1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
17 *
18 * Copyright IBM Corporation, 2001
19 *
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 *	    Manfred Spraul <manfred@colorfullife.com>
22 *
23 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
24 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
25 * Papers:
26 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
27 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
28 *
29 * For detailed explanation of Read-Copy Update mechanism see -
30 *		http://lse.sourceforge.net/locking/rcupdate.html
31 *
32 */
33#include <linux/types.h>
34#include <linux/kernel.h>
35#include <linux/init.h>
36#include <linux/spinlock.h>
37#include <linux/smp.h>
38#include <linux/interrupt.h>
39#include <linux/sched.h>
40#include <linux/atomic.h>
41#include <linux/bitops.h>
42#include <linux/percpu.h>
43#include <linux/notifier.h>
44#include <linux/cpu.h>
45#include <linux/mutex.h>
46#include <linux/export.h>
47#include <linux/hardirq.h>
48#include <linux/delay.h>
49#include <linux/module.h>
50#include <linux/kthread.h>
51#include <linux/tick.h>
52
53#define CREATE_TRACE_POINTS
54
55#include "rcu.h"
56
57MODULE_ALIAS("rcupdate");
58#ifdef MODULE_PARAM_PREFIX
59#undef MODULE_PARAM_PREFIX
60#endif
61#define MODULE_PARAM_PREFIX "rcupdate."
62
63module_param(rcu_expedited, int, 0);
64
65#ifndef CONFIG_TINY_RCU
66
67static atomic_t rcu_expedited_nesting =
68	ATOMIC_INIT(IS_ENABLED(CONFIG_RCU_EXPEDITE_BOOT) ? 1 : 0);
69
70/*
71 * Should normal grace-period primitives be expedited?  Intended for
72 * use within RCU.  Note that this function takes the rcu_expedited
73 * sysfs/boot variable into account as well as the rcu_expedite_gp()
74 * nesting.  So looping on rcu_unexpedite_gp() until rcu_gp_is_expedited()
75 * returns false is a -really- bad idea.
76 */
77bool rcu_gp_is_expedited(void)
78{
79	return rcu_expedited || atomic_read(&rcu_expedited_nesting);
80}
81EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
82
83/**
84 * rcu_expedite_gp - Expedite future RCU grace periods
85 *
86 * After a call to this function, future calls to synchronize_rcu() and
87 * friends act as the corresponding synchronize_rcu_expedited() function
88 * had instead been called.
89 */
90void rcu_expedite_gp(void)
91{
92	atomic_inc(&rcu_expedited_nesting);
93}
94EXPORT_SYMBOL_GPL(rcu_expedite_gp);
95
96/**
97 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
98 *
99 * Undo a prior call to rcu_expedite_gp().  If all prior calls to
100 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
101 * and if the rcu_expedited sysfs/boot parameter is not set, then all
102 * subsequent calls to synchronize_rcu() and friends will return to
103 * their normal non-expedited behavior.
104 */
105void rcu_unexpedite_gp(void)
106{
107	atomic_dec(&rcu_expedited_nesting);
108}
109EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
110
111#endif /* #ifndef CONFIG_TINY_RCU */
112
113/*
114 * Inform RCU of the end of the in-kernel boot sequence.
115 */
116void rcu_end_inkernel_boot(void)
117{
118	if (IS_ENABLED(CONFIG_RCU_EXPEDITE_BOOT))
119		rcu_unexpedite_gp();
120}
121
122#ifdef CONFIG_PREEMPT_RCU
123
124/*
125 * Preemptible RCU implementation for rcu_read_lock().
126 * Just increment ->rcu_read_lock_nesting, shared state will be updated
127 * if we block.
128 */
129void __rcu_read_lock(void)
130{
131	current->rcu_read_lock_nesting++;
132	barrier();  /* critical section after entry code. */
133}
134EXPORT_SYMBOL_GPL(__rcu_read_lock);
135
136/*
137 * Preemptible RCU implementation for rcu_read_unlock().
138 * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
139 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
140 * invoke rcu_read_unlock_special() to clean up after a context switch
141 * in an RCU read-side critical section and other special cases.
142 */
143void __rcu_read_unlock(void)
144{
145	struct task_struct *t = current;
146
147	if (t->rcu_read_lock_nesting != 1) {
148		--t->rcu_read_lock_nesting;
149	} else {
150		barrier();  /* critical section before exit code. */
151		t->rcu_read_lock_nesting = INT_MIN;
152		barrier();  /* assign before ->rcu_read_unlock_special load */
153		if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special.s)))
154			rcu_read_unlock_special(t);
155		barrier();  /* ->rcu_read_unlock_special load before assign */
156		t->rcu_read_lock_nesting = 0;
157	}
158#ifdef CONFIG_PROVE_LOCKING
159	{
160		int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting);
161
162		WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
163	}
164#endif /* #ifdef CONFIG_PROVE_LOCKING */
165}
166EXPORT_SYMBOL_GPL(__rcu_read_unlock);
167
168#endif /* #ifdef CONFIG_PREEMPT_RCU */
169
170#ifdef CONFIG_DEBUG_LOCK_ALLOC
171static struct lock_class_key rcu_lock_key;
172struct lockdep_map rcu_lock_map =
173	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
174EXPORT_SYMBOL_GPL(rcu_lock_map);
175
176static struct lock_class_key rcu_bh_lock_key;
177struct lockdep_map rcu_bh_lock_map =
178	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
179EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
180
181static struct lock_class_key rcu_sched_lock_key;
182struct lockdep_map rcu_sched_lock_map =
183	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
184EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
185
186static struct lock_class_key rcu_callback_key;
187struct lockdep_map rcu_callback_map =
188	STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
189EXPORT_SYMBOL_GPL(rcu_callback_map);
190
191int notrace debug_lockdep_rcu_enabled(void)
192{
193	return rcu_scheduler_active && debug_locks &&
194	       current->lockdep_recursion == 0;
195}
196EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
197
198/**
199 * rcu_read_lock_held() - might we be in RCU read-side critical section?
200 *
201 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
202 * read-side critical section.  In absence of CONFIG_DEBUG_LOCK_ALLOC,
203 * this assumes we are in an RCU read-side critical section unless it can
204 * prove otherwise.  This is useful for debug checks in functions that
205 * require that they be called within an RCU read-side critical section.
206 *
207 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
208 * and while lockdep is disabled.
209 *
210 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
211 * occur in the same context, for example, it is illegal to invoke
212 * rcu_read_unlock() in process context if the matching rcu_read_lock()
213 * was invoked from within an irq handler.
214 *
215 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
216 * offline from an RCU perspective, so check for those as well.
217 */
218int rcu_read_lock_held(void)
219{
220	if (!debug_lockdep_rcu_enabled())
221		return 1;
222	if (!rcu_is_watching())
223		return 0;
224	if (!rcu_lockdep_current_cpu_online())
225		return 0;
226	return lock_is_held(&rcu_lock_map);
227}
228EXPORT_SYMBOL_GPL(rcu_read_lock_held);
229
230/**
231 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
232 *
233 * Check for bottom half being disabled, which covers both the
234 * CONFIG_PROVE_RCU and not cases.  Note that if someone uses
235 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
236 * will show the situation.  This is useful for debug checks in functions
237 * that require that they be called within an RCU read-side critical
238 * section.
239 *
240 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
241 *
242 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
243 * offline from an RCU perspective, so check for those as well.
244 */
245int rcu_read_lock_bh_held(void)
246{
247	if (!debug_lockdep_rcu_enabled())
248		return 1;
249	if (!rcu_is_watching())
250		return 0;
251	if (!rcu_lockdep_current_cpu_online())
252		return 0;
253	return in_softirq() || irqs_disabled();
254}
255EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
256
257#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
258
259/**
260 * wakeme_after_rcu() - Callback function to awaken a task after grace period
261 * @head: Pointer to rcu_head member within rcu_synchronize structure
262 *
263 * Awaken the corresponding task now that a grace period has elapsed.
264 */
265void wakeme_after_rcu(struct rcu_head *head)
266{
267	struct rcu_synchronize *rcu;
268
269	rcu = container_of(head, struct rcu_synchronize, head);
270	complete(&rcu->completion);
271}
272
273void wait_rcu_gp(call_rcu_func_t crf)
274{
275	struct rcu_synchronize rcu;
276
277	init_rcu_head_on_stack(&rcu.head);
278	init_completion(&rcu.completion);
279	/* Will wake me after RCU finished. */
280	crf(&rcu.head, wakeme_after_rcu);
281	/* Wait for it. */
282	wait_for_completion(&rcu.completion);
283	destroy_rcu_head_on_stack(&rcu.head);
284}
285EXPORT_SYMBOL_GPL(wait_rcu_gp);
286
287#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
288void init_rcu_head(struct rcu_head *head)
289{
290	debug_object_init(head, &rcuhead_debug_descr);
291}
292
293void destroy_rcu_head(struct rcu_head *head)
294{
295	debug_object_free(head, &rcuhead_debug_descr);
296}
297
298/*
299 * fixup_activate is called when:
300 * - an active object is activated
301 * - an unknown object is activated (might be a statically initialized object)
302 * Activation is performed internally by call_rcu().
303 */
304static int rcuhead_fixup_activate(void *addr, enum debug_obj_state state)
305{
306	struct rcu_head *head = addr;
307
308	switch (state) {
309
310	case ODEBUG_STATE_NOTAVAILABLE:
311		/*
312		 * This is not really a fixup. We just make sure that it is
313		 * tracked in the object tracker.
314		 */
315		debug_object_init(head, &rcuhead_debug_descr);
316		debug_object_activate(head, &rcuhead_debug_descr);
317		return 0;
318	default:
319		return 1;
320	}
321}
322
323/**
324 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
325 * @head: pointer to rcu_head structure to be initialized
326 *
327 * This function informs debugobjects of a new rcu_head structure that
328 * has been allocated as an auto variable on the stack.  This function
329 * is not required for rcu_head structures that are statically defined or
330 * that are dynamically allocated on the heap.  This function has no
331 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
332 */
333void init_rcu_head_on_stack(struct rcu_head *head)
334{
335	debug_object_init_on_stack(head, &rcuhead_debug_descr);
336}
337EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
338
339/**
340 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
341 * @head: pointer to rcu_head structure to be initialized
342 *
343 * This function informs debugobjects that an on-stack rcu_head structure
344 * is about to go out of scope.  As with init_rcu_head_on_stack(), this
345 * function is not required for rcu_head structures that are statically
346 * defined or that are dynamically allocated on the heap.  Also as with
347 * init_rcu_head_on_stack(), this function has no effect for
348 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
349 */
350void destroy_rcu_head_on_stack(struct rcu_head *head)
351{
352	debug_object_free(head, &rcuhead_debug_descr);
353}
354EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
355
356struct debug_obj_descr rcuhead_debug_descr = {
357	.name = "rcu_head",
358	.fixup_activate = rcuhead_fixup_activate,
359};
360EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
361#endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
362
363#if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
364void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
365			       unsigned long secs,
366			       unsigned long c_old, unsigned long c)
367{
368	trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
369}
370EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
371#else
372#define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
373	do { } while (0)
374#endif
375
376#ifdef CONFIG_RCU_STALL_COMMON
377
378#ifdef CONFIG_PROVE_RCU
379#define RCU_STALL_DELAY_DELTA	       (5 * HZ)
380#else
381#define RCU_STALL_DELAY_DELTA	       0
382#endif
383
384int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
385static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
386
387module_param(rcu_cpu_stall_suppress, int, 0644);
388module_param(rcu_cpu_stall_timeout, int, 0644);
389
390int rcu_jiffies_till_stall_check(void)
391{
392	int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
393
394	/*
395	 * Limit check must be consistent with the Kconfig limits
396	 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
397	 */
398	if (till_stall_check < 3) {
399		ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
400		till_stall_check = 3;
401	} else if (till_stall_check > 300) {
402		ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
403		till_stall_check = 300;
404	}
405	return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
406}
407
408void rcu_sysrq_start(void)
409{
410	if (!rcu_cpu_stall_suppress)
411		rcu_cpu_stall_suppress = 2;
412}
413
414void rcu_sysrq_end(void)
415{
416	if (rcu_cpu_stall_suppress == 2)
417		rcu_cpu_stall_suppress = 0;
418}
419
420static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
421{
422	rcu_cpu_stall_suppress = 1;
423	return NOTIFY_DONE;
424}
425
426static struct notifier_block rcu_panic_block = {
427	.notifier_call = rcu_panic,
428};
429
430static int __init check_cpu_stall_init(void)
431{
432	atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
433	return 0;
434}
435early_initcall(check_cpu_stall_init);
436
437#endif /* #ifdef CONFIG_RCU_STALL_COMMON */
438
439#ifdef CONFIG_TASKS_RCU
440
441/*
442 * Simple variant of RCU whose quiescent states are voluntary context switch,
443 * user-space execution, and idle.  As such, grace periods can take one good
444 * long time.  There are no read-side primitives similar to rcu_read_lock()
445 * and rcu_read_unlock() because this implementation is intended to get
446 * the system into a safe state for some of the manipulations involved in
447 * tracing and the like.  Finally, this implementation does not support
448 * high call_rcu_tasks() rates from multiple CPUs.  If this is required,
449 * per-CPU callback lists will be needed.
450 */
451
452/* Global list of callbacks and associated lock. */
453static struct rcu_head *rcu_tasks_cbs_head;
454static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
455static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
456static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
457
458/* Track exiting tasks in order to allow them to be waited for. */
459DEFINE_SRCU(tasks_rcu_exit_srcu);
460
461/* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
462static int rcu_task_stall_timeout __read_mostly = HZ * 60 * 10;
463module_param(rcu_task_stall_timeout, int, 0644);
464
465static void rcu_spawn_tasks_kthread(void);
466
467/*
468 * Post an RCU-tasks callback.  First call must be from process context
469 * after the scheduler if fully operational.
470 */
471void call_rcu_tasks(struct rcu_head *rhp, void (*func)(struct rcu_head *rhp))
472{
473	unsigned long flags;
474	bool needwake;
475
476	rhp->next = NULL;
477	rhp->func = func;
478	raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
479	needwake = !rcu_tasks_cbs_head;
480	*rcu_tasks_cbs_tail = rhp;
481	rcu_tasks_cbs_tail = &rhp->next;
482	raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
483	if (needwake) {
484		rcu_spawn_tasks_kthread();
485		wake_up(&rcu_tasks_cbs_wq);
486	}
487}
488EXPORT_SYMBOL_GPL(call_rcu_tasks);
489
490/**
491 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
492 *
493 * Control will return to the caller some time after a full rcu-tasks
494 * grace period has elapsed, in other words after all currently
495 * executing rcu-tasks read-side critical sections have elapsed.  These
496 * read-side critical sections are delimited by calls to schedule(),
497 * cond_resched_rcu_qs(), idle execution, userspace execution, calls
498 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
499 *
500 * This is a very specialized primitive, intended only for a few uses in
501 * tracing and other situations requiring manipulation of function
502 * preambles and profiling hooks.  The synchronize_rcu_tasks() function
503 * is not (yet) intended for heavy use from multiple CPUs.
504 *
505 * Note that this guarantee implies further memory-ordering guarantees.
506 * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
507 * each CPU is guaranteed to have executed a full memory barrier since the
508 * end of its last RCU-tasks read-side critical section whose beginning
509 * preceded the call to synchronize_rcu_tasks().  In addition, each CPU
510 * having an RCU-tasks read-side critical section that extends beyond
511 * the return from synchronize_rcu_tasks() is guaranteed to have executed
512 * a full memory barrier after the beginning of synchronize_rcu_tasks()
513 * and before the beginning of that RCU-tasks read-side critical section.
514 * Note that these guarantees include CPUs that are offline, idle, or
515 * executing in user mode, as well as CPUs that are executing in the kernel.
516 *
517 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
518 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
519 * to have executed a full memory barrier during the execution of
520 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
521 * (but again only if the system has more than one CPU).
522 */
523void synchronize_rcu_tasks(void)
524{
525	/* Complain if the scheduler has not started.  */
526	rcu_lockdep_assert(!rcu_scheduler_active,
527			   "synchronize_rcu_tasks called too soon");
528
529	/* Wait for the grace period. */
530	wait_rcu_gp(call_rcu_tasks);
531}
532EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
533
534/**
535 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
536 *
537 * Although the current implementation is guaranteed to wait, it is not
538 * obligated to, for example, if there are no pending callbacks.
539 */
540void rcu_barrier_tasks(void)
541{
542	/* There is only one callback queue, so this is easy.  ;-) */
543	synchronize_rcu_tasks();
544}
545EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
546
547/* See if tasks are still holding out, complain if so. */
548static void check_holdout_task(struct task_struct *t,
549			       bool needreport, bool *firstreport)
550{
551	int cpu;
552
553	if (!ACCESS_ONCE(t->rcu_tasks_holdout) ||
554	    t->rcu_tasks_nvcsw != ACCESS_ONCE(t->nvcsw) ||
555	    !ACCESS_ONCE(t->on_rq) ||
556	    (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
557	     !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
558		ACCESS_ONCE(t->rcu_tasks_holdout) = false;
559		list_del_init(&t->rcu_tasks_holdout_list);
560		put_task_struct(t);
561		return;
562	}
563	if (!needreport)
564		return;
565	if (*firstreport) {
566		pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
567		*firstreport = false;
568	}
569	cpu = task_cpu(t);
570	pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
571		 t, ".I"[is_idle_task(t)],
572		 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
573		 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
574		 t->rcu_tasks_idle_cpu, cpu);
575	sched_show_task(t);
576}
577
578/* RCU-tasks kthread that detects grace periods and invokes callbacks. */
579static int __noreturn rcu_tasks_kthread(void *arg)
580{
581	unsigned long flags;
582	struct task_struct *g, *t;
583	unsigned long lastreport;
584	struct rcu_head *list;
585	struct rcu_head *next;
586	LIST_HEAD(rcu_tasks_holdouts);
587
588	/* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
589	housekeeping_affine(current);
590
591	/*
592	 * Each pass through the following loop makes one check for
593	 * newly arrived callbacks, and, if there are some, waits for
594	 * one RCU-tasks grace period and then invokes the callbacks.
595	 * This loop is terminated by the system going down.  ;-)
596	 */
597	for (;;) {
598
599		/* Pick up any new callbacks. */
600		raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
601		list = rcu_tasks_cbs_head;
602		rcu_tasks_cbs_head = NULL;
603		rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
604		raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
605
606		/* If there were none, wait a bit and start over. */
607		if (!list) {
608			wait_event_interruptible(rcu_tasks_cbs_wq,
609						 rcu_tasks_cbs_head);
610			if (!rcu_tasks_cbs_head) {
611				WARN_ON(signal_pending(current));
612				schedule_timeout_interruptible(HZ/10);
613			}
614			continue;
615		}
616
617		/*
618		 * Wait for all pre-existing t->on_rq and t->nvcsw
619		 * transitions to complete.  Invoking synchronize_sched()
620		 * suffices because all these transitions occur with
621		 * interrupts disabled.  Without this synchronize_sched(),
622		 * a read-side critical section that started before the
623		 * grace period might be incorrectly seen as having started
624		 * after the grace period.
625		 *
626		 * This synchronize_sched() also dispenses with the
627		 * need for a memory barrier on the first store to
628		 * ->rcu_tasks_holdout, as it forces the store to happen
629		 * after the beginning of the grace period.
630		 */
631		synchronize_sched();
632
633		/*
634		 * There were callbacks, so we need to wait for an
635		 * RCU-tasks grace period.  Start off by scanning
636		 * the task list for tasks that are not already
637		 * voluntarily blocked.  Mark these tasks and make
638		 * a list of them in rcu_tasks_holdouts.
639		 */
640		rcu_read_lock();
641		for_each_process_thread(g, t) {
642			if (t != current && ACCESS_ONCE(t->on_rq) &&
643			    !is_idle_task(t)) {
644				get_task_struct(t);
645				t->rcu_tasks_nvcsw = ACCESS_ONCE(t->nvcsw);
646				ACCESS_ONCE(t->rcu_tasks_holdout) = true;
647				list_add(&t->rcu_tasks_holdout_list,
648					 &rcu_tasks_holdouts);
649			}
650		}
651		rcu_read_unlock();
652
653		/*
654		 * Wait for tasks that are in the process of exiting.
655		 * This does only part of the job, ensuring that all
656		 * tasks that were previously exiting reach the point
657		 * where they have disabled preemption, allowing the
658		 * later synchronize_sched() to finish the job.
659		 */
660		synchronize_srcu(&tasks_rcu_exit_srcu);
661
662		/*
663		 * Each pass through the following loop scans the list
664		 * of holdout tasks, removing any that are no longer
665		 * holdouts.  When the list is empty, we are done.
666		 */
667		lastreport = jiffies;
668		while (!list_empty(&rcu_tasks_holdouts)) {
669			bool firstreport;
670			bool needreport;
671			int rtst;
672			struct task_struct *t1;
673
674			schedule_timeout_interruptible(HZ);
675			rtst = ACCESS_ONCE(rcu_task_stall_timeout);
676			needreport = rtst > 0 &&
677				     time_after(jiffies, lastreport + rtst);
678			if (needreport)
679				lastreport = jiffies;
680			firstreport = true;
681			WARN_ON(signal_pending(current));
682			list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
683						rcu_tasks_holdout_list) {
684				check_holdout_task(t, needreport, &firstreport);
685				cond_resched();
686			}
687		}
688
689		/*
690		 * Because ->on_rq and ->nvcsw are not guaranteed
691		 * to have a full memory barriers prior to them in the
692		 * schedule() path, memory reordering on other CPUs could
693		 * cause their RCU-tasks read-side critical sections to
694		 * extend past the end of the grace period.  However,
695		 * because these ->nvcsw updates are carried out with
696		 * interrupts disabled, we can use synchronize_sched()
697		 * to force the needed ordering on all such CPUs.
698		 *
699		 * This synchronize_sched() also confines all
700		 * ->rcu_tasks_holdout accesses to be within the grace
701		 * period, avoiding the need for memory barriers for
702		 * ->rcu_tasks_holdout accesses.
703		 *
704		 * In addition, this synchronize_sched() waits for exiting
705		 * tasks to complete their final preempt_disable() region
706		 * of execution, cleaning up after the synchronize_srcu()
707		 * above.
708		 */
709		synchronize_sched();
710
711		/* Invoke the callbacks. */
712		while (list) {
713			next = list->next;
714			local_bh_disable();
715			list->func(list);
716			local_bh_enable();
717			list = next;
718			cond_resched();
719		}
720		schedule_timeout_uninterruptible(HZ/10);
721	}
722}
723
724/* Spawn rcu_tasks_kthread() at first call to call_rcu_tasks(). */
725static void rcu_spawn_tasks_kthread(void)
726{
727	static DEFINE_MUTEX(rcu_tasks_kthread_mutex);
728	static struct task_struct *rcu_tasks_kthread_ptr;
729	struct task_struct *t;
730
731	if (ACCESS_ONCE(rcu_tasks_kthread_ptr)) {
732		smp_mb(); /* Ensure caller sees full kthread. */
733		return;
734	}
735	mutex_lock(&rcu_tasks_kthread_mutex);
736	if (rcu_tasks_kthread_ptr) {
737		mutex_unlock(&rcu_tasks_kthread_mutex);
738		return;
739	}
740	t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
741	BUG_ON(IS_ERR(t));
742	smp_mb(); /* Ensure others see full kthread. */
743	ACCESS_ONCE(rcu_tasks_kthread_ptr) = t;
744	mutex_unlock(&rcu_tasks_kthread_mutex);
745}
746
747#endif /* #ifdef CONFIG_TASKS_RCU */
748
749#ifdef CONFIG_PROVE_RCU
750
751/*
752 * Early boot self test parameters, one for each flavor
753 */
754static bool rcu_self_test;
755static bool rcu_self_test_bh;
756static bool rcu_self_test_sched;
757
758module_param(rcu_self_test, bool, 0444);
759module_param(rcu_self_test_bh, bool, 0444);
760module_param(rcu_self_test_sched, bool, 0444);
761
762static int rcu_self_test_counter;
763
764static void test_callback(struct rcu_head *r)
765{
766	rcu_self_test_counter++;
767	pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
768}
769
770static void early_boot_test_call_rcu(void)
771{
772	static struct rcu_head head;
773
774	call_rcu(&head, test_callback);
775}
776
777static void early_boot_test_call_rcu_bh(void)
778{
779	static struct rcu_head head;
780
781	call_rcu_bh(&head, test_callback);
782}
783
784static void early_boot_test_call_rcu_sched(void)
785{
786	static struct rcu_head head;
787
788	call_rcu_sched(&head, test_callback);
789}
790
791void rcu_early_boot_tests(void)
792{
793	pr_info("Running RCU self tests\n");
794
795	if (rcu_self_test)
796		early_boot_test_call_rcu();
797	if (rcu_self_test_bh)
798		early_boot_test_call_rcu_bh();
799	if (rcu_self_test_sched)
800		early_boot_test_call_rcu_sched();
801}
802
803static int rcu_verify_early_boot_tests(void)
804{
805	int ret = 0;
806	int early_boot_test_counter = 0;
807
808	if (rcu_self_test) {
809		early_boot_test_counter++;
810		rcu_barrier();
811	}
812	if (rcu_self_test_bh) {
813		early_boot_test_counter++;
814		rcu_barrier_bh();
815	}
816	if (rcu_self_test_sched) {
817		early_boot_test_counter++;
818		rcu_barrier_sched();
819	}
820
821	if (rcu_self_test_counter != early_boot_test_counter) {
822		WARN_ON(1);
823		ret = -1;
824	}
825
826	return ret;
827}
828late_initcall(rcu_verify_early_boot_tests);
829#else
830void rcu_early_boot_tests(void) {}
831#endif /* CONFIG_PROVE_RCU */
832