 1);
 981                delta = time_offset + quot * time_mult +
 982                        ((rem * time_mult) >> time_shift);
 983
 984                enabled += delta;
 985                if (idx)
 986                        running += delta;
 987
 988                quot = count / running;
 989                rem = count % running;
 990                count = quot * enabled + (rem * enabled) / running;
 991        }
 992
 993        return count;
 994}
 995
 996/*
 997 * If the RDPMC instruction faults then signal this back to the test parent task:
 998 */
 999static void segfault_handler(int sig __used, siginfo_t *info __used, void *uc __used)
1000{
1001        exit(-1);
1002}
1003
1004static int __test__rdpmc(void)
1005{
1006        long page_size = sysconf(_SC_PAGE_SIZE);
1007        volatile int tmp = 0;
1008        u64 i, loops = 1000;
1009        int n;
1010        int fd;
1011        void *addr;
1012        struct perf_event_attr attr = {
1013                .type = PERF_TYPE_HARDWARE,
1014                .config = PERF_COUNT_HW_INSTRUCTIONS,
1015                .exclude_kernel = 1,
1016        };
1017        u64 delta_sum = 0;
1018        struct sigaction sa;
1019
1020        sigfillset(&sa.sa_mask);
1021        sa.sa_sigaction = segfault_handler;
1022        sigaction(SIGSEGV, &sa, NULL);
1023
1024        fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
1025        if (fd < 0) {
1026                die("Error: sys_perf_event_open() syscall returned "
1027                    "with %d (%s)\n", fd, strerror(errno));
1028        }
1029
1030        addr = mmap(NULL, page_size, PROT_READ, MAP_SHARED, fd, 0);
1031        if (addr == (void *)(-1)) {
1032                die("Error: mmap() syscall returned "
1033                    "with (%s)\n", strerror(errno));
1034        }
1035
1036        for (n = 0; n < 6; n++) {
1037                u64 stamp, now, delta;
1038
1039                stamp = mmap_read_self(addr);
1040
1041                for (i = 0; i < loops; i++)
1042                        tmp++;
1043
1044                now = mmap_read_self(addr);
1045                loops *= 10;
1046
1047                delta = now - stamp;
1048                pr_debug("%14d: %14Lu\n", n, (long long)delta);
1049
1050                delta_sum += delta;
1051        }
1052
1053        munmap(addr, page_size);
1054        close(fd);
1055
1056        pr_debug("   ");
1057
1058        if (!delta_sum)
1059                return -1;
1060
1061        return 0;
1062}
1063
1064static int test__rdpmc(void)
1065{
1066        int status = 0;
1067        int wret = 0;
1068        int ret;
1069        int pid;
1070
1071        pid = fork();
1072        if (pid < 0)
1073                return -1;
1074
1075        if (!pid) {
1076                ret = __test__rdpmc();
1077
1078                exit(ret);
1079        }
1080
1081        wret = waitpid(pid, &status, 0);
1082        if (wret < 0 || status)
1083                return -1;
1084
1085        return 0;
1086}
1087
1088#endif
1089
1090static int test__perf_pmu(void)
1091{
1092        return perf_pmu__test();
1093}
1094
1095static struct test {
1096        const char *desc;
1097        int (*func)(void);
1098} tests[] = {
1099        {
1100                .desc = "vmlinux symtab matches kallsyms",
1101                .func = test__vmlinux_matches_kallsyms,
1102        },
1103        {
1104                .desc = "detect open syscall event",
1105                .func = test__open_syscall_event,
1106        },
1107        {
1108                .desc = "detect open syscall event on all cpus",
1109                .func = test__open_syscall_event_on_all_cpus,
1110        },
1111        {
1112                .desc = "read samples using the mmap interface",
1113                .func = test__basic_mmap,
1114        },
1115        {
1116                .desc = "parse events tests",
1117                .func = parse_events__test,
1118        },
1119#if defined(__x86_64__) || defined(__i386__)
1120        {
1121                .desc = "x86 rdpmc test",
1122                .func = test__rdpmc,
1123        },
1124#endif
1125        {
1126                .desc = "Validate PERF_RECORD_* events & perf_sample fields",
1127                .func = test__PERF_RECORD,
1128        },
1129        {
1130                .desc = "Test perf pmu format parsing",
1131                .func = test__perf_pmu,
1132        },
1133        {
1134                .desc = "Test dso data interface",
1135                .func = dso__test_data,
1136        },
1137        {
1138                .func = NULL,
1139        },
1140};
1141
1142static bool perf_test__matches(int curr, int argc, const char *argv[])
1143{
1144        int i;
1145
1146        if (argc == 0)
1147                return true;
1148
1149        for (i = 0; i < argc; ++i) {
1150                char *end;
1151                long nr = strtoul(argv[i], &end, 10);
1152
1153                if (*end == '\0') {
1154                        if (nr == curr + 1)
1155                                return true;
1156                        continue;
1157                }
1158
1159                if (strstr(tests[curr].desc, argv[i]))
1160                        return true;
1161        }
1162
1163        return false;
1164}
1165
1166static int __cmd_test(int argc, const char *argv[])
1167{
1168        int i = 0;
1169
1170        while (tests[i].func) {
1171                int curr = i++, err;
1172
1173                if (!perf_test__matches(curr, argc, argv))
1174                        continue;
1175
1176                pr_info("%2d: %s:", i, tests[curr].desc);
1177                pr_debug("\n--- start ---\n");
1178                err = tests[curr].func();
1179                pr_debug("---- end ----\n%s:", tests[curr].desc);
1180                pr_info(" %s\n", err ? "FAILED!\n" : "Ok");
1181        }
1182
1183        return 0;
1184}
1185
1186static int perf_test__list(int argc, const char **argv)
1187{
1188        int i = 0;
1189
1190        while (tests[i].func) {
1191                int curr = i++;
1192
1193                if (argc > 1 && !strstr(tests[curr].desc, argv[1]))
1194                        continue;
1195
1196                pr_info("%2d: %s\n", i, tests[curr].desc);
1197        }
1198
1199        return 0;
1200}
1201
1202int cmd_test(int argc, const char **argv, const char *prefix __used)
1203{
1204        const char * const test_usage[] = {
1205        "perf test [<options>] [{list <test-name-fragment>|[<test-name-fragments>|<test-numbers>]}]",
1206        NULL,
1207        };
1208        const struct option test_options[] = {
1209        OPT_INCR('v', "verbose", &verbose,
1210                    "be more verbose (show symbol address, etc)"),
1211        OPT_END()
1212        };
1213
1214        argc = parse_options(argc, argv, test_options, test_usage, 0);
1215        if (argc >= 1 && !strcmp(argv[0], "list"))
1216                return perf_test__list(argc, argv);
1217
1218        symbol_conf.priv_size = sizeof(int);
1219        symbol_conf.sort_by_name = true;
1220        symbol_conf.try_vmlinux_path = true;
1221
1222        if (symbol__init() < 0)
1223                return -1;
1224
1225        return __cmd_test(argc, argv);
1226}
1227
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.