]>
Commit | Line | Data |
---|---|---|
07800601 IM |
1 | /* |
2 | * kerneltop.c: show top kernel functions - performance counters showcase | |
3 | ||
4 | Build with: | |
5 | ||
38105f02 | 6 | make -C Documentation/perf_counter/ |
07800601 IM |
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 | ||
07800601 IM |
31 | /* |
32 | * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> | |
33 | * | |
34 | * Improvements and fixes by: | |
35 | * | |
36 | * Arjan van de Ven <arjan@linux.intel.com> | |
37 | * Yanmin Zhang <yanmin.zhang@intel.com> | |
38 | * Wu Fengguang <fengguang.wu@intel.com> | |
39 | * Mike Galbraith <efault@gmx.de> | |
40 | * Paul Mackerras <paulus@samba.org> | |
41 | * | |
42 | * Released under the GPL v2. (and only v2, not any later version) | |
43 | */ | |
44 | ||
148be2c1 | 45 | #include "util/util.h" |
07800601 | 46 | |
07800601 IM |
47 | #include <getopt.h> |
48 | #include <assert.h> | |
49 | #include <fcntl.h> | |
50 | #include <stdio.h> | |
51 | #include <errno.h> | |
07800601 IM |
52 | #include <time.h> |
53 | #include <sched.h> | |
54 | #include <pthread.h> | |
55 | ||
56 | #include <sys/syscall.h> | |
57 | #include <sys/ioctl.h> | |
58 | #include <sys/poll.h> | |
59 | #include <sys/prctl.h> | |
60 | #include <sys/wait.h> | |
61 | #include <sys/uio.h> | |
62 | #include <sys/mman.h> | |
63 | ||
64 | #include <linux/unistd.h> | |
65 | #include <linux/types.h> | |
66 | ||
67 | #include "../../include/linux/perf_counter.h" | |
68 | ||
6eda5838 | 69 | #include "perf.h" |
07800601 | 70 | |
07800601 IM |
71 | static int system_wide = 0; |
72 | ||
73 | static int nr_counters = 0; | |
74 | static __u64 event_id[MAX_COUNTERS] = { | |
75 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK), | |
76 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), | |
77 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), | |
78 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), | |
79 | ||
80 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), | |
81 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS), | |
82 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES), | |
83 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES), | |
84 | }; | |
85 | static int default_interval = 100000; | |
86 | static int event_count[MAX_COUNTERS]; | |
87 | static int fd[MAX_NR_CPUS][MAX_COUNTERS]; | |
88 | ||
89 | static __u64 count_filter = 100; | |
90 | ||
91 | static int tid = -1; | |
92 | static int profile_cpu = -1; | |
93 | static int nr_cpus = 0; | |
94 | static int nmi = 1; | |
95 | static unsigned int realtime_prio = 0; | |
96 | static int group = 0; | |
97 | static unsigned int page_size; | |
98 | static unsigned int mmap_pages = 16; | |
99 | static int use_mmap = 0; | |
100 | static int use_munmap = 0; | |
f5456a6b | 101 | static int freq = 0; |
07800601 IM |
102 | |
103 | static char *vmlinux; | |
104 | ||
105 | static char *sym_filter; | |
106 | static unsigned long filter_start; | |
107 | static unsigned long filter_end; | |
108 | ||
109 | static int delay_secs = 2; | |
110 | static int zero; | |
111 | static int dump_symtab; | |
112 | ||
113 | static int scale; | |
114 | ||
115 | struct source_line { | |
116 | uint64_t EIP; | |
117 | unsigned long count; | |
118 | char *line; | |
119 | struct source_line *next; | |
120 | }; | |
121 | ||
122 | static struct source_line *lines; | |
123 | static struct source_line **lines_tail; | |
124 | ||
ddcacfa0 | 125 | static const unsigned int default_count[] = { |
07800601 IM |
126 | 1000000, |
127 | 1000000, | |
128 | 10000, | |
129 | 10000, | |
130 | 1000000, | |
131 | 10000, | |
132 | }; | |
133 | ||
134 | static char *hw_event_names[] = { | |
135 | "CPU cycles", | |
136 | "instructions", | |
137 | "cache references", | |
138 | "cache misses", | |
139 | "branches", | |
140 | "branch misses", | |
141 | "bus cycles", | |
142 | }; | |
143 | ||
144 | static char *sw_event_names[] = { | |
145 | "cpu clock ticks", | |
146 | "task clock ticks", | |
147 | "pagefaults", | |
148 | "context switches", | |
149 | "CPU migrations", | |
150 | "minor faults", | |
151 | "major faults", | |
152 | }; | |
153 | ||
154 | struct event_symbol { | |
155 | __u64 event; | |
156 | char *symbol; | |
157 | }; | |
158 | ||
159 | static struct event_symbol event_symbols[] = { | |
160 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), "cpu-cycles", }, | |
161 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), "cycles", }, | |
162 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS), "instructions", }, | |
163 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES), "cache-references", }, | |
164 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES), "cache-misses", }, | |
165 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_INSTRUCTIONS), "branch-instructions", }, | |
166 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_INSTRUCTIONS), "branches", }, | |
167 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BRANCH_MISSES), "branch-misses", }, | |
168 | {EID(PERF_TYPE_HARDWARE, PERF_COUNT_BUS_CYCLES), "bus-cycles", }, | |
169 | ||
170 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_CLOCK), "cpu-clock", }, | |
171 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK), "task-clock", }, | |
172 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), "page-faults", }, | |
173 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), "faults", }, | |
174 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS_MIN), "minor-faults", }, | |
175 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS_MAJ), "major-faults", }, | |
176 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), "context-switches", }, | |
177 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), "cs", }, | |
178 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), "cpu-migrations", }, | |
179 | {EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), "migrations", }, | |
180 | }; | |
181 | ||
182 | #define __PERF_COUNTER_FIELD(config, name) \ | |
183 | ((config & PERF_COUNTER_##name##_MASK) >> PERF_COUNTER_##name##_SHIFT) | |
184 | ||
185 | #define PERF_COUNTER_RAW(config) __PERF_COUNTER_FIELD(config, RAW) | |
186 | #define PERF_COUNTER_CONFIG(config) __PERF_COUNTER_FIELD(config, CONFIG) | |
187 | #define PERF_COUNTER_TYPE(config) __PERF_COUNTER_FIELD(config, TYPE) | |
188 | #define PERF_COUNTER_ID(config) __PERF_COUNTER_FIELD(config, EVENT) | |
189 | ||
190 | static void display_events_help(void) | |
191 | { | |
192 | unsigned int i; | |
193 | __u64 e; | |
194 | ||
195 | printf( | |
196 | " -e EVENT --event=EVENT # symbolic-name abbreviations"); | |
197 | ||
198 | for (i = 0; i < ARRAY_SIZE(event_symbols); i++) { | |
199 | int type, id; | |
200 | ||
201 | e = event_symbols[i].event; | |
202 | type = PERF_COUNTER_TYPE(e); | |
203 | id = PERF_COUNTER_ID(e); | |
204 | ||
205 | printf("\n %d:%d: %-20s", | |
206 | type, id, event_symbols[i].symbol); | |
207 | } | |
208 | ||
209 | printf("\n" | |
210 | " rNNN: raw PMU events (eventsel+umask)\n\n"); | |
211 | } | |
212 | ||
07800601 IM |
213 | static void display_help(void) |
214 | { | |
07800601 IM |
215 | printf( |
216 | "Usage: kerneltop [<options>]\n" | |
217 | " Or: kerneltop -S [<options>] COMMAND [ARGS]\n\n" | |
218 | "KernelTop Options (up to %d event types can be specified at once):\n\n", | |
219 | MAX_COUNTERS); | |
220 | ||
221 | display_events_help(); | |
222 | ||
223 | printf( | |
07800601 IM |
224 | " -c CNT --count=CNT # event period to sample\n\n" |
225 | " -C CPU --cpu=CPU # CPU (-1 for all) [default: -1]\n" | |
226 | " -p PID --pid=PID # PID of sampled task (-1 for all) [default: -1]\n\n" | |
227 | " -l # show scale factor for RR events\n" | |
228 | " -d delay --delay=<seconds> # sampling/display delay [default: 2]\n" | |
229 | " -f CNT --filter=CNT # min-event-count filter [default: 100]\n\n" | |
230 | " -r prio --realtime=<prio> # event acquisition runs with SCHED_FIFO policy\n" | |
231 | " -s symbol --symbol=<symbol> # function to be showed annotated one-shot\n" | |
232 | " -x path --vmlinux=<path> # the vmlinux binary, required for -s use\n" | |
233 | " -z --zero # zero counts after display\n" | |
234 | " -D --dump_symtab # dump symbol table to stderr on startup\n" | |
235 | " -m pages --mmap_pages=<pages> # number of mmap data pages\n" | |
236 | " -M --mmap_info # print mmap info stream\n" | |
237 | " -U --munmap_info # print munmap info stream\n" | |
238 | ); | |
239 | ||
240 | exit(0); | |
241 | } | |
242 | ||
243 | static char *event_name(int ctr) | |
244 | { | |
245 | __u64 config = event_id[ctr]; | |
246 | int type = PERF_COUNTER_TYPE(config); | |
247 | int id = PERF_COUNTER_ID(config); | |
248 | static char buf[32]; | |
249 | ||
250 | if (PERF_COUNTER_RAW(config)) { | |
251 | sprintf(buf, "raw 0x%llx", PERF_COUNTER_CONFIG(config)); | |
252 | return buf; | |
253 | } | |
254 | ||
255 | switch (type) { | |
256 | case PERF_TYPE_HARDWARE: | |
257 | if (id < PERF_HW_EVENTS_MAX) | |
258 | return hw_event_names[id]; | |
259 | return "unknown-hardware"; | |
260 | ||
261 | case PERF_TYPE_SOFTWARE: | |
262 | if (id < PERF_SW_EVENTS_MAX) | |
263 | return sw_event_names[id]; | |
264 | return "unknown-software"; | |
265 | ||
266 | default: | |
267 | break; | |
268 | } | |
269 | ||
270 | return "unknown"; | |
271 | } | |
272 | ||
273 | /* | |
274 | * Each event can have multiple symbolic names. | |
275 | * Symbolic names are (almost) exactly matched. | |
276 | */ | |
277 | static __u64 match_event_symbols(char *str) | |
278 | { | |
279 | __u64 config, id; | |
280 | int type; | |
281 | unsigned int i; | |
282 | ||
283 | if (sscanf(str, "r%llx", &config) == 1) | |
284 | return config | PERF_COUNTER_RAW_MASK; | |
285 | ||
286 | if (sscanf(str, "%d:%llu", &type, &id) == 2) | |
287 | return EID(type, id); | |
288 | ||
289 | for (i = 0; i < ARRAY_SIZE(event_symbols); i++) { | |
290 | if (!strncmp(str, event_symbols[i].symbol, | |
291 | strlen(event_symbols[i].symbol))) | |
292 | return event_symbols[i].event; | |
293 | } | |
294 | ||
295 | return ~0ULL; | |
296 | } | |
297 | ||
298 | static int parse_events(char *str) | |
299 | { | |
300 | __u64 config; | |
301 | ||
302 | again: | |
303 | if (nr_counters == MAX_COUNTERS) | |
304 | return -1; | |
305 | ||
306 | config = match_event_symbols(str); | |
307 | if (config == ~0ULL) | |
308 | return -1; | |
309 | ||
310 | event_id[nr_counters] = config; | |
311 | nr_counters++; | |
312 | ||
313 | str = strstr(str, ","); | |
314 | if (str) { | |
315 | str++; | |
316 | goto again; | |
317 | } | |
318 | ||
319 | return 0; | |
320 | } | |
321 | ||
07800601 IM |
322 | /* |
323 | * Symbols | |
324 | */ | |
325 | ||
326 | static uint64_t min_ip; | |
327 | static uint64_t max_ip = -1ll; | |
328 | ||
329 | struct sym_entry { | |
330 | unsigned long long addr; | |
331 | char *sym; | |
332 | unsigned long count[MAX_COUNTERS]; | |
333 | int skip; | |
334 | struct source_line *source; | |
335 | }; | |
336 | ||
337 | #define MAX_SYMS 100000 | |
338 | ||
339 | static int sym_table_count; | |
340 | ||
341 | struct sym_entry *sym_filter_entry; | |
342 | ||
343 | static struct sym_entry sym_table[MAX_SYMS]; | |
344 | ||
345 | static void show_details(struct sym_entry *sym); | |
346 | ||
347 | /* | |
348 | * Ordering weight: count-1 * count-2 * ... / count-n | |
349 | */ | |
350 | static double sym_weight(const struct sym_entry *sym) | |
351 | { | |
352 | double weight; | |
353 | int counter; | |
354 | ||
355 | weight = sym->count[0]; | |
356 | ||
357 | for (counter = 1; counter < nr_counters-1; counter++) | |
358 | weight *= sym->count[counter]; | |
359 | ||
360 | weight /= (sym->count[counter] + 1); | |
361 | ||
362 | return weight; | |
363 | } | |
364 | ||
365 | static int compare(const void *__sym1, const void *__sym2) | |
366 | { | |
367 | const struct sym_entry *sym1 = __sym1, *sym2 = __sym2; | |
368 | ||
369 | return sym_weight(sym1) < sym_weight(sym2); | |
370 | } | |
371 | ||
372 | static long events; | |
373 | static long userspace_events; | |
374 | static const char CONSOLE_CLEAR[] = "\e[H\e[2J"; | |
375 | ||
376 | static struct sym_entry tmp[MAX_SYMS]; | |
377 | ||
378 | static void print_sym_table(void) | |
379 | { | |
380 | int i, printed; | |
381 | int counter; | |
382 | float events_per_sec = events/delay_secs; | |
383 | float kevents_per_sec = (events-userspace_events)/delay_secs; | |
384 | float sum_kevents = 0.0; | |
385 | ||
386 | events = userspace_events = 0; | |
387 | memcpy(tmp, sym_table, sizeof(sym_table[0])*sym_table_count); | |
388 | qsort(tmp, sym_table_count, sizeof(tmp[0]), compare); | |
389 | ||
390 | for (i = 0; i < sym_table_count && tmp[i].count[0]; i++) | |
391 | sum_kevents += tmp[i].count[0]; | |
392 | ||
393 | write(1, CONSOLE_CLEAR, strlen(CONSOLE_CLEAR)); | |
394 | ||
395 | printf( | |
396 | "------------------------------------------------------------------------------\n"); | |
397 | printf( " KernelTop:%8.0f irqs/sec kernel:%4.1f%% [%s, ", | |
398 | events_per_sec, | |
399 | 100.0 - (100.0*((events_per_sec-kevents_per_sec)/events_per_sec)), | |
400 | nmi ? "NMI" : "IRQ"); | |
401 | ||
402 | if (nr_counters == 1) | |
403 | printf("%d ", event_count[0]); | |
404 | ||
405 | for (counter = 0; counter < nr_counters; counter++) { | |
406 | if (counter) | |
407 | printf("/"); | |
408 | ||
409 | printf("%s", event_name(counter)); | |
410 | } | |
411 | ||
412 | printf( "], "); | |
413 | ||
414 | if (tid != -1) | |
415 | printf(" (tid: %d", tid); | |
416 | else | |
417 | printf(" (all"); | |
418 | ||
419 | if (profile_cpu != -1) | |
420 | printf(", cpu: %d)\n", profile_cpu); | |
421 | else { | |
422 | if (tid != -1) | |
423 | printf(")\n"); | |
424 | else | |
425 | printf(", %d CPUs)\n", nr_cpus); | |
426 | } | |
427 | ||
428 | printf("------------------------------------------------------------------------------\n\n"); | |
429 | ||
430 | if (nr_counters == 1) | |
431 | printf(" events pcnt"); | |
432 | else | |
433 | printf(" weight events pcnt"); | |
434 | ||
435 | printf(" RIP kernel function\n" | |
436 | " ______ ______ _____ ________________ _______________\n\n" | |
437 | ); | |
438 | ||
439 | for (i = 0, printed = 0; i < sym_table_count; i++) { | |
440 | float pcnt; | |
441 | int count; | |
442 | ||
443 | if (printed <= 18 && tmp[i].count[0] >= count_filter) { | |
444 | pcnt = 100.0 - (100.0*((sum_kevents-tmp[i].count[0])/sum_kevents)); | |
445 | ||
446 | if (nr_counters == 1) | |
447 | printf("%19.2f - %4.1f%% - %016llx : %s\n", | |
448 | sym_weight(tmp + i), | |
449 | pcnt, tmp[i].addr, tmp[i].sym); | |
450 | else | |
451 | printf("%8.1f %10ld - %4.1f%% - %016llx : %s\n", | |
452 | sym_weight(tmp + i), | |
453 | tmp[i].count[0], | |
454 | pcnt, tmp[i].addr, tmp[i].sym); | |
455 | printed++; | |
456 | } | |
457 | /* | |
458 | * Add decay to the counts: | |
459 | */ | |
460 | for (count = 0; count < nr_counters; count++) | |
461 | sym_table[i].count[count] = zero ? 0 : sym_table[i].count[count] * 7 / 8; | |
462 | } | |
463 | ||
464 | if (sym_filter_entry) | |
465 | show_details(sym_filter_entry); | |
466 | ||
467 | { | |
468 | struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; | |
469 | ||
470 | if (poll(&stdin_poll, 1, 0) == 1) { | |
471 | printf("key pressed - exiting.\n"); | |
472 | exit(0); | |
473 | } | |
474 | } | |
475 | } | |
476 | ||
477 | static void *display_thread(void *arg) | |
478 | { | |
479 | printf("KernelTop refresh period: %d seconds\n", delay_secs); | |
480 | ||
481 | while (!sleep(delay_secs)) | |
482 | print_sym_table(); | |
483 | ||
484 | return NULL; | |
485 | } | |
486 | ||
487 | static int read_symbol(FILE *in, struct sym_entry *s) | |
488 | { | |
489 | static int filter_match = 0; | |
490 | char *sym, stype; | |
491 | char str[500]; | |
492 | int rc, pos; | |
493 | ||
494 | rc = fscanf(in, "%llx %c %499s", &s->addr, &stype, str); | |
495 | if (rc == EOF) | |
496 | return -1; | |
497 | ||
498 | assert(rc == 3); | |
499 | ||
500 | /* skip until end of line: */ | |
501 | pos = strlen(str); | |
502 | do { | |
503 | rc = fgetc(in); | |
504 | if (rc == '\n' || rc == EOF || pos >= 499) | |
505 | break; | |
506 | str[pos] = rc; | |
507 | pos++; | |
508 | } while (1); | |
509 | str[pos] = 0; | |
510 | ||
511 | sym = str; | |
512 | ||
513 | /* Filter out known duplicates and non-text symbols. */ | |
514 | if (!strcmp(sym, "_text")) | |
515 | return 1; | |
516 | if (!min_ip && !strcmp(sym, "_stext")) | |
517 | return 1; | |
518 | if (!strcmp(sym, "_etext") || !strcmp(sym, "_sinittext")) | |
519 | return 1; | |
520 | if (stype != 'T' && stype != 't') | |
521 | return 1; | |
522 | if (!strncmp("init_module", sym, 11) || !strncmp("cleanup_module", sym, 14)) | |
523 | return 1; | |
524 | if (strstr(sym, "_text_start") || strstr(sym, "_text_end")) | |
525 | return 1; | |
526 | ||
82afae60 | 527 | s->sym = malloc(strlen(str)+1); |
07800601 IM |
528 | assert(s->sym); |
529 | ||
530 | strcpy((char *)s->sym, str); | |
531 | s->skip = 0; | |
532 | ||
533 | /* Tag events to be skipped. */ | |
534 | if (!strcmp("default_idle", s->sym) || !strcmp("cpu_idle", s->sym)) | |
535 | s->skip = 1; | |
536 | else if (!strcmp("enter_idle", s->sym) || !strcmp("exit_idle", s->sym)) | |
537 | s->skip = 1; | |
538 | else if (!strcmp("mwait_idle", s->sym)) | |
539 | s->skip = 1; | |
540 | ||
541 | if (filter_match == 1) { | |
542 | filter_end = s->addr; | |
543 | filter_match = -1; | |
544 | if (filter_end - filter_start > 10000) { | |
545 | printf("hm, too large filter symbol <%s> - skipping.\n", | |
546 | sym_filter); | |
547 | printf("symbol filter start: %016lx\n", filter_start); | |
548 | printf(" end: %016lx\n", filter_end); | |
549 | filter_end = filter_start = 0; | |
550 | sym_filter = NULL; | |
551 | sleep(1); | |
552 | } | |
553 | } | |
554 | if (filter_match == 0 && sym_filter && !strcmp(s->sym, sym_filter)) { | |
555 | filter_match = 1; | |
556 | filter_start = s->addr; | |
557 | } | |
558 | ||
559 | return 0; | |
560 | } | |
561 | ||
ddcacfa0 | 562 | static int compare_addr(const void *__sym1, const void *__sym2) |
07800601 IM |
563 | { |
564 | const struct sym_entry *sym1 = __sym1, *sym2 = __sym2; | |
565 | ||
566 | return sym1->addr > sym2->addr; | |
567 | } | |
568 | ||
569 | static void sort_symbol_table(void) | |
570 | { | |
571 | int i, dups; | |
572 | ||
573 | do { | |
574 | qsort(sym_table, sym_table_count, sizeof(sym_table[0]), compare_addr); | |
575 | for (i = 0, dups = 0; i < sym_table_count; i++) { | |
576 | if (sym_table[i].addr == sym_table[i+1].addr) { | |
577 | sym_table[i+1].addr = -1ll; | |
578 | dups++; | |
579 | } | |
580 | } | |
581 | sym_table_count -= dups; | |
582 | } while(dups); | |
583 | } | |
584 | ||
585 | static void parse_symbols(void) | |
586 | { | |
587 | struct sym_entry *last; | |
588 | ||
589 | FILE *kallsyms = fopen("/proc/kallsyms", "r"); | |
590 | ||
591 | if (!kallsyms) { | |
592 | printf("Could not open /proc/kallsyms - no CONFIG_KALLSYMS_ALL=y?\n"); | |
593 | exit(-1); | |
594 | } | |
595 | ||
596 | while (!feof(kallsyms)) { | |
597 | if (read_symbol(kallsyms, &sym_table[sym_table_count]) == 0) { | |
598 | sym_table_count++; | |
599 | assert(sym_table_count <= MAX_SYMS); | |
600 | } | |
601 | } | |
602 | ||
603 | sort_symbol_table(); | |
604 | min_ip = sym_table[0].addr; | |
605 | max_ip = sym_table[sym_table_count-1].addr; | |
606 | last = sym_table + sym_table_count++; | |
607 | ||
608 | last->addr = -1ll; | |
609 | last->sym = "<end>"; | |
610 | ||
611 | if (filter_end) { | |
612 | int count; | |
613 | for (count=0; count < sym_table_count; count ++) { | |
614 | if (!strcmp(sym_table[count].sym, sym_filter)) { | |
615 | sym_filter_entry = &sym_table[count]; | |
616 | break; | |
617 | } | |
618 | } | |
619 | } | |
620 | if (dump_symtab) { | |
621 | int i; | |
622 | ||
623 | for (i = 0; i < sym_table_count; i++) | |
624 | fprintf(stderr, "%llx %s\n", | |
625 | sym_table[i].addr, sym_table[i].sym); | |
626 | } | |
627 | } | |
628 | ||
629 | /* | |
630 | * Source lines | |
631 | */ | |
632 | ||
633 | static void parse_vmlinux(char *filename) | |
634 | { | |
635 | FILE *file; | |
636 | char command[PATH_MAX*2]; | |
637 | if (!filename) | |
638 | return; | |
639 | ||
640 | sprintf(command, "objdump --start-address=0x%016lx --stop-address=0x%016lx -dS %s", filter_start, filter_end, filename); | |
641 | ||
642 | file = popen(command, "r"); | |
643 | if (!file) | |
644 | return; | |
645 | ||
646 | lines_tail = &lines; | |
647 | while (!feof(file)) { | |
648 | struct source_line *src; | |
649 | size_t dummy = 0; | |
650 | char *c; | |
651 | ||
652 | src = malloc(sizeof(struct source_line)); | |
653 | assert(src != NULL); | |
654 | memset(src, 0, sizeof(struct source_line)); | |
655 | ||
656 | if (getline(&src->line, &dummy, file) < 0) | |
657 | break; | |
658 | if (!src->line) | |
659 | break; | |
660 | ||
661 | c = strchr(src->line, '\n'); | |
662 | if (c) | |
663 | *c = 0; | |
664 | ||
665 | src->next = NULL; | |
666 | *lines_tail = src; | |
667 | lines_tail = &src->next; | |
668 | ||
669 | if (strlen(src->line)>8 && src->line[8] == ':') | |
670 | src->EIP = strtoull(src->line, NULL, 16); | |
671 | if (strlen(src->line)>8 && src->line[16] == ':') | |
672 | src->EIP = strtoull(src->line, NULL, 16); | |
673 | } | |
674 | pclose(file); | |
675 | } | |
676 | ||
677 | static void record_precise_ip(uint64_t ip) | |
678 | { | |
679 | struct source_line *line; | |
680 | ||
681 | for (line = lines; line; line = line->next) { | |
682 | if (line->EIP == ip) | |
683 | line->count++; | |
684 | if (line->EIP > ip) | |
685 | break; | |
686 | } | |
687 | } | |
688 | ||
689 | static void lookup_sym_in_vmlinux(struct sym_entry *sym) | |
690 | { | |
691 | struct source_line *line; | |
692 | char pattern[PATH_MAX]; | |
693 | sprintf(pattern, "<%s>:", sym->sym); | |
694 | ||
695 | for (line = lines; line; line = line->next) { | |
696 | if (strstr(line->line, pattern)) { | |
697 | sym->source = line; | |
698 | break; | |
699 | } | |
700 | } | |
701 | } | |
702 | ||
703 | static void show_lines(struct source_line *line_queue, int line_queue_count) | |
704 | { | |
705 | int i; | |
706 | struct source_line *line; | |
707 | ||
708 | line = line_queue; | |
709 | for (i = 0; i < line_queue_count; i++) { | |
710 | printf("%8li\t%s\n", line->count, line->line); | |
711 | line = line->next; | |
712 | } | |
713 | } | |
714 | ||
715 | #define TRACE_COUNT 3 | |
716 | ||
717 | static void show_details(struct sym_entry *sym) | |
718 | { | |
719 | struct source_line *line; | |
720 | struct source_line *line_queue = NULL; | |
721 | int displayed = 0; | |
722 | int line_queue_count = 0; | |
723 | ||
724 | if (!sym->source) | |
725 | lookup_sym_in_vmlinux(sym); | |
726 | if (!sym->source) | |
727 | return; | |
728 | ||
729 | printf("Showing details for %s\n", sym->sym); | |
730 | ||
731 | line = sym->source; | |
732 | while (line) { | |
733 | if (displayed && strstr(line->line, ">:")) | |
734 | break; | |
735 | ||
736 | if (!line_queue_count) | |
737 | line_queue = line; | |
738 | line_queue_count ++; | |
739 | ||
740 | if (line->count >= count_filter) { | |
741 | show_lines(line_queue, line_queue_count); | |
742 | line_queue_count = 0; | |
743 | line_queue = NULL; | |
744 | } else if (line_queue_count > TRACE_COUNT) { | |
745 | line_queue = line_queue->next; | |
746 | line_queue_count --; | |
747 | } | |
748 | ||
749 | line->count = 0; | |
750 | displayed++; | |
751 | if (displayed > 300) | |
752 | break; | |
753 | line = line->next; | |
754 | } | |
755 | } | |
756 | ||
757 | /* | |
758 | * Binary search in the histogram table and record the hit: | |
759 | */ | |
760 | static void record_ip(uint64_t ip, int counter) | |
761 | { | |
762 | int left_idx, middle_idx, right_idx, idx; | |
763 | unsigned long left, middle, right; | |
764 | ||
765 | record_precise_ip(ip); | |
766 | ||
767 | left_idx = 0; | |
768 | right_idx = sym_table_count-1; | |
769 | assert(ip <= max_ip && ip >= min_ip); | |
770 | ||
771 | while (left_idx + 1 < right_idx) { | |
772 | middle_idx = (left_idx + right_idx) / 2; | |
773 | ||
774 | left = sym_table[ left_idx].addr; | |
775 | middle = sym_table[middle_idx].addr; | |
776 | right = sym_table[ right_idx].addr; | |
777 | ||
778 | if (!(left <= middle && middle <= right)) { | |
779 | printf("%016lx...\n%016lx...\n%016lx\n", left, middle, right); | |
780 | printf("%d %d %d\n", left_idx, middle_idx, right_idx); | |
781 | } | |
782 | assert(left <= middle && middle <= right); | |
783 | if (!(left <= ip && ip <= right)) { | |
784 | printf(" left: %016lx\n", left); | |
785 | printf(" ip: %016lx\n", (unsigned long)ip); | |
786 | printf("right: %016lx\n", right); | |
787 | } | |
788 | assert(left <= ip && ip <= right); | |
789 | /* | |
790 | * [ left .... target .... middle .... right ] | |
791 | * => right := middle | |
792 | */ | |
793 | if (ip < middle) { | |
794 | right_idx = middle_idx; | |
795 | continue; | |
796 | } | |
797 | /* | |
798 | * [ left .... middle ... target ... right ] | |
799 | * => left := middle | |
800 | */ | |
801 | left_idx = middle_idx; | |
802 | } | |
803 | ||
804 | idx = left_idx; | |
805 | ||
806 | if (!sym_table[idx].skip) | |
807 | sym_table[idx].count[counter]++; | |
808 | else events--; | |
809 | } | |
810 | ||
811 | static void process_event(uint64_t ip, int counter) | |
812 | { | |
813 | events++; | |
814 | ||
815 | if (ip < min_ip || ip > max_ip) { | |
816 | userspace_events++; | |
817 | return; | |
818 | } | |
819 | ||
820 | record_ip(ip, counter); | |
821 | } | |
822 | ||
6f06ccbc | 823 | static void process_options(int argc, char **argv) |
07800601 IM |
824 | { |
825 | int error = 0, counter; | |
826 | ||
07800601 IM |
827 | for (;;) { |
828 | int option_index = 0; | |
829 | /** Options for getopt */ | |
830 | static struct option long_options[] = { | |
831 | {"count", required_argument, NULL, 'c'}, | |
832 | {"cpu", required_argument, NULL, 'C'}, | |
833 | {"delay", required_argument, NULL, 'd'}, | |
834 | {"dump_symtab", no_argument, NULL, 'D'}, | |
835 | {"event", required_argument, NULL, 'e'}, | |
836 | {"filter", required_argument, NULL, 'f'}, | |
837 | {"group", required_argument, NULL, 'g'}, | |
838 | {"help", no_argument, NULL, 'h'}, | |
839 | {"nmi", required_argument, NULL, 'n'}, | |
840 | {"mmap_info", no_argument, NULL, 'M'}, | |
841 | {"mmap_pages", required_argument, NULL, 'm'}, | |
842 | {"munmap_info", no_argument, NULL, 'U'}, | |
843 | {"pid", required_argument, NULL, 'p'}, | |
844 | {"realtime", required_argument, NULL, 'r'}, | |
845 | {"scale", no_argument, NULL, 'l'}, | |
846 | {"symbol", required_argument, NULL, 's'}, | |
847 | {"stat", no_argument, NULL, 'S'}, | |
848 | {"vmlinux", required_argument, NULL, 'x'}, | |
849 | {"zero", no_argument, NULL, 'z'}, | |
f5456a6b | 850 | {"freq", required_argument, NULL, 'F'}, |
07800601 IM |
851 | {NULL, 0, NULL, 0 } |
852 | }; | |
f5456a6b | 853 | int c = getopt_long(argc, argv, "+:ac:C:d:De:f:g:hln:m:p:r:s:Sx:zMUF:", |
07800601 IM |
854 | long_options, &option_index); |
855 | if (c == -1) | |
856 | break; | |
857 | ||
858 | switch (c) { | |
859 | case 'a': system_wide = 1; break; | |
860 | case 'c': default_interval = atoi(optarg); break; | |
861 | case 'C': | |
862 | /* CPU and PID are mutually exclusive */ | |
863 | if (tid != -1) { | |
864 | printf("WARNING: CPU switch overriding PID\n"); | |
865 | sleep(1); | |
866 | tid = -1; | |
867 | } | |
868 | profile_cpu = atoi(optarg); break; | |
869 | case 'd': delay_secs = atoi(optarg); break; | |
870 | case 'D': dump_symtab = 1; break; | |
871 | ||
872 | case 'e': error = parse_events(optarg); break; | |
873 | ||
874 | case 'f': count_filter = atoi(optarg); break; | |
875 | case 'g': group = atoi(optarg); break; | |
876 | case 'h': display_help(); break; | |
877 | case 'l': scale = 1; break; | |
878 | case 'n': nmi = atoi(optarg); break; | |
879 | case 'p': | |
880 | /* CPU and PID are mutually exclusive */ | |
881 | if (profile_cpu != -1) { | |
882 | printf("WARNING: PID switch overriding CPU\n"); | |
883 | sleep(1); | |
884 | profile_cpu = -1; | |
885 | } | |
886 | tid = atoi(optarg); break; | |
887 | case 'r': realtime_prio = atoi(optarg); break; | |
888 | case 's': sym_filter = strdup(optarg); break; | |
07800601 IM |
889 | case 'x': vmlinux = strdup(optarg); break; |
890 | case 'z': zero = 1; break; | |
891 | case 'm': mmap_pages = atoi(optarg); break; | |
892 | case 'M': use_mmap = 1; break; | |
893 | case 'U': use_munmap = 1; break; | |
f5456a6b | 894 | case 'F': freq = 1; default_interval = atoi(optarg); break; |
07800601 IM |
895 | default: error = 1; break; |
896 | } | |
897 | } | |
898 | if (error) | |
899 | display_help(); | |
900 | ||
901 | if (!nr_counters) { | |
ddcacfa0 IM |
902 | nr_counters = 1; |
903 | event_id[0] = 0; | |
07800601 IM |
904 | } |
905 | ||
906 | for (counter = 0; counter < nr_counters; counter++) { | |
907 | if (event_count[counter]) | |
908 | continue; | |
909 | ||
910 | event_count[counter] = default_interval; | |
911 | } | |
912 | } | |
913 | ||
914 | struct mmap_data { | |
915 | int counter; | |
916 | void *base; | |
917 | unsigned int mask; | |
918 | unsigned int prev; | |
919 | }; | |
920 | ||
921 | static unsigned int mmap_read_head(struct mmap_data *md) | |
922 | { | |
923 | struct perf_counter_mmap_page *pc = md->base; | |
924 | int head; | |
925 | ||
926 | head = pc->data_head; | |
927 | rmb(); | |
928 | ||
929 | return head; | |
930 | } | |
931 | ||
932 | struct timeval last_read, this_read; | |
933 | ||
934 | static void mmap_read(struct mmap_data *md) | |
935 | { | |
936 | unsigned int head = mmap_read_head(md); | |
937 | unsigned int old = md->prev; | |
938 | unsigned char *data = md->base + page_size; | |
939 | int diff; | |
940 | ||
941 | gettimeofday(&this_read, NULL); | |
942 | ||
943 | /* | |
944 | * If we're further behind than half the buffer, there's a chance | |
945 | * the writer will bite our tail and screw up the events under us. | |
946 | * | |
947 | * If we somehow ended up ahead of the head, we got messed up. | |
948 | * | |
949 | * In either case, truncate and restart at head. | |
950 | */ | |
951 | diff = head - old; | |
952 | if (diff > md->mask / 2 || diff < 0) { | |
953 | struct timeval iv; | |
954 | unsigned long msecs; | |
955 | ||
956 | timersub(&this_read, &last_read, &iv); | |
957 | msecs = iv.tv_sec*1000 + iv.tv_usec/1000; | |
958 | ||
959 | fprintf(stderr, "WARNING: failed to keep up with mmap data." | |
960 | " Last read %lu msecs ago.\n", msecs); | |
961 | ||
962 | /* | |
963 | * head points to a known good entry, start there. | |
964 | */ | |
965 | old = head; | |
966 | } | |
967 | ||
968 | last_read = this_read; | |
969 | ||
970 | for (; old != head;) { | |
971 | struct ip_event { | |
972 | struct perf_event_header header; | |
973 | __u64 ip; | |
974 | __u32 pid, tid; | |
975 | }; | |
976 | struct mmap_event { | |
977 | struct perf_event_header header; | |
978 | __u32 pid, tid; | |
979 | __u64 start; | |
980 | __u64 len; | |
981 | __u64 pgoff; | |
982 | char filename[PATH_MAX]; | |
983 | }; | |
984 | ||
985 | typedef union event_union { | |
986 | struct perf_event_header header; | |
987 | struct ip_event ip; | |
988 | struct mmap_event mmap; | |
989 | } event_t; | |
990 | ||
991 | event_t *event = (event_t *)&data[old & md->mask]; | |
992 | ||
993 | event_t event_copy; | |
994 | ||
6f06ccbc | 995 | size_t size = event->header.size; |
07800601 IM |
996 | |
997 | /* | |
998 | * Event straddles the mmap boundary -- header should always | |
999 | * be inside due to u64 alignment of output. | |
1000 | */ | |
1001 | if ((old & md->mask) + size != ((old + size) & md->mask)) { | |
1002 | unsigned int offset = old; | |
1003 | unsigned int len = min(sizeof(*event), size), cpy; | |
1004 | void *dst = &event_copy; | |
1005 | ||
1006 | do { | |
1007 | cpy = min(md->mask + 1 - (offset & md->mask), len); | |
1008 | memcpy(dst, &data[offset & md->mask], cpy); | |
1009 | offset += cpy; | |
1010 | dst += cpy; | |
1011 | len -= cpy; | |
1012 | } while (len); | |
1013 | ||
1014 | event = &event_copy; | |
1015 | } | |
1016 | ||
1017 | old += size; | |
1018 | ||
1019 | if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) { | |
1020 | if (event->header.type & PERF_RECORD_IP) | |
1021 | process_event(event->ip.ip, md->counter); | |
1022 | } else { | |
1023 | switch (event->header.type) { | |
1024 | case PERF_EVENT_MMAP: | |
1025 | case PERF_EVENT_MUNMAP: | |
1026 | printf("%s: %Lu %Lu %Lu %s\n", | |
1027 | event->header.type == PERF_EVENT_MMAP | |
1028 | ? "mmap" : "munmap", | |
1029 | event->mmap.start, | |
1030 | event->mmap.len, | |
1031 | event->mmap.pgoff, | |
1032 | event->mmap.filename); | |
1033 | break; | |
1034 | } | |
1035 | } | |
1036 | } | |
1037 | ||
1038 | md->prev = old; | |
1039 | } | |
1040 | ||
6f06ccbc | 1041 | int cmd_top(int argc, char **argv, const char *prefix) |
07800601 IM |
1042 | { |
1043 | struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS]; | |
1044 | struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS]; | |
1045 | struct perf_counter_hw_event hw_event; | |
1046 | pthread_t thread; | |
1047 | int i, counter, group_fd, nr_poll = 0; | |
1048 | unsigned int cpu; | |
1049 | int ret; | |
1050 | ||
1051 | page_size = sysconf(_SC_PAGE_SIZE); | |
1052 | ||
1053 | process_options(argc, argv); | |
1054 | ||
1055 | nr_cpus = sysconf(_SC_NPROCESSORS_ONLN); | |
1056 | assert(nr_cpus <= MAX_NR_CPUS); | |
1057 | assert(nr_cpus >= 0); | |
1058 | ||
07800601 IM |
1059 | if (tid != -1 || profile_cpu != -1) |
1060 | nr_cpus = 1; | |
1061 | ||
1062 | parse_symbols(); | |
1063 | if (vmlinux && sym_filter_entry) | |
1064 | parse_vmlinux(vmlinux); | |
1065 | ||
1066 | for (i = 0; i < nr_cpus; i++) { | |
1067 | group_fd = -1; | |
1068 | for (counter = 0; counter < nr_counters; counter++) { | |
1069 | ||
1070 | cpu = profile_cpu; | |
1071 | if (tid == -1 && profile_cpu == -1) | |
1072 | cpu = i; | |
1073 | ||
1074 | memset(&hw_event, 0, sizeof(hw_event)); | |
1075 | hw_event.config = event_id[counter]; | |
1076 | hw_event.irq_period = event_count[counter]; | |
1077 | hw_event.record_type = PERF_RECORD_IP | PERF_RECORD_TID; | |
1078 | hw_event.nmi = nmi; | |
1079 | hw_event.mmap = use_mmap; | |
1080 | hw_event.munmap = use_munmap; | |
f5456a6b | 1081 | hw_event.freq = freq; |
07800601 IM |
1082 | |
1083 | fd[i][counter] = sys_perf_counter_open(&hw_event, tid, cpu, group_fd, 0); | |
1084 | if (fd[i][counter] < 0) { | |
1085 | int err = errno; | |
1086 | printf("kerneltop error: syscall returned with %d (%s)\n", | |
1087 | fd[i][counter], strerror(err)); | |
1088 | if (err == EPERM) | |
1089 | printf("Are you root?\n"); | |
1090 | exit(-1); | |
1091 | } | |
1092 | assert(fd[i][counter] >= 0); | |
1093 | fcntl(fd[i][counter], F_SETFL, O_NONBLOCK); | |
1094 | ||
1095 | /* | |
1096 | * First counter acts as the group leader: | |
1097 | */ | |
1098 | if (group && group_fd == -1) | |
1099 | group_fd = fd[i][counter]; | |
1100 | ||
1101 | event_array[nr_poll].fd = fd[i][counter]; | |
1102 | event_array[nr_poll].events = POLLIN; | |
1103 | nr_poll++; | |
1104 | ||
1105 | mmap_array[i][counter].counter = counter; | |
1106 | mmap_array[i][counter].prev = 0; | |
1107 | mmap_array[i][counter].mask = mmap_pages*page_size - 1; | |
1108 | mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size, | |
1109 | PROT_READ, MAP_SHARED, fd[i][counter], 0); | |
1110 | if (mmap_array[i][counter].base == MAP_FAILED) { | |
1111 | printf("kerneltop error: failed to mmap with %d (%s)\n", | |
1112 | errno, strerror(errno)); | |
1113 | exit(-1); | |
1114 | } | |
1115 | } | |
1116 | } | |
1117 | ||
1118 | if (pthread_create(&thread, NULL, display_thread, NULL)) { | |
1119 | printf("Could not create display thread.\n"); | |
1120 | exit(-1); | |
1121 | } | |
1122 | ||
1123 | if (realtime_prio) { | |
1124 | struct sched_param param; | |
1125 | ||
1126 | param.sched_priority = realtime_prio; | |
1127 | if (sched_setscheduler(0, SCHED_FIFO, ¶m)) { | |
1128 | printf("Could not set realtime priority.\n"); | |
1129 | exit(-1); | |
1130 | } | |
1131 | } | |
1132 | ||
1133 | while (1) { | |
1134 | int hits = events; | |
1135 | ||
1136 | for (i = 0; i < nr_cpus; i++) { | |
1137 | for (counter = 0; counter < nr_counters; counter++) | |
1138 | mmap_read(&mmap_array[i][counter]); | |
1139 | } | |
1140 | ||
1141 | if (hits == events) | |
1142 | ret = poll(event_array, nr_poll, 100); | |
1143 | } | |
1144 | ||
1145 | return 0; | |
1146 | } |