xref: /linux/kernel/rcu/update.c (revision cc04a46f11ea046ed53e2c832ae29e4790f7e35f)
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 
57 MODULE_ALIAS("rcupdate");
58 #ifdef MODULE_PARAM_PREFIX
59 #undef MODULE_PARAM_PREFIX
60 #endif
61 #define MODULE_PARAM_PREFIX "rcupdate."
62 
63 module_param(rcu_expedited, int, 0);
64 
65 #if defined(CONFIG_DEBUG_LOCK_ALLOC) && defined(CONFIG_PREEMPT_COUNT)
66 /**
67  * rcu_read_lock_sched_held() - might we be in RCU-sched read-side critical section?
68  *
69  * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an
70  * RCU-sched read-side critical section.  In absence of
71  * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
72  * critical section unless it can prove otherwise.  Note that disabling
73  * of preemption (including disabling irqs) counts as an RCU-sched
74  * read-side critical section.  This is useful for debug checks in functions
75  * that required that they be called within an RCU-sched read-side
76  * critical section.
77  *
78  * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
79  * and while lockdep is disabled.
80  *
81  * Note that if the CPU is in the idle loop from an RCU point of
82  * view (ie: that we are in the section between rcu_idle_enter() and
83  * rcu_idle_exit()) then rcu_read_lock_held() returns false even if the CPU
84  * did an rcu_read_lock().  The reason for this is that RCU ignores CPUs
85  * that are in such a section, considering these as in extended quiescent
86  * state, so such a CPU is effectively never in an RCU read-side critical
87  * section regardless of what RCU primitives it invokes.  This state of
88  * affairs is required --- we need to keep an RCU-free window in idle
89  * where the CPU may possibly enter into low power mode. This way we can
90  * notice an extended quiescent state to other CPUs that started a grace
91  * period. Otherwise we would delay any grace period as long as we run in
92  * the idle task.
93  *
94  * Similarly, we avoid claiming an SRCU read lock held if the current
95  * CPU is offline.
96  */
97 int rcu_read_lock_sched_held(void)
98 {
99 	int lockdep_opinion = 0;
100 
101 	if (!debug_lockdep_rcu_enabled())
102 		return 1;
103 	if (!rcu_is_watching())
104 		return 0;
105 	if (!rcu_lockdep_current_cpu_online())
106 		return 0;
107 	if (debug_locks)
108 		lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
109 	return lockdep_opinion || preempt_count() != 0 || irqs_disabled();
110 }
111 EXPORT_SYMBOL(rcu_read_lock_sched_held);
112 #endif
113 
114 #ifndef CONFIG_TINY_RCU
115 
116 static atomic_t rcu_expedited_nesting =
117 	ATOMIC_INIT(IS_ENABLED(CONFIG_RCU_EXPEDITE_BOOT) ? 1 : 0);
118 
119 /*
120  * Should normal grace-period primitives be expedited?  Intended for
121  * use within RCU.  Note that this function takes the rcu_expedited
122  * sysfs/boot variable into account as well as the rcu_expedite_gp()
123  * nesting.  So looping on rcu_unexpedite_gp() until rcu_gp_is_expedited()
124  * returns false is a -really- bad idea.
125  */
126 bool rcu_gp_is_expedited(void)
127 {
128 	return rcu_expedited || atomic_read(&rcu_expedited_nesting);
129 }
130 EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
131 
132 /**
133  * rcu_expedite_gp - Expedite future RCU grace periods
134  *
135  * After a call to this function, future calls to synchronize_rcu() and
136  * friends act as the corresponding synchronize_rcu_expedited() function
137  * had instead been called.
138  */
139 void rcu_expedite_gp(void)
140 {
141 	atomic_inc(&rcu_expedited_nesting);
142 }
143 EXPORT_SYMBOL_GPL(rcu_expedite_gp);
144 
145 /**
146  * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
147  *
148  * Undo a prior call to rcu_expedite_gp().  If all prior calls to
149  * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
150  * and if the rcu_expedited sysfs/boot parameter is not set, then all
151  * subsequent calls to synchronize_rcu() and friends will return to
152  * their normal non-expedited behavior.
153  */
154 void rcu_unexpedite_gp(void)
155 {
156 	atomic_dec(&rcu_expedited_nesting);
157 }
158 EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
159 
160 #endif /* #ifndef CONFIG_TINY_RCU */
161 
162 /*
163  * Inform RCU of the end of the in-kernel boot sequence.
164  */
165 void rcu_end_inkernel_boot(void)
166 {
167 	if (IS_ENABLED(CONFIG_RCU_EXPEDITE_BOOT))
168 		rcu_unexpedite_gp();
169 }
170 
171 #ifdef CONFIG_PREEMPT_RCU
172 
173 /*
174  * Preemptible RCU implementation for rcu_read_lock().
175  * Just increment ->rcu_read_lock_nesting, shared state will be updated
176  * if we block.
177  */
178 void __rcu_read_lock(void)
179 {
180 	current->rcu_read_lock_nesting++;
181 	barrier();  /* critical section after entry code. */
182 }
183 EXPORT_SYMBOL_GPL(__rcu_read_lock);
184 
185 /*
186  * Preemptible RCU implementation for rcu_read_unlock().
187  * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
188  * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
189  * invoke rcu_read_unlock_special() to clean up after a context switch
190  * in an RCU read-side critical section and other special cases.
191  */
192 void __rcu_read_unlock(void)
193 {
194 	struct task_struct *t = current;
195 
196 	if (t->rcu_read_lock_nesting != 1) {
197 		--t->rcu_read_lock_nesting;
198 	} else {
199 		barrier();  /* critical section before exit code. */
200 		t->rcu_read_lock_nesting = INT_MIN;
201 		barrier();  /* assign before ->rcu_read_unlock_special load */
202 		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
203 			rcu_read_unlock_special(t);
204 		barrier();  /* ->rcu_read_unlock_special load before assign */
205 		t->rcu_read_lock_nesting = 0;
206 	}
207 #ifdef CONFIG_PROVE_LOCKING
208 	{
209 		int rrln = READ_ONCE(t->rcu_read_lock_nesting);
210 
211 		WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
212 	}
213 #endif /* #ifdef CONFIG_PROVE_LOCKING */
214 }
215 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
216 
217 #endif /* #ifdef CONFIG_PREEMPT_RCU */
218 
219 #ifdef CONFIG_DEBUG_LOCK_ALLOC
220 static struct lock_class_key rcu_lock_key;
221 struct lockdep_map rcu_lock_map =
222 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
223 EXPORT_SYMBOL_GPL(rcu_lock_map);
224 
225 static struct lock_class_key rcu_bh_lock_key;
226 struct lockdep_map rcu_bh_lock_map =
227 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
228 EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
229 
230 static struct lock_class_key rcu_sched_lock_key;
231 struct lockdep_map rcu_sched_lock_map =
232 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
233 EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
234 
235 static struct lock_class_key rcu_callback_key;
236 struct lockdep_map rcu_callback_map =
237 	STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
238 EXPORT_SYMBOL_GPL(rcu_callback_map);
239 
240 int notrace debug_lockdep_rcu_enabled(void)
241 {
242 	return rcu_scheduler_active && debug_locks &&
243 	       current->lockdep_recursion == 0;
244 }
245 EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
246 
247 /**
248  * rcu_read_lock_held() - might we be in RCU read-side critical section?
249  *
250  * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
251  * read-side critical section.  In absence of CONFIG_DEBUG_LOCK_ALLOC,
252  * this assumes we are in an RCU read-side critical section unless it can
253  * prove otherwise.  This is useful for debug checks in functions that
254  * require that they be called within an RCU read-side critical section.
255  *
256  * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
257  * and while lockdep is disabled.
258  *
259  * Note that rcu_read_lock() and the matching rcu_read_unlock() must
260  * occur in the same context, for example, it is illegal to invoke
261  * rcu_read_unlock() in process context if the matching rcu_read_lock()
262  * was invoked from within an irq handler.
263  *
264  * Note that rcu_read_lock() is disallowed if the CPU is either idle or
265  * offline from an RCU perspective, so check for those as well.
266  */
267 int rcu_read_lock_held(void)
268 {
269 	if (!debug_lockdep_rcu_enabled())
270 		return 1;
271 	if (!rcu_is_watching())
272 		return 0;
273 	if (!rcu_lockdep_current_cpu_online())
274 		return 0;
275 	return lock_is_held(&rcu_lock_map);
276 }
277 EXPORT_SYMBOL_GPL(rcu_read_lock_held);
278 
279 /**
280  * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
281  *
282  * Check for bottom half being disabled, which covers both the
283  * CONFIG_PROVE_RCU and not cases.  Note that if someone uses
284  * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
285  * will show the situation.  This is useful for debug checks in functions
286  * that require that they be called within an RCU read-side critical
287  * section.
288  *
289  * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
290  *
291  * Note that rcu_read_lock() is disallowed if the CPU is either idle or
292  * offline from an RCU perspective, so check for those as well.
293  */
294 int rcu_read_lock_bh_held(void)
295 {
296 	if (!debug_lockdep_rcu_enabled())
297 		return 1;
298 	if (!rcu_is_watching())
299 		return 0;
300 	if (!rcu_lockdep_current_cpu_online())
301 		return 0;
302 	return in_softirq() || irqs_disabled();
303 }
304 EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
305 
306 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
307 
308 /**
309  * wakeme_after_rcu() - Callback function to awaken a task after grace period
310  * @head: Pointer to rcu_head member within rcu_synchronize structure
311  *
312  * Awaken the corresponding task now that a grace period has elapsed.
313  */
314 void wakeme_after_rcu(struct rcu_head *head)
315 {
316 	struct rcu_synchronize *rcu;
317 
318 	rcu = container_of(head, struct rcu_synchronize, head);
319 	complete(&rcu->completion);
320 }
321 EXPORT_SYMBOL_GPL(wakeme_after_rcu);
322 
323 void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
324 		   struct rcu_synchronize *rs_array)
325 {
326 	int i;
327 
328 	/* Initialize and register callbacks for each flavor specified. */
329 	for (i = 0; i < n; i++) {
330 		if (checktiny &&
331 		    (crcu_array[i] == call_rcu ||
332 		     crcu_array[i] == call_rcu_bh)) {
333 			might_sleep();
334 			continue;
335 		}
336 		init_rcu_head_on_stack(&rs_array[i].head);
337 		init_completion(&rs_array[i].completion);
338 		(crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
339 	}
340 
341 	/* Wait for all callbacks to be invoked. */
342 	for (i = 0; i < n; i++) {
343 		if (checktiny &&
344 		    (crcu_array[i] == call_rcu ||
345 		     crcu_array[i] == call_rcu_bh))
346 			continue;
347 		wait_for_completion(&rs_array[i].completion);
348 		destroy_rcu_head_on_stack(&rs_array[i].head);
349 	}
350 }
351 EXPORT_SYMBOL_GPL(__wait_rcu_gp);
352 
353 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
354 void init_rcu_head(struct rcu_head *head)
355 {
356 	debug_object_init(head, &rcuhead_debug_descr);
357 }
358 
359 void destroy_rcu_head(struct rcu_head *head)
360 {
361 	debug_object_free(head, &rcuhead_debug_descr);
362 }
363 
364 /*
365  * fixup_activate is called when:
366  * - an active object is activated
367  * - an unknown object is activated (might be a statically initialized object)
368  * Activation is performed internally by call_rcu().
369  */
370 static int rcuhead_fixup_activate(void *addr, enum debug_obj_state state)
371 {
372 	struct rcu_head *head = addr;
373 
374 	switch (state) {
375 
376 	case ODEBUG_STATE_NOTAVAILABLE:
377 		/*
378 		 * This is not really a fixup. We just make sure that it is
379 		 * tracked in the object tracker.
380 		 */
381 		debug_object_init(head, &rcuhead_debug_descr);
382 		debug_object_activate(head, &rcuhead_debug_descr);
383 		return 0;
384 	default:
385 		return 1;
386 	}
387 }
388 
389 /**
390  * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
391  * @head: pointer to rcu_head structure to be initialized
392  *
393  * This function informs debugobjects of a new rcu_head structure that
394  * has been allocated as an auto variable on the stack.  This function
395  * is not required for rcu_head structures that are statically defined or
396  * that are dynamically allocated on the heap.  This function has no
397  * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
398  */
399 void init_rcu_head_on_stack(struct rcu_head *head)
400 {
401 	debug_object_init_on_stack(head, &rcuhead_debug_descr);
402 }
403 EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
404 
405 /**
406  * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
407  * @head: pointer to rcu_head structure to be initialized
408  *
409  * This function informs debugobjects that an on-stack rcu_head structure
410  * is about to go out of scope.  As with init_rcu_head_on_stack(), this
411  * function is not required for rcu_head structures that are statically
412  * defined or that are dynamically allocated on the heap.  Also as with
413  * init_rcu_head_on_stack(), this function has no effect for
414  * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
415  */
416 void destroy_rcu_head_on_stack(struct rcu_head *head)
417 {
418 	debug_object_free(head, &rcuhead_debug_descr);
419 }
420 EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
421 
422 struct debug_obj_descr rcuhead_debug_descr = {
423 	.name = "rcu_head",
424 	.fixup_activate = rcuhead_fixup_activate,
425 };
426 EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
427 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
428 
429 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
430 void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
431 			       unsigned long secs,
432 			       unsigned long c_old, unsigned long c)
433 {
434 	trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
435 }
436 EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
437 #else
438 #define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
439 	do { } while (0)
440 #endif
441 
442 #ifdef CONFIG_RCU_STALL_COMMON
443 
444 #ifdef CONFIG_PROVE_RCU
445 #define RCU_STALL_DELAY_DELTA	       (5 * HZ)
446 #else
447 #define RCU_STALL_DELAY_DELTA	       0
448 #endif
449 
450 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
451 static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
452 
453 module_param(rcu_cpu_stall_suppress, int, 0644);
454 module_param(rcu_cpu_stall_timeout, int, 0644);
455 
456 int rcu_jiffies_till_stall_check(void)
457 {
458 	int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
459 
460 	/*
461 	 * Limit check must be consistent with the Kconfig limits
462 	 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
463 	 */
464 	if (till_stall_check < 3) {
465 		WRITE_ONCE(rcu_cpu_stall_timeout, 3);
466 		till_stall_check = 3;
467 	} else if (till_stall_check > 300) {
468 		WRITE_ONCE(rcu_cpu_stall_timeout, 300);
469 		till_stall_check = 300;
470 	}
471 	return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
472 }
473 
474 void rcu_sysrq_start(void)
475 {
476 	if (!rcu_cpu_stall_suppress)
477 		rcu_cpu_stall_suppress = 2;
478 }
479 
480 void rcu_sysrq_end(void)
481 {
482 	if (rcu_cpu_stall_suppress == 2)
483 		rcu_cpu_stall_suppress = 0;
484 }
485 
486 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
487 {
488 	rcu_cpu_stall_suppress = 1;
489 	return NOTIFY_DONE;
490 }
491 
492 static struct notifier_block rcu_panic_block = {
493 	.notifier_call = rcu_panic,
494 };
495 
496 static int __init check_cpu_stall_init(void)
497 {
498 	atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
499 	return 0;
500 }
501 early_initcall(check_cpu_stall_init);
502 
503 #endif /* #ifdef CONFIG_RCU_STALL_COMMON */
504 
505 #ifdef CONFIG_TASKS_RCU
506 
507 /*
508  * Simple variant of RCU whose quiescent states are voluntary context switch,
509  * user-space execution, and idle.  As such, grace periods can take one good
510  * long time.  There are no read-side primitives similar to rcu_read_lock()
511  * and rcu_read_unlock() because this implementation is intended to get
512  * the system into a safe state for some of the manipulations involved in
513  * tracing and the like.  Finally, this implementation does not support
514  * high call_rcu_tasks() rates from multiple CPUs.  If this is required,
515  * per-CPU callback lists will be needed.
516  */
517 
518 /* Global list of callbacks and associated lock. */
519 static struct rcu_head *rcu_tasks_cbs_head;
520 static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
521 static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
522 static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
523 
524 /* Track exiting tasks in order to allow them to be waited for. */
525 DEFINE_SRCU(tasks_rcu_exit_srcu);
526 
527 /* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
528 static int rcu_task_stall_timeout __read_mostly = HZ * 60 * 10;
529 module_param(rcu_task_stall_timeout, int, 0644);
530 
531 static void rcu_spawn_tasks_kthread(void);
532 
533 /*
534  * Post an RCU-tasks callback.  First call must be from process context
535  * after the scheduler if fully operational.
536  */
537 void call_rcu_tasks(struct rcu_head *rhp, void (*func)(struct rcu_head *rhp))
538 {
539 	unsigned long flags;
540 	bool needwake;
541 
542 	rhp->next = NULL;
543 	rhp->func = func;
544 	raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
545 	needwake = !rcu_tasks_cbs_head;
546 	*rcu_tasks_cbs_tail = rhp;
547 	rcu_tasks_cbs_tail = &rhp->next;
548 	raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
549 	if (needwake) {
550 		rcu_spawn_tasks_kthread();
551 		wake_up(&rcu_tasks_cbs_wq);
552 	}
553 }
554 EXPORT_SYMBOL_GPL(call_rcu_tasks);
555 
556 /**
557  * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
558  *
559  * Control will return to the caller some time after a full rcu-tasks
560  * grace period has elapsed, in other words after all currently
561  * executing rcu-tasks read-side critical sections have elapsed.  These
562  * read-side critical sections are delimited by calls to schedule(),
563  * cond_resched_rcu_qs(), idle execution, userspace execution, calls
564  * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
565  *
566  * This is a very specialized primitive, intended only for a few uses in
567  * tracing and other situations requiring manipulation of function
568  * preambles and profiling hooks.  The synchronize_rcu_tasks() function
569  * is not (yet) intended for heavy use from multiple CPUs.
570  *
571  * Note that this guarantee implies further memory-ordering guarantees.
572  * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
573  * each CPU is guaranteed to have executed a full memory barrier since the
574  * end of its last RCU-tasks read-side critical section whose beginning
575  * preceded the call to synchronize_rcu_tasks().  In addition, each CPU
576  * having an RCU-tasks read-side critical section that extends beyond
577  * the return from synchronize_rcu_tasks() is guaranteed to have executed
578  * a full memory barrier after the beginning of synchronize_rcu_tasks()
579  * and before the beginning of that RCU-tasks read-side critical section.
580  * Note that these guarantees include CPUs that are offline, idle, or
581  * executing in user mode, as well as CPUs that are executing in the kernel.
582  *
583  * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
584  * to its caller on CPU B, then both CPU A and CPU B are guaranteed
585  * to have executed a full memory barrier during the execution of
586  * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
587  * (but again only if the system has more than one CPU).
588  */
589 void synchronize_rcu_tasks(void)
590 {
591 	/* Complain if the scheduler has not started.  */
592 	RCU_LOCKDEP_WARN(!rcu_scheduler_active,
593 			 "synchronize_rcu_tasks called too soon");
594 
595 	/* Wait for the grace period. */
596 	wait_rcu_gp(call_rcu_tasks);
597 }
598 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
599 
600 /**
601  * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
602  *
603  * Although the current implementation is guaranteed to wait, it is not
604  * obligated to, for example, if there are no pending callbacks.
605  */
606 void rcu_barrier_tasks(void)
607 {
608 	/* There is only one callback queue, so this is easy.  ;-) */
609 	synchronize_rcu_tasks();
610 }
611 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
612 
613 /* See if tasks are still holding out, complain if so. */
614 static void check_holdout_task(struct task_struct *t,
615 			       bool needreport, bool *firstreport)
616 {
617 	int cpu;
618 
619 	if (!READ_ONCE(t->rcu_tasks_holdout) ||
620 	    t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
621 	    !READ_ONCE(t->on_rq) ||
622 	    (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
623 	     !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
624 		WRITE_ONCE(t->rcu_tasks_holdout, false);
625 		list_del_init(&t->rcu_tasks_holdout_list);
626 		put_task_struct(t);
627 		return;
628 	}
629 	if (!needreport)
630 		return;
631 	if (*firstreport) {
632 		pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
633 		*firstreport = false;
634 	}
635 	cpu = task_cpu(t);
636 	pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
637 		 t, ".I"[is_idle_task(t)],
638 		 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
639 		 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
640 		 t->rcu_tasks_idle_cpu, cpu);
641 	sched_show_task(t);
642 }
643 
644 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
645 static int __noreturn rcu_tasks_kthread(void *arg)
646 {
647 	unsigned long flags;
648 	struct task_struct *g, *t;
649 	unsigned long lastreport;
650 	struct rcu_head *list;
651 	struct rcu_head *next;
652 	LIST_HEAD(rcu_tasks_holdouts);
653 
654 	/* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
655 	housekeeping_affine(current);
656 
657 	/*
658 	 * Each pass through the following loop makes one check for
659 	 * newly arrived callbacks, and, if there are some, waits for
660 	 * one RCU-tasks grace period and then invokes the callbacks.
661 	 * This loop is terminated by the system going down.  ;-)
662 	 */
663 	for (;;) {
664 
665 		/* Pick up any new callbacks. */
666 		raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
667 		list = rcu_tasks_cbs_head;
668 		rcu_tasks_cbs_head = NULL;
669 		rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
670 		raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
671 
672 		/* If there were none, wait a bit and start over. */
673 		if (!list) {
674 			wait_event_interruptible(rcu_tasks_cbs_wq,
675 						 rcu_tasks_cbs_head);
676 			if (!rcu_tasks_cbs_head) {
677 				WARN_ON(signal_pending(current));
678 				schedule_timeout_interruptible(HZ/10);
679 			}
680 			continue;
681 		}
682 
683 		/*
684 		 * Wait for all pre-existing t->on_rq and t->nvcsw
685 		 * transitions to complete.  Invoking synchronize_sched()
686 		 * suffices because all these transitions occur with
687 		 * interrupts disabled.  Without this synchronize_sched(),
688 		 * a read-side critical section that started before the
689 		 * grace period might be incorrectly seen as having started
690 		 * after the grace period.
691 		 *
692 		 * This synchronize_sched() also dispenses with the
693 		 * need for a memory barrier on the first store to
694 		 * ->rcu_tasks_holdout, as it forces the store to happen
695 		 * after the beginning of the grace period.
696 		 */
697 		synchronize_sched();
698 
699 		/*
700 		 * There were callbacks, so we need to wait for an
701 		 * RCU-tasks grace period.  Start off by scanning
702 		 * the task list for tasks that are not already
703 		 * voluntarily blocked.  Mark these tasks and make
704 		 * a list of them in rcu_tasks_holdouts.
705 		 */
706 		rcu_read_lock();
707 		for_each_process_thread(g, t) {
708 			if (t != current && READ_ONCE(t->on_rq) &&
709 			    !is_idle_task(t)) {
710 				get_task_struct(t);
711 				t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
712 				WRITE_ONCE(t->rcu_tasks_holdout, true);
713 				list_add(&t->rcu_tasks_holdout_list,
714 					 &rcu_tasks_holdouts);
715 			}
716 		}
717 		rcu_read_unlock();
718 
719 		/*
720 		 * Wait for tasks that are in the process of exiting.
721 		 * This does only part of the job, ensuring that all
722 		 * tasks that were previously exiting reach the point
723 		 * where they have disabled preemption, allowing the
724 		 * later synchronize_sched() to finish the job.
725 		 */
726 		synchronize_srcu(&tasks_rcu_exit_srcu);
727 
728 		/*
729 		 * Each pass through the following loop scans the list
730 		 * of holdout tasks, removing any that are no longer
731 		 * holdouts.  When the list is empty, we are done.
732 		 */
733 		lastreport = jiffies;
734 		while (!list_empty(&rcu_tasks_holdouts)) {
735 			bool firstreport;
736 			bool needreport;
737 			int rtst;
738 			struct task_struct *t1;
739 
740 			schedule_timeout_interruptible(HZ);
741 			rtst = READ_ONCE(rcu_task_stall_timeout);
742 			needreport = rtst > 0 &&
743 				     time_after(jiffies, lastreport + rtst);
744 			if (needreport)
745 				lastreport = jiffies;
746 			firstreport = true;
747 			WARN_ON(signal_pending(current));
748 			list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
749 						rcu_tasks_holdout_list) {
750 				check_holdout_task(t, needreport, &firstreport);
751 				cond_resched();
752 			}
753 		}
754 
755 		/*
756 		 * Because ->on_rq and ->nvcsw are not guaranteed
757 		 * to have a full memory barriers prior to them in the
758 		 * schedule() path, memory reordering on other CPUs could
759 		 * cause their RCU-tasks read-side critical sections to
760 		 * extend past the end of the grace period.  However,
761 		 * because these ->nvcsw updates are carried out with
762 		 * interrupts disabled, we can use synchronize_sched()
763 		 * to force the needed ordering on all such CPUs.
764 		 *
765 		 * This synchronize_sched() also confines all
766 		 * ->rcu_tasks_holdout accesses to be within the grace
767 		 * period, avoiding the need for memory barriers for
768 		 * ->rcu_tasks_holdout accesses.
769 		 *
770 		 * In addition, this synchronize_sched() waits for exiting
771 		 * tasks to complete their final preempt_disable() region
772 		 * of execution, cleaning up after the synchronize_srcu()
773 		 * above.
774 		 */
775 		synchronize_sched();
776 
777 		/* Invoke the callbacks. */
778 		while (list) {
779 			next = list->next;
780 			local_bh_disable();
781 			list->func(list);
782 			local_bh_enable();
783 			list = next;
784 			cond_resched();
785 		}
786 		schedule_timeout_uninterruptible(HZ/10);
787 	}
788 }
789 
790 /* Spawn rcu_tasks_kthread() at first call to call_rcu_tasks(). */
791 static void rcu_spawn_tasks_kthread(void)
792 {
793 	static DEFINE_MUTEX(rcu_tasks_kthread_mutex);
794 	static struct task_struct *rcu_tasks_kthread_ptr;
795 	struct task_struct *t;
796 
797 	if (READ_ONCE(rcu_tasks_kthread_ptr)) {
798 		smp_mb(); /* Ensure caller sees full kthread. */
799 		return;
800 	}
801 	mutex_lock(&rcu_tasks_kthread_mutex);
802 	if (rcu_tasks_kthread_ptr) {
803 		mutex_unlock(&rcu_tasks_kthread_mutex);
804 		return;
805 	}
806 	t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
807 	BUG_ON(IS_ERR(t));
808 	smp_mb(); /* Ensure others see full kthread. */
809 	WRITE_ONCE(rcu_tasks_kthread_ptr, t);
810 	mutex_unlock(&rcu_tasks_kthread_mutex);
811 }
812 
813 #endif /* #ifdef CONFIG_TASKS_RCU */
814 
815 #ifdef CONFIG_PROVE_RCU
816 
817 /*
818  * Early boot self test parameters, one for each flavor
819  */
820 static bool rcu_self_test;
821 static bool rcu_self_test_bh;
822 static bool rcu_self_test_sched;
823 
824 module_param(rcu_self_test, bool, 0444);
825 module_param(rcu_self_test_bh, bool, 0444);
826 module_param(rcu_self_test_sched, bool, 0444);
827 
828 static int rcu_self_test_counter;
829 
830 static void test_callback(struct rcu_head *r)
831 {
832 	rcu_self_test_counter++;
833 	pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
834 }
835 
836 static void early_boot_test_call_rcu(void)
837 {
838 	static struct rcu_head head;
839 
840 	call_rcu(&head, test_callback);
841 }
842 
843 static void early_boot_test_call_rcu_bh(void)
844 {
845 	static struct rcu_head head;
846 
847 	call_rcu_bh(&head, test_callback);
848 }
849 
850 static void early_boot_test_call_rcu_sched(void)
851 {
852 	static struct rcu_head head;
853 
854 	call_rcu_sched(&head, test_callback);
855 }
856 
857 void rcu_early_boot_tests(void)
858 {
859 	pr_info("Running RCU self tests\n");
860 
861 	if (rcu_self_test)
862 		early_boot_test_call_rcu();
863 	if (rcu_self_test_bh)
864 		early_boot_test_call_rcu_bh();
865 	if (rcu_self_test_sched)
866 		early_boot_test_call_rcu_sched();
867 }
868 
869 static int rcu_verify_early_boot_tests(void)
870 {
871 	int ret = 0;
872 	int early_boot_test_counter = 0;
873 
874 	if (rcu_self_test) {
875 		early_boot_test_counter++;
876 		rcu_barrier();
877 	}
878 	if (rcu_self_test_bh) {
879 		early_boot_test_counter++;
880 		rcu_barrier_bh();
881 	}
882 	if (rcu_self_test_sched) {
883 		early_boot_test_counter++;
884 		rcu_barrier_sched();
885 	}
886 
887 	if (rcu_self_test_counter != early_boot_test_counter) {
888 		WARN_ON(1);
889 		ret = -1;
890 	}
891 
892 	return ret;
893 }
894 late_initcall(rcu_verify_early_boot_tests);
895 #else
896 void rcu_early_boot_tests(void) {}
897 #endif /* CONFIG_PROVE_RCU */
898