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, last_addr = start_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                 if (addr < last_addr) {
 105                         pr_debug("addr going backwards, read beyond section?\n");
 106                         break;
 107                 }
 108                 last_addr = addr;
 109
 110                 /* copy it from temporary buffer to 'buf' according
 111                  * to address on current objdump line */
 112                 off = addr - start_addr;
 113                 if (off >= *len)
 114                         break;
 115                 written_bytes = MIN(read_bytes, *len - off);
 116                 memcpy(buf + off, tmp, written_bytes);
 117                 off_last = off + written_bytes;
 118         }
 119
 120         /* len returns number of bytes that could not be read */
 121         *len -= off_last;
 122
 123         free(line);
 124
 125         return err;
 126 }
 127
 128 static int read_via_objdump(const char *filename, u64 addr, void *buf,
 129                             size_t len)
 130 {
 131         char cmd[PATH_MAX * 2];
 132         const char *fmt;
 133         FILE *f;
 134         int ret;
 135
 136         fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
 137         ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
 138                        filename);
 139         if (ret <= 0 || (size_t)ret >= sizeof(cmd))
 140                 return -1;
 141
 142         pr_debug("Objdump command is: %s\n", cmd);
 143
 144         /* Ignore objdump errors */
 145         strcat(cmd, " 2>/dev/null");
 146
 147         f = popen(cmd, "r");
 148         if (!f) {
 149                 pr_debug("popen failed\n");
 150                 return -1;
 151         }
 152
 153         ret = read_objdump_output(f, buf, &len, addr);
 154         if (len) {
 155                 pr_debug("objdump read too few bytes\n");
 156                 if (!ret)
 157                         ret = len;
 158         }
 159
 160         pclose(f);
 161
 162         return ret;
 163 }
 164
 165 static int read_object_code(u64 addr, size_t len, u8 cpumode,
 166                             struct thread *thread, struct state *state)
 167 {
 168         struct addr_location al;
 169         unsigned char buf1[BUFSZ];
 170         unsigned char buf2[BUFSZ];
 171         size_t ret_len;
 172         u64 objdump_addr;
 173         int ret;
 174
 175         pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
 176
 177         thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
 178         if (!al.map || !al.map->dso) {
 179                 pr_debug("thread__find_addr_map failed\n");
 180                 return -1;
 181         }
 182
 183         pr_debug("File is: %s\n", al.map->dso->long_name);
 184
 185         if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 186             !dso__is_kcore(al.map->dso)) {
 187                 pr_debug("Unexpected kernel address - skipping\n");
 188                 return 0;
 189         }
 190
 191         pr_debug("On file address is: %#"PRIx64"\n", al.addr);
 192
 193         if (len > BUFSZ)
 194                 len = BUFSZ;
 195
 196         /* Do not go off the map */
 197         if (addr + len > al.map->end)
 198                 len = al.map->end - addr;
 199
 200         /* Read the object code using perf */
 201         ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
 202                                         al.addr, buf1, len);
 203         if (ret_len != len) {
 204                 pr_debug("dso__data_read_offset failed\n");
 205                 return -1;
 206         }
 207
 208         /*
 209          * Converting addresses for use by objdump requires more information.
 210          * map__load() does that.  See map__rip_2objdump() for details.
 211          */
 212         if (map__load(al.map, NULL))
 213                 return -1;
 214
 215         /* objdump struggles with kcore - try each map only once */
 216         if (dso__is_kcore(al.map->dso)) {
 217                 size_t d;
 218
 219                 for (d = 0; d < state->done_cnt; d++) {
 220                         if (state->done[d] == al.map->start) {
 221                                 pr_debug("kcore map tested already");
 222                                 pr_debug(" - skipping\n");
 223                                 return 0;
 224                         }
 225                 }
 226                 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
 227                         pr_debug("Too many kcore maps - skipping\n");
 228                         return 0;
 229                 }
 230                 state->done[state->done_cnt++] = al.map->start;
 231         }
 232
 233         /* Read the object code using objdump */
 234         objdump_addr = map__rip_2objdump(al.map, al.addr);
 235         ret = read_via_objdump(al.map->dso->long_name, objdump_addr, buf2, len);
 236         if (ret > 0) {
 237                 /*
 238                  * The kernel maps are inaccurate - assume objdump is right in
 239                  * that case.
 240                  */
 241                 if (cpumode == PERF_RECORD_MISC_KERNEL ||
 242                     cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
 243                         len -= ret;
 244                         if (len) {
 245                                 pr_debug("Reducing len to %zu\n", len);
 246                         } else if (dso__is_kcore(al.map->dso)) {
 247                                 /*
 248                                  * objdump cannot handle very large segments
 249                                  * that may be found in kcore.
 250                                  */
 251                                 pr_debug("objdump failed for kcore");
 252                                 pr_debug(" - skipping\n");
 253                                 return 0;
 254                         } else {
 255                                 return -1;
 256                         }
 257                 }
 258         }
 259         if (ret < 0) {
 260                 pr_debug("read_via_objdump failed\n");
 261                 return -1;
 262         }
 263
 264         /* The results should be identical */
 265         if (memcmp(buf1, buf2, len)) {
 266                 pr_debug("Bytes read differ from those read by objdump\n");
 267                 return -1;
 268         }
 269         pr_debug("Bytes read match those read by objdump\n");
 270
 271         return 0;
 272 }
 273
 274 static int process_sample_event(struct machine *machine,
 275                                 struct perf_evlist *evlist,
 276                                 union perf_event *event, struct state *state)
 277 {
 278         struct perf_sample sample;
 279         struct thread *thread;
 280         u8 cpumode;
 281         int ret;
 282
 283         if (perf_evlist__parse_sample(evlist, event, &sample)) {
 284                 pr_debug("perf_evlist__parse_sample failed\n");
 285                 return -1;
 286         }
 287
 288         thread = machine__findnew_thread(machine, sample.pid, sample.tid);
 289         if (!thread) {
 290                 pr_debug("machine__findnew_thread failed\n");
 291                 return -1;
 292         }
 293
 294         cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 295
 296         ret = read_object_code(sample.ip, READLEN, cpumode, thread, state);
 297         thread__put(thread);
 298         return ret;
 299 }
 300
 301 static int process_event(struct machine *machine, struct perf_evlist *evlist,
 302                          union perf_event *event, struct state *state)
 303 {
 304         if (event->header.type == PERF_RECORD_SAMPLE)
 305                 return process_sample_event(machine, evlist, event, state);
 306
 307         if (event->header.type == PERF_RECORD_THROTTLE ||
 308             event->header.type == PERF_RECORD_UNTHROTTLE)
 309                 return 0;
 310
 311         if (event->header.type < PERF_RECORD_MAX) {
 312                 int ret;
 313
 314                 ret = machine__process_event(machine, event, NULL);
 315                 if (ret < 0)
 316                         pr_debug("machine__process_event failed, event type %u\n",
 317                                  event->header.type);
 318                 return ret;
 319         }
 320
 321         return 0;
 322 }
 323
 324 static int process_events(struct machine *machine, struct perf_evlist *evlist,
 325                           struct state *state)
 326 {
 327         union perf_event *event;
 328         int i, ret;
 329
 330         for (i = 0; i < evlist->nr_mmaps; i++) {
 331                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 332                         ret = process_event(machine, evlist, event, state);
 333                         perf_evlist__mmap_consume(evlist, i);
 334                         if (ret < 0)
 335                                 return ret;
 336                 }
 337         }
 338         return 0;
 339 }
 340
 341 static int comp(const void *a, const void *b)
 342 {
 343         return *(int *)a - *(int *)b;
 344 }
 345
 346 static void do_sort_something(void)
 347 {
 348         int buf[40960], i;
 349
 350         for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
 351                 buf[i] = ARRAY_SIZE(buf) - i - 1;
 352
 353         qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
 354
 355         for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
 356                 if (buf[i] != i) {
 357                         pr_debug("qsort failed\n");
 358                         break;
 359                 }
 360         }
 361 }
 362
 363 static void sort_something(void)
 364 {
 365         int i;
 366
 367         for (i = 0; i < 10; i++)
 368                 do_sort_something();
 369 }
 370
 371 static void syscall_something(void)
 372 {
 373         int pipefd[2];
 374         int i;
 375
 376         for (i = 0; i < 1000; i++) {
 377                 if (pipe(pipefd) < 0) {
 378                         pr_debug("pipe failed\n");
 379                         break;
 380                 }
 381                 close(pipefd[1]);
 382                 close(pipefd[0]);
 383         }
 384 }
 385
 386 static void fs_something(void)
 387 {
 388         const char *test_file_name = "temp-perf-code-reading-test-file--";
 389         FILE *f;
 390         int i;
 391
 392         for (i = 0; i < 1000; i++) {
 393                 f = fopen(test_file_name, "w+");
 394                 if (f) {
 395                         fclose(f);
 396                         unlink(test_file_name);
 397                 }
 398         }
 399 }
 400
 401 static void do_something(void)
 402 {
 403         fs_something();
 404
 405         sort_something();
 406
 407         syscall_something();
 408 }
 409
 410 enum {
 411         TEST_CODE_READING_OK,
 412         TEST_CODE_READING_NO_VMLINUX,
 413         TEST_CODE_READING_NO_KCORE,
 414         TEST_CODE_READING_NO_ACCESS,
 415         TEST_CODE_READING_NO_KERNEL_OBJ,
 416 };
 417
 418 static int do_test_code_reading(bool try_kcore)
 419 {
 420         struct machines machines;
 421         struct machine *machine;
 422         struct thread *thread;
 423         struct record_opts opts = {
 424                 .mmap_pages          = UINT_MAX,
 425                 .user_freq           = UINT_MAX,
 426                 .user_interval       = ULLONG_MAX,
 427                 .freq                = 4000,
 428                 .target              = {
 429                         .uses_mmap   = true,
 430                 },
 431         };
 432         struct state state = {
 433                 .done_cnt = 0,
 434         };
 435         struct thread_map *threads = NULL;
 436         struct cpu_map *cpus = NULL;
 437         struct perf_evlist *evlist = NULL;
 438         struct perf_evsel *evsel = NULL;
 439         int err = -1, ret;
 440         pid_t pid;
 441         struct map *map;
 442         bool have_vmlinux, have_kcore, excl_kernel = false;
 443
 444         pid = getpid();
 445
 446         machines__init(&machines);
 447         machine = &machines.host;
 448
 449         ret = machine__create_kernel_maps(machine);
 450         if (ret < 0) {
 451                 pr_debug("machine__create_kernel_maps failed\n");
 452                 goto out_err;
 453         }
 454
 455         /* Force the use of kallsyms instead of vmlinux to try kcore */
 456         if (try_kcore)
 457                 symbol_conf.kallsyms_name = "/proc/kallsyms";
 458
 459         /* Load kernel map */
 460         map = machine->vmlinux_maps[MAP__FUNCTION];
 461         ret = map__load(map, NULL);
 462         if (ret < 0) {
 463                 pr_debug("map__load failed\n");
 464                 goto out_err;
 465         }
 466         have_vmlinux = dso__is_vmlinux(map->dso);
 467         have_kcore = dso__is_kcore(map->dso);
 468
 469         /* 2nd time through we just try kcore */
 470         if (try_kcore && !have_kcore)
 471                 return TEST_CODE_READING_NO_KCORE;
 472
 473         /* No point getting kernel events if there is no kernel object */
 474         if (!have_vmlinux && !have_kcore)
 475                 excl_kernel = true;
 476
 477         threads = thread_map__new_by_tid(pid);
 478         if (!threads) {
 479                 pr_debug("thread_map__new_by_tid failed\n");
 480                 goto out_err;
 481         }
 482
 483         ret = perf_event__synthesize_thread_map(NULL, threads,
 484                                                 perf_event__process, machine, false, 500);
 485         if (ret < 0) {
 486                 pr_debug("perf_event__synthesize_thread_map failed\n");
 487                 goto out_err;
 488         }
 489
 490         thread = machine__findnew_thread(machine, pid, pid);
 491         if (!thread) {
 492                 pr_debug("machine__findnew_thread failed\n");
 493                 goto out_put;
 494         }
 495
 496         cpus = cpu_map__new(NULL);
 497         if (!cpus) {
 498                 pr_debug("cpu_map__new failed\n");
 499                 goto out_put;
 500         }
 501
 502         while (1) {
 503                 const char *str;
 504
 505                 evlist = perf_evlist__new();
 506                 if (!evlist) {
 507                         pr_debug("perf_evlist__new failed\n");
 508                         goto out_put;
 509                 }
 510
 511                 perf_evlist__set_maps(evlist, cpus, threads);
 512
 513                 if (excl_kernel)
 514                         str = "cycles:u";
 515                 else
 516                         str = "cycles";
 517                 pr_debug("Parsing event '%s'\n", str);
 518                 ret = parse_events(evlist, str, NULL);
 519                 if (ret < 0) {
 520                         pr_debug("parse_events failed\n");
 521                         goto out_put;
 522                 }
 523
 524                 perf_evlist__config(evlist, &opts);
 525
 526                 evsel = perf_evlist__first(evlist);
 527
 528                 evsel->attr.comm = 1;
 529                 evsel->attr.disabled = 1;
 530                 evsel->attr.enable_on_exec = 0;
 531
 532                 ret = perf_evlist__open(evlist);
 533                 if (ret < 0) {
 534                         if (!excl_kernel) {
 535                                 excl_kernel = true;
 536                                 perf_evlist__set_maps(evlist, NULL, NULL);
 537                                 perf_evlist__delete(evlist);
 538                                 evlist = NULL;
 539                                 continue;
 540                         }
 541                         pr_debug("perf_evlist__open failed\n");
 542                         goto out_put;
 543                 }
 544                 break;
 545         }
 546
 547         ret = perf_evlist__mmap(evlist, UINT_MAX, false);
 548         if (ret < 0) {
 549                 pr_debug("perf_evlist__mmap failed\n");
 550                 goto out_put;
 551         }
 552
 553         perf_evlist__enable(evlist);
 554
 555         do_something();
 556
 557         perf_evlist__disable(evlist);
 558
 559         ret = process_events(machine, evlist, &state);
 560         if (ret < 0)
 561                 goto out_put;
 562
 563         if (!have_vmlinux && !have_kcore && !try_kcore)
 564                 err = TEST_CODE_READING_NO_KERNEL_OBJ;
 565         else if (!have_vmlinux && !try_kcore)
 566                 err = TEST_CODE_READING_NO_VMLINUX;
 567         else if (excl_kernel)
 568                 err = TEST_CODE_READING_NO_ACCESS;
 569         else
 570                 err = TEST_CODE_READING_OK;
 571 out_put:
 572         thread__put(thread);
 573 out_err:
 574
 575         if (evlist) {
 576                 perf_evlist__delete(evlist);
 577         } else {
 578                 cpu_map__put(cpus);
 579                 thread_map__put(threads);
 580         }
 581         machines__destroy_kernel_maps(&machines);
 582         machine__delete_threads(machine);
 583         machines__exit(&machines);
 584
 585         return err;
 586 }
 587
 588 int test__code_reading(void)
 589 {
 590         int ret;
 591
 592         ret = do_test_code_reading(false);
 593         if (!ret)
 594                 ret = do_test_code_reading(true);
 595
 596         switch (ret) {
 597         case TEST_CODE_READING_OK:
 598                 return 0;
 599         case TEST_CODE_READING_NO_VMLINUX:
 600                 fprintf(stderr, " (no vmlinux)");
 601                 return 0;
 602         case TEST_CODE_READING_NO_KCORE:
 603                 fprintf(stderr, " (no kcore)");
 604                 return 0;
 605         case TEST_CODE_READING_NO_ACCESS:
 606                 fprintf(stderr, " (no access)");
 607                 return 0;
 608         case TEST_CODE_READING_NO_KERNEL_OBJ:
 609                 fprintf(stderr, " (no kernel obj)");
 610                 return 0;
 611         default:
 612                 return -1;
 613         };
 614 }