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
  16static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
  17{
  18	/*
  19	 * In order to read the offloaded state of an rdp in a safe
  20	 * and stable way and prevent from its value to be changed
  21	 * under us, we must either hold the barrier mutex, the cpu
  22	 * hotplug lock (read or write) or the nocb lock. Local
  23	 * non-preemptible reads are also safe. NOCB kthreads and
  24	 * timers have their own means of synchronization against the
  25	 * offloaded state updaters.
  26	 */
  27	RCU_LOCKDEP_WARN(
  28		!(lockdep_is_held(&rcu_state.barrier_mutex) ||
  29		  (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
  30		  rcu_lockdep_is_held_nocb(rdp) ||
  31		  (rdp == this_cpu_ptr(&rcu_data) &&
  32		   !(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible())) ||
  33		  rcu_current_is_nocb_kthread(rdp)),
  34		"Unsafe read of RCU_NOCB offloaded state"
  35	);
  36
  37	return rcu_segcblist_is_offloaded(&rdp->cblist);
  38}
  39
  40/*
  41 * Check the RCU kernel configuration parameters and print informative
  42 * messages about anything out of the ordinary.
  43 */
  44static void __init rcu_bootup_announce_oddness(void)
  45{
  46	if (IS_ENABLED(CONFIG_RCU_TRACE))
  47		pr_info("\tRCU event tracing is enabled.\n");
  48	if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
  49	    (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
  50		pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
  51			RCU_FANOUT);
  52	if (rcu_fanout_exact)
  53		pr_info("\tHierarchical RCU autobalancing is disabled.\n");
  54	if (IS_ENABLED(CONFIG_PROVE_RCU))
  55		pr_info("\tRCU lockdep checking is enabled.\n");
  56	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
  57		pr_info("\tRCU strict (and thus non-scalable) grace periods are enabled.\n");
  58	if (RCU_NUM_LVLS >= 4)
  59		pr_info("\tFour(or more)-level hierarchy is enabled.\n");
  60	if (RCU_FANOUT_LEAF != 16)
  61		pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
  62			RCU_FANOUT_LEAF);
  63	if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
  64		pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
  65			rcu_fanout_leaf);
  66	if (nr_cpu_ids != NR_CPUS)
  67		pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
  68#ifdef CONFIG_RCU_BOOST
  69	pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
  70		kthread_prio, CONFIG_RCU_BOOST_DELAY);
  71#endif
  72	if (blimit != DEFAULT_RCU_BLIMIT)
  73		pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
  74	if (qhimark != DEFAULT_RCU_QHIMARK)
  75		pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
  76	if (qlowmark != DEFAULT_RCU_QLOMARK)
  77		pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
  78	if (qovld != DEFAULT_RCU_QOVLD)
  79		pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
  80	if (jiffies_till_first_fqs != ULONG_MAX)
  81		pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
  82	if (jiffies_till_next_fqs != ULONG_MAX)
  83		pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
  84	if (jiffies_till_sched_qs != ULONG_MAX)
  85		pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
  86	if (rcu_kick_kthreads)
  87		pr_info("\tKick kthreads if too-long grace period.\n");
  88	if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
  89		pr_info("\tRCU callback double-/use-after-free debug is enabled.\n");
  90	if (gp_preinit_delay)
  91		pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
  92	if (gp_init_delay)
  93		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
  94	if (gp_cleanup_delay)
  95		pr_info("\tRCU debug GP cleanup slowdown %d jiffies.\n", gp_cleanup_delay);
  96	if (!use_softirq)
  97		pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
  98	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
  99		pr_info("\tRCU debug extended QS entry/exit.\n");
 100	rcupdate_announce_bootup_oddness();
 101}
 102
 103#ifdef CONFIG_PREEMPT_RCU
 104
 105static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
 106static void rcu_read_unlock_special(struct task_struct *t);
 107
 108/*
 109 * Tell them what RCU they are running.
 110 */
 111static void __init rcu_bootup_announce(void)
 112{
 113	pr_info("Preemptible hierarchical RCU implementation.\n");
 114	rcu_bootup_announce_oddness();
 115}
 116
 117/* Flags for rcu_preempt_ctxt_queue() decision table. */
 118#define RCU_GP_TASKS	0x8
 119#define RCU_EXP_TASKS	0x4
 120#define RCU_GP_BLKD	0x2
 121#define RCU_EXP_BLKD	0x1
 122
 123/*
 124 * Queues a task preempted within an RCU-preempt read-side critical
 125 * section into the appropriate location within the ->blkd_tasks list,
 126 * depending on the states of any ongoing normal and expedited grace
 127 * periods.  The ->gp_tasks pointer indicates which element the normal
 128 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
 129 * indicates which element the expedited grace period is waiting on (again,
 130 * NULL if none).  If a grace period is waiting on a given element in the
 131 * ->blkd_tasks list, it also waits on all subsequent elements.  Thus,
 132 * adding a task to the tail of the list blocks any grace period that is
 133 * already waiting on one of the elements.  In contrast, adding a task
 134 * to the head of the list won't block any grace period that is already
 135 * waiting on one of the elements.
 136 *
 137 * This queuing is imprecise, and can sometimes make an ongoing grace
 138 * period wait for a task that is not strictly speaking blocking it.
 139 * Given the choice, we needlessly block a normal grace period rather than
 140 * blocking an expedited grace period.
 141 *
 142 * Note that an endless sequence of expedited grace periods still cannot
 143 * indefinitely postpone a normal grace period.  Eventually, all of the
 144 * fixed number of preempted tasks blocking the normal grace period that are
 145 * not also blocking the expedited grace period will resume and complete
 146 * their RCU read-side critical sections.  At that point, the ->gp_tasks
 147 * pointer will equal the ->exp_tasks pointer, at which point the end of
 148 * the corresponding expedited grace period will also be the end of the
 149 * normal grace period.
 150 */
 151static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
 152	__releases(rnp->lock) /* But leaves rrupts disabled. */
 153{
 154	int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
 155			 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
 156			 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
 157			 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
 158	struct task_struct *t = current;
 159
 160	raw_lockdep_assert_held_rcu_node(rnp);
 161	WARN_ON_ONCE(rdp->mynode != rnp);
 162	WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 163	/* RCU better not be waiting on newly onlined CPUs! */
 164	WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
 165		     rdp->grpmask);
 166
 167	/*
 168	 * Decide where to queue the newly blocked task.  In theory,
 169	 * this could be an if-statement.  In practice, when I tried
 170	 * that, it was quite messy.
 171	 */
 172	switch (blkd_state) {
 173	case 0:
 174	case                RCU_EXP_TASKS:
 175	case                RCU_EXP_TASKS + RCU_GP_BLKD:
 176	case RCU_GP_TASKS:
 177	case RCU_GP_TASKS + RCU_EXP_TASKS:
 178
 179		/*
 180		 * Blocking neither GP, or first task blocking the normal
 181		 * GP but not blocking the already-waiting expedited GP.
 182		 * Queue at the head of the list to avoid unnecessarily
 183		 * blocking the already-waiting GPs.
 184		 */
 185		list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
 186		break;
 187
 188	case                                              RCU_EXP_BLKD:
 189	case                                RCU_GP_BLKD:
 190	case                                RCU_GP_BLKD + RCU_EXP_BLKD:
 191	case RCU_GP_TASKS +                               RCU_EXP_BLKD:
 192	case RCU_GP_TASKS +                 RCU_GP_BLKD + RCU_EXP_BLKD:
 193	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 194
 195		/*
 196		 * First task arriving that blocks either GP, or first task
 197		 * arriving that blocks the expedited GP (with the normal
 198		 * GP already waiting), or a task arriving that blocks
 199		 * both GPs with both GPs already waiting.  Queue at the
 200		 * tail of the list to avoid any GP waiting on any of the
 201		 * already queued tasks that are not blocking it.
 202		 */
 203		list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
 204		break;
 205
 206	case                RCU_EXP_TASKS +               RCU_EXP_BLKD:
 207	case                RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
 208	case RCU_GP_TASKS + RCU_EXP_TASKS +               RCU_EXP_BLKD:
 209
 210		/*
 211		 * Second or subsequent task blocking the expedited GP.
 212		 * The task either does not block the normal GP, or is the
 213		 * first task blocking the normal GP.  Queue just after
 214		 * the first task blocking the expedited GP.
 215		 */
 216		list_add(&t->rcu_node_entry, rnp->exp_tasks);
 217		break;
 218
 219	case RCU_GP_TASKS +                 RCU_GP_BLKD:
 220	case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
 221
 222		/*
 223		 * Second or subsequent task blocking the normal GP.
 224		 * The task does not block the expedited GP. Queue just
 225		 * after the first task blocking the normal GP.
 226		 */
 227		list_add(&t->rcu_node_entry, rnp->gp_tasks);
 228		break;
 229
 230	default:
 231
 232		/* Yet another exercise in excessive paranoia. */
 233		WARN_ON_ONCE(1);
 234		break;
 235	}
 236
 237	/*
 238	 * We have now queued the task.  If it was the first one to
 239	 * block either grace period, update the ->gp_tasks and/or
 240	 * ->exp_tasks pointers, respectively, to reference the newly
 241	 * blocked tasks.
 242	 */
 243	if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
 244		WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
 245		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
 246	}
 247	if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
 248		WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
 249	WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
 250		     !(rnp->qsmask & rdp->grpmask));
 251	WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
 252		     !(rnp->expmask & rdp->grpmask));
 253	raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
 254
 255	/*
 256	 * Report the quiescent state for the expedited GP.  This expedited
 257	 * GP should not be able to end until we report, so there should be
 258	 * no need to check for a subsequent expedited GP.  (Though we are
 259	 * still in a quiescent state in any case.)
 
 
 260	 */
 261	if (blkd_state & RCU_EXP_BLKD && rdp->cpu_no_qs.b.exp)
 262		rcu_report_exp_rdp(rdp);
 263	else
 264		WARN_ON_ONCE(rdp->cpu_no_qs.b.exp);
 265}
 266
 267/*
 268 * Record a preemptible-RCU quiescent state for the specified CPU.
 269 * Note that this does not necessarily mean that the task currently running
 270 * on the CPU is in a quiescent state:  Instead, it means that the current
 271 * grace period need not wait on any RCU read-side critical section that
 272 * starts later on this CPU.  It also means that if the current task is
 273 * in an RCU read-side critical section, it has already added itself to
 274 * some leaf rcu_node structure's ->blkd_tasks list.  In addition to the
 275 * current task, there might be any number of other tasks blocked while
 276 * in an RCU read-side critical section.
 277 *
 278 * Unlike non-preemptible-RCU, quiescent state reports for expedited
 279 * grace periods are handled separately via deferred quiescent states
 280 * and context switch events.
 281 *
 282 * Callers to this function must disable preemption.
 283 */
 284static void rcu_qs(void)
 285{
 286	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
 287	if (__this_cpu_read(rcu_data.cpu_no_qs.b.norm)) {
 288		trace_rcu_grace_period(TPS("rcu_preempt"),
 289				       __this_cpu_read(rcu_data.gp_seq),
 290				       TPS("cpuqs"));
 291		__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 292		barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
 293		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
 294	}
 295}
 296
 297/*
 298 * We have entered the scheduler, and the current task might soon be
 299 * context-switched away from.  If this task is in an RCU read-side
 300 * critical section, we will no longer be able to rely on the CPU to
 301 * record that fact, so we enqueue the task on the blkd_tasks list.
 302 * The task will dequeue itself when it exits the outermost enclosing
 303 * RCU read-side critical section.  Therefore, the current grace period
 304 * cannot be permitted to complete until the blkd_tasks list entries
 305 * predating the current grace period drain, in other words, until
 306 * rnp->gp_tasks becomes NULL.
 307 *
 308 * Caller must disable interrupts.
 309 */
 310void rcu_note_context_switch(bool preempt)
 311{
 312	struct task_struct *t = current;
 313	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 314	struct rcu_node *rnp;
 315
 316	trace_rcu_utilization(TPS("Start context switch"));
 317	lockdep_assert_irqs_disabled();
 318	WARN_ONCE(!preempt && rcu_preempt_depth() > 0, "Voluntary context switch within RCU read-side critical section!");
 319	if (rcu_preempt_depth() > 0 &&
 320	    !t->rcu_read_unlock_special.b.blocked) {
 321
 322		/* Possibly blocking in an RCU read-side critical section. */
 323		rnp = rdp->mynode;
 324		raw_spin_lock_rcu_node(rnp);
 325		t->rcu_read_unlock_special.b.blocked = true;
 326		t->rcu_blocked_node = rnp;
 327
 328		/*
 329		 * Verify the CPU's sanity, trace the preemption, and
 330		 * then queue the task as required based on the states
 331		 * of any ongoing and expedited grace periods.
 332		 */
 333		WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp));
 334		WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
 335		trace_rcu_preempt_task(rcu_state.name,
 336				       t->pid,
 337				       (rnp->qsmask & rdp->grpmask)
 338				       ? rnp->gp_seq
 339				       : rcu_seq_snap(&rnp->gp_seq));
 340		rcu_preempt_ctxt_queue(rnp, rdp);
 341	} else {
 342		rcu_preempt_deferred_qs(t);
 343	}
 344
 345	/*
 346	 * Either we were not in an RCU read-side critical section to
 347	 * begin with, or we have now recorded that critical section
 348	 * globally.  Either way, we can now note a quiescent state
 349	 * for this CPU.  Again, if we were in an RCU read-side critical
 350	 * section, and if that critical section was blocking the current
 351	 * grace period, then the fact that the task has been enqueued
 352	 * means that we continue to block the current grace period.
 353	 */
 354	rcu_qs();
 355	if (rdp->cpu_no_qs.b.exp)
 356		rcu_report_exp_rdp(rdp);
 357	rcu_tasks_qs(current, preempt);
 358	trace_rcu_utilization(TPS("End context switch"));
 359}
 360EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 361
 362/*
 363 * Check for preempted RCU readers blocking the current grace period
 364 * for the specified rcu_node structure.  If the caller needs a reliable
 365 * answer, it must hold the rcu_node's ->lock.
 366 */
 367static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 368{
 369	return READ_ONCE(rnp->gp_tasks) != NULL;
 370}
 371
 372/* limit value for ->rcu_read_lock_nesting. */
 373#define RCU_NEST_PMAX (INT_MAX / 2)
 374
 375static void rcu_preempt_read_enter(void)
 376{
 377	WRITE_ONCE(current->rcu_read_lock_nesting, READ_ONCE(current->rcu_read_lock_nesting) + 1);
 378}
 379
 380static int rcu_preempt_read_exit(void)
 381{
 382	int ret = READ_ONCE(current->rcu_read_lock_nesting) - 1;
 383
 384	WRITE_ONCE(current->rcu_read_lock_nesting, ret);
 385	return ret;
 386}
 387
 388static void rcu_preempt_depth_set(int val)
 389{
 390	WRITE_ONCE(current->rcu_read_lock_nesting, val);
 391}
 392
 393/*
 394 * Preemptible RCU implementation for rcu_read_lock().
 395 * Just increment ->rcu_read_lock_nesting, shared state will be updated
 396 * if we block.
 397 */
 398void __rcu_read_lock(void)
 399{
 400	rcu_preempt_read_enter();
 401	if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 402		WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
 403	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
 404		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
 405	barrier();  /* critical section after entry code. */
 406}
 407EXPORT_SYMBOL_GPL(__rcu_read_lock);
 408
 409/*
 410 * Preemptible RCU implementation for rcu_read_unlock().
 411 * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
 412 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
 413 * invoke rcu_read_unlock_special() to clean up after a context switch
 414 * in an RCU read-side critical section and other special cases.
 415 */
 416void __rcu_read_unlock(void)
 417{
 418	struct task_struct *t = current;
 419
 420	barrier();  // critical section before exit code.
 421	if (rcu_preempt_read_exit() == 0) {
 422		barrier();  // critical-section exit before .s check.
 423		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
 424			rcu_read_unlock_special(t);
 425	}
 426	if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 427		int rrln = rcu_preempt_depth();
 428
 429		WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
 430	}
 431}
 432EXPORT_SYMBOL_GPL(__rcu_read_unlock);
 433
 434/*
 435 * Advance a ->blkd_tasks-list pointer to the next entry, instead
 436 * returning NULL if at the end of the list.
 437 */
 438static struct list_head *rcu_next_node_entry(struct task_struct *t,
 439					     struct rcu_node *rnp)
 440{
 441	struct list_head *np;
 442
 443	np = t->rcu_node_entry.next;
 444	if (np == &rnp->blkd_tasks)
 445		np = NULL;
 446	return np;
 447}
 448
 449/*
 450 * Return true if the specified rcu_node structure has tasks that were
 451 * preempted within an RCU read-side critical section.
 452 */
 453static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 454{
 455	return !list_empty(&rnp->blkd_tasks);
 456}
 457
 458/*
 459 * Report deferred quiescent states.  The deferral time can
 460 * be quite short, for example, in the case of the call from
 461 * rcu_read_unlock_special().
 462 */
 463static notrace void
 464rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
 465{
 466	bool empty_exp;
 467	bool empty_norm;
 468	bool empty_exp_now;
 469	struct list_head *np;
 470	bool drop_boost_mutex = false;
 471	struct rcu_data *rdp;
 472	struct rcu_node *rnp;
 473	union rcu_special special;
 474
 475	/*
 476	 * If RCU core is waiting for this CPU to exit its critical section,
 477	 * report the fact that it has exited.  Because irqs are disabled,
 478	 * t->rcu_read_unlock_special cannot change.
 479	 */
 480	special = t->rcu_read_unlock_special;
 481	rdp = this_cpu_ptr(&rcu_data);
 482	if (!special.s && !rdp->cpu_no_qs.b.exp) {
 483		local_irq_restore(flags);
 484		return;
 485	}
 486	t->rcu_read_unlock_special.s = 0;
 487	if (special.b.need_qs) {
 488		if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
 489			rdp->cpu_no_qs.b.norm = false;
 490			rcu_report_qs_rdp(rdp);
 491			udelay(rcu_unlock_delay);
 492		} else {
 493			rcu_qs();
 494		}
 495	}
 496
 497	/*
 498	 * Respond to a request by an expedited grace period for a
 499	 * quiescent state from this CPU.  Note that requests from
 500	 * tasks are handled when removing the task from the
 501	 * blocked-tasks list below.
 502	 */
 503	if (rdp->cpu_no_qs.b.exp)
 504		rcu_report_exp_rdp(rdp);
 505
 506	/* Clean up if blocked during RCU read-side critical section. */
 507	if (special.b.blocked) {
 508
 509		/*
 510		 * Remove this task from the list it blocked on.  The task
 511		 * now remains queued on the rcu_node corresponding to the
 512		 * CPU it first blocked on, so there is no longer any need
 513		 * to loop.  Retain a WARN_ON_ONCE() out of sheer paranoia.
 514		 */
 515		rnp = t->rcu_blocked_node;
 516		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
 517		WARN_ON_ONCE(rnp != t->rcu_blocked_node);
 518		WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 519		empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
 520		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
 521			     (!empty_norm || rnp->qsmask));
 522		empty_exp = sync_rcu_exp_done(rnp);
 523		smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
 524		np = rcu_next_node_entry(t, rnp);
 525		list_del_init(&t->rcu_node_entry);
 526		t->rcu_blocked_node = NULL;
 527		trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
 528						rnp->gp_seq, t->pid);
 529		if (&t->rcu_node_entry == rnp->gp_tasks)
 530			WRITE_ONCE(rnp->gp_tasks, np);
 531		if (&t->rcu_node_entry == rnp->exp_tasks)
 532			WRITE_ONCE(rnp->exp_tasks, np);
 533		if (IS_ENABLED(CONFIG_RCU_BOOST)) {
 534			/* Snapshot ->boost_mtx ownership w/rnp->lock held. */
 535			drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx.rtmutex) == t;
 536			if (&t->rcu_node_entry == rnp->boost_tasks)
 537				WRITE_ONCE(rnp->boost_tasks, np);
 538		}
 539
 540		/*
 541		 * If this was the last task on the current list, and if
 542		 * we aren't waiting on any CPUs, report the quiescent state.
 543		 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
 544		 * so we must take a snapshot of the expedited state.
 545		 */
 546		empty_exp_now = sync_rcu_exp_done(rnp);
 547		if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
 548			trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
 549							 rnp->gp_seq,
 550							 0, rnp->qsmask,
 551							 rnp->level,
 552							 rnp->grplo,
 553							 rnp->grphi,
 554							 !!rnp->gp_tasks);
 555			rcu_report_unblock_qs_rnp(rnp, flags);
 556		} else {
 557			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 558		}
 559
 560		/*
 561		 * If this was the last task on the expedited lists,
 562		 * then we need to report up the rcu_node hierarchy.
 563		 */
 564		if (!empty_exp && empty_exp_now)
 565			rcu_report_exp_rnp(rnp, true);
 566
 567		/* Unboost if we were boosted. */
 568		if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
 569			rt_mutex_futex_unlock(&rnp->boost_mtx.rtmutex);
 570	} else {
 571		local_irq_restore(flags);
 572	}
 573}
 574
 575/*
 576 * Is a deferred quiescent-state pending, and are we also not in
 577 * an RCU read-side critical section?  It is the caller's responsibility
 578 * to ensure it is otherwise safe to report any deferred quiescent
 579 * states.  The reason for this is that it is safe to report a
 580 * quiescent state during context switch even though preemption
 581 * is disabled.  This function cannot be expected to understand these
 582 * nuances, so the caller must handle them.
 583 */
 584static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 585{
 586	return (__this_cpu_read(rcu_data.cpu_no_qs.b.exp) ||
 587		READ_ONCE(t->rcu_read_unlock_special.s)) &&
 588	       rcu_preempt_depth() == 0;
 589}
 590
 591/*
 592 * Report a deferred quiescent state if needed and safe to do so.
 593 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
 594 * not being in an RCU read-side critical section.  The caller must
 595 * evaluate safety in terms of interrupt, softirq, and preemption
 596 * disabling.
 597 */
 598notrace void rcu_preempt_deferred_qs(struct task_struct *t)
 599{
 600	unsigned long flags;
 601
 602	if (!rcu_preempt_need_deferred_qs(t))
 603		return;
 604	local_irq_save(flags);
 605	rcu_preempt_deferred_qs_irqrestore(t, flags);
 606}
 607
 608/*
 609 * Minimal handler to give the scheduler a chance to re-evaluate.
 610 */
 611static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
 612{
 613	struct rcu_data *rdp;
 614
 615	rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
 616	rdp->defer_qs_iw_pending = false;
 617}
 618
 619/*
 620 * Handle special cases during rcu_read_unlock(), such as needing to
 621 * notify RCU core processing or task having blocked during the RCU
 622 * read-side critical section.
 623 */
 624static void rcu_read_unlock_special(struct task_struct *t)
 625{
 626	unsigned long flags;
 627	bool irqs_were_disabled;
 628	bool preempt_bh_were_disabled =
 629			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
 630
 631	/* NMI handlers cannot block and cannot safely manipulate state. */
 632	if (in_nmi())
 633		return;
 634
 635	local_irq_save(flags);
 636	irqs_were_disabled = irqs_disabled_flags(flags);
 637	if (preempt_bh_were_disabled || irqs_were_disabled) {
 638		bool expboost; // Expedited GP in flight or possible boosting.
 639		struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 640		struct rcu_node *rnp = rdp->mynode;
 641
 642		expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
 643			   (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
 644			   (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
 645			   ((rdp->grpmask & READ_ONCE(rnp->qsmask)) || t->rcu_blocked_node)) ||
 646			   (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
 647			    t->rcu_blocked_node);
 648		// Need to defer quiescent state until everything is enabled.
 649		if (use_softirq && (in_hardirq() || (expboost && !irqs_were_disabled))) {
 650			// Using softirq, safe to awaken, and either the
 651			// wakeup is free or there is either an expedited
 652			// GP in flight or a potential need to deboost.
 653			raise_softirq_irqoff(RCU_SOFTIRQ);
 654		} else {
 655			// Enabling BH or preempt does reschedule, so...
 656			// Also if no expediting and no possible deboosting,
 657			// slow is OK.  Plus nohz_full CPUs eventually get
 658			// tick enabled.
 659			set_tsk_need_resched(current);
 660			set_preempt_need_resched();
 661			if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
 662			    expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
 663				// Get scheduler to re-evaluate and call hooks.
 664				// If !IRQ_WORK, FQS scan will eventually IPI.
 665				if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
 666				    IS_ENABLED(CONFIG_PREEMPT_RT))
 667					rdp->defer_qs_iw = IRQ_WORK_INIT_HARD(
 668								rcu_preempt_deferred_qs_handler);
 669				else
 670					init_irq_work(&rdp->defer_qs_iw,
 671						      rcu_preempt_deferred_qs_handler);
 672				rdp->defer_qs_iw_pending = true;
 673				irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
 674			}
 675		}
 676		local_irq_restore(flags);
 677		return;
 678	}
 679	rcu_preempt_deferred_qs_irqrestore(t, flags);
 680}
 681
 682/*
 683 * Check that the list of blocked tasks for the newly completed grace
 684 * period is in fact empty.  It is a serious bug to complete a grace
 685 * period that still has RCU readers blocked!  This function must be
 686 * invoked -before- updating this rnp's ->gp_seq.
 687 *
 688 * Also, if there are blocked tasks on the list, they automatically
 689 * block the newly created grace period, so set up ->gp_tasks accordingly.
 690 */
 691static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 692{
 693	struct task_struct *t;
 694
 695	RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
 696	raw_lockdep_assert_held_rcu_node(rnp);
 697	if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
 698		dump_blkd_tasks(rnp, 10);
 699	if (rcu_preempt_has_tasks(rnp) &&
 700	    (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
 701		WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
 702		t = container_of(rnp->gp_tasks, struct task_struct,
 703				 rcu_node_entry);
 704		trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
 705						rnp->gp_seq, t->pid);
 706	}
 707	WARN_ON_ONCE(rnp->qsmask);
 708}
 709
 710/*
 711 * Check for a quiescent state from the current CPU, including voluntary
 712 * context switches for Tasks RCU.  When a task blocks, the task is
 713 * recorded in the corresponding CPU's rcu_node structure, which is checked
 714 * elsewhere, hence this function need only check for quiescent states
 715 * related to the current CPU, not to those related to tasks.
 716 */
 717static void rcu_flavor_sched_clock_irq(int user)
 718{
 719	struct task_struct *t = current;
 720
 721	lockdep_assert_irqs_disabled();
 722	if (rcu_preempt_depth() > 0 ||
 723	    (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
 724		/* No QS, force context switch if deferred. */
 725		if (rcu_preempt_need_deferred_qs(t)) {
 726			set_tsk_need_resched(t);
 727			set_preempt_need_resched();
 728		}
 729	} else if (rcu_preempt_need_deferred_qs(t)) {
 730		rcu_preempt_deferred_qs(t); /* Report deferred QS. */
 731		return;
 732	} else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
 733		rcu_qs(); /* Report immediate QS. */
 734		return;
 735	}
 736
 737	/* If GP is oldish, ask for help from rcu_read_unlock_special(). */
 738	if (rcu_preempt_depth() > 0 &&
 739	    __this_cpu_read(rcu_data.core_needs_qs) &&
 740	    __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
 741	    !t->rcu_read_unlock_special.b.need_qs &&
 742	    time_after(jiffies, rcu_state.gp_start + HZ))
 743		t->rcu_read_unlock_special.b.need_qs = true;
 744}
 745
 746/*
 747 * Check for a task exiting while in a preemptible-RCU read-side
 748 * critical section, clean up if so.  No need to issue warnings, as
 749 * debug_check_no_locks_held() already does this if lockdep is enabled.
 750 * Besides, if this function does anything other than just immediately
 751 * return, there was a bug of some sort.  Spewing warnings from this
 752 * function is like as not to simply obscure important prior warnings.
 753 */
 754void exit_rcu(void)
 755{
 756	struct task_struct *t = current;
 757
 758	if (unlikely(!list_empty(&current->rcu_node_entry))) {
 759		rcu_preempt_depth_set(1);
 760		barrier();
 761		WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
 762	} else if (unlikely(rcu_preempt_depth())) {
 763		rcu_preempt_depth_set(1);
 764	} else {
 765		return;
 766	}
 767	__rcu_read_unlock();
 768	rcu_preempt_deferred_qs(current);
 769}
 770
 771/*
 772 * Dump the blocked-tasks state, but limit the list dump to the
 773 * specified number of elements.
 774 */
 775static void
 776dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 777{
 778	int cpu;
 779	int i;
 780	struct list_head *lhp;
 781	struct rcu_data *rdp;
 782	struct rcu_node *rnp1;
 783
 784	raw_lockdep_assert_held_rcu_node(rnp);
 785	pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
 786		__func__, rnp->grplo, rnp->grphi, rnp->level,
 787		(long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
 788	for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
 789		pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
 790			__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
 791	pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
 792		__func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
 793		READ_ONCE(rnp->exp_tasks));
 794	pr_info("%s: ->blkd_tasks", __func__);
 795	i = 0;
 796	list_for_each(lhp, &rnp->blkd_tasks) {
 797		pr_cont(" %p", lhp);
 798		if (++i >= ncheck)
 799			break;
 800	}
 801	pr_cont("\n");
 802	for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
 803		rdp = per_cpu_ptr(&rcu_data, cpu);
 804		pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
 805			cpu, ".o"[rcu_rdp_cpu_online(rdp)],
 806			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
 807			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
 808	}
 809}
 810
 811#else /* #ifdef CONFIG_PREEMPT_RCU */
 812
 813/*
 814 * If strict grace periods are enabled, and if the calling
 815 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
 816 * report that quiescent state and, if requested, spin for a bit.
 817 */
 818void rcu_read_unlock_strict(void)
 819{
 820	struct rcu_data *rdp;
 821
 822	if (irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
 823		return;
 824	rdp = this_cpu_ptr(&rcu_data);
 825	rdp->cpu_no_qs.b.norm = false;
 826	rcu_report_qs_rdp(rdp);
 827	udelay(rcu_unlock_delay);
 828}
 829EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
 830
 831/*
 832 * Tell them what RCU they are running.
 833 */
 834static void __init rcu_bootup_announce(void)
 835{
 836	pr_info("Hierarchical RCU implementation.\n");
 837	rcu_bootup_announce_oddness();
 838}
 839
 840/*
 841 * Note a quiescent state for PREEMPTION=n.  Because we do not need to know
 842 * how many quiescent states passed, just if there was at least one since
 843 * the start of the grace period, this just sets a flag.  The caller must
 844 * have disabled preemption.
 845 */
 846static void rcu_qs(void)
 847{
 848	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
 849	if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
 850		return;
 851	trace_rcu_grace_period(TPS("rcu_sched"),
 852			       __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
 853	__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 854	if (__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
 855		rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 856}
 857
 858/*
 859 * Register an urgently needed quiescent state.  If there is an
 860 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
 861 * dyntick-idle quiescent state visible to other CPUs, which will in
 862 * some cases serve for expedited as well as normal grace periods.
 863 * Either way, register a lightweight quiescent state.
 864 */
 865void rcu_all_qs(void)
 866{
 867	unsigned long flags;
 868
 869	if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
 870		return;
 871	preempt_disable();  // For CONFIG_PREEMPT_COUNT=y kernels
 872	/* Load rcu_urgent_qs before other flags. */
 873	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
 874		preempt_enable();
 875		return;
 876	}
 877	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 878	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
 879		local_irq_save(flags);
 880		rcu_momentary_dyntick_idle();
 881		local_irq_restore(flags);
 882	}
 883	rcu_qs();
 884	preempt_enable();
 885}
 886EXPORT_SYMBOL_GPL(rcu_all_qs);
 887
 888/*
 889 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
 890 */
 891void rcu_note_context_switch(bool preempt)
 892{
 893	trace_rcu_utilization(TPS("Start context switch"));
 894	rcu_qs();
 895	/* Load rcu_urgent_qs before other flags. */
 896	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
 897		goto out;
 898	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 899	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
 900		rcu_momentary_dyntick_idle();
 901out:
 902	rcu_tasks_qs(current, preempt);
 903	trace_rcu_utilization(TPS("End context switch"));
 904}
 905EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 906
 907/*
 908 * Because preemptible RCU does not exist, there are never any preempted
 909 * RCU readers.
 910 */
 911static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 912{
 913	return 0;
 914}
 915
 916/*
 917 * Because there is no preemptible RCU, there can be no readers blocked.
 918 */
 919static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 920{
 921	return false;
 922}
 923
 924/*
 925 * Because there is no preemptible RCU, there can be no deferred quiescent
 926 * states.
 927 */
 928static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 929{
 930	return false;
 931}
 932
 933// Except that we do need to respond to a request by an expedited
 934// grace period for a quiescent state from this CPU.  Note that in
 935// non-preemptible kernels, there can be no context switches within RCU
 936// read-side critical sections, which in turn means that the leaf rcu_node
 937// structure's blocked-tasks list is always empty.  is therefore no need to
 938// actually check it.  Instead, a quiescent state from this CPU suffices,
 939// and this function is only called from such a quiescent state.
 940notrace void rcu_preempt_deferred_qs(struct task_struct *t)
 941{
 942	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 943
 944	if (rdp->cpu_no_qs.b.exp)
 945		rcu_report_exp_rdp(rdp);
 946}
 947
 948/*
 949 * Because there is no preemptible RCU, there can be no readers blocked,
 950 * so there is no need to check for blocked tasks.  So check only for
 951 * bogus qsmask values.
 952 */
 953static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 954{
 955	WARN_ON_ONCE(rnp->qsmask);
 956}
 957
 958/*
 959 * Check to see if this CPU is in a non-context-switch quiescent state,
 960 * namely user mode and idle loop.
 961 */
 962static void rcu_flavor_sched_clock_irq(int user)
 963{
 964	if (user || rcu_is_cpu_rrupt_from_idle()) {
 965
 966		/*
 967		 * Get here if this CPU took its interrupt from user
 968		 * mode or from the idle loop, and if this is not a
 969		 * nested interrupt.  In this case, the CPU is in
 970		 * a quiescent state, so note it.
 971		 *
 972		 * No memory barrier is required here because rcu_qs()
 973		 * references only CPU-local variables that other CPUs
 974		 * neither access nor modify, at least not while the
 975		 * corresponding CPU is online.
 976		 */
 977		rcu_qs();
 978	}
 979}
 980
 981/*
 982 * Because preemptible RCU does not exist, tasks cannot possibly exit
 983 * while in preemptible RCU read-side critical sections.
 984 */
 985void exit_rcu(void)
 986{
 987}
 988
 989/*
 990 * Dump the guaranteed-empty blocked-tasks state.  Trust but verify.
 991 */
 992static void
 993dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 994{
 995	WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
 996}
 997
 998#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 999
1000/*
1001 * If boosting, set rcuc kthreads to realtime priority.
1002 */
1003static void rcu_cpu_kthread_setup(unsigned int cpu)
1004{
1005	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1006#ifdef CONFIG_RCU_BOOST
1007	struct sched_param sp;
1008
1009	sp.sched_priority = kthread_prio;
1010	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1011#endif /* #ifdef CONFIG_RCU_BOOST */
1012
1013	WRITE_ONCE(rdp->rcuc_activity, jiffies);
1014}
1015
1016static bool rcu_is_callbacks_nocb_kthread(struct rcu_data *rdp)
1017{
1018#ifdef CONFIG_RCU_NOCB_CPU
1019	return rdp->nocb_cb_kthread == current;
1020#else
1021	return false;
1022#endif
1023}
1024
1025/*
1026 * Is the current CPU running the RCU-callbacks kthread?
1027 * Caller must have preemption disabled.
1028 */
1029static bool rcu_is_callbacks_kthread(struct rcu_data *rdp)
1030{
1031	return rdp->rcu_cpu_kthread_task == current ||
1032			rcu_is_callbacks_nocb_kthread(rdp);
1033}
1034
1035#ifdef CONFIG_RCU_BOOST
1036
1037/*
1038 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1039 * or ->boost_tasks, advancing the pointer to the next task in the
1040 * ->blkd_tasks list.
1041 *
1042 * Note that irqs must be enabled: boosting the task can block.
1043 * Returns 1 if there are more tasks needing to be boosted.
1044 */
1045static int rcu_boost(struct rcu_node *rnp)
1046{
1047	unsigned long flags;
1048	struct task_struct *t;
1049	struct list_head *tb;
1050
1051	if (READ_ONCE(rnp->exp_tasks) == NULL &&
1052	    READ_ONCE(rnp->boost_tasks) == NULL)
1053		return 0;  /* Nothing left to boost. */
1054
1055	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1056
1057	/*
1058	 * Recheck under the lock: all tasks in need of boosting
1059	 * might exit their RCU read-side critical sections on their own.
1060	 */
1061	if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1062		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1063		return 0;
1064	}
1065
1066	/*
1067	 * Preferentially boost tasks blocking expedited grace periods.
1068	 * This cannot starve the normal grace periods because a second
1069	 * expedited grace period must boost all blocked tasks, including
1070	 * those blocking the pre-existing normal grace period.
1071	 */
1072	if (rnp->exp_tasks != NULL)
1073		tb = rnp->exp_tasks;
1074	else
1075		tb = rnp->boost_tasks;
1076
1077	/*
1078	 * We boost task t by manufacturing an rt_mutex that appears to
1079	 * be held by task t.  We leave a pointer to that rt_mutex where
1080	 * task t can find it, and task t will release the mutex when it
1081	 * exits its outermost RCU read-side critical section.  Then
1082	 * simply acquiring this artificial rt_mutex will boost task
1083	 * t's priority.  (Thanks to tglx for suggesting this approach!)
1084	 *
1085	 * Note that task t must acquire rnp->lock to remove itself from
1086	 * the ->blkd_tasks list, which it will do from exit() if from
1087	 * nowhere else.  We therefore are guaranteed that task t will
1088	 * stay around at least until we drop rnp->lock.  Note that
1089	 * rnp->lock also resolves races between our priority boosting
1090	 * and task t's exiting its outermost RCU read-side critical
1091	 * section.
1092	 */
1093	t = container_of(tb, struct task_struct, rcu_node_entry);
1094	rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
1095	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1096	/* Lock only for side effect: boosts task t's priority. */
1097	rt_mutex_lock(&rnp->boost_mtx);
1098	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1099	rnp->n_boosts++;
1100
1101	return READ_ONCE(rnp->exp_tasks) != NULL ||
1102	       READ_ONCE(rnp->boost_tasks) != NULL;
1103}
1104
1105/*
1106 * Priority-boosting kthread, one per leaf rcu_node.
1107 */
1108static int rcu_boost_kthread(void *arg)
1109{
1110	struct rcu_node *rnp = (struct rcu_node *)arg;
1111	int spincnt = 0;
1112	int more2boost;
1113
1114	trace_rcu_utilization(TPS("Start boost kthread@init"));
1115	for (;;) {
1116		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1117		trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1118		rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1119			 READ_ONCE(rnp->exp_tasks));
1120		trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1121		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1122		more2boost = rcu_boost(rnp);
1123		if (more2boost)
1124			spincnt++;
1125		else
1126			spincnt = 0;
1127		if (spincnt > 10) {
1128			WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1129			trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1130			schedule_timeout_idle(2);
1131			trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1132			spincnt = 0;
1133		}
1134	}
1135	/* NOTREACHED */
1136	trace_rcu_utilization(TPS("End boost kthread@notreached"));
1137	return 0;
1138}
1139
1140/*
1141 * Check to see if it is time to start boosting RCU readers that are
1142 * blocking the current grace period, and, if so, tell the per-rcu_node
1143 * kthread to start boosting them.  If there is an expedited grace
1144 * period in progress, it is always time to boost.
1145 *
1146 * The caller must hold rnp->lock, which this function releases.
1147 * The ->boost_kthread_task is immortal, so we don't need to worry
1148 * about it going away.
1149 */
1150static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1151	__releases(rnp->lock)
1152{
1153	raw_lockdep_assert_held_rcu_node(rnp);
1154	if (!rnp->boost_kthread_task ||
1155	    (!rcu_preempt_blocked_readers_cgp(rnp) && !rnp->exp_tasks)) {
1156		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1157		return;
1158	}
1159	if (rnp->exp_tasks != NULL ||
1160	    (rnp->gp_tasks != NULL &&
1161	     rnp->boost_tasks == NULL &&
1162	     rnp->qsmask == 0 &&
1163	     (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld ||
1164	      IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)))) {
1165		if (rnp->exp_tasks == NULL)
1166			WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1167		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1168		rcu_wake_cond(rnp->boost_kthread_task,
1169			      READ_ONCE(rnp->boost_kthread_status));
1170	} else {
1171		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1172	}
1173}
1174
1175#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1176
1177/*
1178 * Do priority-boost accounting for the start of a new grace period.
1179 */
1180static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1181{
1182	rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1183}
1184
1185/*
1186 * Create an RCU-boost kthread for the specified node if one does not
1187 * already exist.  We only create this kthread for preemptible RCU.
1188 */
1189static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1190{
1191	unsigned long flags;
1192	int rnp_index = rnp - rcu_get_root();
1193	struct sched_param sp;
1194	struct task_struct *t;
1195
1196	mutex_lock(&rnp->boost_kthread_mutex);
1197	if (rnp->boost_kthread_task || !rcu_scheduler_fully_active)
1198		goto out;
1199
1200	t = kthread_create(rcu_boost_kthread, (void *)rnp,
1201			   "rcub/%d", rnp_index);
1202	if (WARN_ON_ONCE(IS_ERR(t)))
1203		goto out;
1204
1205	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1206	rnp->boost_kthread_task = t;
1207	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1208	sp.sched_priority = kthread_prio;
1209	sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1210	wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1211
1212 out:
1213	mutex_unlock(&rnp->boost_kthread_mutex);
1214}
1215
1216/*
1217 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1218 * served by the rcu_node in question.  The CPU hotplug lock is still
1219 * held, so the value of rnp->qsmaskinit will be stable.
1220 *
1221 * We don't include outgoingcpu in the affinity set, use -1 if there is
1222 * no outgoing CPU.  If there are no CPUs left in the affinity set,
1223 * this function allows the kthread to execute on any CPU.
1224 *
1225 * Any future concurrent calls are serialized via ->boost_kthread_mutex.
1226 */
1227static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1228{
1229	struct task_struct *t = rnp->boost_kthread_task;
1230	unsigned long mask;
1231	cpumask_var_t cm;
1232	int cpu;
1233
1234	if (!t)
1235		return;
1236	if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1237		return;
1238	mutex_lock(&rnp->boost_kthread_mutex);
1239	mask = rcu_rnp_online_cpus(rnp);
1240	for_each_leaf_node_possible_cpu(rnp, cpu)
1241		if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1242		    cpu != outgoingcpu)
1243			cpumask_set_cpu(cpu, cm);
1244	cpumask_and(cm, cm, housekeeping_cpumask(HK_TYPE_RCU));
1245	if (cpumask_empty(cm)) {
1246		cpumask_copy(cm, housekeeping_cpumask(HK_TYPE_RCU));
1247		if (outgoingcpu >= 0)
1248			cpumask_clear_cpu(outgoingcpu, cm);
1249	}
1250	set_cpus_allowed_ptr(t, cm);
1251	mutex_unlock(&rnp->boost_kthread_mutex);
1252	free_cpumask_var(cm);
1253}
1254
1255#else /* #ifdef CONFIG_RCU_BOOST */
1256
1257static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1258	__releases(rnp->lock)
1259{
1260	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1261}
1262
1263static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1264{
1265}
1266
1267static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1268{
1269}
1270
1271static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1272{
1273}
1274
1275#endif /* #else #ifdef CONFIG_RCU_BOOST */
1276
1277/*
1278 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
1279 * grace-period kthread will do force_quiescent_state() processing?
1280 * The idea is to avoid waking up RCU core processing on such a
1281 * CPU unless the grace period has extended for too long.
1282 *
1283 * This code relies on the fact that all NO_HZ_FULL CPUs are also
1284 * RCU_NOCB_CPU CPUs.
1285 */
1286static bool rcu_nohz_full_cpu(void)
1287{
1288#ifdef CONFIG_NO_HZ_FULL
1289	if (tick_nohz_full_cpu(smp_processor_id()) &&
1290	    (!rcu_gp_in_progress() ||
1291	     time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
1292		return true;
1293#endif /* #ifdef CONFIG_NO_HZ_FULL */
1294	return false;
1295}
1296
1297/*
1298 * Bind the RCU grace-period kthreads to the housekeeping CPU.
1299 */
1300static void rcu_bind_gp_kthread(void)
1301{
1302	if (!tick_nohz_full_enabled())
1303		return;
1304	housekeeping_affine(current, HK_TYPE_RCU);
1305}