1 // SPDX-License-Identifier: GPL-2.0+
3 * Read-Copy Update mechanism for mutual exclusion
5 * Copyright IBM Corporation, 2001
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
10 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
11 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
13 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
14 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
16 * For detailed explanation of Read-Copy Update mechanism see -
17 * http://lse.sourceforge.net/locking/rcupdate.html
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>
44 #define CREATE_TRACE_POINTS
48 #ifdef MODULE_PARAM_PREFIX
49 #undef MODULE_PARAM_PREFIX
51 #define MODULE_PARAM_PREFIX "rcupdate."
53 #ifndef CONFIG_TINY_RCU
54 extern int rcu_expedited; /* from sysctl */
55 module_param(rcu_expedited, int, 0);
56 extern int rcu_normal; /* from sysctl */
57 module_param(rcu_normal, int, 0);
58 static int rcu_normal_after_boot;
59 module_param(rcu_normal_after_boot, int, 0);
60 #endif /* #ifndef CONFIG_TINY_RCU */
62 #ifdef CONFIG_DEBUG_LOCK_ALLOC
64 * rcu_read_lock_sched_held() - might we be in RCU-sched read-side critical section?
66 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an
67 * RCU-sched read-side critical section. In absence of
68 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
69 * critical section unless it can prove otherwise. Note that disabling
70 * of preemption (including disabling irqs) counts as an RCU-sched
71 * read-side critical section. This is useful for debug checks in functions
72 * that required that they be called within an RCU-sched read-side
75 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
76 * and while lockdep is disabled.
78 * Note that if the CPU is in the idle loop from an RCU point of
79 * view (ie: that we are in the section between rcu_idle_enter() and
80 * rcu_idle_exit()) then rcu_read_lock_held() returns false even if the CPU
81 * did an rcu_read_lock(). The reason for this is that RCU ignores CPUs
82 * that are in such a section, considering these as in extended quiescent
83 * state, so such a CPU is effectively never in an RCU read-side critical
84 * section regardless of what RCU primitives it invokes. This state of
85 * affairs is required --- we need to keep an RCU-free window in idle
86 * where the CPU may possibly enter into low power mode. This way we can
87 * notice an extended quiescent state to other CPUs that started a grace
88 * period. Otherwise we would delay any grace period as long as we run in
91 * Similarly, we avoid claiming an SRCU read lock held if the current
94 int rcu_read_lock_sched_held(void)
96 int lockdep_opinion = 0;
98 if (!debug_lockdep_rcu_enabled())
100 if (!rcu_is_watching())
102 if (!rcu_lockdep_current_cpu_online())
105 lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
106 return lockdep_opinion || !preemptible();
108 EXPORT_SYMBOL(rcu_read_lock_sched_held);
111 #ifndef CONFIG_TINY_RCU
114 * Should expedited grace-period primitives always fall back to their
115 * non-expedited counterparts? Intended for use within RCU. Note
116 * that if the user specifies both rcu_expedited and rcu_normal, then
117 * rcu_normal wins. (Except during the time period during boot from
118 * when the first task is spawned until the rcu_set_runtime_mode()
119 * core_initcall() is invoked, at which point everything is expedited.)
121 bool rcu_gp_is_normal(void)
123 return READ_ONCE(rcu_normal) &&
124 rcu_scheduler_active != RCU_SCHEDULER_INIT;
126 EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
128 static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1);
131 * Should normal grace-period primitives be expedited? Intended for
132 * use within RCU. Note that this function takes the rcu_expedited
133 * sysfs/boot variable and rcu_scheduler_active into account as well
134 * as the rcu_expedite_gp() nesting. So looping on rcu_unexpedite_gp()
135 * until rcu_gp_is_expedited() returns false is a -really- bad idea.
137 bool rcu_gp_is_expedited(void)
139 return rcu_expedited || atomic_read(&rcu_expedited_nesting) ||
140 rcu_scheduler_active == RCU_SCHEDULER_INIT;
142 EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
145 * rcu_expedite_gp - Expedite future RCU grace periods
147 * After a call to this function, future calls to synchronize_rcu() and
148 * friends act as the corresponding synchronize_rcu_expedited() function
149 * had instead been called.
151 void rcu_expedite_gp(void)
153 atomic_inc(&rcu_expedited_nesting);
155 EXPORT_SYMBOL_GPL(rcu_expedite_gp);
158 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
160 * Undo a prior call to rcu_expedite_gp(). If all prior calls to
161 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
162 * and if the rcu_expedited sysfs/boot parameter is not set, then all
163 * subsequent calls to synchronize_rcu() and friends will return to
164 * their normal non-expedited behavior.
166 void rcu_unexpedite_gp(void)
168 atomic_dec(&rcu_expedited_nesting);
170 EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
173 * Inform RCU of the end of the in-kernel boot sequence.
175 void rcu_end_inkernel_boot(void)
178 if (rcu_normal_after_boot)
179 WRITE_ONCE(rcu_normal, 1);
182 #endif /* #ifndef CONFIG_TINY_RCU */
185 * Test each non-SRCU synchronous grace-period wait API. This is
186 * useful just after a change in mode for these primitives, and
189 void rcu_test_sync_prims(void)
191 if (!IS_ENABLED(CONFIG_PROVE_RCU))
194 synchronize_rcu_expedited();
197 #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU)
200 * Switch to run-time mode once RCU has fully initialized.
202 static int __init rcu_set_runtime_mode(void)
204 rcu_test_sync_prims();
205 rcu_scheduler_active = RCU_SCHEDULER_RUNNING;
206 rcu_test_sync_prims();
209 core_initcall(rcu_set_runtime_mode);
211 #endif /* #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) */
213 #ifdef CONFIG_DEBUG_LOCK_ALLOC
214 static struct lock_class_key rcu_lock_key;
215 struct lockdep_map rcu_lock_map =
216 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
217 EXPORT_SYMBOL_GPL(rcu_lock_map);
219 static struct lock_class_key rcu_bh_lock_key;
220 struct lockdep_map rcu_bh_lock_map =
221 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
222 EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
224 static struct lock_class_key rcu_sched_lock_key;
225 struct lockdep_map rcu_sched_lock_map =
226 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
227 EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
229 static struct lock_class_key rcu_callback_key;
230 struct lockdep_map rcu_callback_map =
231 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
232 EXPORT_SYMBOL_GPL(rcu_callback_map);
234 int notrace debug_lockdep_rcu_enabled(void)
236 return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks &&
237 current->lockdep_recursion == 0;
239 EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
240 NOKPROBE_SYMBOL(debug_lockdep_rcu_enabled);
243 * rcu_read_lock_held() - might we be in RCU read-side critical section?
245 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
246 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC,
247 * this assumes we are in an RCU read-side critical section unless it can
248 * prove otherwise. This is useful for debug checks in functions that
249 * require that they be called within an RCU read-side critical section.
251 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
252 * and while lockdep is disabled.
254 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
255 * occur in the same context, for example, it is illegal to invoke
256 * rcu_read_unlock() in process context if the matching rcu_read_lock()
257 * was invoked from within an irq handler.
259 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
260 * offline from an RCU perspective, so check for those as well.
262 int rcu_read_lock_held(void)
264 if (!debug_lockdep_rcu_enabled())
266 if (!rcu_is_watching())
268 if (!rcu_lockdep_current_cpu_online())
270 return lock_is_held(&rcu_lock_map);
272 EXPORT_SYMBOL_GPL(rcu_read_lock_held);
275 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
277 * Check for bottom half being disabled, which covers both the
278 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
279 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
280 * will show the situation. This is useful for debug checks in functions
281 * that require that they be called within an RCU read-side critical
284 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
286 * Note that rcu_read_lock_bh() is disallowed if the CPU is either idle or
287 * offline from an RCU perspective, so check for those as well.
289 int rcu_read_lock_bh_held(void)
291 if (!debug_lockdep_rcu_enabled())
293 if (!rcu_is_watching())
295 if (!rcu_lockdep_current_cpu_online())
297 return in_softirq() || irqs_disabled();
299 EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
301 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
304 * wakeme_after_rcu() - Callback function to awaken a task after grace period
305 * @head: Pointer to rcu_head member within rcu_synchronize structure
307 * Awaken the corresponding task now that a grace period has elapsed.
309 void wakeme_after_rcu(struct rcu_head *head)
311 struct rcu_synchronize *rcu;
313 rcu = container_of(head, struct rcu_synchronize, head);
314 complete(&rcu->completion);
316 EXPORT_SYMBOL_GPL(wakeme_after_rcu);
318 void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
319 struct rcu_synchronize *rs_array)
324 /* Initialize and register callbacks for each crcu_array element. */
325 for (i = 0; i < n; i++) {
327 (crcu_array[i] == call_rcu)) {
331 init_rcu_head_on_stack(&rs_array[i].head);
332 init_completion(&rs_array[i].completion);
333 for (j = 0; j < i; j++)
334 if (crcu_array[j] == crcu_array[i])
337 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
340 /* Wait for all callbacks to be invoked. */
341 for (i = 0; i < n; i++) {
343 (crcu_array[i] == call_rcu))
345 for (j = 0; j < i; j++)
346 if (crcu_array[j] == crcu_array[i])
349 wait_for_completion(&rs_array[i].completion);
350 destroy_rcu_head_on_stack(&rs_array[i].head);
353 EXPORT_SYMBOL_GPL(__wait_rcu_gp);
355 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
356 void init_rcu_head(struct rcu_head *head)
358 debug_object_init(head, &rcuhead_debug_descr);
360 EXPORT_SYMBOL_GPL(init_rcu_head);
362 void destroy_rcu_head(struct rcu_head *head)
364 debug_object_free(head, &rcuhead_debug_descr);
366 EXPORT_SYMBOL_GPL(destroy_rcu_head);
368 static bool rcuhead_is_static_object(void *addr)
374 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
375 * @head: pointer to rcu_head structure to be initialized
377 * This function informs debugobjects of a new rcu_head structure that
378 * has been allocated as an auto variable on the stack. This function
379 * is not required for rcu_head structures that are statically defined or
380 * that are dynamically allocated on the heap. This function has no
381 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
383 void init_rcu_head_on_stack(struct rcu_head *head)
385 debug_object_init_on_stack(head, &rcuhead_debug_descr);
387 EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
390 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
391 * @head: pointer to rcu_head structure to be initialized
393 * This function informs debugobjects that an on-stack rcu_head structure
394 * is about to go out of scope. As with init_rcu_head_on_stack(), this
395 * function is not required for rcu_head structures that are statically
396 * defined or that are dynamically allocated on the heap. Also as with
397 * init_rcu_head_on_stack(), this function has no effect for
398 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
400 void destroy_rcu_head_on_stack(struct rcu_head *head)
402 debug_object_free(head, &rcuhead_debug_descr);
404 EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
406 struct debug_obj_descr rcuhead_debug_descr = {
408 .is_static_object = rcuhead_is_static_object,
410 EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
411 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
413 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
414 void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
416 unsigned long c_old, unsigned long c)
418 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
420 EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
422 #define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
426 #ifdef CONFIG_RCU_STALL_COMMON
427 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
428 EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress);
429 module_param(rcu_cpu_stall_suppress, int, 0644);
430 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
431 module_param(rcu_cpu_stall_timeout, int, 0644);
432 #endif /* #ifdef CONFIG_RCU_STALL_COMMON */
434 #ifdef CONFIG_TASKS_RCU
437 * Simple variant of RCU whose quiescent states are voluntary context
438 * switch, cond_resched_rcu_qs(), user-space execution, and idle.
439 * As such, grace periods can take one good long time. There are no
440 * read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
441 * because this implementation is intended to get the system into a safe
442 * state for some of the manipulations involved in tracing and the like.
443 * Finally, this implementation does not support high call_rcu_tasks()
444 * rates from multiple CPUs. If this is required, per-CPU callback lists
448 /* Global list of callbacks and associated lock. */
449 static struct rcu_head *rcu_tasks_cbs_head;
450 static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
451 static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
452 static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
454 /* Track exiting tasks in order to allow them to be waited for. */
455 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
457 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
458 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
459 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
460 module_param(rcu_task_stall_timeout, int, 0644);
462 static struct task_struct *rcu_tasks_kthread_ptr;
465 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
466 * @rhp: structure to be used for queueing the RCU updates.
467 * @func: actual callback function to be invoked after the grace period
469 * The callback function will be invoked some time after a full grace
470 * period elapses, in other words after all currently executing RCU
471 * read-side critical sections have completed. call_rcu_tasks() assumes
472 * that the read-side critical sections end at a voluntary context
473 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
474 * or transition to usermode execution. As such, there are no read-side
475 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
476 * this primitive is intended to determine that all tasks have passed
477 * through a safe state, not so much for data-strcuture synchronization.
479 * See the description of call_rcu() for more detailed information on
480 * memory ordering guarantees.
482 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
489 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
490 needwake = !rcu_tasks_cbs_head;
491 *rcu_tasks_cbs_tail = rhp;
492 rcu_tasks_cbs_tail = &rhp->next;
493 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
494 /* We can't create the thread unless interrupts are enabled. */
495 if (needwake && READ_ONCE(rcu_tasks_kthread_ptr))
496 wake_up(&rcu_tasks_cbs_wq);
498 EXPORT_SYMBOL_GPL(call_rcu_tasks);
501 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
503 * Control will return to the caller some time after a full rcu-tasks
504 * grace period has elapsed, in other words after all currently
505 * executing rcu-tasks read-side critical sections have elapsed. These
506 * read-side critical sections are delimited by calls to schedule(),
507 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
508 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
510 * This is a very specialized primitive, intended only for a few uses in
511 * tracing and other situations requiring manipulation of function
512 * preambles and profiling hooks. The synchronize_rcu_tasks() function
513 * is not (yet) intended for heavy use from multiple CPUs.
515 * Note that this guarantee implies further memory-ordering guarantees.
516 * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
517 * each CPU is guaranteed to have executed a full memory barrier since the
518 * end of its last RCU-tasks read-side critical section whose beginning
519 * preceded the call to synchronize_rcu_tasks(). In addition, each CPU
520 * having an RCU-tasks read-side critical section that extends beyond
521 * the return from synchronize_rcu_tasks() is guaranteed to have executed
522 * a full memory barrier after the beginning of synchronize_rcu_tasks()
523 * and before the beginning of that RCU-tasks read-side critical section.
524 * Note that these guarantees include CPUs that are offline, idle, or
525 * executing in user mode, as well as CPUs that are executing in the kernel.
527 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
528 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
529 * to have executed a full memory barrier during the execution of
530 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
531 * (but again only if the system has more than one CPU).
533 void synchronize_rcu_tasks(void)
535 /* Complain if the scheduler has not started. */
536 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
537 "synchronize_rcu_tasks called too soon");
539 /* Wait for the grace period. */
540 wait_rcu_gp(call_rcu_tasks);
542 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
545 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
547 * Although the current implementation is guaranteed to wait, it is not
548 * obligated to, for example, if there are no pending callbacks.
550 void rcu_barrier_tasks(void)
552 /* There is only one callback queue, so this is easy. ;-) */
553 synchronize_rcu_tasks();
555 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
557 /* See if tasks are still holding out, complain if so. */
558 static void check_holdout_task(struct task_struct *t,
559 bool needreport, bool *firstreport)
563 if (!READ_ONCE(t->rcu_tasks_holdout) ||
564 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
565 !READ_ONCE(t->on_rq) ||
566 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
567 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
568 WRITE_ONCE(t->rcu_tasks_holdout, false);
569 list_del_init(&t->rcu_tasks_holdout_list);
573 rcu_request_urgent_qs_task(t);
577 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
578 *firstreport = false;
581 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
582 t, ".I"[is_idle_task(t)],
583 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
584 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
585 t->rcu_tasks_idle_cpu, cpu);
589 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
590 static int __noreturn rcu_tasks_kthread(void *arg)
593 struct task_struct *g, *t;
594 unsigned long lastreport;
595 struct rcu_head *list;
596 struct rcu_head *next;
597 LIST_HEAD(rcu_tasks_holdouts);
600 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
601 housekeeping_affine(current, HK_FLAG_RCU);
604 * Each pass through the following loop makes one check for
605 * newly arrived callbacks, and, if there are some, waits for
606 * one RCU-tasks grace period and then invokes the callbacks.
607 * This loop is terminated by the system going down. ;-)
611 /* Pick up any new callbacks. */
612 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
613 list = rcu_tasks_cbs_head;
614 rcu_tasks_cbs_head = NULL;
615 rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
616 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
618 /* If there were none, wait a bit and start over. */
620 wait_event_interruptible(rcu_tasks_cbs_wq,
622 if (!rcu_tasks_cbs_head) {
623 WARN_ON(signal_pending(current));
624 schedule_timeout_interruptible(HZ/10);
630 * Wait for all pre-existing t->on_rq and t->nvcsw
631 * transitions to complete. Invoking synchronize_rcu()
632 * suffices because all these transitions occur with
633 * interrupts disabled. Without this synchronize_rcu(),
634 * a read-side critical section that started before the
635 * grace period might be incorrectly seen as having started
636 * after the grace period.
638 * This synchronize_rcu() also dispenses with the
639 * need for a memory barrier on the first store to
640 * ->rcu_tasks_holdout, as it forces the store to happen
641 * after the beginning of the grace period.
646 * There were callbacks, so we need to wait for an
647 * RCU-tasks grace period. Start off by scanning
648 * the task list for tasks that are not already
649 * voluntarily blocked. Mark these tasks and make
650 * a list of them in rcu_tasks_holdouts.
653 for_each_process_thread(g, t) {
654 if (t != current && READ_ONCE(t->on_rq) &&
657 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
658 WRITE_ONCE(t->rcu_tasks_holdout, true);
659 list_add(&t->rcu_tasks_holdout_list,
660 &rcu_tasks_holdouts);
666 * Wait for tasks that are in the process of exiting.
667 * This does only part of the job, ensuring that all
668 * tasks that were previously exiting reach the point
669 * where they have disabled preemption, allowing the
670 * later synchronize_rcu() to finish the job.
672 synchronize_srcu(&tasks_rcu_exit_srcu);
675 * Each pass through the following loop scans the list
676 * of holdout tasks, removing any that are no longer
677 * holdouts. When the list is empty, we are done.
679 lastreport = jiffies;
681 /* Start off with HZ/10 wait and slowly back off to 1 HZ wait*/
688 struct task_struct *t1;
690 if (list_empty(&rcu_tasks_holdouts))
693 /* Slowly back off waiting for holdouts */
694 schedule_timeout_interruptible(HZ/fract);
699 rtst = READ_ONCE(rcu_task_stall_timeout);
700 needreport = rtst > 0 &&
701 time_after(jiffies, lastreport + rtst);
703 lastreport = jiffies;
705 WARN_ON(signal_pending(current));
706 list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
707 rcu_tasks_holdout_list) {
708 check_holdout_task(t, needreport, &firstreport);
714 * Because ->on_rq and ->nvcsw are not guaranteed
715 * to have a full memory barriers prior to them in the
716 * schedule() path, memory reordering on other CPUs could
717 * cause their RCU-tasks read-side critical sections to
718 * extend past the end of the grace period. However,
719 * because these ->nvcsw updates are carried out with
720 * interrupts disabled, we can use synchronize_rcu()
721 * to force the needed ordering on all such CPUs.
723 * This synchronize_rcu() also confines all
724 * ->rcu_tasks_holdout accesses to be within the grace
725 * period, avoiding the need for memory barriers for
726 * ->rcu_tasks_holdout accesses.
728 * In addition, this synchronize_rcu() waits for exiting
729 * tasks to complete their final preempt_disable() region
730 * of execution, cleaning up after the synchronize_srcu()
735 /* Invoke the callbacks. */
744 /* Paranoid sleep to keep this from entering a tight loop */
745 schedule_timeout_uninterruptible(HZ/10);
749 /* Spawn rcu_tasks_kthread() at core_initcall() time. */
750 static int __init rcu_spawn_tasks_kthread(void)
752 struct task_struct *t;
754 t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
755 if (WARN_ONCE(IS_ERR(t), "%s: Could not start Tasks-RCU grace-period kthread, OOM is now expected behavior\n", __func__))
757 smp_mb(); /* Ensure others see full kthread. */
758 WRITE_ONCE(rcu_tasks_kthread_ptr, t);
761 core_initcall(rcu_spawn_tasks_kthread);
763 /* Do the srcu_read_lock() for the above synchronize_srcu(). */
764 void exit_tasks_rcu_start(void)
767 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
771 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */
772 void exit_tasks_rcu_finish(void)
775 __srcu_read_unlock(&tasks_rcu_exit_srcu, current->rcu_tasks_idx);
779 #endif /* #ifdef CONFIG_TASKS_RCU */
781 #ifndef CONFIG_TINY_RCU
784 * Print any non-default Tasks RCU settings.
786 static void __init rcu_tasks_bootup_oddness(void)
788 #ifdef CONFIG_TASKS_RCU
789 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
790 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
792 pr_info("\tTasks RCU enabled.\n");
793 #endif /* #ifdef CONFIG_TASKS_RCU */
796 #endif /* #ifndef CONFIG_TINY_RCU */
798 #ifdef CONFIG_PROVE_RCU
801 * Early boot self test parameters.
803 static bool rcu_self_test;
804 module_param(rcu_self_test, bool, 0444);
806 static int rcu_self_test_counter;
808 static void test_callback(struct rcu_head *r)
810 rcu_self_test_counter++;
811 pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
814 DEFINE_STATIC_SRCU(early_srcu);
816 static void early_boot_test_call_rcu(void)
818 static struct rcu_head head;
819 static struct rcu_head shead;
821 call_rcu(&head, test_callback);
822 if (IS_ENABLED(CONFIG_SRCU))
823 call_srcu(&early_srcu, &shead, test_callback);
826 void rcu_early_boot_tests(void)
828 pr_info("Running RCU self tests\n");
831 early_boot_test_call_rcu();
832 rcu_test_sync_prims();
835 static int rcu_verify_early_boot_tests(void)
838 int early_boot_test_counter = 0;
841 early_boot_test_counter++;
843 if (IS_ENABLED(CONFIG_SRCU)) {
844 early_boot_test_counter++;
845 srcu_barrier(&early_srcu);
848 if (rcu_self_test_counter != early_boot_test_counter) {
855 late_initcall(rcu_verify_early_boot_tests);
857 void rcu_early_boot_tests(void) {}
858 #endif /* CONFIG_PROVE_RCU */
860 #ifndef CONFIG_TINY_RCU
863 * Print any significant non-default boot-time settings.
865 void __init rcupdate_announce_bootup_oddness(void)
868 pr_info("\tNo expedited grace period (rcu_normal).\n");
869 else if (rcu_normal_after_boot)
870 pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n");
871 else if (rcu_expedited)
872 pr_info("\tAll grace periods are expedited (rcu_expedited).\n");
873 if (rcu_cpu_stall_suppress)
874 pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n");
875 if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT)
876 pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout);
877 rcu_tasks_bootup_oddness();
880 #endif /* #ifndef CONFIG_TINY_RCU */