Linux Audio

Check our new training course

Loading...
v3.1
   1#include "builtin.h"
   2#include "perf.h"
   3
   4#include "util/util.h"
 
   5#include "util/cache.h"
 
   6#include "util/symbol.h"
   7#include "util/thread.h"
   8#include "util/header.h"
   9#include "util/session.h"
 
 
  10
  11#include "util/parse-options.h"
  12#include "util/trace-event.h"
  13
  14#include "util/debug.h"
  15
  16#include <sys/prctl.h>
 
  17
  18#include <semaphore.h>
  19#include <pthread.h>
  20#include <math.h>
  21
  22static char			const *input_name = "perf.data";
  23
  24static char			default_sort_order[] = "avg, max, switch, runtime";
  25static const char		*sort_order = default_sort_order;
  26
  27static int			profile_cpu = -1;
  28
  29#define PR_SET_NAME		15               /* Set process name */
  30#define MAX_CPUS		4096
  31
  32static u64			run_measurement_overhead;
  33static u64			sleep_measurement_overhead;
  34
  35#define COMM_LEN		20
  36#define SYM_LEN			129
  37
  38#define MAX_PID			65536
  39
  40static unsigned long		nr_tasks;
  41
  42struct sched_atom;
  43
  44struct task_desc {
  45	unsigned long		nr;
  46	unsigned long		pid;
  47	char			comm[COMM_LEN];
  48
  49	unsigned long		nr_events;
  50	unsigned long		curr_event;
  51	struct sched_atom	**atoms;
  52
  53	pthread_t		thread;
  54	sem_t			sleep_sem;
  55
  56	sem_t			ready_for_work;
  57	sem_t			work_done_sem;
  58
  59	u64			cpu_usage;
  60};
  61
  62enum sched_event_type {
  63	SCHED_EVENT_RUN,
  64	SCHED_EVENT_SLEEP,
  65	SCHED_EVENT_WAKEUP,
  66	SCHED_EVENT_MIGRATION,
  67};
  68
  69struct sched_atom {
  70	enum sched_event_type	type;
  71	int			specific_wait;
  72	u64			timestamp;
  73	u64			duration;
  74	unsigned long		nr;
  75	sem_t			*wait_sem;
  76	struct task_desc	*wakee;
  77};
  78
  79static struct task_desc		*pid_to_task[MAX_PID];
  80
  81static struct task_desc		**tasks;
  82
  83static pthread_mutex_t		start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  84static u64			start_time;
  85
  86static pthread_mutex_t		work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  87
  88static unsigned long		nr_run_events;
  89static unsigned long		nr_sleep_events;
  90static unsigned long		nr_wakeup_events;
  91
  92static unsigned long		nr_sleep_corrections;
  93static unsigned long		nr_run_events_optimized;
  94
  95static unsigned long		targetless_wakeups;
  96static unsigned long		multitarget_wakeups;
  97
  98static u64			cpu_usage;
  99static u64			runavg_cpu_usage;
 100static u64			parent_cpu_usage;
 101static u64			runavg_parent_cpu_usage;
 102
 103static unsigned long		nr_runs;
 104static u64			sum_runtime;
 105static u64			sum_fluct;
 106static u64			run_avg;
 107
 108static unsigned int		replay_repeat = 10;
 109static unsigned long		nr_timestamps;
 110static unsigned long		nr_unordered_timestamps;
 111static unsigned long		nr_state_machine_bugs;
 112static unsigned long		nr_context_switch_bugs;
 113static unsigned long		nr_events;
 114static unsigned long		nr_lost_chunks;
 115static unsigned long		nr_lost_events;
 116
 117#define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 118
 119enum thread_state {
 120	THREAD_SLEEPING = 0,
 121	THREAD_WAIT_CPU,
 122	THREAD_SCHED_IN,
 123	THREAD_IGNORE
 124};
 125
 126struct work_atom {
 127	struct list_head	list;
 128	enum thread_state	state;
 129	u64			sched_out_time;
 130	u64			wake_up_time;
 131	u64			sched_in_time;
 132	u64			runtime;
 133};
 134
 135struct work_atoms {
 136	struct list_head	work_list;
 137	struct thread		*thread;
 138	struct rb_node		node;
 139	u64			max_lat;
 140	u64			max_lat_at;
 141	u64			total_lat;
 142	u64			nb_atoms;
 143	u64			total_runtime;
 
 144};
 145
 146typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 147
 148static struct rb_root		atom_root, sorted_atom_root;
 
 
 
 
 
 
 
 149
 150static u64			all_runtime;
 151static u64			all_count;
 152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 153
 154static u64 get_nsecs(void)
 155{
 156	struct timespec ts;
 157
 158	clock_gettime(CLOCK_MONOTONIC, &ts);
 159
 160	return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 161}
 162
 163static void burn_nsecs(u64 nsecs)
 164{
 165	u64 T0 = get_nsecs(), T1;
 166
 167	do {
 168		T1 = get_nsecs();
 169	} while (T1 + run_measurement_overhead < T0 + nsecs);
 170}
 171
 172static void sleep_nsecs(u64 nsecs)
 173{
 174	struct timespec ts;
 175
 176	ts.tv_nsec = nsecs % 999999999;
 177	ts.tv_sec = nsecs / 999999999;
 178
 179	nanosleep(&ts, NULL);
 180}
 181
 182static void calibrate_run_measurement_overhead(void)
 183{
 184	u64 T0, T1, delta, min_delta = 1000000000ULL;
 185	int i;
 186
 187	for (i = 0; i < 10; i++) {
 188		T0 = get_nsecs();
 189		burn_nsecs(0);
 190		T1 = get_nsecs();
 191		delta = T1-T0;
 192		min_delta = min(min_delta, delta);
 193	}
 194	run_measurement_overhead = min_delta;
 195
 196	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 197}
 198
 199static void calibrate_sleep_measurement_overhead(void)
 200{
 201	u64 T0, T1, delta, min_delta = 1000000000ULL;
 202	int i;
 203
 204	for (i = 0; i < 10; i++) {
 205		T0 = get_nsecs();
 206		sleep_nsecs(10000);
 207		T1 = get_nsecs();
 208		delta = T1-T0;
 209		min_delta = min(min_delta, delta);
 210	}
 211	min_delta -= 10000;
 212	sleep_measurement_overhead = min_delta;
 213
 214	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 215}
 216
 217static struct sched_atom *
 218get_new_event(struct task_desc *task, u64 timestamp)
 219{
 220	struct sched_atom *event = zalloc(sizeof(*event));
 221	unsigned long idx = task->nr_events;
 222	size_t size;
 223
 224	event->timestamp = timestamp;
 225	event->nr = idx;
 226
 227	task->nr_events++;
 228	size = sizeof(struct sched_atom *) * task->nr_events;
 229	task->atoms = realloc(task->atoms, size);
 230	BUG_ON(!task->atoms);
 231
 232	task->atoms[idx] = event;
 233
 234	return event;
 235}
 236
 237static struct sched_atom *last_event(struct task_desc *task)
 238{
 239	if (!task->nr_events)
 240		return NULL;
 241
 242	return task->atoms[task->nr_events - 1];
 243}
 244
 245static void
 246add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 247{
 248	struct sched_atom *event, *curr_event = last_event(task);
 249
 250	/*
 251	 * optimize an existing RUN event by merging this one
 252	 * to it:
 253	 */
 254	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 255		nr_run_events_optimized++;
 256		curr_event->duration += duration;
 257		return;
 258	}
 259
 260	event = get_new_event(task, timestamp);
 261
 262	event->type = SCHED_EVENT_RUN;
 263	event->duration = duration;
 264
 265	nr_run_events++;
 266}
 267
 268static void
 269add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 270		       struct task_desc *wakee)
 271{
 272	struct sched_atom *event, *wakee_event;
 273
 274	event = get_new_event(task, timestamp);
 275	event->type = SCHED_EVENT_WAKEUP;
 276	event->wakee = wakee;
 277
 278	wakee_event = last_event(wakee);
 279	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 280		targetless_wakeups++;
 281		return;
 282	}
 283	if (wakee_event->wait_sem) {
 284		multitarget_wakeups++;
 285		return;
 286	}
 287
 288	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 289	sem_init(wakee_event->wait_sem, 0, 0);
 290	wakee_event->specific_wait = 1;
 291	event->wait_sem = wakee_event->wait_sem;
 292
 293	nr_wakeup_events++;
 294}
 295
 296static void
 297add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 298		      u64 task_state __used)
 299{
 300	struct sched_atom *event = get_new_event(task, timestamp);
 301
 302	event->type = SCHED_EVENT_SLEEP;
 303
 304	nr_sleep_events++;
 305}
 306
 307static struct task_desc *register_pid(unsigned long pid, const char *comm)
 
 308{
 309	struct task_desc *task;
 
 310
 311	BUG_ON(pid >= MAX_PID);
 
 
 
 
 
 
 
 
 
 
 312
 313	task = pid_to_task[pid];
 314
 315	if (task)
 316		return task;
 317
 318	task = zalloc(sizeof(*task));
 319	task->pid = pid;
 320	task->nr = nr_tasks;
 321	strcpy(task->comm, comm);
 322	/*
 323	 * every task starts in sleeping state - this gets ignored
 324	 * if there's no wakeup pointing to this sleep state:
 325	 */
 326	add_sched_event_sleep(task, 0, 0);
 327
 328	pid_to_task[pid] = task;
 329	nr_tasks++;
 330	tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 331	BUG_ON(!tasks);
 332	tasks[task->nr] = task;
 333
 334	if (verbose)
 335		printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 336
 337	return task;
 338}
 339
 340
 341static void print_task_traces(void)
 342{
 343	struct task_desc *task;
 344	unsigned long i;
 345
 346	for (i = 0; i < nr_tasks; i++) {
 347		task = tasks[i];
 348		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 349			task->nr, task->comm, task->pid, task->nr_events);
 350	}
 351}
 352
 353static void add_cross_task_wakeups(void)
 354{
 355	struct task_desc *task1, *task2;
 356	unsigned long i, j;
 357
 358	for (i = 0; i < nr_tasks; i++) {
 359		task1 = tasks[i];
 360		j = i + 1;
 361		if (j == nr_tasks)
 362			j = 0;
 363		task2 = tasks[j];
 364		add_sched_event_wakeup(task1, 0, task2);
 365	}
 366}
 367
 368static void
 369process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 370{
 371	int ret = 0;
 372
 373	switch (atom->type) {
 374		case SCHED_EVENT_RUN:
 375			burn_nsecs(atom->duration);
 376			break;
 377		case SCHED_EVENT_SLEEP:
 378			if (atom->wait_sem)
 379				ret = sem_wait(atom->wait_sem);
 380			BUG_ON(ret);
 381			break;
 382		case SCHED_EVENT_WAKEUP:
 383			if (atom->wait_sem)
 384				ret = sem_post(atom->wait_sem);
 385			BUG_ON(ret);
 386			break;
 387		case SCHED_EVENT_MIGRATION:
 388			break;
 389		default:
 390			BUG_ON(1);
 391	}
 392}
 393
 394static u64 get_cpu_usage_nsec_parent(void)
 395{
 396	struct rusage ru;
 397	u64 sum;
 398	int err;
 399
 400	err = getrusage(RUSAGE_SELF, &ru);
 401	BUG_ON(err);
 402
 403	sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 404	sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 405
 406	return sum;
 407}
 408
 409static int self_open_counters(void)
 410{
 411	struct perf_event_attr attr;
 
 412	int fd;
 
 
 413
 414	memset(&attr, 0, sizeof(attr));
 415
 416	attr.type = PERF_TYPE_SOFTWARE;
 417	attr.config = PERF_COUNT_SW_TASK_CLOCK;
 418
 419	fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 420
 421	if (fd < 0)
 422		die("Error: sys_perf_event_open() syscall returned"
 423		    "with %d (%s)\n", fd, strerror(errno));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 424	return fd;
 425}
 426
 427static u64 get_cpu_usage_nsec_self(int fd)
 428{
 429	u64 runtime;
 430	int ret;
 431
 432	ret = read(fd, &runtime, sizeof(runtime));
 433	BUG_ON(ret != sizeof(runtime));
 434
 435	return runtime;
 436}
 437
 
 
 
 
 
 
 438static void *thread_func(void *ctx)
 439{
 440	struct task_desc *this_task = ctx;
 
 
 441	u64 cpu_usage_0, cpu_usage_1;
 442	unsigned long i, ret;
 443	char comm2[22];
 444	int fd;
 
 
 445
 446	sprintf(comm2, ":%s", this_task->comm);
 447	prctl(PR_SET_NAME, comm2);
 448	fd = self_open_counters();
 449
 450again:
 451	ret = sem_post(&this_task->ready_for_work);
 452	BUG_ON(ret);
 453	ret = pthread_mutex_lock(&start_work_mutex);
 454	BUG_ON(ret);
 455	ret = pthread_mutex_unlock(&start_work_mutex);
 456	BUG_ON(ret);
 457
 458	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 459
 460	for (i = 0; i < this_task->nr_events; i++) {
 461		this_task->curr_event = i;
 462		process_sched_event(this_task, this_task->atoms[i]);
 463	}
 464
 465	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 466	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 467	ret = sem_post(&this_task->work_done_sem);
 468	BUG_ON(ret);
 469
 470	ret = pthread_mutex_lock(&work_done_wait_mutex);
 471	BUG_ON(ret);
 472	ret = pthread_mutex_unlock(&work_done_wait_mutex);
 473	BUG_ON(ret);
 474
 475	goto again;
 476}
 477
 478static void create_tasks(void)
 479{
 480	struct task_desc *task;
 481	pthread_attr_t attr;
 482	unsigned long i;
 483	int err;
 484
 485	err = pthread_attr_init(&attr);
 486	BUG_ON(err);
 487	err = pthread_attr_setstacksize(&attr,
 488			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 489	BUG_ON(err);
 490	err = pthread_mutex_lock(&start_work_mutex);
 491	BUG_ON(err);
 492	err = pthread_mutex_lock(&work_done_wait_mutex);
 493	BUG_ON(err);
 494	for (i = 0; i < nr_tasks; i++) {
 495		task = tasks[i];
 
 
 
 
 496		sem_init(&task->sleep_sem, 0, 0);
 497		sem_init(&task->ready_for_work, 0, 0);
 498		sem_init(&task->work_done_sem, 0, 0);
 499		task->curr_event = 0;
 500		err = pthread_create(&task->thread, &attr, thread_func, task);
 501		BUG_ON(err);
 502	}
 503}
 504
 505static void wait_for_tasks(void)
 506{
 507	u64 cpu_usage_0, cpu_usage_1;
 508	struct task_desc *task;
 509	unsigned long i, ret;
 510
 511	start_time = get_nsecs();
 512	cpu_usage = 0;
 513	pthread_mutex_unlock(&work_done_wait_mutex);
 514
 515	for (i = 0; i < nr_tasks; i++) {
 516		task = tasks[i];
 517		ret = sem_wait(&task->ready_for_work);
 518		BUG_ON(ret);
 519		sem_init(&task->ready_for_work, 0, 0);
 520	}
 521	ret = pthread_mutex_lock(&work_done_wait_mutex);
 522	BUG_ON(ret);
 523
 524	cpu_usage_0 = get_cpu_usage_nsec_parent();
 525
 526	pthread_mutex_unlock(&start_work_mutex);
 527
 528	for (i = 0; i < nr_tasks; i++) {
 529		task = tasks[i];
 530		ret = sem_wait(&task->work_done_sem);
 531		BUG_ON(ret);
 532		sem_init(&task->work_done_sem, 0, 0);
 533		cpu_usage += task->cpu_usage;
 534		task->cpu_usage = 0;
 535	}
 536
 537	cpu_usage_1 = get_cpu_usage_nsec_parent();
 538	if (!runavg_cpu_usage)
 539		runavg_cpu_usage = cpu_usage;
 540	runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 541
 542	parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 543	if (!runavg_parent_cpu_usage)
 544		runavg_parent_cpu_usage = parent_cpu_usage;
 545	runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 546				   parent_cpu_usage)/10;
 547
 548	ret = pthread_mutex_lock(&start_work_mutex);
 549	BUG_ON(ret);
 550
 551	for (i = 0; i < nr_tasks; i++) {
 552		task = tasks[i];
 553		sem_init(&task->sleep_sem, 0, 0);
 554		task->curr_event = 0;
 555	}
 556}
 557
 558static void run_one_test(void)
 559{
 560	u64 T0, T1, delta, avg_delta, fluct;
 561
 562	T0 = get_nsecs();
 563	wait_for_tasks();
 564	T1 = get_nsecs();
 565
 566	delta = T1 - T0;
 567	sum_runtime += delta;
 568	nr_runs++;
 569
 570	avg_delta = sum_runtime / nr_runs;
 571	if (delta < avg_delta)
 572		fluct = avg_delta - delta;
 573	else
 574		fluct = delta - avg_delta;
 575	sum_fluct += fluct;
 576	if (!run_avg)
 577		run_avg = delta;
 578	run_avg = (run_avg*9 + delta)/10;
 579
 580	printf("#%-3ld: %0.3f, ",
 581		nr_runs, (double)delta/1000000.0);
 582
 583	printf("ravg: %0.2f, ",
 584		(double)run_avg/1e6);
 585
 586	printf("cpu: %0.2f / %0.2f",
 587		(double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 588
 589#if 0
 590	/*
 591	 * rusage statistics done by the parent, these are less
 592	 * accurate than the sum_exec_runtime based statistics:
 593	 */
 594	printf(" [%0.2f / %0.2f]",
 595		(double)parent_cpu_usage/1e6,
 596		(double)runavg_parent_cpu_usage/1e6);
 597#endif
 598
 599	printf("\n");
 600
 601	if (nr_sleep_corrections)
 602		printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 603	nr_sleep_corrections = 0;
 604}
 605
 606static void test_calibrations(void)
 607{
 608	u64 T0, T1;
 609
 610	T0 = get_nsecs();
 611	burn_nsecs(1e6);
 612	T1 = get_nsecs();
 613
 614	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 615
 616	T0 = get_nsecs();
 617	sleep_nsecs(1e6);
 618	T1 = get_nsecs();
 619
 620	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 621}
 622
 623#define FILL_FIELD(ptr, field, event, data)	\
 624	ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 625
 626#define FILL_ARRAY(ptr, array, event, data)			\
 627do {								\
 628	void *__array = raw_field_ptr(event, #array, data);	\
 629	memcpy(ptr.array, __array, sizeof(ptr.array));	\
 630} while(0)
 631
 632#define FILL_COMMON_FIELDS(ptr, event, data)			\
 633do {								\
 634	FILL_FIELD(ptr, common_type, event, data);		\
 635	FILL_FIELD(ptr, common_flags, event, data);		\
 636	FILL_FIELD(ptr, common_preempt_count, event, data);	\
 637	FILL_FIELD(ptr, common_pid, event, data);		\
 638	FILL_FIELD(ptr, common_tgid, event, data);		\
 639} while (0)
 640
 641
 642
 643struct trace_switch_event {
 644	u32 size;
 645
 646	u16 common_type;
 647	u8 common_flags;
 648	u8 common_preempt_count;
 649	u32 common_pid;
 650	u32 common_tgid;
 651
 652	char prev_comm[16];
 653	u32 prev_pid;
 654	u32 prev_prio;
 655	u64 prev_state;
 656	char next_comm[16];
 657	u32 next_pid;
 658	u32 next_prio;
 659};
 660
 661struct trace_runtime_event {
 662	u32 size;
 663
 664	u16 common_type;
 665	u8 common_flags;
 666	u8 common_preempt_count;
 667	u32 common_pid;
 668	u32 common_tgid;
 669
 670	char comm[16];
 671	u32 pid;
 672	u64 runtime;
 673	u64 vruntime;
 674};
 675
 676struct trace_wakeup_event {
 677	u32 size;
 678
 679	u16 common_type;
 680	u8 common_flags;
 681	u8 common_preempt_count;
 682	u32 common_pid;
 683	u32 common_tgid;
 684
 685	char comm[16];
 686	u32 pid;
 687
 688	u32 prio;
 689	u32 success;
 690	u32 cpu;
 691};
 692
 693struct trace_fork_event {
 694	u32 size;
 695
 696	u16 common_type;
 697	u8 common_flags;
 698	u8 common_preempt_count;
 699	u32 common_pid;
 700	u32 common_tgid;
 701
 702	char parent_comm[16];
 703	u32 parent_pid;
 704	char child_comm[16];
 705	u32 child_pid;
 706};
 707
 708struct trace_migrate_task_event {
 709	u32 size;
 710
 711	u16 common_type;
 712	u8 common_flags;
 713	u8 common_preempt_count;
 714	u32 common_pid;
 715	u32 common_tgid;
 716
 717	char comm[16];
 718	u32 pid;
 719
 720	u32 prio;
 721	u32 cpu;
 722};
 723
 724struct trace_sched_handler {
 725	void (*switch_event)(struct trace_switch_event *,
 726			     struct perf_session *,
 727			     struct event *,
 728			     int cpu,
 729			     u64 timestamp,
 730			     struct thread *thread);
 731
 732	void (*runtime_event)(struct trace_runtime_event *,
 733			      struct perf_session *,
 734			      struct event *,
 735			      int cpu,
 736			      u64 timestamp,
 737			      struct thread *thread);
 738
 739	void (*wakeup_event)(struct trace_wakeup_event *,
 740			     struct perf_session *,
 741			     struct event *,
 742			     int cpu,
 743			     u64 timestamp,
 744			     struct thread *thread);
 745
 746	void (*fork_event)(struct trace_fork_event *,
 747			   struct event *,
 748			   int cpu,
 749			   u64 timestamp,
 750			   struct thread *thread);
 751
 752	void (*migrate_task_event)(struct trace_migrate_task_event *,
 753			   struct perf_session *session,
 754			   struct event *,
 755			   int cpu,
 756			   u64 timestamp,
 757			   struct thread *thread);
 758};
 759
 760
 761static void
 762replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 763		    struct perf_session *session __used,
 764		    struct event *event,
 765		    int cpu __used,
 766		    u64 timestamp __used,
 767		    struct thread *thread __used)
 768{
 
 
 769	struct task_desc *waker, *wakee;
 770
 771	if (verbose) {
 772		printf("sched_wakeup event %p\n", event);
 773
 774		printf(" ... pid %d woke up %s/%d\n",
 775			wakeup_event->common_pid,
 776			wakeup_event->comm,
 777			wakeup_event->pid);
 778	}
 779
 780	waker = register_pid(wakeup_event->common_pid, "<unknown>");
 781	wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 782
 783	add_sched_event_wakeup(waker, timestamp, wakee);
 
 784}
 785
 786static u64 cpu_last_switched[MAX_CPUS];
 787
 788static void
 789replay_switch_event(struct trace_switch_event *switch_event,
 790		    struct perf_session *session __used,
 791		    struct event *event,
 792		    int cpu,
 793		    u64 timestamp,
 794		    struct thread *thread __used)
 795{
 796	struct task_desc *prev, __used *next;
 797	u64 timestamp0;
 
 798	s64 delta;
 799
 800	if (verbose)
 801		printf("sched_switch event %p\n", event);
 802
 803	if (cpu >= MAX_CPUS || cpu < 0)
 804		return;
 805
 806	timestamp0 = cpu_last_switched[cpu];
 807	if (timestamp0)
 808		delta = timestamp - timestamp0;
 809	else
 810		delta = 0;
 811
 812	if (delta < 0)
 813		die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 814
 815	if (verbose) {
 816		printf(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 817			switch_event->prev_comm, switch_event->prev_pid,
 818			switch_event->next_comm, switch_event->next_pid,
 819			delta);
 820	}
 821
 822	prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 823	next = register_pid(switch_event->next_pid, switch_event->next_comm);
 
 
 
 
 
 824
 825	cpu_last_switched[cpu] = timestamp;
 
 826
 827	add_sched_event_run(prev, timestamp, delta);
 828	add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 829}
 830
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 831
 832static void
 833replay_fork_event(struct trace_fork_event *fork_event,
 834		  struct event *event,
 835		  int cpu __used,
 836		  u64 timestamp __used,
 837		  struct thread *thread __used)
 838{
 839	if (verbose) {
 840		printf("sched_fork event %p\n", event);
 841		printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 842		printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 843	}
 844	register_pid(fork_event->parent_pid, fork_event->parent_comm);
 845	register_pid(fork_event->child_pid, fork_event->child_comm);
 846}
 847
 848static struct trace_sched_handler replay_ops  = {
 849	.wakeup_event		= replay_wakeup_event,
 850	.switch_event		= replay_switch_event,
 851	.fork_event		= replay_fork_event,
 852};
 
 
 853
 854struct sort_dimension {
 855	const char		*name;
 856	sort_fn_t		cmp;
 857	struct list_head	list;
 858};
 859
 860static LIST_HEAD(cmp_pid);
 861
 862static int
 863thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 864{
 865	struct sort_dimension *sort;
 866	int ret = 0;
 867
 868	BUG_ON(list_empty(list));
 869
 870	list_for_each_entry(sort, list, list) {
 871		ret = sort->cmp(l, r);
 872		if (ret)
 873			return ret;
 874	}
 875
 876	return ret;
 877}
 878
 879static struct work_atoms *
 880thread_atoms_search(struct rb_root *root, struct thread *thread,
 881			 struct list_head *sort_list)
 882{
 883	struct rb_node *node = root->rb_node;
 884	struct work_atoms key = { .thread = thread };
 885
 886	while (node) {
 887		struct work_atoms *atoms;
 888		int cmp;
 889
 890		atoms = container_of(node, struct work_atoms, node);
 891
 892		cmp = thread_lat_cmp(sort_list, &key, atoms);
 893		if (cmp > 0)
 894			node = node->rb_left;
 895		else if (cmp < 0)
 896			node = node->rb_right;
 897		else {
 898			BUG_ON(thread != atoms->thread);
 899			return atoms;
 900		}
 901	}
 902	return NULL;
 903}
 904
 905static void
 906__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 907			 struct list_head *sort_list)
 908{
 909	struct rb_node **new = &(root->rb_node), *parent = NULL;
 910
 911	while (*new) {
 912		struct work_atoms *this;
 913		int cmp;
 914
 915		this = container_of(*new, struct work_atoms, node);
 916		parent = *new;
 917
 918		cmp = thread_lat_cmp(sort_list, data, this);
 919
 920		if (cmp > 0)
 921			new = &((*new)->rb_left);
 922		else
 923			new = &((*new)->rb_right);
 924	}
 925
 926	rb_link_node(&data->node, parent, new);
 927	rb_insert_color(&data->node, root);
 928}
 929
 930static void thread_atoms_insert(struct thread *thread)
 931{
 932	struct work_atoms *atoms = zalloc(sizeof(*atoms));
 933	if (!atoms)
 934		die("No memory");
 
 
 935
 936	atoms->thread = thread;
 937	INIT_LIST_HEAD(&atoms->work_list);
 938	__thread_latency_insert(&atom_root, atoms, &cmp_pid);
 939}
 940
 941static void
 942latency_fork_event(struct trace_fork_event *fork_event __used,
 943		   struct event *event __used,
 944		   int cpu __used,
 945		   u64 timestamp __used,
 946		   struct thread *thread __used)
 947{
 948	/* should insert the newcomer */
 949}
 950
 951__used
 952static char sched_out_state(struct trace_switch_event *switch_event)
 953{
 954	const char *str = TASK_STATE_TO_CHAR_STR;
 955
 956	return str[switch_event->prev_state];
 957}
 958
 959static void
 960add_sched_out_event(struct work_atoms *atoms,
 961		    char run_state,
 962		    u64 timestamp)
 963{
 964	struct work_atom *atom = zalloc(sizeof(*atom));
 965	if (!atom)
 966		die("Non memory");
 
 
 967
 968	atom->sched_out_time = timestamp;
 969
 970	if (run_state == 'R') {
 971		atom->state = THREAD_WAIT_CPU;
 972		atom->wake_up_time = atom->sched_out_time;
 973	}
 974
 975	list_add_tail(&atom->list, &atoms->work_list);
 
 976}
 977
 978static void
 979add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
 
 980{
 981	struct work_atom *atom;
 982
 983	BUG_ON(list_empty(&atoms->work_list));
 984
 985	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 986
 987	atom->runtime += delta;
 988	atoms->total_runtime += delta;
 989}
 990
 991static void
 992add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 993{
 994	struct work_atom *atom;
 995	u64 delta;
 996
 997	if (list_empty(&atoms->work_list))
 998		return;
 999
1000	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1001
1002	if (atom->state != THREAD_WAIT_CPU)
1003		return;
1004
1005	if (timestamp < atom->wake_up_time) {
1006		atom->state = THREAD_IGNORE;
1007		return;
1008	}
1009
1010	atom->state = THREAD_SCHED_IN;
1011	atom->sched_in_time = timestamp;
1012
1013	delta = atom->sched_in_time - atom->wake_up_time;
1014	atoms->total_lat += delta;
1015	if (delta > atoms->max_lat) {
1016		atoms->max_lat = delta;
1017		atoms->max_lat_at = timestamp;
1018	}
1019	atoms->nb_atoms++;
1020}
1021
1022static void
1023latency_switch_event(struct trace_switch_event *switch_event,
1024		     struct perf_session *session,
1025		     struct event *event __used,
1026		     int cpu,
1027		     u64 timestamp,
1028		     struct thread *thread __used)
1029{
 
 
 
1030	struct work_atoms *out_events, *in_events;
1031	struct thread *sched_out, *sched_in;
1032	u64 timestamp0;
 
1033	s64 delta;
1034
1035	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1036
1037	timestamp0 = cpu_last_switched[cpu];
1038	cpu_last_switched[cpu] = timestamp;
1039	if (timestamp0)
1040		delta = timestamp - timestamp0;
1041	else
1042		delta = 0;
1043
1044	if (delta < 0)
1045		die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1046
 
1047
1048	sched_out = perf_session__findnew(session, switch_event->prev_pid);
1049	sched_in = perf_session__findnew(session, switch_event->next_pid);
 
 
1050
1051	out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1052	if (!out_events) {
1053		thread_atoms_insert(sched_out);
1054		out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1055		if (!out_events)
1056			die("out-event: Internal tree error");
 
 
 
1057	}
1058	add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
 
1059
1060	in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1061	if (!in_events) {
1062		thread_atoms_insert(sched_in);
1063		in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1064		if (!in_events)
1065			die("in-event: Internal tree error");
 
 
 
1066		/*
1067		 * Take came in we have not heard about yet,
1068		 * add in an initial atom in runnable state:
1069		 */
1070		add_sched_out_event(in_events, 'R', timestamp);
 
1071	}
1072	add_sched_in_event(in_events, timestamp);
1073}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1074
1075static void
1076latency_runtime_event(struct trace_runtime_event *runtime_event,
1077		     struct perf_session *session,
1078		     struct event *event __used,
1079		     int cpu,
1080		     u64 timestamp,
1081		     struct thread *this_thread __used)
1082{
1083	struct thread *thread = perf_session__findnew(session, runtime_event->pid);
1084	struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1085
1086	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1087	if (!atoms) {
1088		thread_atoms_insert(thread);
1089		atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1090		if (!atoms)
1091			die("in-event: Internal tree error");
1092		add_sched_out_event(atoms, 'R', timestamp);
 
 
 
 
1093	}
1094
1095	add_runtime_event(atoms, runtime_event->runtime, timestamp);
 
 
 
 
1096}
1097
1098static void
1099latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1100		     struct perf_session *session,
1101		     struct event *__event __used,
1102		     int cpu __used,
1103		     u64 timestamp,
1104		     struct thread *thread __used)
1105{
 
1106	struct work_atoms *atoms;
1107	struct work_atom *atom;
1108	struct thread *wakee;
 
 
1109
1110	/* Note for later, it may be interesting to observe the failing cases */
1111	if (!wakeup_event->success)
1112		return;
1113
1114	wakee = perf_session__findnew(session, wakeup_event->pid);
1115	atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1116	if (!atoms) {
1117		thread_atoms_insert(wakee);
1118		atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1119		if (!atoms)
1120			die("wakeup-event: Internal tree error");
1121		add_sched_out_event(atoms, 'S', timestamp);
 
 
 
 
1122	}
1123
1124	BUG_ON(list_empty(&atoms->work_list));
1125
1126	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1127
1128	/*
 
 
 
 
 
 
1129	 * You WILL be missing events if you've recorded only
1130	 * one CPU, or are only looking at only one, so don't
1131	 * make useless noise.
1132	 */
1133	if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1134		nr_state_machine_bugs++;
1135
1136	nr_timestamps++;
1137	if (atom->sched_out_time > timestamp) {
1138		nr_unordered_timestamps++;
1139		return;
1140	}
1141
1142	atom->state = THREAD_WAIT_CPU;
1143	atom->wake_up_time = timestamp;
 
 
 
 
 
1144}
1145
1146static void
1147latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1148		     struct perf_session *session,
1149		     struct event *__event __used,
1150		     int cpu __used,
1151		     u64 timestamp,
1152		     struct thread *thread __used)
1153{
 
 
1154	struct work_atoms *atoms;
1155	struct work_atom *atom;
1156	struct thread *migrant;
 
1157
1158	/*
1159	 * Only need to worry about migration when profiling one CPU.
1160	 */
1161	if (profile_cpu == -1)
1162		return;
1163
1164	migrant = perf_session__findnew(session, migrate_task_event->pid);
1165	atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
 
 
1166	if (!atoms) {
1167		thread_atoms_insert(migrant);
1168		register_pid(migrant->pid, migrant->comm);
1169		atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1170		if (!atoms)
1171			die("migration-event: Internal tree error");
1172		add_sched_out_event(atoms, 'R', timestamp);
 
 
 
 
1173	}
1174
1175	BUG_ON(list_empty(&atoms->work_list));
1176
1177	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1178	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1179
1180	nr_timestamps++;
1181
1182	if (atom->sched_out_time > timestamp)
1183		nr_unordered_timestamps++;
 
 
 
 
1184}
1185
1186static struct trace_sched_handler lat_ops  = {
1187	.wakeup_event		= latency_wakeup_event,
1188	.switch_event		= latency_switch_event,
1189	.runtime_event		= latency_runtime_event,
1190	.fork_event		= latency_fork_event,
1191	.migrate_task_event	= latency_migrate_task_event,
1192};
1193
1194static void output_lat_thread(struct work_atoms *work_list)
1195{
1196	int i;
1197	int ret;
1198	u64 avg;
1199
1200	if (!work_list->nb_atoms)
1201		return;
1202	/*
1203	 * Ignore idle threads:
1204	 */
1205	if (!strcmp(work_list->thread->comm, "swapper"))
1206		return;
1207
1208	all_runtime += work_list->total_runtime;
1209	all_count += work_list->nb_atoms;
1210
1211	ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
 
 
 
1212
1213	for (i = 0; i < 24 - ret; i++)
1214		printf(" ");
1215
1216	avg = work_list->total_lat / work_list->nb_atoms;
1217
1218	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1219	      (double)work_list->total_runtime / 1e6,
1220		 work_list->nb_atoms, (double)avg / 1e6,
1221		 (double)work_list->max_lat / 1e6,
1222		 (double)work_list->max_lat_at / 1e9);
1223}
1224
1225static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1226{
1227	if (l->thread->pid < r->thread->pid)
 
 
1228		return -1;
1229	if (l->thread->pid > r->thread->pid)
1230		return 1;
1231
1232	return 0;
1233}
1234
1235static struct sort_dimension pid_sort_dimension = {
1236	.name			= "pid",
1237	.cmp			= pid_cmp,
1238};
1239
1240static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1241{
1242	u64 avgl, avgr;
1243
1244	if (!l->nb_atoms)
1245		return -1;
1246
1247	if (!r->nb_atoms)
1248		return 1;
1249
1250	avgl = l->total_lat / l->nb_atoms;
1251	avgr = r->total_lat / r->nb_atoms;
1252
1253	if (avgl < avgr)
1254		return -1;
1255	if (avgl > avgr)
1256		return 1;
1257
1258	return 0;
1259}
1260
1261static struct sort_dimension avg_sort_dimension = {
1262	.name			= "avg",
1263	.cmp			= avg_cmp,
1264};
1265
1266static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1267{
1268	if (l->max_lat < r->max_lat)
1269		return -1;
1270	if (l->max_lat > r->max_lat)
1271		return 1;
1272
1273	return 0;
1274}
1275
1276static struct sort_dimension max_sort_dimension = {
1277	.name			= "max",
1278	.cmp			= max_cmp,
1279};
1280
1281static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1282{
1283	if (l->nb_atoms < r->nb_atoms)
1284		return -1;
1285	if (l->nb_atoms > r->nb_atoms)
1286		return 1;
1287
1288	return 0;
1289}
1290
1291static struct sort_dimension switch_sort_dimension = {
1292	.name			= "switch",
1293	.cmp			= switch_cmp,
1294};
1295
1296static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1297{
1298	if (l->total_runtime < r->total_runtime)
1299		return -1;
1300	if (l->total_runtime > r->total_runtime)
1301		return 1;
1302
1303	return 0;
1304}
1305
1306static struct sort_dimension runtime_sort_dimension = {
1307	.name			= "runtime",
1308	.cmp			= runtime_cmp,
1309};
1310
1311static struct sort_dimension *available_sorts[] = {
1312	&pid_sort_dimension,
1313	&avg_sort_dimension,
1314	&max_sort_dimension,
1315	&switch_sort_dimension,
1316	&runtime_sort_dimension,
1317};
1318
1319#define NB_AVAILABLE_SORTS	(int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1320
1321static LIST_HEAD(sort_list);
1322
1323static int sort_dimension__add(const char *tok, struct list_head *list)
1324{
1325	int i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1326
1327	for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1328		if (!strcmp(available_sorts[i]->name, tok)) {
1329			list_add_tail(&available_sorts[i]->list, list);
1330
1331			return 0;
1332		}
1333	}
1334
1335	return -1;
1336}
1337
1338static void setup_sorting(void);
1339
1340static void sort_lat(void)
1341{
1342	struct rb_node *node;
1343
 
1344	for (;;) {
1345		struct work_atoms *data;
1346		node = rb_first(&atom_root);
1347		if (!node)
1348			break;
1349
1350		rb_erase(node, &atom_root);
1351		data = rb_entry(node, struct work_atoms, node);
1352		__thread_latency_insert(&sorted_atom_root, data, &sort_list);
 
 
 
 
1353	}
1354}
1355
1356static struct trace_sched_handler *trace_handler;
1357
1358static void
1359process_sched_wakeup_event(void *data, struct perf_session *session,
1360			   struct event *event,
1361			   int cpu __used,
1362			   u64 timestamp __used,
1363			   struct thread *thread __used)
1364{
1365	struct trace_wakeup_event wakeup_event;
1366
1367	FILL_COMMON_FIELDS(wakeup_event, event, data);
1368
1369	FILL_ARRAY(wakeup_event, comm, event, data);
1370	FILL_FIELD(wakeup_event, pid, event, data);
1371	FILL_FIELD(wakeup_event, prio, event, data);
1372	FILL_FIELD(wakeup_event, success, event, data);
1373	FILL_FIELD(wakeup_event, cpu, event, data);
1374
1375	if (trace_handler->wakeup_event)
1376		trace_handler->wakeup_event(&wakeup_event, session, event,
1377					    cpu, timestamp, thread);
1378}
1379
1380/*
1381 * Track the current task - that way we can know whether there's any
1382 * weird events, such as a task being switched away that is not current.
1383 */
1384static int max_cpu;
1385
1386static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1387
1388static struct thread *curr_thread[MAX_CPUS];
 
1389
1390static char next_shortname1 = 'A';
1391static char next_shortname2 = '0';
1392
1393static void
1394map_switch_event(struct trace_switch_event *switch_event,
1395		 struct perf_session *session,
1396		 struct event *event __used,
1397		 int this_cpu,
1398		 u64 timestamp,
1399		 struct thread *thread __used)
1400{
1401	struct thread *sched_out __used, *sched_in;
 
1402	int new_shortname;
1403	u64 timestamp0;
1404	s64 delta;
1405	int cpu;
1406
1407	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1408
1409	if (this_cpu > max_cpu)
1410		max_cpu = this_cpu;
1411
1412	timestamp0 = cpu_last_switched[this_cpu];
1413	cpu_last_switched[this_cpu] = timestamp;
1414	if (timestamp0)
1415		delta = timestamp - timestamp0;
1416	else
1417		delta = 0;
1418
1419	if (delta < 0)
1420		die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1421
 
1422
1423	sched_out = perf_session__findnew(session, switch_event->prev_pid);
1424	sched_in = perf_session__findnew(session, switch_event->next_pid);
 
1425
1426	curr_thread[this_cpu] = sched_in;
1427
1428	printf("  ");
1429
1430	new_shortname = 0;
1431	if (!sched_in->shortname[0]) {
1432		sched_in->shortname[0] = next_shortname1;
1433		sched_in->shortname[1] = next_shortname2;
1434
1435		if (next_shortname1 < 'Z') {
1436			next_shortname1++;
 
 
1437		} else {
1438			next_shortname1='A';
1439			if (next_shortname2 < '9') {
1440				next_shortname2++;
 
 
1441			} else {
1442				next_shortname2='0';
 
 
 
 
1443			}
1444		}
1445		new_shortname = 1;
1446	}
1447
1448	for (cpu = 0; cpu <= max_cpu; cpu++) {
1449		if (cpu != this_cpu)
1450			printf(" ");
1451		else
1452			printf("*");
1453
1454		if (curr_thread[cpu]) {
1455			if (curr_thread[cpu]->pid)
1456				printf("%2s ", curr_thread[cpu]->shortname);
1457			else
1458				printf(".  ");
1459		} else
1460			printf("   ");
1461	}
1462
1463	printf("  %12.6f secs ", (double)timestamp/1e9);
1464	if (new_shortname) {
1465		printf("%s => %s:%d\n",
1466			sched_in->shortname, sched_in->comm, sched_in->pid);
1467	} else {
1468		printf("\n");
1469	}
1470}
1471
 
1472
1473static void
1474process_sched_switch_event(void *data, struct perf_session *session,
1475			   struct event *event,
1476			   int this_cpu,
1477			   u64 timestamp __used,
1478			   struct thread *thread __used)
1479{
1480	struct trace_switch_event switch_event;
1481
1482	FILL_COMMON_FIELDS(switch_event, event, data);
1483
1484	FILL_ARRAY(switch_event, prev_comm, event, data);
1485	FILL_FIELD(switch_event, prev_pid, event, data);
1486	FILL_FIELD(switch_event, prev_prio, event, data);
1487	FILL_FIELD(switch_event, prev_state, event, data);
1488	FILL_ARRAY(switch_event, next_comm, event, data);
1489	FILL_FIELD(switch_event, next_pid, event, data);
1490	FILL_FIELD(switch_event, next_prio, event, data);
1491
1492	if (curr_pid[this_cpu] != (u32)-1) {
1493		/*
1494		 * Are we trying to switch away a PID that is
1495		 * not current?
1496		 */
1497		if (curr_pid[this_cpu] != switch_event.prev_pid)
1498			nr_context_switch_bugs++;
1499	}
1500	if (trace_handler->switch_event)
1501		trace_handler->switch_event(&switch_event, session, event,
1502					    this_cpu, timestamp, thread);
1503
1504	curr_pid[this_cpu] = switch_event.next_pid;
 
 
 
 
1505}
1506
1507static void
1508process_sched_runtime_event(void *data, struct perf_session *session,
1509			   struct event *event,
1510			   int cpu __used,
1511			   u64 timestamp __used,
1512			   struct thread *thread __used)
1513{
1514	struct trace_runtime_event runtime_event;
1515
1516	FILL_ARRAY(runtime_event, comm, event, data);
1517	FILL_FIELD(runtime_event, pid, event, data);
1518	FILL_FIELD(runtime_event, runtime, event, data);
1519	FILL_FIELD(runtime_event, vruntime, event, data);
1520
1521	if (trace_handler->runtime_event)
1522		trace_handler->runtime_event(&runtime_event, session, event, cpu, timestamp, thread);
1523}
1524
1525static void
1526process_sched_fork_event(void *data,
1527			 struct event *event,
1528			 int cpu __used,
1529			 u64 timestamp __used,
1530			 struct thread *thread __used)
1531{
1532	struct trace_fork_event fork_event;
1533
1534	FILL_COMMON_FIELDS(fork_event, event, data);
1535
1536	FILL_ARRAY(fork_event, parent_comm, event, data);
1537	FILL_FIELD(fork_event, parent_pid, event, data);
1538	FILL_ARRAY(fork_event, child_comm, event, data);
1539	FILL_FIELD(fork_event, child_pid, event, data);
1540
1541	if (trace_handler->fork_event)
1542		trace_handler->fork_event(&fork_event, event,
1543					  cpu, timestamp, thread);
1544}
1545
1546static void
1547process_sched_exit_event(struct event *event,
1548			 int cpu __used,
1549			 u64 timestamp __used,
1550			 struct thread *thread __used)
1551{
1552	if (verbose)
1553		printf("sched_exit event %p\n", event);
1554}
 
 
 
 
1555
1556static void
1557process_sched_migrate_task_event(void *data, struct perf_session *session,
1558			   struct event *event,
1559			   int cpu __used,
1560			   u64 timestamp __used,
1561			   struct thread *thread __used)
1562{
1563	struct trace_migrate_task_event migrate_task_event;
1564
1565	FILL_COMMON_FIELDS(migrate_task_event, event, data);
1566
1567	FILL_ARRAY(migrate_task_event, comm, event, data);
1568	FILL_FIELD(migrate_task_event, pid, event, data);
1569	FILL_FIELD(migrate_task_event, prio, event, data);
1570	FILL_FIELD(migrate_task_event, cpu, event, data);
1571
1572	if (trace_handler->migrate_task_event)
1573		trace_handler->migrate_task_event(&migrate_task_event, session,
1574						 event, cpu, timestamp, thread);
1575}
1576
1577static void process_raw_event(union perf_event *raw_event __used,
1578			      struct perf_session *session, void *data, int cpu,
1579			      u64 timestamp, struct thread *thread)
1580{
1581	struct event *event;
1582	int type;
1583
1584
1585	type = trace_parse_common_type(data);
1586	event = trace_find_event(type);
1587
1588	if (!strcmp(event->name, "sched_switch"))
1589		process_sched_switch_event(data, session, event, cpu, timestamp, thread);
1590	if (!strcmp(event->name, "sched_stat_runtime"))
1591		process_sched_runtime_event(data, session, event, cpu, timestamp, thread);
1592	if (!strcmp(event->name, "sched_wakeup"))
1593		process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1594	if (!strcmp(event->name, "sched_wakeup_new"))
1595		process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1596	if (!strcmp(event->name, "sched_process_fork"))
1597		process_sched_fork_event(data, event, cpu, timestamp, thread);
1598	if (!strcmp(event->name, "sched_process_exit"))
1599		process_sched_exit_event(event, cpu, timestamp, thread);
1600	if (!strcmp(event->name, "sched_migrate_task"))
1601		process_sched_migrate_task_event(data, session, event, cpu, timestamp, thread);
1602}
1603
1604static int process_sample_event(union perf_event *event,
1605				struct perf_sample *sample,
1606				struct perf_evsel *evsel __used,
1607				struct perf_session *session)
1608{
1609	struct thread *thread;
1610
1611	if (!(session->sample_type & PERF_SAMPLE_RAW))
1612		return 0;
1613
1614	thread = perf_session__findnew(session, sample->pid);
1615	if (thread == NULL) {
1616		pr_debug("problem processing %d event, skipping it.\n",
1617			 event->header.type);
1618		return -1;
1619	}
1620
1621	dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 
 
 
1622
1623	if (profile_cpu != -1 && profile_cpu != (int)sample->cpu)
1624		return 0;
 
 
 
 
 
1625
1626	process_raw_event(event, session, sample->raw_data, sample->cpu,
1627			  sample->time, thread);
 
 
1628
1629	return 0;
1630}
1631
1632static struct perf_event_ops event_ops = {
1633	.sample			= process_sample_event,
1634	.comm			= perf_event__process_comm,
1635	.lost			= perf_event__process_lost,
1636	.fork			= perf_event__process_task,
1637	.ordered_samples	= true,
1638};
1639
1640static void read_events(bool destroy, struct perf_session **psession)
1641{
1642	int err = -EINVAL;
1643	struct perf_session *session = perf_session__new(input_name, O_RDONLY,
1644							 0, false, &event_ops);
1645	if (session == NULL)
1646		die("No Memory");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1647
1648	if (perf_session__has_traces(session, "record -R")) {
1649		err = perf_session__process_events(session, &event_ops);
1650		if (err)
1651			die("Failed to process events, error %d", err);
 
 
1652
1653		nr_events      = session->hists.stats.nr_events[0];
1654		nr_lost_events = session->hists.stats.total_lost;
1655		nr_lost_chunks = session->hists.stats.nr_events[PERF_RECORD_LOST];
1656	}
1657
1658	if (destroy)
1659		perf_session__delete(session);
1660
1661	if (psession)
1662		*psession = session;
1663}
1664
1665static void print_bad_events(void)
1666{
1667	if (nr_unordered_timestamps && nr_timestamps) {
1668		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1669			(double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1670			nr_unordered_timestamps, nr_timestamps);
1671	}
1672	if (nr_lost_events && nr_events) {
1673		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1674			(double)nr_lost_events/(double)nr_events*100.0,
1675			nr_lost_events, nr_events, nr_lost_chunks);
1676	}
1677	if (nr_state_machine_bugs && nr_timestamps) {
1678		printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1679			(double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1680			nr_state_machine_bugs, nr_timestamps);
1681		if (nr_lost_events)
1682			printf(" (due to lost events?)");
1683		printf("\n");
1684	}
1685	if (nr_context_switch_bugs && nr_timestamps) {
1686		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1687			(double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1688			nr_context_switch_bugs, nr_timestamps);
1689		if (nr_lost_events)
1690			printf(" (due to lost events?)");
1691		printf("\n");
1692	}
1693}
1694
1695static void __cmd_lat(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1696{
1697	struct rb_node *next;
1698	struct perf_session *session;
1699
1700	setup_pager();
1701	read_events(false, &session);
1702	sort_lat();
1703
1704	printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1705	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1706	printf(" ---------------------------------------------------------------------------------------------------------------\n");
 
 
 
 
 
 
1707
1708	next = rb_first(&sorted_atom_root);
1709
1710	while (next) {
1711		struct work_atoms *work_list;
1712
1713		work_list = rb_entry(next, struct work_atoms, node);
1714		output_lat_thread(work_list);
1715		next = rb_next(next);
 
1716	}
1717
1718	printf(" -----------------------------------------------------------------------------------------\n");
1719	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1720		(double)all_runtime/1e6, all_count);
1721
1722	printf(" ---------------------------------------------------\n");
1723
1724	print_bad_events();
1725	printf("\n");
1726
1727	perf_session__delete(session);
1728}
1729
1730static struct trace_sched_handler map_ops  = {
1731	.wakeup_event		= NULL,
1732	.switch_event		= map_switch_event,
1733	.runtime_event		= NULL,
1734	.fork_event		= NULL,
1735};
1736
1737static void __cmd_map(void)
1738{
1739	max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1740
1741	setup_pager();
1742	read_events(true, NULL);
1743	print_bad_events();
 
 
1744}
1745
1746static void __cmd_replay(void)
1747{
1748	unsigned long i;
1749
1750	calibrate_run_measurement_overhead();
1751	calibrate_sleep_measurement_overhead();
1752
1753	test_calibrations();
1754
1755	read_events(true, NULL);
 
1756
1757	printf("nr_run_events:        %ld\n", nr_run_events);
1758	printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1759	printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1760
1761	if (targetless_wakeups)
1762		printf("target-less wakeups:  %ld\n", targetless_wakeups);
1763	if (multitarget_wakeups)
1764		printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1765	if (nr_run_events_optimized)
1766		printf("run atoms optimized: %ld\n",
1767			nr_run_events_optimized);
1768
1769	print_task_traces();
1770	add_cross_task_wakeups();
1771
1772	create_tasks();
1773	printf("------------------------------------------------------------\n");
1774	for (i = 0; i < replay_repeat; i++)
1775		run_one_test();
1776}
1777
1778
1779static const char * const sched_usage[] = {
1780	"perf sched [<options>] {record|latency|map|replay|script}",
1781	NULL
1782};
1783
1784static const struct option sched_options[] = {
1785	OPT_STRING('i', "input", &input_name, "file",
1786		    "input file name"),
1787	OPT_INCR('v', "verbose", &verbose,
1788		    "be more verbose (show symbol address, etc)"),
1789	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1790		    "dump raw trace in ASCII"),
1791	OPT_END()
1792};
1793
1794static const char * const latency_usage[] = {
1795	"perf sched latency [<options>]",
1796	NULL
1797};
1798
1799static const struct option latency_options[] = {
1800	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1801		   "sort by key(s): runtime, switch, avg, max"),
1802	OPT_INCR('v', "verbose", &verbose,
1803		    "be more verbose (show symbol address, etc)"),
1804	OPT_INTEGER('C', "CPU", &profile_cpu,
1805		    "CPU to profile on"),
1806	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1807		    "dump raw trace in ASCII"),
1808	OPT_END()
1809};
1810
1811static const char * const replay_usage[] = {
1812	"perf sched replay [<options>]",
1813	NULL
1814};
1815
1816static const struct option replay_options[] = {
1817	OPT_UINTEGER('r', "repeat", &replay_repeat,
1818		     "repeat the workload replay N times (-1: infinite)"),
1819	OPT_INCR('v', "verbose", &verbose,
1820		    "be more verbose (show symbol address, etc)"),
1821	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1822		    "dump raw trace in ASCII"),
1823	OPT_END()
1824};
1825
1826static void setup_sorting(void)
 
1827{
1828	char *tmp, *tok, *str = strdup(sort_order);
1829
1830	for (tok = strtok_r(str, ", ", &tmp);
1831			tok; tok = strtok_r(NULL, ", ", &tmp)) {
1832		if (sort_dimension__add(tok, &sort_list) < 0) {
1833			error("Unknown --sort key: `%s'", tok);
1834			usage_with_options(latency_usage, latency_options);
1835		}
1836	}
1837
1838	free(str);
1839
1840	sort_dimension__add("pid", &cmp_pid);
1841}
1842
1843static const char *record_args[] = {
1844	"record",
1845	"-a",
1846	"-R",
1847	"-f",
1848	"-m", "1024",
1849	"-c", "1",
1850	"-e", "sched:sched_switch",
1851	"-e", "sched:sched_stat_wait",
1852	"-e", "sched:sched_stat_sleep",
1853	"-e", "sched:sched_stat_iowait",
1854	"-e", "sched:sched_stat_runtime",
1855	"-e", "sched:sched_process_exit",
1856	"-e", "sched:sched_process_fork",
1857	"-e", "sched:sched_wakeup",
1858	"-e", "sched:sched_migrate_task",
1859};
1860
1861static int __cmd_record(int argc, const char **argv)
1862{
1863	unsigned int rec_argc, i, j;
1864	const char **rec_argv;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1865
1866	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1867	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1868
1869	if (rec_argv == NULL)
1870		return -ENOMEM;
1871
1872	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1873		rec_argv[i] = strdup(record_args[i]);
1874
1875	for (j = 1; j < (unsigned int)argc; j++, i++)
1876		rec_argv[i] = argv[j];
1877
1878	BUG_ON(i != rec_argc);
1879
1880	return cmd_record(i, rec_argv, NULL);
1881}
1882
1883int cmd_sched(int argc, const char **argv, const char *prefix __used)
1884{
1885	argc = parse_options(argc, argv, sched_options, sched_usage,
1886			     PARSE_OPT_STOP_AT_NON_OPTION);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1887	if (!argc)
1888		usage_with_options(sched_usage, sched_options);
1889
1890	/*
1891	 * Aliased to 'perf script' for now:
1892	 */
1893	if (!strcmp(argv[0], "script"))
1894		return cmd_script(argc, argv, prefix);
1895
1896	symbol__init();
1897	if (!strncmp(argv[0], "rec", 3)) {
1898		return __cmd_record(argc, argv);
1899	} else if (!strncmp(argv[0], "lat", 3)) {
1900		trace_handler = &lat_ops;
1901		if (argc > 1) {
1902			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1903			if (argc)
1904				usage_with_options(latency_usage, latency_options);
1905		}
1906		setup_sorting();
1907		__cmd_lat();
1908	} else if (!strcmp(argv[0], "map")) {
1909		trace_handler = &map_ops;
1910		setup_sorting();
1911		__cmd_map();
1912	} else if (!strncmp(argv[0], "rep", 3)) {
1913		trace_handler = &replay_ops;
1914		if (argc) {
1915			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1916			if (argc)
1917				usage_with_options(replay_usage, replay_options);
1918		}
1919		__cmd_replay();
1920	} else {
1921		usage_with_options(sched_usage, sched_options);
1922	}
1923
1924	return 0;
1925}
v4.6
   1#include "builtin.h"
   2#include "perf.h"
   3
   4#include "util/util.h"
   5#include "util/evlist.h"
   6#include "util/cache.h"
   7#include "util/evsel.h"
   8#include "util/symbol.h"
   9#include "util/thread.h"
  10#include "util/header.h"
  11#include "util/session.h"
  12#include "util/tool.h"
  13#include "util/cloexec.h"
  14
  15#include <subcmd/parse-options.h>
  16#include "util/trace-event.h"
  17
  18#include "util/debug.h"
  19
  20#include <sys/prctl.h>
  21#include <sys/resource.h>
  22
  23#include <semaphore.h>
  24#include <pthread.h>
  25#include <math.h>
  26#include <api/fs/fs.h>
 
 
 
 
 
 
  27
  28#define PR_SET_NAME		15               /* Set process name */
  29#define MAX_CPUS		4096
 
 
 
 
  30#define COMM_LEN		20
  31#define SYM_LEN			129
  32#define MAX_PID			1024000
 
 
 
  33
  34struct sched_atom;
  35
  36struct task_desc {
  37	unsigned long		nr;
  38	unsigned long		pid;
  39	char			comm[COMM_LEN];
  40
  41	unsigned long		nr_events;
  42	unsigned long		curr_event;
  43	struct sched_atom	**atoms;
  44
  45	pthread_t		thread;
  46	sem_t			sleep_sem;
  47
  48	sem_t			ready_for_work;
  49	sem_t			work_done_sem;
  50
  51	u64			cpu_usage;
  52};
  53
  54enum sched_event_type {
  55	SCHED_EVENT_RUN,
  56	SCHED_EVENT_SLEEP,
  57	SCHED_EVENT_WAKEUP,
  58	SCHED_EVENT_MIGRATION,
  59};
  60
  61struct sched_atom {
  62	enum sched_event_type	type;
  63	int			specific_wait;
  64	u64			timestamp;
  65	u64			duration;
  66	unsigned long		nr;
  67	sem_t			*wait_sem;
  68	struct task_desc	*wakee;
  69};
  70
  71#define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  72
  73enum thread_state {
  74	THREAD_SLEEPING = 0,
  75	THREAD_WAIT_CPU,
  76	THREAD_SCHED_IN,
  77	THREAD_IGNORE
  78};
  79
  80struct work_atom {
  81	struct list_head	list;
  82	enum thread_state	state;
  83	u64			sched_out_time;
  84	u64			wake_up_time;
  85	u64			sched_in_time;
  86	u64			runtime;
  87};
  88
  89struct work_atoms {
  90	struct list_head	work_list;
  91	struct thread		*thread;
  92	struct rb_node		node;
  93	u64			max_lat;
  94	u64			max_lat_at;
  95	u64			total_lat;
  96	u64			nb_atoms;
  97	u64			total_runtime;
  98	int			num_merged;
  99};
 100
 101typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 102
 103struct perf_sched;
 104
 105struct trace_sched_handler {
 106	int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 107			    struct perf_sample *sample, struct machine *machine);
 108
 109	int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 110			     struct perf_sample *sample, struct machine *machine);
 111
 112	int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 113			    struct perf_sample *sample, struct machine *machine);
 114
 115	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 116	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 117			  struct machine *machine);
 118
 119	int (*migrate_task_event)(struct perf_sched *sched,
 120				  struct perf_evsel *evsel,
 121				  struct perf_sample *sample,
 122				  struct machine *machine);
 123};
 124
 125struct perf_sched {
 126	struct perf_tool tool;
 127	const char	 *sort_order;
 128	unsigned long	 nr_tasks;
 129	struct task_desc **pid_to_task;
 130	struct task_desc **tasks;
 131	const struct trace_sched_handler *tp_handler;
 132	pthread_mutex_t	 start_work_mutex;
 133	pthread_mutex_t	 work_done_wait_mutex;
 134	int		 profile_cpu;
 135/*
 136 * Track the current task - that way we can know whether there's any
 137 * weird events, such as a task being switched away that is not current.
 138 */
 139	int		 max_cpu;
 140	u32		 curr_pid[MAX_CPUS];
 141	struct thread	 *curr_thread[MAX_CPUS];
 142	char		 next_shortname1;
 143	char		 next_shortname2;
 144	unsigned int	 replay_repeat;
 145	unsigned long	 nr_run_events;
 146	unsigned long	 nr_sleep_events;
 147	unsigned long	 nr_wakeup_events;
 148	unsigned long	 nr_sleep_corrections;
 149	unsigned long	 nr_run_events_optimized;
 150	unsigned long	 targetless_wakeups;
 151	unsigned long	 multitarget_wakeups;
 152	unsigned long	 nr_runs;
 153	unsigned long	 nr_timestamps;
 154	unsigned long	 nr_unordered_timestamps;
 155	unsigned long	 nr_context_switch_bugs;
 156	unsigned long	 nr_events;
 157	unsigned long	 nr_lost_chunks;
 158	unsigned long	 nr_lost_events;
 159	u64		 run_measurement_overhead;
 160	u64		 sleep_measurement_overhead;
 161	u64		 start_time;
 162	u64		 cpu_usage;
 163	u64		 runavg_cpu_usage;
 164	u64		 parent_cpu_usage;
 165	u64		 runavg_parent_cpu_usage;
 166	u64		 sum_runtime;
 167	u64		 sum_fluct;
 168	u64		 run_avg;
 169	u64		 all_runtime;
 170	u64		 all_count;
 171	u64		 cpu_last_switched[MAX_CPUS];
 172	struct rb_root	 atom_root, sorted_atom_root, merged_atom_root;
 173	struct list_head sort_list, cmp_pid;
 174	bool force;
 175	bool skip_merge;
 176};
 177
 178static u64 get_nsecs(void)
 179{
 180	struct timespec ts;
 181
 182	clock_gettime(CLOCK_MONOTONIC, &ts);
 183
 184	return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 185}
 186
 187static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 188{
 189	u64 T0 = get_nsecs(), T1;
 190
 191	do {
 192		T1 = get_nsecs();
 193	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 194}
 195
 196static void sleep_nsecs(u64 nsecs)
 197{
 198	struct timespec ts;
 199
 200	ts.tv_nsec = nsecs % 999999999;
 201	ts.tv_sec = nsecs / 999999999;
 202
 203	nanosleep(&ts, NULL);
 204}
 205
 206static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 207{
 208	u64 T0, T1, delta, min_delta = 1000000000ULL;
 209	int i;
 210
 211	for (i = 0; i < 10; i++) {
 212		T0 = get_nsecs();
 213		burn_nsecs(sched, 0);
 214		T1 = get_nsecs();
 215		delta = T1-T0;
 216		min_delta = min(min_delta, delta);
 217	}
 218	sched->run_measurement_overhead = min_delta;
 219
 220	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 221}
 222
 223static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 224{
 225	u64 T0, T1, delta, min_delta = 1000000000ULL;
 226	int i;
 227
 228	for (i = 0; i < 10; i++) {
 229		T0 = get_nsecs();
 230		sleep_nsecs(10000);
 231		T1 = get_nsecs();
 232		delta = T1-T0;
 233		min_delta = min(min_delta, delta);
 234	}
 235	min_delta -= 10000;
 236	sched->sleep_measurement_overhead = min_delta;
 237
 238	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 239}
 240
 241static struct sched_atom *
 242get_new_event(struct task_desc *task, u64 timestamp)
 243{
 244	struct sched_atom *event = zalloc(sizeof(*event));
 245	unsigned long idx = task->nr_events;
 246	size_t size;
 247
 248	event->timestamp = timestamp;
 249	event->nr = idx;
 250
 251	task->nr_events++;
 252	size = sizeof(struct sched_atom *) * task->nr_events;
 253	task->atoms = realloc(task->atoms, size);
 254	BUG_ON(!task->atoms);
 255
 256	task->atoms[idx] = event;
 257
 258	return event;
 259}
 260
 261static struct sched_atom *last_event(struct task_desc *task)
 262{
 263	if (!task->nr_events)
 264		return NULL;
 265
 266	return task->atoms[task->nr_events - 1];
 267}
 268
 269static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 270				u64 timestamp, u64 duration)
 271{
 272	struct sched_atom *event, *curr_event = last_event(task);
 273
 274	/*
 275	 * optimize an existing RUN event by merging this one
 276	 * to it:
 277	 */
 278	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 279		sched->nr_run_events_optimized++;
 280		curr_event->duration += duration;
 281		return;
 282	}
 283
 284	event = get_new_event(task, timestamp);
 285
 286	event->type = SCHED_EVENT_RUN;
 287	event->duration = duration;
 288
 289	sched->nr_run_events++;
 290}
 291
 292static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 293				   u64 timestamp, struct task_desc *wakee)
 
 294{
 295	struct sched_atom *event, *wakee_event;
 296
 297	event = get_new_event(task, timestamp);
 298	event->type = SCHED_EVENT_WAKEUP;
 299	event->wakee = wakee;
 300
 301	wakee_event = last_event(wakee);
 302	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 303		sched->targetless_wakeups++;
 304		return;
 305	}
 306	if (wakee_event->wait_sem) {
 307		sched->multitarget_wakeups++;
 308		return;
 309	}
 310
 311	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 312	sem_init(wakee_event->wait_sem, 0, 0);
 313	wakee_event->specific_wait = 1;
 314	event->wait_sem = wakee_event->wait_sem;
 315
 316	sched->nr_wakeup_events++;
 317}
 318
 319static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 320				  u64 timestamp, u64 task_state __maybe_unused)
 
 321{
 322	struct sched_atom *event = get_new_event(task, timestamp);
 323
 324	event->type = SCHED_EVENT_SLEEP;
 325
 326	sched->nr_sleep_events++;
 327}
 328
 329static struct task_desc *register_pid(struct perf_sched *sched,
 330				      unsigned long pid, const char *comm)
 331{
 332	struct task_desc *task;
 333	static int pid_max;
 334
 335	if (sched->pid_to_task == NULL) {
 336		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 337			pid_max = MAX_PID;
 338		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 339	}
 340	if (pid >= (unsigned long)pid_max) {
 341		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 342			sizeof(struct task_desc *))) == NULL);
 343		while (pid >= (unsigned long)pid_max)
 344			sched->pid_to_task[pid_max++] = NULL;
 345	}
 346
 347	task = sched->pid_to_task[pid];
 348
 349	if (task)
 350		return task;
 351
 352	task = zalloc(sizeof(*task));
 353	task->pid = pid;
 354	task->nr = sched->nr_tasks;
 355	strcpy(task->comm, comm);
 356	/*
 357	 * every task starts in sleeping state - this gets ignored
 358	 * if there's no wakeup pointing to this sleep state:
 359	 */
 360	add_sched_event_sleep(sched, task, 0, 0);
 361
 362	sched->pid_to_task[pid] = task;
 363	sched->nr_tasks++;
 364	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 365	BUG_ON(!sched->tasks);
 366	sched->tasks[task->nr] = task;
 367
 368	if (verbose)
 369		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 370
 371	return task;
 372}
 373
 374
 375static void print_task_traces(struct perf_sched *sched)
 376{
 377	struct task_desc *task;
 378	unsigned long i;
 379
 380	for (i = 0; i < sched->nr_tasks; i++) {
 381		task = sched->tasks[i];
 382		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 383			task->nr, task->comm, task->pid, task->nr_events);
 384	}
 385}
 386
 387static void add_cross_task_wakeups(struct perf_sched *sched)
 388{
 389	struct task_desc *task1, *task2;
 390	unsigned long i, j;
 391
 392	for (i = 0; i < sched->nr_tasks; i++) {
 393		task1 = sched->tasks[i];
 394		j = i + 1;
 395		if (j == sched->nr_tasks)
 396			j = 0;
 397		task2 = sched->tasks[j];
 398		add_sched_event_wakeup(sched, task1, 0, task2);
 399	}
 400}
 401
 402static void perf_sched__process_event(struct perf_sched *sched,
 403				      struct sched_atom *atom)
 404{
 405	int ret = 0;
 406
 407	switch (atom->type) {
 408		case SCHED_EVENT_RUN:
 409			burn_nsecs(sched, atom->duration);
 410			break;
 411		case SCHED_EVENT_SLEEP:
 412			if (atom->wait_sem)
 413				ret = sem_wait(atom->wait_sem);
 414			BUG_ON(ret);
 415			break;
 416		case SCHED_EVENT_WAKEUP:
 417			if (atom->wait_sem)
 418				ret = sem_post(atom->wait_sem);
 419			BUG_ON(ret);
 420			break;
 421		case SCHED_EVENT_MIGRATION:
 422			break;
 423		default:
 424			BUG_ON(1);
 425	}
 426}
 427
 428static u64 get_cpu_usage_nsec_parent(void)
 429{
 430	struct rusage ru;
 431	u64 sum;
 432	int err;
 433
 434	err = getrusage(RUSAGE_SELF, &ru);
 435	BUG_ON(err);
 436
 437	sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 438	sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 439
 440	return sum;
 441}
 442
 443static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 444{
 445	struct perf_event_attr attr;
 446	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 447	int fd;
 448	struct rlimit limit;
 449	bool need_privilege = false;
 450
 451	memset(&attr, 0, sizeof(attr));
 452
 453	attr.type = PERF_TYPE_SOFTWARE;
 454	attr.config = PERF_COUNT_SW_TASK_CLOCK;
 455
 456force_again:
 457	fd = sys_perf_event_open(&attr, 0, -1, -1,
 458				 perf_event_open_cloexec_flag());
 459
 460	if (fd < 0) {
 461		if (errno == EMFILE) {
 462			if (sched->force) {
 463				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 464				limit.rlim_cur += sched->nr_tasks - cur_task;
 465				if (limit.rlim_cur > limit.rlim_max) {
 466					limit.rlim_max = limit.rlim_cur;
 467					need_privilege = true;
 468				}
 469				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 470					if (need_privilege && errno == EPERM)
 471						strcpy(info, "Need privilege\n");
 472				} else
 473					goto force_again;
 474			} else
 475				strcpy(info, "Have a try with -f option\n");
 476		}
 477		pr_err("Error: sys_perf_event_open() syscall returned "
 478		       "with %d (%s)\n%s", fd,
 479		       strerror_r(errno, sbuf, sizeof(sbuf)), info);
 480		exit(EXIT_FAILURE);
 481	}
 482	return fd;
 483}
 484
 485static u64 get_cpu_usage_nsec_self(int fd)
 486{
 487	u64 runtime;
 488	int ret;
 489
 490	ret = read(fd, &runtime, sizeof(runtime));
 491	BUG_ON(ret != sizeof(runtime));
 492
 493	return runtime;
 494}
 495
 496struct sched_thread_parms {
 497	struct task_desc  *task;
 498	struct perf_sched *sched;
 499	int fd;
 500};
 501
 502static void *thread_func(void *ctx)
 503{
 504	struct sched_thread_parms *parms = ctx;
 505	struct task_desc *this_task = parms->task;
 506	struct perf_sched *sched = parms->sched;
 507	u64 cpu_usage_0, cpu_usage_1;
 508	unsigned long i, ret;
 509	char comm2[22];
 510	int fd = parms->fd;
 511
 512	zfree(&parms);
 513
 514	sprintf(comm2, ":%s", this_task->comm);
 515	prctl(PR_SET_NAME, comm2);
 516	if (fd < 0)
 517		return NULL;
 518again:
 519	ret = sem_post(&this_task->ready_for_work);
 520	BUG_ON(ret);
 521	ret = pthread_mutex_lock(&sched->start_work_mutex);
 522	BUG_ON(ret);
 523	ret = pthread_mutex_unlock(&sched->start_work_mutex);
 524	BUG_ON(ret);
 525
 526	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 527
 528	for (i = 0; i < this_task->nr_events; i++) {
 529		this_task->curr_event = i;
 530		perf_sched__process_event(sched, this_task->atoms[i]);
 531	}
 532
 533	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 534	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 535	ret = sem_post(&this_task->work_done_sem);
 536	BUG_ON(ret);
 537
 538	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 539	BUG_ON(ret);
 540	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 541	BUG_ON(ret);
 542
 543	goto again;
 544}
 545
 546static void create_tasks(struct perf_sched *sched)
 547{
 548	struct task_desc *task;
 549	pthread_attr_t attr;
 550	unsigned long i;
 551	int err;
 552
 553	err = pthread_attr_init(&attr);
 554	BUG_ON(err);
 555	err = pthread_attr_setstacksize(&attr,
 556			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 557	BUG_ON(err);
 558	err = pthread_mutex_lock(&sched->start_work_mutex);
 559	BUG_ON(err);
 560	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 561	BUG_ON(err);
 562	for (i = 0; i < sched->nr_tasks; i++) {
 563		struct sched_thread_parms *parms = malloc(sizeof(*parms));
 564		BUG_ON(parms == NULL);
 565		parms->task = task = sched->tasks[i];
 566		parms->sched = sched;
 567		parms->fd = self_open_counters(sched, i);
 568		sem_init(&task->sleep_sem, 0, 0);
 569		sem_init(&task->ready_for_work, 0, 0);
 570		sem_init(&task->work_done_sem, 0, 0);
 571		task->curr_event = 0;
 572		err = pthread_create(&task->thread, &attr, thread_func, parms);
 573		BUG_ON(err);
 574	}
 575}
 576
 577static void wait_for_tasks(struct perf_sched *sched)
 578{
 579	u64 cpu_usage_0, cpu_usage_1;
 580	struct task_desc *task;
 581	unsigned long i, ret;
 582
 583	sched->start_time = get_nsecs();
 584	sched->cpu_usage = 0;
 585	pthread_mutex_unlock(&sched->work_done_wait_mutex);
 586
 587	for (i = 0; i < sched->nr_tasks; i++) {
 588		task = sched->tasks[i];
 589		ret = sem_wait(&task->ready_for_work);
 590		BUG_ON(ret);
 591		sem_init(&task->ready_for_work, 0, 0);
 592	}
 593	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 594	BUG_ON(ret);
 595
 596	cpu_usage_0 = get_cpu_usage_nsec_parent();
 597
 598	pthread_mutex_unlock(&sched->start_work_mutex);
 599
 600	for (i = 0; i < sched->nr_tasks; i++) {
 601		task = sched->tasks[i];
 602		ret = sem_wait(&task->work_done_sem);
 603		BUG_ON(ret);
 604		sem_init(&task->work_done_sem, 0, 0);
 605		sched->cpu_usage += task->cpu_usage;
 606		task->cpu_usage = 0;
 607	}
 608
 609	cpu_usage_1 = get_cpu_usage_nsec_parent();
 610	if (!sched->runavg_cpu_usage)
 611		sched->runavg_cpu_usage = sched->cpu_usage;
 612	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 613
 614	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 615	if (!sched->runavg_parent_cpu_usage)
 616		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 617	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 618					 sched->parent_cpu_usage)/sched->replay_repeat;
 619
 620	ret = pthread_mutex_lock(&sched->start_work_mutex);
 621	BUG_ON(ret);
 622
 623	for (i = 0; i < sched->nr_tasks; i++) {
 624		task = sched->tasks[i];
 625		sem_init(&task->sleep_sem, 0, 0);
 626		task->curr_event = 0;
 627	}
 628}
 629
 630static void run_one_test(struct perf_sched *sched)
 631{
 632	u64 T0, T1, delta, avg_delta, fluct;
 633
 634	T0 = get_nsecs();
 635	wait_for_tasks(sched);
 636	T1 = get_nsecs();
 637
 638	delta = T1 - T0;
 639	sched->sum_runtime += delta;
 640	sched->nr_runs++;
 641
 642	avg_delta = sched->sum_runtime / sched->nr_runs;
 643	if (delta < avg_delta)
 644		fluct = avg_delta - delta;
 645	else
 646		fluct = delta - avg_delta;
 647	sched->sum_fluct += fluct;
 648	if (!sched->run_avg)
 649		sched->run_avg = delta;
 650	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 651
 652	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
 
 653
 654	printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
 
 655
 656	printf("cpu: %0.2f / %0.2f",
 657		(double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
 658
 659#if 0
 660	/*
 661	 * rusage statistics done by the parent, these are less
 662	 * accurate than the sched->sum_exec_runtime based statistics:
 663	 */
 664	printf(" [%0.2f / %0.2f]",
 665		(double)sched->parent_cpu_usage/1e6,
 666		(double)sched->runavg_parent_cpu_usage/1e6);
 667#endif
 668
 669	printf("\n");
 670
 671	if (sched->nr_sleep_corrections)
 672		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 673	sched->nr_sleep_corrections = 0;
 674}
 675
 676static void test_calibrations(struct perf_sched *sched)
 677{
 678	u64 T0, T1;
 679
 680	T0 = get_nsecs();
 681	burn_nsecs(sched, 1e6);
 682	T1 = get_nsecs();
 683
 684	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 685
 686	T0 = get_nsecs();
 687	sleep_nsecs(1e6);
 688	T1 = get_nsecs();
 689
 690	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 691}
 692
 693static int
 694replay_wakeup_event(struct perf_sched *sched,
 695		    struct perf_evsel *evsel, struct perf_sample *sample,
 696		    struct machine *machine __maybe_unused)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 697{
 698	const char *comm = perf_evsel__strval(evsel, sample, "comm");
 699	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
 700	struct task_desc *waker, *wakee;
 701
 702	if (verbose) {
 703		printf("sched_wakeup event %p\n", evsel);
 704
 705		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 
 
 
 706	}
 707
 708	waker = register_pid(sched, sample->tid, "<unknown>");
 709	wakee = register_pid(sched, pid, comm);
 710
 711	add_sched_event_wakeup(sched, waker, sample->time, wakee);
 712	return 0;
 713}
 714
 715static int replay_switch_event(struct perf_sched *sched,
 716			       struct perf_evsel *evsel,
 717			       struct perf_sample *sample,
 718			       struct machine *machine __maybe_unused)
 719{
 720	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 721		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 722	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 723		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 724	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 725	struct task_desc *prev, __maybe_unused *next;
 726	u64 timestamp0, timestamp = sample->time;
 727	int cpu = sample->cpu;
 728	s64 delta;
 729
 730	if (verbose)
 731		printf("sched_switch event %p\n", evsel);
 732
 733	if (cpu >= MAX_CPUS || cpu < 0)
 734		return 0;
 735
 736	timestamp0 = sched->cpu_last_switched[cpu];
 737	if (timestamp0)
 738		delta = timestamp - timestamp0;
 739	else
 740		delta = 0;
 741
 742	if (delta < 0) {
 743		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 744		return -1;
 
 
 
 
 
 745	}
 746
 747	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 748		 prev_comm, prev_pid, next_comm, next_pid, delta);
 749
 750	prev = register_pid(sched, prev_pid, prev_comm);
 751	next = register_pid(sched, next_pid, next_comm);
 752
 753	sched->cpu_last_switched[cpu] = timestamp;
 754
 755	add_sched_event_run(sched, prev, timestamp, delta);
 756	add_sched_event_sleep(sched, prev, timestamp, prev_state);
 757
 758	return 0;
 
 759}
 760
 761static int replay_fork_event(struct perf_sched *sched,
 762			     union perf_event *event,
 763			     struct machine *machine)
 764{
 765	struct thread *child, *parent;
 766
 767	child = machine__findnew_thread(machine, event->fork.pid,
 768					event->fork.tid);
 769	parent = machine__findnew_thread(machine, event->fork.ppid,
 770					 event->fork.ptid);
 771
 772	if (child == NULL || parent == NULL) {
 773		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 774				 child, parent);
 775		goto out_put;
 776	}
 777
 
 
 
 
 
 
 
 778	if (verbose) {
 779		printf("fork event\n");
 780		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 781		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 782	}
 
 
 
 783
 784	register_pid(sched, parent->tid, thread__comm_str(parent));
 785	register_pid(sched, child->tid, thread__comm_str(child));
 786out_put:
 787	thread__put(child);
 788	thread__put(parent);
 789	return 0;
 790}
 791
 792struct sort_dimension {
 793	const char		*name;
 794	sort_fn_t		cmp;
 795	struct list_head	list;
 796};
 797
 
 
 798static int
 799thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 800{
 801	struct sort_dimension *sort;
 802	int ret = 0;
 803
 804	BUG_ON(list_empty(list));
 805
 806	list_for_each_entry(sort, list, list) {
 807		ret = sort->cmp(l, r);
 808		if (ret)
 809			return ret;
 810	}
 811
 812	return ret;
 813}
 814
 815static struct work_atoms *
 816thread_atoms_search(struct rb_root *root, struct thread *thread,
 817			 struct list_head *sort_list)
 818{
 819	struct rb_node *node = root->rb_node;
 820	struct work_atoms key = { .thread = thread };
 821
 822	while (node) {
 823		struct work_atoms *atoms;
 824		int cmp;
 825
 826		atoms = container_of(node, struct work_atoms, node);
 827
 828		cmp = thread_lat_cmp(sort_list, &key, atoms);
 829		if (cmp > 0)
 830			node = node->rb_left;
 831		else if (cmp < 0)
 832			node = node->rb_right;
 833		else {
 834			BUG_ON(thread != atoms->thread);
 835			return atoms;
 836		}
 837	}
 838	return NULL;
 839}
 840
 841static void
 842__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 843			 struct list_head *sort_list)
 844{
 845	struct rb_node **new = &(root->rb_node), *parent = NULL;
 846
 847	while (*new) {
 848		struct work_atoms *this;
 849		int cmp;
 850
 851		this = container_of(*new, struct work_atoms, node);
 852		parent = *new;
 853
 854		cmp = thread_lat_cmp(sort_list, data, this);
 855
 856		if (cmp > 0)
 857			new = &((*new)->rb_left);
 858		else
 859			new = &((*new)->rb_right);
 860	}
 861
 862	rb_link_node(&data->node, parent, new);
 863	rb_insert_color(&data->node, root);
 864}
 865
 866static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 867{
 868	struct work_atoms *atoms = zalloc(sizeof(*atoms));
 869	if (!atoms) {
 870		pr_err("No memory at %s\n", __func__);
 871		return -1;
 872	}
 873
 874	atoms->thread = thread__get(thread);
 875	INIT_LIST_HEAD(&atoms->work_list);
 876	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 877	return 0;
 
 
 
 
 
 
 
 
 
 878}
 879
 880static char sched_out_state(u64 prev_state)
 
 881{
 882	const char *str = TASK_STATE_TO_CHAR_STR;
 883
 884	return str[prev_state];
 885}
 886
 887static int
 888add_sched_out_event(struct work_atoms *atoms,
 889		    char run_state,
 890		    u64 timestamp)
 891{
 892	struct work_atom *atom = zalloc(sizeof(*atom));
 893	if (!atom) {
 894		pr_err("Non memory at %s", __func__);
 895		return -1;
 896	}
 897
 898	atom->sched_out_time = timestamp;
 899
 900	if (run_state == 'R') {
 901		atom->state = THREAD_WAIT_CPU;
 902		atom->wake_up_time = atom->sched_out_time;
 903	}
 904
 905	list_add_tail(&atom->list, &atoms->work_list);
 906	return 0;
 907}
 908
 909static void
 910add_runtime_event(struct work_atoms *atoms, u64 delta,
 911		  u64 timestamp __maybe_unused)
 912{
 913	struct work_atom *atom;
 914
 915	BUG_ON(list_empty(&atoms->work_list));
 916
 917	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 918
 919	atom->runtime += delta;
 920	atoms->total_runtime += delta;
 921}
 922
 923static void
 924add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 925{
 926	struct work_atom *atom;
 927	u64 delta;
 928
 929	if (list_empty(&atoms->work_list))
 930		return;
 931
 932	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 933
 934	if (atom->state != THREAD_WAIT_CPU)
 935		return;
 936
 937	if (timestamp < atom->wake_up_time) {
 938		atom->state = THREAD_IGNORE;
 939		return;
 940	}
 941
 942	atom->state = THREAD_SCHED_IN;
 943	atom->sched_in_time = timestamp;
 944
 945	delta = atom->sched_in_time - atom->wake_up_time;
 946	atoms->total_lat += delta;
 947	if (delta > atoms->max_lat) {
 948		atoms->max_lat = delta;
 949		atoms->max_lat_at = timestamp;
 950	}
 951	atoms->nb_atoms++;
 952}
 953
 954static int latency_switch_event(struct perf_sched *sched,
 955				struct perf_evsel *evsel,
 956				struct perf_sample *sample,
 957				struct machine *machine)
 
 
 
 958{
 959	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 960		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 961	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 962	struct work_atoms *out_events, *in_events;
 963	struct thread *sched_out, *sched_in;
 964	u64 timestamp0, timestamp = sample->time;
 965	int cpu = sample->cpu, err = -1;
 966	s64 delta;
 967
 968	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 969
 970	timestamp0 = sched->cpu_last_switched[cpu];
 971	sched->cpu_last_switched[cpu] = timestamp;
 972	if (timestamp0)
 973		delta = timestamp - timestamp0;
 974	else
 975		delta = 0;
 976
 977	if (delta < 0) {
 978		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 979		return -1;
 980	}
 981
 982	sched_out = machine__findnew_thread(machine, -1, prev_pid);
 983	sched_in = machine__findnew_thread(machine, -1, next_pid);
 984	if (sched_out == NULL || sched_in == NULL)
 985		goto out_put;
 986
 987	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 988	if (!out_events) {
 989		if (thread_atoms_insert(sched, sched_out))
 990			goto out_put;
 991		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 992		if (!out_events) {
 993			pr_err("out-event: Internal tree error");
 994			goto out_put;
 995		}
 996	}
 997	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
 998		return -1;
 999
1000	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1001	if (!in_events) {
1002		if (thread_atoms_insert(sched, sched_in))
1003			goto out_put;
1004		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1005		if (!in_events) {
1006			pr_err("in-event: Internal tree error");
1007			goto out_put;
1008		}
1009		/*
1010		 * Take came in we have not heard about yet,
1011		 * add in an initial atom in runnable state:
1012		 */
1013		if (add_sched_out_event(in_events, 'R', timestamp))
1014			goto out_put;
1015	}
1016	add_sched_in_event(in_events, timestamp);
1017	err = 0;
1018out_put:
1019	thread__put(sched_out);
1020	thread__put(sched_in);
1021	return err;
1022}
1023
1024static int latency_runtime_event(struct perf_sched *sched,
1025				 struct perf_evsel *evsel,
1026				 struct perf_sample *sample,
1027				 struct machine *machine)
1028{
1029	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1030	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1031	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1032	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1033	u64 timestamp = sample->time;
1034	int cpu = sample->cpu, err = -1;
1035
1036	if (thread == NULL)
1037		return -1;
 
 
 
 
 
 
 
 
1038
1039	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1040	if (!atoms) {
1041		if (thread_atoms_insert(sched, thread))
1042			goto out_put;
1043		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1044		if (!atoms) {
1045			pr_err("in-event: Internal tree error");
1046			goto out_put;
1047		}
1048		if (add_sched_out_event(atoms, 'R', timestamp))
1049			goto out_put;
1050	}
1051
1052	add_runtime_event(atoms, runtime, timestamp);
1053	err = 0;
1054out_put:
1055	thread__put(thread);
1056	return err;
1057}
1058
1059static int latency_wakeup_event(struct perf_sched *sched,
1060				struct perf_evsel *evsel,
1061				struct perf_sample *sample,
1062				struct machine *machine)
 
 
 
1063{
1064	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1065	struct work_atoms *atoms;
1066	struct work_atom *atom;
1067	struct thread *wakee;
1068	u64 timestamp = sample->time;
1069	int err = -1;
1070
1071	wakee = machine__findnew_thread(machine, -1, pid);
1072	if (wakee == NULL)
1073		return -1;
1074	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
 
 
1075	if (!atoms) {
1076		if (thread_atoms_insert(sched, wakee))
1077			goto out_put;
1078		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1079		if (!atoms) {
1080			pr_err("wakeup-event: Internal tree error");
1081			goto out_put;
1082		}
1083		if (add_sched_out_event(atoms, 'S', timestamp))
1084			goto out_put;
1085	}
1086
1087	BUG_ON(list_empty(&atoms->work_list));
1088
1089	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1090
1091	/*
1092	 * As we do not guarantee the wakeup event happens when
1093	 * task is out of run queue, also may happen when task is
1094	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1095	 * then we should not set the ->wake_up_time when wake up a
1096	 * task which is on run queue.
1097	 *
1098	 * You WILL be missing events if you've recorded only
1099	 * one CPU, or are only looking at only one, so don't
1100	 * skip in this case.
1101	 */
1102	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1103		goto out_ok;
1104
1105	sched->nr_timestamps++;
1106	if (atom->sched_out_time > timestamp) {
1107		sched->nr_unordered_timestamps++;
1108		goto out_ok;
1109	}
1110
1111	atom->state = THREAD_WAIT_CPU;
1112	atom->wake_up_time = timestamp;
1113out_ok:
1114	err = 0;
1115out_put:
1116	thread__put(wakee);
1117	return err;
1118}
1119
1120static int latency_migrate_task_event(struct perf_sched *sched,
1121				      struct perf_evsel *evsel,
1122				      struct perf_sample *sample,
1123				      struct machine *machine)
 
 
 
1124{
1125	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1126	u64 timestamp = sample->time;
1127	struct work_atoms *atoms;
1128	struct work_atom *atom;
1129	struct thread *migrant;
1130	int err = -1;
1131
1132	/*
1133	 * Only need to worry about migration when profiling one CPU.
1134	 */
1135	if (sched->profile_cpu == -1)
1136		return 0;
1137
1138	migrant = machine__findnew_thread(machine, -1, pid);
1139	if (migrant == NULL)
1140		return -1;
1141	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1142	if (!atoms) {
1143		if (thread_atoms_insert(sched, migrant))
1144			goto out_put;
1145		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1146		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1147		if (!atoms) {
1148			pr_err("migration-event: Internal tree error");
1149			goto out_put;
1150		}
1151		if (add_sched_out_event(atoms, 'R', timestamp))
1152			goto out_put;
1153	}
1154
1155	BUG_ON(list_empty(&atoms->work_list));
1156
1157	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1158	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1159
1160	sched->nr_timestamps++;
1161
1162	if (atom->sched_out_time > timestamp)
1163		sched->nr_unordered_timestamps++;
1164	err = 0;
1165out_put:
1166	thread__put(migrant);
1167	return err;
1168}
1169
1170static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
 
 
 
 
 
 
 
 
1171{
1172	int i;
1173	int ret;
1174	u64 avg;
1175
1176	if (!work_list->nb_atoms)
1177		return;
1178	/*
1179	 * Ignore idle threads:
1180	 */
1181	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1182		return;
1183
1184	sched->all_runtime += work_list->total_runtime;
1185	sched->all_count   += work_list->nb_atoms;
1186
1187	if (work_list->num_merged > 1)
1188		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1189	else
1190		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1191
1192	for (i = 0; i < 24 - ret; i++)
1193		printf(" ");
1194
1195	avg = work_list->total_lat / work_list->nb_atoms;
1196
1197	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1198	      (double)work_list->total_runtime / 1e6,
1199		 work_list->nb_atoms, (double)avg / 1e6,
1200		 (double)work_list->max_lat / 1e6,
1201		 (double)work_list->max_lat_at / 1e9);
1202}
1203
1204static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1205{
1206	if (l->thread == r->thread)
1207		return 0;
1208	if (l->thread->tid < r->thread->tid)
1209		return -1;
1210	if (l->thread->tid > r->thread->tid)
1211		return 1;
1212	return (int)(l->thread - r->thread);
 
1213}
1214
 
 
 
 
 
1215static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1216{
1217	u64 avgl, avgr;
1218
1219	if (!l->nb_atoms)
1220		return -1;
1221
1222	if (!r->nb_atoms)
1223		return 1;
1224
1225	avgl = l->total_lat / l->nb_atoms;
1226	avgr = r->total_lat / r->nb_atoms;
1227
1228	if (avgl < avgr)
1229		return -1;
1230	if (avgl > avgr)
1231		return 1;
1232
1233	return 0;
1234}
1235
 
 
 
 
 
1236static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1237{
1238	if (l->max_lat < r->max_lat)
1239		return -1;
1240	if (l->max_lat > r->max_lat)
1241		return 1;
1242
1243	return 0;
1244}
1245
 
 
 
 
 
1246static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1247{
1248	if (l->nb_atoms < r->nb_atoms)
1249		return -1;
1250	if (l->nb_atoms > r->nb_atoms)
1251		return 1;
1252
1253	return 0;
1254}
1255
 
 
 
 
 
1256static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1257{
1258	if (l->total_runtime < r->total_runtime)
1259		return -1;
1260	if (l->total_runtime > r->total_runtime)
1261		return 1;
1262
1263	return 0;
1264}
1265
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1266static int sort_dimension__add(const char *tok, struct list_head *list)
1267{
1268	size_t i;
1269	static struct sort_dimension avg_sort_dimension = {
1270		.name = "avg",
1271		.cmp  = avg_cmp,
1272	};
1273	static struct sort_dimension max_sort_dimension = {
1274		.name = "max",
1275		.cmp  = max_cmp,
1276	};
1277	static struct sort_dimension pid_sort_dimension = {
1278		.name = "pid",
1279		.cmp  = pid_cmp,
1280	};
1281	static struct sort_dimension runtime_sort_dimension = {
1282		.name = "runtime",
1283		.cmp  = runtime_cmp,
1284	};
1285	static struct sort_dimension switch_sort_dimension = {
1286		.name = "switch",
1287		.cmp  = switch_cmp,
1288	};
1289	struct sort_dimension *available_sorts[] = {
1290		&pid_sort_dimension,
1291		&avg_sort_dimension,
1292		&max_sort_dimension,
1293		&switch_sort_dimension,
1294		&runtime_sort_dimension,
1295	};
1296
1297	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1298		if (!strcmp(available_sorts[i]->name, tok)) {
1299			list_add_tail(&available_sorts[i]->list, list);
1300
1301			return 0;
1302		}
1303	}
1304
1305	return -1;
1306}
1307
1308static void perf_sched__sort_lat(struct perf_sched *sched)
 
 
1309{
1310	struct rb_node *node;
1311	struct rb_root *root = &sched->atom_root;
1312again:
1313	for (;;) {
1314		struct work_atoms *data;
1315		node = rb_first(root);
1316		if (!node)
1317			break;
1318
1319		rb_erase(node, root);
1320		data = rb_entry(node, struct work_atoms, node);
1321		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1322	}
1323	if (root == &sched->atom_root) {
1324		root = &sched->merged_atom_root;
1325		goto again;
1326	}
1327}
1328
1329static int process_sched_wakeup_event(struct perf_tool *tool,
1330				      struct perf_evsel *evsel,
1331				      struct perf_sample *sample,
1332				      struct machine *machine)
1333{
1334	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1335
1336	if (sched->tp_handler->wakeup_event)
1337		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1338
1339	return 0;
1340}
1341
1342static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1343			    struct perf_sample *sample, struct machine *machine)
 
 
 
 
 
1344{
1345	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1346	struct thread *sched_in;
1347	int new_shortname;
1348	u64 timestamp0, timestamp = sample->time;
1349	s64 delta;
1350	int cpu, this_cpu = sample->cpu;
1351
1352	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1353
1354	if (this_cpu > sched->max_cpu)
1355		sched->max_cpu = this_cpu;
1356
1357	timestamp0 = sched->cpu_last_switched[this_cpu];
1358	sched->cpu_last_switched[this_cpu] = timestamp;
1359	if (timestamp0)
1360		delta = timestamp - timestamp0;
1361	else
1362		delta = 0;
1363
1364	if (delta < 0) {
1365		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1366		return -1;
1367	}
1368
1369	sched_in = machine__findnew_thread(machine, -1, next_pid);
1370	if (sched_in == NULL)
1371		return -1;
1372
1373	sched->curr_thread[this_cpu] = thread__get(sched_in);
1374
1375	printf("  ");
1376
1377	new_shortname = 0;
1378	if (!sched_in->shortname[0]) {
1379		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1380			/*
1381			 * Don't allocate a letter-number for swapper:0
1382			 * as a shortname. Instead, we use '.' for it.
1383			 */
1384			sched_in->shortname[0] = '.';
1385			sched_in->shortname[1] = ' ';
1386		} else {
1387			sched_in->shortname[0] = sched->next_shortname1;
1388			sched_in->shortname[1] = sched->next_shortname2;
1389
1390			if (sched->next_shortname1 < 'Z') {
1391				sched->next_shortname1++;
1392			} else {
1393				sched->next_shortname1 = 'A';
1394				if (sched->next_shortname2 < '9')
1395					sched->next_shortname2++;
1396				else
1397					sched->next_shortname2 = '0';
1398			}
1399		}
1400		new_shortname = 1;
1401	}
1402
1403	for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1404		if (cpu != this_cpu)
1405			printf(" ");
1406		else
1407			printf("*");
1408
1409		if (sched->curr_thread[cpu])
1410			printf("%2s ", sched->curr_thread[cpu]->shortname);
1411		else
 
 
 
1412			printf("   ");
1413	}
1414
1415	printf("  %12.6f secs ", (double)timestamp/1e9);
1416	if (new_shortname) {
1417		printf("%s => %s:%d\n",
1418		       sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1419	} else {
1420		printf("\n");
1421	}
 
1422
1423	thread__put(sched_in);
1424
1425	return 0;
1426}
1427
1428static int process_sched_switch_event(struct perf_tool *tool,
1429				      struct perf_evsel *evsel,
1430				      struct perf_sample *sample,
1431				      struct machine *machine)
1432{
1433	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1434	int this_cpu = sample->cpu, err = 0;
1435	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1436	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 
 
 
 
 
 
1437
1438	if (sched->curr_pid[this_cpu] != (u32)-1) {
1439		/*
1440		 * Are we trying to switch away a PID that is
1441		 * not current?
1442		 */
1443		if (sched->curr_pid[this_cpu] != prev_pid)
1444			sched->nr_context_switch_bugs++;
1445	}
 
 
 
1446
1447	if (sched->tp_handler->switch_event)
1448		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1449
1450	sched->curr_pid[this_cpu] = next_pid;
1451	return err;
1452}
1453
1454static int process_sched_runtime_event(struct perf_tool *tool,
1455				       struct perf_evsel *evsel,
1456				       struct perf_sample *sample,
1457				       struct machine *machine)
1458{
1459	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
 
 
 
 
 
 
 
1460
1461	if (sched->tp_handler->runtime_event)
1462		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
 
1463
1464	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1465}
1466
1467static int perf_sched__process_fork_event(struct perf_tool *tool,
1468					  union perf_event *event,
1469					  struct perf_sample *sample,
1470					  struct machine *machine)
1471{
1472	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1473
1474	/* run the fork event through the perf machineruy */
1475	perf_event__process_fork(tool, event, sample, machine);
1476
1477	/* and then run additional processing needed for this command */
1478	if (sched->tp_handler->fork_event)
1479		return sched->tp_handler->fork_event(sched, event, machine);
1480
1481	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1482}
1483
1484static int process_sched_migrate_task_event(struct perf_tool *tool,
1485					    struct perf_evsel *evsel,
1486					    struct perf_sample *sample,
1487					    struct machine *machine)
1488{
1489	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1490
1491	if (sched->tp_handler->migrate_task_event)
1492		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1493
1494	return 0;
1495}
 
 
 
 
1496
1497typedef int (*tracepoint_handler)(struct perf_tool *tool,
1498				  struct perf_evsel *evsel,
1499				  struct perf_sample *sample,
1500				  struct machine *machine);
1501
1502static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1503						 union perf_event *event __maybe_unused,
1504						 struct perf_sample *sample,
1505						 struct perf_evsel *evsel,
1506						 struct machine *machine)
1507{
1508	int err = 0;
1509
1510	if (evsel->handler != NULL) {
1511		tracepoint_handler f = evsel->handler;
1512		err = f(tool, evsel, sample, machine);
1513	}
1514
1515	return err;
1516}
1517
1518static int perf_sched__read_events(struct perf_sched *sched)
 
 
 
 
 
 
 
 
1519{
1520	const struct perf_evsel_str_handler handlers[] = {
1521		{ "sched:sched_switch",	      process_sched_switch_event, },
1522		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1523		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1524		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1525		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1526	};
1527	struct perf_session *session;
1528	struct perf_data_file file = {
1529		.path = input_name,
1530		.mode = PERF_DATA_MODE_READ,
1531		.force = sched->force,
1532	};
1533	int rc = -1;
1534
1535	session = perf_session__new(&file, false, &sched->tool);
1536	if (session == NULL) {
1537		pr_debug("No Memory for session\n");
1538		return -1;
1539	}
1540
1541	symbol__init(&session->header.env);
1542
1543	if (perf_session__set_tracepoints_handlers(session, handlers))
1544		goto out_delete;
1545
1546	if (perf_session__has_traces(session, "record -R")) {
1547		int err = perf_session__process_events(session);
1548		if (err) {
1549			pr_err("Failed to process events, error %d", err);
1550			goto out_delete;
1551		}
1552
1553		sched->nr_events      = session->evlist->stats.nr_events[0];
1554		sched->nr_lost_events = session->evlist->stats.total_lost;
1555		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1556	}
1557
1558	rc = 0;
1559out_delete:
1560	perf_session__delete(session);
1561	return rc;
 
1562}
1563
1564static void print_bad_events(struct perf_sched *sched)
1565{
1566	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1567		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1568			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1569			sched->nr_unordered_timestamps, sched->nr_timestamps);
1570	}
1571	if (sched->nr_lost_events && sched->nr_events) {
1572		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1573			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1574			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1575	}
1576	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
 
 
 
 
 
 
 
 
1577		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1578			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1579			sched->nr_context_switch_bugs, sched->nr_timestamps);
1580		if (sched->nr_lost_events)
1581			printf(" (due to lost events?)");
1582		printf("\n");
1583	}
1584}
1585
1586static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
1587{
1588	struct rb_node **new = &(root->rb_node), *parent = NULL;
1589	struct work_atoms *this;
1590	const char *comm = thread__comm_str(data->thread), *this_comm;
1591
1592	while (*new) {
1593		int cmp;
1594
1595		this = container_of(*new, struct work_atoms, node);
1596		parent = *new;
1597
1598		this_comm = thread__comm_str(this->thread);
1599		cmp = strcmp(comm, this_comm);
1600		if (cmp > 0) {
1601			new = &((*new)->rb_left);
1602		} else if (cmp < 0) {
1603			new = &((*new)->rb_right);
1604		} else {
1605			this->num_merged++;
1606			this->total_runtime += data->total_runtime;
1607			this->nb_atoms += data->nb_atoms;
1608			this->total_lat += data->total_lat;
1609			list_splice(&data->work_list, &this->work_list);
1610			if (this->max_lat < data->max_lat) {
1611				this->max_lat = data->max_lat;
1612				this->max_lat_at = data->max_lat_at;
1613			}
1614			zfree(&data);
1615			return;
1616		}
1617	}
1618
1619	data->num_merged++;
1620	rb_link_node(&data->node, parent, new);
1621	rb_insert_color(&data->node, root);
1622}
1623
1624static void perf_sched__merge_lat(struct perf_sched *sched)
1625{
1626	struct work_atoms *data;
1627	struct rb_node *node;
1628
1629	if (sched->skip_merge)
1630		return;
1631
1632	while ((node = rb_first(&sched->atom_root))) {
1633		rb_erase(node, &sched->atom_root);
1634		data = rb_entry(node, struct work_atoms, node);
1635		__merge_work_atoms(&sched->merged_atom_root, data);
1636	}
1637}
1638
1639static int perf_sched__lat(struct perf_sched *sched)
1640{
1641	struct rb_node *next;
 
1642
1643	setup_pager();
 
 
1644
1645	if (perf_sched__read_events(sched))
1646		return -1;
1647
1648	perf_sched__merge_lat(sched);
1649	perf_sched__sort_lat(sched);
1650
1651	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1652	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1653	printf(" -----------------------------------------------------------------------------------------------------------------\n");
1654
1655	next = rb_first(&sched->sorted_atom_root);
1656
1657	while (next) {
1658		struct work_atoms *work_list;
1659
1660		work_list = rb_entry(next, struct work_atoms, node);
1661		output_lat_thread(sched, work_list);
1662		next = rb_next(next);
1663		thread__zput(work_list->thread);
1664	}
1665
1666	printf(" -----------------------------------------------------------------------------------------------------------------\n");
1667	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1668		(double)sched->all_runtime / 1e6, sched->all_count);
1669
1670	printf(" ---------------------------------------------------\n");
1671
1672	print_bad_events(sched);
1673	printf("\n");
1674
1675	return 0;
1676}
1677
1678static int perf_sched__map(struct perf_sched *sched)
 
 
 
 
 
 
 
1679{
1680	sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1681
1682	setup_pager();
1683	if (perf_sched__read_events(sched))
1684		return -1;
1685	print_bad_events(sched);
1686	return 0;
1687}
1688
1689static int perf_sched__replay(struct perf_sched *sched)
1690{
1691	unsigned long i;
1692
1693	calibrate_run_measurement_overhead(sched);
1694	calibrate_sleep_measurement_overhead(sched);
1695
1696	test_calibrations(sched);
1697
1698	if (perf_sched__read_events(sched))
1699		return -1;
1700
1701	printf("nr_run_events:        %ld\n", sched->nr_run_events);
1702	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1703	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1704
1705	if (sched->targetless_wakeups)
1706		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1707	if (sched->multitarget_wakeups)
1708		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1709	if (sched->nr_run_events_optimized)
1710		printf("run atoms optimized: %ld\n",
1711			sched->nr_run_events_optimized);
1712
1713	print_task_traces(sched);
1714	add_cross_task_wakeups(sched);
1715
1716	create_tasks(sched);
1717	printf("------------------------------------------------------------\n");
1718	for (i = 0; i < sched->replay_repeat; i++)
1719		run_one_test(sched);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1720
1721	return 0;
1722}
 
 
 
 
 
 
 
1723
1724static void setup_sorting(struct perf_sched *sched, const struct option *options,
1725			  const char * const usage_msg[])
1726{
1727	char *tmp, *tok, *str = strdup(sched->sort_order);
1728
1729	for (tok = strtok_r(str, ", ", &tmp);
1730			tok; tok = strtok_r(NULL, ", ", &tmp)) {
1731		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1732			usage_with_options_msg(usage_msg, options,
1733					"Unknown --sort key: `%s'", tok);
1734		}
1735	}
1736
1737	free(str);
1738
1739	sort_dimension__add("pid", &sched->cmp_pid);
1740}
1741
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1742static int __cmd_record(int argc, const char **argv)
1743{
1744	unsigned int rec_argc, i, j;
1745	const char **rec_argv;
1746	const char * const record_args[] = {
1747		"record",
1748		"-a",
1749		"-R",
1750		"-m", "1024",
1751		"-c", "1",
1752		"-e", "sched:sched_switch",
1753		"-e", "sched:sched_stat_wait",
1754		"-e", "sched:sched_stat_sleep",
1755		"-e", "sched:sched_stat_iowait",
1756		"-e", "sched:sched_stat_runtime",
1757		"-e", "sched:sched_process_fork",
1758		"-e", "sched:sched_wakeup",
1759		"-e", "sched:sched_wakeup_new",
1760		"-e", "sched:sched_migrate_task",
1761	};
1762
1763	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1764	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1765
1766	if (rec_argv == NULL)
1767		return -ENOMEM;
1768
1769	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1770		rec_argv[i] = strdup(record_args[i]);
1771
1772	for (j = 1; j < (unsigned int)argc; j++, i++)
1773		rec_argv[i] = argv[j];
1774
1775	BUG_ON(i != rec_argc);
1776
1777	return cmd_record(i, rec_argv, NULL);
1778}
1779
1780int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1781{
1782	const char default_sort_order[] = "avg, max, switch, runtime";
1783	struct perf_sched sched = {
1784		.tool = {
1785			.sample		 = perf_sched__process_tracepoint_sample,
1786			.comm		 = perf_event__process_comm,
1787			.lost		 = perf_event__process_lost,
1788			.fork		 = perf_sched__process_fork_event,
1789			.ordered_events = true,
1790		},
1791		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
1792		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
1793		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1794		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1795		.sort_order	      = default_sort_order,
1796		.replay_repeat	      = 10,
1797		.profile_cpu	      = -1,
1798		.next_shortname1      = 'A',
1799		.next_shortname2      = '0',
1800		.skip_merge           = 0,
1801	};
1802	const struct option latency_options[] = {
1803	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1804		   "sort by key(s): runtime, switch, avg, max"),
1805	OPT_INCR('v', "verbose", &verbose,
1806		    "be more verbose (show symbol address, etc)"),
1807	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1808		    "CPU to profile on"),
1809	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1810		    "dump raw trace in ASCII"),
1811	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
1812		    "latency stats per pid instead of per comm"),
1813	OPT_END()
1814	};
1815	const struct option replay_options[] = {
1816	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1817		     "repeat the workload replay N times (-1: infinite)"),
1818	OPT_INCR('v', "verbose", &verbose,
1819		    "be more verbose (show symbol address, etc)"),
1820	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1821		    "dump raw trace in ASCII"),
1822	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
1823	OPT_END()
1824	};
1825	const struct option sched_options[] = {
1826	OPT_STRING('i', "input", &input_name, "file",
1827		    "input file name"),
1828	OPT_INCR('v', "verbose", &verbose,
1829		    "be more verbose (show symbol address, etc)"),
1830	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1831		    "dump raw trace in ASCII"),
1832	OPT_END()
1833	};
1834	const char * const latency_usage[] = {
1835		"perf sched latency [<options>]",
1836		NULL
1837	};
1838	const char * const replay_usage[] = {
1839		"perf sched replay [<options>]",
1840		NULL
1841	};
1842	const char *const sched_subcommands[] = { "record", "latency", "map",
1843						  "replay", "script", NULL };
1844	const char *sched_usage[] = {
1845		NULL,
1846		NULL
1847	};
1848	struct trace_sched_handler lat_ops  = {
1849		.wakeup_event	    = latency_wakeup_event,
1850		.switch_event	    = latency_switch_event,
1851		.runtime_event	    = latency_runtime_event,
1852		.migrate_task_event = latency_migrate_task_event,
1853	};
1854	struct trace_sched_handler map_ops  = {
1855		.switch_event	    = map_switch_event,
1856	};
1857	struct trace_sched_handler replay_ops  = {
1858		.wakeup_event	    = replay_wakeup_event,
1859		.switch_event	    = replay_switch_event,
1860		.fork_event	    = replay_fork_event,
1861	};
1862	unsigned int i;
1863
1864	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1865		sched.curr_pid[i] = -1;
1866
1867	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1868					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1869	if (!argc)
1870		usage_with_options(sched_usage, sched_options);
1871
1872	/*
1873	 * Aliased to 'perf script' for now:
1874	 */
1875	if (!strcmp(argv[0], "script"))
1876		return cmd_script(argc, argv, prefix);
1877
 
1878	if (!strncmp(argv[0], "rec", 3)) {
1879		return __cmd_record(argc, argv);
1880	} else if (!strncmp(argv[0], "lat", 3)) {
1881		sched.tp_handler = &lat_ops;
1882		if (argc > 1) {
1883			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1884			if (argc)
1885				usage_with_options(latency_usage, latency_options);
1886		}
1887		setup_sorting(&sched, latency_options, latency_usage);
1888		return perf_sched__lat(&sched);
1889	} else if (!strcmp(argv[0], "map")) {
1890		sched.tp_handler = &map_ops;
1891		setup_sorting(&sched, latency_options, latency_usage);
1892		return perf_sched__map(&sched);
1893	} else if (!strncmp(argv[0], "rep", 3)) {
1894		sched.tp_handler = &replay_ops;
1895		if (argc) {
1896			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1897			if (argc)
1898				usage_with_options(replay_usage, replay_options);
1899		}
1900		return perf_sched__replay(&sched);
1901	} else {
1902		usage_with_options(sched_usage, sched_options);
1903	}
1904
1905	return 0;
1906}