Linux Audio

Check our new training course

Loading...
v3.15
 
   1/*
   2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
   3 * Internal non-public definitions that provide either classic
   4 * or preemptible semantics.
   5 *
   6 * This program is free software; you can redistribute it and/or modify
   7 * it under the terms of the GNU General Public License as published by
   8 * the Free Software Foundation; either version 2 of the License, or
   9 * (at your option) any later version.
  10 *
  11 * This program is distributed in the hope that it will be useful,
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 * GNU General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * along with this program; if not, you can access it online at
  18 * http://www.gnu.org/licenses/gpl-2.0.html.
  19 *
  20 * Copyright Red Hat, 2009
  21 * Copyright IBM Corporation, 2009
  22 *
  23 * Author: Ingo Molnar <mingo@elte.hu>
  24 *	   Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  25 */
  26
  27#include <linux/delay.h>
  28#include <linux/gfp.h>
  29#include <linux/oom.h>
  30#include <linux/smpboot.h>
  31#include "../time/tick-internal.h"
  32
  33#define RCU_KTHREAD_PRIO 1
  34
  35#ifdef CONFIG_RCU_BOOST
  36#define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
  37#else
  38#define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
  39#endif
  40
  41#ifdef CONFIG_RCU_NOCB_CPU
  42static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
  43static bool have_rcu_nocb_mask;	    /* Was rcu_nocb_mask allocated? */
  44static bool __read_mostly rcu_nocb_poll;    /* Offload kthread are to poll. */
  45static char __initdata nocb_buf[NR_CPUS * 5];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  46#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
  47
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  48/*
  49 * Check the RCU kernel configuration parameters and print informative
  50 * messages about anything out of the ordinary.  If you like #ifdef, you
  51 * will love this function.
  52 */
  53static void __init rcu_bootup_announce_oddness(void)
  54{
  55#ifdef CONFIG_RCU_TRACE
  56	pr_info("\tRCU debugfs-based tracing is enabled.\n");
  57#endif
  58#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
  59	pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
  60	       CONFIG_RCU_FANOUT);
  61#endif
  62#ifdef CONFIG_RCU_FANOUT_EXACT
  63	pr_info("\tHierarchical RCU autobalancing is disabled.\n");
  64#endif
  65#ifdef CONFIG_RCU_FAST_NO_HZ
  66	pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
  67#endif
  68#ifdef CONFIG_PROVE_RCU
  69	pr_info("\tRCU lockdep checking is enabled.\n");
  70#endif
  71#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
  72	pr_info("\tRCU torture testing starts during boot.\n");
  73#endif
  74#if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
  75	pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
  76#endif
  77#if defined(CONFIG_RCU_CPU_STALL_INFO)
  78	pr_info("\tAdditional per-CPU info printed with stalls.\n");
  79#endif
  80#if NUM_RCU_LVL_4 != 0
  81	pr_info("\tFour-level hierarchy is enabled.\n");
  82#endif
  83	if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
  84		pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
  85	if (nr_cpu_ids != NR_CPUS)
  86		pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
  87#ifdef CONFIG_RCU_NOCB_CPU
  88#ifndef CONFIG_RCU_NOCB_CPU_NONE
  89	if (!have_rcu_nocb_mask) {
  90		zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL);
  91		have_rcu_nocb_mask = true;
  92	}
  93#ifdef CONFIG_RCU_NOCB_CPU_ZERO
  94	pr_info("\tOffload RCU callbacks from CPU 0\n");
  95	cpumask_set_cpu(0, rcu_nocb_mask);
  96#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
  97#ifdef CONFIG_RCU_NOCB_CPU_ALL
  98	pr_info("\tOffload RCU callbacks from all CPUs\n");
  99	cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
 100#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
 101#endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
 102	if (have_rcu_nocb_mask) {
 103		if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
 104			pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
 105			cpumask_and(rcu_nocb_mask, cpu_possible_mask,
 106				    rcu_nocb_mask);
 107		}
 108		cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
 109		pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
 110		if (rcu_nocb_poll)
 111			pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
 112	}
 113#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 
 
 
 
 
 
 114}
 115
 116#ifdef CONFIG_TREE_PREEMPT_RCU
 117
 118RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
 119static struct rcu_state *rcu_state = &rcu_preempt_state;
 120
 121static int rcu_preempted_readers_exp(struct rcu_node *rnp);
 
 122
 123/*
 124 * Tell them what RCU they are running.
 125 */
 126static void __init rcu_bootup_announce(void)
 127{
 128	pr_info("Preemptible hierarchical RCU implementation.\n");
 129	rcu_bootup_announce_oddness();
 130}
 131
 132/*
 133 * Return the number of RCU-preempt batches processed thus far
 134 * for debug and statistics.
 135 */
 136long rcu_batches_completed_preempt(void)
 137{
 138	return rcu_preempt_state.completed;
 139}
 140EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 141
 142/*
 143 * Return the number of RCU batches processed thus far for debug & stats.
 144 */
 145long rcu_batches_completed(void)
 146{
 147	return rcu_batches_completed_preempt();
 148}
 149EXPORT_SYMBOL_GPL(rcu_batches_completed);
 150
 151/*
 152 * Force a quiescent state for preemptible RCU.
 153 */
 154void rcu_force_quiescent_state(void)
 155{
 156	force_quiescent_state(&rcu_preempt_state);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 157}
 158EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
 159
 160/*
 161 * Record a preemptible-RCU quiescent state for the specified CPU.  Note
 162 * that this just means that the task currently running on the CPU is
 163 * not in a quiescent state.  There might be any number of tasks blocked
 164 * while in an RCU read-side critical section.
 165 *
 166 * Unlike the other rcu_*_qs() functions, callers to this function
 167 * must disable irqs in order to protect the assignment to
 168 * ->rcu_read_unlock_special.
 169 */
 170static void rcu_preempt_qs(int cpu)
 171{
 172	struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
 173
 174	if (rdp->passed_quiesce == 0)
 175		trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs"));
 176	rdp->passed_quiesce = 1;
 177	current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
 
 
 
 
 
 
 178}
 179
 180/*
 181 * We have entered the scheduler, and the current task might soon be
 182 * context-switched away from.  If this task is in an RCU read-side
 183 * critical section, we will no longer be able to rely on the CPU to
 184 * record that fact, so we enqueue the task on the blkd_tasks list.
 185 * The task will dequeue itself when it exits the outermost enclosing
 186 * RCU read-side critical section.  Therefore, the current grace period
 187 * cannot be permitted to complete until the blkd_tasks list entries
 188 * predating the current grace period drain, in other words, until
 189 * rnp->gp_tasks becomes NULL.
 190 *
 191 * Caller must disable preemption.
 192 */
 193static void rcu_preempt_note_context_switch(int cpu)
 194{
 195	struct task_struct *t = current;
 196	unsigned long flags;
 197	struct rcu_data *rdp;
 198	struct rcu_node *rnp;
 199
 200	if (t->rcu_read_lock_nesting > 0 &&
 201	    (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
 
 
 
 202
 203		/* Possibly blocking in an RCU read-side critical section. */
 204		rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
 205		rnp = rdp->mynode;
 206		raw_spin_lock_irqsave(&rnp->lock, flags);
 207		smp_mb__after_unlock_lock();
 208		t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
 209		t->rcu_blocked_node = rnp;
 210
 211		/*
 212		 * If this CPU has already checked in, then this task
 213		 * will hold up the next grace period rather than the
 214		 * current grace period.  Queue the task accordingly.
 215		 * If the task is queued for the current grace period
 216		 * (i.e., this CPU has not yet passed through a quiescent
 217		 * state for the current grace period), then as long
 218		 * as that task remains queued, the current grace period
 219		 * cannot end.  Note that there is some uncertainty as
 220		 * to exactly when the current grace period started.
 221		 * We take a conservative approach, which can result
 222		 * in unnecessarily waiting on tasks that started very
 223		 * slightly after the current grace period began.  C'est
 224		 * la vie!!!
 225		 *
 226		 * But first, note that the current CPU must still be
 227		 * on line!
 228		 */
 229		WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
 230		WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
 231		if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
 232			list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
 233			rnp->gp_tasks = &t->rcu_node_entry;
 234#ifdef CONFIG_RCU_BOOST
 235			if (rnp->boost_tasks != NULL)
 236				rnp->boost_tasks = rnp->gp_tasks;
 237#endif /* #ifdef CONFIG_RCU_BOOST */
 238		} else {
 239			list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
 240			if (rnp->qsmask & rdp->grpmask)
 241				rnp->gp_tasks = &t->rcu_node_entry;
 242		}
 243		trace_rcu_preempt_task(rdp->rsp->name,
 244				       t->pid,
 245				       (rnp->qsmask & rdp->grpmask)
 246				       ? rnp->gpnum
 247				       : rnp->gpnum + 1);
 248		raw_spin_unlock_irqrestore(&rnp->lock, flags);
 249	} else if (t->rcu_read_lock_nesting < 0 &&
 250		   t->rcu_read_unlock_special) {
 251
 252		/*
 253		 * Complete exit from RCU read-side critical section on
 254		 * behalf of preempted instance of __rcu_read_unlock().
 255		 */
 256		rcu_read_unlock_special(t);
 257	}
 258
 259	/*
 260	 * Either we were not in an RCU read-side critical section to
 261	 * begin with, or we have now recorded that critical section
 262	 * globally.  Either way, we can now note a quiescent state
 263	 * for this CPU.  Again, if we were in an RCU read-side critical
 264	 * section, and if that critical section was blocking the current
 265	 * grace period, then the fact that the task has been enqueued
 266	 * means that we continue to block the current grace period.
 267	 */
 268	local_irq_save(flags);
 269	rcu_preempt_qs(cpu);
 270	local_irq_restore(flags);
 
 
 271}
 
 272
 273/*
 274 * Check for preempted RCU readers blocking the current grace period
 275 * for the specified rcu_node structure.  If the caller needs a reliable
 276 * answer, it must hold the rcu_node's ->lock.
 277 */
 278static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 279{
 280	return rnp->gp_tasks != NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 281}
 282
 283/*
 284 * Record a quiescent state for all tasks that were previously queued
 285 * on the specified rcu_node structure and that were blocking the current
 286 * RCU grace period.  The caller must hold the specified rnp->lock with
 287 * irqs disabled, and this lock is released upon return, but irqs remain
 288 * disabled.
 289 */
 290static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
 291	__releases(rnp->lock)
 292{
 293	unsigned long mask;
 294	struct rcu_node *rnp_p;
 
 
 
 
 
 
 295
 296	if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
 297		raw_spin_unlock_irqrestore(&rnp->lock, flags);
 298		return;  /* Still need more quiescent states! */
 299	}
 
 
 
 
 
 
 300
 301	rnp_p = rnp->parent;
 302	if (rnp_p == NULL) {
 303		/*
 304		 * Either there is only one rcu_node in the tree,
 305		 * or tasks were kicked up to root rcu_node due to
 306		 * CPUs going offline.
 307		 */
 308		rcu_report_qs_rsp(&rcu_preempt_state, flags);
 309		return;
 310	}
 
 
 311
 312	/* Report up the rest of the hierarchy. */
 313	mask = rnp->grpmask;
 314	raw_spin_unlock(&rnp->lock);	/* irqs remain disabled. */
 315	raw_spin_lock(&rnp_p->lock);	/* irqs already disabled. */
 316	smp_mb__after_unlock_lock();
 317	rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
 318}
 
 319
 320/*
 321 * Advance a ->blkd_tasks-list pointer to the next entry, instead
 322 * returning NULL if at the end of the list.
 323 */
 324static struct list_head *rcu_next_node_entry(struct task_struct *t,
 325					     struct rcu_node *rnp)
 326{
 327	struct list_head *np;
 328
 329	np = t->rcu_node_entry.next;
 330	if (np == &rnp->blkd_tasks)
 331		np = NULL;
 332	return np;
 333}
 334
 335/*
 336 * Handle special cases during rcu_read_unlock(), such as needing to
 337 * notify RCU core processing or task having blocked during the RCU
 338 * read-side critical section.
 339 */
 340void rcu_read_unlock_special(struct task_struct *t)
 341{
 342	int empty;
 343	int empty_exp;
 344	int empty_exp_now;
 345	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 346	struct list_head *np;
 347#ifdef CONFIG_RCU_BOOST
 348	struct rt_mutex *rbmp = NULL;
 349#endif /* #ifdef CONFIG_RCU_BOOST */
 350	struct rcu_node *rnp;
 351	int special;
 352
 353	/* NMI handlers cannot block and cannot safely manipulate state. */
 354	if (in_nmi())
 355		return;
 356
 357	local_irq_save(flags);
 358
 359	/*
 360	 * If RCU core is waiting for this CPU to exit critical section,
 361	 * let it know that we have done so.
 
 362	 */
 363	special = t->rcu_read_unlock_special;
 364	if (special & RCU_READ_UNLOCK_NEED_QS) {
 365		rcu_preempt_qs(smp_processor_id());
 366		if (!t->rcu_read_unlock_special) {
 367			local_irq_restore(flags);
 368			return;
 369		}
 370	}
 371
 372	/* Hardware IRQ handlers cannot block, complain if they get here. */
 373	if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
 374		local_irq_restore(flags);
 375		return;
 376	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 377
 378	/* Clean up if blocked during RCU read-side critical section. */
 379	if (special & RCU_READ_UNLOCK_BLOCKED) {
 380		t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
 381
 382		/*
 383		 * Remove this task from the list it blocked on.  The
 384		 * task can migrate while we acquire the lock, but at
 385		 * most one time.  So at most two passes through loop.
 
 386		 */
 387		for (;;) {
 388			rnp = t->rcu_blocked_node;
 389			raw_spin_lock(&rnp->lock);  /* irqs already disabled. */
 390			smp_mb__after_unlock_lock();
 391			if (rnp == t->rcu_blocked_node)
 392				break;
 393			raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
 394		}
 395		empty = !rcu_preempt_blocked_readers_cgp(rnp);
 396		empty_exp = !rcu_preempted_readers_exp(rnp);
 397		smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
 398		np = rcu_next_node_entry(t, rnp);
 399		list_del_init(&t->rcu_node_entry);
 400		t->rcu_blocked_node = NULL;
 401		trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
 402						rnp->gpnum, t->pid);
 403		if (&t->rcu_node_entry == rnp->gp_tasks)
 404			rnp->gp_tasks = np;
 405		if (&t->rcu_node_entry == rnp->exp_tasks)
 406			rnp->exp_tasks = np;
 407#ifdef CONFIG_RCU_BOOST
 408		if (&t->rcu_node_entry == rnp->boost_tasks)
 409			rnp->boost_tasks = np;
 410		/* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
 411		if (t->rcu_boost_mutex) {
 412			rbmp = t->rcu_boost_mutex;
 413			t->rcu_boost_mutex = NULL;
 414		}
 415#endif /* #ifdef CONFIG_RCU_BOOST */
 416
 417		/*
 418		 * If this was the last task on the current list, and if
 419		 * we aren't waiting on any CPUs, report the quiescent state.
 420		 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
 421		 * so we must take a snapshot of the expedited state.
 422		 */
 423		empty_exp_now = !rcu_preempted_readers_exp(rnp);
 424		if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
 425			trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
 426							 rnp->gpnum,
 427							 0, rnp->qsmask,
 428							 rnp->level,
 429							 rnp->grplo,
 430							 rnp->grphi,
 431							 !!rnp->gp_tasks);
 432			rcu_report_unblock_qs_rnp(rnp, flags);
 433		} else {
 434			raw_spin_unlock_irqrestore(&rnp->lock, flags);
 435		}
 436
 437#ifdef CONFIG_RCU_BOOST
 438		/* Unboost if we were boosted. */
 439		if (rbmp)
 440			rt_mutex_unlock(rbmp);
 441#endif /* #ifdef CONFIG_RCU_BOOST */
 442
 443		/*
 444		 * If this was the last task on the expedited lists,
 445		 * then we need to report up the rcu_node hierarchy.
 446		 */
 447		if (!empty_exp && empty_exp_now)
 448			rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
 449	} else {
 450		local_irq_restore(flags);
 451	}
 452}
 453
 454#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
 455
 456/*
 457 * Dump detailed information for all tasks blocking the current RCU
 458 * grace period on the specified rcu_node structure.
 
 
 
 
 
 459 */
 460static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 461{
 462	unsigned long flags;
 463	struct task_struct *t;
 464
 465	raw_spin_lock_irqsave(&rnp->lock, flags);
 466	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 467		raw_spin_unlock_irqrestore(&rnp->lock, flags);
 468		return;
 469	}
 470	t = list_entry(rnp->gp_tasks,
 471		       struct task_struct, rcu_node_entry);
 472	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
 473		sched_show_task(t);
 474	raw_spin_unlock_irqrestore(&rnp->lock, flags);
 475}
 476
 477/*
 478 * Dump detailed information for all tasks blocking the current RCU
 479 * grace period.
 
 
 
 480 */
 481static void rcu_print_detail_task_stall(struct rcu_state *rsp)
 482{
 483	struct rcu_node *rnp = rcu_get_root(rsp);
 484
 485	rcu_print_detail_task_stall_rnp(rnp);
 486	rcu_for_each_leaf_node(rsp, rnp)
 487		rcu_print_detail_task_stall_rnp(rnp);
 488}
 489
 490#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
 491
 492static void rcu_print_detail_task_stall(struct rcu_state *rsp)
 493{
 494}
 495
 496#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
 497
 498#ifdef CONFIG_RCU_CPU_STALL_INFO
 499
 500static void rcu_print_task_stall_begin(struct rcu_node *rnp)
 501{
 502	pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
 503	       rnp->level, rnp->grplo, rnp->grphi);
 504}
 505
 506static void rcu_print_task_stall_end(void)
 507{
 508	pr_cont("\n");
 
 509}
 510
 511#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
 512
 513static void rcu_print_task_stall_begin(struct rcu_node *rnp)
 
 514{
 515}
 516
 517static void rcu_print_task_stall_end(void)
 518{
 519}
 520
 521#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
 522
 523/*
 524 * Scan the current list of tasks blocked within RCU read-side critical
 525 * sections, printing out the tid of each.
 
 526 */
 527static int rcu_print_task_stall(struct rcu_node *rnp)
 528{
 529	struct task_struct *t;
 530	int ndetected = 0;
 
 
 531
 532	if (!rcu_preempt_blocked_readers_cgp(rnp))
 533		return 0;
 534	rcu_print_task_stall_begin(rnp);
 535	t = list_entry(rnp->gp_tasks,
 536		       struct task_struct, rcu_node_entry);
 537	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 538		pr_cont(" P%d", t->pid);
 539		ndetected++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 540	}
 541	rcu_print_task_stall_end();
 542	return ndetected;
 543}
 544
 545/*
 546 * Check that the list of blocked tasks for the newly completed grace
 547 * period is in fact empty.  It is a serious bug to complete a grace
 548 * period that still has RCU readers blocked!  This function must be
 549 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
 550 * must be held by the caller.
 551 *
 552 * Also, if there are blocked tasks on the list, they automatically
 553 * block the newly created grace period, so set up ->gp_tasks accordingly.
 554 */
 555static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 556{
 557	WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
 558	if (!list_empty(&rnp->blkd_tasks))
 559		rnp->gp_tasks = rnp->blkd_tasks.next;
 560	WARN_ON_ONCE(rnp->qsmask);
 561}
 562
 563#ifdef CONFIG_HOTPLUG_CPU
 564
 565/*
 566 * Handle tasklist migration for case in which all CPUs covered by the
 567 * specified rcu_node have gone offline.  Move them up to the root
 568 * rcu_node.  The reason for not just moving them to the immediate
 569 * parent is to remove the need for rcu_read_unlock_special() to
 570 * make more than two attempts to acquire the target rcu_node's lock.
 571 * Returns true if there were tasks blocking the current RCU grace
 572 * period.
 573 *
 574 * Returns 1 if there was previously a task blocking the current grace
 575 * period on the specified rcu_node structure.
 576 *
 577 * The caller must hold rnp->lock with irqs disabled.
 578 */
 579static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
 580				     struct rcu_node *rnp,
 581				     struct rcu_data *rdp)
 582{
 583	struct list_head *lp;
 584	struct list_head *lp_root;
 585	int retval = 0;
 586	struct rcu_node *rnp_root = rcu_get_root(rsp);
 587	struct task_struct *t;
 588
 589	if (rnp == rnp_root) {
 590		WARN_ONCE(1, "Last CPU thought to be offlined?");
 591		return 0;  /* Shouldn't happen: at least one CPU online. */
 592	}
 593
 594	/* If we are on an internal node, complain bitterly. */
 595	WARN_ON_ONCE(rnp != rdp->mynode);
 596
 597	/*
 598	 * Move tasks up to root rcu_node.  Don't try to get fancy for
 599	 * this corner-case operation -- just put this node's tasks
 600	 * at the head of the root node's list, and update the root node's
 601	 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
 602	 * if non-NULL.  This might result in waiting for more tasks than
 603	 * absolutely necessary, but this is a good performance/complexity
 604	 * tradeoff.
 605	 */
 606	if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
 607		retval |= RCU_OFL_TASKS_NORM_GP;
 608	if (rcu_preempted_readers_exp(rnp))
 609		retval |= RCU_OFL_TASKS_EXP_GP;
 610	lp = &rnp->blkd_tasks;
 611	lp_root = &rnp_root->blkd_tasks;
 612	while (!list_empty(lp)) {
 613		t = list_entry(lp->next, typeof(*t), rcu_node_entry);
 614		raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
 615		smp_mb__after_unlock_lock();
 616		list_del(&t->rcu_node_entry);
 617		t->rcu_blocked_node = rnp_root;
 618		list_add(&t->rcu_node_entry, lp_root);
 619		if (&t->rcu_node_entry == rnp->gp_tasks)
 620			rnp_root->gp_tasks = rnp->gp_tasks;
 621		if (&t->rcu_node_entry == rnp->exp_tasks)
 622			rnp_root->exp_tasks = rnp->exp_tasks;
 623#ifdef CONFIG_RCU_BOOST
 624		if (&t->rcu_node_entry == rnp->boost_tasks)
 625			rnp_root->boost_tasks = rnp->boost_tasks;
 626#endif /* #ifdef CONFIG_RCU_BOOST */
 627		raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
 628	}
 629
 630	rnp->gp_tasks = NULL;
 631	rnp->exp_tasks = NULL;
 632#ifdef CONFIG_RCU_BOOST
 633	rnp->boost_tasks = NULL;
 634	/*
 635	 * In case root is being boosted and leaf was not.  Make sure
 636	 * that we boost the tasks blocking the current grace period
 637	 * in this case.
 638	 */
 639	raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
 640	smp_mb__after_unlock_lock();
 641	if (rnp_root->boost_tasks != NULL &&
 642	    rnp_root->boost_tasks != rnp_root->gp_tasks &&
 643	    rnp_root->boost_tasks != rnp_root->exp_tasks)
 644		rnp_root->boost_tasks = rnp_root->gp_tasks;
 645	raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
 646#endif /* #ifdef CONFIG_RCU_BOOST */
 647
 648	return retval;
 649}
 650
 651#endif /* #ifdef CONFIG_HOTPLUG_CPU */
 652
 653/*
 654 * Check for a quiescent state from the current CPU.  When a task blocks,
 655 * the task is recorded in the corresponding CPU's rcu_node structure,
 656 * which is checked elsewhere.
 657 *
 658 * Caller must disable hard irqs.
 659 */
 660static void rcu_preempt_check_callbacks(int cpu)
 661{
 662	struct task_struct *t = current;
 663
 664	if (t->rcu_read_lock_nesting == 0) {
 665		rcu_preempt_qs(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 666		return;
 667	}
 668	if (t->rcu_read_lock_nesting > 0 &&
 669	    per_cpu(rcu_preempt_data, cpu).qs_pending)
 670		t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
 671}
 672
 673#ifdef CONFIG_RCU_BOOST
 674
 675static void rcu_preempt_do_callbacks(void)
 676{
 677	rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
 678}
 679
 680#endif /* #ifdef CONFIG_RCU_BOOST */
 681
 682/*
 683 * Queue a preemptible-RCU callback for invocation after a grace period.
 684 */
 685void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 686{
 687	__call_rcu(head, func, &rcu_preempt_state, -1, 0);
 
 688}
 689EXPORT_SYMBOL_GPL(call_rcu);
 690
 691/*
 692 * Queue an RCU callback for lazy invocation after a grace period.
 693 * This will likely be later named something like "call_rcu_lazy()",
 694 * but this change will require some way of tagging the lazy RCU
 695 * callbacks in the list of pending callbacks.  Until then, this
 696 * function may only be called from __kfree_rcu().
 
 697 */
 698void kfree_call_rcu(struct rcu_head *head,
 699		    void (*func)(struct rcu_head *rcu))
 700{
 701	__call_rcu(head, func, &rcu_preempt_state, -1, 1);
 702}
 703EXPORT_SYMBOL_GPL(kfree_call_rcu);
 704
 705/**
 706 * synchronize_rcu - wait until a grace period has elapsed.
 707 *
 708 * Control will return to the caller some time after a full grace
 709 * period has elapsed, in other words after all currently executing RCU
 710 * read-side critical sections have completed.  Note, however, that
 711 * upon return from synchronize_rcu(), the caller might well be executing
 712 * concurrently with new RCU read-side critical sections that began while
 713 * synchronize_rcu() was waiting.  RCU read-side critical sections are
 714 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
 715 *
 716 * See the description of synchronize_sched() for more detailed information
 717 * on memory ordering guarantees.
 718 */
 719void synchronize_rcu(void)
 720{
 721	rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
 722			   !lock_is_held(&rcu_lock_map) &&
 723			   !lock_is_held(&rcu_sched_lock_map),
 724			   "Illegal synchronize_rcu() in RCU read-side critical section");
 725	if (!rcu_scheduler_active)
 726		return;
 727	if (rcu_expedited)
 728		synchronize_rcu_expedited();
 729	else
 730		wait_rcu_gp(call_rcu);
 731}
 732EXPORT_SYMBOL_GPL(synchronize_rcu);
 733
 734static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
 735static unsigned long sync_rcu_preempt_exp_count;
 736static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
 737
 738/*
 739 * Return non-zero if there are any tasks in RCU read-side critical
 740 * sections blocking the current preemptible-RCU expedited grace period.
 741 * If there is no preemptible-RCU expedited grace period currently in
 742 * progress, returns zero unconditionally.
 743 */
 744static int rcu_preempted_readers_exp(struct rcu_node *rnp)
 745{
 746	return rnp->exp_tasks != NULL;
 747}
 748
 749/*
 750 * return non-zero if there is no RCU expedited grace period in progress
 751 * for the specified rcu_node structure, in other words, if all CPUs and
 752 * tasks covered by the specified rcu_node structure have done their bit
 753 * for the current expedited grace period.  Works only for preemptible
 754 * RCU -- other RCU implementation use other means.
 755 *
 756 * Caller must hold sync_rcu_preempt_exp_mutex.
 757 */
 758static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
 759{
 760	return !rcu_preempted_readers_exp(rnp) &&
 761	       ACCESS_ONCE(rnp->expmask) == 0;
 762}
 763
 764/*
 765 * Report the exit from RCU read-side critical section for the last task
 766 * that queued itself during or before the current expedited preemptible-RCU
 767 * grace period.  This event is reported either to the rcu_node structure on
 768 * which the task was queued or to one of that rcu_node structure's ancestors,
 769 * recursively up the tree.  (Calm down, calm down, we do the recursion
 770 * iteratively!)
 771 *
 772 * Most callers will set the "wake" flag, but the task initiating the
 773 * expedited grace period need not wake itself.
 774 *
 775 * Caller must hold sync_rcu_preempt_exp_mutex.
 776 */
 777static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
 778			       bool wake)
 779{
 780	unsigned long flags;
 781	unsigned long mask;
 782
 783	raw_spin_lock_irqsave(&rnp->lock, flags);
 784	smp_mb__after_unlock_lock();
 785	for (;;) {
 786		if (!sync_rcu_preempt_exp_done(rnp)) {
 787			raw_spin_unlock_irqrestore(&rnp->lock, flags);
 788			break;
 789		}
 790		if (rnp->parent == NULL) {
 791			raw_spin_unlock_irqrestore(&rnp->lock, flags);
 792			if (wake) {
 793				smp_mb(); /* EGP done before wake_up(). */
 794				wake_up(&sync_rcu_preempt_exp_wq);
 795			}
 796			break;
 797		}
 798		mask = rnp->grpmask;
 799		raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
 800		rnp = rnp->parent;
 801		raw_spin_lock(&rnp->lock); /* irqs already disabled */
 802		smp_mb__after_unlock_lock();
 803		rnp->expmask &= ~mask;
 804	}
 
 
 805}
 806
 807/*
 808 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
 809 * grace period for the specified rcu_node structure.  If there are no such
 810 * tasks, report it up the rcu_node hierarchy.
 811 *
 812 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
 813 * CPU hotplug operations.
 814 */
 815static void
 816sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
 817{
 818	unsigned long flags;
 819	int must_wait = 0;
 820
 821	raw_spin_lock_irqsave(&rnp->lock, flags);
 822	smp_mb__after_unlock_lock();
 823	if (list_empty(&rnp->blkd_tasks)) {
 824		raw_spin_unlock_irqrestore(&rnp->lock, flags);
 825	} else {
 826		rnp->exp_tasks = rnp->blkd_tasks.next;
 827		rcu_initiate_boost(rnp, flags);  /* releases rnp->lock */
 828		must_wait = 1;
 829	}
 830	if (!must_wait)
 831		rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
 832}
 833
 834/**
 835 * synchronize_rcu_expedited - Brute-force RCU grace period
 836 *
 837 * Wait for an RCU-preempt grace period, but expedite it.  The basic
 838 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
 839 * the ->blkd_tasks lists and wait for this list to drain.  This consumes
 840 * significant time on all CPUs and is unfriendly to real-time workloads,
 841 * so is thus not recommended for any sort of common-case code.
 842 * In fact, if you are using synchronize_rcu_expedited() in a loop,
 843 * please restructure your code to batch your updates, and then Use a
 844 * single synchronize_rcu() instead.
 845 *
 846 * Note that it is illegal to call this function while holding any lock
 847 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
 848 * to call this function from a CPU-hotplug notifier.  Failing to observe
 849 * these restriction will result in deadlock.
 850 */
 851void synchronize_rcu_expedited(void)
 852{
 853	unsigned long flags;
 854	struct rcu_node *rnp;
 855	struct rcu_state *rsp = &rcu_preempt_state;
 856	unsigned long snap;
 857	int trycount = 0;
 858
 859	smp_mb(); /* Caller's modifications seen first by other CPUs. */
 860	snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
 861	smp_mb(); /* Above access cannot bleed into critical section. */
 862
 863	/*
 864	 * Block CPU-hotplug operations.  This means that any CPU-hotplug
 865	 * operation that finds an rcu_node structure with tasks in the
 866	 * process of being boosted will know that all tasks blocking
 867	 * this expedited grace period will already be in the process of
 868	 * being boosted.  This simplifies the process of moving tasks
 869	 * from leaf to root rcu_node structures.
 870	 */
 871	get_online_cpus();
 872
 873	/*
 874	 * Acquire lock, falling back to synchronize_rcu() if too many
 875	 * lock-acquisition failures.  Of course, if someone does the
 876	 * expedited grace period for us, just leave.
 877	 */
 878	while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
 879		if (ULONG_CMP_LT(snap,
 880		    ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
 881			put_online_cpus();
 882			goto mb_ret; /* Others did our work for us. */
 883		}
 884		if (trycount++ < 10) {
 885			udelay(trycount * num_online_cpus());
 886		} else {
 887			put_online_cpus();
 888			wait_rcu_gp(call_rcu);
 889			return;
 890		}
 891	}
 892	if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
 893		put_online_cpus();
 894		goto unlock_mb_ret; /* Others did our work for us. */
 895	}
 896
 897	/* force all RCU readers onto ->blkd_tasks lists. */
 898	synchronize_sched_expedited();
 899
 900	/* Initialize ->expmask for all non-leaf rcu_node structures. */
 901	rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
 902		raw_spin_lock_irqsave(&rnp->lock, flags);
 903		smp_mb__after_unlock_lock();
 904		rnp->expmask = rnp->qsmaskinit;
 905		raw_spin_unlock_irqrestore(&rnp->lock, flags);
 906	}
 907
 908	/* Snapshot current state of ->blkd_tasks lists. */
 909	rcu_for_each_leaf_node(rsp, rnp)
 910		sync_rcu_preempt_exp_init(rsp, rnp);
 911	if (NUM_RCU_NODES > 1)
 912		sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
 913
 914	put_online_cpus();
 915
 916	/* Wait for snapshotted ->blkd_tasks lists to drain. */
 917	rnp = rcu_get_root(rsp);
 918	wait_event(sync_rcu_preempt_exp_wq,
 919		   sync_rcu_preempt_exp_done(rnp));
 920
 921	/* Clean up and exit. */
 922	smp_mb(); /* ensure expedited GP seen before counter increment. */
 923	ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
 924unlock_mb_ret:
 925	mutex_unlock(&sync_rcu_preempt_exp_mutex);
 926mb_ret:
 927	smp_mb(); /* ensure subsequent action seen after grace period. */
 928}
 929EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 930
 931/**
 932 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
 933 *
 934 * Note that this primitive does not necessarily wait for an RCU grace period
 935 * to complete.  For example, if there are no RCU callbacks queued anywhere
 936 * in the system, then rcu_barrier() is within its rights to return
 937 * immediately, without waiting for anything, much less an RCU grace period.
 938 */
 939void rcu_barrier(void)
 940{
 941	_rcu_barrier(&rcu_preempt_state);
 942}
 943EXPORT_SYMBOL_GPL(rcu_barrier);
 944
 945/*
 946 * Initialize preemptible RCU's state structures.
 947 */
 948static void __init __rcu_init_preempt(void)
 949{
 950	rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
 951}
 952
 953/*
 954 * Check for a task exiting while in a preemptible-RCU read-side
 955 * critical section, clean up if so.  No need to issue warnings,
 956 * as debug_check_no_locks_held() already does this if lockdep
 957 * is enabled.
 958 */
 959void exit_rcu(void)
 960{
 961	struct task_struct *t = current;
 962
 963	if (likely(list_empty(&current->rcu_node_entry)))
 
 964		return;
 965	t->rcu_read_lock_nesting = 1;
 966	barrier();
 967	t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED;
 968	__rcu_read_unlock();
 969}
 970
 971#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
 972
 973static struct rcu_state *rcu_state = &rcu_sched_state;
 974
 975/*
 976 * Tell them what RCU they are running.
 977 */
 978static void __init rcu_bootup_announce(void)
 979{
 980	pr_info("Hierarchical RCU implementation.\n");
 981	rcu_bootup_announce_oddness();
 982}
 983
 984/*
 985 * Return the number of RCU batches processed thus far for debug & stats.
 
 
 
 986 */
 987long rcu_batches_completed(void)
 988{
 989	return rcu_batches_completed_sched();
 
 
 
 
 
 
 
 
 
 990}
 991EXPORT_SYMBOL_GPL(rcu_batches_completed);
 992
 993/*
 994 * Force a quiescent state for RCU, which, because there is no preemptible
 995 * RCU, becomes the same as rcu-sched.
 
 
 
 996 */
 997void rcu_force_quiescent_state(void)
 998{
 999	rcu_sched_force_quiescent_state();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1000}
1001EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
1002
1003/*
1004 * Because preemptible RCU does not exist, we never have to check for
1005 * CPUs being in quiescent states.
1006 */
1007static void rcu_preempt_note_context_switch(int cpu)
1008{
 
 
 
 
 
 
 
 
 
 
 
1009}
 
1010
1011/*
1012 * Because preemptible RCU does not exist, there are never any preempted
1013 * RCU readers.
1014 */
1015static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
1016{
1017	return 0;
1018}
1019
1020#ifdef CONFIG_HOTPLUG_CPU
1021
1022/* Because preemptible RCU does not exist, no quieting of tasks. */
1023static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1024{
1025	raw_spin_unlock_irqrestore(&rnp->lock, flags);
1026}
1027
1028#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1029
1030/*
1031 * Because preemptible RCU does not exist, we never have to check for
1032 * tasks blocked within RCU read-side critical sections.
1033 */
1034static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1035{
 
1036}
1037
1038/*
1039 * Because preemptible RCU does not exist, we never have to check for
1040 * tasks blocked within RCU read-side critical sections.
1041 */
1042static int rcu_print_task_stall(struct rcu_node *rnp)
1043{
1044	return 0;
1045}
 
1046
1047/*
1048 * Because there is no preemptible RCU, there can be no readers blocked,
1049 * so there is no need to check for blocked tasks.  So check only for
1050 * bogus qsmask values.
1051 */
1052static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1053{
1054	WARN_ON_ONCE(rnp->qsmask);
1055}
1056
1057#ifdef CONFIG_HOTPLUG_CPU
1058
1059/*
1060 * Because preemptible RCU does not exist, it never needs to migrate
1061 * tasks that were blocked within RCU read-side critical sections, and
1062 * such non-existent tasks cannot possibly have been blocking the current
1063 * grace period.
1064 */
1065static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1066				     struct rcu_node *rnp,
1067				     struct rcu_data *rdp)
1068{
1069	return 0;
1070}
1071
1072#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1073
1074/*
1075 * Because preemptible RCU does not exist, it never has any callbacks
1076 * to check.
1077 */
1078static void rcu_preempt_check_callbacks(int cpu)
1079{
1080}
1081
1082/*
1083 * Queue an RCU callback for lazy invocation after a grace period.
1084 * This will likely be later named something like "call_rcu_lazy()",
1085 * but this change will require some way of tagging the lazy RCU
1086 * callbacks in the list of pending callbacks.  Until then, this
1087 * function may only be called from __kfree_rcu().
1088 *
1089 * Because there is no preemptible RCU, we use RCU-sched instead.
1090 */
1091void kfree_call_rcu(struct rcu_head *head,
1092		    void (*func)(struct rcu_head *rcu))
1093{
1094	__call_rcu(head, func, &rcu_sched_state, -1, 1);
1095}
1096EXPORT_SYMBOL_GPL(kfree_call_rcu);
1097
1098/*
1099 * Wait for an rcu-preempt grace period, but make it happen quickly.
1100 * But because preemptible RCU does not exist, map to rcu-sched.
1101 */
1102void synchronize_rcu_expedited(void)
1103{
1104	synchronize_sched_expedited();
1105}
1106EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1107
1108#ifdef CONFIG_HOTPLUG_CPU
1109
1110/*
1111 * Because preemptible RCU does not exist, there is never any need to
1112 * report on tasks preempted in RCU read-side critical sections during
1113 * expedited RCU grace periods.
1114 */
1115static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1116			       bool wake)
1117{
1118}
1119
1120#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1121
1122/*
1123 * Because preemptible RCU does not exist, rcu_barrier() is just
1124 * another name for rcu_barrier_sched().
1125 */
1126void rcu_barrier(void)
 
1127{
1128	rcu_barrier_sched();
1129}
1130EXPORT_SYMBOL_GPL(rcu_barrier);
1131
1132/*
1133 * Because preemptible RCU does not exist, it need not be initialized.
1134 */
1135static void __init __rcu_init_preempt(void)
1136{
1137}
1138
1139/*
1140 * Because preemptible RCU does not exist, tasks cannot possibly exit
1141 * while in preemptible RCU read-side critical sections.
1142 */
1143void exit_rcu(void)
1144{
1145}
1146
1147#endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1148
1149#ifdef CONFIG_RCU_BOOST
 
1150
1151#include "../locking/rtmutex_common.h"
1152
1153#ifdef CONFIG_RCU_TRACE
1154
1155static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1156{
1157	if (list_empty(&rnp->blkd_tasks))
1158		rnp->n_balk_blkd_tasks++;
1159	else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1160		rnp->n_balk_exp_gp_tasks++;
1161	else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1162		rnp->n_balk_boost_tasks++;
1163	else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1164		rnp->n_balk_notblocked++;
1165	else if (rnp->gp_tasks != NULL &&
1166		 ULONG_CMP_LT(jiffies, rnp->boost_time))
1167		rnp->n_balk_notyet++;
1168	else
1169		rnp->n_balk_nos++;
1170}
1171
1172#else /* #ifdef CONFIG_RCU_TRACE */
1173
1174static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1175{
1176}
1177
1178#endif /* #else #ifdef CONFIG_RCU_TRACE */
1179
1180static void rcu_wake_cond(struct task_struct *t, int status)
1181{
1182	/*
1183	 * If the thread is yielding, only wake it when this
1184	 * is invoked from idle
1185	 */
1186	if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1187		wake_up_process(t);
1188}
1189
1190/*
1191 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1192 * or ->boost_tasks, advancing the pointer to the next task in the
1193 * ->blkd_tasks list.
1194 *
1195 * Note that irqs must be enabled: boosting the task can block.
1196 * Returns 1 if there are more tasks needing to be boosted.
1197 */
1198static int rcu_boost(struct rcu_node *rnp)
1199{
1200	unsigned long flags;
1201	struct rt_mutex mtx;
1202	struct task_struct *t;
1203	struct list_head *tb;
1204
1205	if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
 
1206		return 0;  /* Nothing left to boost. */
1207
1208	raw_spin_lock_irqsave(&rnp->lock, flags);
1209	smp_mb__after_unlock_lock();
1210
1211	/*
1212	 * Recheck under the lock: all tasks in need of boosting
1213	 * might exit their RCU read-side critical sections on their own.
1214	 */
1215	if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1216		raw_spin_unlock_irqrestore(&rnp->lock, flags);
1217		return 0;
1218	}
1219
1220	/*
1221	 * Preferentially boost tasks blocking expedited grace periods.
1222	 * This cannot starve the normal grace periods because a second
1223	 * expedited grace period must boost all blocked tasks, including
1224	 * those blocking the pre-existing normal grace period.
1225	 */
1226	if (rnp->exp_tasks != NULL) {
1227		tb = rnp->exp_tasks;
1228		rnp->n_exp_boosts++;
1229	} else {
1230		tb = rnp->boost_tasks;
1231		rnp->n_normal_boosts++;
1232	}
1233	rnp->n_tasks_boosted++;
1234
1235	/*
1236	 * We boost task t by manufacturing an rt_mutex that appears to
1237	 * be held by task t.  We leave a pointer to that rt_mutex where
1238	 * task t can find it, and task t will release the mutex when it
1239	 * exits its outermost RCU read-side critical section.  Then
1240	 * simply acquiring this artificial rt_mutex will boost task
1241	 * t's priority.  (Thanks to tglx for suggesting this approach!)
1242	 *
1243	 * Note that task t must acquire rnp->lock to remove itself from
1244	 * the ->blkd_tasks list, which it will do from exit() if from
1245	 * nowhere else.  We therefore are guaranteed that task t will
1246	 * stay around at least until we drop rnp->lock.  Note that
1247	 * rnp->lock also resolves races between our priority boosting
1248	 * and task t's exiting its outermost RCU read-side critical
1249	 * section.
1250	 */
1251	t = container_of(tb, struct task_struct, rcu_node_entry);
1252	rt_mutex_init_proxy_locked(&mtx, t);
1253	t->rcu_boost_mutex = &mtx;
1254	raw_spin_unlock_irqrestore(&rnp->lock, flags);
1255	rt_mutex_lock(&mtx);  /* Side effect: boosts task t's priority. */
1256	rt_mutex_unlock(&mtx);  /* Keep lockdep happy. */
 
1257
1258	return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1259	       ACCESS_ONCE(rnp->boost_tasks) != NULL;
1260}
1261
1262/*
1263 * Priority-boosting kthread.  One per leaf rcu_node and one for the
1264 * root rcu_node.
1265 */
1266static int rcu_boost_kthread(void *arg)
1267{
1268	struct rcu_node *rnp = (struct rcu_node *)arg;
1269	int spincnt = 0;
1270	int more2boost;
1271
1272	trace_rcu_utilization(TPS("Start boost kthread@init"));
1273	for (;;) {
1274		rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1275		trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1276		rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
 
1277		trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1278		rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1279		more2boost = rcu_boost(rnp);
1280		if (more2boost)
1281			spincnt++;
1282		else
1283			spincnt = 0;
1284		if (spincnt > 10) {
1285			rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1286			trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1287			schedule_timeout_interruptible(2);
1288			trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1289			spincnt = 0;
1290		}
1291	}
1292	/* NOTREACHED */
1293	trace_rcu_utilization(TPS("End boost kthread@notreached"));
1294	return 0;
1295}
1296
1297/*
1298 * Check to see if it is time to start boosting RCU readers that are
1299 * blocking the current grace period, and, if so, tell the per-rcu_node
1300 * kthread to start boosting them.  If there is an expedited grace
1301 * period in progress, it is always time to boost.
1302 *
1303 * The caller must hold rnp->lock, which this function releases.
1304 * The ->boost_kthread_task is immortal, so we don't need to worry
1305 * about it going away.
1306 */
1307static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
 
1308{
1309	struct task_struct *t;
1310
1311	if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1312		rnp->n_balk_exp_gp_tasks++;
1313		raw_spin_unlock_irqrestore(&rnp->lock, flags);
1314		return;
1315	}
1316	if (rnp->exp_tasks != NULL ||
1317	    (rnp->gp_tasks != NULL &&
1318	     rnp->boost_tasks == NULL &&
1319	     rnp->qsmask == 0 &&
1320	     ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1321		if (rnp->exp_tasks == NULL)
1322			rnp->boost_tasks = rnp->gp_tasks;
1323		raw_spin_unlock_irqrestore(&rnp->lock, flags);
1324		t = rnp->boost_kthread_task;
1325		if (t)
1326			rcu_wake_cond(t, rnp->boost_kthread_status);
1327	} else {
1328		rcu_initiate_boost_trace(rnp);
1329		raw_spin_unlock_irqrestore(&rnp->lock, flags);
1330	}
1331}
1332
1333/*
1334 * Wake up the per-CPU kthread to invoke RCU callbacks.
1335 */
1336static void invoke_rcu_callbacks_kthread(void)
1337{
1338	unsigned long flags;
1339
1340	local_irq_save(flags);
1341	__this_cpu_write(rcu_cpu_has_work, 1);
1342	if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1343	    current != __this_cpu_read(rcu_cpu_kthread_task)) {
1344		rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1345			      __this_cpu_read(rcu_cpu_kthread_status));
1346	}
1347	local_irq_restore(flags);
1348}
1349
1350/*
1351 * Is the current CPU running the RCU-callbacks kthread?
1352 * Caller must have preemption disabled.
1353 */
1354static bool rcu_is_callbacks_kthread(void)
1355{
1356	return __this_cpu_read(rcu_cpu_kthread_task) == current;
1357}
1358
1359#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1360
1361/*
1362 * Do priority-boost accounting for the start of a new grace period.
1363 */
1364static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1365{
1366	rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1367}
1368
1369/*
1370 * Create an RCU-boost kthread for the specified node if one does not
1371 * already exist.  We only create this kthread for preemptible RCU.
1372 * Returns zero if all is well, a negated errno otherwise.
1373 */
1374static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1375						 struct rcu_node *rnp)
1376{
1377	int rnp_index = rnp - &rsp->node[0];
1378	unsigned long flags;
 
1379	struct sched_param sp;
1380	struct task_struct *t;
1381
1382	if (&rcu_preempt_state != rsp)
1383		return 0;
1384
1385	if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1386		return 0;
1387
1388	rsp->boost = 1;
1389	if (rnp->boost_kthread_task != NULL)
1390		return 0;
1391	t = kthread_create(rcu_boost_kthread, (void *)rnp,
1392			   "rcub/%d", rnp_index);
1393	if (IS_ERR(t))
1394		return PTR_ERR(t);
1395	raw_spin_lock_irqsave(&rnp->lock, flags);
1396	smp_mb__after_unlock_lock();
1397	rnp->boost_kthread_task = t;
1398	raw_spin_unlock_irqrestore(&rnp->lock, flags);
1399	sp.sched_priority = RCU_BOOST_PRIO;
1400	sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1401	wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1402	return 0;
1403}
1404
1405static void rcu_kthread_do_work(void)
1406{
1407	rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1408	rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1409	rcu_preempt_do_callbacks();
1410}
1411
1412static void rcu_cpu_kthread_setup(unsigned int cpu)
1413{
1414	struct sched_param sp;
1415
1416	sp.sched_priority = RCU_KTHREAD_PRIO;
1417	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1418}
1419
1420static void rcu_cpu_kthread_park(unsigned int cpu)
1421{
1422	per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1423}
1424
1425static int rcu_cpu_kthread_should_run(unsigned int cpu)
1426{
1427	return __this_cpu_read(rcu_cpu_has_work);
1428}
1429
1430/*
1431 * Per-CPU kernel thread that invokes RCU callbacks.  This replaces the
1432 * RCU softirq used in flavors and configurations of RCU that do not
1433 * support RCU priority boosting.
1434 */
1435static void rcu_cpu_kthread(unsigned int cpu)
1436{
1437	unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1438	char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1439	int spincnt;
1440
1441	for (spincnt = 0; spincnt < 10; spincnt++) {
1442		trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1443		local_bh_disable();
1444		*statusp = RCU_KTHREAD_RUNNING;
1445		this_cpu_inc(rcu_cpu_kthread_loops);
1446		local_irq_disable();
1447		work = *workp;
1448		*workp = 0;
1449		local_irq_enable();
1450		if (work)
1451			rcu_kthread_do_work();
1452		local_bh_enable();
1453		if (*workp == 0) {
1454			trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1455			*statusp = RCU_KTHREAD_WAITING;
1456			return;
1457		}
1458	}
1459	*statusp = RCU_KTHREAD_YIELDING;
1460	trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1461	schedule_timeout_interruptible(2);
1462	trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1463	*statusp = RCU_KTHREAD_WAITING;
1464}
1465
1466/*
1467 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1468 * served by the rcu_node in question.  The CPU hotplug lock is still
1469 * held, so the value of rnp->qsmaskinit will be stable.
1470 *
1471 * We don't include outgoingcpu in the affinity set, use -1 if there is
1472 * no outgoing CPU.  If there are no CPUs left in the affinity set,
1473 * this function allows the kthread to execute on any CPU.
1474 */
1475static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1476{
1477	struct task_struct *t = rnp->boost_kthread_task;
1478	unsigned long mask = rnp->qsmaskinit;
1479	cpumask_var_t cm;
1480	int cpu;
1481
1482	if (!t)
1483		return;
1484	if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1485		return;
1486	for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1487		if ((mask & 0x1) && cpu != outgoingcpu)
 
1488			cpumask_set_cpu(cpu, cm);
1489	if (cpumask_weight(cm) == 0) {
1490		cpumask_setall(cm);
1491		for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1492			cpumask_clear_cpu(cpu, cm);
1493		WARN_ON_ONCE(cpumask_weight(cm) == 0);
1494	}
1495	set_cpus_allowed_ptr(t, cm);
1496	free_cpumask_var(cm);
1497}
1498
1499static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1500	.store			= &rcu_cpu_kthread_task,
1501	.thread_should_run	= rcu_cpu_kthread_should_run,
1502	.thread_fn		= rcu_cpu_kthread,
1503	.thread_comm		= "rcuc/%u",
1504	.setup			= rcu_cpu_kthread_setup,
1505	.park			= rcu_cpu_kthread_park,
1506};
1507
1508/*
1509 * Spawn all kthreads -- called as soon as the scheduler is running.
1510 */
1511static int __init rcu_spawn_kthreads(void)
1512{
1513	struct rcu_node *rnp;
1514	int cpu;
1515
1516	rcu_scheduler_fully_active = 1;
1517	for_each_possible_cpu(cpu)
1518		per_cpu(rcu_cpu_has_work, cpu) = 0;
1519	BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1520	rnp = rcu_get_root(rcu_state);
1521	(void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1522	if (NUM_RCU_NODES > 1) {
1523		rcu_for_each_leaf_node(rcu_state, rnp)
1524			(void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1525	}
1526	return 0;
1527}
1528early_initcall(rcu_spawn_kthreads);
1529
1530static void rcu_prepare_kthreads(int cpu)
1531{
1532	struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1533	struct rcu_node *rnp = rdp->mynode;
1534
1535	/* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1536	if (rcu_scheduler_fully_active)
1537		(void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1538}
1539
1540#else /* #ifdef CONFIG_RCU_BOOST */
1541
1542static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
 
1543{
1544	raw_spin_unlock_irqrestore(&rnp->lock, flags);
1545}
1546
1547static void invoke_rcu_callbacks_kthread(void)
1548{
1549	WARN_ON_ONCE(1);
1550}
1551
1552static bool rcu_is_callbacks_kthread(void)
1553{
1554	return false;
1555}
1556
1557static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1558{
1559}
1560
1561static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1562{
1563}
1564
1565static int __init rcu_scheduler_really_started(void)
1566{
1567	rcu_scheduler_fully_active = 1;
1568	return 0;
1569}
1570early_initcall(rcu_scheduler_really_started);
1571
1572static void rcu_prepare_kthreads(int cpu)
1573{
1574}
1575
1576#endif /* #else #ifdef CONFIG_RCU_BOOST */
1577
1578#if !defined(CONFIG_RCU_FAST_NO_HZ)
1579
1580/*
1581 * Check to see if any future RCU-related work will need to be done
1582 * by the current CPU, even if none need be done immediately, returning
1583 * 1 if so.  This function is part of the RCU implementation; it is -not-
1584 * an exported member of the RCU API.
1585 *
1586 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1587 * any flavor of RCU.
1588 */
1589#ifndef CONFIG_RCU_NOCB_CPU_ALL
1590int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1591{
1592	*delta_jiffies = ULONG_MAX;
1593	return rcu_cpu_has_callbacks(cpu, NULL);
1594}
1595#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1596
1597/*
1598 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1599 * after it.
1600 */
1601static void rcu_cleanup_after_idle(int cpu)
1602{
1603}
1604
1605/*
1606 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1607 * is nothing.
1608 */
1609static void rcu_prepare_for_idle(int cpu)
1610{
1611}
1612
1613/*
1614 * Don't bother keeping a running count of the number of RCU callbacks
1615 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1616 */
1617static void rcu_idle_count_callbacks_posted(void)
1618{
1619}
1620
1621#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1622
1623/*
1624 * This code is invoked when a CPU goes idle, at which point we want
1625 * to have the CPU do everything required for RCU so that it can enter
1626 * the energy-efficient dyntick-idle mode.  This is handled by a
1627 * state machine implemented by rcu_prepare_for_idle() below.
1628 *
1629 * The following three proprocessor symbols control this state machine:
1630 *
1631 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1632 *	to sleep in dyntick-idle mode with RCU callbacks pending.  This
1633 *	is sized to be roughly one RCU grace period.  Those energy-efficiency
1634 *	benchmarkers who might otherwise be tempted to set this to a large
1635 *	number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1636 *	system.  And if you are -that- concerned about energy efficiency,
1637 *	just power the system down and be done with it!
1638 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1639 *	permitted to sleep in dyntick-idle mode with only lazy RCU
1640 *	callbacks pending.  Setting this too high can OOM your system.
1641 *
1642 * The values below work well in practice.  If future workloads require
1643 * adjustment, they can be converted into kernel config parameters, though
1644 * making the state machine smarter might be a better option.
1645 */
1646#define RCU_IDLE_GP_DELAY 4		/* Roughly one grace period. */
1647#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ)	/* Roughly six seconds. */
1648
1649static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1650module_param(rcu_idle_gp_delay, int, 0644);
1651static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1652module_param(rcu_idle_lazy_gp_delay, int, 0644);
1653
1654extern int tick_nohz_active;
1655
1656/*
1657 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1658 * only if it has been awhile since the last time we did so.  Afterwards,
1659 * if there are any callbacks ready for immediate invocation, return true.
1660 */
1661static bool __maybe_unused rcu_try_advance_all_cbs(void)
1662{
1663	bool cbs_ready = false;
1664	struct rcu_data *rdp;
1665	struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1666	struct rcu_node *rnp;
1667	struct rcu_state *rsp;
1668
1669	/* Exit early if we advanced recently. */
1670	if (jiffies == rdtp->last_advance_all)
1671		return 0;
1672	rdtp->last_advance_all = jiffies;
1673
1674	for_each_rcu_flavor(rsp) {
1675		rdp = this_cpu_ptr(rsp->rda);
1676		rnp = rdp->mynode;
1677
1678		/*
1679		 * Don't bother checking unless a grace period has
1680		 * completed since we last checked and there are
1681		 * callbacks not yet ready to invoke.
1682		 */
1683		if (rdp->completed != rnp->completed &&
1684		    rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1685			note_gp_changes(rsp, rdp);
 
 
1686
1687		if (cpu_has_callbacks_ready_to_invoke(rdp))
1688			cbs_ready = true;
1689	}
1690	return cbs_ready;
1691}
1692
1693/*
1694 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1695 * to invoke.  If the CPU has callbacks, try to advance them.  Tell the
1696 * caller to set the timeout based on whether or not there are non-lazy
1697 * callbacks.
1698 *
1699 * The caller must have disabled interrupts.
1700 */
1701#ifndef CONFIG_RCU_NOCB_CPU_ALL
1702int rcu_needs_cpu(int cpu, unsigned long *dj)
1703{
1704	struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
 
1705
1706	/* Snapshot to detect later posting of non-lazy callback. */
1707	rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1708
1709	/* If no callbacks, RCU doesn't need the CPU. */
1710	if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1711		*dj = ULONG_MAX;
 
1712		return 0;
1713	}
1714
1715	/* Attempt to advance callbacks. */
1716	if (rcu_try_advance_all_cbs()) {
1717		/* Some ready to invoke, so initiate later invocation. */
1718		invoke_rcu_core();
1719		return 1;
1720	}
1721	rdtp->last_accelerate = jiffies;
1722
1723	/* Request timer delay depending on laziness, and round. */
1724	if (!rdtp->all_lazy) {
1725		*dj = round_up(rcu_idle_gp_delay + jiffies,
1726			       rcu_idle_gp_delay) - jiffies;
1727	} else {
1728		*dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1729	}
1730	return 0;
1731}
1732#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1733
1734/*
1735 * Prepare a CPU for idle from an RCU perspective.  The first major task
1736 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1737 * The second major task is to check to see if a non-lazy callback has
1738 * arrived at a CPU that previously had only lazy callbacks.  The third
1739 * major task is to accelerate (that is, assign grace-period numbers to)
1740 * any recently arrived callbacks.
1741 *
1742 * The caller must have disabled interrupts.
1743 */
1744static void rcu_prepare_for_idle(int cpu)
1745{
1746#ifndef CONFIG_RCU_NOCB_CPU_ALL
1747	struct rcu_data *rdp;
1748	struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1749	struct rcu_node *rnp;
1750	struct rcu_state *rsp;
1751	int tne;
1752
 
 
 
 
1753	/* Handle nohz enablement switches conservatively. */
1754	tne = ACCESS_ONCE(tick_nohz_active);
1755	if (tne != rdtp->tick_nohz_enabled_snap) {
1756		if (rcu_cpu_has_callbacks(cpu, NULL))
1757			invoke_rcu_core(); /* force nohz to see update. */
1758		rdtp->tick_nohz_enabled_snap = tne;
1759		return;
1760	}
1761	if (!tne)
1762		return;
1763
1764	/* If this is a no-CBs CPU, no callbacks, just return. */
1765	if (rcu_is_nocb_cpu(cpu))
1766		return;
1767
1768	/*
1769	 * If a non-lazy callback arrived at a CPU having only lazy
1770	 * callbacks, invoke RCU core for the side-effect of recalculating
1771	 * idle duration on re-entry to idle.
1772	 */
1773	if (rdtp->all_lazy &&
1774	    rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1775		rdtp->all_lazy = false;
1776		rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1777		invoke_rcu_core();
1778		return;
1779	}
1780
1781	/*
1782	 * If we have not yet accelerated this jiffy, accelerate all
1783	 * callbacks on this CPU.
1784	 */
1785	if (rdtp->last_accelerate == jiffies)
1786		return;
1787	rdtp->last_accelerate = jiffies;
1788	for_each_rcu_flavor(rsp) {
1789		rdp = per_cpu_ptr(rsp->rda, cpu);
1790		if (!*rdp->nxttail[RCU_DONE_TAIL])
1791			continue;
1792		rnp = rdp->mynode;
1793		raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1794		smp_mb__after_unlock_lock();
1795		rcu_accelerate_cbs(rsp, rnp, rdp);
1796		raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
 
1797	}
1798#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1799}
1800
1801/*
1802 * Clean up for exit from idle.  Attempt to advance callbacks based on
1803 * any grace periods that elapsed while the CPU was idle, and if any
1804 * callbacks are now ready to invoke, initiate invocation.
1805 */
1806static void rcu_cleanup_after_idle(int cpu)
1807{
1808#ifndef CONFIG_RCU_NOCB_CPU_ALL
1809	if (rcu_is_nocb_cpu(cpu))
 
 
1810		return;
1811	if (rcu_try_advance_all_cbs())
1812		invoke_rcu_core();
1813#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1814}
1815
 
 
 
 
1816/*
1817 * Keep a running count of the number of non-lazy callbacks posted
1818 * on this CPU.  This running counter (which is never decremented) allows
1819 * rcu_prepare_for_idle() to detect when something out of the idle loop
1820 * posts a callback, even if an equal number of callbacks are invoked.
1821 * Of course, callbacks should only be posted from within a trace event
1822 * designed to be called from idle or from within RCU_NONIDLE().
 
 
 
 
 
 
 
 
 
 
 
 
 
1823 */
1824static void rcu_idle_count_callbacks_posted(void)
 
 
 
 
 
 
1825{
1826	__this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
 
 
 
 
 
 
 
 
 
 
 
 
1827}
 
1828
1829/*
1830 * Data for flushing lazy RCU callbacks at OOM time.
 
 
1831 */
1832static atomic_t oom_callback_count;
1833static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1834
1835/*
1836 * RCU OOM callback -- decrement the outstanding count and deliver the
1837 * wake-up if we are the last one.
 
1838 */
1839static void rcu_oom_callback(struct rcu_head *rhp)
 
1840{
1841	if (atomic_dec_and_test(&oom_callback_count))
1842		wake_up(&oom_callback_wq);
 
 
 
 
 
 
 
1843}
1844
1845/*
1846 * Post an rcu_oom_notify callback on the current CPU if it has at
1847 * least one lazy callback.  This will unnecessarily post callbacks
1848 * to CPUs that already have a non-lazy callback at the end of their
1849 * callback list, but this is an infrequent operation, so accept some
1850 * extra overhead to keep things simple.
 
 
 
1851 */
1852static void rcu_oom_notify_cpu(void *unused)
1853{
1854	struct rcu_state *rsp;
1855	struct rcu_data *rdp;
1856
1857	for_each_rcu_flavor(rsp) {
1858		rdp = __this_cpu_ptr(rsp->rda);
1859		if (rdp->qlen_lazy != 0) {
1860			atomic_inc(&oom_callback_count);
1861			rsp->call(&rdp->oom_head, rcu_oom_callback);
1862		}
1863	}
1864}
1865
1866/*
1867 * If low on memory, ensure that each CPU has a non-lazy callback.
1868 * This will wake up CPUs that have only lazy callbacks, in turn
1869 * ensuring that they free up the corresponding memory in a timely manner.
1870 * Because an uncertain amount of memory will be freed in some uncertain
1871 * timeframe, we do not claim to have freed anything.
1872 */
1873static int rcu_oom_notify(struct notifier_block *self,
1874			  unsigned long notused, void *nfreed)
1875{
1876	int cpu;
1877
1878	/* Wait for callbacks from earlier instance to complete. */
1879	wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1880	smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1881
1882	/*
1883	 * Prevent premature wakeup: ensure that all increments happen
1884	 * before there is a chance of the counter reaching zero.
1885	 */
1886	atomic_set(&oom_callback_count, 1);
1887
1888	get_online_cpus();
1889	for_each_online_cpu(cpu) {
1890		smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1891		cond_resched();
1892	}
1893	put_online_cpus();
1894
1895	/* Unconditionally decrement: no need to wake ourselves up. */
1896	atomic_dec(&oom_callback_count);
1897
1898	return NOTIFY_OK;
1899}
1900
1901static struct notifier_block rcu_oom_nb = {
1902	.notifier_call = rcu_oom_notify
1903};
1904
1905static int __init rcu_register_oom_notifier(void)
1906{
1907	register_oom_notifier(&rcu_oom_nb);
1908	return 0;
1909}
1910early_initcall(rcu_register_oom_notifier);
1911
1912#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1913
1914#ifdef CONFIG_RCU_CPU_STALL_INFO
1915
1916#ifdef CONFIG_RCU_FAST_NO_HZ
1917
1918static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1919{
1920	struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1921	unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1922
1923	sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1924		rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1925		ulong2long(nlpd),
1926		rdtp->all_lazy ? 'L' : '.',
1927		rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1928}
1929
1930#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1931
1932static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 
 
1933{
1934	*cp = '\0';
 
 
 
1935}
1936
1937#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1938
1939/* Initiate the stall-info list. */
1940static void print_cpu_stall_info_begin(void)
 
1941{
1942	pr_cont("\n");
 
 
 
1943}
1944
1945/*
1946 * Print out diagnostic information for the specified stalled CPU.
1947 *
1948 * If the specified CPU is aware of the current RCU grace period
1949 * (flavor specified by rsp), then print the number of scheduling
1950 * clock interrupts the CPU has taken during the time that it has
1951 * been aware.  Otherwise, print the number of RCU grace periods
1952 * that this CPU is ignorant of, for example, "1" if the CPU was
1953 * aware of the previous grace period.
1954 *
1955 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1956 */
1957static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1958{
1959	char fast_no_hz[72];
1960	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1961	struct rcu_dynticks *rdtp = rdp->dynticks;
1962	char *ticks_title;
1963	unsigned long ticks_value;
1964
1965	if (rsp->gpnum == rdp->gpnum) {
1966		ticks_title = "ticks this GP";
1967		ticks_value = rdp->ticks_this_gp;
1968	} else {
1969		ticks_title = "GPs behind";
1970		ticks_value = rsp->gpnum - rdp->gpnum;
1971	}
1972	print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1973	pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1974	       cpu, ticks_value, ticks_title,
1975	       atomic_read(&rdtp->dynticks) & 0xfff,
1976	       rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1977	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1978	       fast_no_hz);
1979}
1980
1981/* Terminate the stall-info list. */
1982static void print_cpu_stall_info_end(void)
1983{
1984	pr_err("\t");
 
 
1985}
1986
1987/* Zero ->ticks_this_gp for all flavors of RCU. */
1988static void zero_cpu_stall_ticks(struct rcu_data *rdp)
 
 
 
1989{
1990	rdp->ticks_this_gp = 0;
1991	rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1992}
1993
1994/* Increment ->ticks_this_gp for all flavors of RCU. */
1995static void increment_cpu_stall_ticks(void)
1996{
1997	struct rcu_state *rsp;
1998
1999	for_each_rcu_flavor(rsp)
2000		__this_cpu_ptr(rsp->rda)->ticks_this_gp++;
2001}
2002
2003#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
2004
2005static void print_cpu_stall_info_begin(void)
2006{
2007	pr_cont(" {");
 
2008}
2009
2010static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
 
2011{
2012	pr_cont(" %d", cpu);
 
 
2013}
2014
2015static void print_cpu_stall_info_end(void)
2016{
2017	pr_cont("} ");
2018}
 
 
 
 
 
 
 
 
 
2019
2020static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2021{
2022}
 
2023
2024static void increment_cpu_stall_ticks(void)
2025{
 
 
 
 
 
 
 
 
 
2026}
2027
2028#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
 
 
 
 
 
 
2029
2030#ifdef CONFIG_RCU_NOCB_CPU
 
 
2031
2032/*
2033 * Offload callback processing from the boot-time-specified set of CPUs
2034 * specified by rcu_nocb_mask.  For each CPU in the set, there is a
2035 * kthread created that pulls the callbacks from the corresponding CPU,
2036 * waits for a grace period to elapse, and invokes the callbacks.
2037 * The no-CBs CPUs do a wake_up() on their kthread when they insert
2038 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2039 * has been specified, in which case each kthread actively polls its
2040 * CPU.  (Which isn't so great for energy efficiency, but which does
2041 * reduce RCU's overhead on that CPU.)
2042 *
2043 * This is intended to be used in conjunction with Frederic Weisbecker's
2044 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2045 * running CPU-bound user-mode computations.
2046 *
2047 * Offloading of callback processing could also in theory be used as
2048 * an energy-efficiency measure because CPUs with no RCU callbacks
2049 * queued are more aggressive about entering dyntick-idle mode.
2050 */
 
 
 
 
 
2051
 
2052
2053/* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2054static int __init rcu_nocb_setup(char *str)
2055{
2056	alloc_bootmem_cpumask_var(&rcu_nocb_mask);
2057	have_rcu_nocb_mask = true;
2058	cpulist_parse(str, rcu_nocb_mask);
2059	return 1;
2060}
2061__setup("rcu_nocbs=", rcu_nocb_setup);
 
 
 
 
2062
2063static int __init parse_rcu_nocb_poll(char *arg)
2064{
2065	rcu_nocb_poll = 1;
2066	return 0;
2067}
2068early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2069
2070/*
2071 * Do any no-CBs CPUs need another grace period?
 
 
 
2072 *
2073 * Interrupts must be disabled.  If the caller does not hold the root
2074 * rnp_node structure's ->lock, the results are advisory only.
2075 */
2076static int rcu_nocb_needs_gp(struct rcu_state *rsp)
 
2077{
2078	struct rcu_node *rnp = rcu_get_root(rsp);
2079
2080	return rnp->need_future_gp[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2081}
2082
2083/*
2084 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2085 * grace period.
 
 
 
 
2086 */
2087static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
 
2088{
2089	wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
 
 
 
 
2090}
2091
2092/*
2093 * Set the root rcu_node structure's ->need_future_gp field
2094 * based on the sum of those of all rcu_node structures.  This does
2095 * double-count the root rcu_node structure's requests, but this
2096 * is necessary to handle the possibility of a rcu_nocb_kthread()
2097 * having awakened during the time that the rcu_node structures
2098 * were being updated for the end of the previous grace period.
2099 */
2100static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2101{
2102	rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
 
 
 
 
2103}
2104
2105static void rcu_init_one_nocb(struct rcu_node *rnp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2106{
2107	init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2108	init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2109}
 
 
 
 
 
 
 
 
 
 
2110
2111#ifndef CONFIG_RCU_NOCB_CPU_ALL
2112/* Is the specified CPU a no-CPUs CPU? */
2113bool rcu_is_nocb_cpu(int cpu)
2114{
2115	if (have_rcu_nocb_mask)
2116		return cpumask_test_cpu(cpu, rcu_nocb_mask);
2117	return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2118}
2119#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
2120
2121/*
2122 * Enqueue the specified string of rcu_head structures onto the specified
2123 * CPU's no-CBs lists.  The CPU is specified by rdp, the head of the
2124 * string by rhp, and the tail of the string by rhtp.  The non-lazy/lazy
2125 * counts are supplied by rhcount and rhcount_lazy.
2126 *
2127 * If warranted, also wake up the kthread servicing this CPUs queues.
2128 */
2129static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2130				    struct rcu_head *rhp,
2131				    struct rcu_head **rhtp,
2132				    int rhcount, int rhcount_lazy,
2133				    unsigned long flags)
2134{
2135	int len;
2136	struct rcu_head **old_rhpp;
 
2137	struct task_struct *t;
2138
2139	/* Enqueue the callback on the nocb list and update counts. */
2140	old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2141	ACCESS_ONCE(*old_rhpp) = rhp;
2142	atomic_long_add(rhcount, &rdp->nocb_q_count);
2143	atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2144
2145	/* If we are not being polled and there is a kthread, awaken it ... */
2146	t = ACCESS_ONCE(rdp->nocb_kthread);
2147	if (rcu_nocb_poll || !t) {
2148		trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
 
2149				    TPS("WakeNotPoll"));
2150		return;
2151	}
2152	len = atomic_long_read(&rdp->nocb_q_count);
2153	if (old_rhpp == &rdp->nocb_head) {
 
 
2154		if (!irqs_disabled_flags(flags)) {
2155			wake_up(&rdp->nocb_wq); /* ... if queue was empty ... */
2156			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
 
 
2157					    TPS("WakeEmpty"));
2158		} else {
2159			rdp->nocb_defer_wakeup = true;
2160			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2161					    TPS("WakeEmptyIsDeferred"));
2162		}
2163		rdp->qlen_last_fqs_check = 0;
2164	} else if (len > rdp->qlen_last_fqs_check + qhimark) {
2165		wake_up_process(t); /* ... or if many callbacks queued. */
2166		rdp->qlen_last_fqs_check = LONG_MAX / 2;
2167		trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf"));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2168	} else {
2169		trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
 
2170	}
2171	return;
2172}
2173
2174/*
2175 * This is a helper for __call_rcu(), which invokes this when the normal
2176 * callback queue is inoperable.  If this is not a no-CBs CPU, this
2177 * function returns failure back to __call_rcu(), which can complain
2178 * appropriately.
2179 *
2180 * Otherwise, this function queues the callback where the corresponding
2181 * "rcuo" kthread can find it.
 
 
 
 
 
 
 
 
2182 */
2183static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2184			    bool lazy, unsigned long flags)
2185{
 
2186
2187	if (!rcu_is_nocb_cpu(rdp->cpu))
2188		return 0;
2189	__call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2190	if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2191		trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2192					 (unsigned long)rhp->func,
2193					 -atomic_long_read(&rdp->nocb_q_count_lazy),
2194					 -atomic_long_read(&rdp->nocb_q_count));
2195	else
2196		trace_rcu_callback(rdp->rsp->name, rhp,
2197				   -atomic_long_read(&rdp->nocb_q_count_lazy),
2198				   -atomic_long_read(&rdp->nocb_q_count));
2199	return 1;
2200}
2201
2202/*
2203 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2204 * not a no-CBs CPU.
2205 */
2206static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2207						     struct rcu_data *rdp,
2208						     unsigned long flags)
2209{
2210	long ql = rsp->qlen;
2211	long qll = rsp->qlen_lazy;
2212
2213	/* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2214	if (!rcu_is_nocb_cpu(smp_processor_id()))
2215		return 0;
2216	rsp->qlen = 0;
2217	rsp->qlen_lazy = 0;
2218
2219	/* First, enqueue the donelist, if any.  This preserves CB ordering. */
2220	if (rsp->orphan_donelist != NULL) {
2221		__call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2222					rsp->orphan_donetail, ql, qll, flags);
2223		ql = qll = 0;
2224		rsp->orphan_donelist = NULL;
2225		rsp->orphan_donetail = &rsp->orphan_donelist;
2226	}
2227	if (rsp->orphan_nxtlist != NULL) {
2228		__call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2229					rsp->orphan_nxttail, ql, qll, flags);
2230		ql = qll = 0;
2231		rsp->orphan_nxtlist = NULL;
2232		rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2233	}
2234	return 1;
 
 
 
 
 
 
 
 
 
2235}
2236
 
2237/*
2238 * If necessary, kick off a new grace period, and either way wait
2239 * for a subsequent grace period to complete.
2240 */
2241static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2242{
2243	unsigned long c;
2244	bool d;
 
 
2245	unsigned long flags;
2246	struct rcu_node *rnp = rdp->mynode;
2247
2248	raw_spin_lock_irqsave(&rnp->lock, flags);
2249	smp_mb__after_unlock_lock();
2250	c = rcu_start_future_gp(rnp, rdp);
2251	raw_spin_unlock_irqrestore(&rnp->lock, flags);
 
 
 
2252
2253	/*
2254	 * Wait for the grace period.  Do so interruptibly to avoid messing
2255	 * up the load average.
 
2256	 */
2257	trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2258	for (;;) {
2259		wait_event_interruptible(
2260			rnp->nocb_gp_wq[c & 0x1],
2261			(d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2262		if (likely(d))
2263			break;
2264		flush_signals(current);
2265		trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2266	}
2267	trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2268	smp_mb(); /* Ensure that CB invocation happens after GP end. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2269}
2270
2271/*
2272 * Per-rcu_data kthread, but only for no-CBs CPUs.  Each kthread invokes
2273 * callbacks queued by the corresponding no-CBs CPU.
 
 
 
 
2274 */
2275static int rcu_nocb_kthread(void *arg)
2276{
2277	int c, cl;
2278	bool firsttime = 1;
2279	struct rcu_head *list;
2280	struct rcu_head *next;
2281	struct rcu_head **tail;
2282	struct rcu_data *rdp = arg;
2283
2284	/* Each pass through this loop invokes one batch of callbacks */
2285	for (;;) {
2286		/* If not polling, wait for next batch of callbacks. */
2287		if (!rcu_nocb_poll) {
2288			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2289					    TPS("Sleep"));
2290			wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head);
2291			/* Memory barrier provide by xchg() below. */
2292		} else if (firsttime) {
2293			firsttime = 0;
2294			trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2295					    TPS("Poll"));
2296		}
2297		list = ACCESS_ONCE(rdp->nocb_head);
2298		if (!list) {
2299			if (!rcu_nocb_poll)
2300				trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2301						    TPS("WokeEmpty"));
2302			schedule_timeout_interruptible(1);
2303			flush_signals(current);
2304			continue;
2305		}
2306		firsttime = 1;
2307		trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2308				    TPS("WokeNonEmpty"));
2309
2310		/*
2311		 * Extract queued callbacks, update counts, and wait
2312		 * for a grace period to elapse.
2313		 */
2314		ACCESS_ONCE(rdp->nocb_head) = NULL;
2315		tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2316		c = atomic_long_xchg(&rdp->nocb_q_count, 0);
2317		cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2318		ACCESS_ONCE(rdp->nocb_p_count) += c;
2319		ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl;
2320		rcu_nocb_wait_gp(rdp);
2321
2322		/* Each pass through the following loop invokes a callback. */
2323		trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2324		c = cl = 0;
2325		while (list) {
2326			next = list->next;
2327			/* Wait for enqueuing to complete, if needed. */
2328			while (next == NULL && &list->next != tail) {
2329				trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2330						    TPS("WaitQueue"));
2331				schedule_timeout_interruptible(1);
2332				trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2333						    TPS("WokeQueue"));
2334				next = list->next;
2335			}
2336			debug_rcu_head_unqueue(list);
2337			local_bh_disable();
2338			if (__rcu_reclaim(rdp->rsp->name, list))
2339				cl++;
2340			c++;
2341			local_bh_enable();
2342			list = next;
2343		}
2344		trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2345		ACCESS_ONCE(rdp->nocb_p_count) -= c;
2346		ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2347		rdp->n_nocbs_invoked += c;
2348	}
2349	return 0;
2350}
2351
2352/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2353static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2354{
2355	return ACCESS_ONCE(rdp->nocb_defer_wakeup);
 
2356}
2357
2358/* Do a deferred wakeup of rcu_nocb_kthread(). */
2359static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2360{
2361	if (!rcu_nocb_need_deferred_wakeup(rdp))
2362		return;
2363	ACCESS_ONCE(rdp->nocb_defer_wakeup) = false;
2364	wake_up(&rdp->nocb_wq);
2365	trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWakeEmpty"));
2366}
2367
2368/* Initialize per-rcu_data variables for no-CBs CPUs. */
2369static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2370{
2371	rdp->nocb_tail = &rdp->nocb_head;
2372	init_waitqueue_head(&rdp->nocb_wq);
2373}
2374
2375/* Create a kthread for each RCU flavor for each no-CBs CPU. */
2376static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2377{
2378	int cpu;
2379	struct rcu_data *rdp;
2380	struct task_struct *t;
 
 
 
 
2381
2382	if (rcu_nocb_mask == NULL)
2383		return;
2384	for_each_cpu(cpu, rcu_nocb_mask) {
2385		rdp = per_cpu_ptr(rsp->rda, cpu);
2386		t = kthread_run(rcu_nocb_kthread, rdp,
2387				"rcuo%c/%d", rsp->abbr, cpu);
2388		BUG_ON(IS_ERR(t));
2389		ACCESS_ONCE(rdp->nocb_kthread) = t;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2390	}
2391}
2392
2393/* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2394static bool init_nocb_callback_list(struct rcu_data *rdp)
2395{
2396	if (rcu_nocb_mask == NULL ||
2397	    !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2398		return false;
2399	rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2400	return true;
2401}
2402
2403#else /* #ifdef CONFIG_RCU_NOCB_CPU */
 
2404
2405static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2406{
2407	return 0;
2408}
2409
2410static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2411{
2412}
2413
2414static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2415{
2416}
2417
2418static void rcu_init_one_nocb(struct rcu_node *rnp)
2419{
 
 
 
 
2420}
2421
2422static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2423			    bool lazy, unsigned long flags)
 
 
 
2424{
 
 
 
 
 
 
 
 
2425	return 0;
2426}
2427
2428static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2429						     struct rcu_data *rdp,
2430						     unsigned long flags)
2431{
2432	return 0;
2433}
2434
2435static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
 
 
 
 
2436{
 
 
 
 
 
 
 
 
 
 
 
 
 
2437}
2438
2439static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
 
2440{
2441	return false;
 
 
 
 
 
 
 
 
2442}
2443
2444static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
 
 
 
 
 
2445{
 
 
 
 
 
 
 
 
2446}
2447
2448static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2449{
 
2450}
 
2451
2452static bool init_nocb_callback_list(struct rcu_data *rdp)
 
 
2453{
2454	return false;
2455}
 
2456
2457#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2458
2459/*
2460 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2461 * arbitrarily long period of time with the scheduling-clock tick turned
2462 * off.  RCU will be paying attention to this CPU because it is in the
2463 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2464 * machine because the scheduling-clock tick has been disabled.  Therefore,
2465 * if an adaptive-ticks CPU is failing to respond to the current grace
2466 * period and has not be idle from an RCU perspective, kick it.
2467 */
2468static void rcu_kick_nohz_cpu(int cpu)
2469{
2470#ifdef CONFIG_NO_HZ_FULL
2471	if (tick_nohz_full_cpu(cpu))
2472		smp_send_reschedule(cpu);
2473#endif /* #ifdef CONFIG_NO_HZ_FULL */
 
 
 
 
 
 
2474}
2475
 
 
 
 
 
 
2476
2477#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2478
2479/*
2480 * Define RCU flavor that holds sysidle state.  This needs to be the
2481 * most active flavor of RCU.
2482 */
2483#ifdef CONFIG_PREEMPT_RCU
2484static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state;
2485#else /* #ifdef CONFIG_PREEMPT_RCU */
2486static struct rcu_state *rcu_sysidle_state = &rcu_sched_state;
2487#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2488
2489static int full_sysidle_state;		/* Current system-idle state. */
2490#define RCU_SYSIDLE_NOT		0	/* Some CPU is not idle. */
2491#define RCU_SYSIDLE_SHORT	1	/* All CPUs idle for brief period. */
2492#define RCU_SYSIDLE_LONG	2	/* All CPUs idle for long enough. */
2493#define RCU_SYSIDLE_FULL	3	/* All CPUs idle, ready for sysidle. */
2494#define RCU_SYSIDLE_FULL_NOTED	4	/* Actually entered sysidle state. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2495
2496/*
2497 * Invoked to note exit from irq or task transition to idle.  Note that
2498 * usermode execution does -not- count as idle here!  After all, we want
2499 * to detect full-system idle states, not RCU quiescent states and grace
2500 * periods.  The caller must have disabled interrupts.
2501 */
2502static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
 
2503{
2504	unsigned long j;
 
2505
2506	/* Adjust nesting, check for fully idle. */
2507	if (irq) {
2508		rdtp->dynticks_idle_nesting--;
2509		WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2510		if (rdtp->dynticks_idle_nesting != 0)
2511			return;  /* Still not fully idle. */
2512	} else {
2513		if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2514		    DYNTICK_TASK_NEST_VALUE) {
2515			rdtp->dynticks_idle_nesting = 0;
2516		} else {
2517			rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2518			WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2519			return;  /* Still not fully idle. */
2520		}
2521	}
 
 
2522
2523	/* Record start of fully idle period. */
2524	j = jiffies;
2525	ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2526	smp_mb__before_atomic_inc();
2527	atomic_inc(&rdtp->dynticks_idle);
2528	smp_mb__after_atomic_inc();
2529	WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2530}
 
2531
2532/*
2533 * Unconditionally force exit from full system-idle state.  This is
2534 * invoked when a normal CPU exits idle, but must be called separately
2535 * for the timekeeping CPU (tick_do_timer_cpu).  The reason for this
2536 * is that the timekeeping CPU is permitted to take scheduling-clock
2537 * interrupts while the system is in system-idle state, and of course
2538 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2539 * interrupt from any other type of interrupt.
2540 */
2541void rcu_sysidle_force_exit(void)
2542{
2543	int oldstate = ACCESS_ONCE(full_sysidle_state);
2544	int newoldstate;
 
 
2545
 
2546	/*
2547	 * Each pass through the following loop attempts to exit full
2548	 * system-idle state.  If contention proves to be a problem,
2549	 * a trylock-based contention tree could be used here.
2550	 */
2551	while (oldstate > RCU_SYSIDLE_SHORT) {
2552		newoldstate = cmpxchg(&full_sysidle_state,
2553				      oldstate, RCU_SYSIDLE_NOT);
2554		if (oldstate == newoldstate &&
2555		    oldstate == RCU_SYSIDLE_FULL_NOTED) {
2556			rcu_kick_nohz_cpu(tick_do_timer_cpu);
2557			return; /* We cleared it, done! */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2558		}
2559		oldstate = newoldstate;
2560	}
2561	smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
 
 
 
2562}
 
2563
2564/*
2565 * Invoked to note entry to irq or task transition from idle.  Note that
2566 * usermode execution does -not- count as idle here!  The caller must
2567 * have disabled interrupts.
2568 */
2569static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2570{
2571	/* Adjust nesting, check for already non-idle. */
2572	if (irq) {
2573		rdtp->dynticks_idle_nesting++;
2574		WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2575		if (rdtp->dynticks_idle_nesting != 1)
2576			return; /* Already non-idle. */
2577	} else {
2578		/*
2579		 * Allow for irq misnesting.  Yes, it really is possible
2580		 * to enter an irq handler then never leave it, and maybe
2581		 * also vice versa.  Handle both possibilities.
2582		 */
2583		if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2584			rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2585			WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2586			return; /* Already non-idle. */
2587		} else {
2588			rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2589		}
2590	}
 
 
2591
2592	/* Record end of idle period. */
2593	smp_mb__before_atomic_inc();
2594	atomic_inc(&rdtp->dynticks_idle);
2595	smp_mb__after_atomic_inc();
2596	WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2597
2598	/*
2599	 * If we are the timekeeping CPU, we are permitted to be non-idle
2600	 * during a system-idle state.  This must be the case, because
2601	 * the timekeeping CPU has to take scheduling-clock interrupts
2602	 * during the time that the system is transitioning to full
2603	 * system-idle state.  This means that the timekeeping CPU must
2604	 * invoke rcu_sysidle_force_exit() directly if it does anything
2605	 * more than take a scheduling-clock interrupt.
2606	 */
2607	if (smp_processor_id() == tick_do_timer_cpu)
2608		return;
 
 
 
 
 
 
 
 
 
 
 
 
2609
2610	/* Update system-idle state: We are clearly no longer fully idle! */
2611	rcu_sysidle_force_exit();
 
 
 
 
 
 
 
 
 
2612}
2613
2614/*
2615 * Check to see if the current CPU is idle.  Note that usermode execution
2616 * does not count as idle.  The caller must have disabled interrupts.
 
2617 */
2618static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2619				  unsigned long *maxj)
2620{
2621	int cur;
2622	unsigned long j;
2623	struct rcu_dynticks *rdtp = rdp->dynticks;
2624
2625	/*
2626	 * If some other CPU has already reported non-idle, if this is
2627	 * not the flavor of RCU that tracks sysidle state, or if this
2628	 * is an offline or the timekeeping CPU, nothing to do.
2629	 */
2630	if (!*isidle || rdp->rsp != rcu_sysidle_state ||
2631	    cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2632		return;
2633	if (rcu_gp_in_progress(rdp->rsp))
2634		WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2635
2636	/* Pick up current idle and NMI-nesting counter and check. */
2637	cur = atomic_read(&rdtp->dynticks_idle);
2638	if (cur & 0x1) {
2639		*isidle = false; /* We are not idle! */
2640		return;
 
 
 
2641	}
2642	smp_mb(); /* Read counters before timestamps. */
2643
2644	/* Pick up timestamps. */
2645	j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2646	/* If this CPU entered idle more recently, update maxj timestamp. */
2647	if (ULONG_CMP_LT(*maxj, j))
2648		*maxj = j;
 
 
2649}
2650
2651/*
2652 * Is this the flavor of RCU that is handling full-system idle?
 
2653 */
2654static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2655{
2656	return rsp == rcu_sysidle_state;
 
2657}
2658
2659/*
2660 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2661 * timekeeping CPU.
 
 
2662 */
2663static void rcu_bind_gp_kthread(void)
2664{
2665	int cpu = ACCESS_ONCE(tick_do_timer_cpu);
2666
2667	if (cpu < 0 || cpu >= nr_cpu_ids)
2668		return;
2669	if (raw_smp_processor_id() != cpu)
2670		set_cpus_allowed_ptr(current, cpumask_of(cpu));
2671}
2672
2673/*
2674 * Return a delay in jiffies based on the number of CPUs, rcu_node
2675 * leaf fanout, and jiffies tick rate.  The idea is to allow larger
2676 * systems more time to transition to full-idle state in order to
2677 * avoid the cache thrashing that otherwise occur on the state variable.
2678 * Really small systems (less than a couple of tens of CPUs) should
2679 * instead use a single global atomically incremented counter, and later
2680 * versions of this will automatically reconfigure themselves accordingly.
2681 */
2682static unsigned long rcu_sysidle_delay(void)
2683{
2684	if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2685		return 0;
2686	return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2687}
2688
2689/*
2690 * Advance the full-system-idle state.  This is invoked when all of
2691 * the non-timekeeping CPUs are idle.
2692 */
2693static void rcu_sysidle(unsigned long j)
2694{
2695	/* Check the current state. */
2696	switch (ACCESS_ONCE(full_sysidle_state)) {
2697	case RCU_SYSIDLE_NOT:
2698
2699		/* First time all are idle, so note a short idle period. */
2700		ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2701		break;
2702
2703	case RCU_SYSIDLE_SHORT:
2704
2705		/*
2706		 * Idle for a bit, time to advance to next state?
2707		 * cmpxchg failure means race with non-idle, let them win.
2708		 */
2709		if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2710			(void)cmpxchg(&full_sysidle_state,
2711				      RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2712		break;
2713
2714	case RCU_SYSIDLE_LONG:
2715
2716		/*
2717		 * Do an additional check pass before advancing to full.
2718		 * cmpxchg failure means race with non-idle, let them win.
2719		 */
2720		if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2721			(void)cmpxchg(&full_sysidle_state,
2722				      RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2723		break;
2724
2725	default:
2726		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2727	}
 
 
2728}
2729
2730/*
2731 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2732 * back to the beginning.
2733 */
2734static void rcu_sysidle_cancel(void)
2735{
2736	smp_mb();
2737	ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2738}
 
2739
2740/*
2741 * Update the sysidle state based on the results of a force-quiescent-state
2742 * scan of the CPUs' dyntick-idle state.
2743 */
2744static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2745			       unsigned long maxj, bool gpkt)
2746{
2747	if (rsp != rcu_sysidle_state)
2748		return;  /* Wrong flavor, ignore. */
2749	if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2750		return;  /* Running state machine from timekeeping CPU. */
2751	if (isidle)
2752		rcu_sysidle(maxj);    /* More idle! */
2753	else
2754		rcu_sysidle_cancel(); /* Idle is over. */
2755}
 
 
 
 
 
 
2756
2757/*
2758 * Wrapper for rcu_sysidle_report() when called from the grace-period
2759 * kthread's context.
2760 */
2761static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2762				  unsigned long maxj)
2763{
2764	rcu_sysidle_report(rsp, isidle, maxj, true);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2765}
2766
2767/* Callback and function for forcing an RCU grace period. */
2768struct rcu_sysidle_head {
2769	struct rcu_head rh;
2770	int inuse;
2771};
2772
2773static void rcu_sysidle_cb(struct rcu_head *rhp)
 
2774{
2775	struct rcu_sysidle_head *rshp;
2776
2777	/*
2778	 * The following memory barrier is needed to replace the
2779	 * memory barriers that would normally be in the memory
2780	 * allocator.
2781	 */
2782	smp_mb();  /* grace period precedes setting inuse. */
2783
2784	rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2785	ACCESS_ONCE(rshp->inuse) = 0;
2786}
2787
2788/*
2789 * Check to see if the system is fully idle, other than the timekeeping CPU.
2790 * The caller must have disabled interrupts.
2791 */
2792bool rcu_sys_is_idle(void)
2793{
2794	static struct rcu_sysidle_head rsh;
2795	int rss = ACCESS_ONCE(full_sysidle_state);
2796
2797	if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2798		return false;
2799
2800	/* Handle small-system case by doing a full scan of CPUs. */
2801	if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2802		int oldrss = rss - 1;
2803
2804		/*
2805		 * One pass to advance to each state up to _FULL.
2806		 * Give up if any pass fails to advance the state.
2807		 */
2808		while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2809			int cpu;
2810			bool isidle = true;
2811			unsigned long maxj = jiffies - ULONG_MAX / 4;
2812			struct rcu_data *rdp;
2813
2814			/* Scan all the CPUs looking for nonidle CPUs. */
2815			for_each_possible_cpu(cpu) {
2816				rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu);
2817				rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2818				if (!isidle)
2819					break;
2820			}
2821			rcu_sysidle_report(rcu_sysidle_state,
2822					   isidle, maxj, false);
2823			oldrss = rss;
2824			rss = ACCESS_ONCE(full_sysidle_state);
2825		}
2826	}
2827
2828	/* If this is the first observation of an idle period, record it. */
2829	if (rss == RCU_SYSIDLE_FULL) {
2830		rss = cmpxchg(&full_sysidle_state,
2831			      RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2832		return rss == RCU_SYSIDLE_FULL;
2833	}
2834
2835	smp_mb(); /* ensure rss load happens before later caller actions. */
 
 
2836
2837	/* If already fully idle, tell the caller (in case of races). */
2838	if (rss == RCU_SYSIDLE_FULL_NOTED)
2839		return true;
 
2840
2841	/*
2842	 * If we aren't there yet, and a grace period is not in flight,
2843	 * initiate a grace period.  Either way, tell the caller that
2844	 * we are not there yet.  We use an xchg() rather than an assignment
2845	 * to make up for the memory barriers that would otherwise be
2846	 * provided by the memory allocator.
2847	 */
2848	if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2849	    !rcu_gp_in_progress(rcu_sysidle_state) &&
2850	    !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2851		call_rcu(&rsh.rh, rcu_sysidle_cb);
2852	return false;
2853}
2854
2855/*
2856 * Initialize dynticks sysidle state for CPUs coming online.
2857 */
2858static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2859{
2860	rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2861}
2862
2863#else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
 
 
 
 
2864
2865static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
 
2866{
 
2867}
2868
2869static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2870{
2871}
2872
2873static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2874				  unsigned long *maxj)
2875{
 
2876}
2877
2878static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2879{
2880	return false;
2881}
2882
2883static void rcu_bind_gp_kthread(void)
2884{
2885}
2886
2887static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2888				  unsigned long maxj)
2889{
2890}
2891
2892static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2893{
2894}
2895
2896#endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2897
2898/*
2899 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2900 * grace-period kthread will do force_quiescent_state() processing?
2901 * The idea is to avoid waking up RCU core processing on such a
2902 * CPU unless the grace period has extended for too long.
2903 *
2904 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2905 * CONFIG_RCU_NOCB_CPU CPUs.
2906 */
2907static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2908{
2909#ifdef CONFIG_NO_HZ_FULL
2910	if (tick_nohz_full_cpu(smp_processor_id()) &&
2911	    (!rcu_gp_in_progress(rsp) ||
2912	     ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
2913		return 1;
2914#endif /* #ifdef CONFIG_NO_HZ_FULL */
2915	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2916}
v5.14.15
   1/* SPDX-License-Identifier: GPL-2.0+ */
   2/*
   3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
   4 * Internal non-public definitions that provide either classic
   5 * or preemptible semantics.
   6 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   7 * Copyright Red Hat, 2009
   8 * Copyright IBM Corporation, 2009
   9 *
  10 * Author: Ingo Molnar <mingo@elte.hu>
  11 *	   Paul E. McKenney <paulmck@linux.ibm.com>
  12 */
  13
  14#include "../locking/rtmutex_common.h"
 
 
 
 
 
 
 
 
 
 
 
 
  15
  16#ifdef CONFIG_RCU_NOCB_CPU
  17static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
 
  18static bool __read_mostly rcu_nocb_poll;    /* Offload kthread are to poll. */
  19static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
  20{
  21	return lockdep_is_held(&rdp->nocb_lock);
  22}
  23
  24static inline bool rcu_current_is_nocb_kthread(struct rcu_data *rdp)
  25{
  26	/* Race on early boot between thread creation and assignment */
  27	if (!rdp->nocb_cb_kthread || !rdp->nocb_gp_kthread)
  28		return true;
  29
  30	if (current == rdp->nocb_cb_kthread || current == rdp->nocb_gp_kthread)
  31		if (in_task())
  32			return true;
  33	return false;
  34}
  35
  36#else
  37static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
  38{
  39	return 0;
  40}
  41
  42static inline bool rcu_current_is_nocb_kthread(struct rcu_data *rdp)
  43{
  44	return false;
  45}
  46
  47#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
  48
  49static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
  50{
  51	/*
  52	 * In order to read the offloaded state of an rdp is a safe
  53	 * and stable way and prevent from its value to be changed
  54	 * under us, we must either hold the barrier mutex, the cpu
  55	 * hotplug lock (read or write) or the nocb lock. Local
  56	 * non-preemptible reads are also safe. NOCB kthreads and
  57	 * timers have their own means of synchronization against the
  58	 * offloaded state updaters.
  59	 */
  60	RCU_LOCKDEP_WARN(
  61		!(lockdep_is_held(&rcu_state.barrier_mutex) ||
  62		  (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
  63		  rcu_lockdep_is_held_nocb(rdp) ||
  64		  (rdp == this_cpu_ptr(&rcu_data) &&
  65		   !(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible())) ||
  66		  rcu_current_is_nocb_kthread(rdp)),
  67		"Unsafe read of RCU_NOCB offloaded state"
  68	);
  69
  70	return rcu_segcblist_is_offloaded(&rdp->cblist);
  71}
  72
  73/*
  74 * Check the RCU kernel configuration parameters and print informative
  75 * messages about anything out of the ordinary.
 
  76 */
  77static void __init rcu_bootup_announce_oddness(void)
  78{
  79	if (IS_ENABLED(CONFIG_RCU_TRACE))
  80		pr_info("\tRCU event tracing is enabled.\n");
  81	if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
  82	    (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
  83		pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
  84			RCU_FANOUT);
  85	if (rcu_fanout_exact)
  86		pr_info("\tHierarchical RCU autobalancing is disabled.\n");
  87	if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
  88		pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
  89	if (IS_ENABLED(CONFIG_PROVE_RCU))
  90		pr_info("\tRCU lockdep checking is enabled.\n");
  91	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
  92		pr_info("\tRCU strict (and thus non-scalable) grace periods enabled.\n");
  93	if (RCU_NUM_LVLS >= 4)
  94		pr_info("\tFour(or more)-level hierarchy is enabled.\n");
  95	if (RCU_FANOUT_LEAF != 16)
  96		pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
  97			RCU_FANOUT_LEAF);
  98	if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
  99		pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
 100			rcu_fanout_leaf);
 
 
 
 
 
 
 
 
 101	if (nr_cpu_ids != NR_CPUS)
 102		pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
 103#ifdef CONFIG_RCU_BOOST
 104	pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
 105		kthread_prio, CONFIG_RCU_BOOST_DELAY);
 106#endif
 107	if (blimit != DEFAULT_RCU_BLIMIT)
 108		pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
 109	if (qhimark != DEFAULT_RCU_QHIMARK)
 110		pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
 111	if (qlowmark != DEFAULT_RCU_QLOMARK)
 112		pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
 113	if (qovld != DEFAULT_RCU_QOVLD)
 114		pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
 115	if (jiffies_till_first_fqs != ULONG_MAX)
 116		pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
 117	if (jiffies_till_next_fqs != ULONG_MAX)
 118		pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
 119	if (jiffies_till_sched_qs != ULONG_MAX)
 120		pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
 121	if (rcu_kick_kthreads)
 122		pr_info("\tKick kthreads if too-long grace period.\n");
 123	if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
 124		pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
 125	if (gp_preinit_delay)
 126		pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
 127	if (gp_init_delay)
 128		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
 129	if (gp_cleanup_delay)
 130		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
 131	if (!use_softirq)
 132		pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
 133	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
 134		pr_info("\tRCU debug extended QS entry/exit.\n");
 135	rcupdate_announce_bootup_oddness();
 136}
 137
 138#ifdef CONFIG_PREEMPT_RCU
 
 
 
 139
 140static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
 141static void rcu_read_unlock_special(struct task_struct *t);
 142
 143/*
 144 * Tell them what RCU they are running.
 145 */
 146static void __init rcu_bootup_announce(void)
 147{
 148	pr_info("Preemptible hierarchical RCU implementation.\n");
 149	rcu_bootup_announce_oddness();
 150}
 151
 152/* Flags for rcu_preempt_ctxt_queue() decision table. */
 153#define RCU_GP_TASKS	0x8
 154#define RCU_EXP_TASKS	0x4
 155#define RCU_GP_BLKD	0x2
 156#define RCU_EXP_BLKD	0x1
 157
 158/*
 159 * Queues a task preempted within an RCU-preempt read-side critical
 160 * section into the appropriate location within the ->blkd_tasks list,
 161 * depending on the states of any ongoing normal and expedited grace
 162 * periods.  The ->gp_tasks pointer indicates which element the normal
 163 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
 164 * indicates which element the expedited grace period is waiting on (again,
 165 * NULL if none).  If a grace period is waiting on a given element in the
 166 * ->blkd_tasks list, it also waits on all subsequent elements.  Thus,
 167 * adding a task to the tail of the list blocks any grace period that is
 168 * already waiting on one of the elements.  In contrast, adding a task
 169 * to the head of the list won't block any grace period that is already
 170 * waiting on one of the elements.
 171 *
 172 * This queuing is imprecise, and can sometimes make an ongoing grace
 173 * period wait for a task that is not strictly speaking blocking it.
 174 * Given the choice, we needlessly block a normal grace period rather than
 175 * blocking an expedited grace period.
 176 *
 177 * Note that an endless sequence of expedited grace periods still cannot
 178 * indefinitely postpone a normal grace period.  Eventually, all of the
 179 * fixed number of preempted tasks blocking the normal grace period that are
 180 * not also blocking the expedited grace period will resume and complete
 181 * their RCU read-side critical sections.  At that point, the ->gp_tasks
 182 * pointer will equal the ->exp_tasks pointer, at which point the end of
 183 * the corresponding expedited grace period will also be the end of the
 184 * normal grace period.
 185 */
 186static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
 187	__releases(rnp->lock) /* But leaves rrupts disabled. */
 188{
 189	int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
 190			 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
 191			 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
 192			 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
 193	struct task_struct *t = current;
 194
 195	raw_lockdep_assert_held_rcu_node(rnp);
 196	WARN_ON_ONCE(rdp->mynode != rnp);
 197	WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 198	/* RCU better not be waiting on newly onlined CPUs! */
 199	WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
 200		     rdp->grpmask);
 
 
 201
 202	/*
 203	 * Decide where to queue the newly blocked task.  In theory,
 204	 * this could be an if-statement.  In practice, when I tried
 205	 * that, it was quite messy.
 206	 */
 207	switch (blkd_state) {
 208	case 0:
 209	case                RCU_EXP_TASKS:
 210	case                RCU_EXP_TASKS + RCU_GP_BLKD:
 211	case RCU_GP_TASKS:
 212	case RCU_GP_TASKS + RCU_EXP_TASKS:
 213
 214		/*
 215		 * Blocking neither GP, or first task blocking the normal
 216		 * GP but not blocking the already-waiting expedited GP.
 217		 * Queue at the head of the list to avoid unnecessarily
 218		 * blocking the already-waiting GPs.
 219		 */
 220		list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
 221		break;
 222
 223	case                                              RCU_EXP_BLKD:
 224	case                                RCU_GP_BLKD:
 225	case                                RCU_GP_BLKD + RCU_EXP_BLKD:
 226	case RCU_GP_TASKS +                               RCU_EXP_BLKD:
 227	case RCU_GP_TASKS +                 RCU_GP_BLKD + RCU_EXP_BLKD:
 228	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 229
 230		/*
 231		 * First task arriving that blocks either GP, or first task
 232		 * arriving that blocks the expedited GP (with the normal
 233		 * GP already waiting), or a task arriving that blocks
 234		 * both GPs with both GPs already waiting.  Queue at the
 235		 * tail of the list to avoid any GP waiting on any of the
 236		 * already queued tasks that are not blocking it.
 237		 */
 238		list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
 239		break;
 240
 241	case                RCU_EXP_TASKS +               RCU_EXP_BLKD:
 242	case                RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 243	case RCU_GP_TASKS + RCU_EXP_TASKS +               RCU_EXP_BLKD:
 244
 245		/*
 246		 * Second or subsequent task blocking the expedited GP.
 247		 * The task either does not block the normal GP, or is the
 248		 * first task blocking the normal GP.  Queue just after
 249		 * the first task blocking the expedited GP.
 250		 */
 251		list_add(&t->rcu_node_entry, rnp->exp_tasks);
 252		break;
 253
 254	case RCU_GP_TASKS +                 RCU_GP_BLKD:
 255	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
 256
 257		/*
 258		 * Second or subsequent task blocking the normal GP.
 259		 * The task does not block the expedited GP. Queue just
 260		 * after the first task blocking the normal GP.
 261		 */
 262		list_add(&t->rcu_node_entry, rnp->gp_tasks);
 263		break;
 264
 265	default:
 266
 267		/* Yet another exercise in excessive paranoia. */
 268		WARN_ON_ONCE(1);
 269		break;
 270	}
 271
 272	/*
 273	 * We have now queued the task.  If it was the first one to
 274	 * block either grace period, update the ->gp_tasks and/or
 275	 * ->exp_tasks pointers, respectively, to reference the newly
 276	 * blocked tasks.
 277	 */
 278	if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
 279		WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
 280		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
 281	}
 282	if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
 283		WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
 284	WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
 285		     !(rnp->qsmask & rdp->grpmask));
 286	WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
 287		     !(rnp->expmask & rdp->grpmask));
 288	raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
 289
 290	/*
 291	 * Report the quiescent state for the expedited GP.  This expedited
 292	 * GP should not be able to end until we report, so there should be
 293	 * no need to check for a subsequent expedited GP.  (Though we are
 294	 * still in a quiescent state in any case.)
 295	 */
 296	if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
 297		rcu_report_exp_rdp(rdp);
 298	else
 299		WARN_ON_ONCE(rdp->exp_deferred_qs);
 300}
 
 301
 302/*
 303 * Record a preemptible-RCU quiescent state for the specified CPU.
 304 * Note that this does not necessarily mean that the task currently running
 305 * on the CPU is in a quiescent state:  Instead, it means that the current
 306 * grace period need not wait on any RCU read-side critical section that
 307 * starts later on this CPU.  It also means that if the current task is
 308 * in an RCU read-side critical section, it has already added itself to
 309 * some leaf rcu_node structure's ->blkd_tasks list.  In addition to the
 310 * current task, there might be any number of other tasks blocked while
 311 * in an RCU read-side critical section.
 312 *
 313 * Callers to this function must disable preemption.
 314 */
 315static void rcu_qs(void)
 316{
 317	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
 318	if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
 319		trace_rcu_grace_period(TPS("rcu_preempt"),
 320				       __this_cpu_read(rcu_data.gp_seq),
 321				       TPS("cpuqs"));
 322		__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 323		barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
 324		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
 325	}
 326}
 327
 328/*
 329 * We have entered the scheduler, and the current task might soon be
 330 * context-switched away from.  If this task is in an RCU read-side
 331 * critical section, we will no longer be able to rely on the CPU to
 332 * record that fact, so we enqueue the task on the blkd_tasks list.
 333 * The task will dequeue itself when it exits the outermost enclosing
 334 * RCU read-side critical section.  Therefore, the current grace period
 335 * cannot be permitted to complete until the blkd_tasks list entries
 336 * predating the current grace period drain, in other words, until
 337 * rnp->gp_tasks becomes NULL.
 338 *
 339 * Caller must disable interrupts.
 340 */
 341void rcu_note_context_switch(bool preempt)
 342{
 343	struct task_struct *t = current;
 344	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
 345	struct rcu_node *rnp;
 346
 347	trace_rcu_utilization(TPS("Start context switch"));
 348	lockdep_assert_irqs_disabled();
 349	WARN_ON_ONCE(!preempt && rcu_preempt_depth() > 0);
 350	if (rcu_preempt_depth() > 0 &&
 351	    !t->rcu_read_unlock_special.b.blocked) {
 352
 353		/* Possibly blocking in an RCU read-side critical section. */
 
 354		rnp = rdp->mynode;
 355		raw_spin_lock_rcu_node(rnp);
 356		t->rcu_read_unlock_special.b.blocked = true;
 
 357		t->rcu_blocked_node = rnp;
 358
 359		/*
 360		 * Verify the CPU's sanity, trace the preemption, and
 361		 * then queue the task as required based on the states
 362		 * of any ongoing and expedited grace periods.
 
 
 
 
 
 
 
 
 
 
 
 
 
 363		 */
 364		WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
 365		WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
 366		trace_rcu_preempt_task(rcu_state.name,
 
 
 
 
 
 
 
 
 
 
 
 
 367				       t->pid,
 368				       (rnp->qsmask & rdp->grpmask)
 369				       ? rnp->gp_seq
 370				       : rcu_seq_snap(&rnp->gp_seq));
 371		rcu_preempt_ctxt_queue(rnp, rdp);
 372	} else {
 373		rcu_preempt_deferred_qs(t);
 
 
 
 
 
 
 374	}
 375
 376	/*
 377	 * Either we were not in an RCU read-side critical section to
 378	 * begin with, or we have now recorded that critical section
 379	 * globally.  Either way, we can now note a quiescent state
 380	 * for this CPU.  Again, if we were in an RCU read-side critical
 381	 * section, and if that critical section was blocking the current
 382	 * grace period, then the fact that the task has been enqueued
 383	 * means that we continue to block the current grace period.
 384	 */
 385	rcu_qs();
 386	if (rdp->exp_deferred_qs)
 387		rcu_report_exp_rdp(rdp);
 388	rcu_tasks_qs(current, preempt);
 389	trace_rcu_utilization(TPS("End context switch"));
 390}
 391EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 392
 393/*
 394 * Check for preempted RCU readers blocking the current grace period
 395 * for the specified rcu_node structure.  If the caller needs a reliable
 396 * answer, it must hold the rcu_node's ->lock.
 397 */
 398static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 399{
 400	return READ_ONCE(rnp->gp_tasks) != NULL;
 401}
 402
 403/* limit value for ->rcu_read_lock_nesting. */
 404#define RCU_NEST_PMAX (INT_MAX / 2)
 405
 406static void rcu_preempt_read_enter(void)
 407{
 408	current->rcu_read_lock_nesting++;
 409}
 410
 411static int rcu_preempt_read_exit(void)
 412{
 413	return --current->rcu_read_lock_nesting;
 414}
 415
 416static void rcu_preempt_depth_set(int val)
 417{
 418	current->rcu_read_lock_nesting = val;
 419}
 420
 421/*
 422 * Preemptible RCU implementation for rcu_read_lock().
 423 * Just increment ->rcu_read_lock_nesting, shared state will be updated
 424 * if we block.
 
 
 425 */
 426void __rcu_read_lock(void)
 
 427{
 428	rcu_preempt_read_enter();
 429	if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 430		WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
 431	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
 432		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
 433	barrier();  /* critical section after entry code. */
 434}
 435EXPORT_SYMBOL_GPL(__rcu_read_lock);
 436
 437/*
 438 * Preemptible RCU implementation for rcu_read_unlock().
 439 * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
 440 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
 441 * invoke rcu_read_unlock_special() to clean up after a context switch
 442 * in an RCU read-side critical section and other special cases.
 443 */
 444void __rcu_read_unlock(void)
 445{
 446	struct task_struct *t = current;
 447
 448	barrier();  // critical section before exit code.
 449	if (rcu_preempt_read_exit() == 0) {
 450		barrier();  // critical-section exit before .s check.
 451		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
 452			rcu_read_unlock_special(t);
 
 
 
 
 453	}
 454	if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 455		int rrln = rcu_preempt_depth();
 456
 457		WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
 458	}
 
 
 
 
 459}
 460EXPORT_SYMBOL_GPL(__rcu_read_unlock);
 461
 462/*
 463 * Advance a ->blkd_tasks-list pointer to the next entry, instead
 464 * returning NULL if at the end of the list.
 465 */
 466static struct list_head *rcu_next_node_entry(struct task_struct *t,
 467					     struct rcu_node *rnp)
 468{
 469	struct list_head *np;
 470
 471	np = t->rcu_node_entry.next;
 472	if (np == &rnp->blkd_tasks)
 473		np = NULL;
 474	return np;
 475}
 476
 477/*
 478 * Return true if the specified rcu_node structure has tasks that were
 479 * preempted within an RCU read-side critical section.
 
 480 */
 481static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 482{
 483	return !list_empty(&rnp->blkd_tasks);
 484}
 485
 486/*
 487 * Report deferred quiescent states.  The deferral time can
 488 * be quite short, for example, in the case of the call from
 489 * rcu_read_unlock_special().
 490 */
 491static void
 492rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
 493{
 494	bool empty_exp;
 495	bool empty_norm;
 496	bool empty_exp_now;
 497	struct list_head *np;
 498	bool drop_boost_mutex = false;
 499	struct rcu_data *rdp;
 
 500	struct rcu_node *rnp;
 501	union rcu_special special;
 
 
 
 
 
 
 502
 503	/*
 504	 * If RCU core is waiting for this CPU to exit its critical section,
 505	 * report the fact that it has exited.  Because irqs are disabled,
 506	 * t->rcu_read_unlock_special cannot change.
 507	 */
 508	special = t->rcu_read_unlock_special;
 509	rdp = this_cpu_ptr(&rcu_data);
 510	if (!special.s && !rdp->exp_deferred_qs) {
 
 
 
 
 
 
 
 
 511		local_irq_restore(flags);
 512		return;
 513	}
 514	t->rcu_read_unlock_special.s = 0;
 515	if (special.b.need_qs) {
 516		if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
 517			rcu_report_qs_rdp(rdp);
 518			udelay(rcu_unlock_delay);
 519		} else {
 520			rcu_qs();
 521		}
 522	}
 523
 524	/*
 525	 * Respond to a request by an expedited grace period for a
 526	 * quiescent state from this CPU.  Note that requests from
 527	 * tasks are handled when removing the task from the
 528	 * blocked-tasks list below.
 529	 */
 530	if (rdp->exp_deferred_qs)
 531		rcu_report_exp_rdp(rdp);
 532
 533	/* Clean up if blocked during RCU read-side critical section. */
 534	if (special.b.blocked) {
 
 535
 536		/*
 537		 * Remove this task from the list it blocked on.  The task
 538		 * now remains queued on the rcu_node corresponding to the
 539		 * CPU it first blocked on, so there is no longer any need
 540		 * to loop.  Retain a WARN_ON_ONCE() out of sheer paranoia.
 541		 */
 542		rnp = t->rcu_blocked_node;
 543		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
 544		WARN_ON_ONCE(rnp != t->rcu_blocked_node);
 545		WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 546		empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
 547		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
 548			     (!empty_norm || rnp->qsmask));
 549		empty_exp = sync_rcu_exp_done(rnp);
 
 
 550		smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
 551		np = rcu_next_node_entry(t, rnp);
 552		list_del_init(&t->rcu_node_entry);
 553		t->rcu_blocked_node = NULL;
 554		trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
 555						rnp->gp_seq, t->pid);
 556		if (&t->rcu_node_entry == rnp->gp_tasks)
 557			WRITE_ONCE(rnp->gp_tasks, np);
 558		if (&t->rcu_node_entry == rnp->exp_tasks)
 559			WRITE_ONCE(rnp->exp_tasks, np);
 560		if (IS_ENABLED(CONFIG_RCU_BOOST)) {
 561			/* Snapshot ->boost_mtx ownership w/rnp->lock held. */
 562			drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
 563			if (&t->rcu_node_entry == rnp->boost_tasks)
 564				WRITE_ONCE(rnp->boost_tasks, np);
 
 
 565		}
 
 566
 567		/*
 568		 * If this was the last task on the current list, and if
 569		 * we aren't waiting on any CPUs, report the quiescent state.
 570		 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
 571		 * so we must take a snapshot of the expedited state.
 572		 */
 573		empty_exp_now = sync_rcu_exp_done(rnp);
 574		if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
 575			trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
 576							 rnp->gp_seq,
 577							 0, rnp->qsmask,
 578							 rnp->level,
 579							 rnp->grplo,
 580							 rnp->grphi,
 581							 !!rnp->gp_tasks);
 582			rcu_report_unblock_qs_rnp(rnp, flags);
 583		} else {
 584			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 585		}
 586
 
 587		/* Unboost if we were boosted. */
 588		if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
 589			rt_mutex_futex_unlock(&rnp->boost_mtx);
 
 590
 591		/*
 592		 * If this was the last task on the expedited lists,
 593		 * then we need to report up the rcu_node hierarchy.
 594		 */
 595		if (!empty_exp && empty_exp_now)
 596			rcu_report_exp_rnp(rnp, true);
 597	} else {
 598		local_irq_restore(flags);
 599	}
 600}
 601
 
 
 602/*
 603 * Is a deferred quiescent-state pending, and are we also not in
 604 * an RCU read-side critical section?  It is the caller's responsibility
 605 * to ensure it is otherwise safe to report any deferred quiescent
 606 * states.  The reason for this is that it is safe to report a
 607 * quiescent state during context switch even though preemption
 608 * is disabled.  This function cannot be expected to understand these
 609 * nuances, so the caller must handle them.
 610 */
 611static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 612{
 613	return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
 614		READ_ONCE(t->rcu_read_unlock_special.s)) &&
 615	       rcu_preempt_depth() == 0;
 
 
 
 
 
 
 
 
 
 
 616}
 617
 618/*
 619 * Report a deferred quiescent state if needed and safe to do so.
 620 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
 621 * not being in an RCU read-side critical section.  The caller must
 622 * evaluate safety in terms of interrupt, softirq, and preemption
 623 * disabling.
 624 */
 625static void rcu_preempt_deferred_qs(struct task_struct *t)
 
 
 
 
 
 
 
 
 
 
 
 626{
 627	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 628
 629	if (!rcu_preempt_need_deferred_qs(t))
 630		return;
 631	local_irq_save(flags);
 632	rcu_preempt_deferred_qs_irqrestore(t, flags);
 633}
 634
 635/*
 636 * Minimal handler to give the scheduler a chance to re-evaluate.
 637 */
 638static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
 639{
 640	struct rcu_data *rdp;
 641
 642	rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
 643	rdp->defer_qs_iw_pending = false;
 644}
 645
 
 
 646/*
 647 * Handle special cases during rcu_read_unlock(), such as needing to
 648 * notify RCU core processing or task having blocked during the RCU
 649 * read-side critical section.
 650 */
 651static void rcu_read_unlock_special(struct task_struct *t)
 652{
 653	unsigned long flags;
 654	bool irqs_were_disabled;
 655	bool preempt_bh_were_disabled =
 656			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
 657
 658	/* NMI handlers cannot block and cannot safely manipulate state. */
 659	if (in_nmi())
 660		return;
 661
 662	local_irq_save(flags);
 663	irqs_were_disabled = irqs_disabled_flags(flags);
 664	if (preempt_bh_were_disabled || irqs_were_disabled) {
 665		bool expboost; // Expedited GP in flight or possible boosting.
 666		struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 667		struct rcu_node *rnp = rdp->mynode;
 668
 669		expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
 670			   (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
 671			   IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ||
 672			   (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
 673			    t->rcu_blocked_node);
 674		// Need to defer quiescent state until everything is enabled.
 675		if (use_softirq && (in_irq() || (expboost && !irqs_were_disabled))) {
 676			// Using softirq, safe to awaken, and either the
 677			// wakeup is free or there is either an expedited
 678			// GP in flight or a potential need to deboost.
 679			raise_softirq_irqoff(RCU_SOFTIRQ);
 680		} else {
 681			// Enabling BH or preempt does reschedule, so...
 682			// Also if no expediting and no possible deboosting,
 683			// slow is OK.  Plus nohz_full CPUs eventually get
 684			// tick enabled.
 685			set_tsk_need_resched(current);
 686			set_preempt_need_resched();
 687			if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
 688			    expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
 689				// Get scheduler to re-evaluate and call hooks.
 690				// If !IRQ_WORK, FQS scan will eventually IPI.
 691				init_irq_work(&rdp->defer_qs_iw, rcu_preempt_deferred_qs_handler);
 692				rdp->defer_qs_iw_pending = true;
 693				irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
 694			}
 695		}
 696		local_irq_restore(flags);
 697		return;
 698	}
 699	rcu_preempt_deferred_qs_irqrestore(t, flags);
 
 700}
 701
 702/*
 703 * Check that the list of blocked tasks for the newly completed grace
 704 * period is in fact empty.  It is a serious bug to complete a grace
 705 * period that still has RCU readers blocked!  This function must be
 706 * invoked -before- updating this rnp's ->gp_seq.
 
 707 *
 708 * Also, if there are blocked tasks on the list, they automatically
 709 * block the newly created grace period, so set up ->gp_tasks accordingly.
 710 */
 711static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 712{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 713	struct task_struct *t;
 714
 715	RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
 716	raw_lockdep_assert_held_rcu_node(rnp);
 717	if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
 718		dump_blkd_tasks(rnp, 10);
 719	if (rcu_preempt_has_tasks(rnp) &&
 720	    (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
 721		WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
 722		t = container_of(rnp->gp_tasks, struct task_struct,
 723				 rcu_node_entry);
 724		trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
 725						rnp->gp_seq, t->pid);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 726	}
 727	WARN_ON_ONCE(rnp->qsmask);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 728}
 729
 
 
 730/*
 731 * Check for a quiescent state from the current CPU, including voluntary
 732 * context switches for Tasks RCU.  When a task blocks, the task is
 733 * recorded in the corresponding CPU's rcu_node structure, which is checked
 734 * elsewhere, hence this function need only check for quiescent states
 735 * related to the current CPU, not to those related to tasks.
 736 */
 737static void rcu_flavor_sched_clock_irq(int user)
 738{
 739	struct task_struct *t = current;
 740
 741	lockdep_assert_irqs_disabled();
 742	if (user || rcu_is_cpu_rrupt_from_idle()) {
 743		rcu_note_voluntary_context_switch(current);
 744	}
 745	if (rcu_preempt_depth() > 0 ||
 746	    (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
 747		/* No QS, force context switch if deferred. */
 748		if (rcu_preempt_need_deferred_qs(t)) {
 749			set_tsk_need_resched(t);
 750			set_preempt_need_resched();
 751		}
 752	} else if (rcu_preempt_need_deferred_qs(t)) {
 753		rcu_preempt_deferred_qs(t); /* Report deferred QS. */
 754		return;
 755	} else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
 756		rcu_qs(); /* Report immediate QS. */
 757		return;
 758	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 759
 760	/* If GP is oldish, ask for help from rcu_read_unlock_special(). */
 761	if (rcu_preempt_depth() > 0 &&
 762	    __this_cpu_read(rcu_data.core_needs_qs) &&
 763	    __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
 764	    !t->rcu_read_unlock_special.b.need_qs &&
 765	    time_after(jiffies, rcu_state.gp_start + HZ))
 766		t->rcu_read_unlock_special.b.need_qs = true;
 767}
 
 768
 769/*
 770 * Check for a task exiting while in a preemptible-RCU read-side
 771 * critical section, clean up if so.  No need to issue warnings, as
 772 * debug_check_no_locks_held() already does this if lockdep is enabled.
 773 * Besides, if this function does anything other than just immediately
 774 * return, there was a bug of some sort.  Spewing warnings from this
 775 * function is like as not to simply obscure important prior warnings.
 776 */
 777void exit_rcu(void)
 
 778{
 779	struct task_struct *t = current;
 
 
 780
 781	if (unlikely(!list_empty(&current->rcu_node_entry))) {
 782		rcu_preempt_depth_set(1);
 783		barrier();
 784		WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
 785	} else if (unlikely(rcu_preempt_depth())) {
 786		rcu_preempt_depth_set(1);
 787	} else {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 788		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 789	}
 790	__rcu_read_unlock();
 791	rcu_preempt_deferred_qs(current);
 792}
 793
 794/*
 795 * Dump the blocked-tasks state, but limit the list dump to the
 796 * specified number of elements.
 
 
 
 
 797 */
 798static void
 799dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 800{
 801	int cpu;
 802	int i;
 803	struct list_head *lhp;
 804	bool onl;
 805	struct rcu_data *rdp;
 806	struct rcu_node *rnp1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 807
 808	raw_lockdep_assert_held_rcu_node(rnp);
 809	pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
 810		__func__, rnp->grplo, rnp->grphi, rnp->level,
 811		(long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
 812	for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
 813		pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
 814			__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
 815	pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
 816		__func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
 817		READ_ONCE(rnp->exp_tasks));
 818	pr_info("%s: ->blkd_tasks", __func__);
 819	i = 0;
 820	list_for_each(lhp, &rnp->blkd_tasks) {
 821		pr_cont(" %p", lhp);
 822		if (++i >= ncheck)
 823			break;
 
 
 
 
 
 
 824	}
 825	pr_cont("\n");
 826	for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
 827		rdp = per_cpu_ptr(&rcu_data, cpu);
 828		onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
 829		pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
 830			cpu, ".o"[onl],
 831			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
 832			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
 
 
 833	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 834}
 
 835
 836#else /* #ifdef CONFIG_PREEMPT_RCU */
 
 
 
 
 
 
 837
 838/*
 839 * If strict grace periods are enabled, and if the calling
 840 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
 841 * report that quiescent state and, if requested, spin for a bit.
 
 842 */
 843void rcu_read_unlock_strict(void)
 844{
 845	struct rcu_data *rdp;
 846
 847	if (!IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ||
 848	   irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
 849		return;
 850	rdp = this_cpu_ptr(&rcu_data);
 851	rcu_report_qs_rdp(rdp);
 852	udelay(rcu_unlock_delay);
 
 853}
 854EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
 
 
 
 855
 856/*
 857 * Tell them what RCU they are running.
 858 */
 859static void __init rcu_bootup_announce(void)
 860{
 861	pr_info("Hierarchical RCU implementation.\n");
 862	rcu_bootup_announce_oddness();
 863}
 864
 865/*
 866 * Note a quiescent state for PREEMPTION=n.  Because we do not need to know
 867 * how many quiescent states passed, just if there was at least one since
 868 * the start of the grace period, this just sets a flag.  The caller must
 869 * have disabled preemption.
 870 */
 871static void rcu_qs(void)
 872{
 873	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
 874	if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
 875		return;
 876	trace_rcu_grace_period(TPS("rcu_sched"),
 877			       __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
 878	__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 879	if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
 880		return;
 881	__this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
 882	rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 883}
 
 884
 885/*
 886 * Register an urgently needed quiescent state.  If there is an
 887 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
 888 * dyntick-idle quiescent state visible to other CPUs, which will in
 889 * some cases serve for expedited as well as normal grace periods.
 890 * Either way, register a lightweight quiescent state.
 891 */
 892void rcu_all_qs(void)
 893{
 894	unsigned long flags;
 895
 896	if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
 897		return;
 898	preempt_disable();
 899	/* Load rcu_urgent_qs before other flags. */
 900	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
 901		preempt_enable();
 902		return;
 903	}
 904	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 905	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
 906		local_irq_save(flags);
 907		rcu_momentary_dyntick_idle();
 908		local_irq_restore(flags);
 909	}
 910	rcu_qs();
 911	preempt_enable();
 912}
 913EXPORT_SYMBOL_GPL(rcu_all_qs);
 914
 915/*
 916 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
 
 917 */
 918void rcu_note_context_switch(bool preempt)
 919{
 920	trace_rcu_utilization(TPS("Start context switch"));
 921	rcu_qs();
 922	/* Load rcu_urgent_qs before other flags. */
 923	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
 924		goto out;
 925	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 926	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
 927		rcu_momentary_dyntick_idle();
 928	rcu_tasks_qs(current, preempt);
 929out:
 930	trace_rcu_utilization(TPS("End context switch"));
 931}
 932EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 933
 934/*
 935 * Because preemptible RCU does not exist, there are never any preempted
 936 * RCU readers.
 937 */
 938static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 939{
 940	return 0;
 941}
 942
 
 
 
 
 
 
 
 
 
 
 943/*
 944 * Because there is no preemptible RCU, there can be no readers blocked.
 
 945 */
 946static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 947{
 948	return false;
 949}
 950
 951/*
 952 * Because there is no preemptible RCU, there can be no deferred quiescent
 953 * states.
 954 */
 955static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 956{
 957	return false;
 958}
 959static void rcu_preempt_deferred_qs(struct task_struct *t) { }
 960
 961/*
 962 * Because there is no preemptible RCU, there can be no readers blocked,
 963 * so there is no need to check for blocked tasks.  So check only for
 964 * bogus qsmask values.
 965 */
 966static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 967{
 968	WARN_ON_ONCE(rnp->qsmask);
 969}
 970
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 971/*
 972 * Check to see if this CPU is in a non-context-switch quiescent state,
 973 * namely user mode and idle loop.
 974 */
 975static void rcu_flavor_sched_clock_irq(int user)
 976{
 977	if (user || rcu_is_cpu_rrupt_from_idle()) {
 978
 979		/*
 980		 * Get here if this CPU took its interrupt from user
 981		 * mode or from the idle loop, and if this is not a
 982		 * nested interrupt.  In this case, the CPU is in
 983		 * a quiescent state, so note it.
 984		 *
 985		 * No memory barrier is required here because rcu_qs()
 986		 * references only CPU-local variables that other CPUs
 987		 * neither access nor modify, at least not while the
 988		 * corresponding CPU is online.
 989		 */
 
 
 
 
 990
 991		rcu_qs();
 992	}
 
 
 
 
 
 993}
 
 
 
 994
 995/*
 996 * Because preemptible RCU does not exist, tasks cannot possibly exit
 997 * while in preemptible RCU read-side critical sections.
 
 998 */
 999void exit_rcu(void)
 
1000{
1001}
1002
 
 
1003/*
1004 * Dump the guaranteed-empty blocked-tasks state.  Trust but verify.
 
1005 */
1006static void
1007dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
1008{
1009	WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
1010}
 
1011
1012#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 
 
 
 
 
1013
1014/*
1015 * If boosting, set rcuc kthreads to realtime priority.
 
1016 */
1017static void rcu_cpu_kthread_setup(unsigned int cpu)
1018{
 
 
 
 
1019#ifdef CONFIG_RCU_BOOST
1020	struct sched_param sp;
1021
1022	sp.sched_priority = kthread_prio;
1023	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1024#endif /* #ifdef CONFIG_RCU_BOOST */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1025}
1026
1027#ifdef CONFIG_RCU_BOOST
 
 
 
 
 
 
 
 
 
 
1028
1029/*
1030 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1031 * or ->boost_tasks, advancing the pointer to the next task in the
1032 * ->blkd_tasks list.
1033 *
1034 * Note that irqs must be enabled: boosting the task can block.
1035 * Returns 1 if there are more tasks needing to be boosted.
1036 */
1037static int rcu_boost(struct rcu_node *rnp)
1038{
1039	unsigned long flags;
 
1040	struct task_struct *t;
1041	struct list_head *tb;
1042
1043	if (READ_ONCE(rnp->exp_tasks) == NULL &&
1044	    READ_ONCE(rnp->boost_tasks) == NULL)
1045		return 0;  /* Nothing left to boost. */
1046
1047	raw_spin_lock_irqsave_rcu_node(rnp, flags);
 
1048
1049	/*
1050	 * Recheck under the lock: all tasks in need of boosting
1051	 * might exit their RCU read-side critical sections on their own.
1052	 */
1053	if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1054		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1055		return 0;
1056	}
1057
1058	/*
1059	 * Preferentially boost tasks blocking expedited grace periods.
1060	 * This cannot starve the normal grace periods because a second
1061	 * expedited grace period must boost all blocked tasks, including
1062	 * those blocking the pre-existing normal grace period.
1063	 */
1064	if (rnp->exp_tasks != NULL)
1065		tb = rnp->exp_tasks;
1066	else
 
1067		tb = rnp->boost_tasks;
 
 
 
1068
1069	/*
1070	 * We boost task t by manufacturing an rt_mutex that appears to
1071	 * be held by task t.  We leave a pointer to that rt_mutex where
1072	 * task t can find it, and task t will release the mutex when it
1073	 * exits its outermost RCU read-side critical section.  Then
1074	 * simply acquiring this artificial rt_mutex will boost task
1075	 * t's priority.  (Thanks to tglx for suggesting this approach!)
1076	 *
1077	 * Note that task t must acquire rnp->lock to remove itself from
1078	 * the ->blkd_tasks list, which it will do from exit() if from
1079	 * nowhere else.  We therefore are guaranteed that task t will
1080	 * stay around at least until we drop rnp->lock.  Note that
1081	 * rnp->lock also resolves races between our priority boosting
1082	 * and task t's exiting its outermost RCU read-side critical
1083	 * section.
1084	 */
1085	t = container_of(tb, struct task_struct, rcu_node_entry);
1086	rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1087	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1088	/* Lock only for side effect: boosts task t's priority. */
1089	rt_mutex_lock(&rnp->boost_mtx);
1090	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1091	rnp->n_boosts++;
1092
1093	return READ_ONCE(rnp->exp_tasks) != NULL ||
1094	       READ_ONCE(rnp->boost_tasks) != NULL;
1095}
1096
1097/*
1098 * Priority-boosting kthread, one per leaf rcu_node.
 
1099 */
1100static int rcu_boost_kthread(void *arg)
1101{
1102	struct rcu_node *rnp = (struct rcu_node *)arg;
1103	int spincnt = 0;
1104	int more2boost;
1105
1106	trace_rcu_utilization(TPS("Start boost kthread@init"));
1107	for (;;) {
1108		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1109		trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1110		rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1111			 READ_ONCE(rnp->exp_tasks));
1112		trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1113		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1114		more2boost = rcu_boost(rnp);
1115		if (more2boost)
1116			spincnt++;
1117		else
1118			spincnt = 0;
1119		if (spincnt > 10) {
1120			WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1121			trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1122			schedule_timeout_idle(2);
1123			trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1124			spincnt = 0;
1125		}
1126	}
1127	/* NOTREACHED */
1128	trace_rcu_utilization(TPS("End boost kthread@notreached"));
1129	return 0;
1130}
1131
1132/*
1133 * Check to see if it is time to start boosting RCU readers that are
1134 * blocking the current grace period, and, if so, tell the per-rcu_node
1135 * kthread to start boosting them.  If there is an expedited grace
1136 * period in progress, it is always time to boost.
1137 *
1138 * The caller must hold rnp->lock, which this function releases.
1139 * The ->boost_kthread_task is immortal, so we don't need to worry
1140 * about it going away.
1141 */
1142static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1143	__releases(rnp->lock)
1144{
1145	raw_lockdep_assert_held_rcu_node(rnp);
 
1146	if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1147		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 
1148		return;
1149	}
1150	if (rnp->exp_tasks != NULL ||
1151	    (rnp->gp_tasks != NULL &&
1152	     rnp->boost_tasks == NULL &&
1153	     rnp->qsmask == 0 &&
1154	     (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld))) {
1155		if (rnp->exp_tasks == NULL)
1156			WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1157		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1158		rcu_wake_cond(rnp->boost_kthread_task,
1159			      READ_ONCE(rnp->boost_kthread_status));
 
1160	} else {
1161		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 
1162	}
1163}
1164
1165/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1166 * Is the current CPU running the RCU-callbacks kthread?
1167 * Caller must have preemption disabled.
1168 */
1169static bool rcu_is_callbacks_kthread(void)
1170{
1171	return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1172}
1173
1174#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1175
1176/*
1177 * Do priority-boost accounting for the start of a new grace period.
1178 */
1179static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1180{
1181	rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1182}
1183
1184/*
1185 * Create an RCU-boost kthread for the specified node if one does not
1186 * already exist.  We only create this kthread for preemptible RCU.
1187 * Returns zero if all is well, a negated errno otherwise.
1188 */
1189static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
 
1190{
 
1191	unsigned long flags;
1192	int rnp_index = rnp - rcu_get_root();
1193	struct sched_param sp;
1194	struct task_struct *t;
1195
1196	if (rnp->boost_kthread_task || !rcu_scheduler_fully_active)
1197		return;
1198
1199	rcu_state.boost = 1;
 
1200
 
 
 
1201	t = kthread_create(rcu_boost_kthread, (void *)rnp,
1202			   "rcub/%d", rnp_index);
1203	if (WARN_ON_ONCE(IS_ERR(t)))
1204		return;
1205
1206	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1207	rnp->boost_kthread_task = t;
1208	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1209	sp.sched_priority = kthread_prio;
1210	sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1211	wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1212}
1213
1214/*
1215 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1216 * served by the rcu_node in question.  The CPU hotplug lock is still
1217 * held, so the value of rnp->qsmaskinit will be stable.
1218 *
1219 * We don't include outgoingcpu in the affinity set, use -1 if there is
1220 * no outgoing CPU.  If there are no CPUs left in the affinity set,
1221 * this function allows the kthread to execute on any CPU.
1222 */
1223static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1224{
1225	struct task_struct *t = rnp->boost_kthread_task;
1226	unsigned long mask = rcu_rnp_online_cpus(rnp);
1227	cpumask_var_t cm;
1228	int cpu;
1229
1230	if (!t)
1231		return;
1232	if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1233		return;
1234	for_each_leaf_node_possible_cpu(rnp, cpu)
1235		if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1236		    cpu != outgoingcpu)
1237			cpumask_set_cpu(cpu, cm);
1238	if (cpumask_weight(cm) == 0)
1239		cpumask_setall(cm);
 
 
 
 
1240	set_cpus_allowed_ptr(t, cm);
1241	free_cpumask_var(cm);
1242}
1243
 
 
 
 
 
 
 
 
 
1244/*
1245 * Spawn boost kthreads -- called as soon as the scheduler is running.
1246 */
1247static void __init rcu_spawn_boost_kthreads(void)
1248{
1249	struct rcu_node *rnp;
 
1250
1251	rcu_for_each_leaf_node(rnp)
1252		if (rcu_rnp_online_cpus(rnp))
1253			rcu_spawn_one_boost_kthread(rnp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1254}
1255
1256#else /* #ifdef CONFIG_RCU_BOOST */
1257
1258static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1259	__releases(rnp->lock)
1260{
1261	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 
 
 
 
 
1262}
1263
1264static bool rcu_is_callbacks_kthread(void)
1265{
1266	return false;
1267}
1268
1269static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1270{
1271}
1272
1273static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1274{
1275}
1276
1277static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1278{
 
 
1279}
 
1280
1281static void __init rcu_spawn_boost_kthreads(void)
1282{
1283}
1284
1285#endif /* #else #ifdef CONFIG_RCU_BOOST */
1286
1287#if !defined(CONFIG_RCU_FAST_NO_HZ)
1288
1289/*
1290 * Check to see if any future non-offloaded RCU-related work will need
1291 * to be done by the current CPU, even if none need be done immediately,
1292 * returning 1 if so.  This function is part of the RCU implementation;
1293 * it is -not- an exported member of the RCU API.
1294 *
1295 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1296 * CPU has RCU callbacks queued.
1297 */
1298int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1299{
1300	*nextevt = KTIME_MAX;
1301	return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1302		!rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
1303}
 
1304
1305/*
1306 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1307 * after it.
1308 */
1309static void rcu_cleanup_after_idle(void)
1310{
1311}
1312
1313/*
1314 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1315 * is nothing.
1316 */
1317static void rcu_prepare_for_idle(void)
 
 
 
 
 
 
 
 
1318{
1319}
1320
1321#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1322
1323/*
1324 * This code is invoked when a CPU goes idle, at which point we want
1325 * to have the CPU do everything required for RCU so that it can enter
1326 * the energy-efficient dyntick-idle mode.
 
1327 *
1328 * The following preprocessor symbol controls this:
1329 *
1330 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1331 *	to sleep in dyntick-idle mode with RCU callbacks pending.  This
1332 *	is sized to be roughly one RCU grace period.  Those energy-efficiency
1333 *	benchmarkers who might otherwise be tempted to set this to a large
1334 *	number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1335 *	system.  And if you are -that- concerned about energy efficiency,
1336 *	just power the system down and be done with it!
 
 
 
1337 *
1338 * The value below works well in practice.  If future workloads require
1339 * adjustment, they can be converted into kernel config parameters, though
1340 * making the state machine smarter might be a better option.
1341 */
1342#define RCU_IDLE_GP_DELAY 4		/* Roughly one grace period. */
 
1343
1344static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1345module_param(rcu_idle_gp_delay, int, 0644);
 
 
 
 
1346
1347/*
1348 * Try to advance callbacks on the current CPU, but only if it has been
1349 * awhile since the last time we did so.  Afterwards, if there are any
1350 * callbacks ready for immediate invocation, return true.
1351 */
1352static bool __maybe_unused rcu_try_advance_all_cbs(void)
1353{
1354	bool cbs_ready = false;
1355	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
1356	struct rcu_node *rnp;
 
1357
1358	/* Exit early if we advanced recently. */
1359	if (jiffies == rdp->last_advance_all)
1360		return false;
1361	rdp->last_advance_all = jiffies;
1362
1363	rnp = rdp->mynode;
 
 
1364
1365	/*
1366	 * Don't bother checking unless a grace period has
1367	 * completed since we last checked and there are
1368	 * callbacks not yet ready to invoke.
1369	 */
1370	if ((rcu_seq_completed_gp(rdp->gp_seq,
1371				  rcu_seq_current(&rnp->gp_seq)) ||
1372	     unlikely(READ_ONCE(rdp->gpwrap))) &&
1373	    rcu_segcblist_pend_cbs(&rdp->cblist))
1374		note_gp_changes(rdp);
1375
1376	if (rcu_segcblist_ready_cbs(&rdp->cblist))
1377		cbs_ready = true;
 
1378	return cbs_ready;
1379}
1380
1381/*
1382 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1383 * to invoke.  If the CPU has callbacks, try to advance them.  Tell the
1384 * caller about what to set the timeout.
 
1385 *
1386 * The caller must have disabled interrupts.
1387 */
1388int rcu_needs_cpu(u64 basemono, u64 *nextevt)
 
1389{
1390	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1391	unsigned long dj;
1392
1393	lockdep_assert_irqs_disabled();
 
1394
1395	/* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1396	if (rcu_segcblist_empty(&rdp->cblist) ||
1397	    rcu_rdp_is_offloaded(rdp)) {
1398		*nextevt = KTIME_MAX;
1399		return 0;
1400	}
1401
1402	/* Attempt to advance callbacks. */
1403	if (rcu_try_advance_all_cbs()) {
1404		/* Some ready to invoke, so initiate later invocation. */
1405		invoke_rcu_core();
1406		return 1;
1407	}
1408	rdp->last_accelerate = jiffies;
1409
1410	/* Request timer and round. */
1411	dj = round_up(rcu_idle_gp_delay + jiffies, rcu_idle_gp_delay) - jiffies;
1412
1413	*nextevt = basemono + dj * TICK_NSEC;
 
 
 
1414	return 0;
1415}
 
1416
1417/*
1418 * Prepare a CPU for idle from an RCU perspective.  The first major task is to
1419 * sense whether nohz mode has been enabled or disabled via sysfs.  The second
1420 * major task is to accelerate (that is, assign grace-period numbers to) any
1421 * recently arrived callbacks.
 
 
1422 *
1423 * The caller must have disabled interrupts.
1424 */
1425static void rcu_prepare_for_idle(void)
1426{
1427	bool needwake;
1428	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
1429	struct rcu_node *rnp;
 
1430	int tne;
1431
1432	lockdep_assert_irqs_disabled();
1433	if (rcu_rdp_is_offloaded(rdp))
1434		return;
1435
1436	/* Handle nohz enablement switches conservatively. */
1437	tne = READ_ONCE(tick_nohz_active);
1438	if (tne != rdp->tick_nohz_enabled_snap) {
1439		if (!rcu_segcblist_empty(&rdp->cblist))
1440			invoke_rcu_core(); /* force nohz to see update. */
1441		rdp->tick_nohz_enabled_snap = tne;
1442		return;
1443	}
1444	if (!tne)
1445		return;
1446
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1447	/*
1448	 * If we have not yet accelerated this jiffy, accelerate all
1449	 * callbacks on this CPU.
1450	 */
1451	if (rdp->last_accelerate == jiffies)
1452		return;
1453	rdp->last_accelerate = jiffies;
1454	if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
 
 
 
1455		rnp = rdp->mynode;
1456		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1457		needwake = rcu_accelerate_cbs(rnp, rdp);
1458		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1459		if (needwake)
1460			rcu_gp_kthread_wake();
1461	}
 
1462}
1463
1464/*
1465 * Clean up for exit from idle.  Attempt to advance callbacks based on
1466 * any grace periods that elapsed while the CPU was idle, and if any
1467 * callbacks are now ready to invoke, initiate invocation.
1468 */
1469static void rcu_cleanup_after_idle(void)
1470{
1471	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1472
1473	lockdep_assert_irqs_disabled();
1474	if (rcu_rdp_is_offloaded(rdp))
1475		return;
1476	if (rcu_try_advance_all_cbs())
1477		invoke_rcu_core();
 
1478}
1479
1480#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1481
1482#ifdef CONFIG_RCU_NOCB_CPU
1483
1484/*
1485 * Offload callback processing from the boot-time-specified set of CPUs
1486 * specified by rcu_nocb_mask.  For the CPUs in the set, there are kthreads
1487 * created that pull the callbacks from the corresponding CPU, wait for
1488 * a grace period to elapse, and invoke the callbacks.  These kthreads
1489 * are organized into GP kthreads, which manage incoming callbacks, wait for
1490 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1491 * invoke callbacks.  Each GP kthread invokes its own CBs.  The no-CBs CPUs
1492 * do a wake_up() on their GP kthread when they insert a callback into any
1493 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1494 * in which case each kthread actively polls its CPU.  (Which isn't so great
1495 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1496 *
1497 * This is intended to be used in conjunction with Frederic Weisbecker's
1498 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1499 * running CPU-bound user-mode computations.
1500 *
1501 * Offloading of callbacks can also be used as an energy-efficiency
1502 * measure because CPUs with no RCU callbacks queued are more aggressive
1503 * about entering dyntick-idle mode.
1504 */
1505
1506
1507/*
1508 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1509 * If the list is invalid, a warning is emitted and all CPUs are offloaded.
1510 */
1511static int __init rcu_nocb_setup(char *str)
1512{
1513	alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1514	if (cpulist_parse(str, rcu_nocb_mask)) {
1515		pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1516		cpumask_setall(rcu_nocb_mask);
1517	}
1518	return 1;
1519}
1520__setup("rcu_nocbs=", rcu_nocb_setup);
1521
1522static int __init parse_rcu_nocb_poll(char *arg)
1523{
1524	rcu_nocb_poll = true;
1525	return 0;
1526}
1527early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1528
1529/*
1530 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1531 * After all, the main point of bypassing is to avoid lock contention
1532 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1533 */
1534static int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1535module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1536
1537/*
1538 * Acquire the specified rcu_data structure's ->nocb_bypass_lock.  If the
1539 * lock isn't immediately available, increment ->nocb_lock_contended to
1540 * flag the contention.
1541 */
1542static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1543	__acquires(&rdp->nocb_bypass_lock)
1544{
1545	lockdep_assert_irqs_disabled();
1546	if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1547		return;
1548	atomic_inc(&rdp->nocb_lock_contended);
1549	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1550	smp_mb__after_atomic(); /* atomic_inc() before lock. */
1551	raw_spin_lock(&rdp->nocb_bypass_lock);
1552	smp_mb__before_atomic(); /* atomic_dec() after lock. */
1553	atomic_dec(&rdp->nocb_lock_contended);
1554}
1555
1556/*
1557 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1558 * not contended.  Please note that this is extremely special-purpose,
1559 * relying on the fact that at most two kthreads and one CPU contend for
1560 * this lock, and also that the two kthreads are guaranteed to have frequent
1561 * grace-period-duration time intervals between successive acquisitions
1562 * of the lock.  This allows us to use an extremely simple throttling
1563 * mechanism, and further to apply it only to the CPU doing floods of
1564 * call_rcu() invocations.  Don't try this at home!
1565 */
1566static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1567{
1568	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1569	while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1570		cpu_relax();
 
 
 
 
 
 
 
1571}
1572
1573/*
1574 * Conditionally acquire the specified rcu_data structure's
1575 * ->nocb_bypass_lock.
 
 
 
1576 */
1577static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1578{
1579	lockdep_assert_irqs_disabled();
1580	return raw_spin_trylock(&rdp->nocb_bypass_lock);
1581}
 
 
 
1582
1583/*
1584 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1585 */
1586static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1587	__releases(&rdp->nocb_bypass_lock)
1588{
1589	lockdep_assert_irqs_disabled();
1590	raw_spin_unlock(&rdp->nocb_bypass_lock);
 
 
 
 
 
 
1591}
1592
1593/*
1594 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1595 * if it corresponds to a no-CBs CPU.
1596 */
1597static void rcu_nocb_lock(struct rcu_data *rdp)
1598{
1599	lockdep_assert_irqs_disabled();
1600	if (!rcu_rdp_is_offloaded(rdp))
1601		return;
1602	raw_spin_lock(&rdp->nocb_lock);
1603}
1604
1605/*
1606 * Release the specified rcu_data structure's ->nocb_lock, but only
1607 * if it corresponds to a no-CBs CPU.
1608 */
1609static void rcu_nocb_unlock(struct rcu_data *rdp)
1610{
1611	if (rcu_rdp_is_offloaded(rdp)) {
1612		lockdep_assert_irqs_disabled();
1613		raw_spin_unlock(&rdp->nocb_lock);
1614	}
1615}
1616
1617/*
1618 * Release the specified rcu_data structure's ->nocb_lock and restore
1619 * interrupts, but only if it corresponds to a no-CBs CPU.
1620 */
1621static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1622				       unsigned long flags)
1623{
1624	if (rcu_rdp_is_offloaded(rdp)) {
1625		lockdep_assert_irqs_disabled();
1626		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
1627	} else {
1628		local_irq_restore(flags);
 
1629	}
 
 
 
 
 
 
 
1630}
1631
1632/* Lockdep check that ->cblist may be safely accessed. */
1633static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1634{
1635	lockdep_assert_irqs_disabled();
1636	if (rcu_rdp_is_offloaded(rdp))
1637		lockdep_assert_held(&rdp->nocb_lock);
1638}
1639
1640/*
1641 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1642 * grace period.
1643 */
1644static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1645{
1646	swake_up_all(sq);
 
1647}
1648
1649static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
 
1650{
1651	return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
 
 
 
1652}
1653
1654static void rcu_init_one_nocb(struct rcu_node *rnp)
 
 
1655{
1656	init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1657	init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1658}
1659
1660/* Is the specified CPU a no-CBs CPU? */
1661bool rcu_is_nocb_cpu(int cpu)
1662{
1663	if (cpumask_available(rcu_nocb_mask))
1664		return cpumask_test_cpu(cpu, rcu_nocb_mask);
1665	return false;
1666}
1667
1668static bool __wake_nocb_gp(struct rcu_data *rdp_gp,
1669			   struct rcu_data *rdp,
1670			   bool force, unsigned long flags)
1671	__releases(rdp_gp->nocb_gp_lock)
1672{
1673	bool needwake = false;
1674
1675	if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1676		raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1677		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1678				    TPS("AlreadyAwake"));
1679		return false;
1680	}
1681
1682	if (rdp_gp->nocb_defer_wakeup > RCU_NOCB_WAKE_NOT) {
1683		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
1684		del_timer(&rdp_gp->nocb_timer);
1685	}
1686
1687	if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1688		WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1689		needwake = true;
1690	}
1691	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1692	if (needwake) {
1693		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1694		wake_up_process(rdp_gp->nocb_gp_kthread);
1695	}
1696
1697	return needwake;
1698}
1699
1700/*
1701 * Kick the GP kthread for this NOCB group.
1702 */
1703static bool wake_nocb_gp(struct rcu_data *rdp, bool force)
1704{
1705	unsigned long flags;
1706	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1707
1708	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1709	return __wake_nocb_gp(rdp_gp, rdp, force, flags);
1710}
1711
1712/*
1713 * Arrange to wake the GP kthread for this NOCB group at some future
1714 * time when it is safe to do so.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1715 */
1716static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1717			       const char *reason)
1718{
1719	unsigned long flags;
1720	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1721
1722	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1723
1724	/*
1725	 * Bypass wakeup overrides previous deferments. In case
1726	 * of callback storm, no need to wake up too early.
1727	 */
1728	if (waketype == RCU_NOCB_WAKE_BYPASS) {
1729		mod_timer(&rdp_gp->nocb_timer, jiffies + 2);
1730		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
1731	} else {
1732		if (rdp_gp->nocb_defer_wakeup < RCU_NOCB_WAKE)
1733			mod_timer(&rdp_gp->nocb_timer, jiffies + 1);
1734		if (rdp_gp->nocb_defer_wakeup < waketype)
1735			WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
1736	}
1737
1738	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1739
1740	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
 
1741}
 
1742
1743/*
1744 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1745 * However, if there is a callback to be enqueued and if ->nocb_bypass
1746 * proves to be initially empty, just return false because the no-CB GP
1747 * kthread may need to be awakened in this case.
1748 *
1749 * Note that this function always returns true if rhp is NULL.
 
1750 */
1751static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1752				     unsigned long j)
1753{
1754	struct rcu_cblist rcl;
1755
1756	WARN_ON_ONCE(!rcu_rdp_is_offloaded(rdp));
1757	rcu_lockdep_assert_cblist_protected(rdp);
1758	lockdep_assert_held(&rdp->nocb_bypass_lock);
1759	if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1760		raw_spin_unlock(&rdp->nocb_bypass_lock);
1761		return false;
1762	}
1763	/* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1764	if (rhp)
1765		rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1766	rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1767	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1768	WRITE_ONCE(rdp->nocb_bypass_first, j);
1769	rcu_nocb_bypass_unlock(rdp);
1770	return true;
1771}
1772
1773/*
1774 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1775 * However, if there is a callback to be enqueued and if ->nocb_bypass
1776 * proves to be initially empty, just return false because the no-CB GP
1777 * kthread may need to be awakened in this case.
1778 *
1779 * Note that this function always returns true if rhp is NULL.
1780 */
1781static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1782				  unsigned long j)
1783{
1784	if (!rcu_rdp_is_offloaded(rdp))
1785		return true;
1786	rcu_lockdep_assert_cblist_protected(rdp);
1787	rcu_nocb_bypass_lock(rdp);
1788	return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1789}
1790
1791/*
1792 * If the ->nocb_bypass_lock is immediately available, flush the
1793 * ->nocb_bypass queue into ->cblist.
 
 
 
 
1794 */
1795static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1796{
1797	rcu_lockdep_assert_cblist_protected(rdp);
1798	if (!rcu_rdp_is_offloaded(rdp) ||
1799	    !rcu_nocb_bypass_trylock(rdp))
1800		return;
1801	WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1802}
1803
1804/*
1805 * See whether it is appropriate to use the ->nocb_bypass list in order
1806 * to control contention on ->nocb_lock.  A limited number of direct
1807 * enqueues are permitted into ->cblist per jiffy.  If ->nocb_bypass
1808 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1809 * otherwise rcu_barrier() breaks.  Use rcu_nocb_flush_bypass() to switch
1810 * back to direct use of ->cblist.  However, ->nocb_bypass should not be
1811 * used if ->cblist is empty, because otherwise callbacks can be stranded
1812 * on ->nocb_bypass because we cannot count on the current CPU ever again
1813 * invoking call_rcu().  The general rule is that if ->nocb_bypass is
1814 * non-empty, the corresponding no-CBs grace-period kthread must not be
1815 * in an indefinite sleep state.
1816 *
1817 * Finally, it is not permitted to use the bypass during early boot,
1818 * as doing so would confuse the auto-initialization code.  Besides
1819 * which, there is no point in worrying about lock contention while
1820 * there is only one CPU in operation.
1821 */
1822static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1823				bool *was_alldone, unsigned long flags)
1824{
1825	unsigned long c;
1826	unsigned long cur_gp_seq;
1827	unsigned long j = jiffies;
1828	long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1829
1830	lockdep_assert_irqs_disabled();
1831
1832	// Pure softirq/rcuc based processing: no bypassing, no
1833	// locking.
1834	if (!rcu_rdp_is_offloaded(rdp)) {
1835		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1836		return false;
1837	}
1838
1839	// In the process of (de-)offloading: no bypassing, but
1840	// locking.
1841	if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
1842		rcu_nocb_lock(rdp);
1843		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1844		return false; /* Not offloaded, no bypassing. */
1845	}
1846
1847	// Don't use ->nocb_bypass during early boot.
1848	if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1849		rcu_nocb_lock(rdp);
1850		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1851		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1852		return false;
1853	}
1854
1855	// If we have advanced to a new jiffy, reset counts to allow
1856	// moving back from ->nocb_bypass to ->cblist.
1857	if (j == rdp->nocb_nobypass_last) {
1858		c = rdp->nocb_nobypass_count + 1;
1859	} else {
1860		WRITE_ONCE(rdp->nocb_nobypass_last, j);
1861		c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1862		if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1863				 nocb_nobypass_lim_per_jiffy))
1864			c = 0;
1865		else if (c > nocb_nobypass_lim_per_jiffy)
1866			c = nocb_nobypass_lim_per_jiffy;
1867	}
1868	WRITE_ONCE(rdp->nocb_nobypass_count, c);
1869
1870	// If there hasn't yet been all that many ->cblist enqueues
1871	// this jiffy, tell the caller to enqueue onto ->cblist.  But flush
1872	// ->nocb_bypass first.
1873	if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1874		rcu_nocb_lock(rdp);
1875		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1876		if (*was_alldone)
1877			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1878					    TPS("FirstQ"));
1879		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1880		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1881		return false; // Caller must enqueue the callback.
1882	}
1883
1884	// If ->nocb_bypass has been used too long or is too full,
1885	// flush ->nocb_bypass to ->cblist.
1886	if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1887	    ncbs >= qhimark) {
1888		rcu_nocb_lock(rdp);
1889		if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1890			*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1891			if (*was_alldone)
1892				trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1893						    TPS("FirstQ"));
1894			WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1895			return false; // Caller must enqueue the callback.
1896		}
1897		if (j != rdp->nocb_gp_adv_time &&
1898		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1899		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1900			rcu_advance_cbs_nowake(rdp->mynode, rdp);
1901			rdp->nocb_gp_adv_time = j;
1902		}
1903		rcu_nocb_unlock_irqrestore(rdp, flags);
1904		return true; // Callback already enqueued.
1905	}
1906
1907	// We need to use the bypass.
1908	rcu_nocb_wait_contended(rdp);
1909	rcu_nocb_bypass_lock(rdp);
1910	ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1911	rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1912	rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1913	if (!ncbs) {
1914		WRITE_ONCE(rdp->nocb_bypass_first, j);
1915		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1916	}
1917	rcu_nocb_bypass_unlock(rdp);
1918	smp_mb(); /* Order enqueue before wake. */
1919	if (ncbs) {
1920		local_irq_restore(flags);
1921	} else {
1922		// No-CBs GP kthread might be indefinitely asleep, if so, wake.
1923		rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1924		if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1925			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1926					    TPS("FirstBQwake"));
1927			__call_rcu_nocb_wake(rdp, true, flags);
1928		} else {
1929			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1930					    TPS("FirstBQnoWake"));
1931			rcu_nocb_unlock_irqrestore(rdp, flags);
1932		}
1933	}
1934	return true; // Callback already enqueued.
1935}
 
1936
1937/*
1938 * Awaken the no-CBs grace-period kthread if needed, either due to it
1939 * legitimately being asleep or due to overload conditions.
 
 
1940 *
1941 * If warranted, also wake up the kthread servicing this CPUs queues.
1942 */
1943static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1944				 unsigned long flags)
1945				 __releases(rdp->nocb_lock)
 
 
1946{
1947	unsigned long cur_gp_seq;
1948	unsigned long j;
1949	long len;
1950	struct task_struct *t;
1951
1952	// If we are being polled or there is no kthread, just leave.
1953	t = READ_ONCE(rdp->nocb_gp_kthread);
 
 
 
 
 
 
1954	if (rcu_nocb_poll || !t) {
1955		rcu_nocb_unlock_irqrestore(rdp, flags);
1956		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1957				    TPS("WakeNotPoll"));
1958		return;
1959	}
1960	// Need to actually to a wakeup.
1961	len = rcu_segcblist_n_cbs(&rdp->cblist);
1962	if (was_alldone) {
1963		rdp->qlen_last_fqs_check = len;
1964		if (!irqs_disabled_flags(flags)) {
1965			/* ... if queue was empty ... */
1966			rcu_nocb_unlock_irqrestore(rdp, flags);
1967			wake_nocb_gp(rdp, false);
1968			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1969					    TPS("WakeEmpty"));
1970		} else {
1971			rcu_nocb_unlock_irqrestore(rdp, flags);
1972			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1973					   TPS("WakeEmptyIsDeferred"));
1974		}
 
1975	} else if (len > rdp->qlen_last_fqs_check + qhimark) {
1976		/* ... or if many callbacks queued. */
1977		rdp->qlen_last_fqs_check = len;
1978		j = jiffies;
1979		if (j != rdp->nocb_gp_adv_time &&
1980		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1981		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1982			rcu_advance_cbs_nowake(rdp->mynode, rdp);
1983			rdp->nocb_gp_adv_time = j;
1984		}
1985		smp_mb(); /* Enqueue before timer_pending(). */
1986		if ((rdp->nocb_cb_sleep ||
1987		     !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1988		    !timer_pending(&rdp->nocb_timer)) {
1989			rcu_nocb_unlock_irqrestore(rdp, flags);
1990			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1991					   TPS("WakeOvfIsDeferred"));
1992		} else {
1993			rcu_nocb_unlock_irqrestore(rdp, flags);
1994			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1995		}
1996	} else {
1997		rcu_nocb_unlock_irqrestore(rdp, flags);
1998		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1999	}
2000	return;
2001}
2002
2003/*
2004 * Check if we ignore this rdp.
 
 
 
2005 *
2006 * We check that without holding the nocb lock but
2007 * we make sure not to miss a freshly offloaded rdp
2008 * with the current ordering:
2009 *
2010 *  rdp_offload_toggle()        nocb_gp_enabled_cb()
2011 * -------------------------   ----------------------------
2012 *    WRITE flags                 LOCK nocb_gp_lock
2013 *    LOCK nocb_gp_lock           READ/WRITE nocb_gp_sleep
2014 *    READ/WRITE nocb_gp_sleep    UNLOCK nocb_gp_lock
2015 *    UNLOCK nocb_gp_lock         READ flags
2016 */
2017static inline bool nocb_gp_enabled_cb(struct rcu_data *rdp)
 
2018{
2019	u8 flags = SEGCBLIST_OFFLOADED | SEGCBLIST_KTHREAD_GP;
2020
2021	return rcu_segcblist_test_flags(&rdp->cblist, flags);
 
 
 
 
 
 
 
 
 
 
 
 
2022}
2023
2024static inline bool nocb_gp_update_state_deoffloading(struct rcu_data *rdp,
2025						     bool *needwake_state)
 
 
 
 
 
2026{
2027	struct rcu_segcblist *cblist = &rdp->cblist;
 
 
 
 
 
 
 
2028
2029	if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED)) {
2030		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP)) {
2031			rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_GP);
2032			if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
2033				*needwake_state = true;
2034		}
2035		return false;
 
 
 
 
 
 
 
2036	}
2037
2038	/*
2039	 * De-offloading. Clear our flag and notify the de-offload worker.
2040	 * We will ignore this rdp until it ever gets re-offloaded.
2041	 */
2042	WARN_ON_ONCE(!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP));
2043	rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_GP);
2044	if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
2045		*needwake_state = true;
2046	return true;
2047}
2048
2049
2050/*
2051 * No-CBs GP kthreads come here to wait for additional callbacks to show up
2052 * or for grace periods to end.
2053 */
2054static void nocb_gp_wait(struct rcu_data *my_rdp)
2055{
2056	bool bypass = false;
2057	long bypass_ncbs;
2058	int __maybe_unused cpu = my_rdp->cpu;
2059	unsigned long cur_gp_seq;
2060	unsigned long flags;
2061	bool gotcbs = false;
2062	unsigned long j = jiffies;
2063	bool needwait_gp = false; // This prevents actual uninitialized use.
2064	bool needwake;
2065	bool needwake_gp;
2066	struct rcu_data *rdp;
2067	struct rcu_node *rnp;
2068	unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
2069	bool wasempty = false;
2070
2071	/*
2072	 * Each pass through the following loop checks for CBs and for the
2073	 * nearest grace period (if any) to wait for next.  The CB kthreads
2074	 * and the global grace-period kthread are awakened if needed.
2075	 */
2076	WARN_ON_ONCE(my_rdp->nocb_gp_rdp != my_rdp);
2077	for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
2078		bool needwake_state = false;
2079
2080		if (!nocb_gp_enabled_cb(rdp))
2081			continue;
2082		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
2083		rcu_nocb_lock_irqsave(rdp, flags);
2084		if (nocb_gp_update_state_deoffloading(rdp, &needwake_state)) {
2085			rcu_nocb_unlock_irqrestore(rdp, flags);
2086			if (needwake_state)
2087				swake_up_one(&rdp->nocb_state_wq);
2088			continue;
2089		}
2090		bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
2091		if (bypass_ncbs &&
2092		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
2093		     bypass_ncbs > 2 * qhimark)) {
2094			// Bypass full or old, so flush it.
2095			(void)rcu_nocb_try_flush_bypass(rdp, j);
2096			bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
2097		} else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
2098			rcu_nocb_unlock_irqrestore(rdp, flags);
2099			if (needwake_state)
2100				swake_up_one(&rdp->nocb_state_wq);
2101			continue; /* No callbacks here, try next. */
2102		}
2103		if (bypass_ncbs) {
2104			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2105					    TPS("Bypass"));
2106			bypass = true;
2107		}
2108		rnp = rdp->mynode;
2109
2110		// Advance callbacks if helpful and low contention.
2111		needwake_gp = false;
2112		if (!rcu_segcblist_restempty(&rdp->cblist,
2113					     RCU_NEXT_READY_TAIL) ||
2114		    (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2115		     rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
2116			raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
2117			needwake_gp = rcu_advance_cbs(rnp, rdp);
2118			wasempty = rcu_segcblist_restempty(&rdp->cblist,
2119							   RCU_NEXT_READY_TAIL);
2120			raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
2121		}
2122		// Need to wait on some grace period?
2123		WARN_ON_ONCE(wasempty &&
2124			     !rcu_segcblist_restempty(&rdp->cblist,
2125						      RCU_NEXT_READY_TAIL));
2126		if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
2127			if (!needwait_gp ||
2128			    ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
2129				wait_gp_seq = cur_gp_seq;
2130			needwait_gp = true;
2131			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2132					    TPS("NeedWaitGP"));
2133		}
2134		if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2135			needwake = rdp->nocb_cb_sleep;
2136			WRITE_ONCE(rdp->nocb_cb_sleep, false);
2137			smp_mb(); /* CB invocation -after- GP end. */
2138		} else {
2139			needwake = false;
2140		}
2141		rcu_nocb_unlock_irqrestore(rdp, flags);
2142		if (needwake) {
2143			swake_up_one(&rdp->nocb_cb_wq);
2144			gotcbs = true;
2145		}
2146		if (needwake_gp)
2147			rcu_gp_kthread_wake();
2148		if (needwake_state)
2149			swake_up_one(&rdp->nocb_state_wq);
2150	}
2151
2152	my_rdp->nocb_gp_bypass = bypass;
2153	my_rdp->nocb_gp_gp = needwait_gp;
2154	my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
2155
2156	if (bypass && !rcu_nocb_poll) {
2157		// At least one child with non-empty ->nocb_bypass, so set
2158		// timer in order to avoid stranding its callbacks.
2159		wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_BYPASS,
2160				   TPS("WakeBypassIsDeferred"));
2161	}
2162	if (rcu_nocb_poll) {
2163		/* Polling, so trace if first poll in the series. */
2164		if (gotcbs)
2165			trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2166		schedule_timeout_idle(1);
2167	} else if (!needwait_gp) {
2168		/* Wait for callbacks to appear. */
2169		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2170		swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2171				!READ_ONCE(my_rdp->nocb_gp_sleep));
2172		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2173	} else {
2174		rnp = my_rdp->mynode;
2175		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2176		swait_event_interruptible_exclusive(
2177			rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2178			rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2179			!READ_ONCE(my_rdp->nocb_gp_sleep));
2180		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2181	}
2182	if (!rcu_nocb_poll) {
2183		raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2184		if (my_rdp->nocb_defer_wakeup > RCU_NOCB_WAKE_NOT) {
2185			WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2186			del_timer(&my_rdp->nocb_timer);
2187		}
2188		WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2189		raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2190	}
2191	my_rdp->nocb_gp_seq = -1;
2192	WARN_ON(signal_pending(current));
2193}
2194
2195/*
2196 * No-CBs grace-period-wait kthread.  There is one of these per group
2197 * of CPUs, but only once at least one CPU in that group has come online
2198 * at least once since boot.  This kthread checks for newly posted
2199 * callbacks from any of the CPUs it is responsible for, waits for a
2200 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2201 * that then have callback-invocation work to do.
2202 */
2203static int rcu_nocb_gp_kthread(void *arg)
2204{
 
 
 
 
 
2205	struct rcu_data *rdp = arg;
2206
 
2207	for (;;) {
2208		WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2209		nocb_gp_wait(rdp);
2210		cond_resched_tasks_rcu_qs();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2211	}
2212	return 0;
2213}
2214
2215static inline bool nocb_cb_can_run(struct rcu_data *rdp)
 
2216{
2217	u8 flags = SEGCBLIST_OFFLOADED | SEGCBLIST_KTHREAD_CB;
2218	return rcu_segcblist_test_flags(&rdp->cblist, flags);
2219}
2220
2221static inline bool nocb_cb_wait_cond(struct rcu_data *rdp)
 
2222{
2223	return nocb_cb_can_run(rdp) && !READ_ONCE(rdp->nocb_cb_sleep);
 
 
 
 
2224}
2225
2226/*
2227 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2228 * then, if there are no more, wait for more to appear.
2229 */
2230static void nocb_cb_wait(struct rcu_data *rdp)
 
 
 
 
2231{
2232	struct rcu_segcblist *cblist = &rdp->cblist;
2233	unsigned long cur_gp_seq;
2234	unsigned long flags;
2235	bool needwake_state = false;
2236	bool needwake_gp = false;
2237	bool can_sleep = true;
2238	struct rcu_node *rnp = rdp->mynode;
2239
2240	local_irq_save(flags);
2241	rcu_momentary_dyntick_idle();
2242	local_irq_restore(flags);
2243	/*
2244	 * Disable BH to provide the expected environment.  Also, when
2245	 * transitioning to/from NOCB mode, a self-requeuing callback might
2246	 * be invoked from softirq.  A short grace period could cause both
2247	 * instances of this callback would execute concurrently.
2248	 */
2249	local_bh_disable();
2250	rcu_do_batch(rdp);
2251	local_bh_enable();
2252	lockdep_assert_irqs_enabled();
2253	rcu_nocb_lock_irqsave(rdp, flags);
2254	if (rcu_segcblist_nextgp(cblist, &cur_gp_seq) &&
2255	    rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2256	    raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2257		needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2258		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2259	}
2260
2261	if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED)) {
2262		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB)) {
2263			rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_CB);
2264			if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
2265				needwake_state = true;
2266		}
2267		if (rcu_segcblist_ready_cbs(cblist))
2268			can_sleep = false;
2269	} else {
2270		/*
2271		 * De-offloading. Clear our flag and notify the de-offload worker.
2272		 * We won't touch the callbacks and keep sleeping until we ever
2273		 * get re-offloaded.
2274		 */
2275		WARN_ON_ONCE(!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB));
2276		rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_CB);
2277		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
2278			needwake_state = true;
2279	}
 
2280
2281	WRITE_ONCE(rdp->nocb_cb_sleep, can_sleep);
 
 
 
 
 
 
 
 
2282
2283	if (rdp->nocb_cb_sleep)
2284		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2285
2286	rcu_nocb_unlock_irqrestore(rdp, flags);
2287	if (needwake_gp)
2288		rcu_gp_kthread_wake();
 
2289
2290	if (needwake_state)
2291		swake_up_one(&rdp->nocb_state_wq);
 
2292
2293	do {
2294		swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2295						    nocb_cb_wait_cond(rdp));
2296
2297		// VVV Ensure CB invocation follows _sleep test.
2298		if (smp_load_acquire(&rdp->nocb_cb_sleep)) { // ^^^
2299			WARN_ON(signal_pending(current));
2300			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2301		}
2302	} while (!nocb_cb_can_run(rdp));
2303}
2304
2305/*
2306 * Per-rcu_data kthread, but only for no-CBs CPUs.  Repeatedly invoke
2307 * nocb_cb_wait() to do the dirty work.
2308 */
2309static int rcu_nocb_cb_kthread(void *arg)
2310{
2311	struct rcu_data *rdp = arg;
2312
2313	// Each pass through this loop does one callback batch, and,
2314	// if there are no more ready callbacks, waits for them.
2315	for (;;) {
2316		nocb_cb_wait(rdp);
2317		cond_resched_tasks_rcu_qs();
2318	}
2319	return 0;
2320}
2321
2322/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2323static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
 
2324{
2325	return READ_ONCE(rdp->nocb_defer_wakeup) >= level;
2326}
2327
2328/* Do a deferred wakeup of rcu_nocb_kthread(). */
2329static bool do_nocb_deferred_wakeup_common(struct rcu_data *rdp_gp,
2330					   struct rcu_data *rdp, int level,
2331					   unsigned long flags)
2332	__releases(rdp_gp->nocb_gp_lock)
2333{
2334	int ndw;
2335	int ret;
2336
2337	if (!rcu_nocb_need_deferred_wakeup(rdp_gp, level)) {
2338		raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
2339		return false;
2340	}
2341
2342	ndw = rdp_gp->nocb_defer_wakeup;
2343	ret = __wake_nocb_gp(rdp_gp, rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2344	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2345
2346	return ret;
2347}
2348
2349/* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2350static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2351{
2352	unsigned long flags;
2353	struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2354
2355	WARN_ON_ONCE(rdp->nocb_gp_rdp != rdp);
2356	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
2357
2358	raw_spin_lock_irqsave(&rdp->nocb_gp_lock, flags);
2359	smp_mb__after_spinlock(); /* Timer expire before wakeup. */
2360	do_nocb_deferred_wakeup_common(rdp, rdp, RCU_NOCB_WAKE_BYPASS, flags);
2361}
2362
2363/*
2364 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2365 * This means we do an inexact common-case check.  Note that if
2366 * we miss, ->nocb_timer will eventually clean things up.
2367 */
2368static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
2369{
2370	unsigned long flags;
2371	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
2372
2373	if (!rdp_gp || !rcu_nocb_need_deferred_wakeup(rdp_gp, RCU_NOCB_WAKE))
2374		return false;
2375
2376	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
2377	return do_nocb_deferred_wakeup_common(rdp_gp, rdp, RCU_NOCB_WAKE, flags);
2378}
2379
2380void rcu_nocb_flush_deferred_wakeup(void)
2381{
2382	do_nocb_deferred_wakeup(this_cpu_ptr(&rcu_data));
2383}
2384EXPORT_SYMBOL_GPL(rcu_nocb_flush_deferred_wakeup);
2385
2386static int rdp_offload_toggle(struct rcu_data *rdp,
2387			       bool offload, unsigned long flags)
2388	__releases(rdp->nocb_lock)
2389{
2390	struct rcu_segcblist *cblist = &rdp->cblist;
2391	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
2392	bool wake_gp = false;
2393
2394	rcu_segcblist_offload(cblist, offload);
2395
2396	if (rdp->nocb_cb_sleep)
2397		rdp->nocb_cb_sleep = false;
2398	rcu_nocb_unlock_irqrestore(rdp, flags);
2399
2400	/*
2401	 * Ignore former value of nocb_cb_sleep and force wake up as it could
2402	 * have been spuriously set to false already.
2403	 */
2404	swake_up_one(&rdp->nocb_cb_wq);
2405
2406	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
2407	if (rdp_gp->nocb_gp_sleep) {
2408		rdp_gp->nocb_gp_sleep = false;
2409		wake_gp = true;
2410	}
2411	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
2412
2413	if (wake_gp)
2414		wake_up_process(rdp_gp->nocb_gp_kthread);
2415
2416	return 0;
2417}
2418
2419static long rcu_nocb_rdp_deoffload(void *arg)
2420{
2421	struct rcu_data *rdp = arg;
2422	struct rcu_segcblist *cblist = &rdp->cblist;
2423	unsigned long flags;
2424	int ret;
2425
2426	WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
2427
2428	pr_info("De-offloading %d\n", rdp->cpu);
 
 
 
 
 
 
 
 
2429
2430	rcu_nocb_lock_irqsave(rdp, flags);
2431	/*
2432	 * Flush once and for all now. This suffices because we are
2433	 * running on the target CPU holding ->nocb_lock (thus having
2434	 * interrupts disabled), and because rdp_offload_toggle()
2435	 * invokes rcu_segcblist_offload(), which clears SEGCBLIST_OFFLOADED.
2436	 * Thus future calls to rcu_segcblist_completely_offloaded() will
2437	 * return false, which means that future calls to rcu_nocb_try_bypass()
2438	 * will refuse to put anything into the bypass.
2439	 */
2440	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
2441	ret = rdp_offload_toggle(rdp, false, flags);
2442	swait_event_exclusive(rdp->nocb_state_wq,
2443			      !rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB |
2444							SEGCBLIST_KTHREAD_GP));
2445	/*
2446	 * Lock one last time to acquire latest callback updates from kthreads
2447	 * so we can later handle callbacks locally without locking.
2448	 */
2449	rcu_nocb_lock_irqsave(rdp, flags);
2450	/*
2451	 * Theoretically we could set SEGCBLIST_SOFTIRQ_ONLY after the nocb
2452	 * lock is released but how about being paranoid for once?
2453	 */
2454	rcu_segcblist_set_flags(cblist, SEGCBLIST_SOFTIRQ_ONLY);
2455	/*
2456	 * With SEGCBLIST_SOFTIRQ_ONLY, we can't use
2457	 * rcu_nocb_unlock_irqrestore() anymore.
2458	 */
2459	raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2460
2461	/* Sanity check */
2462	WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
2463
2464
2465	return ret;
2466}
2467
2468int rcu_nocb_cpu_deoffload(int cpu)
2469{
2470	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2471	int ret = 0;
2472
2473	mutex_lock(&rcu_state.barrier_mutex);
2474	cpus_read_lock();
2475	if (rcu_rdp_is_offloaded(rdp)) {
2476		if (cpu_online(cpu)) {
2477			ret = work_on_cpu(cpu, rcu_nocb_rdp_deoffload, rdp);
2478			if (!ret)
2479				cpumask_clear_cpu(cpu, rcu_nocb_mask);
 
 
 
2480		} else {
2481			pr_info("NOCB: Can't CB-deoffload an offline CPU\n");
2482			ret = -EINVAL;
 
2483		}
2484	}
2485	cpus_read_unlock();
2486	mutex_unlock(&rcu_state.barrier_mutex);
2487
2488	return ret;
 
 
 
 
 
 
2489}
2490EXPORT_SYMBOL_GPL(rcu_nocb_cpu_deoffload);
2491
2492static long rcu_nocb_rdp_offload(void *arg)
 
 
 
 
 
 
 
 
 
2493{
2494	struct rcu_data *rdp = arg;
2495	struct rcu_segcblist *cblist = &rdp->cblist;
2496	unsigned long flags;
2497	int ret;
2498
2499	WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
2500	/*
2501	 * For now we only support re-offload, ie: the rdp must have been
2502	 * offloaded on boot first.
 
2503	 */
2504	if (!rdp->nocb_gp_rdp)
2505		return -EINVAL;
2506
2507	pr_info("Offloading %d\n", rdp->cpu);
2508	/*
2509	 * Can't use rcu_nocb_lock_irqsave() while we are in
2510	 * SEGCBLIST_SOFTIRQ_ONLY mode.
2511	 */
2512	raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2513
2514	/*
2515	 * We didn't take the nocb lock while working on the
2516	 * rdp->cblist in SEGCBLIST_SOFTIRQ_ONLY mode.
2517	 * Every modifications that have been done previously on
2518	 * rdp->cblist must be visible remotely by the nocb kthreads
2519	 * upon wake up after reading the cblist flags.
2520	 *
2521	 * The layout against nocb_lock enforces that ordering:
2522	 *
2523	 *  __rcu_nocb_rdp_offload()   nocb_cb_wait()/nocb_gp_wait()
2524	 * -------------------------   ----------------------------
2525	 *      WRITE callbacks           rcu_nocb_lock()
2526	 *      rcu_nocb_lock()           READ flags
2527	 *      WRITE flags               READ callbacks
2528	 *      rcu_nocb_unlock()         rcu_nocb_unlock()
2529	 */
2530	ret = rdp_offload_toggle(rdp, true, flags);
2531	swait_event_exclusive(rdp->nocb_state_wq,
2532			      rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB) &&
2533			      rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP));
2534
2535	return ret;
2536}
2537
2538int rcu_nocb_cpu_offload(int cpu)
2539{
2540	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2541	int ret = 0;
2542
2543	mutex_lock(&rcu_state.barrier_mutex);
2544	cpus_read_lock();
2545	if (!rcu_rdp_is_offloaded(rdp)) {
2546		if (cpu_online(cpu)) {
2547			ret = work_on_cpu(cpu, rcu_nocb_rdp_offload, rdp);
2548			if (!ret)
2549				cpumask_set_cpu(cpu, rcu_nocb_mask);
2550		} else {
2551			pr_info("NOCB: Can't CB-offload an offline CPU\n");
2552			ret = -EINVAL;
2553		}
 
2554	}
2555	cpus_read_unlock();
2556	mutex_unlock(&rcu_state.barrier_mutex);
2557
2558	return ret;
2559}
2560EXPORT_SYMBOL_GPL(rcu_nocb_cpu_offload);
2561
2562void __init rcu_init_nohz(void)
 
 
 
 
 
2563{
2564	int cpu;
2565	bool need_rcu_nocb_mask = false;
2566	struct rcu_data *rdp;
2567
2568#if defined(CONFIG_NO_HZ_FULL)
2569	if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2570		need_rcu_nocb_mask = true;
2571#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2572
2573	if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2574		if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2575			pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2576			return;
 
 
 
 
 
2577		}
2578	}
2579	if (!cpumask_available(rcu_nocb_mask))
2580		return;
2581
2582#if defined(CONFIG_NO_HZ_FULL)
2583	if (tick_nohz_full_running)
2584		cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2585#endif /* #if defined(CONFIG_NO_HZ_FULL) */
 
2586
2587	if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2588		pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2589		cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2590			    rcu_nocb_mask);
2591	}
2592	if (cpumask_empty(rcu_nocb_mask))
2593		pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2594	else
2595		pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2596			cpumask_pr_args(rcu_nocb_mask));
2597	if (rcu_nocb_poll)
2598		pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2599
2600	for_each_cpu(cpu, rcu_nocb_mask) {
2601		rdp = per_cpu_ptr(&rcu_data, cpu);
2602		if (rcu_segcblist_empty(&rdp->cblist))
2603			rcu_segcblist_init(&rdp->cblist);
2604		rcu_segcblist_offload(&rdp->cblist, true);
2605		rcu_segcblist_set_flags(&rdp->cblist, SEGCBLIST_KTHREAD_CB);
2606		rcu_segcblist_set_flags(&rdp->cblist, SEGCBLIST_KTHREAD_GP);
2607	}
2608	rcu_organize_nocb_kthreads();
2609}
2610
2611/* Initialize per-rcu_data variables for no-CBs CPUs. */
2612static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2613{
2614	init_swait_queue_head(&rdp->nocb_cb_wq);
2615	init_swait_queue_head(&rdp->nocb_gp_wq);
2616	init_swait_queue_head(&rdp->nocb_state_wq);
2617	raw_spin_lock_init(&rdp->nocb_lock);
2618	raw_spin_lock_init(&rdp->nocb_bypass_lock);
2619	raw_spin_lock_init(&rdp->nocb_gp_lock);
2620	timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2621	rcu_cblist_init(&rdp->nocb_bypass);
2622}
2623
2624/*
2625 * If the specified CPU is a no-CBs CPU that does not already have its
2626 * rcuo CB kthread, spawn it.  Additionally, if the rcuo GP kthread
2627 * for this CPU's group has not yet been created, spawn it as well.
2628 */
2629static void rcu_spawn_one_nocb_kthread(int cpu)
 
2630{
2631	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2632	struct rcu_data *rdp_gp;
2633	struct task_struct *t;
2634
2635	/*
2636	 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2637	 * then nothing to do.
 
2638	 */
2639	if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
 
2640		return;
 
 
2641
2642	/* If we didn't spawn the GP kthread first, reorganize! */
2643	rdp_gp = rdp->nocb_gp_rdp;
2644	if (!rdp_gp->nocb_gp_kthread) {
2645		t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2646				"rcuog/%d", rdp_gp->cpu);
2647		if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2648			return;
2649		WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2650	}
 
2651
2652	/* Spawn the kthread for this CPU. */
2653	t = kthread_run(rcu_nocb_cb_kthread, rdp,
2654			"rcuo%c/%d", rcu_state.abbr, cpu);
2655	if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2656		return;
2657	WRITE_ONCE(rdp->nocb_cb_kthread, t);
2658	WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2659}
2660
2661/*
2662 * If the specified CPU is a no-CBs CPU that does not already have its
2663 * rcuo kthread, spawn it.
2664 */
2665static void rcu_spawn_cpu_nocb_kthread(int cpu)
2666{
2667	if (rcu_scheduler_fully_active)
2668		rcu_spawn_one_nocb_kthread(cpu);
2669}
2670
2671/*
2672 * Once the scheduler is running, spawn rcuo kthreads for all online
2673 * no-CBs CPUs.  This assumes that the early_initcall()s happen before
2674 * non-boot CPUs come online -- if this changes, we will need to add
2675 * some mutual exclusion.
2676 */
2677static void __init rcu_spawn_nocb_kthreads(void)
2678{
2679	int cpu;
2680
2681	for_each_online_cpu(cpu)
2682		rcu_spawn_cpu_nocb_kthread(cpu);
 
 
2683}
2684
2685/* How many CB CPU IDs per GP kthread?  Default of -1 for sqrt(nr_cpu_ids). */
2686static int rcu_nocb_gp_stride = -1;
2687module_param(rcu_nocb_gp_stride, int, 0444);
 
 
 
 
 
 
 
 
 
 
 
 
2688
2689/*
2690 * Initialize GP-CB relationships for all no-CBs CPU.
 
2691 */
2692static void __init rcu_organize_nocb_kthreads(void)
2693{
2694	int cpu;
2695	bool firsttime = true;
2696	bool gotnocbs = false;
2697	bool gotnocbscbs = true;
2698	int ls = rcu_nocb_gp_stride;
2699	int nl = 0;  /* Next GP kthread. */
2700	struct rcu_data *rdp;
2701	struct rcu_data *rdp_gp = NULL;  /* Suppress misguided gcc warn. */
2702	struct rcu_data *rdp_prev = NULL;
 
 
 
 
 
 
 
 
 
 
 
2703
2704	if (!cpumask_available(rcu_nocb_mask))
2705		return;
2706	if (ls == -1) {
2707		ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2708		rcu_nocb_gp_stride = ls;
2709	}
 
 
2710
2711	/*
2712	 * Each pass through this loop sets up one rcu_data structure.
2713	 * Should the corresponding CPU come online in the future, then
2714	 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2715	 */
2716	for_each_cpu(cpu, rcu_nocb_mask) {
2717		rdp = per_cpu_ptr(&rcu_data, cpu);
2718		if (rdp->cpu >= nl) {
2719			/* New GP kthread, set up for CBs & next GP. */
2720			gotnocbs = true;
2721			nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2722			rdp->nocb_gp_rdp = rdp;
2723			rdp_gp = rdp;
2724			if (dump_tree) {
2725				if (!firsttime)
2726					pr_cont("%s\n", gotnocbscbs
2727							? "" : " (self only)");
2728				gotnocbscbs = false;
2729				firsttime = false;
2730				pr_alert("%s: No-CB GP kthread CPU %d:",
2731					 __func__, cpu);
2732			}
2733		} else {
2734			/* Another CB kthread, link to previous GP kthread. */
2735			gotnocbscbs = true;
2736			rdp->nocb_gp_rdp = rdp_gp;
2737			rdp_prev->nocb_next_cb_rdp = rdp;
2738			if (dump_tree)
2739				pr_cont(" %d", cpu);
2740		}
2741		rdp_prev = rdp;
2742	}
2743	if (gotnocbs && dump_tree)
2744		pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
2745}
2746
2747/*
2748 * Bind the current task to the offloaded CPUs.  If there are no offloaded
2749 * CPUs, leave the task unbound.  Splat if the bind attempt fails.
2750 */
2751void rcu_bind_current_to_nocb(void)
2752{
2753	if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2754		WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2755}
2756EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2757
2758// The ->on_cpu field is available only in CONFIG_SMP=y, so...
2759#ifdef CONFIG_SMP
2760static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
 
 
 
2761{
2762	return tsp && task_is_running(tsp) && !tsp->on_cpu ? "!" : "";
 
 
 
 
 
 
 
2763}
2764#else // #ifdef CONFIG_SMP
2765static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
2766{
2767	return "";
2768}
2769#endif // #else #ifdef CONFIG_SMP
2770
2771/*
2772 * Dump out nocb grace-period kthread state for the specified rcu_data
2773 * structure.
2774 */
2775static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
 
2776{
2777	struct rcu_node *rnp = rdp->mynode;
2778
2779	pr_info("nocb GP %d %c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu %c CPU %d%s\n",
2780		rdp->cpu,
2781		"kK"[!!rdp->nocb_gp_kthread],
2782		"lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2783		"dD"[!!rdp->nocb_defer_wakeup],
2784		"tT"[timer_pending(&rdp->nocb_timer)],
2785		"sS"[!!rdp->nocb_gp_sleep],
2786		".W"[swait_active(&rdp->nocb_gp_wq)],
2787		".W"[swait_active(&rnp->nocb_gp_wq[0])],
2788		".W"[swait_active(&rnp->nocb_gp_wq[1])],
2789		".B"[!!rdp->nocb_gp_bypass],
2790		".G"[!!rdp->nocb_gp_gp],
2791		(long)rdp->nocb_gp_seq,
2792		rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops),
2793		rdp->nocb_gp_kthread ? task_state_to_char(rdp->nocb_gp_kthread) : '.',
2794		rdp->nocb_cb_kthread ? (int)task_cpu(rdp->nocb_gp_kthread) : -1,
2795		show_rcu_should_be_on_cpu(rdp->nocb_cb_kthread));
2796}
2797
2798/* Dump out nocb kthread state for the specified rcu_data structure. */
2799static void show_rcu_nocb_state(struct rcu_data *rdp)
2800{
2801	char bufw[20];
2802	char bufr[20];
2803	struct rcu_segcblist *rsclp = &rdp->cblist;
2804	bool waslocked;
2805	bool wassleep;
2806
2807	if (rdp->nocb_gp_rdp == rdp)
2808		show_rcu_nocb_gp_state(rdp);
2809
2810	sprintf(bufw, "%ld", rsclp->gp_seq[RCU_WAIT_TAIL]);
2811	sprintf(bufr, "%ld", rsclp->gp_seq[RCU_NEXT_READY_TAIL]);
2812	pr_info("   CB %d^%d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%s%c%s%c%c q%ld %c CPU %d%s\n",
2813		rdp->cpu, rdp->nocb_gp_rdp->cpu,
2814		rdp->nocb_next_cb_rdp ? rdp->nocb_next_cb_rdp->cpu : -1,
2815		"kK"[!!rdp->nocb_cb_kthread],
2816		"bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2817		"cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2818		"lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2819		"sS"[!!rdp->nocb_cb_sleep],
2820		".W"[swait_active(&rdp->nocb_cb_wq)],
2821		jiffies - rdp->nocb_bypass_first,
2822		jiffies - rdp->nocb_nobypass_last,
2823		rdp->nocb_nobypass_count,
2824		".D"[rcu_segcblist_ready_cbs(rsclp)],
2825		".W"[!rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL)],
2826		rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL) ? "" : bufw,
2827		".R"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL)],
2828		rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL) ? "" : bufr,
2829		".N"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_TAIL)],
2830		".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2831		rcu_segcblist_n_cbs(&rdp->cblist),
2832		rdp->nocb_cb_kthread ? task_state_to_char(rdp->nocb_cb_kthread) : '.',
2833		rdp->nocb_cb_kthread ? (int)task_cpu(rdp->nocb_gp_kthread) : -1,
2834		show_rcu_should_be_on_cpu(rdp->nocb_cb_kthread));
2835
2836	/* It is OK for GP kthreads to have GP state. */
2837	if (rdp->nocb_gp_rdp == rdp)
2838		return;
2839
2840	waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2841	wassleep = swait_active(&rdp->nocb_gp_wq);
2842	if (!rdp->nocb_gp_sleep && !waslocked && !wassleep)
2843		return;  /* Nothing untoward. */
2844
2845	pr_info("   nocb GP activity on CB-only CPU!!! %c%c%c %c\n",
2846		"lL"[waslocked],
2847		"dD"[!!rdp->nocb_defer_wakeup],
2848		"sS"[!!rdp->nocb_gp_sleep],
2849		".W"[wassleep]);
2850}
2851
2852#else /* #ifdef CONFIG_RCU_NOCB_CPU */
 
 
 
 
2853
2854/* No ->nocb_lock to acquire.  */
2855static void rcu_nocb_lock(struct rcu_data *rdp)
2856{
 
 
 
 
 
 
 
 
 
 
 
2857}
2858
2859/* No ->nocb_lock to release.  */
2860static void rcu_nocb_unlock(struct rcu_data *rdp)
 
 
 
2861{
2862}
 
 
 
 
 
 
 
 
2863
2864/* No ->nocb_lock to release.  */
2865static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2866				       unsigned long flags)
2867{
2868	local_irq_restore(flags);
2869}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2870
2871/* Lockdep check that ->cblist may be safely accessed. */
2872static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2873{
2874	lockdep_assert_irqs_disabled();
2875}
 
2876
2877static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2878{
2879}
2880
2881static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2882{
2883	return NULL;
2884}
2885
2886static void rcu_init_one_nocb(struct rcu_node *rnp)
2887{
 
 
 
 
 
 
 
 
 
 
2888}
2889
2890static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2891				  unsigned long j)
 
 
2892{
2893	return true;
2894}
2895
2896static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2897				bool *was_alldone, unsigned long flags)
2898{
2899	return false;
2900}
2901
2902static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2903				 unsigned long flags)
2904{
2905	WARN_ON_ONCE(1);  /* Should be dead code! */
2906}
2907
2908static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2909{
2910}
2911
2912static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
 
2913{
2914	return false;
2915}
2916
2917static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
2918{
2919	return false;
2920}
2921
2922static void rcu_spawn_cpu_nocb_kthread(int cpu)
2923{
2924}
2925
2926static void __init rcu_spawn_nocb_kthreads(void)
 
2927{
2928}
2929
2930static void show_rcu_nocb_state(struct rcu_data *rdp)
2931{
2932}
2933
2934#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2935
2936/*
2937 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2938 * grace-period kthread will do force_quiescent_state() processing?
2939 * The idea is to avoid waking up RCU core processing on such a
2940 * CPU unless the grace period has extended for too long.
2941 *
2942 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2943 * CONFIG_RCU_NOCB_CPU CPUs.
2944 */
2945static bool rcu_nohz_full_cpu(void)
2946{
2947#ifdef CONFIG_NO_HZ_FULL
2948	if (tick_nohz_full_cpu(smp_processor_id()) &&
2949	    (!rcu_gp_in_progress() ||
2950	     time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2951		return true;
2952#endif /* #ifdef CONFIG_NO_HZ_FULL */
2953	return false;
2954}
2955
2956/*
2957 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2958 */
2959static void rcu_bind_gp_kthread(void)
2960{
2961	if (!tick_nohz_full_enabled())
2962		return;
2963	housekeeping_affine(current, HK_FLAG_RCU);
2964}
2965
2966/* Record the current task on dyntick-idle entry. */
2967static void noinstr rcu_dynticks_task_enter(void)
2968{
2969#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2970	WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2971#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2972}
2973
2974/* Record no current task on dyntick-idle exit. */
2975static void noinstr rcu_dynticks_task_exit(void)
2976{
2977#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2978	WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2979#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2980}
2981
2982/* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
2983static void rcu_dynticks_task_trace_enter(void)
2984{
2985#ifdef CONFIG_TASKS_TRACE_RCU
2986	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2987		current->trc_reader_special.b.need_mb = true;
2988#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
2989}
2990
2991/* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
2992static void rcu_dynticks_task_trace_exit(void)
2993{
2994#ifdef CONFIG_TASKS_TRACE_RCU
2995	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2996		current->trc_reader_special.b.need_mb = false;
2997#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
2998}