Linux Audio

Check our new training course

Loading...
v5.4
   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. */
  19#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
  20
  21/*
  22 * Check the RCU kernel configuration parameters and print informative
  23 * messages about anything out of the ordinary.
  24 */
  25static void __init rcu_bootup_announce_oddness(void)
  26{
  27	if (IS_ENABLED(CONFIG_RCU_TRACE))
  28		pr_info("\tRCU event tracing is enabled.\n");
  29	if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
  30	    (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
  31		pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
  32			RCU_FANOUT);
  33	if (rcu_fanout_exact)
  34		pr_info("\tHierarchical RCU autobalancing is disabled.\n");
  35	if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
  36		pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
  37	if (IS_ENABLED(CONFIG_PROVE_RCU))
  38		pr_info("\tRCU lockdep checking is enabled.\n");
  39	if (RCU_NUM_LVLS >= 4)
  40		pr_info("\tFour(or more)-level hierarchy is enabled.\n");
  41	if (RCU_FANOUT_LEAF != 16)
  42		pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
  43			RCU_FANOUT_LEAF);
  44	if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
  45		pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
  46			rcu_fanout_leaf);
  47	if (nr_cpu_ids != NR_CPUS)
  48		pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
  49#ifdef CONFIG_RCU_BOOST
  50	pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
  51		kthread_prio, CONFIG_RCU_BOOST_DELAY);
  52#endif
  53	if (blimit != DEFAULT_RCU_BLIMIT)
  54		pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
  55	if (qhimark != DEFAULT_RCU_QHIMARK)
  56		pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
  57	if (qlowmark != DEFAULT_RCU_QLOMARK)
  58		pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
 
 
  59	if (jiffies_till_first_fqs != ULONG_MAX)
  60		pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
  61	if (jiffies_till_next_fqs != ULONG_MAX)
  62		pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
  63	if (jiffies_till_sched_qs != ULONG_MAX)
  64		pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
  65	if (rcu_kick_kthreads)
  66		pr_info("\tKick kthreads if too-long grace period.\n");
  67	if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
  68		pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
  69	if (gp_preinit_delay)
  70		pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
  71	if (gp_init_delay)
  72		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
  73	if (gp_cleanup_delay)
  74		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
  75	if (!use_softirq)
  76		pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
  77	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
  78		pr_info("\tRCU debug extended QS entry/exit.\n");
  79	rcupdate_announce_bootup_oddness();
  80}
  81
  82#ifdef CONFIG_PREEMPT_RCU
  83
  84static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
  85static void rcu_read_unlock_special(struct task_struct *t);
  86
  87/*
  88 * Tell them what RCU they are running.
  89 */
  90static void __init rcu_bootup_announce(void)
  91{
  92	pr_info("Preemptible hierarchical RCU implementation.\n");
  93	rcu_bootup_announce_oddness();
  94}
  95
  96/* Flags for rcu_preempt_ctxt_queue() decision table. */
  97#define RCU_GP_TASKS	0x8
  98#define RCU_EXP_TASKS	0x4
  99#define RCU_GP_BLKD	0x2
 100#define RCU_EXP_BLKD	0x1
 101
 102/*
 103 * Queues a task preempted within an RCU-preempt read-side critical
 104 * section into the appropriate location within the ->blkd_tasks list,
 105 * depending on the states of any ongoing normal and expedited grace
 106 * periods.  The ->gp_tasks pointer indicates which element the normal
 107 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
 108 * indicates which element the expedited grace period is waiting on (again,
 109 * NULL if none).  If a grace period is waiting on a given element in the
 110 * ->blkd_tasks list, it also waits on all subsequent elements.  Thus,
 111 * adding a task to the tail of the list blocks any grace period that is
 112 * already waiting on one of the elements.  In contrast, adding a task
 113 * to the head of the list won't block any grace period that is already
 114 * waiting on one of the elements.
 115 *
 116 * This queuing is imprecise, and can sometimes make an ongoing grace
 117 * period wait for a task that is not strictly speaking blocking it.
 118 * Given the choice, we needlessly block a normal grace period rather than
 119 * blocking an expedited grace period.
 120 *
 121 * Note that an endless sequence of expedited grace periods still cannot
 122 * indefinitely postpone a normal grace period.  Eventually, all of the
 123 * fixed number of preempted tasks blocking the normal grace period that are
 124 * not also blocking the expedited grace period will resume and complete
 125 * their RCU read-side critical sections.  At that point, the ->gp_tasks
 126 * pointer will equal the ->exp_tasks pointer, at which point the end of
 127 * the corresponding expedited grace period will also be the end of the
 128 * normal grace period.
 129 */
 130static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
 131	__releases(rnp->lock) /* But leaves rrupts disabled. */
 132{
 133	int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
 134			 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
 135			 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
 136			 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
 137	struct task_struct *t = current;
 138
 139	raw_lockdep_assert_held_rcu_node(rnp);
 140	WARN_ON_ONCE(rdp->mynode != rnp);
 141	WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 142	/* RCU better not be waiting on newly onlined CPUs! */
 143	WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
 144		     rdp->grpmask);
 145
 146	/*
 147	 * Decide where to queue the newly blocked task.  In theory,
 148	 * this could be an if-statement.  In practice, when I tried
 149	 * that, it was quite messy.
 150	 */
 151	switch (blkd_state) {
 152	case 0:
 153	case                RCU_EXP_TASKS:
 154	case                RCU_EXP_TASKS + RCU_GP_BLKD:
 155	case RCU_GP_TASKS:
 156	case RCU_GP_TASKS + RCU_EXP_TASKS:
 157
 158		/*
 159		 * Blocking neither GP, or first task blocking the normal
 160		 * GP but not blocking the already-waiting expedited GP.
 161		 * Queue at the head of the list to avoid unnecessarily
 162		 * blocking the already-waiting GPs.
 163		 */
 164		list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
 165		break;
 166
 167	case                                              RCU_EXP_BLKD:
 168	case                                RCU_GP_BLKD:
 169	case                                RCU_GP_BLKD + RCU_EXP_BLKD:
 170	case RCU_GP_TASKS +                               RCU_EXP_BLKD:
 171	case RCU_GP_TASKS +                 RCU_GP_BLKD + RCU_EXP_BLKD:
 172	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 173
 174		/*
 175		 * First task arriving that blocks either GP, or first task
 176		 * arriving that blocks the expedited GP (with the normal
 177		 * GP already waiting), or a task arriving that blocks
 178		 * both GPs with both GPs already waiting.  Queue at the
 179		 * tail of the list to avoid any GP waiting on any of the
 180		 * already queued tasks that are not blocking it.
 181		 */
 182		list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
 183		break;
 184
 185	case                RCU_EXP_TASKS +               RCU_EXP_BLKD:
 186	case                RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 187	case RCU_GP_TASKS + RCU_EXP_TASKS +               RCU_EXP_BLKD:
 188
 189		/*
 190		 * Second or subsequent task blocking the expedited GP.
 191		 * The task either does not block the normal GP, or is the
 192		 * first task blocking the normal GP.  Queue just after
 193		 * the first task blocking the expedited GP.
 194		 */
 195		list_add(&t->rcu_node_entry, rnp->exp_tasks);
 196		break;
 197
 198	case RCU_GP_TASKS +                 RCU_GP_BLKD:
 199	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
 200
 201		/*
 202		 * Second or subsequent task blocking the normal GP.
 203		 * The task does not block the expedited GP. Queue just
 204		 * after the first task blocking the normal GP.
 205		 */
 206		list_add(&t->rcu_node_entry, rnp->gp_tasks);
 207		break;
 208
 209	default:
 210
 211		/* Yet another exercise in excessive paranoia. */
 212		WARN_ON_ONCE(1);
 213		break;
 214	}
 215
 216	/*
 217	 * We have now queued the task.  If it was the first one to
 218	 * block either grace period, update the ->gp_tasks and/or
 219	 * ->exp_tasks pointers, respectively, to reference the newly
 220	 * blocked tasks.
 221	 */
 222	if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
 223		rnp->gp_tasks = &t->rcu_node_entry;
 224		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
 225	}
 226	if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
 227		rnp->exp_tasks = &t->rcu_node_entry;
 228	WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
 229		     !(rnp->qsmask & rdp->grpmask));
 230	WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
 231		     !(rnp->expmask & rdp->grpmask));
 232	raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
 233
 234	/*
 235	 * Report the quiescent state for the expedited GP.  This expedited
 236	 * GP should not be able to end until we report, so there should be
 237	 * no need to check for a subsequent expedited GP.  (Though we are
 238	 * still in a quiescent state in any case.)
 239	 */
 240	if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
 241		rcu_report_exp_rdp(rdp);
 242	else
 243		WARN_ON_ONCE(rdp->exp_deferred_qs);
 244}
 245
 246/*
 247 * Record a preemptible-RCU quiescent state for the specified CPU.
 248 * Note that this does not necessarily mean that the task currently running
 249 * on the CPU is in a quiescent state:  Instead, it means that the current
 250 * grace period need not wait on any RCU read-side critical section that
 251 * starts later on this CPU.  It also means that if the current task is
 252 * in an RCU read-side critical section, it has already added itself to
 253 * some leaf rcu_node structure's ->blkd_tasks list.  In addition to the
 254 * current task, there might be any number of other tasks blocked while
 255 * in an RCU read-side critical section.
 256 *
 257 * Callers to this function must disable preemption.
 258 */
 259static void rcu_qs(void)
 260{
 261	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
 262	if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
 263		trace_rcu_grace_period(TPS("rcu_preempt"),
 264				       __this_cpu_read(rcu_data.gp_seq),
 265				       TPS("cpuqs"));
 266		__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 267		barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
 268		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
 269	}
 270}
 271
 272/*
 273 * We have entered the scheduler, and the current task might soon be
 274 * context-switched away from.  If this task is in an RCU read-side
 275 * critical section, we will no longer be able to rely on the CPU to
 276 * record that fact, so we enqueue the task on the blkd_tasks list.
 277 * The task will dequeue itself when it exits the outermost enclosing
 278 * RCU read-side critical section.  Therefore, the current grace period
 279 * cannot be permitted to complete until the blkd_tasks list entries
 280 * predating the current grace period drain, in other words, until
 281 * rnp->gp_tasks becomes NULL.
 282 *
 283 * Caller must disable interrupts.
 284 */
 285void rcu_note_context_switch(bool preempt)
 286{
 287	struct task_struct *t = current;
 288	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 289	struct rcu_node *rnp;
 290
 291	trace_rcu_utilization(TPS("Start context switch"));
 292	lockdep_assert_irqs_disabled();
 293	WARN_ON_ONCE(!preempt && t->rcu_read_lock_nesting > 0);
 294	if (t->rcu_read_lock_nesting > 0 &&
 295	    !t->rcu_read_unlock_special.b.blocked) {
 296
 297		/* Possibly blocking in an RCU read-side critical section. */
 298		rnp = rdp->mynode;
 299		raw_spin_lock_rcu_node(rnp);
 300		t->rcu_read_unlock_special.b.blocked = true;
 301		t->rcu_blocked_node = rnp;
 302
 303		/*
 304		 * Verify the CPU's sanity, trace the preemption, and
 305		 * then queue the task as required based on the states
 306		 * of any ongoing and expedited grace periods.
 307		 */
 308		WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
 309		WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
 310		trace_rcu_preempt_task(rcu_state.name,
 311				       t->pid,
 312				       (rnp->qsmask & rdp->grpmask)
 313				       ? rnp->gp_seq
 314				       : rcu_seq_snap(&rnp->gp_seq));
 315		rcu_preempt_ctxt_queue(rnp, rdp);
 316	} else {
 317		rcu_preempt_deferred_qs(t);
 318	}
 319
 320	/*
 321	 * Either we were not in an RCU read-side critical section to
 322	 * begin with, or we have now recorded that critical section
 323	 * globally.  Either way, we can now note a quiescent state
 324	 * for this CPU.  Again, if we were in an RCU read-side critical
 325	 * section, and if that critical section was blocking the current
 326	 * grace period, then the fact that the task has been enqueued
 327	 * means that we continue to block the current grace period.
 328	 */
 329	rcu_qs();
 330	if (rdp->exp_deferred_qs)
 331		rcu_report_exp_rdp(rdp);
 
 332	trace_rcu_utilization(TPS("End context switch"));
 333}
 334EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 335
 336/*
 337 * Check for preempted RCU readers blocking the current grace period
 338 * for the specified rcu_node structure.  If the caller needs a reliable
 339 * answer, it must hold the rcu_node's ->lock.
 340 */
 341static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 342{
 343	return rnp->gp_tasks != NULL;
 344}
 345
 346/* Bias and limit values for ->rcu_read_lock_nesting. */
 347#define RCU_NEST_BIAS INT_MAX
 348#define RCU_NEST_NMAX (-INT_MAX / 2)
 349#define RCU_NEST_PMAX (INT_MAX / 2)
 350
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 351/*
 352 * Preemptible RCU implementation for rcu_read_lock().
 353 * Just increment ->rcu_read_lock_nesting, shared state will be updated
 354 * if we block.
 355 */
 356void __rcu_read_lock(void)
 357{
 358	current->rcu_read_lock_nesting++;
 359	if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 360		WARN_ON_ONCE(current->rcu_read_lock_nesting > RCU_NEST_PMAX);
 361	barrier();  /* critical section after entry code. */
 362}
 363EXPORT_SYMBOL_GPL(__rcu_read_lock);
 364
 365/*
 366 * Preemptible RCU implementation for rcu_read_unlock().
 367 * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
 368 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
 369 * invoke rcu_read_unlock_special() to clean up after a context switch
 370 * in an RCU read-side critical section and other special cases.
 371 */
 372void __rcu_read_unlock(void)
 373{
 374	struct task_struct *t = current;
 375
 376	if (t->rcu_read_lock_nesting != 1) {
 377		--t->rcu_read_lock_nesting;
 378	} else {
 379		barrier();  /* critical section before exit code. */
 380		t->rcu_read_lock_nesting = -RCU_NEST_BIAS;
 381		barrier();  /* assign before ->rcu_read_unlock_special load */
 382		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
 383			rcu_read_unlock_special(t);
 384		barrier();  /* ->rcu_read_unlock_special load before assign */
 385		t->rcu_read_lock_nesting = 0;
 386	}
 387	if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 388		int rrln = t->rcu_read_lock_nesting;
 389
 390		WARN_ON_ONCE(rrln < 0 && rrln > RCU_NEST_NMAX);
 391	}
 392}
 393EXPORT_SYMBOL_GPL(__rcu_read_unlock);
 394
 395/*
 396 * Advance a ->blkd_tasks-list pointer to the next entry, instead
 397 * returning NULL if at the end of the list.
 398 */
 399static struct list_head *rcu_next_node_entry(struct task_struct *t,
 400					     struct rcu_node *rnp)
 401{
 402	struct list_head *np;
 403
 404	np = t->rcu_node_entry.next;
 405	if (np == &rnp->blkd_tasks)
 406		np = NULL;
 407	return np;
 408}
 409
 410/*
 411 * Return true if the specified rcu_node structure has tasks that were
 412 * preempted within an RCU read-side critical section.
 413 */
 414static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 415{
 416	return !list_empty(&rnp->blkd_tasks);
 417}
 418
 419/*
 420 * Report deferred quiescent states.  The deferral time can
 421 * be quite short, for example, in the case of the call from
 422 * rcu_read_unlock_special().
 423 */
 424static void
 425rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
 426{
 427	bool empty_exp;
 428	bool empty_norm;
 429	bool empty_exp_now;
 430	struct list_head *np;
 431	bool drop_boost_mutex = false;
 432	struct rcu_data *rdp;
 433	struct rcu_node *rnp;
 434	union rcu_special special;
 435
 436	/*
 437	 * If RCU core is waiting for this CPU to exit its critical section,
 438	 * report the fact that it has exited.  Because irqs are disabled,
 439	 * t->rcu_read_unlock_special cannot change.
 440	 */
 441	special = t->rcu_read_unlock_special;
 442	rdp = this_cpu_ptr(&rcu_data);
 443	if (!special.s && !rdp->exp_deferred_qs) {
 444		local_irq_restore(flags);
 445		return;
 446	}
 447	t->rcu_read_unlock_special.b.deferred_qs = false;
 448	if (special.b.need_qs) {
 449		rcu_qs();
 450		t->rcu_read_unlock_special.b.need_qs = false;
 451		if (!t->rcu_read_unlock_special.s && !rdp->exp_deferred_qs) {
 452			local_irq_restore(flags);
 453			return;
 454		}
 455	}
 456
 457	/*
 458	 * Respond to a request by an expedited grace period for a
 459	 * quiescent state from this CPU.  Note that requests from
 460	 * tasks are handled when removing the task from the
 461	 * blocked-tasks list below.
 462	 */
 463	if (rdp->exp_deferred_qs) {
 464		rcu_report_exp_rdp(rdp);
 465		if (!t->rcu_read_unlock_special.s) {
 466			local_irq_restore(flags);
 467			return;
 468		}
 469	}
 470
 471	/* Clean up if blocked during RCU read-side critical section. */
 472	if (special.b.blocked) {
 473		t->rcu_read_unlock_special.b.blocked = false;
 474
 475		/*
 476		 * Remove this task from the list it blocked on.  The task
 477		 * now remains queued on the rcu_node corresponding to the
 478		 * CPU it first blocked on, so there is no longer any need
 479		 * to loop.  Retain a WARN_ON_ONCE() out of sheer paranoia.
 480		 */
 481		rnp = t->rcu_blocked_node;
 482		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
 483		WARN_ON_ONCE(rnp != t->rcu_blocked_node);
 484		WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 485		empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
 486		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
 487			     (!empty_norm || rnp->qsmask));
 488		empty_exp = sync_rcu_preempt_exp_done(rnp);
 489		smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
 490		np = rcu_next_node_entry(t, rnp);
 491		list_del_init(&t->rcu_node_entry);
 492		t->rcu_blocked_node = NULL;
 493		trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
 494						rnp->gp_seq, t->pid);
 495		if (&t->rcu_node_entry == rnp->gp_tasks)
 496			rnp->gp_tasks = np;
 497		if (&t->rcu_node_entry == rnp->exp_tasks)
 498			rnp->exp_tasks = np;
 499		if (IS_ENABLED(CONFIG_RCU_BOOST)) {
 500			/* Snapshot ->boost_mtx ownership w/rnp->lock held. */
 501			drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
 502			if (&t->rcu_node_entry == rnp->boost_tasks)
 503				rnp->boost_tasks = np;
 504		}
 505
 506		/*
 507		 * If this was the last task on the current list, and if
 508		 * we aren't waiting on any CPUs, report the quiescent state.
 509		 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
 510		 * so we must take a snapshot of the expedited state.
 511		 */
 512		empty_exp_now = sync_rcu_preempt_exp_done(rnp);
 513		if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
 514			trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
 515							 rnp->gp_seq,
 516							 0, rnp->qsmask,
 517							 rnp->level,
 518							 rnp->grplo,
 519							 rnp->grphi,
 520							 !!rnp->gp_tasks);
 521			rcu_report_unblock_qs_rnp(rnp, flags);
 522		} else {
 523			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 524		}
 525
 526		/* Unboost if we were boosted. */
 527		if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
 528			rt_mutex_futex_unlock(&rnp->boost_mtx);
 529
 530		/*
 531		 * If this was the last task on the expedited lists,
 532		 * then we need to report up the rcu_node hierarchy.
 533		 */
 534		if (!empty_exp && empty_exp_now)
 535			rcu_report_exp_rnp(rnp, true);
 536	} else {
 537		local_irq_restore(flags);
 538	}
 539}
 540
 541/*
 542 * Is a deferred quiescent-state pending, and are we also not in
 543 * an RCU read-side critical section?  It is the caller's responsibility
 544 * to ensure it is otherwise safe to report any deferred quiescent
 545 * states.  The reason for this is that it is safe to report a
 546 * quiescent state during context switch even though preemption
 547 * is disabled.  This function cannot be expected to understand these
 548 * nuances, so the caller must handle them.
 549 */
 550static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 551{
 552	return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
 553		READ_ONCE(t->rcu_read_unlock_special.s)) &&
 554	       t->rcu_read_lock_nesting <= 0;
 555}
 556
 557/*
 558 * Report a deferred quiescent state if needed and safe to do so.
 559 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
 560 * not being in an RCU read-side critical section.  The caller must
 561 * evaluate safety in terms of interrupt, softirq, and preemption
 562 * disabling.
 563 */
 564static void rcu_preempt_deferred_qs(struct task_struct *t)
 565{
 566	unsigned long flags;
 567	bool couldrecurse = t->rcu_read_lock_nesting >= 0;
 568
 569	if (!rcu_preempt_need_deferred_qs(t))
 570		return;
 571	if (couldrecurse)
 572		t->rcu_read_lock_nesting -= RCU_NEST_BIAS;
 573	local_irq_save(flags);
 574	rcu_preempt_deferred_qs_irqrestore(t, flags);
 575	if (couldrecurse)
 576		t->rcu_read_lock_nesting += RCU_NEST_BIAS;
 577}
 578
 579/*
 580 * Minimal handler to give the scheduler a chance to re-evaluate.
 581 */
 582static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
 583{
 584	struct rcu_data *rdp;
 585
 586	rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
 587	rdp->defer_qs_iw_pending = false;
 588}
 589
 590/*
 591 * Handle special cases during rcu_read_unlock(), such as needing to
 592 * notify RCU core processing or task having blocked during the RCU
 593 * read-side critical section.
 594 */
 595static void rcu_read_unlock_special(struct task_struct *t)
 596{
 597	unsigned long flags;
 598	bool preempt_bh_were_disabled =
 599			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
 600	bool irqs_were_disabled;
 601
 602	/* NMI handlers cannot block and cannot safely manipulate state. */
 603	if (in_nmi())
 604		return;
 605
 606	local_irq_save(flags);
 607	irqs_were_disabled = irqs_disabled_flags(flags);
 608	if (preempt_bh_were_disabled || irqs_were_disabled) {
 609		bool exp;
 610		struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 611		struct rcu_node *rnp = rdp->mynode;
 612
 613		t->rcu_read_unlock_special.b.exp_hint = false;
 614		exp = (t->rcu_blocked_node && t->rcu_blocked_node->exp_tasks) ||
 615		      (rdp->grpmask & rnp->expmask) ||
 616		      tick_nohz_full_cpu(rdp->cpu);
 617		// Need to defer quiescent state until everything is enabled.
 618		if (irqs_were_disabled && use_softirq &&
 619		    (in_interrupt() ||
 620		     (exp && !t->rcu_read_unlock_special.b.deferred_qs))) {
 621			// Using softirq, safe to awaken, and we get
 622			// no help from enabling irqs, unlike bh/preempt.
 623			raise_softirq_irqoff(RCU_SOFTIRQ);
 624		} else {
 625			// Enabling BH or preempt does reschedule, so...
 626			// Also if no expediting or NO_HZ_FULL, slow is OK.
 
 627			set_tsk_need_resched(current);
 628			set_preempt_need_resched();
 629			if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
 630			    !rdp->defer_qs_iw_pending && exp) {
 631				// Get scheduler to re-evaluate and call hooks.
 632				// If !IRQ_WORK, FQS scan will eventually IPI.
 633				init_irq_work(&rdp->defer_qs_iw,
 634					      rcu_preempt_deferred_qs_handler);
 635				rdp->defer_qs_iw_pending = true;
 636				irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
 637			}
 638		}
 639		t->rcu_read_unlock_special.b.deferred_qs = true;
 640		local_irq_restore(flags);
 641		return;
 642	}
 643	WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
 644	rcu_preempt_deferred_qs_irqrestore(t, flags);
 645}
 646
 647/*
 648 * Check that the list of blocked tasks for the newly completed grace
 649 * period is in fact empty.  It is a serious bug to complete a grace
 650 * period that still has RCU readers blocked!  This function must be
 651 * invoked -before- updating this rnp's ->gp_seq, and the rnp's ->lock
 652 * must be held by the caller.
 653 *
 654 * Also, if there are blocked tasks on the list, they automatically
 655 * block the newly created grace period, so set up ->gp_tasks accordingly.
 656 */
 657static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 658{
 659	struct task_struct *t;
 660
 661	RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
 
 662	if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
 663		dump_blkd_tasks(rnp, 10);
 664	if (rcu_preempt_has_tasks(rnp) &&
 665	    (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
 666		rnp->gp_tasks = rnp->blkd_tasks.next;
 667		t = container_of(rnp->gp_tasks, struct task_struct,
 668				 rcu_node_entry);
 669		trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
 670						rnp->gp_seq, t->pid);
 671	}
 672	WARN_ON_ONCE(rnp->qsmask);
 673}
 674
 675/*
 676 * Check for a quiescent state from the current CPU, including voluntary
 677 * context switches for Tasks RCU.  When a task blocks, the task is
 678 * recorded in the corresponding CPU's rcu_node structure, which is checked
 679 * elsewhere, hence this function need only check for quiescent states
 680 * related to the current CPU, not to those related to tasks.
 681 */
 682static void rcu_flavor_sched_clock_irq(int user)
 683{
 684	struct task_struct *t = current;
 685
 686	if (user || rcu_is_cpu_rrupt_from_idle()) {
 687		rcu_note_voluntary_context_switch(current);
 688	}
 689	if (t->rcu_read_lock_nesting > 0 ||
 690	    (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
 691		/* No QS, force context switch if deferred. */
 692		if (rcu_preempt_need_deferred_qs(t)) {
 693			set_tsk_need_resched(t);
 694			set_preempt_need_resched();
 695		}
 696	} else if (rcu_preempt_need_deferred_qs(t)) {
 697		rcu_preempt_deferred_qs(t); /* Report deferred QS. */
 698		return;
 699	} else if (!t->rcu_read_lock_nesting) {
 700		rcu_qs(); /* Report immediate QS. */
 701		return;
 702	}
 703
 704	/* If GP is oldish, ask for help from rcu_read_unlock_special(). */
 705	if (t->rcu_read_lock_nesting > 0 &&
 706	    __this_cpu_read(rcu_data.core_needs_qs) &&
 707	    __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
 708	    !t->rcu_read_unlock_special.b.need_qs &&
 709	    time_after(jiffies, rcu_state.gp_start + HZ))
 710		t->rcu_read_unlock_special.b.need_qs = true;
 711}
 712
 713/*
 714 * Check for a task exiting while in a preemptible-RCU read-side
 715 * critical section, clean up if so.  No need to issue warnings, as
 716 * debug_check_no_locks_held() already does this if lockdep is enabled.
 717 * Besides, if this function does anything other than just immediately
 718 * return, there was a bug of some sort.  Spewing warnings from this
 719 * function is like as not to simply obscure important prior warnings.
 720 */
 721void exit_rcu(void)
 722{
 723	struct task_struct *t = current;
 724
 725	if (unlikely(!list_empty(&current->rcu_node_entry))) {
 726		t->rcu_read_lock_nesting = 1;
 727		barrier();
 728		WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
 729	} else if (unlikely(t->rcu_read_lock_nesting)) {
 730		t->rcu_read_lock_nesting = 1;
 731	} else {
 732		return;
 733	}
 734	__rcu_read_unlock();
 735	rcu_preempt_deferred_qs(current);
 736}
 737
 738/*
 739 * Dump the blocked-tasks state, but limit the list dump to the
 740 * specified number of elements.
 741 */
 742static void
 743dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 744{
 745	int cpu;
 746	int i;
 747	struct list_head *lhp;
 748	bool onl;
 749	struct rcu_data *rdp;
 750	struct rcu_node *rnp1;
 751
 752	raw_lockdep_assert_held_rcu_node(rnp);
 753	pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
 754		__func__, rnp->grplo, rnp->grphi, rnp->level,
 755		(long)rnp->gp_seq, (long)rnp->completedqs);
 756	for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
 757		pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
 758			__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
 759	pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
 760		__func__, rnp->gp_tasks, rnp->boost_tasks, rnp->exp_tasks);
 
 761	pr_info("%s: ->blkd_tasks", __func__);
 762	i = 0;
 763	list_for_each(lhp, &rnp->blkd_tasks) {
 764		pr_cont(" %p", lhp);
 765		if (++i >= ncheck)
 766			break;
 767	}
 768	pr_cont("\n");
 769	for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
 770		rdp = per_cpu_ptr(&rcu_data, cpu);
 771		onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
 772		pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
 773			cpu, ".o"[onl],
 774			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
 775			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
 776	}
 777}
 778
 779#else /* #ifdef CONFIG_PREEMPT_RCU */
 780
 781/*
 782 * Tell them what RCU they are running.
 783 */
 784static void __init rcu_bootup_announce(void)
 785{
 786	pr_info("Hierarchical RCU implementation.\n");
 787	rcu_bootup_announce_oddness();
 788}
 789
 790/*
 791 * Note a quiescent state for PREEMPT=n.  Because we do not need to know
 792 * how many quiescent states passed, just if there was at least one since
 793 * the start of the grace period, this just sets a flag.  The caller must
 794 * have disabled preemption.
 795 */
 796static void rcu_qs(void)
 797{
 798	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
 799	if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
 800		return;
 801	trace_rcu_grace_period(TPS("rcu_sched"),
 802			       __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
 803	__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 804	if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
 805		return;
 806	__this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
 807	rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 808}
 809
 810/*
 811 * Register an urgently needed quiescent state.  If there is an
 812 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
 813 * dyntick-idle quiescent state visible to other CPUs, which will in
 814 * some cases serve for expedited as well as normal grace periods.
 815 * Either way, register a lightweight quiescent state.
 816 */
 817void rcu_all_qs(void)
 818{
 819	unsigned long flags;
 820
 821	if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
 822		return;
 823	preempt_disable();
 824	/* Load rcu_urgent_qs before other flags. */
 825	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
 826		preempt_enable();
 827		return;
 828	}
 829	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 830	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
 831		local_irq_save(flags);
 832		rcu_momentary_dyntick_idle();
 833		local_irq_restore(flags);
 834	}
 835	rcu_qs();
 836	preempt_enable();
 837}
 838EXPORT_SYMBOL_GPL(rcu_all_qs);
 839
 840/*
 841 * Note a PREEMPT=n context switch.  The caller must have disabled interrupts.
 842 */
 843void rcu_note_context_switch(bool preempt)
 844{
 845	trace_rcu_utilization(TPS("Start context switch"));
 846	rcu_qs();
 847	/* Load rcu_urgent_qs before other flags. */
 848	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
 849		goto out;
 850	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 851	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
 852		rcu_momentary_dyntick_idle();
 853	if (!preempt)
 854		rcu_tasks_qs(current);
 855out:
 856	trace_rcu_utilization(TPS("End context switch"));
 857}
 858EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 859
 860/*
 861 * Because preemptible RCU does not exist, there are never any preempted
 862 * RCU readers.
 863 */
 864static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 865{
 866	return 0;
 867}
 868
 869/*
 870 * Because there is no preemptible RCU, there can be no readers blocked.
 871 */
 872static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 873{
 874	return false;
 875}
 876
 877/*
 878 * Because there is no preemptible RCU, there can be no deferred quiescent
 879 * states.
 880 */
 881static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 882{
 883	return false;
 884}
 885static void rcu_preempt_deferred_qs(struct task_struct *t) { }
 886
 887/*
 888 * Because there is no preemptible RCU, there can be no readers blocked,
 889 * so there is no need to check for blocked tasks.  So check only for
 890 * bogus qsmask values.
 891 */
 892static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 893{
 894	WARN_ON_ONCE(rnp->qsmask);
 895}
 896
 897/*
 898 * Check to see if this CPU is in a non-context-switch quiescent state,
 899 * namely user mode and idle loop.
 900 */
 901static void rcu_flavor_sched_clock_irq(int user)
 902{
 903	if (user || rcu_is_cpu_rrupt_from_idle()) {
 904
 905		/*
 906		 * Get here if this CPU took its interrupt from user
 907		 * mode or from the idle loop, and if this is not a
 908		 * nested interrupt.  In this case, the CPU is in
 909		 * a quiescent state, so note it.
 910		 *
 911		 * No memory barrier is required here because rcu_qs()
 912		 * references only CPU-local variables that other CPUs
 913		 * neither access nor modify, at least not while the
 914		 * corresponding CPU is online.
 915		 */
 916
 917		rcu_qs();
 918	}
 919}
 920
 921/*
 922 * Because preemptible RCU does not exist, tasks cannot possibly exit
 923 * while in preemptible RCU read-side critical sections.
 924 */
 925void exit_rcu(void)
 926{
 927}
 928
 929/*
 930 * Dump the guaranteed-empty blocked-tasks state.  Trust but verify.
 931 */
 932static void
 933dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 934{
 935	WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
 936}
 937
 938#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 939
 940/*
 941 * If boosting, set rcuc kthreads to realtime priority.
 942 */
 943static void rcu_cpu_kthread_setup(unsigned int cpu)
 944{
 945#ifdef CONFIG_RCU_BOOST
 946	struct sched_param sp;
 947
 948	sp.sched_priority = kthread_prio;
 949	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 950#endif /* #ifdef CONFIG_RCU_BOOST */
 951}
 952
 953#ifdef CONFIG_RCU_BOOST
 954
 955/*
 956 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
 957 * or ->boost_tasks, advancing the pointer to the next task in the
 958 * ->blkd_tasks list.
 959 *
 960 * Note that irqs must be enabled: boosting the task can block.
 961 * Returns 1 if there are more tasks needing to be boosted.
 962 */
 963static int rcu_boost(struct rcu_node *rnp)
 964{
 965	unsigned long flags;
 966	struct task_struct *t;
 967	struct list_head *tb;
 968
 969	if (READ_ONCE(rnp->exp_tasks) == NULL &&
 970	    READ_ONCE(rnp->boost_tasks) == NULL)
 971		return 0;  /* Nothing left to boost. */
 972
 973	raw_spin_lock_irqsave_rcu_node(rnp, flags);
 974
 975	/*
 976	 * Recheck under the lock: all tasks in need of boosting
 977	 * might exit their RCU read-side critical sections on their own.
 978	 */
 979	if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
 980		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 981		return 0;
 982	}
 983
 984	/*
 985	 * Preferentially boost tasks blocking expedited grace periods.
 986	 * This cannot starve the normal grace periods because a second
 987	 * expedited grace period must boost all blocked tasks, including
 988	 * those blocking the pre-existing normal grace period.
 989	 */
 990	if (rnp->exp_tasks != NULL)
 991		tb = rnp->exp_tasks;
 992	else
 993		tb = rnp->boost_tasks;
 994
 995	/*
 996	 * We boost task t by manufacturing an rt_mutex that appears to
 997	 * be held by task t.  We leave a pointer to that rt_mutex where
 998	 * task t can find it, and task t will release the mutex when it
 999	 * exits its outermost RCU read-side critical section.  Then
1000	 * simply acquiring this artificial rt_mutex will boost task
1001	 * t's priority.  (Thanks to tglx for suggesting this approach!)
1002	 *
1003	 * Note that task t must acquire rnp->lock to remove itself from
1004	 * the ->blkd_tasks list, which it will do from exit() if from
1005	 * nowhere else.  We therefore are guaranteed that task t will
1006	 * stay around at least until we drop rnp->lock.  Note that
1007	 * rnp->lock also resolves races between our priority boosting
1008	 * and task t's exiting its outermost RCU read-side critical
1009	 * section.
1010	 */
1011	t = container_of(tb, struct task_struct, rcu_node_entry);
1012	rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1013	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1014	/* Lock only for side effect: boosts task t's priority. */
1015	rt_mutex_lock(&rnp->boost_mtx);
1016	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1017
1018	return READ_ONCE(rnp->exp_tasks) != NULL ||
1019	       READ_ONCE(rnp->boost_tasks) != NULL;
1020}
1021
1022/*
1023 * Priority-boosting kthread, one per leaf rcu_node.
1024 */
1025static int rcu_boost_kthread(void *arg)
1026{
1027	struct rcu_node *rnp = (struct rcu_node *)arg;
1028	int spincnt = 0;
1029	int more2boost;
1030
1031	trace_rcu_utilization(TPS("Start boost kthread@init"));
1032	for (;;) {
1033		rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1034		trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1035		rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
 
1036		trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1037		rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1038		more2boost = rcu_boost(rnp);
1039		if (more2boost)
1040			spincnt++;
1041		else
1042			spincnt = 0;
1043		if (spincnt > 10) {
1044			rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1045			trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1046			schedule_timeout_interruptible(2);
1047			trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1048			spincnt = 0;
1049		}
1050	}
1051	/* NOTREACHED */
1052	trace_rcu_utilization(TPS("End boost kthread@notreached"));
1053	return 0;
1054}
1055
1056/*
1057 * Check to see if it is time to start boosting RCU readers that are
1058 * blocking the current grace period, and, if so, tell the per-rcu_node
1059 * kthread to start boosting them.  If there is an expedited grace
1060 * period in progress, it is always time to boost.
1061 *
1062 * The caller must hold rnp->lock, which this function releases.
1063 * The ->boost_kthread_task is immortal, so we don't need to worry
1064 * about it going away.
1065 */
1066static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1067	__releases(rnp->lock)
1068{
1069	raw_lockdep_assert_held_rcu_node(rnp);
1070	if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1071		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1072		return;
1073	}
1074	if (rnp->exp_tasks != NULL ||
1075	    (rnp->gp_tasks != NULL &&
1076	     rnp->boost_tasks == NULL &&
1077	     rnp->qsmask == 0 &&
1078	     ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1079		if (rnp->exp_tasks == NULL)
1080			rnp->boost_tasks = rnp->gp_tasks;
1081		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1082		rcu_wake_cond(rnp->boost_kthread_task,
1083			      rnp->boost_kthread_status);
1084	} else {
1085		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1086	}
1087}
1088
1089/*
1090 * Is the current CPU running the RCU-callbacks kthread?
1091 * Caller must have preemption disabled.
1092 */
1093static bool rcu_is_callbacks_kthread(void)
1094{
1095	return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1096}
1097
1098#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1099
1100/*
1101 * Do priority-boost accounting for the start of a new grace period.
1102 */
1103static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1104{
1105	rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1106}
1107
1108/*
1109 * Create an RCU-boost kthread for the specified node if one does not
1110 * already exist.  We only create this kthread for preemptible RCU.
1111 * Returns zero if all is well, a negated errno otherwise.
1112 */
1113static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1114{
1115	int rnp_index = rnp - rcu_get_root();
1116	unsigned long flags;
1117	struct sched_param sp;
1118	struct task_struct *t;
1119
1120	if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1121		return;
1122
1123	if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1124		return;
1125
1126	rcu_state.boost = 1;
1127
1128	if (rnp->boost_kthread_task != NULL)
1129		return;
1130
1131	t = kthread_create(rcu_boost_kthread, (void *)rnp,
1132			   "rcub/%d", rnp_index);
1133	if (WARN_ON_ONCE(IS_ERR(t)))
1134		return;
1135
1136	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1137	rnp->boost_kthread_task = t;
1138	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1139	sp.sched_priority = kthread_prio;
1140	sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1141	wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1142}
1143
1144/*
1145 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1146 * served by the rcu_node in question.  The CPU hotplug lock is still
1147 * held, so the value of rnp->qsmaskinit will be stable.
1148 *
1149 * We don't include outgoingcpu in the affinity set, use -1 if there is
1150 * no outgoing CPU.  If there are no CPUs left in the affinity set,
1151 * this function allows the kthread to execute on any CPU.
1152 */
1153static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1154{
1155	struct task_struct *t = rnp->boost_kthread_task;
1156	unsigned long mask = rcu_rnp_online_cpus(rnp);
1157	cpumask_var_t cm;
1158	int cpu;
1159
1160	if (!t)
1161		return;
1162	if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1163		return;
1164	for_each_leaf_node_possible_cpu(rnp, cpu)
1165		if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1166		    cpu != outgoingcpu)
1167			cpumask_set_cpu(cpu, cm);
1168	if (cpumask_weight(cm) == 0)
1169		cpumask_setall(cm);
1170	set_cpus_allowed_ptr(t, cm);
1171	free_cpumask_var(cm);
1172}
1173
1174/*
1175 * Spawn boost kthreads -- called as soon as the scheduler is running.
1176 */
1177static void __init rcu_spawn_boost_kthreads(void)
1178{
1179	struct rcu_node *rnp;
1180
1181	rcu_for_each_leaf_node(rnp)
1182		rcu_spawn_one_boost_kthread(rnp);
1183}
1184
1185static void rcu_prepare_kthreads(int cpu)
1186{
1187	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1188	struct rcu_node *rnp = rdp->mynode;
1189
1190	/* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1191	if (rcu_scheduler_fully_active)
1192		rcu_spawn_one_boost_kthread(rnp);
1193}
1194
1195#else /* #ifdef CONFIG_RCU_BOOST */
1196
1197static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1198	__releases(rnp->lock)
1199{
1200	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1201}
1202
1203static bool rcu_is_callbacks_kthread(void)
1204{
1205	return false;
1206}
1207
1208static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1209{
1210}
1211
1212static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1213{
1214}
1215
1216static void __init rcu_spawn_boost_kthreads(void)
1217{
1218}
1219
1220static void rcu_prepare_kthreads(int cpu)
1221{
1222}
1223
1224#endif /* #else #ifdef CONFIG_RCU_BOOST */
1225
1226#if !defined(CONFIG_RCU_FAST_NO_HZ)
1227
1228/*
1229 * Check to see if any future non-offloaded RCU-related work will need
1230 * to be done by the current CPU, even if none need be done immediately,
1231 * returning 1 if so.  This function is part of the RCU implementation;
1232 * it is -not- an exported member of the RCU API.
1233 *
1234 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1235 * CPU has RCU callbacks queued.
1236 */
1237int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1238{
1239	*nextevt = KTIME_MAX;
1240	return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1241	       !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist);
1242}
1243
1244/*
1245 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1246 * after it.
1247 */
1248static void rcu_cleanup_after_idle(void)
1249{
1250}
1251
1252/*
1253 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1254 * is nothing.
1255 */
1256static void rcu_prepare_for_idle(void)
1257{
1258}
1259
1260#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1261
1262/*
1263 * This code is invoked when a CPU goes idle, at which point we want
1264 * to have the CPU do everything required for RCU so that it can enter
1265 * the energy-efficient dyntick-idle mode.  This is handled by a
1266 * state machine implemented by rcu_prepare_for_idle() below.
1267 *
1268 * The following three proprocessor symbols control this state machine:
1269 *
1270 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1271 *	to sleep in dyntick-idle mode with RCU callbacks pending.  This
1272 *	is sized to be roughly one RCU grace period.  Those energy-efficiency
1273 *	benchmarkers who might otherwise be tempted to set this to a large
1274 *	number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1275 *	system.  And if you are -that- concerned about energy efficiency,
1276 *	just power the system down and be done with it!
1277 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1278 *	permitted to sleep in dyntick-idle mode with only lazy RCU
1279 *	callbacks pending.  Setting this too high can OOM your system.
1280 *
1281 * The values below work well in practice.  If future workloads require
1282 * adjustment, they can be converted into kernel config parameters, though
1283 * making the state machine smarter might be a better option.
1284 */
1285#define RCU_IDLE_GP_DELAY 4		/* Roughly one grace period. */
1286#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ)	/* Roughly six seconds. */
1287
1288static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1289module_param(rcu_idle_gp_delay, int, 0644);
1290static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1291module_param(rcu_idle_lazy_gp_delay, int, 0644);
1292
1293/*
1294 * Try to advance callbacks on the current CPU, but only if it has been
1295 * awhile since the last time we did so.  Afterwards, if there are any
1296 * callbacks ready for immediate invocation, return true.
1297 */
1298static bool __maybe_unused rcu_try_advance_all_cbs(void)
1299{
1300	bool cbs_ready = false;
1301	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1302	struct rcu_node *rnp;
1303
1304	/* Exit early if we advanced recently. */
1305	if (jiffies == rdp->last_advance_all)
1306		return false;
1307	rdp->last_advance_all = jiffies;
1308
1309	rnp = rdp->mynode;
1310
1311	/*
1312	 * Don't bother checking unless a grace period has
1313	 * completed since we last checked and there are
1314	 * callbacks not yet ready to invoke.
1315	 */
1316	if ((rcu_seq_completed_gp(rdp->gp_seq,
1317				  rcu_seq_current(&rnp->gp_seq)) ||
1318	     unlikely(READ_ONCE(rdp->gpwrap))) &&
1319	    rcu_segcblist_pend_cbs(&rdp->cblist))
1320		note_gp_changes(rdp);
1321
1322	if (rcu_segcblist_ready_cbs(&rdp->cblist))
1323		cbs_ready = true;
1324	return cbs_ready;
1325}
1326
1327/*
1328 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1329 * to invoke.  If the CPU has callbacks, try to advance them.  Tell the
1330 * caller to set the timeout based on whether or not there are non-lazy
1331 * callbacks.
1332 *
1333 * The caller must have disabled interrupts.
1334 */
1335int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1336{
1337	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1338	unsigned long dj;
1339
1340	lockdep_assert_irqs_disabled();
1341
1342	/* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1343	if (rcu_segcblist_empty(&rdp->cblist) ||
1344	    rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist)) {
1345		*nextevt = KTIME_MAX;
1346		return 0;
1347	}
1348
1349	/* Attempt to advance callbacks. */
1350	if (rcu_try_advance_all_cbs()) {
1351		/* Some ready to invoke, so initiate later invocation. */
1352		invoke_rcu_core();
1353		return 1;
1354	}
1355	rdp->last_accelerate = jiffies;
1356
1357	/* Request timer delay depending on laziness, and round. */
1358	rdp->all_lazy = !rcu_segcblist_n_nonlazy_cbs(&rdp->cblist);
1359	if (rdp->all_lazy) {
1360		dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1361	} else {
1362		dj = round_up(rcu_idle_gp_delay + jiffies,
1363			       rcu_idle_gp_delay) - jiffies;
1364	}
1365	*nextevt = basemono + dj * TICK_NSEC;
1366	return 0;
1367}
1368
1369/*
1370 * Prepare a CPU for idle from an RCU perspective.  The first major task
1371 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1372 * The second major task is to check to see if a non-lazy callback has
1373 * arrived at a CPU that previously had only lazy callbacks.  The third
1374 * major task is to accelerate (that is, assign grace-period numbers to)
1375 * any recently arrived callbacks.
1376 *
1377 * The caller must have disabled interrupts.
1378 */
1379static void rcu_prepare_for_idle(void)
1380{
1381	bool needwake;
1382	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1383	struct rcu_node *rnp;
1384	int tne;
1385
1386	lockdep_assert_irqs_disabled();
1387	if (rcu_segcblist_is_offloaded(&rdp->cblist))
1388		return;
1389
1390	/* Handle nohz enablement switches conservatively. */
1391	tne = READ_ONCE(tick_nohz_active);
1392	if (tne != rdp->tick_nohz_enabled_snap) {
1393		if (!rcu_segcblist_empty(&rdp->cblist))
1394			invoke_rcu_core(); /* force nohz to see update. */
1395		rdp->tick_nohz_enabled_snap = tne;
1396		return;
1397	}
1398	if (!tne)
1399		return;
1400
1401	/*
1402	 * If a non-lazy callback arrived at a CPU having only lazy
1403	 * callbacks, invoke RCU core for the side-effect of recalculating
1404	 * idle duration on re-entry to idle.
1405	 */
1406	if (rdp->all_lazy && rcu_segcblist_n_nonlazy_cbs(&rdp->cblist)) {
1407		rdp->all_lazy = false;
1408		invoke_rcu_core();
1409		return;
1410	}
1411
1412	/*
1413	 * If we have not yet accelerated this jiffy, accelerate all
1414	 * callbacks on this CPU.
1415	 */
1416	if (rdp->last_accelerate == jiffies)
1417		return;
1418	rdp->last_accelerate = jiffies;
1419	if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1420		rnp = rdp->mynode;
1421		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1422		needwake = rcu_accelerate_cbs(rnp, rdp);
1423		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1424		if (needwake)
1425			rcu_gp_kthread_wake();
1426	}
1427}
1428
1429/*
1430 * Clean up for exit from idle.  Attempt to advance callbacks based on
1431 * any grace periods that elapsed while the CPU was idle, and if any
1432 * callbacks are now ready to invoke, initiate invocation.
1433 */
1434static void rcu_cleanup_after_idle(void)
1435{
1436	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1437
1438	lockdep_assert_irqs_disabled();
1439	if (rcu_segcblist_is_offloaded(&rdp->cblist))
1440		return;
1441	if (rcu_try_advance_all_cbs())
1442		invoke_rcu_core();
1443}
1444
1445#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1446
1447#ifdef CONFIG_RCU_NOCB_CPU
1448
1449/*
1450 * Offload callback processing from the boot-time-specified set of CPUs
1451 * specified by rcu_nocb_mask.  For the CPUs in the set, there are kthreads
1452 * created that pull the callbacks from the corresponding CPU, wait for
1453 * a grace period to elapse, and invoke the callbacks.  These kthreads
1454 * are organized into GP kthreads, which manage incoming callbacks, wait for
1455 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1456 * invoke callbacks.  Each GP kthread invokes its own CBs.  The no-CBs CPUs
1457 * do a wake_up() on their GP kthread when they insert a callback into any
1458 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1459 * in which case each kthread actively polls its CPU.  (Which isn't so great
1460 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1461 *
1462 * This is intended to be used in conjunction with Frederic Weisbecker's
1463 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1464 * running CPU-bound user-mode computations.
1465 *
1466 * Offloading of callbacks can also be used as an energy-efficiency
1467 * measure because CPUs with no RCU callbacks queued are more aggressive
1468 * about entering dyntick-idle mode.
1469 */
1470
1471
1472/*
1473 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1474 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1475 * comma-separated list of CPUs and/or CPU ranges.  If an invalid list is
1476 * given, a warning is emitted and all CPUs are offloaded.
1477 */
1478static int __init rcu_nocb_setup(char *str)
1479{
1480	alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1481	if (!strcasecmp(str, "all"))
1482		cpumask_setall(rcu_nocb_mask);
1483	else
1484		if (cpulist_parse(str, rcu_nocb_mask)) {
1485			pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1486			cpumask_setall(rcu_nocb_mask);
1487		}
1488	return 1;
1489}
1490__setup("rcu_nocbs=", rcu_nocb_setup);
1491
1492static int __init parse_rcu_nocb_poll(char *arg)
1493{
1494	rcu_nocb_poll = true;
1495	return 0;
1496}
1497early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1498
1499/*
1500 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1501 * After all, the main point of bypassing is to avoid lock contention
1502 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1503 */
1504int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1505module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1506
1507/*
1508 * Acquire the specified rcu_data structure's ->nocb_bypass_lock.  If the
1509 * lock isn't immediately available, increment ->nocb_lock_contended to
1510 * flag the contention.
1511 */
1512static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
 
1513{
1514	lockdep_assert_irqs_disabled();
1515	if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1516		return;
1517	atomic_inc(&rdp->nocb_lock_contended);
1518	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1519	smp_mb__after_atomic(); /* atomic_inc() before lock. */
1520	raw_spin_lock(&rdp->nocb_bypass_lock);
1521	smp_mb__before_atomic(); /* atomic_dec() after lock. */
1522	atomic_dec(&rdp->nocb_lock_contended);
1523}
1524
1525/*
1526 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1527 * not contended.  Please note that this is extremely special-purpose,
1528 * relying on the fact that at most two kthreads and one CPU contend for
1529 * this lock, and also that the two kthreads are guaranteed to have frequent
1530 * grace-period-duration time intervals between successive acquisitions
1531 * of the lock.  This allows us to use an extremely simple throttling
1532 * mechanism, and further to apply it only to the CPU doing floods of
1533 * call_rcu() invocations.  Don't try this at home!
1534 */
1535static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1536{
1537	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1538	while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1539		cpu_relax();
1540}
1541
1542/*
1543 * Conditionally acquire the specified rcu_data structure's
1544 * ->nocb_bypass_lock.
1545 */
1546static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1547{
1548	lockdep_assert_irqs_disabled();
1549	return raw_spin_trylock(&rdp->nocb_bypass_lock);
1550}
1551
1552/*
1553 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1554 */
1555static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
 
1556{
1557	lockdep_assert_irqs_disabled();
1558	raw_spin_unlock(&rdp->nocb_bypass_lock);
1559}
1560
1561/*
1562 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1563 * if it corresponds to a no-CBs CPU.
1564 */
1565static void rcu_nocb_lock(struct rcu_data *rdp)
1566{
1567	lockdep_assert_irqs_disabled();
1568	if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1569		return;
1570	raw_spin_lock(&rdp->nocb_lock);
1571}
1572
1573/*
1574 * Release the specified rcu_data structure's ->nocb_lock, but only
1575 * if it corresponds to a no-CBs CPU.
1576 */
1577static void rcu_nocb_unlock(struct rcu_data *rdp)
1578{
1579	if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1580		lockdep_assert_irqs_disabled();
1581		raw_spin_unlock(&rdp->nocb_lock);
1582	}
1583}
1584
1585/*
1586 * Release the specified rcu_data structure's ->nocb_lock and restore
1587 * interrupts, but only if it corresponds to a no-CBs CPU.
1588 */
1589static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1590				       unsigned long flags)
1591{
1592	if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1593		lockdep_assert_irqs_disabled();
1594		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1595	} else {
1596		local_irq_restore(flags);
1597	}
1598}
1599
1600/* Lockdep check that ->cblist may be safely accessed. */
1601static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1602{
1603	lockdep_assert_irqs_disabled();
1604	if (rcu_segcblist_is_offloaded(&rdp->cblist) &&
1605	    cpu_online(rdp->cpu))
1606		lockdep_assert_held(&rdp->nocb_lock);
1607}
1608
1609/*
1610 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1611 * grace period.
1612 */
1613static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1614{
1615	swake_up_all(sq);
1616}
1617
1618static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1619{
1620	return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1621}
1622
1623static void rcu_init_one_nocb(struct rcu_node *rnp)
1624{
1625	init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1626	init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1627}
1628
1629/* Is the specified CPU a no-CBs CPU? */
1630bool rcu_is_nocb_cpu(int cpu)
1631{
1632	if (cpumask_available(rcu_nocb_mask))
1633		return cpumask_test_cpu(cpu, rcu_nocb_mask);
1634	return false;
1635}
1636
1637/*
1638 * Kick the GP kthread for this NOCB group.  Caller holds ->nocb_lock
1639 * and this function releases it.
1640 */
1641static void wake_nocb_gp(struct rcu_data *rdp, bool force,
1642			   unsigned long flags)
1643	__releases(rdp->nocb_lock)
1644{
1645	bool needwake = false;
1646	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1647
1648	lockdep_assert_held(&rdp->nocb_lock);
1649	if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1650		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1651				    TPS("AlreadyAwake"));
1652		rcu_nocb_unlock_irqrestore(rdp, flags);
1653		return;
1654	}
1655	del_timer(&rdp->nocb_timer);
1656	rcu_nocb_unlock_irqrestore(rdp, flags);
1657	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1658	if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1659		WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1660		needwake = true;
1661		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1662	}
1663	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1664	if (needwake)
1665		wake_up_process(rdp_gp->nocb_gp_kthread);
1666}
1667
1668/*
1669 * Arrange to wake the GP kthread for this NOCB group at some future
1670 * time when it is safe to do so.
1671 */
1672static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1673			       const char *reason)
1674{
1675	if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1676		mod_timer(&rdp->nocb_timer, jiffies + 1);
1677	if (rdp->nocb_defer_wakeup < waketype)
1678		WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1679	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1680}
1681
1682/*
1683 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1684 * However, if there is a callback to be enqueued and if ->nocb_bypass
1685 * proves to be initially empty, just return false because the no-CB GP
1686 * kthread may need to be awakened in this case.
1687 *
1688 * Note that this function always returns true if rhp is NULL.
1689 */
1690static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1691				     unsigned long j)
1692{
1693	struct rcu_cblist rcl;
1694
1695	WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp->cblist));
1696	rcu_lockdep_assert_cblist_protected(rdp);
1697	lockdep_assert_held(&rdp->nocb_bypass_lock);
1698	if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1699		raw_spin_unlock(&rdp->nocb_bypass_lock);
1700		return false;
1701	}
1702	/* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1703	if (rhp)
1704		rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1705	rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1706	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1707	WRITE_ONCE(rdp->nocb_bypass_first, j);
1708	rcu_nocb_bypass_unlock(rdp);
1709	return true;
1710}
1711
1712/*
1713 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1714 * However, if there is a callback to be enqueued and if ->nocb_bypass
1715 * proves to be initially empty, just return false because the no-CB GP
1716 * kthread may need to be awakened in this case.
1717 *
1718 * Note that this function always returns true if rhp is NULL.
1719 */
1720static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1721				  unsigned long j)
1722{
1723	if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1724		return true;
1725	rcu_lockdep_assert_cblist_protected(rdp);
1726	rcu_nocb_bypass_lock(rdp);
1727	return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1728}
1729
1730/*
1731 * If the ->nocb_bypass_lock is immediately available, flush the
1732 * ->nocb_bypass queue into ->cblist.
1733 */
1734static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1735{
1736	rcu_lockdep_assert_cblist_protected(rdp);
1737	if (!rcu_segcblist_is_offloaded(&rdp->cblist) ||
1738	    !rcu_nocb_bypass_trylock(rdp))
1739		return;
1740	WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1741}
1742
1743/*
1744 * See whether it is appropriate to use the ->nocb_bypass list in order
1745 * to control contention on ->nocb_lock.  A limited number of direct
1746 * enqueues are permitted into ->cblist per jiffy.  If ->nocb_bypass
1747 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1748 * otherwise rcu_barrier() breaks.  Use rcu_nocb_flush_bypass() to switch
1749 * back to direct use of ->cblist.  However, ->nocb_bypass should not be
1750 * used if ->cblist is empty, because otherwise callbacks can be stranded
1751 * on ->nocb_bypass because we cannot count on the current CPU ever again
1752 * invoking call_rcu().  The general rule is that if ->nocb_bypass is
1753 * non-empty, the corresponding no-CBs grace-period kthread must not be
1754 * in an indefinite sleep state.
1755 *
1756 * Finally, it is not permitted to use the bypass during early boot,
1757 * as doing so would confuse the auto-initialization code.  Besides
1758 * which, there is no point in worrying about lock contention while
1759 * there is only one CPU in operation.
1760 */
1761static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1762				bool *was_alldone, unsigned long flags)
1763{
1764	unsigned long c;
1765	unsigned long cur_gp_seq;
1766	unsigned long j = jiffies;
1767	long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1768
1769	if (!rcu_segcblist_is_offloaded(&rdp->cblist)) {
1770		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1771		return false; /* Not offloaded, no bypassing. */
1772	}
1773	lockdep_assert_irqs_disabled();
1774
1775	// Don't use ->nocb_bypass during early boot.
1776	if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1777		rcu_nocb_lock(rdp);
1778		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1779		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1780		return false;
1781	}
1782
1783	// If we have advanced to a new jiffy, reset counts to allow
1784	// moving back from ->nocb_bypass to ->cblist.
1785	if (j == rdp->nocb_nobypass_last) {
1786		c = rdp->nocb_nobypass_count + 1;
1787	} else {
1788		WRITE_ONCE(rdp->nocb_nobypass_last, j);
1789		c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1790		if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1791				 nocb_nobypass_lim_per_jiffy))
1792			c = 0;
1793		else if (c > nocb_nobypass_lim_per_jiffy)
1794			c = nocb_nobypass_lim_per_jiffy;
1795	}
1796	WRITE_ONCE(rdp->nocb_nobypass_count, c);
1797
1798	// If there hasn't yet been all that many ->cblist enqueues
1799	// this jiffy, tell the caller to enqueue onto ->cblist.  But flush
1800	// ->nocb_bypass first.
1801	if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1802		rcu_nocb_lock(rdp);
1803		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1804		if (*was_alldone)
1805			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1806					    TPS("FirstQ"));
1807		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1808		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1809		return false; // Caller must enqueue the callback.
1810	}
1811
1812	// If ->nocb_bypass has been used too long or is too full,
1813	// flush ->nocb_bypass to ->cblist.
1814	if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1815	    ncbs >= qhimark) {
1816		rcu_nocb_lock(rdp);
1817		if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1818			*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1819			if (*was_alldone)
1820				trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1821						    TPS("FirstQ"));
1822			WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1823			return false; // Caller must enqueue the callback.
1824		}
1825		if (j != rdp->nocb_gp_adv_time &&
1826		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1827		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1828			rcu_advance_cbs_nowake(rdp->mynode, rdp);
1829			rdp->nocb_gp_adv_time = j;
1830		}
1831		rcu_nocb_unlock_irqrestore(rdp, flags);
1832		return true; // Callback already enqueued.
1833	}
1834
1835	// We need to use the bypass.
1836	rcu_nocb_wait_contended(rdp);
1837	rcu_nocb_bypass_lock(rdp);
1838	ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1839	rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1840	rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1841	if (!ncbs) {
1842		WRITE_ONCE(rdp->nocb_bypass_first, j);
1843		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1844	}
1845	rcu_nocb_bypass_unlock(rdp);
1846	smp_mb(); /* Order enqueue before wake. */
1847	if (ncbs) {
1848		local_irq_restore(flags);
1849	} else {
1850		// No-CBs GP kthread might be indefinitely asleep, if so, wake.
1851		rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1852		if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1853			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1854					    TPS("FirstBQwake"));
1855			__call_rcu_nocb_wake(rdp, true, flags);
1856		} else {
1857			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1858					    TPS("FirstBQnoWake"));
1859			rcu_nocb_unlock_irqrestore(rdp, flags);
1860		}
1861	}
1862	return true; // Callback already enqueued.
1863}
1864
1865/*
1866 * Awaken the no-CBs grace-period kthead if needed, either due to it
1867 * legitimately being asleep or due to overload conditions.
1868 *
1869 * If warranted, also wake up the kthread servicing this CPUs queues.
1870 */
1871static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1872				 unsigned long flags)
1873				 __releases(rdp->nocb_lock)
1874{
1875	unsigned long cur_gp_seq;
1876	unsigned long j;
1877	long len;
1878	struct task_struct *t;
1879
1880	// If we are being polled or there is no kthread, just leave.
1881	t = READ_ONCE(rdp->nocb_gp_kthread);
1882	if (rcu_nocb_poll || !t) {
1883		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1884				    TPS("WakeNotPoll"));
1885		rcu_nocb_unlock_irqrestore(rdp, flags);
1886		return;
1887	}
1888	// Need to actually to a wakeup.
1889	len = rcu_segcblist_n_cbs(&rdp->cblist);
1890	if (was_alldone) {
1891		rdp->qlen_last_fqs_check = len;
1892		if (!irqs_disabled_flags(flags)) {
1893			/* ... if queue was empty ... */
1894			wake_nocb_gp(rdp, false, flags);
1895			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1896					    TPS("WakeEmpty"));
1897		} else {
1898			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1899					   TPS("WakeEmptyIsDeferred"));
1900			rcu_nocb_unlock_irqrestore(rdp, flags);
1901		}
1902	} else if (len > rdp->qlen_last_fqs_check + qhimark) {
1903		/* ... or if many callbacks queued. */
1904		rdp->qlen_last_fqs_check = len;
1905		j = jiffies;
1906		if (j != rdp->nocb_gp_adv_time &&
1907		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1908		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1909			rcu_advance_cbs_nowake(rdp->mynode, rdp);
1910			rdp->nocb_gp_adv_time = j;
1911		}
1912		smp_mb(); /* Enqueue before timer_pending(). */
1913		if ((rdp->nocb_cb_sleep ||
1914		     !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1915		    !timer_pending(&rdp->nocb_bypass_timer))
1916			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1917					   TPS("WakeOvfIsDeferred"));
1918		rcu_nocb_unlock_irqrestore(rdp, flags);
1919	} else {
1920		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1921		rcu_nocb_unlock_irqrestore(rdp, flags);
1922	}
1923	return;
1924}
1925
1926/* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1927static void do_nocb_bypass_wakeup_timer(struct timer_list *t)
1928{
1929	unsigned long flags;
1930	struct rcu_data *rdp = from_timer(rdp, t, nocb_bypass_timer);
1931
1932	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1933	rcu_nocb_lock_irqsave(rdp, flags);
1934	smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1935	__call_rcu_nocb_wake(rdp, true, flags);
1936}
1937
1938/*
1939 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1940 * or for grace periods to end.
1941 */
1942static void nocb_gp_wait(struct rcu_data *my_rdp)
1943{
1944	bool bypass = false;
1945	long bypass_ncbs;
1946	int __maybe_unused cpu = my_rdp->cpu;
1947	unsigned long cur_gp_seq;
1948	unsigned long flags;
1949	bool gotcbs;
1950	unsigned long j = jiffies;
1951	bool needwait_gp = false; // This prevents actual uninitialized use.
1952	bool needwake;
1953	bool needwake_gp;
1954	struct rcu_data *rdp;
1955	struct rcu_node *rnp;
1956	unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
 
1957
1958	/*
1959	 * Each pass through the following loop checks for CBs and for the
1960	 * nearest grace period (if any) to wait for next.  The CB kthreads
1961	 * and the global grace-period kthread are awakened if needed.
1962	 */
1963	for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1964		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
1965		rcu_nocb_lock_irqsave(rdp, flags);
1966		bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1967		if (bypass_ncbs &&
1968		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
1969		     bypass_ncbs > 2 * qhimark)) {
1970			// Bypass full or old, so flush it.
1971			(void)rcu_nocb_try_flush_bypass(rdp, j);
1972			bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1973		} else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
1974			rcu_nocb_unlock_irqrestore(rdp, flags);
1975			continue; /* No callbacks here, try next. */
1976		}
1977		if (bypass_ncbs) {
1978			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1979					    TPS("Bypass"));
1980			bypass = true;
1981		}
1982		rnp = rdp->mynode;
1983		if (bypass) {  // Avoid race with first bypass CB.
1984			WRITE_ONCE(my_rdp->nocb_defer_wakeup,
1985				   RCU_NOCB_WAKE_NOT);
1986			del_timer(&my_rdp->nocb_timer);
1987		}
1988		// Advance callbacks if helpful and low contention.
1989		needwake_gp = false;
1990		if (!rcu_segcblist_restempty(&rdp->cblist,
1991					     RCU_NEXT_READY_TAIL) ||
1992		    (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1993		     rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
1994			raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
1995			needwake_gp = rcu_advance_cbs(rnp, rdp);
 
 
1996			raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
1997		}
1998		// Need to wait on some grace period?
1999		WARN_ON_ONCE(!rcu_segcblist_restempty(&rdp->cblist,
 
2000						      RCU_NEXT_READY_TAIL));
2001		if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
2002			if (!needwait_gp ||
2003			    ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
2004				wait_gp_seq = cur_gp_seq;
2005			needwait_gp = true;
2006			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2007					    TPS("NeedWaitGP"));
2008		}
2009		if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2010			needwake = rdp->nocb_cb_sleep;
2011			WRITE_ONCE(rdp->nocb_cb_sleep, false);
2012			smp_mb(); /* CB invocation -after- GP end. */
2013		} else {
2014			needwake = false;
2015		}
2016		rcu_nocb_unlock_irqrestore(rdp, flags);
2017		if (needwake) {
2018			swake_up_one(&rdp->nocb_cb_wq);
2019			gotcbs = true;
2020		}
2021		if (needwake_gp)
2022			rcu_gp_kthread_wake();
2023	}
2024
2025	my_rdp->nocb_gp_bypass = bypass;
2026	my_rdp->nocb_gp_gp = needwait_gp;
2027	my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
2028	if (bypass && !rcu_nocb_poll) {
2029		// At least one child with non-empty ->nocb_bypass, so set
2030		// timer in order to avoid stranding its callbacks.
2031		raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2032		mod_timer(&my_rdp->nocb_bypass_timer, j + 2);
2033		raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2034	}
2035	if (rcu_nocb_poll) {
2036		/* Polling, so trace if first poll in the series. */
2037		if (gotcbs)
2038			trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2039		schedule_timeout_interruptible(1);
2040	} else if (!needwait_gp) {
2041		/* Wait for callbacks to appear. */
2042		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2043		swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2044				!READ_ONCE(my_rdp->nocb_gp_sleep));
2045		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2046	} else {
2047		rnp = my_rdp->mynode;
2048		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2049		swait_event_interruptible_exclusive(
2050			rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2051			rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2052			!READ_ONCE(my_rdp->nocb_gp_sleep));
2053		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2054	}
2055	if (!rcu_nocb_poll) {
2056		raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2057		if (bypass)
2058			del_timer(&my_rdp->nocb_bypass_timer);
2059		WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2060		raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2061	}
2062	my_rdp->nocb_gp_seq = -1;
2063	WARN_ON(signal_pending(current));
2064}
2065
2066/*
2067 * No-CBs grace-period-wait kthread.  There is one of these per group
2068 * of CPUs, but only once at least one CPU in that group has come online
2069 * at least once since boot.  This kthread checks for newly posted
2070 * callbacks from any of the CPUs it is responsible for, waits for a
2071 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2072 * that then have callback-invocation work to do.
2073 */
2074static int rcu_nocb_gp_kthread(void *arg)
2075{
2076	struct rcu_data *rdp = arg;
2077
2078	for (;;) {
2079		WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2080		nocb_gp_wait(rdp);
2081		cond_resched_tasks_rcu_qs();
2082	}
2083	return 0;
2084}
2085
2086/*
2087 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2088 * then, if there are no more, wait for more to appear.
2089 */
2090static void nocb_cb_wait(struct rcu_data *rdp)
2091{
2092	unsigned long cur_gp_seq;
2093	unsigned long flags;
2094	bool needwake_gp = false;
2095	struct rcu_node *rnp = rdp->mynode;
2096
2097	local_irq_save(flags);
2098	rcu_momentary_dyntick_idle();
2099	local_irq_restore(flags);
2100	local_bh_disable();
2101	rcu_do_batch(rdp);
2102	local_bh_enable();
2103	lockdep_assert_irqs_enabled();
2104	rcu_nocb_lock_irqsave(rdp, flags);
2105	if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2106	    rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2107	    raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2108		needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2109		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2110	}
2111	if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2112		rcu_nocb_unlock_irqrestore(rdp, flags);
2113		if (needwake_gp)
2114			rcu_gp_kthread_wake();
2115		return;
2116	}
2117
2118	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2119	WRITE_ONCE(rdp->nocb_cb_sleep, true);
2120	rcu_nocb_unlock_irqrestore(rdp, flags);
2121	if (needwake_gp)
2122		rcu_gp_kthread_wake();
2123	swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2124				 !READ_ONCE(rdp->nocb_cb_sleep));
2125	if (!smp_load_acquire(&rdp->nocb_cb_sleep)) { /* VVV */
2126		/* ^^^ Ensure CB invocation follows _sleep test. */
2127		return;
2128	}
2129	WARN_ON(signal_pending(current));
2130	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2131}
2132
2133/*
2134 * Per-rcu_data kthread, but only for no-CBs CPUs.  Repeatedly invoke
2135 * nocb_cb_wait() to do the dirty work.
2136 */
2137static int rcu_nocb_cb_kthread(void *arg)
2138{
2139	struct rcu_data *rdp = arg;
2140
2141	// Each pass through this loop does one callback batch, and,
2142	// if there are no more ready callbacks, waits for them.
2143	for (;;) {
2144		nocb_cb_wait(rdp);
2145		cond_resched_tasks_rcu_qs();
2146	}
2147	return 0;
2148}
2149
2150/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2151static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2152{
2153	return READ_ONCE(rdp->nocb_defer_wakeup);
2154}
2155
2156/* Do a deferred wakeup of rcu_nocb_kthread(). */
2157static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2158{
2159	unsigned long flags;
2160	int ndw;
2161
2162	rcu_nocb_lock_irqsave(rdp, flags);
2163	if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2164		rcu_nocb_unlock_irqrestore(rdp, flags);
2165		return;
2166	}
2167	ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2168	WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2169	wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2170	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2171}
2172
2173/* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2174static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2175{
2176	struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2177
2178	do_nocb_deferred_wakeup_common(rdp);
2179}
2180
2181/*
2182 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2183 * This means we do an inexact common-case check.  Note that if
2184 * we miss, ->nocb_timer will eventually clean things up.
2185 */
2186static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2187{
2188	if (rcu_nocb_need_deferred_wakeup(rdp))
2189		do_nocb_deferred_wakeup_common(rdp);
2190}
2191
2192void __init rcu_init_nohz(void)
2193{
2194	int cpu;
2195	bool need_rcu_nocb_mask = false;
2196	struct rcu_data *rdp;
2197
2198#if defined(CONFIG_NO_HZ_FULL)
2199	if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2200		need_rcu_nocb_mask = true;
2201#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2202
2203	if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2204		if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2205			pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2206			return;
2207		}
2208	}
2209	if (!cpumask_available(rcu_nocb_mask))
2210		return;
2211
2212#if defined(CONFIG_NO_HZ_FULL)
2213	if (tick_nohz_full_running)
2214		cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2215#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2216
2217	if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2218		pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2219		cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2220			    rcu_nocb_mask);
2221	}
2222	if (cpumask_empty(rcu_nocb_mask))
2223		pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2224	else
2225		pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2226			cpumask_pr_args(rcu_nocb_mask));
2227	if (rcu_nocb_poll)
2228		pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2229
2230	for_each_cpu(cpu, rcu_nocb_mask) {
2231		rdp = per_cpu_ptr(&rcu_data, cpu);
2232		if (rcu_segcblist_empty(&rdp->cblist))
2233			rcu_segcblist_init(&rdp->cblist);
2234		rcu_segcblist_offload(&rdp->cblist);
2235	}
2236	rcu_organize_nocb_kthreads();
2237}
2238
2239/* Initialize per-rcu_data variables for no-CBs CPUs. */
2240static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2241{
2242	init_swait_queue_head(&rdp->nocb_cb_wq);
2243	init_swait_queue_head(&rdp->nocb_gp_wq);
2244	raw_spin_lock_init(&rdp->nocb_lock);
2245	raw_spin_lock_init(&rdp->nocb_bypass_lock);
2246	raw_spin_lock_init(&rdp->nocb_gp_lock);
2247	timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2248	timer_setup(&rdp->nocb_bypass_timer, do_nocb_bypass_wakeup_timer, 0);
2249	rcu_cblist_init(&rdp->nocb_bypass);
2250}
2251
2252/*
2253 * If the specified CPU is a no-CBs CPU that does not already have its
2254 * rcuo CB kthread, spawn it.  Additionally, if the rcuo GP kthread
2255 * for this CPU's group has not yet been created, spawn it as well.
2256 */
2257static void rcu_spawn_one_nocb_kthread(int cpu)
2258{
2259	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2260	struct rcu_data *rdp_gp;
2261	struct task_struct *t;
2262
2263	/*
2264	 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2265	 * then nothing to do.
2266	 */
2267	if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2268		return;
2269
2270	/* If we didn't spawn the GP kthread first, reorganize! */
2271	rdp_gp = rdp->nocb_gp_rdp;
2272	if (!rdp_gp->nocb_gp_kthread) {
2273		t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2274				"rcuog/%d", rdp_gp->cpu);
2275		if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2276			return;
2277		WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2278	}
2279
2280	/* Spawn the kthread for this CPU. */
2281	t = kthread_run(rcu_nocb_cb_kthread, rdp,
2282			"rcuo%c/%d", rcu_state.abbr, cpu);
2283	if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2284		return;
2285	WRITE_ONCE(rdp->nocb_cb_kthread, t);
2286	WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2287}
2288
2289/*
2290 * If the specified CPU is a no-CBs CPU that does not already have its
2291 * rcuo kthread, spawn it.
2292 */
2293static void rcu_spawn_cpu_nocb_kthread(int cpu)
2294{
2295	if (rcu_scheduler_fully_active)
2296		rcu_spawn_one_nocb_kthread(cpu);
2297}
2298
2299/*
2300 * Once the scheduler is running, spawn rcuo kthreads for all online
2301 * no-CBs CPUs.  This assumes that the early_initcall()s happen before
2302 * non-boot CPUs come online -- if this changes, we will need to add
2303 * some mutual exclusion.
2304 */
2305static void __init rcu_spawn_nocb_kthreads(void)
2306{
2307	int cpu;
2308
2309	for_each_online_cpu(cpu)
2310		rcu_spawn_cpu_nocb_kthread(cpu);
2311}
2312
2313/* How many CB CPU IDs per GP kthread?  Default of -1 for sqrt(nr_cpu_ids). */
2314static int rcu_nocb_gp_stride = -1;
2315module_param(rcu_nocb_gp_stride, int, 0444);
2316
2317/*
2318 * Initialize GP-CB relationships for all no-CBs CPU.
2319 */
2320static void __init rcu_organize_nocb_kthreads(void)
2321{
2322	int cpu;
2323	bool firsttime = true;
 
 
2324	int ls = rcu_nocb_gp_stride;
2325	int nl = 0;  /* Next GP kthread. */
2326	struct rcu_data *rdp;
2327	struct rcu_data *rdp_gp = NULL;  /* Suppress misguided gcc warn. */
2328	struct rcu_data *rdp_prev = NULL;
2329
2330	if (!cpumask_available(rcu_nocb_mask))
2331		return;
2332	if (ls == -1) {
2333		ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2334		rcu_nocb_gp_stride = ls;
2335	}
2336
2337	/*
2338	 * Each pass through this loop sets up one rcu_data structure.
2339	 * Should the corresponding CPU come online in the future, then
2340	 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2341	 */
2342	for_each_cpu(cpu, rcu_nocb_mask) {
2343		rdp = per_cpu_ptr(&rcu_data, cpu);
2344		if (rdp->cpu >= nl) {
2345			/* New GP kthread, set up for CBs & next GP. */
 
2346			nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2347			rdp->nocb_gp_rdp = rdp;
2348			rdp_gp = rdp;
2349			if (!firsttime && dump_tree)
2350				pr_cont("\n");
2351			firsttime = false;
2352			pr_alert("%s: No-CB GP kthread CPU %d:", __func__, cpu);
 
 
 
 
 
2353		} else {
2354			/* Another CB kthread, link to previous GP kthread. */
 
2355			rdp->nocb_gp_rdp = rdp_gp;
2356			rdp_prev->nocb_next_cb_rdp = rdp;
2357			pr_alert(" %d", cpu);
 
2358		}
2359		rdp_prev = rdp;
2360	}
 
 
2361}
2362
2363/*
2364 * Bind the current task to the offloaded CPUs.  If there are no offloaded
2365 * CPUs, leave the task unbound.  Splat if the bind attempt fails.
2366 */
2367void rcu_bind_current_to_nocb(void)
2368{
2369	if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2370		WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2371}
2372EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2373
2374/*
2375 * Dump out nocb grace-period kthread state for the specified rcu_data
2376 * structure.
2377 */
2378static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2379{
2380	struct rcu_node *rnp = rdp->mynode;
2381
2382	pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2383		rdp->cpu,
2384		"kK"[!!rdp->nocb_gp_kthread],
2385		"lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2386		"dD"[!!rdp->nocb_defer_wakeup],
2387		"tT"[timer_pending(&rdp->nocb_timer)],
2388		"bB"[timer_pending(&rdp->nocb_bypass_timer)],
2389		"sS"[!!rdp->nocb_gp_sleep],
2390		".W"[swait_active(&rdp->nocb_gp_wq)],
2391		".W"[swait_active(&rnp->nocb_gp_wq[0])],
2392		".W"[swait_active(&rnp->nocb_gp_wq[1])],
2393		".B"[!!rdp->nocb_gp_bypass],
2394		".G"[!!rdp->nocb_gp_gp],
2395		(long)rdp->nocb_gp_seq,
2396		rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops));
2397}
2398
2399/* Dump out nocb kthread state for the specified rcu_data structure. */
2400static void show_rcu_nocb_state(struct rcu_data *rdp)
2401{
2402	struct rcu_segcblist *rsclp = &rdp->cblist;
2403	bool waslocked;
2404	bool wastimer;
2405	bool wassleep;
2406
2407	if (rdp->nocb_gp_rdp == rdp)
2408		show_rcu_nocb_gp_state(rdp);
2409
2410	pr_info("   CB %d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%c%c%c q%ld\n",
2411		rdp->cpu, rdp->nocb_gp_rdp->cpu,
2412		"kK"[!!rdp->nocb_cb_kthread],
2413		"bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2414		"cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2415		"lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2416		"sS"[!!rdp->nocb_cb_sleep],
2417		".W"[swait_active(&rdp->nocb_cb_wq)],
2418		jiffies - rdp->nocb_bypass_first,
2419		jiffies - rdp->nocb_nobypass_last,
2420		rdp->nocb_nobypass_count,
2421		".D"[rcu_segcblist_ready_cbs(rsclp)],
2422		".W"[!rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL)],
2423		".R"[!rcu_segcblist_restempty(rsclp, RCU_WAIT_TAIL)],
2424		".N"[!rcu_segcblist_restempty(rsclp, RCU_NEXT_READY_TAIL)],
2425		".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2426		rcu_segcblist_n_cbs(&rdp->cblist));
2427
2428	/* It is OK for GP kthreads to have GP state. */
2429	if (rdp->nocb_gp_rdp == rdp)
2430		return;
2431
2432	waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2433	wastimer = timer_pending(&rdp->nocb_timer);
2434	wassleep = swait_active(&rdp->nocb_gp_wq);
2435	if (!rdp->nocb_defer_wakeup && !rdp->nocb_gp_sleep &&
2436	    !waslocked && !wastimer && !wassleep)
2437		return;  /* Nothing untowards. */
2438
2439	pr_info("   !!! %c%c%c%c %c\n",
2440		"lL"[waslocked],
2441		"dD"[!!rdp->nocb_defer_wakeup],
2442		"tT"[wastimer],
2443		"sS"[!!rdp->nocb_gp_sleep],
2444		".W"[wassleep]);
2445}
2446
2447#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2448
2449/* No ->nocb_lock to acquire.  */
2450static void rcu_nocb_lock(struct rcu_data *rdp)
2451{
2452}
2453
2454/* No ->nocb_lock to release.  */
2455static void rcu_nocb_unlock(struct rcu_data *rdp)
2456{
2457}
2458
2459/* No ->nocb_lock to release.  */
2460static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2461				       unsigned long flags)
2462{
2463	local_irq_restore(flags);
2464}
2465
2466/* Lockdep check that ->cblist may be safely accessed. */
2467static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2468{
2469	lockdep_assert_irqs_disabled();
2470}
2471
2472static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2473{
2474}
2475
2476static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2477{
2478	return NULL;
2479}
2480
2481static void rcu_init_one_nocb(struct rcu_node *rnp)
2482{
2483}
2484
2485static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2486				  unsigned long j)
2487{
2488	return true;
2489}
2490
2491static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2492				bool *was_alldone, unsigned long flags)
2493{
2494	return false;
2495}
2496
2497static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2498				 unsigned long flags)
2499{
2500	WARN_ON_ONCE(1);  /* Should be dead code! */
2501}
2502
2503static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2504{
2505}
2506
2507static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2508{
2509	return false;
2510}
2511
2512static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2513{
2514}
2515
2516static void rcu_spawn_cpu_nocb_kthread(int cpu)
2517{
2518}
2519
2520static void __init rcu_spawn_nocb_kthreads(void)
2521{
2522}
2523
2524static void show_rcu_nocb_state(struct rcu_data *rdp)
2525{
2526}
2527
2528#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2529
2530/*
2531 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2532 * grace-period kthread will do force_quiescent_state() processing?
2533 * The idea is to avoid waking up RCU core processing on such a
2534 * CPU unless the grace period has extended for too long.
2535 *
2536 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2537 * CONFIG_RCU_NOCB_CPU CPUs.
2538 */
2539static bool rcu_nohz_full_cpu(void)
2540{
2541#ifdef CONFIG_NO_HZ_FULL
2542	if (tick_nohz_full_cpu(smp_processor_id()) &&
2543	    (!rcu_gp_in_progress() ||
2544	     ULONG_CMP_LT(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2545		return true;
2546#endif /* #ifdef CONFIG_NO_HZ_FULL */
2547	return false;
2548}
2549
2550/*
2551 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2552 */
2553static void rcu_bind_gp_kthread(void)
2554{
2555	if (!tick_nohz_full_enabled())
2556		return;
2557	housekeeping_affine(current, HK_FLAG_RCU);
2558}
2559
2560/* Record the current task on dyntick-idle entry. */
2561static void rcu_dynticks_task_enter(void)
2562{
2563#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2564	WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2565#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2566}
2567
2568/* Record no current task on dyntick-idle exit. */
2569static void rcu_dynticks_task_exit(void)
2570{
2571#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2572	WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2573#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2574}
v5.9
   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. */
  19#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
  20
  21/*
  22 * Check the RCU kernel configuration parameters and print informative
  23 * messages about anything out of the ordinary.
  24 */
  25static void __init rcu_bootup_announce_oddness(void)
  26{
  27	if (IS_ENABLED(CONFIG_RCU_TRACE))
  28		pr_info("\tRCU event tracing is enabled.\n");
  29	if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
  30	    (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
  31		pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
  32			RCU_FANOUT);
  33	if (rcu_fanout_exact)
  34		pr_info("\tHierarchical RCU autobalancing is disabled.\n");
  35	if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
  36		pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
  37	if (IS_ENABLED(CONFIG_PROVE_RCU))
  38		pr_info("\tRCU lockdep checking is enabled.\n");
  39	if (RCU_NUM_LVLS >= 4)
  40		pr_info("\tFour(or more)-level hierarchy is enabled.\n");
  41	if (RCU_FANOUT_LEAF != 16)
  42		pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
  43			RCU_FANOUT_LEAF);
  44	if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
  45		pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
  46			rcu_fanout_leaf);
  47	if (nr_cpu_ids != NR_CPUS)
  48		pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
  49#ifdef CONFIG_RCU_BOOST
  50	pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
  51		kthread_prio, CONFIG_RCU_BOOST_DELAY);
  52#endif
  53	if (blimit != DEFAULT_RCU_BLIMIT)
  54		pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
  55	if (qhimark != DEFAULT_RCU_QHIMARK)
  56		pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
  57	if (qlowmark != DEFAULT_RCU_QLOMARK)
  58		pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
  59	if (qovld != DEFAULT_RCU_QOVLD)
  60		pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
  61	if (jiffies_till_first_fqs != ULONG_MAX)
  62		pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
  63	if (jiffies_till_next_fqs != ULONG_MAX)
  64		pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
  65	if (jiffies_till_sched_qs != ULONG_MAX)
  66		pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
  67	if (rcu_kick_kthreads)
  68		pr_info("\tKick kthreads if too-long grace period.\n");
  69	if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
  70		pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
  71	if (gp_preinit_delay)
  72		pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
  73	if (gp_init_delay)
  74		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
  75	if (gp_cleanup_delay)
  76		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
  77	if (!use_softirq)
  78		pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
  79	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
  80		pr_info("\tRCU debug extended QS entry/exit.\n");
  81	rcupdate_announce_bootup_oddness();
  82}
  83
  84#ifdef CONFIG_PREEMPT_RCU
  85
  86static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
  87static void rcu_read_unlock_special(struct task_struct *t);
  88
  89/*
  90 * Tell them what RCU they are running.
  91 */
  92static void __init rcu_bootup_announce(void)
  93{
  94	pr_info("Preemptible hierarchical RCU implementation.\n");
  95	rcu_bootup_announce_oddness();
  96}
  97
  98/* Flags for rcu_preempt_ctxt_queue() decision table. */
  99#define RCU_GP_TASKS	0x8
 100#define RCU_EXP_TASKS	0x4
 101#define RCU_GP_BLKD	0x2
 102#define RCU_EXP_BLKD	0x1
 103
 104/*
 105 * Queues a task preempted within an RCU-preempt read-side critical
 106 * section into the appropriate location within the ->blkd_tasks list,
 107 * depending on the states of any ongoing normal and expedited grace
 108 * periods.  The ->gp_tasks pointer indicates which element the normal
 109 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
 110 * indicates which element the expedited grace period is waiting on (again,
 111 * NULL if none).  If a grace period is waiting on a given element in the
 112 * ->blkd_tasks list, it also waits on all subsequent elements.  Thus,
 113 * adding a task to the tail of the list blocks any grace period that is
 114 * already waiting on one of the elements.  In contrast, adding a task
 115 * to the head of the list won't block any grace period that is already
 116 * waiting on one of the elements.
 117 *
 118 * This queuing is imprecise, and can sometimes make an ongoing grace
 119 * period wait for a task that is not strictly speaking blocking it.
 120 * Given the choice, we needlessly block a normal grace period rather than
 121 * blocking an expedited grace period.
 122 *
 123 * Note that an endless sequence of expedited grace periods still cannot
 124 * indefinitely postpone a normal grace period.  Eventually, all of the
 125 * fixed number of preempted tasks blocking the normal grace period that are
 126 * not also blocking the expedited grace period will resume and complete
 127 * their RCU read-side critical sections.  At that point, the ->gp_tasks
 128 * pointer will equal the ->exp_tasks pointer, at which point the end of
 129 * the corresponding expedited grace period will also be the end of the
 130 * normal grace period.
 131 */
 132static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
 133	__releases(rnp->lock) /* But leaves rrupts disabled. */
 134{
 135	int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
 136			 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
 137			 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
 138			 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
 139	struct task_struct *t = current;
 140
 141	raw_lockdep_assert_held_rcu_node(rnp);
 142	WARN_ON_ONCE(rdp->mynode != rnp);
 143	WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 144	/* RCU better not be waiting on newly onlined CPUs! */
 145	WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
 146		     rdp->grpmask);
 147
 148	/*
 149	 * Decide where to queue the newly blocked task.  In theory,
 150	 * this could be an if-statement.  In practice, when I tried
 151	 * that, it was quite messy.
 152	 */
 153	switch (blkd_state) {
 154	case 0:
 155	case                RCU_EXP_TASKS:
 156	case                RCU_EXP_TASKS + RCU_GP_BLKD:
 157	case RCU_GP_TASKS:
 158	case RCU_GP_TASKS + RCU_EXP_TASKS:
 159
 160		/*
 161		 * Blocking neither GP, or first task blocking the normal
 162		 * GP but not blocking the already-waiting expedited GP.
 163		 * Queue at the head of the list to avoid unnecessarily
 164		 * blocking the already-waiting GPs.
 165		 */
 166		list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
 167		break;
 168
 169	case                                              RCU_EXP_BLKD:
 170	case                                RCU_GP_BLKD:
 171	case                                RCU_GP_BLKD + RCU_EXP_BLKD:
 172	case RCU_GP_TASKS +                               RCU_EXP_BLKD:
 173	case RCU_GP_TASKS +                 RCU_GP_BLKD + RCU_EXP_BLKD:
 174	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 175
 176		/*
 177		 * First task arriving that blocks either GP, or first task
 178		 * arriving that blocks the expedited GP (with the normal
 179		 * GP already waiting), or a task arriving that blocks
 180		 * both GPs with both GPs already waiting.  Queue at the
 181		 * tail of the list to avoid any GP waiting on any of the
 182		 * already queued tasks that are not blocking it.
 183		 */
 184		list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
 185		break;
 186
 187	case                RCU_EXP_TASKS +               RCU_EXP_BLKD:
 188	case                RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 189	case RCU_GP_TASKS + RCU_EXP_TASKS +               RCU_EXP_BLKD:
 190
 191		/*
 192		 * Second or subsequent task blocking the expedited GP.
 193		 * The task either does not block the normal GP, or is the
 194		 * first task blocking the normal GP.  Queue just after
 195		 * the first task blocking the expedited GP.
 196		 */
 197		list_add(&t->rcu_node_entry, rnp->exp_tasks);
 198		break;
 199
 200	case RCU_GP_TASKS +                 RCU_GP_BLKD:
 201	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
 202
 203		/*
 204		 * Second or subsequent task blocking the normal GP.
 205		 * The task does not block the expedited GP. Queue just
 206		 * after the first task blocking the normal GP.
 207		 */
 208		list_add(&t->rcu_node_entry, rnp->gp_tasks);
 209		break;
 210
 211	default:
 212
 213		/* Yet another exercise in excessive paranoia. */
 214		WARN_ON_ONCE(1);
 215		break;
 216	}
 217
 218	/*
 219	 * We have now queued the task.  If it was the first one to
 220	 * block either grace period, update the ->gp_tasks and/or
 221	 * ->exp_tasks pointers, respectively, to reference the newly
 222	 * blocked tasks.
 223	 */
 224	if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
 225		WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
 226		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
 227	}
 228	if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
 229		WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
 230	WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
 231		     !(rnp->qsmask & rdp->grpmask));
 232	WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
 233		     !(rnp->expmask & rdp->grpmask));
 234	raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
 235
 236	/*
 237	 * Report the quiescent state for the expedited GP.  This expedited
 238	 * GP should not be able to end until we report, so there should be
 239	 * no need to check for a subsequent expedited GP.  (Though we are
 240	 * still in a quiescent state in any case.)
 241	 */
 242	if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
 243		rcu_report_exp_rdp(rdp);
 244	else
 245		WARN_ON_ONCE(rdp->exp_deferred_qs);
 246}
 247
 248/*
 249 * Record a preemptible-RCU quiescent state for the specified CPU.
 250 * Note that this does not necessarily mean that the task currently running
 251 * on the CPU is in a quiescent state:  Instead, it means that the current
 252 * grace period need not wait on any RCU read-side critical section that
 253 * starts later on this CPU.  It also means that if the current task is
 254 * in an RCU read-side critical section, it has already added itself to
 255 * some leaf rcu_node structure's ->blkd_tasks list.  In addition to the
 256 * current task, there might be any number of other tasks blocked while
 257 * in an RCU read-side critical section.
 258 *
 259 * Callers to this function must disable preemption.
 260 */
 261static void rcu_qs(void)
 262{
 263	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
 264	if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
 265		trace_rcu_grace_period(TPS("rcu_preempt"),
 266				       __this_cpu_read(rcu_data.gp_seq),
 267				       TPS("cpuqs"));
 268		__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 269		barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
 270		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
 271	}
 272}
 273
 274/*
 275 * We have entered the scheduler, and the current task might soon be
 276 * context-switched away from.  If this task is in an RCU read-side
 277 * critical section, we will no longer be able to rely on the CPU to
 278 * record that fact, so we enqueue the task on the blkd_tasks list.
 279 * The task will dequeue itself when it exits the outermost enclosing
 280 * RCU read-side critical section.  Therefore, the current grace period
 281 * cannot be permitted to complete until the blkd_tasks list entries
 282 * predating the current grace period drain, in other words, until
 283 * rnp->gp_tasks becomes NULL.
 284 *
 285 * Caller must disable interrupts.
 286 */
 287void rcu_note_context_switch(bool preempt)
 288{
 289	struct task_struct *t = current;
 290	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 291	struct rcu_node *rnp;
 292
 293	trace_rcu_utilization(TPS("Start context switch"));
 294	lockdep_assert_irqs_disabled();
 295	WARN_ON_ONCE(!preempt && rcu_preempt_depth() > 0);
 296	if (rcu_preempt_depth() > 0 &&
 297	    !t->rcu_read_unlock_special.b.blocked) {
 298
 299		/* Possibly blocking in an RCU read-side critical section. */
 300		rnp = rdp->mynode;
 301		raw_spin_lock_rcu_node(rnp);
 302		t->rcu_read_unlock_special.b.blocked = true;
 303		t->rcu_blocked_node = rnp;
 304
 305		/*
 306		 * Verify the CPU's sanity, trace the preemption, and
 307		 * then queue the task as required based on the states
 308		 * of any ongoing and expedited grace periods.
 309		 */
 310		WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
 311		WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
 312		trace_rcu_preempt_task(rcu_state.name,
 313				       t->pid,
 314				       (rnp->qsmask & rdp->grpmask)
 315				       ? rnp->gp_seq
 316				       : rcu_seq_snap(&rnp->gp_seq));
 317		rcu_preempt_ctxt_queue(rnp, rdp);
 318	} else {
 319		rcu_preempt_deferred_qs(t);
 320	}
 321
 322	/*
 323	 * Either we were not in an RCU read-side critical section to
 324	 * begin with, or we have now recorded that critical section
 325	 * globally.  Either way, we can now note a quiescent state
 326	 * for this CPU.  Again, if we were in an RCU read-side critical
 327	 * section, and if that critical section was blocking the current
 328	 * grace period, then the fact that the task has been enqueued
 329	 * means that we continue to block the current grace period.
 330	 */
 331	rcu_qs();
 332	if (rdp->exp_deferred_qs)
 333		rcu_report_exp_rdp(rdp);
 334	rcu_tasks_qs(current, preempt);
 335	trace_rcu_utilization(TPS("End context switch"));
 336}
 337EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 338
 339/*
 340 * Check for preempted RCU readers blocking the current grace period
 341 * for the specified rcu_node structure.  If the caller needs a reliable
 342 * answer, it must hold the rcu_node's ->lock.
 343 */
 344static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 345{
 346	return READ_ONCE(rnp->gp_tasks) != NULL;
 347}
 348
 349/* limit value for ->rcu_read_lock_nesting. */
 
 
 350#define RCU_NEST_PMAX (INT_MAX / 2)
 351
 352static void rcu_preempt_read_enter(void)
 353{
 354	current->rcu_read_lock_nesting++;
 355}
 356
 357static int rcu_preempt_read_exit(void)
 358{
 359	return --current->rcu_read_lock_nesting;
 360}
 361
 362static void rcu_preempt_depth_set(int val)
 363{
 364	current->rcu_read_lock_nesting = val;
 365}
 366
 367/*
 368 * Preemptible RCU implementation for rcu_read_lock().
 369 * Just increment ->rcu_read_lock_nesting, shared state will be updated
 370 * if we block.
 371 */
 372void __rcu_read_lock(void)
 373{
 374	rcu_preempt_read_enter();
 375	if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 376		WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
 377	barrier();  /* critical section after entry code. */
 378}
 379EXPORT_SYMBOL_GPL(__rcu_read_lock);
 380
 381/*
 382 * Preemptible RCU implementation for rcu_read_unlock().
 383 * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
 384 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
 385 * invoke rcu_read_unlock_special() to clean up after a context switch
 386 * in an RCU read-side critical section and other special cases.
 387 */
 388void __rcu_read_unlock(void)
 389{
 390	struct task_struct *t = current;
 391
 392	if (rcu_preempt_read_exit() == 0) {
 
 
 393		barrier();  /* critical section before exit code. */
 
 
 394		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
 395			rcu_read_unlock_special(t);
 
 
 396	}
 397	if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 398		int rrln = rcu_preempt_depth();
 399
 400		WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
 401	}
 402}
 403EXPORT_SYMBOL_GPL(__rcu_read_unlock);
 404
 405/*
 406 * Advance a ->blkd_tasks-list pointer to the next entry, instead
 407 * returning NULL if at the end of the list.
 408 */
 409static struct list_head *rcu_next_node_entry(struct task_struct *t,
 410					     struct rcu_node *rnp)
 411{
 412	struct list_head *np;
 413
 414	np = t->rcu_node_entry.next;
 415	if (np == &rnp->blkd_tasks)
 416		np = NULL;
 417	return np;
 418}
 419
 420/*
 421 * Return true if the specified rcu_node structure has tasks that were
 422 * preempted within an RCU read-side critical section.
 423 */
 424static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 425{
 426	return !list_empty(&rnp->blkd_tasks);
 427}
 428
 429/*
 430 * Report deferred quiescent states.  The deferral time can
 431 * be quite short, for example, in the case of the call from
 432 * rcu_read_unlock_special().
 433 */
 434static void
 435rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
 436{
 437	bool empty_exp;
 438	bool empty_norm;
 439	bool empty_exp_now;
 440	struct list_head *np;
 441	bool drop_boost_mutex = false;
 442	struct rcu_data *rdp;
 443	struct rcu_node *rnp;
 444	union rcu_special special;
 445
 446	/*
 447	 * If RCU core is waiting for this CPU to exit its critical section,
 448	 * report the fact that it has exited.  Because irqs are disabled,
 449	 * t->rcu_read_unlock_special cannot change.
 450	 */
 451	special = t->rcu_read_unlock_special;
 452	rdp = this_cpu_ptr(&rcu_data);
 453	if (!special.s && !rdp->exp_deferred_qs) {
 454		local_irq_restore(flags);
 455		return;
 456	}
 457	t->rcu_read_unlock_special.s = 0;
 458	if (special.b.need_qs)
 459		rcu_qs();
 
 
 
 
 
 
 460
 461	/*
 462	 * Respond to a request by an expedited grace period for a
 463	 * quiescent state from this CPU.  Note that requests from
 464	 * tasks are handled when removing the task from the
 465	 * blocked-tasks list below.
 466	 */
 467	if (rdp->exp_deferred_qs)
 468		rcu_report_exp_rdp(rdp);
 
 
 
 
 
 469
 470	/* Clean up if blocked during RCU read-side critical section. */
 471	if (special.b.blocked) {
 
 472
 473		/*
 474		 * Remove this task from the list it blocked on.  The task
 475		 * now remains queued on the rcu_node corresponding to the
 476		 * CPU it first blocked on, so there is no longer any need
 477		 * to loop.  Retain a WARN_ON_ONCE() out of sheer paranoia.
 478		 */
 479		rnp = t->rcu_blocked_node;
 480		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
 481		WARN_ON_ONCE(rnp != t->rcu_blocked_node);
 482		WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 483		empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
 484		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
 485			     (!empty_norm || rnp->qsmask));
 486		empty_exp = sync_rcu_exp_done(rnp);
 487		smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
 488		np = rcu_next_node_entry(t, rnp);
 489		list_del_init(&t->rcu_node_entry);
 490		t->rcu_blocked_node = NULL;
 491		trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
 492						rnp->gp_seq, t->pid);
 493		if (&t->rcu_node_entry == rnp->gp_tasks)
 494			WRITE_ONCE(rnp->gp_tasks, np);
 495		if (&t->rcu_node_entry == rnp->exp_tasks)
 496			WRITE_ONCE(rnp->exp_tasks, np);
 497		if (IS_ENABLED(CONFIG_RCU_BOOST)) {
 498			/* Snapshot ->boost_mtx ownership w/rnp->lock held. */
 499			drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
 500			if (&t->rcu_node_entry == rnp->boost_tasks)
 501				WRITE_ONCE(rnp->boost_tasks, np);
 502		}
 503
 504		/*
 505		 * If this was the last task on the current list, and if
 506		 * we aren't waiting on any CPUs, report the quiescent state.
 507		 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
 508		 * so we must take a snapshot of the expedited state.
 509		 */
 510		empty_exp_now = sync_rcu_exp_done(rnp);
 511		if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
 512			trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
 513							 rnp->gp_seq,
 514							 0, rnp->qsmask,
 515							 rnp->level,
 516							 rnp->grplo,
 517							 rnp->grphi,
 518							 !!rnp->gp_tasks);
 519			rcu_report_unblock_qs_rnp(rnp, flags);
 520		} else {
 521			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 522		}
 523
 524		/* Unboost if we were boosted. */
 525		if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
 526			rt_mutex_futex_unlock(&rnp->boost_mtx);
 527
 528		/*
 529		 * If this was the last task on the expedited lists,
 530		 * then we need to report up the rcu_node hierarchy.
 531		 */
 532		if (!empty_exp && empty_exp_now)
 533			rcu_report_exp_rnp(rnp, true);
 534	} else {
 535		local_irq_restore(flags);
 536	}
 537}
 538
 539/*
 540 * Is a deferred quiescent-state pending, and are we also not in
 541 * an RCU read-side critical section?  It is the caller's responsibility
 542 * to ensure it is otherwise safe to report any deferred quiescent
 543 * states.  The reason for this is that it is safe to report a
 544 * quiescent state during context switch even though preemption
 545 * is disabled.  This function cannot be expected to understand these
 546 * nuances, so the caller must handle them.
 547 */
 548static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 549{
 550	return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
 551		READ_ONCE(t->rcu_read_unlock_special.s)) &&
 552	       rcu_preempt_depth() == 0;
 553}
 554
 555/*
 556 * Report a deferred quiescent state if needed and safe to do so.
 557 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
 558 * not being in an RCU read-side critical section.  The caller must
 559 * evaluate safety in terms of interrupt, softirq, and preemption
 560 * disabling.
 561 */
 562static void rcu_preempt_deferred_qs(struct task_struct *t)
 563{
 564	unsigned long flags;
 
 565
 566	if (!rcu_preempt_need_deferred_qs(t))
 567		return;
 
 
 568	local_irq_save(flags);
 569	rcu_preempt_deferred_qs_irqrestore(t, flags);
 
 
 570}
 571
 572/*
 573 * Minimal handler to give the scheduler a chance to re-evaluate.
 574 */
 575static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
 576{
 577	struct rcu_data *rdp;
 578
 579	rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
 580	rdp->defer_qs_iw_pending = false;
 581}
 582
 583/*
 584 * Handle special cases during rcu_read_unlock(), such as needing to
 585 * notify RCU core processing or task having blocked during the RCU
 586 * read-side critical section.
 587 */
 588static void rcu_read_unlock_special(struct task_struct *t)
 589{
 590	unsigned long flags;
 591	bool preempt_bh_were_disabled =
 592			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
 593	bool irqs_were_disabled;
 594
 595	/* NMI handlers cannot block and cannot safely manipulate state. */
 596	if (in_nmi())
 597		return;
 598
 599	local_irq_save(flags);
 600	irqs_were_disabled = irqs_disabled_flags(flags);
 601	if (preempt_bh_were_disabled || irqs_were_disabled) {
 602		bool exp;
 603		struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 604		struct rcu_node *rnp = rdp->mynode;
 605
 606		exp = (t->rcu_blocked_node &&
 607		       READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
 608		      (rdp->grpmask & READ_ONCE(rnp->expmask));
 
 609		// Need to defer quiescent state until everything is enabled.
 610		if (use_softirq && (in_irq() || (exp && !irqs_were_disabled))) {
 611			// Using softirq, safe to awaken, and either the
 612			// wakeup is free or there is an expedited GP.
 
 
 613			raise_softirq_irqoff(RCU_SOFTIRQ);
 614		} else {
 615			// Enabling BH or preempt does reschedule, so...
 616			// Also if no expediting, slow is OK.
 617			// Plus nohz_full CPUs eventually get tick enabled.
 618			set_tsk_need_resched(current);
 619			set_preempt_need_resched();
 620			if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
 621			    !rdp->defer_qs_iw_pending && exp) {
 622				// Get scheduler to re-evaluate and call hooks.
 623				// If !IRQ_WORK, FQS scan will eventually IPI.
 624				init_irq_work(&rdp->defer_qs_iw,
 625					      rcu_preempt_deferred_qs_handler);
 626				rdp->defer_qs_iw_pending = true;
 627				irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
 628			}
 629		}
 
 630		local_irq_restore(flags);
 631		return;
 632	}
 
 633	rcu_preempt_deferred_qs_irqrestore(t, flags);
 634}
 635
 636/*
 637 * Check that the list of blocked tasks for the newly completed grace
 638 * period is in fact empty.  It is a serious bug to complete a grace
 639 * period that still has RCU readers blocked!  This function must be
 640 * invoked -before- updating this rnp's ->gp_seq.
 
 641 *
 642 * Also, if there are blocked tasks on the list, they automatically
 643 * block the newly created grace period, so set up ->gp_tasks accordingly.
 644 */
 645static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 646{
 647	struct task_struct *t;
 648
 649	RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
 650	raw_lockdep_assert_held_rcu_node(rnp);
 651	if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
 652		dump_blkd_tasks(rnp, 10);
 653	if (rcu_preempt_has_tasks(rnp) &&
 654	    (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
 655		WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
 656		t = container_of(rnp->gp_tasks, struct task_struct,
 657				 rcu_node_entry);
 658		trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
 659						rnp->gp_seq, t->pid);
 660	}
 661	WARN_ON_ONCE(rnp->qsmask);
 662}
 663
 664/*
 665 * Check for a quiescent state from the current CPU, including voluntary
 666 * context switches for Tasks RCU.  When a task blocks, the task is
 667 * recorded in the corresponding CPU's rcu_node structure, which is checked
 668 * elsewhere, hence this function need only check for quiescent states
 669 * related to the current CPU, not to those related to tasks.
 670 */
 671static void rcu_flavor_sched_clock_irq(int user)
 672{
 673	struct task_struct *t = current;
 674
 675	if (user || rcu_is_cpu_rrupt_from_idle()) {
 676		rcu_note_voluntary_context_switch(current);
 677	}
 678	if (rcu_preempt_depth() > 0 ||
 679	    (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
 680		/* No QS, force context switch if deferred. */
 681		if (rcu_preempt_need_deferred_qs(t)) {
 682			set_tsk_need_resched(t);
 683			set_preempt_need_resched();
 684		}
 685	} else if (rcu_preempt_need_deferred_qs(t)) {
 686		rcu_preempt_deferred_qs(t); /* Report deferred QS. */
 687		return;
 688	} else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
 689		rcu_qs(); /* Report immediate QS. */
 690		return;
 691	}
 692
 693	/* If GP is oldish, ask for help from rcu_read_unlock_special(). */
 694	if (rcu_preempt_depth() > 0 &&
 695	    __this_cpu_read(rcu_data.core_needs_qs) &&
 696	    __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
 697	    !t->rcu_read_unlock_special.b.need_qs &&
 698	    time_after(jiffies, rcu_state.gp_start + HZ))
 699		t->rcu_read_unlock_special.b.need_qs = true;
 700}
 701
 702/*
 703 * Check for a task exiting while in a preemptible-RCU read-side
 704 * critical section, clean up if so.  No need to issue warnings, as
 705 * debug_check_no_locks_held() already does this if lockdep is enabled.
 706 * Besides, if this function does anything other than just immediately
 707 * return, there was a bug of some sort.  Spewing warnings from this
 708 * function is like as not to simply obscure important prior warnings.
 709 */
 710void exit_rcu(void)
 711{
 712	struct task_struct *t = current;
 713
 714	if (unlikely(!list_empty(&current->rcu_node_entry))) {
 715		rcu_preempt_depth_set(1);
 716		barrier();
 717		WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
 718	} else if (unlikely(rcu_preempt_depth())) {
 719		rcu_preempt_depth_set(1);
 720	} else {
 721		return;
 722	}
 723	__rcu_read_unlock();
 724	rcu_preempt_deferred_qs(current);
 725}
 726
 727/*
 728 * Dump the blocked-tasks state, but limit the list dump to the
 729 * specified number of elements.
 730 */
 731static void
 732dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 733{
 734	int cpu;
 735	int i;
 736	struct list_head *lhp;
 737	bool onl;
 738	struct rcu_data *rdp;
 739	struct rcu_node *rnp1;
 740
 741	raw_lockdep_assert_held_rcu_node(rnp);
 742	pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
 743		__func__, rnp->grplo, rnp->grphi, rnp->level,
 744		(long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
 745	for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
 746		pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
 747			__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
 748	pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
 749		__func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
 750		READ_ONCE(rnp->exp_tasks));
 751	pr_info("%s: ->blkd_tasks", __func__);
 752	i = 0;
 753	list_for_each(lhp, &rnp->blkd_tasks) {
 754		pr_cont(" %p", lhp);
 755		if (++i >= ncheck)
 756			break;
 757	}
 758	pr_cont("\n");
 759	for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
 760		rdp = per_cpu_ptr(&rcu_data, cpu);
 761		onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
 762		pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
 763			cpu, ".o"[onl],
 764			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
 765			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
 766	}
 767}
 768
 769#else /* #ifdef CONFIG_PREEMPT_RCU */
 770
 771/*
 772 * Tell them what RCU they are running.
 773 */
 774static void __init rcu_bootup_announce(void)
 775{
 776	pr_info("Hierarchical RCU implementation.\n");
 777	rcu_bootup_announce_oddness();
 778}
 779
 780/*
 781 * Note a quiescent state for PREEMPTION=n.  Because we do not need to know
 782 * how many quiescent states passed, just if there was at least one since
 783 * the start of the grace period, this just sets a flag.  The caller must
 784 * have disabled preemption.
 785 */
 786static void rcu_qs(void)
 787{
 788	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
 789	if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
 790		return;
 791	trace_rcu_grace_period(TPS("rcu_sched"),
 792			       __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
 793	__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 794	if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
 795		return;
 796	__this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
 797	rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 798}
 799
 800/*
 801 * Register an urgently needed quiescent state.  If there is an
 802 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
 803 * dyntick-idle quiescent state visible to other CPUs, which will in
 804 * some cases serve for expedited as well as normal grace periods.
 805 * Either way, register a lightweight quiescent state.
 806 */
 807void rcu_all_qs(void)
 808{
 809	unsigned long flags;
 810
 811	if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
 812		return;
 813	preempt_disable();
 814	/* Load rcu_urgent_qs before other flags. */
 815	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
 816		preempt_enable();
 817		return;
 818	}
 819	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 820	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
 821		local_irq_save(flags);
 822		rcu_momentary_dyntick_idle();
 823		local_irq_restore(flags);
 824	}
 825	rcu_qs();
 826	preempt_enable();
 827}
 828EXPORT_SYMBOL_GPL(rcu_all_qs);
 829
 830/*
 831 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
 832 */
 833void rcu_note_context_switch(bool preempt)
 834{
 835	trace_rcu_utilization(TPS("Start context switch"));
 836	rcu_qs();
 837	/* Load rcu_urgent_qs before other flags. */
 838	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
 839		goto out;
 840	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 841	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
 842		rcu_momentary_dyntick_idle();
 843	rcu_tasks_qs(current, preempt);
 
 844out:
 845	trace_rcu_utilization(TPS("End context switch"));
 846}
 847EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 848
 849/*
 850 * Because preemptible RCU does not exist, there are never any preempted
 851 * RCU readers.
 852 */
 853static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 854{
 855	return 0;
 856}
 857
 858/*
 859 * Because there is no preemptible RCU, there can be no readers blocked.
 860 */
 861static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 862{
 863	return false;
 864}
 865
 866/*
 867 * Because there is no preemptible RCU, there can be no deferred quiescent
 868 * states.
 869 */
 870static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 871{
 872	return false;
 873}
 874static void rcu_preempt_deferred_qs(struct task_struct *t) { }
 875
 876/*
 877 * Because there is no preemptible RCU, there can be no readers blocked,
 878 * so there is no need to check for blocked tasks.  So check only for
 879 * bogus qsmask values.
 880 */
 881static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 882{
 883	WARN_ON_ONCE(rnp->qsmask);
 884}
 885
 886/*
 887 * Check to see if this CPU is in a non-context-switch quiescent state,
 888 * namely user mode and idle loop.
 889 */
 890static void rcu_flavor_sched_clock_irq(int user)
 891{
 892	if (user || rcu_is_cpu_rrupt_from_idle()) {
 893
 894		/*
 895		 * Get here if this CPU took its interrupt from user
 896		 * mode or from the idle loop, and if this is not a
 897		 * nested interrupt.  In this case, the CPU is in
 898		 * a quiescent state, so note it.
 899		 *
 900		 * No memory barrier is required here because rcu_qs()
 901		 * references only CPU-local variables that other CPUs
 902		 * neither access nor modify, at least not while the
 903		 * corresponding CPU is online.
 904		 */
 905
 906		rcu_qs();
 907	}
 908}
 909
 910/*
 911 * Because preemptible RCU does not exist, tasks cannot possibly exit
 912 * while in preemptible RCU read-side critical sections.
 913 */
 914void exit_rcu(void)
 915{
 916}
 917
 918/*
 919 * Dump the guaranteed-empty blocked-tasks state.  Trust but verify.
 920 */
 921static void
 922dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 923{
 924	WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
 925}
 926
 927#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 928
 929/*
 930 * If boosting, set rcuc kthreads to realtime priority.
 931 */
 932static void rcu_cpu_kthread_setup(unsigned int cpu)
 933{
 934#ifdef CONFIG_RCU_BOOST
 935	struct sched_param sp;
 936
 937	sp.sched_priority = kthread_prio;
 938	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 939#endif /* #ifdef CONFIG_RCU_BOOST */
 940}
 941
 942#ifdef CONFIG_RCU_BOOST
 943
 944/*
 945 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
 946 * or ->boost_tasks, advancing the pointer to the next task in the
 947 * ->blkd_tasks list.
 948 *
 949 * Note that irqs must be enabled: boosting the task can block.
 950 * Returns 1 if there are more tasks needing to be boosted.
 951 */
 952static int rcu_boost(struct rcu_node *rnp)
 953{
 954	unsigned long flags;
 955	struct task_struct *t;
 956	struct list_head *tb;
 957
 958	if (READ_ONCE(rnp->exp_tasks) == NULL &&
 959	    READ_ONCE(rnp->boost_tasks) == NULL)
 960		return 0;  /* Nothing left to boost. */
 961
 962	raw_spin_lock_irqsave_rcu_node(rnp, flags);
 963
 964	/*
 965	 * Recheck under the lock: all tasks in need of boosting
 966	 * might exit their RCU read-side critical sections on their own.
 967	 */
 968	if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
 969		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 970		return 0;
 971	}
 972
 973	/*
 974	 * Preferentially boost tasks blocking expedited grace periods.
 975	 * This cannot starve the normal grace periods because a second
 976	 * expedited grace period must boost all blocked tasks, including
 977	 * those blocking the pre-existing normal grace period.
 978	 */
 979	if (rnp->exp_tasks != NULL)
 980		tb = rnp->exp_tasks;
 981	else
 982		tb = rnp->boost_tasks;
 983
 984	/*
 985	 * We boost task t by manufacturing an rt_mutex that appears to
 986	 * be held by task t.  We leave a pointer to that rt_mutex where
 987	 * task t can find it, and task t will release the mutex when it
 988	 * exits its outermost RCU read-side critical section.  Then
 989	 * simply acquiring this artificial rt_mutex will boost task
 990	 * t's priority.  (Thanks to tglx for suggesting this approach!)
 991	 *
 992	 * Note that task t must acquire rnp->lock to remove itself from
 993	 * the ->blkd_tasks list, which it will do from exit() if from
 994	 * nowhere else.  We therefore are guaranteed that task t will
 995	 * stay around at least until we drop rnp->lock.  Note that
 996	 * rnp->lock also resolves races between our priority boosting
 997	 * and task t's exiting its outermost RCU read-side critical
 998	 * section.
 999	 */
1000	t = container_of(tb, struct task_struct, rcu_node_entry);
1001	rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1002	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1003	/* Lock only for side effect: boosts task t's priority. */
1004	rt_mutex_lock(&rnp->boost_mtx);
1005	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1006
1007	return READ_ONCE(rnp->exp_tasks) != NULL ||
1008	       READ_ONCE(rnp->boost_tasks) != NULL;
1009}
1010
1011/*
1012 * Priority-boosting kthread, one per leaf rcu_node.
1013 */
1014static int rcu_boost_kthread(void *arg)
1015{
1016	struct rcu_node *rnp = (struct rcu_node *)arg;
1017	int spincnt = 0;
1018	int more2boost;
1019
1020	trace_rcu_utilization(TPS("Start boost kthread@init"));
1021	for (;;) {
1022		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1023		trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1024		rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1025			 READ_ONCE(rnp->exp_tasks));
1026		trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1027		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1028		more2boost = rcu_boost(rnp);
1029		if (more2boost)
1030			spincnt++;
1031		else
1032			spincnt = 0;
1033		if (spincnt > 10) {
1034			WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1035			trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1036			schedule_timeout_idle(2);
1037			trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1038			spincnt = 0;
1039		}
1040	}
1041	/* NOTREACHED */
1042	trace_rcu_utilization(TPS("End boost kthread@notreached"));
1043	return 0;
1044}
1045
1046/*
1047 * Check to see if it is time to start boosting RCU readers that are
1048 * blocking the current grace period, and, if so, tell the per-rcu_node
1049 * kthread to start boosting them.  If there is an expedited grace
1050 * period in progress, it is always time to boost.
1051 *
1052 * The caller must hold rnp->lock, which this function releases.
1053 * The ->boost_kthread_task is immortal, so we don't need to worry
1054 * about it going away.
1055 */
1056static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1057	__releases(rnp->lock)
1058{
1059	raw_lockdep_assert_held_rcu_node(rnp);
1060	if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1061		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1062		return;
1063	}
1064	if (rnp->exp_tasks != NULL ||
1065	    (rnp->gp_tasks != NULL &&
1066	     rnp->boost_tasks == NULL &&
1067	     rnp->qsmask == 0 &&
1068	     (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld))) {
1069		if (rnp->exp_tasks == NULL)
1070			WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1071		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1072		rcu_wake_cond(rnp->boost_kthread_task,
1073			      READ_ONCE(rnp->boost_kthread_status));
1074	} else {
1075		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1076	}
1077}
1078
1079/*
1080 * Is the current CPU running the RCU-callbacks kthread?
1081 * Caller must have preemption disabled.
1082 */
1083static bool rcu_is_callbacks_kthread(void)
1084{
1085	return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1086}
1087
1088#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1089
1090/*
1091 * Do priority-boost accounting for the start of a new grace period.
1092 */
1093static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1094{
1095	rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1096}
1097
1098/*
1099 * Create an RCU-boost kthread for the specified node if one does not
1100 * already exist.  We only create this kthread for preemptible RCU.
1101 * Returns zero if all is well, a negated errno otherwise.
1102 */
1103static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1104{
1105	int rnp_index = rnp - rcu_get_root();
1106	unsigned long flags;
1107	struct sched_param sp;
1108	struct task_struct *t;
1109
1110	if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1111		return;
1112
1113	if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1114		return;
1115
1116	rcu_state.boost = 1;
1117
1118	if (rnp->boost_kthread_task != NULL)
1119		return;
1120
1121	t = kthread_create(rcu_boost_kthread, (void *)rnp,
1122			   "rcub/%d", rnp_index);
1123	if (WARN_ON_ONCE(IS_ERR(t)))
1124		return;
1125
1126	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1127	rnp->boost_kthread_task = t;
1128	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1129	sp.sched_priority = kthread_prio;
1130	sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1131	wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1132}
1133
1134/*
1135 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1136 * served by the rcu_node in question.  The CPU hotplug lock is still
1137 * held, so the value of rnp->qsmaskinit will be stable.
1138 *
1139 * We don't include outgoingcpu in the affinity set, use -1 if there is
1140 * no outgoing CPU.  If there are no CPUs left in the affinity set,
1141 * this function allows the kthread to execute on any CPU.
1142 */
1143static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1144{
1145	struct task_struct *t = rnp->boost_kthread_task;
1146	unsigned long mask = rcu_rnp_online_cpus(rnp);
1147	cpumask_var_t cm;
1148	int cpu;
1149
1150	if (!t)
1151		return;
1152	if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1153		return;
1154	for_each_leaf_node_possible_cpu(rnp, cpu)
1155		if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1156		    cpu != outgoingcpu)
1157			cpumask_set_cpu(cpu, cm);
1158	if (cpumask_weight(cm) == 0)
1159		cpumask_setall(cm);
1160	set_cpus_allowed_ptr(t, cm);
1161	free_cpumask_var(cm);
1162}
1163
1164/*
1165 * Spawn boost kthreads -- called as soon as the scheduler is running.
1166 */
1167static void __init rcu_spawn_boost_kthreads(void)
1168{
1169	struct rcu_node *rnp;
1170
1171	rcu_for_each_leaf_node(rnp)
1172		rcu_spawn_one_boost_kthread(rnp);
1173}
1174
1175static void rcu_prepare_kthreads(int cpu)
1176{
1177	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1178	struct rcu_node *rnp = rdp->mynode;
1179
1180	/* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1181	if (rcu_scheduler_fully_active)
1182		rcu_spawn_one_boost_kthread(rnp);
1183}
1184
1185#else /* #ifdef CONFIG_RCU_BOOST */
1186
1187static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1188	__releases(rnp->lock)
1189{
1190	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1191}
1192
1193static bool rcu_is_callbacks_kthread(void)
1194{
1195	return false;
1196}
1197
1198static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1199{
1200}
1201
1202static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1203{
1204}
1205
1206static void __init rcu_spawn_boost_kthreads(void)
1207{
1208}
1209
1210static void rcu_prepare_kthreads(int cpu)
1211{
1212}
1213
1214#endif /* #else #ifdef CONFIG_RCU_BOOST */
1215
1216#if !defined(CONFIG_RCU_FAST_NO_HZ)
1217
1218/*
1219 * Check to see if any future non-offloaded RCU-related work will need
1220 * to be done by the current CPU, even if none need be done immediately,
1221 * returning 1 if so.  This function is part of the RCU implementation;
1222 * it is -not- an exported member of the RCU API.
1223 *
1224 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1225 * CPU has RCU callbacks queued.
1226 */
1227int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1228{
1229	*nextevt = KTIME_MAX;
1230	return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1231	       !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist);
1232}
1233
1234/*
1235 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1236 * after it.
1237 */
1238static void rcu_cleanup_after_idle(void)
1239{
1240}
1241
1242/*
1243 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1244 * is nothing.
1245 */
1246static void rcu_prepare_for_idle(void)
1247{
1248}
1249
1250#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1251
1252/*
1253 * This code is invoked when a CPU goes idle, at which point we want
1254 * to have the CPU do everything required for RCU so that it can enter
1255 * the energy-efficient dyntick-idle mode.
 
1256 *
1257 * The following preprocessor symbol controls this:
1258 *
1259 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1260 *	to sleep in dyntick-idle mode with RCU callbacks pending.  This
1261 *	is sized to be roughly one RCU grace period.  Those energy-efficiency
1262 *	benchmarkers who might otherwise be tempted to set this to a large
1263 *	number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1264 *	system.  And if you are -that- concerned about energy efficiency,
1265 *	just power the system down and be done with it!
 
 
 
1266 *
1267 * The value below works well in practice.  If future workloads require
1268 * adjustment, they can be converted into kernel config parameters, though
1269 * making the state machine smarter might be a better option.
1270 */
1271#define RCU_IDLE_GP_DELAY 4		/* Roughly one grace period. */
 
1272
1273static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1274module_param(rcu_idle_gp_delay, int, 0644);
 
 
1275
1276/*
1277 * Try to advance callbacks on the current CPU, but only if it has been
1278 * awhile since the last time we did so.  Afterwards, if there are any
1279 * callbacks ready for immediate invocation, return true.
1280 */
1281static bool __maybe_unused rcu_try_advance_all_cbs(void)
1282{
1283	bool cbs_ready = false;
1284	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1285	struct rcu_node *rnp;
1286
1287	/* Exit early if we advanced recently. */
1288	if (jiffies == rdp->last_advance_all)
1289		return false;
1290	rdp->last_advance_all = jiffies;
1291
1292	rnp = rdp->mynode;
1293
1294	/*
1295	 * Don't bother checking unless a grace period has
1296	 * completed since we last checked and there are
1297	 * callbacks not yet ready to invoke.
1298	 */
1299	if ((rcu_seq_completed_gp(rdp->gp_seq,
1300				  rcu_seq_current(&rnp->gp_seq)) ||
1301	     unlikely(READ_ONCE(rdp->gpwrap))) &&
1302	    rcu_segcblist_pend_cbs(&rdp->cblist))
1303		note_gp_changes(rdp);
1304
1305	if (rcu_segcblist_ready_cbs(&rdp->cblist))
1306		cbs_ready = true;
1307	return cbs_ready;
1308}
1309
1310/*
1311 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1312 * to invoke.  If the CPU has callbacks, try to advance them.  Tell the
1313 * caller about what to set the timeout.
 
1314 *
1315 * The caller must have disabled interrupts.
1316 */
1317int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1318{
1319	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1320	unsigned long dj;
1321
1322	lockdep_assert_irqs_disabled();
1323
1324	/* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1325	if (rcu_segcblist_empty(&rdp->cblist) ||
1326	    rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist)) {
1327		*nextevt = KTIME_MAX;
1328		return 0;
1329	}
1330
1331	/* Attempt to advance callbacks. */
1332	if (rcu_try_advance_all_cbs()) {
1333		/* Some ready to invoke, so initiate later invocation. */
1334		invoke_rcu_core();
1335		return 1;
1336	}
1337	rdp->last_accelerate = jiffies;
1338
1339	/* Request timer and round. */
1340	dj = round_up(rcu_idle_gp_delay + jiffies, rcu_idle_gp_delay) - jiffies;
1341
 
 
 
 
 
1342	*nextevt = basemono + dj * TICK_NSEC;
1343	return 0;
1344}
1345
1346/*
1347 * Prepare a CPU for idle from an RCU perspective.  The first major task is to
1348 * sense whether nohz mode has been enabled or disabled via sysfs.  The second
1349 * major task is to accelerate (that is, assign grace-period numbers to) any
1350 * recently arrived callbacks.
 
 
1351 *
1352 * The caller must have disabled interrupts.
1353 */
1354static void rcu_prepare_for_idle(void)
1355{
1356	bool needwake;
1357	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1358	struct rcu_node *rnp;
1359	int tne;
1360
1361	lockdep_assert_irqs_disabled();
1362	if (rcu_segcblist_is_offloaded(&rdp->cblist))
1363		return;
1364
1365	/* Handle nohz enablement switches conservatively. */
1366	tne = READ_ONCE(tick_nohz_active);
1367	if (tne != rdp->tick_nohz_enabled_snap) {
1368		if (!rcu_segcblist_empty(&rdp->cblist))
1369			invoke_rcu_core(); /* force nohz to see update. */
1370		rdp->tick_nohz_enabled_snap = tne;
1371		return;
1372	}
1373	if (!tne)
1374		return;
1375
1376	/*
 
 
 
 
 
 
 
 
 
 
 
1377	 * If we have not yet accelerated this jiffy, accelerate all
1378	 * callbacks on this CPU.
1379	 */
1380	if (rdp->last_accelerate == jiffies)
1381		return;
1382	rdp->last_accelerate = jiffies;
1383	if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1384		rnp = rdp->mynode;
1385		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1386		needwake = rcu_accelerate_cbs(rnp, rdp);
1387		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1388		if (needwake)
1389			rcu_gp_kthread_wake();
1390	}
1391}
1392
1393/*
1394 * Clean up for exit from idle.  Attempt to advance callbacks based on
1395 * any grace periods that elapsed while the CPU was idle, and if any
1396 * callbacks are now ready to invoke, initiate invocation.
1397 */
1398static void rcu_cleanup_after_idle(void)
1399{
1400	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1401
1402	lockdep_assert_irqs_disabled();
1403	if (rcu_segcblist_is_offloaded(&rdp->cblist))
1404		return;
1405	if (rcu_try_advance_all_cbs())
1406		invoke_rcu_core();
1407}
1408
1409#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1410
1411#ifdef CONFIG_RCU_NOCB_CPU
1412
1413/*
1414 * Offload callback processing from the boot-time-specified set of CPUs
1415 * specified by rcu_nocb_mask.  For the CPUs in the set, there are kthreads
1416 * created that pull the callbacks from the corresponding CPU, wait for
1417 * a grace period to elapse, and invoke the callbacks.  These kthreads
1418 * are organized into GP kthreads, which manage incoming callbacks, wait for
1419 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1420 * invoke callbacks.  Each GP kthread invokes its own CBs.  The no-CBs CPUs
1421 * do a wake_up() on their GP kthread when they insert a callback into any
1422 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1423 * in which case each kthread actively polls its CPU.  (Which isn't so great
1424 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1425 *
1426 * This is intended to be used in conjunction with Frederic Weisbecker's
1427 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1428 * running CPU-bound user-mode computations.
1429 *
1430 * Offloading of callbacks can also be used as an energy-efficiency
1431 * measure because CPUs with no RCU callbacks queued are more aggressive
1432 * about entering dyntick-idle mode.
1433 */
1434
1435
1436/*
1437 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1438 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1439 * comma-separated list of CPUs and/or CPU ranges.  If an invalid list is
1440 * given, a warning is emitted and all CPUs are offloaded.
1441 */
1442static int __init rcu_nocb_setup(char *str)
1443{
1444	alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1445	if (!strcasecmp(str, "all"))
1446		cpumask_setall(rcu_nocb_mask);
1447	else
1448		if (cpulist_parse(str, rcu_nocb_mask)) {
1449			pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1450			cpumask_setall(rcu_nocb_mask);
1451		}
1452	return 1;
1453}
1454__setup("rcu_nocbs=", rcu_nocb_setup);
1455
1456static int __init parse_rcu_nocb_poll(char *arg)
1457{
1458	rcu_nocb_poll = true;
1459	return 0;
1460}
1461early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1462
1463/*
1464 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1465 * After all, the main point of bypassing is to avoid lock contention
1466 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1467 */
1468int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1469module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1470
1471/*
1472 * Acquire the specified rcu_data structure's ->nocb_bypass_lock.  If the
1473 * lock isn't immediately available, increment ->nocb_lock_contended to
1474 * flag the contention.
1475 */
1476static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1477	__acquires(&rdp->nocb_bypass_lock)
1478{
1479	lockdep_assert_irqs_disabled();
1480	if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1481		return;
1482	atomic_inc(&rdp->nocb_lock_contended);
1483	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1484	smp_mb__after_atomic(); /* atomic_inc() before lock. */
1485	raw_spin_lock(&rdp->nocb_bypass_lock);
1486	smp_mb__before_atomic(); /* atomic_dec() after lock. */
1487	atomic_dec(&rdp->nocb_lock_contended);
1488}
1489
1490/*
1491 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1492 * not contended.  Please note that this is extremely special-purpose,
1493 * relying on the fact that at most two kthreads and one CPU contend for
1494 * this lock, and also that the two kthreads are guaranteed to have frequent
1495 * grace-period-duration time intervals between successive acquisitions
1496 * of the lock.  This allows us to use an extremely simple throttling
1497 * mechanism, and further to apply it only to the CPU doing floods of
1498 * call_rcu() invocations.  Don't try this at home!
1499 */
1500static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1501{
1502	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1503	while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1504		cpu_relax();
1505}
1506
1507/*
1508 * Conditionally acquire the specified rcu_data structure's
1509 * ->nocb_bypass_lock.
1510 */
1511static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1512{
1513	lockdep_assert_irqs_disabled();
1514	return raw_spin_trylock(&rdp->nocb_bypass_lock);
1515}
1516
1517/*
1518 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1519 */
1520static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1521	__releases(&rdp->nocb_bypass_lock)
1522{
1523	lockdep_assert_irqs_disabled();
1524	raw_spin_unlock(&rdp->nocb_bypass_lock);
1525}
1526
1527/*
1528 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1529 * if it corresponds to a no-CBs CPU.
1530 */
1531static void rcu_nocb_lock(struct rcu_data *rdp)
1532{
1533	lockdep_assert_irqs_disabled();
1534	if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1535		return;
1536	raw_spin_lock(&rdp->nocb_lock);
1537}
1538
1539/*
1540 * Release the specified rcu_data structure's ->nocb_lock, but only
1541 * if it corresponds to a no-CBs CPU.
1542 */
1543static void rcu_nocb_unlock(struct rcu_data *rdp)
1544{
1545	if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1546		lockdep_assert_irqs_disabled();
1547		raw_spin_unlock(&rdp->nocb_lock);
1548	}
1549}
1550
1551/*
1552 * Release the specified rcu_data structure's ->nocb_lock and restore
1553 * interrupts, but only if it corresponds to a no-CBs CPU.
1554 */
1555static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1556				       unsigned long flags)
1557{
1558	if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1559		lockdep_assert_irqs_disabled();
1560		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1561	} else {
1562		local_irq_restore(flags);
1563	}
1564}
1565
1566/* Lockdep check that ->cblist may be safely accessed. */
1567static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1568{
1569	lockdep_assert_irqs_disabled();
1570	if (rcu_segcblist_is_offloaded(&rdp->cblist))
 
1571		lockdep_assert_held(&rdp->nocb_lock);
1572}
1573
1574/*
1575 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1576 * grace period.
1577 */
1578static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1579{
1580	swake_up_all(sq);
1581}
1582
1583static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1584{
1585	return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1586}
1587
1588static void rcu_init_one_nocb(struct rcu_node *rnp)
1589{
1590	init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1591	init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1592}
1593
1594/* Is the specified CPU a no-CBs CPU? */
1595bool rcu_is_nocb_cpu(int cpu)
1596{
1597	if (cpumask_available(rcu_nocb_mask))
1598		return cpumask_test_cpu(cpu, rcu_nocb_mask);
1599	return false;
1600}
1601
1602/*
1603 * Kick the GP kthread for this NOCB group.  Caller holds ->nocb_lock
1604 * and this function releases it.
1605 */
1606static void wake_nocb_gp(struct rcu_data *rdp, bool force,
1607			   unsigned long flags)
1608	__releases(rdp->nocb_lock)
1609{
1610	bool needwake = false;
1611	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1612
1613	lockdep_assert_held(&rdp->nocb_lock);
1614	if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1615		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1616				    TPS("AlreadyAwake"));
1617		rcu_nocb_unlock_irqrestore(rdp, flags);
1618		return;
1619	}
1620	del_timer(&rdp->nocb_timer);
1621	rcu_nocb_unlock_irqrestore(rdp, flags);
1622	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1623	if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1624		WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1625		needwake = true;
1626		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1627	}
1628	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1629	if (needwake)
1630		wake_up_process(rdp_gp->nocb_gp_kthread);
1631}
1632
1633/*
1634 * Arrange to wake the GP kthread for this NOCB group at some future
1635 * time when it is safe to do so.
1636 */
1637static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1638			       const char *reason)
1639{
1640	if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1641		mod_timer(&rdp->nocb_timer, jiffies + 1);
1642	if (rdp->nocb_defer_wakeup < waketype)
1643		WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1644	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1645}
1646
1647/*
1648 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1649 * However, if there is a callback to be enqueued and if ->nocb_bypass
1650 * proves to be initially empty, just return false because the no-CB GP
1651 * kthread may need to be awakened in this case.
1652 *
1653 * Note that this function always returns true if rhp is NULL.
1654 */
1655static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1656				     unsigned long j)
1657{
1658	struct rcu_cblist rcl;
1659
1660	WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp->cblist));
1661	rcu_lockdep_assert_cblist_protected(rdp);
1662	lockdep_assert_held(&rdp->nocb_bypass_lock);
1663	if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1664		raw_spin_unlock(&rdp->nocb_bypass_lock);
1665		return false;
1666	}
1667	/* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1668	if (rhp)
1669		rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1670	rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1671	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1672	WRITE_ONCE(rdp->nocb_bypass_first, j);
1673	rcu_nocb_bypass_unlock(rdp);
1674	return true;
1675}
1676
1677/*
1678 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1679 * However, if there is a callback to be enqueued and if ->nocb_bypass
1680 * proves to be initially empty, just return false because the no-CB GP
1681 * kthread may need to be awakened in this case.
1682 *
1683 * Note that this function always returns true if rhp is NULL.
1684 */
1685static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1686				  unsigned long j)
1687{
1688	if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1689		return true;
1690	rcu_lockdep_assert_cblist_protected(rdp);
1691	rcu_nocb_bypass_lock(rdp);
1692	return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1693}
1694
1695/*
1696 * If the ->nocb_bypass_lock is immediately available, flush the
1697 * ->nocb_bypass queue into ->cblist.
1698 */
1699static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1700{
1701	rcu_lockdep_assert_cblist_protected(rdp);
1702	if (!rcu_segcblist_is_offloaded(&rdp->cblist) ||
1703	    !rcu_nocb_bypass_trylock(rdp))
1704		return;
1705	WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1706}
1707
1708/*
1709 * See whether it is appropriate to use the ->nocb_bypass list in order
1710 * to control contention on ->nocb_lock.  A limited number of direct
1711 * enqueues are permitted into ->cblist per jiffy.  If ->nocb_bypass
1712 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1713 * otherwise rcu_barrier() breaks.  Use rcu_nocb_flush_bypass() to switch
1714 * back to direct use of ->cblist.  However, ->nocb_bypass should not be
1715 * used if ->cblist is empty, because otherwise callbacks can be stranded
1716 * on ->nocb_bypass because we cannot count on the current CPU ever again
1717 * invoking call_rcu().  The general rule is that if ->nocb_bypass is
1718 * non-empty, the corresponding no-CBs grace-period kthread must not be
1719 * in an indefinite sleep state.
1720 *
1721 * Finally, it is not permitted to use the bypass during early boot,
1722 * as doing so would confuse the auto-initialization code.  Besides
1723 * which, there is no point in worrying about lock contention while
1724 * there is only one CPU in operation.
1725 */
1726static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1727				bool *was_alldone, unsigned long flags)
1728{
1729	unsigned long c;
1730	unsigned long cur_gp_seq;
1731	unsigned long j = jiffies;
1732	long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1733
1734	if (!rcu_segcblist_is_offloaded(&rdp->cblist)) {
1735		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1736		return false; /* Not offloaded, no bypassing. */
1737	}
1738	lockdep_assert_irqs_disabled();
1739
1740	// Don't use ->nocb_bypass during early boot.
1741	if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1742		rcu_nocb_lock(rdp);
1743		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1744		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1745		return false;
1746	}
1747
1748	// If we have advanced to a new jiffy, reset counts to allow
1749	// moving back from ->nocb_bypass to ->cblist.
1750	if (j == rdp->nocb_nobypass_last) {
1751		c = rdp->nocb_nobypass_count + 1;
1752	} else {
1753		WRITE_ONCE(rdp->nocb_nobypass_last, j);
1754		c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1755		if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1756				 nocb_nobypass_lim_per_jiffy))
1757			c = 0;
1758		else if (c > nocb_nobypass_lim_per_jiffy)
1759			c = nocb_nobypass_lim_per_jiffy;
1760	}
1761	WRITE_ONCE(rdp->nocb_nobypass_count, c);
1762
1763	// If there hasn't yet been all that many ->cblist enqueues
1764	// this jiffy, tell the caller to enqueue onto ->cblist.  But flush
1765	// ->nocb_bypass first.
1766	if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1767		rcu_nocb_lock(rdp);
1768		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1769		if (*was_alldone)
1770			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1771					    TPS("FirstQ"));
1772		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1773		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1774		return false; // Caller must enqueue the callback.
1775	}
1776
1777	// If ->nocb_bypass has been used too long or is too full,
1778	// flush ->nocb_bypass to ->cblist.
1779	if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1780	    ncbs >= qhimark) {
1781		rcu_nocb_lock(rdp);
1782		if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1783			*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1784			if (*was_alldone)
1785				trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1786						    TPS("FirstQ"));
1787			WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1788			return false; // Caller must enqueue the callback.
1789		}
1790		if (j != rdp->nocb_gp_adv_time &&
1791		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1792		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1793			rcu_advance_cbs_nowake(rdp->mynode, rdp);
1794			rdp->nocb_gp_adv_time = j;
1795		}
1796		rcu_nocb_unlock_irqrestore(rdp, flags);
1797		return true; // Callback already enqueued.
1798	}
1799
1800	// We need to use the bypass.
1801	rcu_nocb_wait_contended(rdp);
1802	rcu_nocb_bypass_lock(rdp);
1803	ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1804	rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1805	rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1806	if (!ncbs) {
1807		WRITE_ONCE(rdp->nocb_bypass_first, j);
1808		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1809	}
1810	rcu_nocb_bypass_unlock(rdp);
1811	smp_mb(); /* Order enqueue before wake. */
1812	if (ncbs) {
1813		local_irq_restore(flags);
1814	} else {
1815		// No-CBs GP kthread might be indefinitely asleep, if so, wake.
1816		rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1817		if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1818			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1819					    TPS("FirstBQwake"));
1820			__call_rcu_nocb_wake(rdp, true, flags);
1821		} else {
1822			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1823					    TPS("FirstBQnoWake"));
1824			rcu_nocb_unlock_irqrestore(rdp, flags);
1825		}
1826	}
1827	return true; // Callback already enqueued.
1828}
1829
1830/*
1831 * Awaken the no-CBs grace-period kthead if needed, either due to it
1832 * legitimately being asleep or due to overload conditions.
1833 *
1834 * If warranted, also wake up the kthread servicing this CPUs queues.
1835 */
1836static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1837				 unsigned long flags)
1838				 __releases(rdp->nocb_lock)
1839{
1840	unsigned long cur_gp_seq;
1841	unsigned long j;
1842	long len;
1843	struct task_struct *t;
1844
1845	// If we are being polled or there is no kthread, just leave.
1846	t = READ_ONCE(rdp->nocb_gp_kthread);
1847	if (rcu_nocb_poll || !t) {
1848		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1849				    TPS("WakeNotPoll"));
1850		rcu_nocb_unlock_irqrestore(rdp, flags);
1851		return;
1852	}
1853	// Need to actually to a wakeup.
1854	len = rcu_segcblist_n_cbs(&rdp->cblist);
1855	if (was_alldone) {
1856		rdp->qlen_last_fqs_check = len;
1857		if (!irqs_disabled_flags(flags)) {
1858			/* ... if queue was empty ... */
1859			wake_nocb_gp(rdp, false, flags);
1860			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1861					    TPS("WakeEmpty"));
1862		} else {
1863			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1864					   TPS("WakeEmptyIsDeferred"));
1865			rcu_nocb_unlock_irqrestore(rdp, flags);
1866		}
1867	} else if (len > rdp->qlen_last_fqs_check + qhimark) {
1868		/* ... or if many callbacks queued. */
1869		rdp->qlen_last_fqs_check = len;
1870		j = jiffies;
1871		if (j != rdp->nocb_gp_adv_time &&
1872		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1873		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1874			rcu_advance_cbs_nowake(rdp->mynode, rdp);
1875			rdp->nocb_gp_adv_time = j;
1876		}
1877		smp_mb(); /* Enqueue before timer_pending(). */
1878		if ((rdp->nocb_cb_sleep ||
1879		     !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1880		    !timer_pending(&rdp->nocb_bypass_timer))
1881			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1882					   TPS("WakeOvfIsDeferred"));
1883		rcu_nocb_unlock_irqrestore(rdp, flags);
1884	} else {
1885		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1886		rcu_nocb_unlock_irqrestore(rdp, flags);
1887	}
1888	return;
1889}
1890
1891/* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1892static void do_nocb_bypass_wakeup_timer(struct timer_list *t)
1893{
1894	unsigned long flags;
1895	struct rcu_data *rdp = from_timer(rdp, t, nocb_bypass_timer);
1896
1897	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1898	rcu_nocb_lock_irqsave(rdp, flags);
1899	smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1900	__call_rcu_nocb_wake(rdp, true, flags);
1901}
1902
1903/*
1904 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1905 * or for grace periods to end.
1906 */
1907static void nocb_gp_wait(struct rcu_data *my_rdp)
1908{
1909	bool bypass = false;
1910	long bypass_ncbs;
1911	int __maybe_unused cpu = my_rdp->cpu;
1912	unsigned long cur_gp_seq;
1913	unsigned long flags;
1914	bool gotcbs = false;
1915	unsigned long j = jiffies;
1916	bool needwait_gp = false; // This prevents actual uninitialized use.
1917	bool needwake;
1918	bool needwake_gp;
1919	struct rcu_data *rdp;
1920	struct rcu_node *rnp;
1921	unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
1922	bool wasempty = false;
1923
1924	/*
1925	 * Each pass through the following loop checks for CBs and for the
1926	 * nearest grace period (if any) to wait for next.  The CB kthreads
1927	 * and the global grace-period kthread are awakened if needed.
1928	 */
1929	for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1930		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
1931		rcu_nocb_lock_irqsave(rdp, flags);
1932		bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1933		if (bypass_ncbs &&
1934		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
1935		     bypass_ncbs > 2 * qhimark)) {
1936			// Bypass full or old, so flush it.
1937			(void)rcu_nocb_try_flush_bypass(rdp, j);
1938			bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1939		} else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
1940			rcu_nocb_unlock_irqrestore(rdp, flags);
1941			continue; /* No callbacks here, try next. */
1942		}
1943		if (bypass_ncbs) {
1944			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1945					    TPS("Bypass"));
1946			bypass = true;
1947		}
1948		rnp = rdp->mynode;
1949		if (bypass) {  // Avoid race with first bypass CB.
1950			WRITE_ONCE(my_rdp->nocb_defer_wakeup,
1951				   RCU_NOCB_WAKE_NOT);
1952			del_timer(&my_rdp->nocb_timer);
1953		}
1954		// Advance callbacks if helpful and low contention.
1955		needwake_gp = false;
1956		if (!rcu_segcblist_restempty(&rdp->cblist,
1957					     RCU_NEXT_READY_TAIL) ||
1958		    (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1959		     rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
1960			raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
1961			needwake_gp = rcu_advance_cbs(rnp, rdp);
1962			wasempty = rcu_segcblist_restempty(&rdp->cblist,
1963							   RCU_NEXT_READY_TAIL);
1964			raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
1965		}
1966		// Need to wait on some grace period?
1967		WARN_ON_ONCE(wasempty &&
1968			     !rcu_segcblist_restempty(&rdp->cblist,
1969						      RCU_NEXT_READY_TAIL));
1970		if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
1971			if (!needwait_gp ||
1972			    ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
1973				wait_gp_seq = cur_gp_seq;
1974			needwait_gp = true;
1975			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1976					    TPS("NeedWaitGP"));
1977		}
1978		if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
1979			needwake = rdp->nocb_cb_sleep;
1980			WRITE_ONCE(rdp->nocb_cb_sleep, false);
1981			smp_mb(); /* CB invocation -after- GP end. */
1982		} else {
1983			needwake = false;
1984		}
1985		rcu_nocb_unlock_irqrestore(rdp, flags);
1986		if (needwake) {
1987			swake_up_one(&rdp->nocb_cb_wq);
1988			gotcbs = true;
1989		}
1990		if (needwake_gp)
1991			rcu_gp_kthread_wake();
1992	}
1993
1994	my_rdp->nocb_gp_bypass = bypass;
1995	my_rdp->nocb_gp_gp = needwait_gp;
1996	my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
1997	if (bypass && !rcu_nocb_poll) {
1998		// At least one child with non-empty ->nocb_bypass, so set
1999		// timer in order to avoid stranding its callbacks.
2000		raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2001		mod_timer(&my_rdp->nocb_bypass_timer, j + 2);
2002		raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2003	}
2004	if (rcu_nocb_poll) {
2005		/* Polling, so trace if first poll in the series. */
2006		if (gotcbs)
2007			trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2008		schedule_timeout_idle(1);
2009	} else if (!needwait_gp) {
2010		/* Wait for callbacks to appear. */
2011		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2012		swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2013				!READ_ONCE(my_rdp->nocb_gp_sleep));
2014		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2015	} else {
2016		rnp = my_rdp->mynode;
2017		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2018		swait_event_interruptible_exclusive(
2019			rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2020			rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2021			!READ_ONCE(my_rdp->nocb_gp_sleep));
2022		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2023	}
2024	if (!rcu_nocb_poll) {
2025		raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2026		if (bypass)
2027			del_timer(&my_rdp->nocb_bypass_timer);
2028		WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2029		raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2030	}
2031	my_rdp->nocb_gp_seq = -1;
2032	WARN_ON(signal_pending(current));
2033}
2034
2035/*
2036 * No-CBs grace-period-wait kthread.  There is one of these per group
2037 * of CPUs, but only once at least one CPU in that group has come online
2038 * at least once since boot.  This kthread checks for newly posted
2039 * callbacks from any of the CPUs it is responsible for, waits for a
2040 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2041 * that then have callback-invocation work to do.
2042 */
2043static int rcu_nocb_gp_kthread(void *arg)
2044{
2045	struct rcu_data *rdp = arg;
2046
2047	for (;;) {
2048		WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2049		nocb_gp_wait(rdp);
2050		cond_resched_tasks_rcu_qs();
2051	}
2052	return 0;
2053}
2054
2055/*
2056 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2057 * then, if there are no more, wait for more to appear.
2058 */
2059static void nocb_cb_wait(struct rcu_data *rdp)
2060{
2061	unsigned long cur_gp_seq;
2062	unsigned long flags;
2063	bool needwake_gp = false;
2064	struct rcu_node *rnp = rdp->mynode;
2065
2066	local_irq_save(flags);
2067	rcu_momentary_dyntick_idle();
2068	local_irq_restore(flags);
2069	local_bh_disable();
2070	rcu_do_batch(rdp);
2071	local_bh_enable();
2072	lockdep_assert_irqs_enabled();
2073	rcu_nocb_lock_irqsave(rdp, flags);
2074	if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2075	    rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2076	    raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2077		needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2078		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2079	}
2080	if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2081		rcu_nocb_unlock_irqrestore(rdp, flags);
2082		if (needwake_gp)
2083			rcu_gp_kthread_wake();
2084		return;
2085	}
2086
2087	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2088	WRITE_ONCE(rdp->nocb_cb_sleep, true);
2089	rcu_nocb_unlock_irqrestore(rdp, flags);
2090	if (needwake_gp)
2091		rcu_gp_kthread_wake();
2092	swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2093				 !READ_ONCE(rdp->nocb_cb_sleep));
2094	if (!smp_load_acquire(&rdp->nocb_cb_sleep)) { /* VVV */
2095		/* ^^^ Ensure CB invocation follows _sleep test. */
2096		return;
2097	}
2098	WARN_ON(signal_pending(current));
2099	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2100}
2101
2102/*
2103 * Per-rcu_data kthread, but only for no-CBs CPUs.  Repeatedly invoke
2104 * nocb_cb_wait() to do the dirty work.
2105 */
2106static int rcu_nocb_cb_kthread(void *arg)
2107{
2108	struct rcu_data *rdp = arg;
2109
2110	// Each pass through this loop does one callback batch, and,
2111	// if there are no more ready callbacks, waits for them.
2112	for (;;) {
2113		nocb_cb_wait(rdp);
2114		cond_resched_tasks_rcu_qs();
2115	}
2116	return 0;
2117}
2118
2119/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2120static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2121{
2122	return READ_ONCE(rdp->nocb_defer_wakeup);
2123}
2124
2125/* Do a deferred wakeup of rcu_nocb_kthread(). */
2126static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2127{
2128	unsigned long flags;
2129	int ndw;
2130
2131	rcu_nocb_lock_irqsave(rdp, flags);
2132	if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2133		rcu_nocb_unlock_irqrestore(rdp, flags);
2134		return;
2135	}
2136	ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2137	WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2138	wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2139	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2140}
2141
2142/* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2143static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2144{
2145	struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2146
2147	do_nocb_deferred_wakeup_common(rdp);
2148}
2149
2150/*
2151 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2152 * This means we do an inexact common-case check.  Note that if
2153 * we miss, ->nocb_timer will eventually clean things up.
2154 */
2155static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2156{
2157	if (rcu_nocb_need_deferred_wakeup(rdp))
2158		do_nocb_deferred_wakeup_common(rdp);
2159}
2160
2161void __init rcu_init_nohz(void)
2162{
2163	int cpu;
2164	bool need_rcu_nocb_mask = false;
2165	struct rcu_data *rdp;
2166
2167#if defined(CONFIG_NO_HZ_FULL)
2168	if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2169		need_rcu_nocb_mask = true;
2170#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2171
2172	if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2173		if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2174			pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2175			return;
2176		}
2177	}
2178	if (!cpumask_available(rcu_nocb_mask))
2179		return;
2180
2181#if defined(CONFIG_NO_HZ_FULL)
2182	if (tick_nohz_full_running)
2183		cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2184#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2185
2186	if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2187		pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2188		cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2189			    rcu_nocb_mask);
2190	}
2191	if (cpumask_empty(rcu_nocb_mask))
2192		pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2193	else
2194		pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2195			cpumask_pr_args(rcu_nocb_mask));
2196	if (rcu_nocb_poll)
2197		pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2198
2199	for_each_cpu(cpu, rcu_nocb_mask) {
2200		rdp = per_cpu_ptr(&rcu_data, cpu);
2201		if (rcu_segcblist_empty(&rdp->cblist))
2202			rcu_segcblist_init(&rdp->cblist);
2203		rcu_segcblist_offload(&rdp->cblist);
2204	}
2205	rcu_organize_nocb_kthreads();
2206}
2207
2208/* Initialize per-rcu_data variables for no-CBs CPUs. */
2209static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2210{
2211	init_swait_queue_head(&rdp->nocb_cb_wq);
2212	init_swait_queue_head(&rdp->nocb_gp_wq);
2213	raw_spin_lock_init(&rdp->nocb_lock);
2214	raw_spin_lock_init(&rdp->nocb_bypass_lock);
2215	raw_spin_lock_init(&rdp->nocb_gp_lock);
2216	timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2217	timer_setup(&rdp->nocb_bypass_timer, do_nocb_bypass_wakeup_timer, 0);
2218	rcu_cblist_init(&rdp->nocb_bypass);
2219}
2220
2221/*
2222 * If the specified CPU is a no-CBs CPU that does not already have its
2223 * rcuo CB kthread, spawn it.  Additionally, if the rcuo GP kthread
2224 * for this CPU's group has not yet been created, spawn it as well.
2225 */
2226static void rcu_spawn_one_nocb_kthread(int cpu)
2227{
2228	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2229	struct rcu_data *rdp_gp;
2230	struct task_struct *t;
2231
2232	/*
2233	 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2234	 * then nothing to do.
2235	 */
2236	if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2237		return;
2238
2239	/* If we didn't spawn the GP kthread first, reorganize! */
2240	rdp_gp = rdp->nocb_gp_rdp;
2241	if (!rdp_gp->nocb_gp_kthread) {
2242		t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2243				"rcuog/%d", rdp_gp->cpu);
2244		if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2245			return;
2246		WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2247	}
2248
2249	/* Spawn the kthread for this CPU. */
2250	t = kthread_run(rcu_nocb_cb_kthread, rdp,
2251			"rcuo%c/%d", rcu_state.abbr, cpu);
2252	if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2253		return;
2254	WRITE_ONCE(rdp->nocb_cb_kthread, t);
2255	WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2256}
2257
2258/*
2259 * If the specified CPU is a no-CBs CPU that does not already have its
2260 * rcuo kthread, spawn it.
2261 */
2262static void rcu_spawn_cpu_nocb_kthread(int cpu)
2263{
2264	if (rcu_scheduler_fully_active)
2265		rcu_spawn_one_nocb_kthread(cpu);
2266}
2267
2268/*
2269 * Once the scheduler is running, spawn rcuo kthreads for all online
2270 * no-CBs CPUs.  This assumes that the early_initcall()s happen before
2271 * non-boot CPUs come online -- if this changes, we will need to add
2272 * some mutual exclusion.
2273 */
2274static void __init rcu_spawn_nocb_kthreads(void)
2275{
2276	int cpu;
2277
2278	for_each_online_cpu(cpu)
2279		rcu_spawn_cpu_nocb_kthread(cpu);
2280}
2281
2282/* How many CB CPU IDs per GP kthread?  Default of -1 for sqrt(nr_cpu_ids). */
2283static int rcu_nocb_gp_stride = -1;
2284module_param(rcu_nocb_gp_stride, int, 0444);
2285
2286/*
2287 * Initialize GP-CB relationships for all no-CBs CPU.
2288 */
2289static void __init rcu_organize_nocb_kthreads(void)
2290{
2291	int cpu;
2292	bool firsttime = true;
2293	bool gotnocbs = false;
2294	bool gotnocbscbs = true;
2295	int ls = rcu_nocb_gp_stride;
2296	int nl = 0;  /* Next GP kthread. */
2297	struct rcu_data *rdp;
2298	struct rcu_data *rdp_gp = NULL;  /* Suppress misguided gcc warn. */
2299	struct rcu_data *rdp_prev = NULL;
2300
2301	if (!cpumask_available(rcu_nocb_mask))
2302		return;
2303	if (ls == -1) {
2304		ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2305		rcu_nocb_gp_stride = ls;
2306	}
2307
2308	/*
2309	 * Each pass through this loop sets up one rcu_data structure.
2310	 * Should the corresponding CPU come online in the future, then
2311	 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2312	 */
2313	for_each_cpu(cpu, rcu_nocb_mask) {
2314		rdp = per_cpu_ptr(&rcu_data, cpu);
2315		if (rdp->cpu >= nl) {
2316			/* New GP kthread, set up for CBs & next GP. */
2317			gotnocbs = true;
2318			nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2319			rdp->nocb_gp_rdp = rdp;
2320			rdp_gp = rdp;
2321			if (dump_tree) {
2322				if (!firsttime)
2323					pr_cont("%s\n", gotnocbscbs
2324							? "" : " (self only)");
2325				gotnocbscbs = false;
2326				firsttime = false;
2327				pr_alert("%s: No-CB GP kthread CPU %d:",
2328					 __func__, cpu);
2329			}
2330		} else {
2331			/* Another CB kthread, link to previous GP kthread. */
2332			gotnocbscbs = true;
2333			rdp->nocb_gp_rdp = rdp_gp;
2334			rdp_prev->nocb_next_cb_rdp = rdp;
2335			if (dump_tree)
2336				pr_cont(" %d", cpu);
2337		}
2338		rdp_prev = rdp;
2339	}
2340	if (gotnocbs && dump_tree)
2341		pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
2342}
2343
2344/*
2345 * Bind the current task to the offloaded CPUs.  If there are no offloaded
2346 * CPUs, leave the task unbound.  Splat if the bind attempt fails.
2347 */
2348void rcu_bind_current_to_nocb(void)
2349{
2350	if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2351		WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2352}
2353EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2354
2355/*
2356 * Dump out nocb grace-period kthread state for the specified rcu_data
2357 * structure.
2358 */
2359static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2360{
2361	struct rcu_node *rnp = rdp->mynode;
2362
2363	pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2364		rdp->cpu,
2365		"kK"[!!rdp->nocb_gp_kthread],
2366		"lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2367		"dD"[!!rdp->nocb_defer_wakeup],
2368		"tT"[timer_pending(&rdp->nocb_timer)],
2369		"bB"[timer_pending(&rdp->nocb_bypass_timer)],
2370		"sS"[!!rdp->nocb_gp_sleep],
2371		".W"[swait_active(&rdp->nocb_gp_wq)],
2372		".W"[swait_active(&rnp->nocb_gp_wq[0])],
2373		".W"[swait_active(&rnp->nocb_gp_wq[1])],
2374		".B"[!!rdp->nocb_gp_bypass],
2375		".G"[!!rdp->nocb_gp_gp],
2376		(long)rdp->nocb_gp_seq,
2377		rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops));
2378}
2379
2380/* Dump out nocb kthread state for the specified rcu_data structure. */
2381static void show_rcu_nocb_state(struct rcu_data *rdp)
2382{
2383	struct rcu_segcblist *rsclp = &rdp->cblist;
2384	bool waslocked;
2385	bool wastimer;
2386	bool wassleep;
2387
2388	if (rdp->nocb_gp_rdp == rdp)
2389		show_rcu_nocb_gp_state(rdp);
2390
2391	pr_info("   CB %d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%c%c%c q%ld\n",
2392		rdp->cpu, rdp->nocb_gp_rdp->cpu,
2393		"kK"[!!rdp->nocb_cb_kthread],
2394		"bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2395		"cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2396		"lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2397		"sS"[!!rdp->nocb_cb_sleep],
2398		".W"[swait_active(&rdp->nocb_cb_wq)],
2399		jiffies - rdp->nocb_bypass_first,
2400		jiffies - rdp->nocb_nobypass_last,
2401		rdp->nocb_nobypass_count,
2402		".D"[rcu_segcblist_ready_cbs(rsclp)],
2403		".W"[!rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL)],
2404		".R"[!rcu_segcblist_restempty(rsclp, RCU_WAIT_TAIL)],
2405		".N"[!rcu_segcblist_restempty(rsclp, RCU_NEXT_READY_TAIL)],
2406		".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2407		rcu_segcblist_n_cbs(&rdp->cblist));
2408
2409	/* It is OK for GP kthreads to have GP state. */
2410	if (rdp->nocb_gp_rdp == rdp)
2411		return;
2412
2413	waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2414	wastimer = timer_pending(&rdp->nocb_timer);
2415	wassleep = swait_active(&rdp->nocb_gp_wq);
2416	if (!rdp->nocb_defer_wakeup && !rdp->nocb_gp_sleep &&
2417	    !waslocked && !wastimer && !wassleep)
2418		return;  /* Nothing untowards. */
2419
2420	pr_info("   !!! %c%c%c%c %c\n",
2421		"lL"[waslocked],
2422		"dD"[!!rdp->nocb_defer_wakeup],
2423		"tT"[wastimer],
2424		"sS"[!!rdp->nocb_gp_sleep],
2425		".W"[wassleep]);
2426}
2427
2428#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2429
2430/* No ->nocb_lock to acquire.  */
2431static void rcu_nocb_lock(struct rcu_data *rdp)
2432{
2433}
2434
2435/* No ->nocb_lock to release.  */
2436static void rcu_nocb_unlock(struct rcu_data *rdp)
2437{
2438}
2439
2440/* No ->nocb_lock to release.  */
2441static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2442				       unsigned long flags)
2443{
2444	local_irq_restore(flags);
2445}
2446
2447/* Lockdep check that ->cblist may be safely accessed. */
2448static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2449{
2450	lockdep_assert_irqs_disabled();
2451}
2452
2453static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2454{
2455}
2456
2457static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2458{
2459	return NULL;
2460}
2461
2462static void rcu_init_one_nocb(struct rcu_node *rnp)
2463{
2464}
2465
2466static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2467				  unsigned long j)
2468{
2469	return true;
2470}
2471
2472static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2473				bool *was_alldone, unsigned long flags)
2474{
2475	return false;
2476}
2477
2478static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2479				 unsigned long flags)
2480{
2481	WARN_ON_ONCE(1);  /* Should be dead code! */
2482}
2483
2484static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2485{
2486}
2487
2488static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2489{
2490	return false;
2491}
2492
2493static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2494{
2495}
2496
2497static void rcu_spawn_cpu_nocb_kthread(int cpu)
2498{
2499}
2500
2501static void __init rcu_spawn_nocb_kthreads(void)
2502{
2503}
2504
2505static void show_rcu_nocb_state(struct rcu_data *rdp)
2506{
2507}
2508
2509#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2510
2511/*
2512 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2513 * grace-period kthread will do force_quiescent_state() processing?
2514 * The idea is to avoid waking up RCU core processing on such a
2515 * CPU unless the grace period has extended for too long.
2516 *
2517 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2518 * CONFIG_RCU_NOCB_CPU CPUs.
2519 */
2520static bool rcu_nohz_full_cpu(void)
2521{
2522#ifdef CONFIG_NO_HZ_FULL
2523	if (tick_nohz_full_cpu(smp_processor_id()) &&
2524	    (!rcu_gp_in_progress() ||
2525	     time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2526		return true;
2527#endif /* #ifdef CONFIG_NO_HZ_FULL */
2528	return false;
2529}
2530
2531/*
2532 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2533 */
2534static void rcu_bind_gp_kthread(void)
2535{
2536	if (!tick_nohz_full_enabled())
2537		return;
2538	housekeeping_affine(current, HK_FLAG_RCU);
2539}
2540
2541/* Record the current task on dyntick-idle entry. */
2542static void noinstr rcu_dynticks_task_enter(void)
2543{
2544#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2545	WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2546#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2547}
2548
2549/* Record no current task on dyntick-idle exit. */
2550static void noinstr rcu_dynticks_task_exit(void)
2551{
2552#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2553	WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2554#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2555}
2556
2557/* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
2558static void rcu_dynticks_task_trace_enter(void)
2559{
2560#ifdef CONFIG_TASKS_RCU_TRACE
2561	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2562		current->trc_reader_special.b.need_mb = true;
2563#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */
2564}
2565
2566/* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
2567static void rcu_dynticks_task_trace_exit(void)
2568{
2569#ifdef CONFIG_TASKS_RCU_TRACE
2570	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2571		current->trc_reader_special.b.need_mb = false;
2572#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */
2573}