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  */
   7
   8 #include "builtin.h"
   9
  10 #include <errno.h>
  11 #include <unistd.h>
  12 #include <signal.h>
  13 #include <stdlib.h>
  14 #include <fcntl.h>
  15 #include <poll.h>
  16 #include <linux/capability.h>
  17 #include <linux/string.h>
  18
  19 #include "debug.h"
  20 #include <subcmd/pager.h>
  21 #include <subcmd/parse-options.h>
  22 #include <api/fs/tracing_path.h>
  23 #include "evlist.h"
  24 #include "target.h"
  25 #include "cpumap.h"
  26 #include "thread_map.h"
  27 #include "util/cap.h"
  28 #include "util/config.h"
  29
  30 #define DEFAULT_TRACER  "function_graph"
  31
  32 struct perf_ftrace {
  33         struct evlist           *evlist;
  34         struct target           target;
  35         const char              *tracer;
  36         bool                    list_avail_functions;
  37         struct list_head        filters;
  38         struct list_head        notrace;
  39         struct list_head        graph_funcs;
  40         struct list_head        nograph_funcs;
  41         int                     graph_depth;
  42 };
  43
  44 struct filter_entry {
  45         struct list_head        list;
  46         char                    name[];
  47 };
  48
  49 static volatile int workload_exec_errno;
  50 static bool done;
  51
  52 static void sig_handler(int sig __maybe_unused)
  53 {
  54         done = true;
  55 }
  56
  57 /*
  58  * perf_evlist__prepare_workload will send a SIGUSR1 if the fork fails, since
  59  * we asked by setting its exec_error to the function below,
  60  * ftrace__workload_exec_failed_signal.
  61  *
  62  * XXX We need to handle this more appropriately, emitting an error, etc.
  63  */
  64 static void ftrace__workload_exec_failed_signal(int signo __maybe_unused,
  65                                                 siginfo_t *info __maybe_unused,
  66                                                 void *ucontext __maybe_unused)
  67 {
  68         workload_exec_errno = info->si_value.sival_int;
  69         done = true;
  70 }
  71
  72 static int __write_tracing_file(const char *name, const char *val, bool append)
  73 {
  74         char *file;
  75         int fd, ret = -1;
  76         ssize_t size = strlen(val);
  77         int flags = O_WRONLY;
  78         char errbuf[512];
  79         char *val_copy;
  80
  81         file = get_tracing_file(name);
  82         if (!file) {
  83                 pr_debug("cannot get tracing file: %s\n", name);
  84                 return -1;
  85         }
  86
  87         if (append)
  88                 flags |= O_APPEND;
  89         else
  90                 flags |= O_TRUNC;
  91
  92         fd = open(file, flags);
  93         if (fd < 0) {
  94                 pr_debug("cannot open tracing file: %s: %s\n",
  95                          name, str_error_r(errno, errbuf, sizeof(errbuf)));
  96                 goto out;
  97         }
  98
  99         /*
 100          * Copy the original value and append a '\n'. Without this,
 101          * the kernel can hide possible errors.
 102          */
 103         val_copy = strdup(val);
 104         if (!val_copy)
 105                 goto out_close;
 106         val_copy[size] = '\n';
 107
 108         if (write(fd, val_copy, size + 1) == size + 1)
 109                 ret = 0;
 110         else
 111                 pr_debug("write '%s' to tracing/%s failed: %s\n",
 112                          val, name, str_error_r(errno, errbuf, sizeof(errbuf)));
 113
 114         free(val_copy);
 115 out_close:
 116         close(fd);
 117 out:
 118         put_tracing_file(file);
 119         return ret;
 120 }
 121
 122 static int write_tracing_file(const char *name, const char *val)
 123 {
 124         return __write_tracing_file(name, val, false);
 125 }
 126
 127 static int append_tracing_file(const char *name, const char *val)
 128 {
 129         return __write_tracing_file(name, val, true);
 130 }
 131
 132 static int read_tracing_file_to_stdout(const char *name)
 133 {
 134         char buf[4096];
 135         char *file;
 136         int fd;
 137         int ret = -1;
 138
 139         file = get_tracing_file(name);
 140         if (!file) {
 141                 pr_debug("cannot get tracing file: %s\n", name);
 142                 return -1;
 143         }
 144
 145         fd = open(file, O_RDONLY);
 146         if (fd < 0) {
 147                 pr_debug("cannot open tracing file: %s: %s\n",
 148                          name, str_error_r(errno, buf, sizeof(buf)));
 149                 goto out;
 150         }
 151
 152         /* read contents to stdout */
 153         while (true) {
 154                 int n = read(fd, buf, sizeof(buf));
 155                 if (n == 0)
 156                         break;
 157                 else if (n < 0)
 158                         goto out_close;
 159
 160                 if (fwrite(buf, n, 1, stdout) != 1)
 161                         goto out_close;
 162         }
 163         ret = 0;
 164
 165 out_close:
 166         close(fd);
 167 out:
 168         put_tracing_file(file);
 169         return ret;
 170 }
 171
 172 static int reset_tracing_cpu(void);
 173 static void reset_tracing_filters(void);
 174
 175 static int reset_tracing_files(struct perf_ftrace *ftrace __maybe_unused)
 176 {
 177         if (write_tracing_file("tracing_on", "0") < 0)
 178                 return -1;
 179
 180         if (write_tracing_file("current_tracer", "nop") < 0)
 181                 return -1;
 182
 183         if (write_tracing_file("set_ftrace_pid", " ") < 0)
 184                 return -1;
 185
 186         if (reset_tracing_cpu() < 0)
 187                 return -1;
 188
 189         if (write_tracing_file("max_graph_depth", "0") < 0)
 190                 return -1;
 191
 192         reset_tracing_filters();
 193         return 0;
 194 }
 195
 196 static int set_tracing_pid(struct perf_ftrace *ftrace)
 197 {
 198         int i;
 199         char buf[16];
 200
 201         if (target__has_cpu(&ftrace->target))
 202                 return 0;
 203
 204         for (i = 0; i < perf_thread_map__nr(ftrace->evlist->core.threads); i++) {
 205                 scnprintf(buf, sizeof(buf), "%d",
 206                           ftrace->evlist->core.threads->map[i]);
 207                 if (append_tracing_file("set_ftrace_pid", buf) < 0)
 208                         return -1;
 209         }
 210         return 0;
 211 }
 212
 213 static int set_tracing_cpumask(struct perf_cpu_map *cpumap)
 214 {
 215         char *cpumask;
 216         size_t mask_size;
 217         int ret;
 218         int last_cpu;
 219
 220         last_cpu = cpu_map__cpu(cpumap, cpumap->nr - 1);
 221         mask_size = last_cpu / 4 + 2; /* one more byte for EOS */
 222         mask_size += last_cpu / 32; /* ',' is needed for every 32th cpus */
 223
 224         cpumask = malloc(mask_size);
 225         if (cpumask == NULL) {
 226                 pr_debug("failed to allocate cpu mask\n");
 227                 return -1;
 228         }
 229
 230         cpu_map__snprint_mask(cpumap, cpumask, mask_size);
 231
 232         ret = write_tracing_file("tracing_cpumask", cpumask);
 233
 234         free(cpumask);
 235         return ret;
 236 }
 237
 238 static int set_tracing_cpu(struct perf_ftrace *ftrace)
 239 {
 240         struct perf_cpu_map *cpumap = ftrace->evlist->core.cpus;
 241
 242         if (!target__has_cpu(&ftrace->target))
 243                 return 0;
 244
 245         return set_tracing_cpumask(cpumap);
 246 }
 247
 248 static int reset_tracing_cpu(void)
 249 {
 250         struct perf_cpu_map *cpumap = perf_cpu_map__new(NULL);
 251         int ret;
 252
 253         ret = set_tracing_cpumask(cpumap);
 254         perf_cpu_map__put(cpumap);
 255         return ret;
 256 }
 257
 258 static int __set_tracing_filter(const char *filter_file, struct list_head *funcs)
 259 {
 260         struct filter_entry *pos;
 261
 262         list_for_each_entry(pos, funcs, list) {
 263                 if (append_tracing_file(filter_file, pos->name) < 0)
 264                         return -1;
 265         }
 266
 267         return 0;
 268 }
 269
 270 static int set_tracing_filters(struct perf_ftrace *ftrace)
 271 {
 272         int ret;
 273
 274         ret = __set_tracing_filter("set_ftrace_filter", &ftrace->filters);
 275         if (ret < 0)
 276                 return ret;
 277
 278         ret = __set_tracing_filter("set_ftrace_notrace", &ftrace->notrace);
 279         if (ret < 0)
 280                 return ret;
 281
 282         ret = __set_tracing_filter("set_graph_function", &ftrace->graph_funcs);
 283         if (ret < 0)
 284                 return ret;
 285
 286         /* old kernels do not have this filter */
 287         __set_tracing_filter("set_graph_notrace", &ftrace->nograph_funcs);
 288
 289         return ret;
 290 }
 291
 292 static void reset_tracing_filters(void)
 293 {
 294         write_tracing_file("set_ftrace_filter", " ");
 295         write_tracing_file("set_ftrace_notrace", " ");
 296         write_tracing_file("set_graph_function", " ");
 297         write_tracing_file("set_graph_notrace", " ");
 298 }
 299
 300 static int set_tracing_depth(struct perf_ftrace *ftrace)
 301 {
 302         char buf[16];
 303
 304         if (ftrace->graph_depth == 0)
 305                 return 0;
 306
 307         if (ftrace->graph_depth < 0) {
 308                 pr_err("invalid graph depth: %d\n", ftrace->graph_depth);
 309                 return -1;
 310         }
 311
 312         snprintf(buf, sizeof(buf), "%d", ftrace->graph_depth);
 313
 314         if (write_tracing_file("max_graph_depth", buf) < 0)
 315                 return -1;
 316
 317         return 0;
 318 }
 319
 320 static int __cmd_ftrace(struct perf_ftrace *ftrace, int argc, const char **argv)
 321 {
 322         char *trace_file;
 323         int trace_fd;
 324         char buf[4096];
 325         struct pollfd pollfd = {
 326                 .events = POLLIN,
 327         };
 328
 329         if (!(perf_cap__capable(CAP_PERFMON) ||
 330               perf_cap__capable(CAP_SYS_ADMIN))) {
 331                 pr_err("ftrace only works for %s!\n",
 332 #ifdef HAVE_LIBCAP_SUPPORT
 333                 "users with the CAP_PERFMON or CAP_SYS_ADMIN capability"
 334 #else
 335                 "root"
 336 #endif
 337                 );
 338                 return -1;
 339         }
 340
 341         signal(SIGINT, sig_handler);
 342         signal(SIGUSR1, sig_handler);
 343         signal(SIGCHLD, sig_handler);
 344         signal(SIGPIPE, sig_handler);
 345
 346         if (ftrace->list_avail_functions)
 347                 return read_tracing_file_to_stdout("available_filter_functions");
 348
 349         if (reset_tracing_files(ftrace) < 0) {
 350                 pr_err("failed to reset ftrace\n");
 351                 goto out;
 352         }
 353
 354         /* reset ftrace buffer */
 355         if (write_tracing_file("trace", "0") < 0)
 356                 goto out;
 357
 358         if (argc && perf_evlist__prepare_workload(ftrace->evlist,
 359                                 &ftrace->target, argv, false,
 360                                 ftrace__workload_exec_failed_signal) < 0) {
 361                 goto out;
 362         }
 363
 364         if (set_tracing_pid(ftrace) < 0) {
 365                 pr_err("failed to set ftrace pid\n");
 366                 goto out_reset;
 367         }
 368
 369         if (set_tracing_cpu(ftrace) < 0) {
 370                 pr_err("failed to set tracing cpumask\n");
 371                 goto out_reset;
 372         }
 373
 374         if (set_tracing_filters(ftrace) < 0) {
 375                 pr_err("failed to set tracing filters\n");
 376                 goto out_reset;
 377         }
 378
 379         if (set_tracing_depth(ftrace) < 0) {
 380                 pr_err("failed to set graph depth\n");
 381                 goto out_reset;
 382         }
 383
 384         if (write_tracing_file("current_tracer", ftrace->tracer) < 0) {
 385                 pr_err("failed to set current_tracer to %s\n", ftrace->tracer);
 386                 goto out_reset;
 387         }
 388
 389         setup_pager();
 390
 391         trace_file = get_tracing_file("trace_pipe");
 392         if (!trace_file) {
 393                 pr_err("failed to open trace_pipe\n");
 394                 goto out_reset;
 395         }
 396
 397         trace_fd = open(trace_file, O_RDONLY);
 398
 399         put_tracing_file(trace_file);
 400
 401         if (trace_fd < 0) {
 402                 pr_err("failed to open trace_pipe\n");
 403                 goto out_reset;
 404         }
 405
 406         fcntl(trace_fd, F_SETFL, O_NONBLOCK);
 407         pollfd.fd = trace_fd;
 408
 409         if (write_tracing_file("tracing_on", "1") < 0) {
 410                 pr_err("can't enable tracing\n");
 411                 goto out_close_fd;
 412         }
 413
 414         perf_evlist__start_workload(ftrace->evlist);
 415
 416         while (!done) {
 417                 if (poll(&pollfd, 1, -1) < 0)
 418                         break;
 419
 420                 if (pollfd.revents & POLLIN) {
 421                         int n = read(trace_fd, buf, sizeof(buf));
 422                         if (n < 0)
 423                                 break;
 424                         if (fwrite(buf, n, 1, stdout) != 1)
 425                                 break;
 426                 }
 427         }
 428
 429         write_tracing_file("tracing_on", "0");
 430
 431         if (workload_exec_errno) {
 432                 const char *emsg = str_error_r(workload_exec_errno, buf, sizeof(buf));
 433                 /* flush stdout first so below error msg appears at the end. */
 434                 fflush(stdout);
 435                 pr_err("workload failed: %s\n", emsg);
 436                 goto out_close_fd;
 437         }
 438
 439         /* read remaining buffer contents */
 440         while (true) {
 441                 int n = read(trace_fd, buf, sizeof(buf));
 442                 if (n <= 0)
 443                         break;
 444                 if (fwrite(buf, n, 1, stdout) != 1)
 445                         break;
 446         }
 447
 448 out_close_fd:
 449         close(trace_fd);
 450 out_reset:
 451         reset_tracing_files(ftrace);
 452 out:
 453         return (done && !workload_exec_errno) ? 0 : -1;
 454 }
 455
 456 static int perf_ftrace_config(const char *var, const char *value, void *cb)
 457 {
 458         struct perf_ftrace *ftrace = cb;
 459
 460         if (!strstarts(var, "ftrace."))
 461                 return 0;
 462
 463         if (strcmp(var, "ftrace.tracer"))
 464                 return -1;
 465
 466         if (!strcmp(value, "function_graph") ||
 467             !strcmp(value, "function")) {
 468                 ftrace->tracer = value;
 469                 return 0;
 470         }
 471
 472         pr_err("Please select \"function_graph\" (default) or \"function\"\n");
 473         return -1;
 474 }
 475
 476 static int parse_filter_func(const struct option *opt, const char *str,
 477                              int unset __maybe_unused)
 478 {
 479         struct list_head *head = opt->value;
 480         struct filter_entry *entry;
 481
 482         entry = malloc(sizeof(*entry) + strlen(str) + 1);
 483         if (entry == NULL)
 484                 return -ENOMEM;
 485
 486         strcpy(entry->name, str);
 487         list_add_tail(&entry->list, head);
 488
 489         return 0;
 490 }
 491
 492 static void delete_filter_func(struct list_head *head)
 493 {
 494         struct filter_entry *pos, *tmp;
 495
 496         list_for_each_entry_safe(pos, tmp, head, list) {
 497                 list_del_init(&pos->list);
 498                 free(pos);
 499         }
 500 }
 501
 502 static void select_tracer(struct perf_ftrace *ftrace)
 503 {
 504         bool graph = !list_empty(&ftrace->graph_funcs) ||
 505                      !list_empty(&ftrace->nograph_funcs);
 506         bool func = !list_empty(&ftrace->filters) ||
 507                     !list_empty(&ftrace->notrace);
 508
 509         /* The function_graph has priority over function tracer. */
 510         if (graph)
 511                 ftrace->tracer = "function_graph";
 512         else if (func)
 513                 ftrace->tracer = "function";
 514         /* Otherwise, the default tracer is used. */
 515
 516         pr_debug("%s tracer is used\n", ftrace->tracer);
 517 }
 518
 519 int cmd_ftrace(int argc, const char **argv)
 520 {
 521         int ret;
 522         struct perf_ftrace ftrace = {
 523                 .tracer = DEFAULT_TRACER,
 524                 .target = { .uid = UINT_MAX, },
 525         };
 526         const char * const ftrace_usage[] = {
 527                 "perf ftrace [<options>] [<command>]",
 528                 "perf ftrace [<options>] -- <command> [<options>]",
 529                 NULL
 530         };
 531         const struct option ftrace_options[] = {
 532         OPT_STRING('t', "tracer", &ftrace.tracer, "tracer",
 533                    "tracer to use: function_graph(default) or function"),
 534         OPT_BOOLEAN('F', "funcs", &ftrace.list_avail_functions,
 535                     "Show available functions to filter"),
 536         OPT_STRING('p', "pid", &ftrace.target.pid, "pid",
 537                    "trace on existing process id"),
 538         OPT_INCR('v', "verbose", &verbose,
 539                  "be more verbose"),
 540         OPT_BOOLEAN('a', "all-cpus", &ftrace.target.system_wide,
 541                     "system-wide collection from all CPUs"),
 542         OPT_STRING('C', "cpu", &ftrace.target.cpu_list, "cpu",
 543                     "list of cpus to monitor"),
 544         OPT_CALLBACK('T', "trace-funcs", &ftrace.filters, "func",
 545                      "trace given functions using function tracer",
 546                      parse_filter_func),
 547         OPT_CALLBACK('N', "notrace-funcs", &ftrace.notrace, "func",
 548                      "do not trace given functions", parse_filter_func),
 549         OPT_CALLBACK('G', "graph-funcs", &ftrace.graph_funcs, "func",
 550                      "trace given functions using function_graph tracer",
 551                      parse_filter_func),
 552         OPT_CALLBACK('g', "nograph-funcs", &ftrace.nograph_funcs, "func",
 553                      "Set nograph filter on given functions", parse_filter_func),
 554         OPT_INTEGER('D', "graph-depth", &ftrace.graph_depth,
 555                     "Max depth for function graph tracer"),
 556         OPT_END()
 557         };
 558
 559         INIT_LIST_HEAD(&ftrace.filters);
 560         INIT_LIST_HEAD(&ftrace.notrace);
 561         INIT_LIST_HEAD(&ftrace.graph_funcs);
 562         INIT_LIST_HEAD(&ftrace.nograph_funcs);
 563
 564         ret = perf_config(perf_ftrace_config, &ftrace);
 565         if (ret < 0)
 566                 return -1;
 567
 568         argc = parse_options(argc, argv, ftrace_options, ftrace_usage,
 569                             PARSE_OPT_STOP_AT_NON_OPTION);
 570         if (!argc && target__none(&ftrace.target))
 571                 ftrace.target.system_wide = true;
 572
 573         select_tracer(&ftrace);
 574
 575         ret = target__validate(&ftrace.target);
 576         if (ret) {
 577                 char errbuf[512];
 578
 579                 target__strerror(&ftrace.target, ret, errbuf, 512);
 580                 pr_err("%s\n", errbuf);
 581                 goto out_delete_filters;
 582         }
 583
 584         ftrace.evlist = evlist__new();
 585         if (ftrace.evlist == NULL) {
 586                 ret = -ENOMEM;
 587                 goto out_delete_filters;
 588         }
 589
 590         ret = perf_evlist__create_maps(ftrace.evlist, &ftrace.target);
 591         if (ret < 0)
 592                 goto out_delete_evlist;
 593
 594         ret = __cmd_ftrace(&ftrace, argc, argv);
 595
 596 out_delete_evlist:
 597         evlist__delete(ftrace.evlist);
 598
 599 out_delete_filters:
 600         delete_filter_func(&ftrace.filters);
 601         delete_filter_func(&ftrace.notrace);
 602         delete_filter_func(&ftrace.graph_funcs);
 603         delete_filter_func(&ftrace.nograph_funcs);
 604
 605         return ret;
 606 }