xref: /openbmc/linux/samples/bpf/trace_output_user.c (revision 389dcb9df652856294774f0ddc66a3c5bb5f1794)
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.
4  */
5 #include <stdio.h>
6 #include <unistd.h>
7 #include <stdlib.h>
8 #include <stdbool.h>
9 #include <string.h>
10 #include <fcntl.h>
11 #include <poll.h>
12 #include <sys/ioctl.h>
13 #include <linux/perf_event.h>
14 #include <linux/bpf.h>
15 #include <errno.h>
16 #include <assert.h>
17 #include <sys/syscall.h>
18 #include <sys/ioctl.h>
19 #include <sys/mman.h>
20 #include <time.h>
21 #include <signal.h>
22 #include "libbpf.h"
23 #include "bpf_load.h"
24 #include "perf-sys.h"
25 
26 static int pmu_fd;
27 
28 int page_size;
29 int page_cnt = 8;
30 volatile struct perf_event_mmap_page *header;
31 
32 typedef void (*print_fn)(void *data, int size);
33 
34 static int perf_event_mmap(int fd)
35 {
36 	void *base;
37 	int mmap_size;
38 
39 	page_size = getpagesize();
40 	mmap_size = page_size * (page_cnt + 1);
41 
42 	base = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
43 	if (base == MAP_FAILED) {
44 		printf("mmap err\n");
45 		return -1;
46 	}
47 
48 	header = base;
49 	return 0;
50 }
51 
52 static int perf_event_poll(int fd)
53 {
54 	struct pollfd pfd = { .fd = fd, .events = POLLIN };
55 
56 	return poll(&pfd, 1, 1000);
57 }
58 
59 struct perf_event_sample {
60 	struct perf_event_header header;
61 	__u32 size;
62 	char data[];
63 };
64 
65 static void perf_event_read(print_fn fn)
66 {
67 	__u64 data_tail = header->data_tail;
68 	__u64 data_head = header->data_head;
69 	__u64 buffer_size = page_cnt * page_size;
70 	void *base, *begin, *end;
71 	char buf[256];
72 
73 	asm volatile("" ::: "memory"); /* in real code it should be smp_rmb() */
74 	if (data_head == data_tail)
75 		return;
76 
77 	base = ((char *)header) + page_size;
78 
79 	begin = base + data_tail % buffer_size;
80 	end = base + data_head % buffer_size;
81 
82 	while (begin != end) {
83 		struct perf_event_sample *e;
84 
85 		e = begin;
86 		if (begin + e->header.size > base + buffer_size) {
87 			long len = base + buffer_size - begin;
88 
89 			assert(len < e->header.size);
90 			memcpy(buf, begin, len);
91 			memcpy(buf + len, base, e->header.size - len);
92 			e = (void *) buf;
93 			begin = base + e->header.size - len;
94 		} else if (begin + e->header.size == base + buffer_size) {
95 			begin = base;
96 		} else {
97 			begin += e->header.size;
98 		}
99 
100 		if (e->header.type == PERF_RECORD_SAMPLE) {
101 			fn(e->data, e->size);
102 		} else if (e->header.type == PERF_RECORD_LOST) {
103 			struct {
104 				struct perf_event_header header;
105 				__u64 id;
106 				__u64 lost;
107 			} *lost = (void *) e;
108 			printf("lost %lld events\n", lost->lost);
109 		} else {
110 			printf("unknown event type=%d size=%d\n",
111 			       e->header.type, e->header.size);
112 		}
113 	}
114 
115 	__sync_synchronize(); /* smp_mb() */
116 	header->data_tail = data_head;
117 }
118 
119 static __u64 time_get_ns(void)
120 {
121 	struct timespec ts;
122 
123 	clock_gettime(CLOCK_MONOTONIC, &ts);
124 	return ts.tv_sec * 1000000000ull + ts.tv_nsec;
125 }
126 
127 static __u64 start_time;
128 
129 #define MAX_CNT 100000ll
130 
131 static void print_bpf_output(void *data, int size)
132 {
133 	static __u64 cnt;
134 	struct {
135 		__u64 pid;
136 		__u64 cookie;
137 	} *e = data;
138 
139 	if (e->cookie != 0x12345678) {
140 		printf("BUG pid %llx cookie %llx sized %d\n",
141 		       e->pid, e->cookie, size);
142 		kill(0, SIGINT);
143 	}
144 
145 	cnt++;
146 
147 	if (cnt == MAX_CNT) {
148 		printf("recv %lld events per sec\n",
149 		       MAX_CNT * 1000000000ll / (time_get_ns() - start_time));
150 		kill(0, SIGINT);
151 	}
152 }
153 
154 static void test_bpf_perf_event(void)
155 {
156 	struct perf_event_attr attr = {
157 		.sample_type = PERF_SAMPLE_RAW,
158 		.type = PERF_TYPE_SOFTWARE,
159 		.config = PERF_COUNT_SW_BPF_OUTPUT,
160 	};
161 	int key = 0;
162 
163 	pmu_fd = sys_perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
164 
165 	assert(pmu_fd >= 0);
166 	assert(bpf_map_update_elem(map_fd[0], &key, &pmu_fd, BPF_ANY) == 0);
167 	ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE, 0);
168 }
169 
170 int main(int argc, char **argv)
171 {
172 	char filename[256];
173 	FILE *f;
174 
175 	snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
176 
177 	if (load_bpf_file(filename)) {
178 		printf("%s", bpf_log_buf);
179 		return 1;
180 	}
181 
182 	test_bpf_perf_event();
183 
184 	if (perf_event_mmap(pmu_fd) < 0)
185 		return 1;
186 
187 	f = popen("taskset 1 dd if=/dev/zero of=/dev/null", "r");
188 	(void) f;
189 
190 	start_time = time_get_ns();
191 	for (;;) {
192 		perf_event_poll(pmu_fd);
193 		perf_event_read(print_bpf_output);
194 	}
195 
196 	return 0;
197 }
198