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 #include <perf/cpumap.h>
  12 #include <perf/evlist.h>
  13
  14 #include "debug.h"
  15 #include "dso.h"
  16 #include "env.h"
  17 #include "parse-events.h"
  18 #include "trace-event.h"
  19 #include "evlist.h"
  20 #include "evsel.h"
  21 #include "thread_map.h"
  22 #include "machine.h"
  23 #include "map.h"
  24 #include "symbol.h"
  25 #include "event.h"
  26 #include "record.h"
  27 #include "util/synthetic-events.h"
  28 #include "thread.h"
  29
  30 #include "tests.h"
  31
  32 #include <linux/ctype.h>
  33
  34 #define BUFSZ   1024
  35 #define READLEN 128
  36
  37 struct state {
  38         u64 done[1024];
  39         size_t done_cnt;
  40 };
  41
  42 static unsigned int hex(char c)
  43 {
  44         if (c >= '0' && c <= '9')
  45                 return c - '0';
  46         if (c >= 'a' && c <= 'f')
  47                 return c - 'a' + 10;
  48         return c - 'A' + 10;
  49 }
  50
  51 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
  52                                  size_t *buf_len)
  53 {
  54         size_t bytes_read = 0;
  55         unsigned char *chunk_start = *buf;
  56
  57         /* Read bytes */
  58         while (*buf_len > 0) {
  59                 char c1, c2;
  60
  61                 /* Get 2 hex digits */
  62                 c1 = *(*line)++;
  63                 if (!isxdigit(c1))
  64                         break;
  65                 c2 = *(*line)++;
  66                 if (!isxdigit(c2))
  67                         break;
  68
  69                 /* Store byte and advance buf */
  70                 **buf = (hex(c1) << 4) | hex(c2);
  71                 (*buf)++;
  72                 (*buf_len)--;
  73                 bytes_read++;
  74
  75                 /* End of chunk? */
  76                 if (isspace(**line))
  77                         break;
  78         }
  79
  80         /*
  81          * objdump will display raw insn as LE if code endian
  82          * is LE and bytes_per_chunk > 1. In that case reverse
  83          * the chunk we just read.
  84          *
  85          * see disassemble_bytes() at binutils/objdump.c for details
  86          * how objdump chooses display endian)
  87          */
  88         if (bytes_read > 1 && !bigendian()) {
  89                 unsigned char *chunk_end = chunk_start + bytes_read - 1;
  90                 unsigned char tmp;
  91
  92                 while (chunk_start < chunk_end) {
  93                         tmp = *chunk_start;
  94                         *chunk_start = *chunk_end;
  95                         *chunk_end = tmp;
  96                         chunk_start++;
  97                         chunk_end--;
  98                 }
  99         }
 100
 101         return bytes_read;
 102 }
 103
 104 static size_t read_objdump_line(const char *line, unsigned char *buf,
 105                                 size_t buf_len)
 106 {
 107         const char *p;
 108         size_t ret, bytes_read = 0;
 109
 110         /* Skip to a colon */
 111         p = strchr(line, ':');
 112         if (!p)
 113                 return 0;
 114         p++;
 115
 116         /* Skip initial spaces */
 117         while (*p) {
 118                 if (!isspace(*p))
 119                         break;
 120                 p++;
 121         }
 122
 123         do {
 124                 ret = read_objdump_chunk(&p, &buf, &buf_len);
 125                 bytes_read += ret;
 126                 p++;
 127         } while (ret > 0);
 128
 129         /* return number of successfully read bytes */
 130         return bytes_read;
 131 }
 132
 133 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
 134 {
 135         char *line = NULL;
 136         size_t line_len, off_last = 0;
 137         ssize_t ret;
 138         int err = 0;
 139         u64 addr, last_addr = start_addr;
 140
 141         while (off_last < *len) {
 142                 size_t off, read_bytes, written_bytes;
 143                 unsigned char tmp[BUFSZ];
 144
 145                 ret = getline(&line, &line_len, f);
 146                 if (feof(f))
 147                         break;
 148                 if (ret < 0) {
 149                         pr_debug("getline failed\n");
 150                         err = -1;
 151                         break;
 152                 }
 153
 154                 /* read objdump data into temporary buffer */
 155                 read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
 156                 if (!read_bytes)
 157                         continue;
 158
 159                 if (sscanf(line, "%"PRIx64, &addr) != 1)
 160                         continue;
 161                 if (addr < last_addr) {
 162                         pr_debug("addr going backwards, read beyond section?\n");
 163                         break;
 164                 }
 165                 last_addr = addr;
 166
 167                 /* copy it from temporary buffer to 'buf' according
 168                  * to address on current objdump line */
 169                 off = addr - start_addr;
 170                 if (off >= *len)
 171                         break;
 172                 written_bytes = MIN(read_bytes, *len - off);
 173                 memcpy(buf + off, tmp, written_bytes);
 174                 off_last = off + written_bytes;
 175         }
 176
 177         /* len returns number of bytes that could not be read */
 178         *len -= off_last;
 179
 180         free(line);
 181
 182         return err;
 183 }
 184
 185 static int read_via_objdump(const char *filename, u64 addr, void *buf,
 186                             size_t len)
 187 {
 188         char cmd[PATH_MAX * 2];
 189         const char *fmt;
 190         FILE *f;
 191         int ret;
 192
 193         fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
 194         ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
 195                        filename);
 196         if (ret <= 0 || (size_t)ret >= sizeof(cmd))
 197                 return -1;
 198
 199         pr_debug("Objdump command is: %s\n", cmd);
 200
 201         /* Ignore objdump errors */
 202         strcat(cmd, " 2>/dev/null");
 203
 204         f = popen(cmd, "r");
 205         if (!f) {
 206                 pr_debug("popen failed\n");
 207                 return -1;
 208         }
 209
 210         ret = read_objdump_output(f, buf, &len, addr);
 211         if (len) {
 212                 pr_debug("objdump read too few bytes: %zd\n", len);
 213                 if (!ret)
 214                         ret = len;
 215         }
 216
 217         pclose(f);
 218
 219         return ret;
 220 }
 221
 222 static void dump_buf(unsigned char *buf, size_t len)
 223 {
 224         size_t i;
 225
 226         for (i = 0; i < len; i++) {
 227                 pr_debug("0x%02x ", buf[i]);
 228                 if (i % 16 == 15)
 229                         pr_debug("\n");
 230         }
 231         pr_debug("\n");
 232 }
 233
 234 static int read_object_code(u64 addr, size_t len, u8 cpumode,
 235                             struct thread *thread, struct state *state)
 236 {
 237         struct addr_location al;
 238         unsigned char buf1[BUFSZ];
 239         unsigned char buf2[BUFSZ];
 240         size_t ret_len;
 241         u64 objdump_addr;
 242         const char *objdump_name;
 243         char decomp_name[KMOD_DECOMP_LEN];
 244         bool decomp = false;
 245         int ret;
 246
 247         pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
 248
 249         if (!thread__find_map(thread, cpumode, addr, &al) || !al.map->dso) {
 250                 if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
 251                         pr_debug("Hypervisor address can not be resolved - skipping\n");
 252                         return 0;
 253                 }
 254
 255                 pr_debug("thread__find_map failed\n");
 256                 return -1;
 257         }
 258
 259         pr_debug("File is: %s\n", al.map->dso->long_name);
 260
 261         if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 262             !dso__is_kcore(al.map->dso)) {
 263                 pr_debug("Unexpected kernel address - skipping\n");
 264                 return 0;
 265         }
 266
 267         pr_debug("On file address is: %#"PRIx64"\n", al.addr);
 268
 269         if (len > BUFSZ)
 270                 len = BUFSZ;
 271
 272         /* Do not go off the map */
 273         if (addr + len > al.map->end)
 274                 len = al.map->end - addr;
 275
 276         /* Read the object code using perf */
 277         ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
 278                                         al.addr, buf1, len);
 279         if (ret_len != len) {
 280                 pr_debug("dso__data_read_offset failed\n");
 281                 return -1;
 282         }
 283
 284         /*
 285          * Converting addresses for use by objdump requires more information.
 286          * map__load() does that.  See map__rip_2objdump() for details.
 287          */
 288         if (map__load(al.map))
 289                 return -1;
 290
 291         /* objdump struggles with kcore - try each map only once */
 292         if (dso__is_kcore(al.map->dso)) {
 293                 size_t d;
 294
 295                 for (d = 0; d < state->done_cnt; d++) {
 296                         if (state->done[d] == al.map->start) {
 297                                 pr_debug("kcore map tested already");
 298                                 pr_debug(" - skipping\n");
 299                                 return 0;
 300                         }
 301                 }
 302                 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
 303                         pr_debug("Too many kcore maps - skipping\n");
 304                         return 0;
 305                 }
 306                 state->done[state->done_cnt++] = al.map->start;
 307         }
 308
 309         objdump_name = al.map->dso->long_name;
 310         if (dso__needs_decompress(al.map->dso)) {
 311                 if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
 312                                                  decomp_name,
 313                                                  sizeof(decomp_name)) < 0) {
 314                         pr_debug("decompression failed\n");
 315                         return -1;
 316                 }
 317
 318                 decomp = true;
 319                 objdump_name = decomp_name;
 320         }
 321
 322         /* Read the object code using objdump */
 323         objdump_addr = map__rip_2objdump(al.map, al.addr);
 324         ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
 325
 326         if (decomp)
 327                 unlink(objdump_name);
 328
 329         if (ret > 0) {
 330                 /*
 331                  * The kernel maps are inaccurate - assume objdump is right in
 332                  * that case.
 333                  */
 334                 if (cpumode == PERF_RECORD_MISC_KERNEL ||
 335                     cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
 336                         len -= ret;
 337                         if (len) {
 338                                 pr_debug("Reducing len to %zu\n", len);
 339                         } else if (dso__is_kcore(al.map->dso)) {
 340                                 /*
 341                                  * objdump cannot handle very large segments
 342                                  * that may be found in kcore.
 343                                  */
 344                                 pr_debug("objdump failed for kcore");
 345                                 pr_debug(" - skipping\n");
 346                                 return 0;
 347                         } else {
 348                                 return -1;
 349                         }
 350                 }
 351         }
 352         if (ret < 0) {
 353                 pr_debug("read_via_objdump failed\n");
 354                 return -1;
 355         }
 356
 357         /* The results should be identical */
 358         if (memcmp(buf1, buf2, len)) {
 359                 pr_debug("Bytes read differ from those read by objdump\n");
 360                 pr_debug("buf1 (dso):\n");
 361                 dump_buf(buf1, len);
 362                 pr_debug("buf2 (objdump):\n");
 363                 dump_buf(buf2, len);
 364                 return -1;
 365         }
 366         pr_debug("Bytes read match those read by objdump\n");
 367
 368         return 0;
 369 }
 370
 371 static int process_sample_event(struct machine *machine,
 372                                 struct evlist *evlist,
 373                                 union perf_event *event, struct state *state)
 374 {
 375         struct perf_sample sample;
 376         struct thread *thread;
 377         int ret;
 378
 379         if (perf_evlist__parse_sample(evlist, event, &sample)) {
 380                 pr_debug("perf_evlist__parse_sample failed\n");
 381                 return -1;
 382         }
 383
 384         thread = machine__findnew_thread(machine, sample.pid, sample.tid);
 385         if (!thread) {
 386                 pr_debug("machine__findnew_thread failed\n");
 387                 return -1;
 388         }
 389
 390         ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
 391         thread__put(thread);
 392         return ret;
 393 }
 394
 395 static int process_event(struct machine *machine, struct evlist *evlist,
 396                          union perf_event *event, struct state *state)
 397 {
 398         if (event->header.type == PERF_RECORD_SAMPLE)
 399                 return process_sample_event(machine, evlist, event, state);
 400
 401         if (event->header.type == PERF_RECORD_THROTTLE ||
 402             event->header.type == PERF_RECORD_UNTHROTTLE)
 403                 return 0;
 404
 405         if (event->header.type < PERF_RECORD_MAX) {
 406                 int ret;
 407
 408                 ret = machine__process_event(machine, event, NULL);
 409                 if (ret < 0)
 410                         pr_debug("machine__process_event failed, event type %u\n",
 411                                  event->header.type);
 412                 return ret;
 413         }
 414
 415         return 0;
 416 }
 417
 418 static int process_events(struct machine *machine, struct evlist *evlist,
 419                           struct state *state)
 420 {
 421         union perf_event *event;
 422         struct mmap *md;
 423         int i, ret;
 424
 425         for (i = 0; i < evlist->nr_mmaps; i++) {
 426                 md = &evlist->mmap[i];
 427                 if (perf_mmap__read_init(md) < 0)
 428                         continue;
 429
 430                 while ((event = perf_mmap__read_event(md)) != NULL) {
 431                         ret = process_event(machine, evlist, event, state);
 432                         perf_mmap__consume(md);
 433                         if (ret < 0)
 434                                 return ret;
 435                 }
 436                 perf_mmap__read_done(md);
 437         }
 438         return 0;
 439 }
 440
 441 static int comp(const void *a, const void *b)
 442 {
 443         return *(int *)a - *(int *)b;
 444 }
 445
 446 static void do_sort_something(void)
 447 {
 448         int buf[40960], i;
 449
 450         for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
 451                 buf[i] = ARRAY_SIZE(buf) - i - 1;
 452
 453         qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
 454
 455         for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
 456                 if (buf[i] != i) {
 457                         pr_debug("qsort failed\n");
 458                         break;
 459                 }
 460         }
 461 }
 462
 463 static void sort_something(void)
 464 {
 465         int i;
 466
 467         for (i = 0; i < 10; i++)
 468                 do_sort_something();
 469 }
 470
 471 static void syscall_something(void)
 472 {
 473         int pipefd[2];
 474         int i;
 475
 476         for (i = 0; i < 1000; i++) {
 477                 if (pipe(pipefd) < 0) {
 478                         pr_debug("pipe failed\n");
 479                         break;
 480                 }
 481                 close(pipefd[1]);
 482                 close(pipefd[0]);
 483         }
 484 }
 485
 486 static void fs_something(void)
 487 {
 488         const char *test_file_name = "temp-perf-code-reading-test-file--";
 489         FILE *f;
 490         int i;
 491
 492         for (i = 0; i < 1000; i++) {
 493                 f = fopen(test_file_name, "w+");
 494                 if (f) {
 495                         fclose(f);
 496                         unlink(test_file_name);
 497                 }
 498         }
 499 }
 500
 501 #ifdef __s390x__
 502 #include "header.h" // for get_cpuid()
 503 #endif
 504
 505 static const char *do_determine_event(bool excl_kernel)
 506 {
 507         const char *event = excl_kernel ? "cycles:u" : "cycles";
 508
 509 #ifdef __s390x__
 510         char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
 511         unsigned int family;
 512         int ret, cpum_cf_a;
 513
 514         if (get_cpuid(cpuid, sizeof(cpuid)))
 515                 goto out_clocks;
 516         ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
 517                      model, cpum_cf_v, &cpum_cf_a);
 518         if (ret != 5)            /* Not available */
 519                 goto out_clocks;
 520         if (excl_kernel && (cpum_cf_a & 4))
 521                 return event;
 522         if (!excl_kernel && (cpum_cf_a & 2))
 523                 return event;
 524
 525         /* Fall through: missing authorization */
 526 out_clocks:
 527         event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
 528
 529 #endif
 530         return event;
 531 }
 532
 533 static void do_something(void)
 534 {
 535         fs_something();
 536
 537         sort_something();
 538
 539         syscall_something();
 540 }
 541
 542 enum {
 543         TEST_CODE_READING_OK,
 544         TEST_CODE_READING_NO_VMLINUX,
 545         TEST_CODE_READING_NO_KCORE,
 546         TEST_CODE_READING_NO_ACCESS,
 547         TEST_CODE_READING_NO_KERNEL_OBJ,
 548 };
 549
 550 static int do_test_code_reading(bool try_kcore)
 551 {
 552         struct machine *machine;
 553         struct thread *thread;
 554         struct record_opts opts = {
 555                 .mmap_pages          = UINT_MAX,
 556                 .user_freq           = UINT_MAX,
 557                 .user_interval       = ULLONG_MAX,
 558                 .freq                = 500,
 559                 .target              = {
 560                         .uses_mmap   = true,
 561                 },
 562         };
 563         struct state state = {
 564                 .done_cnt = 0,
 565         };
 566         struct perf_thread_map *threads = NULL;
 567         struct perf_cpu_map *cpus = NULL;
 568         struct evlist *evlist = NULL;
 569         struct evsel *evsel = NULL;
 570         int err = -1, ret;
 571         pid_t pid;
 572         struct map *map;
 573         bool have_vmlinux, have_kcore, excl_kernel = false;
 574
 575         pid = getpid();
 576
 577         machine = machine__new_host();
 578         machine->env = &perf_env;
 579
 580         ret = machine__create_kernel_maps(machine);
 581         if (ret < 0) {
 582                 pr_debug("machine__create_kernel_maps failed\n");
 583                 goto out_err;
 584         }
 585
 586         /* Force the use of kallsyms instead of vmlinux to try kcore */
 587         if (try_kcore)
 588                 symbol_conf.kallsyms_name = "/proc/kallsyms";
 589
 590         /* Load kernel map */
 591         map = machine__kernel_map(machine);
 592         ret = map__load(map);
 593         if (ret < 0) {
 594                 pr_debug("map__load failed\n");
 595                 goto out_err;
 596         }
 597         have_vmlinux = dso__is_vmlinux(map->dso);
 598         have_kcore = dso__is_kcore(map->dso);
 599
 600         /* 2nd time through we just try kcore */
 601         if (try_kcore && !have_kcore)
 602                 return TEST_CODE_READING_NO_KCORE;
 603
 604         /* No point getting kernel events if there is no kernel object */
 605         if (!have_vmlinux && !have_kcore)
 606                 excl_kernel = true;
 607
 608         threads = thread_map__new_by_tid(pid);
 609         if (!threads) {
 610                 pr_debug("thread_map__new_by_tid failed\n");
 611                 goto out_err;
 612         }
 613
 614         ret = perf_event__synthesize_thread_map(NULL, threads,
 615                                                 perf_event__process, machine, false);
 616         if (ret < 0) {
 617                 pr_debug("perf_event__synthesize_thread_map failed\n");
 618                 goto out_err;
 619         }
 620
 621         thread = machine__findnew_thread(machine, pid, pid);
 622         if (!thread) {
 623                 pr_debug("machine__findnew_thread failed\n");
 624                 goto out_put;
 625         }
 626
 627         cpus = perf_cpu_map__new(NULL);
 628         if (!cpus) {
 629                 pr_debug("perf_cpu_map__new failed\n");
 630                 goto out_put;
 631         }
 632
 633         while (1) {
 634                 const char *str;
 635
 636                 evlist = evlist__new();
 637                 if (!evlist) {
 638                         pr_debug("perf_evlist__new failed\n");
 639                         goto out_put;
 640                 }
 641
 642                 perf_evlist__set_maps(&evlist->core, cpus, threads);
 643
 644                 str = do_determine_event(excl_kernel);
 645                 pr_debug("Parsing event '%s'\n", str);
 646                 ret = parse_events(evlist, str, NULL);
 647                 if (ret < 0) {
 648                         pr_debug("parse_events failed\n");
 649                         goto out_put;
 650                 }
 651
 652                 perf_evlist__config(evlist, &opts, NULL);
 653
 654                 evsel = perf_evlist__first(evlist);
 655
 656                 evsel->core.attr.comm = 1;
 657                 evsel->core.attr.disabled = 1;
 658                 evsel->core.attr.enable_on_exec = 0;
 659
 660                 ret = evlist__open(evlist);
 661                 if (ret < 0) {
 662                         if (!excl_kernel) {
 663                                 excl_kernel = true;
 664                                 /*
 665                                  * Both cpus and threads are now owned by evlist
 666                                  * and will be freed by following perf_evlist__set_maps
 667                                  * call. Getting refference to keep them alive.
 668                                  */
 669                                 perf_cpu_map__get(cpus);
 670                                 perf_thread_map__get(threads);
 671                                 perf_evlist__set_maps(&evlist->core, NULL, NULL);
 672                                 evlist__delete(evlist);
 673                                 evlist = NULL;
 674                                 continue;
 675                         }
 676
 677                         if (verbose > 0) {
 678                                 char errbuf[512];
 679                                 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
 680                                 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
 681                         }
 682
 683                         goto out_put;
 684                 }
 685                 break;
 686         }
 687
 688         ret = evlist__mmap(evlist, UINT_MAX);
 689         if (ret < 0) {
 690                 pr_debug("evlist__mmap failed\n");
 691                 goto out_put;
 692         }
 693
 694         evlist__enable(evlist);
 695
 696         do_something();
 697
 698         evlist__disable(evlist);
 699
 700         ret = process_events(machine, evlist, &state);
 701         if (ret < 0)
 702                 goto out_put;
 703
 704         if (!have_vmlinux && !have_kcore && !try_kcore)
 705                 err = TEST_CODE_READING_NO_KERNEL_OBJ;
 706         else if (!have_vmlinux && !try_kcore)
 707                 err = TEST_CODE_READING_NO_VMLINUX;
 708         else if (excl_kernel)
 709                 err = TEST_CODE_READING_NO_ACCESS;
 710         else
 711                 err = TEST_CODE_READING_OK;
 712 out_put:
 713         thread__put(thread);
 714 out_err:
 715
 716         if (evlist) {
 717                 evlist__delete(evlist);
 718         } else {
 719                 perf_cpu_map__put(cpus);
 720                 perf_thread_map__put(threads);
 721         }
 722         machine__delete_threads(machine);
 723         machine__delete(machine);
 724
 725         return err;
 726 }
 727
 728 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
 729 {
 730         int ret;
 731
 732         ret = do_test_code_reading(false);
 733         if (!ret)
 734                 ret = do_test_code_reading(true);
 735
 736         switch (ret) {
 737         case TEST_CODE_READING_OK:
 738                 return 0;
 739         case TEST_CODE_READING_NO_VMLINUX:
 740                 pr_debug("no vmlinux\n");
 741                 return 0;
 742         case TEST_CODE_READING_NO_KCORE:
 743                 pr_debug("no kcore\n");
 744                 return 0;
 745         case TEST_CODE_READING_NO_ACCESS:
 746                 pr_debug("no access\n");
 747                 return 0;
 748         case TEST_CODE_READING_NO_KERNEL_OBJ:
 749                 pr_debug("no kernel obj\n");
 750                 return 0;
 751         default:
 752                 return -1;
 753         };
 754 }