tools/perf/builtin-ftrace.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-ftrace.c
   4  *
   5  * Copyright (c) 2013  LG Electronics,  Namhyung Kim <namhyung@kernel.org>
   6  * Copyright (c) 2020  Changbin Du <changbin.du@gmail.com>, significant enhancement.
   7  */
   8
   9 #include "builtin.h"
  10
  11 #include <errno.h>
  12 #include <unistd.h>
  13 #include <signal.h>
  14 #include <stdlib.h>
  15 #include <fcntl.h>
  16 #include <math.h>
  17 #include <poll.h>
  18 #include <ctype.h>
  19 #include <linux/capability.h>
  20 #include <linux/string.h>
  21
  22 #include "debug.h"
  23 #include <subcmd/pager.h>
  24 #include <subcmd/parse-options.h>
  25 #include <api/fs/tracing_path.h>
  26 #include "evlist.h"
  27 #include "target.h"
  28 #include "cpumap.h"
  29 #include "thread_map.h"
  30 #include "strfilter.h"
  31 #include "util/cap.h"
  32 #include "util/config.h"
  33 #include "util/ftrace.h"
  34 #include "util/units.h"
  35 #include "util/parse-sublevel-options.h"
  36
  37 #define DEFAULT_TRACER  "function_graph"
  38
  39 static volatile sig_atomic_t workload_exec_errno;
  40 static volatile sig_atomic_t done;
  41
  42 static void sig_handler(int sig __maybe_unused)
  43 {
  44         done = true;
  45 }
  46
  47 /*
  48  * evlist__prepare_workload will send a SIGUSR1 if the fork fails, since
  49  * we asked by setting its exec_error to the function below,
  50  * ftrace__workload_exec_failed_signal.
  51  *
  52  * XXX We need to handle this more appropriately, emitting an error, etc.
  53  */
  54 static void ftrace__workload_exec_failed_signal(int signo __maybe_unused,
  55                                                 siginfo_t *info __maybe_unused,
  56                                                 void *ucontext __maybe_unused)
  57 {
  58         workload_exec_errno = info->si_value.sival_int;
  59         done = true;
  60 }
  61
  62 static int __write_tracing_file(const char *name, const char *val, bool append)
  63 {
  64         char *file;
  65         int fd, ret = -1;
  66         ssize_t size = strlen(val);
  67         int flags = O_WRONLY;
  68         char errbuf[512];
  69         char *val_copy;
  70
  71         file = get_tracing_file(name);
  72         if (!file) {
  73                 pr_debug("cannot get tracing file: %s\n", name);
  74                 return -1;
  75         }
  76
  77         if (append)
  78                 flags |= O_APPEND;
  79         else
  80                 flags |= O_TRUNC;
  81
  82         fd = open(file, flags);
  83         if (fd < 0) {
  84                 pr_debug("cannot open tracing file: %s: %s\n",
  85                          name, str_error_r(errno, errbuf, sizeof(errbuf)));
  86                 goto out;
  87         }
  88
  89         /*
  90          * Copy the original value and append a '\n'. Without this,
  91          * the kernel can hide possible errors.
  92          */
  93         val_copy = strdup(val);
  94         if (!val_copy)
  95                 goto out_close;
  96         val_copy[size] = '\n';
  97
  98         if (write(fd, val_copy, size + 1) == size + 1)
  99                 ret = 0;
 100         else
 101                 pr_debug("write '%s' to tracing/%s failed: %s\n",
 102                          val, name, str_error_r(errno, errbuf, sizeof(errbuf)));
 103
 104         free(val_copy);
 105 out_close:
 106         close(fd);
 107 out:
 108         put_tracing_file(file);
 109         return ret;
 110 }
 111
 112 static int write_tracing_file(const char *name, const char *val)
 113 {
 114         return __write_tracing_file(name, val, false);
 115 }
 116
 117 static int append_tracing_file(const char *name, const char *val)
 118 {
 119         return __write_tracing_file(name, val, true);
 120 }
 121
 122 static int read_tracing_file_to_stdout(const char *name)
 123 {
 124         char buf[4096];
 125         char *file;
 126         int fd;
 127         int ret = -1;
 128
 129         file = get_tracing_file(name);
 130         if (!file) {
 131                 pr_debug("cannot get tracing file: %s\n", name);
 132                 return -1;
 133         }
 134
 135         fd = open(file, O_RDONLY);
 136         if (fd < 0) {
 137                 pr_debug("cannot open tracing file: %s: %s\n",
 138                          name, str_error_r(errno, buf, sizeof(buf)));
 139                 goto out;
 140         }
 141
 142         /* read contents to stdout */
 143         while (true) {
 144                 int n = read(fd, buf, sizeof(buf));
 145                 if (n == 0)
 146                         break;
 147                 else if (n < 0)
 148                         goto out_close;
 149
 150                 if (fwrite(buf, n, 1, stdout) != 1)
 151                         goto out_close;
 152         }
 153         ret = 0;
 154
 155 out_close:
 156         close(fd);
 157 out:
 158         put_tracing_file(file);
 159         return ret;
 160 }
 161
 162 static int read_tracing_file_by_line(const char *name,
 163                                      void (*cb)(char *str, void *arg),
 164                                      void *cb_arg)
 165 {
 166         char *line = NULL;
 167         size_t len = 0;
 168         char *file;
 169         FILE *fp;
 170
 171         file = get_tracing_file(name);
 172         if (!file) {
 173                 pr_debug("cannot get tracing file: %s\n", name);
 174                 return -1;
 175         }
 176
 177         fp = fopen(file, "r");
 178         if (fp == NULL) {
 179                 pr_debug("cannot open tracing file: %s\n", name);
 180                 put_tracing_file(file);
 181                 return -1;
 182         }
 183
 184         while (getline(&line, &len, fp) != -1) {
 185                 cb(line, cb_arg);
 186         }
 187
 188         if (line)
 189                 free(line);
 190
 191         fclose(fp);
 192         put_tracing_file(file);
 193         return 0;
 194 }
 195
 196 static int write_tracing_file_int(const char *name, int value)
 197 {
 198         char buf[16];
 199
 200         snprintf(buf, sizeof(buf), "%d", value);
 201         if (write_tracing_file(name, buf) < 0)
 202                 return -1;
 203
 204         return 0;
 205 }
 206
 207 static int write_tracing_option_file(const char *name, const char *val)
 208 {
 209         char *file;
 210         int ret;
 211
 212         if (asprintf(&file, "options/%s", name) < 0)
 213                 return -1;
 214
 215         ret = __write_tracing_file(file, val, false);
 216         free(file);
 217         return ret;
 218 }
 219
 220 static int reset_tracing_cpu(void);
 221 static void reset_tracing_filters(void);
 222
 223 static void reset_tracing_options(struct perf_ftrace *ftrace __maybe_unused)
 224 {
 225         write_tracing_option_file("function-fork", "0");
 226         write_tracing_option_file("func_stack_trace", "0");
 227         write_tracing_option_file("sleep-time", "1");
 228         write_tracing_option_file("funcgraph-irqs", "1");
 229         write_tracing_option_file("funcgraph-proc", "0");
 230         write_tracing_option_file("funcgraph-abstime", "0");
 231         write_tracing_option_file("latency-format", "0");
 232         write_tracing_option_file("irq-info", "0");
 233 }
 234
 235 static int reset_tracing_files(struct perf_ftrace *ftrace __maybe_unused)
 236 {
 237         if (write_tracing_file("tracing_on", "0") < 0)
 238                 return -1;
 239
 240         if (write_tracing_file("current_tracer", "nop") < 0)
 241                 return -1;
 242
 243         if (write_tracing_file("set_ftrace_pid", " ") < 0)
 244                 return -1;
 245
 246         if (reset_tracing_cpu() < 0)
 247                 return -1;
 248
 249         if (write_tracing_file("max_graph_depth", "0") < 0)
 250                 return -1;
 251
 252         if (write_tracing_file("tracing_thresh", "0") < 0)
 253                 return -1;
 254
 255         reset_tracing_filters();
 256         reset_tracing_options(ftrace);
 257         return 0;
 258 }
 259
 260 static int set_tracing_pid(struct perf_ftrace *ftrace)
 261 {
 262         int i;
 263         char buf[16];
 264
 265         if (target__has_cpu(&ftrace->target))
 266                 return 0;
 267
 268         for (i = 0; i < perf_thread_map__nr(ftrace->evlist->core.threads); i++) {
 269                 scnprintf(buf, sizeof(buf), "%d",
 270                           perf_thread_map__pid(ftrace->evlist->core.threads, i));
 271                 if (append_tracing_file("set_ftrace_pid", buf) < 0)
 272                         return -1;
 273         }
 274         return 0;
 275 }
 276
 277 static int set_tracing_cpumask(struct perf_cpu_map *cpumap)
 278 {
 279         char *cpumask;
 280         size_t mask_size;
 281         int ret;
 282         int last_cpu;
 283
 284         last_cpu = perf_cpu_map__cpu(cpumap, perf_cpu_map__nr(cpumap) - 1).cpu;
 285         mask_size = last_cpu / 4 + 2; /* one more byte for EOS */
 286         mask_size += last_cpu / 32; /* ',' is needed for every 32th cpus */
 287
 288         cpumask = malloc(mask_size);
 289         if (cpumask == NULL) {
 290                 pr_debug("failed to allocate cpu mask\n");
 291                 return -1;
 292         }
 293
 294         cpu_map__snprint_mask(cpumap, cpumask, mask_size);
 295
 296         ret = write_tracing_file("tracing_cpumask", cpumask);
 297
 298         free(cpumask);
 299         return ret;
 300 }
 301
 302 static int set_tracing_cpu(struct perf_ftrace *ftrace)
 303 {
 304         struct perf_cpu_map *cpumap = ftrace->evlist->core.user_requested_cpus;
 305
 306         if (!target__has_cpu(&ftrace->target))
 307                 return 0;
 308
 309         return set_tracing_cpumask(cpumap);
 310 }
 311
 312 static int set_tracing_func_stack_trace(struct perf_ftrace *ftrace)
 313 {
 314         if (!ftrace->func_stack_trace)
 315                 return 0;
 316
 317         if (write_tracing_option_file("func_stack_trace", "1") < 0)
 318                 return -1;
 319
 320         return 0;
 321 }
 322
 323 static int set_tracing_func_irqinfo(struct perf_ftrace *ftrace)
 324 {
 325         if (!ftrace->func_irq_info)
 326                 return 0;
 327
 328         if (write_tracing_option_file("irq-info", "1") < 0)
 329                 return -1;
 330
 331         return 0;
 332 }
 333
 334 static int reset_tracing_cpu(void)
 335 {
 336         struct perf_cpu_map *cpumap = perf_cpu_map__new(NULL);
 337         int ret;
 338
 339         ret = set_tracing_cpumask(cpumap);
 340         perf_cpu_map__put(cpumap);
 341         return ret;
 342 }
 343
 344 static int __set_tracing_filter(const char *filter_file, struct list_head *funcs)
 345 {
 346         struct filter_entry *pos;
 347
 348         list_for_each_entry(pos, funcs, list) {
 349                 if (append_tracing_file(filter_file, pos->name) < 0)
 350                         return -1;
 351         }
 352
 353         return 0;
 354 }
 355
 356 static int set_tracing_filters(struct perf_ftrace *ftrace)
 357 {
 358         int ret;
 359
 360         ret = __set_tracing_filter("set_ftrace_filter", &ftrace->filters);
 361         if (ret < 0)
 362                 return ret;
 363
 364         ret = __set_tracing_filter("set_ftrace_notrace", &ftrace->notrace);
 365         if (ret < 0)
 366                 return ret;
 367
 368         ret = __set_tracing_filter("set_graph_function", &ftrace->graph_funcs);
 369         if (ret < 0)
 370                 return ret;
 371
 372         /* old kernels do not have this filter */
 373         __set_tracing_filter("set_graph_notrace", &ftrace->nograph_funcs);
 374
 375         return ret;
 376 }
 377
 378 static void reset_tracing_filters(void)
 379 {
 380         write_tracing_file("set_ftrace_filter", " ");
 381         write_tracing_file("set_ftrace_notrace", " ");
 382         write_tracing_file("set_graph_function", " ");
 383         write_tracing_file("set_graph_notrace", " ");
 384 }
 385
 386 static int set_tracing_depth(struct perf_ftrace *ftrace)
 387 {
 388         if (ftrace->graph_depth == 0)
 389                 return 0;
 390
 391         if (ftrace->graph_depth < 0) {
 392                 pr_err("invalid graph depth: %d\n", ftrace->graph_depth);
 393                 return -1;
 394         }
 395
 396         if (write_tracing_file_int("max_graph_depth", ftrace->graph_depth) < 0)
 397                 return -1;
 398
 399         return 0;
 400 }
 401
 402 static int set_tracing_percpu_buffer_size(struct perf_ftrace *ftrace)
 403 {
 404         int ret;
 405
 406         if (ftrace->percpu_buffer_size == 0)
 407                 return 0;
 408
 409         ret = write_tracing_file_int("buffer_size_kb",
 410                                      ftrace->percpu_buffer_size / 1024);
 411         if (ret < 0)
 412                 return ret;
 413
 414         return 0;
 415 }
 416
 417 static int set_tracing_trace_inherit(struct perf_ftrace *ftrace)
 418 {
 419         if (!ftrace->inherit)
 420                 return 0;
 421
 422         if (write_tracing_option_file("function-fork", "1") < 0)
 423                 return -1;
 424
 425         return 0;
 426 }
 427
 428 static int set_tracing_sleep_time(struct perf_ftrace *ftrace)
 429 {
 430         if (!ftrace->graph_nosleep_time)
 431                 return 0;
 432
 433         if (write_tracing_option_file("sleep-time", "0") < 0)
 434                 return -1;
 435
 436         return 0;
 437 }
 438
 439 static int set_tracing_funcgraph_irqs(struct perf_ftrace *ftrace)
 440 {
 441         if (!ftrace->graph_noirqs)
 442                 return 0;
 443
 444         if (write_tracing_option_file("funcgraph-irqs", "0") < 0)
 445                 return -1;
 446
 447         return 0;
 448 }
 449
 450 static int set_tracing_funcgraph_verbose(struct perf_ftrace *ftrace)
 451 {
 452         if (!ftrace->graph_verbose)
 453                 return 0;
 454
 455         if (write_tracing_option_file("funcgraph-proc", "1") < 0)
 456                 return -1;
 457
 458         if (write_tracing_option_file("funcgraph-abstime", "1") < 0)
 459                 return -1;
 460
 461         if (write_tracing_option_file("latency-format", "1") < 0)
 462                 return -1;
 463
 464         return 0;
 465 }
 466
 467 static int set_tracing_thresh(struct perf_ftrace *ftrace)
 468 {
 469         int ret;
 470
 471         if (ftrace->graph_thresh == 0)
 472                 return 0;
 473
 474         ret = write_tracing_file_int("tracing_thresh", ftrace->graph_thresh);
 475         if (ret < 0)
 476                 return ret;
 477
 478         return 0;
 479 }
 480
 481 static int set_tracing_options(struct perf_ftrace *ftrace)
 482 {
 483         if (set_tracing_pid(ftrace) < 0) {
 484                 pr_err("failed to set ftrace pid\n");
 485                 return -1;
 486         }
 487
 488         if (set_tracing_cpu(ftrace) < 0) {
 489                 pr_err("failed to set tracing cpumask\n");
 490                 return -1;
 491         }
 492
 493         if (set_tracing_func_stack_trace(ftrace) < 0) {
 494                 pr_err("failed to set tracing option func_stack_trace\n");
 495                 return -1;
 496         }
 497
 498         if (set_tracing_func_irqinfo(ftrace) < 0) {
 499                 pr_err("failed to set tracing option irq-info\n");
 500                 return -1;
 501         }
 502
 503         if (set_tracing_filters(ftrace) < 0) {
 504                 pr_err("failed to set tracing filters\n");
 505                 return -1;
 506         }
 507
 508         if (set_tracing_depth(ftrace) < 0) {
 509                 pr_err("failed to set graph depth\n");
 510                 return -1;
 511         }
 512
 513         if (set_tracing_percpu_buffer_size(ftrace) < 0) {
 514                 pr_err("failed to set tracing per-cpu buffer size\n");
 515                 return -1;
 516         }
 517
 518         if (set_tracing_trace_inherit(ftrace) < 0) {
 519                 pr_err("failed to set tracing option function-fork\n");
 520                 return -1;
 521         }
 522
 523         if (set_tracing_sleep_time(ftrace) < 0) {
 524                 pr_err("failed to set tracing option sleep-time\n");
 525                 return -1;
 526         }
 527
 528         if (set_tracing_funcgraph_irqs(ftrace) < 0) {
 529                 pr_err("failed to set tracing option funcgraph-irqs\n");
 530                 return -1;
 531         }
 532
 533         if (set_tracing_funcgraph_verbose(ftrace) < 0) {
 534                 pr_err("failed to set tracing option funcgraph-proc/funcgraph-abstime\n");
 535                 return -1;
 536         }
 537
 538         if (set_tracing_thresh(ftrace) < 0) {
 539                 pr_err("failed to set tracing thresh\n");
 540                 return -1;
 541         }
 542
 543         return 0;
 544 }
 545
 546 static void select_tracer(struct perf_ftrace *ftrace)
 547 {
 548         bool graph = !list_empty(&ftrace->graph_funcs) ||
 549                      !list_empty(&ftrace->nograph_funcs);
 550         bool func = !list_empty(&ftrace->filters) ||
 551                     !list_empty(&ftrace->notrace);
 552
 553         /* The function_graph has priority over function tracer. */
 554         if (graph)
 555                 ftrace->tracer = "function_graph";
 556         else if (func)
 557                 ftrace->tracer = "function";
 558         /* Otherwise, the default tracer is used. */
 559
 560         pr_debug("%s tracer is used\n", ftrace->tracer);
 561 }
 562
 563 static int __cmd_ftrace(struct perf_ftrace *ftrace)
 564 {
 565         char *trace_file;
 566         int trace_fd;
 567         char buf[4096];
 568         struct pollfd pollfd = {
 569                 .events = POLLIN,
 570         };
 571
 572         if (!(perf_cap__capable(CAP_PERFMON) ||
 573               perf_cap__capable(CAP_SYS_ADMIN))) {
 574                 pr_err("ftrace only works for %s!\n",
 575 #ifdef HAVE_LIBCAP_SUPPORT
 576                 "users with the CAP_PERFMON or CAP_SYS_ADMIN capability"
 577 #else
 578                 "root"
 579 #endif
 580                 );
 581                 return -1;
 582         }
 583
 584         select_tracer(ftrace);
 585
 586         if (reset_tracing_files(ftrace) < 0) {
 587                 pr_err("failed to reset ftrace\n");
 588                 goto out;
 589         }
 590
 591         /* reset ftrace buffer */
 592         if (write_tracing_file("trace", "0") < 0)
 593                 goto out;
 594
 595         if (set_tracing_options(ftrace) < 0)
 596                 goto out_reset;
 597
 598         if (write_tracing_file("current_tracer", ftrace->tracer) < 0) {
 599                 pr_err("failed to set current_tracer to %s\n", ftrace->tracer);
 600                 goto out_reset;
 601         }
 602
 603         setup_pager();
 604
 605         trace_file = get_tracing_file("trace_pipe");
 606         if (!trace_file) {
 607                 pr_err("failed to open trace_pipe\n");
 608                 goto out_reset;
 609         }
 610
 611         trace_fd = open(trace_file, O_RDONLY);
 612
 613         put_tracing_file(trace_file);
 614
 615         if (trace_fd < 0) {
 616                 pr_err("failed to open trace_pipe\n");
 617                 goto out_reset;
 618         }
 619
 620         fcntl(trace_fd, F_SETFL, O_NONBLOCK);
 621         pollfd.fd = trace_fd;
 622
 623         /* display column headers */
 624         read_tracing_file_to_stdout("trace");
 625
 626         if (!ftrace->target.initial_delay) {
 627                 if (write_tracing_file("tracing_on", "1") < 0) {
 628                         pr_err("can't enable tracing\n");
 629                         goto out_close_fd;
 630                 }
 631         }
 632
 633         evlist__start_workload(ftrace->evlist);
 634
 635         if (ftrace->target.initial_delay > 0) {
 636                 usleep(ftrace->target.initial_delay * 1000);
 637                 if (write_tracing_file("tracing_on", "1") < 0) {
 638                         pr_err("can't enable tracing\n");
 639                         goto out_close_fd;
 640                 }
 641         }
 642
 643         while (!done) {
 644                 if (poll(&pollfd, 1, -1) < 0)
 645                         break;
 646
 647                 if (pollfd.revents & POLLIN) {
 648                         int n = read(trace_fd, buf, sizeof(buf));
 649                         if (n < 0)
 650                                 break;
 651                         if (fwrite(buf, n, 1, stdout) != 1)
 652                                 break;
 653                         /* flush output since stdout is in full buffering mode due to pager */
 654                         fflush(stdout);
 655                 }
 656         }
 657
 658         write_tracing_file("tracing_on", "0");
 659
 660         if (workload_exec_errno) {
 661                 const char *emsg = str_error_r(workload_exec_errno, buf, sizeof(buf));
 662                 /* flush stdout first so below error msg appears at the end. */
 663                 fflush(stdout);
 664                 pr_err("workload failed: %s\n", emsg);
 665                 goto out_close_fd;
 666         }
 667
 668         /* read remaining buffer contents */
 669         while (true) {
 670                 int n = read(trace_fd, buf, sizeof(buf));
 671                 if (n <= 0)
 672                         break;
 673                 if (fwrite(buf, n, 1, stdout) != 1)
 674                         break;
 675         }
 676
 677 out_close_fd:
 678         close(trace_fd);
 679 out_reset:
 680         reset_tracing_files(ftrace);
 681 out:
 682         return (done && !workload_exec_errno) ? 0 : -1;
 683 }
 684
 685 static void make_histogram(int buckets[], char *buf, size_t len, char *linebuf,
 686                            bool use_nsec)
 687 {
 688         char *p, *q;
 689         char *unit;
 690         double num;
 691         int i;
 692
 693         /* ensure NUL termination */
 694         buf[len] = '\0';
 695
 696         /* handle data line by line */
 697         for (p = buf; (q = strchr(p, '\n')) != NULL; p = q + 1) {
 698                 *q = '\0';
 699                 /* move it to the line buffer */
 700                 strcat(linebuf, p);
 701
 702                 /*
 703                  * parse trace output to get function duration like in
 704                  *
 705                  * # tracer: function_graph
 706                  * #
 707                  * # CPU  DURATION                  FUNCTION CALLS
 708                  * # |     |   |                     |   |   |   |
 709                  *  1) + 10.291 us   |  do_filp_open();
 710                  *  1)   4.889 us    |  do_filp_open();
 711                  *  1)   6.086 us    |  do_filp_open();
 712                  *
 713                  */
 714                 if (linebuf[0] == '#')
 715                         goto next;
 716
 717                 /* ignore CPU */
 718                 p = strchr(linebuf, ')');
 719                 if (p == NULL)
 720                         p = linebuf;
 721
 722                 while (*p && !isdigit(*p) && (*p != '|'))
 723                         p++;
 724
 725                 /* no duration */
 726                 if (*p == '\0' || *p == '|')
 727                         goto next;
 728
 729                 num = strtod(p, &unit);
 730                 if (!unit || strncmp(unit, " us", 3))
 731                         goto next;
 732
 733                 if (use_nsec)
 734                         num *= 1000;
 735
 736                 i = log2(num);
 737                 if (i < 0)
 738                         i = 0;
 739                 if (i >= NUM_BUCKET)
 740                         i = NUM_BUCKET - 1;
 741
 742                 buckets[i]++;
 743
 744 next:
 745                 /* empty the line buffer for the next output  */
 746                 linebuf[0] = '\0';
 747         }
 748
 749         /* preserve any remaining output (before newline) */
 750         strcat(linebuf, p);
 751 }
 752
 753 static void display_histogram(int buckets[], bool use_nsec)
 754 {
 755         int i;
 756         int total = 0;
 757         int bar_total = 46;  /* to fit in 80 column */
 758         char bar[] = "###############################################";
 759         int bar_len;
 760
 761         for (i = 0; i < NUM_BUCKET; i++)
 762                 total += buckets[i];
 763
 764         if (total == 0) {
 765                 printf("No data found\n");
 766                 return;
 767         }
 768
 769         printf("# %14s | %10s | %-*s |\n",
 770                "  DURATION    ", "COUNT", bar_total, "GRAPH");
 771
 772         bar_len = buckets[0] * bar_total / total;
 773         printf("  %4d - %-4d %s | %10d | %.*s%*s |\n",
 774                0, 1, "us", buckets[0], bar_len, bar, bar_total - bar_len, "");
 775
 776         for (i = 1; i < NUM_BUCKET - 1; i++) {
 777                 int start = (1 << (i - 1));
 778                 int stop = 1 << i;
 779                 const char *unit = use_nsec ? "ns" : "us";
 780
 781                 if (start >= 1024) {
 782                         start >>= 10;
 783                         stop >>= 10;
 784                         unit = use_nsec ? "us" : "ms";
 785                 }
 786                 bar_len = buckets[i] * bar_total / total;
 787                 printf("  %4d - %-4d %s | %10d | %.*s%*s |\n",
 788                        start, stop, unit, buckets[i], bar_len, bar,
 789                        bar_total - bar_len, "");
 790         }
 791
 792         bar_len = buckets[NUM_BUCKET - 1] * bar_total / total;
 793         printf("  %4d - %-4s %s | %10d | %.*s%*s |\n",
 794                1, "...", use_nsec ? "ms" : " s", buckets[NUM_BUCKET - 1],
 795                bar_len, bar, bar_total - bar_len, "");
 796
 797 }
 798
 799 static int prepare_func_latency(struct perf_ftrace *ftrace)
 800 {
 801         char *trace_file;
 802         int fd;
 803
 804         if (ftrace->target.use_bpf)
 805                 return perf_ftrace__latency_prepare_bpf(ftrace);
 806
 807         if (reset_tracing_files(ftrace) < 0) {
 808                 pr_err("failed to reset ftrace\n");
 809                 return -1;
 810         }
 811
 812         /* reset ftrace buffer */
 813         if (write_tracing_file("trace", "0") < 0)
 814                 return -1;
 815
 816         if (set_tracing_options(ftrace) < 0)
 817                 return -1;
 818
 819         /* force to use the function_graph tracer to track duration */
 820         if (write_tracing_file("current_tracer", "function_graph") < 0) {
 821                 pr_err("failed to set current_tracer to function_graph\n");
 822                 return -1;
 823         }
 824
 825         trace_file = get_tracing_file("trace_pipe");
 826         if (!trace_file) {
 827                 pr_err("failed to open trace_pipe\n");
 828                 return -1;
 829         }
 830
 831         fd = open(trace_file, O_RDONLY);
 832         if (fd < 0)
 833                 pr_err("failed to open trace_pipe\n");
 834
 835         put_tracing_file(trace_file);
 836         return fd;
 837 }
 838
 839 static int start_func_latency(struct perf_ftrace *ftrace)
 840 {
 841         if (ftrace->target.use_bpf)
 842                 return perf_ftrace__latency_start_bpf(ftrace);
 843
 844         if (write_tracing_file("tracing_on", "1") < 0) {
 845                 pr_err("can't enable tracing\n");
 846                 return -1;
 847         }
 848
 849         return 0;
 850 }
 851
 852 static int stop_func_latency(struct perf_ftrace *ftrace)
 853 {
 854         if (ftrace->target.use_bpf)
 855                 return perf_ftrace__latency_stop_bpf(ftrace);
 856
 857         write_tracing_file("tracing_on", "0");
 858         return 0;
 859 }
 860
 861 static int read_func_latency(struct perf_ftrace *ftrace, int buckets[])
 862 {
 863         if (ftrace->target.use_bpf)
 864                 return perf_ftrace__latency_read_bpf(ftrace, buckets);
 865
 866         return 0;
 867 }
 868
 869 static int cleanup_func_latency(struct perf_ftrace *ftrace)
 870 {
 871         if (ftrace->target.use_bpf)
 872                 return perf_ftrace__latency_cleanup_bpf(ftrace);
 873
 874         reset_tracing_files(ftrace);
 875         return 0;
 876 }
 877
 878 static int __cmd_latency(struct perf_ftrace *ftrace)
 879 {
 880         int trace_fd;
 881         char buf[4096];
 882         char line[256];
 883         struct pollfd pollfd = {
 884                 .events = POLLIN,
 885         };
 886         int buckets[NUM_BUCKET] = { };
 887
 888         if (!(perf_cap__capable(CAP_PERFMON) ||
 889               perf_cap__capable(CAP_SYS_ADMIN))) {
 890                 pr_err("ftrace only works for %s!\n",
 891 #ifdef HAVE_LIBCAP_SUPPORT
 892                 "users with the CAP_PERFMON or CAP_SYS_ADMIN capability"
 893 #else
 894                 "root"
 895 #endif
 896                 );
 897                 return -1;
 898         }
 899
 900         trace_fd = prepare_func_latency(ftrace);
 901         if (trace_fd < 0)
 902                 goto out;
 903
 904         fcntl(trace_fd, F_SETFL, O_NONBLOCK);
 905         pollfd.fd = trace_fd;
 906
 907         if (start_func_latency(ftrace) < 0)
 908                 goto out;
 909
 910         evlist__start_workload(ftrace->evlist);
 911
 912         line[0] = '\0';
 913         while (!done) {
 914                 if (poll(&pollfd, 1, -1) < 0)
 915                         break;
 916
 917                 if (pollfd.revents & POLLIN) {
 918                         int n = read(trace_fd, buf, sizeof(buf) - 1);
 919                         if (n < 0)
 920                                 break;
 921
 922                         make_histogram(buckets, buf, n, line, ftrace->use_nsec);
 923                 }
 924         }
 925
 926         stop_func_latency(ftrace);
 927
 928         if (workload_exec_errno) {
 929                 const char *emsg = str_error_r(workload_exec_errno, buf, sizeof(buf));
 930                 pr_err("workload failed: %s\n", emsg);
 931                 goto out;
 932         }
 933
 934         /* read remaining buffer contents */
 935         while (!ftrace->target.use_bpf) {
 936                 int n = read(trace_fd, buf, sizeof(buf) - 1);
 937                 if (n <= 0)
 938                         break;
 939                 make_histogram(buckets, buf, n, line, ftrace->use_nsec);
 940         }
 941
 942         read_func_latency(ftrace, buckets);
 943
 944         display_histogram(buckets, ftrace->use_nsec);
 945
 946 out:
 947         close(trace_fd);
 948         cleanup_func_latency(ftrace);
 949
 950         return (done && !workload_exec_errno) ? 0 : -1;
 951 }
 952
 953 static int perf_ftrace_config(const char *var, const char *value, void *cb)
 954 {
 955         struct perf_ftrace *ftrace = cb;
 956
 957         if (!strstarts(var, "ftrace."))
 958                 return 0;
 959
 960         if (strcmp(var, "ftrace.tracer"))
 961                 return -1;
 962
 963         if (!strcmp(value, "function_graph") ||
 964             !strcmp(value, "function")) {
 965                 ftrace->tracer = value;
 966                 return 0;
 967         }
 968
 969         pr_err("Please select \"function_graph\" (default) or \"function\"\n");
 970         return -1;
 971 }
 972
 973 static void list_function_cb(char *str, void *arg)
 974 {
 975         struct strfilter *filter = (struct strfilter *)arg;
 976
 977         if (strfilter__compare(filter, str))
 978                 printf("%s", str);
 979 }
 980
 981 static int opt_list_avail_functions(const struct option *opt __maybe_unused,
 982                                     const char *str, int unset)
 983 {
 984         struct strfilter *filter;
 985         const char *err = NULL;
 986         int ret;
 987
 988         if (unset || !str)
 989                 return -1;
 990
 991         filter = strfilter__new(str, &err);
 992         if (!filter)
 993                 return err ? -EINVAL : -ENOMEM;
 994
 995         ret = strfilter__or(filter, str, &err);
 996         if (ret == -EINVAL) {
 997                 pr_err("Filter parse error at %td.\n", err - str + 1);
 998                 pr_err("Source: \"%s\"\n", str);
 999                 pr_err("         %*c\n", (int)(err - str + 1), '^');
1000                 strfilter__delete(filter);
1001                 return ret;
1002         }
1003
1004         ret = read_tracing_file_by_line("available_filter_functions",
1005                                         list_function_cb, filter);
1006         strfilter__delete(filter);
1007         if (ret < 0)
1008                 return ret;
1009
1010         exit(0);
1011 }
1012
1013 static int parse_filter_func(const struct option *opt, const char *str,
1014                              int unset __maybe_unused)
1015 {
1016         struct list_head *head = opt->value;
1017         struct filter_entry *entry;
1018
1019         entry = malloc(sizeof(*entry) + strlen(str) + 1);
1020         if (entry == NULL)
1021                 return -ENOMEM;
1022
1023         strcpy(entry->name, str);
1024         list_add_tail(&entry->list, head);
1025
1026         return 0;
1027 }
1028
1029 static void delete_filter_func(struct list_head *head)
1030 {
1031         struct filter_entry *pos, *tmp;
1032
1033         list_for_each_entry_safe(pos, tmp, head, list) {
1034                 list_del_init(&pos->list);
1035                 free(pos);
1036         }
1037 }
1038
1039 static int parse_buffer_size(const struct option *opt,
1040                              const char *str, int unset)
1041 {
1042         unsigned long *s = (unsigned long *)opt->value;
1043         static struct parse_tag tags_size[] = {
1044                 { .tag  = 'B', .mult = 1       },
1045                 { .tag  = 'K', .mult = 1 << 10 },
1046                 { .tag  = 'M', .mult = 1 << 20 },
1047                 { .tag  = 'G', .mult = 1 << 30 },
1048                 { .tag  = 0 },
1049         };
1050         unsigned long val;
1051
1052         if (unset) {
1053                 *s = 0;
1054                 return 0;
1055         }
1056
1057         val = parse_tag_value(str, tags_size);
1058         if (val != (unsigned long) -1) {
1059                 if (val < 1024) {
1060                         pr_err("buffer size too small, must larger than 1KB.");
1061                         return -1;
1062                 }
1063                 *s = val;
1064                 return 0;
1065         }
1066
1067         return -1;
1068 }
1069
1070 static int parse_func_tracer_opts(const struct option *opt,
1071                                   const char *str, int unset)
1072 {
1073         int ret;
1074         struct perf_ftrace *ftrace = (struct perf_ftrace *) opt->value;
1075         struct sublevel_option func_tracer_opts[] = {
1076                 { .name = "call-graph", .value_ptr = &ftrace->func_stack_trace },
1077                 { .name = "irq-info",   .value_ptr = &ftrace->func_irq_info },
1078                 { .name = NULL, }
1079         };
1080
1081         if (unset)
1082                 return 0;
1083
1084         ret = perf_parse_sublevel_options(str, func_tracer_opts);
1085         if (ret)
1086                 return ret;
1087
1088         return 0;
1089 }
1090
1091 static int parse_graph_tracer_opts(const struct option *opt,
1092                                   const char *str, int unset)
1093 {
1094         int ret;
1095         struct perf_ftrace *ftrace = (struct perf_ftrace *) opt->value;
1096         struct sublevel_option graph_tracer_opts[] = {
1097                 { .name = "nosleep-time",       .value_ptr = &ftrace->graph_nosleep_time },
1098                 { .name = "noirqs",             .value_ptr = &ftrace->graph_noirqs },
1099                 { .name = "verbose",            .value_ptr = &ftrace->graph_verbose },
1100                 { .name = "thresh",             .value_ptr = &ftrace->graph_thresh },
1101                 { .name = "depth",              .value_ptr = &ftrace->graph_depth },
1102                 { .name = NULL, }
1103         };
1104
1105         if (unset)
1106                 return 0;
1107
1108         ret = perf_parse_sublevel_options(str, graph_tracer_opts);
1109         if (ret)
1110                 return ret;
1111
1112         return 0;
1113 }
1114
1115 enum perf_ftrace_subcommand {
1116         PERF_FTRACE_NONE,
1117         PERF_FTRACE_TRACE,
1118         PERF_FTRACE_LATENCY,
1119 };
1120
1121 int cmd_ftrace(int argc, const char **argv)
1122 {
1123         int ret;
1124         int (*cmd_func)(struct perf_ftrace *) = NULL;
1125         struct perf_ftrace ftrace = {
1126                 .tracer = DEFAULT_TRACER,
1127                 .target = { .uid = UINT_MAX, },
1128         };
1129         const struct option common_options[] = {
1130         OPT_STRING('p', "pid", &ftrace.target.pid, "pid",
1131                    "Trace on existing process id"),
1132         /* TODO: Add short option -t after -t/--tracer can be removed. */
1133         OPT_STRING(0, "tid", &ftrace.target.tid, "tid",
1134                    "Trace on existing thread id (exclusive to --pid)"),
1135         OPT_INCR('v', "verbose", &verbose,
1136                  "Be more verbose"),
1137         OPT_BOOLEAN('a', "all-cpus", &ftrace.target.system_wide,
1138                     "System-wide collection from all CPUs"),
1139         OPT_STRING('C', "cpu", &ftrace.target.cpu_list, "cpu",
1140                     "List of cpus to monitor"),
1141         OPT_END()
1142         };
1143         const struct option ftrace_options[] = {
1144         OPT_STRING('t', "tracer", &ftrace.tracer, "tracer",
1145                    "Tracer to use: function_graph(default) or function"),
1146         OPT_CALLBACK_DEFAULT('F', "funcs", NULL, "[FILTER]",
1147                              "Show available functions to filter",
1148                              opt_list_avail_functions, "*"),
1149         OPT_CALLBACK('T', "trace-funcs", &ftrace.filters, "func",
1150                      "Trace given functions using function tracer",
1151                      parse_filter_func),
1152         OPT_CALLBACK('N', "notrace-funcs", &ftrace.notrace, "func",
1153                      "Do not trace given functions", parse_filter_func),
1154         OPT_CALLBACK(0, "func-opts", &ftrace, "options",
1155                      "Function tracer options, available options: call-graph,irq-info",
1156                      parse_func_tracer_opts),
1157         OPT_CALLBACK('G', "graph-funcs", &ftrace.graph_funcs, "func",
1158                      "Trace given functions using function_graph tracer",
1159                      parse_filter_func),
1160         OPT_CALLBACK('g', "nograph-funcs", &ftrace.nograph_funcs, "func",
1161                      "Set nograph filter on given functions", parse_filter_func),
1162         OPT_CALLBACK(0, "graph-opts", &ftrace, "options",
1163                      "Graph tracer options, available options: nosleep-time,noirqs,verbose,thresh=<n>,depth=<n>",
1164                      parse_graph_tracer_opts),
1165         OPT_CALLBACK('m', "buffer-size", &ftrace.percpu_buffer_size, "size",
1166                      "Size of per cpu buffer, needs to use a B, K, M or G suffix.", parse_buffer_size),
1167         OPT_BOOLEAN(0, "inherit", &ftrace.inherit,
1168                     "Trace children processes"),
1169         OPT_INTEGER('D', "delay", &ftrace.target.initial_delay,
1170                     "Number of milliseconds to wait before starting tracing after program start"),
1171         OPT_PARENT(common_options),
1172         };
1173         const struct option latency_options[] = {
1174         OPT_CALLBACK('T', "trace-funcs", &ftrace.filters, "func",
1175                      "Show latency of given function", parse_filter_func),
1176 #ifdef HAVE_BPF_SKEL
1177         OPT_BOOLEAN('b', "use-bpf", &ftrace.target.use_bpf,
1178                     "Use BPF to measure function latency"),
1179 #endif
1180         OPT_BOOLEAN('n', "use-nsec", &ftrace.use_nsec,
1181                     "Use nano-second histogram"),
1182         OPT_PARENT(common_options),
1183         };
1184         const struct option *options = ftrace_options;
1185
1186         const char * const ftrace_usage[] = {
1187                 "perf ftrace [<options>] [<command>]",
1188                 "perf ftrace [<options>] -- [<command>] [<options>]",
1189                 "perf ftrace {trace|latency} [<options>] [<command>]",
1190                 "perf ftrace {trace|latency} [<options>] -- [<command>] [<options>]",
1191                 NULL
1192         };
1193         enum perf_ftrace_subcommand subcmd = PERF_FTRACE_NONE;
1194
1195         INIT_LIST_HEAD(&ftrace.filters);
1196         INIT_LIST_HEAD(&ftrace.notrace);
1197         INIT_LIST_HEAD(&ftrace.graph_funcs);
1198         INIT_LIST_HEAD(&ftrace.nograph_funcs);
1199
1200         signal(SIGINT, sig_handler);
1201         signal(SIGUSR1, sig_handler);
1202         signal(SIGCHLD, sig_handler);
1203         signal(SIGPIPE, sig_handler);
1204
1205         ret = perf_config(perf_ftrace_config, &ftrace);
1206         if (ret < 0)
1207                 return -1;
1208
1209         if (argc > 1) {
1210                 if (!strcmp(argv[1], "trace")) {
1211                         subcmd = PERF_FTRACE_TRACE;
1212                 } else if (!strcmp(argv[1], "latency")) {
1213                         subcmd = PERF_FTRACE_LATENCY;
1214                         options = latency_options;
1215                 }
1216
1217                 if (subcmd != PERF_FTRACE_NONE) {
1218                         argc--;
1219                         argv++;
1220                 }
1221         }
1222         /* for backward compatibility */
1223         if (subcmd == PERF_FTRACE_NONE)
1224                 subcmd = PERF_FTRACE_TRACE;
1225
1226         argc = parse_options(argc, argv, options, ftrace_usage,
1227                             PARSE_OPT_STOP_AT_NON_OPTION);
1228         if (argc < 0) {
1229                 ret = -EINVAL;
1230                 goto out_delete_filters;
1231         }
1232
1233         /* Make system wide (-a) the default target. */
1234         if (!argc && target__none(&ftrace.target))
1235                 ftrace.target.system_wide = true;
1236
1237         switch (subcmd) {
1238         case PERF_FTRACE_TRACE:
1239                 cmd_func = __cmd_ftrace;
1240                 break;
1241         case PERF_FTRACE_LATENCY:
1242                 if (list_empty(&ftrace.filters)) {
1243                         pr_err("Should provide a function to measure\n");
1244                         parse_options_usage(ftrace_usage, options, "T", 1);
1245                         ret = -EINVAL;
1246                         goto out_delete_filters;
1247                 }
1248                 cmd_func = __cmd_latency;
1249                 break;
1250         case PERF_FTRACE_NONE:
1251         default:
1252                 pr_err("Invalid subcommand\n");
1253                 ret = -EINVAL;
1254                 goto out_delete_filters;
1255         }
1256
1257         ret = target__validate(&ftrace.target);
1258         if (ret) {
1259                 char errbuf[512];
1260
1261                 target__strerror(&ftrace.target, ret, errbuf, 512);
1262                 pr_err("%s\n", errbuf);
1263                 goto out_delete_filters;
1264         }
1265
1266         ftrace.evlist = evlist__new();
1267         if (ftrace.evlist == NULL) {
1268                 ret = -ENOMEM;
1269                 goto out_delete_filters;
1270         }
1271
1272         ret = evlist__create_maps(ftrace.evlist, &ftrace.target);
1273         if (ret < 0)
1274                 goto out_delete_evlist;
1275
1276         if (argc) {
1277                 ret = evlist__prepare_workload(ftrace.evlist, &ftrace.target,
1278                                                argv, false,
1279                                                ftrace__workload_exec_failed_signal);
1280                 if (ret < 0)
1281                         goto out_delete_evlist;
1282         }
1283
1284         ret = cmd_func(&ftrace);
1285
1286 out_delete_evlist:
1287         evlist__delete(ftrace.evlist);
1288
1289 out_delete_filters:
1290         delete_filter_func(&ftrace.filters);
1291         delete_filter_func(&ftrace.notrace);
1292         delete_filter_func(&ftrace.graph_funcs);
1293         delete_filter_func(&ftrace.nograph_funcs);
1294
1295         return ret;
1296 }