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 state *state)
 137 {
 138         struct addr_location al;
 139         unsigned char buf1[BUFSZ];
 140         unsigned char buf2[BUFSZ];
 141         size_t ret_len;
 142         u64 objdump_addr;
 143         int ret;
 144
 145         pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
 146
 147         thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
 148         if (!al.map || !al.map->dso) {
 149                 pr_debug("thread__find_addr_map failed\n");
 150                 return -1;
 151         }
 152
 153         pr_debug("File is: %s\n", al.map->dso->long_name);
 154
 155         if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 156             !dso__is_kcore(al.map->dso)) {
 157                 pr_debug("Unexpected kernel address - skipping\n");
 158                 return 0;
 159         }
 160
 161         pr_debug("On file address is: %#"PRIx64"\n", al.addr);
 162
 163         if (len > BUFSZ)
 164                 len = BUFSZ;
 165
 166         /* Do not go off the map */
 167         if (addr + len > al.map->end)
 168                 len = al.map->end - addr;
 169
 170         /* Read the object code using perf */
 171         ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
 172                                         al.addr, buf1, len);
 173         if (ret_len != len) {
 174                 pr_debug("dso__data_read_offset failed\n");
 175                 return -1;
 176         }
 177
 178         /*
 179          * Converting addresses for use by objdump requires more information.
 180          * map__load() does that.  See map__rip_2objdump() for details.
 181          */
 182         if (map__load(al.map, NULL))
 183                 return -1;
 184
 185         /* objdump struggles with kcore - try each map only once */
 186         if (dso__is_kcore(al.map->dso)) {
 187                 size_t d;
 188
 189                 for (d = 0; d < state->done_cnt; d++) {
 190                         if (state->done[d] == al.map->start) {
 191                                 pr_debug("kcore map tested already");
 192                                 pr_debug(" - skipping\n");
 193                                 return 0;
 194                         }
 195                 }
 196                 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
 197                         pr_debug("Too many kcore maps - skipping\n");
 198                         return 0;
 199                 }
 200                 state->done[state->done_cnt++] = al.map->start;
 201         }
 202
 203         /* Read the object code using objdump */
 204         objdump_addr = map__rip_2objdump(al.map, al.addr);
 205         ret = read_via_objdump(al.map->dso->long_name, objdump_addr, buf2, len);
 206         if (ret > 0) {
 207                 /*
 208                  * The kernel maps are inaccurate - assume objdump is right in
 209                  * that case.
 210                  */
 211                 if (cpumode == PERF_RECORD_MISC_KERNEL ||
 212                     cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
 213                         len -= ret;
 214                         if (len) {
 215                                 pr_debug("Reducing len to %zu\n", len);
 216                         } else if (dso__is_kcore(al.map->dso)) {
 217                                 /*
 218                                  * objdump cannot handle very large segments
 219                                  * that may be found in kcore.
 220                                  */
 221                                 pr_debug("objdump failed for kcore");
 222                                 pr_debug(" - skipping\n");
 223                                 return 0;
 224                         } else {
 225                                 return -1;
 226                         }
 227                 }
 228         }
 229         if (ret < 0) {
 230                 pr_debug("read_via_objdump failed\n");
 231                 return -1;
 232         }
 233
 234         /* The results should be identical */
 235         if (memcmp(buf1, buf2, len)) {
 236                 pr_debug("Bytes read differ from those read by objdump\n");
 237                 return -1;
 238         }
 239         pr_debug("Bytes read match those read by objdump\n");
 240
 241         return 0;
 242 }
 243
 244 static int process_sample_event(struct machine *machine,
 245                                 struct perf_evlist *evlist,
 246                                 union perf_event *event, struct state *state)
 247 {
 248         struct perf_sample sample;
 249         struct thread *thread;
 250         u8 cpumode;
 251         int ret;
 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         ret = read_object_code(sample.ip, READLEN, cpumode, thread, state);
 267         thread__put(thread);
 268         return ret;
 269 }
 270
 271 static int process_event(struct machine *machine, struct perf_evlist *evlist,
 272                          union perf_event *event, struct state *state)
 273 {
 274         if (event->header.type == PERF_RECORD_SAMPLE)
 275                 return process_sample_event(machine, evlist, event, state);
 276
 277         if (event->header.type == PERF_RECORD_THROTTLE ||
 278             event->header.type == PERF_RECORD_UNTHROTTLE)
 279                 return 0;
 280
 281         if (event->header.type < PERF_RECORD_MAX) {
 282                 int ret;
 283
 284                 ret = machine__process_event(machine, event, NULL);
 285                 if (ret < 0)
 286                         pr_debug("machine__process_event failed, event type %u\n",
 287                                  event->header.type);
 288                 return ret;
 289         }
 290
 291         return 0;
 292 }
 293
 294 static int process_events(struct machine *machine, struct perf_evlist *evlist,
 295                           struct state *state)
 296 {
 297         union perf_event *event;
 298         int i, ret;
 299
 300         for (i = 0; i < evlist->nr_mmaps; i++) {
 301                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 302                         ret = process_event(machine, evlist, event, state);
 303                         perf_evlist__mmap_consume(evlist, i);
 304                         if (ret < 0)
 305                                 return ret;
 306                 }
 307         }
 308         return 0;
 309 }
 310
 311 static int comp(const void *a, const void *b)
 312 {
 313         return *(int *)a - *(int *)b;
 314 }
 315
 316 static void do_sort_something(void)
 317 {
 318         int buf[40960], i;
 319
 320         for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
 321                 buf[i] = ARRAY_SIZE(buf) - i - 1;
 322
 323         qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
 324
 325         for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
 326                 if (buf[i] != i) {
 327                         pr_debug("qsort failed\n");
 328                         break;
 329                 }
 330         }
 331 }
 332
 333 static void sort_something(void)
 334 {
 335         int i;
 336
 337         for (i = 0; i < 10; i++)
 338                 do_sort_something();
 339 }
 340
 341 static void syscall_something(void)
 342 {
 343         int pipefd[2];
 344         int i;
 345
 346         for (i = 0; i < 1000; i++) {
 347                 if (pipe(pipefd) < 0) {
 348                         pr_debug("pipe failed\n");
 349                         break;
 350                 }
 351                 close(pipefd[1]);
 352                 close(pipefd[0]);
 353         }
 354 }
 355
 356 static void fs_something(void)
 357 {
 358         const char *test_file_name = "temp-perf-code-reading-test-file--";
 359         FILE *f;
 360         int i;
 361
 362         for (i = 0; i < 1000; i++) {
 363                 f = fopen(test_file_name, "w+");
 364                 if (f) {
 365                         fclose(f);
 366                         unlink(test_file_name);
 367                 }
 368         }
 369 }
 370
 371 static void do_something(void)
 372 {
 373         fs_something();
 374
 375         sort_something();
 376
 377         syscall_something();
 378 }
 379
 380 enum {
 381         TEST_CODE_READING_OK,
 382         TEST_CODE_READING_NO_VMLINUX,
 383         TEST_CODE_READING_NO_KCORE,
 384         TEST_CODE_READING_NO_ACCESS,
 385         TEST_CODE_READING_NO_KERNEL_OBJ,
 386 };
 387
 388 static int do_test_code_reading(bool try_kcore)
 389 {
 390         struct machines machines;
 391         struct machine *machine;
 392         struct thread *thread;
 393         struct record_opts opts = {
 394                 .mmap_pages          = UINT_MAX,
 395                 .user_freq           = UINT_MAX,
 396                 .user_interval       = ULLONG_MAX,
 397                 .freq                = 4000,
 398                 .target              = {
 399                         .uses_mmap   = true,
 400                 },
 401         };
 402         struct state state = {
 403                 .done_cnt = 0,
 404         };
 405         struct thread_map *threads = NULL;
 406         struct cpu_map *cpus = NULL;
 407         struct perf_evlist *evlist = NULL;
 408         struct perf_evsel *evsel = NULL;
 409         int err = -1, ret;
 410         pid_t pid;
 411         struct map *map;
 412         bool have_vmlinux, have_kcore, excl_kernel = false;
 413
 414         pid = getpid();
 415
 416         machines__init(&machines);
 417         machine = &machines.host;
 418
 419         ret = machine__create_kernel_maps(machine);
 420         if (ret < 0) {
 421                 pr_debug("machine__create_kernel_maps failed\n");
 422                 goto out_err;
 423         }
 424
 425         /* Force the use of kallsyms instead of vmlinux to try kcore */
 426         if (try_kcore)
 427                 symbol_conf.kallsyms_name = "/proc/kallsyms";
 428
 429         /* Load kernel map */
 430         map = machine->vmlinux_maps[MAP__FUNCTION];
 431         ret = map__load(map, NULL);
 432         if (ret < 0) {
 433                 pr_debug("map__load failed\n");
 434                 goto out_err;
 435         }
 436         have_vmlinux = dso__is_vmlinux(map->dso);
 437         have_kcore = dso__is_kcore(map->dso);
 438
 439         /* 2nd time through we just try kcore */
 440         if (try_kcore && !have_kcore)
 441                 return TEST_CODE_READING_NO_KCORE;
 442
 443         /* No point getting kernel events if there is no kernel object */
 444         if (!have_vmlinux && !have_kcore)
 445                 excl_kernel = true;
 446
 447         threads = thread_map__new_by_tid(pid);
 448         if (!threads) {
 449                 pr_debug("thread_map__new_by_tid failed\n");
 450                 goto out_err;
 451         }
 452
 453         ret = perf_event__synthesize_thread_map(NULL, threads,
 454                                                 perf_event__process, machine, false, 500);
 455         if (ret < 0) {
 456                 pr_debug("perf_event__synthesize_thread_map failed\n");
 457                 goto out_err;
 458         }
 459
 460         thread = machine__findnew_thread(machine, pid, pid);
 461         if (!thread) {
 462                 pr_debug("machine__findnew_thread failed\n");
 463                 goto out_put;
 464         }
 465
 466         cpus = cpu_map__new(NULL);
 467         if (!cpus) {
 468                 pr_debug("cpu_map__new failed\n");
 469                 goto out_put;
 470         }
 471
 472         while (1) {
 473                 const char *str;
 474
 475                 evlist = perf_evlist__new();
 476                 if (!evlist) {
 477                         pr_debug("perf_evlist__new failed\n");
 478                         goto out_put;
 479                 }
 480
 481                 perf_evlist__set_maps(evlist, cpus, threads);
 482
 483                 if (excl_kernel)
 484                         str = "cycles:u";
 485                 else
 486                         str = "cycles";
 487                 pr_debug("Parsing event '%s'\n", str);
 488                 ret = parse_events(evlist, str, NULL);
 489                 if (ret < 0) {
 490                         pr_debug("parse_events failed\n");
 491                         goto out_put;
 492                 }
 493
 494                 perf_evlist__config(evlist, &opts);
 495
 496                 evsel = perf_evlist__first(evlist);
 497
 498                 evsel->attr.comm = 1;
 499                 evsel->attr.disabled = 1;
 500                 evsel->attr.enable_on_exec = 0;
 501
 502                 ret = perf_evlist__open(evlist);
 503                 if (ret < 0) {
 504                         if (!excl_kernel) {
 505                                 excl_kernel = true;
 506                                 perf_evlist__set_maps(evlist, NULL, NULL);
 507                                 perf_evlist__delete(evlist);
 508                                 evlist = NULL;
 509                                 continue;
 510                         }
 511                         pr_debug("perf_evlist__open failed\n");
 512                         goto out_put;
 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_put;
 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_put;
 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_put:
 542         thread__put(thread);
 543 out_err:
 544
 545         if (evlist) {
 546                 perf_evlist__delete(evlist);
 547         } else {
 548                 cpu_map__put(cpus);
 549                 thread_map__put(threads);
 550         }
 551         machines__destroy_kernel_maps(&machines);
 552         machine__delete_threads(machine);
 553         machines__exit(&machines);
 554
 555         return err;
 556 }
 557
 558 int test__code_reading(void)
 559 {
 560         int ret;
 561
 562         ret = do_test_code_reading(false);
 563         if (!ret)
 564                 ret = do_test_code_reading(true);
 565
 566         switch (ret) {
 567         case TEST_CODE_READING_OK:
 568                 return 0;
 569         case TEST_CODE_READING_NO_VMLINUX:
 570                 fprintf(stderr, " (no vmlinux)");
 571                 return 0;
 572         case TEST_CODE_READING_NO_KCORE:
 573                 fprintf(stderr, " (no kcore)");
 574                 return 0;
 575         case TEST_CODE_READING_NO_ACCESS:
 576                 fprintf(stderr, " (no access)");
 577                 return 0;
 578         case TEST_CODE_READING_NO_KERNEL_OBJ:
 579                 fprintf(stderr, " (no kernel obj)");
 580                 return 0;
 581         default:
 582                 return -1;
 583         };
 584 }