xref: /linux/tools/perf/util/event.c (revision 5499b45190237ca90dd2ac86395cf464fe1f4cc7)
1 #include <linux/types.h>
2 #include "event.h"
3 #include "debug.h"
4 #include "session.h"
5 #include "sort.h"
6 #include "string.h"
7 #include "strlist.h"
8 #include "thread.h"
9 
10 static pid_t event__synthesize_comm(pid_t pid, int full,
11 				    event__handler_t process,
12 				    struct perf_session *session)
13 {
14 	event_t ev;
15 	char filename[PATH_MAX];
16 	char bf[BUFSIZ];
17 	FILE *fp;
18 	size_t size = 0;
19 	DIR *tasks;
20 	struct dirent dirent, *next;
21 	pid_t tgid = 0;
22 
23 	snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
24 
25 	fp = fopen(filename, "r");
26 	if (fp == NULL) {
27 out_race:
28 		/*
29 		 * We raced with a task exiting - just return:
30 		 */
31 		pr_debug("couldn't open %s\n", filename);
32 		return 0;
33 	}
34 
35 	memset(&ev.comm, 0, sizeof(ev.comm));
36 	while (!ev.comm.comm[0] || !ev.comm.pid) {
37 		if (fgets(bf, sizeof(bf), fp) == NULL)
38 			goto out_failure;
39 
40 		if (memcmp(bf, "Name:", 5) == 0) {
41 			char *name = bf + 5;
42 			while (*name && isspace(*name))
43 				++name;
44 			size = strlen(name) - 1;
45 			memcpy(ev.comm.comm, name, size++);
46 		} else if (memcmp(bf, "Tgid:", 5) == 0) {
47 			char *tgids = bf + 5;
48 			while (*tgids && isspace(*tgids))
49 				++tgids;
50 			tgid = ev.comm.pid = atoi(tgids);
51 		}
52 	}
53 
54 	ev.comm.header.type = PERF_RECORD_COMM;
55 	size = ALIGN(size, sizeof(u64));
56 	ev.comm.header.size = sizeof(ev.comm) - (sizeof(ev.comm.comm) - size);
57 
58 	if (!full) {
59 		ev.comm.tid = pid;
60 
61 		process(&ev, session);
62 		goto out_fclose;
63 	}
64 
65 	snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
66 
67 	tasks = opendir(filename);
68 	if (tasks == NULL)
69 		goto out_race;
70 
71 	while (!readdir_r(tasks, &dirent, &next) && next) {
72 		char *end;
73 		pid = strtol(dirent.d_name, &end, 10);
74 		if (*end)
75 			continue;
76 
77 		ev.comm.tid = pid;
78 
79 		process(&ev, session);
80 	}
81 	closedir(tasks);
82 
83 out_fclose:
84 	fclose(fp);
85 	return tgid;
86 
87 out_failure:
88 	pr_warning("couldn't get COMM and pgid, malformed %s\n", filename);
89 	return -1;
90 }
91 
92 static int event__synthesize_mmap_events(pid_t pid, pid_t tgid,
93 					 event__handler_t process,
94 					 struct perf_session *session)
95 {
96 	char filename[PATH_MAX];
97 	FILE *fp;
98 
99 	snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
100 
101 	fp = fopen(filename, "r");
102 	if (fp == NULL) {
103 		/*
104 		 * We raced with a task exiting - just return:
105 		 */
106 		pr_debug("couldn't open %s\n", filename);
107 		return -1;
108 	}
109 
110 	while (1) {
111 		char bf[BUFSIZ], *pbf = bf;
112 		event_t ev = {
113 			.header = {
114 				.type = PERF_RECORD_MMAP,
115 				.misc = 0, /* Just like the kernel, see kernel/perf_event.c __perf_event_mmap */
116 			 },
117 		};
118 		int n;
119 		size_t size;
120 		if (fgets(bf, sizeof(bf), fp) == NULL)
121 			break;
122 
123 		/* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
124 		n = hex2u64(pbf, &ev.mmap.start);
125 		if (n < 0)
126 			continue;
127 		pbf += n + 1;
128 		n = hex2u64(pbf, &ev.mmap.len);
129 		if (n < 0)
130 			continue;
131 		pbf += n + 3;
132 		if (*pbf == 'x') { /* vm_exec */
133 			char *execname = strchr(bf, '/');
134 
135 			/* Catch VDSO */
136 			if (execname == NULL)
137 				execname = strstr(bf, "[vdso]");
138 
139 			if (execname == NULL)
140 				continue;
141 
142 			size = strlen(execname);
143 			execname[size - 1] = '\0'; /* Remove \n */
144 			memcpy(ev.mmap.filename, execname, size);
145 			size = ALIGN(size, sizeof(u64));
146 			ev.mmap.len -= ev.mmap.start;
147 			ev.mmap.header.size = (sizeof(ev.mmap) -
148 					       (sizeof(ev.mmap.filename) - size));
149 			ev.mmap.pid = tgid;
150 			ev.mmap.tid = pid;
151 
152 			process(&ev, session);
153 		}
154 	}
155 
156 	fclose(fp);
157 	return 0;
158 }
159 
160 int event__synthesize_modules(event__handler_t process,
161 			      struct perf_session *session)
162 {
163 	struct rb_node *nd;
164 
165 	for (nd = rb_first(&session->kmaps.maps[MAP__FUNCTION]);
166 	     nd; nd = rb_next(nd)) {
167 		event_t ev;
168 		size_t size;
169 		struct map *pos = rb_entry(nd, struct map, rb_node);
170 
171 		if (pos->dso->kernel)
172 			continue;
173 
174 		size = ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
175 		memset(&ev, 0, sizeof(ev));
176 		ev.mmap.header.misc = 1; /* kernel uses 0 for user space maps, see kernel/perf_event.c __perf_event_mmap */
177 		ev.mmap.header.type = PERF_RECORD_MMAP;
178 		ev.mmap.header.size = (sizeof(ev.mmap) -
179 				        (sizeof(ev.mmap.filename) - size));
180 		ev.mmap.start = pos->start;
181 		ev.mmap.len   = pos->end - pos->start;
182 
183 		memcpy(ev.mmap.filename, pos->dso->long_name,
184 		       pos->dso->long_name_len + 1);
185 		process(&ev, session);
186 	}
187 
188 	return 0;
189 }
190 
191 int event__synthesize_thread(pid_t pid, event__handler_t process,
192 			     struct perf_session *session)
193 {
194 	pid_t tgid = event__synthesize_comm(pid, 1, process, session);
195 	if (tgid == -1)
196 		return -1;
197 	return event__synthesize_mmap_events(pid, tgid, process, session);
198 }
199 
200 void event__synthesize_threads(event__handler_t process,
201 			       struct perf_session *session)
202 {
203 	DIR *proc;
204 	struct dirent dirent, *next;
205 
206 	proc = opendir("/proc");
207 
208 	while (!readdir_r(proc, &dirent, &next) && next) {
209 		char *end;
210 		pid_t pid = strtol(dirent.d_name, &end, 10);
211 
212 		if (*end) /* only interested in proper numerical dirents */
213 			continue;
214 
215 		event__synthesize_thread(pid, process, session);
216 	}
217 
218 	closedir(proc);
219 }
220 
221 struct process_symbol_args {
222 	const char *name;
223 	u64	   start;
224 };
225 
226 static int find_symbol_cb(void *arg, const char *name, char type, u64 start)
227 {
228 	struct process_symbol_args *args = arg;
229 
230 	/*
231 	 * Must be a function or at least an alias, as in PARISC64, where "_text" is
232 	 * an 'A' to the same address as "_stext".
233 	 */
234 	if (!(symbol_type__is_a(type, MAP__FUNCTION) ||
235 	      type == 'A') || strcmp(name, args->name))
236 		return 0;
237 
238 	args->start = start;
239 	return 1;
240 }
241 
242 int event__synthesize_kernel_mmap(event__handler_t process,
243 				  struct perf_session *session,
244 				  const char *symbol_name)
245 {
246 	size_t size;
247 	event_t ev = {
248 		.header = {
249 			.type = PERF_RECORD_MMAP,
250 			.misc = 1, /* kernel uses 0 for user space maps, see kernel/perf_event.c __perf_event_mmap */
251 		},
252 	};
253 	/*
254 	 * We should get this from /sys/kernel/sections/.text, but till that is
255 	 * available use this, and after it is use this as a fallback for older
256 	 * kernels.
257 	 */
258 	struct process_symbol_args args = { .name = symbol_name, };
259 
260 	if (kallsyms__parse("/proc/kallsyms", &args, find_symbol_cb) <= 0)
261 		return -ENOENT;
262 
263 	size = snprintf(ev.mmap.filename, sizeof(ev.mmap.filename),
264 			"[kernel.kallsyms.%s]", symbol_name) + 1;
265 	size = ALIGN(size, sizeof(u64));
266 	ev.mmap.header.size = (sizeof(ev.mmap) - (sizeof(ev.mmap.filename) - size));
267 	ev.mmap.pgoff = args.start;
268 	ev.mmap.start = session->vmlinux_maps[MAP__FUNCTION]->start;
269 	ev.mmap.len   = session->vmlinux_maps[MAP__FUNCTION]->end - ev.mmap.start ;
270 
271 	return process(&ev, session);
272 }
273 
274 static void thread__comm_adjust(struct thread *self)
275 {
276 	char *comm = self->comm;
277 
278 	if (!symbol_conf.col_width_list_str && !symbol_conf.field_sep &&
279 	    (!symbol_conf.comm_list ||
280 	     strlist__has_entry(symbol_conf.comm_list, comm))) {
281 		unsigned int slen = strlen(comm);
282 
283 		if (slen > comms__col_width) {
284 			comms__col_width = slen;
285 			threads__col_width = slen + 6;
286 		}
287 	}
288 }
289 
290 static int thread__set_comm_adjust(struct thread *self, const char *comm)
291 {
292 	int ret = thread__set_comm(self, comm);
293 
294 	if (ret)
295 		return ret;
296 
297 	thread__comm_adjust(self);
298 
299 	return 0;
300 }
301 
302 int event__process_comm(event_t *self, struct perf_session *session)
303 {
304 	struct thread *thread = perf_session__findnew(session, self->comm.pid);
305 
306 	dump_printf(": %s:%d\n", self->comm.comm, self->comm.pid);
307 
308 	if (thread == NULL || thread__set_comm_adjust(thread, self->comm.comm)) {
309 		dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
310 		return -1;
311 	}
312 
313 	return 0;
314 }
315 
316 int event__process_lost(event_t *self, struct perf_session *session)
317 {
318 	dump_printf(": id:%Ld: lost:%Ld\n", self->lost.id, self->lost.lost);
319 	session->events_stats.lost += self->lost.lost;
320 	return 0;
321 }
322 
323 int event__process_mmap(event_t *self, struct perf_session *session)
324 {
325 	struct thread *thread;
326 	struct map *map;
327 
328 	dump_printf(" %d/%d: [%#Lx(%#Lx) @ %#Lx]: %s\n",
329 		    self->mmap.pid, self->mmap.tid, self->mmap.start,
330 		    self->mmap.len, self->mmap.pgoff, self->mmap.filename);
331 
332 	if (self->mmap.pid == 0) {
333 		static const char kmmap_prefix[] = "[kernel.kallsyms.";
334 
335 		if (self->mmap.filename[0] == '/') {
336 			char short_module_name[1024];
337 			char *name = strrchr(self->mmap.filename, '/'), *dot;
338 
339 			if (name == NULL)
340 				goto out_problem;
341 
342 			++name; /* skip / */
343 			dot = strrchr(name, '.');
344 			if (dot == NULL)
345 				goto out_problem;
346 
347 			snprintf(short_module_name, sizeof(short_module_name),
348 				 "[%.*s]", (int)(dot - name), name);
349 			strxfrchar(short_module_name, '-', '_');
350 
351 			map = perf_session__new_module_map(session,
352 							   self->mmap.start,
353 							   self->mmap.filename);
354 			if (map == NULL)
355 				goto out_problem;
356 
357 			name = strdup(short_module_name);
358 			if (name == NULL)
359 				goto out_problem;
360 
361 			map->dso->short_name = name;
362 			map->end = map->start + self->mmap.len;
363 		} else if (memcmp(self->mmap.filename, kmmap_prefix,
364 				sizeof(kmmap_prefix) - 1) == 0) {
365 			const char *symbol_name = (self->mmap.filename +
366 						   sizeof(kmmap_prefix) - 1);
367 			/*
368 			 * Should be there already, from the build-id table in
369 			 * the header.
370 			 */
371 			struct dso *kernel = __dsos__findnew(&dsos__kernel,
372 							     "[kernel.kallsyms]");
373 			if (kernel == NULL)
374 				goto out_problem;
375 
376 			kernel->kernel = 1;
377 			if (__perf_session__create_kernel_maps(session, kernel) < 0)
378 				goto out_problem;
379 
380 			session->vmlinux_maps[MAP__FUNCTION]->start = self->mmap.start;
381 			session->vmlinux_maps[MAP__FUNCTION]->end   = self->mmap.start + self->mmap.len;
382 			/*
383 			 * Be a bit paranoid here, some perf.data file came with
384 			 * a zero sized synthesized MMAP event for the kernel.
385 			 */
386 			if (session->vmlinux_maps[MAP__FUNCTION]->end == 0)
387 				session->vmlinux_maps[MAP__FUNCTION]->end = ~0UL;
388 
389 			perf_session__set_kallsyms_ref_reloc_sym(session, symbol_name,
390 								 self->mmap.pgoff);
391 		}
392 		return 0;
393 	}
394 
395 	thread = perf_session__findnew(session, self->mmap.pid);
396 	map = map__new(&self->mmap, MAP__FUNCTION,
397 		       session->cwd, session->cwdlen);
398 
399 	if (thread == NULL || map == NULL)
400 		goto out_problem;
401 
402 	thread__insert_map(thread, map);
403 	return 0;
404 
405 out_problem:
406 	dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
407 	return 0;
408 }
409 
410 int event__process_task(event_t *self, struct perf_session *session)
411 {
412 	struct thread *thread = perf_session__findnew(session, self->fork.pid);
413 	struct thread *parent = perf_session__findnew(session, self->fork.ppid);
414 
415 	dump_printf("(%d:%d):(%d:%d)\n", self->fork.pid, self->fork.tid,
416 		    self->fork.ppid, self->fork.ptid);
417 	/*
418 	 * A thread clone will have the same PID for both parent and child.
419 	 */
420 	if (thread == parent)
421 		return 0;
422 
423 	if (self->header.type == PERF_RECORD_EXIT)
424 		return 0;
425 
426 	if (thread == NULL || parent == NULL ||
427 	    thread__fork(thread, parent) < 0) {
428 		dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
429 		return -1;
430 	}
431 
432 	return 0;
433 }
434 
435 void thread__find_addr_map(struct thread *self,
436 			   struct perf_session *session, u8 cpumode,
437 			   enum map_type type, u64 addr,
438 			   struct addr_location *al)
439 {
440 	struct map_groups *mg = &self->mg;
441 
442 	al->thread = self;
443 	al->addr = addr;
444 
445 	if (cpumode == PERF_RECORD_MISC_KERNEL) {
446 		al->level = 'k';
447 		mg = &session->kmaps;
448 	} else if (cpumode == PERF_RECORD_MISC_USER)
449 		al->level = '.';
450 	else {
451 		al->level = 'H';
452 		al->map = NULL;
453 		return;
454 	}
455 try_again:
456 	al->map = map_groups__find(mg, type, al->addr);
457 	if (al->map == NULL) {
458 		/*
459 		 * If this is outside of all known maps, and is a negative
460 		 * address, try to look it up in the kernel dso, as it might be
461 		 * a vsyscall or vdso (which executes in user-mode).
462 		 *
463 		 * XXX This is nasty, we should have a symbol list in the
464 		 * "[vdso]" dso, but for now lets use the old trick of looking
465 		 * in the whole kernel symbol list.
466 		 */
467 		if ((long long)al->addr < 0 && mg != &session->kmaps) {
468 			mg = &session->kmaps;
469 			goto try_again;
470 		}
471 	} else
472 		al->addr = al->map->map_ip(al->map, al->addr);
473 }
474 
475 void thread__find_addr_location(struct thread *self,
476 				struct perf_session *session, u8 cpumode,
477 				enum map_type type, u64 addr,
478 				struct addr_location *al,
479 				symbol_filter_t filter)
480 {
481 	thread__find_addr_map(self, session, cpumode, type, addr, al);
482 	if (al->map != NULL)
483 		al->sym = map__find_symbol(al->map, al->addr, filter);
484 	else
485 		al->sym = NULL;
486 }
487 
488 static void dso__calc_col_width(struct dso *self)
489 {
490 	if (!symbol_conf.col_width_list_str && !symbol_conf.field_sep &&
491 	    (!symbol_conf.dso_list ||
492 	     strlist__has_entry(symbol_conf.dso_list, self->name))) {
493 		unsigned int slen = strlen(self->name);
494 		if (slen > dsos__col_width)
495 			dsos__col_width = slen;
496 	}
497 
498 	self->slen_calculated = 1;
499 }
500 
501 int event__preprocess_sample(const event_t *self, struct perf_session *session,
502 			     struct addr_location *al, symbol_filter_t filter)
503 {
504 	u8 cpumode = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
505 	struct thread *thread = perf_session__findnew(session, self->ip.pid);
506 
507 	if (thread == NULL)
508 		return -1;
509 
510 	if (symbol_conf.comm_list &&
511 	    !strlist__has_entry(symbol_conf.comm_list, thread->comm))
512 		goto out_filtered;
513 
514 	dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
515 
516 	thread__find_addr_location(thread, session, cpumode, MAP__FUNCTION,
517 				   self->ip.ip, al, filter);
518 	dump_printf(" ...... dso: %s\n",
519 		    al->map ? al->map->dso->long_name :
520 			al->level == 'H' ? "[hypervisor]" : "<not found>");
521 	/*
522 	 * We have to do this here as we may have a dso with no symbol hit that
523 	 * has a name longer than the ones with symbols sampled.
524 	 */
525 	if (al->map && !sort_dso.elide && !al->map->dso->slen_calculated)
526 		dso__calc_col_width(al->map->dso);
527 
528 	if (symbol_conf.dso_list &&
529 	    (!al->map || !al->map->dso ||
530 	     !(strlist__has_entry(symbol_conf.dso_list, al->map->dso->short_name) ||
531 	       (al->map->dso->short_name != al->map->dso->long_name &&
532 		strlist__has_entry(symbol_conf.dso_list, al->map->dso->long_name)))))
533 		goto out_filtered;
534 
535 	if (symbol_conf.sym_list && al->sym &&
536 	    !strlist__has_entry(symbol_conf.sym_list, al->sym->name))
537 		goto out_filtered;
538 
539 	al->filtered = false;
540 	return 0;
541 
542 out_filtered:
543 	al->filtered = true;
544 	return 0;
545 }
546 
547 int event__parse_sample(event_t *event, u64 type, struct sample_data *data)
548 {
549 	u64 *array = event->sample.array;
550 
551 	if (type & PERF_SAMPLE_IP) {
552 		data->ip = event->ip.ip;
553 		array++;
554 	}
555 
556 	if (type & PERF_SAMPLE_TID) {
557 		u32 *p = (u32 *)array;
558 		data->pid = p[0];
559 		data->tid = p[1];
560 		array++;
561 	}
562 
563 	if (type & PERF_SAMPLE_TIME) {
564 		data->time = *array;
565 		array++;
566 	}
567 
568 	if (type & PERF_SAMPLE_ADDR) {
569 		data->addr = *array;
570 		array++;
571 	}
572 
573 	if (type & PERF_SAMPLE_ID) {
574 		data->id = *array;
575 		array++;
576 	}
577 
578 	if (type & PERF_SAMPLE_STREAM_ID) {
579 		data->stream_id = *array;
580 		array++;
581 	}
582 
583 	if (type & PERF_SAMPLE_CPU) {
584 		u32 *p = (u32 *)array;
585 		data->cpu = *p;
586 		array++;
587 	}
588 
589 	if (type & PERF_SAMPLE_PERIOD) {
590 		data->period = *array;
591 		array++;
592 	}
593 
594 	if (type & PERF_SAMPLE_READ) {
595 		pr_debug("PERF_SAMPLE_READ is unsuported for now\n");
596 		return -1;
597 	}
598 
599 	if (type & PERF_SAMPLE_CALLCHAIN) {
600 		data->callchain = (struct ip_callchain *)array;
601 		array += 1 + data->callchain->nr;
602 	}
603 
604 	if (type & PERF_SAMPLE_RAW) {
605 		u32 *p = (u32 *)array;
606 		data->raw_size = *p;
607 		p++;
608 		data->raw_data = p;
609 	}
610 
611 	return 0;
612 }
613