xref: /openbmc/linux/tools/perf/util/session.c (revision 7405ed10)
1 #define _FILE_OFFSET_BITS 64
2 
3 #include <linux/kernel.h>
4 
5 #include <byteswap.h>
6 #include <unistd.h>
7 #include <sys/types.h>
8 #include <sys/mman.h>
9 
10 #include "evlist.h"
11 #include "evsel.h"
12 #include "session.h"
13 #include "tool.h"
14 #include "sort.h"
15 #include "util.h"
16 #include "cpumap.h"
17 #include "event-parse.h"
18 
19 static int perf_session__open(struct perf_session *self, bool force)
20 {
21 	struct stat input_stat;
22 
23 	if (!strcmp(self->filename, "-")) {
24 		self->fd_pipe = true;
25 		self->fd = STDIN_FILENO;
26 
27 		if (perf_session__read_header(self, self->fd) < 0)
28 			pr_err("incompatible file format (rerun with -v to learn more)");
29 
30 		return 0;
31 	}
32 
33 	self->fd = open(self->filename, O_RDONLY);
34 	if (self->fd < 0) {
35 		int err = errno;
36 
37 		pr_err("failed to open %s: %s", self->filename, strerror(err));
38 		if (err == ENOENT && !strcmp(self->filename, "perf.data"))
39 			pr_err("  (try 'perf record' first)");
40 		pr_err("\n");
41 		return -errno;
42 	}
43 
44 	if (fstat(self->fd, &input_stat) < 0)
45 		goto out_close;
46 
47 	if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
48 		pr_err("file %s not owned by current user or root\n",
49 		       self->filename);
50 		goto out_close;
51 	}
52 
53 	if (!input_stat.st_size) {
54 		pr_info("zero-sized file (%s), nothing to do!\n",
55 			self->filename);
56 		goto out_close;
57 	}
58 
59 	if (perf_session__read_header(self, self->fd) < 0) {
60 		pr_err("incompatible file format (rerun with -v to learn more)");
61 		goto out_close;
62 	}
63 
64 	if (!perf_evlist__valid_sample_type(self->evlist)) {
65 		pr_err("non matching sample_type");
66 		goto out_close;
67 	}
68 
69 	if (!perf_evlist__valid_sample_id_all(self->evlist)) {
70 		pr_err("non matching sample_id_all");
71 		goto out_close;
72 	}
73 
74 	self->size = input_stat.st_size;
75 	return 0;
76 
77 out_close:
78 	close(self->fd);
79 	self->fd = -1;
80 	return -1;
81 }
82 
83 void perf_session__set_id_hdr_size(struct perf_session *session)
84 {
85 	u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
86 
87 	session->host_machine.id_hdr_size = id_hdr_size;
88 	machines__set_id_hdr_size(&session->machines, id_hdr_size);
89 }
90 
91 int perf_session__create_kernel_maps(struct perf_session *self)
92 {
93 	int ret = machine__create_kernel_maps(&self->host_machine);
94 
95 	if (ret >= 0)
96 		ret = machines__create_guest_kernel_maps(&self->machines);
97 	return ret;
98 }
99 
100 static void perf_session__destroy_kernel_maps(struct perf_session *self)
101 {
102 	machine__destroy_kernel_maps(&self->host_machine);
103 	machines__destroy_guest_kernel_maps(&self->machines);
104 }
105 
106 struct perf_session *perf_session__new(const char *filename, int mode,
107 				       bool force, bool repipe,
108 				       struct perf_tool *tool)
109 {
110 	struct perf_session *self;
111 	struct stat st;
112 	size_t len;
113 
114 	if (!filename || !strlen(filename)) {
115 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
116 			filename = "-";
117 		else
118 			filename = "perf.data";
119 	}
120 
121 	len = strlen(filename);
122 	self = zalloc(sizeof(*self) + len);
123 
124 	if (self == NULL)
125 		goto out;
126 
127 	memcpy(self->filename, filename, len);
128 	/*
129 	 * On 64bit we can mmap the data file in one go. No need for tiny mmap
130 	 * slices. On 32bit we use 32MB.
131 	 */
132 #if BITS_PER_LONG == 64
133 	self->mmap_window = ULLONG_MAX;
134 #else
135 	self->mmap_window = 32 * 1024 * 1024ULL;
136 #endif
137 	self->machines = RB_ROOT;
138 	self->repipe = repipe;
139 	INIT_LIST_HEAD(&self->ordered_samples.samples);
140 	INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
141 	INIT_LIST_HEAD(&self->ordered_samples.to_free);
142 	machine__init(&self->host_machine, "", HOST_KERNEL_ID);
143 	hists__init(&self->hists);
144 
145 	if (mode == O_RDONLY) {
146 		if (perf_session__open(self, force) < 0)
147 			goto out_delete;
148 		perf_session__set_id_hdr_size(self);
149 	} else if (mode == O_WRONLY) {
150 		/*
151 		 * In O_RDONLY mode this will be performed when reading the
152 		 * kernel MMAP event, in perf_event__process_mmap().
153 		 */
154 		if (perf_session__create_kernel_maps(self) < 0)
155 			goto out_delete;
156 	}
157 
158 	if (tool && tool->ordering_requires_timestamps &&
159 	    tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
160 		dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
161 		tool->ordered_samples = false;
162 	}
163 
164 out:
165 	return self;
166 out_delete:
167 	perf_session__delete(self);
168 	return NULL;
169 }
170 
171 static void machine__delete_dead_threads(struct machine *machine)
172 {
173 	struct thread *n, *t;
174 
175 	list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
176 		list_del(&t->node);
177 		thread__delete(t);
178 	}
179 }
180 
181 static void perf_session__delete_dead_threads(struct perf_session *session)
182 {
183 	machine__delete_dead_threads(&session->host_machine);
184 }
185 
186 static void machine__delete_threads(struct machine *self)
187 {
188 	struct rb_node *nd = rb_first(&self->threads);
189 
190 	while (nd) {
191 		struct thread *t = rb_entry(nd, struct thread, rb_node);
192 
193 		rb_erase(&t->rb_node, &self->threads);
194 		nd = rb_next(nd);
195 		thread__delete(t);
196 	}
197 }
198 
199 static void perf_session__delete_threads(struct perf_session *session)
200 {
201 	machine__delete_threads(&session->host_machine);
202 }
203 
204 void perf_session__delete(struct perf_session *self)
205 {
206 	perf_session__destroy_kernel_maps(self);
207 	perf_session__delete_dead_threads(self);
208 	perf_session__delete_threads(self);
209 	machine__exit(&self->host_machine);
210 	close(self->fd);
211 	free(self);
212 }
213 
214 void machine__remove_thread(struct machine *self, struct thread *th)
215 {
216 	self->last_match = NULL;
217 	rb_erase(&th->rb_node, &self->threads);
218 	/*
219 	 * We may have references to this thread, for instance in some hist_entry
220 	 * instances, so just move them to a separate list.
221 	 */
222 	list_add_tail(&th->node, &self->dead_threads);
223 }
224 
225 static bool symbol__match_parent_regex(struct symbol *sym)
226 {
227 	if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
228 		return 1;
229 
230 	return 0;
231 }
232 
233 static const u8 cpumodes[] = {
234 	PERF_RECORD_MISC_USER,
235 	PERF_RECORD_MISC_KERNEL,
236 	PERF_RECORD_MISC_GUEST_USER,
237 	PERF_RECORD_MISC_GUEST_KERNEL
238 };
239 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
240 
241 static void ip__resolve_ams(struct machine *self, struct thread *thread,
242 			    struct addr_map_symbol *ams,
243 			    u64 ip)
244 {
245 	struct addr_location al;
246 	size_t i;
247 	u8 m;
248 
249 	memset(&al, 0, sizeof(al));
250 
251 	for (i = 0; i < NCPUMODES; i++) {
252 		m = cpumodes[i];
253 		/*
254 		 * We cannot use the header.misc hint to determine whether a
255 		 * branch stack address is user, kernel, guest, hypervisor.
256 		 * Branches may straddle the kernel/user/hypervisor boundaries.
257 		 * Thus, we have to try consecutively until we find a match
258 		 * or else, the symbol is unknown
259 		 */
260 		thread__find_addr_location(thread, self, m, MAP__FUNCTION,
261 				ip, &al, NULL);
262 		if (al.sym)
263 			goto found;
264 	}
265 found:
266 	ams->addr = ip;
267 	ams->al_addr = al.addr;
268 	ams->sym = al.sym;
269 	ams->map = al.map;
270 }
271 
272 struct branch_info *machine__resolve_bstack(struct machine *self,
273 					    struct thread *thr,
274 					    struct branch_stack *bs)
275 {
276 	struct branch_info *bi;
277 	unsigned int i;
278 
279 	bi = calloc(bs->nr, sizeof(struct branch_info));
280 	if (!bi)
281 		return NULL;
282 
283 	for (i = 0; i < bs->nr; i++) {
284 		ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
285 		ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
286 		bi[i].flags = bs->entries[i].flags;
287 	}
288 	return bi;
289 }
290 
291 int machine__resolve_callchain(struct machine *self,
292 			       struct thread *thread,
293 			       struct ip_callchain *chain,
294 			       struct symbol **parent)
295 {
296 	u8 cpumode = PERF_RECORD_MISC_USER;
297 	unsigned int i;
298 	int err;
299 
300 	callchain_cursor_reset(&callchain_cursor);
301 
302 	if (chain->nr > PERF_MAX_STACK_DEPTH) {
303 		pr_warning("corrupted callchain. skipping...\n");
304 		return 0;
305 	}
306 
307 	for (i = 0; i < chain->nr; i++) {
308 		u64 ip;
309 		struct addr_location al;
310 
311 		if (callchain_param.order == ORDER_CALLEE)
312 			ip = chain->ips[i];
313 		else
314 			ip = chain->ips[chain->nr - i - 1];
315 
316 		if (ip >= PERF_CONTEXT_MAX) {
317 			switch (ip) {
318 			case PERF_CONTEXT_HV:
319 				cpumode = PERF_RECORD_MISC_HYPERVISOR;	break;
320 			case PERF_CONTEXT_KERNEL:
321 				cpumode = PERF_RECORD_MISC_KERNEL;	break;
322 			case PERF_CONTEXT_USER:
323 				cpumode = PERF_RECORD_MISC_USER;	break;
324 			default:
325 				pr_debug("invalid callchain context: "
326 					 "%"PRId64"\n", (s64) ip);
327 				/*
328 				 * It seems the callchain is corrupted.
329 				 * Discard all.
330 				 */
331 				callchain_cursor_reset(&callchain_cursor);
332 				return 0;
333 			}
334 			continue;
335 		}
336 
337 		al.filtered = false;
338 		thread__find_addr_location(thread, self, cpumode,
339 					   MAP__FUNCTION, ip, &al, NULL);
340 		if (al.sym != NULL) {
341 			if (sort__has_parent && !*parent &&
342 			    symbol__match_parent_regex(al.sym))
343 				*parent = al.sym;
344 			if (!symbol_conf.use_callchain)
345 				break;
346 		}
347 
348 		err = callchain_cursor_append(&callchain_cursor,
349 					      ip, al.map, al.sym);
350 		if (err)
351 			return err;
352 	}
353 
354 	return 0;
355 }
356 
357 static int process_event_synth_tracing_data_stub(union perf_event *event __used,
358 						 struct perf_session *session __used)
359 {
360 	dump_printf(": unhandled!\n");
361 	return 0;
362 }
363 
364 static int process_event_synth_attr_stub(union perf_event *event __used,
365 					 struct perf_evlist **pevlist __used)
366 {
367 	dump_printf(": unhandled!\n");
368 	return 0;
369 }
370 
371 static int process_event_sample_stub(struct perf_tool *tool __used,
372 				     union perf_event *event __used,
373 				     struct perf_sample *sample __used,
374 				     struct perf_evsel *evsel __used,
375 				     struct machine *machine __used)
376 {
377 	dump_printf(": unhandled!\n");
378 	return 0;
379 }
380 
381 static int process_event_stub(struct perf_tool *tool __used,
382 			      union perf_event *event __used,
383 			      struct perf_sample *sample __used,
384 			      struct machine *machine __used)
385 {
386 	dump_printf(": unhandled!\n");
387 	return 0;
388 }
389 
390 static int process_finished_round_stub(struct perf_tool *tool __used,
391 				       union perf_event *event __used,
392 				       struct perf_session *perf_session __used)
393 {
394 	dump_printf(": unhandled!\n");
395 	return 0;
396 }
397 
398 static int process_event_type_stub(struct perf_tool *tool __used,
399 				   union perf_event *event __used)
400 {
401 	dump_printf(": unhandled!\n");
402 	return 0;
403 }
404 
405 static int process_finished_round(struct perf_tool *tool,
406 				  union perf_event *event,
407 				  struct perf_session *session);
408 
409 static void perf_tool__fill_defaults(struct perf_tool *tool)
410 {
411 	if (tool->sample == NULL)
412 		tool->sample = process_event_sample_stub;
413 	if (tool->mmap == NULL)
414 		tool->mmap = process_event_stub;
415 	if (tool->comm == NULL)
416 		tool->comm = process_event_stub;
417 	if (tool->fork == NULL)
418 		tool->fork = process_event_stub;
419 	if (tool->exit == NULL)
420 		tool->exit = process_event_stub;
421 	if (tool->lost == NULL)
422 		tool->lost = perf_event__process_lost;
423 	if (tool->read == NULL)
424 		tool->read = process_event_sample_stub;
425 	if (tool->throttle == NULL)
426 		tool->throttle = process_event_stub;
427 	if (tool->unthrottle == NULL)
428 		tool->unthrottle = process_event_stub;
429 	if (tool->attr == NULL)
430 		tool->attr = process_event_synth_attr_stub;
431 	if (tool->event_type == NULL)
432 		tool->event_type = process_event_type_stub;
433 	if (tool->tracing_data == NULL)
434 		tool->tracing_data = process_event_synth_tracing_data_stub;
435 	if (tool->build_id == NULL)
436 		tool->build_id = process_finished_round_stub;
437 	if (tool->finished_round == NULL) {
438 		if (tool->ordered_samples)
439 			tool->finished_round = process_finished_round;
440 		else
441 			tool->finished_round = process_finished_round_stub;
442 	}
443 }
444 
445 void mem_bswap_32(void *src, int byte_size)
446 {
447 	u32 *m = src;
448 	while (byte_size > 0) {
449 		*m = bswap_32(*m);
450 		byte_size -= sizeof(u32);
451 		++m;
452 	}
453 }
454 
455 void mem_bswap_64(void *src, int byte_size)
456 {
457 	u64 *m = src;
458 
459 	while (byte_size > 0) {
460 		*m = bswap_64(*m);
461 		byte_size -= sizeof(u64);
462 		++m;
463 	}
464 }
465 
466 static void swap_sample_id_all(union perf_event *event, void *data)
467 {
468 	void *end = (void *) event + event->header.size;
469 	int size = end - data;
470 
471 	BUG_ON(size % sizeof(u64));
472 	mem_bswap_64(data, size);
473 }
474 
475 static void perf_event__all64_swap(union perf_event *event,
476 				   bool sample_id_all __used)
477 {
478 	struct perf_event_header *hdr = &event->header;
479 	mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
480 }
481 
482 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
483 {
484 	event->comm.pid = bswap_32(event->comm.pid);
485 	event->comm.tid = bswap_32(event->comm.tid);
486 
487 	if (sample_id_all) {
488 		void *data = &event->comm.comm;
489 
490 		data += ALIGN(strlen(data) + 1, sizeof(u64));
491 		swap_sample_id_all(event, data);
492 	}
493 }
494 
495 static void perf_event__mmap_swap(union perf_event *event,
496 				  bool sample_id_all)
497 {
498 	event->mmap.pid	  = bswap_32(event->mmap.pid);
499 	event->mmap.tid	  = bswap_32(event->mmap.tid);
500 	event->mmap.start = bswap_64(event->mmap.start);
501 	event->mmap.len	  = bswap_64(event->mmap.len);
502 	event->mmap.pgoff = bswap_64(event->mmap.pgoff);
503 
504 	if (sample_id_all) {
505 		void *data = &event->mmap.filename;
506 
507 		data += ALIGN(strlen(data) + 1, sizeof(u64));
508 		swap_sample_id_all(event, data);
509 	}
510 }
511 
512 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
513 {
514 	event->fork.pid	 = bswap_32(event->fork.pid);
515 	event->fork.tid	 = bswap_32(event->fork.tid);
516 	event->fork.ppid = bswap_32(event->fork.ppid);
517 	event->fork.ptid = bswap_32(event->fork.ptid);
518 	event->fork.time = bswap_64(event->fork.time);
519 
520 	if (sample_id_all)
521 		swap_sample_id_all(event, &event->fork + 1);
522 }
523 
524 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
525 {
526 	event->read.pid		 = bswap_32(event->read.pid);
527 	event->read.tid		 = bswap_32(event->read.tid);
528 	event->read.value	 = bswap_64(event->read.value);
529 	event->read.time_enabled = bswap_64(event->read.time_enabled);
530 	event->read.time_running = bswap_64(event->read.time_running);
531 	event->read.id		 = bswap_64(event->read.id);
532 
533 	if (sample_id_all)
534 		swap_sample_id_all(event, &event->read + 1);
535 }
536 
537 static u8 revbyte(u8 b)
538 {
539 	int rev = (b >> 4) | ((b & 0xf) << 4);
540 	rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
541 	rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
542 	return (u8) rev;
543 }
544 
545 /*
546  * XXX this is hack in attempt to carry flags bitfield
547  * throught endian village. ABI says:
548  *
549  * Bit-fields are allocated from right to left (least to most significant)
550  * on little-endian implementations and from left to right (most to least
551  * significant) on big-endian implementations.
552  *
553  * The above seems to be byte specific, so we need to reverse each
554  * byte of the bitfield. 'Internet' also says this might be implementation
555  * specific and we probably need proper fix and carry perf_event_attr
556  * bitfield flags in separate data file FEAT_ section. Thought this seems
557  * to work for now.
558  */
559 static void swap_bitfield(u8 *p, unsigned len)
560 {
561 	unsigned i;
562 
563 	for (i = 0; i < len; i++) {
564 		*p = revbyte(*p);
565 		p++;
566 	}
567 }
568 
569 /* exported for swapping attributes in file header */
570 void perf_event__attr_swap(struct perf_event_attr *attr)
571 {
572 	attr->type		= bswap_32(attr->type);
573 	attr->size		= bswap_32(attr->size);
574 	attr->config		= bswap_64(attr->config);
575 	attr->sample_period	= bswap_64(attr->sample_period);
576 	attr->sample_type	= bswap_64(attr->sample_type);
577 	attr->read_format	= bswap_64(attr->read_format);
578 	attr->wakeup_events	= bswap_32(attr->wakeup_events);
579 	attr->bp_type		= bswap_32(attr->bp_type);
580 	attr->bp_addr		= bswap_64(attr->bp_addr);
581 	attr->bp_len		= bswap_64(attr->bp_len);
582 
583 	swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
584 }
585 
586 static void perf_event__hdr_attr_swap(union perf_event *event,
587 				      bool sample_id_all __used)
588 {
589 	size_t size;
590 
591 	perf_event__attr_swap(&event->attr.attr);
592 
593 	size = event->header.size;
594 	size -= (void *)&event->attr.id - (void *)event;
595 	mem_bswap_64(event->attr.id, size);
596 }
597 
598 static void perf_event__event_type_swap(union perf_event *event,
599 					bool sample_id_all __used)
600 {
601 	event->event_type.event_type.event_id =
602 		bswap_64(event->event_type.event_type.event_id);
603 }
604 
605 static void perf_event__tracing_data_swap(union perf_event *event,
606 					  bool sample_id_all __used)
607 {
608 	event->tracing_data.size = bswap_32(event->tracing_data.size);
609 }
610 
611 typedef void (*perf_event__swap_op)(union perf_event *event,
612 				    bool sample_id_all);
613 
614 static perf_event__swap_op perf_event__swap_ops[] = {
615 	[PERF_RECORD_MMAP]		  = perf_event__mmap_swap,
616 	[PERF_RECORD_COMM]		  = perf_event__comm_swap,
617 	[PERF_RECORD_FORK]		  = perf_event__task_swap,
618 	[PERF_RECORD_EXIT]		  = perf_event__task_swap,
619 	[PERF_RECORD_LOST]		  = perf_event__all64_swap,
620 	[PERF_RECORD_READ]		  = perf_event__read_swap,
621 	[PERF_RECORD_SAMPLE]		  = perf_event__all64_swap,
622 	[PERF_RECORD_HEADER_ATTR]	  = perf_event__hdr_attr_swap,
623 	[PERF_RECORD_HEADER_EVENT_TYPE]	  = perf_event__event_type_swap,
624 	[PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
625 	[PERF_RECORD_HEADER_BUILD_ID]	  = NULL,
626 	[PERF_RECORD_HEADER_MAX]	  = NULL,
627 };
628 
629 struct sample_queue {
630 	u64			timestamp;
631 	u64			file_offset;
632 	union perf_event	*event;
633 	struct list_head	list;
634 };
635 
636 static void perf_session_free_sample_buffers(struct perf_session *session)
637 {
638 	struct ordered_samples *os = &session->ordered_samples;
639 
640 	while (!list_empty(&os->to_free)) {
641 		struct sample_queue *sq;
642 
643 		sq = list_entry(os->to_free.next, struct sample_queue, list);
644 		list_del(&sq->list);
645 		free(sq);
646 	}
647 }
648 
649 static int perf_session_deliver_event(struct perf_session *session,
650 				      union perf_event *event,
651 				      struct perf_sample *sample,
652 				      struct perf_tool *tool,
653 				      u64 file_offset);
654 
655 static void flush_sample_queue(struct perf_session *s,
656 			       struct perf_tool *tool)
657 {
658 	struct ordered_samples *os = &s->ordered_samples;
659 	struct list_head *head = &os->samples;
660 	struct sample_queue *tmp, *iter;
661 	struct perf_sample sample;
662 	u64 limit = os->next_flush;
663 	u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
664 	unsigned idx = 0, progress_next = os->nr_samples / 16;
665 	int ret;
666 
667 	if (!tool->ordered_samples || !limit)
668 		return;
669 
670 	list_for_each_entry_safe(iter, tmp, head, list) {
671 		if (iter->timestamp > limit)
672 			break;
673 
674 		ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample,
675 						s->header.needs_swap);
676 		if (ret)
677 			pr_err("Can't parse sample, err = %d\n", ret);
678 		else
679 			perf_session_deliver_event(s, iter->event, &sample, tool,
680 						   iter->file_offset);
681 
682 		os->last_flush = iter->timestamp;
683 		list_del(&iter->list);
684 		list_add(&iter->list, &os->sample_cache);
685 		if (++idx >= progress_next) {
686 			progress_next += os->nr_samples / 16;
687 			ui_progress__update(idx, os->nr_samples,
688 					    "Processing time ordered events...");
689 		}
690 	}
691 
692 	if (list_empty(head)) {
693 		os->last_sample = NULL;
694 	} else if (last_ts <= limit) {
695 		os->last_sample =
696 			list_entry(head->prev, struct sample_queue, list);
697 	}
698 
699 	os->nr_samples = 0;
700 }
701 
702 /*
703  * When perf record finishes a pass on every buffers, it records this pseudo
704  * event.
705  * We record the max timestamp t found in the pass n.
706  * Assuming these timestamps are monotonic across cpus, we know that if
707  * a buffer still has events with timestamps below t, they will be all
708  * available and then read in the pass n + 1.
709  * Hence when we start to read the pass n + 2, we can safely flush every
710  * events with timestamps below t.
711  *
712  *    ============ PASS n =================
713  *       CPU 0         |   CPU 1
714  *                     |
715  *    cnt1 timestamps  |   cnt2 timestamps
716  *          1          |         2
717  *          2          |         3
718  *          -          |         4  <--- max recorded
719  *
720  *    ============ PASS n + 1 ==============
721  *       CPU 0         |   CPU 1
722  *                     |
723  *    cnt1 timestamps  |   cnt2 timestamps
724  *          3          |         5
725  *          4          |         6
726  *          5          |         7 <---- max recorded
727  *
728  *      Flush every events below timestamp 4
729  *
730  *    ============ PASS n + 2 ==============
731  *       CPU 0         |   CPU 1
732  *                     |
733  *    cnt1 timestamps  |   cnt2 timestamps
734  *          6          |         8
735  *          7          |         9
736  *          -          |         10
737  *
738  *      Flush every events below timestamp 7
739  *      etc...
740  */
741 static int process_finished_round(struct perf_tool *tool,
742 				  union perf_event *event __used,
743 				  struct perf_session *session)
744 {
745 	flush_sample_queue(session, tool);
746 	session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
747 
748 	return 0;
749 }
750 
751 /* The queue is ordered by time */
752 static void __queue_event(struct sample_queue *new, struct perf_session *s)
753 {
754 	struct ordered_samples *os = &s->ordered_samples;
755 	struct sample_queue *sample = os->last_sample;
756 	u64 timestamp = new->timestamp;
757 	struct list_head *p;
758 
759 	++os->nr_samples;
760 	os->last_sample = new;
761 
762 	if (!sample) {
763 		list_add(&new->list, &os->samples);
764 		os->max_timestamp = timestamp;
765 		return;
766 	}
767 
768 	/*
769 	 * last_sample might point to some random place in the list as it's
770 	 * the last queued event. We expect that the new event is close to
771 	 * this.
772 	 */
773 	if (sample->timestamp <= timestamp) {
774 		while (sample->timestamp <= timestamp) {
775 			p = sample->list.next;
776 			if (p == &os->samples) {
777 				list_add_tail(&new->list, &os->samples);
778 				os->max_timestamp = timestamp;
779 				return;
780 			}
781 			sample = list_entry(p, struct sample_queue, list);
782 		}
783 		list_add_tail(&new->list, &sample->list);
784 	} else {
785 		while (sample->timestamp > timestamp) {
786 			p = sample->list.prev;
787 			if (p == &os->samples) {
788 				list_add(&new->list, &os->samples);
789 				return;
790 			}
791 			sample = list_entry(p, struct sample_queue, list);
792 		}
793 		list_add(&new->list, &sample->list);
794 	}
795 }
796 
797 #define MAX_SAMPLE_BUFFER	(64 * 1024 / sizeof(struct sample_queue))
798 
799 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
800 				    struct perf_sample *sample, u64 file_offset)
801 {
802 	struct ordered_samples *os = &s->ordered_samples;
803 	struct list_head *sc = &os->sample_cache;
804 	u64 timestamp = sample->time;
805 	struct sample_queue *new;
806 
807 	if (!timestamp || timestamp == ~0ULL)
808 		return -ETIME;
809 
810 	if (timestamp < s->ordered_samples.last_flush) {
811 		printf("Warning: Timestamp below last timeslice flush\n");
812 		return -EINVAL;
813 	}
814 
815 	if (!list_empty(sc)) {
816 		new = list_entry(sc->next, struct sample_queue, list);
817 		list_del(&new->list);
818 	} else if (os->sample_buffer) {
819 		new = os->sample_buffer + os->sample_buffer_idx;
820 		if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
821 			os->sample_buffer = NULL;
822 	} else {
823 		os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
824 		if (!os->sample_buffer)
825 			return -ENOMEM;
826 		list_add(&os->sample_buffer->list, &os->to_free);
827 		os->sample_buffer_idx = 2;
828 		new = os->sample_buffer + 1;
829 	}
830 
831 	new->timestamp = timestamp;
832 	new->file_offset = file_offset;
833 	new->event = event;
834 
835 	__queue_event(new, s);
836 
837 	return 0;
838 }
839 
840 static void callchain__printf(struct perf_sample *sample)
841 {
842 	unsigned int i;
843 
844 	printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
845 
846 	for (i = 0; i < sample->callchain->nr; i++)
847 		printf("..... %2d: %016" PRIx64 "\n",
848 		       i, sample->callchain->ips[i]);
849 }
850 
851 static void branch_stack__printf(struct perf_sample *sample)
852 {
853 	uint64_t i;
854 
855 	printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
856 
857 	for (i = 0; i < sample->branch_stack->nr; i++)
858 		printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
859 			i, sample->branch_stack->entries[i].from,
860 			sample->branch_stack->entries[i].to);
861 }
862 
863 static void perf_session__print_tstamp(struct perf_session *session,
864 				       union perf_event *event,
865 				       struct perf_sample *sample)
866 {
867 	u64 sample_type = perf_evlist__sample_type(session->evlist);
868 
869 	if (event->header.type != PERF_RECORD_SAMPLE &&
870 	    !perf_evlist__sample_id_all(session->evlist)) {
871 		fputs("-1 -1 ", stdout);
872 		return;
873 	}
874 
875 	if ((sample_type & PERF_SAMPLE_CPU))
876 		printf("%u ", sample->cpu);
877 
878 	if (sample_type & PERF_SAMPLE_TIME)
879 		printf("%" PRIu64 " ", sample->time);
880 }
881 
882 static void dump_event(struct perf_session *session, union perf_event *event,
883 		       u64 file_offset, struct perf_sample *sample)
884 {
885 	if (!dump_trace)
886 		return;
887 
888 	printf("\n%#" PRIx64 " [%#x]: event: %d\n",
889 	       file_offset, event->header.size, event->header.type);
890 
891 	trace_event(event);
892 
893 	if (sample)
894 		perf_session__print_tstamp(session, event, sample);
895 
896 	printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
897 	       event->header.size, perf_event__name(event->header.type));
898 }
899 
900 static void dump_sample(struct perf_session *session, union perf_event *event,
901 			struct perf_sample *sample)
902 {
903 	u64 sample_type;
904 
905 	if (!dump_trace)
906 		return;
907 
908 	printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
909 	       event->header.misc, sample->pid, sample->tid, sample->ip,
910 	       sample->period, sample->addr);
911 
912 	sample_type = perf_evlist__sample_type(session->evlist);
913 
914 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
915 		callchain__printf(sample);
916 
917 	if (sample_type & PERF_SAMPLE_BRANCH_STACK)
918 		branch_stack__printf(sample);
919 }
920 
921 static struct machine *
922 	perf_session__find_machine_for_cpumode(struct perf_session *session,
923 					       union perf_event *event)
924 {
925 	const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
926 
927 	if (perf_guest &&
928 	    ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
929 	     (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
930 		u32 pid;
931 
932 		if (event->header.type == PERF_RECORD_MMAP)
933 			pid = event->mmap.pid;
934 		else
935 			pid = event->ip.pid;
936 
937 		return perf_session__findnew_machine(session, pid);
938 	}
939 
940 	return perf_session__find_host_machine(session);
941 }
942 
943 static int perf_session_deliver_event(struct perf_session *session,
944 				      union perf_event *event,
945 				      struct perf_sample *sample,
946 				      struct perf_tool *tool,
947 				      u64 file_offset)
948 {
949 	struct perf_evsel *evsel;
950 	struct machine *machine;
951 
952 	dump_event(session, event, file_offset, sample);
953 
954 	evsel = perf_evlist__id2evsel(session->evlist, sample->id);
955 	if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
956 		/*
957 		 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
958 		 * because the tools right now may apply filters, discarding
959 		 * some of the samples. For consistency, in the future we
960 		 * should have something like nr_filtered_samples and remove
961 		 * the sample->period from total_sample_period, etc, KISS for
962 		 * now tho.
963 		 *
964 		 * Also testing against NULL allows us to handle files without
965 		 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
966 		 * future probably it'll be a good idea to restrict event
967 		 * processing via perf_session to files with both set.
968 		 */
969 		hists__inc_nr_events(&evsel->hists, event->header.type);
970 	}
971 
972 	machine = perf_session__find_machine_for_cpumode(session, event);
973 
974 	switch (event->header.type) {
975 	case PERF_RECORD_SAMPLE:
976 		dump_sample(session, event, sample);
977 		if (evsel == NULL) {
978 			++session->hists.stats.nr_unknown_id;
979 			return 0;
980 		}
981 		if (machine == NULL) {
982 			++session->hists.stats.nr_unprocessable_samples;
983 			return 0;
984 		}
985 		return tool->sample(tool, event, sample, evsel, machine);
986 	case PERF_RECORD_MMAP:
987 		return tool->mmap(tool, event, sample, machine);
988 	case PERF_RECORD_COMM:
989 		return tool->comm(tool, event, sample, machine);
990 	case PERF_RECORD_FORK:
991 		return tool->fork(tool, event, sample, machine);
992 	case PERF_RECORD_EXIT:
993 		return tool->exit(tool, event, sample, machine);
994 	case PERF_RECORD_LOST:
995 		if (tool->lost == perf_event__process_lost)
996 			session->hists.stats.total_lost += event->lost.lost;
997 		return tool->lost(tool, event, sample, machine);
998 	case PERF_RECORD_READ:
999 		return tool->read(tool, event, sample, evsel, machine);
1000 	case PERF_RECORD_THROTTLE:
1001 		return tool->throttle(tool, event, sample, machine);
1002 	case PERF_RECORD_UNTHROTTLE:
1003 		return tool->unthrottle(tool, event, sample, machine);
1004 	default:
1005 		++session->hists.stats.nr_unknown_events;
1006 		return -1;
1007 	}
1008 }
1009 
1010 static int perf_session__preprocess_sample(struct perf_session *session,
1011 					   union perf_event *event, struct perf_sample *sample)
1012 {
1013 	if (event->header.type != PERF_RECORD_SAMPLE ||
1014 	    !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
1015 		return 0;
1016 
1017 	if (!ip_callchain__valid(sample->callchain, event)) {
1018 		pr_debug("call-chain problem with event, skipping it.\n");
1019 		++session->hists.stats.nr_invalid_chains;
1020 		session->hists.stats.total_invalid_chains += sample->period;
1021 		return -EINVAL;
1022 	}
1023 	return 0;
1024 }
1025 
1026 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1027 					    struct perf_tool *tool, u64 file_offset)
1028 {
1029 	int err;
1030 
1031 	dump_event(session, event, file_offset, NULL);
1032 
1033 	/* These events are processed right away */
1034 	switch (event->header.type) {
1035 	case PERF_RECORD_HEADER_ATTR:
1036 		err = tool->attr(event, &session->evlist);
1037 		if (err == 0)
1038 			perf_session__set_id_hdr_size(session);
1039 		return err;
1040 	case PERF_RECORD_HEADER_EVENT_TYPE:
1041 		return tool->event_type(tool, event);
1042 	case PERF_RECORD_HEADER_TRACING_DATA:
1043 		/* setup for reading amidst mmap */
1044 		lseek(session->fd, file_offset, SEEK_SET);
1045 		return tool->tracing_data(event, session);
1046 	case PERF_RECORD_HEADER_BUILD_ID:
1047 		return tool->build_id(tool, event, session);
1048 	case PERF_RECORD_FINISHED_ROUND:
1049 		return tool->finished_round(tool, event, session);
1050 	default:
1051 		return -EINVAL;
1052 	}
1053 }
1054 
1055 static void event_swap(union perf_event *event, bool sample_id_all)
1056 {
1057 	perf_event__swap_op swap;
1058 
1059 	swap = perf_event__swap_ops[event->header.type];
1060 	if (swap)
1061 		swap(event, sample_id_all);
1062 }
1063 
1064 static int perf_session__process_event(struct perf_session *session,
1065 				       union perf_event *event,
1066 				       struct perf_tool *tool,
1067 				       u64 file_offset)
1068 {
1069 	struct perf_sample sample;
1070 	int ret;
1071 
1072 	if (session->header.needs_swap)
1073 		event_swap(event, perf_evlist__sample_id_all(session->evlist));
1074 
1075 	if (event->header.type >= PERF_RECORD_HEADER_MAX)
1076 		return -EINVAL;
1077 
1078 	hists__inc_nr_events(&session->hists, event->header.type);
1079 
1080 	if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1081 		return perf_session__process_user_event(session, event, tool, file_offset);
1082 
1083 	/*
1084 	 * For all kernel events we get the sample data
1085 	 */
1086 	ret = perf_evlist__parse_sample(session->evlist, event, &sample,
1087 					session->header.needs_swap);
1088 	if (ret)
1089 		return ret;
1090 
1091 	/* Preprocess sample records - precheck callchains */
1092 	if (perf_session__preprocess_sample(session, event, &sample))
1093 		return 0;
1094 
1095 	if (tool->ordered_samples) {
1096 		ret = perf_session_queue_event(session, event, &sample,
1097 					       file_offset);
1098 		if (ret != -ETIME)
1099 			return ret;
1100 	}
1101 
1102 	return perf_session_deliver_event(session, event, &sample, tool,
1103 					  file_offset);
1104 }
1105 
1106 void perf_event_header__bswap(struct perf_event_header *self)
1107 {
1108 	self->type = bswap_32(self->type);
1109 	self->misc = bswap_16(self->misc);
1110 	self->size = bswap_16(self->size);
1111 }
1112 
1113 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1114 {
1115 	return machine__findnew_thread(&session->host_machine, pid);
1116 }
1117 
1118 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1119 {
1120 	struct thread *thread = perf_session__findnew(self, 0);
1121 
1122 	if (thread == NULL || thread__set_comm(thread, "swapper")) {
1123 		pr_err("problem inserting idle task.\n");
1124 		thread = NULL;
1125 	}
1126 
1127 	return thread;
1128 }
1129 
1130 static void perf_session__warn_about_errors(const struct perf_session *session,
1131 					    const struct perf_tool *tool)
1132 {
1133 	if (tool->lost == perf_event__process_lost &&
1134 	    session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1135 		ui__warning("Processed %d events and lost %d chunks!\n\n"
1136 			    "Check IO/CPU overload!\n\n",
1137 			    session->hists.stats.nr_events[0],
1138 			    session->hists.stats.nr_events[PERF_RECORD_LOST]);
1139 	}
1140 
1141 	if (session->hists.stats.nr_unknown_events != 0) {
1142 		ui__warning("Found %u unknown events!\n\n"
1143 			    "Is this an older tool processing a perf.data "
1144 			    "file generated by a more recent tool?\n\n"
1145 			    "If that is not the case, consider "
1146 			    "reporting to linux-kernel@vger.kernel.org.\n\n",
1147 			    session->hists.stats.nr_unknown_events);
1148 	}
1149 
1150 	if (session->hists.stats.nr_unknown_id != 0) {
1151 		ui__warning("%u samples with id not present in the header\n",
1152 			    session->hists.stats.nr_unknown_id);
1153 	}
1154 
1155  	if (session->hists.stats.nr_invalid_chains != 0) {
1156  		ui__warning("Found invalid callchains!\n\n"
1157  			    "%u out of %u events were discarded for this reason.\n\n"
1158  			    "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1159  			    session->hists.stats.nr_invalid_chains,
1160  			    session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1161  	}
1162 
1163 	if (session->hists.stats.nr_unprocessable_samples != 0) {
1164 		ui__warning("%u unprocessable samples recorded.\n"
1165 			    "Do you have a KVM guest running and not using 'perf kvm'?\n",
1166 			    session->hists.stats.nr_unprocessable_samples);
1167 	}
1168 }
1169 
1170 #define session_done()	(*(volatile int *)(&session_done))
1171 volatile int session_done;
1172 
1173 static int __perf_session__process_pipe_events(struct perf_session *self,
1174 					       struct perf_tool *tool)
1175 {
1176 	union perf_event *event;
1177 	uint32_t size, cur_size = 0;
1178 	void *buf = NULL;
1179 	int skip = 0;
1180 	u64 head;
1181 	int err;
1182 	void *p;
1183 
1184 	perf_tool__fill_defaults(tool);
1185 
1186 	head = 0;
1187 	cur_size = sizeof(union perf_event);
1188 
1189 	buf = malloc(cur_size);
1190 	if (!buf)
1191 		return -errno;
1192 more:
1193 	event = buf;
1194 	err = readn(self->fd, event, sizeof(struct perf_event_header));
1195 	if (err <= 0) {
1196 		if (err == 0)
1197 			goto done;
1198 
1199 		pr_err("failed to read event header\n");
1200 		goto out_err;
1201 	}
1202 
1203 	if (self->header.needs_swap)
1204 		perf_event_header__bswap(&event->header);
1205 
1206 	size = event->header.size;
1207 	if (size == 0)
1208 		size = 8;
1209 
1210 	if (size > cur_size) {
1211 		void *new = realloc(buf, size);
1212 		if (!new) {
1213 			pr_err("failed to allocate memory to read event\n");
1214 			goto out_err;
1215 		}
1216 		buf = new;
1217 		cur_size = size;
1218 		event = buf;
1219 	}
1220 	p = event;
1221 	p += sizeof(struct perf_event_header);
1222 
1223 	if (size - sizeof(struct perf_event_header)) {
1224 		err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1225 		if (err <= 0) {
1226 			if (err == 0) {
1227 				pr_err("unexpected end of event stream\n");
1228 				goto done;
1229 			}
1230 
1231 			pr_err("failed to read event data\n");
1232 			goto out_err;
1233 		}
1234 	}
1235 
1236 	if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1237 		pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1238 		       head, event->header.size, event->header.type);
1239 		err = -EINVAL;
1240 		goto out_err;
1241 	}
1242 
1243 	head += size;
1244 
1245 	if (skip > 0)
1246 		head += skip;
1247 
1248 	if (!session_done())
1249 		goto more;
1250 done:
1251 	err = 0;
1252 out_err:
1253 	free(buf);
1254 	perf_session__warn_about_errors(self, tool);
1255 	perf_session_free_sample_buffers(self);
1256 	return err;
1257 }
1258 
1259 static union perf_event *
1260 fetch_mmaped_event(struct perf_session *session,
1261 		   u64 head, size_t mmap_size, char *buf)
1262 {
1263 	union perf_event *event;
1264 
1265 	/*
1266 	 * Ensure we have enough space remaining to read
1267 	 * the size of the event in the headers.
1268 	 */
1269 	if (head + sizeof(event->header) > mmap_size)
1270 		return NULL;
1271 
1272 	event = (union perf_event *)(buf + head);
1273 
1274 	if (session->header.needs_swap)
1275 		perf_event_header__bswap(&event->header);
1276 
1277 	if (head + event->header.size > mmap_size)
1278 		return NULL;
1279 
1280 	return event;
1281 }
1282 
1283 int __perf_session__process_events(struct perf_session *session,
1284 				   u64 data_offset, u64 data_size,
1285 				   u64 file_size, struct perf_tool *tool)
1286 {
1287 	u64 head, page_offset, file_offset, file_pos, progress_next;
1288 	int err, mmap_prot, mmap_flags, map_idx = 0;
1289 	size_t	page_size, mmap_size;
1290 	char *buf, *mmaps[8];
1291 	union perf_event *event;
1292 	uint32_t size;
1293 
1294 	perf_tool__fill_defaults(tool);
1295 
1296 	page_size = sysconf(_SC_PAGESIZE);
1297 
1298 	page_offset = page_size * (data_offset / page_size);
1299 	file_offset = page_offset;
1300 	head = data_offset - page_offset;
1301 
1302 	if (data_offset + data_size < file_size)
1303 		file_size = data_offset + data_size;
1304 
1305 	progress_next = file_size / 16;
1306 
1307 	mmap_size = session->mmap_window;
1308 	if (mmap_size > file_size)
1309 		mmap_size = file_size;
1310 
1311 	memset(mmaps, 0, sizeof(mmaps));
1312 
1313 	mmap_prot  = PROT_READ;
1314 	mmap_flags = MAP_SHARED;
1315 
1316 	if (session->header.needs_swap) {
1317 		mmap_prot  |= PROT_WRITE;
1318 		mmap_flags = MAP_PRIVATE;
1319 	}
1320 remap:
1321 	buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1322 		   file_offset);
1323 	if (buf == MAP_FAILED) {
1324 		pr_err("failed to mmap file\n");
1325 		err = -errno;
1326 		goto out_err;
1327 	}
1328 	mmaps[map_idx] = buf;
1329 	map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1330 	file_pos = file_offset + head;
1331 
1332 more:
1333 	event = fetch_mmaped_event(session, head, mmap_size, buf);
1334 	if (!event) {
1335 		if (mmaps[map_idx]) {
1336 			munmap(mmaps[map_idx], mmap_size);
1337 			mmaps[map_idx] = NULL;
1338 		}
1339 
1340 		page_offset = page_size * (head / page_size);
1341 		file_offset += page_offset;
1342 		head -= page_offset;
1343 		goto remap;
1344 	}
1345 
1346 	size = event->header.size;
1347 
1348 	if (size == 0 ||
1349 	    perf_session__process_event(session, event, tool, file_pos) < 0) {
1350 		pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1351 		       file_offset + head, event->header.size,
1352 		       event->header.type);
1353 		err = -EINVAL;
1354 		goto out_err;
1355 	}
1356 
1357 	head += size;
1358 	file_pos += size;
1359 
1360 	if (file_pos >= progress_next) {
1361 		progress_next += file_size / 16;
1362 		ui_progress__update(file_pos, file_size,
1363 				    "Processing events...");
1364 	}
1365 
1366 	if (file_pos < file_size)
1367 		goto more;
1368 
1369 	err = 0;
1370 	/* do the final flush for ordered samples */
1371 	session->ordered_samples.next_flush = ULLONG_MAX;
1372 	flush_sample_queue(session, tool);
1373 out_err:
1374 	perf_session__warn_about_errors(session, tool);
1375 	perf_session_free_sample_buffers(session);
1376 	return err;
1377 }
1378 
1379 int perf_session__process_events(struct perf_session *self,
1380 				 struct perf_tool *tool)
1381 {
1382 	int err;
1383 
1384 	if (perf_session__register_idle_thread(self) == NULL)
1385 		return -ENOMEM;
1386 
1387 	if (!self->fd_pipe)
1388 		err = __perf_session__process_events(self,
1389 						     self->header.data_offset,
1390 						     self->header.data_size,
1391 						     self->size, tool);
1392 	else
1393 		err = __perf_session__process_pipe_events(self, tool);
1394 
1395 	return err;
1396 }
1397 
1398 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1399 {
1400 	if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1401 		pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1402 		return false;
1403 	}
1404 
1405 	return true;
1406 }
1407 
1408 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1409 				     const char *symbol_name, u64 addr)
1410 {
1411 	char *bracket;
1412 	enum map_type i;
1413 	struct ref_reloc_sym *ref;
1414 
1415 	ref = zalloc(sizeof(struct ref_reloc_sym));
1416 	if (ref == NULL)
1417 		return -ENOMEM;
1418 
1419 	ref->name = strdup(symbol_name);
1420 	if (ref->name == NULL) {
1421 		free(ref);
1422 		return -ENOMEM;
1423 	}
1424 
1425 	bracket = strchr(ref->name, ']');
1426 	if (bracket)
1427 		*bracket = '\0';
1428 
1429 	ref->addr = addr;
1430 
1431 	for (i = 0; i < MAP__NR_TYPES; ++i) {
1432 		struct kmap *kmap = map__kmap(maps[i]);
1433 		kmap->ref_reloc_sym = ref;
1434 	}
1435 
1436 	return 0;
1437 }
1438 
1439 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1440 {
1441 	return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1442 	       __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1443 	       machines__fprintf_dsos(&self->machines, fp);
1444 }
1445 
1446 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1447 					  bool with_hits)
1448 {
1449 	size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1450 	return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1451 }
1452 
1453 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1454 {
1455 	struct perf_evsel *pos;
1456 	size_t ret = fprintf(fp, "Aggregated stats:\n");
1457 
1458 	ret += hists__fprintf_nr_events(&session->hists, fp);
1459 
1460 	list_for_each_entry(pos, &session->evlist->entries, node) {
1461 		ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1462 		ret += hists__fprintf_nr_events(&pos->hists, fp);
1463 	}
1464 
1465 	return ret;
1466 }
1467 
1468 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1469 {
1470 	/*
1471 	 * FIXME: Here we have to actually print all the machines in this
1472 	 * session, not just the host...
1473 	 */
1474 	return machine__fprintf(&session->host_machine, fp);
1475 }
1476 
1477 void perf_session__remove_thread(struct perf_session *session,
1478 				 struct thread *th)
1479 {
1480 	/*
1481 	 * FIXME: This one makes no sense, we need to remove the thread from
1482 	 * the machine it belongs to, perf_session can have many machines, so
1483 	 * doing it always on ->host_machine is wrong.  Fix when auditing all
1484 	 * the 'perf kvm' code.
1485 	 */
1486 	machine__remove_thread(&session->host_machine, th);
1487 }
1488 
1489 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1490 					      unsigned int type)
1491 {
1492 	struct perf_evsel *pos;
1493 
1494 	list_for_each_entry(pos, &session->evlist->entries, node) {
1495 		if (pos->attr.type == type)
1496 			return pos;
1497 	}
1498 	return NULL;
1499 }
1500 
1501 void perf_event__print_ip(union perf_event *event, struct perf_sample *sample,
1502 			  struct machine *machine, int print_sym,
1503 			  int print_dso, int print_symoffset)
1504 {
1505 	struct addr_location al;
1506 	struct callchain_cursor_node *node;
1507 
1508 	if (perf_event__preprocess_sample(event, machine, &al, sample,
1509 					  NULL) < 0) {
1510 		error("problem processing %d event, skipping it.\n",
1511 			event->header.type);
1512 		return;
1513 	}
1514 
1515 	if (symbol_conf.use_callchain && sample->callchain) {
1516 
1517 		if (machine__resolve_callchain(machine, al.thread,
1518 						sample->callchain, NULL) != 0) {
1519 			if (verbose)
1520 				error("Failed to resolve callchain. Skipping\n");
1521 			return;
1522 		}
1523 		callchain_cursor_commit(&callchain_cursor);
1524 
1525 		while (1) {
1526 			node = callchain_cursor_current(&callchain_cursor);
1527 			if (!node)
1528 				break;
1529 
1530 			printf("\t%16" PRIx64, node->ip);
1531 			if (print_sym) {
1532 				printf(" ");
1533 				symbol__fprintf_symname(node->sym, stdout);
1534 			}
1535 			if (print_dso) {
1536 				printf(" (");
1537 				map__fprintf_dsoname(node->map, stdout);
1538 				printf(")");
1539 			}
1540 			printf("\n");
1541 
1542 			callchain_cursor_advance(&callchain_cursor);
1543 		}
1544 
1545 	} else {
1546 		printf("%16" PRIx64, sample->ip);
1547 		if (print_sym) {
1548 			printf(" ");
1549 			if (print_symoffset)
1550 				symbol__fprintf_symname_offs(al.sym, &al,
1551 							     stdout);
1552 			else
1553 				symbol__fprintf_symname(al.sym, stdout);
1554 		}
1555 
1556 		if (print_dso) {
1557 			printf(" (");
1558 			map__fprintf_dsoname(al.map, stdout);
1559 			printf(")");
1560 		}
1561 	}
1562 }
1563 
1564 int perf_session__cpu_bitmap(struct perf_session *session,
1565 			     const char *cpu_list, unsigned long *cpu_bitmap)
1566 {
1567 	int i;
1568 	struct cpu_map *map;
1569 
1570 	for (i = 0; i < PERF_TYPE_MAX; ++i) {
1571 		struct perf_evsel *evsel;
1572 
1573 		evsel = perf_session__find_first_evtype(session, i);
1574 		if (!evsel)
1575 			continue;
1576 
1577 		if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1578 			pr_err("File does not contain CPU events. "
1579 			       "Remove -c option to proceed.\n");
1580 			return -1;
1581 		}
1582 	}
1583 
1584 	map = cpu_map__new(cpu_list);
1585 	if (map == NULL) {
1586 		pr_err("Invalid cpu_list\n");
1587 		return -1;
1588 	}
1589 
1590 	for (i = 0; i < map->nr; i++) {
1591 		int cpu = map->map[i];
1592 
1593 		if (cpu >= MAX_NR_CPUS) {
1594 			pr_err("Requested CPU %d too large. "
1595 			       "Consider raising MAX_NR_CPUS\n", cpu);
1596 			return -1;
1597 		}
1598 
1599 		set_bit(cpu, cpu_bitmap);
1600 	}
1601 
1602 	return 0;
1603 }
1604 
1605 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1606 				bool full)
1607 {
1608 	struct stat st;
1609 	int ret;
1610 
1611 	if (session == NULL || fp == NULL)
1612 		return;
1613 
1614 	ret = fstat(session->fd, &st);
1615 	if (ret == -1)
1616 		return;
1617 
1618 	fprintf(fp, "# ========\n");
1619 	fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1620 	perf_header__fprintf_info(session, fp, full);
1621 	fprintf(fp, "# ========\n#\n");
1622 }
1623 
1624 
1625 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1626 					     const struct perf_evsel_str_handler *assocs,
1627 					     size_t nr_assocs)
1628 {
1629 	struct perf_evlist *evlist = session->evlist;
1630 	struct event_format *format;
1631 	struct perf_evsel *evsel;
1632 	char *tracepoint, *name;
1633 	size_t i;
1634 	int err;
1635 
1636 	for (i = 0; i < nr_assocs; i++) {
1637 		err = -ENOMEM;
1638 		tracepoint = strdup(assocs[i].name);
1639 		if (tracepoint == NULL)
1640 			goto out;
1641 
1642 		err = -ENOENT;
1643 		name = strchr(tracepoint, ':');
1644 		if (name == NULL)
1645 			goto out_free;
1646 
1647 		*name++ = '\0';
1648 		format = pevent_find_event_by_name(session->pevent,
1649 						   tracepoint, name);
1650 		if (format == NULL) {
1651 			/*
1652 			 * Adding a handler for an event not in the session,
1653 			 * just ignore it.
1654 			 */
1655 			goto next;
1656 		}
1657 
1658 		evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1659 		if (evsel == NULL)
1660 			goto next;
1661 
1662 		err = -EEXIST;
1663 		if (evsel->handler.func != NULL)
1664 			goto out_free;
1665 		evsel->handler.func = assocs[i].handler;
1666 next:
1667 		free(tracepoint);
1668 	}
1669 
1670 	err = 0;
1671 out:
1672 	return err;
1673 
1674 out_free:
1675 	free(tracepoint);
1676 	goto out;
1677 }
1678