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