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