xref: /openbmc/linux/tools/perf/builtin-kwork.c (revision 278d3ba6)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * builtin-kwork.c
4  *
5  * Copyright (c) 2022  Huawei Inc,  Yang Jihong <yangjihong1@huawei.com>
6  */
7 
8 #include "builtin.h"
9 
10 #include "util/data.h"
11 #include "util/kwork.h"
12 #include "util/debug.h"
13 #include "util/symbol.h"
14 #include "util/thread.h"
15 #include "util/string2.h"
16 #include "util/callchain.h"
17 #include "util/evsel_fprintf.h"
18 
19 #include <subcmd/pager.h>
20 #include <subcmd/parse-options.h>
21 
22 #include <errno.h>
23 #include <inttypes.h>
24 #include <linux/err.h>
25 #include <linux/time64.h>
26 #include <linux/zalloc.h>
27 
28 /*
29  * report header elements width
30  */
31 #define PRINT_CPU_WIDTH 4
32 #define PRINT_COUNT_WIDTH 9
33 #define PRINT_RUNTIME_WIDTH 10
34 #define PRINT_LATENCY_WIDTH 10
35 #define PRINT_TIMESTAMP_WIDTH 17
36 #define PRINT_KWORK_NAME_WIDTH 30
37 #define RPINT_DECIMAL_WIDTH 3
38 #define PRINT_BRACKETPAIR_WIDTH 2
39 #define PRINT_TIME_UNIT_SEC_WIDTH 2
40 #define PRINT_TIME_UNIT_MESC_WIDTH 3
41 #define PRINT_RUNTIME_HEADER_WIDTH (PRINT_RUNTIME_WIDTH + PRINT_TIME_UNIT_MESC_WIDTH)
42 #define PRINT_LATENCY_HEADER_WIDTH (PRINT_LATENCY_WIDTH + PRINT_TIME_UNIT_MESC_WIDTH)
43 #define PRINT_TIMEHIST_CPU_WIDTH (PRINT_CPU_WIDTH + PRINT_BRACKETPAIR_WIDTH)
44 #define PRINT_TIMESTAMP_HEADER_WIDTH (PRINT_TIMESTAMP_WIDTH + PRINT_TIME_UNIT_SEC_WIDTH)
45 
46 struct sort_dimension {
47 	const char      *name;
48 	int             (*cmp)(struct kwork_work *l, struct kwork_work *r);
49 	struct          list_head list;
50 };
51 
52 static int id_cmp(struct kwork_work *l, struct kwork_work *r)
53 {
54 	if (l->cpu > r->cpu)
55 		return 1;
56 	if (l->cpu < r->cpu)
57 		return -1;
58 
59 	if (l->id > r->id)
60 		return 1;
61 	if (l->id < r->id)
62 		return -1;
63 
64 	return 0;
65 }
66 
67 static int count_cmp(struct kwork_work *l, struct kwork_work *r)
68 {
69 	if (l->nr_atoms > r->nr_atoms)
70 		return 1;
71 	if (l->nr_atoms < r->nr_atoms)
72 		return -1;
73 
74 	return 0;
75 }
76 
77 static int runtime_cmp(struct kwork_work *l, struct kwork_work *r)
78 {
79 	if (l->total_runtime > r->total_runtime)
80 		return 1;
81 	if (l->total_runtime < r->total_runtime)
82 		return -1;
83 
84 	return 0;
85 }
86 
87 static int max_runtime_cmp(struct kwork_work *l, struct kwork_work *r)
88 {
89 	if (l->max_runtime > r->max_runtime)
90 		return 1;
91 	if (l->max_runtime < r->max_runtime)
92 		return -1;
93 
94 	return 0;
95 }
96 
97 static int avg_latency_cmp(struct kwork_work *l, struct kwork_work *r)
98 {
99 	u64 avgl, avgr;
100 
101 	if (!r->nr_atoms)
102 		return 1;
103 	if (!l->nr_atoms)
104 		return -1;
105 
106 	avgl = l->total_latency / l->nr_atoms;
107 	avgr = r->total_latency / r->nr_atoms;
108 
109 	if (avgl > avgr)
110 		return 1;
111 	if (avgl < avgr)
112 		return -1;
113 
114 	return 0;
115 }
116 
117 static int max_latency_cmp(struct kwork_work *l, struct kwork_work *r)
118 {
119 	if (l->max_latency > r->max_latency)
120 		return 1;
121 	if (l->max_latency < r->max_latency)
122 		return -1;
123 
124 	return 0;
125 }
126 
127 static int sort_dimension__add(struct perf_kwork *kwork __maybe_unused,
128 			       const char *tok, struct list_head *list)
129 {
130 	size_t i;
131 	static struct sort_dimension max_sort_dimension = {
132 		.name = "max",
133 		.cmp  = max_runtime_cmp,
134 	};
135 	static struct sort_dimension id_sort_dimension = {
136 		.name = "id",
137 		.cmp  = id_cmp,
138 	};
139 	static struct sort_dimension runtime_sort_dimension = {
140 		.name = "runtime",
141 		.cmp  = runtime_cmp,
142 	};
143 	static struct sort_dimension count_sort_dimension = {
144 		.name = "count",
145 		.cmp  = count_cmp,
146 	};
147 	static struct sort_dimension avg_sort_dimension = {
148 		.name = "avg",
149 		.cmp  = avg_latency_cmp,
150 	};
151 	struct sort_dimension *available_sorts[] = {
152 		&id_sort_dimension,
153 		&max_sort_dimension,
154 		&count_sort_dimension,
155 		&runtime_sort_dimension,
156 		&avg_sort_dimension,
157 	};
158 
159 	if (kwork->report == KWORK_REPORT_LATENCY)
160 		max_sort_dimension.cmp = max_latency_cmp;
161 
162 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
163 		if (!strcmp(available_sorts[i]->name, tok)) {
164 			list_add_tail(&available_sorts[i]->list, list);
165 			return 0;
166 		}
167 	}
168 
169 	return -1;
170 }
171 
172 static void setup_sorting(struct perf_kwork *kwork,
173 			  const struct option *options,
174 			  const char * const usage_msg[])
175 {
176 	char *tmp, *tok, *str = strdup(kwork->sort_order);
177 
178 	for (tok = strtok_r(str, ", ", &tmp);
179 	     tok; tok = strtok_r(NULL, ", ", &tmp)) {
180 		if (sort_dimension__add(kwork, tok, &kwork->sort_list) < 0)
181 			usage_with_options_msg(usage_msg, options,
182 					       "Unknown --sort key: `%s'", tok);
183 	}
184 
185 	pr_debug("Sort order: %s\n", kwork->sort_order);
186 	free(str);
187 }
188 
189 static struct kwork_atom *atom_new(struct perf_kwork *kwork,
190 				   struct perf_sample *sample)
191 {
192 	unsigned long i;
193 	struct kwork_atom_page *page;
194 	struct kwork_atom *atom = NULL;
195 
196 	list_for_each_entry(page, &kwork->atom_page_list, list) {
197 		if (!bitmap_full(page->bitmap, NR_ATOM_PER_PAGE)) {
198 			i = find_first_zero_bit(page->bitmap, NR_ATOM_PER_PAGE);
199 			BUG_ON(i >= NR_ATOM_PER_PAGE);
200 			atom = &page->atoms[i];
201 			goto found_atom;
202 		}
203 	}
204 
205 	/*
206 	 * new page
207 	 */
208 	page = zalloc(sizeof(*page));
209 	if (page == NULL) {
210 		pr_err("Failed to zalloc kwork atom page\n");
211 		return NULL;
212 	}
213 
214 	i = 0;
215 	atom = &page->atoms[0];
216 	list_add_tail(&page->list, &kwork->atom_page_list);
217 
218 found_atom:
219 	set_bit(i, page->bitmap);
220 	atom->time = sample->time;
221 	atom->prev = NULL;
222 	atom->page_addr = page;
223 	atom->bit_inpage = i;
224 	return atom;
225 }
226 
227 static void atom_free(struct kwork_atom *atom)
228 {
229 	if (atom->prev != NULL)
230 		atom_free(atom->prev);
231 
232 	clear_bit(atom->bit_inpage,
233 		  ((struct kwork_atom_page *)atom->page_addr)->bitmap);
234 }
235 
236 static void atom_del(struct kwork_atom *atom)
237 {
238 	list_del(&atom->list);
239 	atom_free(atom);
240 }
241 
242 static int work_cmp(struct list_head *list,
243 		    struct kwork_work *l, struct kwork_work *r)
244 {
245 	int ret = 0;
246 	struct sort_dimension *sort;
247 
248 	BUG_ON(list_empty(list));
249 
250 	list_for_each_entry(sort, list, list) {
251 		ret = sort->cmp(l, r);
252 		if (ret)
253 			return ret;
254 	}
255 
256 	return ret;
257 }
258 
259 static struct kwork_work *work_search(struct rb_root_cached *root,
260 				      struct kwork_work *key,
261 				      struct list_head *sort_list)
262 {
263 	int cmp;
264 	struct kwork_work *work;
265 	struct rb_node *node = root->rb_root.rb_node;
266 
267 	while (node) {
268 		work = container_of(node, struct kwork_work, node);
269 		cmp = work_cmp(sort_list, key, work);
270 		if (cmp > 0)
271 			node = node->rb_left;
272 		else if (cmp < 0)
273 			node = node->rb_right;
274 		else {
275 			if (work->name == NULL)
276 				work->name = key->name;
277 			return work;
278 		}
279 	}
280 	return NULL;
281 }
282 
283 static void work_insert(struct rb_root_cached *root,
284 			struct kwork_work *key, struct list_head *sort_list)
285 {
286 	int cmp;
287 	bool leftmost = true;
288 	struct kwork_work *cur;
289 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
290 
291 	while (*new) {
292 		cur = container_of(*new, struct kwork_work, node);
293 		parent = *new;
294 		cmp = work_cmp(sort_list, key, cur);
295 
296 		if (cmp > 0)
297 			new = &((*new)->rb_left);
298 		else {
299 			new = &((*new)->rb_right);
300 			leftmost = false;
301 		}
302 	}
303 
304 	rb_link_node(&key->node, parent, new);
305 	rb_insert_color_cached(&key->node, root, leftmost);
306 }
307 
308 static struct kwork_work *work_new(struct kwork_work *key)
309 {
310 	int i;
311 	struct kwork_work *work = zalloc(sizeof(*work));
312 
313 	if (work == NULL) {
314 		pr_err("Failed to zalloc kwork work\n");
315 		return NULL;
316 	}
317 
318 	for (i = 0; i < KWORK_TRACE_MAX; i++)
319 		INIT_LIST_HEAD(&work->atom_list[i]);
320 
321 	work->id = key->id;
322 	work->cpu = key->cpu;
323 	work->name = key->name;
324 	work->class = key->class;
325 	return work;
326 }
327 
328 static struct kwork_work *work_findnew(struct rb_root_cached *root,
329 				       struct kwork_work *key,
330 				       struct list_head *sort_list)
331 {
332 	struct kwork_work *work = work_search(root, key, sort_list);
333 
334 	if (work != NULL)
335 		return work;
336 
337 	work = work_new(key);
338 	if (work)
339 		work_insert(root, work, sort_list);
340 
341 	return work;
342 }
343 
344 static void profile_update_timespan(struct perf_kwork *kwork,
345 				    struct perf_sample *sample)
346 {
347 	if (!kwork->summary)
348 		return;
349 
350 	if ((kwork->timestart == 0) || (kwork->timestart > sample->time))
351 		kwork->timestart = sample->time;
352 
353 	if (kwork->timeend < sample->time)
354 		kwork->timeend = sample->time;
355 }
356 
357 static bool profile_event_match(struct perf_kwork *kwork,
358 				struct kwork_work *work,
359 				struct perf_sample *sample)
360 {
361 	int cpu = work->cpu;
362 	u64 time = sample->time;
363 	struct perf_time_interval *ptime = &kwork->ptime;
364 
365 	if ((kwork->cpu_list != NULL) && !test_bit(cpu, kwork->cpu_bitmap))
366 		return false;
367 
368 	if (((ptime->start != 0) && (ptime->start > time)) ||
369 	    ((ptime->end != 0) && (ptime->end < time)))
370 		return false;
371 
372 	if ((kwork->profile_name != NULL) &&
373 	    (work->name != NULL) &&
374 	    (strcmp(work->name, kwork->profile_name) != 0))
375 		return false;
376 
377 	profile_update_timespan(kwork, sample);
378 	return true;
379 }
380 
381 static int work_push_atom(struct perf_kwork *kwork,
382 			  struct kwork_class *class,
383 			  enum kwork_trace_type src_type,
384 			  enum kwork_trace_type dst_type,
385 			  struct evsel *evsel,
386 			  struct perf_sample *sample,
387 			  struct machine *machine,
388 			  struct kwork_work **ret_work)
389 {
390 	struct kwork_atom *atom, *dst_atom;
391 	struct kwork_work *work, key;
392 
393 	BUG_ON(class->work_init == NULL);
394 	class->work_init(class, &key, evsel, sample, machine);
395 
396 	atom = atom_new(kwork, sample);
397 	if (atom == NULL)
398 		return -1;
399 
400 	work = work_findnew(&class->work_root, &key, &kwork->cmp_id);
401 	if (work == NULL) {
402 		free(atom);
403 		return -1;
404 	}
405 
406 	if (!profile_event_match(kwork, work, sample))
407 		return 0;
408 
409 	if (dst_type < KWORK_TRACE_MAX) {
410 		dst_atom = list_last_entry_or_null(&work->atom_list[dst_type],
411 						   struct kwork_atom, list);
412 		if (dst_atom != NULL) {
413 			atom->prev = dst_atom;
414 			list_del(&dst_atom->list);
415 		}
416 	}
417 
418 	if (ret_work != NULL)
419 		*ret_work = work;
420 
421 	list_add_tail(&atom->list, &work->atom_list[src_type]);
422 
423 	return 0;
424 }
425 
426 static struct kwork_atom *work_pop_atom(struct perf_kwork *kwork,
427 					struct kwork_class *class,
428 					enum kwork_trace_type src_type,
429 					enum kwork_trace_type dst_type,
430 					struct evsel *evsel,
431 					struct perf_sample *sample,
432 					struct machine *machine,
433 					struct kwork_work **ret_work)
434 {
435 	struct kwork_atom *atom, *src_atom;
436 	struct kwork_work *work, key;
437 
438 	BUG_ON(class->work_init == NULL);
439 	class->work_init(class, &key, evsel, sample, machine);
440 
441 	work = work_findnew(&class->work_root, &key, &kwork->cmp_id);
442 	if (ret_work != NULL)
443 		*ret_work = work;
444 
445 	if (work == NULL)
446 		return NULL;
447 
448 	if (!profile_event_match(kwork, work, sample))
449 		return NULL;
450 
451 	atom = list_last_entry_or_null(&work->atom_list[dst_type],
452 				       struct kwork_atom, list);
453 	if (atom != NULL)
454 		return atom;
455 
456 	src_atom = atom_new(kwork, sample);
457 	if (src_atom != NULL)
458 		list_add_tail(&src_atom->list, &work->atom_list[src_type]);
459 	else {
460 		if (ret_work != NULL)
461 			*ret_work = NULL;
462 	}
463 
464 	return NULL;
465 }
466 
467 static void report_update_exit_event(struct kwork_work *work,
468 				     struct kwork_atom *atom,
469 				     struct perf_sample *sample)
470 {
471 	u64 delta;
472 	u64 exit_time = sample->time;
473 	u64 entry_time = atom->time;
474 
475 	if ((entry_time != 0) && (exit_time >= entry_time)) {
476 		delta = exit_time - entry_time;
477 		if ((delta > work->max_runtime) ||
478 		    (work->max_runtime == 0)) {
479 			work->max_runtime = delta;
480 			work->max_runtime_start = entry_time;
481 			work->max_runtime_end = exit_time;
482 		}
483 		work->total_runtime += delta;
484 		work->nr_atoms++;
485 	}
486 }
487 
488 static int report_entry_event(struct perf_kwork *kwork,
489 			      struct kwork_class *class,
490 			      struct evsel *evsel,
491 			      struct perf_sample *sample,
492 			      struct machine *machine)
493 {
494 	return work_push_atom(kwork, class, KWORK_TRACE_ENTRY,
495 			      KWORK_TRACE_MAX, evsel, sample,
496 			      machine, NULL);
497 }
498 
499 static int report_exit_event(struct perf_kwork *kwork,
500 			     struct kwork_class *class,
501 			     struct evsel *evsel,
502 			     struct perf_sample *sample,
503 			     struct machine *machine)
504 {
505 	struct kwork_atom *atom = NULL;
506 	struct kwork_work *work = NULL;
507 
508 	atom = work_pop_atom(kwork, class, KWORK_TRACE_EXIT,
509 			     KWORK_TRACE_ENTRY, evsel, sample,
510 			     machine, &work);
511 	if (work == NULL)
512 		return -1;
513 
514 	if (atom != NULL) {
515 		report_update_exit_event(work, atom, sample);
516 		atom_del(atom);
517 	}
518 
519 	return 0;
520 }
521 
522 static void latency_update_entry_event(struct kwork_work *work,
523 				       struct kwork_atom *atom,
524 				       struct perf_sample *sample)
525 {
526 	u64 delta;
527 	u64 entry_time = sample->time;
528 	u64 raise_time = atom->time;
529 
530 	if ((raise_time != 0) && (entry_time >= raise_time)) {
531 		delta = entry_time - raise_time;
532 		if ((delta > work->max_latency) ||
533 		    (work->max_latency == 0)) {
534 			work->max_latency = delta;
535 			work->max_latency_start = raise_time;
536 			work->max_latency_end = entry_time;
537 		}
538 		work->total_latency += delta;
539 		work->nr_atoms++;
540 	}
541 }
542 
543 static int latency_raise_event(struct perf_kwork *kwork,
544 			       struct kwork_class *class,
545 			       struct evsel *evsel,
546 			       struct perf_sample *sample,
547 			       struct machine *machine)
548 {
549 	return work_push_atom(kwork, class, KWORK_TRACE_RAISE,
550 			      KWORK_TRACE_MAX, evsel, sample,
551 			      machine, NULL);
552 }
553 
554 static int latency_entry_event(struct perf_kwork *kwork,
555 			       struct kwork_class *class,
556 			       struct evsel *evsel,
557 			       struct perf_sample *sample,
558 			       struct machine *machine)
559 {
560 	struct kwork_atom *atom = NULL;
561 	struct kwork_work *work = NULL;
562 
563 	atom = work_pop_atom(kwork, class, KWORK_TRACE_ENTRY,
564 			     KWORK_TRACE_RAISE, evsel, sample,
565 			     machine, &work);
566 	if (work == NULL)
567 		return -1;
568 
569 	if (atom != NULL) {
570 		latency_update_entry_event(work, atom, sample);
571 		atom_del(atom);
572 	}
573 
574 	return 0;
575 }
576 
577 static void timehist_save_callchain(struct perf_kwork *kwork,
578 				    struct perf_sample *sample,
579 				    struct evsel *evsel,
580 				    struct machine *machine)
581 {
582 	struct symbol *sym;
583 	struct thread *thread;
584 	struct callchain_cursor_node *node;
585 	struct callchain_cursor *cursor = &callchain_cursor;
586 
587 	if (!kwork->show_callchain || sample->callchain == NULL)
588 		return;
589 
590 	/* want main thread for process - has maps */
591 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
592 	if (thread == NULL) {
593 		pr_debug("Failed to get thread for pid %d\n", sample->pid);
594 		return;
595 	}
596 
597 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
598 				      NULL, NULL, kwork->max_stack + 2) != 0) {
599 		pr_debug("Failed to resolve callchain, skipping\n");
600 		goto out_put;
601 	}
602 
603 	callchain_cursor_commit(cursor);
604 
605 	while (true) {
606 		node = callchain_cursor_current(cursor);
607 		if (node == NULL)
608 			break;
609 
610 		sym = node->ms.sym;
611 		if (sym) {
612 			if (!strcmp(sym->name, "__softirqentry_text_start") ||
613 			    !strcmp(sym->name, "__do_softirq"))
614 				sym->ignore = 1;
615 		}
616 
617 		callchain_cursor_advance(cursor);
618 	}
619 
620 out_put:
621 	thread__put(thread);
622 }
623 
624 static void timehist_print_event(struct perf_kwork *kwork,
625 				 struct kwork_work *work,
626 				 struct kwork_atom *atom,
627 				 struct perf_sample *sample,
628 				 struct addr_location *al)
629 {
630 	char entrytime[32], exittime[32];
631 	char kwork_name[PRINT_KWORK_NAME_WIDTH];
632 
633 	/*
634 	 * runtime start
635 	 */
636 	timestamp__scnprintf_usec(atom->time,
637 				  entrytime, sizeof(entrytime));
638 	printf(" %*s ", PRINT_TIMESTAMP_WIDTH, entrytime);
639 
640 	/*
641 	 * runtime end
642 	 */
643 	timestamp__scnprintf_usec(sample->time,
644 				  exittime, sizeof(exittime));
645 	printf(" %*s ", PRINT_TIMESTAMP_WIDTH, exittime);
646 
647 	/*
648 	 * cpu
649 	 */
650 	printf(" [%0*d] ", PRINT_CPU_WIDTH, work->cpu);
651 
652 	/*
653 	 * kwork name
654 	 */
655 	if (work->class && work->class->work_name) {
656 		work->class->work_name(work, kwork_name,
657 				       PRINT_KWORK_NAME_WIDTH);
658 		printf(" %-*s ", PRINT_KWORK_NAME_WIDTH, kwork_name);
659 	} else
660 		printf(" %-*s ", PRINT_KWORK_NAME_WIDTH, "");
661 
662 	/*
663 	 *runtime
664 	 */
665 	printf(" %*.*f ",
666 	       PRINT_RUNTIME_WIDTH, RPINT_DECIMAL_WIDTH,
667 	       (double)(sample->time - atom->time) / NSEC_PER_MSEC);
668 
669 	/*
670 	 * delaytime
671 	 */
672 	if (atom->prev != NULL)
673 		printf(" %*.*f ", PRINT_LATENCY_WIDTH, RPINT_DECIMAL_WIDTH,
674 		       (double)(atom->time - atom->prev->time) / NSEC_PER_MSEC);
675 	else
676 		printf(" %*s ", PRINT_LATENCY_WIDTH, " ");
677 
678 	/*
679 	 * callchain
680 	 */
681 	if (kwork->show_callchain) {
682 		printf(" ");
683 		sample__fprintf_sym(sample, al, 0,
684 				    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
685 				    EVSEL__PRINT_CALLCHAIN_ARROW |
686 				    EVSEL__PRINT_SKIP_IGNORED,
687 				    &callchain_cursor, symbol_conf.bt_stop_list,
688 				    stdout);
689 	}
690 
691 	printf("\n");
692 }
693 
694 static int timehist_raise_event(struct perf_kwork *kwork,
695 				struct kwork_class *class,
696 				struct evsel *evsel,
697 				struct perf_sample *sample,
698 				struct machine *machine)
699 {
700 	return work_push_atom(kwork, class, KWORK_TRACE_RAISE,
701 			      KWORK_TRACE_MAX, evsel, sample,
702 			      machine, NULL);
703 }
704 
705 static int timehist_entry_event(struct perf_kwork *kwork,
706 				struct kwork_class *class,
707 				struct evsel *evsel,
708 				struct perf_sample *sample,
709 				struct machine *machine)
710 {
711 	int ret;
712 	struct kwork_work *work = NULL;
713 
714 	ret = work_push_atom(kwork, class, KWORK_TRACE_ENTRY,
715 			     KWORK_TRACE_RAISE, evsel, sample,
716 			     machine, &work);
717 	if (ret)
718 		return ret;
719 
720 	if (work != NULL)
721 		timehist_save_callchain(kwork, sample, evsel, machine);
722 
723 	return 0;
724 }
725 
726 static int timehist_exit_event(struct perf_kwork *kwork,
727 			       struct kwork_class *class,
728 			       struct evsel *evsel,
729 			       struct perf_sample *sample,
730 			       struct machine *machine)
731 {
732 	struct kwork_atom *atom = NULL;
733 	struct kwork_work *work = NULL;
734 	struct addr_location al;
735 
736 	if (machine__resolve(machine, &al, sample) < 0) {
737 		pr_debug("Problem processing event, skipping it\n");
738 		return -1;
739 	}
740 
741 	atom = work_pop_atom(kwork, class, KWORK_TRACE_EXIT,
742 			     KWORK_TRACE_ENTRY, evsel, sample,
743 			     machine, &work);
744 	if (work == NULL)
745 		return -1;
746 
747 	if (atom != NULL) {
748 		work->nr_atoms++;
749 		timehist_print_event(kwork, work, atom, sample, &al);
750 		atom_del(atom);
751 	}
752 
753 	return 0;
754 }
755 
756 static struct kwork_class kwork_irq;
757 static int process_irq_handler_entry_event(struct perf_tool *tool,
758 					   struct evsel *evsel,
759 					   struct perf_sample *sample,
760 					   struct machine *machine)
761 {
762 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
763 
764 	if (kwork->tp_handler->entry_event)
765 		return kwork->tp_handler->entry_event(kwork, &kwork_irq,
766 						      evsel, sample, machine);
767 	return 0;
768 }
769 
770 static int process_irq_handler_exit_event(struct perf_tool *tool,
771 					  struct evsel *evsel,
772 					  struct perf_sample *sample,
773 					  struct machine *machine)
774 {
775 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
776 
777 	if (kwork->tp_handler->exit_event)
778 		return kwork->tp_handler->exit_event(kwork, &kwork_irq,
779 						     evsel, sample, machine);
780 	return 0;
781 }
782 
783 const struct evsel_str_handler irq_tp_handlers[] = {
784 	{ "irq:irq_handler_entry", process_irq_handler_entry_event, },
785 	{ "irq:irq_handler_exit",  process_irq_handler_exit_event,  },
786 };
787 
788 static int irq_class_init(struct kwork_class *class,
789 			  struct perf_session *session)
790 {
791 	if (perf_session__set_tracepoints_handlers(session, irq_tp_handlers)) {
792 		pr_err("Failed to set irq tracepoints handlers\n");
793 		return -1;
794 	}
795 
796 	class->work_root = RB_ROOT_CACHED;
797 	return 0;
798 }
799 
800 static void irq_work_init(struct kwork_class *class,
801 			  struct kwork_work *work,
802 			  struct evsel *evsel,
803 			  struct perf_sample *sample,
804 			  struct machine *machine __maybe_unused)
805 {
806 	work->class = class;
807 	work->cpu = sample->cpu;
808 	work->id = evsel__intval(evsel, sample, "irq");
809 	work->name = evsel__strval(evsel, sample, "name");
810 }
811 
812 static void irq_work_name(struct kwork_work *work, char *buf, int len)
813 {
814 	snprintf(buf, len, "%s:%" PRIu64 "", work->name, work->id);
815 }
816 
817 static struct kwork_class kwork_irq = {
818 	.name           = "irq",
819 	.type           = KWORK_CLASS_IRQ,
820 	.nr_tracepoints = 2,
821 	.tp_handlers    = irq_tp_handlers,
822 	.class_init     = irq_class_init,
823 	.work_init      = irq_work_init,
824 	.work_name      = irq_work_name,
825 };
826 
827 static struct kwork_class kwork_softirq;
828 static int process_softirq_raise_event(struct perf_tool *tool,
829 				       struct evsel *evsel,
830 				       struct perf_sample *sample,
831 				       struct machine *machine)
832 {
833 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
834 
835 	if (kwork->tp_handler->raise_event)
836 		return kwork->tp_handler->raise_event(kwork, &kwork_softirq,
837 						      evsel, sample, machine);
838 
839 	return 0;
840 }
841 
842 static int process_softirq_entry_event(struct perf_tool *tool,
843 				       struct evsel *evsel,
844 				       struct perf_sample *sample,
845 				       struct machine *machine)
846 {
847 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
848 
849 	if (kwork->tp_handler->entry_event)
850 		return kwork->tp_handler->entry_event(kwork, &kwork_softirq,
851 						      evsel, sample, machine);
852 
853 	return 0;
854 }
855 
856 static int process_softirq_exit_event(struct perf_tool *tool,
857 				      struct evsel *evsel,
858 				      struct perf_sample *sample,
859 				      struct machine *machine)
860 {
861 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
862 
863 	if (kwork->tp_handler->exit_event)
864 		return kwork->tp_handler->exit_event(kwork, &kwork_softirq,
865 						     evsel, sample, machine);
866 
867 	return 0;
868 }
869 
870 const struct evsel_str_handler softirq_tp_handlers[] = {
871 	{ "irq:softirq_raise", process_softirq_raise_event, },
872 	{ "irq:softirq_entry", process_softirq_entry_event, },
873 	{ "irq:softirq_exit",  process_softirq_exit_event,  },
874 };
875 
876 static int softirq_class_init(struct kwork_class *class,
877 			      struct perf_session *session)
878 {
879 	if (perf_session__set_tracepoints_handlers(session,
880 						   softirq_tp_handlers)) {
881 		pr_err("Failed to set softirq tracepoints handlers\n");
882 		return -1;
883 	}
884 
885 	class->work_root = RB_ROOT_CACHED;
886 	return 0;
887 }
888 
889 static char *evsel__softirq_name(struct evsel *evsel, u64 num)
890 {
891 	char *name = NULL;
892 	bool found = false;
893 	struct tep_print_flag_sym *sym = NULL;
894 	struct tep_print_arg *args = evsel->tp_format->print_fmt.args;
895 
896 	if ((args == NULL) || (args->next == NULL))
897 		return NULL;
898 
899 	/* skip softirq field: "REC->vec" */
900 	for (sym = args->next->symbol.symbols; sym != NULL; sym = sym->next) {
901 		if ((eval_flag(sym->value) == (unsigned long long)num) &&
902 		    (strlen(sym->str) != 0)) {
903 			found = true;
904 			break;
905 		}
906 	}
907 
908 	if (!found)
909 		return NULL;
910 
911 	name = strdup(sym->str);
912 	if (name == NULL) {
913 		pr_err("Failed to copy symbol name\n");
914 		return NULL;
915 	}
916 	return name;
917 }
918 
919 static void softirq_work_init(struct kwork_class *class,
920 			      struct kwork_work *work,
921 			      struct evsel *evsel,
922 			      struct perf_sample *sample,
923 			      struct machine *machine __maybe_unused)
924 {
925 	u64 num = evsel__intval(evsel, sample, "vec");
926 
927 	work->id = num;
928 	work->class = class;
929 	work->cpu = sample->cpu;
930 	work->name = evsel__softirq_name(evsel, num);
931 }
932 
933 static void softirq_work_name(struct kwork_work *work, char *buf, int len)
934 {
935 	snprintf(buf, len, "(s)%s:%" PRIu64 "", work->name, work->id);
936 }
937 
938 static struct kwork_class kwork_softirq = {
939 	.name           = "softirq",
940 	.type           = KWORK_CLASS_SOFTIRQ,
941 	.nr_tracepoints = 3,
942 	.tp_handlers    = softirq_tp_handlers,
943 	.class_init     = softirq_class_init,
944 	.work_init      = softirq_work_init,
945 	.work_name      = softirq_work_name,
946 };
947 
948 static struct kwork_class kwork_workqueue;
949 static int process_workqueue_activate_work_event(struct perf_tool *tool,
950 						 struct evsel *evsel,
951 						 struct perf_sample *sample,
952 						 struct machine *machine)
953 {
954 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
955 
956 	if (kwork->tp_handler->raise_event)
957 		return kwork->tp_handler->raise_event(kwork, &kwork_workqueue,
958 						    evsel, sample, machine);
959 
960 	return 0;
961 }
962 
963 static int process_workqueue_execute_start_event(struct perf_tool *tool,
964 						 struct evsel *evsel,
965 						 struct perf_sample *sample,
966 						 struct machine *machine)
967 {
968 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
969 
970 	if (kwork->tp_handler->entry_event)
971 		return kwork->tp_handler->entry_event(kwork, &kwork_workqueue,
972 						    evsel, sample, machine);
973 
974 	return 0;
975 }
976 
977 static int process_workqueue_execute_end_event(struct perf_tool *tool,
978 					       struct evsel *evsel,
979 					       struct perf_sample *sample,
980 					       struct machine *machine)
981 {
982 	struct perf_kwork *kwork = container_of(tool, struct perf_kwork, tool);
983 
984 	if (kwork->tp_handler->exit_event)
985 		return kwork->tp_handler->exit_event(kwork, &kwork_workqueue,
986 						   evsel, sample, machine);
987 
988 	return 0;
989 }
990 
991 const struct evsel_str_handler workqueue_tp_handlers[] = {
992 	{ "workqueue:workqueue_activate_work", process_workqueue_activate_work_event, },
993 	{ "workqueue:workqueue_execute_start", process_workqueue_execute_start_event, },
994 	{ "workqueue:workqueue_execute_end",   process_workqueue_execute_end_event,   },
995 };
996 
997 static int workqueue_class_init(struct kwork_class *class,
998 				struct perf_session *session)
999 {
1000 	if (perf_session__set_tracepoints_handlers(session,
1001 						   workqueue_tp_handlers)) {
1002 		pr_err("Failed to set workqueue tracepoints handlers\n");
1003 		return -1;
1004 	}
1005 
1006 	class->work_root = RB_ROOT_CACHED;
1007 	return 0;
1008 }
1009 
1010 static void workqueue_work_init(struct kwork_class *class,
1011 				struct kwork_work *work,
1012 				struct evsel *evsel,
1013 				struct perf_sample *sample,
1014 				struct machine *machine)
1015 {
1016 	char *modp = NULL;
1017 	unsigned long long function_addr = evsel__intval(evsel,
1018 							 sample, "function");
1019 
1020 	work->class = class;
1021 	work->cpu = sample->cpu;
1022 	work->id = evsel__intval(evsel, sample, "work");
1023 	work->name = function_addr == 0 ? NULL :
1024 		machine__resolve_kernel_addr(machine, &function_addr, &modp);
1025 }
1026 
1027 static void workqueue_work_name(struct kwork_work *work, char *buf, int len)
1028 {
1029 	if (work->name != NULL)
1030 		snprintf(buf, len, "(w)%s", work->name);
1031 	else
1032 		snprintf(buf, len, "(w)0x%" PRIx64, work->id);
1033 }
1034 
1035 static struct kwork_class kwork_workqueue = {
1036 	.name           = "workqueue",
1037 	.type           = KWORK_CLASS_WORKQUEUE,
1038 	.nr_tracepoints = 3,
1039 	.tp_handlers    = workqueue_tp_handlers,
1040 	.class_init     = workqueue_class_init,
1041 	.work_init      = workqueue_work_init,
1042 	.work_name      = workqueue_work_name,
1043 };
1044 
1045 static struct kwork_class *kwork_class_supported_list[KWORK_CLASS_MAX] = {
1046 	[KWORK_CLASS_IRQ]       = &kwork_irq,
1047 	[KWORK_CLASS_SOFTIRQ]   = &kwork_softirq,
1048 	[KWORK_CLASS_WORKQUEUE] = &kwork_workqueue,
1049 };
1050 
1051 static void print_separator(int len)
1052 {
1053 	printf(" %.*s\n", len, graph_dotted_line);
1054 }
1055 
1056 static int report_print_work(struct perf_kwork *kwork, struct kwork_work *work)
1057 {
1058 	int ret = 0;
1059 	char kwork_name[PRINT_KWORK_NAME_WIDTH];
1060 	char max_runtime_start[32], max_runtime_end[32];
1061 	char max_latency_start[32], max_latency_end[32];
1062 
1063 	printf(" ");
1064 
1065 	/*
1066 	 * kwork name
1067 	 */
1068 	if (work->class && work->class->work_name) {
1069 		work->class->work_name(work, kwork_name,
1070 				       PRINT_KWORK_NAME_WIDTH);
1071 		ret += printf(" %-*s |", PRINT_KWORK_NAME_WIDTH, kwork_name);
1072 	} else {
1073 		ret += printf(" %-*s |", PRINT_KWORK_NAME_WIDTH, "");
1074 	}
1075 
1076 	/*
1077 	 * cpu
1078 	 */
1079 	ret += printf(" %0*d |", PRINT_CPU_WIDTH, work->cpu);
1080 
1081 	/*
1082 	 * total runtime
1083 	 */
1084 	if (kwork->report == KWORK_REPORT_RUNTIME) {
1085 		ret += printf(" %*.*f ms |",
1086 			      PRINT_RUNTIME_WIDTH, RPINT_DECIMAL_WIDTH,
1087 			      (double)work->total_runtime / NSEC_PER_MSEC);
1088 	} else if (kwork->report == KWORK_REPORT_LATENCY) { // avg delay
1089 		ret += printf(" %*.*f ms |",
1090 			      PRINT_LATENCY_WIDTH, RPINT_DECIMAL_WIDTH,
1091 			      (double)work->total_latency /
1092 			      work->nr_atoms / NSEC_PER_MSEC);
1093 	}
1094 
1095 	/*
1096 	 * count
1097 	 */
1098 	ret += printf(" %*" PRIu64 " |", PRINT_COUNT_WIDTH, work->nr_atoms);
1099 
1100 	/*
1101 	 * max runtime, max runtime start, max runtime end
1102 	 */
1103 	if (kwork->report == KWORK_REPORT_RUNTIME) {
1104 		timestamp__scnprintf_usec(work->max_runtime_start,
1105 					  max_runtime_start,
1106 					  sizeof(max_runtime_start));
1107 		timestamp__scnprintf_usec(work->max_runtime_end,
1108 					  max_runtime_end,
1109 					  sizeof(max_runtime_end));
1110 		ret += printf(" %*.*f ms | %*s s | %*s s |",
1111 			      PRINT_RUNTIME_WIDTH, RPINT_DECIMAL_WIDTH,
1112 			      (double)work->max_runtime / NSEC_PER_MSEC,
1113 			      PRINT_TIMESTAMP_WIDTH, max_runtime_start,
1114 			      PRINT_TIMESTAMP_WIDTH, max_runtime_end);
1115 	}
1116 	/*
1117 	 * max delay, max delay start, max delay end
1118 	 */
1119 	else if (kwork->report == KWORK_REPORT_LATENCY) {
1120 		timestamp__scnprintf_usec(work->max_latency_start,
1121 					  max_latency_start,
1122 					  sizeof(max_latency_start));
1123 		timestamp__scnprintf_usec(work->max_latency_end,
1124 					  max_latency_end,
1125 					  sizeof(max_latency_end));
1126 		ret += printf(" %*.*f ms | %*s s | %*s s |",
1127 			      PRINT_LATENCY_WIDTH, RPINT_DECIMAL_WIDTH,
1128 			      (double)work->max_latency / NSEC_PER_MSEC,
1129 			      PRINT_TIMESTAMP_WIDTH, max_latency_start,
1130 			      PRINT_TIMESTAMP_WIDTH, max_latency_end);
1131 	}
1132 
1133 	printf("\n");
1134 	return ret;
1135 }
1136 
1137 static int report_print_header(struct perf_kwork *kwork)
1138 {
1139 	int ret;
1140 
1141 	printf("\n ");
1142 	ret = printf(" %-*s | %-*s |",
1143 		     PRINT_KWORK_NAME_WIDTH, "Kwork Name",
1144 		     PRINT_CPU_WIDTH, "Cpu");
1145 
1146 	if (kwork->report == KWORK_REPORT_RUNTIME) {
1147 		ret += printf(" %-*s |",
1148 			      PRINT_RUNTIME_HEADER_WIDTH, "Total Runtime");
1149 	} else if (kwork->report == KWORK_REPORT_LATENCY) {
1150 		ret += printf(" %-*s |",
1151 			      PRINT_LATENCY_HEADER_WIDTH, "Avg delay");
1152 	}
1153 
1154 	ret += printf(" %-*s |", PRINT_COUNT_WIDTH, "Count");
1155 
1156 	if (kwork->report == KWORK_REPORT_RUNTIME) {
1157 		ret += printf(" %-*s | %-*s | %-*s |",
1158 			      PRINT_RUNTIME_HEADER_WIDTH, "Max runtime",
1159 			      PRINT_TIMESTAMP_HEADER_WIDTH, "Max runtime start",
1160 			      PRINT_TIMESTAMP_HEADER_WIDTH, "Max runtime end");
1161 	} else if (kwork->report == KWORK_REPORT_LATENCY) {
1162 		ret += printf(" %-*s | %-*s | %-*s |",
1163 			      PRINT_LATENCY_HEADER_WIDTH, "Max delay",
1164 			      PRINT_TIMESTAMP_HEADER_WIDTH, "Max delay start",
1165 			      PRINT_TIMESTAMP_HEADER_WIDTH, "Max delay end");
1166 	}
1167 
1168 	printf("\n");
1169 	print_separator(ret);
1170 	return ret;
1171 }
1172 
1173 static void timehist_print_header(void)
1174 {
1175 	/*
1176 	 * header row
1177 	 */
1178 	printf(" %-*s  %-*s  %-*s  %-*s  %-*s  %-*s\n",
1179 	       PRINT_TIMESTAMP_WIDTH, "Runtime start",
1180 	       PRINT_TIMESTAMP_WIDTH, "Runtime end",
1181 	       PRINT_TIMEHIST_CPU_WIDTH, "Cpu",
1182 	       PRINT_KWORK_NAME_WIDTH, "Kwork name",
1183 	       PRINT_RUNTIME_WIDTH, "Runtime",
1184 	       PRINT_RUNTIME_WIDTH, "Delaytime");
1185 
1186 	/*
1187 	 * units row
1188 	 */
1189 	printf(" %-*s  %-*s  %-*s  %-*s  %-*s  %-*s\n",
1190 	       PRINT_TIMESTAMP_WIDTH, "",
1191 	       PRINT_TIMESTAMP_WIDTH, "",
1192 	       PRINT_TIMEHIST_CPU_WIDTH, "",
1193 	       PRINT_KWORK_NAME_WIDTH, "(TYPE)NAME:NUM",
1194 	       PRINT_RUNTIME_WIDTH, "(msec)",
1195 	       PRINT_RUNTIME_WIDTH, "(msec)");
1196 
1197 	/*
1198 	 * separator
1199 	 */
1200 	printf(" %.*s  %.*s  %.*s  %.*s  %.*s  %.*s\n",
1201 	       PRINT_TIMESTAMP_WIDTH, graph_dotted_line,
1202 	       PRINT_TIMESTAMP_WIDTH, graph_dotted_line,
1203 	       PRINT_TIMEHIST_CPU_WIDTH, graph_dotted_line,
1204 	       PRINT_KWORK_NAME_WIDTH, graph_dotted_line,
1205 	       PRINT_RUNTIME_WIDTH, graph_dotted_line,
1206 	       PRINT_RUNTIME_WIDTH, graph_dotted_line);
1207 }
1208 
1209 static void print_summary(struct perf_kwork *kwork)
1210 {
1211 	u64 time = kwork->timeend - kwork->timestart;
1212 
1213 	printf("  Total count            : %9" PRIu64 "\n", kwork->all_count);
1214 	printf("  Total runtime   (msec) : %9.3f (%.3f%% load average)\n",
1215 	       (double)kwork->all_runtime / NSEC_PER_MSEC,
1216 	       time == 0 ? 0 : (double)kwork->all_runtime / time);
1217 	printf("  Total time span (msec) : %9.3f\n",
1218 	       (double)time / NSEC_PER_MSEC);
1219 }
1220 
1221 static unsigned long long nr_list_entry(struct list_head *head)
1222 {
1223 	struct list_head *pos;
1224 	unsigned long long n = 0;
1225 
1226 	list_for_each(pos, head)
1227 		n++;
1228 
1229 	return n;
1230 }
1231 
1232 static void print_skipped_events(struct perf_kwork *kwork)
1233 {
1234 	int i;
1235 	const char *const kwork_event_str[] = {
1236 		[KWORK_TRACE_RAISE] = "raise",
1237 		[KWORK_TRACE_ENTRY] = "entry",
1238 		[KWORK_TRACE_EXIT]  = "exit",
1239 	};
1240 
1241 	if ((kwork->nr_skipped_events[KWORK_TRACE_MAX] != 0) &&
1242 	    (kwork->nr_events != 0)) {
1243 		printf("  INFO: %.3f%% skipped events (%" PRIu64 " including ",
1244 		       (double)kwork->nr_skipped_events[KWORK_TRACE_MAX] /
1245 		       (double)kwork->nr_events * 100.0,
1246 		       kwork->nr_skipped_events[KWORK_TRACE_MAX]);
1247 
1248 		for (i = 0; i < KWORK_TRACE_MAX; i++) {
1249 			printf("%" PRIu64 " %s%s",
1250 			       kwork->nr_skipped_events[i],
1251 			       kwork_event_str[i],
1252 			       (i == KWORK_TRACE_MAX - 1) ? ")\n" : ", ");
1253 		}
1254 	}
1255 
1256 	if (verbose > 0)
1257 		printf("  INFO: use %lld atom pages\n",
1258 		       nr_list_entry(&kwork->atom_page_list));
1259 }
1260 
1261 static void print_bad_events(struct perf_kwork *kwork)
1262 {
1263 	if ((kwork->nr_lost_events != 0) && (kwork->nr_events != 0)) {
1264 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1265 		       (double)kwork->nr_lost_events /
1266 		       (double)kwork->nr_events * 100.0,
1267 		       kwork->nr_lost_events, kwork->nr_events,
1268 		       kwork->nr_lost_chunks);
1269 	}
1270 }
1271 
1272 static void work_sort(struct perf_kwork *kwork, struct kwork_class *class)
1273 {
1274 	struct rb_node *node;
1275 	struct kwork_work *data;
1276 	struct rb_root_cached *root = &class->work_root;
1277 
1278 	pr_debug("Sorting %s ...\n", class->name);
1279 	for (;;) {
1280 		node = rb_first_cached(root);
1281 		if (!node)
1282 			break;
1283 
1284 		rb_erase_cached(node, root);
1285 		data = rb_entry(node, struct kwork_work, node);
1286 		work_insert(&kwork->sorted_work_root,
1287 			       data, &kwork->sort_list);
1288 	}
1289 }
1290 
1291 static void perf_kwork__sort(struct perf_kwork *kwork)
1292 {
1293 	struct kwork_class *class;
1294 
1295 	list_for_each_entry(class, &kwork->class_list, list)
1296 		work_sort(kwork, class);
1297 }
1298 
1299 static int perf_kwork__check_config(struct perf_kwork *kwork,
1300 				    struct perf_session *session)
1301 {
1302 	int ret;
1303 	struct evsel *evsel;
1304 	struct kwork_class *class;
1305 
1306 	static struct trace_kwork_handler report_ops = {
1307 		.entry_event = report_entry_event,
1308 		.exit_event  = report_exit_event,
1309 	};
1310 	static struct trace_kwork_handler latency_ops = {
1311 		.raise_event = latency_raise_event,
1312 		.entry_event = latency_entry_event,
1313 	};
1314 	static struct trace_kwork_handler timehist_ops = {
1315 		.raise_event = timehist_raise_event,
1316 		.entry_event = timehist_entry_event,
1317 		.exit_event  = timehist_exit_event,
1318 	};
1319 
1320 	switch (kwork->report) {
1321 	case KWORK_REPORT_RUNTIME:
1322 		kwork->tp_handler = &report_ops;
1323 		break;
1324 	case KWORK_REPORT_LATENCY:
1325 		kwork->tp_handler = &latency_ops;
1326 		break;
1327 	case KWORK_REPORT_TIMEHIST:
1328 		kwork->tp_handler = &timehist_ops;
1329 		break;
1330 	default:
1331 		pr_debug("Invalid report type %d\n", kwork->report);
1332 		return -1;
1333 	}
1334 
1335 	list_for_each_entry(class, &kwork->class_list, list)
1336 		if ((class->class_init != NULL) &&
1337 		    (class->class_init(class, session) != 0))
1338 			return -1;
1339 
1340 	if (kwork->cpu_list != NULL) {
1341 		ret = perf_session__cpu_bitmap(session,
1342 					       kwork->cpu_list,
1343 					       kwork->cpu_bitmap);
1344 		if (ret < 0) {
1345 			pr_err("Invalid cpu bitmap\n");
1346 			return -1;
1347 		}
1348 	}
1349 
1350 	if (kwork->time_str != NULL) {
1351 		ret = perf_time__parse_str(&kwork->ptime, kwork->time_str);
1352 		if (ret != 0) {
1353 			pr_err("Invalid time span\n");
1354 			return -1;
1355 		}
1356 	}
1357 
1358 	list_for_each_entry(evsel, &session->evlist->core.entries, core.node) {
1359 		if (kwork->show_callchain && !evsel__has_callchain(evsel)) {
1360 			pr_debug("Samples do not have callchains\n");
1361 			kwork->show_callchain = 0;
1362 			symbol_conf.use_callchain = 0;
1363 		}
1364 	}
1365 
1366 	return 0;
1367 }
1368 
1369 static int perf_kwork__read_events(struct perf_kwork *kwork)
1370 {
1371 	int ret = -1;
1372 	struct perf_session *session = NULL;
1373 
1374 	struct perf_data data = {
1375 		.path  = input_name,
1376 		.mode  = PERF_DATA_MODE_READ,
1377 		.force = kwork->force,
1378 	};
1379 
1380 	session = perf_session__new(&data, &kwork->tool);
1381 	if (IS_ERR(session)) {
1382 		pr_debug("Error creating perf session\n");
1383 		return PTR_ERR(session);
1384 	}
1385 
1386 	symbol__init(&session->header.env);
1387 
1388 	if (perf_kwork__check_config(kwork, session) != 0)
1389 		goto out_delete;
1390 
1391 	if (session->tevent.pevent &&
1392 	    tep_set_function_resolver(session->tevent.pevent,
1393 				      machine__resolve_kernel_addr,
1394 				      &session->machines.host) < 0) {
1395 		pr_err("Failed to set libtraceevent function resolver\n");
1396 		goto out_delete;
1397 	}
1398 
1399 	if (kwork->report == KWORK_REPORT_TIMEHIST)
1400 		timehist_print_header();
1401 
1402 	ret = perf_session__process_events(session);
1403 	if (ret) {
1404 		pr_debug("Failed to process events, error %d\n", ret);
1405 		goto out_delete;
1406 	}
1407 
1408 	kwork->nr_events      = session->evlist->stats.nr_events[0];
1409 	kwork->nr_lost_events = session->evlist->stats.total_lost;
1410 	kwork->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1411 
1412 out_delete:
1413 	perf_session__delete(session);
1414 	return ret;
1415 }
1416 
1417 static void process_skipped_events(struct perf_kwork *kwork,
1418 				   struct kwork_work *work)
1419 {
1420 	int i;
1421 	unsigned long long count;
1422 
1423 	for (i = 0; i < KWORK_TRACE_MAX; i++) {
1424 		count = nr_list_entry(&work->atom_list[i]);
1425 		kwork->nr_skipped_events[i] += count;
1426 		kwork->nr_skipped_events[KWORK_TRACE_MAX] += count;
1427 	}
1428 }
1429 
1430 struct kwork_work *perf_kwork_add_work(struct perf_kwork *kwork,
1431 				       struct kwork_class *class,
1432 				       struct kwork_work *key)
1433 {
1434 	struct kwork_work *work = NULL;
1435 
1436 	work = work_new(key);
1437 	if (work == NULL)
1438 		return NULL;
1439 
1440 	work_insert(&class->work_root, work, &kwork->cmp_id);
1441 	return work;
1442 }
1443 
1444 static void sig_handler(int sig)
1445 {
1446 	/*
1447 	 * Simply capture termination signal so that
1448 	 * the program can continue after pause returns
1449 	 */
1450 	pr_debug("Captuer signal %d\n", sig);
1451 }
1452 
1453 static int perf_kwork__report_bpf(struct perf_kwork *kwork)
1454 {
1455 	int ret;
1456 
1457 	signal(SIGINT, sig_handler);
1458 	signal(SIGTERM, sig_handler);
1459 
1460 	ret = perf_kwork__trace_prepare_bpf(kwork);
1461 	if (ret)
1462 		return -1;
1463 
1464 	printf("Starting trace, Hit <Ctrl+C> to stop and report\n");
1465 
1466 	perf_kwork__trace_start();
1467 
1468 	/*
1469 	 * a simple pause, wait here for stop signal
1470 	 */
1471 	pause();
1472 
1473 	perf_kwork__trace_finish();
1474 
1475 	perf_kwork__report_read_bpf(kwork);
1476 
1477 	perf_kwork__report_cleanup_bpf();
1478 
1479 	return 0;
1480 }
1481 
1482 static int perf_kwork__report(struct perf_kwork *kwork)
1483 {
1484 	int ret;
1485 	struct rb_node *next;
1486 	struct kwork_work *work;
1487 
1488 	if (kwork->use_bpf)
1489 		ret = perf_kwork__report_bpf(kwork);
1490 	else
1491 		ret = perf_kwork__read_events(kwork);
1492 
1493 	if (ret != 0)
1494 		return -1;
1495 
1496 	perf_kwork__sort(kwork);
1497 
1498 	setup_pager();
1499 
1500 	ret = report_print_header(kwork);
1501 	next = rb_first_cached(&kwork->sorted_work_root);
1502 	while (next) {
1503 		work = rb_entry(next, struct kwork_work, node);
1504 		process_skipped_events(kwork, work);
1505 
1506 		if (work->nr_atoms != 0) {
1507 			report_print_work(kwork, work);
1508 			if (kwork->summary) {
1509 				kwork->all_runtime += work->total_runtime;
1510 				kwork->all_count += work->nr_atoms;
1511 			}
1512 		}
1513 		next = rb_next(next);
1514 	}
1515 	print_separator(ret);
1516 
1517 	if (kwork->summary) {
1518 		print_summary(kwork);
1519 		print_separator(ret);
1520 	}
1521 
1522 	print_bad_events(kwork);
1523 	print_skipped_events(kwork);
1524 	printf("\n");
1525 
1526 	return 0;
1527 }
1528 
1529 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1530 				  struct evsel *evsel,
1531 				  struct perf_sample *sample,
1532 				  struct machine *machine);
1533 
1534 static int perf_kwork__process_tracepoint_sample(struct perf_tool *tool,
1535 						 union perf_event *event __maybe_unused,
1536 						 struct perf_sample *sample,
1537 						 struct evsel *evsel,
1538 						 struct machine *machine)
1539 {
1540 	int err = 0;
1541 
1542 	if (evsel->handler != NULL) {
1543 		tracepoint_handler f = evsel->handler;
1544 
1545 		err = f(tool, evsel, sample, machine);
1546 	}
1547 
1548 	return err;
1549 }
1550 
1551 static int perf_kwork__timehist(struct perf_kwork *kwork)
1552 {
1553 	/*
1554 	 * event handlers for timehist option
1555 	 */
1556 	kwork->tool.comm	 = perf_event__process_comm;
1557 	kwork->tool.exit	 = perf_event__process_exit;
1558 	kwork->tool.fork	 = perf_event__process_fork;
1559 	kwork->tool.attr	 = perf_event__process_attr;
1560 	kwork->tool.tracing_data = perf_event__process_tracing_data;
1561 	kwork->tool.build_id	 = perf_event__process_build_id;
1562 	kwork->tool.ordered_events = true;
1563 	kwork->tool.ordering_requires_timestamps = true;
1564 	symbol_conf.use_callchain = kwork->show_callchain;
1565 
1566 	if (symbol__validate_sym_arguments()) {
1567 		pr_err("Failed to validate sym arguments\n");
1568 		return -1;
1569 	}
1570 
1571 	setup_pager();
1572 
1573 	return perf_kwork__read_events(kwork);
1574 }
1575 
1576 static void setup_event_list(struct perf_kwork *kwork,
1577 			     const struct option *options,
1578 			     const char * const usage_msg[])
1579 {
1580 	int i;
1581 	struct kwork_class *class;
1582 	char *tmp, *tok, *str;
1583 
1584 	if (kwork->event_list_str == NULL)
1585 		goto null_event_list_str;
1586 
1587 	str = strdup(kwork->event_list_str);
1588 	for (tok = strtok_r(str, ", ", &tmp);
1589 	     tok; tok = strtok_r(NULL, ", ", &tmp)) {
1590 		for (i = 0; i < KWORK_CLASS_MAX; i++) {
1591 			class = kwork_class_supported_list[i];
1592 			if (strcmp(tok, class->name) == 0) {
1593 				list_add_tail(&class->list, &kwork->class_list);
1594 				break;
1595 			}
1596 		}
1597 		if (i == KWORK_CLASS_MAX) {
1598 			usage_with_options_msg(usage_msg, options,
1599 					       "Unknown --event key: `%s'", tok);
1600 		}
1601 	}
1602 	free(str);
1603 
1604 null_event_list_str:
1605 	/*
1606 	 * config all kwork events if not specified
1607 	 */
1608 	if (list_empty(&kwork->class_list)) {
1609 		for (i = 0; i < KWORK_CLASS_MAX; i++) {
1610 			list_add_tail(&kwork_class_supported_list[i]->list,
1611 				      &kwork->class_list);
1612 		}
1613 	}
1614 
1615 	pr_debug("Config event list:");
1616 	list_for_each_entry(class, &kwork->class_list, list)
1617 		pr_debug(" %s", class->name);
1618 	pr_debug("\n");
1619 }
1620 
1621 static int perf_kwork__record(struct perf_kwork *kwork,
1622 			      int argc, const char **argv)
1623 {
1624 	const char **rec_argv;
1625 	unsigned int rec_argc, i, j;
1626 	struct kwork_class *class;
1627 
1628 	const char *const record_args[] = {
1629 		"record",
1630 		"-a",
1631 		"-R",
1632 		"-m", "1024",
1633 		"-c", "1",
1634 	};
1635 
1636 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1637 
1638 	list_for_each_entry(class, &kwork->class_list, list)
1639 		rec_argc += 2 * class->nr_tracepoints;
1640 
1641 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1642 	if (rec_argv == NULL)
1643 		return -ENOMEM;
1644 
1645 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1646 		rec_argv[i] = strdup(record_args[i]);
1647 
1648 	list_for_each_entry(class, &kwork->class_list, list) {
1649 		for (j = 0; j < class->nr_tracepoints; j++) {
1650 			rec_argv[i++] = strdup("-e");
1651 			rec_argv[i++] = strdup(class->tp_handlers[j].name);
1652 		}
1653 	}
1654 
1655 	for (j = 1; j < (unsigned int)argc; j++, i++)
1656 		rec_argv[i] = argv[j];
1657 
1658 	BUG_ON(i != rec_argc);
1659 
1660 	pr_debug("record comm: ");
1661 	for (j = 0; j < rec_argc; j++)
1662 		pr_debug("%s ", rec_argv[j]);
1663 	pr_debug("\n");
1664 
1665 	return cmd_record(i, rec_argv);
1666 }
1667 
1668 int cmd_kwork(int argc, const char **argv)
1669 {
1670 	static struct perf_kwork kwork = {
1671 		.class_list          = LIST_HEAD_INIT(kwork.class_list),
1672 		.tool = {
1673 			.mmap    = perf_event__process_mmap,
1674 			.mmap2   = perf_event__process_mmap2,
1675 			.sample  = perf_kwork__process_tracepoint_sample,
1676 		},
1677 		.atom_page_list      = LIST_HEAD_INIT(kwork.atom_page_list),
1678 		.sort_list           = LIST_HEAD_INIT(kwork.sort_list),
1679 		.cmp_id              = LIST_HEAD_INIT(kwork.cmp_id),
1680 		.sorted_work_root    = RB_ROOT_CACHED,
1681 		.tp_handler          = NULL,
1682 		.profile_name        = NULL,
1683 		.cpu_list            = NULL,
1684 		.time_str            = NULL,
1685 		.force               = false,
1686 		.event_list_str      = NULL,
1687 		.summary             = false,
1688 		.sort_order          = NULL,
1689 		.show_callchain      = false,
1690 		.max_stack           = 5,
1691 		.timestart           = 0,
1692 		.timeend             = 0,
1693 		.nr_events           = 0,
1694 		.nr_lost_chunks      = 0,
1695 		.nr_lost_events      = 0,
1696 		.all_runtime         = 0,
1697 		.all_count           = 0,
1698 		.nr_skipped_events   = { 0 },
1699 	};
1700 	static const char default_report_sort_order[] = "runtime, max, count";
1701 	static const char default_latency_sort_order[] = "avg, max, count";
1702 	const struct option kwork_options[] = {
1703 	OPT_INCR('v', "verbose", &verbose,
1704 		 "be more verbose (show symbol address, etc)"),
1705 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1706 		    "dump raw trace in ASCII"),
1707 	OPT_STRING('k', "kwork", &kwork.event_list_str, "kwork",
1708 		   "list of kwork to profile (irq, softirq, workqueue, etc)"),
1709 	OPT_BOOLEAN('f', "force", &kwork.force, "don't complain, do it"),
1710 	OPT_END()
1711 	};
1712 	const struct option report_options[] = {
1713 	OPT_STRING('s', "sort", &kwork.sort_order, "key[,key2...]",
1714 		   "sort by key(s): runtime, max, count"),
1715 	OPT_STRING('C', "cpu", &kwork.cpu_list, "cpu",
1716 		   "list of cpus to profile"),
1717 	OPT_STRING('n', "name", &kwork.profile_name, "name",
1718 		   "event name to profile"),
1719 	OPT_STRING(0, "time", &kwork.time_str, "str",
1720 		   "Time span for analysis (start,stop)"),
1721 	OPT_STRING('i', "input", &input_name, "file",
1722 		   "input file name"),
1723 	OPT_BOOLEAN('S', "with-summary", &kwork.summary,
1724 		    "Show summary with statistics"),
1725 #ifdef HAVE_BPF_SKEL
1726 	OPT_BOOLEAN('b', "use-bpf", &kwork.use_bpf,
1727 		    "Use BPF to measure kwork runtime"),
1728 #endif
1729 	OPT_PARENT(kwork_options)
1730 	};
1731 	const struct option latency_options[] = {
1732 	OPT_STRING('s', "sort", &kwork.sort_order, "key[,key2...]",
1733 		   "sort by key(s): avg, max, count"),
1734 	OPT_STRING('C', "cpu", &kwork.cpu_list, "cpu",
1735 		   "list of cpus to profile"),
1736 	OPT_STRING('n', "name", &kwork.profile_name, "name",
1737 		   "event name to profile"),
1738 	OPT_STRING(0, "time", &kwork.time_str, "str",
1739 		   "Time span for analysis (start,stop)"),
1740 	OPT_STRING('i', "input", &input_name, "file",
1741 		   "input file name"),
1742 #ifdef HAVE_BPF_SKEL
1743 	OPT_BOOLEAN('b', "use-bpf", &kwork.use_bpf,
1744 		    "Use BPF to measure kwork latency"),
1745 #endif
1746 	OPT_PARENT(kwork_options)
1747 	};
1748 	const struct option timehist_options[] = {
1749 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1750 		   "file", "vmlinux pathname"),
1751 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1752 		   "file", "kallsyms pathname"),
1753 	OPT_BOOLEAN('g', "call-graph", &kwork.show_callchain,
1754 		    "Display call chains if present"),
1755 	OPT_UINTEGER(0, "max-stack", &kwork.max_stack,
1756 		   "Maximum number of functions to display backtrace."),
1757 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
1758 		    "Look for files with symbols relative to this directory"),
1759 	OPT_STRING(0, "time", &kwork.time_str, "str",
1760 		   "Time span for analysis (start,stop)"),
1761 	OPT_STRING('C', "cpu", &kwork.cpu_list, "cpu",
1762 		   "list of cpus to profile"),
1763 	OPT_STRING('n', "name", &kwork.profile_name, "name",
1764 		   "event name to profile"),
1765 	OPT_STRING('i', "input", &input_name, "file",
1766 		   "input file name"),
1767 	OPT_PARENT(kwork_options)
1768 	};
1769 	const char *kwork_usage[] = {
1770 		NULL,
1771 		NULL
1772 	};
1773 	const char * const report_usage[] = {
1774 		"perf kwork report [<options>]",
1775 		NULL
1776 	};
1777 	const char * const latency_usage[] = {
1778 		"perf kwork latency [<options>]",
1779 		NULL
1780 	};
1781 	const char * const timehist_usage[] = {
1782 		"perf kwork timehist [<options>]",
1783 		NULL
1784 	};
1785 	const char *const kwork_subcommands[] = {
1786 		"record", "report", "latency", "timehist", NULL
1787 	};
1788 
1789 	argc = parse_options_subcommand(argc, argv, kwork_options,
1790 					kwork_subcommands, kwork_usage,
1791 					PARSE_OPT_STOP_AT_NON_OPTION);
1792 	if (!argc)
1793 		usage_with_options(kwork_usage, kwork_options);
1794 
1795 	setup_event_list(&kwork, kwork_options, kwork_usage);
1796 	sort_dimension__add(&kwork, "id", &kwork.cmp_id);
1797 
1798 	if (strlen(argv[0]) > 2 && strstarts("record", argv[0]))
1799 		return perf_kwork__record(&kwork, argc, argv);
1800 	else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) {
1801 		kwork.sort_order = default_report_sort_order;
1802 		if (argc > 1) {
1803 			argc = parse_options(argc, argv, report_options, report_usage, 0);
1804 			if (argc)
1805 				usage_with_options(report_usage, report_options);
1806 		}
1807 		kwork.report = KWORK_REPORT_RUNTIME;
1808 		setup_sorting(&kwork, report_options, report_usage);
1809 		return perf_kwork__report(&kwork);
1810 	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
1811 		kwork.sort_order = default_latency_sort_order;
1812 		if (argc > 1) {
1813 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1814 			if (argc)
1815 				usage_with_options(latency_usage, latency_options);
1816 		}
1817 		kwork.report = KWORK_REPORT_LATENCY;
1818 		setup_sorting(&kwork, latency_options, latency_usage);
1819 		return perf_kwork__report(&kwork);
1820 	} else if (strlen(argv[0]) > 2 && strstarts("timehist", argv[0])) {
1821 		if (argc > 1) {
1822 			argc = parse_options(argc, argv, timehist_options, timehist_usage, 0);
1823 			if (argc)
1824 				usage_with_options(timehist_usage, timehist_options);
1825 		}
1826 		kwork.report = KWORK_REPORT_TIMEHIST;
1827 		return perf_kwork__timehist(&kwork);
1828 	} else
1829 		usage_with_options(kwork_usage, kwork_options);
1830 
1831 	return 0;
1832 }
1833