aboutsummaryrefslogtreecommitdiffstats
path: root/Documentation
diff options
context:
space:
mode:
authorIngo Molnar <mingo@elte.hu>2009-03-23 16:49:25 -0400
committerIngo Molnar <mingo@elte.hu>2009-04-06 03:30:29 -0400
commit383c5f8cd7d253be0017d8d5a39cbb78aaf70206 (patch)
treea93de6088c2748daf92cdc91426dd0ec47f25b29 /Documentation
parentbcbcb37cdb67d8100acfa66df40c4d636c28c4d1 (diff)
perf_counter tools: tidy up in-kernel dependencies
Remove now unified perfstat.c and perf_counter.h, and link to the in-kernel perf_counter.h. Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Paul Mackerras <paulus@samba.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Orig-LKML-Reference: <20090323172417.677932499@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'Documentation')
-rw-r--r--Documentation/perf_counter/Makefile2
-rw-r--r--Documentation/perf_counter/kerneltop.c2
-rw-r--r--Documentation/perf_counter/perfcounters.h142
-rw-r--r--Documentation/perf_counter/perfstat.c251
4 files changed, 2 insertions, 395 deletions
diff --git a/Documentation/perf_counter/Makefile b/Documentation/perf_counter/Makefile
index b45749753fcb..666da95a7877 100644
--- a/Documentation/perf_counter/Makefile
+++ b/Documentation/perf_counter/Makefile
@@ -2,7 +2,7 @@ BINS = kerneltop perfstat
2 2
3all: $(BINS) 3all: $(BINS)
4 4
5kerneltop: kerneltop.c perfcounters.h 5kerneltop: kerneltop.c ../../include/linux/perf_counter.h
6 cc -O6 -Wall -lrt `pkg-config --cflags --libs glib-2.0` -o $@ $< 6 cc -O6 -Wall -lrt `pkg-config --cflags --libs glib-2.0` -o $@ $<
7 7
8perfstat: kerneltop 8perfstat: kerneltop
diff --git a/Documentation/perf_counter/kerneltop.c b/Documentation/perf_counter/kerneltop.c
index 80b790553ec8..25e80bc44556 100644
--- a/Documentation/perf_counter/kerneltop.c
+++ b/Documentation/perf_counter/kerneltop.c
@@ -88,7 +88,7 @@
88 88
89#include <linux/unistd.h> 89#include <linux/unistd.h>
90 90
91#include "include/linux/perf_counter.h" 91#include "../../include/linux/perf_counter.h"
92 92
93 93
94/* 94/*
diff --git a/Documentation/perf_counter/perfcounters.h b/Documentation/perf_counter/perfcounters.h
deleted file mode 100644
index 32e24b9154ab..000000000000
--- a/Documentation/perf_counter/perfcounters.h
+++ /dev/null
@@ -1,142 +0,0 @@
1/*
2 * Ioctls that can be done on a perf counter fd:
3 */
4#define PERF_COUNTER_IOC_ENABLE _IO('$', 0)
5#define PERF_COUNTER_IOC_DISABLE _IO('$', 1)
6
7/*
8 * prctl(PR_TASK_PERF_COUNTERS_DISABLE) will (cheaply) disable all
9 * counters in the current task.
10 */
11#define PR_TASK_PERF_COUNTERS_DISABLE 31
12#define PR_TASK_PERF_COUNTERS_ENABLE 32
13
14#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
15
16#define rdclock() \
17({ \
18 struct timespec ts; \
19 \
20 clock_gettime(CLOCK_MONOTONIC, &ts); \
21 ts.tv_sec * 1000000000ULL + ts.tv_nsec; \
22})
23
24/*
25 * Pick up some kernel type conventions:
26 */
27#define __user
28#define asmlinkage
29
30typedef unsigned int __u32;
31typedef unsigned long long __u64;
32typedef long long __s64;
33
34/*
35 * User-space ABI bits:
36 */
37
38/*
39 * Generalized performance counter event types, used by the hw_event.type
40 * parameter of the sys_perf_counter_open() syscall:
41 */
42enum hw_event_types {
43 /*
44 * Common hardware events, generalized by the kernel:
45 */
46 PERF_COUNT_CPU_CYCLES = 0,
47 PERF_COUNT_INSTRUCTIONS = 1,
48 PERF_COUNT_CACHE_REFERENCES = 2,
49 PERF_COUNT_CACHE_MISSES = 3,
50 PERF_COUNT_BRANCH_INSTRUCTIONS = 4,
51 PERF_COUNT_BRANCH_MISSES = 5,
52 PERF_COUNT_BUS_CYCLES = 6,
53
54 PERF_HW_EVENTS_MAX = 7,
55
56 /*
57 * Special "software" counters provided by the kernel, even if
58 * the hardware does not support performance counters. These
59 * counters measure various physical and sw events of the
60 * kernel (and allow the profiling of them as well):
61 */
62 PERF_COUNT_CPU_CLOCK = -1,
63 PERF_COUNT_TASK_CLOCK = -2,
64 PERF_COUNT_PAGE_FAULTS = -3,
65 PERF_COUNT_CONTEXT_SWITCHES = -4,
66 PERF_COUNT_CPU_MIGRATIONS = -5,
67
68 PERF_SW_EVENTS_MIN = -6,
69};
70
71/*
72 * IRQ-notification data record type:
73 */
74enum perf_counter_record_type {
75 PERF_RECORD_SIMPLE = 0,
76 PERF_RECORD_IRQ = 1,
77 PERF_RECORD_GROUP = 2,
78};
79
80/*
81 * Hardware event to monitor via a performance monitoring counter:
82 */
83struct perf_counter_hw_event {
84 __s64 type;
85
86 __u64 irq_period;
87 __u64 record_type;
88 __u64 read_format;
89
90 __u64 disabled : 1, /* off by default */
91 nmi : 1, /* NMI sampling */
92 raw : 1, /* raw event type */
93 inherit : 1, /* children inherit it */
94 pinned : 1, /* must always be on PMU */
95 exclusive : 1, /* only group on PMU */
96 exclude_user : 1, /* don't count user */
97 exclude_kernel : 1, /* ditto kernel */
98 exclude_hv : 1, /* ditto hypervisor */
99 exclude_idle : 1, /* don't count when idle */
100
101 __reserved_1 : 54;
102
103 __u32 extra_config_len;
104 __u32 __reserved_4;
105
106 __u64 __reserved_2;
107 __u64 __reserved_3;
108};
109
110
111#ifdef __x86_64__
112# define __NR_perf_counter_open 295
113#endif
114
115#ifdef __i386__
116# define __NR_perf_counter_open 333
117#endif
118
119#ifdef __powerpc__
120#define __NR_perf_counter_open 319
121#endif
122
123asmlinkage int sys_perf_counter_open(
124
125 struct perf_counter_hw_event *hw_event_uptr __user,
126 pid_t pid,
127 int cpu,
128 int group_fd,
129 unsigned long flags)
130{
131 int ret;
132
133 ret = syscall(
134 __NR_perf_counter_open, hw_event_uptr, pid, cpu, group_fd, flags);
135#if defined(__x86_64__) || defined(__i386__)
136 if (ret < 0 && ret > -4096) {
137 errno = -ret;
138 ret = -1;
139 }
140#endif
141 return ret;
142}
diff --git a/Documentation/perf_counter/perfstat.c b/Documentation/perf_counter/perfstat.c
deleted file mode 100644
index fd594468e655..000000000000
--- a/Documentation/perf_counter/perfstat.c
+++ /dev/null
@@ -1,251 +0,0 @@
1/*
2 * perfstat: /usr/bin/time -alike performance counter statistics utility
3 *
4 * It summarizes the counter events of all tasks (and child tasks),
5 * covering all CPUs that the command (or workload) executes on.
6 * It only counts the per-task events of the workload started,
7 * independent of how many other tasks run on those CPUs.
8 *
9 * Build with: cc -O2 -g -lrt -Wall -W -o perfstat perfstat.c
10 *
11 * Sample output:
12 *
13
14 $ ./perfstat -e 1 -e 3 -e 5 ls -lR /usr/include/ >/dev/null
15
16 Performance counter stats for 'ls':
17
18 163516953 instructions
19 2295 cache-misses
20 2855182 branch-misses
21
22 *
23 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
24 *
25 * Released under the GPLv2 (not later).
26 *
27 * Percpu counter support by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
28 * Symbolic event options by: Wu Fengguang <fengguang.wu@intel.com>
29 */
30#define _GNU_SOURCE
31
32#include <assert.h>
33#include <getopt.h>
34#include <stdint.h>
35#include <stdlib.h>
36#include <string.h>
37#include <unistd.h>
38#include <ctype.h>
39#include <errno.h>
40#include <fcntl.h>
41#include <stdio.h>
42#include <time.h>
43
44#include <sys/syscall.h>
45#include <sys/ioctl.h>
46#include <sys/prctl.h>
47#include <sys/types.h>
48#include <sys/stat.h>
49#include <sys/time.h>
50#include <sys/wait.h>
51#include <sys/uio.h>
52
53#include <linux/unistd.h>
54
55#include "perfcounters.h"
56
57static int nr_cpus = 0;
58
59static int system_wide = 0;
60
61static void display_help(void)
62{
63 unsigned int i;
64 int e;
65
66 printf(
67 "Usage: perfstat [<events...>] <cmd...>\n\n"
68 "PerfStat Options (up to %d event types can be specified):\n\n",
69 MAX_COUNTERS);
70 printf(
71 " -e EVENT --event=EVENT # symbolic-name abbreviations");
72
73 for (i = 0, e = PERF_HW_EVENTS_MAX; i < ARRAY_SIZE(event_symbols); i++) {
74 if (e != event_symbols[i].event) {
75 e = event_symbols[i].event;
76 printf(
77 "\n %2d: %-20s", e, event_symbols[i].symbol);
78 } else
79 printf(" %s", event_symbols[i].symbol);
80 }
81
82 printf("\n"
83 " rNNN: raw event type\n\n"
84 " -s # system-wide collection\n\n"
85 " -c <cmd..> --command=<cmd..> # command+arguments to be timed.\n"
86 "\n");
87 exit(0);
88}
89
90static void process_options(int argc, char *argv[])
91{
92 for (;;) {
93 int option_index = 0;
94 /** Options for getopt */
95 static struct option long_options[] = {
96 {"event", required_argument, NULL, 'e'},
97 {"help", no_argument, NULL, 'h'},
98 {"command", no_argument, NULL, 'c'},
99 {NULL, 0, NULL, 0 }
100 };
101 int c = getopt_long(argc, argv, "+:e:c:s",
102 long_options, &option_index);
103 if (c == -1)
104 break;
105
106 switch (c) {
107 case 'c':
108 break;
109 case 's':
110 system_wide = 1;
111 break;
112 case 'e':
113 parse_events(optarg);
114 break;
115 default:
116 break;
117 }
118 }
119 if (optind == argc)
120 goto err;
121
122 if (!nr_counters)
123 nr_counters = 8;
124 return;
125
126err:
127 display_help();
128}
129
130char fault_here[1000000];
131
132static int fd[MAX_NR_CPUS][MAX_COUNTERS];
133
134static void create_counter(int counter)
135{
136 struct perf_counter_hw_event hw_event;
137
138 memset(&hw_event, 0, sizeof(hw_event));
139 hw_event.type = event_id[counter];
140 hw_event.raw = event_raw[counter];
141 hw_event.record_type = PERF_RECORD_SIMPLE;
142 hw_event.nmi = 0;
143
144 if (system_wide) {
145 int cpu;
146 for (cpu = 0; cpu < nr_cpus; cpu ++) {
147 fd[cpu][counter] = sys_perf_counter_open(&hw_event, -1, cpu, -1, 0);
148 if (fd[cpu][counter] < 0) {
149 printf("perfstat error: syscall returned with %d (%s)\n",
150 fd[cpu][counter], strerror(errno));
151 exit(-1);
152 }
153
154 }
155 } else {
156 hw_event.inherit = 1;
157 hw_event.disabled = 1;
158
159 fd[0][counter] = sys_perf_counter_open(&hw_event, 0, -1, -1, 0);
160 if (fd[0][counter] < 0) {
161 printf("perfstat error: syscall returned with %d (%s)\n",
162 fd[0][counter], strerror(errno));
163 exit(-1);
164 }
165 }
166}
167
168
169int main(int argc, char *argv[])
170{
171 unsigned long long t0, t1;
172 int counter;
173 ssize_t res;
174 int status;
175 int pid;
176
177 process_options(argc, argv);
178
179 if (system_wide) {
180 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
181 assert(nr_cpus <= MAX_NR_CPUS);
182 assert(nr_cpus >= 0);
183 } else
184 nr_cpus = 1;
185
186 for (counter = 0; counter < nr_counters; counter++)
187 create_counter(counter);
188
189 argc -= optind;
190 argv += optind;
191
192 /*
193 * Enable counters and exec the command:
194 */
195 t0 = rdclock();
196 prctl(PR_TASK_PERF_COUNTERS_ENABLE);
197
198 if ((pid = fork()) < 0)
199 perror("failed to fork");
200 if (!pid) {
201 if (execvp(argv[0], argv)) {
202 perror(argv[0]);
203 exit(-1);
204 }
205 }
206 while (wait(&status) >= 0)
207 ;
208 prctl(PR_TASK_PERF_COUNTERS_DISABLE);
209 t1 = rdclock();
210
211 fflush(stdout);
212
213 fprintf(stderr, "\n");
214 fprintf(stderr, " Performance counter stats for \'%s\':\n",
215 argv[0]);
216 fprintf(stderr, "\n");
217
218 for (counter = 0; counter < nr_counters; counter++) {
219 int cpu;
220 __u64 count, single_count;
221
222 count = 0;
223 for (cpu = 0; cpu < nr_cpus; cpu ++) {
224 res = read(fd[cpu][counter],
225 (char *) &single_count, sizeof(single_count));
226 assert(res == sizeof(single_count));
227 count += single_count;
228 }
229
230 if (!event_raw[counter] &&
231 (event_id[counter] == PERF_COUNT_CPU_CLOCK ||
232 event_id[counter] == PERF_COUNT_TASK_CLOCK)) {
233
234 double msecs = (double)count / 1000000;
235
236 fprintf(stderr, " %14.6f %-20s (msecs)\n",
237 msecs, event_name(counter));
238 } else {
239 fprintf(stderr, " %14Ld %-20s (events)\n",
240 count, event_name(counter));
241 }
242 if (!counter)
243 fprintf(stderr, "\n");
244 }
245 fprintf(stderr, "\n");
246 fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n",
247 (double)(t1-t0)/1e6);
248 fprintf(stderr, "\n");
249
250 return 0;
251}
generated by cgit v1.2.2 (git 2.25.0) at 2025-01-03 19:04:44 -0500