1 /* This program is free software; you can redistribute it and/or
2 * modify it under the terms of version 2 of the GNU General Public
3 * License as published by the Free Software Foundation.
12 #include <linux/perf_event.h>
13 #include <linux/bpf.h>
16 #include <sys/syscall.h>
17 #include <sys/ioctl.h>
29 volatile struct perf_event_mmap_page *header;
31 typedef void (*print_fn)(void *data, int size);
33 static int perf_event_mmap(int fd)
38 page_size = getpagesize();
39 mmap_size = page_size * (page_cnt + 1);
41 base = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
42 if (base == MAP_FAILED) {
51 static int perf_event_poll(int fd)
53 struct pollfd pfd = { .fd = fd, .events = POLLIN };
55 return poll(&pfd, 1, 1000);
58 struct perf_event_sample {
59 struct perf_event_header header;
64 static void perf_event_read(print_fn fn)
66 __u64 data_tail = header->data_tail;
67 __u64 data_head = header->data_head;
68 __u64 buffer_size = page_cnt * page_size;
69 void *base, *begin, *end;
72 asm volatile("" ::: "memory"); /* in real code it should be smp_rmb() */
73 if (data_head == data_tail)
76 base = ((char *)header) + page_size;
78 begin = base + data_tail % buffer_size;
79 end = base + data_head % buffer_size;
81 while (begin != end) {
82 struct perf_event_sample *e;
85 if (begin + e->header.size > base + buffer_size) {
86 long len = base + buffer_size - begin;
88 assert(len < e->header.size);
89 memcpy(buf, begin, len);
90 memcpy(buf + len, base, e->header.size - len);
92 begin = base + e->header.size - len;
93 } else if (begin + e->header.size == base + buffer_size) {
96 begin += e->header.size;
99 if (e->header.type == PERF_RECORD_SAMPLE) {
100 fn(e->data, e->size);
101 } else if (e->header.type == PERF_RECORD_LOST) {
103 struct perf_event_header header;
106 } *lost = (void *) e;
107 printf("lost %lld events\n", lost->lost);
109 printf("unknown event type=%d size=%d\n",
110 e->header.type, e->header.size);
114 __sync_synchronize(); /* smp_mb() */
115 header->data_tail = data_head;
118 static __u64 time_get_ns(void)
122 clock_gettime(CLOCK_MONOTONIC, &ts);
123 return ts.tv_sec * 1000000000ull + ts.tv_nsec;
126 static __u64 start_time;
128 #define MAX_CNT 100000ll
130 static void print_bpf_output(void *data, int size)
138 if (e->cookie != 0x12345678) {
139 printf("BUG pid %llx cookie %llx sized %d\n",
140 e->pid, e->cookie, size);
146 if (cnt == MAX_CNT) {
147 printf("recv %lld events per sec\n",
148 MAX_CNT * 1000000000ll / (time_get_ns() - start_time));
153 static void test_bpf_perf_event(void)
155 struct perf_event_attr attr = {
156 .sample_type = PERF_SAMPLE_RAW,
157 .type = PERF_TYPE_SOFTWARE,
158 .config = PERF_COUNT_SW_BPF_OUTPUT,
162 pmu_fd = sys_perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
165 assert(bpf_map_update_elem(map_fd[0], &key, &pmu_fd, BPF_ANY) == 0);
166 ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE, 0);
169 int main(int argc, char **argv)
174 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
176 if (load_bpf_file(filename)) {
177 printf("%s", bpf_log_buf);
181 test_bpf_perf_event();
183 if (perf_event_mmap(pmu_fd) < 0)
186 f = popen("taskset 1 dd if=/dev/zero of=/dev/null", "r");
189 start_time = time_get_ns();
191 perf_event_poll(pmu_fd);
192 perf_event_read(print_bpf_output);