1 // SPDX-License-Identifier: GPL-2.0
3 #include <subcmd/parse-options.h>
4 #include <linux/hw_breakpoint.h>
5 #include <linux/perf_event.h>
6 #include <linux/time64.h>
7 #include <sys/syscall.h>
20 unsigned int nbreakpoints;
21 unsigned int nparallel;
22 unsigned int nthreads;
29 static const struct option thread_options[] = {
30 OPT_UINTEGER('b', "breakpoints", &thread_params.nbreakpoints,
31 "Specify amount of breakpoints"),
32 OPT_UINTEGER('p', "parallelism", &thread_params.nparallel, "Specify amount of parallelism"),
33 OPT_UINTEGER('t', "threads", &thread_params.nthreads, "Specify amount of threads"),
37 static const char * const thread_usage[] = {
38 "perf bench breakpoint thread <options>",
47 static int breakpoint_setup(void *addr)
49 struct perf_event_attr attr = { .size = 0, };
52 attr.type = PERF_TYPE_BREAKPOINT;
53 attr.size = sizeof(attr);
55 attr.exclude_kernel = 1;
57 attr.bp_addr = (unsigned long)addr;
58 attr.bp_type = HW_BREAKPOINT_RW;
59 attr.bp_len = HW_BREAKPOINT_LEN_1;
60 fd = syscall(SYS_perf_event_open, &attr, 0, -1, -1, 0);
68 static void *passive_thread(void *arg)
70 unsigned int *done = (unsigned int *)arg;
72 while (!__atomic_load_n(done, __ATOMIC_RELAXED))
73 futex_wait(done, 0, NULL, 0);
77 static void *active_thread(void *arg)
79 unsigned int *done = (unsigned int *)arg;
81 while (!__atomic_load_n(done, __ATOMIC_RELAXED));
85 static void *breakpoint_thread(void *arg)
88 int *repeat = (int *)arg;
91 threads = calloc(thread_params.nthreads, sizeof(threads[0]));
93 exit((perror("calloc"), EXIT_FAILURE));
95 while (__atomic_fetch_sub(repeat, 1, __ATOMIC_RELAXED) > 0) {
97 for (i = 0; i < thread_params.nthreads; i++) {
98 if (pthread_create(&threads[i], NULL, passive_thread, &done))
99 exit((perror("pthread_create"), EXIT_FAILURE));
101 __atomic_store_n(&done, 1, __ATOMIC_RELAXED);
102 futex_wake(&done, thread_params.nthreads, 0);
103 for (i = 0; i < thread_params.nthreads; i++)
104 pthread_join(threads[i], NULL);
110 // The benchmark creates nbreakpoints inheritable breakpoints,
111 // then starts nparallel threads which create and join bench_repeat batches of nthreads threads.
112 int bench_breakpoint_thread(int argc, const char **argv)
114 unsigned int i, result_usec;
115 int repeat = bench_repeat;
116 struct breakpoint *breakpoints;
118 struct timeval start, stop, diff;
120 if (parse_options(argc, argv, thread_options, thread_usage, 0)) {
121 usage_with_options(thread_usage, thread_options);
124 breakpoints = calloc(thread_params.nbreakpoints, sizeof(breakpoints[0]));
125 parallel = calloc(thread_params.nparallel, sizeof(parallel[0]));
126 if (!breakpoints || !parallel)
127 exit((perror("calloc"), EXIT_FAILURE));
129 for (i = 0; i < thread_params.nbreakpoints; i++) {
130 breakpoints[i].fd = breakpoint_setup(&breakpoints[i].watched);
132 if (breakpoints[i].fd < 0) {
133 if (breakpoints[i].fd == -ENODEV) {
134 printf("Skipping perf bench breakpoint thread: No hardware support\n");
137 exit((perror("perf_event_open"), EXIT_FAILURE));
140 gettimeofday(&start, NULL);
141 for (i = 0; i < thread_params.nparallel; i++) {
142 if (pthread_create(¶llel[i], NULL, breakpoint_thread, &repeat))
143 exit((perror("pthread_create"), EXIT_FAILURE));
145 for (i = 0; i < thread_params.nparallel; i++)
146 pthread_join(parallel[i], NULL);
147 gettimeofday(&stop, NULL);
148 timersub(&stop, &start, &diff);
149 for (i = 0; i < thread_params.nbreakpoints; i++)
150 close(breakpoints[i].fd);
153 switch (bench_format) {
154 case BENCH_FORMAT_DEFAULT:
155 printf("# Created/joined %d threads with %d breakpoints and %d parallelism\n",
156 bench_repeat, thread_params.nbreakpoints, thread_params.nparallel);
157 printf(" %14s: %lu.%03lu [sec]\n\n", "Total time",
158 (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
159 result_usec = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
160 printf(" %14lf usecs/op\n",
161 (double)result_usec / bench_repeat / thread_params.nthreads);
162 printf(" %14lf usecs/op/cpu\n",
163 (double)result_usec / bench_repeat /
164 thread_params.nthreads * thread_params.nparallel);
166 case BENCH_FORMAT_SIMPLE:
167 printf("%lu.%03lu\n", (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
170 fprintf(stderr, "Unknown format: %d\n", bench_format);
177 unsigned int npassive;
178 unsigned int nactive;
184 static const struct option enable_options[] = {
185 OPT_UINTEGER('p', "passive", &enable_params.npassive, "Specify amount of passive threads"),
186 OPT_UINTEGER('a', "active", &enable_params.nactive, "Specify amount of active threads"),
190 static const char * const enable_usage[] = {
191 "perf bench breakpoint enable <options>",
195 // The benchmark creates an inheritable breakpoint,
196 // then starts npassive threads that block and nactive threads that actively spin
197 // and then disables and enables the breakpoint bench_repeat times.
198 int bench_breakpoint_enable(int argc, const char **argv)
200 unsigned int i, nthreads, result_usec, done = 0;
204 struct timeval start, stop, diff;
206 if (parse_options(argc, argv, enable_options, enable_usage, 0)) {
207 usage_with_options(enable_usage, enable_options);
210 fd = breakpoint_setup(&watched);
214 printf("Skipping perf bench breakpoint enable: No hardware support\n");
217 exit((perror("perf_event_open"), EXIT_FAILURE));
219 nthreads = enable_params.npassive + enable_params.nactive;
220 threads = calloc(nthreads, sizeof(threads[0]));
222 exit((perror("calloc"), EXIT_FAILURE));
224 for (i = 0; i < nthreads; i++) {
225 if (pthread_create(&threads[i], NULL,
226 i < enable_params.npassive ? passive_thread : active_thread, &done))
227 exit((perror("pthread_create"), EXIT_FAILURE));
229 usleep(10000); // let the threads block
230 gettimeofday(&start, NULL);
231 for (i = 0; i < bench_repeat; i++) {
232 if (ioctl(fd, PERF_EVENT_IOC_DISABLE, 0))
233 exit((perror("ioctl(PERF_EVENT_IOC_DISABLE)"), EXIT_FAILURE));
234 if (ioctl(fd, PERF_EVENT_IOC_ENABLE, 0))
235 exit((perror("ioctl(PERF_EVENT_IOC_ENABLE)"), EXIT_FAILURE));
237 gettimeofday(&stop, NULL);
238 timersub(&stop, &start, &diff);
239 __atomic_store_n(&done, 1, __ATOMIC_RELAXED);
240 futex_wake(&done, enable_params.npassive, 0);
241 for (i = 0; i < nthreads; i++)
242 pthread_join(threads[i], NULL);
245 switch (bench_format) {
246 case BENCH_FORMAT_DEFAULT:
247 printf("# Enabled/disabled breakpoint %d time with %d passive and %d active threads\n",
248 bench_repeat, enable_params.npassive, enable_params.nactive);
249 printf(" %14s: %lu.%03lu [sec]\n\n", "Total time",
250 (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
251 result_usec = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
252 printf(" %14lf usecs/op\n", (double)result_usec / bench_repeat);
254 case BENCH_FORMAT_SIMPLE:
255 printf("%lu.%03lu\n", (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
258 fprintf(stderr, "Unknown format: %d\n", bench_format);