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 size_t 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, j = 0;
  41
  42         /* Skip to a colon */
  43         p = strchr(line, ':');
  44         if (!p)
  45                 return 0;
  46         i = p + 1 - line;
  47
  48         /* Read bytes */
  49         while (j < 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                 j++;
  71         }
  72         /* return number of successfully read bytes */
  73         return j;
  74 }
  75
  76 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
  77 {
  78         char *line = NULL;
  79         size_t line_len, off_last = 0;
  80         ssize_t ret;
  81         int err = 0;
  82         u64 addr;
  83
  84         while (off_last < *len) {
  85                 size_t off, read_bytes, written_bytes;
  86                 unsigned char tmp[BUFSZ];
  87
  88                 ret = getline(&line, &line_len, f);
  89                 if (feof(f))
  90                         break;
  91                 if (ret < 0) {
  92                         pr_debug("getline failed\n");
  93                         err = -1;
  94                         break;
  95                 }
  96
  97                 /* read objdump data into temporary buffer */
  98                 read_bytes = read_objdump_line(line, ret, tmp, sizeof(tmp));
  99                 if (!read_bytes)
 100                         continue;
 101
 102                 if (sscanf(line, "%"PRIx64, &addr) != 1)
 103                         continue;
 104
 105                 /* copy it from temporary buffer to 'buf' according
 106                  * to address on current objdump line */
 107                 off = addr - start_addr;
 108                 if (off >= *len)
 109                         break;
 110                 written_bytes = MIN(read_bytes, *len - off);
 111                 memcpy(buf + off, tmp, written_bytes);
 112                 off_last = off + written_bytes;
 113         }
 114
 115         /* len returns number of bytes that could not be read */
 116         *len -= off_last;
 117
 118         free(line);
 119
 120         return err;
 121 }
 122
 123 static int read_via_objdump(const char *filename, u64 addr, void *buf,
 124                             size_t len)
 125 {
 126         char cmd[PATH_MAX * 2];
 127         const char *fmt;
 128         FILE *f;
 129         int ret;
 130
 131         fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
 132         ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
 133                        filename);
 134         if (ret <= 0 || (size_t)ret >= sizeof(cmd))
 135                 return -1;
 136
 137         pr_debug("Objdump command is: %s\n", cmd);
 138
 139         /* Ignore objdump errors */
 140         strcat(cmd, " 2>/dev/null");
 141
 142         f = popen(cmd, "r");
 143         if (!f) {
 144                 pr_debug("popen failed\n");
 145                 return -1;
 146         }
 147
 148         ret = read_objdump_output(f, buf, &len, addr);
 149         if (len) {
 150                 pr_debug("objdump read too few bytes\n");
 151                 if (!ret)
 152                         ret = len;
 153         }
 154
 155         pclose(f);
 156
 157         return ret;
 158 }
 159
 160 static int read_object_code(u64 addr, size_t len, u8 cpumode,
 161                             struct thread *thread, struct state *state)
 162 {
 163         struct addr_location al;
 164         unsigned char buf1[BUFSZ];
 165         unsigned char buf2[BUFSZ];
 166         size_t ret_len;
 167         u64 objdump_addr;
 168         int ret;
 169
 170         pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
 171
 172         thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
 173         if (!al.map || !al.map->dso) {
 174                 pr_debug("thread__find_addr_map failed\n");
 175                 return -1;
 176         }
 177
 178         pr_debug("File is: %s\n", al.map->dso->long_name);
 179
 180         if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 181             !dso__is_kcore(al.map->dso)) {
 182                 pr_debug("Unexpected kernel address - skipping\n");
 183                 return 0;
 184         }
 185
 186         pr_debug("On file address is: %#"PRIx64"\n", al.addr);
 187
 188         if (len > BUFSZ)
 189                 len = BUFSZ;
 190
 191         /* Do not go off the map */
 192         if (addr + len > al.map->end)
 193                 len = al.map->end - addr;
 194
 195         /* Read the object code using perf */
 196         ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
 197                                         al.addr, buf1, len);
 198         if (ret_len != len) {
 199                 pr_debug("dso__data_read_offset failed\n");
 200                 return -1;
 201         }
 202
 203         /*
 204          * Converting addresses for use by objdump requires more information.
 205          * map__load() does that.  See map__rip_2objdump() for details.
 206          */
 207         if (map__load(al.map, NULL))
 208                 return -1;
 209
 210         /* objdump struggles with kcore - try each map only once */
 211         if (dso__is_kcore(al.map->dso)) {
 212                 size_t d;
 213
 214                 for (d = 0; d < state->done_cnt; d++) {
 215                         if (state->done[d] == al.map->start) {
 216                                 pr_debug("kcore map tested already");
 217                                 pr_debug(" - skipping\n");
 218                                 return 0;
 219                         }
 220                 }
 221                 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
 222                         pr_debug("Too many kcore maps - skipping\n");
 223                         return 0;
 224                 }
 225                 state->done[state->done_cnt++] = al.map->start;
 226         }
 227
 228         /* Read the object code using objdump */
 229         objdump_addr = map__rip_2objdump(al.map, al.addr);
 230         ret = read_via_objdump(al.map->dso->long_name, objdump_addr, buf2, len);
 231         if (ret > 0) {
 232                 /*
 233                  * The kernel maps are inaccurate - assume objdump is right in
 234                  * that case.
 235                  */
 236                 if (cpumode == PERF_RECORD_MISC_KERNEL ||
 237                     cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
 238                         len -= ret;
 239                         if (len) {
 240                                 pr_debug("Reducing len to %zu\n", len);
 241                         } else if (dso__is_kcore(al.map->dso)) {
 242                                 /*
 243                                  * objdump cannot handle very large segments
 244                                  * that may be found in kcore.
 245                                  */
 246                                 pr_debug("objdump failed for kcore");
 247                                 pr_debug(" - skipping\n");
 248                                 return 0;
 249                         } else {
 250                                 return -1;
 251                         }
 252                 }
 253         }
 254         if (ret < 0) {
 255                 pr_debug("read_via_objdump failed\n");
 256                 return -1;
 257         }
 258
 259         /* The results should be identical */
 260         if (memcmp(buf1, buf2, len)) {
 261                 pr_debug("Bytes read differ from those read by objdump\n");
 262                 return -1;
 263         }
 264         pr_debug("Bytes read match those read by objdump\n");
 265
 266         return 0;
 267 }
 268
 269 static int process_sample_event(struct machine *machine,
 270                                 struct perf_evlist *evlist,
 271                                 union perf_event *event, struct state *state)
 272 {
 273         struct perf_sample sample;
 274         struct thread *thread;
 275         u8 cpumode;
 276         int ret;
 277
 278         if (perf_evlist__parse_sample(evlist, event, &sample)) {
 279                 pr_debug("perf_evlist__parse_sample failed\n");
 280                 return -1;
 281         }
 282
 283         thread = machine__findnew_thread(machine, sample.pid, sample.tid);
 284         if (!thread) {
 285                 pr_debug("machine__findnew_thread failed\n");
 286                 return -1;
 287         }
 288
 289         cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 290
 291         ret = read_object_code(sample.ip, READLEN, cpumode, thread, state);
 292         thread__put(thread);
 293         return ret;
 294 }
 295
 296 static int process_event(struct machine *machine, struct perf_evlist *evlist,
 297                          union perf_event *event, struct state *state)
 298 {
 299         if (event->header.type == PERF_RECORD_SAMPLE)
 300                 return process_sample_event(machine, evlist, event, state);
 301
 302         if (event->header.type == PERF_RECORD_THROTTLE ||
 303             event->header.type == PERF_RECORD_UNTHROTTLE)
 304                 return 0;
 305
 306         if (event->header.type < PERF_RECORD_MAX) {
 307                 int ret;
 308
 309                 ret = machine__process_event(machine, event, NULL);
 310                 if (ret < 0)
 311                         pr_debug("machine__process_event failed, event type %u\n",
 312                                  event->header.type);
 313                 return ret;
 314         }
 315
 316         return 0;
 317 }
 318
 319 static int process_events(struct machine *machine, struct perf_evlist *evlist,
 320                           struct state *state)
 321 {
 322         union perf_event *event;
 323         int i, ret;
 324
 325         for (i = 0; i < evlist->nr_mmaps; i++) {
 326                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 327                         ret = process_event(machine, evlist, event, state);
 328                         perf_evlist__mmap_consume(evlist, i);
 329                         if (ret < 0)
 330                                 return ret;
 331                 }
 332         }
 333         return 0;
 334 }
 335
 336 static int comp(const void *a, const void *b)
 337 {
 338         return *(int *)a - *(int *)b;
 339 }
 340
 341 static void do_sort_something(void)
 342 {
 343         int buf[40960], i;
 344
 345         for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
 346                 buf[i] = ARRAY_SIZE(buf) - i - 1;
 347
 348         qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
 349
 350         for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
 351                 if (buf[i] != i) {
 352                         pr_debug("qsort failed\n");
 353                         break;
 354                 }
 355         }
 356 }
 357
 358 static void sort_something(void)
 359 {
 360         int i;
 361
 362         for (i = 0; i < 10; i++)
 363                 do_sort_something();
 364 }
 365
 366 static void syscall_something(void)
 367 {
 368         int pipefd[2];
 369         int i;
 370
 371         for (i = 0; i < 1000; i++) {
 372                 if (pipe(pipefd) < 0) {
 373                         pr_debug("pipe failed\n");
 374                         break;
 375                 }
 376                 close(pipefd[1]);
 377                 close(pipefd[0]);
 378         }
 379 }
 380
 381 static void fs_something(void)
 382 {
 383         const char *test_file_name = "temp-perf-code-reading-test-file--";
 384         FILE *f;
 385         int i;
 386
 387         for (i = 0; i < 1000; i++) {
 388                 f = fopen(test_file_name, "w+");
 389                 if (f) {
 390                         fclose(f);
 391                         unlink(test_file_name);
 392                 }
 393         }
 394 }
 395
 396 static void do_something(void)
 397 {
 398         fs_something();
 399
 400         sort_something();
 401
 402         syscall_something();
 403 }
 404
 405 enum {
 406         TEST_CODE_READING_OK,
 407         TEST_CODE_READING_NO_VMLINUX,
 408         TEST_CODE_READING_NO_KCORE,
 409         TEST_CODE_READING_NO_ACCESS,
 410         TEST_CODE_READING_NO_KERNEL_OBJ,
 411 };
 412
 413 static int do_test_code_reading(bool try_kcore)
 414 {
 415         struct machines machines;
 416         struct machine *machine;
 417         struct thread *thread;
 418         struct record_opts opts = {
 419                 .mmap_pages          = UINT_MAX,
 420                 .user_freq           = UINT_MAX,
 421                 .user_interval       = ULLONG_MAX,
 422                 .freq                = 4000,
 423                 .target              = {
 424                         .uses_mmap   = true,
 425                 },
 426         };
 427         struct state state = {
 428                 .done_cnt = 0,
 429         };
 430         struct thread_map *threads = NULL;
 431         struct cpu_map *cpus = NULL;
 432         struct perf_evlist *evlist = NULL;
 433         struct perf_evsel *evsel = NULL;
 434         int err = -1, ret;
 435         pid_t pid;
 436         struct map *map;
 437         bool have_vmlinux, have_kcore, excl_kernel = false;
 438
 439         pid = getpid();
 440
 441         machines__init(&machines);
 442         machine = &machines.host;
 443
 444         ret = machine__create_kernel_maps(machine);
 445         if (ret < 0) {
 446                 pr_debug("machine__create_kernel_maps failed\n");
 447                 goto out_err;
 448         }
 449
 450         /* Force the use of kallsyms instead of vmlinux to try kcore */
 451         if (try_kcore)
 452                 symbol_conf.kallsyms_name = "/proc/kallsyms";
 453
 454         /* Load kernel map */
 455         map = machine->vmlinux_maps[MAP__FUNCTION];
 456         ret = map__load(map, NULL);
 457         if (ret < 0) {
 458                 pr_debug("map__load failed\n");
 459                 goto out_err;
 460         }
 461         have_vmlinux = dso__is_vmlinux(map->dso);
 462         have_kcore = dso__is_kcore(map->dso);
 463
 464         /* 2nd time through we just try kcore */
 465         if (try_kcore && !have_kcore)
 466                 return TEST_CODE_READING_NO_KCORE;
 467
 468         /* No point getting kernel events if there is no kernel object */
 469         if (!have_vmlinux && !have_kcore)
 470                 excl_kernel = true;
 471
 472         threads = thread_map__new_by_tid(pid);
 473         if (!threads) {
 474                 pr_debug("thread_map__new_by_tid failed\n");
 475                 goto out_err;
 476         }
 477
 478         ret = perf_event__synthesize_thread_map(NULL, threads,
 479                                                 perf_event__process, machine, false, 500);
 480         if (ret < 0) {
 481                 pr_debug("perf_event__synthesize_thread_map failed\n");
 482                 goto out_err;
 483         }
 484
 485         thread = machine__findnew_thread(machine, pid, pid);
 486         if (!thread) {
 487                 pr_debug("machine__findnew_thread failed\n");
 488                 goto out_put;
 489         }
 490
 491         cpus = cpu_map__new(NULL);
 492         if (!cpus) {
 493                 pr_debug("cpu_map__new failed\n");
 494                 goto out_put;
 495         }
 496
 497         while (1) {
 498                 const char *str;
 499
 500                 evlist = perf_evlist__new();
 501                 if (!evlist) {
 502                         pr_debug("perf_evlist__new failed\n");
 503                         goto out_put;
 504                 }
 505
 506                 perf_evlist__set_maps(evlist, cpus, threads);
 507
 508                 if (excl_kernel)
 509                         str = "cycles:u";
 510                 else
 511                         str = "cycles";
 512                 pr_debug("Parsing event '%s'\n", str);
 513                 ret = parse_events(evlist, str, NULL);
 514                 if (ret < 0) {
 515                         pr_debug("parse_events failed\n");
 516                         goto out_put;
 517                 }
 518
 519                 perf_evlist__config(evlist, &opts);
 520
 521                 evsel = perf_evlist__first(evlist);
 522
 523                 evsel->attr.comm = 1;
 524                 evsel->attr.disabled = 1;
 525                 evsel->attr.enable_on_exec = 0;
 526
 527                 ret = perf_evlist__open(evlist);
 528                 if (ret < 0) {
 529                         if (!excl_kernel) {
 530                                 excl_kernel = true;
 531                                 perf_evlist__set_maps(evlist, NULL, NULL);
 532                                 perf_evlist__delete(evlist);
 533                                 evlist = NULL;
 534                                 continue;
 535                         }
 536                         pr_debug("perf_evlist__open failed\n");
 537                         goto out_put;
 538                 }
 539                 break;
 540         }
 541
 542         ret = perf_evlist__mmap(evlist, UINT_MAX, false);
 543         if (ret < 0) {
 544                 pr_debug("perf_evlist__mmap failed\n");
 545                 goto out_put;
 546         }
 547
 548         perf_evlist__enable(evlist);
 549
 550         do_something();
 551
 552         perf_evlist__disable(evlist);
 553
 554         ret = process_events(machine, evlist, &state);
 555         if (ret < 0)
 556                 goto out_put;
 557
 558         if (!have_vmlinux && !have_kcore && !try_kcore)
 559                 err = TEST_CODE_READING_NO_KERNEL_OBJ;
 560         else if (!have_vmlinux && !try_kcore)
 561                 err = TEST_CODE_READING_NO_VMLINUX;
 562         else if (excl_kernel)
 563                 err = TEST_CODE_READING_NO_ACCESS;
 564         else
 565                 err = TEST_CODE_READING_OK;
 566 out_put:
 567         thread__put(thread);
 568 out_err:
 569
 570         if (evlist) {
 571                 perf_evlist__delete(evlist);
 572         } else {
 573                 cpu_map__put(cpus);
 574                 thread_map__put(threads);
 575         }
 576         machines__destroy_kernel_maps(&machines);
 577         machine__delete_threads(machine);
 578         machines__exit(&machines);
 579
 580         return err;
 581 }
 582
 583 int test__code_reading(void)
 584 {
 585         int ret;
 586
 587         ret = do_test_code_reading(false);
 588         if (!ret)
 589                 ret = do_test_code_reading(true);
 590
 591         switch (ret) {
 592         case TEST_CODE_READING_OK:
 593                 return 0;
 594         case TEST_CODE_READING_NO_VMLINUX:
 595                 fprintf(stderr, " (no vmlinux)");
 596                 return 0;
 597         case TEST_CODE_READING_NO_KCORE:
 598                 fprintf(stderr, " (no kcore)");
 599                 return 0;
 600         case TEST_CODE_READING_NO_ACCESS:
 601                 fprintf(stderr, " (no access)");
 602                 return 0;
 603         case TEST_CODE_READING_NO_KERNEL_OBJ:
 604                 fprintf(stderr, " (no kernel obj)");
 605                 return 0;
 606         default:
 607                 return -1;
 608         };
 609 }