xref: /openbmc/linux/tools/perf/builtin-trace.c (revision e8e0929d)
1 #include "builtin.h"
2 
3 #include "util/util.h"
4 #include "util/cache.h"
5 #include "util/symbol.h"
6 #include "util/thread.h"
7 #include "util/header.h"
8 
9 #include "util/parse-options.h"
10 
11 #include "perf.h"
12 #include "util/debug.h"
13 
14 #include "util/trace-event.h"
15 
16 static char		const *input_name = "perf.data";
17 static int		input;
18 static unsigned long	page_size;
19 static unsigned long	mmap_window = 32;
20 
21 static unsigned long	total = 0;
22 static unsigned long	total_comm = 0;
23 
24 static struct rb_root	threads;
25 static struct thread	*last_match;
26 
27 static struct perf_header *header;
28 static u64		sample_type;
29 
30 
31 static int
32 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
33 {
34 	struct thread *thread;
35 
36 	thread = threads__findnew(event->comm.pid, &threads, &last_match);
37 
38 	dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
39 		(void *)(offset + head),
40 		(void *)(long)(event->header.size),
41 		event->comm.comm, event->comm.pid);
42 
43 	if (thread == NULL ||
44 	    thread__set_comm(thread, event->comm.comm)) {
45 		dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
46 		return -1;
47 	}
48 	total_comm++;
49 
50 	return 0;
51 }
52 
53 static int
54 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
55 {
56 	char level;
57 	int show = 0;
58 	struct dso *dso = NULL;
59 	struct thread *thread;
60 	u64 ip = event->ip.ip;
61 	u64 timestamp = -1;
62 	u32 cpu = -1;
63 	u64 period = 1;
64 	void *more_data = event->ip.__more_data;
65 	int cpumode;
66 
67 	thread = threads__findnew(event->ip.pid, &threads, &last_match);
68 
69 	if (sample_type & PERF_SAMPLE_TIME) {
70 		timestamp = *(u64 *)more_data;
71 		more_data += sizeof(u64);
72 	}
73 
74 	if (sample_type & PERF_SAMPLE_CPU) {
75 		cpu = *(u32 *)more_data;
76 		more_data += sizeof(u32);
77 		more_data += sizeof(u32); /* reserved */
78 	}
79 
80 	if (sample_type & PERF_SAMPLE_PERIOD) {
81 		period = *(u64 *)more_data;
82 		more_data += sizeof(u64);
83 	}
84 
85 	dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
86 		(void *)(offset + head),
87 		(void *)(long)(event->header.size),
88 		event->header.misc,
89 		event->ip.pid, event->ip.tid,
90 		(void *)(long)ip,
91 		(long long)period);
92 
93 	dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
94 
95 	if (thread == NULL) {
96 		eprintf("problem processing %d event, skipping it.\n",
97 			event->header.type);
98 		return -1;
99 	}
100 
101 	cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
102 
103 	if (cpumode == PERF_RECORD_MISC_KERNEL) {
104 		show = SHOW_KERNEL;
105 		level = 'k';
106 
107 		dso = kernel_dso;
108 
109 		dump_printf(" ...... dso: %s\n", dso->name);
110 
111 	} else if (cpumode == PERF_RECORD_MISC_USER) {
112 
113 		show = SHOW_USER;
114 		level = '.';
115 
116 	} else {
117 		show = SHOW_HV;
118 		level = 'H';
119 
120 		dso = hypervisor_dso;
121 
122 		dump_printf(" ...... dso: [hypervisor]\n");
123 	}
124 
125 	if (sample_type & PERF_SAMPLE_RAW) {
126 		struct {
127 			u32 size;
128 			char data[0];
129 		} *raw = more_data;
130 
131 		/*
132 		 * FIXME: better resolve from pid from the struct trace_entry
133 		 * field, although it should be the same than this perf
134 		 * event pid
135 		 */
136 		print_event(cpu, raw->data, raw->size, timestamp, thread->comm);
137 	}
138 	total += period;
139 
140 	return 0;
141 }
142 
143 static int
144 process_event(event_t *event, unsigned long offset, unsigned long head)
145 {
146 	trace_event(event);
147 
148 	switch (event->header.type) {
149 	case PERF_RECORD_MMAP ... PERF_RECORD_LOST:
150 		return 0;
151 
152 	case PERF_RECORD_COMM:
153 		return process_comm_event(event, offset, head);
154 
155 	case PERF_RECORD_EXIT ... PERF_RECORD_READ:
156 		return 0;
157 
158 	case PERF_RECORD_SAMPLE:
159 		return process_sample_event(event, offset, head);
160 
161 	case PERF_RECORD_MAX:
162 	default:
163 		return -1;
164 	}
165 
166 	return 0;
167 }
168 
169 static int __cmd_trace(void)
170 {
171 	int ret, rc = EXIT_FAILURE;
172 	unsigned long offset = 0;
173 	unsigned long head = 0;
174 	struct stat perf_stat;
175 	event_t *event;
176 	uint32_t size;
177 	char *buf;
178 
179 	trace_report();
180 	register_idle_thread(&threads, &last_match);
181 
182 	input = open(input_name, O_RDONLY);
183 	if (input < 0) {
184 		perror("failed to open file");
185 		exit(-1);
186 	}
187 
188 	ret = fstat(input, &perf_stat);
189 	if (ret < 0) {
190 		perror("failed to stat file");
191 		exit(-1);
192 	}
193 
194 	if (!perf_stat.st_size) {
195 		fprintf(stderr, "zero-sized file, nothing to do!\n");
196 		exit(0);
197 	}
198 	header = perf_header__read(input);
199 	head = header->data_offset;
200 	sample_type = perf_header__sample_type(header);
201 
202 	if (!(sample_type & PERF_SAMPLE_RAW))
203 		die("No trace sample to read. Did you call perf record "
204 		    "without -R?");
205 
206 	if (load_kernel() < 0) {
207 		perror("failed to load kernel symbols");
208 		return EXIT_FAILURE;
209 	}
210 
211 remap:
212 	buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
213 			   MAP_SHARED, input, offset);
214 	if (buf == MAP_FAILED) {
215 		perror("failed to mmap file");
216 		exit(-1);
217 	}
218 
219 more:
220 	event = (event_t *)(buf + head);
221 
222 	size = event->header.size;
223 	if (!size)
224 		size = 8;
225 
226 	if (head + event->header.size >= page_size * mmap_window) {
227 		unsigned long shift = page_size * (head / page_size);
228 		int res;
229 
230 		res = munmap(buf, page_size * mmap_window);
231 		assert(res == 0);
232 
233 		offset += shift;
234 		head -= shift;
235 		goto remap;
236 	}
237 
238 	size = event->header.size;
239 
240 
241 	if (!size || process_event(event, offset, head) < 0) {
242 
243 		/*
244 		 * assume we lost track of the stream, check alignment, and
245 		 * increment a single u64 in the hope to catch on again 'soon'.
246 		 */
247 
248 		if (unlikely(head & 7))
249 			head &= ~7ULL;
250 
251 		size = 8;
252 	}
253 
254 	head += size;
255 
256 	if (offset + head < (unsigned long)perf_stat.st_size)
257 		goto more;
258 
259 	rc = EXIT_SUCCESS;
260 	close(input);
261 
262 	return rc;
263 }
264 
265 static const char * const annotate_usage[] = {
266 	"perf trace [<options>] <command>",
267 	NULL
268 };
269 
270 static const struct option options[] = {
271 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
272 		    "dump raw trace in ASCII"),
273 	OPT_BOOLEAN('v', "verbose", &verbose,
274 		    "be more verbose (show symbol address, etc)"),
275 	OPT_END()
276 };
277 
278 int cmd_trace(int argc, const char **argv, const char *prefix __used)
279 {
280 	symbol__init();
281 	page_size = getpagesize();
282 
283 	argc = parse_options(argc, argv, options, annotate_usage, 0);
284 	if (argc) {
285 		/*
286 		 * Special case: if there's an argument left then assume tha
287 		 * it's a symbol filter:
288 		 */
289 		if (argc > 1)
290 			usage_with_options(annotate_usage, options);
291 	}
292 
293 
294 	setup_pager();
295 
296 	return __cmd_trace();
297 }
298