Linux Audio

Check our new training course

Loading...
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Simple CPU accounting cgroup controller
   4 */
 
 
   5#include "sched.h"
 
 
 
 
   6
   7#ifdef CONFIG_IRQ_TIME_ACCOUNTING
   8
   9/*
  10 * There are no locks covering percpu hardirq/softirq time.
  11 * They are only modified in vtime_account, on corresponding CPU
  12 * with interrupts disabled. So, writes are safe.
  13 * They are read and saved off onto struct rq in update_rq_clock().
  14 * This may result in other CPU reading this CPU's irq time and can
  15 * race with irq/vtime_account on this CPU. We would either get old
  16 * or new value with a side effect of accounting a slice of irq time to wrong
  17 * task when irq is in progress while we read rq->clock. That is a worthy
  18 * compromise in place of having locks on each irq in account_system_time.
  19 */
  20DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
  21
  22static int sched_clock_irqtime;
  23
  24void enable_sched_clock_irqtime(void)
  25{
  26	sched_clock_irqtime = 1;
  27}
  28
  29void disable_sched_clock_irqtime(void)
  30{
  31	sched_clock_irqtime = 0;
  32}
  33
  34static void irqtime_account_delta(struct irqtime *irqtime, u64 delta,
  35				  enum cpu_usage_stat idx)
  36{
  37	u64 *cpustat = kcpustat_this_cpu->cpustat;
  38
  39	u64_stats_update_begin(&irqtime->sync);
  40	cpustat[idx] += delta;
  41	irqtime->total += delta;
  42	irqtime->tick_delta += delta;
  43	u64_stats_update_end(&irqtime->sync);
  44}
  45
  46/*
  47 * Called after incrementing preempt_count on {soft,}irq_enter
  48 * and before decrementing preempt_count on {soft,}irq_exit.
  49 */
  50void irqtime_account_irq(struct task_struct *curr, unsigned int offset)
  51{
  52	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
  53	unsigned int pc;
  54	s64 delta;
  55	int cpu;
  56
  57	if (!sched_clock_irqtime)
  58		return;
  59
  60	cpu = smp_processor_id();
  61	delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
  62	irqtime->irq_start_time += delta;
  63	pc = irq_count() - offset;
  64
 
  65	/*
  66	 * We do not account for softirq time from ksoftirqd here.
  67	 * We want to continue accounting softirq time to ksoftirqd thread
  68	 * in that case, so as not to confuse scheduler with a special task
  69	 * that do not consume any time, but still wants to run.
  70	 */
  71	if (pc & HARDIRQ_MASK)
  72		irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
  73	else if ((pc & SOFTIRQ_OFFSET) && curr != this_cpu_ksoftirqd())
  74		irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
 
 
  75}
 
  76
  77static u64 irqtime_tick_accounted(u64 maxtime)
  78{
  79	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
  80	u64 delta;
  81
  82	delta = min(irqtime->tick_delta, maxtime);
  83	irqtime->tick_delta -= delta;
 
  84
  85	return delta;
 
 
 
 
 
 
 
 
 
 
 
 
  86}
  87
  88#else /* CONFIG_IRQ_TIME_ACCOUNTING */
  89
  90#define sched_clock_irqtime	(0)
  91
  92static u64 irqtime_tick_accounted(u64 dummy)
 
 
 
 
 
  93{
  94	return 0;
  95}
  96
  97#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
  98
  99static inline void task_group_account_field(struct task_struct *p, int index,
 100					    u64 tmp)
 101{
 102	/*
 103	 * Since all updates are sure to touch the root cgroup, we
 104	 * get ourselves ahead and touch it first. If the root cgroup
 105	 * is the only cgroup, then nothing else should be necessary.
 106	 *
 107	 */
 108	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
 109
 110	cgroup_account_cputime_field(p, index, tmp);
 111}
 112
 113/*
 114 * Account user CPU time to a process.
 115 * @p: the process that the CPU time gets accounted to
 116 * @cputime: the CPU time spent in user space since the last update
 117 */
 118void account_user_time(struct task_struct *p, u64 cputime)
 119{
 120	int index;
 121
 122	/* Add user time to process. */
 123	p->utime += cputime;
 124	account_group_user_time(p, cputime);
 125
 126	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
 127
 128	/* Add user time to cpustat. */
 129	task_group_account_field(p, index, cputime);
 130
 131	/* Account for user time used */
 132	acct_account_cputime(p);
 133}
 134
 135/*
 136 * Account guest CPU time to a process.
 137 * @p: the process that the CPU time gets accounted to
 138 * @cputime: the CPU time spent in virtual machine since the last update
 139 */
 140void account_guest_time(struct task_struct *p, u64 cputime)
 141{
 142	u64 *cpustat = kcpustat_this_cpu->cpustat;
 143
 144	/* Add guest time to process. */
 145	p->utime += cputime;
 146	account_group_user_time(p, cputime);
 147	p->gtime += cputime;
 148
 149	/* Add guest time to cpustat. */
 150	if (task_nice(p) > 0) {
 151		cpustat[CPUTIME_NICE] += cputime;
 152		cpustat[CPUTIME_GUEST_NICE] += cputime;
 153	} else {
 154		cpustat[CPUTIME_USER] += cputime;
 155		cpustat[CPUTIME_GUEST] += cputime;
 156	}
 157}
 158
 159/*
 160 * Account system CPU time to a process and desired cpustat field
 161 * @p: the process that the CPU time gets accounted to
 162 * @cputime: the CPU time spent in kernel space since the last update
 163 * @index: pointer to cpustat field that has to be updated
 164 */
 165void account_system_index_time(struct task_struct *p,
 166			       u64 cputime, enum cpu_usage_stat index)
 167{
 168	/* Add system time to process. */
 169	p->stime += cputime;
 170	account_group_system_time(p, cputime);
 171
 172	/* Add system time to cpustat. */
 173	task_group_account_field(p, index, cputime);
 174
 175	/* Account for system time used */
 176	acct_account_cputime(p);
 177}
 178
 179/*
 180 * Account system CPU time to a process.
 181 * @p: the process that the CPU time gets accounted to
 182 * @hardirq_offset: the offset to subtract from hardirq_count()
 183 * @cputime: the CPU time spent in kernel space since the last update
 184 */
 185void account_system_time(struct task_struct *p, int hardirq_offset, u64 cputime)
 
 186{
 187	int index;
 188
 189	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
 190		account_guest_time(p, cputime);
 191		return;
 192	}
 193
 194	if (hardirq_count() - hardirq_offset)
 195		index = CPUTIME_IRQ;
 196	else if (in_serving_softirq())
 197		index = CPUTIME_SOFTIRQ;
 198	else
 199		index = CPUTIME_SYSTEM;
 200
 201	account_system_index_time(p, cputime, index);
 202}
 203
 204/*
 205 * Account for involuntary wait time.
 206 * @cputime: the CPU time spent in involuntary wait
 207 */
 208void account_steal_time(u64 cputime)
 209{
 210	u64 *cpustat = kcpustat_this_cpu->cpustat;
 211
 212	cpustat[CPUTIME_STEAL] += cputime;
 213}
 214
 215/*
 216 * Account for idle time.
 217 * @cputime: the CPU time spent in idle wait
 218 */
 219void account_idle_time(u64 cputime)
 220{
 221	u64 *cpustat = kcpustat_this_cpu->cpustat;
 222	struct rq *rq = this_rq();
 223
 224	if (atomic_read(&rq->nr_iowait) > 0)
 225		cpustat[CPUTIME_IOWAIT] += cputime;
 226	else
 227		cpustat[CPUTIME_IDLE] += cputime;
 228}
 229
 230/*
 231 * When a guest is interrupted for a longer amount of time, missed clock
 232 * ticks are not redelivered later. Due to that, this function may on
 233 * occasion account more time than the calling functions think elapsed.
 234 */
 235static __always_inline u64 steal_account_process_time(u64 maxtime)
 236{
 237#ifdef CONFIG_PARAVIRT
 238	if (static_key_false(&paravirt_steal_enabled)) {
 
 239		u64 steal;
 240
 241		steal = paravirt_steal_clock(smp_processor_id());
 242		steal -= this_rq()->prev_steal_time;
 243		steal = min(steal, maxtime);
 244		account_steal_time(steal);
 245		this_rq()->prev_steal_time += steal;
 246
 247		return steal;
 
 
 
 
 248	}
 249#endif
 250	return 0;
 251}
 252
 253/*
 254 * Account how much elapsed time was spent in steal, irq, or softirq time.
 255 */
 256static inline u64 account_other_time(u64 max)
 257{
 258	u64 accounted;
 259
 260	lockdep_assert_irqs_disabled();
 
 261
 262	accounted = steal_account_process_time(max);
 263
 264	if (accounted < max)
 265		accounted += irqtime_tick_accounted(max - accounted);
 
 
 
 266
 267	return accounted;
 268}
 269
 270#ifdef CONFIG_64BIT
 271static inline u64 read_sum_exec_runtime(struct task_struct *t)
 272{
 273	return t->se.sum_exec_runtime;
 274}
 275#else
 276static u64 read_sum_exec_runtime(struct task_struct *t)
 277{
 278	u64 ns;
 279	struct rq_flags rf;
 280	struct rq *rq;
 281
 282	rq = task_rq_lock(t, &rf);
 283	ns = t->se.sum_exec_runtime;
 284	task_rq_unlock(rq, t, &rf);
 285
 286	return ns;
 287}
 288#endif
 289
 290/*
 291 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
 292 * tasks (sum on group iteration) belonging to @tsk's group.
 293 */
 294void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
 295{
 296	struct signal_struct *sig = tsk->signal;
 297	u64 utime, stime;
 298	struct task_struct *t;
 299	unsigned int seq, nextseq;
 300	unsigned long flags;
 301
 302	/*
 303	 * Update current task runtime to account pending time since last
 304	 * scheduler action or thread_group_cputime() call. This thread group
 305	 * might have other running tasks on different CPUs, but updating
 306	 * their runtime can affect syscall performance, so we skip account
 307	 * those pending times and rely only on values updated on tick or
 308	 * other scheduler action.
 309	 */
 310	if (same_thread_group(current, tsk))
 311		(void) task_sched_runtime(current);
 312
 313	rcu_read_lock();
 314	/* Attempt a lockless read on the first round. */
 315	nextseq = 0;
 316	do {
 317		seq = nextseq;
 318		flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
 319		times->utime = sig->utime;
 320		times->stime = sig->stime;
 321		times->sum_exec_runtime = sig->sum_sched_runtime;
 322
 323		for_each_thread(tsk, t) {
 324			task_cputime(t, &utime, &stime);
 325			times->utime += utime;
 326			times->stime += stime;
 327			times->sum_exec_runtime += read_sum_exec_runtime(t);
 328		}
 329		/* If lockless access failed, take the lock. */
 330		nextseq = 1;
 331	} while (need_seqretry(&sig->stats_lock, seq));
 332	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
 333	rcu_read_unlock();
 334}
 335
 336#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 337/*
 338 * Account a tick to a process and cpustat
 339 * @p: the process that the CPU time gets accounted to
 340 * @user_tick: is the tick from userspace
 341 * @rq: the pointer to rq
 342 *
 343 * Tick demultiplexing follows the order
 344 * - pending hardirq update
 345 * - pending softirq update
 346 * - user_time
 347 * - idle_time
 348 * - system time
 349 *   - check for guest_time
 350 *   - else account as system_time
 351 *
 352 * Check for hardirq is done both for system and user time as there is
 353 * no timer going off while we are on hardirq and hence we may never get an
 354 * opportunity to update it solely in system time.
 355 * p->stime and friends are only updated on system time and not on irq
 356 * softirq as those do not count in task exec_runtime any more.
 357 */
 358static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 359					 int ticks)
 360{
 361	u64 other, cputime = TICK_NSEC * ticks;
 
 362
 363	/*
 364	 * When returning from idle, many ticks can get accounted at
 365	 * once, including some ticks of steal, irq, and softirq time.
 366	 * Subtract those ticks from the amount of time accounted to
 367	 * idle, or potentially user or system time. Due to rounding,
 368	 * other time can exceed ticks occasionally.
 369	 */
 370	other = account_other_time(ULONG_MAX);
 371	if (other >= cputime)
 372		return;
 373
 374	cputime -= other;
 375
 376	if (this_cpu_ksoftirqd() == p) {
 377		/*
 378		 * ksoftirqd time do not get accounted in cpu_softirq_time.
 379		 * So, we have to handle it separately here.
 380		 * Also, p->stime needs to be updated for ksoftirqd.
 381		 */
 382		account_system_index_time(p, cputime, CPUTIME_SOFTIRQ);
 383	} else if (user_tick) {
 384		account_user_time(p, cputime);
 385	} else if (p == this_rq()->idle) {
 386		account_idle_time(cputime);
 387	} else if (p->flags & PF_VCPU) { /* System time or guest time */
 388		account_guest_time(p, cputime);
 389	} else {
 390		account_system_index_time(p, cputime, CPUTIME_SYSTEM);
 391	}
 392}
 393
 394static void irqtime_account_idle_ticks(int ticks)
 395{
 396	irqtime_account_process_tick(current, 0, ticks);
 
 
 397}
 398#else /* CONFIG_IRQ_TIME_ACCOUNTING */
 399static inline void irqtime_account_idle_ticks(int ticks) { }
 400static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
 401						int nr_ticks) { }
 402#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
 403
 404/*
 405 * Use precise platform statistics if available:
 406 */
 407#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 408
 409# ifndef __ARCH_HAS_VTIME_TASK_SWITCH
 410void vtime_task_switch(struct task_struct *prev)
 411{
 412	if (is_idle_task(prev))
 413		vtime_account_idle(prev);
 414	else
 415		vtime_account_kernel(prev);
 416
 417	vtime_flush(prev);
 
 
 418	arch_vtime_task_switch(prev);
 419}
 420# endif
 421
 422void vtime_account_irq(struct task_struct *tsk, unsigned int offset)
 423{
 424	unsigned int pc = irq_count() - offset;
 425
 426	if (pc & HARDIRQ_OFFSET) {
 427		vtime_account_hardirq(tsk);
 428	} else if (pc & SOFTIRQ_OFFSET) {
 429		vtime_account_softirq(tsk);
 430	} else if (!IS_ENABLED(CONFIG_HAVE_VIRT_CPU_ACCOUNTING_IDLE) &&
 431		   is_idle_task(tsk)) {
 432		vtime_account_idle(tsk);
 433	} else {
 434		vtime_account_kernel(tsk);
 435	}
 436}
 437
 438void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
 439		    u64 *ut, u64 *st)
 
 
 
 
 
 
 
 
 
 440{
 441	*ut = curr->utime;
 442	*st = curr->stime;
 
 
 443}
 
 
 444
 445void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 446{
 447	*ut = p->utime;
 448	*st = p->stime;
 449}
 450EXPORT_SYMBOL_GPL(task_cputime_adjusted);
 451
 452void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 453{
 454	struct task_cputime cputime;
 455
 456	thread_group_cputime(p, &cputime);
 457
 458	*ut = cputime.utime;
 459	*st = cputime.stime;
 460}
 461
 462#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE: */
 463
 464/*
 465 * Account a single tick of CPU time.
 466 * @p: the process that the CPU time gets accounted to
 467 * @user_tick: indicates if the tick is a user or a system tick
 468 */
 469void account_process_tick(struct task_struct *p, int user_tick)
 470{
 471	u64 cputime, steal;
 
 472
 473	if (vtime_accounting_enabled_this_cpu())
 474		return;
 475
 476	if (sched_clock_irqtime) {
 477		irqtime_account_process_tick(p, user_tick, 1);
 478		return;
 479	}
 480
 481	cputime = TICK_NSEC;
 482	steal = steal_account_process_time(ULONG_MAX);
 483
 484	if (steal >= cputime)
 485		return;
 486
 487	cputime -= steal;
 488
 489	if (user_tick)
 490		account_user_time(p, cputime);
 491	else if ((p != this_rq()->idle) || (irq_count() != HARDIRQ_OFFSET))
 492		account_system_time(p, HARDIRQ_OFFSET, cputime);
 493	else
 494		account_idle_time(cputime);
 495}
 496
 497/*
 498 * Account multiple ticks of idle time.
 499 * @ticks: number of stolen ticks
 500 */
 501void account_idle_ticks(unsigned long ticks)
 502{
 503	u64 cputime, steal;
 504
 505	if (sched_clock_irqtime) {
 506		irqtime_account_idle_ticks(ticks);
 507		return;
 508	}
 509
 510	cputime = ticks * TICK_NSEC;
 511	steal = steal_account_process_time(ULONG_MAX);
 512
 513	if (steal >= cputime)
 514		return;
 515
 516	cputime -= steal;
 517	account_idle_time(cputime);
 518}
 519
 520/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 521 * Adjust tick based cputime random precision against scheduler runtime
 522 * accounting.
 523 *
 524 * Tick based cputime accounting depend on random scheduling timeslices of a
 525 * task to be interrupted or not by the timer.  Depending on these
 526 * circumstances, the number of these interrupts may be over or
 527 * under-optimistic, matching the real user and system cputime with a variable
 528 * precision.
 529 *
 530 * Fix this by scaling these tick based values against the total runtime
 531 * accounted by the CFS scheduler.
 532 *
 533 * This code provides the following guarantees:
 534 *
 535 *   stime + utime == rtime
 536 *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
 537 *
 538 * Assuming that rtime_i+1 >= rtime_i.
 539 */
 540void cputime_adjust(struct task_cputime *curr, struct prev_cputime *prev,
 541		    u64 *ut, u64 *st)
 
 542{
 543	u64 rtime, stime, utime;
 544	unsigned long flags;
 545
 546	/* Serialize concurrent callers such that we can honour our guarantees */
 547	raw_spin_lock_irqsave(&prev->lock, flags);
 548	rtime = curr->sum_exec_runtime;
 549
 550	/*
 551	 * This is possible under two circumstances:
 552	 *  - rtime isn't monotonic after all (a bug);
 553	 *  - we got reordered by the lock.
 554	 *
 555	 * In both cases this acts as a filter such that the rest of the code
 556	 * can assume it is monotonic regardless of anything else.
 557	 */
 558	if (prev->stime + prev->utime >= rtime)
 559		goto out;
 560
 561	stime = curr->stime;
 562	utime = curr->utime;
 563
 564	/*
 565	 * If either stime or utime are 0, assume all runtime is userspace.
 566	 * Once a task gets some ticks, the monotonicity code at 'update:'
 567	 * will ensure things converge to the observed ratio.
 568	 */
 569	if (stime == 0) {
 570		utime = rtime;
 571		goto update;
 572	}
 573
 574	if (utime == 0) {
 575		stime = rtime;
 576		goto update;
 577	}
 578
 579	stime = mul_u64_u64_div_u64(stime, rtime, stime + utime);
 
 580
 581update:
 582	/*
 583	 * Make sure stime doesn't go backwards; this preserves monotonicity
 584	 * for utime because rtime is monotonic.
 585	 *
 586	 *  utime_i+1 = rtime_i+1 - stime_i
 587	 *            = rtime_i+1 - (rtime_i - utime_i)
 588	 *            = (rtime_i+1 - rtime_i) + utime_i
 589	 *            >= utime_i
 590	 */
 591	if (stime < prev->stime)
 592		stime = prev->stime;
 593	utime = rtime - stime;
 594
 595	/*
 596	 * Make sure utime doesn't go backwards; this still preserves
 597	 * monotonicity for stime, analogous argument to above.
 598	 */
 599	if (utime < prev->utime) {
 600		utime = prev->utime;
 601		stime = rtime - utime;
 602	}
 603
 604	prev->stime = stime;
 605	prev->utime = utime;
 606out:
 607	*ut = prev->utime;
 608	*st = prev->stime;
 609	raw_spin_unlock_irqrestore(&prev->lock, flags);
 610}
 611
 612void task_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 613{
 614	struct task_cputime cputime = {
 615		.sum_exec_runtime = p->se.sum_exec_runtime,
 616	};
 617
 618	task_cputime(p, &cputime.utime, &cputime.stime);
 619	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
 620}
 621EXPORT_SYMBOL_GPL(task_cputime_adjusted);
 622
 623void thread_group_cputime_adjusted(struct task_struct *p, u64 *ut, u64 *st)
 624{
 625	struct task_cputime cputime;
 626
 627	thread_group_cputime(p, &cputime);
 628	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
 629}
 630#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 631
 632#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
 633static u64 vtime_delta(struct vtime *vtime)
 634{
 635	unsigned long long clock;
 636
 637	clock = sched_clock();
 638	if (clock < vtime->starttime)
 639		return 0;
 640
 641	return clock - vtime->starttime;
 642}
 643
 644static u64 get_vtime_delta(struct vtime *vtime)
 645{
 646	u64 delta = vtime_delta(vtime);
 647	u64 other;
 648
 649	/*
 650	 * Unlike tick based timing, vtime based timing never has lost
 651	 * ticks, and no need for steal time accounting to make up for
 652	 * lost ticks. Vtime accounts a rounded version of actual
 653	 * elapsed time. Limit account_other_time to prevent rounding
 654	 * errors from causing elapsed vtime to go negative.
 655	 */
 
 656	other = account_other_time(delta);
 657	WARN_ON_ONCE(vtime->state == VTIME_INACTIVE);
 658	vtime->starttime += delta;
 659
 660	return delta - other;
 661}
 662
 663static void vtime_account_system(struct task_struct *tsk,
 664				 struct vtime *vtime)
 665{
 666	vtime->stime += get_vtime_delta(vtime);
 667	if (vtime->stime >= TICK_NSEC) {
 668		account_system_time(tsk, irq_count(), vtime->stime);
 669		vtime->stime = 0;
 670	}
 671}
 672
 673static void vtime_account_guest(struct task_struct *tsk,
 674				struct vtime *vtime)
 675{
 676	vtime->gtime += get_vtime_delta(vtime);
 677	if (vtime->gtime >= TICK_NSEC) {
 678		account_guest_time(tsk, vtime->gtime);
 679		vtime->gtime = 0;
 680	}
 681}
 682
 683static void __vtime_account_kernel(struct task_struct *tsk,
 684				   struct vtime *vtime)
 685{
 686	/* We might have scheduled out from guest path */
 687	if (vtime->state == VTIME_GUEST)
 688		vtime_account_guest(tsk, vtime);
 689	else
 690		vtime_account_system(tsk, vtime);
 691}
 692
 693void vtime_account_kernel(struct task_struct *tsk)
 694{
 695	struct vtime *vtime = &tsk->vtime;
 696
 697	if (!vtime_delta(vtime))
 698		return;
 699
 700	write_seqcount_begin(&vtime->seqcount);
 701	__vtime_account_kernel(tsk, vtime);
 702	write_seqcount_end(&vtime->seqcount);
 703}
 704
 705void vtime_user_enter(struct task_struct *tsk)
 706{
 707	struct vtime *vtime = &tsk->vtime;
 708
 709	write_seqcount_begin(&vtime->seqcount);
 710	vtime_account_system(tsk, vtime);
 711	vtime->state = VTIME_USER;
 712	write_seqcount_end(&vtime->seqcount);
 
 
 
 713}
 714
 715void vtime_user_exit(struct task_struct *tsk)
 716{
 717	struct vtime *vtime = &tsk->vtime;
 718
 719	write_seqcount_begin(&vtime->seqcount);
 720	vtime->utime += get_vtime_delta(vtime);
 721	if (vtime->utime >= TICK_NSEC) {
 722		account_user_time(tsk, vtime->utime);
 723		vtime->utime = 0;
 724	}
 725	vtime->state = VTIME_SYS;
 726	write_seqcount_end(&vtime->seqcount);
 727}
 728
 729void vtime_guest_enter(struct task_struct *tsk)
 730{
 731	struct vtime *vtime = &tsk->vtime;
 732	/*
 733	 * The flags must be updated under the lock with
 734	 * the vtime_starttime flush and update.
 735	 * That enforces a right ordering and update sequence
 736	 * synchronization against the reader (task_gtime())
 737	 * that can thus safely catch up with a tickless delta.
 738	 */
 739	write_seqcount_begin(&vtime->seqcount);
 740	vtime_account_system(tsk, vtime);
 741	tsk->flags |= PF_VCPU;
 742	vtime->state = VTIME_GUEST;
 743	write_seqcount_end(&vtime->seqcount);
 744}
 745EXPORT_SYMBOL_GPL(vtime_guest_enter);
 746
 747void vtime_guest_exit(struct task_struct *tsk)
 748{
 749	struct vtime *vtime = &tsk->vtime;
 750
 751	write_seqcount_begin(&vtime->seqcount);
 752	vtime_account_guest(tsk, vtime);
 753	tsk->flags &= ~PF_VCPU;
 754	vtime->state = VTIME_SYS;
 755	write_seqcount_end(&vtime->seqcount);
 756}
 757EXPORT_SYMBOL_GPL(vtime_guest_exit);
 758
 759void vtime_account_idle(struct task_struct *tsk)
 760{
 761	account_idle_time(get_vtime_delta(&tsk->vtime));
 
 
 762}
 763
 764void vtime_task_switch_generic(struct task_struct *prev)
 765{
 766	struct vtime *vtime = &prev->vtime;
 
 
 767
 768	write_seqcount_begin(&vtime->seqcount);
 769	if (vtime->state == VTIME_IDLE)
 770		vtime_account_idle(prev);
 771	else
 772		__vtime_account_kernel(prev, vtime);
 773	vtime->state = VTIME_INACTIVE;
 774	vtime->cpu = -1;
 775	write_seqcount_end(&vtime->seqcount);
 776
 777	vtime = &current->vtime;
 778
 779	write_seqcount_begin(&vtime->seqcount);
 780	if (is_idle_task(current))
 781		vtime->state = VTIME_IDLE;
 782	else if (current->flags & PF_VCPU)
 783		vtime->state = VTIME_GUEST;
 784	else
 785		vtime->state = VTIME_SYS;
 786	vtime->starttime = sched_clock();
 787	vtime->cpu = smp_processor_id();
 788	write_seqcount_end(&vtime->seqcount);
 789}
 790
 791void vtime_init_idle(struct task_struct *t, int cpu)
 792{
 793	struct vtime *vtime = &t->vtime;
 794	unsigned long flags;
 795
 796	local_irq_save(flags);
 797	write_seqcount_begin(&vtime->seqcount);
 798	vtime->state = VTIME_IDLE;
 799	vtime->starttime = sched_clock();
 800	vtime->cpu = cpu;
 801	write_seqcount_end(&vtime->seqcount);
 802	local_irq_restore(flags);
 803}
 804
 805u64 task_gtime(struct task_struct *t)
 806{
 807	struct vtime *vtime = &t->vtime;
 808	unsigned int seq;
 809	u64 gtime;
 810
 811	if (!vtime_accounting_enabled())
 812		return t->gtime;
 813
 814	do {
 815		seq = read_seqcount_begin(&vtime->seqcount);
 816
 817		gtime = t->gtime;
 818		if (vtime->state == VTIME_GUEST)
 819			gtime += vtime->gtime + vtime_delta(vtime);
 820
 821	} while (read_seqcount_retry(&vtime->seqcount, seq));
 822
 823	return gtime;
 824}
 825
 826/*
 827 * Fetch cputime raw values from fields of task_struct and
 828 * add up the pending nohz execution time since the last
 829 * cputime snapshot.
 830 */
 831void task_cputime(struct task_struct *t, u64 *utime, u64 *stime)
 832{
 833	struct vtime *vtime = &t->vtime;
 834	unsigned int seq;
 835	u64 delta;
 836
 837	if (!vtime_accounting_enabled()) {
 838		*utime = t->utime;
 839		*stime = t->stime;
 840		return;
 841	}
 842
 843	do {
 844		seq = read_seqcount_begin(&vtime->seqcount);
 845
 846		*utime = t->utime;
 847		*stime = t->stime;
 848
 849		/* Task is sleeping or idle, nothing to add */
 850		if (vtime->state < VTIME_SYS)
 851			continue;
 852
 853		delta = vtime_delta(vtime);
 854
 855		/*
 856		 * Task runs either in user (including guest) or kernel space,
 857		 * add pending nohz time to the right place.
 858		 */
 859		if (vtime->state == VTIME_SYS)
 860			*stime += vtime->stime + delta;
 861		else
 862			*utime += vtime->utime + delta;
 863	} while (read_seqcount_retry(&vtime->seqcount, seq));
 864}
 865
 866static int vtime_state_fetch(struct vtime *vtime, int cpu)
 867{
 868	int state = READ_ONCE(vtime->state);
 869
 870	/*
 871	 * We raced against a context switch, fetch the
 872	 * kcpustat task again.
 873	 */
 874	if (vtime->cpu != cpu && vtime->cpu != -1)
 875		return -EAGAIN;
 876
 877	/*
 878	 * Two possible things here:
 879	 * 1) We are seeing the scheduling out task (prev) or any past one.
 880	 * 2) We are seeing the scheduling in task (next) but it hasn't
 881	 *    passed though vtime_task_switch() yet so the pending
 882	 *    cputime of the prev task may not be flushed yet.
 883	 *
 884	 * Case 1) is ok but 2) is not. So wait for a safe VTIME state.
 885	 */
 886	if (state == VTIME_INACTIVE)
 887		return -EAGAIN;
 888
 889	return state;
 890}
 891
 892static u64 kcpustat_user_vtime(struct vtime *vtime)
 893{
 894	if (vtime->state == VTIME_USER)
 895		return vtime->utime + vtime_delta(vtime);
 896	else if (vtime->state == VTIME_GUEST)
 897		return vtime->gtime + vtime_delta(vtime);
 898	return 0;
 899}
 900
 901static int kcpustat_field_vtime(u64 *cpustat,
 902				struct task_struct *tsk,
 903				enum cpu_usage_stat usage,
 904				int cpu, u64 *val)
 905{
 906	struct vtime *vtime = &tsk->vtime;
 907	unsigned int seq;
 908
 909	do {
 910		int state;
 911
 912		seq = read_seqcount_begin(&vtime->seqcount);
 913
 914		state = vtime_state_fetch(vtime, cpu);
 915		if (state < 0)
 916			return state;
 917
 918		*val = cpustat[usage];
 919
 920		/*
 921		 * Nice VS unnice cputime accounting may be inaccurate if
 922		 * the nice value has changed since the last vtime update.
 923		 * But proper fix would involve interrupting target on nice
 924		 * updates which is a no go on nohz_full (although the scheduler
 925		 * may still interrupt the target if rescheduling is needed...)
 926		 */
 927		switch (usage) {
 928		case CPUTIME_SYSTEM:
 929			if (state == VTIME_SYS)
 930				*val += vtime->stime + vtime_delta(vtime);
 931			break;
 932		case CPUTIME_USER:
 933			if (task_nice(tsk) <= 0)
 934				*val += kcpustat_user_vtime(vtime);
 935			break;
 936		case CPUTIME_NICE:
 937			if (task_nice(tsk) > 0)
 938				*val += kcpustat_user_vtime(vtime);
 939			break;
 940		case CPUTIME_GUEST:
 941			if (state == VTIME_GUEST && task_nice(tsk) <= 0)
 942				*val += vtime->gtime + vtime_delta(vtime);
 943			break;
 944		case CPUTIME_GUEST_NICE:
 945			if (state == VTIME_GUEST && task_nice(tsk) > 0)
 946				*val += vtime->gtime + vtime_delta(vtime);
 947			break;
 948		default:
 949			break;
 950		}
 951	} while (read_seqcount_retry(&vtime->seqcount, seq));
 952
 953	return 0;
 954}
 955
 956u64 kcpustat_field(struct kernel_cpustat *kcpustat,
 957		   enum cpu_usage_stat usage, int cpu)
 958{
 959	u64 *cpustat = kcpustat->cpustat;
 960	u64 val = cpustat[usage];
 961	struct rq *rq;
 962	int err;
 963
 964	if (!vtime_accounting_enabled_cpu(cpu))
 965		return val;
 966
 967	rq = cpu_rq(cpu);
 968
 969	for (;;) {
 970		struct task_struct *curr;
 971
 972		rcu_read_lock();
 973		curr = rcu_dereference(rq->curr);
 974		if (WARN_ON_ONCE(!curr)) {
 975			rcu_read_unlock();
 976			return cpustat[usage];
 977		}
 978
 979		err = kcpustat_field_vtime(cpustat, curr, usage, cpu, &val);
 980		rcu_read_unlock();
 981
 982		if (!err)
 983			return val;
 984
 985		cpu_relax();
 986	}
 987}
 988EXPORT_SYMBOL_GPL(kcpustat_field);
 989
 990static int kcpustat_cpu_fetch_vtime(struct kernel_cpustat *dst,
 991				    const struct kernel_cpustat *src,
 992				    struct task_struct *tsk, int cpu)
 993{
 994	struct vtime *vtime = &tsk->vtime;
 995	unsigned int seq;
 996
 997	do {
 998		u64 *cpustat;
 999		u64 delta;
1000		int state;
1001
1002		seq = read_seqcount_begin(&vtime->seqcount);
1003
1004		state = vtime_state_fetch(vtime, cpu);
1005		if (state < 0)
1006			return state;
1007
1008		*dst = *src;
1009		cpustat = dst->cpustat;
1010
1011		/* Task is sleeping, dead or idle, nothing to add */
1012		if (state < VTIME_SYS)
1013			continue;
1014
1015		delta = vtime_delta(vtime);
1016
1017		/*
1018		 * Task runs either in user (including guest) or kernel space,
1019		 * add pending nohz time to the right place.
1020		 */
1021		if (state == VTIME_SYS) {
1022			cpustat[CPUTIME_SYSTEM] += vtime->stime + delta;
1023		} else if (state == VTIME_USER) {
1024			if (task_nice(tsk) > 0)
1025				cpustat[CPUTIME_NICE] += vtime->utime + delta;
1026			else
1027				cpustat[CPUTIME_USER] += vtime->utime + delta;
1028		} else {
1029			WARN_ON_ONCE(state != VTIME_GUEST);
1030			if (task_nice(tsk) > 0) {
1031				cpustat[CPUTIME_GUEST_NICE] += vtime->gtime + delta;
1032				cpustat[CPUTIME_NICE] += vtime->gtime + delta;
1033			} else {
1034				cpustat[CPUTIME_GUEST] += vtime->gtime + delta;
1035				cpustat[CPUTIME_USER] += vtime->gtime + delta;
1036			}
1037		}
1038	} while (read_seqcount_retry(&vtime->seqcount, seq));
1039
1040	return 0;
1041}
1042
1043void kcpustat_cpu_fetch(struct kernel_cpustat *dst, int cpu)
1044{
1045	const struct kernel_cpustat *src = &kcpustat_cpu(cpu);
1046	struct rq *rq;
1047	int err;
1048
1049	if (!vtime_accounting_enabled_cpu(cpu)) {
1050		*dst = *src;
1051		return;
1052	}
1053
1054	rq = cpu_rq(cpu);
1055
1056	for (;;) {
1057		struct task_struct *curr;
1058
1059		rcu_read_lock();
1060		curr = rcu_dereference(rq->curr);
1061		if (WARN_ON_ONCE(!curr)) {
1062			rcu_read_unlock();
1063			*dst = *src;
1064			return;
1065		}
1066
1067		err = kcpustat_cpu_fetch_vtime(dst, src, curr, cpu);
1068		rcu_read_unlock();
1069
1070		if (!err)
1071			return;
1072
1073		cpu_relax();
1074	}
1075}
1076EXPORT_SYMBOL_GPL(kcpustat_cpu_fetch);
1077
1078#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
v4.10.11
  1#include <linux/export.h>
  2#include <linux/sched.h>
  3#include <linux/tsacct_kern.h>
  4#include <linux/kernel_stat.h>
  5#include <linux/static_key.h>
  6#include <linux/context_tracking.h>
  7#include "sched.h"
  8#ifdef CONFIG_PARAVIRT
  9#include <asm/paravirt.h>
 10#endif
 11
 12
 13#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 14
 15/*
 16 * There are no locks covering percpu hardirq/softirq time.
 17 * They are only modified in vtime_account, on corresponding CPU
 18 * with interrupts disabled. So, writes are safe.
 19 * They are read and saved off onto struct rq in update_rq_clock().
 20 * This may result in other CPU reading this CPU's irq time and can
 21 * race with irq/vtime_account on this CPU. We would either get old
 22 * or new value with a side effect of accounting a slice of irq time to wrong
 23 * task when irq is in progress while we read rq->clock. That is a worthy
 24 * compromise in place of having locks on each irq in account_system_time.
 25 */
 26DEFINE_PER_CPU(struct irqtime, cpu_irqtime);
 27
 28static int sched_clock_irqtime;
 29
 30void enable_sched_clock_irqtime(void)
 31{
 32	sched_clock_irqtime = 1;
 33}
 34
 35void disable_sched_clock_irqtime(void)
 36{
 37	sched_clock_irqtime = 0;
 38}
 39
 
 
 
 
 
 
 
 
 
 
 
 
 40/*
 41 * Called before incrementing preempt_count on {soft,}irq_enter
 42 * and before decrementing preempt_count on {soft,}irq_exit.
 43 */
 44void irqtime_account_irq(struct task_struct *curr)
 45{
 46	struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
 
 47	s64 delta;
 48	int cpu;
 49
 50	if (!sched_clock_irqtime)
 51		return;
 52
 53	cpu = smp_processor_id();
 54	delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
 55	irqtime->irq_start_time += delta;
 
 56
 57	u64_stats_update_begin(&irqtime->sync);
 58	/*
 59	 * We do not account for softirq time from ksoftirqd here.
 60	 * We want to continue accounting softirq time to ksoftirqd thread
 61	 * in that case, so as not to confuse scheduler with a special task
 62	 * that do not consume any time, but still wants to run.
 63	 */
 64	if (hardirq_count())
 65		irqtime->hardirq_time += delta;
 66	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
 67		irqtime->softirq_time += delta;
 68
 69	u64_stats_update_end(&irqtime->sync);
 70}
 71EXPORT_SYMBOL_GPL(irqtime_account_irq);
 72
 73static cputime_t irqtime_account_update(u64 irqtime, int idx, cputime_t maxtime)
 74{
 75	u64 *cpustat = kcpustat_this_cpu->cpustat;
 76	cputime_t irq_cputime;
 77
 78	irq_cputime = nsecs_to_cputime64(irqtime) - cpustat[idx];
 79	irq_cputime = min(irq_cputime, maxtime);
 80	cpustat[idx] += irq_cputime;
 81
 82	return irq_cputime;
 83}
 84
 85static cputime_t irqtime_account_hi_update(cputime_t maxtime)
 86{
 87	return irqtime_account_update(__this_cpu_read(cpu_irqtime.hardirq_time),
 88				      CPUTIME_IRQ, maxtime);
 89}
 90
 91static cputime_t irqtime_account_si_update(cputime_t maxtime)
 92{
 93	return irqtime_account_update(__this_cpu_read(cpu_irqtime.softirq_time),
 94				      CPUTIME_SOFTIRQ, maxtime);
 95}
 96
 97#else /* CONFIG_IRQ_TIME_ACCOUNTING */
 98
 99#define sched_clock_irqtime	(0)
100
101static cputime_t irqtime_account_hi_update(cputime_t dummy)
102{
103	return 0;
104}
105
106static cputime_t irqtime_account_si_update(cputime_t dummy)
107{
108	return 0;
109}
110
111#endif /* !CONFIG_IRQ_TIME_ACCOUNTING */
112
113static inline void task_group_account_field(struct task_struct *p, int index,
114					    u64 tmp)
115{
116	/*
117	 * Since all updates are sure to touch the root cgroup, we
118	 * get ourselves ahead and touch it first. If the root cgroup
119	 * is the only cgroup, then nothing else should be necessary.
120	 *
121	 */
122	__this_cpu_add(kernel_cpustat.cpustat[index], tmp);
123
124	cpuacct_account_field(p, index, tmp);
125}
126
127/*
128 * Account user cpu time to a process.
129 * @p: the process that the cpu time gets accounted to
130 * @cputime: the cpu time spent in user space since the last update
131 */
132void account_user_time(struct task_struct *p, cputime_t cputime)
133{
134	int index;
135
136	/* Add user time to process. */
137	p->utime += cputime;
138	account_group_user_time(p, cputime);
139
140	index = (task_nice(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
141
142	/* Add user time to cpustat. */
143	task_group_account_field(p, index, (__force u64) cputime);
144
145	/* Account for user time used */
146	acct_account_cputime(p);
147}
148
149/*
150 * Account guest cpu time to a process.
151 * @p: the process that the cpu time gets accounted to
152 * @cputime: the cpu time spent in virtual machine since the last update
153 */
154static void account_guest_time(struct task_struct *p, cputime_t cputime)
155{
156	u64 *cpustat = kcpustat_this_cpu->cpustat;
157
158	/* Add guest time to process. */
159	p->utime += cputime;
160	account_group_user_time(p, cputime);
161	p->gtime += cputime;
162
163	/* Add guest time to cpustat. */
164	if (task_nice(p) > 0) {
165		cpustat[CPUTIME_NICE] += (__force u64) cputime;
166		cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
167	} else {
168		cpustat[CPUTIME_USER] += (__force u64) cputime;
169		cpustat[CPUTIME_GUEST] += (__force u64) cputime;
170	}
171}
172
173/*
174 * Account system cpu time to a process and desired cpustat field
175 * @p: the process that the cpu time gets accounted to
176 * @cputime: the cpu time spent in kernel space since the last update
177 * @index: pointer to cpustat field that has to be updated
178 */
179static inline
180void __account_system_time(struct task_struct *p, cputime_t cputime, int index)
181{
182	/* Add system time to process. */
183	p->stime += cputime;
184	account_group_system_time(p, cputime);
185
186	/* Add system time to cpustat. */
187	task_group_account_field(p, index, (__force u64) cputime);
188
189	/* Account for system time used */
190	acct_account_cputime(p);
191}
192
193/*
194 * Account system cpu time to a process.
195 * @p: the process that the cpu time gets accounted to
196 * @hardirq_offset: the offset to subtract from hardirq_count()
197 * @cputime: the cpu time spent in kernel space since the last update
198 */
199void account_system_time(struct task_struct *p, int hardirq_offset,
200			 cputime_t cputime)
201{
202	int index;
203
204	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
205		account_guest_time(p, cputime);
206		return;
207	}
208
209	if (hardirq_count() - hardirq_offset)
210		index = CPUTIME_IRQ;
211	else if (in_serving_softirq())
212		index = CPUTIME_SOFTIRQ;
213	else
214		index = CPUTIME_SYSTEM;
215
216	__account_system_time(p, cputime, index);
217}
218
219/*
220 * Account for involuntary wait time.
221 * @cputime: the cpu time spent in involuntary wait
222 */
223void account_steal_time(cputime_t cputime)
224{
225	u64 *cpustat = kcpustat_this_cpu->cpustat;
226
227	cpustat[CPUTIME_STEAL] += (__force u64) cputime;
228}
229
230/*
231 * Account for idle time.
232 * @cputime: the cpu time spent in idle wait
233 */
234void account_idle_time(cputime_t cputime)
235{
236	u64 *cpustat = kcpustat_this_cpu->cpustat;
237	struct rq *rq = this_rq();
238
239	if (atomic_read(&rq->nr_iowait) > 0)
240		cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
241	else
242		cpustat[CPUTIME_IDLE] += (__force u64) cputime;
243}
244
245/*
246 * When a guest is interrupted for a longer amount of time, missed clock
247 * ticks are not redelivered later. Due to that, this function may on
248 * occasion account more time than the calling functions think elapsed.
249 */
250static __always_inline cputime_t steal_account_process_time(cputime_t maxtime)
251{
252#ifdef CONFIG_PARAVIRT
253	if (static_key_false(&paravirt_steal_enabled)) {
254		cputime_t steal_cputime;
255		u64 steal;
256
257		steal = paravirt_steal_clock(smp_processor_id());
258		steal -= this_rq()->prev_steal_time;
 
 
 
259
260		steal_cputime = min(nsecs_to_cputime(steal), maxtime);
261		account_steal_time(steal_cputime);
262		this_rq()->prev_steal_time += cputime_to_nsecs(steal_cputime);
263
264		return steal_cputime;
265	}
266#endif
267	return 0;
268}
269
270/*
271 * Account how much elapsed time was spent in steal, irq, or softirq time.
272 */
273static inline cputime_t account_other_time(cputime_t max)
274{
275	cputime_t accounted;
276
277	/* Shall be converted to a lockdep-enabled lightweight check */
278	WARN_ON_ONCE(!irqs_disabled());
279
280	accounted = steal_account_process_time(max);
281
282	if (accounted < max)
283		accounted += irqtime_account_hi_update(max - accounted);
284
285	if (accounted < max)
286		accounted += irqtime_account_si_update(max - accounted);
287
288	return accounted;
289}
290
291#ifdef CONFIG_64BIT
292static inline u64 read_sum_exec_runtime(struct task_struct *t)
293{
294	return t->se.sum_exec_runtime;
295}
296#else
297static u64 read_sum_exec_runtime(struct task_struct *t)
298{
299	u64 ns;
300	struct rq_flags rf;
301	struct rq *rq;
302
303	rq = task_rq_lock(t, &rf);
304	ns = t->se.sum_exec_runtime;
305	task_rq_unlock(rq, t, &rf);
306
307	return ns;
308}
309#endif
310
311/*
312 * Accumulate raw cputime values of dead tasks (sig->[us]time) and live
313 * tasks (sum on group iteration) belonging to @tsk's group.
314 */
315void thread_group_cputime(struct task_struct *tsk, struct task_cputime *times)
316{
317	struct signal_struct *sig = tsk->signal;
318	cputime_t utime, stime;
319	struct task_struct *t;
320	unsigned int seq, nextseq;
321	unsigned long flags;
322
323	/*
324	 * Update current task runtime to account pending time since last
325	 * scheduler action or thread_group_cputime() call. This thread group
326	 * might have other running tasks on different CPUs, but updating
327	 * their runtime can affect syscall performance, so we skip account
328	 * those pending times and rely only on values updated on tick or
329	 * other scheduler action.
330	 */
331	if (same_thread_group(current, tsk))
332		(void) task_sched_runtime(current);
333
334	rcu_read_lock();
335	/* Attempt a lockless read on the first round. */
336	nextseq = 0;
337	do {
338		seq = nextseq;
339		flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
340		times->utime = sig->utime;
341		times->stime = sig->stime;
342		times->sum_exec_runtime = sig->sum_sched_runtime;
343
344		for_each_thread(tsk, t) {
345			task_cputime(t, &utime, &stime);
346			times->utime += utime;
347			times->stime += stime;
348			times->sum_exec_runtime += read_sum_exec_runtime(t);
349		}
350		/* If lockless access failed, take the lock. */
351		nextseq = 1;
352	} while (need_seqretry(&sig->stats_lock, seq));
353	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
354	rcu_read_unlock();
355}
356
357#ifdef CONFIG_IRQ_TIME_ACCOUNTING
358/*
359 * Account a tick to a process and cpustat
360 * @p: the process that the cpu time gets accounted to
361 * @user_tick: is the tick from userspace
362 * @rq: the pointer to rq
363 *
364 * Tick demultiplexing follows the order
365 * - pending hardirq update
366 * - pending softirq update
367 * - user_time
368 * - idle_time
369 * - system time
370 *   - check for guest_time
371 *   - else account as system_time
372 *
373 * Check for hardirq is done both for system and user time as there is
374 * no timer going off while we are on hardirq and hence we may never get an
375 * opportunity to update it solely in system time.
376 * p->stime and friends are only updated on system time and not on irq
377 * softirq as those do not count in task exec_runtime any more.
378 */
379static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
380					 struct rq *rq, int ticks)
381{
382	u64 cputime = (__force u64) cputime_one_jiffy * ticks;
383	cputime_t other;
384
385	/*
386	 * When returning from idle, many ticks can get accounted at
387	 * once, including some ticks of steal, irq, and softirq time.
388	 * Subtract those ticks from the amount of time accounted to
389	 * idle, or potentially user or system time. Due to rounding,
390	 * other time can exceed ticks occasionally.
391	 */
392	other = account_other_time(ULONG_MAX);
393	if (other >= cputime)
394		return;
 
395	cputime -= other;
396
397	if (this_cpu_ksoftirqd() == p) {
398		/*
399		 * ksoftirqd time do not get accounted in cpu_softirq_time.
400		 * So, we have to handle it separately here.
401		 * Also, p->stime needs to be updated for ksoftirqd.
402		 */
403		__account_system_time(p, cputime, CPUTIME_SOFTIRQ);
404	} else if (user_tick) {
405		account_user_time(p, cputime);
406	} else if (p == rq->idle) {
407		account_idle_time(cputime);
408	} else if (p->flags & PF_VCPU) { /* System time or guest time */
409		account_guest_time(p, cputime);
410	} else {
411		__account_system_time(p, cputime, CPUTIME_SYSTEM);
412	}
413}
414
415static void irqtime_account_idle_ticks(int ticks)
416{
417	struct rq *rq = this_rq();
418
419	irqtime_account_process_tick(current, 0, rq, ticks);
420}
421#else /* CONFIG_IRQ_TIME_ACCOUNTING */
422static inline void irqtime_account_idle_ticks(int ticks) {}
423static inline void irqtime_account_process_tick(struct task_struct *p, int user_tick,
424						struct rq *rq, int nr_ticks) {}
425#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
426
427/*
428 * Use precise platform statistics if available:
429 */
430#ifdef CONFIG_VIRT_CPU_ACCOUNTING
431
432#ifndef __ARCH_HAS_VTIME_TASK_SWITCH
433void vtime_common_task_switch(struct task_struct *prev)
434{
435	if (is_idle_task(prev))
436		vtime_account_idle(prev);
437	else
438		vtime_account_system(prev);
439
440#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
441	vtime_account_user(prev);
442#endif
443	arch_vtime_task_switch(prev);
444}
445#endif
446
447#endif /* CONFIG_VIRT_CPU_ACCOUNTING */
 
 
448
 
 
 
 
 
 
 
 
 
 
 
449
450#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
451/*
452 * Archs that account the whole time spent in the idle task
453 * (outside irq) as idle time can rely on this and just implement
454 * vtime_account_system() and vtime_account_idle(). Archs that
455 * have other meaning of the idle time (s390 only includes the
456 * time spent by the CPU when it's in low power mode) must override
457 * vtime_account().
458 */
459#ifndef __ARCH_HAS_VTIME_ACCOUNT
460void vtime_account_irq_enter(struct task_struct *tsk)
461{
462	if (!in_interrupt() && is_idle_task(tsk))
463		vtime_account_idle(tsk);
464	else
465		vtime_account_system(tsk);
466}
467EXPORT_SYMBOL_GPL(vtime_account_irq_enter);
468#endif /* __ARCH_HAS_VTIME_ACCOUNT */
469
470void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
471{
472	*ut = p->utime;
473	*st = p->stime;
474}
475EXPORT_SYMBOL_GPL(task_cputime_adjusted);
476
477void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
478{
479	struct task_cputime cputime;
480
481	thread_group_cputime(p, &cputime);
482
483	*ut = cputime.utime;
484	*st = cputime.stime;
485}
486#else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 
 
487/*
488 * Account a single tick of cpu time.
489 * @p: the process that the cpu time gets accounted to
490 * @user_tick: indicates if the tick is a user or a system tick
491 */
492void account_process_tick(struct task_struct *p, int user_tick)
493{
494	cputime_t cputime, steal;
495	struct rq *rq = this_rq();
496
497	if (vtime_accounting_cpu_enabled())
498		return;
499
500	if (sched_clock_irqtime) {
501		irqtime_account_process_tick(p, user_tick, rq, 1);
502		return;
503	}
504
505	cputime = cputime_one_jiffy;
506	steal = steal_account_process_time(ULONG_MAX);
507
508	if (steal >= cputime)
509		return;
510
511	cputime -= steal;
512
513	if (user_tick)
514		account_user_time(p, cputime);
515	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
516		account_system_time(p, HARDIRQ_OFFSET, cputime);
517	else
518		account_idle_time(cputime);
519}
520
521/*
522 * Account multiple ticks of idle time.
523 * @ticks: number of stolen ticks
524 */
525void account_idle_ticks(unsigned long ticks)
526{
527	cputime_t cputime, steal;
528
529	if (sched_clock_irqtime) {
530		irqtime_account_idle_ticks(ticks);
531		return;
532	}
533
534	cputime = jiffies_to_cputime(ticks);
535	steal = steal_account_process_time(ULONG_MAX);
536
537	if (steal >= cputime)
538		return;
539
540	cputime -= steal;
541	account_idle_time(cputime);
542}
543
544/*
545 * Perform (stime * rtime) / total, but avoid multiplication overflow by
546 * loosing precision when the numbers are big.
547 */
548static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
549{
550	u64 scaled;
551
552	for (;;) {
553		/* Make sure "rtime" is the bigger of stime/rtime */
554		if (stime > rtime)
555			swap(rtime, stime);
556
557		/* Make sure 'total' fits in 32 bits */
558		if (total >> 32)
559			goto drop_precision;
560
561		/* Does rtime (and thus stime) fit in 32 bits? */
562		if (!(rtime >> 32))
563			break;
564
565		/* Can we just balance rtime/stime rather than dropping bits? */
566		if (stime >> 31)
567			goto drop_precision;
568
569		/* We can grow stime and shrink rtime and try to make them both fit */
570		stime <<= 1;
571		rtime >>= 1;
572		continue;
573
574drop_precision:
575		/* We drop from rtime, it has more bits than stime */
576		rtime >>= 1;
577		total >>= 1;
578	}
579
580	/*
581	 * Make sure gcc understands that this is a 32x32->64 multiply,
582	 * followed by a 64/32->64 divide.
583	 */
584	scaled = div_u64((u64) (u32) stime * (u64) (u32) rtime, (u32)total);
585	return (__force cputime_t) scaled;
586}
587
588/*
589 * Adjust tick based cputime random precision against scheduler runtime
590 * accounting.
591 *
592 * Tick based cputime accounting depend on random scheduling timeslices of a
593 * task to be interrupted or not by the timer.  Depending on these
594 * circumstances, the number of these interrupts may be over or
595 * under-optimistic, matching the real user and system cputime with a variable
596 * precision.
597 *
598 * Fix this by scaling these tick based values against the total runtime
599 * accounted by the CFS scheduler.
600 *
601 * This code provides the following guarantees:
602 *
603 *   stime + utime == rtime
604 *   stime_i+1 >= stime_i, utime_i+1 >= utime_i
605 *
606 * Assuming that rtime_i+1 >= rtime_i.
607 */
608static void cputime_adjust(struct task_cputime *curr,
609			   struct prev_cputime *prev,
610			   cputime_t *ut, cputime_t *st)
611{
612	cputime_t rtime, stime, utime;
613	unsigned long flags;
614
615	/* Serialize concurrent callers such that we can honour our guarantees */
616	raw_spin_lock_irqsave(&prev->lock, flags);
617	rtime = nsecs_to_cputime(curr->sum_exec_runtime);
618
619	/*
620	 * This is possible under two circumstances:
621	 *  - rtime isn't monotonic after all (a bug);
622	 *  - we got reordered by the lock.
623	 *
624	 * In both cases this acts as a filter such that the rest of the code
625	 * can assume it is monotonic regardless of anything else.
626	 */
627	if (prev->stime + prev->utime >= rtime)
628		goto out;
629
630	stime = curr->stime;
631	utime = curr->utime;
632
633	/*
634	 * If either stime or both stime and utime are 0, assume all runtime is
635	 * userspace. Once a task gets some ticks, the monotonicy code at
636	 * 'update' will ensure things converge to the observed ratio.
637	 */
638	if (stime == 0) {
639		utime = rtime;
640		goto update;
641	}
642
643	if (utime == 0) {
644		stime = rtime;
645		goto update;
646	}
647
648	stime = scale_stime((__force u64)stime, (__force u64)rtime,
649			    (__force u64)(stime + utime));
650
651update:
652	/*
653	 * Make sure stime doesn't go backwards; this preserves monotonicity
654	 * for utime because rtime is monotonic.
655	 *
656	 *  utime_i+1 = rtime_i+1 - stime_i
657	 *            = rtime_i+1 - (rtime_i - utime_i)
658	 *            = (rtime_i+1 - rtime_i) + utime_i
659	 *            >= utime_i
660	 */
661	if (stime < prev->stime)
662		stime = prev->stime;
663	utime = rtime - stime;
664
665	/*
666	 * Make sure utime doesn't go backwards; this still preserves
667	 * monotonicity for stime, analogous argument to above.
668	 */
669	if (utime < prev->utime) {
670		utime = prev->utime;
671		stime = rtime - utime;
672	}
673
674	prev->stime = stime;
675	prev->utime = utime;
676out:
677	*ut = prev->utime;
678	*st = prev->stime;
679	raw_spin_unlock_irqrestore(&prev->lock, flags);
680}
681
682void task_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
683{
684	struct task_cputime cputime = {
685		.sum_exec_runtime = p->se.sum_exec_runtime,
686	};
687
688	task_cputime(p, &cputime.utime, &cputime.stime);
689	cputime_adjust(&cputime, &p->prev_cputime, ut, st);
690}
691EXPORT_SYMBOL_GPL(task_cputime_adjusted);
692
693void thread_group_cputime_adjusted(struct task_struct *p, cputime_t *ut, cputime_t *st)
694{
695	struct task_cputime cputime;
696
697	thread_group_cputime(p, &cputime);
698	cputime_adjust(&cputime, &p->signal->prev_cputime, ut, st);
699}
700#endif /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
701
702#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
703static cputime_t vtime_delta(struct task_struct *tsk)
704{
705	unsigned long now = READ_ONCE(jiffies);
706
707	if (time_before(now, (unsigned long)tsk->vtime_snap))
 
708		return 0;
709
710	return jiffies_to_cputime(now - tsk->vtime_snap);
711}
712
713static cputime_t get_vtime_delta(struct task_struct *tsk)
714{
715	unsigned long now = READ_ONCE(jiffies);
716	cputime_t delta, other;
717
718	/*
719	 * Unlike tick based timing, vtime based timing never has lost
720	 * ticks, and no need for steal time accounting to make up for
721	 * lost ticks. Vtime accounts a rounded version of actual
722	 * elapsed time. Limit account_other_time to prevent rounding
723	 * errors from causing elapsed vtime to go negative.
724	 */
725	delta = jiffies_to_cputime(now - tsk->vtime_snap);
726	other = account_other_time(delta);
727	WARN_ON_ONCE(tsk->vtime_snap_whence == VTIME_INACTIVE);
728	tsk->vtime_snap = now;
729
730	return delta - other;
731}
732
733static void __vtime_account_system(struct task_struct *tsk)
 
734{
735	cputime_t delta_cpu = get_vtime_delta(tsk);
 
 
 
 
 
736
737	account_system_time(tsk, irq_count(), delta_cpu);
 
 
 
 
 
 
 
738}
739
740void vtime_account_system(struct task_struct *tsk)
 
741{
742	if (!vtime_delta(tsk))
 
 
 
 
 
 
 
 
 
 
 
743		return;
744
745	write_seqcount_begin(&tsk->vtime_seqcount);
746	__vtime_account_system(tsk);
747	write_seqcount_end(&tsk->vtime_seqcount);
748}
749
750void vtime_account_user(struct task_struct *tsk)
751{
752	cputime_t delta_cpu;
753
754	write_seqcount_begin(&tsk->vtime_seqcount);
755	tsk->vtime_snap_whence = VTIME_SYS;
756	if (vtime_delta(tsk)) {
757		delta_cpu = get_vtime_delta(tsk);
758		account_user_time(tsk, delta_cpu);
759	}
760	write_seqcount_end(&tsk->vtime_seqcount);
761}
762
763void vtime_user_enter(struct task_struct *tsk)
764{
765	write_seqcount_begin(&tsk->vtime_seqcount);
766	if (vtime_delta(tsk))
767		__vtime_account_system(tsk);
768	tsk->vtime_snap_whence = VTIME_USER;
769	write_seqcount_end(&tsk->vtime_seqcount);
 
 
 
 
 
770}
771
772void vtime_guest_enter(struct task_struct *tsk)
773{
 
774	/*
775	 * The flags must be updated under the lock with
776	 * the vtime_snap flush and update.
777	 * That enforces a right ordering and update sequence
778	 * synchronization against the reader (task_gtime())
779	 * that can thus safely catch up with a tickless delta.
780	 */
781	write_seqcount_begin(&tsk->vtime_seqcount);
782	if (vtime_delta(tsk))
783		__vtime_account_system(tsk);
784	current->flags |= PF_VCPU;
785	write_seqcount_end(&tsk->vtime_seqcount);
786}
787EXPORT_SYMBOL_GPL(vtime_guest_enter);
788
789void vtime_guest_exit(struct task_struct *tsk)
790{
791	write_seqcount_begin(&tsk->vtime_seqcount);
792	__vtime_account_system(tsk);
793	current->flags &= ~PF_VCPU;
794	write_seqcount_end(&tsk->vtime_seqcount);
 
 
 
795}
796EXPORT_SYMBOL_GPL(vtime_guest_exit);
797
798void vtime_account_idle(struct task_struct *tsk)
799{
800	cputime_t delta_cpu = get_vtime_delta(tsk);
801
802	account_idle_time(delta_cpu);
803}
804
805void arch_vtime_task_switch(struct task_struct *prev)
806{
807	write_seqcount_begin(&prev->vtime_seqcount);
808	prev->vtime_snap_whence = VTIME_INACTIVE;
809	write_seqcount_end(&prev->vtime_seqcount);
810
811	write_seqcount_begin(&current->vtime_seqcount);
812	current->vtime_snap_whence = VTIME_SYS;
813	current->vtime_snap = jiffies;
814	write_seqcount_end(&current->vtime_seqcount);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
815}
816
817void vtime_init_idle(struct task_struct *t, int cpu)
818{
 
819	unsigned long flags;
820
821	local_irq_save(flags);
822	write_seqcount_begin(&t->vtime_seqcount);
823	t->vtime_snap_whence = VTIME_SYS;
824	t->vtime_snap = jiffies;
825	write_seqcount_end(&t->vtime_seqcount);
 
826	local_irq_restore(flags);
827}
828
829cputime_t task_gtime(struct task_struct *t)
830{
 
831	unsigned int seq;
832	cputime_t gtime;
833
834	if (!vtime_accounting_enabled())
835		return t->gtime;
836
837	do {
838		seq = read_seqcount_begin(&t->vtime_seqcount);
839
840		gtime = t->gtime;
841		if (t->vtime_snap_whence == VTIME_SYS && t->flags & PF_VCPU)
842			gtime += vtime_delta(t);
843
844	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
845
846	return gtime;
847}
848
849/*
850 * Fetch cputime raw values from fields of task_struct and
851 * add up the pending nohz execution time since the last
852 * cputime snapshot.
853 */
854void task_cputime(struct task_struct *t, cputime_t *utime, cputime_t *stime)
855{
856	cputime_t delta;
857	unsigned int seq;
 
858
859	if (!vtime_accounting_enabled()) {
860		*utime = t->utime;
861		*stime = t->stime;
862		return;
863	}
864
865	do {
866		seq = read_seqcount_begin(&t->vtime_seqcount);
867
868		*utime = t->utime;
869		*stime = t->stime;
870
871		/* Task is sleeping, nothing to add */
872		if (t->vtime_snap_whence == VTIME_INACTIVE || is_idle_task(t))
873			continue;
874
875		delta = vtime_delta(t);
876
877		/*
878		 * Task runs either in user or kernel space, add pending nohz time to
879		 * the right place.
880		 */
881		if (t->vtime_snap_whence == VTIME_USER || t->flags & PF_VCPU)
882			*utime += delta;
883		else if (t->vtime_snap_whence == VTIME_SYS)
884			*stime += delta;
885	} while (read_seqcount_retry(&t->vtime_seqcount, seq));
886}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
887#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */