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 write_tracing_file_int(const char *name, int value)
 173 {
 174         char buf[16];
 175
 176         snprintf(buf, sizeof(buf), "%d", value);
 177         if (write_tracing_file(name, buf) < 0)
 178                 return -1;
 179
 180         return 0;
 181 }
 182
 183 static int reset_tracing_cpu(void);
 184 static void reset_tracing_filters(void);
 185
 186 static int reset_tracing_files(struct perf_ftrace *ftrace __maybe_unused)
 187 {
 188         if (write_tracing_file("tracing_on", "0") < 0)
 189                 return -1;
 190
 191         if (write_tracing_file("current_tracer", "nop") < 0)
 192                 return -1;
 193
 194         if (write_tracing_file("set_ftrace_pid", " ") < 0)
 195                 return -1;
 196
 197         if (reset_tracing_cpu() < 0)
 198                 return -1;
 199
 200         if (write_tracing_file("max_graph_depth", "0") < 0)
 201                 return -1;
 202
 203         reset_tracing_filters();
 204         return 0;
 205 }
 206
 207 static int set_tracing_pid(struct perf_ftrace *ftrace)
 208 {
 209         int i;
 210         char buf[16];
 211
 212         if (target__has_cpu(&ftrace->target))
 213                 return 0;
 214
 215         for (i = 0; i < perf_thread_map__nr(ftrace->evlist->core.threads); i++) {
 216                 scnprintf(buf, sizeof(buf), "%d",
 217                           ftrace->evlist->core.threads->map[i]);
 218                 if (append_tracing_file("set_ftrace_pid", buf) < 0)
 219                         return -1;
 220         }
 221         return 0;
 222 }
 223
 224 static int set_tracing_cpumask(struct perf_cpu_map *cpumap)
 225 {
 226         char *cpumask;
 227         size_t mask_size;
 228         int ret;
 229         int last_cpu;
 230
 231         last_cpu = cpu_map__cpu(cpumap, cpumap->nr - 1);
 232         mask_size = last_cpu / 4 + 2; /* one more byte for EOS */
 233         mask_size += last_cpu / 32; /* ',' is needed for every 32th cpus */
 234
 235         cpumask = malloc(mask_size);
 236         if (cpumask == NULL) {
 237                 pr_debug("failed to allocate cpu mask\n");
 238                 return -1;
 239         }
 240
 241         cpu_map__snprint_mask(cpumap, cpumask, mask_size);
 242
 243         ret = write_tracing_file("tracing_cpumask", cpumask);
 244
 245         free(cpumask);
 246         return ret;
 247 }
 248
 249 static int set_tracing_cpu(struct perf_ftrace *ftrace)
 250 {
 251         struct perf_cpu_map *cpumap = ftrace->evlist->core.cpus;
 252
 253         if (!target__has_cpu(&ftrace->target))
 254                 return 0;
 255
 256         return set_tracing_cpumask(cpumap);
 257 }
 258
 259 static int reset_tracing_cpu(void)
 260 {
 261         struct perf_cpu_map *cpumap = perf_cpu_map__new(NULL);
 262         int ret;
 263
 264         ret = set_tracing_cpumask(cpumap);
 265         perf_cpu_map__put(cpumap);
 266         return ret;
 267 }
 268
 269 static int __set_tracing_filter(const char *filter_file, struct list_head *funcs)
 270 {
 271         struct filter_entry *pos;
 272
 273         list_for_each_entry(pos, funcs, list) {
 274                 if (append_tracing_file(filter_file, pos->name) < 0)
 275                         return -1;
 276         }
 277
 278         return 0;
 279 }
 280
 281 static int set_tracing_filters(struct perf_ftrace *ftrace)
 282 {
 283         int ret;
 284
 285         ret = __set_tracing_filter("set_ftrace_filter", &ftrace->filters);
 286         if (ret < 0)
 287                 return ret;
 288
 289         ret = __set_tracing_filter("set_ftrace_notrace", &ftrace->notrace);
 290         if (ret < 0)
 291                 return ret;
 292
 293         ret = __set_tracing_filter("set_graph_function", &ftrace->graph_funcs);
 294         if (ret < 0)
 295                 return ret;
 296
 297         /* old kernels do not have this filter */
 298         __set_tracing_filter("set_graph_notrace", &ftrace->nograph_funcs);
 299
 300         return ret;
 301 }
 302
 303 static void reset_tracing_filters(void)
 304 {
 305         write_tracing_file("set_ftrace_filter", " ");
 306         write_tracing_file("set_ftrace_notrace", " ");
 307         write_tracing_file("set_graph_function", " ");
 308         write_tracing_file("set_graph_notrace", " ");
 309 }
 310
 311 static int set_tracing_depth(struct perf_ftrace *ftrace)
 312 {
 313         if (ftrace->graph_depth == 0)
 314                 return 0;
 315
 316         if (ftrace->graph_depth < 0) {
 317                 pr_err("invalid graph depth: %d\n", ftrace->graph_depth);
 318                 return -1;
 319         }
 320
 321         if (write_tracing_file_int("max_graph_depth", ftrace->graph_depth) < 0)
 322                 return -1;
 323
 324         return 0;
 325 }
 326
 327 static int __cmd_ftrace(struct perf_ftrace *ftrace, int argc, const char **argv)
 328 {
 329         char *trace_file;
 330         int trace_fd;
 331         char buf[4096];
 332         struct pollfd pollfd = {
 333                 .events = POLLIN,
 334         };
 335
 336         if (!(perf_cap__capable(CAP_PERFMON) ||
 337               perf_cap__capable(CAP_SYS_ADMIN))) {
 338                 pr_err("ftrace only works for %s!\n",
 339 #ifdef HAVE_LIBCAP_SUPPORT
 340                 "users with the CAP_PERFMON or CAP_SYS_ADMIN capability"
 341 #else
 342                 "root"
 343 #endif
 344                 );
 345                 return -1;
 346         }
 347
 348         signal(SIGINT, sig_handler);
 349         signal(SIGUSR1, sig_handler);
 350         signal(SIGCHLD, sig_handler);
 351         signal(SIGPIPE, sig_handler);
 352
 353         if (ftrace->list_avail_functions)
 354                 return read_tracing_file_to_stdout("available_filter_functions");
 355
 356         if (reset_tracing_files(ftrace) < 0) {
 357                 pr_err("failed to reset ftrace\n");
 358                 goto out;
 359         }
 360
 361         /* reset ftrace buffer */
 362         if (write_tracing_file("trace", "0") < 0)
 363                 goto out;
 364
 365         if (argc && perf_evlist__prepare_workload(ftrace->evlist,
 366                                 &ftrace->target, argv, false,
 367                                 ftrace__workload_exec_failed_signal) < 0) {
 368                 goto out;
 369         }
 370
 371         if (set_tracing_pid(ftrace) < 0) {
 372                 pr_err("failed to set ftrace pid\n");
 373                 goto out_reset;
 374         }
 375
 376         if (set_tracing_cpu(ftrace) < 0) {
 377                 pr_err("failed to set tracing cpumask\n");
 378                 goto out_reset;
 379         }
 380
 381         if (set_tracing_filters(ftrace) < 0) {
 382                 pr_err("failed to set tracing filters\n");
 383                 goto out_reset;
 384         }
 385
 386         if (set_tracing_depth(ftrace) < 0) {
 387                 pr_err("failed to set graph depth\n");
 388                 goto out_reset;
 389         }
 390
 391         if (write_tracing_file("current_tracer", ftrace->tracer) < 0) {
 392                 pr_err("failed to set current_tracer to %s\n", ftrace->tracer);
 393                 goto out_reset;
 394         }
 395
 396         setup_pager();
 397
 398         trace_file = get_tracing_file("trace_pipe");
 399         if (!trace_file) {
 400                 pr_err("failed to open trace_pipe\n");
 401                 goto out_reset;
 402         }
 403
 404         trace_fd = open(trace_file, O_RDONLY);
 405
 406         put_tracing_file(trace_file);
 407
 408         if (trace_fd < 0) {
 409                 pr_err("failed to open trace_pipe\n");
 410                 goto out_reset;
 411         }
 412
 413         fcntl(trace_fd, F_SETFL, O_NONBLOCK);
 414         pollfd.fd = trace_fd;
 415
 416         if (write_tracing_file("tracing_on", "1") < 0) {
 417                 pr_err("can't enable tracing\n");
 418                 goto out_close_fd;
 419         }
 420
 421         perf_evlist__start_workload(ftrace->evlist);
 422
 423         while (!done) {
 424                 if (poll(&pollfd, 1, -1) < 0)
 425                         break;
 426
 427                 if (pollfd.revents & POLLIN) {
 428                         int n = read(trace_fd, buf, sizeof(buf));
 429                         if (n < 0)
 430                                 break;
 431                         if (fwrite(buf, n, 1, stdout) != 1)
 432                                 break;
 433                 }
 434         }
 435
 436         write_tracing_file("tracing_on", "0");
 437
 438         if (workload_exec_errno) {
 439                 const char *emsg = str_error_r(workload_exec_errno, buf, sizeof(buf));
 440                 /* flush stdout first so below error msg appears at the end. */
 441                 fflush(stdout);
 442                 pr_err("workload failed: %s\n", emsg);
 443                 goto out_close_fd;
 444         }
 445
 446         /* read remaining buffer contents */
 447         while (true) {
 448                 int n = read(trace_fd, buf, sizeof(buf));
 449                 if (n <= 0)
 450                         break;
 451                 if (fwrite(buf, n, 1, stdout) != 1)
 452                         break;
 453         }
 454
 455 out_close_fd:
 456         close(trace_fd);
 457 out_reset:
 458         reset_tracing_files(ftrace);
 459 out:
 460         return (done && !workload_exec_errno) ? 0 : -1;
 461 }
 462
 463 static int perf_ftrace_config(const char *var, const char *value, void *cb)
 464 {
 465         struct perf_ftrace *ftrace = cb;
 466
 467         if (!strstarts(var, "ftrace."))
 468                 return 0;
 469
 470         if (strcmp(var, "ftrace.tracer"))
 471                 return -1;
 472
 473         if (!strcmp(value, "function_graph") ||
 474             !strcmp(value, "function")) {
 475                 ftrace->tracer = value;
 476                 return 0;
 477         }
 478
 479         pr_err("Please select \"function_graph\" (default) or \"function\"\n");
 480         return -1;
 481 }
 482
 483 static int parse_filter_func(const struct option *opt, const char *str,
 484                              int unset __maybe_unused)
 485 {
 486         struct list_head *head = opt->value;
 487         struct filter_entry *entry;
 488
 489         entry = malloc(sizeof(*entry) + strlen(str) + 1);
 490         if (entry == NULL)
 491                 return -ENOMEM;
 492
 493         strcpy(entry->name, str);
 494         list_add_tail(&entry->list, head);
 495
 496         return 0;
 497 }
 498
 499 static void delete_filter_func(struct list_head *head)
 500 {
 501         struct filter_entry *pos, *tmp;
 502
 503         list_for_each_entry_safe(pos, tmp, head, list) {
 504                 list_del_init(&pos->list);
 505                 free(pos);
 506         }
 507 }
 508
 509 static void select_tracer(struct perf_ftrace *ftrace)
 510 {
 511         bool graph = !list_empty(&ftrace->graph_funcs) ||
 512                      !list_empty(&ftrace->nograph_funcs);
 513         bool func = !list_empty(&ftrace->filters) ||
 514                     !list_empty(&ftrace->notrace);
 515
 516         /* The function_graph has priority over function tracer. */
 517         if (graph)
 518                 ftrace->tracer = "function_graph";
 519         else if (func)
 520                 ftrace->tracer = "function";
 521         /* Otherwise, the default tracer is used. */
 522
 523         pr_debug("%s tracer is used\n", ftrace->tracer);
 524 }
 525
 526 int cmd_ftrace(int argc, const char **argv)
 527 {
 528         int ret;
 529         struct perf_ftrace ftrace = {
 530                 .tracer = DEFAULT_TRACER,
 531                 .target = { .uid = UINT_MAX, },
 532         };
 533         const char * const ftrace_usage[] = {
 534                 "perf ftrace [<options>] [<command>]",
 535                 "perf ftrace [<options>] -- <command> [<options>]",
 536                 NULL
 537         };
 538         const struct option ftrace_options[] = {
 539         OPT_STRING('t', "tracer", &ftrace.tracer, "tracer",
 540                    "tracer to use: function_graph(default) or function"),
 541         OPT_BOOLEAN('F', "funcs", &ftrace.list_avail_functions,
 542                     "Show available functions to filter"),
 543         OPT_STRING('p', "pid", &ftrace.target.pid, "pid",
 544                    "trace on existing process id"),
 545         OPT_INCR('v', "verbose", &verbose,
 546                  "be more verbose"),
 547         OPT_BOOLEAN('a', "all-cpus", &ftrace.target.system_wide,
 548                     "system-wide collection from all CPUs"),
 549         OPT_STRING('C', "cpu", &ftrace.target.cpu_list, "cpu",
 550                     "list of cpus to monitor"),
 551         OPT_CALLBACK('T', "trace-funcs", &ftrace.filters, "func",
 552                      "trace given functions using function tracer",
 553                      parse_filter_func),
 554         OPT_CALLBACK('N', "notrace-funcs", &ftrace.notrace, "func",
 555                      "do not trace given functions", parse_filter_func),
 556         OPT_CALLBACK('G', "graph-funcs", &ftrace.graph_funcs, "func",
 557                      "trace given functions using function_graph tracer",
 558                      parse_filter_func),
 559         OPT_CALLBACK('g', "nograph-funcs", &ftrace.nograph_funcs, "func",
 560                      "Set nograph filter on given functions", parse_filter_func),
 561         OPT_INTEGER('D', "graph-depth", &ftrace.graph_depth,
 562                     "Max depth for function graph tracer"),
 563         OPT_END()
 564         };
 565
 566         INIT_LIST_HEAD(&ftrace.filters);
 567         INIT_LIST_HEAD(&ftrace.notrace);
 568         INIT_LIST_HEAD(&ftrace.graph_funcs);
 569         INIT_LIST_HEAD(&ftrace.nograph_funcs);
 570
 571         ret = perf_config(perf_ftrace_config, &ftrace);
 572         if (ret < 0)
 573                 return -1;
 574
 575         argc = parse_options(argc, argv, ftrace_options, ftrace_usage,
 576                             PARSE_OPT_STOP_AT_NON_OPTION);
 577         if (!argc && target__none(&ftrace.target))
 578                 ftrace.target.system_wide = true;
 579
 580         select_tracer(&ftrace);
 581
 582         ret = target__validate(&ftrace.target);
 583         if (ret) {
 584                 char errbuf[512];
 585
 586                 target__strerror(&ftrace.target, ret, errbuf, 512);
 587                 pr_err("%s\n", errbuf);
 588                 goto out_delete_filters;
 589         }
 590
 591         ftrace.evlist = evlist__new();
 592         if (ftrace.evlist == NULL) {
 593                 ret = -ENOMEM;
 594                 goto out_delete_filters;
 595         }
 596
 597         ret = perf_evlist__create_maps(ftrace.evlist, &ftrace.target);
 598         if (ret < 0)
 599                 goto out_delete_evlist;
 600
 601         ret = __cmd_ftrace(&ftrace, argc, argv);
 602
 603 out_delete_evlist:
 604         evlist__delete(ftrace.evlist);
 605
 606 out_delete_filters:
 607         delete_filter_func(&ftrace.filters);
 608         delete_filter_func(&ftrace.notrace);
 609         delete_filter_func(&ftrace.graph_funcs);
 610         delete_filter_func(&ftrace.nograph_funcs);
 611
 612         return ret;
 613 }