aboutsummaryrefslogtreecommitdiffstats
path: root/Documentation
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation')
-rw-r--r--Documentation/perf_counter/Makefile12
-rw-r--r--Documentation/perf_counter/kerneltop.c956
-rw-r--r--Documentation/perf_counter/perfstat.c521
3 files changed, 1489 insertions, 0 deletions
diff --git a/Documentation/perf_counter/Makefile b/Documentation/perf_counter/Makefile
new file mode 100644
index 000000000000..b45749753fcb
--- /dev/null
+++ b/Documentation/perf_counter/Makefile
@@ -0,0 +1,12 @@
1BINS = kerneltop perfstat
2
3all: $(BINS)
4
5kerneltop: kerneltop.c perfcounters.h
6 cc -O6 -Wall -lrt `pkg-config --cflags --libs glib-2.0` -o $@ $<
7
8perfstat: kerneltop
9 ln -sf kerneltop perfstat
10
11clean:
12 rm $(BINS)
diff --git a/Documentation/perf_counter/kerneltop.c b/Documentation/perf_counter/kerneltop.c
new file mode 100644
index 000000000000..cf0e30bab5d5
--- /dev/null
+++ b/Documentation/perf_counter/kerneltop.c
@@ -0,0 +1,956 @@
1/*
2 * kerneltop.c: show top kernel functions - performance counters showcase
3
4 Build with:
5
6 cc -O6 -Wall `pkg-config --cflags --libs glib-2.0` -o kerneltop kerneltop.c
7
8 Sample output:
9
10------------------------------------------------------------------------------
11 KernelTop: 2669 irqs/sec [NMI, cache-misses/cache-refs], (all, cpu: 2)
12------------------------------------------------------------------------------
13
14 weight RIP kernel function
15 ______ ________________ _______________
16
17 35.20 - ffffffff804ce74b : skb_copy_and_csum_dev
18 33.00 - ffffffff804cb740 : sock_alloc_send_skb
19 31.26 - ffffffff804ce808 : skb_push
20 22.43 - ffffffff80510004 : tcp_established_options
21 19.00 - ffffffff8027d250 : find_get_page
22 15.76 - ffffffff804e4fc9 : eth_type_trans
23 15.20 - ffffffff804d8baa : dst_release
24 14.86 - ffffffff804cf5d8 : skb_release_head_state
25 14.00 - ffffffff802217d5 : read_hpet
26 12.00 - ffffffff804ffb7f : __ip_local_out
27 11.97 - ffffffff804fc0c8 : ip_local_deliver_finish
28 8.54 - ffffffff805001a3 : ip_queue_xmit
29
30 Started by Ingo Molnar <mingo@redhat.com>
31
32 Improvements and fixes by:
33
34 Arjan van de Ven <arjan@linux.intel.com>
35 Yanmin Zhang <yanmin.zhang@intel.com>
36 Mike Galbraith <efault@gmx.de>
37
38 Released under the GPL v2. (and only v2, not any later version)
39
40 */
41#define _GNU_SOURCE
42#include <sys/types.h>
43#include <sys/stat.h>
44#include <sys/time.h>
45#include <unistd.h>
46#include <stdint.h>
47#include <stdlib.h>
48#include <string.h>
49#include <getopt.h>
50#include <assert.h>
51#include <fcntl.h>
52#include <stdio.h>
53#include <errno.h>
54#include <ctype.h>
55#include <time.h>
56
57#include <glib.h>
58
59#include <sys/syscall.h>
60#include <sys/ioctl.h>
61#include <sys/poll.h>
62#include <sys/prctl.h>
63#include <sys/wait.h>
64#include <sys/uio.h>
65
66#include <linux/unistd.h>
67
68#ifdef __x86_64__
69# define __NR_perf_counter_open 295
70#endif
71
72#ifdef __i386__
73# define __NR_perf_counter_open 333
74#endif
75
76/*
77 * Pick up some kernel type conventions:
78 */
79#define __user
80#define asmlinkage
81
82typedef unsigned int __u32;
83typedef unsigned long long __u64;
84typedef long long __s64;
85
86/*
87 * User-space ABI bits:
88 */
89
90/*
91 * Generalized performance counter event types, used by the hw_event.type
92 * parameter of the sys_perf_counter_open() syscall:
93 */
94enum hw_event_types {
95 /*
96 * Common hardware events, generalized by the kernel:
97 */
98 PERF_COUNT_CPU_CYCLES = 0,
99 PERF_COUNT_INSTRUCTIONS = 1,
100 PERF_COUNT_CACHE_REFERENCES = 2,
101 PERF_COUNT_CACHE_MISSES = 3,
102 PERF_COUNT_BRANCH_INSTRUCTIONS = 4,
103 PERF_COUNT_BRANCH_MISSES = 5,
104 PERF_COUNT_BUS_CYCLES = 6,
105
106 PERF_HW_EVENTS_MAX = 7,
107
108 /*
109 * Special "software" counters provided by the kernel, even if
110 * the hardware does not support performance counters. These
111 * counters measure various physical and sw events of the
112 * kernel (and allow the profiling of them as well):
113 */
114 PERF_COUNT_CPU_CLOCK = -1,
115 PERF_COUNT_TASK_CLOCK = -2,
116 PERF_COUNT_PAGE_FAULTS = -3,
117 PERF_COUNT_CONTEXT_SWITCHES = -4,
118 PERF_COUNT_CPU_MIGRATIONS = -5,
119
120 PERF_SW_EVENTS_MIN = -6,
121};
122
123/*
124 * IRQ-notification data record type:
125 */
126enum perf_counter_record_type {
127 PERF_RECORD_SIMPLE = 0,
128 PERF_RECORD_IRQ = 1,
129 PERF_RECORD_GROUP = 2,
130};
131
132/*
133 * Hardware event to monitor via a performance monitoring counter:
134 */
135struct perf_counter_hw_event {
136 __s64 type;
137
138 __u64 irq_period;
139 __u64 record_type;
140 __u64 read_format;
141
142 __u64 disabled : 1, /* off by default */
143 nmi : 1, /* NMI sampling */
144 raw : 1, /* raw event type */
145 inherit : 1, /* children inherit it */
146 pinned : 1, /* must always be on PMU */
147 exclusive : 1, /* only group on PMU */
148 exclude_user : 1, /* don't count user */
149 exclude_kernel : 1, /* ditto kernel */
150 exclude_hv : 1, /* ditto hypervisor */
151 exclude_idle : 1, /* don't count when idle */
152
153 __reserved_1 : 54;
154
155 __u32 extra_config_len;
156 __u32 __reserved_4;
157
158 __u64 __reserved_2;
159 __u64 __reserved_3;
160};
161
162/*
163 * Ioctls that can be done on a perf counter fd:
164 */
165#define PERF_COUNTER_IOC_ENABLE _IO('$', 0)
166#define PERF_COUNTER_IOC_DISABLE _IO('$', 1)
167
168asmlinkage int sys_perf_counter_open(
169
170 struct perf_counter_hw_event *hw_event_uptr __user,
171 pid_t pid,
172 int cpu,
173 int group_fd,
174 unsigned long flags)
175{
176 int ret;
177
178 ret = syscall(
179 __NR_perf_counter_open, hw_event_uptr, pid, cpu, group_fd, flags);
180#if defined(__x86_64__) || defined(__i386__)
181 if (ret < 0 && ret > -4096) {
182 errno = -ret;
183 ret = -1;
184 }
185#endif
186 return ret;
187}
188
189const char *event_types [] = {
190 "CPU cycles",
191 "instructions",
192 "cache-refs",
193 "cache-misses",
194 "branches",
195 "branch-misses",
196 "bus cycles"
197};
198
199const unsigned int default_count[] = {
200 1000000,
201 1000000,
202 10000,
203 10000,
204 1000000,
205 10000,
206};
207
208/*
209 * prctl(PR_TASK_PERF_COUNTERS_DISABLE) will (cheaply) disable all
210 * counters in the current task.
211 */
212#define PR_TASK_PERF_COUNTERS_DISABLE 31
213#define PR_TASK_PERF_COUNTERS_ENABLE 32
214
215#define MAX_COUNTERS 8
216
217static int nr_counters = -1;
218
219static __u64 count_filter = 100;
220
221#define MAX_NR_CPUS 256
222
223static int event_count[MAX_COUNTERS];
224static unsigned long event_id[MAX_COUNTERS];
225static int event_raw[MAX_COUNTERS];
226
227static int tid = -1;
228static int profile_cpu = -1;
229static int nr_cpus = 0;
230static int nmi = 1;
231static int group = 0;
232
233static char *vmlinux;
234
235static char *sym_filter;
236static unsigned long filter_start;
237static unsigned long filter_end;
238
239static int delay_secs = 2;
240static int zero;
241static int dump_symtab;
242
243struct source_line {
244 uint64_t EIP;
245 unsigned long count;
246 char *line;
247};
248
249static GList *lines;
250
251static void display_help(void)
252{
253 printf(
254 "Usage: kerneltop [<options>]\n\n"
255 "KernelTop Options (up to %d event types can be specified at once):\n\n",
256 MAX_COUNTERS);
257 printf(
258 " -e EID --event_id=EID # event type ID [default: 0]\n"
259 " 0: CPU cycles\n"
260 " 1: instructions\n"
261 " 2: cache accesses\n"
262 " 3: cache misses\n"
263 " 4: branch instructions\n"
264 " 5: branch prediction misses\n"
265 " 6: bus cycles\n\n"
266 " rNNN: raw PMU events (eventsel+umask)\n\n"
267 " -c CNT --count=CNT # event period to sample\n\n"
268 " -C CPU --cpu=CPU # CPU (-1 for all) [default: -1]\n"
269 " -p PID --pid=PID # PID of sampled task (-1 for all) [default: -1]\n\n"
270 " -d delay --delay=<seconds> # sampling/display delay [default: 2]\n"
271 " -f CNT --filter=CNT # min-event-count filter [default: 100]\n\n"
272 " -s symbol --symbol=<symbol> # function to be showed annotated one-shot\n"
273 " -x path --vmlinux=<path> # the vmlinux binary, required for -s use:\n"
274 " -z --zero # zero counts after display\n"
275 " -D --dump_symtab # dump symbol table to stderr on startup\n"
276 "\n");
277
278 exit(0);
279}
280
281static void process_options(int argc, char *argv[])
282{
283 int error = 0, counter;
284
285 for (;;) {
286 int option_index = 0;
287 /** Options for getopt */
288 static struct option long_options[] = {
289 {"count", required_argument, NULL, 'c'},
290 {"cpu", required_argument, NULL, 'C'},
291 {"delay", required_argument, NULL, 'd'},
292 {"dump_symtab", no_argument, NULL, 'D'},
293 {"event_id", required_argument, NULL, 'e'},
294 {"filter", required_argument, NULL, 'f'},
295 {"group", required_argument, NULL, 'g'},
296 {"help", no_argument, NULL, 'h'},
297 {"nmi", required_argument, NULL, 'n'},
298 {"pid", required_argument, NULL, 'p'},
299 {"vmlinux", required_argument, NULL, 'x'},
300 {"symbol", required_argument, NULL, 's'},
301 {"zero", no_argument, NULL, 'z'},
302 {NULL, 0, NULL, 0 }
303 };
304 int c = getopt_long(argc, argv, "c:C:d:De:f:g:hn:p:s:x:z",
305 long_options, &option_index);
306 if (c == -1)
307 break;
308
309 switch (c) {
310 case 'c':
311 if (nr_counters == -1)
312 nr_counters = 0;
313 event_count[nr_counters] = atoi(optarg); break;
314 case 'C':
315 /* CPU and PID are mutually exclusive */
316 if (tid != -1) {
317 printf("WARNING: CPU switch overriding PID\n");
318 sleep(1);
319 tid = -1;
320 }
321 profile_cpu = atoi(optarg); break;
322 case 'd': delay_secs = atoi(optarg); break;
323 case 'D': dump_symtab = 1; break;
324
325 case 'e':
326 nr_counters++;
327 if (nr_counters == MAX_COUNTERS) {
328 error = 1;
329 break;
330 }
331 if (*optarg == 'r') {
332 event_raw[nr_counters] = 1;
333 ++optarg;
334 }
335 event_id[nr_counters] = strtol(optarg, NULL, 16);
336 break;
337
338 case 'f': count_filter = atoi(optarg); break;
339 case 'g': group = atoi(optarg); break;
340 case 'h': display_help(); break;
341 case 'n': nmi = atoi(optarg); break;
342 case 'p':
343 /* CPU and PID are mutually exclusive */
344 if (profile_cpu != -1) {
345 printf("WARNING: PID switch overriding CPU\n");
346 sleep(1);
347 profile_cpu = -1;
348 }
349 tid = atoi(optarg); break;
350 case 's': sym_filter = strdup(optarg); break;
351 case 'x': vmlinux = strdup(optarg); break;
352 case 'z': zero = 1; break;
353 default: error = 1; break;
354 }
355 }
356 if (error)
357 display_help();
358
359 nr_counters++;
360 if (nr_counters < 1)
361 nr_counters = 1;
362
363 for (counter = 0; counter < nr_counters; counter++) {
364 if (event_count[counter])
365 continue;
366
367 if (event_id[counter] < PERF_HW_EVENTS_MAX)
368 event_count[counter] = default_count[event_id[counter]];
369 else
370 event_count[counter] = 100000;
371 }
372}
373
374static uint64_t min_ip;
375static uint64_t max_ip = -1ll;
376
377struct sym_entry {
378 unsigned long long addr;
379 char *sym;
380 unsigned long count[MAX_COUNTERS];
381 int skip;
382 GList *source;
383};
384
385#define MAX_SYMS 100000
386
387static int sym_table_count;
388
389struct sym_entry *sym_filter_entry;
390
391static struct sym_entry sym_table[MAX_SYMS];
392
393static void show_details(struct sym_entry *sym);
394
395/*
396 * Ordering weight: count-1 * count-1 * ... / count-n
397 */
398static double sym_weight(const struct sym_entry *sym)
399{
400 double weight;
401 int counter;
402
403 weight = sym->count[0];
404
405 for (counter = 1; counter < nr_counters-1; counter++)
406 weight *= sym->count[counter];
407
408 weight /= (sym->count[counter] + 1);
409
410 return weight;
411}
412
413static int compare(const void *__sym1, const void *__sym2)
414{
415 const struct sym_entry *sym1 = __sym1, *sym2 = __sym2;
416
417 return sym_weight(sym1) < sym_weight(sym2);
418}
419
420static time_t last_refresh;
421static long events;
422static long userspace_events;
423static const char CONSOLE_CLEAR[] = "";
424
425static struct sym_entry tmp[MAX_SYMS];
426
427static void print_sym_table(void)
428{
429 int i, printed;
430 int counter;
431 float events_per_sec = events/delay_secs;
432 float kevents_per_sec = (events-userspace_events)/delay_secs;
433
434 memcpy(tmp, sym_table, sizeof(sym_table[0])*sym_table_count);
435 qsort(tmp, sym_table_count, sizeof(tmp[0]), compare);
436
437 write(1, CONSOLE_CLEAR, strlen(CONSOLE_CLEAR));
438
439 printf(
440"------------------------------------------------------------------------------\n");
441 printf( " KernelTop:%8.0f irqs/sec kernel:%3.1f%% [%s, ",
442 events_per_sec,
443 100.0 - (100.0*((events_per_sec-kevents_per_sec)/events_per_sec)),
444 nmi ? "NMI" : "IRQ");
445
446 if (nr_counters == 1)
447 printf("%d ", event_count[0]);
448
449 for (counter = 0; counter < nr_counters; counter++) {
450 if (counter)
451 printf("/");
452
453 if (event_id[counter] < PERF_HW_EVENTS_MAX)
454 printf( "%s", event_types[event_id[counter]]);
455 else
456 printf( "raw:%04lx", event_id[counter]);
457 }
458
459 printf( "], ");
460
461 if (tid != -1)
462 printf(" (tid: %d", tid);
463 else
464 printf(" (all");
465
466 if (profile_cpu != -1)
467 printf(", cpu: %d)\n", profile_cpu);
468 else {
469 if (tid != -1)
470 printf(")\n");
471 else
472 printf(", %d CPUs)\n", nr_cpus);
473 }
474
475 printf("------------------------------------------------------------------------------\n\n");
476
477 if (nr_counters == 1)
478 printf(" events");
479 else
480 printf(" weight events");
481
482 printf(" RIP kernel function\n"
483 " ______ ______ ________________ _______________\n\n"
484 );
485
486 printed = 0;
487 for (i = 0; i < sym_table_count; i++) {
488 int count;
489
490 if (nr_counters == 1) {
491 if (printed <= 18 &&
492 tmp[i].count[0] >= count_filter) {
493 printf("%19.2f - %016llx : %s\n",
494 sym_weight(tmp + i), tmp[i].addr, tmp[i].sym);
495 printed++;
496 }
497 } else {
498 if (printed <= 18 &&
499 tmp[i].count[0] >= count_filter) {
500 printf("%8.1f %10ld - %016llx : %s\n",
501 sym_weight(tmp + i),
502 tmp[i].count[0],
503 tmp[i].addr, tmp[i].sym);
504 printed++;
505 }
506 }
507 /*
508 * Add decay to the counts:
509 */
510 for (count = 0; count < nr_counters; count++)
511 sym_table[i].count[count] = zero ? 0 : sym_table[i].count[count] * 7 / 8;
512 }
513
514 if (sym_filter_entry)
515 show_details(sym_filter_entry);
516
517 last_refresh = time(NULL);
518
519 {
520 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
521
522 if (poll(&stdin_poll, 1, 0) == 1) {
523 printf("key pressed - exiting.\n");
524 exit(0);
525 }
526 }
527}
528
529static int read_symbol(FILE *in, struct sym_entry *s)
530{
531 static int filter_match = 0;
532 char *sym, stype;
533 char str[500];
534 int rc, pos;
535
536 rc = fscanf(in, "%llx %c %499s", &s->addr, &stype, str);
537 if (rc == EOF)
538 return -1;
539
540 assert(rc == 3);
541
542 /* skip until end of line: */
543 pos = strlen(str);
544 do {
545 rc = fgetc(in);
546 if (rc == '\n' || rc == EOF || pos >= 499)
547 break;
548 str[pos] = rc;
549 pos++;
550 } while (1);
551 str[pos] = 0;
552
553 sym = str;
554
555 /* Filter out known duplicates and non-text symbols. */
556 if (!strcmp(sym, "_text"))
557 return 1;
558 if (!min_ip && !strcmp(sym, "_stext"))
559 return 1;
560 if (!strcmp(sym, "_etext") || !strcmp(sym, "_sinittext"))
561 return 1;
562 if (stype != 'T' && stype != 't')
563 return 1;
564 if (!strncmp("init_module", sym, 11) || !strncmp("cleanup_module", sym, 14))
565 return 1;
566 if (strstr(sym, "_text_start") || strstr(sym, "_text_end"))
567 return 1;
568
569 s->sym = malloc(strlen(str));
570 assert(s->sym);
571
572 strcpy((char *)s->sym, str);
573 s->skip = 0;
574
575 /* Tag events to be skipped. */
576 if (!strcmp("default_idle", s->sym) || !strcmp("cpu_idle", s->sym))
577 s->skip = 1;
578 if (!strcmp("enter_idle", s->sym) || !strcmp("exit_idle", s->sym))
579 s->skip = 1;
580
581 if (filter_match == 1) {
582 filter_end = s->addr;
583 filter_match = -1;
584 if (filter_end - filter_start > 10000) {
585 printf("hm, too large filter symbol <%s> - skipping.\n",
586 sym_filter);
587 printf("symbol filter start: %016lx\n", filter_start);
588 printf(" end: %016lx\n", filter_end);
589 filter_end = filter_start = 0;
590 sym_filter = NULL;
591 sleep(1);
592 }
593 }
594 if (filter_match == 0 && sym_filter && !strcmp(s->sym, sym_filter)) {
595 filter_match = 1;
596 filter_start = s->addr;
597 }
598
599 return 0;
600}
601
602int compare_addr(const void *__sym1, const void *__sym2)
603{
604 const struct sym_entry *sym1 = __sym1, *sym2 = __sym2;
605
606 return sym1->addr > sym2->addr;
607}
608
609static void sort_symbol_table(void)
610{
611 int i, dups;
612
613 do {
614 qsort(sym_table, sym_table_count, sizeof(sym_table[0]), compare_addr);
615 for (i = 0, dups = 0; i < sym_table_count; i++) {
616 if (sym_table[i].addr == sym_table[i+1].addr) {
617 sym_table[i+1].addr = -1ll;
618 dups++;
619 }
620 }
621 sym_table_count -= dups;
622 } while(dups);
623}
624
625static void parse_symbols(void)
626{
627 struct sym_entry *last;
628
629 FILE *kallsyms = fopen("/proc/kallsyms", "r");
630
631 if (!kallsyms) {
632 printf("Could not open /proc/kallsyms - no CONFIG_KALLSYMS_ALL=y?\n");
633 exit(-1);
634 }
635
636 while (!feof(kallsyms)) {
637 if (read_symbol(kallsyms, &sym_table[sym_table_count]) == 0) {
638 sym_table_count++;
639 assert(sym_table_count <= MAX_SYMS);
640 }
641 }
642
643 sort_symbol_table();
644 min_ip = sym_table[0].addr;
645 max_ip = sym_table[sym_table_count-1].addr;
646 last = sym_table + sym_table_count++;
647
648 last->addr = -1ll;
649 last->sym = "<end>";
650
651 if (filter_end) {
652 int count;
653 for (count=0; count < sym_table_count; count ++) {
654 if (!strcmp(sym_table[count].sym, sym_filter)) {
655 sym_filter_entry = &sym_table[count];
656 break;
657 }
658 }
659 }
660 if (dump_symtab) {
661 int i;
662
663 for (i = 0; i < sym_table_count; i++)
664 fprintf(stderr, "%llx %s\n",
665 sym_table[i].addr, sym_table[i].sym);
666 }
667}
668
669
670static void parse_vmlinux(char *filename)
671{
672 FILE *file;
673 char command[PATH_MAX*2];
674 if (!filename)
675 return;
676
677 sprintf(command, "objdump --start-address=0x%016lx --stop-address=0x%016lx -dS %s", filter_start, filter_end, filename);
678
679 file = popen(command, "r");
680 if (!file)
681 return;
682
683 while (!feof(file)) {
684 struct source_line *src;
685 size_t dummy = 0;
686 char *c;
687
688 src = malloc(sizeof(struct source_line));
689 assert(src != NULL);
690 memset(src, 0, sizeof(struct source_line));
691
692 if (getline(&src->line, &dummy, file) < 0)
693 break;
694 if (!src->line)
695 break;
696
697 c = strchr(src->line, '\n');
698 if (c)
699 *c = 0;
700
701 lines = g_list_prepend(lines, src);
702
703 if (strlen(src->line)>8 && src->line[8] == ':')
704 src->EIP = strtoull(src->line, NULL, 16);
705 if (strlen(src->line)>8 && src->line[16] == ':')
706 src->EIP = strtoull(src->line, NULL, 16);
707 }
708 pclose(file);
709 lines = g_list_reverse(lines);
710}
711
712static void record_precise_ip(uint64_t ip)
713{
714 struct source_line *line;
715 GList *item;
716
717 item = g_list_first(lines);
718 while (item) {
719 line = item->data;
720 if (line->EIP == ip)
721 line->count++;
722 if (line->EIP > ip)
723 break;
724 item = g_list_next(item);
725 }
726}
727
728static void lookup_sym_in_vmlinux(struct sym_entry *sym)
729{
730 struct source_line *line;
731 GList *item;
732 char pattern[PATH_MAX];
733 sprintf(pattern, "<%s>:", sym->sym);
734
735 item = g_list_first(lines);
736 while (item) {
737 line = item->data;
738 if (strstr(line->line, pattern)) {
739 sym->source = item;
740 break;
741 }
742 item = g_list_next(item);
743 }
744}
745
746void show_lines(GList *item_queue, int item_queue_count)
747{
748 int i;
749 struct source_line *line;
750
751 for (i = 0; i < item_queue_count; i++) {
752 line = item_queue->data;
753 printf("%8li\t%s\n", line->count, line->line);
754 item_queue = g_list_next(item_queue);
755 }
756}
757
758#define TRACE_COUNT 3
759
760static void show_details(struct sym_entry *sym)
761{
762 struct source_line *line;
763 GList *item;
764 int displayed = 0;
765 GList *item_queue = NULL;
766 int item_queue_count = 0;
767
768 if (!sym->source)
769 lookup_sym_in_vmlinux(sym);
770 if (!sym->source)
771 return;
772
773 printf("Showing details for %s\n", sym->sym);
774
775 item = sym->source;
776 while (item) {
777 line = item->data;
778 if (displayed && strstr(line->line, ">:"))
779 break;
780
781 if (!item_queue_count)
782 item_queue = item;
783 item_queue_count ++;
784
785 if (line->count >= count_filter) {
786 show_lines(item_queue, item_queue_count);
787 item_queue_count = 0;
788 item_queue = NULL;
789 } else if (item_queue_count > TRACE_COUNT) {
790 item_queue = g_list_next(item_queue);
791 item_queue_count --;
792 }
793
794 line->count = 0;
795 displayed++;
796 if (displayed > 300)
797 break;
798 item = g_list_next(item);
799 }
800}
801
802/*
803 * Binary search in the histogram table and record the hit:
804 */
805static void record_ip(uint64_t ip, int counter)
806{
807 int left_idx, middle_idx, right_idx, idx;
808 unsigned long left, middle, right;
809
810 record_precise_ip(ip);
811
812 left_idx = 0;
813 right_idx = sym_table_count-1;
814 assert(ip <= max_ip && ip >= min_ip);
815
816 while (left_idx + 1 < right_idx) {
817 middle_idx = (left_idx + right_idx) / 2;
818
819 left = sym_table[ left_idx].addr;
820 middle = sym_table[middle_idx].addr;
821 right = sym_table[ right_idx].addr;
822
823 if (!(left <= middle && middle <= right)) {
824 printf("%016lx...\n%016lx...\n%016lx\n", left, middle, right);
825 printf("%d %d %d\n", left_idx, middle_idx, right_idx);
826 }
827 assert(left <= middle && middle <= right);
828 if (!(left <= ip && ip <= right)) {
829 printf(" left: %016lx\n", left);
830 printf(" ip: %016lx\n", ip);
831 printf("right: %016lx\n", right);
832 }
833 assert(left <= ip && ip <= right);
834 /*
835 * [ left .... target .... middle .... right ]
836 * => right := middle
837 */
838 if (ip < middle) {
839 right_idx = middle_idx;
840 continue;
841 }
842 /*
843 * [ left .... middle ... target ... right ]
844 * => left := middle
845 */
846 left_idx = middle_idx;
847 }
848
849 idx = left_idx;
850
851 if (!sym_table[idx].skip)
852 sym_table[idx].count[counter]++;
853 else events--;
854}
855
856static void process_event(uint64_t ip, int counter)
857{
858 events++;
859
860 if (ip < min_ip || ip > max_ip) {
861 userspace_events++;
862 return;
863 }
864
865 record_ip(ip, counter);
866}
867
868int main(int argc, char *argv[])
869{
870 struct pollfd event_array[MAX_NR_CPUS][MAX_COUNTERS];
871 struct perf_counter_hw_event hw_event;
872 int fd[MAX_NR_CPUS][MAX_COUNTERS];
873 int i, counter, group_fd;
874 unsigned int cpu;
875 uint64_t ip;
876 ssize_t res;
877 int ret;
878
879 process_options(argc, argv);
880
881 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
882 if (tid != -1 || profile_cpu != -1)
883 nr_cpus = 1;
884
885 assert(nr_cpus <= MAX_NR_CPUS);
886
887 for (i = 0; i < nr_cpus; i++) {
888 group_fd = -1;
889 for (counter = 0; counter < nr_counters; counter++) {
890
891 cpu = profile_cpu;
892 if (tid == -1 && profile_cpu == -1)
893 cpu = i;
894
895 memset(&hw_event, 0, sizeof(hw_event));
896 hw_event.type = event_id[counter];
897 hw_event.raw = event_raw[counter];
898 hw_event.irq_period = event_count[counter];
899 hw_event.record_type = PERF_RECORD_IRQ;
900 hw_event.nmi = nmi;
901
902 fd[i][counter] = sys_perf_counter_open(&hw_event, tid, cpu, group_fd, 0);
903 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
904 if (fd[i][counter] < 0) {
905 printf("kerneltop error: syscall returned with %d (%s)\n",
906 fd[i][counter], strerror(-fd[i][counter]));
907 if (fd[i][counter] == -1)
908 printf("Are you root?\n");
909 exit(-1);
910 }
911 assert(fd[i][counter] >= 0);
912
913 /*
914 * First counter acts as the group leader:
915 */
916 if (group && group_fd == -1)
917 group_fd = fd[i][counter];
918
919 event_array[i][counter].fd = fd[i][counter];
920 event_array[i][counter].events = POLLIN;
921 }
922 }
923
924 parse_symbols();
925 if (vmlinux && sym_filter_entry)
926 parse_vmlinux(vmlinux);
927
928 printf("KernelTop refresh period: %d seconds\n", delay_secs);
929 last_refresh = time(NULL);
930
931 while (1) {
932 int hits = events;
933
934 for (i = 0; i < nr_cpus; i++) {
935 for (counter = 0; counter < nr_counters; counter++) {
936 res = read(fd[i][counter], (char *) &ip, sizeof(ip));
937 if (res > 0) {
938 assert(res == sizeof(ip));
939
940 process_event(ip, counter);
941 }
942 }
943 }
944
945 if (time(NULL) >= last_refresh + delay_secs) {
946 print_sym_table();
947 events = userspace_events = 0;
948 }
949
950 if (hits == events)
951 ret = poll(event_array[0], nr_cpus, 1000);
952 hits = events;
953 }
954
955 return 0;
956}
diff --git a/Documentation/perf_counter/perfstat.c b/Documentation/perf_counter/perfstat.c
new file mode 100644
index 000000000000..9a5808fbcf90
--- /dev/null
+++ b/Documentation/perf_counter/perfstat.c
@@ -0,0 +1,521 @@
1/*
2 * perfstat: /usr/bin/time -alike performance counter statistics utility
3 *
4 * It summarizes the counter events of all tasks (and child tasks),
5 * covering all CPUs that the command (or workload) executes on.
6 * It only counts the per-task events of the workload started,
7 * independent of how many other tasks run on those CPUs.
8 *
9 * Build with: cc -O2 -g -lrt -Wall -W -o perfstat perfstat.c
10 *
11 * Sample output:
12 *
13
14 $ ./perfstat -e 1 -e 3 -e 5 ls -lR /usr/include/ >/dev/null
15
16 Performance counter stats for 'ls':
17
18 163516953 instructions
19 2295 cache-misses
20 2855182 branch-misses
21
22 *
23 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
24 *
25 * Released under the GPLv2 (not later).
26 *
27 * Percpu counter support by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
28 * Symbolic event options by: Wu Fengguang <fengguang.wu@intel.com>
29 */
30#define _GNU_SOURCE
31
32#include <assert.h>
33#include <getopt.h>
34#include <stdint.h>
35#include <stdlib.h>
36#include <string.h>
37#include <unistd.h>
38#include <ctype.h>
39#include <errno.h>
40#include <fcntl.h>
41#include <stdio.h>
42#include <time.h>
43
44#include <sys/syscall.h>
45#include <sys/ioctl.h>
46#include <sys/prctl.h>
47#include <sys/types.h>
48#include <sys/stat.h>
49#include <sys/time.h>
50#include <sys/wait.h>
51#include <sys/uio.h>
52
53#include <linux/unistd.h>
54
55#ifdef __x86_64__
56# define __NR_perf_counter_open 295
57#endif
58
59#ifdef __i386__
60# define __NR_perf_counter_open 333
61#endif
62
63#ifdef __powerpc__
64#define __NR_perf_counter_open 319
65#endif
66
67/*
68 * Pick up some kernel type conventions:
69 */
70#define __user
71#define asmlinkage
72
73typedef unsigned int __u32;
74typedef unsigned long long __u64;
75typedef long long __s64;
76
77#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
78
79/*
80 * User-space ABI bits:
81 */
82
83/*
84 * Generalized performance counter event types, used by the hw_event.type
85 * parameter of the sys_perf_counter_open() syscall:
86 */
87enum hw_event_types {
88 /*
89 * Common hardware events, generalized by the kernel:
90 */
91 PERF_COUNT_CPU_CYCLES = 0,
92 PERF_COUNT_INSTRUCTIONS = 1,
93 PERF_COUNT_CACHE_REFERENCES = 2,
94 PERF_COUNT_CACHE_MISSES = 3,
95 PERF_COUNT_BRANCH_INSTRUCTIONS = 4,
96 PERF_COUNT_BRANCH_MISSES = 5,
97 PERF_COUNT_BUS_CYCLES = 6,
98
99 PERF_HW_EVENTS_MAX = 7,
100
101 /*
102 * Special "software" counters provided by the kernel, even if
103 * the hardware does not support performance counters. These
104 * counters measure various physical and sw events of the
105 * kernel (and allow the profiling of them as well):
106 */
107 PERF_COUNT_CPU_CLOCK = -1,
108 PERF_COUNT_TASK_CLOCK = -2,
109 PERF_COUNT_PAGE_FAULTS = -3,
110 PERF_COUNT_CONTEXT_SWITCHES = -4,
111 PERF_COUNT_CPU_MIGRATIONS = -5,
112
113 PERF_SW_EVENTS_MIN = -6,
114};
115
116/*
117 * IRQ-notification data record type:
118 */
119enum perf_counter_record_type {
120 PERF_RECORD_SIMPLE = 0,
121 PERF_RECORD_IRQ = 1,
122 PERF_RECORD_GROUP = 2,
123};
124
125/*
126 * Hardware event to monitor via a performance monitoring counter:
127 */
128struct perf_counter_hw_event {
129 __s64 type;
130
131 __u64 irq_period;
132 __u64 record_type;
133 __u64 read_format;
134
135 __u64 disabled : 1, /* off by default */
136 nmi : 1, /* NMI sampling */
137 raw : 1, /* raw event type */
138 inherit : 1, /* children inherit it */
139 pinned : 1, /* must always be on PMU */
140 exclusive : 1, /* only group on PMU */
141 exclude_user : 1, /* don't count user */
142 exclude_kernel : 1, /* ditto kernel */
143 exclude_hv : 1, /* ditto hypervisor */
144 exclude_idle : 1, /* don't count when idle */
145
146 __reserved_1 : 54;
147
148 __u32 extra_config_len;
149 __u32 __reserved_4;
150
151 __u64 __reserved_2;
152 __u64 __reserved_3;
153};
154
155/*
156 * Ioctls that can be done on a perf counter fd:
157 */
158#define PERF_COUNTER_IOC_ENABLE _IO('$', 0)
159#define PERF_COUNTER_IOC_DISABLE _IO('$', 1)
160
161asmlinkage int sys_perf_counter_open(
162
163 struct perf_counter_hw_event *hw_event_uptr __user,
164 pid_t pid,
165 int cpu,
166 int group_fd,
167 unsigned long flags)
168{
169 int ret;
170
171 ret = syscall(
172 __NR_perf_counter_open, hw_event_uptr, pid, cpu, group_fd, flags);
173#if defined(__x86_64__) || defined(__i386__)
174 if (ret < 0 && ret > -4096) {
175 errno = -ret;
176 ret = -1;
177 }
178#endif
179 return ret;
180}
181
182
183static char *hw_event_names [] = {
184 "CPU cycles",
185 "instructions",
186 "cache references",
187 "cache misses",
188 "branches",
189 "branch misses",
190 "bus cycles",
191};
192
193static char *sw_event_names [] = {
194 "cpu clock ticks",
195 "task clock ticks",
196 "pagefaults",
197 "context switches",
198 "CPU migrations",
199};
200
201struct event_symbol {
202 int event;
203 char *symbol;
204};
205
206static struct event_symbol event_symbols [] = {
207 {PERF_COUNT_CPU_CYCLES, "cpu-cycles", },
208 {PERF_COUNT_CPU_CYCLES, "cycles", },
209 {PERF_COUNT_INSTRUCTIONS, "instructions", },
210 {PERF_COUNT_CACHE_REFERENCES, "cache-references", },
211 {PERF_COUNT_CACHE_MISSES, "cache-misses", },
212 {PERF_COUNT_BRANCH_INSTRUCTIONS, "branch-instructions", },
213 {PERF_COUNT_BRANCH_INSTRUCTIONS, "branches", },
214 {PERF_COUNT_BRANCH_MISSES, "branch-misses", },
215 {PERF_COUNT_BUS_CYCLES, "bus-cycles", },
216 {PERF_COUNT_CPU_CLOCK, "cpu-ticks", },
217 {PERF_COUNT_CPU_CLOCK, "ticks", },
218 {PERF_COUNT_TASK_CLOCK, "task-ticks", },
219 {PERF_COUNT_PAGE_FAULTS, "page-faults", },
220 {PERF_COUNT_PAGE_FAULTS, "faults", },
221 {PERF_COUNT_CONTEXT_SWITCHES, "context-switches", },
222 {PERF_COUNT_CONTEXT_SWITCHES, "cs", },
223 {PERF_COUNT_CPU_MIGRATIONS, "cpu-migrations", },
224 {PERF_COUNT_CPU_MIGRATIONS, "migrations", },
225};
226
227#define MAX_COUNTERS 64
228#define MAX_NR_CPUS 256
229
230static int nr_counters = 0;
231static int nr_cpus = 0;
232
233static int event_id[MAX_COUNTERS] =
234 { -2, -5, -4, -3, 0, 1, 2, 3};
235
236static int event_raw[MAX_COUNTERS];
237
238static int system_wide = 0;
239
240static void display_help(void)
241{
242 unsigned int i;
243 int e;
244
245 printf(
246 "Usage: perfstat [<events...>] <cmd...>\n\n"
247 "PerfStat Options (up to %d event types can be specified):\n\n",
248 MAX_COUNTERS);
249 printf(
250 " -e EVENT --event=EVENT # symbolic-name abbreviations");
251
252 for (i = 0, e = PERF_HW_EVENTS_MAX; i < ARRAY_SIZE(event_symbols); i++) {
253 if (e != event_symbols[i].event) {
254 e = event_symbols[i].event;
255 printf(
256 "\n %2d: %-20s", e, event_symbols[i].symbol);
257 } else
258 printf(" %s", event_symbols[i].symbol);
259 }
260
261 printf("\n"
262 " rNNN: raw event type\n\n"
263 " -s # system-wide collection\n\n"
264 " -c <cmd..> --command=<cmd..> # command+arguments to be timed.\n"
265 "\n");
266 exit(0);
267}
268
269static int type_valid(int type)
270{
271 if (type >= PERF_HW_EVENTS_MAX)
272 return 0;
273 if (type <= PERF_SW_EVENTS_MIN)
274 return 0;
275
276 return 1;
277}
278
279static char *event_name(int ctr)
280{
281 int type = event_id[ctr];
282 static char buf[32];
283
284 if (event_raw[ctr]) {
285 sprintf(buf, "raw 0x%x", type);
286 return buf;
287 }
288 if (!type_valid(type))
289 return "unknown";
290
291 if (type >= 0)
292 return hw_event_names[type];
293
294 return sw_event_names[-type-1];
295}
296
297/*
298 * Each event can have multiple symbolic names.
299 * Symbolic names are (almost) exactly matched.
300 */
301static int match_event_symbols(char *str)
302{
303 unsigned int i;
304
305 if (isdigit(str[0]) || str[0] == '-')
306 return atoi(str);
307
308 for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
309 if (!strncmp(str, event_symbols[i].symbol,
310 strlen(event_symbols[i].symbol)))
311 return event_symbols[i].event;
312 }
313
314 return PERF_HW_EVENTS_MAX;
315}
316
317static void parse_events(char *str)
318{
319 int type, raw;
320
321again:
322 nr_counters++;
323 if (nr_counters == MAX_COUNTERS)
324 display_help();
325
326 raw = 0;
327 if (*str == 'r') {
328 raw = 1;
329 ++str;
330 type = strtol(str, NULL, 16);
331 } else {
332 type = match_event_symbols(str);
333 if (!type_valid(type))
334 display_help();
335 }
336
337 event_id[nr_counters] = type;
338 event_raw[nr_counters] = raw;
339
340 str = strstr(str, ",");
341 if (str) {
342 str++;
343 goto again;
344 }
345}
346
347static void process_options(int argc, char *argv[])
348{
349 for (;;) {
350 int option_index = 0;
351 /** Options for getopt */
352 static struct option long_options[] = {
353 {"event", required_argument, NULL, 'e'},
354 {"help", no_argument, NULL, 'h'},
355 {"command", no_argument, NULL, 'c'},
356 {NULL, 0, NULL, 0 }
357 };
358 int c = getopt_long(argc, argv, "+:e:c:s",
359 long_options, &option_index);
360 if (c == -1)
361 break;
362
363 switch (c) {
364 case 'c':
365 break;
366 case 's':
367 system_wide = 1;
368 break;
369 case 'e':
370 parse_events(optarg);
371 break;
372 default:
373 break;
374 }
375 }
376 if (optind == argc)
377 goto err;
378
379 if (!nr_counters)
380 nr_counters = 8;
381 else
382 nr_counters++;
383 return;
384
385err:
386 display_help();
387}
388
389char fault_here[1000000];
390
391#define PR_TASK_PERF_COUNTERS_DISABLE 31
392#define PR_TASK_PERF_COUNTERS_ENABLE 32
393
394static int fd[MAX_NR_CPUS][MAX_COUNTERS];
395
396static void create_counter(int counter)
397{
398 struct perf_counter_hw_event hw_event;
399
400 memset(&hw_event, 0, sizeof(hw_event));
401 hw_event.type = event_id[counter];
402 hw_event.raw = event_raw[counter];
403 hw_event.record_type = PERF_RECORD_SIMPLE;
404 hw_event.nmi = 0;
405
406 if (system_wide) {
407 int cpu;
408 for (cpu = 0; cpu < nr_cpus; cpu ++) {
409 fd[cpu][counter] = sys_perf_counter_open(&hw_event, -1, cpu, -1, 0);
410 if (fd[cpu][counter] < 0) {
411 printf("perfstat error: syscall returned with %d (%s)\n",
412 fd[cpu][counter], strerror(errno));
413 exit(-1);
414 }
415
416 }
417 } else {
418 hw_event.inherit = 1;
419 hw_event.disabled = 1;
420
421 fd[0][counter] = sys_perf_counter_open(&hw_event, 0, -1, -1, 0);
422 if (fd[0][counter] < 0) {
423 printf("perfstat error: syscall returned with %d (%s)\n",
424 fd[0][counter], strerror(errno));
425 exit(-1);
426 }
427 }
428}
429
430
431#define rdclock() \
432({ \
433 struct timespec ts; \
434 \
435 clock_gettime(CLOCK_MONOTONIC, &ts); \
436 ts.tv_sec * 1000000000ULL + ts.tv_nsec; \
437})
438
439int main(int argc, char *argv[])
440{
441 unsigned long long t0, t1;
442 int counter;
443 ssize_t res;
444 int status;
445 int pid;
446
447 process_options(argc, argv);
448
449 if (system_wide) {
450 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
451 assert(nr_cpus <= MAX_NR_CPUS);
452 assert(nr_cpus >= 0);
453 } else
454 nr_cpus = 1;
455
456 for (counter = 0; counter < nr_counters; counter++)
457 create_counter(counter);
458
459 argc -= optind;
460 argv += optind;
461
462 /*
463 * Enable counters and exec the command:
464 */
465 t0 = rdclock();
466 prctl(PR_TASK_PERF_COUNTERS_ENABLE);
467
468 if ((pid = fork()) < 0)
469 perror("failed to fork");
470 if (!pid) {
471 if (execvp(argv[0], argv)) {
472 perror(argv[0]);
473 exit(-1);
474 }
475 }
476 while (wait(&status) >= 0)
477 ;
478 prctl(PR_TASK_PERF_COUNTERS_DISABLE);
479 t1 = rdclock();
480
481 fflush(stdout);
482
483 fprintf(stderr, "\n");
484 fprintf(stderr, " Performance counter stats for \'%s\':\n",
485 argv[0]);
486 fprintf(stderr, "\n");
487
488 for (counter = 0; counter < nr_counters; counter++) {
489 int cpu;
490 __u64 count, single_count;
491
492 count = 0;
493 for (cpu = 0; cpu < nr_cpus; cpu ++) {
494 res = read(fd[cpu][counter],
495 (char *) &single_count, sizeof(single_count));
496 assert(res == sizeof(single_count));
497 count += single_count;
498 }
499
500 if (!event_raw[counter] &&
501 (event_id[counter] == PERF_COUNT_CPU_CLOCK ||
502 event_id[counter] == PERF_COUNT_TASK_CLOCK)) {
503
504 double msecs = (double)count / 1000000;
505
506 fprintf(stderr, " %14.6f %-20s (msecs)\n",
507 msecs, event_name(counter));
508 } else {
509 fprintf(stderr, " %14Ld %-20s (events)\n",
510 count, event_name(counter));
511 }
512 if (!counter)
513 fprintf(stderr, "\n");
514 }
515 fprintf(stderr, "\n");
516 fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n",
517 (double)(t1-t0)/1e6);
518 fprintf(stderr, "\n");
519
520 return 0;
521}
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-28 14:31:15 -0500