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_NOCB_LOCKDEP_WARN(
  28		!(lockdep_is_held(&rcu_state.barrier_mutex) ||
  29		  (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
  30		  lockdep_is_held(&rdp->nocb_lock) ||
  31		  lockdep_is_held(&rcu_state.nocb_mutex) ||
  32		  (!(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible()) &&
  33		   rdp == this_cpu_ptr(&rcu_data)) ||
  34		  rcu_current_is_nocb_kthread(rdp)),
  35		"Unsafe read of RCU_NOCB offloaded state"
  36	);
  37
  38	return rcu_segcblist_is_offloaded(&rdp->cblist);
  39}
  40
  41/*
  42 * Check the RCU kernel configuration parameters and print informative
  43 * messages about anything out of the ordinary.
  44 */
  45static void __init rcu_bootup_announce_oddness(void)
  46{
  47	if (IS_ENABLED(CONFIG_RCU_TRACE))
  48		pr_info("\tRCU event tracing is enabled.\n");
  49	if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
  50	    (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
  51		pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
  52			RCU_FANOUT);
  53	if (rcu_fanout_exact)
  54		pr_info("\tHierarchical RCU autobalancing is disabled.\n");
  55	if (IS_ENABLED(CONFIG_PROVE_RCU))
  56		pr_info("\tRCU lockdep checking is enabled.\n");
  57	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
  58		pr_info("\tRCU strict (and thus non-scalable) grace periods are enabled.\n");
  59	if (RCU_NUM_LVLS >= 4)
  60		pr_info("\tFour(or more)-level hierarchy is enabled.\n");
  61	if (RCU_FANOUT_LEAF != 16)
  62		pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
  63			RCU_FANOUT_LEAF);
  64	if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
  65		pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
  66			rcu_fanout_leaf);
  67	if (nr_cpu_ids != NR_CPUS)
  68		pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
  69#ifdef CONFIG_RCU_BOOST
  70	pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
  71		kthread_prio, CONFIG_RCU_BOOST_DELAY);
  72#endif
  73	if (blimit != DEFAULT_RCU_BLIMIT)
  74		pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
  75	if (qhimark != DEFAULT_RCU_QHIMARK)
  76		pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
  77	if (qlowmark != DEFAULT_RCU_QLOMARK)
  78		pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
  79	if (qovld != DEFAULT_RCU_QOVLD)
  80		pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
  81	if (jiffies_till_first_fqs != ULONG_MAX)
  82		pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
  83	if (jiffies_till_next_fqs != ULONG_MAX)
  84		pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
  85	if (jiffies_till_sched_qs != ULONG_MAX)
  86		pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
  87	if (rcu_kick_kthreads)
  88		pr_info("\tKick kthreads if too-long grace period.\n");
  89	if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
  90		pr_info("\tRCU callback double-/use-after-free debug is enabled.\n");
  91	if (gp_preinit_delay)
  92		pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
  93	if (gp_init_delay)
  94		pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
  95	if (gp_cleanup_delay)
  96		pr_info("\tRCU debug GP cleanup slowdown %d jiffies.\n", gp_cleanup_delay);
  97	if (nohz_full_patience_delay < 0) {
  98		pr_info("\tRCU NOCB CPU patience negative (%d), resetting to zero.\n", nohz_full_patience_delay);
  99		nohz_full_patience_delay = 0;
 100	} else if (nohz_full_patience_delay > 5 * MSEC_PER_SEC) {
 101		pr_info("\tRCU NOCB CPU patience too large (%d), resetting to %ld.\n", nohz_full_patience_delay, 5 * MSEC_PER_SEC);
 102		nohz_full_patience_delay = 5 * MSEC_PER_SEC;
 103	} else if (nohz_full_patience_delay) {
 104		pr_info("\tRCU NOCB CPU patience set to %d milliseconds.\n", nohz_full_patience_delay);
 105	}
 106	nohz_full_patience_delay_jiffies = msecs_to_jiffies(nohz_full_patience_delay);
 107	if (!use_softirq)
 108		pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
 109	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
 110		pr_info("\tRCU debug extended QS entry/exit.\n");
 111	rcupdate_announce_bootup_oddness();
 112}
 113
 114#ifdef CONFIG_PREEMPT_RCU
 115
 116static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
 117static void rcu_read_unlock_special(struct task_struct *t);
 118
 119/*
 120 * Tell them what RCU they are running.
 121 */
 122static void __init rcu_bootup_announce(void)
 123{
 124	pr_info("Preemptible hierarchical RCU implementation.\n");
 125	rcu_bootup_announce_oddness();
 126}
 127
 128/* Flags for rcu_preempt_ctxt_queue() decision table. */
 129#define RCU_GP_TASKS	0x8
 130#define RCU_EXP_TASKS	0x4
 131#define RCU_GP_BLKD	0x2
 132#define RCU_EXP_BLKD	0x1
 133
 134/*
 135 * Queues a task preempted within an RCU-preempt read-side critical
 136 * section into the appropriate location within the ->blkd_tasks list,
 137 * depending on the states of any ongoing normal and expedited grace
 138 * periods.  The ->gp_tasks pointer indicates which element the normal
 139 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
 140 * indicates which element the expedited grace period is waiting on (again,
 141 * NULL if none).  If a grace period is waiting on a given element in the
 142 * ->blkd_tasks list, it also waits on all subsequent elements.  Thus,
 143 * adding a task to the tail of the list blocks any grace period that is
 144 * already waiting on one of the elements.  In contrast, adding a task
 145 * to the head of the list won't block any grace period that is already
 146 * waiting on one of the elements.
 147 *
 148 * This queuing is imprecise, and can sometimes make an ongoing grace
 149 * period wait for a task that is not strictly speaking blocking it.
 150 * Given the choice, we needlessly block a normal grace period rather than
 151 * blocking an expedited grace period.
 152 *
 153 * Note that an endless sequence of expedited grace periods still cannot
 154 * indefinitely postpone a normal grace period.  Eventually, all of the
 155 * fixed number of preempted tasks blocking the normal grace period that are
 156 * not also blocking the expedited grace period will resume and complete
 157 * their RCU read-side critical sections.  At that point, the ->gp_tasks
 158 * pointer will equal the ->exp_tasks pointer, at which point the end of
 159 * the corresponding expedited grace period will also be the end of the
 160 * normal grace period.
 161 */
 162static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
 163	__releases(rnp->lock) /* But leaves rrupts disabled. */
 164{
 165	int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
 166			 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
 167			 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
 168			 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
 169	struct task_struct *t = current;
 170
 171	raw_lockdep_assert_held_rcu_node(rnp);
 172	WARN_ON_ONCE(rdp->mynode != rnp);
 173	WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 174	/* RCU better not be waiting on newly onlined CPUs! */
 175	WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
 176		     rdp->grpmask);
 177
 178	/*
 179	 * Decide where to queue the newly blocked task.  In theory,
 180	 * this could be an if-statement.  In practice, when I tried
 181	 * that, it was quite messy.
 182	 */
 183	switch (blkd_state) {
 184	case 0:
 185	case                RCU_EXP_TASKS:
 186	case                RCU_EXP_TASKS | RCU_GP_BLKD:
 187	case RCU_GP_TASKS:
 188	case RCU_GP_TASKS | RCU_EXP_TASKS:
 189
 190		/*
 191		 * Blocking neither GP, or first task blocking the normal
 192		 * GP but not blocking the already-waiting expedited GP.
 193		 * Queue at the head of the list to avoid unnecessarily
 194		 * blocking the already-waiting GPs.
 195		 */
 196		list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
 197		break;
 198
 199	case                                              RCU_EXP_BLKD:
 200	case                                RCU_GP_BLKD:
 201	case                                RCU_GP_BLKD | RCU_EXP_BLKD:
 202	case RCU_GP_TASKS |                               RCU_EXP_BLKD:
 203	case RCU_GP_TASKS |                 RCU_GP_BLKD | RCU_EXP_BLKD:
 204	case RCU_GP_TASKS | RCU_EXP_TASKS | RCU_GP_BLKD | RCU_EXP_BLKD:
 205
 206		/*
 207		 * First task arriving that blocks either GP, or first task
 208		 * arriving that blocks the expedited GP (with the normal
 209		 * GP already waiting), or a task arriving that blocks
 210		 * both GPs with both GPs already waiting.  Queue at the
 211		 * tail of the list to avoid any GP waiting on any of the
 212		 * already queued tasks that are not blocking it.
 213		 */
 214		list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
 215		break;
 216
 217	case                RCU_EXP_TASKS |               RCU_EXP_BLKD:
 218	case                RCU_EXP_TASKS | RCU_GP_BLKD | RCU_EXP_BLKD:
 219	case RCU_GP_TASKS | RCU_EXP_TASKS |               RCU_EXP_BLKD:
 220
 221		/*
 222		 * Second or subsequent task blocking the expedited GP.
 223		 * The task either does not block the normal GP, or is the
 224		 * first task blocking the normal GP.  Queue just after
 225		 * the first task blocking the expedited GP.
 226		 */
 227		list_add(&t->rcu_node_entry, rnp->exp_tasks);
 228		break;
 229
 230	case RCU_GP_TASKS |                 RCU_GP_BLKD:
 231	case RCU_GP_TASKS | RCU_EXP_TASKS | RCU_GP_BLKD:
 232
 233		/*
 234		 * Second or subsequent task blocking the normal GP.
 235		 * The task does not block the expedited GP. Queue just
 236		 * after the first task blocking the normal GP.
 237		 */
 238		list_add(&t->rcu_node_entry, rnp->gp_tasks);
 239		break;
 240
 241	default:
 242
 243		/* Yet another exercise in excessive paranoia. */
 244		WARN_ON_ONCE(1);
 245		break;
 246	}
 247
 248	/*
 249	 * We have now queued the task.  If it was the first one to
 250	 * block either grace period, update the ->gp_tasks and/or
 251	 * ->exp_tasks pointers, respectively, to reference the newly
 252	 * blocked tasks.
 253	 */
 254	if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
 255		WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
 256		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
 257	}
 258	if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
 259		WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
 260	WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
 261		     !(rnp->qsmask & rdp->grpmask));
 262	WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
 263		     !(rnp->expmask & rdp->grpmask));
 264	raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
 265
 266	/*
 267	 * Report the quiescent state for the expedited GP.  This expedited
 268	 * GP should not be able to end until we report, so there should be
 269	 * no need to check for a subsequent expedited GP.  (Though we are
 270	 * still in a quiescent state in any case.)
 271	 *
 272	 * Interrupts are disabled, so ->cpu_no_qs.b.exp cannot change.
 273	 */
 274	if (blkd_state & RCU_EXP_BLKD && rdp->cpu_no_qs.b.exp)
 275		rcu_report_exp_rdp(rdp);
 276	else
 277		WARN_ON_ONCE(rdp->cpu_no_qs.b.exp);
 278}
 279
 280/*
 281 * Record a preemptible-RCU quiescent state for the specified CPU.
 282 * Note that this does not necessarily mean that the task currently running
 283 * on the CPU is in a quiescent state:  Instead, it means that the current
 284 * grace period need not wait on any RCU read-side critical section that
 285 * starts later on this CPU.  It also means that if the current task is
 286 * in an RCU read-side critical section, it has already added itself to
 287 * some leaf rcu_node structure's ->blkd_tasks list.  In addition to the
 288 * current task, there might be any number of other tasks blocked while
 289 * in an RCU read-side critical section.
 290 *
 291 * Unlike non-preemptible-RCU, quiescent state reports for expedited
 292 * grace periods are handled separately via deferred quiescent states
 293 * and context switch events.
 294 *
 295 * Callers to this function must disable preemption.
 296 */
 297static void rcu_qs(void)
 298{
 299	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
 300	if (__this_cpu_read(rcu_data.cpu_no_qs.b.norm)) {
 301		trace_rcu_grace_period(TPS("rcu_preempt"),
 302				       __this_cpu_read(rcu_data.gp_seq),
 303				       TPS("cpuqs"));
 304		__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 305		barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
 306		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
 307	}
 308}
 309
 310/*
 311 * We have entered the scheduler, and the current task might soon be
 312 * context-switched away from.  If this task is in an RCU read-side
 313 * critical section, we will no longer be able to rely on the CPU to
 314 * record that fact, so we enqueue the task on the blkd_tasks list.
 315 * The task will dequeue itself when it exits the outermost enclosing
 316 * RCU read-side critical section.  Therefore, the current grace period
 317 * cannot be permitted to complete until the blkd_tasks list entries
 318 * predating the current grace period drain, in other words, until
 319 * rnp->gp_tasks becomes NULL.
 320 *
 321 * Caller must disable interrupts.
 322 */
 323void rcu_note_context_switch(bool preempt)
 324{
 325	struct task_struct *t = current;
 326	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 327	struct rcu_node *rnp;
 328
 329	trace_rcu_utilization(TPS("Start context switch"));
 330	lockdep_assert_irqs_disabled();
 331	WARN_ONCE(!preempt && rcu_preempt_depth() > 0, "Voluntary context switch within RCU read-side critical section!");
 332	if (rcu_preempt_depth() > 0 &&
 333	    !t->rcu_read_unlock_special.b.blocked) {
 334
 335		/* Possibly blocking in an RCU read-side critical section. */
 336		rnp = rdp->mynode;
 337		raw_spin_lock_rcu_node(rnp);
 338		t->rcu_read_unlock_special.b.blocked = true;
 339		t->rcu_blocked_node = rnp;
 340
 341		/*
 342		 * Verify the CPU's sanity, trace the preemption, and
 343		 * then queue the task as required based on the states
 344		 * of any ongoing and expedited grace periods.
 345		 */
 346		WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp));
 347		WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
 348		trace_rcu_preempt_task(rcu_state.name,
 349				       t->pid,
 350				       (rnp->qsmask & rdp->grpmask)
 351				       ? rnp->gp_seq
 352				       : rcu_seq_snap(&rnp->gp_seq));
 353		rcu_preempt_ctxt_queue(rnp, rdp);
 354	} else {
 355		rcu_preempt_deferred_qs(t);
 356	}
 357
 358	/*
 359	 * Either we were not in an RCU read-side critical section to
 360	 * begin with, or we have now recorded that critical section
 361	 * globally.  Either way, we can now note a quiescent state
 362	 * for this CPU.  Again, if we were in an RCU read-side critical
 363	 * section, and if that critical section was blocking the current
 364	 * grace period, then the fact that the task has been enqueued
 365	 * means that we continue to block the current grace period.
 366	 */
 367	rcu_qs();
 368	if (rdp->cpu_no_qs.b.exp)
 369		rcu_report_exp_rdp(rdp);
 370	rcu_tasks_qs(current, preempt);
 371	trace_rcu_utilization(TPS("End context switch"));
 372}
 373EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 374
 375/*
 376 * Check for preempted RCU readers blocking the current grace period
 377 * for the specified rcu_node structure.  If the caller needs a reliable
 378 * answer, it must hold the rcu_node's ->lock.
 379 */
 380static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 381{
 382	return READ_ONCE(rnp->gp_tasks) != NULL;
 383}
 384
 385/* limit value for ->rcu_read_lock_nesting. */
 386#define RCU_NEST_PMAX (INT_MAX / 2)
 387
 388static void rcu_preempt_read_enter(void)
 389{
 390	WRITE_ONCE(current->rcu_read_lock_nesting, READ_ONCE(current->rcu_read_lock_nesting) + 1);
 391}
 392
 393static int rcu_preempt_read_exit(void)
 394{
 395	int ret = READ_ONCE(current->rcu_read_lock_nesting) - 1;
 396
 397	WRITE_ONCE(current->rcu_read_lock_nesting, ret);
 398	return ret;
 399}
 400
 401static void rcu_preempt_depth_set(int val)
 402{
 403	WRITE_ONCE(current->rcu_read_lock_nesting, val);
 404}
 405
 406/*
 407 * Preemptible RCU implementation for rcu_read_lock().
 408 * Just increment ->rcu_read_lock_nesting, shared state will be updated
 409 * if we block.
 410 */
 411void __rcu_read_lock(void)
 412{
 413	rcu_preempt_read_enter();
 414	if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 415		WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
 416	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
 417		WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
 418	barrier();  /* critical section after entry code. */
 419}
 420EXPORT_SYMBOL_GPL(__rcu_read_lock);
 421
 422/*
 423 * Preemptible RCU implementation for rcu_read_unlock().
 424 * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
 425 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
 426 * invoke rcu_read_unlock_special() to clean up after a context switch
 427 * in an RCU read-side critical section and other special cases.
 428 */
 429void __rcu_read_unlock(void)
 430{
 431	struct task_struct *t = current;
 432
 433	barrier();  // critical section before exit code.
 434	if (rcu_preempt_read_exit() == 0) {
 435		barrier();  // critical-section exit before .s check.
 436		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
 437			rcu_read_unlock_special(t);
 438	}
 439	if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 440		int rrln = rcu_preempt_depth();
 441
 442		WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
 443	}
 444}
 445EXPORT_SYMBOL_GPL(__rcu_read_unlock);
 446
 447/*
 448 * Advance a ->blkd_tasks-list pointer to the next entry, instead
 449 * returning NULL if at the end of the list.
 450 */
 451static struct list_head *rcu_next_node_entry(struct task_struct *t,
 452					     struct rcu_node *rnp)
 453{
 454	struct list_head *np;
 455
 456	np = t->rcu_node_entry.next;
 457	if (np == &rnp->blkd_tasks)
 458		np = NULL;
 459	return np;
 460}
 461
 462/*
 463 * Return true if the specified rcu_node structure has tasks that were
 464 * preempted within an RCU read-side critical section.
 465 */
 466static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 467{
 468	return !list_empty(&rnp->blkd_tasks);
 469}
 470
 471/*
 472 * Report deferred quiescent states.  The deferral time can
 473 * be quite short, for example, in the case of the call from
 474 * rcu_read_unlock_special().
 475 */
 476static notrace void
 477rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
 478{
 479	bool empty_exp;
 480	bool empty_norm;
 481	bool empty_exp_now;
 482	struct list_head *np;
 483	bool drop_boost_mutex = false;
 484	struct rcu_data *rdp;
 485	struct rcu_node *rnp;
 486	union rcu_special special;
 487
 488	/*
 489	 * If RCU core is waiting for this CPU to exit its critical section,
 490	 * report the fact that it has exited.  Because irqs are disabled,
 491	 * t->rcu_read_unlock_special cannot change.
 492	 */
 493	special = t->rcu_read_unlock_special;
 494	rdp = this_cpu_ptr(&rcu_data);
 495	if (!special.s && !rdp->cpu_no_qs.b.exp) {
 496		local_irq_restore(flags);
 497		return;
 498	}
 499	t->rcu_read_unlock_special.s = 0;
 500	if (special.b.need_qs) {
 501		if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
 502			rdp->cpu_no_qs.b.norm = false;
 503			rcu_report_qs_rdp(rdp);
 504			udelay(rcu_unlock_delay);
 505		} else {
 506			rcu_qs();
 507		}
 508	}
 509
 510	/*
 511	 * Respond to a request by an expedited grace period for a
 512	 * quiescent state from this CPU.  Note that requests from
 513	 * tasks are handled when removing the task from the
 514	 * blocked-tasks list below.
 515	 */
 516	if (rdp->cpu_no_qs.b.exp)
 517		rcu_report_exp_rdp(rdp);
 518
 519	/* Clean up if blocked during RCU read-side critical section. */
 520	if (special.b.blocked) {
 521
 522		/*
 523		 * Remove this task from the list it blocked on.  The task
 524		 * now remains queued on the rcu_node corresponding to the
 525		 * CPU it first blocked on, so there is no longer any need
 526		 * to loop.  Retain a WARN_ON_ONCE() out of sheer paranoia.
 527		 */
 528		rnp = t->rcu_blocked_node;
 529		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
 530		WARN_ON_ONCE(rnp != t->rcu_blocked_node);
 531		WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
 532		empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
 533		WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
 534			     (!empty_norm || rnp->qsmask));
 535		empty_exp = sync_rcu_exp_done(rnp);
 536		smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
 537		np = rcu_next_node_entry(t, rnp);
 538		list_del_init(&t->rcu_node_entry);
 539		t->rcu_blocked_node = NULL;
 540		trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
 541						rnp->gp_seq, t->pid);
 542		if (&t->rcu_node_entry == rnp->gp_tasks)
 543			WRITE_ONCE(rnp->gp_tasks, np);
 544		if (&t->rcu_node_entry == rnp->exp_tasks)
 545			WRITE_ONCE(rnp->exp_tasks, np);
 546		if (IS_ENABLED(CONFIG_RCU_BOOST)) {
 547			/* Snapshot ->boost_mtx ownership w/rnp->lock held. */
 548			drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx.rtmutex) == t;
 549			if (&t->rcu_node_entry == rnp->boost_tasks)
 550				WRITE_ONCE(rnp->boost_tasks, np);
 551		}
 552
 553		/*
 554		 * If this was the last task on the current list, and if
 555		 * we aren't waiting on any CPUs, report the quiescent state.
 556		 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
 557		 * so we must take a snapshot of the expedited state.
 558		 */
 559		empty_exp_now = sync_rcu_exp_done(rnp);
 560		if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
 561			trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
 562							 rnp->gp_seq,
 563							 0, rnp->qsmask,
 564							 rnp->level,
 565							 rnp->grplo,
 566							 rnp->grphi,
 567							 !!rnp->gp_tasks);
 568			rcu_report_unblock_qs_rnp(rnp, flags);
 569		} else {
 570			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 571		}
 572
 573		/*
 574		 * If this was the last task on the expedited lists,
 575		 * then we need to report up the rcu_node hierarchy.
 576		 */
 577		if (!empty_exp && empty_exp_now)
 578			rcu_report_exp_rnp(rnp, true);
 579
 580		/* Unboost if we were boosted. */
 581		if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
 582			rt_mutex_futex_unlock(&rnp->boost_mtx.rtmutex);
 583	} else {
 584		local_irq_restore(flags);
 585	}
 586}
 587
 588/*
 589 * Is a deferred quiescent-state pending, and are we also not in
 590 * an RCU read-side critical section?  It is the caller's responsibility
 591 * to ensure it is otherwise safe to report any deferred quiescent
 592 * states.  The reason for this is that it is safe to report a
 593 * quiescent state during context switch even though preemption
 594 * is disabled.  This function cannot be expected to understand these
 595 * nuances, so the caller must handle them.
 596 */
 597static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 598{
 599	return (__this_cpu_read(rcu_data.cpu_no_qs.b.exp) ||
 600		READ_ONCE(t->rcu_read_unlock_special.s)) &&
 601	       rcu_preempt_depth() == 0;
 602}
 603
 604/*
 605 * Report a deferred quiescent state if needed and safe to do so.
 606 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
 607 * not being in an RCU read-side critical section.  The caller must
 608 * evaluate safety in terms of interrupt, softirq, and preemption
 609 * disabling.
 610 */
 611notrace void rcu_preempt_deferred_qs(struct task_struct *t)
 612{
 613	unsigned long flags;
 614
 615	if (!rcu_preempt_need_deferred_qs(t))
 616		return;
 617	local_irq_save(flags);
 618	rcu_preempt_deferred_qs_irqrestore(t, flags);
 619}
 620
 621/*
 622 * Minimal handler to give the scheduler a chance to re-evaluate.
 623 */
 624static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
 625{
 626	struct rcu_data *rdp;
 627
 628	rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
 629	rdp->defer_qs_iw_pending = false;
 630}
 631
 632/*
 633 * Handle special cases during rcu_read_unlock(), such as needing to
 634 * notify RCU core processing or task having blocked during the RCU
 635 * read-side critical section.
 636 */
 637static void rcu_read_unlock_special(struct task_struct *t)
 638{
 639	unsigned long flags;
 640	bool irqs_were_disabled;
 641	bool preempt_bh_were_disabled =
 642			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
 643
 644	/* NMI handlers cannot block and cannot safely manipulate state. */
 645	if (in_nmi())
 646		return;
 647
 648	local_irq_save(flags);
 649	irqs_were_disabled = irqs_disabled_flags(flags);
 650	if (preempt_bh_were_disabled || irqs_were_disabled) {
 651		bool expboost; // Expedited GP in flight or possible boosting.
 652		struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 653		struct rcu_node *rnp = rdp->mynode;
 654
 655		expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
 656			   (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
 657			   (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
 658			   ((rdp->grpmask & READ_ONCE(rnp->qsmask)) || t->rcu_blocked_node)) ||
 659			   (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
 660			    t->rcu_blocked_node);
 661		// Need to defer quiescent state until everything is enabled.
 662		if (use_softirq && (in_hardirq() || (expboost && !irqs_were_disabled))) {
 663			// Using softirq, safe to awaken, and either the
 664			// wakeup is free or there is either an expedited
 665			// GP in flight or a potential need to deboost.
 666			raise_softirq_irqoff(RCU_SOFTIRQ);
 667		} else {
 668			// Enabling BH or preempt does reschedule, so...
 669			// Also if no expediting and no possible deboosting,
 670			// slow is OK.  Plus nohz_full CPUs eventually get
 671			// tick enabled.
 672			set_tsk_need_resched(current);
 673			set_preempt_need_resched();
 674			if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
 675			    expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
 676				// Get scheduler to re-evaluate and call hooks.
 677				// If !IRQ_WORK, FQS scan will eventually IPI.
 678				if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
 679				    IS_ENABLED(CONFIG_PREEMPT_RT))
 680					rdp->defer_qs_iw = IRQ_WORK_INIT_HARD(
 681								rcu_preempt_deferred_qs_handler);
 682				else
 683					init_irq_work(&rdp->defer_qs_iw,
 684						      rcu_preempt_deferred_qs_handler);
 685				rdp->defer_qs_iw_pending = true;
 686				irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
 687			}
 688		}
 689		local_irq_restore(flags);
 690		return;
 691	}
 692	rcu_preempt_deferred_qs_irqrestore(t, flags);
 693}
 694
 695/*
 696 * Check that the list of blocked tasks for the newly completed grace
 697 * period is in fact empty.  It is a serious bug to complete a grace
 698 * period that still has RCU readers blocked!  This function must be
 699 * invoked -before- updating this rnp's ->gp_seq.
 700 *
 701 * Also, if there are blocked tasks on the list, they automatically
 702 * block the newly created grace period, so set up ->gp_tasks accordingly.
 703 */
 704static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 705{
 706	struct task_struct *t;
 707
 708	RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
 709	raw_lockdep_assert_held_rcu_node(rnp);
 710	if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
 711		dump_blkd_tasks(rnp, 10);
 712	if (rcu_preempt_has_tasks(rnp) &&
 713	    (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
 714		WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
 715		t = container_of(rnp->gp_tasks, struct task_struct,
 716				 rcu_node_entry);
 717		trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
 718						rnp->gp_seq, t->pid);
 719	}
 720	WARN_ON_ONCE(rnp->qsmask);
 721}
 722
 723/*
 724 * Check for a quiescent state from the current CPU, including voluntary
 725 * context switches for Tasks RCU.  When a task blocks, the task is
 726 * recorded in the corresponding CPU's rcu_node structure, which is checked
 727 * elsewhere, hence this function need only check for quiescent states
 728 * related to the current CPU, not to those related to tasks.
 729 */
 730static void rcu_flavor_sched_clock_irq(int user)
 731{
 732	struct task_struct *t = current;
 733
 734	lockdep_assert_irqs_disabled();
 735	if (rcu_preempt_depth() > 0 ||
 736	    (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
 737		/* No QS, force context switch if deferred. */
 738		if (rcu_preempt_need_deferred_qs(t)) {
 739			set_tsk_need_resched(t);
 740			set_preempt_need_resched();
 741		}
 742	} else if (rcu_preempt_need_deferred_qs(t)) {
 743		rcu_preempt_deferred_qs(t); /* Report deferred QS. */
 744		return;
 745	} else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
 746		rcu_qs(); /* Report immediate QS. */
 747		return;
 748	}
 749
 750	/* If GP is oldish, ask for help from rcu_read_unlock_special(). */
 751	if (rcu_preempt_depth() > 0 &&
 752	    __this_cpu_read(rcu_data.core_needs_qs) &&
 753	    __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
 754	    !t->rcu_read_unlock_special.b.need_qs &&
 755	    time_after(jiffies, rcu_state.gp_start + HZ))
 756		t->rcu_read_unlock_special.b.need_qs = true;
 757}
 758
 759/*
 760 * Check for a task exiting while in a preemptible-RCU read-side
 761 * critical section, clean up if so.  No need to issue warnings, as
 762 * debug_check_no_locks_held() already does this if lockdep is enabled.
 763 * Besides, if this function does anything other than just immediately
 764 * return, there was a bug of some sort.  Spewing warnings from this
 765 * function is like as not to simply obscure important prior warnings.
 766 */
 767void exit_rcu(void)
 768{
 769	struct task_struct *t = current;
 770
 771	if (unlikely(!list_empty(&current->rcu_node_entry))) {
 772		rcu_preempt_depth_set(1);
 773		barrier();
 774		WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
 775	} else if (unlikely(rcu_preempt_depth())) {
 776		rcu_preempt_depth_set(1);
 777	} else {
 778		return;
 779	}
 780	__rcu_read_unlock();
 781	rcu_preempt_deferred_qs(current);
 782}
 783
 784/*
 785 * Dump the blocked-tasks state, but limit the list dump to the
 786 * specified number of elements.
 787 */
 788static void
 789dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
 790{
 791	int cpu;
 792	int i;
 793	struct list_head *lhp;
 794	struct rcu_data *rdp;
 795	struct rcu_node *rnp1;
 796
 797	raw_lockdep_assert_held_rcu_node(rnp);
 798	pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
 799		__func__, rnp->grplo, rnp->grphi, rnp->level,
 800		(long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
 801	for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
 802		pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
 803			__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
 804	pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
 805		__func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
 806		READ_ONCE(rnp->exp_tasks));
 807	pr_info("%s: ->blkd_tasks", __func__);
 808	i = 0;
 809	list_for_each(lhp, &rnp->blkd_tasks) {
 810		pr_cont(" %p", lhp);
 811		if (++i >= ncheck)
 812			break;
 813	}
 814	pr_cont("\n");
 815	for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
 816		rdp = per_cpu_ptr(&rcu_data, cpu);
 817		pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
 818			cpu, ".o"[rcu_rdp_cpu_online(rdp)],
 819			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_state,
 820			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_state);
 821	}
 822}
 823
 824#else /* #ifdef CONFIG_PREEMPT_RCU */
 825
 826/*
 827 * If strict grace periods are enabled, and if the calling
 828 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
 829 * report that quiescent state and, if requested, spin for a bit.
 830 */
 831void rcu_read_unlock_strict(void)
 832{
 833	struct rcu_data *rdp;
 834
 835	if (irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
 836		return;
 837	rdp = this_cpu_ptr(&rcu_data);
 838	rdp->cpu_no_qs.b.norm = false;
 839	rcu_report_qs_rdp(rdp);
 840	udelay(rcu_unlock_delay);
 841}
 842EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
 843
 844/*
 845 * Tell them what RCU they are running.
 846 */
 847static void __init rcu_bootup_announce(void)
 848{
 849	pr_info("Hierarchical RCU implementation.\n");
 850	rcu_bootup_announce_oddness();
 851}
 852
 853/*
 854 * Note a quiescent state for PREEMPTION=n.  Because we do not need to know
 855 * how many quiescent states passed, just if there was at least one since
 856 * the start of the grace period, this just sets a flag.  The caller must
 857 * have disabled preemption.
 858 */
 859static void rcu_qs(void)
 860{
 861	RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
 862	if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
 863		return;
 864	trace_rcu_grace_period(TPS("rcu_sched"),
 865			       __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
 866	__this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
 867	if (__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
 868		rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 869}
 870
 871/*
 872 * Register an urgently needed quiescent state.  If there is an
 873 * emergency, invoke rcu_momentary_eqs() to do a heavy-weight
 874 * dyntick-idle quiescent state visible to other CPUs, which will in
 875 * some cases serve for expedited as well as normal grace periods.
 876 * Either way, register a lightweight quiescent state.
 877 */
 878void rcu_all_qs(void)
 879{
 880	unsigned long flags;
 881
 882	if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
 883		return;
 884	preempt_disable();  // For CONFIG_PREEMPT_COUNT=y kernels
 885	/* Load rcu_urgent_qs before other flags. */
 886	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
 887		preempt_enable();
 888		return;
 889	}
 890	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 891	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
 892		local_irq_save(flags);
 893		rcu_momentary_eqs();
 894		local_irq_restore(flags);
 895	}
 896	rcu_qs();
 897	preempt_enable();
 898}
 899EXPORT_SYMBOL_GPL(rcu_all_qs);
 900
 901/*
 902 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
 903 */
 904void rcu_note_context_switch(bool preempt)
 905{
 906	trace_rcu_utilization(TPS("Start context switch"));
 907	rcu_qs();
 908	/* Load rcu_urgent_qs before other flags. */
 909	if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
 910		goto out;
 911	this_cpu_write(rcu_data.rcu_urgent_qs, false);
 912	if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
 913		rcu_momentary_eqs();
 914out:
 915	rcu_tasks_qs(current, preempt);
 916	trace_rcu_utilization(TPS("End context switch"));
 917}
 918EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 919
 920/*
 921 * Because preemptible RCU does not exist, there are never any preempted
 922 * RCU readers.
 923 */
 924static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 925{
 926	return 0;
 927}
 928
 929/*
 930 * Because there is no preemptible RCU, there can be no readers blocked.
 931 */
 932static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
 933{
 934	return false;
 935}
 936
 937/*
 938 * Because there is no preemptible RCU, there can be no deferred quiescent
 939 * states.
 940 */
 941static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
 942{
 943	return false;
 944}
 945
 946// Except that we do need to respond to a request by an expedited
 947// grace period for a quiescent state from this CPU.  Note that in
 948// non-preemptible kernels, there can be no context switches within RCU
 949// read-side critical sections, which in turn means that the leaf rcu_node
 950// structure's blocked-tasks list is always empty.  is therefore no need to
 951// actually check it.  Instead, a quiescent state from this CPU suffices,
 952// and this function is only called from such a quiescent state.
 953notrace void rcu_preempt_deferred_qs(struct task_struct *t)
 954{
 955	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 956
 957	if (READ_ONCE(rdp->cpu_no_qs.b.exp))
 958		rcu_report_exp_rdp(rdp);
 959}
 960
 961/*
 962 * Because there is no preemptible RCU, there can be no readers blocked,
 963 * so there is no need to check for blocked tasks.  So check only for
 964 * bogus qsmask values.
 965 */
 966static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 967{
 968	WARN_ON_ONCE(rnp->qsmask);
 969}
 970
 971/*
 972 * Check to see if this CPU is in a non-context-switch quiescent state,
 973 * namely user mode and idle loop.
 974 */
 975static void rcu_flavor_sched_clock_irq(int user)
 976{
 977	if (user || rcu_is_cpu_rrupt_from_idle()) {
 978
 979		/*
 980		 * Get here if this CPU took its interrupt from user
 981		 * mode or from the idle loop, and if this is not a
 982		 * nested interrupt.  In this case, the CPU is in
 983		 * a quiescent state, so note it.
 984		 *
 985		 * No memory barrier is required here because rcu_qs()
 986		 * references only CPU-local variables that other CPUs
 987		 * neither access nor modify, at least not while the
 988		 * corresponding CPU is online.
 989		 */
 990		rcu_qs();
 991	}
 992}
 993
 994/*
 995 * Because preemptible RCU does not exist, tasks cannot possibly exit
 996 * while in preemptible RCU read-side critical sections.
 997 */
 998void exit_rcu(void)
 999{
1000}
1001
1002/*
1003 * Dump the guaranteed-empty blocked-tasks state.  Trust but verify.
1004 */
1005static void
1006dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
1007{
1008	WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
1009}
1010
1011#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
1012
1013/*
1014 * If boosting, set rcuc kthreads to realtime priority.
1015 */
1016static void rcu_cpu_kthread_setup(unsigned int cpu)
1017{
1018	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1019#ifdef CONFIG_RCU_BOOST
1020	struct sched_param sp;
1021
1022	sp.sched_priority = kthread_prio;
1023	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1024#endif /* #ifdef CONFIG_RCU_BOOST */
1025
1026	WRITE_ONCE(rdp->rcuc_activity, jiffies);
1027}
1028
1029static bool rcu_is_callbacks_nocb_kthread(struct rcu_data *rdp)
1030{
1031#ifdef CONFIG_RCU_NOCB_CPU
1032	return rdp->nocb_cb_kthread == current;
1033#else
1034	return false;
1035#endif
1036}
1037
1038/*
1039 * Is the current CPU running the RCU-callbacks kthread?
1040 * Caller must have preemption disabled.
1041 */
1042static bool rcu_is_callbacks_kthread(struct rcu_data *rdp)
1043{
1044	return rdp->rcu_cpu_kthread_task == current ||
1045			rcu_is_callbacks_nocb_kthread(rdp);
1046}
1047
1048#ifdef CONFIG_RCU_BOOST
1049
1050/*
1051 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1052 * or ->boost_tasks, advancing the pointer to the next task in the
1053 * ->blkd_tasks list.
1054 *
1055 * Note that irqs must be enabled: boosting the task can block.
1056 * Returns 1 if there are more tasks needing to be boosted.
1057 */
1058static int rcu_boost(struct rcu_node *rnp)
1059{
1060	unsigned long flags;
1061	struct task_struct *t;
1062	struct list_head *tb;
1063
1064	if (READ_ONCE(rnp->exp_tasks) == NULL &&
1065	    READ_ONCE(rnp->boost_tasks) == NULL)
1066		return 0;  /* Nothing left to boost. */
1067
1068	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1069
1070	/*
1071	 * Recheck under the lock: all tasks in need of boosting
1072	 * might exit their RCU read-side critical sections on their own.
1073	 */
1074	if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1075		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1076		return 0;
1077	}
1078
1079	/*
1080	 * Preferentially boost tasks blocking expedited grace periods.
1081	 * This cannot starve the normal grace periods because a second
1082	 * expedited grace period must boost all blocked tasks, including
1083	 * those blocking the pre-existing normal grace period.
1084	 */
1085	if (rnp->exp_tasks != NULL)
1086		tb = rnp->exp_tasks;
1087	else
1088		tb = rnp->boost_tasks;
1089
1090	/*
1091	 * We boost task t by manufacturing an rt_mutex that appears to
1092	 * be held by task t.  We leave a pointer to that rt_mutex where
1093	 * task t can find it, and task t will release the mutex when it
1094	 * exits its outermost RCU read-side critical section.  Then
1095	 * simply acquiring this artificial rt_mutex will boost task
1096	 * t's priority.  (Thanks to tglx for suggesting this approach!)
1097	 *
1098	 * Note that task t must acquire rnp->lock to remove itself from
1099	 * the ->blkd_tasks list, which it will do from exit() if from
1100	 * nowhere else.  We therefore are guaranteed that task t will
1101	 * stay around at least until we drop rnp->lock.  Note that
1102	 * rnp->lock also resolves races between our priority boosting
1103	 * and task t's exiting its outermost RCU read-side critical
1104	 * section.
1105	 */
1106	t = container_of(tb, struct task_struct, rcu_node_entry);
1107	rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
1108	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1109	/* Lock only for side effect: boosts task t's priority. */
1110	rt_mutex_lock(&rnp->boost_mtx);
1111	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1112	rnp->n_boosts++;
1113
1114	return READ_ONCE(rnp->exp_tasks) != NULL ||
1115	       READ_ONCE(rnp->boost_tasks) != NULL;
1116}
1117
1118/*
1119 * Priority-boosting kthread, one per leaf rcu_node.
1120 */
1121static int rcu_boost_kthread(void *arg)
1122{
1123	struct rcu_node *rnp = (struct rcu_node *)arg;
1124	int spincnt = 0;
1125	int more2boost;
1126
1127	trace_rcu_utilization(TPS("Start boost kthread@init"));
1128	for (;;) {
1129		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1130		trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1131		rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1132			 READ_ONCE(rnp->exp_tasks));
1133		trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1134		WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1135		more2boost = rcu_boost(rnp);
1136		if (more2boost)
1137			spincnt++;
1138		else
1139			spincnt = 0;
1140		if (spincnt > 10) {
1141			WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1142			trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1143			schedule_timeout_idle(2);
1144			trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1145			spincnt = 0;
1146		}
1147	}
1148	/* NOTREACHED */
1149	trace_rcu_utilization(TPS("End boost kthread@notreached"));
1150	return 0;
1151}
1152
1153/*
1154 * Check to see if it is time to start boosting RCU readers that are
1155 * blocking the current grace period, and, if so, tell the per-rcu_node
1156 * kthread to start boosting them.  If there is an expedited grace
1157 * period in progress, it is always time to boost.
1158 *
1159 * The caller must hold rnp->lock, which this function releases.
1160 * The ->boost_kthread_task is immortal, so we don't need to worry
1161 * about it going away.
1162 */
1163static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1164	__releases(rnp->lock)
1165{
1166	raw_lockdep_assert_held_rcu_node(rnp);
1167	if (!rnp->boost_kthread_task ||
1168	    (!rcu_preempt_blocked_readers_cgp(rnp) && !rnp->exp_tasks)) {
1169		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1170		return;
1171	}
1172	if (rnp->exp_tasks != NULL ||
1173	    (rnp->gp_tasks != NULL &&
1174	     rnp->boost_tasks == NULL &&
1175	     rnp->qsmask == 0 &&
1176	     (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld ||
1177	      IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)))) {
1178		if (rnp->exp_tasks == NULL)
1179			WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1180		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1181		rcu_wake_cond(rnp->boost_kthread_task,
1182			      READ_ONCE(rnp->boost_kthread_status));
1183	} else {
1184		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1185	}
1186}
1187
1188#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1189
1190/*
1191 * Do priority-boost accounting for the start of a new grace period.
1192 */
1193static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1194{
1195	rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1196}
1197
1198/*
1199 * Create an RCU-boost kthread for the specified node if one does not
1200 * already exist.  We only create this kthread for preemptible RCU.
1201 */
1202static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1203{
1204	unsigned long flags;
1205	int rnp_index = rnp - rcu_get_root();
1206	struct sched_param sp;
1207	struct task_struct *t;
1208
1209	if (rnp->boost_kthread_task)
1210		return;
 
1211
1212	t = kthread_create(rcu_boost_kthread, (void *)rnp,
1213			   "rcub/%d", rnp_index);
1214	if (WARN_ON_ONCE(IS_ERR(t)))
1215		return;
1216
1217	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1218	rnp->boost_kthread_task = t;
1219	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1220	sp.sched_priority = kthread_prio;
1221	sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1222	wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
 
 
 
1223}
1224
1225static struct task_struct *rcu_boost_task(struct rcu_node *rnp)
 
 
 
 
 
 
 
 
 
 
 
1226{
1227	return READ_ONCE(rnp->boost_kthread_task);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1228}
1229
1230#else /* #ifdef CONFIG_RCU_BOOST */
1231
1232static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1233	__releases(rnp->lock)
1234{
1235	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1236}
1237
1238static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1239{
1240}
1241
1242static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1243{
1244}
1245
1246static struct task_struct *rcu_boost_task(struct rcu_node *rnp)
1247{
1248	return NULL;
1249}
 
1250#endif /* #else #ifdef CONFIG_RCU_BOOST */
1251
1252/*
1253 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
1254 * grace-period kthread will do force_quiescent_state() processing?
1255 * The idea is to avoid waking up RCU core processing on such a
1256 * CPU unless the grace period has extended for too long.
1257 *
1258 * This code relies on the fact that all NO_HZ_FULL CPUs are also
1259 * RCU_NOCB_CPU CPUs.
1260 */
1261static bool rcu_nohz_full_cpu(void)
1262{
1263#ifdef CONFIG_NO_HZ_FULL
1264	if (tick_nohz_full_cpu(smp_processor_id()) &&
1265	    (!rcu_gp_in_progress() ||
1266	     time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
1267		return true;
1268#endif /* #ifdef CONFIG_NO_HZ_FULL */
1269	return false;
1270}
1271
1272/*
1273 * Bind the RCU grace-period kthreads to the housekeeping CPU.
1274 */
1275static void rcu_bind_gp_kthread(void)
1276{
1277	if (!tick_nohz_full_enabled())
1278		return;
1279	housekeeping_affine(current, HK_TYPE_RCU);
1280}