xref: /linux/kernel/rcu/update.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
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/signal.h>
40 #include <linux/sched/debug.h>
41 #include <linux/atomic.h>
42 #include <linux/bitops.h>
43 #include <linux/percpu.h>
44 #include <linux/notifier.h>
45 #include <linux/cpu.h>
46 #include <linux/mutex.h>
47 #include <linux/export.h>
48 #include <linux/hardirq.h>
49 #include <linux/delay.h>
50 #include <linux/moduleparam.h>
51 #include <linux/kthread.h>
52 #include <linux/tick.h>
53 #include <linux/rcupdate_wait.h>
54 #include <linux/sched/isolation.h>
55 
56 #define CREATE_TRACE_POINTS
57 
58 #include "rcu.h"
59 
60 #ifdef MODULE_PARAM_PREFIX
61 #undef MODULE_PARAM_PREFIX
62 #endif
63 #define MODULE_PARAM_PREFIX "rcupdate."
64 
65 #ifndef CONFIG_TINY_RCU
66 extern int rcu_expedited; /* from sysctl */
67 module_param(rcu_expedited, int, 0);
68 extern int rcu_normal; /* from sysctl */
69 module_param(rcu_normal, int, 0);
70 static int rcu_normal_after_boot;
71 module_param(rcu_normal_after_boot, int, 0);
72 #endif /* #ifndef CONFIG_TINY_RCU */
73 
74 #ifdef CONFIG_DEBUG_LOCK_ALLOC
75 /**
76  * rcu_read_lock_sched_held() - might we be in RCU-sched read-side critical section?
77  *
78  * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an
79  * RCU-sched read-side critical section.  In absence of
80  * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
81  * critical section unless it can prove otherwise.  Note that disabling
82  * of preemption (including disabling irqs) counts as an RCU-sched
83  * read-side critical section.  This is useful for debug checks in functions
84  * that required that they be called within an RCU-sched read-side
85  * critical section.
86  *
87  * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
88  * and while lockdep is disabled.
89  *
90  * Note that if the CPU is in the idle loop from an RCU point of
91  * view (ie: that we are in the section between rcu_idle_enter() and
92  * rcu_idle_exit()) then rcu_read_lock_held() returns false even if the CPU
93  * did an rcu_read_lock().  The reason for this is that RCU ignores CPUs
94  * that are in such a section, considering these as in extended quiescent
95  * state, so such a CPU is effectively never in an RCU read-side critical
96  * section regardless of what RCU primitives it invokes.  This state of
97  * affairs is required --- we need to keep an RCU-free window in idle
98  * where the CPU may possibly enter into low power mode. This way we can
99  * notice an extended quiescent state to other CPUs that started a grace
100  * period. Otherwise we would delay any grace period as long as we run in
101  * the idle task.
102  *
103  * Similarly, we avoid claiming an SRCU read lock held if the current
104  * CPU is offline.
105  */
106 int rcu_read_lock_sched_held(void)
107 {
108 	int lockdep_opinion = 0;
109 
110 	if (!debug_lockdep_rcu_enabled())
111 		return 1;
112 	if (!rcu_is_watching())
113 		return 0;
114 	if (!rcu_lockdep_current_cpu_online())
115 		return 0;
116 	if (debug_locks)
117 		lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
118 	return lockdep_opinion || !preemptible();
119 }
120 EXPORT_SYMBOL(rcu_read_lock_sched_held);
121 #endif
122 
123 #ifndef CONFIG_TINY_RCU
124 
125 /*
126  * Should expedited grace-period primitives always fall back to their
127  * non-expedited counterparts?  Intended for use within RCU.  Note
128  * that if the user specifies both rcu_expedited and rcu_normal, then
129  * rcu_normal wins.  (Except during the time period during boot from
130  * when the first task is spawned until the rcu_set_runtime_mode()
131  * core_initcall() is invoked, at which point everything is expedited.)
132  */
133 bool rcu_gp_is_normal(void)
134 {
135 	return READ_ONCE(rcu_normal) &&
136 	       rcu_scheduler_active != RCU_SCHEDULER_INIT;
137 }
138 EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
139 
140 static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1);
141 
142 /*
143  * Should normal grace-period primitives be expedited?  Intended for
144  * use within RCU.  Note that this function takes the rcu_expedited
145  * sysfs/boot variable and rcu_scheduler_active into account as well
146  * as the rcu_expedite_gp() nesting.  So looping on rcu_unexpedite_gp()
147  * until rcu_gp_is_expedited() returns false is a -really- bad idea.
148  */
149 bool rcu_gp_is_expedited(void)
150 {
151 	return rcu_expedited || atomic_read(&rcu_expedited_nesting) ||
152 	       rcu_scheduler_active == RCU_SCHEDULER_INIT;
153 }
154 EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
155 
156 /**
157  * rcu_expedite_gp - Expedite future RCU grace periods
158  *
159  * After a call to this function, future calls to synchronize_rcu() and
160  * friends act as the corresponding synchronize_rcu_expedited() function
161  * had instead been called.
162  */
163 void rcu_expedite_gp(void)
164 {
165 	atomic_inc(&rcu_expedited_nesting);
166 }
167 EXPORT_SYMBOL_GPL(rcu_expedite_gp);
168 
169 /**
170  * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
171  *
172  * Undo a prior call to rcu_expedite_gp().  If all prior calls to
173  * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
174  * and if the rcu_expedited sysfs/boot parameter is not set, then all
175  * subsequent calls to synchronize_rcu() and friends will return to
176  * their normal non-expedited behavior.
177  */
178 void rcu_unexpedite_gp(void)
179 {
180 	atomic_dec(&rcu_expedited_nesting);
181 }
182 EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
183 
184 /*
185  * Inform RCU of the end of the in-kernel boot sequence.
186  */
187 void rcu_end_inkernel_boot(void)
188 {
189 	rcu_unexpedite_gp();
190 	if (rcu_normal_after_boot)
191 		WRITE_ONCE(rcu_normal, 1);
192 }
193 
194 #endif /* #ifndef CONFIG_TINY_RCU */
195 
196 /*
197  * Test each non-SRCU synchronous grace-period wait API.  This is
198  * useful just after a change in mode for these primitives, and
199  * during early boot.
200  */
201 void rcu_test_sync_prims(void)
202 {
203 	if (!IS_ENABLED(CONFIG_PROVE_RCU))
204 		return;
205 	synchronize_rcu();
206 	synchronize_rcu_bh();
207 	synchronize_sched();
208 	synchronize_rcu_expedited();
209 	synchronize_rcu_bh_expedited();
210 	synchronize_sched_expedited();
211 }
212 
213 #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU)
214 
215 /*
216  * Switch to run-time mode once RCU has fully initialized.
217  */
218 static int __init rcu_set_runtime_mode(void)
219 {
220 	rcu_test_sync_prims();
221 	rcu_scheduler_active = RCU_SCHEDULER_RUNNING;
222 	rcu_test_sync_prims();
223 	return 0;
224 }
225 core_initcall(rcu_set_runtime_mode);
226 
227 #endif /* #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) */
228 
229 #ifdef CONFIG_PREEMPT_RCU
230 
231 /*
232  * Preemptible RCU implementation for rcu_read_lock().
233  * Just increment ->rcu_read_lock_nesting, shared state will be updated
234  * if we block.
235  */
236 void __rcu_read_lock(void)
237 {
238 	current->rcu_read_lock_nesting++;
239 	barrier();  /* critical section after entry code. */
240 }
241 EXPORT_SYMBOL_GPL(__rcu_read_lock);
242 
243 /*
244  * Preemptible RCU implementation for rcu_read_unlock().
245  * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
246  * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
247  * invoke rcu_read_unlock_special() to clean up after a context switch
248  * in an RCU read-side critical section and other special cases.
249  */
250 void __rcu_read_unlock(void)
251 {
252 	struct task_struct *t = current;
253 
254 	if (t->rcu_read_lock_nesting != 1) {
255 		--t->rcu_read_lock_nesting;
256 	} else {
257 		barrier();  /* critical section before exit code. */
258 		t->rcu_read_lock_nesting = INT_MIN;
259 		barrier();  /* assign before ->rcu_read_unlock_special load */
260 		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
261 			rcu_read_unlock_special(t);
262 		barrier();  /* ->rcu_read_unlock_special load before assign */
263 		t->rcu_read_lock_nesting = 0;
264 	}
265 #ifdef CONFIG_PROVE_LOCKING
266 	{
267 		int rrln = READ_ONCE(t->rcu_read_lock_nesting);
268 
269 		WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
270 	}
271 #endif /* #ifdef CONFIG_PROVE_LOCKING */
272 }
273 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
274 
275 #endif /* #ifdef CONFIG_PREEMPT_RCU */
276 
277 #ifdef CONFIG_DEBUG_LOCK_ALLOC
278 static struct lock_class_key rcu_lock_key;
279 struct lockdep_map rcu_lock_map =
280 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
281 EXPORT_SYMBOL_GPL(rcu_lock_map);
282 
283 static struct lock_class_key rcu_bh_lock_key;
284 struct lockdep_map rcu_bh_lock_map =
285 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
286 EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
287 
288 static struct lock_class_key rcu_sched_lock_key;
289 struct lockdep_map rcu_sched_lock_map =
290 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
291 EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
292 
293 static struct lock_class_key rcu_callback_key;
294 struct lockdep_map rcu_callback_map =
295 	STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
296 EXPORT_SYMBOL_GPL(rcu_callback_map);
297 
298 int notrace debug_lockdep_rcu_enabled(void)
299 {
300 	return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks &&
301 	       current->lockdep_recursion == 0;
302 }
303 EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
304 
305 /**
306  * rcu_read_lock_held() - might we be in RCU read-side critical section?
307  *
308  * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
309  * read-side critical section.  In absence of CONFIG_DEBUG_LOCK_ALLOC,
310  * this assumes we are in an RCU read-side critical section unless it can
311  * prove otherwise.  This is useful for debug checks in functions that
312  * require that they be called within an RCU read-side critical section.
313  *
314  * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
315  * and while lockdep is disabled.
316  *
317  * Note that rcu_read_lock() and the matching rcu_read_unlock() must
318  * occur in the same context, for example, it is illegal to invoke
319  * rcu_read_unlock() in process context if the matching rcu_read_lock()
320  * was invoked from within an irq handler.
321  *
322  * Note that rcu_read_lock() is disallowed if the CPU is either idle or
323  * offline from an RCU perspective, so check for those as well.
324  */
325 int rcu_read_lock_held(void)
326 {
327 	if (!debug_lockdep_rcu_enabled())
328 		return 1;
329 	if (!rcu_is_watching())
330 		return 0;
331 	if (!rcu_lockdep_current_cpu_online())
332 		return 0;
333 	return lock_is_held(&rcu_lock_map);
334 }
335 EXPORT_SYMBOL_GPL(rcu_read_lock_held);
336 
337 /**
338  * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
339  *
340  * Check for bottom half being disabled, which covers both the
341  * CONFIG_PROVE_RCU and not cases.  Note that if someone uses
342  * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
343  * will show the situation.  This is useful for debug checks in functions
344  * that require that they be called within an RCU read-side critical
345  * section.
346  *
347  * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
348  *
349  * Note that rcu_read_lock() is disallowed if the CPU is either idle or
350  * offline from an RCU perspective, so check for those as well.
351  */
352 int rcu_read_lock_bh_held(void)
353 {
354 	if (!debug_lockdep_rcu_enabled())
355 		return 1;
356 	if (!rcu_is_watching())
357 		return 0;
358 	if (!rcu_lockdep_current_cpu_online())
359 		return 0;
360 	return in_softirq() || irqs_disabled();
361 }
362 EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
363 
364 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
365 
366 /**
367  * wakeme_after_rcu() - Callback function to awaken a task after grace period
368  * @head: Pointer to rcu_head member within rcu_synchronize structure
369  *
370  * Awaken the corresponding task now that a grace period has elapsed.
371  */
372 void wakeme_after_rcu(struct rcu_head *head)
373 {
374 	struct rcu_synchronize *rcu;
375 
376 	rcu = container_of(head, struct rcu_synchronize, head);
377 	complete(&rcu->completion);
378 }
379 EXPORT_SYMBOL_GPL(wakeme_after_rcu);
380 
381 void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
382 		   struct rcu_synchronize *rs_array)
383 {
384 	int i;
385 	int j;
386 
387 	/* Initialize and register callbacks for each flavor specified. */
388 	for (i = 0; i < n; i++) {
389 		if (checktiny &&
390 		    (crcu_array[i] == call_rcu ||
391 		     crcu_array[i] == call_rcu_bh)) {
392 			might_sleep();
393 			continue;
394 		}
395 		init_rcu_head_on_stack(&rs_array[i].head);
396 		init_completion(&rs_array[i].completion);
397 		for (j = 0; j < i; j++)
398 			if (crcu_array[j] == crcu_array[i])
399 				break;
400 		if (j == i)
401 			(crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
402 	}
403 
404 	/* Wait for all callbacks to be invoked. */
405 	for (i = 0; i < n; i++) {
406 		if (checktiny &&
407 		    (crcu_array[i] == call_rcu ||
408 		     crcu_array[i] == call_rcu_bh))
409 			continue;
410 		for (j = 0; j < i; j++)
411 			if (crcu_array[j] == crcu_array[i])
412 				break;
413 		if (j == i)
414 			wait_for_completion(&rs_array[i].completion);
415 		destroy_rcu_head_on_stack(&rs_array[i].head);
416 	}
417 }
418 EXPORT_SYMBOL_GPL(__wait_rcu_gp);
419 
420 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
421 void init_rcu_head(struct rcu_head *head)
422 {
423 	debug_object_init(head, &rcuhead_debug_descr);
424 }
425 
426 void destroy_rcu_head(struct rcu_head *head)
427 {
428 	debug_object_free(head, &rcuhead_debug_descr);
429 }
430 
431 static bool rcuhead_is_static_object(void *addr)
432 {
433 	return true;
434 }
435 
436 /**
437  * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
438  * @head: pointer to rcu_head structure to be initialized
439  *
440  * This function informs debugobjects of a new rcu_head structure that
441  * has been allocated as an auto variable on the stack.  This function
442  * is not required for rcu_head structures that are statically defined or
443  * that are dynamically allocated on the heap.  This function has no
444  * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
445  */
446 void init_rcu_head_on_stack(struct rcu_head *head)
447 {
448 	debug_object_init_on_stack(head, &rcuhead_debug_descr);
449 }
450 EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
451 
452 /**
453  * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
454  * @head: pointer to rcu_head structure to be initialized
455  *
456  * This function informs debugobjects that an on-stack rcu_head structure
457  * is about to go out of scope.  As with init_rcu_head_on_stack(), this
458  * function is not required for rcu_head structures that are statically
459  * defined or that are dynamically allocated on the heap.  Also as with
460  * init_rcu_head_on_stack(), this function has no effect for
461  * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
462  */
463 void destroy_rcu_head_on_stack(struct rcu_head *head)
464 {
465 	debug_object_free(head, &rcuhead_debug_descr);
466 }
467 EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
468 
469 struct debug_obj_descr rcuhead_debug_descr = {
470 	.name = "rcu_head",
471 	.is_static_object = rcuhead_is_static_object,
472 };
473 EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
474 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
475 
476 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
477 void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
478 			       unsigned long secs,
479 			       unsigned long c_old, unsigned long c)
480 {
481 	trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
482 }
483 EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
484 #else
485 #define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
486 	do { } while (0)
487 #endif
488 
489 #ifdef CONFIG_RCU_STALL_COMMON
490 
491 #ifdef CONFIG_PROVE_RCU
492 #define RCU_STALL_DELAY_DELTA	       (5 * HZ)
493 #else
494 #define RCU_STALL_DELAY_DELTA	       0
495 #endif
496 
497 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
498 EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress);
499 static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
500 
501 module_param(rcu_cpu_stall_suppress, int, 0644);
502 module_param(rcu_cpu_stall_timeout, int, 0644);
503 
504 int rcu_jiffies_till_stall_check(void)
505 {
506 	int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
507 
508 	/*
509 	 * Limit check must be consistent with the Kconfig limits
510 	 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
511 	 */
512 	if (till_stall_check < 3) {
513 		WRITE_ONCE(rcu_cpu_stall_timeout, 3);
514 		till_stall_check = 3;
515 	} else if (till_stall_check > 300) {
516 		WRITE_ONCE(rcu_cpu_stall_timeout, 300);
517 		till_stall_check = 300;
518 	}
519 	return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
520 }
521 
522 void rcu_sysrq_start(void)
523 {
524 	if (!rcu_cpu_stall_suppress)
525 		rcu_cpu_stall_suppress = 2;
526 }
527 
528 void rcu_sysrq_end(void)
529 {
530 	if (rcu_cpu_stall_suppress == 2)
531 		rcu_cpu_stall_suppress = 0;
532 }
533 
534 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
535 {
536 	rcu_cpu_stall_suppress = 1;
537 	return NOTIFY_DONE;
538 }
539 
540 static struct notifier_block rcu_panic_block = {
541 	.notifier_call = rcu_panic,
542 };
543 
544 static int __init check_cpu_stall_init(void)
545 {
546 	atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
547 	return 0;
548 }
549 early_initcall(check_cpu_stall_init);
550 
551 #endif /* #ifdef CONFIG_RCU_STALL_COMMON */
552 
553 #ifdef CONFIG_TASKS_RCU
554 
555 /*
556  * Simple variant of RCU whose quiescent states are voluntary context switch,
557  * user-space execution, and idle.  As such, grace periods can take one good
558  * long time.  There are no read-side primitives similar to rcu_read_lock()
559  * and rcu_read_unlock() because this implementation is intended to get
560  * the system into a safe state for some of the manipulations involved in
561  * tracing and the like.  Finally, this implementation does not support
562  * high call_rcu_tasks() rates from multiple CPUs.  If this is required,
563  * per-CPU callback lists will be needed.
564  */
565 
566 /* Global list of callbacks and associated lock. */
567 static struct rcu_head *rcu_tasks_cbs_head;
568 static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
569 static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
570 static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
571 
572 /* Track exiting tasks in order to allow them to be waited for. */
573 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
574 
575 /* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
576 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
577 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
578 module_param(rcu_task_stall_timeout, int, 0644);
579 
580 static struct task_struct *rcu_tasks_kthread_ptr;
581 
582 /**
583  * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
584  * @rhp: structure to be used for queueing the RCU updates.
585  * @func: actual callback function to be invoked after the grace period
586  *
587  * The callback function will be invoked some time after a full grace
588  * period elapses, in other words after all currently executing RCU
589  * read-side critical sections have completed. call_rcu_tasks() assumes
590  * that the read-side critical sections end at a voluntary context
591  * switch (not a preemption!), entry into idle, or transition to usermode
592  * execution.  As such, there are no read-side primitives analogous to
593  * rcu_read_lock() and rcu_read_unlock() because this primitive is intended
594  * to determine that all tasks have passed through a safe state, not so
595  * much for data-strcuture synchronization.
596  *
597  * See the description of call_rcu() for more detailed information on
598  * memory ordering guarantees.
599  */
600 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
601 {
602 	unsigned long flags;
603 	bool needwake;
604 
605 	rhp->next = NULL;
606 	rhp->func = func;
607 	raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
608 	needwake = !rcu_tasks_cbs_head;
609 	*rcu_tasks_cbs_tail = rhp;
610 	rcu_tasks_cbs_tail = &rhp->next;
611 	raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
612 	/* We can't create the thread unless interrupts are enabled. */
613 	if (needwake && READ_ONCE(rcu_tasks_kthread_ptr))
614 		wake_up(&rcu_tasks_cbs_wq);
615 }
616 EXPORT_SYMBOL_GPL(call_rcu_tasks);
617 
618 /**
619  * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
620  *
621  * Control will return to the caller some time after a full rcu-tasks
622  * grace period has elapsed, in other words after all currently
623  * executing rcu-tasks read-side critical sections have elapsed.  These
624  * read-side critical sections are delimited by calls to schedule(),
625  * cond_resched_rcu_qs(), idle execution, userspace execution, calls
626  * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
627  *
628  * This is a very specialized primitive, intended only for a few uses in
629  * tracing and other situations requiring manipulation of function
630  * preambles and profiling hooks.  The synchronize_rcu_tasks() function
631  * is not (yet) intended for heavy use from multiple CPUs.
632  *
633  * Note that this guarantee implies further memory-ordering guarantees.
634  * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
635  * each CPU is guaranteed to have executed a full memory barrier since the
636  * end of its last RCU-tasks read-side critical section whose beginning
637  * preceded the call to synchronize_rcu_tasks().  In addition, each CPU
638  * having an RCU-tasks read-side critical section that extends beyond
639  * the return from synchronize_rcu_tasks() is guaranteed to have executed
640  * a full memory barrier after the beginning of synchronize_rcu_tasks()
641  * and before the beginning of that RCU-tasks read-side critical section.
642  * Note that these guarantees include CPUs that are offline, idle, or
643  * executing in user mode, as well as CPUs that are executing in the kernel.
644  *
645  * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
646  * to its caller on CPU B, then both CPU A and CPU B are guaranteed
647  * to have executed a full memory barrier during the execution of
648  * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
649  * (but again only if the system has more than one CPU).
650  */
651 void synchronize_rcu_tasks(void)
652 {
653 	/* Complain if the scheduler has not started.  */
654 	RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
655 			 "synchronize_rcu_tasks called too soon");
656 
657 	/* Wait for the grace period. */
658 	wait_rcu_gp(call_rcu_tasks);
659 }
660 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
661 
662 /**
663  * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
664  *
665  * Although the current implementation is guaranteed to wait, it is not
666  * obligated to, for example, if there are no pending callbacks.
667  */
668 void rcu_barrier_tasks(void)
669 {
670 	/* There is only one callback queue, so this is easy.  ;-) */
671 	synchronize_rcu_tasks();
672 }
673 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
674 
675 /* See if tasks are still holding out, complain if so. */
676 static void check_holdout_task(struct task_struct *t,
677 			       bool needreport, bool *firstreport)
678 {
679 	int cpu;
680 
681 	if (!READ_ONCE(t->rcu_tasks_holdout) ||
682 	    t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
683 	    !READ_ONCE(t->on_rq) ||
684 	    (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
685 	     !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
686 		WRITE_ONCE(t->rcu_tasks_holdout, false);
687 		list_del_init(&t->rcu_tasks_holdout_list);
688 		put_task_struct(t);
689 		return;
690 	}
691 	rcu_request_urgent_qs_task(t);
692 	if (!needreport)
693 		return;
694 	if (*firstreport) {
695 		pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
696 		*firstreport = false;
697 	}
698 	cpu = task_cpu(t);
699 	pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
700 		 t, ".I"[is_idle_task(t)],
701 		 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
702 		 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
703 		 t->rcu_tasks_idle_cpu, cpu);
704 	sched_show_task(t);
705 }
706 
707 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
708 static int __noreturn rcu_tasks_kthread(void *arg)
709 {
710 	unsigned long flags;
711 	struct task_struct *g, *t;
712 	unsigned long lastreport;
713 	struct rcu_head *list;
714 	struct rcu_head *next;
715 	LIST_HEAD(rcu_tasks_holdouts);
716 
717 	/* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
718 	housekeeping_affine(current, HK_FLAG_RCU);
719 
720 	/*
721 	 * Each pass through the following loop makes one check for
722 	 * newly arrived callbacks, and, if there are some, waits for
723 	 * one RCU-tasks grace period and then invokes the callbacks.
724 	 * This loop is terminated by the system going down.  ;-)
725 	 */
726 	for (;;) {
727 
728 		/* Pick up any new callbacks. */
729 		raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
730 		list = rcu_tasks_cbs_head;
731 		rcu_tasks_cbs_head = NULL;
732 		rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
733 		raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
734 
735 		/* If there were none, wait a bit and start over. */
736 		if (!list) {
737 			wait_event_interruptible(rcu_tasks_cbs_wq,
738 						 rcu_tasks_cbs_head);
739 			if (!rcu_tasks_cbs_head) {
740 				WARN_ON(signal_pending(current));
741 				schedule_timeout_interruptible(HZ/10);
742 			}
743 			continue;
744 		}
745 
746 		/*
747 		 * Wait for all pre-existing t->on_rq and t->nvcsw
748 		 * transitions to complete.  Invoking synchronize_sched()
749 		 * suffices because all these transitions occur with
750 		 * interrupts disabled.  Without this synchronize_sched(),
751 		 * a read-side critical section that started before the
752 		 * grace period might be incorrectly seen as having started
753 		 * after the grace period.
754 		 *
755 		 * This synchronize_sched() also dispenses with the
756 		 * need for a memory barrier on the first store to
757 		 * ->rcu_tasks_holdout, as it forces the store to happen
758 		 * after the beginning of the grace period.
759 		 */
760 		synchronize_sched();
761 
762 		/*
763 		 * There were callbacks, so we need to wait for an
764 		 * RCU-tasks grace period.  Start off by scanning
765 		 * the task list for tasks that are not already
766 		 * voluntarily blocked.  Mark these tasks and make
767 		 * a list of them in rcu_tasks_holdouts.
768 		 */
769 		rcu_read_lock();
770 		for_each_process_thread(g, t) {
771 			if (t != current && READ_ONCE(t->on_rq) &&
772 			    !is_idle_task(t)) {
773 				get_task_struct(t);
774 				t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
775 				WRITE_ONCE(t->rcu_tasks_holdout, true);
776 				list_add(&t->rcu_tasks_holdout_list,
777 					 &rcu_tasks_holdouts);
778 			}
779 		}
780 		rcu_read_unlock();
781 
782 		/*
783 		 * Wait for tasks that are in the process of exiting.
784 		 * This does only part of the job, ensuring that all
785 		 * tasks that were previously exiting reach the point
786 		 * where they have disabled preemption, allowing the
787 		 * later synchronize_sched() to finish the job.
788 		 */
789 		synchronize_srcu(&tasks_rcu_exit_srcu);
790 
791 		/*
792 		 * Each pass through the following loop scans the list
793 		 * of holdout tasks, removing any that are no longer
794 		 * holdouts.  When the list is empty, we are done.
795 		 */
796 		lastreport = jiffies;
797 		while (!list_empty(&rcu_tasks_holdouts)) {
798 			bool firstreport;
799 			bool needreport;
800 			int rtst;
801 			struct task_struct *t1;
802 
803 			schedule_timeout_interruptible(HZ);
804 			rtst = READ_ONCE(rcu_task_stall_timeout);
805 			needreport = rtst > 0 &&
806 				     time_after(jiffies, lastreport + rtst);
807 			if (needreport)
808 				lastreport = jiffies;
809 			firstreport = true;
810 			WARN_ON(signal_pending(current));
811 			list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
812 						rcu_tasks_holdout_list) {
813 				check_holdout_task(t, needreport, &firstreport);
814 				cond_resched();
815 			}
816 		}
817 
818 		/*
819 		 * Because ->on_rq and ->nvcsw are not guaranteed
820 		 * to have a full memory barriers prior to them in the
821 		 * schedule() path, memory reordering on other CPUs could
822 		 * cause their RCU-tasks read-side critical sections to
823 		 * extend past the end of the grace period.  However,
824 		 * because these ->nvcsw updates are carried out with
825 		 * interrupts disabled, we can use synchronize_sched()
826 		 * to force the needed ordering on all such CPUs.
827 		 *
828 		 * This synchronize_sched() also confines all
829 		 * ->rcu_tasks_holdout accesses to be within the grace
830 		 * period, avoiding the need for memory barriers for
831 		 * ->rcu_tasks_holdout accesses.
832 		 *
833 		 * In addition, this synchronize_sched() waits for exiting
834 		 * tasks to complete their final preempt_disable() region
835 		 * of execution, cleaning up after the synchronize_srcu()
836 		 * above.
837 		 */
838 		synchronize_sched();
839 
840 		/* Invoke the callbacks. */
841 		while (list) {
842 			next = list->next;
843 			local_bh_disable();
844 			list->func(list);
845 			local_bh_enable();
846 			list = next;
847 			cond_resched();
848 		}
849 		schedule_timeout_uninterruptible(HZ/10);
850 	}
851 }
852 
853 /* Spawn rcu_tasks_kthread() at core_initcall() time. */
854 static int __init rcu_spawn_tasks_kthread(void)
855 {
856 	struct task_struct *t;
857 
858 	t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
859 	BUG_ON(IS_ERR(t));
860 	smp_mb(); /* Ensure others see full kthread. */
861 	WRITE_ONCE(rcu_tasks_kthread_ptr, t);
862 	return 0;
863 }
864 core_initcall(rcu_spawn_tasks_kthread);
865 
866 /* Do the srcu_read_lock() for the above synchronize_srcu().  */
867 void exit_tasks_rcu_start(void)
868 {
869 	preempt_disable();
870 	current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
871 	preempt_enable();
872 }
873 
874 /* Do the srcu_read_unlock() for the above synchronize_srcu().  */
875 void exit_tasks_rcu_finish(void)
876 {
877 	preempt_disable();
878 	__srcu_read_unlock(&tasks_rcu_exit_srcu, current->rcu_tasks_idx);
879 	preempt_enable();
880 }
881 
882 #endif /* #ifdef CONFIG_TASKS_RCU */
883 
884 #ifndef CONFIG_TINY_RCU
885 
886 /*
887  * Print any non-default Tasks RCU settings.
888  */
889 static void __init rcu_tasks_bootup_oddness(void)
890 {
891 #ifdef CONFIG_TASKS_RCU
892 	if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
893 		pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
894 	else
895 		pr_info("\tTasks RCU enabled.\n");
896 #endif /* #ifdef CONFIG_TASKS_RCU */
897 }
898 
899 #endif /* #ifndef CONFIG_TINY_RCU */
900 
901 #ifdef CONFIG_PROVE_RCU
902 
903 /*
904  * Early boot self test parameters, one for each flavor
905  */
906 static bool rcu_self_test;
907 static bool rcu_self_test_bh;
908 static bool rcu_self_test_sched;
909 
910 module_param(rcu_self_test, bool, 0444);
911 module_param(rcu_self_test_bh, bool, 0444);
912 module_param(rcu_self_test_sched, bool, 0444);
913 
914 static int rcu_self_test_counter;
915 
916 static void test_callback(struct rcu_head *r)
917 {
918 	rcu_self_test_counter++;
919 	pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
920 }
921 
922 static void early_boot_test_call_rcu(void)
923 {
924 	static struct rcu_head head;
925 
926 	call_rcu(&head, test_callback);
927 }
928 
929 static void early_boot_test_call_rcu_bh(void)
930 {
931 	static struct rcu_head head;
932 
933 	call_rcu_bh(&head, test_callback);
934 }
935 
936 static void early_boot_test_call_rcu_sched(void)
937 {
938 	static struct rcu_head head;
939 
940 	call_rcu_sched(&head, test_callback);
941 }
942 
943 void rcu_early_boot_tests(void)
944 {
945 	pr_info("Running RCU self tests\n");
946 
947 	if (rcu_self_test)
948 		early_boot_test_call_rcu();
949 	if (rcu_self_test_bh)
950 		early_boot_test_call_rcu_bh();
951 	if (rcu_self_test_sched)
952 		early_boot_test_call_rcu_sched();
953 	rcu_test_sync_prims();
954 }
955 
956 static int rcu_verify_early_boot_tests(void)
957 {
958 	int ret = 0;
959 	int early_boot_test_counter = 0;
960 
961 	if (rcu_self_test) {
962 		early_boot_test_counter++;
963 		rcu_barrier();
964 	}
965 	if (rcu_self_test_bh) {
966 		early_boot_test_counter++;
967 		rcu_barrier_bh();
968 	}
969 	if (rcu_self_test_sched) {
970 		early_boot_test_counter++;
971 		rcu_barrier_sched();
972 	}
973 
974 	if (rcu_self_test_counter != early_boot_test_counter) {
975 		WARN_ON(1);
976 		ret = -1;
977 	}
978 
979 	return ret;
980 }
981 late_initcall(rcu_verify_early_boot_tests);
982 #else
983 void rcu_early_boot_tests(void) {}
984 #endif /* CONFIG_PROVE_RCU */
985 
986 #ifndef CONFIG_TINY_RCU
987 
988 /*
989  * Print any significant non-default boot-time settings.
990  */
991 void __init rcupdate_announce_bootup_oddness(void)
992 {
993 	if (rcu_normal)
994 		pr_info("\tNo expedited grace period (rcu_normal).\n");
995 	else if (rcu_normal_after_boot)
996 		pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n");
997 	else if (rcu_expedited)
998 		pr_info("\tAll grace periods are expedited (rcu_expedited).\n");
999 	if (rcu_cpu_stall_suppress)
1000 		pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n");
1001 	if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT)
1002 		pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout);
1003 	rcu_tasks_bootup_oddness();
1004 }
1005 
1006 #endif /* #ifndef CONFIG_TINY_RCU */
1007