Loading...
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
63#ifndef CONFIG_TINY_RCU
64module_param(rcu_expedited, int, 0);
65module_param(rcu_normal, int, 0);
66static int rcu_normal_after_boot;
67module_param(rcu_normal_after_boot, int, 0);
68#endif /* #ifndef CONFIG_TINY_RCU */
69
70#if defined(CONFIG_DEBUG_LOCK_ALLOC) && defined(CONFIG_PREEMPT_COUNT)
71/**
72 * rcu_read_lock_sched_held() - might we be in RCU-sched read-side critical section?
73 *
74 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an
75 * RCU-sched read-side critical section. In absence of
76 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
77 * critical section unless it can prove otherwise. Note that disabling
78 * of preemption (including disabling irqs) counts as an RCU-sched
79 * read-side critical section. This is useful for debug checks in functions
80 * that required that they be called within an RCU-sched read-side
81 * critical section.
82 *
83 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
84 * and while lockdep is disabled.
85 *
86 * Note that if the CPU is in the idle loop from an RCU point of
87 * view (ie: that we are in the section between rcu_idle_enter() and
88 * rcu_idle_exit()) then rcu_read_lock_held() returns false even if the CPU
89 * did an rcu_read_lock(). The reason for this is that RCU ignores CPUs
90 * that are in such a section, considering these as in extended quiescent
91 * state, so such a CPU is effectively never in an RCU read-side critical
92 * section regardless of what RCU primitives it invokes. This state of
93 * affairs is required --- we need to keep an RCU-free window in idle
94 * where the CPU may possibly enter into low power mode. This way we can
95 * notice an extended quiescent state to other CPUs that started a grace
96 * period. Otherwise we would delay any grace period as long as we run in
97 * the idle task.
98 *
99 * Similarly, we avoid claiming an SRCU read lock held if the current
100 * CPU is offline.
101 */
102int rcu_read_lock_sched_held(void)
103{
104 int lockdep_opinion = 0;
105
106 if (!debug_lockdep_rcu_enabled())
107 return 1;
108 if (!rcu_is_watching())
109 return 0;
110 if (!rcu_lockdep_current_cpu_online())
111 return 0;
112 if (debug_locks)
113 lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
114 return lockdep_opinion || preempt_count() != 0 || irqs_disabled();
115}
116EXPORT_SYMBOL(rcu_read_lock_sched_held);
117#endif
118
119#ifndef CONFIG_TINY_RCU
120
121/*
122 * Should expedited grace-period primitives always fall back to their
123 * non-expedited counterparts? Intended for use within RCU. Note
124 * that if the user specifies both rcu_expedited and rcu_normal, then
125 * rcu_normal wins.
126 */
127bool rcu_gp_is_normal(void)
128{
129 return READ_ONCE(rcu_normal);
130}
131EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
132
133static atomic_t rcu_expedited_nesting =
134 ATOMIC_INIT(IS_ENABLED(CONFIG_RCU_EXPEDITE_BOOT) ? 1 : 0);
135
136/*
137 * Should normal grace-period primitives be expedited? Intended for
138 * use within RCU. Note that this function takes the rcu_expedited
139 * sysfs/boot variable into account as well as the rcu_expedite_gp()
140 * nesting. So looping on rcu_unexpedite_gp() until rcu_gp_is_expedited()
141 * returns false is a -really- bad idea.
142 */
143bool rcu_gp_is_expedited(void)
144{
145 return rcu_expedited || atomic_read(&rcu_expedited_nesting);
146}
147EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
148
149/**
150 * rcu_expedite_gp - Expedite future RCU grace periods
151 *
152 * After a call to this function, future calls to synchronize_rcu() and
153 * friends act as the corresponding synchronize_rcu_expedited() function
154 * had instead been called.
155 */
156void rcu_expedite_gp(void)
157{
158 atomic_inc(&rcu_expedited_nesting);
159}
160EXPORT_SYMBOL_GPL(rcu_expedite_gp);
161
162/**
163 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
164 *
165 * Undo a prior call to rcu_expedite_gp(). If all prior calls to
166 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
167 * and if the rcu_expedited sysfs/boot parameter is not set, then all
168 * subsequent calls to synchronize_rcu() and friends will return to
169 * their normal non-expedited behavior.
170 */
171void rcu_unexpedite_gp(void)
172{
173 atomic_dec(&rcu_expedited_nesting);
174}
175EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
176
177/*
178 * Inform RCU of the end of the in-kernel boot sequence.
179 */
180void rcu_end_inkernel_boot(void)
181{
182 if (IS_ENABLED(CONFIG_RCU_EXPEDITE_BOOT))
183 rcu_unexpedite_gp();
184 if (rcu_normal_after_boot)
185 WRITE_ONCE(rcu_normal, 1);
186}
187
188#endif /* #ifndef CONFIG_TINY_RCU */
189
190#ifdef CONFIG_PREEMPT_RCU
191
192/*
193 * Preemptible RCU implementation for rcu_read_lock().
194 * Just increment ->rcu_read_lock_nesting, shared state will be updated
195 * if we block.
196 */
197void __rcu_read_lock(void)
198{
199 current->rcu_read_lock_nesting++;
200 barrier(); /* critical section after entry code. */
201}
202EXPORT_SYMBOL_GPL(__rcu_read_lock);
203
204/*
205 * Preemptible RCU implementation for rcu_read_unlock().
206 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
207 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
208 * invoke rcu_read_unlock_special() to clean up after a context switch
209 * in an RCU read-side critical section and other special cases.
210 */
211void __rcu_read_unlock(void)
212{
213 struct task_struct *t = current;
214
215 if (t->rcu_read_lock_nesting != 1) {
216 --t->rcu_read_lock_nesting;
217 } else {
218 barrier(); /* critical section before exit code. */
219 t->rcu_read_lock_nesting = INT_MIN;
220 barrier(); /* assign before ->rcu_read_unlock_special load */
221 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
222 rcu_read_unlock_special(t);
223 barrier(); /* ->rcu_read_unlock_special load before assign */
224 t->rcu_read_lock_nesting = 0;
225 }
226#ifdef CONFIG_PROVE_LOCKING
227 {
228 int rrln = READ_ONCE(t->rcu_read_lock_nesting);
229
230 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
231 }
232#endif /* #ifdef CONFIG_PROVE_LOCKING */
233}
234EXPORT_SYMBOL_GPL(__rcu_read_unlock);
235
236#endif /* #ifdef CONFIG_PREEMPT_RCU */
237
238#ifdef CONFIG_DEBUG_LOCK_ALLOC
239static struct lock_class_key rcu_lock_key;
240struct lockdep_map rcu_lock_map =
241 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
242EXPORT_SYMBOL_GPL(rcu_lock_map);
243
244static struct lock_class_key rcu_bh_lock_key;
245struct lockdep_map rcu_bh_lock_map =
246 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
247EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
248
249static struct lock_class_key rcu_sched_lock_key;
250struct lockdep_map rcu_sched_lock_map =
251 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
252EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
253
254static struct lock_class_key rcu_callback_key;
255struct lockdep_map rcu_callback_map =
256 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
257EXPORT_SYMBOL_GPL(rcu_callback_map);
258
259int notrace debug_lockdep_rcu_enabled(void)
260{
261 return rcu_scheduler_active && debug_locks &&
262 current->lockdep_recursion == 0;
263}
264EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
265
266/**
267 * rcu_read_lock_held() - might we be in RCU read-side critical section?
268 *
269 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
270 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC,
271 * this assumes we are in an RCU read-side critical section unless it can
272 * prove otherwise. This is useful for debug checks in functions that
273 * require that they be called within an RCU read-side critical section.
274 *
275 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
276 * and while lockdep is disabled.
277 *
278 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
279 * occur in the same context, for example, it is illegal to invoke
280 * rcu_read_unlock() in process context if the matching rcu_read_lock()
281 * was invoked from within an irq handler.
282 *
283 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
284 * offline from an RCU perspective, so check for those as well.
285 */
286int rcu_read_lock_held(void)
287{
288 if (!debug_lockdep_rcu_enabled())
289 return 1;
290 if (!rcu_is_watching())
291 return 0;
292 if (!rcu_lockdep_current_cpu_online())
293 return 0;
294 return lock_is_held(&rcu_lock_map);
295}
296EXPORT_SYMBOL_GPL(rcu_read_lock_held);
297
298/**
299 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
300 *
301 * Check for bottom half being disabled, which covers both the
302 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
303 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
304 * will show the situation. This is useful for debug checks in functions
305 * that require that they be called within an RCU read-side critical
306 * section.
307 *
308 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
309 *
310 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
311 * offline from an RCU perspective, so check for those as well.
312 */
313int rcu_read_lock_bh_held(void)
314{
315 if (!debug_lockdep_rcu_enabled())
316 return 1;
317 if (!rcu_is_watching())
318 return 0;
319 if (!rcu_lockdep_current_cpu_online())
320 return 0;
321 return in_softirq() || irqs_disabled();
322}
323EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
324
325#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
326
327/**
328 * wakeme_after_rcu() - Callback function to awaken a task after grace period
329 * @head: Pointer to rcu_head member within rcu_synchronize structure
330 *
331 * Awaken the corresponding task now that a grace period has elapsed.
332 */
333void wakeme_after_rcu(struct rcu_head *head)
334{
335 struct rcu_synchronize *rcu;
336
337 rcu = container_of(head, struct rcu_synchronize, head);
338 complete(&rcu->completion);
339}
340EXPORT_SYMBOL_GPL(wakeme_after_rcu);
341
342void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
343 struct rcu_synchronize *rs_array)
344{
345 int i;
346
347 /* Initialize and register callbacks for each flavor specified. */
348 for (i = 0; i < n; i++) {
349 if (checktiny &&
350 (crcu_array[i] == call_rcu ||
351 crcu_array[i] == call_rcu_bh)) {
352 might_sleep();
353 continue;
354 }
355 init_rcu_head_on_stack(&rs_array[i].head);
356 init_completion(&rs_array[i].completion);
357 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
358 }
359
360 /* Wait for all callbacks to be invoked. */
361 for (i = 0; i < n; i++) {
362 if (checktiny &&
363 (crcu_array[i] == call_rcu ||
364 crcu_array[i] == call_rcu_bh))
365 continue;
366 wait_for_completion(&rs_array[i].completion);
367 destroy_rcu_head_on_stack(&rs_array[i].head);
368 }
369}
370EXPORT_SYMBOL_GPL(__wait_rcu_gp);
371
372#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
373void init_rcu_head(struct rcu_head *head)
374{
375 debug_object_init(head, &rcuhead_debug_descr);
376}
377
378void destroy_rcu_head(struct rcu_head *head)
379{
380 debug_object_free(head, &rcuhead_debug_descr);
381}
382
383/*
384 * fixup_activate is called when:
385 * - an active object is activated
386 * - an unknown object is activated (might be a statically initialized object)
387 * Activation is performed internally by call_rcu().
388 */
389static int rcuhead_fixup_activate(void *addr, enum debug_obj_state state)
390{
391 struct rcu_head *head = addr;
392
393 switch (state) {
394
395 case ODEBUG_STATE_NOTAVAILABLE:
396 /*
397 * This is not really a fixup. We just make sure that it is
398 * tracked in the object tracker.
399 */
400 debug_object_init(head, &rcuhead_debug_descr);
401 debug_object_activate(head, &rcuhead_debug_descr);
402 return 0;
403 default:
404 return 1;
405 }
406}
407
408/**
409 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
410 * @head: pointer to rcu_head structure to be initialized
411 *
412 * This function informs debugobjects of a new rcu_head structure that
413 * has been allocated as an auto variable on the stack. This function
414 * is not required for rcu_head structures that are statically defined or
415 * that are dynamically allocated on the heap. This function has no
416 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
417 */
418void init_rcu_head_on_stack(struct rcu_head *head)
419{
420 debug_object_init_on_stack(head, &rcuhead_debug_descr);
421}
422EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
423
424/**
425 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
426 * @head: pointer to rcu_head structure to be initialized
427 *
428 * This function informs debugobjects that an on-stack rcu_head structure
429 * is about to go out of scope. As with init_rcu_head_on_stack(), this
430 * function is not required for rcu_head structures that are statically
431 * defined or that are dynamically allocated on the heap. Also as with
432 * init_rcu_head_on_stack(), this function has no effect for
433 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
434 */
435void destroy_rcu_head_on_stack(struct rcu_head *head)
436{
437 debug_object_free(head, &rcuhead_debug_descr);
438}
439EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
440
441struct debug_obj_descr rcuhead_debug_descr = {
442 .name = "rcu_head",
443 .fixup_activate = rcuhead_fixup_activate,
444};
445EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
446#endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
447
448#if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
449void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
450 unsigned long secs,
451 unsigned long c_old, unsigned long c)
452{
453 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
454}
455EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
456#else
457#define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
458 do { } while (0)
459#endif
460
461#ifdef CONFIG_RCU_STALL_COMMON
462
463#ifdef CONFIG_PROVE_RCU
464#define RCU_STALL_DELAY_DELTA (5 * HZ)
465#else
466#define RCU_STALL_DELAY_DELTA 0
467#endif
468
469int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
470static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
471
472module_param(rcu_cpu_stall_suppress, int, 0644);
473module_param(rcu_cpu_stall_timeout, int, 0644);
474
475int rcu_jiffies_till_stall_check(void)
476{
477 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
478
479 /*
480 * Limit check must be consistent with the Kconfig limits
481 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
482 */
483 if (till_stall_check < 3) {
484 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
485 till_stall_check = 3;
486 } else if (till_stall_check > 300) {
487 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
488 till_stall_check = 300;
489 }
490 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
491}
492
493void rcu_sysrq_start(void)
494{
495 if (!rcu_cpu_stall_suppress)
496 rcu_cpu_stall_suppress = 2;
497}
498
499void rcu_sysrq_end(void)
500{
501 if (rcu_cpu_stall_suppress == 2)
502 rcu_cpu_stall_suppress = 0;
503}
504
505static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
506{
507 rcu_cpu_stall_suppress = 1;
508 return NOTIFY_DONE;
509}
510
511static struct notifier_block rcu_panic_block = {
512 .notifier_call = rcu_panic,
513};
514
515static int __init check_cpu_stall_init(void)
516{
517 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
518 return 0;
519}
520early_initcall(check_cpu_stall_init);
521
522#endif /* #ifdef CONFIG_RCU_STALL_COMMON */
523
524#ifdef CONFIG_TASKS_RCU
525
526/*
527 * Simple variant of RCU whose quiescent states are voluntary context switch,
528 * user-space execution, and idle. As such, grace periods can take one good
529 * long time. There are no read-side primitives similar to rcu_read_lock()
530 * and rcu_read_unlock() because this implementation is intended to get
531 * the system into a safe state for some of the manipulations involved in
532 * tracing and the like. Finally, this implementation does not support
533 * high call_rcu_tasks() rates from multiple CPUs. If this is required,
534 * per-CPU callback lists will be needed.
535 */
536
537/* Global list of callbacks and associated lock. */
538static struct rcu_head *rcu_tasks_cbs_head;
539static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
540static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
541static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
542
543/* Track exiting tasks in order to allow them to be waited for. */
544DEFINE_SRCU(tasks_rcu_exit_srcu);
545
546/* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
547static int rcu_task_stall_timeout __read_mostly = HZ * 60 * 10;
548module_param(rcu_task_stall_timeout, int, 0644);
549
550static void rcu_spawn_tasks_kthread(void);
551
552/*
553 * Post an RCU-tasks callback. First call must be from process context
554 * after the scheduler if fully operational.
555 */
556void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
557{
558 unsigned long flags;
559 bool needwake;
560
561 rhp->next = NULL;
562 rhp->func = func;
563 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
564 needwake = !rcu_tasks_cbs_head;
565 *rcu_tasks_cbs_tail = rhp;
566 rcu_tasks_cbs_tail = &rhp->next;
567 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
568 if (needwake) {
569 rcu_spawn_tasks_kthread();
570 wake_up(&rcu_tasks_cbs_wq);
571 }
572}
573EXPORT_SYMBOL_GPL(call_rcu_tasks);
574
575/**
576 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
577 *
578 * Control will return to the caller some time after a full rcu-tasks
579 * grace period has elapsed, in other words after all currently
580 * executing rcu-tasks read-side critical sections have elapsed. These
581 * read-side critical sections are delimited by calls to schedule(),
582 * cond_resched_rcu_qs(), idle execution, userspace execution, calls
583 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
584 *
585 * This is a very specialized primitive, intended only for a few uses in
586 * tracing and other situations requiring manipulation of function
587 * preambles and profiling hooks. The synchronize_rcu_tasks() function
588 * is not (yet) intended for heavy use from multiple CPUs.
589 *
590 * Note that this guarantee implies further memory-ordering guarantees.
591 * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
592 * each CPU is guaranteed to have executed a full memory barrier since the
593 * end of its last RCU-tasks read-side critical section whose beginning
594 * preceded the call to synchronize_rcu_tasks(). In addition, each CPU
595 * having an RCU-tasks read-side critical section that extends beyond
596 * the return from synchronize_rcu_tasks() is guaranteed to have executed
597 * a full memory barrier after the beginning of synchronize_rcu_tasks()
598 * and before the beginning of that RCU-tasks read-side critical section.
599 * Note that these guarantees include CPUs that are offline, idle, or
600 * executing in user mode, as well as CPUs that are executing in the kernel.
601 *
602 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
603 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
604 * to have executed a full memory barrier during the execution of
605 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
606 * (but again only if the system has more than one CPU).
607 */
608void synchronize_rcu_tasks(void)
609{
610 /* Complain if the scheduler has not started. */
611 RCU_LOCKDEP_WARN(!rcu_scheduler_active,
612 "synchronize_rcu_tasks called too soon");
613
614 /* Wait for the grace period. */
615 wait_rcu_gp(call_rcu_tasks);
616}
617EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
618
619/**
620 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
621 *
622 * Although the current implementation is guaranteed to wait, it is not
623 * obligated to, for example, if there are no pending callbacks.
624 */
625void rcu_barrier_tasks(void)
626{
627 /* There is only one callback queue, so this is easy. ;-) */
628 synchronize_rcu_tasks();
629}
630EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
631
632/* See if tasks are still holding out, complain if so. */
633static void check_holdout_task(struct task_struct *t,
634 bool needreport, bool *firstreport)
635{
636 int cpu;
637
638 if (!READ_ONCE(t->rcu_tasks_holdout) ||
639 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
640 !READ_ONCE(t->on_rq) ||
641 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
642 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
643 WRITE_ONCE(t->rcu_tasks_holdout, false);
644 list_del_init(&t->rcu_tasks_holdout_list);
645 put_task_struct(t);
646 return;
647 }
648 if (!needreport)
649 return;
650 if (*firstreport) {
651 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
652 *firstreport = false;
653 }
654 cpu = task_cpu(t);
655 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
656 t, ".I"[is_idle_task(t)],
657 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
658 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
659 t->rcu_tasks_idle_cpu, cpu);
660 sched_show_task(t);
661}
662
663/* RCU-tasks kthread that detects grace periods and invokes callbacks. */
664static int __noreturn rcu_tasks_kthread(void *arg)
665{
666 unsigned long flags;
667 struct task_struct *g, *t;
668 unsigned long lastreport;
669 struct rcu_head *list;
670 struct rcu_head *next;
671 LIST_HEAD(rcu_tasks_holdouts);
672
673 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
674 housekeeping_affine(current);
675
676 /*
677 * Each pass through the following loop makes one check for
678 * newly arrived callbacks, and, if there are some, waits for
679 * one RCU-tasks grace period and then invokes the callbacks.
680 * This loop is terminated by the system going down. ;-)
681 */
682 for (;;) {
683
684 /* Pick up any new callbacks. */
685 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
686 list = rcu_tasks_cbs_head;
687 rcu_tasks_cbs_head = NULL;
688 rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
689 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
690
691 /* If there were none, wait a bit and start over. */
692 if (!list) {
693 wait_event_interruptible(rcu_tasks_cbs_wq,
694 rcu_tasks_cbs_head);
695 if (!rcu_tasks_cbs_head) {
696 WARN_ON(signal_pending(current));
697 schedule_timeout_interruptible(HZ/10);
698 }
699 continue;
700 }
701
702 /*
703 * Wait for all pre-existing t->on_rq and t->nvcsw
704 * transitions to complete. Invoking synchronize_sched()
705 * suffices because all these transitions occur with
706 * interrupts disabled. Without this synchronize_sched(),
707 * a read-side critical section that started before the
708 * grace period might be incorrectly seen as having started
709 * after the grace period.
710 *
711 * This synchronize_sched() also dispenses with the
712 * need for a memory barrier on the first store to
713 * ->rcu_tasks_holdout, as it forces the store to happen
714 * after the beginning of the grace period.
715 */
716 synchronize_sched();
717
718 /*
719 * There were callbacks, so we need to wait for an
720 * RCU-tasks grace period. Start off by scanning
721 * the task list for tasks that are not already
722 * voluntarily blocked. Mark these tasks and make
723 * a list of them in rcu_tasks_holdouts.
724 */
725 rcu_read_lock();
726 for_each_process_thread(g, t) {
727 if (t != current && READ_ONCE(t->on_rq) &&
728 !is_idle_task(t)) {
729 get_task_struct(t);
730 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
731 WRITE_ONCE(t->rcu_tasks_holdout, true);
732 list_add(&t->rcu_tasks_holdout_list,
733 &rcu_tasks_holdouts);
734 }
735 }
736 rcu_read_unlock();
737
738 /*
739 * Wait for tasks that are in the process of exiting.
740 * This does only part of the job, ensuring that all
741 * tasks that were previously exiting reach the point
742 * where they have disabled preemption, allowing the
743 * later synchronize_sched() to finish the job.
744 */
745 synchronize_srcu(&tasks_rcu_exit_srcu);
746
747 /*
748 * Each pass through the following loop scans the list
749 * of holdout tasks, removing any that are no longer
750 * holdouts. When the list is empty, we are done.
751 */
752 lastreport = jiffies;
753 while (!list_empty(&rcu_tasks_holdouts)) {
754 bool firstreport;
755 bool needreport;
756 int rtst;
757 struct task_struct *t1;
758
759 schedule_timeout_interruptible(HZ);
760 rtst = READ_ONCE(rcu_task_stall_timeout);
761 needreport = rtst > 0 &&
762 time_after(jiffies, lastreport + rtst);
763 if (needreport)
764 lastreport = jiffies;
765 firstreport = true;
766 WARN_ON(signal_pending(current));
767 list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
768 rcu_tasks_holdout_list) {
769 check_holdout_task(t, needreport, &firstreport);
770 cond_resched();
771 }
772 }
773
774 /*
775 * Because ->on_rq and ->nvcsw are not guaranteed
776 * to have a full memory barriers prior to them in the
777 * schedule() path, memory reordering on other CPUs could
778 * cause their RCU-tasks read-side critical sections to
779 * extend past the end of the grace period. However,
780 * because these ->nvcsw updates are carried out with
781 * interrupts disabled, we can use synchronize_sched()
782 * to force the needed ordering on all such CPUs.
783 *
784 * This synchronize_sched() also confines all
785 * ->rcu_tasks_holdout accesses to be within the grace
786 * period, avoiding the need for memory barriers for
787 * ->rcu_tasks_holdout accesses.
788 *
789 * In addition, this synchronize_sched() waits for exiting
790 * tasks to complete their final preempt_disable() region
791 * of execution, cleaning up after the synchronize_srcu()
792 * above.
793 */
794 synchronize_sched();
795
796 /* Invoke the callbacks. */
797 while (list) {
798 next = list->next;
799 local_bh_disable();
800 list->func(list);
801 local_bh_enable();
802 list = next;
803 cond_resched();
804 }
805 schedule_timeout_uninterruptible(HZ/10);
806 }
807}
808
809/* Spawn rcu_tasks_kthread() at first call to call_rcu_tasks(). */
810static void rcu_spawn_tasks_kthread(void)
811{
812 static DEFINE_MUTEX(rcu_tasks_kthread_mutex);
813 static struct task_struct *rcu_tasks_kthread_ptr;
814 struct task_struct *t;
815
816 if (READ_ONCE(rcu_tasks_kthread_ptr)) {
817 smp_mb(); /* Ensure caller sees full kthread. */
818 return;
819 }
820 mutex_lock(&rcu_tasks_kthread_mutex);
821 if (rcu_tasks_kthread_ptr) {
822 mutex_unlock(&rcu_tasks_kthread_mutex);
823 return;
824 }
825 t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
826 BUG_ON(IS_ERR(t));
827 smp_mb(); /* Ensure others see full kthread. */
828 WRITE_ONCE(rcu_tasks_kthread_ptr, t);
829 mutex_unlock(&rcu_tasks_kthread_mutex);
830}
831
832#endif /* #ifdef CONFIG_TASKS_RCU */
833
834#ifdef CONFIG_PROVE_RCU
835
836/*
837 * Early boot self test parameters, one for each flavor
838 */
839static bool rcu_self_test;
840static bool rcu_self_test_bh;
841static bool rcu_self_test_sched;
842
843module_param(rcu_self_test, bool, 0444);
844module_param(rcu_self_test_bh, bool, 0444);
845module_param(rcu_self_test_sched, bool, 0444);
846
847static int rcu_self_test_counter;
848
849static void test_callback(struct rcu_head *r)
850{
851 rcu_self_test_counter++;
852 pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
853}
854
855static void early_boot_test_call_rcu(void)
856{
857 static struct rcu_head head;
858
859 call_rcu(&head, test_callback);
860}
861
862static void early_boot_test_call_rcu_bh(void)
863{
864 static struct rcu_head head;
865
866 call_rcu_bh(&head, test_callback);
867}
868
869static void early_boot_test_call_rcu_sched(void)
870{
871 static struct rcu_head head;
872
873 call_rcu_sched(&head, test_callback);
874}
875
876void rcu_early_boot_tests(void)
877{
878 pr_info("Running RCU self tests\n");
879
880 if (rcu_self_test)
881 early_boot_test_call_rcu();
882 if (rcu_self_test_bh)
883 early_boot_test_call_rcu_bh();
884 if (rcu_self_test_sched)
885 early_boot_test_call_rcu_sched();
886}
887
888static int rcu_verify_early_boot_tests(void)
889{
890 int ret = 0;
891 int early_boot_test_counter = 0;
892
893 if (rcu_self_test) {
894 early_boot_test_counter++;
895 rcu_barrier();
896 }
897 if (rcu_self_test_bh) {
898 early_boot_test_counter++;
899 rcu_barrier_bh();
900 }
901 if (rcu_self_test_sched) {
902 early_boot_test_counter++;
903 rcu_barrier_sched();
904 }
905
906 if (rcu_self_test_counter != early_boot_test_counter) {
907 WARN_ON(1);
908 ret = -1;
909 }
910
911 return ret;
912}
913late_initcall(rcu_verify_early_boot_tests);
914#else
915void rcu_early_boot_tests(void) {}
916#endif /* CONFIG_PROVE_RCU */
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update mechanism for mutual exclusion
4 *
5 * Copyright IBM Corporation, 2001
6 *
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 *
10 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
11 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
12 * Papers:
13 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
14 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
15 *
16 * For detailed explanation of Read-Copy Update mechanism see -
17 * http://lse.sourceforge.net/locking/rcupdate.html
18 *
19 */
20#include <linux/types.h>
21#include <linux/kernel.h>
22#include <linux/init.h>
23#include <linux/spinlock.h>
24#include <linux/smp.h>
25#include <linux/interrupt.h>
26#include <linux/sched/signal.h>
27#include <linux/sched/debug.h>
28#include <linux/atomic.h>
29#include <linux/bitops.h>
30#include <linux/percpu.h>
31#include <linux/notifier.h>
32#include <linux/cpu.h>
33#include <linux/mutex.h>
34#include <linux/export.h>
35#include <linux/hardirq.h>
36#include <linux/delay.h>
37#include <linux/moduleparam.h>
38#include <linux/kthread.h>
39#include <linux/tick.h>
40#include <linux/rcupdate_wait.h>
41#include <linux/sched/isolation.h>
42#include <linux/kprobes.h>
43
44#define CREATE_TRACE_POINTS
45
46#include "rcu.h"
47
48#ifdef MODULE_PARAM_PREFIX
49#undef MODULE_PARAM_PREFIX
50#endif
51#define MODULE_PARAM_PREFIX "rcupdate."
52
53#ifndef CONFIG_TINY_RCU
54extern int rcu_expedited; /* from sysctl */
55module_param(rcu_expedited, int, 0);
56extern int rcu_normal; /* from sysctl */
57module_param(rcu_normal, int, 0);
58static int rcu_normal_after_boot;
59module_param(rcu_normal_after_boot, int, 0);
60#endif /* #ifndef CONFIG_TINY_RCU */
61
62#ifdef CONFIG_DEBUG_LOCK_ALLOC
63/**
64 * rcu_read_lock_held_common() - might we be in RCU-sched read-side critical section?
65 * @ret: Best guess answer if lockdep cannot be relied on
66 *
67 * Returns true if lockdep must be ignored, in which case *ret contains
68 * the best guess described below. Otherwise returns false, in which
69 * case *ret tells the caller nothing and the caller should instead
70 * consult lockdep.
71 *
72 * If CONFIG_DEBUG_LOCK_ALLOC is selected, set *ret to nonzero iff in an
73 * RCU-sched read-side critical section. In absence of
74 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
75 * critical section unless it can prove otherwise. Note that disabling
76 * of preemption (including disabling irqs) counts as an RCU-sched
77 * read-side critical section. This is useful for debug checks in functions
78 * that required that they be called within an RCU-sched read-side
79 * critical section.
80 *
81 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
82 * and while lockdep is disabled.
83 *
84 * Note that if the CPU is in the idle loop from an RCU point of view (ie:
85 * that we are in the section between rcu_idle_enter() and rcu_idle_exit())
86 * then rcu_read_lock_held() sets *ret to false even if the CPU did an
87 * rcu_read_lock(). The reason for this is that RCU ignores CPUs that are
88 * in such a section, considering these as in extended quiescent state,
89 * so such a CPU is effectively never in an RCU read-side critical section
90 * regardless of what RCU primitives it invokes. This state of affairs is
91 * required --- we need to keep an RCU-free window in idle where the CPU may
92 * possibly enter into low power mode. This way we can notice an extended
93 * quiescent state to other CPUs that started a grace period. Otherwise
94 * we would delay any grace period as long as we run in the idle task.
95 *
96 * Similarly, we avoid claiming an RCU read lock held if the current
97 * CPU is offline.
98 */
99static bool rcu_read_lock_held_common(bool *ret)
100{
101 if (!debug_lockdep_rcu_enabled()) {
102 *ret = 1;
103 return true;
104 }
105 if (!rcu_is_watching()) {
106 *ret = 0;
107 return true;
108 }
109 if (!rcu_lockdep_current_cpu_online()) {
110 *ret = 0;
111 return true;
112 }
113 return false;
114}
115
116int rcu_read_lock_sched_held(void)
117{
118 bool ret;
119
120 if (rcu_read_lock_held_common(&ret))
121 return ret;
122 return lock_is_held(&rcu_sched_lock_map) || !preemptible();
123}
124EXPORT_SYMBOL(rcu_read_lock_sched_held);
125#endif
126
127#ifndef CONFIG_TINY_RCU
128
129/*
130 * Should expedited grace-period primitives always fall back to their
131 * non-expedited counterparts? Intended for use within RCU. Note
132 * that if the user specifies both rcu_expedited and rcu_normal, then
133 * rcu_normal wins. (Except during the time period during boot from
134 * when the first task is spawned until the rcu_set_runtime_mode()
135 * core_initcall() is invoked, at which point everything is expedited.)
136 */
137bool rcu_gp_is_normal(void)
138{
139 return READ_ONCE(rcu_normal) &&
140 rcu_scheduler_active != RCU_SCHEDULER_INIT;
141}
142EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
143
144static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1);
145
146/*
147 * Should normal grace-period primitives be expedited? Intended for
148 * use within RCU. Note that this function takes the rcu_expedited
149 * sysfs/boot variable and rcu_scheduler_active into account as well
150 * as the rcu_expedite_gp() nesting. So looping on rcu_unexpedite_gp()
151 * until rcu_gp_is_expedited() returns false is a -really- bad idea.
152 */
153bool rcu_gp_is_expedited(void)
154{
155 return rcu_expedited || atomic_read(&rcu_expedited_nesting);
156}
157EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
158
159/**
160 * rcu_expedite_gp - Expedite future RCU grace periods
161 *
162 * After a call to this function, future calls to synchronize_rcu() and
163 * friends act as the corresponding synchronize_rcu_expedited() function
164 * had instead been called.
165 */
166void rcu_expedite_gp(void)
167{
168 atomic_inc(&rcu_expedited_nesting);
169}
170EXPORT_SYMBOL_GPL(rcu_expedite_gp);
171
172/**
173 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
174 *
175 * Undo a prior call to rcu_expedite_gp(). If all prior calls to
176 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
177 * and if the rcu_expedited sysfs/boot parameter is not set, then all
178 * subsequent calls to synchronize_rcu() and friends will return to
179 * their normal non-expedited behavior.
180 */
181void rcu_unexpedite_gp(void)
182{
183 atomic_dec(&rcu_expedited_nesting);
184}
185EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
186
187/*
188 * Inform RCU of the end of the in-kernel boot sequence.
189 */
190void rcu_end_inkernel_boot(void)
191{
192 rcu_unexpedite_gp();
193 if (rcu_normal_after_boot)
194 WRITE_ONCE(rcu_normal, 1);
195}
196
197#endif /* #ifndef CONFIG_TINY_RCU */
198
199/*
200 * Test each non-SRCU synchronous grace-period wait API. This is
201 * useful just after a change in mode for these primitives, and
202 * during early boot.
203 */
204void rcu_test_sync_prims(void)
205{
206 if (!IS_ENABLED(CONFIG_PROVE_RCU))
207 return;
208 synchronize_rcu();
209 synchronize_rcu_expedited();
210}
211
212#if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU)
213
214/*
215 * Switch to run-time mode once RCU has fully initialized.
216 */
217static int __init rcu_set_runtime_mode(void)
218{
219 rcu_test_sync_prims();
220 rcu_scheduler_active = RCU_SCHEDULER_RUNNING;
221 rcu_test_sync_prims();
222 return 0;
223}
224core_initcall(rcu_set_runtime_mode);
225
226#endif /* #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) */
227
228#ifdef CONFIG_DEBUG_LOCK_ALLOC
229static struct lock_class_key rcu_lock_key;
230struct lockdep_map rcu_lock_map =
231 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
232EXPORT_SYMBOL_GPL(rcu_lock_map);
233
234static struct lock_class_key rcu_bh_lock_key;
235struct lockdep_map rcu_bh_lock_map =
236 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
237EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
238
239static struct lock_class_key rcu_sched_lock_key;
240struct lockdep_map rcu_sched_lock_map =
241 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
242EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
243
244static struct lock_class_key rcu_callback_key;
245struct lockdep_map rcu_callback_map =
246 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
247EXPORT_SYMBOL_GPL(rcu_callback_map);
248
249int notrace debug_lockdep_rcu_enabled(void)
250{
251 return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks &&
252 current->lockdep_recursion == 0;
253}
254EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
255NOKPROBE_SYMBOL(debug_lockdep_rcu_enabled);
256
257/**
258 * rcu_read_lock_held() - might we be in RCU read-side critical section?
259 *
260 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
261 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC,
262 * this assumes we are in an RCU read-side critical section unless it can
263 * prove otherwise. This is useful for debug checks in functions that
264 * require that they be called within an RCU read-side critical section.
265 *
266 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
267 * and while lockdep is disabled.
268 *
269 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
270 * occur in the same context, for example, it is illegal to invoke
271 * rcu_read_unlock() in process context if the matching rcu_read_lock()
272 * was invoked from within an irq handler.
273 *
274 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
275 * offline from an RCU perspective, so check for those as well.
276 */
277int rcu_read_lock_held(void)
278{
279 bool ret;
280
281 if (rcu_read_lock_held_common(&ret))
282 return ret;
283 return lock_is_held(&rcu_lock_map);
284}
285EXPORT_SYMBOL_GPL(rcu_read_lock_held);
286
287/**
288 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
289 *
290 * Check for bottom half being disabled, which covers both the
291 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
292 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
293 * will show the situation. This is useful for debug checks in functions
294 * that require that they be called within an RCU read-side critical
295 * section.
296 *
297 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
298 *
299 * Note that rcu_read_lock_bh() is disallowed if the CPU is either idle or
300 * offline from an RCU perspective, so check for those as well.
301 */
302int rcu_read_lock_bh_held(void)
303{
304 bool ret;
305
306 if (rcu_read_lock_held_common(&ret))
307 return ret;
308 return in_softirq() || irqs_disabled();
309}
310EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
311
312int rcu_read_lock_any_held(void)
313{
314 bool ret;
315
316 if (rcu_read_lock_held_common(&ret))
317 return ret;
318 if (lock_is_held(&rcu_lock_map) ||
319 lock_is_held(&rcu_bh_lock_map) ||
320 lock_is_held(&rcu_sched_lock_map))
321 return 1;
322 return !preemptible();
323}
324EXPORT_SYMBOL_GPL(rcu_read_lock_any_held);
325
326#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
327
328/**
329 * wakeme_after_rcu() - Callback function to awaken a task after grace period
330 * @head: Pointer to rcu_head member within rcu_synchronize structure
331 *
332 * Awaken the corresponding task now that a grace period has elapsed.
333 */
334void wakeme_after_rcu(struct rcu_head *head)
335{
336 struct rcu_synchronize *rcu;
337
338 rcu = container_of(head, struct rcu_synchronize, head);
339 complete(&rcu->completion);
340}
341EXPORT_SYMBOL_GPL(wakeme_after_rcu);
342
343void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
344 struct rcu_synchronize *rs_array)
345{
346 int i;
347 int j;
348
349 /* Initialize and register callbacks for each crcu_array element. */
350 for (i = 0; i < n; i++) {
351 if (checktiny &&
352 (crcu_array[i] == call_rcu)) {
353 might_sleep();
354 continue;
355 }
356 init_rcu_head_on_stack(&rs_array[i].head);
357 init_completion(&rs_array[i].completion);
358 for (j = 0; j < i; j++)
359 if (crcu_array[j] == crcu_array[i])
360 break;
361 if (j == i)
362 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
363 }
364
365 /* Wait for all callbacks to be invoked. */
366 for (i = 0; i < n; i++) {
367 if (checktiny &&
368 (crcu_array[i] == call_rcu))
369 continue;
370 for (j = 0; j < i; j++)
371 if (crcu_array[j] == crcu_array[i])
372 break;
373 if (j == i)
374 wait_for_completion(&rs_array[i].completion);
375 destroy_rcu_head_on_stack(&rs_array[i].head);
376 }
377}
378EXPORT_SYMBOL_GPL(__wait_rcu_gp);
379
380#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
381void init_rcu_head(struct rcu_head *head)
382{
383 debug_object_init(head, &rcuhead_debug_descr);
384}
385EXPORT_SYMBOL_GPL(init_rcu_head);
386
387void destroy_rcu_head(struct rcu_head *head)
388{
389 debug_object_free(head, &rcuhead_debug_descr);
390}
391EXPORT_SYMBOL_GPL(destroy_rcu_head);
392
393static bool rcuhead_is_static_object(void *addr)
394{
395 return true;
396}
397
398/**
399 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
400 * @head: pointer to rcu_head structure to be initialized
401 *
402 * This function informs debugobjects of a new rcu_head structure that
403 * has been allocated as an auto variable on the stack. This function
404 * is not required for rcu_head structures that are statically defined or
405 * that are dynamically allocated on the heap. This function has no
406 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
407 */
408void init_rcu_head_on_stack(struct rcu_head *head)
409{
410 debug_object_init_on_stack(head, &rcuhead_debug_descr);
411}
412EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
413
414/**
415 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
416 * @head: pointer to rcu_head structure to be initialized
417 *
418 * This function informs debugobjects that an on-stack rcu_head structure
419 * is about to go out of scope. As with init_rcu_head_on_stack(), this
420 * function is not required for rcu_head structures that are statically
421 * defined or that are dynamically allocated on the heap. Also as with
422 * init_rcu_head_on_stack(), this function has no effect for
423 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
424 */
425void destroy_rcu_head_on_stack(struct rcu_head *head)
426{
427 debug_object_free(head, &rcuhead_debug_descr);
428}
429EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
430
431struct debug_obj_descr rcuhead_debug_descr = {
432 .name = "rcu_head",
433 .is_static_object = rcuhead_is_static_object,
434};
435EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
436#endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
437
438#if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
439void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
440 unsigned long secs,
441 unsigned long c_old, unsigned long c)
442{
443 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
444}
445EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
446#else
447#define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
448 do { } while (0)
449#endif
450
451#if IS_ENABLED(CONFIG_RCU_TORTURE_TEST) || IS_MODULE(CONFIG_RCU_TORTURE_TEST)
452/* Get rcutorture access to sched_setaffinity(). */
453long rcutorture_sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
454{
455 int ret;
456
457 ret = sched_setaffinity(pid, in_mask);
458 WARN_ONCE(ret, "%s: sched_setaffinity() returned %d\n", __func__, ret);
459 return ret;
460}
461EXPORT_SYMBOL_GPL(rcutorture_sched_setaffinity);
462#endif
463
464#ifdef CONFIG_RCU_STALL_COMMON
465int rcu_cpu_stall_ftrace_dump __read_mostly;
466module_param(rcu_cpu_stall_ftrace_dump, int, 0644);
467int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
468EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress);
469module_param(rcu_cpu_stall_suppress, int, 0644);
470int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
471module_param(rcu_cpu_stall_timeout, int, 0644);
472#endif /* #ifdef CONFIG_RCU_STALL_COMMON */
473
474#ifdef CONFIG_TASKS_RCU
475
476/*
477 * Simple variant of RCU whose quiescent states are voluntary context
478 * switch, cond_resched_rcu_qs(), user-space execution, and idle.
479 * As such, grace periods can take one good long time. There are no
480 * read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
481 * because this implementation is intended to get the system into a safe
482 * state for some of the manipulations involved in tracing and the like.
483 * Finally, this implementation does not support high call_rcu_tasks()
484 * rates from multiple CPUs. If this is required, per-CPU callback lists
485 * will be needed.
486 */
487
488/* Global list of callbacks and associated lock. */
489static struct rcu_head *rcu_tasks_cbs_head;
490static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
491static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
492static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
493
494/* Track exiting tasks in order to allow them to be waited for. */
495DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
496
497/* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
498#define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
499static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
500module_param(rcu_task_stall_timeout, int, 0644);
501
502static struct task_struct *rcu_tasks_kthread_ptr;
503
504/**
505 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
506 * @rhp: structure to be used for queueing the RCU updates.
507 * @func: actual callback function to be invoked after the grace period
508 *
509 * The callback function will be invoked some time after a full grace
510 * period elapses, in other words after all currently executing RCU
511 * read-side critical sections have completed. call_rcu_tasks() assumes
512 * that the read-side critical sections end at a voluntary context
513 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
514 * or transition to usermode execution. As such, there are no read-side
515 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
516 * this primitive is intended to determine that all tasks have passed
517 * through a safe state, not so much for data-strcuture synchronization.
518 *
519 * See the description of call_rcu() for more detailed information on
520 * memory ordering guarantees.
521 */
522void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
523{
524 unsigned long flags;
525 bool needwake;
526
527 rhp->next = NULL;
528 rhp->func = func;
529 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
530 needwake = !rcu_tasks_cbs_head;
531 *rcu_tasks_cbs_tail = rhp;
532 rcu_tasks_cbs_tail = &rhp->next;
533 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
534 /* We can't create the thread unless interrupts are enabled. */
535 if (needwake && READ_ONCE(rcu_tasks_kthread_ptr))
536 wake_up(&rcu_tasks_cbs_wq);
537}
538EXPORT_SYMBOL_GPL(call_rcu_tasks);
539
540/**
541 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
542 *
543 * Control will return to the caller some time after a full rcu-tasks
544 * grace period has elapsed, in other words after all currently
545 * executing rcu-tasks read-side critical sections have elapsed. These
546 * read-side critical sections are delimited by calls to schedule(),
547 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
548 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
549 *
550 * This is a very specialized primitive, intended only for a few uses in
551 * tracing and other situations requiring manipulation of function
552 * preambles and profiling hooks. The synchronize_rcu_tasks() function
553 * is not (yet) intended for heavy use from multiple CPUs.
554 *
555 * Note that this guarantee implies further memory-ordering guarantees.
556 * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
557 * each CPU is guaranteed to have executed a full memory barrier since the
558 * end of its last RCU-tasks read-side critical section whose beginning
559 * preceded the call to synchronize_rcu_tasks(). In addition, each CPU
560 * having an RCU-tasks read-side critical section that extends beyond
561 * the return from synchronize_rcu_tasks() is guaranteed to have executed
562 * a full memory barrier after the beginning of synchronize_rcu_tasks()
563 * and before the beginning of that RCU-tasks read-side critical section.
564 * Note that these guarantees include CPUs that are offline, idle, or
565 * executing in user mode, as well as CPUs that are executing in the kernel.
566 *
567 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
568 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
569 * to have executed a full memory barrier during the execution of
570 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
571 * (but again only if the system has more than one CPU).
572 */
573void synchronize_rcu_tasks(void)
574{
575 /* Complain if the scheduler has not started. */
576 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
577 "synchronize_rcu_tasks called too soon");
578
579 /* Wait for the grace period. */
580 wait_rcu_gp(call_rcu_tasks);
581}
582EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
583
584/**
585 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
586 *
587 * Although the current implementation is guaranteed to wait, it is not
588 * obligated to, for example, if there are no pending callbacks.
589 */
590void rcu_barrier_tasks(void)
591{
592 /* There is only one callback queue, so this is easy. ;-) */
593 synchronize_rcu_tasks();
594}
595EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
596
597/* See if tasks are still holding out, complain if so. */
598static void check_holdout_task(struct task_struct *t,
599 bool needreport, bool *firstreport)
600{
601 int cpu;
602
603 if (!READ_ONCE(t->rcu_tasks_holdout) ||
604 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
605 !READ_ONCE(t->on_rq) ||
606 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
607 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
608 WRITE_ONCE(t->rcu_tasks_holdout, false);
609 list_del_init(&t->rcu_tasks_holdout_list);
610 put_task_struct(t);
611 return;
612 }
613 rcu_request_urgent_qs_task(t);
614 if (!needreport)
615 return;
616 if (*firstreport) {
617 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
618 *firstreport = false;
619 }
620 cpu = task_cpu(t);
621 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
622 t, ".I"[is_idle_task(t)],
623 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
624 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
625 t->rcu_tasks_idle_cpu, cpu);
626 sched_show_task(t);
627}
628
629/* RCU-tasks kthread that detects grace periods and invokes callbacks. */
630static int __noreturn rcu_tasks_kthread(void *arg)
631{
632 unsigned long flags;
633 struct task_struct *g, *t;
634 unsigned long lastreport;
635 struct rcu_head *list;
636 struct rcu_head *next;
637 LIST_HEAD(rcu_tasks_holdouts);
638 int fract;
639
640 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
641 housekeeping_affine(current, HK_FLAG_RCU);
642
643 /*
644 * Each pass through the following loop makes one check for
645 * newly arrived callbacks, and, if there are some, waits for
646 * one RCU-tasks grace period and then invokes the callbacks.
647 * This loop is terminated by the system going down. ;-)
648 */
649 for (;;) {
650
651 /* Pick up any new callbacks. */
652 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
653 list = rcu_tasks_cbs_head;
654 rcu_tasks_cbs_head = NULL;
655 rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
656 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
657
658 /* If there were none, wait a bit and start over. */
659 if (!list) {
660 wait_event_interruptible(rcu_tasks_cbs_wq,
661 rcu_tasks_cbs_head);
662 if (!rcu_tasks_cbs_head) {
663 WARN_ON(signal_pending(current));
664 schedule_timeout_interruptible(HZ/10);
665 }
666 continue;
667 }
668
669 /*
670 * Wait for all pre-existing t->on_rq and t->nvcsw
671 * transitions to complete. Invoking synchronize_rcu()
672 * suffices because all these transitions occur with
673 * interrupts disabled. Without this synchronize_rcu(),
674 * a read-side critical section that started before the
675 * grace period might be incorrectly seen as having started
676 * after the grace period.
677 *
678 * This synchronize_rcu() also dispenses with the
679 * need for a memory barrier on the first store to
680 * ->rcu_tasks_holdout, as it forces the store to happen
681 * after the beginning of the grace period.
682 */
683 synchronize_rcu();
684
685 /*
686 * There were callbacks, so we need to wait for an
687 * RCU-tasks grace period. Start off by scanning
688 * the task list for tasks that are not already
689 * voluntarily blocked. Mark these tasks and make
690 * a list of them in rcu_tasks_holdouts.
691 */
692 rcu_read_lock();
693 for_each_process_thread(g, t) {
694 if (t != current && READ_ONCE(t->on_rq) &&
695 !is_idle_task(t)) {
696 get_task_struct(t);
697 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
698 WRITE_ONCE(t->rcu_tasks_holdout, true);
699 list_add(&t->rcu_tasks_holdout_list,
700 &rcu_tasks_holdouts);
701 }
702 }
703 rcu_read_unlock();
704
705 /*
706 * Wait for tasks that are in the process of exiting.
707 * This does only part of the job, ensuring that all
708 * tasks that were previously exiting reach the point
709 * where they have disabled preemption, allowing the
710 * later synchronize_rcu() to finish the job.
711 */
712 synchronize_srcu(&tasks_rcu_exit_srcu);
713
714 /*
715 * Each pass through the following loop scans the list
716 * of holdout tasks, removing any that are no longer
717 * holdouts. When the list is empty, we are done.
718 */
719 lastreport = jiffies;
720
721 /* Start off with HZ/10 wait and slowly back off to 1 HZ wait*/
722 fract = 10;
723
724 for (;;) {
725 bool firstreport;
726 bool needreport;
727 int rtst;
728 struct task_struct *t1;
729
730 if (list_empty(&rcu_tasks_holdouts))
731 break;
732
733 /* Slowly back off waiting for holdouts */
734 schedule_timeout_interruptible(HZ/fract);
735
736 if (fract > 1)
737 fract--;
738
739 rtst = READ_ONCE(rcu_task_stall_timeout);
740 needreport = rtst > 0 &&
741 time_after(jiffies, lastreport + rtst);
742 if (needreport)
743 lastreport = jiffies;
744 firstreport = true;
745 WARN_ON(signal_pending(current));
746 list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
747 rcu_tasks_holdout_list) {
748 check_holdout_task(t, needreport, &firstreport);
749 cond_resched();
750 }
751 }
752
753 /*
754 * Because ->on_rq and ->nvcsw are not guaranteed
755 * to have a full memory barriers prior to them in the
756 * schedule() path, memory reordering on other CPUs could
757 * cause their RCU-tasks read-side critical sections to
758 * extend past the end of the grace period. However,
759 * because these ->nvcsw updates are carried out with
760 * interrupts disabled, we can use synchronize_rcu()
761 * to force the needed ordering on all such CPUs.
762 *
763 * This synchronize_rcu() also confines all
764 * ->rcu_tasks_holdout accesses to be within the grace
765 * period, avoiding the need for memory barriers for
766 * ->rcu_tasks_holdout accesses.
767 *
768 * In addition, this synchronize_rcu() waits for exiting
769 * tasks to complete their final preempt_disable() region
770 * of execution, cleaning up after the synchronize_srcu()
771 * above.
772 */
773 synchronize_rcu();
774
775 /* Invoke the callbacks. */
776 while (list) {
777 next = list->next;
778 local_bh_disable();
779 list->func(list);
780 local_bh_enable();
781 list = next;
782 cond_resched();
783 }
784 /* Paranoid sleep to keep this from entering a tight loop */
785 schedule_timeout_uninterruptible(HZ/10);
786 }
787}
788
789/* Spawn rcu_tasks_kthread() at core_initcall() time. */
790static int __init rcu_spawn_tasks_kthread(void)
791{
792 struct task_struct *t;
793
794 t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
795 if (WARN_ONCE(IS_ERR(t), "%s: Could not start Tasks-RCU grace-period kthread, OOM is now expected behavior\n", __func__))
796 return 0;
797 smp_mb(); /* Ensure others see full kthread. */
798 WRITE_ONCE(rcu_tasks_kthread_ptr, t);
799 return 0;
800}
801core_initcall(rcu_spawn_tasks_kthread);
802
803/* Do the srcu_read_lock() for the above synchronize_srcu(). */
804void exit_tasks_rcu_start(void)
805{
806 preempt_disable();
807 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
808 preempt_enable();
809}
810
811/* Do the srcu_read_unlock() for the above synchronize_srcu(). */
812void exit_tasks_rcu_finish(void)
813{
814 preempt_disable();
815 __srcu_read_unlock(&tasks_rcu_exit_srcu, current->rcu_tasks_idx);
816 preempt_enable();
817}
818
819#endif /* #ifdef CONFIG_TASKS_RCU */
820
821#ifndef CONFIG_TINY_RCU
822
823/*
824 * Print any non-default Tasks RCU settings.
825 */
826static void __init rcu_tasks_bootup_oddness(void)
827{
828#ifdef CONFIG_TASKS_RCU
829 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
830 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
831 else
832 pr_info("\tTasks RCU enabled.\n");
833#endif /* #ifdef CONFIG_TASKS_RCU */
834}
835
836#endif /* #ifndef CONFIG_TINY_RCU */
837
838#ifdef CONFIG_PROVE_RCU
839
840/*
841 * Early boot self test parameters.
842 */
843static bool rcu_self_test;
844module_param(rcu_self_test, bool, 0444);
845
846static int rcu_self_test_counter;
847
848static void test_callback(struct rcu_head *r)
849{
850 rcu_self_test_counter++;
851 pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
852}
853
854DEFINE_STATIC_SRCU(early_srcu);
855
856static void early_boot_test_call_rcu(void)
857{
858 static struct rcu_head head;
859 static struct rcu_head shead;
860
861 call_rcu(&head, test_callback);
862 if (IS_ENABLED(CONFIG_SRCU))
863 call_srcu(&early_srcu, &shead, test_callback);
864}
865
866void rcu_early_boot_tests(void)
867{
868 pr_info("Running RCU self tests\n");
869
870 if (rcu_self_test)
871 early_boot_test_call_rcu();
872 rcu_test_sync_prims();
873}
874
875static int rcu_verify_early_boot_tests(void)
876{
877 int ret = 0;
878 int early_boot_test_counter = 0;
879
880 if (rcu_self_test) {
881 early_boot_test_counter++;
882 rcu_barrier();
883 if (IS_ENABLED(CONFIG_SRCU)) {
884 early_boot_test_counter++;
885 srcu_barrier(&early_srcu);
886 }
887 }
888 if (rcu_self_test_counter != early_boot_test_counter) {
889 WARN_ON(1);
890 ret = -1;
891 }
892
893 return ret;
894}
895late_initcall(rcu_verify_early_boot_tests);
896#else
897void rcu_early_boot_tests(void) {}
898#endif /* CONFIG_PROVE_RCU */
899
900#ifndef CONFIG_TINY_RCU
901
902/*
903 * Print any significant non-default boot-time settings.
904 */
905void __init rcupdate_announce_bootup_oddness(void)
906{
907 if (rcu_normal)
908 pr_info("\tNo expedited grace period (rcu_normal).\n");
909 else if (rcu_normal_after_boot)
910 pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n");
911 else if (rcu_expedited)
912 pr_info("\tAll grace periods are expedited (rcu_expedited).\n");
913 if (rcu_cpu_stall_suppress)
914 pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n");
915 if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT)
916 pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout);
917 rcu_tasks_bootup_oddness();
918}
919
920#endif /* #ifndef CONFIG_TINY_RCU */