aboutsummaryrefslogtreecommitdiffstats
path: root/tools/perf
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2009-06-30 22:02:59 -0400
committerLinus Torvalds <torvalds@linux-foundation.org>2009-06-30 22:02:59 -0400
commit55bcab46955644d5a8149a9b3cc9752a336e02f8 (patch)
treea45593603de9265cb0624043ab36f4e0f98823fc /tools/perf
parent58580c86450bc09ff101f0d23fd8a162c146bc64 (diff)
parent7bec7a9134c25cecb0d7029199b59f7b1bef35b8 (diff)
Merge branch 'perfcounters-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip
* 'perfcounters-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (47 commits) perf report: Add --symbols parameter perf report: Add --comms parameter perf report: Add --dsos parameter perf_counter tools: Adjust only prelinked symbol's addresses perf_counter: Provide a way to enable counters on exec perf_counter tools: Reduce perf stat measurement overhead/skew perf stat: Use percentages for scaling output perf_counter, x86: Update x86_pmu after WARN() perf stat: Micro-optimize the code: memcpy is only required if no event is selected and !null_run perf stat: Improve output perf stat: Fix multi-run stats perf stat: Add -n/--null option to run without counters perf_counter tools: Remove dead code perf_counter: Complete counter swap perf report: Print sorted callchains per histogram entries perf_counter tools: Prepare a small callchain framework perf record: Fix unhandled io return value perf_counter tools: Add alias for 'l1d' and 'l1i' perf-report: Add bare minimum PERF_EVENT_READ parsing perf-report: Add modes for inherited stats and no-samples ...
Diffstat (limited to 'tools/perf')
-rw-r--r--tools/perf/CREDITS30
-rw-r--r--tools/perf/Documentation/perf-report.txt14
-rw-r--r--tools/perf/Documentation/perf-stat.txt6
-rw-r--r--tools/perf/Makefile6
-rw-r--r--tools/perf/builtin-annotate.c8
-rw-r--r--tools/perf/builtin-record.c127
-rw-r--r--tools/perf/builtin-report.c236
-rw-r--r--tools/perf/builtin-stat.c171
-rw-r--r--tools/perf/builtin-top.c11
-rw-r--r--tools/perf/perf.h9
-rw-r--r--tools/perf/util/callchain.c174
-rw-r--r--tools/perf/util/callchain.h33
-rw-r--r--tools/perf/util/header.c242
-rw-r--r--tools/perf/util/header.h37
-rw-r--r--tools/perf/util/help.c15
-rw-r--r--tools/perf/util/pager.c5
-rw-r--r--tools/perf/util/parse-events.c153
-rw-r--r--tools/perf/util/run-command.c95
-rw-r--r--tools/perf/util/run-command.h5
-rw-r--r--tools/perf/util/strbuf.c2
-rw-r--r--tools/perf/util/string.h2
-rw-r--r--tools/perf/util/strlist.c184
-rw-r--r--tools/perf/util/strlist.h32
-rw-r--r--tools/perf/util/symbol.c16
-rw-r--r--tools/perf/util/symbol.h5
-rw-r--r--tools/perf/util/types.h (renamed from tools/perf/types.h)0
-rw-r--r--tools/perf/util/util.h15
27 files changed, 1279 insertions, 354 deletions
diff --git a/tools/perf/CREDITS b/tools/perf/CREDITS
new file mode 100644
index 000000000000..c2ddcb3acbd0
--- /dev/null
+++ b/tools/perf/CREDITS
@@ -0,0 +1,30 @@
1Most of the infrastructure that 'perf' uses here has been reused
2from the Git project, as of version:
3
4 66996ec: Sync with 1.6.2.4
5
6Here is an (incomplete!) list of main contributors to those files
7in util/* and elsewhere:
8
9 Alex Riesen
10 Christian Couder
11 Dmitry Potapov
12 Jeff King
13 Johannes Schindelin
14 Johannes Sixt
15 Junio C Hamano
16 Linus Torvalds
17 Matthias Kestenholz
18 Michal Ostrowski
19 Miklos Vajna
20 Petr Baudis
21 Pierre Habouzit
22 René Scharfe
23 Samuel Tardieu
24 Shawn O. Pearce
25 Steffen Prohaska
26 Steve Haslam
27
28Thanks guys!
29
30The full history of the files can be found in the upstream Git commits.
diff --git a/tools/perf/Documentation/perf-report.txt b/tools/perf/Documentation/perf-report.txt
index 52d3fc6846a9..8aa3f8c88707 100644
--- a/tools/perf/Documentation/perf-report.txt
+++ b/tools/perf/Documentation/perf-report.txt
@@ -13,13 +13,25 @@ SYNOPSIS
13DESCRIPTION 13DESCRIPTION
14----------- 14-----------
15This command displays the performance counter profile information recorded 15This command displays the performance counter profile information recorded
16via perf report. 16via perf record.
17 17
18OPTIONS 18OPTIONS
19------- 19-------
20-i:: 20-i::
21--input=:: 21--input=::
22 Input file name. (default: perf.data) 22 Input file name. (default: perf.data)
23-d::
24--dsos=::
25 Only consider symbols in these dsos. CSV that understands
26 file://filename entries.
27-C::
28--comms=::
29 Only consider symbols in these comms. CSV that understands
30 file://filename entries.
31-S::
32--symbols=::
33 Only consider these symbols. CSV that understands
34 file://filename entries.
23 35
24SEE ALSO 36SEE ALSO
25-------- 37--------
diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt
index c368a72721d7..0d74346d21ab 100644
--- a/tools/perf/Documentation/perf-stat.txt
+++ b/tools/perf/Documentation/perf-stat.txt
@@ -8,8 +8,8 @@ perf-stat - Run a command and gather performance counter statistics
8SYNOPSIS 8SYNOPSIS
9-------- 9--------
10[verse] 10[verse]
11'perf stat' [-e <EVENT> | --event=EVENT] [-l] [-a] <command> 11'perf stat' [-e <EVENT> | --event=EVENT] [-S] [-a] <command>
12'perf stat' [-e <EVENT> | --event=EVENT] [-l] [-a] -- <command> [<options>] 12'perf stat' [-e <EVENT> | --event=EVENT] [-S] [-a] -- <command> [<options>]
13 13
14DESCRIPTION 14DESCRIPTION
15----------- 15-----------
@@ -40,7 +40,7 @@ OPTIONS
40-a:: 40-a::
41 system-wide collection 41 system-wide collection
42 42
43-l:: 43-S::
44 scale counter values 44 scale counter values
45 45
46EXAMPLES 46EXAMPLES
diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index 36d7eef49913..9c6d0ae3708e 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -290,7 +290,7 @@ LIB_FILE=libperf.a
290 290
291LIB_H += ../../include/linux/perf_counter.h 291LIB_H += ../../include/linux/perf_counter.h
292LIB_H += perf.h 292LIB_H += perf.h
293LIB_H += types.h 293LIB_H += util/types.h
294LIB_H += util/list.h 294LIB_H += util/list.h
295LIB_H += util/rbtree.h 295LIB_H += util/rbtree.h
296LIB_H += util/levenshtein.h 296LIB_H += util/levenshtein.h
@@ -301,6 +301,7 @@ LIB_H += util/util.h
301LIB_H += util/help.h 301LIB_H += util/help.h
302LIB_H += util/strbuf.h 302LIB_H += util/strbuf.h
303LIB_H += util/string.h 303LIB_H += util/string.h
304LIB_H += util/strlist.h
304LIB_H += util/run-command.h 305LIB_H += util/run-command.h
305LIB_H += util/sigchain.h 306LIB_H += util/sigchain.h
306LIB_H += util/symbol.h 307LIB_H += util/symbol.h
@@ -322,12 +323,15 @@ LIB_OBJS += util/run-command.o
322LIB_OBJS += util/quote.o 323LIB_OBJS += util/quote.o
323LIB_OBJS += util/strbuf.o 324LIB_OBJS += util/strbuf.o
324LIB_OBJS += util/string.o 325LIB_OBJS += util/string.o
326LIB_OBJS += util/strlist.o
325LIB_OBJS += util/usage.o 327LIB_OBJS += util/usage.o
326LIB_OBJS += util/wrapper.o 328LIB_OBJS += util/wrapper.o
327LIB_OBJS += util/sigchain.o 329LIB_OBJS += util/sigchain.o
328LIB_OBJS += util/symbol.o 330LIB_OBJS += util/symbol.o
329LIB_OBJS += util/color.o 331LIB_OBJS += util/color.o
330LIB_OBJS += util/pager.o 332LIB_OBJS += util/pager.o
333LIB_OBJS += util/header.o
334LIB_OBJS += util/callchain.o
331 335
332BUILTIN_OBJS += builtin-annotate.o 336BUILTIN_OBJS += builtin-annotate.o
333BUILTIN_OBJS += builtin-help.o 337BUILTIN_OBJS += builtin-help.o
diff --git a/tools/perf/builtin-annotate.c b/tools/perf/builtin-annotate.c
index 7e58e3ad1508..722c0f54e549 100644
--- a/tools/perf/builtin-annotate.c
+++ b/tools/perf/builtin-annotate.c
@@ -855,7 +855,7 @@ static unsigned long total = 0,
855 total_unknown = 0; 855 total_unknown = 0;
856 856
857static int 857static int
858process_overflow_event(event_t *event, unsigned long offset, unsigned long head) 858process_sample_event(event_t *event, unsigned long offset, unsigned long head)
859{ 859{
860 char level; 860 char level;
861 int show = 0; 861 int show = 0;
@@ -1013,10 +1013,10 @@ process_period_event(event_t *event, unsigned long offset, unsigned long head)
1013static int 1013static int
1014process_event(event_t *event, unsigned long offset, unsigned long head) 1014process_event(event_t *event, unsigned long offset, unsigned long head)
1015{ 1015{
1016 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1017 return process_overflow_event(event, offset, head);
1018
1019 switch (event->header.type) { 1016 switch (event->header.type) {
1017 case PERF_EVENT_SAMPLE:
1018 return process_sample_event(event, offset, head);
1019
1020 case PERF_EVENT_MMAP: 1020 case PERF_EVENT_MMAP:
1021 return process_mmap_event(event, offset, head); 1021 return process_mmap_event(event, offset, head);
1022 1022
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index d7ebbd757543..d18546f37d7c 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -14,6 +14,8 @@
14#include "util/parse-events.h" 14#include "util/parse-events.h"
15#include "util/string.h" 15#include "util/string.h"
16 16
17#include "util/header.h"
18
17#include <unistd.h> 19#include <unistd.h>
18#include <sched.h> 20#include <sched.h>
19 21
@@ -39,6 +41,8 @@ static int force = 0;
39static int append_file = 0; 41static int append_file = 0;
40static int call_graph = 0; 42static int call_graph = 0;
41static int verbose = 0; 43static int verbose = 0;
44static int inherit_stat = 0;
45static int no_samples = 0;
42 46
43static long samples; 47static long samples;
44static struct timeval last_read; 48static struct timeval last_read;
@@ -52,7 +56,8 @@ static int nr_poll;
52static int nr_cpu; 56static int nr_cpu;
53 57
54static int file_new = 1; 58static int file_new = 1;
55static struct perf_file_header file_header; 59
60struct perf_header *header;
56 61
57struct mmap_event { 62struct mmap_event {
58 struct perf_event_header header; 63 struct perf_event_header header;
@@ -306,12 +311,11 @@ static void pid_synthesize_mmap_samples(pid_t pid)
306 continue; 311 continue;
307 pbf += n + 3; 312 pbf += n + 3;
308 if (*pbf == 'x') { /* vm_exec */ 313 if (*pbf == 'x') { /* vm_exec */
309 char *execname = strrchr(bf, ' '); 314 char *execname = strchr(bf, '/');
310 315
311 if (execname == NULL || execname[1] != '/') 316 if (execname == NULL)
312 continue; 317 continue;
313 318
314 execname += 1;
315 size = strlen(execname); 319 size = strlen(execname);
316 execname[size - 1] = '\0'; /* Remove \n */ 320 execname[size - 1] = '\0'; /* Remove \n */
317 memcpy(mmap_ev.filename, execname, size); 321 memcpy(mmap_ev.filename, execname, size);
@@ -329,7 +333,7 @@ static void pid_synthesize_mmap_samples(pid_t pid)
329 fclose(fp); 333 fclose(fp);
330} 334}
331 335
332static void synthesize_samples(void) 336static void synthesize_all(void)
333{ 337{
334 DIR *proc; 338 DIR *proc;
335 struct dirent dirent, *next; 339 struct dirent dirent, *next;
@@ -353,10 +357,35 @@ static void synthesize_samples(void)
353 357
354static int group_fd; 358static int group_fd;
355 359
360static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
361{
362 struct perf_header_attr *h_attr;
363
364 if (nr < header->attrs) {
365 h_attr = header->attr[nr];
366 } else {
367 h_attr = perf_header_attr__new(a);
368 perf_header__add_attr(header, h_attr);
369 }
370
371 return h_attr;
372}
373
356static void create_counter(int counter, int cpu, pid_t pid) 374static void create_counter(int counter, int cpu, pid_t pid)
357{ 375{
358 struct perf_counter_attr *attr = attrs + counter; 376 struct perf_counter_attr *attr = attrs + counter;
359 int track = 1; 377 struct perf_header_attr *h_attr;
378 int track = !counter; /* only the first counter needs these */
379 struct {
380 u64 count;
381 u64 time_enabled;
382 u64 time_running;
383 u64 id;
384 } read_data;
385
386 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
387 PERF_FORMAT_TOTAL_TIME_RUNNING |
388 PERF_FORMAT_ID;
360 389
361 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID; 390 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
362 391
@@ -366,25 +395,20 @@ static void create_counter(int counter, int cpu, pid_t pid)
366 attr->sample_freq = freq; 395 attr->sample_freq = freq;
367 } 396 }
368 397
398 if (no_samples)
399 attr->sample_freq = 0;
400
401 if (inherit_stat)
402 attr->inherit_stat = 1;
403
369 if (call_graph) 404 if (call_graph)
370 attr->sample_type |= PERF_SAMPLE_CALLCHAIN; 405 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
371 406
372 if (file_new) {
373 file_header.sample_type = attr->sample_type;
374 } else {
375 if (file_header.sample_type != attr->sample_type) {
376 fprintf(stderr, "incompatible append\n");
377 exit(-1);
378 }
379 }
380
381 attr->mmap = track; 407 attr->mmap = track;
382 attr->comm = track; 408 attr->comm = track;
383 attr->inherit = (cpu < 0) && inherit; 409 attr->inherit = (cpu < 0) && inherit;
384 attr->disabled = 1; 410 attr->disabled = 1;
385 411
386 track = 0; /* only the first counter needs these */
387
388try_again: 412try_again:
389 fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0); 413 fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
390 414
@@ -415,6 +439,22 @@ try_again:
415 exit(-1); 439 exit(-1);
416 } 440 }
417 441
442 h_attr = get_header_attr(attr, counter);
443
444 if (!file_new) {
445 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
446 fprintf(stderr, "incompatible append\n");
447 exit(-1);
448 }
449 }
450
451 if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
452 perror("Unable to read perf file descriptor\n");
453 exit(-1);
454 }
455
456 perf_header_attr__add_id(h_attr, read_data.id);
457
418 assert(fd[nr_cpu][counter] >= 0); 458 assert(fd[nr_cpu][counter] >= 0);
419 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK); 459 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
420 460
@@ -445,11 +485,6 @@ static void open_counters(int cpu, pid_t pid)
445{ 485{
446 int counter; 486 int counter;
447 487
448 if (pid > 0) {
449 pid_synthesize_comm_event(pid, 0);
450 pid_synthesize_mmap_samples(pid);
451 }
452
453 group_fd = -1; 488 group_fd = -1;
454 for (counter = 0; counter < nr_counters; counter++) 489 for (counter = 0; counter < nr_counters; counter++)
455 create_counter(counter, cpu, pid); 490 create_counter(counter, cpu, pid);
@@ -459,17 +494,16 @@ static void open_counters(int cpu, pid_t pid)
459 494
460static void atexit_header(void) 495static void atexit_header(void)
461{ 496{
462 file_header.data_size += bytes_written; 497 header->data_size += bytes_written;
463 498
464 if (pwrite(output, &file_header, sizeof(file_header), 0) == -1) 499 perf_header__write(header, output);
465 perror("failed to write on file headers");
466} 500}
467 501
468static int __cmd_record(int argc, const char **argv) 502static int __cmd_record(int argc, const char **argv)
469{ 503{
470 int i, counter; 504 int i, counter;
471 struct stat st; 505 struct stat st;
472 pid_t pid; 506 pid_t pid = 0;
473 int flags; 507 int flags;
474 int ret; 508 int ret;
475 509
@@ -500,22 +534,31 @@ static int __cmd_record(int argc, const char **argv)
500 exit(-1); 534 exit(-1);
501 } 535 }
502 536
503 if (!file_new) { 537 if (!file_new)
504 if (read(output, &file_header, sizeof(file_header)) == -1) { 538 header = perf_header__read(output);
505 perror("failed to read file headers"); 539 else
506 exit(-1); 540 header = perf_header__new();
507 }
508
509 lseek(output, file_header.data_size, SEEK_CUR);
510 }
511 541
512 atexit(atexit_header); 542 atexit(atexit_header);
513 543
514 if (!system_wide) { 544 if (!system_wide) {
515 open_counters(-1, target_pid != -1 ? target_pid : getpid()); 545 pid = target_pid;
546 if (pid == -1)
547 pid = getpid();
548
549 open_counters(-1, pid);
516 } else for (i = 0; i < nr_cpus; i++) 550 } else for (i = 0; i < nr_cpus; i++)
517 open_counters(i, target_pid); 551 open_counters(i, target_pid);
518 552
553 if (file_new)
554 perf_header__write(header, output);
555
556 if (!system_wide) {
557 pid_synthesize_comm_event(pid, 0);
558 pid_synthesize_mmap_samples(pid);
559 } else
560 synthesize_all();
561
519 if (target_pid == -1 && argc) { 562 if (target_pid == -1 && argc) {
520 pid = fork(); 563 pid = fork();
521 if (pid < 0) 564 if (pid < 0)
@@ -539,10 +582,7 @@ static int __cmd_record(int argc, const char **argv)
539 } 582 }
540 } 583 }
541 584
542 if (system_wide) 585 for (;;) {
543 synthesize_samples();
544
545 while (!done) {
546 int hits = samples; 586 int hits = samples;
547 587
548 for (i = 0; i < nr_cpu; i++) { 588 for (i = 0; i < nr_cpu; i++) {
@@ -550,8 +590,11 @@ static int __cmd_record(int argc, const char **argv)
550 mmap_read(&mmap_array[i][counter]); 590 mmap_read(&mmap_array[i][counter]);
551 } 591 }
552 592
553 if (hits == samples) 593 if (hits == samples) {
594 if (done)
595 break;
554 ret = poll(event_array, nr_poll, 100); 596 ret = poll(event_array, nr_poll, 100);
597 }
555 } 598 }
556 599
557 /* 600 /*
@@ -600,6 +643,10 @@ static const struct option options[] = {
600 "do call-graph (stack chain/backtrace) recording"), 643 "do call-graph (stack chain/backtrace) recording"),
601 OPT_BOOLEAN('v', "verbose", &verbose, 644 OPT_BOOLEAN('v', "verbose", &verbose,
602 "be more verbose (show counter open errors, etc)"), 645 "be more verbose (show counter open errors, etc)"),
646 OPT_BOOLEAN('s', "stat", &inherit_stat,
647 "per thread counts"),
648 OPT_BOOLEAN('n', "no-samples", &no_samples,
649 "don't sample"),
603 OPT_END() 650 OPT_END()
604}; 651};
605 652
diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
index 5eb5566f0c95..135b7837e6bf 100644
--- a/tools/perf/builtin-report.c
+++ b/tools/perf/builtin-report.c
@@ -15,8 +15,11 @@
15#include "util/rbtree.h" 15#include "util/rbtree.h"
16#include "util/symbol.h" 16#include "util/symbol.h"
17#include "util/string.h" 17#include "util/string.h"
18#include "util/callchain.h"
19#include "util/strlist.h"
18 20
19#include "perf.h" 21#include "perf.h"
22#include "util/header.h"
20 23
21#include "util/parse-options.h" 24#include "util/parse-options.h"
22#include "util/parse-events.h" 25#include "util/parse-events.h"
@@ -30,6 +33,8 @@ static char *vmlinux = NULL;
30 33
31static char default_sort_order[] = "comm,dso"; 34static char default_sort_order[] = "comm,dso";
32static char *sort_order = default_sort_order; 35static char *sort_order = default_sort_order;
36static char *dso_list_str, *comm_list_str, *sym_list_str;
37static struct strlist *dso_list, *comm_list, *sym_list;
33 38
34static int input; 39static int input;
35static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV; 40static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
@@ -51,6 +56,9 @@ static char *parent_pattern = default_parent_pattern;
51static regex_t parent_regex; 56static regex_t parent_regex;
52 57
53static int exclude_other = 1; 58static int exclude_other = 1;
59static int callchain;
60
61static u64 sample_type;
54 62
55struct ip_event { 63struct ip_event {
56 struct perf_event_header header; 64 struct perf_event_header header;
@@ -59,11 +67,6 @@ struct ip_event {
59 unsigned char __more_data[]; 67 unsigned char __more_data[];
60}; 68};
61 69
62struct ip_callchain {
63 u64 nr;
64 u64 ips[0];
65};
66
67struct mmap_event { 70struct mmap_event {
68 struct perf_event_header header; 71 struct perf_event_header header;
69 u32 pid, tid; 72 u32 pid, tid;
@@ -97,6 +100,13 @@ struct lost_event {
97 u64 lost; 100 u64 lost;
98}; 101};
99 102
103struct read_event {
104 struct perf_event_header header;
105 u32 pid,tid;
106 u64 value;
107 u64 format[3];
108};
109
100typedef union event_union { 110typedef union event_union {
101 struct perf_event_header header; 111 struct perf_event_header header;
102 struct ip_event ip; 112 struct ip_event ip;
@@ -105,6 +115,7 @@ typedef union event_union {
105 struct fork_event fork; 115 struct fork_event fork;
106 struct period_event period; 116 struct period_event period;
107 struct lost_event lost; 117 struct lost_event lost;
118 struct read_event read;
108} event_t; 119} event_t;
109 120
110static LIST_HEAD(dsos); 121static LIST_HEAD(dsos);
@@ -229,7 +240,7 @@ static u64 vdso__map_ip(struct map *map, u64 ip)
229 240
230static inline int is_anon_memory(const char *filename) 241static inline int is_anon_memory(const char *filename)
231{ 242{
232 return strcmp(filename, "//anon") == 0; 243 return strcmp(filename, "//anon") == 0;
233} 244}
234 245
235static struct map *map__new(struct mmap_event *event) 246static struct map *map__new(struct mmap_event *event)
@@ -400,9 +411,27 @@ static void thread__insert_map(struct thread *self, struct map *map)
400 411
401 list_for_each_entry_safe(pos, tmp, &self->maps, node) { 412 list_for_each_entry_safe(pos, tmp, &self->maps, node) {
402 if (map__overlap(pos, map)) { 413 if (map__overlap(pos, map)) {
403 list_del_init(&pos->node); 414 if (verbose >= 2) {
404 /* XXX leaks dsos */ 415 printf("overlapping maps:\n");
405 free(pos); 416 map__fprintf(map, stdout);
417 map__fprintf(pos, stdout);
418 }
419
420 if (map->start <= pos->start && map->end > pos->start)
421 pos->start = map->end;
422
423 if (map->end >= pos->end && map->start < pos->end)
424 pos->end = map->start;
425
426 if (verbose >= 2) {
427 printf("after collision:\n");
428 map__fprintf(pos, stdout);
429 }
430
431 if (pos->start >= pos->end) {
432 list_del_init(&pos->node);
433 free(pos);
434 }
406 } 435 }
407 } 436 }
408 437
@@ -464,17 +493,19 @@ static size_t threads__fprintf(FILE *fp)
464static struct rb_root hist; 493static struct rb_root hist;
465 494
466struct hist_entry { 495struct hist_entry {
467 struct rb_node rb_node; 496 struct rb_node rb_node;
468 497
469 struct thread *thread; 498 struct thread *thread;
470 struct map *map; 499 struct map *map;
471 struct dso *dso; 500 struct dso *dso;
472 struct symbol *sym; 501 struct symbol *sym;
473 struct symbol *parent; 502 struct symbol *parent;
474 u64 ip; 503 u64 ip;
475 char level; 504 char level;
476 505 struct callchain_node callchain;
477 u64 count; 506 struct rb_root sorted_chain;
507
508 u64 count;
478}; 509};
479 510
480/* 511/*
@@ -745,6 +776,48 @@ hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
745} 776}
746 777
747static size_t 778static size_t
779callchain__fprintf(FILE *fp, struct callchain_node *self, u64 total_samples)
780{
781 struct callchain_list *chain;
782 size_t ret = 0;
783
784 if (!self)
785 return 0;
786
787 ret += callchain__fprintf(fp, self->parent, total_samples);
788
789
790 list_for_each_entry(chain, &self->val, list)
791 ret += fprintf(fp, " %p\n", (void *)chain->ip);
792
793 return ret;
794}
795
796static size_t
797hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
798 u64 total_samples)
799{
800 struct rb_node *rb_node;
801 struct callchain_node *chain;
802 size_t ret = 0;
803
804 rb_node = rb_first(&self->sorted_chain);
805 while (rb_node) {
806 double percent;
807
808 chain = rb_entry(rb_node, struct callchain_node, rb_node);
809 percent = chain->hit * 100.0 / total_samples;
810 ret += fprintf(fp, " %6.2f%%\n", percent);
811 ret += callchain__fprintf(fp, chain, total_samples);
812 ret += fprintf(fp, "\n");
813 rb_node = rb_next(rb_node);
814 }
815
816 return ret;
817}
818
819
820static size_t
748hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples) 821hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
749{ 822{
750 struct sort_entry *se; 823 struct sort_entry *se;
@@ -784,6 +857,9 @@ hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
784 857
785 ret += fprintf(fp, "\n"); 858 ret += fprintf(fp, "\n");
786 859
860 if (callchain)
861 hist_entry_callchain__fprintf(fp, self, total_samples);
862
787 return ret; 863 return ret;
788} 864}
789 865
@@ -797,7 +873,7 @@ resolve_symbol(struct thread *thread, struct map **mapp,
797{ 873{
798 struct dso *dso = dsop ? *dsop : NULL; 874 struct dso *dso = dsop ? *dsop : NULL;
799 struct map *map = mapp ? *mapp : NULL; 875 struct map *map = mapp ? *mapp : NULL;
800 uint64_t ip = *ipp; 876 u64 ip = *ipp;
801 877
802 if (!thread) 878 if (!thread)
803 return NULL; 879 return NULL;
@@ -814,7 +890,6 @@ resolve_symbol(struct thread *thread, struct map **mapp,
814 *mapp = map; 890 *mapp = map;
815got_map: 891got_map:
816 ip = map->map_ip(map, ip); 892 ip = map->map_ip(map, ip);
817 *ipp = ip;
818 893
819 dso = map->dso; 894 dso = map->dso;
820 } else { 895 } else {
@@ -828,6 +903,8 @@ got_map:
828 dso = kernel_dso; 903 dso = kernel_dso;
829 } 904 }
830 dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>"); 905 dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
906 dprintf(" ...... map: %Lx -> %Lx\n", *ipp, ip);
907 *ipp = ip;
831 908
832 if (dsop) 909 if (dsop)
833 *dsop = dso; 910 *dsop = dso;
@@ -867,6 +944,7 @@ hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
867 .level = level, 944 .level = level,
868 .count = count, 945 .count = count,
869 .parent = NULL, 946 .parent = NULL,
947 .sorted_chain = RB_ROOT
870 }; 948 };
871 int cmp; 949 int cmp;
872 950
@@ -909,6 +987,8 @@ hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
909 987
910 if (!cmp) { 988 if (!cmp) {
911 he->count += count; 989 he->count += count;
990 if (callchain)
991 append_chain(&he->callchain, chain);
912 return 0; 992 return 0;
913 } 993 }
914 994
@@ -922,6 +1002,10 @@ hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
922 if (!he) 1002 if (!he)
923 return -ENOMEM; 1003 return -ENOMEM;
924 *he = entry; 1004 *he = entry;
1005 if (callchain) {
1006 callchain_init(&he->callchain);
1007 append_chain(&he->callchain, chain);
1008 }
925 rb_link_node(&he->rb_node, parent, p); 1009 rb_link_node(&he->rb_node, parent, p);
926 rb_insert_color(&he->rb_node, &hist); 1010 rb_insert_color(&he->rb_node, &hist);
927 1011
@@ -998,6 +1082,9 @@ static void output__insert_entry(struct hist_entry *he)
998 struct rb_node *parent = NULL; 1082 struct rb_node *parent = NULL;
999 struct hist_entry *iter; 1083 struct hist_entry *iter;
1000 1084
1085 if (callchain)
1086 sort_chain_to_rbtree(&he->sorted_chain, &he->callchain);
1087
1001 while (*p != NULL) { 1088 while (*p != NULL) {
1002 parent = *p; 1089 parent = *p;
1003 iter = rb_entry(parent, struct hist_entry, rb_node); 1090 iter = rb_entry(parent, struct hist_entry, rb_node);
@@ -1115,7 +1202,7 @@ static int validate_chain(struct ip_callchain *chain, event_t *event)
1115} 1202}
1116 1203
1117static int 1204static int
1118process_overflow_event(event_t *event, unsigned long offset, unsigned long head) 1205process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1119{ 1206{
1120 char level; 1207 char level;
1121 int show = 0; 1208 int show = 0;
@@ -1127,12 +1214,12 @@ process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1127 void *more_data = event->ip.__more_data; 1214 void *more_data = event->ip.__more_data;
1128 struct ip_callchain *chain = NULL; 1215 struct ip_callchain *chain = NULL;
1129 1216
1130 if (event->header.type & PERF_SAMPLE_PERIOD) { 1217 if (sample_type & PERF_SAMPLE_PERIOD) {
1131 period = *(u64 *)more_data; 1218 period = *(u64 *)more_data;
1132 more_data += sizeof(u64); 1219 more_data += sizeof(u64);
1133 } 1220 }
1134 1221
1135 dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p period: %Ld\n", 1222 dprintf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d: %p period: %Ld\n",
1136 (void *)(offset + head), 1223 (void *)(offset + head),
1137 (void *)(long)(event->header.size), 1224 (void *)(long)(event->header.size),
1138 event->header.misc, 1225 event->header.misc,
@@ -1140,7 +1227,7 @@ process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1140 (void *)(long)ip, 1227 (void *)(long)ip,
1141 (long long)period); 1228 (long long)period);
1142 1229
1143 if (event->header.type & PERF_SAMPLE_CALLCHAIN) { 1230 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
1144 int i; 1231 int i;
1145 1232
1146 chain = (void *)more_data; 1233 chain = (void *)more_data;
@@ -1166,6 +1253,9 @@ process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1166 return -1; 1253 return -1;
1167 } 1254 }
1168 1255
1256 if (comm_list && !strlist__has_entry(comm_list, thread->comm))
1257 return 0;
1258
1169 if (event->header.misc & PERF_EVENT_MISC_KERNEL) { 1259 if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
1170 show = SHOW_KERNEL; 1260 show = SHOW_KERNEL;
1171 level = 'k'; 1261 level = 'k';
@@ -1188,6 +1278,12 @@ process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1188 if (show & show_mask) { 1278 if (show & show_mask) {
1189 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip); 1279 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1190 1280
1281 if (dso_list && dso && dso->name && !strlist__has_entry(dso_list, dso->name))
1282 return 0;
1283
1284 if (sym_list && sym && !strlist__has_entry(sym_list, sym->name))
1285 return 0;
1286
1191 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) { 1287 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1192 eprintf("problem incrementing symbol count, skipping event\n"); 1288 eprintf("problem incrementing symbol count, skipping event\n");
1193 return -1; 1289 return -1;
@@ -1328,14 +1424,27 @@ static void trace_event(event_t *event)
1328} 1424}
1329 1425
1330static int 1426static int
1427process_read_event(event_t *event, unsigned long offset, unsigned long head)
1428{
1429 dprintf("%p [%p]: PERF_EVENT_READ: %d %d %Lu\n",
1430 (void *)(offset + head),
1431 (void *)(long)(event->header.size),
1432 event->read.pid,
1433 event->read.tid,
1434 event->read.value);
1435
1436 return 0;
1437}
1438
1439static int
1331process_event(event_t *event, unsigned long offset, unsigned long head) 1440process_event(event_t *event, unsigned long offset, unsigned long head)
1332{ 1441{
1333 trace_event(event); 1442 trace_event(event);
1334 1443
1335 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1336 return process_overflow_event(event, offset, head);
1337
1338 switch (event->header.type) { 1444 switch (event->header.type) {
1445 case PERF_EVENT_SAMPLE:
1446 return process_sample_event(event, offset, head);
1447
1339 case PERF_EVENT_MMAP: 1448 case PERF_EVENT_MMAP:
1340 return process_mmap_event(event, offset, head); 1449 return process_mmap_event(event, offset, head);
1341 1450
@@ -1351,6 +1460,9 @@ process_event(event_t *event, unsigned long offset, unsigned long head)
1351 case PERF_EVENT_LOST: 1460 case PERF_EVENT_LOST:
1352 return process_lost_event(event, offset, head); 1461 return process_lost_event(event, offset, head);
1353 1462
1463 case PERF_EVENT_READ:
1464 return process_read_event(event, offset, head);
1465
1354 /* 1466 /*
1355 * We dont process them right now but they are fine: 1467 * We dont process them right now but they are fine:
1356 */ 1468 */
@@ -1366,13 +1478,30 @@ process_event(event_t *event, unsigned long offset, unsigned long head)
1366 return 0; 1478 return 0;
1367} 1479}
1368 1480
1369static struct perf_file_header file_header; 1481static struct perf_header *header;
1482
1483static u64 perf_header__sample_type(void)
1484{
1485 u64 sample_type = 0;
1486 int i;
1487
1488 for (i = 0; i < header->attrs; i++) {
1489 struct perf_header_attr *attr = header->attr[i];
1490
1491 if (!sample_type)
1492 sample_type = attr->attr.sample_type;
1493 else if (sample_type != attr->attr.sample_type)
1494 die("non matching sample_type");
1495 }
1496
1497 return sample_type;
1498}
1370 1499
1371static int __cmd_report(void) 1500static int __cmd_report(void)
1372{ 1501{
1373 int ret, rc = EXIT_FAILURE; 1502 int ret, rc = EXIT_FAILURE;
1374 unsigned long offset = 0; 1503 unsigned long offset = 0;
1375 unsigned long head = sizeof(file_header); 1504 unsigned long head, shift;
1376 struct stat stat; 1505 struct stat stat;
1377 event_t *event; 1506 event_t *event;
1378 uint32_t size; 1507 uint32_t size;
@@ -1400,13 +1529,12 @@ static int __cmd_report(void)
1400 exit(0); 1529 exit(0);
1401 } 1530 }
1402 1531
1403 if (read(input, &file_header, sizeof(file_header)) == -1) { 1532 header = perf_header__read(input);
1404 perror("failed to read file headers"); 1533 head = header->data_offset;
1405 exit(-1);
1406 }
1407 1534
1408 if (sort__has_parent && 1535 sample_type = perf_header__sample_type();
1409 !(file_header.sample_type & PERF_SAMPLE_CALLCHAIN)) { 1536
1537 if (sort__has_parent && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1410 fprintf(stderr, "selected --sort parent, but no callchain data\n"); 1538 fprintf(stderr, "selected --sort parent, but no callchain data\n");
1411 exit(-1); 1539 exit(-1);
1412 } 1540 }
@@ -1426,6 +1554,11 @@ static int __cmd_report(void)
1426 cwd = NULL; 1554 cwd = NULL;
1427 cwdlen = 0; 1555 cwdlen = 0;
1428 } 1556 }
1557
1558 shift = page_size * (head / page_size);
1559 offset += shift;
1560 head -= shift;
1561
1429remap: 1562remap:
1430 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ, 1563 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1431 MAP_SHARED, input, offset); 1564 MAP_SHARED, input, offset);
@@ -1442,9 +1575,10 @@ more:
1442 size = 8; 1575 size = 8;
1443 1576
1444 if (head + event->header.size >= page_size * mmap_window) { 1577 if (head + event->header.size >= page_size * mmap_window) {
1445 unsigned long shift = page_size * (head / page_size);
1446 int ret; 1578 int ret;
1447 1579
1580 shift = page_size * (head / page_size);
1581
1448 ret = munmap(buf, page_size * mmap_window); 1582 ret = munmap(buf, page_size * mmap_window);
1449 assert(ret == 0); 1583 assert(ret == 0);
1450 1584
@@ -1482,7 +1616,7 @@ more:
1482 1616
1483 head += size; 1617 head += size;
1484 1618
1485 if (offset + head >= sizeof(file_header) + file_header.data_size) 1619 if (offset + head >= header->data_offset + header->data_size)
1486 goto done; 1620 goto done;
1487 1621
1488 if (offset + head < stat.st_size) 1622 if (offset + head < stat.st_size)
@@ -1536,6 +1670,13 @@ static const struct option options[] = {
1536 "regex filter to identify parent, see: '--sort parent'"), 1670 "regex filter to identify parent, see: '--sort parent'"),
1537 OPT_BOOLEAN('x', "exclude-other", &exclude_other, 1671 OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1538 "Only display entries with parent-match"), 1672 "Only display entries with parent-match"),
1673 OPT_BOOLEAN('c', "callchain", &callchain, "Display callchains"),
1674 OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
1675 "only consider symbols in these dsos"),
1676 OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
1677 "only consider symbols in these comms"),
1678 OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
1679 "only consider these symbols"),
1539 OPT_END() 1680 OPT_END()
1540}; 1681};
1541 1682
@@ -1554,6 +1695,19 @@ static void setup_sorting(void)
1554 free(str); 1695 free(str);
1555} 1696}
1556 1697
1698static void setup_list(struct strlist **list, const char *list_str,
1699 const char *list_name)
1700{
1701 if (list_str) {
1702 *list = strlist__new(true, list_str);
1703 if (!*list) {
1704 fprintf(stderr, "problems parsing %s list\n",
1705 list_name);
1706 exit(129);
1707 }
1708 }
1709}
1710
1557int cmd_report(int argc, const char **argv, const char *prefix) 1711int cmd_report(int argc, const char **argv, const char *prefix)
1558{ 1712{
1559 symbol__init(); 1713 symbol__init();
@@ -1575,6 +1729,10 @@ int cmd_report(int argc, const char **argv, const char *prefix)
1575 if (argc) 1729 if (argc)
1576 usage_with_options(report_usage, options); 1730 usage_with_options(report_usage, options);
1577 1731
1732 setup_list(&dso_list, dso_list_str, "dso");
1733 setup_list(&comm_list, comm_list_str, "comm");
1734 setup_list(&sym_list, sym_list_str, "symbol");
1735
1578 setup_pager(); 1736 setup_pager();
1579 1737
1580 return __cmd_report(); 1738 return __cmd_report();
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index 6d3eeac1ea25..2e03524a1de0 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -32,6 +32,7 @@
32 * Wu Fengguang <fengguang.wu@intel.com> 32 * Wu Fengguang <fengguang.wu@intel.com>
33 * Mike Galbraith <efault@gmx.de> 33 * Mike Galbraith <efault@gmx.de>
34 * Paul Mackerras <paulus@samba.org> 34 * Paul Mackerras <paulus@samba.org>
35 * Jaswinder Singh Rajput <jaswinder@kernel.org>
35 * 36 *
36 * Released under the GPL v2. (and only v2, not any later version) 37 * Released under the GPL v2. (and only v2, not any later version)
37 */ 38 */
@@ -45,7 +46,7 @@
45#include <sys/prctl.h> 46#include <sys/prctl.h>
46#include <math.h> 47#include <math.h>
47 48
48static struct perf_counter_attr default_attrs[MAX_COUNTERS] = { 49static struct perf_counter_attr default_attrs[] = {
49 50
50 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 51 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
51 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES}, 52 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
@@ -59,42 +60,28 @@ static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
59 60
60}; 61};
61 62
63#define MAX_RUN 100
64
62static int system_wide = 0; 65static int system_wide = 0;
63static int inherit = 1;
64static int verbose = 0; 66static int verbose = 0;
65
66static int fd[MAX_NR_CPUS][MAX_COUNTERS];
67
68static int target_pid = -1;
69static int nr_cpus = 0; 67static int nr_cpus = 0;
70static unsigned int page_size; 68static int run_idx = 0;
71 69
70static int run_count = 1;
71static int inherit = 1;
72static int scale = 1; 72static int scale = 1;
73static int target_pid = -1;
74static int null_run = 0;
73 75
74static const unsigned int default_count[] = { 76static int fd[MAX_NR_CPUS][MAX_COUNTERS];
75 1000000,
76 1000000,
77 10000,
78 10000,
79 1000000,
80 10000,
81};
82
83#define MAX_RUN 100
84
85static int run_count = 1;
86static int run_idx = 0;
87
88static u64 event_res[MAX_RUN][MAX_COUNTERS][3];
89static u64 event_scaled[MAX_RUN][MAX_COUNTERS];
90
91//static u64 event_hist[MAX_RUN][MAX_COUNTERS][3];
92
93 77
94static u64 runtime_nsecs[MAX_RUN]; 78static u64 runtime_nsecs[MAX_RUN];
95static u64 walltime_nsecs[MAX_RUN]; 79static u64 walltime_nsecs[MAX_RUN];
96static u64 runtime_cycles[MAX_RUN]; 80static u64 runtime_cycles[MAX_RUN];
97 81
82static u64 event_res[MAX_RUN][MAX_COUNTERS][3];
83static u64 event_scaled[MAX_RUN][MAX_COUNTERS];
84
98static u64 event_res_avg[MAX_COUNTERS][3]; 85static u64 event_res_avg[MAX_COUNTERS][3];
99static u64 event_res_noise[MAX_COUNTERS][3]; 86static u64 event_res_noise[MAX_COUNTERS][3];
100 87
@@ -109,7 +96,10 @@ static u64 walltime_nsecs_noise;
109static u64 runtime_cycles_avg; 96static u64 runtime_cycles_avg;
110static u64 runtime_cycles_noise; 97static u64 runtime_cycles_noise;
111 98
112static void create_perf_stat_counter(int counter) 99#define ERR_PERF_OPEN \
100"Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
101
102static void create_perf_stat_counter(int counter, int pid)
113{ 103{
114 struct perf_counter_attr *attr = attrs + counter; 104 struct perf_counter_attr *attr = attrs + counter;
115 105
@@ -119,20 +109,21 @@ static void create_perf_stat_counter(int counter)
119 109
120 if (system_wide) { 110 if (system_wide) {
121 int cpu; 111 int cpu;
122 for (cpu = 0; cpu < nr_cpus; cpu ++) { 112 for (cpu = 0; cpu < nr_cpus; cpu++) {
123 fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0); 113 fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
124 if (fd[cpu][counter] < 0 && verbose) { 114 if (fd[cpu][counter] < 0 && verbose)
125 printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[cpu][counter], strerror(errno)); 115 fprintf(stderr, ERR_PERF_OPEN, counter,
126 } 116 fd[cpu][counter], strerror(errno));
127 } 117 }
128 } else { 118 } else {
129 attr->inherit = inherit; 119 attr->inherit = inherit;
130 attr->disabled = 1; 120 attr->disabled = 1;
131 121 attr->enable_on_exec = 1;
132 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0); 122
133 if (fd[0][counter] < 0 && verbose) { 123 fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
134 printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[0][counter], strerror(errno)); 124 if (fd[0][counter] < 0 && verbose)
135 } 125 fprintf(stderr, ERR_PERF_OPEN, counter,
126 fd[0][counter], strerror(errno));
136 } 127 }
137} 128}
138 129
@@ -168,7 +159,7 @@ static void read_counter(int counter)
168 count[0] = count[1] = count[2] = 0; 159 count[0] = count[1] = count[2] = 0;
169 160
170 nv = scale ? 3 : 1; 161 nv = scale ? 3 : 1;
171 for (cpu = 0; cpu < nr_cpus; cpu ++) { 162 for (cpu = 0; cpu < nr_cpus; cpu++) {
172 if (fd[cpu][counter] < 0) 163 if (fd[cpu][counter] < 0)
173 continue; 164 continue;
174 165
@@ -215,32 +206,67 @@ static int run_perf_stat(int argc, const char **argv)
215 int status = 0; 206 int status = 0;
216 int counter; 207 int counter;
217 int pid; 208 int pid;
209 int child_ready_pipe[2], go_pipe[2];
210 char buf;
218 211
219 if (!system_wide) 212 if (!system_wide)
220 nr_cpus = 1; 213 nr_cpus = 1;
221 214
222 for (counter = 0; counter < nr_counters; counter++) 215 if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
223 create_perf_stat_counter(counter); 216 perror("failed to create pipes");
224 217 exit(1);
225 /* 218 }
226 * Enable counters and exec the command:
227 */
228 t0 = rdclock();
229 prctl(PR_TASK_PERF_COUNTERS_ENABLE);
230 219
231 if ((pid = fork()) < 0) 220 if ((pid = fork()) < 0)
232 perror("failed to fork"); 221 perror("failed to fork");
233 222
234 if (!pid) { 223 if (!pid) {
235 if (execvp(argv[0], (char **)argv)) { 224 close(child_ready_pipe[0]);
236 perror(argv[0]); 225 close(go_pipe[1]);
237 exit(-1); 226 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
238 } 227
228 /*
229 * Do a dummy execvp to get the PLT entry resolved,
230 * so we avoid the resolver overhead on the real
231 * execvp call.
232 */
233 execvp("", (char **)argv);
234
235 /*
236 * Tell the parent we're ready to go
237 */
238 close(child_ready_pipe[1]);
239
240 /*
241 * Wait until the parent tells us to go.
242 */
243 read(go_pipe[0], &buf, 1);
244
245 execvp(argv[0], (char **)argv);
246
247 perror(argv[0]);
248 exit(-1);
239 } 249 }
240 250
251 /*
252 * Wait for the child to be ready to exec.
253 */
254 close(child_ready_pipe[1]);
255 close(go_pipe[0]);
256 read(child_ready_pipe[0], &buf, 1);
257 close(child_ready_pipe[0]);
258
259 for (counter = 0; counter < nr_counters; counter++)
260 create_perf_stat_counter(counter, pid);
261
262 /*
263 * Enable counters and exec the command:
264 */
265 t0 = rdclock();
266
267 close(go_pipe[1]);
241 wait(&status); 268 wait(&status);
242 269
243 prctl(PR_TASK_PERF_COUNTERS_DISABLE);
244 t1 = rdclock(); 270 t1 = rdclock();
245 271
246 walltime_nsecs[run_idx] = t1 - t0; 272 walltime_nsecs[run_idx] = t1 - t0;
@@ -262,7 +288,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise)
262{ 288{
263 double msecs = (double)count[0] / 1000000; 289 double msecs = (double)count[0] / 1000000;
264 290
265 fprintf(stderr, " %14.6f %-20s", msecs, event_name(counter)); 291 fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter));
266 292
267 if (attrs[counter].type == PERF_TYPE_SOFTWARE && 293 if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
268 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) { 294 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
@@ -276,7 +302,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise)
276 302
277static void abs_printout(int counter, u64 *count, u64 *noise) 303static void abs_printout(int counter, u64 *count, u64 *noise)
278{ 304{
279 fprintf(stderr, " %14Ld %-20s", count[0], event_name(counter)); 305 fprintf(stderr, " %14Ld %-24s", count[0], event_name(counter));
280 306
281 if (runtime_cycles_avg && 307 if (runtime_cycles_avg &&
282 attrs[counter].type == PERF_TYPE_HARDWARE && 308 attrs[counter].type == PERF_TYPE_HARDWARE &&
@@ -306,7 +332,7 @@ static void print_counter(int counter)
306 scaled = event_scaled_avg[counter]; 332 scaled = event_scaled_avg[counter];
307 333
308 if (scaled == -1) { 334 if (scaled == -1) {
309 fprintf(stderr, " %14s %-20s\n", 335 fprintf(stderr, " %14s %-24s\n",
310 "<not counted>", event_name(counter)); 336 "<not counted>", event_name(counter));
311 return; 337 return;
312 } 338 }
@@ -364,8 +390,11 @@ static void calc_avg(void)
364 event_res_avg[j]+1, event_res[i][j]+1); 390 event_res_avg[j]+1, event_res[i][j]+1);
365 update_avg("counter/2", j, 391 update_avg("counter/2", j,
366 event_res_avg[j]+2, event_res[i][j]+2); 392 event_res_avg[j]+2, event_res[i][j]+2);
367 update_avg("scaled", j, 393 if (event_scaled[i][j] != -1)
368 event_scaled_avg + j, event_scaled[i]+j); 394 update_avg("scaled", j,
395 event_scaled_avg + j, event_scaled[i]+j);
396 else
397 event_scaled_avg[j] = -1;
369 } 398 }
370 } 399 }
371 runtime_nsecs_avg /= run_count; 400 runtime_nsecs_avg /= run_count;
@@ -429,11 +458,14 @@ static void print_stat(int argc, const char **argv)
429 for (counter = 0; counter < nr_counters; counter++) 458 for (counter = 0; counter < nr_counters; counter++)
430 print_counter(counter); 459 print_counter(counter);
431 460
432
433 fprintf(stderr, "\n"); 461 fprintf(stderr, "\n");
434 fprintf(stderr, " %14.9f seconds time elapsed.\n", 462 fprintf(stderr, " %14.9f seconds time elapsed",
435 (double)walltime_nsecs_avg/1e9); 463 (double)walltime_nsecs_avg/1e9);
436 fprintf(stderr, "\n"); 464 if (run_count > 1) {
465 fprintf(stderr, " ( +- %7.3f%% )",
466 100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
467 }
468 fprintf(stderr, "\n\n");
437} 469}
438 470
439static volatile int signr = -1; 471static volatile int signr = -1;
@@ -466,13 +498,15 @@ static const struct option options[] = {
466 OPT_INTEGER('p', "pid", &target_pid, 498 OPT_INTEGER('p', "pid", &target_pid,
467 "stat events on existing pid"), 499 "stat events on existing pid"),
468 OPT_BOOLEAN('a', "all-cpus", &system_wide, 500 OPT_BOOLEAN('a', "all-cpus", &system_wide,
469 "system-wide collection from all CPUs"), 501 "system-wide collection from all CPUs"),
470 OPT_BOOLEAN('S', "scale", &scale, 502 OPT_BOOLEAN('S', "scale", &scale,
471 "scale/normalize counters"), 503 "scale/normalize counters"),
472 OPT_BOOLEAN('v', "verbose", &verbose, 504 OPT_BOOLEAN('v', "verbose", &verbose,
473 "be more verbose (show counter open errors, etc)"), 505 "be more verbose (show counter open errors, etc)"),
474 OPT_INTEGER('r', "repeat", &run_count, 506 OPT_INTEGER('r', "repeat", &run_count,
475 "repeat command and print average + stddev (max: 100)"), 507 "repeat command and print average + stddev (max: 100)"),
508 OPT_BOOLEAN('n', "null", &null_run,
509 "null run - dont start any counters"),
476 OPT_END() 510 OPT_END()
477}; 511};
478 512
@@ -480,18 +514,17 @@ int cmd_stat(int argc, const char **argv, const char *prefix)
480{ 514{
481 int status; 515 int status;
482 516
483 page_size = sysconf(_SC_PAGE_SIZE);
484
485 memcpy(attrs, default_attrs, sizeof(attrs));
486
487 argc = parse_options(argc, argv, options, stat_usage, 0); 517 argc = parse_options(argc, argv, options, stat_usage, 0);
488 if (!argc) 518 if (!argc)
489 usage_with_options(stat_usage, options); 519 usage_with_options(stat_usage, options);
490 if (run_count <= 0 || run_count > MAX_RUN) 520 if (run_count <= 0 || run_count > MAX_RUN)
491 usage_with_options(stat_usage, options); 521 usage_with_options(stat_usage, options);
492 522
493 if (!nr_counters) 523 /* Set attrs and nr_counters if no event is selected and !null_run */
494 nr_counters = 8; 524 if (!null_run && !nr_counters) {
525 memcpy(attrs, default_attrs, sizeof(default_attrs));
526 nr_counters = ARRAY_SIZE(default_attrs);
527 }
495 528
496 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN); 529 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
497 assert(nr_cpus <= MAX_NR_CPUS); 530 assert(nr_cpus <= MAX_NR_CPUS);
@@ -511,7 +544,7 @@ int cmd_stat(int argc, const char **argv, const char *prefix)
511 status = 0; 544 status = 0;
512 for (run_idx = 0; run_idx < run_count; run_idx++) { 545 for (run_idx = 0; run_idx < run_count; run_idx++) {
513 if (run_count != 1 && verbose) 546 if (run_count != 1 && verbose)
514 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx+1); 547 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
515 status = run_perf_stat(argc, argv); 548 status = run_perf_stat(argc, argv);
516 } 549 }
517 550
diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
index 5352b5e352ed..cf0d21f1ae10 100644
--- a/tools/perf/builtin-top.c
+++ b/tools/perf/builtin-top.c
@@ -392,11 +392,11 @@ static void record_ip(u64 ip, int counter)
392 samples--; 392 samples--;
393} 393}
394 394
395static void process_event(u64 ip, int counter) 395static void process_event(u64 ip, int counter, int user)
396{ 396{
397 samples++; 397 samples++;
398 398
399 if (ip < min_ip || ip > max_ip) { 399 if (user) {
400 userspace_samples++; 400 userspace_samples++;
401 return; 401 return;
402 } 402 }
@@ -509,9 +509,10 @@ static void mmap_read_counter(struct mmap_data *md)
509 509
510 old += size; 510 old += size;
511 511
512 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) { 512 if (event->header.type == PERF_EVENT_SAMPLE) {
513 if (event->header.type & PERF_SAMPLE_IP) 513 int user =
514 process_event(event->ip.ip, md->counter); 514 (event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK) == PERF_EVENT_MISC_USER;
515 process_event(event->ip.ip, md->counter, user);
515 } 516 }
516 } 517 }
517 518
diff --git a/tools/perf/perf.h b/tools/perf/perf.h
index ceb68aa51f7f..ce394192c85a 100644
--- a/tools/perf/perf.h
+++ b/tools/perf/perf.h
@@ -25,7 +25,7 @@
25#include <sys/syscall.h> 25#include <sys/syscall.h>
26 26
27#include "../../include/linux/perf_counter.h" 27#include "../../include/linux/perf_counter.h"
28#include "types.h" 28#include "util/types.h"
29 29
30/* 30/*
31 * prctl(PR_TASK_PERF_COUNTERS_DISABLE) will (cheaply) disable all 31 * prctl(PR_TASK_PERF_COUNTERS_DISABLE) will (cheaply) disable all
@@ -72,10 +72,9 @@ sys_perf_counter_open(struct perf_counter_attr *attr,
72#define MAX_COUNTERS 256 72#define MAX_COUNTERS 256
73#define MAX_NR_CPUS 256 73#define MAX_NR_CPUS 256
74 74
75struct perf_file_header { 75struct ip_callchain {
76 u64 version; 76 u64 nr;
77 u64 sample_type; 77 u64 ips[0];
78 u64 data_size;
79}; 78};
80 79
81#endif 80#endif
diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
new file mode 100644
index 000000000000..ad3c28578961
--- /dev/null
+++ b/tools/perf/util/callchain.c
@@ -0,0 +1,174 @@
1/*
2 * Copyright (C) 2009, Frederic Weisbecker <fweisbec@gmail.com>
3 *
4 * Handle the callchains from the stream in an ad-hoc radix tree and then
5 * sort them in an rbtree.
6 *
7 */
8
9#include <stdlib.h>
10#include <stdio.h>
11#include <stdbool.h>
12#include <errno.h>
13
14#include "callchain.h"
15
16
17static void rb_insert_callchain(struct rb_root *root, struct callchain_node *chain)
18{
19 struct rb_node **p = &root->rb_node;
20 struct rb_node *parent = NULL;
21 struct callchain_node *rnode;
22
23 while (*p) {
24 parent = *p;
25 rnode = rb_entry(parent, struct callchain_node, rb_node);
26
27 if (rnode->hit < chain->hit)
28 p = &(*p)->rb_left;
29 else
30 p = &(*p)->rb_right;
31 }
32
33 rb_link_node(&chain->rb_node, parent, p);
34 rb_insert_color(&chain->rb_node, root);
35}
36
37/*
38 * Once we get every callchains from the stream, we can now
39 * sort them by hit
40 */
41void sort_chain_to_rbtree(struct rb_root *rb_root, struct callchain_node *node)
42{
43 struct callchain_node *child;
44
45 list_for_each_entry(child, &node->children, brothers)
46 sort_chain_to_rbtree(rb_root, child);
47
48 if (node->hit)
49 rb_insert_callchain(rb_root, node);
50}
51
52static struct callchain_node *create_child(struct callchain_node *parent)
53{
54 struct callchain_node *new;
55
56 new = malloc(sizeof(*new));
57 if (!new) {
58 perror("not enough memory to create child for code path tree");
59 return NULL;
60 }
61 new->parent = parent;
62 INIT_LIST_HEAD(&new->children);
63 INIT_LIST_HEAD(&new->val);
64 list_add_tail(&new->brothers, &parent->children);
65
66 return new;
67}
68
69static void
70fill_node(struct callchain_node *node, struct ip_callchain *chain, int start)
71{
72 int i;
73
74 for (i = start; i < chain->nr; i++) {
75 struct callchain_list *call;
76
77 call = malloc(sizeof(*chain));
78 if (!call) {
79 perror("not enough memory for the code path tree");
80 return;
81 }
82 call->ip = chain->ips[i];
83 list_add_tail(&call->list, &node->val);
84 }
85 node->val_nr = i - start;
86}
87
88static void add_child(struct callchain_node *parent, struct ip_callchain *chain)
89{
90 struct callchain_node *new;
91
92 new = create_child(parent);
93 fill_node(new, chain, parent->val_nr);
94
95 new->hit = 1;
96}
97
98static void
99split_add_child(struct callchain_node *parent, struct ip_callchain *chain,
100 struct callchain_list *to_split, int idx)
101{
102 struct callchain_node *new;
103
104 /* split */
105 new = create_child(parent);
106 list_move_tail(&to_split->list, &new->val);
107 new->hit = parent->hit;
108 parent->hit = 0;
109 parent->val_nr = idx;
110
111 /* create the new one */
112 add_child(parent, chain);
113}
114
115static int
116__append_chain(struct callchain_node *root, struct ip_callchain *chain,
117 int start);
118
119static int
120__append_chain_children(struct callchain_node *root, struct ip_callchain *chain)
121{
122 struct callchain_node *rnode;
123
124 /* lookup in childrens */
125 list_for_each_entry(rnode, &root->children, brothers) {
126 int ret = __append_chain(rnode, chain, root->val_nr);
127 if (!ret)
128 return 0;
129 }
130 return -1;
131}
132
133static int
134__append_chain(struct callchain_node *root, struct ip_callchain *chain,
135 int start)
136{
137 struct callchain_list *cnode;
138 int i = start;
139 bool found = false;
140
141 /* lookup in the current node */
142 list_for_each_entry(cnode, &root->val, list) {
143 if (cnode->ip != chain->ips[i++])
144 break;
145 if (!found)
146 found = true;
147 if (i == chain->nr)
148 break;
149 }
150
151 /* matches not, relay on the parent */
152 if (!found)
153 return -1;
154
155 /* we match only a part of the node. Split it and add the new chain */
156 if (i < root->val_nr) {
157 split_add_child(root, chain, cnode, i);
158 return 0;
159 }
160
161 /* we match 100% of the path, increment the hit */
162 if (i == root->val_nr) {
163 root->hit++;
164 return 0;
165 }
166
167 return __append_chain_children(root, chain);
168}
169
170void append_chain(struct callchain_node *root, struct ip_callchain *chain)
171{
172 if (__append_chain_children(root, chain) == -1)
173 add_child(root, chain);
174}
diff --git a/tools/perf/util/callchain.h b/tools/perf/util/callchain.h
new file mode 100644
index 000000000000..fa1cd2f71fd3
--- /dev/null
+++ b/tools/perf/util/callchain.h
@@ -0,0 +1,33 @@
1#ifndef __PERF_CALLCHAIN_H
2#define __PERF_CALLCHAIN_H
3
4#include "../perf.h"
5#include "list.h"
6#include "rbtree.h"
7
8
9struct callchain_node {
10 struct callchain_node *parent;
11 struct list_head brothers;
12 struct list_head children;
13 struct list_head val;
14 struct rb_node rb_node;
15 int val_nr;
16 int hit;
17};
18
19struct callchain_list {
20 unsigned long ip;
21 struct list_head list;
22};
23
24static inline void callchain_init(struct callchain_node *node)
25{
26 INIT_LIST_HEAD(&node->brothers);
27 INIT_LIST_HEAD(&node->children);
28 INIT_LIST_HEAD(&node->val);
29}
30
31void append_chain(struct callchain_node *root, struct ip_callchain *chain);
32void sort_chain_to_rbtree(struct rb_root *rb_root, struct callchain_node *node);
33#endif
diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c
new file mode 100644
index 000000000000..450384b3bbe5
--- /dev/null
+++ b/tools/perf/util/header.c
@@ -0,0 +1,242 @@
1#include <sys/types.h>
2#include <unistd.h>
3#include <stdio.h>
4#include <stdlib.h>
5
6#include "util.h"
7#include "header.h"
8
9/*
10 *
11 */
12
13struct perf_header_attr *perf_header_attr__new(struct perf_counter_attr *attr)
14{
15 struct perf_header_attr *self = malloc(sizeof(*self));
16
17 if (!self)
18 die("nomem");
19
20 self->attr = *attr;
21 self->ids = 0;
22 self->size = 1;
23 self->id = malloc(sizeof(u64));
24
25 if (!self->id)
26 die("nomem");
27
28 return self;
29}
30
31void perf_header_attr__add_id(struct perf_header_attr *self, u64 id)
32{
33 int pos = self->ids;
34
35 self->ids++;
36 if (self->ids > self->size) {
37 self->size *= 2;
38 self->id = realloc(self->id, self->size * sizeof(u64));
39 if (!self->id)
40 die("nomem");
41 }
42 self->id[pos] = id;
43}
44
45/*
46 *
47 */
48
49struct perf_header *perf_header__new(void)
50{
51 struct perf_header *self = malloc(sizeof(*self));
52
53 if (!self)
54 die("nomem");
55
56 self->frozen = 0;
57
58 self->attrs = 0;
59 self->size = 1;
60 self->attr = malloc(sizeof(void *));
61
62 if (!self->attr)
63 die("nomem");
64
65 self->data_offset = 0;
66 self->data_size = 0;
67
68 return self;
69}
70
71void perf_header__add_attr(struct perf_header *self,
72 struct perf_header_attr *attr)
73{
74 int pos = self->attrs;
75
76 if (self->frozen)
77 die("frozen");
78
79 self->attrs++;
80 if (self->attrs > self->size) {
81 self->size *= 2;
82 self->attr = realloc(self->attr, self->size * sizeof(void *));
83 if (!self->attr)
84 die("nomem");
85 }
86 self->attr[pos] = attr;
87}
88
89static const char *__perf_magic = "PERFFILE";
90
91#define PERF_MAGIC (*(u64 *)__perf_magic)
92
93struct perf_file_section {
94 u64 offset;
95 u64 size;
96};
97
98struct perf_file_attr {
99 struct perf_counter_attr attr;
100 struct perf_file_section ids;
101};
102
103struct perf_file_header {
104 u64 magic;
105 u64 size;
106 u64 attr_size;
107 struct perf_file_section attrs;
108 struct perf_file_section data;
109};
110
111static void do_write(int fd, void *buf, size_t size)
112{
113 while (size) {
114 int ret = write(fd, buf, size);
115
116 if (ret < 0)
117 die("failed to write");
118
119 size -= ret;
120 buf += ret;
121 }
122}
123
124void perf_header__write(struct perf_header *self, int fd)
125{
126 struct perf_file_header f_header;
127 struct perf_file_attr f_attr;
128 struct perf_header_attr *attr;
129 int i;
130
131 lseek(fd, sizeof(f_header), SEEK_SET);
132
133
134 for (i = 0; i < self->attrs; i++) {
135 attr = self->attr[i];
136
137 attr->id_offset = lseek(fd, 0, SEEK_CUR);
138 do_write(fd, attr->id, attr->ids * sizeof(u64));
139 }
140
141
142 self->attr_offset = lseek(fd, 0, SEEK_CUR);
143
144 for (i = 0; i < self->attrs; i++) {
145 attr = self->attr[i];
146
147 f_attr = (struct perf_file_attr){
148 .attr = attr->attr,
149 .ids = {
150 .offset = attr->id_offset,
151 .size = attr->ids * sizeof(u64),
152 }
153 };
154 do_write(fd, &f_attr, sizeof(f_attr));
155 }
156
157
158 self->data_offset = lseek(fd, 0, SEEK_CUR);
159
160 f_header = (struct perf_file_header){
161 .magic = PERF_MAGIC,
162 .size = sizeof(f_header),
163 .attr_size = sizeof(f_attr),
164 .attrs = {
165 .offset = self->attr_offset,
166 .size = self->attrs * sizeof(f_attr),
167 },
168 .data = {
169 .offset = self->data_offset,
170 .size = self->data_size,
171 },
172 };
173
174 lseek(fd, 0, SEEK_SET);
175 do_write(fd, &f_header, sizeof(f_header));
176 lseek(fd, self->data_offset + self->data_size, SEEK_SET);
177
178 self->frozen = 1;
179}
180
181static void do_read(int fd, void *buf, size_t size)
182{
183 while (size) {
184 int ret = read(fd, buf, size);
185
186 if (ret < 0)
187 die("failed to read");
188
189 size -= ret;
190 buf += ret;
191 }
192}
193
194struct perf_header *perf_header__read(int fd)
195{
196 struct perf_header *self = perf_header__new();
197 struct perf_file_header f_header;
198 struct perf_file_attr f_attr;
199 u64 f_id;
200
201 int nr_attrs, nr_ids, i, j;
202
203 lseek(fd, 0, SEEK_SET);
204 do_read(fd, &f_header, sizeof(f_header));
205
206 if (f_header.magic != PERF_MAGIC ||
207 f_header.size != sizeof(f_header) ||
208 f_header.attr_size != sizeof(f_attr))
209 die("incompatible file format");
210
211 nr_attrs = f_header.attrs.size / sizeof(f_attr);
212 lseek(fd, f_header.attrs.offset, SEEK_SET);
213
214 for (i = 0; i < nr_attrs; i++) {
215 struct perf_header_attr *attr;
216 off_t tmp = lseek(fd, 0, SEEK_CUR);
217
218 do_read(fd, &f_attr, sizeof(f_attr));
219
220 attr = perf_header_attr__new(&f_attr.attr);
221
222 nr_ids = f_attr.ids.size / sizeof(u64);
223 lseek(fd, f_attr.ids.offset, SEEK_SET);
224
225 for (j = 0; j < nr_ids; j++) {
226 do_read(fd, &f_id, sizeof(f_id));
227
228 perf_header_attr__add_id(attr, f_id);
229 }
230 perf_header__add_attr(self, attr);
231 lseek(fd, tmp, SEEK_SET);
232 }
233
234 self->data_offset = f_header.data.offset;
235 self->data_size = f_header.data.size;
236
237 lseek(fd, self->data_offset + self->data_size, SEEK_SET);
238
239 self->frozen = 1;
240
241 return self;
242}
diff --git a/tools/perf/util/header.h b/tools/perf/util/header.h
new file mode 100644
index 000000000000..b5ef53ad4c7a
--- /dev/null
+++ b/tools/perf/util/header.h
@@ -0,0 +1,37 @@
1#ifndef _PERF_HEADER_H
2#define _PERF_HEADER_H
3
4#include "../../../include/linux/perf_counter.h"
5#include <sys/types.h>
6#include "types.h"
7
8struct perf_header_attr {
9 struct perf_counter_attr attr;
10 int ids, size;
11 u64 *id;
12 off_t id_offset;
13};
14
15struct perf_header {
16 int frozen;
17 int attrs, size;
18 struct perf_header_attr **attr;
19 off_t attr_offset;
20 u64 data_offset;
21 u64 data_size;
22};
23
24struct perf_header *perf_header__read(int fd);
25void perf_header__write(struct perf_header *self, int fd);
26
27void perf_header__add_attr(struct perf_header *self,
28 struct perf_header_attr *attr);
29
30struct perf_header_attr *
31perf_header_attr__new(struct perf_counter_attr *attr);
32void perf_header_attr__add_id(struct perf_header_attr *self, u64 id);
33
34
35struct perf_header *perf_header__new(void);
36
37#endif /* _PERF_HEADER_H */
diff --git a/tools/perf/util/help.c b/tools/perf/util/help.c
index 6653f7dd1d78..17a00e0df2c4 100644
--- a/tools/perf/util/help.c
+++ b/tools/perf/util/help.c
@@ -126,21 +126,6 @@ static int is_executable(const char *name)
126 !S_ISREG(st.st_mode)) 126 !S_ISREG(st.st_mode))
127 return 0; 127 return 0;
128 128
129#ifdef __MINGW32__
130 /* cannot trust the executable bit, peek into the file instead */
131 char buf[3] = { 0 };
132 int n;
133 int fd = open(name, O_RDONLY);
134 st.st_mode &= ~S_IXUSR;
135 if (fd >= 0) {
136 n = read(fd, buf, 2);
137 if (n == 2)
138 /* DOS executables start with "MZ" */
139 if (!strcmp(buf, "#!") || !strcmp(buf, "MZ"))
140 st.st_mode |= S_IXUSR;
141 close(fd);
142 }
143#endif
144 return st.st_mode & S_IXUSR; 129 return st.st_mode & S_IXUSR;
145} 130}
146 131
diff --git a/tools/perf/util/pager.c b/tools/perf/util/pager.c
index a28bccae5458..1915de20dcac 100644
--- a/tools/perf/util/pager.c
+++ b/tools/perf/util/pager.c
@@ -9,7 +9,6 @@
9 9
10static int spawned_pager; 10static int spawned_pager;
11 11
12#ifndef __MINGW32__
13static void pager_preexec(void) 12static void pager_preexec(void)
14{ 13{
15 /* 14 /*
@@ -24,7 +23,6 @@ static void pager_preexec(void)
24 23
25 setenv("LESS", "FRSX", 0); 24 setenv("LESS", "FRSX", 0);
26} 25}
27#endif
28 26
29static const char *pager_argv[] = { "sh", "-c", NULL, NULL }; 27static const char *pager_argv[] = { "sh", "-c", NULL, NULL };
30static struct child_process pager_process; 28static struct child_process pager_process;
@@ -70,9 +68,8 @@ void setup_pager(void)
70 pager_argv[2] = pager; 68 pager_argv[2] = pager;
71 pager_process.argv = pager_argv; 69 pager_process.argv = pager_argv;
72 pager_process.in = -1; 70 pager_process.in = -1;
73#ifndef __MINGW32__
74 pager_process.preexec_cb = pager_preexec; 71 pager_process.preexec_cb = pager_preexec;
75#endif 72
76 if (start_command(&pager_process)) 73 if (start_command(&pager_process))
77 return; 74 return;
78 75
diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index 35d04da38d6a..4d042f104cdc 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -16,32 +16,28 @@ struct event_symbol {
16 u8 type; 16 u8 type;
17 u64 config; 17 u64 config;
18 char *symbol; 18 char *symbol;
19 char *alias;
19}; 20};
20 21
21#define C(x, y) .type = PERF_TYPE_##x, .config = PERF_COUNT_##y 22#define CHW(x) .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_##x
22#define CR(x, y) .type = PERF_TYPE_##x, .config = y 23#define CSW(x) .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_##x
23 24
24static struct event_symbol event_symbols[] = { 25static struct event_symbol event_symbols[] = {
25 { C(HARDWARE, HW_CPU_CYCLES), "cpu-cycles", }, 26 { CHW(CPU_CYCLES), "cpu-cycles", "cycles" },
26 { C(HARDWARE, HW_CPU_CYCLES), "cycles", }, 27 { CHW(INSTRUCTIONS), "instructions", "" },
27 { C(HARDWARE, HW_INSTRUCTIONS), "instructions", }, 28 { CHW(CACHE_REFERENCES), "cache-references", "" },
28 { C(HARDWARE, HW_CACHE_REFERENCES), "cache-references", }, 29 { CHW(CACHE_MISSES), "cache-misses", "" },
29 { C(HARDWARE, HW_CACHE_MISSES), "cache-misses", }, 30 { CHW(BRANCH_INSTRUCTIONS), "branch-instructions", "branches" },
30 { C(HARDWARE, HW_BRANCH_INSTRUCTIONS),"branch-instructions", }, 31 { CHW(BRANCH_MISSES), "branch-misses", "" },
31 { C(HARDWARE, HW_BRANCH_INSTRUCTIONS),"branches", }, 32 { CHW(BUS_CYCLES), "bus-cycles", "" },
32 { C(HARDWARE, HW_BRANCH_MISSES), "branch-misses", }, 33
33 { C(HARDWARE, HW_BUS_CYCLES), "bus-cycles", }, 34 { CSW(CPU_CLOCK), "cpu-clock", "" },
34 35 { CSW(TASK_CLOCK), "task-clock", "" },
35 { C(SOFTWARE, SW_CPU_CLOCK), "cpu-clock", }, 36 { CSW(PAGE_FAULTS), "page-faults", "faults" },
36 { C(SOFTWARE, SW_TASK_CLOCK), "task-clock", }, 37 { CSW(PAGE_FAULTS_MIN), "minor-faults", "" },
37 { C(SOFTWARE, SW_PAGE_FAULTS), "page-faults", }, 38 { CSW(PAGE_FAULTS_MAJ), "major-faults", "" },
38 { C(SOFTWARE, SW_PAGE_FAULTS), "faults", }, 39 { CSW(CONTEXT_SWITCHES), "context-switches", "cs" },
39 { C(SOFTWARE, SW_PAGE_FAULTS_MIN), "minor-faults", }, 40 { CSW(CPU_MIGRATIONS), "cpu-migrations", "migrations" },
40 { C(SOFTWARE, SW_PAGE_FAULTS_MAJ), "major-faults", },
41 { C(SOFTWARE, SW_CONTEXT_SWITCHES), "context-switches", },
42 { C(SOFTWARE, SW_CONTEXT_SWITCHES), "cs", },
43 { C(SOFTWARE, SW_CPU_MIGRATIONS), "cpu-migrations", },
44 { C(SOFTWARE, SW_CPU_MIGRATIONS), "migrations", },
45}; 41};
46 42
47#define __PERF_COUNTER_FIELD(config, name) \ 43#define __PERF_COUNTER_FIELD(config, name) \
@@ -74,26 +70,70 @@ static char *sw_event_names[] = {
74 70
75#define MAX_ALIASES 8 71#define MAX_ALIASES 8
76 72
77static char *hw_cache [][MAX_ALIASES] = { 73static char *hw_cache[][MAX_ALIASES] = {
78 { "L1-data" , "l1-d", "l1d" }, 74 { "L1-d$", "l1-d", "l1d", "L1-data", },
79 { "L1-instruction" , "l1-i", "l1i" }, 75 { "L1-i$", "l1-i", "l1i", "L1-instruction", },
80 { "L2" , "l2" }, 76 { "LLC", "L2" },
81 { "Data-TLB" , "dtlb", "d-tlb" }, 77 { "dTLB", "d-tlb", "Data-TLB", },
82 { "Instruction-TLB" , "itlb", "i-tlb" }, 78 { "iTLB", "i-tlb", "Instruction-TLB", },
83 { "Branch" , "bpu" , "btb", "bpc" }, 79 { "branch", "branches", "bpu", "btb", "bpc", },
84}; 80};
85 81
86static char *hw_cache_op [][MAX_ALIASES] = { 82static char *hw_cache_op[][MAX_ALIASES] = {
87 { "Load" , "read" }, 83 { "load", "loads", "read", },
88 { "Store" , "write" }, 84 { "store", "stores", "write", },
89 { "Prefetch" , "speculative-read", "speculative-load" }, 85 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
90}; 86};
91 87
92static char *hw_cache_result [][MAX_ALIASES] = { 88static char *hw_cache_result[][MAX_ALIASES] = {
93 { "Reference" , "ops", "access" }, 89 { "refs", "Reference", "ops", "access", },
94 { "Miss" }, 90 { "misses", "miss", },
95}; 91};
96 92
93#define C(x) PERF_COUNT_HW_CACHE_##x
94#define CACHE_READ (1 << C(OP_READ))
95#define CACHE_WRITE (1 << C(OP_WRITE))
96#define CACHE_PREFETCH (1 << C(OP_PREFETCH))
97#define COP(x) (1 << x)
98
99/*
100 * cache operartion stat
101 * L1I : Read and prefetch only
102 * ITLB and BPU : Read-only
103 */
104static unsigned long hw_cache_stat[C(MAX)] = {
105 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
106 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
107 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
108 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
109 [C(ITLB)] = (CACHE_READ),
110 [C(BPU)] = (CACHE_READ),
111};
112
113static int is_cache_op_valid(u8 cache_type, u8 cache_op)
114{
115 if (hw_cache_stat[cache_type] & COP(cache_op))
116 return 1; /* valid */
117 else
118 return 0; /* invalid */
119}
120
121static char *event_cache_name(u8 cache_type, u8 cache_op, u8 cache_result)
122{
123 static char name[50];
124
125 if (cache_result) {
126 sprintf(name, "%s-%s-%s", hw_cache[cache_type][0],
127 hw_cache_op[cache_op][0],
128 hw_cache_result[cache_result][0]);
129 } else {
130 sprintf(name, "%s-%s", hw_cache[cache_type][0],
131 hw_cache_op[cache_op][1]);
132 }
133
134 return name;
135}
136
97char *event_name(int counter) 137char *event_name(int counter)
98{ 138{
99 u64 config = attrs[counter].config; 139 u64 config = attrs[counter].config;
@@ -113,7 +153,6 @@ char *event_name(int counter)
113 153
114 case PERF_TYPE_HW_CACHE: { 154 case PERF_TYPE_HW_CACHE: {
115 u8 cache_type, cache_op, cache_result; 155 u8 cache_type, cache_op, cache_result;
116 static char name[100];
117 156
118 cache_type = (config >> 0) & 0xff; 157 cache_type = (config >> 0) & 0xff;
119 if (cache_type > PERF_COUNT_HW_CACHE_MAX) 158 if (cache_type > PERF_COUNT_HW_CACHE_MAX)
@@ -127,12 +166,10 @@ char *event_name(int counter)
127 if (cache_result > PERF_COUNT_HW_CACHE_RESULT_MAX) 166 if (cache_result > PERF_COUNT_HW_CACHE_RESULT_MAX)
128 return "unknown-ext-hardware-cache-result"; 167 return "unknown-ext-hardware-cache-result";
129 168
130 sprintf(name, "%s-Cache-%s-%ses", 169 if (!is_cache_op_valid(cache_type, cache_op))
131 hw_cache[cache_type][0], 170 return "invalid-cache";
132 hw_cache_op[cache_op][0],
133 hw_cache_result[cache_result][0]);
134 171
135 return name; 172 return event_cache_name(cache_type, cache_op, cache_result);
136 } 173 }
137 174
138 case PERF_TYPE_SOFTWARE: 175 case PERF_TYPE_SOFTWARE:
@@ -163,7 +200,8 @@ static int parse_aliases(const char *str, char *names[][MAX_ALIASES], int size)
163 return -1; 200 return -1;
164} 201}
165 202
166static int parse_generic_hw_symbols(const char *str, struct perf_counter_attr *attr) 203static int
204parse_generic_hw_symbols(const char *str, struct perf_counter_attr *attr)
167{ 205{
168 int cache_type = -1, cache_op = 0, cache_result = 0; 206 int cache_type = -1, cache_op = 0, cache_result = 0;
169 207
@@ -182,6 +220,9 @@ static int parse_generic_hw_symbols(const char *str, struct perf_counter_attr *a
182 if (cache_op == -1) 220 if (cache_op == -1)
183 cache_op = PERF_COUNT_HW_CACHE_OP_READ; 221 cache_op = PERF_COUNT_HW_CACHE_OP_READ;
184 222
223 if (!is_cache_op_valid(cache_type, cache_op))
224 return -EINVAL;
225
185 cache_result = parse_aliases(str, hw_cache_result, 226 cache_result = parse_aliases(str, hw_cache_result,
186 PERF_COUNT_HW_CACHE_RESULT_MAX); 227 PERF_COUNT_HW_CACHE_RESULT_MAX);
187 /* 228 /*
@@ -196,6 +237,19 @@ static int parse_generic_hw_symbols(const char *str, struct perf_counter_attr *a
196 return 0; 237 return 0;
197} 238}
198 239
240static int check_events(const char *str, unsigned int i)
241{
242 if (!strncmp(str, event_symbols[i].symbol,
243 strlen(event_symbols[i].symbol)))
244 return 1;
245
246 if (strlen(event_symbols[i].alias))
247 if (!strncmp(str, event_symbols[i].alias,
248 strlen(event_symbols[i].alias)))
249 return 1;
250 return 0;
251}
252
199/* 253/*
200 * Each event can have multiple symbolic names. 254 * Each event can have multiple symbolic names.
201 * Symbolic names are (almost) exactly matched. 255 * Symbolic names are (almost) exactly matched.
@@ -235,9 +289,7 @@ static int parse_event_symbols(const char *str, struct perf_counter_attr *attr)
235 } 289 }
236 290
237 for (i = 0; i < ARRAY_SIZE(event_symbols); i++) { 291 for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
238 if (!strncmp(str, event_symbols[i].symbol, 292 if (check_events(str, i)) {
239 strlen(event_symbols[i].symbol))) {
240
241 attr->type = event_symbols[i].type; 293 attr->type = event_symbols[i].type;
242 attr->config = event_symbols[i].config; 294 attr->config = event_symbols[i].config;
243 295
@@ -289,6 +341,7 @@ void print_events(void)
289{ 341{
290 struct event_symbol *syms = event_symbols; 342 struct event_symbol *syms = event_symbols;
291 unsigned int i, type, prev_type = -1; 343 unsigned int i, type, prev_type = -1;
344 char name[40];
292 345
293 fprintf(stderr, "\n"); 346 fprintf(stderr, "\n");
294 fprintf(stderr, "List of pre-defined events (to be used in -e):\n"); 347 fprintf(stderr, "List of pre-defined events (to be used in -e):\n");
@@ -301,14 +354,18 @@ void print_events(void)
301 if (type != prev_type) 354 if (type != prev_type)
302 fprintf(stderr, "\n"); 355 fprintf(stderr, "\n");
303 356
304 fprintf(stderr, " %-30s [%s]\n", syms->symbol, 357 if (strlen(syms->alias))
358 sprintf(name, "%s OR %s", syms->symbol, syms->alias);
359 else
360 strcpy(name, syms->symbol);
361 fprintf(stderr, " %-40s [%s]\n", name,
305 event_type_descriptors[type]); 362 event_type_descriptors[type]);
306 363
307 prev_type = type; 364 prev_type = type;
308 } 365 }
309 366
310 fprintf(stderr, "\n"); 367 fprintf(stderr, "\n");
311 fprintf(stderr, " %-30s [raw hardware event descriptor]\n", 368 fprintf(stderr, " %-40s [raw hardware event descriptor]\n",
312 "rNNN"); 369 "rNNN");
313 fprintf(stderr, "\n"); 370 fprintf(stderr, "\n");
314 371
diff --git a/tools/perf/util/run-command.c b/tools/perf/util/run-command.c
index b2f5e854f40a..a3935343091a 100644
--- a/tools/perf/util/run-command.c
+++ b/tools/perf/util/run-command.c
@@ -65,7 +65,6 @@ int start_command(struct child_process *cmd)
65 cmd->err = fderr[0]; 65 cmd->err = fderr[0];
66 } 66 }
67 67
68#ifndef __MINGW32__
69 fflush(NULL); 68 fflush(NULL);
70 cmd->pid = fork(); 69 cmd->pid = fork();
71 if (!cmd->pid) { 70 if (!cmd->pid) {
@@ -118,71 +117,6 @@ int start_command(struct child_process *cmd)
118 } 117 }
119 exit(127); 118 exit(127);
120 } 119 }
121#else
122 int s0 = -1, s1 = -1, s2 = -1; /* backups of stdin, stdout, stderr */
123 const char **sargv = cmd->argv;
124 char **env = environ;
125
126 if (cmd->no_stdin) {
127 s0 = dup(0);
128 dup_devnull(0);
129 } else if (need_in) {
130 s0 = dup(0);
131 dup2(fdin[0], 0);
132 } else if (cmd->in) {
133 s0 = dup(0);
134 dup2(cmd->in, 0);
135 }
136
137 if (cmd->no_stderr) {
138 s2 = dup(2);
139 dup_devnull(2);
140 } else if (need_err) {
141 s2 = dup(2);
142 dup2(fderr[1], 2);
143 }
144
145 if (cmd->no_stdout) {
146 s1 = dup(1);
147 dup_devnull(1);
148 } else if (cmd->stdout_to_stderr) {
149 s1 = dup(1);
150 dup2(2, 1);
151 } else if (need_out) {
152 s1 = dup(1);
153 dup2(fdout[1], 1);
154 } else if (cmd->out > 1) {
155 s1 = dup(1);
156 dup2(cmd->out, 1);
157 }
158
159 if (cmd->dir)
160 die("chdir in start_command() not implemented");
161 if (cmd->env) {
162 env = copy_environ();
163 for (; *cmd->env; cmd->env++)
164 env = env_setenv(env, *cmd->env);
165 }
166
167 if (cmd->perf_cmd) {
168 cmd->argv = prepare_perf_cmd(cmd->argv);
169 }
170
171 cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, env);
172
173 if (cmd->env)
174 free_environ(env);
175 if (cmd->perf_cmd)
176 free(cmd->argv);
177
178 cmd->argv = sargv;
179 if (s0 >= 0)
180 dup2(s0, 0), close(s0);
181 if (s1 >= 0)
182 dup2(s1, 1), close(s1);
183 if (s2 >= 0)
184 dup2(s2, 2), close(s2);
185#endif
186 120
187 if (cmd->pid < 0) { 121 if (cmd->pid < 0) {
188 int err = errno; 122 int err = errno;
@@ -288,14 +222,6 @@ int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const
288 return run_command(&cmd); 222 return run_command(&cmd);
289} 223}
290 224
291#ifdef __MINGW32__
292static __stdcall unsigned run_thread(void *data)
293{
294 struct async *async = data;
295 return async->proc(async->fd_for_proc, async->data);
296}
297#endif
298
299int start_async(struct async *async) 225int start_async(struct async *async)
300{ 226{
301 int pipe_out[2]; 227 int pipe_out[2];
@@ -304,7 +230,6 @@ int start_async(struct async *async)
304 return error("cannot create pipe: %s", strerror(errno)); 230 return error("cannot create pipe: %s", strerror(errno));
305 async->out = pipe_out[0]; 231 async->out = pipe_out[0];
306 232
307#ifndef __MINGW32__
308 /* Flush stdio before fork() to avoid cloning buffers */ 233 /* Flush stdio before fork() to avoid cloning buffers */
309 fflush(NULL); 234 fflush(NULL);
310 235
@@ -319,33 +244,17 @@ int start_async(struct async *async)
319 exit(!!async->proc(pipe_out[1], async->data)); 244 exit(!!async->proc(pipe_out[1], async->data));
320 } 245 }
321 close(pipe_out[1]); 246 close(pipe_out[1]);
322#else 247
323 async->fd_for_proc = pipe_out[1];
324 async->tid = (HANDLE) _beginthreadex(NULL, 0, run_thread, async, 0, NULL);
325 if (!async->tid) {
326 error("cannot create thread: %s", strerror(errno));
327 close_pair(pipe_out);
328 return -1;
329 }
330#endif
331 return 0; 248 return 0;
332} 249}
333 250
334int finish_async(struct async *async) 251int finish_async(struct async *async)
335{ 252{
336#ifndef __MINGW32__
337 int ret = 0; 253 int ret = 0;
338 254
339 if (wait_or_whine(async->pid)) 255 if (wait_or_whine(async->pid))
340 ret = error("waitpid (async) failed"); 256 ret = error("waitpid (async) failed");
341#else 257
342 DWORD ret = 0;
343 if (WaitForSingleObject(async->tid, INFINITE) != WAIT_OBJECT_0)
344 ret = error("waiting for thread failed: %lu", GetLastError());
345 else if (!GetExitCodeThread(async->tid, &ret))
346 ret = error("cannot get thread exit code: %lu", GetLastError());
347 CloseHandle(async->tid);
348#endif
349 return ret; 258 return ret;
350} 259}
351 260
diff --git a/tools/perf/util/run-command.h b/tools/perf/util/run-command.h
index 328289f23669..cc1837deba88 100644
--- a/tools/perf/util/run-command.h
+++ b/tools/perf/util/run-command.h
@@ -79,12 +79,7 @@ struct async {
79 int (*proc)(int fd, void *data); 79 int (*proc)(int fd, void *data);
80 void *data; 80 void *data;
81 int out; /* caller reads from here and closes it */ 81 int out; /* caller reads from here and closes it */
82#ifndef __MINGW32__
83 pid_t pid; 82 pid_t pid;
84#else
85 HANDLE tid;
86 int fd_for_proc;
87#endif
88}; 83};
89 84
90int start_async(struct async *async); 85int start_async(struct async *async);
diff --git a/tools/perf/util/strbuf.c b/tools/perf/util/strbuf.c
index eaba09306802..464e7ca898cf 100644
--- a/tools/perf/util/strbuf.c
+++ b/tools/perf/util/strbuf.c
@@ -259,7 +259,7 @@ size_t strbuf_fread(struct strbuf *sb, size_t size, FILE *f)
259 res = fread(sb->buf + sb->len, 1, size, f); 259 res = fread(sb->buf + sb->len, 1, size, f);
260 if (res > 0) 260 if (res > 0)
261 strbuf_setlen(sb, sb->len + res); 261 strbuf_setlen(sb, sb->len + res);
262 else if (res < 0 && oldalloc == 0) 262 else if (oldalloc == 0)
263 strbuf_release(sb); 263 strbuf_release(sb);
264 return res; 264 return res;
265} 265}
diff --git a/tools/perf/util/string.h b/tools/perf/util/string.h
index 37b03255b425..3dca2f654cd0 100644
--- a/tools/perf/util/string.h
+++ b/tools/perf/util/string.h
@@ -1,7 +1,7 @@
1#ifndef _PERF_STRING_H_ 1#ifndef _PERF_STRING_H_
2#define _PERF_STRING_H_ 2#define _PERF_STRING_H_
3 3
4#include "../types.h" 4#include "types.h"
5 5
6int hex2u64(const char *ptr, u64 *val); 6int hex2u64(const char *ptr, u64 *val);
7 7
diff --git a/tools/perf/util/strlist.c b/tools/perf/util/strlist.c
new file mode 100644
index 000000000000..025a78edfffe
--- /dev/null
+++ b/tools/perf/util/strlist.c
@@ -0,0 +1,184 @@
1/*
2 * (c) 2009 Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Licensed under the GPLv2.
5 */
6
7#include "strlist.h"
8#include <errno.h>
9#include <stdio.h>
10#include <stdlib.h>
11#include <string.h>
12
13static struct str_node *str_node__new(const char *s, bool dupstr)
14{
15 struct str_node *self = malloc(sizeof(*self));
16
17 if (self != NULL) {
18 if (dupstr) {
19 s = strdup(s);
20 if (s == NULL)
21 goto out_delete;
22 }
23 self->s = s;
24 }
25
26 return self;
27
28out_delete:
29 free(self);
30 return NULL;
31}
32
33static void str_node__delete(struct str_node *self, bool dupstr)
34{
35 if (dupstr)
36 free((void *)self->s);
37 free(self);
38}
39
40int strlist__add(struct strlist *self, const char *new_entry)
41{
42 struct rb_node **p = &self->entries.rb_node;
43 struct rb_node *parent = NULL;
44 struct str_node *sn;
45
46 while (*p != NULL) {
47 int rc;
48
49 parent = *p;
50 sn = rb_entry(parent, struct str_node, rb_node);
51 rc = strcmp(sn->s, new_entry);
52
53 if (rc > 0)
54 p = &(*p)->rb_left;
55 else if (rc < 0)
56 p = &(*p)->rb_right;
57 else
58 return -EEXIST;
59 }
60
61 sn = str_node__new(new_entry, self->dupstr);
62 if (sn == NULL)
63 return -ENOMEM;
64
65 rb_link_node(&sn->rb_node, parent, p);
66 rb_insert_color(&sn->rb_node, &self->entries);
67
68 return 0;
69}
70
71int strlist__load(struct strlist *self, const char *filename)
72{
73 char entry[1024];
74 int err;
75 FILE *fp = fopen(filename, "r");
76
77 if (fp == NULL)
78 return errno;
79
80 while (fgets(entry, sizeof(entry), fp) != NULL) {
81 const size_t len = strlen(entry);
82
83 if (len == 0)
84 continue;
85 entry[len - 1] = '\0';
86
87 err = strlist__add(self, entry);
88 if (err != 0)
89 goto out;
90 }
91
92 err = 0;
93out:
94 fclose(fp);
95 return err;
96}
97
98void strlist__remove(struct strlist *self, struct str_node *sn)
99{
100 rb_erase(&sn->rb_node, &self->entries);
101 str_node__delete(sn, self->dupstr);
102}
103
104bool strlist__has_entry(struct strlist *self, const char *entry)
105{
106 struct rb_node **p = &self->entries.rb_node;
107 struct rb_node *parent = NULL;
108
109 while (*p != NULL) {
110 struct str_node *sn;
111 int rc;
112
113 parent = *p;
114 sn = rb_entry(parent, struct str_node, rb_node);
115 rc = strcmp(sn->s, entry);
116
117 if (rc > 0)
118 p = &(*p)->rb_left;
119 else if (rc < 0)
120 p = &(*p)->rb_right;
121 else
122 return true;
123 }
124
125 return false;
126}
127
128static int strlist__parse_list_entry(struct strlist *self, const char *s)
129{
130 if (strncmp(s, "file://", 7) == 0)
131 return strlist__load(self, s + 7);
132
133 return strlist__add(self, s);
134}
135
136int strlist__parse_list(struct strlist *self, const char *s)
137{
138 char *sep;
139 int err;
140
141 while ((sep = strchr(s, ',')) != NULL) {
142 *sep = '\0';
143 err = strlist__parse_list_entry(self, s);
144 *sep = ',';
145 if (err != 0)
146 return err;
147 s = sep + 1;
148 }
149
150 return *s ? strlist__parse_list_entry(self, s) : 0;
151}
152
153struct strlist *strlist__new(bool dupstr, const char *slist)
154{
155 struct strlist *self = malloc(sizeof(*self));
156
157 if (self != NULL) {
158 self->entries = RB_ROOT;
159 self->dupstr = dupstr;
160 if (slist && strlist__parse_list(self, slist) != 0)
161 goto out_error;
162 }
163
164 return self;
165out_error:
166 free(self);
167 return NULL;
168}
169
170void strlist__delete(struct strlist *self)
171{
172 if (self != NULL) {
173 struct str_node *pos;
174 struct rb_node *next = rb_first(&self->entries);
175
176 while (next) {
177 pos = rb_entry(next, struct str_node, rb_node);
178 next = rb_next(&pos->rb_node);
179 strlist__remove(self, pos);
180 }
181 self->entries = RB_ROOT;
182 free(self);
183 }
184}
diff --git a/tools/perf/util/strlist.h b/tools/perf/util/strlist.h
new file mode 100644
index 000000000000..2fb117fb4b67
--- /dev/null
+++ b/tools/perf/util/strlist.h
@@ -0,0 +1,32 @@
1#ifndef STRLIST_H_
2#define STRLIST_H_
3
4#include "rbtree.h"
5#include <stdbool.h>
6
7struct str_node {
8 struct rb_node rb_node;
9 const char *s;
10};
11
12struct strlist {
13 struct rb_root entries;
14 bool dupstr;
15};
16
17struct strlist *strlist__new(bool dupstr, const char *slist);
18void strlist__delete(struct strlist *self);
19
20void strlist__remove(struct strlist *self, struct str_node *sn);
21int strlist__load(struct strlist *self, const char *filename);
22int strlist__add(struct strlist *self, const char *str);
23
24bool strlist__has_entry(struct strlist *self, const char *entry);
25
26static inline bool strlist__empty(const struct strlist *self)
27{
28 return rb_first(&self->entries) == NULL;
29}
30
31int strlist__parse_list(struct strlist *self, const char *s);
32#endif /* STRLIST_H_ */
diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c
index 86e14375e74e..78c2efde01b7 100644
--- a/tools/perf/util/symbol.c
+++ b/tools/perf/util/symbol.c
@@ -520,7 +520,9 @@ static int dso__load_sym(struct dso *self, int fd, const char *name,
520 nr_syms = shdr.sh_size / shdr.sh_entsize; 520 nr_syms = shdr.sh_size / shdr.sh_entsize;
521 521
522 memset(&sym, 0, sizeof(sym)); 522 memset(&sym, 0, sizeof(sym));
523 523 self->prelinked = elf_section_by_name(elf, &ehdr, &shdr,
524 ".gnu.prelink_undo",
525 NULL) != NULL;
524 elf_symtab__for_each_symbol(syms, nr_syms, index, sym) { 526 elf_symtab__for_each_symbol(syms, nr_syms, index, sym) {
525 struct symbol *f; 527 struct symbol *f;
526 u64 obj_start; 528 u64 obj_start;
@@ -535,7 +537,13 @@ static int dso__load_sym(struct dso *self, int fd, const char *name,
535 gelf_getshdr(sec, &shdr); 537 gelf_getshdr(sec, &shdr);
536 obj_start = sym.st_value; 538 obj_start = sym.st_value;
537 539
538 sym.st_value -= shdr.sh_addr - shdr.sh_offset; 540 if (self->prelinked) {
541 if (verbose >= 2)
542 printf("adjusting symbol: st_value: %Lx sh_addr: %Lx sh_offset: %Lx\n",
543 (u64)sym.st_value, (u64)shdr.sh_addr, (u64)shdr.sh_offset);
544
545 sym.st_value -= shdr.sh_addr - shdr.sh_offset;
546 }
539 547
540 f = symbol__new(sym.st_value, sym.st_size, 548 f = symbol__new(sym.st_value, sym.st_size,
541 elf_sym__name(&sym, symstrs), 549 elf_sym__name(&sym, symstrs),
@@ -569,6 +577,8 @@ int dso__load(struct dso *self, symbol_filter_t filter, int verbose)
569 if (!name) 577 if (!name)
570 return -1; 578 return -1;
571 579
580 self->prelinked = 0;
581
572 if (strncmp(self->name, "/tmp/perf-", 10) == 0) 582 if (strncmp(self->name, "/tmp/perf-", 10) == 0)
573 return dso__load_perf_map(self, filter, verbose); 583 return dso__load_perf_map(self, filter, verbose);
574 584
@@ -629,7 +639,7 @@ int dso__load_kernel(struct dso *self, const char *vmlinux,
629 if (vmlinux) 639 if (vmlinux)
630 err = dso__load_vmlinux(self, vmlinux, filter, verbose); 640 err = dso__load_vmlinux(self, vmlinux, filter, verbose);
631 641
632 if (err) 642 if (err < 0)
633 err = dso__load_kallsyms(self, filter, verbose); 643 err = dso__load_kallsyms(self, filter, verbose);
634 644
635 return err; 645 return err;
diff --git a/tools/perf/util/symbol.h b/tools/perf/util/symbol.h
index ea332e56e458..2c48ace8203b 100644
--- a/tools/perf/util/symbol.h
+++ b/tools/perf/util/symbol.h
@@ -2,7 +2,7 @@
2#define _PERF_SYMBOL_ 1 2#define _PERF_SYMBOL_ 1
3 3
4#include <linux/types.h> 4#include <linux/types.h>
5#include "../types.h" 5#include "types.h"
6#include "list.h" 6#include "list.h"
7#include "rbtree.h" 7#include "rbtree.h"
8 8
@@ -20,8 +20,9 @@ struct symbol {
20struct dso { 20struct dso {
21 struct list_head node; 21 struct list_head node;
22 struct rb_root syms; 22 struct rb_root syms;
23 unsigned int sym_priv_size;
24 struct symbol *(*find_symbol)(struct dso *, u64 ip); 23 struct symbol *(*find_symbol)(struct dso *, u64 ip);
24 unsigned int sym_priv_size;
25 unsigned char prelinked;
25 char name[0]; 26 char name[0];
26}; 27};
27 28
diff --git a/tools/perf/types.h b/tools/perf/util/types.h
index 5e75f9005940..5e75f9005940 100644
--- a/tools/perf/types.h
+++ b/tools/perf/util/types.h
diff --git a/tools/perf/util/util.h b/tools/perf/util/util.h
index b8cfed776d81..b4be6071c105 100644
--- a/tools/perf/util/util.h
+++ b/tools/perf/util/util.h
@@ -67,7 +67,6 @@
67#include <assert.h> 67#include <assert.h>
68#include <regex.h> 68#include <regex.h>
69#include <utime.h> 69#include <utime.h>
70#ifndef __MINGW32__
71#include <sys/wait.h> 70#include <sys/wait.h>
72#include <sys/poll.h> 71#include <sys/poll.h>
73#include <sys/socket.h> 72#include <sys/socket.h>
@@ -81,20 +80,6 @@
81#include <netdb.h> 80#include <netdb.h>
82#include <pwd.h> 81#include <pwd.h>
83#include <inttypes.h> 82#include <inttypes.h>
84#if defined(__CYGWIN__)
85#undef _XOPEN_SOURCE
86#include <grp.h>
87#define _XOPEN_SOURCE 600
88#include "compat/cygwin.h"
89#else
90#undef _ALL_SOURCE /* AIX 5.3L defines a struct list with _ALL_SOURCE. */
91#include <grp.h>
92#define _ALL_SOURCE 1
93#endif
94#else /* __MINGW32__ */
95/* pull in Windows compatibility stuff */
96#include "compat/mingw.h"
97#endif /* __MINGW32__ */
98 83
99#ifndef NO_ICONV 84#ifndef NO_ICONV
100#include <iconv.h> 85#include <iconv.h>
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-11 18:13:33 -0500