1 // SPDX-License-Identifier: GPL-2.0
2
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>
8 #include <sys/ioctl.h>
9 #include <sys/time.h>
10 #include <pthread.h>
11 #include <stddef.h>
12 #include <stdlib.h>
13 #include <unistd.h>
14 #include <stdio.h>
15 #include <errno.h>
16 #include "bench.h"
17 #include "futex.h"
18
19 struct {
20 unsigned int nbreakpoints;
21 unsigned int nparallel;
22 unsigned int nthreads;
23 } thread_params = {
24 .nbreakpoints = 1,
25 .nparallel = 1,
26 .nthreads = 1,
27 };
28
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"),
34 OPT_END()
35 };
36
37 static const char * const thread_usage[] = {
38 "perf bench breakpoint thread <options>",
39 NULL
40 };
41
42 struct breakpoint {
43 int fd;
44 char watched;
45 };
46
breakpoint_setup(void * addr)47 static int breakpoint_setup(void *addr)
48 {
49 struct perf_event_attr attr = { .size = 0, };
50
51 attr.type = PERF_TYPE_BREAKPOINT;
52 attr.size = sizeof(attr);
53 attr.inherit = 1;
54 attr.exclude_kernel = 1;
55 attr.exclude_hv = 1;
56 attr.bp_addr = (unsigned long)addr;
57 attr.bp_type = HW_BREAKPOINT_RW;
58 attr.bp_len = HW_BREAKPOINT_LEN_1;
59 return syscall(SYS_perf_event_open, &attr, 0, -1, -1, 0);
60 }
61
passive_thread(void * arg)62 static void *passive_thread(void *arg)
63 {
64 unsigned int *done = (unsigned int *)arg;
65
66 while (!__atomic_load_n(done, __ATOMIC_RELAXED))
67 futex_wait(done, 0, NULL, 0);
68 return NULL;
69 }
70
active_thread(void * arg)71 static void *active_thread(void *arg)
72 {
73 unsigned int *done = (unsigned int *)arg;
74
75 while (!__atomic_load_n(done, __ATOMIC_RELAXED));
76 return NULL;
77 }
78
breakpoint_thread(void * arg)79 static void *breakpoint_thread(void *arg)
80 {
81 unsigned int i, done;
82 int *repeat = (int *)arg;
83 pthread_t *threads;
84
85 threads = calloc(thread_params.nthreads, sizeof(threads[0]));
86 if (!threads)
87 exit((perror("calloc"), EXIT_FAILURE));
88
89 while (__atomic_fetch_sub(repeat, 1, __ATOMIC_RELAXED) > 0) {
90 done = 0;
91 for (i = 0; i < thread_params.nthreads; i++) {
92 if (pthread_create(&threads[i], NULL, passive_thread, &done))
93 exit((perror("pthread_create"), EXIT_FAILURE));
94 }
95 __atomic_store_n(&done, 1, __ATOMIC_RELAXED);
96 futex_wake(&done, thread_params.nthreads, 0);
97 for (i = 0; i < thread_params.nthreads; i++)
98 pthread_join(threads[i], NULL);
99 }
100 free(threads);
101 return NULL;
102 }
103
104 // The benchmark creates nbreakpoints inheritable breakpoints,
105 // then starts nparallel threads which create and join bench_repeat batches of nthreads threads.
bench_breakpoint_thread(int argc,const char ** argv)106 int bench_breakpoint_thread(int argc, const char **argv)
107 {
108 unsigned int i, result_usec;
109 int repeat = bench_repeat;
110 struct breakpoint *breakpoints;
111 pthread_t *parallel;
112 struct timeval start, stop, diff;
113
114 if (parse_options(argc, argv, thread_options, thread_usage, 0)) {
115 usage_with_options(thread_usage, thread_options);
116 exit(EXIT_FAILURE);
117 }
118 breakpoints = calloc(thread_params.nbreakpoints, sizeof(breakpoints[0]));
119 parallel = calloc(thread_params.nparallel, sizeof(parallel[0]));
120 if (!breakpoints || !parallel)
121 exit((perror("calloc"), EXIT_FAILURE));
122
123 for (i = 0; i < thread_params.nbreakpoints; i++) {
124 breakpoints[i].fd = breakpoint_setup(&breakpoints[i].watched);
125 if (breakpoints[i].fd == -1)
126 exit((perror("perf_event_open"), EXIT_FAILURE));
127 }
128 gettimeofday(&start, NULL);
129 for (i = 0; i < thread_params.nparallel; i++) {
130 if (pthread_create(¶llel[i], NULL, breakpoint_thread, &repeat))
131 exit((perror("pthread_create"), EXIT_FAILURE));
132 }
133 for (i = 0; i < thread_params.nparallel; i++)
134 pthread_join(parallel[i], NULL);
135 gettimeofday(&stop, NULL);
136 timersub(&stop, &start, &diff);
137 for (i = 0; i < thread_params.nbreakpoints; i++)
138 close(breakpoints[i].fd);
139 free(parallel);
140 free(breakpoints);
141 switch (bench_format) {
142 case BENCH_FORMAT_DEFAULT:
143 printf("# Created/joined %d threads with %d breakpoints and %d parallelism\n",
144 bench_repeat, thread_params.nbreakpoints, thread_params.nparallel);
145 printf(" %14s: %lu.%03lu [sec]\n\n", "Total time",
146 (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
147 result_usec = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
148 printf(" %14lf usecs/op\n",
149 (double)result_usec / bench_repeat / thread_params.nthreads);
150 printf(" %14lf usecs/op/cpu\n",
151 (double)result_usec / bench_repeat /
152 thread_params.nthreads * thread_params.nparallel);
153 break;
154 case BENCH_FORMAT_SIMPLE:
155 printf("%lu.%03lu\n", (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
156 break;
157 default:
158 fprintf(stderr, "Unknown format: %d\n", bench_format);
159 exit(EXIT_FAILURE);
160 }
161 return 0;
162 }
163
164 struct {
165 unsigned int npassive;
166 unsigned int nactive;
167 } enable_params = {
168 .nactive = 0,
169 .npassive = 0,
170 };
171
172 static const struct option enable_options[] = {
173 OPT_UINTEGER('p', "passive", &enable_params.npassive, "Specify amount of passive threads"),
174 OPT_UINTEGER('a', "active", &enable_params.nactive, "Specify amount of active threads"),
175 OPT_END()
176 };
177
178 static const char * const enable_usage[] = {
179 "perf bench breakpoint enable <options>",
180 NULL
181 };
182
183 // The benchmark creates an inheritable breakpoint,
184 // then starts npassive threads that block and nactive threads that actively spin
185 // and then disables and enables the breakpoint bench_repeat times.
bench_breakpoint_enable(int argc,const char ** argv)186 int bench_breakpoint_enable(int argc, const char **argv)
187 {
188 unsigned int i, nthreads, result_usec, done = 0;
189 char watched;
190 int fd;
191 pthread_t *threads;
192 struct timeval start, stop, diff;
193
194 if (parse_options(argc, argv, enable_options, enable_usage, 0)) {
195 usage_with_options(enable_usage, enable_options);
196 exit(EXIT_FAILURE);
197 }
198 fd = breakpoint_setup(&watched);
199 if (fd == -1)
200 exit((perror("perf_event_open"), EXIT_FAILURE));
201 nthreads = enable_params.npassive + enable_params.nactive;
202 threads = calloc(nthreads, sizeof(threads[0]));
203 if (!threads)
204 exit((perror("calloc"), EXIT_FAILURE));
205
206 for (i = 0; i < nthreads; i++) {
207 if (pthread_create(&threads[i], NULL,
208 i < enable_params.npassive ? passive_thread : active_thread, &done))
209 exit((perror("pthread_create"), EXIT_FAILURE));
210 }
211 usleep(10000); // let the threads block
212 gettimeofday(&start, NULL);
213 for (i = 0; i < bench_repeat; i++) {
214 if (ioctl(fd, PERF_EVENT_IOC_DISABLE, 0))
215 exit((perror("ioctl(PERF_EVENT_IOC_DISABLE)"), EXIT_FAILURE));
216 if (ioctl(fd, PERF_EVENT_IOC_ENABLE, 0))
217 exit((perror("ioctl(PERF_EVENT_IOC_ENABLE)"), EXIT_FAILURE));
218 }
219 gettimeofday(&stop, NULL);
220 timersub(&stop, &start, &diff);
221 __atomic_store_n(&done, 1, __ATOMIC_RELAXED);
222 futex_wake(&done, enable_params.npassive, 0);
223 for (i = 0; i < nthreads; i++)
224 pthread_join(threads[i], NULL);
225 free(threads);
226 close(fd);
227 switch (bench_format) {
228 case BENCH_FORMAT_DEFAULT:
229 printf("# Enabled/disabled breakpoint %d time with %d passive and %d active threads\n",
230 bench_repeat, enable_params.npassive, enable_params.nactive);
231 printf(" %14s: %lu.%03lu [sec]\n\n", "Total time",
232 (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
233 result_usec = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
234 printf(" %14lf usecs/op\n", (double)result_usec / bench_repeat);
235 break;
236 case BENCH_FORMAT_SIMPLE:
237 printf("%lu.%03lu\n", (long)diff.tv_sec, (long)(diff.tv_usec / USEC_PER_MSEC));
238 break;
239 default:
240 fprintf(stderr, "Unknown format: %d\n", bench_format);
241 exit(EXIT_FAILURE);
242 }
243 return 0;
244 }
245