tools/perf/tests/code-reading.c

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