tools/perf/tests/code-reading.c

   1 #include <sys/types.h>
   2 #include <stdlib.h>
   3 #include <unistd.h>
   4 #include <stdio.h>
   5 #include <inttypes.h>
   6 #include <ctype.h>
   7 #include <string.h>
   8
   9 #include "parse-events.h"
  10 #include "evlist.h"
  11 #include "evsel.h"
  12 #include "thread_map.h"
  13 #include "cpumap.h"
  14 #include "machine.h"
  15 #include "event.h"
  16 #include "thread.h"
  17
  18 #include "tests.h"
  19
  20 #define BUFSZ   1024
  21 #define READLEN 128
  22
  23 struct state {
  24         u64 done[1024];
  25         size_t done_cnt;
  26 };
  27
  28 static unsigned int hex(char c)
  29 {
  30         if (c >= '0' && c <= '9')
  31                 return c - '0';
  32         if (c >= 'a' && c <= 'f')
  33                 return c - 'a' + 10;
  34         return c - 'A' + 10;
  35 }
  36
  37 static void read_objdump_line(const char *line, size_t line_len, void **buf,
  38                               size_t *len)
  39 {
  40         const char *p;
  41         size_t i;
  42
  43         /* Skip to a colon */
  44         p = strchr(line, ':');
  45         if (!p)
  46                 return;
  47         i = p + 1 - line;
  48
  49         /* Read bytes */
  50         while (*len) {
  51                 char c1, c2;
  52
  53                 /* Skip spaces */
  54                 for (; i < line_len; i++) {
  55                         if (!isspace(line[i]))
  56                                 break;
  57                 }
  58                 /* Get 2 hex digits */
  59                 if (i >= line_len || !isxdigit(line[i]))
  60                         break;
  61                 c1 = line[i++];
  62                 if (i >= line_len || !isxdigit(line[i]))
  63                         break;
  64                 c2 = line[i++];
  65                 /* Followed by a space */
  66                 if (i < line_len && line[i] && !isspace(line[i]))
  67                         break;
  68                 /* Store byte */
  69                 *(unsigned char *)*buf = (hex(c1) << 4) | hex(c2);
  70                 *buf += 1;
  71                 *len -= 1;
  72         }
  73 }
  74
  75 static int read_objdump_output(FILE *f, void **buf, size_t *len)
  76 {
  77         char *line = NULL;
  78         size_t line_len;
  79         ssize_t ret;
  80         int err = 0;
  81
  82         while (1) {
  83                 ret = getline(&line, &line_len, f);
  84                 if (feof(f))
  85                         break;
  86                 if (ret < 0) {
  87                         pr_debug("getline failed\n");
  88                         err = -1;
  89                         break;
  90                 }
  91                 read_objdump_line(line, ret, buf, len);
  92         }
  93
  94         free(line);
  95
  96         return err;
  97 }
  98
  99 static int read_via_objdump(const char *filename, u64 addr, void *buf,
 100                             size_t len)
 101 {
 102         char cmd[PATH_MAX * 2];
 103         const char *fmt;
 104         FILE *f;
 105         int ret;
 106
 107         fmt = "%s -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
 108         ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
 109                        filename);
 110         if (ret <= 0 || (size_t)ret >= sizeof(cmd))
 111                 return -1;
 112
 113         pr_debug("Objdump command is: %s\n", cmd);
 114
 115         /* Ignore objdump errors */
 116         strcat(cmd, " 2>/dev/null");
 117
 118         f = popen(cmd, "r");
 119         if (!f) {
 120                 pr_debug("popen failed\n");
 121                 return -1;
 122         }
 123
 124         ret = read_objdump_output(f, &buf, &len);
 125         if (len) {
 126                 pr_debug("objdump read too few bytes\n");
 127                 if (!ret)
 128                         ret = len;
 129         }
 130
 131         pclose(f);
 132
 133         return ret;
 134 }
 135
 136 static int read_object_code(u64 addr, size_t len, u8 cpumode,
 137                             struct thread *thread, struct machine *machine,
 138                             struct state *state)
 139 {
 140         struct addr_location al;
 141         unsigned char buf1[BUFSZ];
 142         unsigned char buf2[BUFSZ];
 143         size_t ret_len;
 144         u64 objdump_addr;
 145         int ret;
 146
 147         pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
 148
 149         thread__find_addr_map(thread, machine, cpumode, MAP__FUNCTION, addr,
 150                               &al);
 151         if (!al.map || !al.map->dso) {
 152                 pr_debug("thread__find_addr_map failed\n");
 153                 return -1;
 154         }
 155
 156         pr_debug("File is: %s\n", al.map->dso->long_name);
 157
 158         if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 159             !dso__is_kcore(al.map->dso)) {
 160                 pr_debug("Unexpected kernel address - skipping\n");
 161                 return 0;
 162         }
 163
 164         pr_debug("On file address is: %#"PRIx64"\n", al.addr);
 165
 166         if (len > BUFSZ)
 167                 len = BUFSZ;
 168
 169         /* Do not go off the map */
 170         if (addr + len > al.map->end)
 171                 len = al.map->end - addr;
 172
 173         /* Read the object code using perf */
 174         ret_len = dso__data_read_offset(al.map->dso, machine, al.addr, buf1,
 175                                         len);
 176         if (ret_len != len) {
 177                 pr_debug("dso__data_read_offset failed\n");
 178                 return -1;
 179         }
 180
 181         /*
 182          * Converting addresses for use by objdump requires more information.
 183          * map__load() does that.  See map__rip_2objdump() for details.
 184          */
 185         if (map__load(al.map, NULL))
 186                 return -1;
 187
 188         /* objdump struggles with kcore - try each map only once */
 189         if (dso__is_kcore(al.map->dso)) {
 190                 size_t d;
 191
 192                 for (d = 0; d < state->done_cnt; d++) {
 193                         if (state->done[d] == al.map->start) {
 194                                 pr_debug("kcore map tested already");
 195                                 pr_debug(" - skipping\n");
 196                                 return 0;
 197                         }
 198                 }
 199                 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
 200                         pr_debug("Too many kcore maps - skipping\n");
 201                         return 0;
 202                 }
 203                 state->done[state->done_cnt++] = al.map->start;
 204         }
 205
 206         /* Read the object code using objdump */
 207         objdump_addr = map__rip_2objdump(al.map, al.addr);
 208         ret = read_via_objdump(al.map->dso->long_name, objdump_addr, buf2, len);
 209         if (ret > 0) {
 210                 /*
 211                  * The kernel maps are inaccurate - assume objdump is right in
 212                  * that case.
 213                  */
 214                 if (cpumode == PERF_RECORD_MISC_KERNEL ||
 215                     cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
 216                         len -= ret;
 217                         if (len) {
 218                                 pr_debug("Reducing len to %zu\n", len);
 219                         } else if (dso__is_kcore(al.map->dso)) {
 220                                 /*
 221                                  * objdump cannot handle very large segments
 222                                  * that may be found in kcore.
 223                                  */
 224                                 pr_debug("objdump failed for kcore");
 225                                 pr_debug(" - skipping\n");
 226                                 return 0;
 227                         } else {
 228                                 return -1;
 229                         }
 230                 }
 231         }
 232         if (ret < 0) {
 233                 pr_debug("read_via_objdump failed\n");
 234                 return -1;
 235         }
 236
 237         /* The results should be identical */
 238         if (memcmp(buf1, buf2, len)) {
 239                 pr_debug("Bytes read differ from those read by objdump\n");
 240                 return -1;
 241         }
 242         pr_debug("Bytes read match those read by objdump\n");
 243
 244         return 0;
 245 }
 246
 247 static int process_sample_event(struct machine *machine,
 248                                 struct perf_evlist *evlist,
 249                                 union perf_event *event, struct state *state)
 250 {
 251         struct perf_sample sample;
 252         struct thread *thread;
 253         u8 cpumode;
 254
 255         if (perf_evlist__parse_sample(evlist, event, &sample)) {
 256                 pr_debug("perf_evlist__parse_sample failed\n");
 257                 return -1;
 258         }
 259
 260         thread = machine__findnew_thread(machine, sample.pid, sample.pid);
 261         if (!thread) {
 262                 pr_debug("machine__findnew_thread failed\n");
 263                 return -1;
 264         }
 265
 266         cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 267
 268         return read_object_code(sample.ip, READLEN, cpumode, thread, machine,
 269                                 state);
 270 }
 271
 272 static int process_event(struct machine *machine, struct perf_evlist *evlist,
 273                          union perf_event *event, struct state *state)
 274 {
 275         if (event->header.type == PERF_RECORD_SAMPLE)
 276                 return process_sample_event(machine, evlist, event, state);
 277
 278         if (event->header.type == PERF_RECORD_THROTTLE ||
 279             event->header.type == PERF_RECORD_UNTHROTTLE)
 280                 return 0;
 281
 282         if (event->header.type < PERF_RECORD_MAX) {
 283                 int ret;
 284
 285                 ret = machine__process_event(machine, event, NULL);
 286                 if (ret < 0)
 287                         pr_debug("machine__process_event failed, event type %u\n",
 288                                  event->header.type);
 289                 return ret;
 290         }
 291
 292         return 0;
 293 }
 294
 295 static int process_events(struct machine *machine, struct perf_evlist *evlist,
 296                           struct state *state)
 297 {
 298         union perf_event *event;
 299         int i, ret;
 300
 301         for (i = 0; i < evlist->nr_mmaps; i++) {
 302                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 303                         ret = process_event(machine, evlist, event, state);
 304                         perf_evlist__mmap_consume(evlist, i);
 305                         if (ret < 0)
 306                                 return ret;
 307                 }
 308         }
 309         return 0;
 310 }
 311
 312 static int comp(const void *a, const void *b)
 313 {
 314         return *(int *)a - *(int *)b;
 315 }
 316
 317 static void do_sort_something(void)
 318 {
 319         int buf[40960], i;
 320
 321         for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
 322                 buf[i] = ARRAY_SIZE(buf) - i - 1;
 323
 324         qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
 325
 326         for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
 327                 if (buf[i] != i) {
 328                         pr_debug("qsort failed\n");
 329                         break;
 330                 }
 331         }
 332 }
 333
 334 static void sort_something(void)
 335 {
 336         int i;
 337
 338         for (i = 0; i < 10; i++)
 339                 do_sort_something();
 340 }
 341
 342 static void syscall_something(void)
 343 {
 344         int pipefd[2];
 345         int i;
 346
 347         for (i = 0; i < 1000; i++) {
 348                 if (pipe(pipefd) < 0) {
 349                         pr_debug("pipe failed\n");
 350                         break;
 351                 }
 352                 close(pipefd[1]);
 353                 close(pipefd[0]);
 354         }
 355 }
 356
 357 static void fs_something(void)
 358 {
 359         const char *test_file_name = "temp-perf-code-reading-test-file--";
 360         FILE *f;
 361         int i;
 362
 363         for (i = 0; i < 1000; i++) {
 364                 f = fopen(test_file_name, "w+");
 365                 if (f) {
 366                         fclose(f);
 367                         unlink(test_file_name);
 368                 }
 369         }
 370 }
 371
 372 static void do_something(void)
 373 {
 374         fs_something();
 375
 376         sort_something();
 377
 378         syscall_something();
 379 }
 380
 381 enum {
 382         TEST_CODE_READING_OK,
 383         TEST_CODE_READING_NO_VMLINUX,
 384         TEST_CODE_READING_NO_KCORE,
 385         TEST_CODE_READING_NO_ACCESS,
 386         TEST_CODE_READING_NO_KERNEL_OBJ,
 387 };
 388
 389 static int do_test_code_reading(bool try_kcore)
 390 {
 391         struct machines machines;
 392         struct machine *machine;
 393         struct thread *thread;
 394         struct record_opts opts = {
 395                 .mmap_pages          = UINT_MAX,
 396                 .user_freq           = UINT_MAX,
 397                 .user_interval       = ULLONG_MAX,
 398                 .freq                = 4000,
 399                 .target              = {
 400                         .uses_mmap   = true,
 401                 },
 402         };
 403         struct state state = {
 404                 .done_cnt = 0,
 405         };
 406         struct thread_map *threads = NULL;
 407         struct cpu_map *cpus = NULL;
 408         struct perf_evlist *evlist = NULL;
 409         struct perf_evsel *evsel = NULL;
 410         int err = -1, ret;
 411         pid_t pid;
 412         struct map *map;
 413         bool have_vmlinux, have_kcore, excl_kernel = false;
 414
 415         pid = getpid();
 416
 417         machines__init(&machines);
 418         machine = &machines.host;
 419
 420         ret = machine__create_kernel_maps(machine);
 421         if (ret < 0) {
 422                 pr_debug("machine__create_kernel_maps failed\n");
 423                 goto out_err;
 424         }
 425
 426         /* Force the use of kallsyms instead of vmlinux to try kcore */
 427         if (try_kcore)
 428                 symbol_conf.kallsyms_name = "/proc/kallsyms";
 429
 430         /* Load kernel map */
 431         map = machine->vmlinux_maps[MAP__FUNCTION];
 432         ret = map__load(map, NULL);
 433         if (ret < 0) {
 434                 pr_debug("map__load failed\n");
 435                 goto out_err;
 436         }
 437         have_vmlinux = dso__is_vmlinux(map->dso);
 438         have_kcore = dso__is_kcore(map->dso);
 439
 440         /* 2nd time through we just try kcore */
 441         if (try_kcore && !have_kcore)
 442                 return TEST_CODE_READING_NO_KCORE;
 443
 444         /* No point getting kernel events if there is no kernel object */
 445         if (!have_vmlinux && !have_kcore)
 446                 excl_kernel = true;
 447
 448         threads = thread_map__new_by_tid(pid);
 449         if (!threads) {
 450                 pr_debug("thread_map__new_by_tid failed\n");
 451                 goto out_err;
 452         }
 453
 454         ret = perf_event__synthesize_thread_map(NULL, threads,
 455                                                 perf_event__process, machine, false);
 456         if (ret < 0) {
 457                 pr_debug("perf_event__synthesize_thread_map failed\n");
 458                 goto out_err;
 459         }
 460
 461         thread = machine__findnew_thread(machine, pid, pid);
 462         if (!thread) {
 463                 pr_debug("machine__findnew_thread failed\n");
 464                 goto out_err;
 465         }
 466
 467         cpus = cpu_map__new(NULL);
 468         if (!cpus) {
 469                 pr_debug("cpu_map__new failed\n");
 470                 goto out_err;
 471         }
 472
 473         while (1) {
 474                 const char *str;
 475
 476                 evlist = perf_evlist__new();
 477                 if (!evlist) {
 478                         pr_debug("perf_evlist__new failed\n");
 479                         goto out_err;
 480                 }
 481
 482                 perf_evlist__set_maps(evlist, cpus, threads);
 483
 484                 if (excl_kernel)
 485                         str = "cycles:u";
 486                 else
 487                         str = "cycles";
 488                 pr_debug("Parsing event '%s'\n", str);
 489                 ret = parse_events(evlist, str);
 490                 if (ret < 0) {
 491                         pr_debug("parse_events failed\n");
 492                         goto out_err;
 493                 }
 494
 495                 perf_evlist__config(evlist, &opts);
 496
 497                 evsel = perf_evlist__first(evlist);
 498
 499                 evsel->attr.comm = 1;
 500                 evsel->attr.disabled = 1;
 501                 evsel->attr.enable_on_exec = 0;
 502
 503                 ret = perf_evlist__open(evlist);
 504                 if (ret < 0) {
 505                         if (!excl_kernel) {
 506                                 excl_kernel = true;
 507                                 perf_evlist__delete(evlist);
 508                                 evlist = NULL;
 509                                 continue;
 510                         }
 511                         pr_debug("perf_evlist__open failed\n");
 512                         goto out_err;
 513                 }
 514                 break;
 515         }
 516
 517         ret = perf_evlist__mmap(evlist, UINT_MAX, false);
 518         if (ret < 0) {
 519                 pr_debug("perf_evlist__mmap failed\n");
 520                 goto out_err;
 521         }
 522
 523         perf_evlist__enable(evlist);
 524
 525         do_something();
 526
 527         perf_evlist__disable(evlist);
 528
 529         ret = process_events(machine, evlist, &state);
 530         if (ret < 0)
 531                 goto out_err;
 532
 533         if (!have_vmlinux && !have_kcore && !try_kcore)
 534                 err = TEST_CODE_READING_NO_KERNEL_OBJ;
 535         else if (!have_vmlinux && !try_kcore)
 536                 err = TEST_CODE_READING_NO_VMLINUX;
 537         else if (excl_kernel)
 538                 err = TEST_CODE_READING_NO_ACCESS;
 539         else
 540                 err = TEST_CODE_READING_OK;
 541 out_err:
 542         if (evlist) {
 543                 perf_evlist__munmap(evlist);
 544                 perf_evlist__close(evlist);
 545                 perf_evlist__delete(evlist);
 546         } else {
 547                 cpu_map__delete(cpus);
 548                 thread_map__delete(threads);
 549         }
 550         machines__destroy_kernel_maps(&machines);
 551         machine__delete_threads(machine);
 552         machines__exit(&machines);
 553
 554         return err;
 555 }
 556
 557 int test__code_reading(void)
 558 {
 559         int ret;
 560
 561         ret = do_test_code_reading(false);
 562         if (!ret)
 563                 ret = do_test_code_reading(true);
 564
 565         switch (ret) {
 566         case TEST_CODE_READING_OK:
 567                 return 0;
 568         case TEST_CODE_READING_NO_VMLINUX:
 569                 fprintf(stderr, " (no vmlinux)");
 570                 return 0;
 571         case TEST_CODE_READING_NO_KCORE:
 572                 fprintf(stderr, " (no kcore)");
 573                 return 0;
 574         case TEST_CODE_READING_NO_ACCESS:
 575                 fprintf(stderr, " (no access)");
 576                 return 0;
 577         case TEST_CODE_READING_NO_KERNEL_OBJ:
 578                 fprintf(stderr, " (no kernel obj)");
 579                 return 0;
 580         default:
 581                 return -1;
 582         };
 583 }