aboutsummaryrefslogtreecommitdiffstats
path: root/tools/perf/builtin-stat.c
diff options
context:
space:
mode:
Diffstat (limited to 'tools/perf/builtin-stat.c')
-rw-r--r--tools/perf/builtin-stat.c1075
1 files changed, 842 insertions, 233 deletions
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index a6b4d44f9502..a9f06715e44d 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -6,24 +6,28 @@
6 * 6 *
7 * Sample output: 7 * Sample output:
8 8
9 $ perf stat ~/hackbench 10 9 $ perf stat ./hackbench 10
10 Time: 0.104
11 10
12 Performance counter stats for '/home/mingo/hackbench': 11 Time: 0.118
13 12
14 1255.538611 task clock ticks # 10.143 CPU utilization factor 13 Performance counter stats for './hackbench 10':
15 54011 context switches # 0.043 M/sec
16 385 CPU migrations # 0.000 M/sec
17 17755 pagefaults # 0.014 M/sec
18 3808323185 CPU cycles # 3033.219 M/sec
19 1575111190 instructions # 1254.530 M/sec
20 17367895 cache references # 13.833 M/sec
21 7674421 cache misses # 6.112 M/sec
22 14
23 Wall-clock time elapsed: 123.786620 msecs 15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
24 28
25 * 29 *
26 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
27 * 31 *
28 * Improvements and fixes by: 32 * Improvements and fixes by:
29 * 33 *
@@ -43,15 +47,21 @@
43#include "util/parse-options.h" 47#include "util/parse-options.h"
44#include "util/parse-events.h" 48#include "util/parse-events.h"
45#include "util/event.h" 49#include "util/event.h"
50#include "util/evlist.h"
51#include "util/evsel.h"
46#include "util/debug.h" 52#include "util/debug.h"
53#include "util/color.h"
47#include "util/header.h" 54#include "util/header.h"
48#include "util/cpumap.h" 55#include "util/cpumap.h"
49#include "util/thread.h" 56#include "util/thread.h"
57#include "util/thread_map.h"
50 58
51#include <sys/prctl.h> 59#include <sys/prctl.h>
52#include <math.h> 60#include <math.h>
53#include <locale.h> 61#include <locale.h>
54 62
63#define DEFAULT_SEPARATOR " "
64
55static struct perf_event_attr default_attrs[] = { 65static struct perf_event_attr default_attrs[] = {
56 66
57 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 67 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
@@ -60,34 +70,127 @@ static struct perf_event_attr default_attrs[] = {
60 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 70 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
61 71
62 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 72 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
73 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
74 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
63 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 75 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
64 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 76 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
65 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 77 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
66 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES },
67 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
68 78
69}; 79};
70 80
81/*
82 * Detailed stats (-d), covering the L1 and last level data caches:
83 */
84static struct perf_event_attr detailed_attrs[] = {
85
86 { .type = PERF_TYPE_HW_CACHE,
87 .config =
88 PERF_COUNT_HW_CACHE_L1D << 0 |
89 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
90 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
91
92 { .type = PERF_TYPE_HW_CACHE,
93 .config =
94 PERF_COUNT_HW_CACHE_L1D << 0 |
95 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
96 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
97
98 { .type = PERF_TYPE_HW_CACHE,
99 .config =
100 PERF_COUNT_HW_CACHE_LL << 0 |
101 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
102 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
103
104 { .type = PERF_TYPE_HW_CACHE,
105 .config =
106 PERF_COUNT_HW_CACHE_LL << 0 |
107 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
108 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
109};
110
111/*
112 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
113 */
114static struct perf_event_attr very_detailed_attrs[] = {
115
116 { .type = PERF_TYPE_HW_CACHE,
117 .config =
118 PERF_COUNT_HW_CACHE_L1I << 0 |
119 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
120 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
121
122 { .type = PERF_TYPE_HW_CACHE,
123 .config =
124 PERF_COUNT_HW_CACHE_L1I << 0 |
125 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
126 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
127
128 { .type = PERF_TYPE_HW_CACHE,
129 .config =
130 PERF_COUNT_HW_CACHE_DTLB << 0 |
131 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
132 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
133
134 { .type = PERF_TYPE_HW_CACHE,
135 .config =
136 PERF_COUNT_HW_CACHE_DTLB << 0 |
137 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
138 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
139
140 { .type = PERF_TYPE_HW_CACHE,
141 .config =
142 PERF_COUNT_HW_CACHE_ITLB << 0 |
143 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
144 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
145
146 { .type = PERF_TYPE_HW_CACHE,
147 .config =
148 PERF_COUNT_HW_CACHE_ITLB << 0 |
149 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
150 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
151
152};
153
154/*
155 * Very, very detailed stats (-d -d -d), adding prefetch events:
156 */
157static struct perf_event_attr very_very_detailed_attrs[] = {
158
159 { .type = PERF_TYPE_HW_CACHE,
160 .config =
161 PERF_COUNT_HW_CACHE_L1D << 0 |
162 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
163 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
164
165 { .type = PERF_TYPE_HW_CACHE,
166 .config =
167 PERF_COUNT_HW_CACHE_L1D << 0 |
168 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
169 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
170};
171
172
173
174struct perf_evlist *evsel_list;
175
71static bool system_wide = false; 176static bool system_wide = false;
72static int nr_cpus = 0;
73static int run_idx = 0; 177static int run_idx = 0;
74 178
75static int run_count = 1; 179static int run_count = 1;
76static bool no_inherit = false; 180static bool no_inherit = false;
77static bool scale = true; 181static bool scale = true;
182static bool no_aggr = false;
78static pid_t target_pid = -1; 183static pid_t target_pid = -1;
79static pid_t target_tid = -1; 184static pid_t target_tid = -1;
80static pid_t *all_tids = NULL;
81static int thread_num = 0;
82static pid_t child_pid = -1; 185static pid_t child_pid = -1;
83static bool null_run = false; 186static bool null_run = false;
84static bool big_num = false; 187static int detailed_run = 0;
188static bool sync_run = false;
189static bool big_num = true;
190static int big_num_opt = -1;
85static const char *cpu_list; 191static const char *cpu_list;
86 192static const char *csv_sep = NULL;
87 193static bool csv_output = false;
88static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
89
90static int event_scaled[MAX_COUNTERS];
91 194
92static volatile int done = 0; 195static volatile int done = 0;
93 196
@@ -96,6 +199,22 @@ struct stats
96 double n, mean, M2; 199 double n, mean, M2;
97}; 200};
98 201
202struct perf_stat {
203 struct stats res_stats[3];
204};
205
206static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
207{
208 evsel->priv = zalloc(sizeof(struct perf_stat));
209 return evsel->priv == NULL ? -ENOMEM : 0;
210}
211
212static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
213{
214 free(evsel->priv);
215 evsel->priv = NULL;
216}
217
99static void update_stats(struct stats *stats, u64 val) 218static void update_stats(struct stats *stats, u64 val)
100{ 219{
101 double delta; 220 double delta;
@@ -135,154 +254,143 @@ static double stddev_stats(struct stats *stats)
135 return sqrt(variance_mean); 254 return sqrt(variance_mean);
136} 255}
137 256
138struct stats event_res_stats[MAX_COUNTERS][3]; 257struct stats runtime_nsecs_stats[MAX_NR_CPUS];
139struct stats runtime_nsecs_stats; 258struct stats runtime_cycles_stats[MAX_NR_CPUS];
259struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
260struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
261struct stats runtime_branches_stats[MAX_NR_CPUS];
262struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
263struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
264struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
265struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
266struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
267struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
140struct stats walltime_nsecs_stats; 268struct stats walltime_nsecs_stats;
141struct stats runtime_cycles_stats;
142struct stats runtime_branches_stats;
143
144#define MATCH_EVENT(t, c, counter) \
145 (attrs[counter].type == PERF_TYPE_##t && \
146 attrs[counter].config == PERF_COUNT_##c)
147 269
148#define ERR_PERF_OPEN \ 270static int create_perf_stat_counter(struct perf_evsel *evsel)
149"Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
150
151static int create_perf_stat_counter(int counter)
152{ 271{
153 struct perf_event_attr *attr = attrs + counter; 272 struct perf_event_attr *attr = &evsel->attr;
154 int thread;
155 int ncreated = 0;
156 273
157 if (scale) 274 if (scale)
158 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 275 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
159 PERF_FORMAT_TOTAL_TIME_RUNNING; 276 PERF_FORMAT_TOTAL_TIME_RUNNING;
160 277
161 if (system_wide) { 278 attr->inherit = !no_inherit;
162 int cpu; 279
163 280 if (system_wide)
164 for (cpu = 0; cpu < nr_cpus; cpu++) { 281 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
165 fd[cpu][counter][0] = sys_perf_event_open(attr, 282
166 -1, cpumap[cpu], -1, 0); 283 if (target_pid == -1 && target_tid == -1) {
167 if (fd[cpu][counter][0] < 0) 284 attr->disabled = 1;
168 pr_debug(ERR_PERF_OPEN, counter, 285 attr->enable_on_exec = 1;
169 fd[cpu][counter][0], strerror(errno));
170 else
171 ++ncreated;
172 }
173 } else {
174 attr->inherit = !no_inherit;
175 if (target_pid == -1 && target_tid == -1) {
176 attr->disabled = 1;
177 attr->enable_on_exec = 1;
178 }
179 for (thread = 0; thread < thread_num; thread++) {
180 fd[0][counter][thread] = sys_perf_event_open(attr,
181 all_tids[thread], -1, -1, 0);
182 if (fd[0][counter][thread] < 0)
183 pr_debug(ERR_PERF_OPEN, counter,
184 fd[0][counter][thread],
185 strerror(errno));
186 else
187 ++ncreated;
188 }
189 } 286 }
190 287
191 return ncreated; 288 return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
192} 289}
193 290
194/* 291/*
195 * Does the counter have nsecs as a unit? 292 * Does the counter have nsecs as a unit?
196 */ 293 */
197static inline int nsec_counter(int counter) 294static inline int nsec_counter(struct perf_evsel *evsel)
198{ 295{
199 if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) || 296 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
200 MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) 297 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
201 return 1; 298 return 1;
202 299
203 return 0; 300 return 0;
204} 301}
205 302
206/* 303/*
304 * Update various tracking values we maintain to print
305 * more semantic information such as miss/hit ratios,
306 * instruction rates, etc:
307 */
308static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
309{
310 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
311 update_stats(&runtime_nsecs_stats[0], count[0]);
312 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
313 update_stats(&runtime_cycles_stats[0], count[0]);
314 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
315 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
316 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
317 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
318 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
319 update_stats(&runtime_branches_stats[0], count[0]);
320 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
321 update_stats(&runtime_cacherefs_stats[0], count[0]);
322 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
323 update_stats(&runtime_l1_dcache_stats[0], count[0]);
324 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
325 update_stats(&runtime_l1_icache_stats[0], count[0]);
326 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
327 update_stats(&runtime_ll_cache_stats[0], count[0]);
328 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
329 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
330 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
331 update_stats(&runtime_itlb_cache_stats[0], count[0]);
332}
333
334/*
207 * Read out the results of a single counter: 335 * Read out the results of a single counter:
336 * aggregate counts across CPUs in system-wide mode
208 */ 337 */
209static void read_counter(int counter) 338static int read_counter_aggr(struct perf_evsel *counter)
210{ 339{
211 u64 count[3], single_count[3]; 340 struct perf_stat *ps = counter->priv;
212 int cpu; 341 u64 *count = counter->counts->aggr.values;
213 size_t res, nv; 342 int i;
214 int scaled;
215 int i, thread;
216 343
217 count[0] = count[1] = count[2] = 0; 344 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
345 evsel_list->threads->nr, scale) < 0)
346 return -1;
218 347
219 nv = scale ? 3 : 1; 348 for (i = 0; i < 3; i++)
220 for (cpu = 0; cpu < nr_cpus; cpu++) { 349 update_stats(&ps->res_stats[i], count[i]);
221 for (thread = 0; thread < thread_num; thread++) {
222 if (fd[cpu][counter][thread] < 0)
223 continue;
224 350
225 res = read(fd[cpu][counter][thread], 351 if (verbose) {
226 single_count, nv * sizeof(u64)); 352 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
227 assert(res == nv * sizeof(u64)); 353 event_name(counter), count[0], count[1], count[2]);
354 }
228 355
229 close(fd[cpu][counter][thread]); 356 /*
230 fd[cpu][counter][thread] = -1; 357 * Save the full runtime - to allow normalization during printout:
358 */
359 update_shadow_stats(counter, count);
231 360
232 count[0] += single_count[0]; 361 return 0;
233 if (scale) { 362}
234 count[1] += single_count[1];
235 count[2] += single_count[2];
236 }
237 }
238 }
239 363
240 scaled = 0; 364/*
241 if (scale) { 365 * Read out the results of a single counter:
242 if (count[2] == 0) { 366 * do not aggregate counts across CPUs in system-wide mode
243 event_scaled[counter] = -1; 367 */
244 count[0] = 0; 368static int read_counter(struct perf_evsel *counter)
245 return; 369{
246 } 370 u64 *count;
371 int cpu;
247 372
248 if (count[2] < count[1]) { 373 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
249 event_scaled[counter] = 1; 374 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
250 count[0] = (unsigned long long) 375 return -1;
251 ((double)count[0] * count[1] / count[2] + 0.5);
252 }
253 }
254 376
255 for (i = 0; i < 3; i++) 377 count = counter->counts->cpu[cpu].values;
256 update_stats(&event_res_stats[counter][i], count[i]);
257 378
258 if (verbose) { 379 update_shadow_stats(counter, count);
259 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
260 count[0], count[1], count[2]);
261 } 380 }
262 381
263 /* 382 return 0;
264 * Save the full runtime - to allow normalization during printout:
265 */
266 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
267 update_stats(&runtime_nsecs_stats, count[0]);
268 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
269 update_stats(&runtime_cycles_stats, count[0]);
270 if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
271 update_stats(&runtime_branches_stats, count[0]);
272} 383}
273 384
274static int run_perf_stat(int argc __used, const char **argv) 385static int run_perf_stat(int argc __used, const char **argv)
275{ 386{
276 unsigned long long t0, t1; 387 unsigned long long t0, t1;
388 struct perf_evsel *counter;
277 int status = 0; 389 int status = 0;
278 int counter, ncreated = 0;
279 int child_ready_pipe[2], go_pipe[2]; 390 int child_ready_pipe[2], go_pipe[2];
280 const bool forks = (argc > 0); 391 const bool forks = (argc > 0);
281 char buf; 392 char buf;
282 393
283 if (!system_wide)
284 nr_cpus = 1;
285
286 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 394 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
287 perror("failed to create pipes"); 395 perror("failed to create pipes");
288 exit(1); 396 exit(1);
@@ -322,7 +430,7 @@ static int run_perf_stat(int argc __used, const char **argv)
322 } 430 }
323 431
324 if (target_tid == -1 && target_pid == -1 && !system_wide) 432 if (target_tid == -1 && target_pid == -1 && !system_wide)
325 all_tids[0] = child_pid; 433 evsel_list->threads->map[0] = child_pid;
326 434
327 /* 435 /*
328 * Wait for the child to be ready to exec. 436 * Wait for the child to be ready to exec.
@@ -334,15 +442,35 @@ static int run_perf_stat(int argc __used, const char **argv)
334 close(child_ready_pipe[0]); 442 close(child_ready_pipe[0]);
335 } 443 }
336 444
337 for (counter = 0; counter < nr_counters; counter++) 445 list_for_each_entry(counter, &evsel_list->entries, node) {
338 ncreated += create_perf_stat_counter(counter); 446 if (create_perf_stat_counter(counter) < 0) {
447 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
448 if (verbose)
449 ui__warning("%s event is not supported by the kernel.\n",
450 event_name(counter));
451 continue;
452 }
453
454 if (errno == EPERM || errno == EACCES) {
455 error("You may not have permission to collect %sstats.\n"
456 "\t Consider tweaking"
457 " /proc/sys/kernel/perf_event_paranoid or running as root.",
458 system_wide ? "system-wide " : "");
459 } else {
460 error("open_counter returned with %d (%s). "
461 "/bin/dmesg may provide additional information.\n",
462 errno, strerror(errno));
463 }
464 if (child_pid != -1)
465 kill(child_pid, SIGTERM);
466 die("Not all events could be opened.\n");
467 return -1;
468 }
469 }
339 470
340 if (ncreated == 0) { 471 if (perf_evlist__set_filters(evsel_list)) {
341 pr_err("No permission to collect %sstats.\n" 472 error("failed to set filter with %d (%s)\n", errno,
342 "Consider tweaking /proc/sys/kernel/perf_event_paranoid.\n", 473 strerror(errno));
343 system_wide ? "system-wide " : "");
344 if (child_pid != -1)
345 kill(child_pid, SIGTERM);
346 return -1; 474 return -1;
347 } 475 }
348 476
@@ -362,136 +490,501 @@ static int run_perf_stat(int argc __used, const char **argv)
362 490
363 update_stats(&walltime_nsecs_stats, t1 - t0); 491 update_stats(&walltime_nsecs_stats, t1 - t0);
364 492
365 for (counter = 0; counter < nr_counters; counter++) 493 if (no_aggr) {
366 read_counter(counter); 494 list_for_each_entry(counter, &evsel_list->entries, node) {
495 read_counter(counter);
496 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
497 }
498 } else {
499 list_for_each_entry(counter, &evsel_list->entries, node) {
500 read_counter_aggr(counter);
501 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
502 evsel_list->threads->nr);
503 }
504 }
367 505
368 return WEXITSTATUS(status); 506 return WEXITSTATUS(status);
369} 507}
370 508
371static void print_noise(int counter, double avg) 509static void print_noise_pct(double total, double avg)
510{
511 double pct = 0.0;
512
513 if (avg)
514 pct = 100.0*total/avg;
515
516 fprintf(stderr, " ( +-%6.2f%% )", pct);
517}
518
519static void print_noise(struct perf_evsel *evsel, double avg)
372{ 520{
521 struct perf_stat *ps;
522
373 if (run_count == 1) 523 if (run_count == 1)
374 return; 524 return;
375 525
376 fprintf(stderr, " ( +- %7.3f%% )", 526 ps = evsel->priv;
377 100 * stddev_stats(&event_res_stats[counter][0]) / avg); 527 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
378} 528}
379 529
380static void nsec_printout(int counter, double avg) 530static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
381{ 531{
382 double msecs = avg / 1e6; 532 double msecs = avg / 1e6;
533 char cpustr[16] = { '\0', };
534 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
383 535
384 fprintf(stderr, " %18.6f %-24s", msecs, event_name(counter)); 536 if (no_aggr)
537 sprintf(cpustr, "CPU%*d%s",
538 csv_output ? 0 : -4,
539 evsel_list->cpus->map[cpu], csv_sep);
385 540
386 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) { 541 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
387 fprintf(stderr, " # %10.3f CPUs ", 542
388 avg / avg_stats(&walltime_nsecs_stats)); 543 if (evsel->cgrp)
389 } 544 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
545
546 if (csv_output)
547 return;
548
549 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
550 fprintf(stderr, " # %8.3f CPUs utilized ", avg / avg_stats(&walltime_nsecs_stats));
551}
552
553static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
554{
555 double total, ratio = 0.0;
556 const char *color;
557
558 total = avg_stats(&runtime_cycles_stats[cpu]);
559
560 if (total)
561 ratio = avg / total * 100.0;
562
563 color = PERF_COLOR_NORMAL;
564 if (ratio > 50.0)
565 color = PERF_COLOR_RED;
566 else if (ratio > 30.0)
567 color = PERF_COLOR_MAGENTA;
568 else if (ratio > 10.0)
569 color = PERF_COLOR_YELLOW;
570
571 fprintf(stderr, " # ");
572 color_fprintf(stderr, color, "%6.2f%%", ratio);
573 fprintf(stderr, " frontend cycles idle ");
574}
575
576static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
577{
578 double total, ratio = 0.0;
579 const char *color;
580
581 total = avg_stats(&runtime_cycles_stats[cpu]);
582
583 if (total)
584 ratio = avg / total * 100.0;
585
586 color = PERF_COLOR_NORMAL;
587 if (ratio > 75.0)
588 color = PERF_COLOR_RED;
589 else if (ratio > 50.0)
590 color = PERF_COLOR_MAGENTA;
591 else if (ratio > 20.0)
592 color = PERF_COLOR_YELLOW;
593
594 fprintf(stderr, " # ");
595 color_fprintf(stderr, color, "%6.2f%%", ratio);
596 fprintf(stderr, " backend cycles idle ");
597}
598
599static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
600{
601 double total, ratio = 0.0;
602 const char *color;
603
604 total = avg_stats(&runtime_branches_stats[cpu]);
605
606 if (total)
607 ratio = avg / total * 100.0;
608
609 color = PERF_COLOR_NORMAL;
610 if (ratio > 20.0)
611 color = PERF_COLOR_RED;
612 else if (ratio > 10.0)
613 color = PERF_COLOR_MAGENTA;
614 else if (ratio > 5.0)
615 color = PERF_COLOR_YELLOW;
616
617 fprintf(stderr, " # ");
618 color_fprintf(stderr, color, "%6.2f%%", ratio);
619 fprintf(stderr, " of all branches ");
620}
621
622static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
623{
624 double total, ratio = 0.0;
625 const char *color;
626
627 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
628
629 if (total)
630 ratio = avg / total * 100.0;
631
632 color = PERF_COLOR_NORMAL;
633 if (ratio > 20.0)
634 color = PERF_COLOR_RED;
635 else if (ratio > 10.0)
636 color = PERF_COLOR_MAGENTA;
637 else if (ratio > 5.0)
638 color = PERF_COLOR_YELLOW;
639
640 fprintf(stderr, " # ");
641 color_fprintf(stderr, color, "%6.2f%%", ratio);
642 fprintf(stderr, " of all L1-dcache hits ");
390} 643}
391 644
392static void abs_printout(int counter, double avg) 645static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
393{ 646{
394 double total, ratio = 0.0; 647 double total, ratio = 0.0;
648 const char *color;
395 649
396 if (big_num) 650 total = avg_stats(&runtime_l1_icache_stats[cpu]);
397 fprintf(stderr, " %'18.0f %-24s", avg, event_name(counter)); 651
652 if (total)
653 ratio = avg / total * 100.0;
654
655 color = PERF_COLOR_NORMAL;
656 if (ratio > 20.0)
657 color = PERF_COLOR_RED;
658 else if (ratio > 10.0)
659 color = PERF_COLOR_MAGENTA;
660 else if (ratio > 5.0)
661 color = PERF_COLOR_YELLOW;
662
663 fprintf(stderr, " # ");
664 color_fprintf(stderr, color, "%6.2f%%", ratio);
665 fprintf(stderr, " of all L1-icache hits ");
666}
667
668static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
669{
670 double total, ratio = 0.0;
671 const char *color;
672
673 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
674
675 if (total)
676 ratio = avg / total * 100.0;
677
678 color = PERF_COLOR_NORMAL;
679 if (ratio > 20.0)
680 color = PERF_COLOR_RED;
681 else if (ratio > 10.0)
682 color = PERF_COLOR_MAGENTA;
683 else if (ratio > 5.0)
684 color = PERF_COLOR_YELLOW;
685
686 fprintf(stderr, " # ");
687 color_fprintf(stderr, color, "%6.2f%%", ratio);
688 fprintf(stderr, " of all dTLB cache hits ");
689}
690
691static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
692{
693 double total, ratio = 0.0;
694 const char *color;
695
696 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
697
698 if (total)
699 ratio = avg / total * 100.0;
700
701 color = PERF_COLOR_NORMAL;
702 if (ratio > 20.0)
703 color = PERF_COLOR_RED;
704 else if (ratio > 10.0)
705 color = PERF_COLOR_MAGENTA;
706 else if (ratio > 5.0)
707 color = PERF_COLOR_YELLOW;
708
709 fprintf(stderr, " # ");
710 color_fprintf(stderr, color, "%6.2f%%", ratio);
711 fprintf(stderr, " of all iTLB cache hits ");
712}
713
714static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
715{
716 double total, ratio = 0.0;
717 const char *color;
718
719 total = avg_stats(&runtime_ll_cache_stats[cpu]);
720
721 if (total)
722 ratio = avg / total * 100.0;
723
724 color = PERF_COLOR_NORMAL;
725 if (ratio > 20.0)
726 color = PERF_COLOR_RED;
727 else if (ratio > 10.0)
728 color = PERF_COLOR_MAGENTA;
729 else if (ratio > 5.0)
730 color = PERF_COLOR_YELLOW;
731
732 fprintf(stderr, " # ");
733 color_fprintf(stderr, color, "%6.2f%%", ratio);
734 fprintf(stderr, " of all LL-cache hits ");
735}
736
737static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
738{
739 double total, ratio = 0.0;
740 char cpustr[16] = { '\0', };
741 const char *fmt;
742
743 if (csv_output)
744 fmt = "%s%.0f%s%s";
745 else if (big_num)
746 fmt = "%s%'18.0f%s%-25s";
398 else 747 else
399 fprintf(stderr, " %18.0f %-24s", avg, event_name(counter)); 748 fmt = "%s%18.0f%s%-25s";
400 749
401 if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) { 750 if (no_aggr)
402 total = avg_stats(&runtime_cycles_stats); 751 sprintf(cpustr, "CPU%*d%s",
752 csv_output ? 0 : -4,
753 evsel_list->cpus->map[cpu], csv_sep);
754 else
755 cpu = 0;
756
757 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
758
759 if (evsel->cgrp)
760 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
761
762 if (csv_output)
763 return;
764
765 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
766 total = avg_stats(&runtime_cycles_stats[cpu]);
403 767
404 if (total) 768 if (total)
405 ratio = avg / total; 769 ratio = avg / total;
406 770
407 fprintf(stderr, " # %10.3f IPC ", ratio); 771 fprintf(stderr, " # %5.2f insns per cycle ", ratio);
408 } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) && 772
409 runtime_branches_stats.n != 0) { 773 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
410 total = avg_stats(&runtime_branches_stats); 774 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
775
776 if (total && avg) {
777 ratio = total / avg;
778 fprintf(stderr, "\n # %5.2f stalled cycles per insn", ratio);
779 }
780
781 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
782 runtime_branches_stats[cpu].n != 0) {
783 print_branch_misses(cpu, evsel, avg);
784 } else if (
785 evsel->attr.type == PERF_TYPE_HW_CACHE &&
786 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
787 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
788 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
789 runtime_l1_dcache_stats[cpu].n != 0) {
790 print_l1_dcache_misses(cpu, evsel, avg);
791 } else if (
792 evsel->attr.type == PERF_TYPE_HW_CACHE &&
793 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
794 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
795 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
796 runtime_l1_icache_stats[cpu].n != 0) {
797 print_l1_icache_misses(cpu, evsel, avg);
798 } else if (
799 evsel->attr.type == PERF_TYPE_HW_CACHE &&
800 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
801 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
802 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
803 runtime_dtlb_cache_stats[cpu].n != 0) {
804 print_dtlb_cache_misses(cpu, evsel, avg);
805 } else if (
806 evsel->attr.type == PERF_TYPE_HW_CACHE &&
807 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
808 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
809 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
810 runtime_itlb_cache_stats[cpu].n != 0) {
811 print_itlb_cache_misses(cpu, evsel, avg);
812 } else if (
813 evsel->attr.type == PERF_TYPE_HW_CACHE &&
814 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
815 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
816 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
817 runtime_ll_cache_stats[cpu].n != 0) {
818 print_ll_cache_misses(cpu, evsel, avg);
819 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
820 runtime_cacherefs_stats[cpu].n != 0) {
821 total = avg_stats(&runtime_cacherefs_stats[cpu]);
411 822
412 if (total) 823 if (total)
413 ratio = avg * 100 / total; 824 ratio = avg * 100 / total;
414 825
415 fprintf(stderr, " # %10.3f %% ", ratio); 826 fprintf(stderr, " # %8.3f %% of all cache refs ", ratio);
827
828 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
829 print_stalled_cycles_frontend(cpu, evsel, avg);
830 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
831 print_stalled_cycles_backend(cpu, evsel, avg);
832 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
833 total = avg_stats(&runtime_nsecs_stats[cpu]);
834
835 if (total)
836 ratio = 1.0 * avg / total;
416 837
417 } else if (runtime_nsecs_stats.n != 0) { 838 fprintf(stderr, " # %8.3f GHz ", ratio);
418 total = avg_stats(&runtime_nsecs_stats); 839 } else if (runtime_nsecs_stats[cpu].n != 0) {
840 total = avg_stats(&runtime_nsecs_stats[cpu]);
419 841
420 if (total) 842 if (total)
421 ratio = 1000.0 * avg / total; 843 ratio = 1000.0 * avg / total;
422 844
423 fprintf(stderr, " # %10.3f M/sec", ratio); 845 fprintf(stderr, " # %8.3f M/sec ", ratio);
846 } else {
847 fprintf(stderr, " ");
424 } 848 }
425} 849}
426 850
427/* 851/*
428 * Print out the results of a single counter: 852 * Print out the results of a single counter:
853 * aggregated counts in system-wide mode
429 */ 854 */
430static void print_counter(int counter) 855static void print_counter_aggr(struct perf_evsel *counter)
431{ 856{
432 double avg = avg_stats(&event_res_stats[counter][0]); 857 struct perf_stat *ps = counter->priv;
433 int scaled = event_scaled[counter]; 858 double avg = avg_stats(&ps->res_stats[0]);
859 int scaled = counter->counts->scaled;
434 860
435 if (scaled == -1) { 861 if (scaled == -1) {
436 fprintf(stderr, " %18s %-24s\n", 862 fprintf(stderr, "%*s%s%*s",
437 "<not counted>", event_name(counter)); 863 csv_output ? 0 : 18,
864 "<not counted>",
865 csv_sep,
866 csv_output ? 0 : -24,
867 event_name(counter));
868
869 if (counter->cgrp)
870 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
871
872 fputc('\n', stderr);
438 return; 873 return;
439 } 874 }
440 875
441 if (nsec_counter(counter)) 876 if (nsec_counter(counter))
442 nsec_printout(counter, avg); 877 nsec_printout(-1, counter, avg);
443 else 878 else
444 abs_printout(counter, avg); 879 abs_printout(-1, counter, avg);
880
881 if (csv_output) {
882 fputc('\n', stderr);
883 return;
884 }
445 885
446 print_noise(counter, avg); 886 print_noise(counter, avg);
447 887
448 if (scaled) { 888 if (scaled) {
449 double avg_enabled, avg_running; 889 double avg_enabled, avg_running;
450 890
451 avg_enabled = avg_stats(&event_res_stats[counter][1]); 891 avg_enabled = avg_stats(&ps->res_stats[1]);
452 avg_running = avg_stats(&event_res_stats[counter][2]); 892 avg_running = avg_stats(&ps->res_stats[2]);
453 893
454 fprintf(stderr, " (scaled from %.2f%%)", 894 fprintf(stderr, " [%5.2f%%]", 100 * avg_running / avg_enabled);
455 100 * avg_running / avg_enabled);
456 } 895 }
457
458 fprintf(stderr, "\n"); 896 fprintf(stderr, "\n");
459} 897}
460 898
899/*
900 * Print out the results of a single counter:
901 * does not use aggregated count in system-wide
902 */
903static void print_counter(struct perf_evsel *counter)
904{
905 u64 ena, run, val;
906 int cpu;
907
908 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
909 val = counter->counts->cpu[cpu].val;
910 ena = counter->counts->cpu[cpu].ena;
911 run = counter->counts->cpu[cpu].run;
912 if (run == 0 || ena == 0) {
913 fprintf(stderr, "CPU%*d%s%*s%s%*s",
914 csv_output ? 0 : -4,
915 evsel_list->cpus->map[cpu], csv_sep,
916 csv_output ? 0 : 18,
917 "<not counted>", csv_sep,
918 csv_output ? 0 : -24,
919 event_name(counter));
920
921 if (counter->cgrp)
922 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
923
924 fputc('\n', stderr);
925 continue;
926 }
927
928 if (nsec_counter(counter))
929 nsec_printout(cpu, counter, val);
930 else
931 abs_printout(cpu, counter, val);
932
933 if (!csv_output) {
934 print_noise(counter, 1.0);
935
936 if (run != ena)
937 fprintf(stderr, " (%.2f%%)", 100.0 * run / ena);
938 }
939 fputc('\n', stderr);
940 }
941}
942
461static void print_stat(int argc, const char **argv) 943static void print_stat(int argc, const char **argv)
462{ 944{
463 int i, counter; 945 struct perf_evsel *counter;
946 int i;
464 947
465 fflush(stdout); 948 fflush(stdout);
466 949
467 fprintf(stderr, "\n"); 950 if (!csv_output) {
468 fprintf(stderr, " Performance counter stats for "); 951 fprintf(stderr, "\n");
469 if(target_pid == -1 && target_tid == -1) { 952 fprintf(stderr, " Performance counter stats for ");
470 fprintf(stderr, "\'%s", argv[0]); 953 if(target_pid == -1 && target_tid == -1) {
471 for (i = 1; i < argc; i++) 954 fprintf(stderr, "\'%s", argv[0]);
472 fprintf(stderr, " %s", argv[i]); 955 for (i = 1; i < argc; i++)
473 } else if (target_pid != -1) 956 fprintf(stderr, " %s", argv[i]);
474 fprintf(stderr, "process id \'%d", target_pid); 957 } else if (target_pid != -1)
475 else 958 fprintf(stderr, "process id \'%d", target_pid);
476 fprintf(stderr, "thread id \'%d", target_tid); 959 else
477 960 fprintf(stderr, "thread id \'%d", target_tid);
478 fprintf(stderr, "\'"); 961
479 if (run_count > 1) 962 fprintf(stderr, "\'");
480 fprintf(stderr, " (%d runs)", run_count); 963 if (run_count > 1)
481 fprintf(stderr, ":\n\n"); 964 fprintf(stderr, " (%d runs)", run_count);
965 fprintf(stderr, ":\n\n");
966 }
482 967
483 for (counter = 0; counter < nr_counters; counter++) 968 if (no_aggr) {
484 print_counter(counter); 969 list_for_each_entry(counter, &evsel_list->entries, node)
970 print_counter(counter);
971 } else {
972 list_for_each_entry(counter, &evsel_list->entries, node)
973 print_counter_aggr(counter);
974 }
485 975
486 fprintf(stderr, "\n"); 976 if (!csv_output) {
487 fprintf(stderr, " %18.9f seconds time elapsed", 977 if (!null_run)
488 avg_stats(&walltime_nsecs_stats)/1e9); 978 fprintf(stderr, "\n");
489 if (run_count > 1) { 979 fprintf(stderr, " %17.9f seconds time elapsed",
490 fprintf(stderr, " ( +- %7.3f%% )", 980 avg_stats(&walltime_nsecs_stats)/1e9);
491 100*stddev_stats(&walltime_nsecs_stats) / 981 if (run_count > 1) {
492 avg_stats(&walltime_nsecs_stats)); 982 fprintf(stderr, " ");
983 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
984 avg_stats(&walltime_nsecs_stats));
985 }
986 fprintf(stderr, "\n\n");
493 } 987 }
494 fprintf(stderr, "\n\n");
495} 988}
496 989
497static volatile int signr = -1; 990static volatile int signr = -1;
@@ -521,10 +1014,19 @@ static const char * const stat_usage[] = {
521 NULL 1014 NULL
522}; 1015};
523 1016
1017static int stat__set_big_num(const struct option *opt __used,
1018 const char *s __used, int unset)
1019{
1020 big_num_opt = unset ? 0 : 1;
1021 return 0;
1022}
1023
524static const struct option options[] = { 1024static const struct option options[] = {
525 OPT_CALLBACK('e', "event", NULL, "event", 1025 OPT_CALLBACK('e', "event", &evsel_list, "event",
526 "event selector. use 'perf list' to list available events", 1026 "event selector. use 'perf list' to list available events",
527 parse_events), 1027 parse_events),
1028 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1029 "event filter", parse_filter),
528 OPT_BOOLEAN('i', "no-inherit", &no_inherit, 1030 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
529 "child tasks do not inherit counters"), 1031 "child tasks do not inherit counters"),
530 OPT_INTEGER('p', "pid", &target_pid, 1032 OPT_INTEGER('p', "pid", &target_pid,
@@ -541,64 +1043,162 @@ static const struct option options[] = {
541 "repeat command and print average + stddev (max: 100)"), 1043 "repeat command and print average + stddev (max: 100)"),
542 OPT_BOOLEAN('n', "null", &null_run, 1044 OPT_BOOLEAN('n', "null", &null_run,
543 "null run - dont start any counters"), 1045 "null run - dont start any counters"),
544 OPT_BOOLEAN('B', "big-num", &big_num, 1046 OPT_INCR('d', "detailed", &detailed_run,
545 "print large numbers with thousands\' separators"), 1047 "detailed run - start a lot of events"),
1048 OPT_BOOLEAN('S', "sync", &sync_run,
1049 "call sync() before starting a run"),
1050 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1051 "print large numbers with thousands\' separators",
1052 stat__set_big_num),
546 OPT_STRING('C', "cpu", &cpu_list, "cpu", 1053 OPT_STRING('C', "cpu", &cpu_list, "cpu",
547 "list of cpus to monitor in system-wide"), 1054 "list of cpus to monitor in system-wide"),
1055 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1056 "disable CPU count aggregation"),
1057 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1058 "print counts with custom separator"),
1059 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1060 "monitor event in cgroup name only",
1061 parse_cgroups),
548 OPT_END() 1062 OPT_END()
549}; 1063};
550 1064
1065/*
1066 * Add default attributes, if there were no attributes specified or
1067 * if -d/--detailed, -d -d or -d -d -d is used:
1068 */
1069static int add_default_attributes(void)
1070{
1071 struct perf_evsel *pos;
1072 size_t attr_nr = 0;
1073 size_t c;
1074
1075 /* Set attrs if no event is selected and !null_run: */
1076 if (null_run)
1077 return 0;
1078
1079 if (!evsel_list->nr_entries) {
1080 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1081 pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1082 if (pos == NULL)
1083 return -1;
1084 perf_evlist__add(evsel_list, pos);
1085 }
1086 attr_nr += c;
1087 }
1088
1089 /* Detailed events get appended to the event list: */
1090
1091 if (detailed_run < 1)
1092 return 0;
1093
1094 /* Append detailed run extra attributes: */
1095 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1096 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1097 if (pos == NULL)
1098 return -1;
1099 perf_evlist__add(evsel_list, pos);
1100 }
1101 attr_nr += c;
1102
1103 if (detailed_run < 2)
1104 return 0;
1105
1106 /* Append very detailed run extra attributes: */
1107 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1108 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1109 if (pos == NULL)
1110 return -1;
1111 perf_evlist__add(evsel_list, pos);
1112 }
1113
1114 if (detailed_run < 3)
1115 return 0;
1116
1117 /* Append very, very detailed run extra attributes: */
1118 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1119 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1120 if (pos == NULL)
1121 return -1;
1122 perf_evlist__add(evsel_list, pos);
1123 }
1124
1125
1126 return 0;
1127}
1128
551int cmd_stat(int argc, const char **argv, const char *prefix __used) 1129int cmd_stat(int argc, const char **argv, const char *prefix __used)
552{ 1130{
553 int status; 1131 struct perf_evsel *pos;
554 int i,j; 1132 int status = -ENOMEM;
555 1133
556 setlocale(LC_ALL, ""); 1134 setlocale(LC_ALL, "");
557 1135
1136 evsel_list = perf_evlist__new(NULL, NULL);
1137 if (evsel_list == NULL)
1138 return -ENOMEM;
1139
558 argc = parse_options(argc, argv, options, stat_usage, 1140 argc = parse_options(argc, argv, options, stat_usage,
559 PARSE_OPT_STOP_AT_NON_OPTION); 1141 PARSE_OPT_STOP_AT_NON_OPTION);
1142
1143 if (csv_sep)
1144 csv_output = true;
1145 else
1146 csv_sep = DEFAULT_SEPARATOR;
1147
1148 /*
1149 * let the spreadsheet do the pretty-printing
1150 */
1151 if (csv_output) {
1152 /* User explicitely passed -B? */
1153 if (big_num_opt == 1) {
1154 fprintf(stderr, "-B option not supported with -x\n");
1155 usage_with_options(stat_usage, options);
1156 } else /* Nope, so disable big number formatting */
1157 big_num = false;
1158 } else if (big_num_opt == 0) /* User passed --no-big-num */
1159 big_num = false;
1160
560 if (!argc && target_pid == -1 && target_tid == -1) 1161 if (!argc && target_pid == -1 && target_tid == -1)
561 usage_with_options(stat_usage, options); 1162 usage_with_options(stat_usage, options);
562 if (run_count <= 0) 1163 if (run_count <= 0)
563 usage_with_options(stat_usage, options); 1164 usage_with_options(stat_usage, options);
564 1165
565 /* Set attrs and nr_counters if no event is selected and !null_run */ 1166 /* no_aggr, cgroup are for system-wide only */
566 if (!null_run && !nr_counters) { 1167 if ((no_aggr || nr_cgroups) && !system_wide) {
567 memcpy(attrs, default_attrs, sizeof(default_attrs)); 1168 fprintf(stderr, "both cgroup and no-aggregation "
568 nr_counters = ARRAY_SIZE(default_attrs); 1169 "modes only available in system-wide mode\n");
1170
1171 usage_with_options(stat_usage, options);
569 } 1172 }
570 1173
571 if (system_wide) 1174 if (add_default_attributes())
572 nr_cpus = read_cpu_map(cpu_list); 1175 goto out;
573 else
574 nr_cpus = 1;
575 1176
576 if (nr_cpus < 1) 1177 if (target_pid != -1)
1178 target_tid = target_pid;
1179
1180 evsel_list->threads = thread_map__new(target_pid, target_tid);
1181 if (evsel_list->threads == NULL) {
1182 pr_err("Problems finding threads of monitor\n");
577 usage_with_options(stat_usage, options); 1183 usage_with_options(stat_usage, options);
1184 }
578 1185
579 if (target_pid != -1) { 1186 if (system_wide)
580 target_tid = target_pid; 1187 evsel_list->cpus = cpu_map__new(cpu_list);
581 thread_num = find_all_tid(target_pid, &all_tids); 1188 else
582 if (thread_num <= 0) { 1189 evsel_list->cpus = cpu_map__dummy_new();
583 fprintf(stderr, "Can't find all threads of pid %d\n",
584 target_pid);
585 usage_with_options(stat_usage, options);
586 }
587 } else {
588 all_tids=malloc(sizeof(pid_t));
589 if (!all_tids)
590 return -ENOMEM;
591 1190
592 all_tids[0] = target_tid; 1191 if (evsel_list->cpus == NULL) {
593 thread_num = 1; 1192 perror("failed to parse CPUs map");
1193 usage_with_options(stat_usage, options);
1194 return -1;
594 } 1195 }
595 1196
596 for (i = 0; i < MAX_NR_CPUS; i++) { 1197 list_for_each_entry(pos, &evsel_list->entries, node) {
597 for (j = 0; j < MAX_COUNTERS; j++) { 1198 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
598 fd[i][j] = malloc(sizeof(int)*thread_num); 1199 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
599 if (!fd[i][j]) 1200 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
600 return -ENOMEM; 1201 goto out_free_fd;
601 }
602 } 1202 }
603 1203
604 /* 1204 /*
@@ -616,11 +1216,20 @@ int cmd_stat(int argc, const char **argv, const char *prefix __used)
616 for (run_idx = 0; run_idx < run_count; run_idx++) { 1216 for (run_idx = 0; run_idx < run_count; run_idx++) {
617 if (run_count != 1 && verbose) 1217 if (run_count != 1 && verbose)
618 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); 1218 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
1219
1220 if (sync_run)
1221 sync();
1222
619 status = run_perf_stat(argc, argv); 1223 status = run_perf_stat(argc, argv);
620 } 1224 }
621 1225
622 if (status != -1) 1226 if (status != -1)
623 print_stat(argc, argv); 1227 print_stat(argc, argv);
624 1228out_free_fd:
1229 list_for_each_entry(pos, &evsel_list->entries, node)
1230 perf_evsel__free_stat_priv(pos);
1231 perf_evlist__delete_maps(evsel_list);
1232out:
1233 perf_evlist__delete(evsel_list);
625 return status; 1234 return status;
626} 1235}
generated by cgit v1.2.2 (git 2.25.0) at 2025-07-12 21:10:02 -0400