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 <inttypes.h>
17#include <math.h>
18#include <poll.h>
19#include <ctype.h>
20#include <linux/capability.h>
21#include <linux/string.h>
22
23#include "debug.h"
24#include <subcmd/pager.h>
25#include <subcmd/parse-options.h>
26#include <api/io.h>
27#include <api/fs/tracing_path.h>
28#include "evlist.h"
29#include "target.h"
30#include "cpumap.h"
31#include "hashmap.h"
32#include "thread_map.h"
33#include "strfilter.h"
34#include "util/cap.h"
35#include "util/config.h"
36#include "util/ftrace.h"
37#include "util/stat.h"
38#include "util/units.h"
39#include "util/parse-sublevel-options.h"
40
41#define DEFAULT_TRACER "function_graph"
42
43static volatile sig_atomic_t workload_exec_errno;
44static volatile sig_atomic_t done;
45
46static void sig_handler(int sig __maybe_unused)
47{
48 done = true;
49}
50
51/*
52 * evlist__prepare_workload will send a SIGUSR1 if the fork fails, since
53 * we asked by setting its exec_error to the function below,
54 * ftrace__workload_exec_failed_signal.
55 *
56 * XXX We need to handle this more appropriately, emitting an error, etc.
57 */
58static void ftrace__workload_exec_failed_signal(int signo __maybe_unused,
59 siginfo_t *info __maybe_unused,
60 void *ucontext __maybe_unused)
61{
62 workload_exec_errno = info->si_value.sival_int;
63 done = true;
64}
65
66static bool check_ftrace_capable(void)
67{
68 bool used_root;
69
70 if (perf_cap__capable(CAP_PERFMON, &used_root))
71 return true;
72
73 if (!used_root && perf_cap__capable(CAP_SYS_ADMIN, &used_root))
74 return true;
75
76 pr_err("ftrace only works for %s!\n",
77 used_root ? "root"
78 : "users with the CAP_PERFMON or CAP_SYS_ADMIN capability"
79 );
80 return false;
81}
82
83static bool is_ftrace_supported(void)
84{
85 char *file;
86 bool supported = false;
87
88 file = get_tracing_file("set_ftrace_pid");
89 if (!file) {
90 pr_debug("cannot get tracing file set_ftrace_pid\n");
91 return false;
92 }
93
94 if (!access(file, F_OK))
95 supported = true;
96
97 put_tracing_file(file);
98 return supported;
99}
100
101static int __write_tracing_file(const char *name, const char *val, bool append)
102{
103 char *file;
104 int fd, ret = -1;
105 ssize_t size = strlen(val);
106 int flags = O_WRONLY;
107 char errbuf[512];
108 char *val_copy;
109
110 file = get_tracing_file(name);
111 if (!file) {
112 pr_debug("cannot get tracing file: %s\n", name);
113 return -1;
114 }
115
116 if (append)
117 flags |= O_APPEND;
118 else
119 flags |= O_TRUNC;
120
121 fd = open(file, flags);
122 if (fd < 0) {
123 pr_debug("cannot open tracing file: %s: %s\n",
124 name, str_error_r(errno, errbuf, sizeof(errbuf)));
125 goto out;
126 }
127
128 /*
129 * Copy the original value and append a '\n'. Without this,
130 * the kernel can hide possible errors.
131 */
132 val_copy = strdup(val);
133 if (!val_copy)
134 goto out_close;
135 val_copy[size] = '\n';
136
137 if (write(fd, val_copy, size + 1) == size + 1)
138 ret = 0;
139 else
140 pr_debug("write '%s' to tracing/%s failed: %s\n",
141 val, name, str_error_r(errno, errbuf, sizeof(errbuf)));
142
143 free(val_copy);
144out_close:
145 close(fd);
146out:
147 put_tracing_file(file);
148 return ret;
149}
150
151static int write_tracing_file(const char *name, const char *val)
152{
153 return __write_tracing_file(name, val, false);
154}
155
156static int append_tracing_file(const char *name, const char *val)
157{
158 return __write_tracing_file(name, val, true);
159}
160
161static int read_tracing_file_to_stdout(const char *name)
162{
163 char buf[4096];
164 char *file;
165 int fd;
166 int ret = -1;
167
168 file = get_tracing_file(name);
169 if (!file) {
170 pr_debug("cannot get tracing file: %s\n", name);
171 return -1;
172 }
173
174 fd = open(file, O_RDONLY);
175 if (fd < 0) {
176 pr_debug("cannot open tracing file: %s: %s\n",
177 name, str_error_r(errno, buf, sizeof(buf)));
178 goto out;
179 }
180
181 /* read contents to stdout */
182 while (true) {
183 int n = read(fd, buf, sizeof(buf));
184 if (n == 0)
185 break;
186 else if (n < 0)
187 goto out_close;
188
189 if (fwrite(buf, n, 1, stdout) != 1)
190 goto out_close;
191 }
192 ret = 0;
193
194out_close:
195 close(fd);
196out:
197 put_tracing_file(file);
198 return ret;
199}
200
201static int read_tracing_file_by_line(const char *name,
202 void (*cb)(char *str, void *arg),
203 void *cb_arg)
204{
205 char *line = NULL;
206 size_t len = 0;
207 char *file;
208 FILE *fp;
209
210 file = get_tracing_file(name);
211 if (!file) {
212 pr_debug("cannot get tracing file: %s\n", name);
213 return -1;
214 }
215
216 fp = fopen(file, "r");
217 if (fp == NULL) {
218 pr_debug("cannot open tracing file: %s\n", name);
219 put_tracing_file(file);
220 return -1;
221 }
222
223 while (getline(&line, &len, fp) != -1) {
224 cb(line, cb_arg);
225 }
226
227 if (line)
228 free(line);
229
230 fclose(fp);
231 put_tracing_file(file);
232 return 0;
233}
234
235static int write_tracing_file_int(const char *name, int value)
236{
237 char buf[16];
238
239 snprintf(buf, sizeof(buf), "%d", value);
240 if (write_tracing_file(name, buf) < 0)
241 return -1;
242
243 return 0;
244}
245
246static int write_tracing_option_file(const char *name, const char *val)
247{
248 char *file;
249 int ret;
250
251 if (asprintf(&file, "options/%s", name) < 0)
252 return -1;
253
254 ret = __write_tracing_file(file, val, false);
255 free(file);
256 return ret;
257}
258
259static int reset_tracing_cpu(void);
260static void reset_tracing_filters(void);
261
262static void reset_tracing_options(struct perf_ftrace *ftrace __maybe_unused)
263{
264 write_tracing_option_file("function-fork", "0");
265 write_tracing_option_file("func_stack_trace", "0");
266 write_tracing_option_file("sleep-time", "1");
267 write_tracing_option_file("funcgraph-irqs", "1");
268 write_tracing_option_file("funcgraph-proc", "0");
269 write_tracing_option_file("funcgraph-abstime", "0");
270 write_tracing_option_file("funcgraph-tail", "0");
271 write_tracing_option_file("latency-format", "0");
272 write_tracing_option_file("irq-info", "0");
273}
274
275static int reset_tracing_files(struct perf_ftrace *ftrace __maybe_unused)
276{
277 if (write_tracing_file("tracing_on", "0") < 0)
278 return -1;
279
280 if (write_tracing_file("current_tracer", "nop") < 0)
281 return -1;
282
283 if (write_tracing_file("set_ftrace_pid", " ") < 0)
284 return -1;
285
286 if (reset_tracing_cpu() < 0)
287 return -1;
288
289 if (write_tracing_file("max_graph_depth", "0") < 0)
290 return -1;
291
292 if (write_tracing_file("tracing_thresh", "0") < 0)
293 return -1;
294
295 reset_tracing_filters();
296 reset_tracing_options(ftrace);
297 return 0;
298}
299
300static int set_tracing_pid(struct perf_ftrace *ftrace)
301{
302 int i;
303 char buf[16];
304
305 if (target__has_cpu(&ftrace->target))
306 return 0;
307
308 for (i = 0; i < perf_thread_map__nr(ftrace->evlist->core.threads); i++) {
309 scnprintf(buf, sizeof(buf), "%d",
310 perf_thread_map__pid(ftrace->evlist->core.threads, i));
311 if (append_tracing_file("set_ftrace_pid", buf) < 0)
312 return -1;
313 }
314 return 0;
315}
316
317static int set_tracing_cpumask(struct perf_cpu_map *cpumap)
318{
319 char *cpumask;
320 size_t mask_size;
321 int ret;
322 int last_cpu;
323
324 last_cpu = perf_cpu_map__cpu(cpumap, perf_cpu_map__nr(cpumap) - 1).cpu;
325 mask_size = last_cpu / 4 + 2; /* one more byte for EOS */
326 mask_size += last_cpu / 32; /* ',' is needed for every 32th cpus */
327
328 cpumask = malloc(mask_size);
329 if (cpumask == NULL) {
330 pr_debug("failed to allocate cpu mask\n");
331 return -1;
332 }
333
334 cpu_map__snprint_mask(cpumap, cpumask, mask_size);
335
336 ret = write_tracing_file("tracing_cpumask", cpumask);
337
338 free(cpumask);
339 return ret;
340}
341
342static int set_tracing_cpu(struct perf_ftrace *ftrace)
343{
344 struct perf_cpu_map *cpumap = ftrace->evlist->core.user_requested_cpus;
345
346 if (!target__has_cpu(&ftrace->target))
347 return 0;
348
349 return set_tracing_cpumask(cpumap);
350}
351
352static int set_tracing_func_stack_trace(struct perf_ftrace *ftrace)
353{
354 if (!ftrace->func_stack_trace)
355 return 0;
356
357 if (write_tracing_option_file("func_stack_trace", "1") < 0)
358 return -1;
359
360 return 0;
361}
362
363static int set_tracing_func_irqinfo(struct perf_ftrace *ftrace)
364{
365 if (!ftrace->func_irq_info)
366 return 0;
367
368 if (write_tracing_option_file("irq-info", "1") < 0)
369 return -1;
370
371 return 0;
372}
373
374static int reset_tracing_cpu(void)
375{
376 struct perf_cpu_map *cpumap = perf_cpu_map__new_online_cpus();
377 int ret;
378
379 ret = set_tracing_cpumask(cpumap);
380 perf_cpu_map__put(cpumap);
381 return ret;
382}
383
384static int __set_tracing_filter(const char *filter_file, struct list_head *funcs)
385{
386 struct filter_entry *pos;
387
388 list_for_each_entry(pos, funcs, list) {
389 if (append_tracing_file(filter_file, pos->name) < 0)
390 return -1;
391 }
392
393 return 0;
394}
395
396static int set_tracing_filters(struct perf_ftrace *ftrace)
397{
398 int ret;
399
400 ret = __set_tracing_filter("set_ftrace_filter", &ftrace->filters);
401 if (ret < 0)
402 return ret;
403
404 ret = __set_tracing_filter("set_ftrace_notrace", &ftrace->notrace);
405 if (ret < 0)
406 return ret;
407
408 ret = __set_tracing_filter("set_graph_function", &ftrace->graph_funcs);
409 if (ret < 0)
410 return ret;
411
412 /* old kernels do not have this filter */
413 __set_tracing_filter("set_graph_notrace", &ftrace->nograph_funcs);
414
415 return ret;
416}
417
418static void reset_tracing_filters(void)
419{
420 write_tracing_file("set_ftrace_filter", " ");
421 write_tracing_file("set_ftrace_notrace", " ");
422 write_tracing_file("set_graph_function", " ");
423 write_tracing_file("set_graph_notrace", " ");
424}
425
426static int set_tracing_depth(struct perf_ftrace *ftrace)
427{
428 if (ftrace->graph_depth == 0)
429 return 0;
430
431 if (ftrace->graph_depth < 0) {
432 pr_err("invalid graph depth: %d\n", ftrace->graph_depth);
433 return -1;
434 }
435
436 if (write_tracing_file_int("max_graph_depth", ftrace->graph_depth) < 0)
437 return -1;
438
439 return 0;
440}
441
442static int set_tracing_percpu_buffer_size(struct perf_ftrace *ftrace)
443{
444 int ret;
445
446 if (ftrace->percpu_buffer_size == 0)
447 return 0;
448
449 ret = write_tracing_file_int("buffer_size_kb",
450 ftrace->percpu_buffer_size / 1024);
451 if (ret < 0)
452 return ret;
453
454 return 0;
455}
456
457static int set_tracing_trace_inherit(struct perf_ftrace *ftrace)
458{
459 if (!ftrace->inherit)
460 return 0;
461
462 if (write_tracing_option_file("function-fork", "1") < 0)
463 return -1;
464
465 return 0;
466}
467
468static int set_tracing_sleep_time(struct perf_ftrace *ftrace)
469{
470 if (!ftrace->graph_nosleep_time)
471 return 0;
472
473 if (write_tracing_option_file("sleep-time", "0") < 0)
474 return -1;
475
476 return 0;
477}
478
479static int set_tracing_funcgraph_irqs(struct perf_ftrace *ftrace)
480{
481 if (!ftrace->graph_noirqs)
482 return 0;
483
484 if (write_tracing_option_file("funcgraph-irqs", "0") < 0)
485 return -1;
486
487 return 0;
488}
489
490static int set_tracing_funcgraph_verbose(struct perf_ftrace *ftrace)
491{
492 if (!ftrace->graph_verbose)
493 return 0;
494
495 if (write_tracing_option_file("funcgraph-proc", "1") < 0)
496 return -1;
497
498 if (write_tracing_option_file("funcgraph-abstime", "1") < 0)
499 return -1;
500
501 if (write_tracing_option_file("latency-format", "1") < 0)
502 return -1;
503
504 return 0;
505}
506
507static int set_tracing_funcgraph_tail(struct perf_ftrace *ftrace)
508{
509 if (!ftrace->graph_tail)
510 return 0;
511
512 if (write_tracing_option_file("funcgraph-tail", "1") < 0)
513 return -1;
514
515 return 0;
516}
517
518static int set_tracing_thresh(struct perf_ftrace *ftrace)
519{
520 int ret;
521
522 if (ftrace->graph_thresh == 0)
523 return 0;
524
525 ret = write_tracing_file_int("tracing_thresh", ftrace->graph_thresh);
526 if (ret < 0)
527 return ret;
528
529 return 0;
530}
531
532static int set_tracing_options(struct perf_ftrace *ftrace)
533{
534 if (set_tracing_pid(ftrace) < 0) {
535 pr_err("failed to set ftrace pid\n");
536 return -1;
537 }
538
539 if (set_tracing_cpu(ftrace) < 0) {
540 pr_err("failed to set tracing cpumask\n");
541 return -1;
542 }
543
544 if (set_tracing_func_stack_trace(ftrace) < 0) {
545 pr_err("failed to set tracing option func_stack_trace\n");
546 return -1;
547 }
548
549 if (set_tracing_func_irqinfo(ftrace) < 0) {
550 pr_err("failed to set tracing option irq-info\n");
551 return -1;
552 }
553
554 if (set_tracing_filters(ftrace) < 0) {
555 pr_err("failed to set tracing filters\n");
556 return -1;
557 }
558
559 if (set_tracing_depth(ftrace) < 0) {
560 pr_err("failed to set graph depth\n");
561 return -1;
562 }
563
564 if (set_tracing_percpu_buffer_size(ftrace) < 0) {
565 pr_err("failed to set tracing per-cpu buffer size\n");
566 return -1;
567 }
568
569 if (set_tracing_trace_inherit(ftrace) < 0) {
570 pr_err("failed to set tracing option function-fork\n");
571 return -1;
572 }
573
574 if (set_tracing_sleep_time(ftrace) < 0) {
575 pr_err("failed to set tracing option sleep-time\n");
576 return -1;
577 }
578
579 if (set_tracing_funcgraph_irqs(ftrace) < 0) {
580 pr_err("failed to set tracing option funcgraph-irqs\n");
581 return -1;
582 }
583
584 if (set_tracing_funcgraph_verbose(ftrace) < 0) {
585 pr_err("failed to set tracing option funcgraph-proc/funcgraph-abstime\n");
586 return -1;
587 }
588
589 if (set_tracing_thresh(ftrace) < 0) {
590 pr_err("failed to set tracing thresh\n");
591 return -1;
592 }
593
594 if (set_tracing_funcgraph_tail(ftrace) < 0) {
595 pr_err("failed to set tracing option funcgraph-tail\n");
596 return -1;
597 }
598
599 return 0;
600}
601
602static void select_tracer(struct perf_ftrace *ftrace)
603{
604 bool graph = !list_empty(&ftrace->graph_funcs) ||
605 !list_empty(&ftrace->nograph_funcs);
606 bool func = !list_empty(&ftrace->filters) ||
607 !list_empty(&ftrace->notrace);
608
609 /* The function_graph has priority over function tracer. */
610 if (graph)
611 ftrace->tracer = "function_graph";
612 else if (func)
613 ftrace->tracer = "function";
614 /* Otherwise, the default tracer is used. */
615
616 pr_debug("%s tracer is used\n", ftrace->tracer);
617}
618
619static int __cmd_ftrace(struct perf_ftrace *ftrace)
620{
621 char *trace_file;
622 int trace_fd;
623 char buf[4096];
624 struct pollfd pollfd = {
625 .events = POLLIN,
626 };
627
628 select_tracer(ftrace);
629
630 if (reset_tracing_files(ftrace) < 0) {
631 pr_err("failed to reset ftrace\n");
632 goto out;
633 }
634
635 /* reset ftrace buffer */
636 if (write_tracing_file("trace", "0") < 0)
637 goto out;
638
639 if (set_tracing_options(ftrace) < 0)
640 goto out_reset;
641
642 if (write_tracing_file("current_tracer", ftrace->tracer) < 0) {
643 pr_err("failed to set current_tracer to %s\n", ftrace->tracer);
644 goto out_reset;
645 }
646
647 setup_pager();
648
649 trace_file = get_tracing_file("trace_pipe");
650 if (!trace_file) {
651 pr_err("failed to open trace_pipe\n");
652 goto out_reset;
653 }
654
655 trace_fd = open(trace_file, O_RDONLY);
656
657 put_tracing_file(trace_file);
658
659 if (trace_fd < 0) {
660 pr_err("failed to open trace_pipe\n");
661 goto out_reset;
662 }
663
664 fcntl(trace_fd, F_SETFL, O_NONBLOCK);
665 pollfd.fd = trace_fd;
666
667 /* display column headers */
668 read_tracing_file_to_stdout("trace");
669
670 if (!ftrace->target.initial_delay) {
671 if (write_tracing_file("tracing_on", "1") < 0) {
672 pr_err("can't enable tracing\n");
673 goto out_close_fd;
674 }
675 }
676
677 evlist__start_workload(ftrace->evlist);
678
679 if (ftrace->target.initial_delay > 0) {
680 usleep(ftrace->target.initial_delay * 1000);
681 if (write_tracing_file("tracing_on", "1") < 0) {
682 pr_err("can't enable tracing\n");
683 goto out_close_fd;
684 }
685 }
686
687 while (!done) {
688 if (poll(&pollfd, 1, -1) < 0)
689 break;
690
691 if (pollfd.revents & POLLIN) {
692 int n = read(trace_fd, buf, sizeof(buf));
693 if (n < 0)
694 break;
695 if (fwrite(buf, n, 1, stdout) != 1)
696 break;
697 /* flush output since stdout is in full buffering mode due to pager */
698 fflush(stdout);
699 }
700 }
701
702 write_tracing_file("tracing_on", "0");
703
704 if (workload_exec_errno) {
705 const char *emsg = str_error_r(workload_exec_errno, buf, sizeof(buf));
706 /* flush stdout first so below error msg appears at the end. */
707 fflush(stdout);
708 pr_err("workload failed: %s\n", emsg);
709 goto out_close_fd;
710 }
711
712 /* read remaining buffer contents */
713 while (true) {
714 int n = read(trace_fd, buf, sizeof(buf));
715 if (n <= 0)
716 break;
717 if (fwrite(buf, n, 1, stdout) != 1)
718 break;
719 }
720
721out_close_fd:
722 close(trace_fd);
723out_reset:
724 reset_tracing_files(ftrace);
725out:
726 return (done && !workload_exec_errno) ? 0 : -1;
727}
728
729static void make_histogram(int buckets[], char *buf, size_t len, char *linebuf,
730 bool use_nsec)
731{
732 char *p, *q;
733 char *unit;
734 double num;
735 int i;
736
737 /* ensure NUL termination */
738 buf[len] = '\0';
739
740 /* handle data line by line */
741 for (p = buf; (q = strchr(p, '\n')) != NULL; p = q + 1) {
742 *q = '\0';
743 /* move it to the line buffer */
744 strcat(linebuf, p);
745
746 /*
747 * parse trace output to get function duration like in
748 *
749 * # tracer: function_graph
750 * #
751 * # CPU DURATION FUNCTION CALLS
752 * # | | | | | | |
753 * 1) + 10.291 us | do_filp_open();
754 * 1) 4.889 us | do_filp_open();
755 * 1) 6.086 us | do_filp_open();
756 *
757 */
758 if (linebuf[0] == '#')
759 goto next;
760
761 /* ignore CPU */
762 p = strchr(linebuf, ')');
763 if (p == NULL)
764 p = linebuf;
765
766 while (*p && !isdigit(*p) && (*p != '|'))
767 p++;
768
769 /* no duration */
770 if (*p == '\0' || *p == '|')
771 goto next;
772
773 num = strtod(p, &unit);
774 if (!unit || strncmp(unit, " us", 3))
775 goto next;
776
777 if (use_nsec)
778 num *= 1000;
779
780 i = log2(num);
781 if (i < 0)
782 i = 0;
783 if (i >= NUM_BUCKET)
784 i = NUM_BUCKET - 1;
785
786 buckets[i]++;
787
788next:
789 /* empty the line buffer for the next output */
790 linebuf[0] = '\0';
791 }
792
793 /* preserve any remaining output (before newline) */
794 strcat(linebuf, p);
795}
796
797static void display_histogram(int buckets[], bool use_nsec)
798{
799 int i;
800 int total = 0;
801 int bar_total = 46; /* to fit in 80 column */
802 char bar[] = "###############################################";
803 int bar_len;
804
805 for (i = 0; i < NUM_BUCKET; i++)
806 total += buckets[i];
807
808 if (total == 0) {
809 printf("No data found\n");
810 return;
811 }
812
813 printf("# %14s | %10s | %-*s |\n",
814 " DURATION ", "COUNT", bar_total, "GRAPH");
815
816 bar_len = buckets[0] * bar_total / total;
817 printf(" %4d - %-4d %s | %10d | %.*s%*s |\n",
818 0, 1, use_nsec ? "ns" : "us", buckets[0], bar_len, bar, bar_total - bar_len, "");
819
820 for (i = 1; i < NUM_BUCKET - 1; i++) {
821 int start = (1 << (i - 1));
822 int stop = 1 << i;
823 const char *unit = use_nsec ? "ns" : "us";
824
825 if (start >= 1024) {
826 start >>= 10;
827 stop >>= 10;
828 unit = use_nsec ? "us" : "ms";
829 }
830 bar_len = buckets[i] * bar_total / total;
831 printf(" %4d - %-4d %s | %10d | %.*s%*s |\n",
832 start, stop, unit, buckets[i], bar_len, bar,
833 bar_total - bar_len, "");
834 }
835
836 bar_len = buckets[NUM_BUCKET - 1] * bar_total / total;
837 printf(" %4d - %-4s %s | %10d | %.*s%*s |\n",
838 1, "...", use_nsec ? "ms" : " s", buckets[NUM_BUCKET - 1],
839 bar_len, bar, bar_total - bar_len, "");
840
841}
842
843static int prepare_func_latency(struct perf_ftrace *ftrace)
844{
845 char *trace_file;
846 int fd;
847
848 if (ftrace->target.use_bpf)
849 return perf_ftrace__latency_prepare_bpf(ftrace);
850
851 if (reset_tracing_files(ftrace) < 0) {
852 pr_err("failed to reset ftrace\n");
853 return -1;
854 }
855
856 /* reset ftrace buffer */
857 if (write_tracing_file("trace", "0") < 0)
858 return -1;
859
860 if (set_tracing_options(ftrace) < 0)
861 return -1;
862
863 /* force to use the function_graph tracer to track duration */
864 if (write_tracing_file("current_tracer", "function_graph") < 0) {
865 pr_err("failed to set current_tracer to function_graph\n");
866 return -1;
867 }
868
869 trace_file = get_tracing_file("trace_pipe");
870 if (!trace_file) {
871 pr_err("failed to open trace_pipe\n");
872 return -1;
873 }
874
875 fd = open(trace_file, O_RDONLY);
876 if (fd < 0)
877 pr_err("failed to open trace_pipe\n");
878
879 put_tracing_file(trace_file);
880 return fd;
881}
882
883static int start_func_latency(struct perf_ftrace *ftrace)
884{
885 if (ftrace->target.use_bpf)
886 return perf_ftrace__latency_start_bpf(ftrace);
887
888 if (write_tracing_file("tracing_on", "1") < 0) {
889 pr_err("can't enable tracing\n");
890 return -1;
891 }
892
893 return 0;
894}
895
896static int stop_func_latency(struct perf_ftrace *ftrace)
897{
898 if (ftrace->target.use_bpf)
899 return perf_ftrace__latency_stop_bpf(ftrace);
900
901 write_tracing_file("tracing_on", "0");
902 return 0;
903}
904
905static int read_func_latency(struct perf_ftrace *ftrace, int buckets[])
906{
907 if (ftrace->target.use_bpf)
908 return perf_ftrace__latency_read_bpf(ftrace, buckets);
909
910 return 0;
911}
912
913static int cleanup_func_latency(struct perf_ftrace *ftrace)
914{
915 if (ftrace->target.use_bpf)
916 return perf_ftrace__latency_cleanup_bpf(ftrace);
917
918 reset_tracing_files(ftrace);
919 return 0;
920}
921
922static int __cmd_latency(struct perf_ftrace *ftrace)
923{
924 int trace_fd;
925 char buf[4096];
926 char line[256];
927 struct pollfd pollfd = {
928 .events = POLLIN,
929 };
930 int buckets[NUM_BUCKET] = { };
931
932 trace_fd = prepare_func_latency(ftrace);
933 if (trace_fd < 0)
934 goto out;
935
936 fcntl(trace_fd, F_SETFL, O_NONBLOCK);
937 pollfd.fd = trace_fd;
938
939 if (start_func_latency(ftrace) < 0)
940 goto out;
941
942 evlist__start_workload(ftrace->evlist);
943
944 line[0] = '\0';
945 while (!done) {
946 if (poll(&pollfd, 1, -1) < 0)
947 break;
948
949 if (pollfd.revents & POLLIN) {
950 int n = read(trace_fd, buf, sizeof(buf) - 1);
951 if (n < 0)
952 break;
953
954 make_histogram(buckets, buf, n, line, ftrace->use_nsec);
955 }
956 }
957
958 stop_func_latency(ftrace);
959
960 if (workload_exec_errno) {
961 const char *emsg = str_error_r(workload_exec_errno, buf, sizeof(buf));
962 pr_err("workload failed: %s\n", emsg);
963 goto out;
964 }
965
966 /* read remaining buffer contents */
967 while (!ftrace->target.use_bpf) {
968 int n = read(trace_fd, buf, sizeof(buf) - 1);
969 if (n <= 0)
970 break;
971 make_histogram(buckets, buf, n, line, ftrace->use_nsec);
972 }
973
974 read_func_latency(ftrace, buckets);
975
976 display_histogram(buckets, ftrace->use_nsec);
977
978out:
979 close(trace_fd);
980 cleanup_func_latency(ftrace);
981
982 return (done && !workload_exec_errno) ? 0 : -1;
983}
984
985static size_t profile_hash(long func, void *ctx __maybe_unused)
986{
987 return str_hash((char *)func);
988}
989
990static bool profile_equal(long func1, long func2, void *ctx __maybe_unused)
991{
992 return !strcmp((char *)func1, (char *)func2);
993}
994
995static int prepare_func_profile(struct perf_ftrace *ftrace)
996{
997 ftrace->tracer = "function_graph";
998 ftrace->graph_tail = 1;
999
1000 ftrace->profile_hash = hashmap__new(profile_hash, profile_equal, NULL);
1001 if (ftrace->profile_hash == NULL)
1002 return -ENOMEM;
1003
1004 return 0;
1005}
1006
1007/* This is saved in a hashmap keyed by the function name */
1008struct ftrace_profile_data {
1009 struct stats st;
1010};
1011
1012static int add_func_duration(struct perf_ftrace *ftrace, char *func, double time_ns)
1013{
1014 struct ftrace_profile_data *prof = NULL;
1015
1016 if (!hashmap__find(ftrace->profile_hash, func, &prof)) {
1017 char *key = strdup(func);
1018
1019 if (key == NULL)
1020 return -ENOMEM;
1021
1022 prof = zalloc(sizeof(*prof));
1023 if (prof == NULL) {
1024 free(key);
1025 return -ENOMEM;
1026 }
1027
1028 init_stats(&prof->st);
1029 hashmap__add(ftrace->profile_hash, key, prof);
1030 }
1031
1032 update_stats(&prof->st, time_ns);
1033 return 0;
1034}
1035
1036/*
1037 * The ftrace function_graph text output normally looks like below:
1038 *
1039 * CPU DURATION FUNCTION
1040 *
1041 * 0) | syscall_trace_enter.isra.0() {
1042 * 0) | __audit_syscall_entry() {
1043 * 0) | auditd_test_task() {
1044 * 0) 0.271 us | __rcu_read_lock();
1045 * 0) 0.275 us | __rcu_read_unlock();
1046 * 0) 1.254 us | } /\* auditd_test_task *\/
1047 * 0) 0.279 us | ktime_get_coarse_real_ts64();
1048 * 0) 2.227 us | } /\* __audit_syscall_entry *\/
1049 * 0) 2.713 us | } /\* syscall_trace_enter.isra.0 *\/
1050 *
1051 * Parse the line and get the duration and function name.
1052 */
1053static int parse_func_duration(struct perf_ftrace *ftrace, char *line, size_t len)
1054{
1055 char *p;
1056 char *func;
1057 double duration;
1058
1059 /* skip CPU */
1060 p = strchr(line, ')');
1061 if (p == NULL)
1062 return 0;
1063
1064 /* get duration */
1065 p = skip_spaces(p + 1);
1066
1067 /* no duration? */
1068 if (p == NULL || *p == '|')
1069 return 0;
1070
1071 /* skip markers like '*' or '!' for longer than ms */
1072 if (!isdigit(*p))
1073 p++;
1074
1075 duration = strtod(p, &p);
1076
1077 if (strncmp(p, " us", 3)) {
1078 pr_debug("non-usec time found.. ignoring\n");
1079 return 0;
1080 }
1081
1082 /*
1083 * profile stat keeps the max and min values as integer,
1084 * convert to nsec time so that we can have accurate max.
1085 */
1086 duration *= 1000;
1087
1088 /* skip to the pipe */
1089 while (p < line + len && *p != '|')
1090 p++;
1091
1092 if (*p++ != '|')
1093 return -EINVAL;
1094
1095 /* get function name */
1096 func = skip_spaces(p);
1097
1098 /* skip the closing bracket and the start of comment */
1099 if (*func == '}')
1100 func += 5;
1101
1102 /* remove semi-colon or end of comment at the end */
1103 p = line + len - 1;
1104 while (!isalnum(*p) && *p != ']') {
1105 *p = '\0';
1106 --p;
1107 }
1108
1109 return add_func_duration(ftrace, func, duration);
1110}
1111
1112enum perf_ftrace_profile_sort_key {
1113 PFP_SORT_TOTAL = 0,
1114 PFP_SORT_AVG,
1115 PFP_SORT_MAX,
1116 PFP_SORT_COUNT,
1117 PFP_SORT_NAME,
1118};
1119
1120static enum perf_ftrace_profile_sort_key profile_sort = PFP_SORT_TOTAL;
1121
1122static int cmp_profile_data(const void *a, const void *b)
1123{
1124 const struct hashmap_entry *e1 = *(const struct hashmap_entry **)a;
1125 const struct hashmap_entry *e2 = *(const struct hashmap_entry **)b;
1126 struct ftrace_profile_data *p1 = e1->pvalue;
1127 struct ftrace_profile_data *p2 = e2->pvalue;
1128 double v1, v2;
1129
1130 switch (profile_sort) {
1131 case PFP_SORT_NAME:
1132 return strcmp(e1->pkey, e2->pkey);
1133 case PFP_SORT_AVG:
1134 v1 = p1->st.mean;
1135 v2 = p2->st.mean;
1136 break;
1137 case PFP_SORT_MAX:
1138 v1 = p1->st.max;
1139 v2 = p2->st.max;
1140 break;
1141 case PFP_SORT_COUNT:
1142 v1 = p1->st.n;
1143 v2 = p2->st.n;
1144 break;
1145 case PFP_SORT_TOTAL:
1146 default:
1147 v1 = p1->st.n * p1->st.mean;
1148 v2 = p2->st.n * p2->st.mean;
1149 break;
1150 }
1151
1152 if (v1 > v2)
1153 return -1;
1154 if (v1 < v2)
1155 return 1;
1156 return 0;
1157}
1158
1159static void print_profile_result(struct perf_ftrace *ftrace)
1160{
1161 struct hashmap_entry *entry, **profile;
1162 size_t i, nr, bkt;
1163
1164 nr = hashmap__size(ftrace->profile_hash);
1165 if (nr == 0)
1166 return;
1167
1168 profile = calloc(nr, sizeof(*profile));
1169 if (profile == NULL) {
1170 pr_err("failed to allocate memory for the result\n");
1171 return;
1172 }
1173
1174 i = 0;
1175 hashmap__for_each_entry(ftrace->profile_hash, entry, bkt)
1176 profile[i++] = entry;
1177
1178 assert(i == nr);
1179
1180 //cmp_profile_data(profile[0], profile[1]);
1181 qsort(profile, nr, sizeof(*profile), cmp_profile_data);
1182
1183 printf("# %10s %10s %10s %10s %s\n",
1184 "Total (us)", "Avg (us)", "Max (us)", "Count", "Function");
1185
1186 for (i = 0; i < nr; i++) {
1187 const char *name = profile[i]->pkey;
1188 struct ftrace_profile_data *p = profile[i]->pvalue;
1189
1190 printf("%12.3f %10.3f %6"PRIu64".%03"PRIu64" %10.0f %s\n",
1191 p->st.n * p->st.mean / 1000, p->st.mean / 1000,
1192 p->st.max / 1000, p->st.max % 1000, p->st.n, name);
1193 }
1194
1195 free(profile);
1196
1197 hashmap__for_each_entry(ftrace->profile_hash, entry, bkt) {
1198 free((char *)entry->pkey);
1199 free(entry->pvalue);
1200 }
1201
1202 hashmap__free(ftrace->profile_hash);
1203 ftrace->profile_hash = NULL;
1204}
1205
1206static int __cmd_profile(struct perf_ftrace *ftrace)
1207{
1208 char *trace_file;
1209 int trace_fd;
1210 char buf[4096];
1211 struct io io;
1212 char *line = NULL;
1213 size_t line_len = 0;
1214
1215 if (prepare_func_profile(ftrace) < 0) {
1216 pr_err("failed to prepare func profiler\n");
1217 goto out;
1218 }
1219
1220 if (reset_tracing_files(ftrace) < 0) {
1221 pr_err("failed to reset ftrace\n");
1222 goto out;
1223 }
1224
1225 /* reset ftrace buffer */
1226 if (write_tracing_file("trace", "0") < 0)
1227 goto out;
1228
1229 if (set_tracing_options(ftrace) < 0)
1230 return -1;
1231
1232 if (write_tracing_file("current_tracer", ftrace->tracer) < 0) {
1233 pr_err("failed to set current_tracer to %s\n", ftrace->tracer);
1234 goto out_reset;
1235 }
1236
1237 setup_pager();
1238
1239 trace_file = get_tracing_file("trace_pipe");
1240 if (!trace_file) {
1241 pr_err("failed to open trace_pipe\n");
1242 goto out_reset;
1243 }
1244
1245 trace_fd = open(trace_file, O_RDONLY);
1246
1247 put_tracing_file(trace_file);
1248
1249 if (trace_fd < 0) {
1250 pr_err("failed to open trace_pipe\n");
1251 goto out_reset;
1252 }
1253
1254 fcntl(trace_fd, F_SETFL, O_NONBLOCK);
1255
1256 if (write_tracing_file("tracing_on", "1") < 0) {
1257 pr_err("can't enable tracing\n");
1258 goto out_close_fd;
1259 }
1260
1261 evlist__start_workload(ftrace->evlist);
1262
1263 io__init(&io, trace_fd, buf, sizeof(buf));
1264 io.timeout_ms = -1;
1265
1266 while (!done && !io.eof) {
1267 if (io__getline(&io, &line, &line_len) < 0)
1268 break;
1269
1270 if (parse_func_duration(ftrace, line, line_len) < 0)
1271 break;
1272 }
1273
1274 write_tracing_file("tracing_on", "0");
1275
1276 if (workload_exec_errno) {
1277 const char *emsg = str_error_r(workload_exec_errno, buf, sizeof(buf));
1278 /* flush stdout first so below error msg appears at the end. */
1279 fflush(stdout);
1280 pr_err("workload failed: %s\n", emsg);
1281 goto out_free_line;
1282 }
1283
1284 /* read remaining buffer contents */
1285 io.timeout_ms = 0;
1286 while (!io.eof) {
1287 if (io__getline(&io, &line, &line_len) < 0)
1288 break;
1289
1290 if (parse_func_duration(ftrace, line, line_len) < 0)
1291 break;
1292 }
1293
1294 print_profile_result(ftrace);
1295
1296out_free_line:
1297 free(line);
1298out_close_fd:
1299 close(trace_fd);
1300out_reset:
1301 reset_tracing_files(ftrace);
1302out:
1303 return (done && !workload_exec_errno) ? 0 : -1;
1304}
1305
1306static int perf_ftrace_config(const char *var, const char *value, void *cb)
1307{
1308 struct perf_ftrace *ftrace = cb;
1309
1310 if (!strstarts(var, "ftrace."))
1311 return 0;
1312
1313 if (strcmp(var, "ftrace.tracer"))
1314 return -1;
1315
1316 if (!strcmp(value, "function_graph") ||
1317 !strcmp(value, "function")) {
1318 ftrace->tracer = value;
1319 return 0;
1320 }
1321
1322 pr_err("Please select \"function_graph\" (default) or \"function\"\n");
1323 return -1;
1324}
1325
1326static void list_function_cb(char *str, void *arg)
1327{
1328 struct strfilter *filter = (struct strfilter *)arg;
1329
1330 if (strfilter__compare(filter, str))
1331 printf("%s", str);
1332}
1333
1334static int opt_list_avail_functions(const struct option *opt __maybe_unused,
1335 const char *str, int unset)
1336{
1337 struct strfilter *filter;
1338 const char *err = NULL;
1339 int ret;
1340
1341 if (unset || !str)
1342 return -1;
1343
1344 filter = strfilter__new(str, &err);
1345 if (!filter)
1346 return err ? -EINVAL : -ENOMEM;
1347
1348 ret = strfilter__or(filter, str, &err);
1349 if (ret == -EINVAL) {
1350 pr_err("Filter parse error at %td.\n", err - str + 1);
1351 pr_err("Source: \"%s\"\n", str);
1352 pr_err(" %*c\n", (int)(err - str + 1), '^');
1353 strfilter__delete(filter);
1354 return ret;
1355 }
1356
1357 ret = read_tracing_file_by_line("available_filter_functions",
1358 list_function_cb, filter);
1359 strfilter__delete(filter);
1360 if (ret < 0)
1361 return ret;
1362
1363 exit(0);
1364}
1365
1366static int parse_filter_func(const struct option *opt, const char *str,
1367 int unset __maybe_unused)
1368{
1369 struct list_head *head = opt->value;
1370 struct filter_entry *entry;
1371
1372 entry = malloc(sizeof(*entry) + strlen(str) + 1);
1373 if (entry == NULL)
1374 return -ENOMEM;
1375
1376 strcpy(entry->name, str);
1377 list_add_tail(&entry->list, head);
1378
1379 return 0;
1380}
1381
1382static void delete_filter_func(struct list_head *head)
1383{
1384 struct filter_entry *pos, *tmp;
1385
1386 list_for_each_entry_safe(pos, tmp, head, list) {
1387 list_del_init(&pos->list);
1388 free(pos);
1389 }
1390}
1391
1392static int parse_buffer_size(const struct option *opt,
1393 const char *str, int unset)
1394{
1395 unsigned long *s = (unsigned long *)opt->value;
1396 static struct parse_tag tags_size[] = {
1397 { .tag = 'B', .mult = 1 },
1398 { .tag = 'K', .mult = 1 << 10 },
1399 { .tag = 'M', .mult = 1 << 20 },
1400 { .tag = 'G', .mult = 1 << 30 },
1401 { .tag = 0 },
1402 };
1403 unsigned long val;
1404
1405 if (unset) {
1406 *s = 0;
1407 return 0;
1408 }
1409
1410 val = parse_tag_value(str, tags_size);
1411 if (val != (unsigned long) -1) {
1412 if (val < 1024) {
1413 pr_err("buffer size too small, must larger than 1KB.");
1414 return -1;
1415 }
1416 *s = val;
1417 return 0;
1418 }
1419
1420 return -1;
1421}
1422
1423static int parse_func_tracer_opts(const struct option *opt,
1424 const char *str, int unset)
1425{
1426 int ret;
1427 struct perf_ftrace *ftrace = (struct perf_ftrace *) opt->value;
1428 struct sublevel_option func_tracer_opts[] = {
1429 { .name = "call-graph", .value_ptr = &ftrace->func_stack_trace },
1430 { .name = "irq-info", .value_ptr = &ftrace->func_irq_info },
1431 { .name = NULL, }
1432 };
1433
1434 if (unset)
1435 return 0;
1436
1437 ret = perf_parse_sublevel_options(str, func_tracer_opts);
1438 if (ret)
1439 return ret;
1440
1441 return 0;
1442}
1443
1444static int parse_graph_tracer_opts(const struct option *opt,
1445 const char *str, int unset)
1446{
1447 int ret;
1448 struct perf_ftrace *ftrace = (struct perf_ftrace *) opt->value;
1449 struct sublevel_option graph_tracer_opts[] = {
1450 { .name = "nosleep-time", .value_ptr = &ftrace->graph_nosleep_time },
1451 { .name = "noirqs", .value_ptr = &ftrace->graph_noirqs },
1452 { .name = "verbose", .value_ptr = &ftrace->graph_verbose },
1453 { .name = "thresh", .value_ptr = &ftrace->graph_thresh },
1454 { .name = "depth", .value_ptr = &ftrace->graph_depth },
1455 { .name = "tail", .value_ptr = &ftrace->graph_tail },
1456 { .name = NULL, }
1457 };
1458
1459 if (unset)
1460 return 0;
1461
1462 ret = perf_parse_sublevel_options(str, graph_tracer_opts);
1463 if (ret)
1464 return ret;
1465
1466 return 0;
1467}
1468
1469static int parse_sort_key(const struct option *opt, const char *str, int unset)
1470{
1471 enum perf_ftrace_profile_sort_key *key = (void *)opt->value;
1472
1473 if (unset)
1474 return 0;
1475
1476 if (!strcmp(str, "total"))
1477 *key = PFP_SORT_TOTAL;
1478 else if (!strcmp(str, "avg"))
1479 *key = PFP_SORT_AVG;
1480 else if (!strcmp(str, "max"))
1481 *key = PFP_SORT_MAX;
1482 else if (!strcmp(str, "count"))
1483 *key = PFP_SORT_COUNT;
1484 else if (!strcmp(str, "name"))
1485 *key = PFP_SORT_NAME;
1486 else {
1487 pr_err("Unknown sort key: %s\n", str);
1488 return -1;
1489 }
1490 return 0;
1491}
1492
1493enum perf_ftrace_subcommand {
1494 PERF_FTRACE_NONE,
1495 PERF_FTRACE_TRACE,
1496 PERF_FTRACE_LATENCY,
1497 PERF_FTRACE_PROFILE,
1498};
1499
1500int cmd_ftrace(int argc, const char **argv)
1501{
1502 int ret;
1503 int (*cmd_func)(struct perf_ftrace *) = NULL;
1504 struct perf_ftrace ftrace = {
1505 .tracer = DEFAULT_TRACER,
1506 .target = { .uid = UINT_MAX, },
1507 };
1508 const struct option common_options[] = {
1509 OPT_STRING('p', "pid", &ftrace.target.pid, "pid",
1510 "Trace on existing process id"),
1511 /* TODO: Add short option -t after -t/--tracer can be removed. */
1512 OPT_STRING(0, "tid", &ftrace.target.tid, "tid",
1513 "Trace on existing thread id (exclusive to --pid)"),
1514 OPT_INCR('v', "verbose", &verbose,
1515 "Be more verbose"),
1516 OPT_BOOLEAN('a', "all-cpus", &ftrace.target.system_wide,
1517 "System-wide collection from all CPUs"),
1518 OPT_STRING('C', "cpu", &ftrace.target.cpu_list, "cpu",
1519 "List of cpus to monitor"),
1520 OPT_END()
1521 };
1522 const struct option ftrace_options[] = {
1523 OPT_STRING('t', "tracer", &ftrace.tracer, "tracer",
1524 "Tracer to use: function_graph(default) or function"),
1525 OPT_CALLBACK_DEFAULT('F', "funcs", NULL, "[FILTER]",
1526 "Show available functions to filter",
1527 opt_list_avail_functions, "*"),
1528 OPT_CALLBACK('T', "trace-funcs", &ftrace.filters, "func",
1529 "Trace given functions using function tracer",
1530 parse_filter_func),
1531 OPT_CALLBACK('N', "notrace-funcs", &ftrace.notrace, "func",
1532 "Do not trace given functions", parse_filter_func),
1533 OPT_CALLBACK(0, "func-opts", &ftrace, "options",
1534 "Function tracer options, available options: call-graph,irq-info",
1535 parse_func_tracer_opts),
1536 OPT_CALLBACK('G', "graph-funcs", &ftrace.graph_funcs, "func",
1537 "Trace given functions using function_graph tracer",
1538 parse_filter_func),
1539 OPT_CALLBACK('g', "nograph-funcs", &ftrace.nograph_funcs, "func",
1540 "Set nograph filter on given functions", parse_filter_func),
1541 OPT_CALLBACK(0, "graph-opts", &ftrace, "options",
1542 "Graph tracer options, available options: nosleep-time,noirqs,verbose,thresh=<n>,depth=<n>",
1543 parse_graph_tracer_opts),
1544 OPT_CALLBACK('m', "buffer-size", &ftrace.percpu_buffer_size, "size",
1545 "Size of per cpu buffer, needs to use a B, K, M or G suffix.", parse_buffer_size),
1546 OPT_BOOLEAN(0, "inherit", &ftrace.inherit,
1547 "Trace children processes"),
1548 OPT_INTEGER('D', "delay", &ftrace.target.initial_delay,
1549 "Number of milliseconds to wait before starting tracing after program start"),
1550 OPT_PARENT(common_options),
1551 };
1552 const struct option latency_options[] = {
1553 OPT_CALLBACK('T', "trace-funcs", &ftrace.filters, "func",
1554 "Show latency of given function", parse_filter_func),
1555#ifdef HAVE_BPF_SKEL
1556 OPT_BOOLEAN('b', "use-bpf", &ftrace.target.use_bpf,
1557 "Use BPF to measure function latency"),
1558#endif
1559 OPT_BOOLEAN('n', "use-nsec", &ftrace.use_nsec,
1560 "Use nano-second histogram"),
1561 OPT_PARENT(common_options),
1562 };
1563 const struct option profile_options[] = {
1564 OPT_CALLBACK('T', "trace-funcs", &ftrace.filters, "func",
1565 "Trace given functions using function tracer",
1566 parse_filter_func),
1567 OPT_CALLBACK('N', "notrace-funcs", &ftrace.notrace, "func",
1568 "Do not trace given functions", parse_filter_func),
1569 OPT_CALLBACK('G', "graph-funcs", &ftrace.graph_funcs, "func",
1570 "Trace given functions using function_graph tracer",
1571 parse_filter_func),
1572 OPT_CALLBACK('g', "nograph-funcs", &ftrace.nograph_funcs, "func",
1573 "Set nograph filter on given functions", parse_filter_func),
1574 OPT_CALLBACK('m', "buffer-size", &ftrace.percpu_buffer_size, "size",
1575 "Size of per cpu buffer, needs to use a B, K, M or G suffix.", parse_buffer_size),
1576 OPT_CALLBACK('s', "sort", &profile_sort, "key",
1577 "Sort result by key: total (default), avg, max, count, name.",
1578 parse_sort_key),
1579 OPT_PARENT(common_options),
1580 };
1581 const struct option *options = ftrace_options;
1582
1583 const char * const ftrace_usage[] = {
1584 "perf ftrace [<options>] [<command>]",
1585 "perf ftrace [<options>] -- [<command>] [<options>]",
1586 "perf ftrace {trace|latency|profile} [<options>] [<command>]",
1587 "perf ftrace {trace|latency|profile} [<options>] -- [<command>] [<options>]",
1588 NULL
1589 };
1590 enum perf_ftrace_subcommand subcmd = PERF_FTRACE_NONE;
1591
1592 INIT_LIST_HEAD(&ftrace.filters);
1593 INIT_LIST_HEAD(&ftrace.notrace);
1594 INIT_LIST_HEAD(&ftrace.graph_funcs);
1595 INIT_LIST_HEAD(&ftrace.nograph_funcs);
1596
1597 signal(SIGINT, sig_handler);
1598 signal(SIGUSR1, sig_handler);
1599 signal(SIGCHLD, sig_handler);
1600 signal(SIGPIPE, sig_handler);
1601
1602 if (!check_ftrace_capable())
1603 return -1;
1604
1605 if (!is_ftrace_supported()) {
1606 pr_err("ftrace is not supported on this system\n");
1607 return -ENOTSUP;
1608 }
1609
1610 ret = perf_config(perf_ftrace_config, &ftrace);
1611 if (ret < 0)
1612 return -1;
1613
1614 if (argc > 1) {
1615 if (!strcmp(argv[1], "trace")) {
1616 subcmd = PERF_FTRACE_TRACE;
1617 } else if (!strcmp(argv[1], "latency")) {
1618 subcmd = PERF_FTRACE_LATENCY;
1619 options = latency_options;
1620 } else if (!strcmp(argv[1], "profile")) {
1621 subcmd = PERF_FTRACE_PROFILE;
1622 options = profile_options;
1623 }
1624
1625 if (subcmd != PERF_FTRACE_NONE) {
1626 argc--;
1627 argv++;
1628 }
1629 }
1630 /* for backward compatibility */
1631 if (subcmd == PERF_FTRACE_NONE)
1632 subcmd = PERF_FTRACE_TRACE;
1633
1634 argc = parse_options(argc, argv, options, ftrace_usage,
1635 PARSE_OPT_STOP_AT_NON_OPTION);
1636 if (argc < 0) {
1637 ret = -EINVAL;
1638 goto out_delete_filters;
1639 }
1640
1641 /* Make system wide (-a) the default target. */
1642 if (!argc && target__none(&ftrace.target))
1643 ftrace.target.system_wide = true;
1644
1645 switch (subcmd) {
1646 case PERF_FTRACE_TRACE:
1647 cmd_func = __cmd_ftrace;
1648 break;
1649 case PERF_FTRACE_LATENCY:
1650 if (list_empty(&ftrace.filters)) {
1651 pr_err("Should provide a function to measure\n");
1652 parse_options_usage(ftrace_usage, options, "T", 1);
1653 ret = -EINVAL;
1654 goto out_delete_filters;
1655 }
1656 cmd_func = __cmd_latency;
1657 break;
1658 case PERF_FTRACE_PROFILE:
1659 cmd_func = __cmd_profile;
1660 break;
1661 case PERF_FTRACE_NONE:
1662 default:
1663 pr_err("Invalid subcommand\n");
1664 ret = -EINVAL;
1665 goto out_delete_filters;
1666 }
1667
1668 ret = target__validate(&ftrace.target);
1669 if (ret) {
1670 char errbuf[512];
1671
1672 target__strerror(&ftrace.target, ret, errbuf, 512);
1673 pr_err("%s\n", errbuf);
1674 goto out_delete_filters;
1675 }
1676
1677 ftrace.evlist = evlist__new();
1678 if (ftrace.evlist == NULL) {
1679 ret = -ENOMEM;
1680 goto out_delete_filters;
1681 }
1682
1683 ret = evlist__create_maps(ftrace.evlist, &ftrace.target);
1684 if (ret < 0)
1685 goto out_delete_evlist;
1686
1687 if (argc) {
1688 ret = evlist__prepare_workload(ftrace.evlist, &ftrace.target,
1689 argv, false,
1690 ftrace__workload_exec_failed_signal);
1691 if (ret < 0)
1692 goto out_delete_evlist;
1693 }
1694
1695 ret = cmd_func(&ftrace);
1696
1697out_delete_evlist:
1698 evlist__delete(ftrace.evlist);
1699
1700out_delete_filters:
1701 delete_filter_func(&ftrace.filters);
1702 delete_filter_func(&ftrace.notrace);
1703 delete_filter_func(&ftrace.graph_funcs);
1704 delete_filter_func(&ftrace.nograph_funcs);
1705
1706 return ret;
1707}