1 // SPDX-License-Identifier: GPL-2.0
11 #include <linux/kernel.h>
12 #include <linux/time64.h>
13 #include <linux/list.h>
14 #include <linux/err.h>
15 #include <linux/zalloc.h>
16 #include <internal/lib.h>
17 #include <subcmd/parse-options.h>
20 #include "util/data.h"
21 #include "util/stat.h"
22 #include "util/debug.h"
23 #include "util/symbol.h"
24 #include "util/session.h"
25 #include "util/build-id.h"
26 #include "util/sample.h"
27 #include "util/synthetic-events.h"
29 #define MMAP_DEV_MAJOR 8
30 #define DSO_MMAP_RATIO 4
32 static unsigned int iterations = 100;
33 static unsigned int nr_mmaps = 100;
34 static unsigned int nr_samples = 100; /* samples per mmap */
36 static u64 bench_sample_type;
37 static u16 bench_id_hdr_size;
47 struct list_head list;
53 static struct bench_dso *dsos;
55 extern int cmd_inject(int argc, const char *argv[]);
57 static const struct option options[] = {
58 OPT_UINTEGER('i', "iterations", &iterations,
59 "Number of iterations used to compute average (default: 100)"),
60 OPT_UINTEGER('m', "nr-mmaps", &nr_mmaps,
61 "Number of mmap events for each iteration (default: 100)"),
62 OPT_UINTEGER('n', "nr-samples", &nr_samples,
63 "Number of sample events per mmap event (default: 100)"),
64 OPT_INCR('v', "verbose", &verbose,
65 "be more verbose (show iteration count, DSO name, etc)"),
69 static const char *const bench_usage[] = {
70 "perf bench internals inject-build-id <options>",
75 * Helper for collect_dso that adds the given file as a dso to dso_list
76 * if it contains a build-id. Stops after collecting 4 times more than
77 * we need (for MMAP2 events).
79 static int add_dso(const char *fpath, const struct stat *sb __maybe_unused,
80 int typeflag, struct FTW *ftwbuf __maybe_unused)
82 struct bench_dso *dso = &dsos[nr_dsos];
85 if (typeflag == FTW_D || typeflag == FTW_SL)
88 if (filename__read_build_id(fpath, &bid) < 0)
91 dso->name = realpath(fpath, NULL);
92 if (dso->name == NULL)
96 pr_debug2(" Adding DSO: %s\n", fpath);
98 /* stop if we collected enough DSOs */
99 if ((unsigned int)nr_dsos == DSO_MMAP_RATIO * nr_mmaps)
105 static void collect_dso(void)
107 dsos = calloc(nr_mmaps * DSO_MMAP_RATIO, sizeof(*dsos));
109 printf(" Memory allocation failed\n");
113 if (nftw("/usr/lib/", add_dso, 10, FTW_PHYS) < 0)
116 pr_debug(" Collected %d DSOs\n", nr_dsos);
119 static void release_dso(void)
123 for (i = 0; i < nr_dsos; i++) {
124 struct bench_dso *dso = &dsos[i];
131 /* Fake address used by mmap and sample events */
132 static u64 dso_map_addr(struct bench_dso *dso)
134 return 0x400000ULL + dso->ino * 8192ULL;
137 static ssize_t synthesize_attr(struct bench_data *data)
139 union perf_event event;
141 memset(&event, 0, sizeof(event.attr) + sizeof(u64));
143 event.header.type = PERF_RECORD_HEADER_ATTR;
144 event.header.size = sizeof(event.attr) + sizeof(u64);
146 event.attr.attr.type = PERF_TYPE_SOFTWARE;
147 event.attr.attr.config = PERF_COUNT_SW_TASK_CLOCK;
148 event.attr.attr.exclude_kernel = 1;
149 event.attr.attr.sample_id_all = 1;
150 event.attr.attr.sample_type = bench_sample_type;
152 return writen(data->input_pipe[1], &event, event.header.size);
155 static ssize_t synthesize_fork(struct bench_data *data)
157 union perf_event event;
159 memset(&event, 0, sizeof(event.fork) + bench_id_hdr_size);
161 event.header.type = PERF_RECORD_FORK;
162 event.header.misc = PERF_RECORD_MISC_FORK_EXEC;
163 event.header.size = sizeof(event.fork) + bench_id_hdr_size;
167 event.fork.pid = data->pid;
168 event.fork.tid = data->pid;
170 return writen(data->input_pipe[1], &event, event.header.size);
173 static ssize_t synthesize_mmap(struct bench_data *data, struct bench_dso *dso, u64 timestamp)
175 union perf_event event;
176 size_t len = offsetof(struct perf_record_mmap2, filename);
177 u64 *id_hdr_ptr = (void *)&event;
180 len += roundup(strlen(dso->name) + 1, 8) + bench_id_hdr_size;
182 memset(&event, 0, min(len, sizeof(event.mmap2)));
184 event.header.type = PERF_RECORD_MMAP2;
185 event.header.misc = PERF_RECORD_MISC_USER;
186 event.header.size = len;
188 event.mmap2.pid = data->pid;
189 event.mmap2.tid = data->pid;
190 event.mmap2.maj = MMAP_DEV_MAJOR;
191 event.mmap2.ino = dso->ino;
193 strcpy(event.mmap2.filename, dso->name);
195 event.mmap2.start = dso_map_addr(dso);
196 event.mmap2.len = 4096;
197 event.mmap2.prot = PROT_EXEC;
199 if (len > sizeof(event.mmap2)) {
200 /* write mmap2 event first */
201 if (writen(data->input_pipe[1], &event, len - bench_id_hdr_size) < 0)
203 /* zero-fill sample id header */
204 memset(id_hdr_ptr, 0, bench_id_hdr_size);
205 /* put timestamp in the right position */
206 ts_idx = (bench_id_hdr_size / sizeof(u64)) - 2;
207 id_hdr_ptr[ts_idx] = timestamp;
208 if (writen(data->input_pipe[1], id_hdr_ptr, bench_id_hdr_size) < 0)
214 ts_idx = (len / sizeof(u64)) - 2;
215 id_hdr_ptr[ts_idx] = timestamp;
216 return writen(data->input_pipe[1], &event, len);
219 static ssize_t synthesize_sample(struct bench_data *data, struct bench_dso *dso, u64 timestamp)
221 union perf_event event;
222 struct perf_sample sample = {
225 .ip = dso_map_addr(dso),
229 event.header.type = PERF_RECORD_SAMPLE;
230 event.header.misc = PERF_RECORD_MISC_USER;
231 event.header.size = perf_event__sample_event_size(&sample, bench_sample_type, 0);
233 perf_event__synthesize_sample(&event, bench_sample_type, 0, &sample);
235 return writen(data->input_pipe[1], &event, event.header.size);
238 static ssize_t synthesize_flush(struct bench_data *data)
240 struct perf_event_header header = {
241 .size = sizeof(header),
242 .type = PERF_RECORD_FINISHED_ROUND,
245 return writen(data->input_pipe[1], &header, header.size);
248 static void *data_reader(void *arg)
250 struct bench_data *data = arg;
255 flag = fcntl(data->output_pipe[0], F_GETFL);
256 fcntl(data->output_pipe[0], F_SETFL, flag | O_NONBLOCK);
258 /* read out data from child */
260 n = read(data->output_pipe[0], buf, sizeof(buf));
266 if (errno != EINTR && errno != EAGAIN)
272 close(data->output_pipe[0]);
276 static int setup_injection(struct bench_data *data, bool build_id_all)
282 if (pipe(ready_pipe) < 0)
285 if (pipe(data->input_pipe) < 0)
288 if (pipe(data->output_pipe) < 0)
295 if (data->pid == 0) {
296 const char **inject_argv;
299 close(data->input_pipe[1]);
300 close(data->output_pipe[0]);
301 close(ready_pipe[0]);
303 dup2(data->input_pipe[0], STDIN_FILENO);
304 close(data->input_pipe[0]);
305 dup2(data->output_pipe[1], STDOUT_FILENO);
306 close(data->output_pipe[1]);
308 dev_null_fd = open("/dev/null", O_WRONLY);
312 dup2(dev_null_fd, STDERR_FILENO);
317 inject_argv = calloc(inject_argc + 1, sizeof(*inject_argv));
318 if (inject_argv == NULL)
321 inject_argv[0] = strdup("inject");
322 inject_argv[1] = strdup("-b");
324 inject_argv[2] = strdup("--buildid-all");
326 /* signal that we're ready to go */
327 close(ready_pipe[1]);
329 cmd_inject(inject_argc, inject_argv);
334 pthread_create(&data->th, NULL, data_reader, data);
336 close(ready_pipe[1]);
337 close(data->input_pipe[0]);
338 close(data->output_pipe[1]);
340 /* wait for child ready */
341 if (read(ready_pipe[0], &buf, 1) < 0)
343 close(ready_pipe[0]);
348 static int inject_build_id(struct bench_data *data, u64 *max_rss)
352 struct rusage rusage;
354 /* this makes the child to run */
355 if (perf_header__write_pipe(data->input_pipe[1]) < 0)
358 if (synthesize_attr(data) < 0)
361 if (synthesize_fork(data) < 0)
364 for (i = 0; i < nr_mmaps; i++) {
365 int idx = rand() % (nr_dsos - 1);
366 struct bench_dso *dso = &dsos[idx];
367 u64 timestamp = rand() % 1000000;
369 pr_debug2(" [%d] injecting: %s\n", i+1, dso->name);
370 if (synthesize_mmap(data, dso, timestamp) < 0)
373 for (k = 0; k < nr_samples; k++) {
374 if (synthesize_sample(data, dso, timestamp + k * 1000) < 0)
378 if ((i + 1) % 10 == 0) {
379 if (synthesize_flush(data) < 0)
384 /* this makes the child to finish */
385 close(data->input_pipe[1]);
387 wait4(data->pid, &status, 0, &rusage);
388 *max_rss = rusage.ru_maxrss;
390 pr_debug(" Child %d exited with %d\n", data->pid, status);
395 static void do_inject_loop(struct bench_data *data, bool build_id_all)
398 struct stats time_stats, mem_stats;
399 double time_average, time_stddev;
400 double mem_average, mem_stddev;
402 init_stats(&time_stats);
403 init_stats(&mem_stats);
405 pr_debug(" Build-id%s injection benchmark\n", build_id_all ? "-all" : "");
407 for (i = 0; i < iterations; i++) {
408 struct timeval start, end, diff;
409 u64 runtime_us, max_rss;
411 pr_debug(" Iteration #%d\n", i+1);
413 if (setup_injection(data, build_id_all) < 0) {
414 printf(" Build-id injection setup failed\n");
418 gettimeofday(&start, NULL);
419 if (inject_build_id(data, &max_rss) < 0) {
420 printf(" Build-id injection failed\n");
424 gettimeofday(&end, NULL);
425 timersub(&end, &start, &diff);
426 runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
427 update_stats(&time_stats, runtime_us);
428 update_stats(&mem_stats, max_rss);
430 pthread_join(data->th, NULL);
433 time_average = avg_stats(&time_stats) / USEC_PER_MSEC;
434 time_stddev = stddev_stats(&time_stats) / USEC_PER_MSEC;
435 printf(" Average build-id%s injection took: %.3f msec (+- %.3f msec)\n",
436 build_id_all ? "-all" : "", time_average, time_stddev);
438 /* each iteration, it processes MMAP2 + BUILD_ID + nr_samples * SAMPLE */
439 time_average = avg_stats(&time_stats) / (nr_mmaps * (nr_samples + 2));
440 time_stddev = stddev_stats(&time_stats) / (nr_mmaps * (nr_samples + 2));
441 printf(" Average time per event: %.3f usec (+- %.3f usec)\n",
442 time_average, time_stddev);
444 mem_average = avg_stats(&mem_stats);
445 mem_stddev = stddev_stats(&mem_stats);
446 printf(" Average memory usage: %.0f KB (+- %.0f KB)\n",
447 mem_average, mem_stddev);
450 static int do_inject_loops(struct bench_data *data)
456 bench_sample_type = PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_IP;
457 bench_sample_type |= PERF_SAMPLE_TID | PERF_SAMPLE_TIME;
458 bench_id_hdr_size = 32;
462 printf(" Cannot collect DSOs for injection\n");
466 do_inject_loop(data, false);
467 do_inject_loop(data, true);
473 int bench_inject_build_id(int argc, const char **argv)
475 struct bench_data data;
477 argc = parse_options(argc, argv, options, bench_usage, 0);
479 usage_with_options(bench_usage, options);
483 return do_inject_loops(&data);