tools/perf/tests/code-reading.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <errno.h>
   3 #include <linux/kernel.h>
   4 #include <linux/types.h>
   5 #include <inttypes.h>
   6 #include <stdlib.h>
   7 #include <unistd.h>
   8 #include <stdio.h>
   9 #include <string.h>
  10 #include <sys/param.h>
  11
  12 #include "parse-events.h"
  13 #include "evlist.h"
  14 #include "evsel.h"
  15 #include "thread_map.h"
  16 #include "cpumap.h"
  17 #include "machine.h"
  18 #include "event.h"
  19 #include "thread.h"
  20
  21 #include "tests.h"
  22
  23 #include "sane_ctype.h"
  24
  25 #define BUFSZ   1024
  26 #define READLEN 128
  27
  28 struct state {
  29         u64 done[1024];
  30         size_t done_cnt;
  31 };
  32
  33 static unsigned int hex(char c)
  34 {
  35         if (c >= '0' && c <= '9')
  36                 return c - '0';
  37         if (c >= 'a' && c <= 'f')
  38                 return c - 'a' + 10;
  39         return c - 'A' + 10;
  40 }
  41
  42 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
  43                                  size_t *buf_len)
  44 {
  45         size_t bytes_read = 0;
  46         unsigned char *chunk_start = *buf;
  47
  48         /* Read bytes */
  49         while (*buf_len > 0) {
  50                 char c1, c2;
  51
  52                 /* Get 2 hex digits */
  53                 c1 = *(*line)++;
  54                 if (!isxdigit(c1))
  55                         break;
  56                 c2 = *(*line)++;
  57                 if (!isxdigit(c2))
  58                         break;
  59
  60                 /* Store byte and advance buf */
  61                 **buf = (hex(c1) << 4) | hex(c2);
  62                 (*buf)++;
  63                 (*buf_len)--;
  64                 bytes_read++;
  65
  66                 /* End of chunk? */
  67                 if (isspace(**line))
  68                         break;
  69         }
  70
  71         /*
  72          * objdump will display raw insn as LE if code endian
  73          * is LE and bytes_per_chunk > 1. In that case reverse
  74          * the chunk we just read.
  75          *
  76          * see disassemble_bytes() at binutils/objdump.c for details
  77          * how objdump chooses display endian)
  78          */
  79         if (bytes_read > 1 && !bigendian()) {
  80                 unsigned char *chunk_end = chunk_start + bytes_read - 1;
  81                 unsigned char tmp;
  82
  83                 while (chunk_start < chunk_end) {
  84                         tmp = *chunk_start;
  85                         *chunk_start = *chunk_end;
  86                         *chunk_end = tmp;
  87                         chunk_start++;
  88                         chunk_end--;
  89                 }
  90         }
  91
  92         return bytes_read;
  93 }
  94
  95 static size_t read_objdump_line(const char *line, unsigned char *buf,
  96                                 size_t buf_len)
  97 {
  98         const char *p;
  99         size_t ret, bytes_read = 0;
 100
 101         /* Skip to a colon */
 102         p = strchr(line, ':');
 103         if (!p)
 104                 return 0;
 105         p++;
 106
 107         /* Skip initial spaces */
 108         while (*p) {
 109                 if (!isspace(*p))
 110                         break;
 111                 p++;
 112         }
 113
 114         do {
 115                 ret = read_objdump_chunk(&p, &buf, &buf_len);
 116                 bytes_read += ret;
 117                 p++;
 118         } while (ret > 0);
 119
 120         /* return number of successfully read bytes */
 121         return bytes_read;
 122 }
 123
 124 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
 125 {
 126         char *line = NULL;
 127         size_t line_len, off_last = 0;
 128         ssize_t ret;
 129         int err = 0;
 130         u64 addr, last_addr = start_addr;
 131
 132         while (off_last < *len) {
 133                 size_t off, read_bytes, written_bytes;
 134                 unsigned char tmp[BUFSZ];
 135
 136                 ret = getline(&line, &line_len, f);
 137                 if (feof(f))
 138                         break;
 139                 if (ret < 0) {
 140                         pr_debug("getline failed\n");
 141                         err = -1;
 142                         break;
 143                 }
 144
 145                 /* read objdump data into temporary buffer */
 146                 read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
 147                 if (!read_bytes)
 148                         continue;
 149
 150                 if (sscanf(line, "%"PRIx64, &addr) != 1)
 151                         continue;
 152                 if (addr < last_addr) {
 153                         pr_debug("addr going backwards, read beyond section?\n");
 154                         break;
 155                 }
 156                 last_addr = addr;
 157
 158                 /* copy it from temporary buffer to 'buf' according
 159                  * to address on current objdump line */
 160                 off = addr - start_addr;
 161                 if (off >= *len)
 162                         break;
 163                 written_bytes = MIN(read_bytes, *len - off);
 164                 memcpy(buf + off, tmp, written_bytes);
 165                 off_last = off + written_bytes;
 166         }
 167
 168         /* len returns number of bytes that could not be read */
 169         *len -= off_last;
 170
 171         free(line);
 172
 173         return err;
 174 }
 175
 176 static int read_via_objdump(const char *filename, u64 addr, void *buf,
 177                             size_t len)
 178 {
 179         char cmd[PATH_MAX * 2];
 180         const char *fmt;
 181         FILE *f;
 182         int ret;
 183
 184         fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
 185         ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
 186                        filename);
 187         if (ret <= 0 || (size_t)ret >= sizeof(cmd))
 188                 return -1;
 189
 190         pr_debug("Objdump command is: %s\n", cmd);
 191
 192         /* Ignore objdump errors */
 193         strcat(cmd, " 2>/dev/null");
 194
 195         f = popen(cmd, "r");
 196         if (!f) {
 197                 pr_debug("popen failed\n");
 198                 return -1;
 199         }
 200
 201         ret = read_objdump_output(f, buf, &len, addr);
 202         if (len) {
 203                 pr_debug("objdump read too few bytes: %zd\n", len);
 204                 if (!ret)
 205                         ret = len;
 206         }
 207
 208         pclose(f);
 209
 210         return ret;
 211 }
 212
 213 static void dump_buf(unsigned char *buf, size_t len)
 214 {
 215         size_t i;
 216
 217         for (i = 0; i < len; i++) {
 218                 pr_debug("0x%02x ", buf[i]);
 219                 if (i % 16 == 15)
 220                         pr_debug("\n");
 221         }
 222         pr_debug("\n");
 223 }
 224
 225 static int read_object_code(u64 addr, size_t len, u8 cpumode,
 226                             struct thread *thread, struct state *state)
 227 {
 228         struct addr_location al;
 229         unsigned char buf1[BUFSZ];
 230         unsigned char buf2[BUFSZ];
 231         size_t ret_len;
 232         u64 objdump_addr;
 233         const char *objdump_name;
 234         char decomp_name[KMOD_DECOMP_LEN];
 235         int ret;
 236
 237         pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
 238
 239         thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
 240         if (!al.map || !al.map->dso) {
 241                 if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
 242                         pr_debug("Hypervisor address can not be resolved - skipping\n");
 243                         return 0;
 244                 }
 245
 246                 pr_debug("thread__find_addr_map failed\n");
 247                 return -1;
 248         }
 249
 250         pr_debug("File is: %s\n", al.map->dso->long_name);
 251
 252         if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 253             !dso__is_kcore(al.map->dso)) {
 254                 pr_debug("Unexpected kernel address - skipping\n");
 255                 return 0;
 256         }
 257
 258         pr_debug("On file address is: %#"PRIx64"\n", al.addr);
 259
 260         if (len > BUFSZ)
 261                 len = BUFSZ;
 262
 263         /* Do not go off the map */
 264         if (addr + len > al.map->end)
 265                 len = al.map->end - addr;
 266
 267         /* Read the object code using perf */
 268         ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
 269                                         al.addr, buf1, len);
 270         if (ret_len != len) {
 271                 pr_debug("dso__data_read_offset failed\n");
 272                 return -1;
 273         }
 274
 275         /*
 276          * Converting addresses for use by objdump requires more information.
 277          * map__load() does that.  See map__rip_2objdump() for details.
 278          */
 279         if (map__load(al.map))
 280                 return -1;
 281
 282         /* objdump struggles with kcore - try each map only once */
 283         if (dso__is_kcore(al.map->dso)) {
 284                 size_t d;
 285
 286                 for (d = 0; d < state->done_cnt; d++) {
 287                         if (state->done[d] == al.map->start) {
 288                                 pr_debug("kcore map tested already");
 289                                 pr_debug(" - skipping\n");
 290                                 return 0;
 291                         }
 292                 }
 293                 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
 294                         pr_debug("Too many kcore maps - skipping\n");
 295                         return 0;
 296                 }
 297                 state->done[state->done_cnt++] = al.map->start;
 298         }
 299
 300         objdump_name = al.map->dso->long_name;
 301         if (dso__needs_decompress(al.map->dso)) {
 302                 if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
 303                                                  decomp_name,
 304                                                  sizeof(decomp_name)) < 0) {
 305                         pr_debug("decompression failed\n");
 306                         return -1;
 307                 }
 308
 309                 objdump_name = decomp_name;
 310         }
 311
 312         /* Read the object code using objdump */
 313         objdump_addr = map__rip_2objdump(al.map, al.addr);
 314         ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
 315
 316         if (dso__needs_decompress(al.map->dso))
 317                 unlink(objdump_name);
 318
 319         if (ret > 0) {
 320                 /*
 321                  * The kernel maps are inaccurate - assume objdump is right in
 322                  * that case.
 323                  */
 324                 if (cpumode == PERF_RECORD_MISC_KERNEL ||
 325                     cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
 326                         len -= ret;
 327                         if (len) {
 328                                 pr_debug("Reducing len to %zu\n", len);
 329                         } else if (dso__is_kcore(al.map->dso)) {
 330                                 /*
 331                                  * objdump cannot handle very large segments
 332                                  * that may be found in kcore.
 333                                  */
 334                                 pr_debug("objdump failed for kcore");
 335                                 pr_debug(" - skipping\n");
 336                                 return 0;
 337                         } else {
 338                                 return -1;
 339                         }
 340                 }
 341         }
 342         if (ret < 0) {
 343                 pr_debug("read_via_objdump failed\n");
 344                 return -1;
 345         }
 346
 347         /* The results should be identical */
 348         if (memcmp(buf1, buf2, len)) {
 349                 pr_debug("Bytes read differ from those read by objdump\n");
 350                 pr_debug("buf1 (dso):\n");
 351                 dump_buf(buf1, len);
 352                 pr_debug("buf2 (objdump):\n");
 353                 dump_buf(buf2, len);
 354                 return -1;
 355         }
 356         pr_debug("Bytes read match those read by objdump\n");
 357
 358         return 0;
 359 }
 360
 361 static int process_sample_event(struct machine *machine,
 362                                 struct perf_evlist *evlist,
 363                                 union perf_event *event, struct state *state)
 364 {
 365         struct perf_sample sample;
 366         struct thread *thread;
 367         int ret;
 368
 369         if (perf_evlist__parse_sample(evlist, event, &sample)) {
 370                 pr_debug("perf_evlist__parse_sample failed\n");
 371                 return -1;
 372         }
 373
 374         thread = machine__findnew_thread(machine, sample.pid, sample.tid);
 375         if (!thread) {
 376                 pr_debug("machine__findnew_thread failed\n");
 377                 return -1;
 378         }
 379
 380         ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
 381         thread__put(thread);
 382         return ret;
 383 }
 384
 385 static int process_event(struct machine *machine, struct perf_evlist *evlist,
 386                          union perf_event *event, struct state *state)
 387 {
 388         if (event->header.type == PERF_RECORD_SAMPLE)
 389                 return process_sample_event(machine, evlist, event, state);
 390
 391         if (event->header.type == PERF_RECORD_THROTTLE ||
 392             event->header.type == PERF_RECORD_UNTHROTTLE)
 393                 return 0;
 394
 395         if (event->header.type < PERF_RECORD_MAX) {
 396                 int ret;
 397
 398                 ret = machine__process_event(machine, event, NULL);
 399                 if (ret < 0)
 400                         pr_debug("machine__process_event failed, event type %u\n",
 401                                  event->header.type);
 402                 return ret;
 403         }
 404
 405         return 0;
 406 }
 407
 408 static int process_events(struct machine *machine, struct perf_evlist *evlist,
 409                           struct state *state)
 410 {
 411         union perf_event *event;
 412         int i, ret;
 413
 414         for (i = 0; i < evlist->nr_mmaps; i++) {
 415                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 416                         ret = process_event(machine, evlist, event, state);
 417                         perf_evlist__mmap_consume(evlist, i);
 418                         if (ret < 0)
 419                                 return ret;
 420                 }
 421         }
 422         return 0;
 423 }
 424
 425 static int comp(const void *a, const void *b)
 426 {
 427         return *(int *)a - *(int *)b;
 428 }
 429
 430 static void do_sort_something(void)
 431 {
 432         int buf[40960], i;
 433
 434         for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
 435                 buf[i] = ARRAY_SIZE(buf) - i - 1;
 436
 437         qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
 438
 439         for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
 440                 if (buf[i] != i) {
 441                         pr_debug("qsort failed\n");
 442                         break;
 443                 }
 444         }
 445 }
 446
 447 static void sort_something(void)
 448 {
 449         int i;
 450
 451         for (i = 0; i < 10; i++)
 452                 do_sort_something();
 453 }
 454
 455 static void syscall_something(void)
 456 {
 457         int pipefd[2];
 458         int i;
 459
 460         for (i = 0; i < 1000; i++) {
 461                 if (pipe(pipefd) < 0) {
 462                         pr_debug("pipe failed\n");
 463                         break;
 464                 }
 465                 close(pipefd[1]);
 466                 close(pipefd[0]);
 467         }
 468 }
 469
 470 static void fs_something(void)
 471 {
 472         const char *test_file_name = "temp-perf-code-reading-test-file--";
 473         FILE *f;
 474         int i;
 475
 476         for (i = 0; i < 1000; i++) {
 477                 f = fopen(test_file_name, "w+");
 478                 if (f) {
 479                         fclose(f);
 480                         unlink(test_file_name);
 481                 }
 482         }
 483 }
 484
 485 static void do_something(void)
 486 {
 487         fs_something();
 488
 489         sort_something();
 490
 491         syscall_something();
 492 }
 493
 494 enum {
 495         TEST_CODE_READING_OK,
 496         TEST_CODE_READING_NO_VMLINUX,
 497         TEST_CODE_READING_NO_KCORE,
 498         TEST_CODE_READING_NO_ACCESS,
 499         TEST_CODE_READING_NO_KERNEL_OBJ,
 500 };
 501
 502 static int do_test_code_reading(bool try_kcore)
 503 {
 504         struct machine *machine;
 505         struct thread *thread;
 506         struct record_opts opts = {
 507                 .mmap_pages          = UINT_MAX,
 508                 .user_freq           = UINT_MAX,
 509                 .user_interval       = ULLONG_MAX,
 510                 .freq                = 500,
 511                 .target              = {
 512                         .uses_mmap   = true,
 513                 },
 514         };
 515         struct state state = {
 516                 .done_cnt = 0,
 517         };
 518         struct thread_map *threads = NULL;
 519         struct cpu_map *cpus = NULL;
 520         struct perf_evlist *evlist = NULL;
 521         struct perf_evsel *evsel = NULL;
 522         int err = -1, ret;
 523         pid_t pid;
 524         struct map *map;
 525         bool have_vmlinux, have_kcore, excl_kernel = false;
 526
 527         pid = getpid();
 528
 529         machine = machine__new_host();
 530
 531         ret = machine__create_kernel_maps(machine);
 532         if (ret < 0) {
 533                 pr_debug("machine__create_kernel_maps failed\n");
 534                 goto out_err;
 535         }
 536
 537         /* Force the use of kallsyms instead of vmlinux to try kcore */
 538         if (try_kcore)
 539                 symbol_conf.kallsyms_name = "/proc/kallsyms";
 540
 541         /* Load kernel map */
 542         map = machine__kernel_map(machine);
 543         ret = map__load(map);
 544         if (ret < 0) {
 545                 pr_debug("map__load failed\n");
 546                 goto out_err;
 547         }
 548         have_vmlinux = dso__is_vmlinux(map->dso);
 549         have_kcore = dso__is_kcore(map->dso);
 550
 551         /* 2nd time through we just try kcore */
 552         if (try_kcore && !have_kcore)
 553                 return TEST_CODE_READING_NO_KCORE;
 554
 555         /* No point getting kernel events if there is no kernel object */
 556         if (!have_vmlinux && !have_kcore)
 557                 excl_kernel = true;
 558
 559         threads = thread_map__new_by_tid(pid);
 560         if (!threads) {
 561                 pr_debug("thread_map__new_by_tid failed\n");
 562                 goto out_err;
 563         }
 564
 565         ret = perf_event__synthesize_thread_map(NULL, threads,
 566                                                 perf_event__process, machine, false, 500);
 567         if (ret < 0) {
 568                 pr_debug("perf_event__synthesize_thread_map failed\n");
 569                 goto out_err;
 570         }
 571
 572         thread = machine__findnew_thread(machine, pid, pid);
 573         if (!thread) {
 574                 pr_debug("machine__findnew_thread failed\n");
 575                 goto out_put;
 576         }
 577
 578         cpus = cpu_map__new(NULL);
 579         if (!cpus) {
 580                 pr_debug("cpu_map__new failed\n");
 581                 goto out_put;
 582         }
 583
 584         while (1) {
 585                 const char *str;
 586
 587                 evlist = perf_evlist__new();
 588                 if (!evlist) {
 589                         pr_debug("perf_evlist__new failed\n");
 590                         goto out_put;
 591                 }
 592
 593                 perf_evlist__set_maps(evlist, cpus, threads);
 594
 595                 if (excl_kernel)
 596                         str = "cycles:u";
 597                 else
 598                         str = "cycles";
 599                 pr_debug("Parsing event '%s'\n", str);
 600                 ret = parse_events(evlist, str, NULL);
 601                 if (ret < 0) {
 602                         pr_debug("parse_events failed\n");
 603                         goto out_put;
 604                 }
 605
 606                 perf_evlist__config(evlist, &opts, NULL);
 607
 608                 evsel = perf_evlist__first(evlist);
 609
 610                 evsel->attr.comm = 1;
 611                 evsel->attr.disabled = 1;
 612                 evsel->attr.enable_on_exec = 0;
 613
 614                 ret = perf_evlist__open(evlist);
 615                 if (ret < 0) {
 616                         if (!excl_kernel) {
 617                                 excl_kernel = true;
 618                                 /*
 619                                  * Both cpus and threads are now owned by evlist
 620                                  * and will be freed by following perf_evlist__set_maps
 621                                  * call. Getting refference to keep them alive.
 622                                  */
 623                                 cpu_map__get(cpus);
 624                                 thread_map__get(threads);
 625                                 perf_evlist__set_maps(evlist, NULL, NULL);
 626                                 perf_evlist__delete(evlist);
 627                                 evlist = NULL;
 628                                 continue;
 629                         }
 630
 631                         if (verbose > 0) {
 632                                 char errbuf[512];
 633                                 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
 634                                 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
 635                         }
 636
 637                         goto out_put;
 638                 }
 639                 break;
 640         }
 641
 642         ret = perf_evlist__mmap(evlist, UINT_MAX, false);
 643         if (ret < 0) {
 644                 pr_debug("perf_evlist__mmap failed\n");
 645                 goto out_put;
 646         }
 647
 648         perf_evlist__enable(evlist);
 649
 650         do_something();
 651
 652         perf_evlist__disable(evlist);
 653
 654         ret = process_events(machine, evlist, &state);
 655         if (ret < 0)
 656                 goto out_put;
 657
 658         if (!have_vmlinux && !have_kcore && !try_kcore)
 659                 err = TEST_CODE_READING_NO_KERNEL_OBJ;
 660         else if (!have_vmlinux && !try_kcore)
 661                 err = TEST_CODE_READING_NO_VMLINUX;
 662         else if (excl_kernel)
 663                 err = TEST_CODE_READING_NO_ACCESS;
 664         else
 665                 err = TEST_CODE_READING_OK;
 666 out_put:
 667         thread__put(thread);
 668 out_err:
 669
 670         if (evlist) {
 671                 perf_evlist__delete(evlist);
 672         } else {
 673                 cpu_map__put(cpus);
 674                 thread_map__put(threads);
 675         }
 676         machine__delete_threads(machine);
 677         machine__delete(machine);
 678
 679         return err;
 680 }
 681
 682 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
 683 {
 684         int ret;
 685
 686         ret = do_test_code_reading(false);
 687         if (!ret)
 688                 ret = do_test_code_reading(true);
 689
 690         switch (ret) {
 691         case TEST_CODE_READING_OK:
 692                 return 0;
 693         case TEST_CODE_READING_NO_VMLINUX:
 694                 pr_debug("no vmlinux\n");
 695                 return 0;
 696         case TEST_CODE_READING_NO_KCORE:
 697                 pr_debug("no kcore\n");
 698                 return 0;
 699         case TEST_CODE_READING_NO_ACCESS:
 700                 pr_debug("no access\n");
 701                 return 0;
 702         case TEST_CODE_READING_NO_KERNEL_OBJ:
 703                 pr_debug("no kernel obj\n");
 704                 return 0;
 705         default:
 706                 return -1;
 707         };
 708 }