2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
31 #include "../time/tick-internal.h"
33 #ifdef CONFIG_RCU_BOOST
35 #include "../locking/rtmutex_common.h"
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
41 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
42 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
43 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
44 DEFINE_PER_CPU(char, rcu_cpu_has_work);
46 #endif /* #ifdef CONFIG_RCU_BOOST */
48 #ifdef CONFIG_RCU_NOCB_CPU
49 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
50 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
51 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
52 static char __initdata nocb_buf[NR_CPUS * 5];
53 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
56 * Check the RCU kernel configuration parameters and print informative
57 * messages about anything out of the ordinary. If you like #ifdef, you
58 * will love this function.
60 static void __init rcu_bootup_announce_oddness(void)
62 #ifdef CONFIG_RCU_TRACE
63 pr_info("\tRCU debugfs-based tracing is enabled.\n");
65 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
66 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
69 #ifdef CONFIG_RCU_FANOUT_EXACT
70 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
72 #ifdef CONFIG_RCU_FAST_NO_HZ
73 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
75 #ifdef CONFIG_PROVE_RCU
76 pr_info("\tRCU lockdep checking is enabled.\n");
78 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
79 pr_info("\tRCU torture testing starts during boot.\n");
81 #if defined(CONFIG_RCU_CPU_STALL_INFO)
82 pr_info("\tAdditional per-CPU info printed with stalls.\n");
84 #if NUM_RCU_LVL_4 != 0
85 pr_info("\tFour-level hierarchy is enabled.\n");
87 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
88 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
89 if (nr_cpu_ids != NR_CPUS)
90 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
91 #ifdef CONFIG_RCU_BOOST
92 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
96 #ifdef CONFIG_PREEMPT_RCU
98 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
99 static struct rcu_state *rcu_state_p = &rcu_preempt_state;
101 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
102 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
106 * Tell them what RCU they are running.
108 static void __init rcu_bootup_announce(void)
110 pr_info("Preemptible hierarchical RCU implementation.\n");
111 rcu_bootup_announce_oddness();
115 * Record a preemptible-RCU quiescent state for the specified CPU. Note
116 * that this just means that the task currently running on the CPU is
117 * not in a quiescent state. There might be any number of tasks blocked
118 * while in an RCU read-side critical section.
120 * As with the other rcu_*_qs() functions, callers to this function
121 * must disable preemption.
123 static void rcu_preempt_qs(void)
125 if (!__this_cpu_read(rcu_preempt_data.passed_quiesce)) {
126 trace_rcu_grace_period(TPS("rcu_preempt"),
127 __this_cpu_read(rcu_preempt_data.gpnum),
129 __this_cpu_write(rcu_preempt_data.passed_quiesce, 1);
130 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
131 current->rcu_read_unlock_special.b.need_qs = false;
136 * We have entered the scheduler, and the current task might soon be
137 * context-switched away from. If this task is in an RCU read-side
138 * critical section, we will no longer be able to rely on the CPU to
139 * record that fact, so we enqueue the task on the blkd_tasks list.
140 * The task will dequeue itself when it exits the outermost enclosing
141 * RCU read-side critical section. Therefore, the current grace period
142 * cannot be permitted to complete until the blkd_tasks list entries
143 * predating the current grace period drain, in other words, until
144 * rnp->gp_tasks becomes NULL.
146 * Caller must disable preemption.
148 static void rcu_preempt_note_context_switch(void)
150 struct task_struct *t = current;
152 struct rcu_data *rdp;
153 struct rcu_node *rnp;
155 if (t->rcu_read_lock_nesting > 0 &&
156 !t->rcu_read_unlock_special.b.blocked) {
158 /* Possibly blocking in an RCU read-side critical section. */
159 rdp = this_cpu_ptr(rcu_preempt_state.rda);
161 raw_spin_lock_irqsave(&rnp->lock, flags);
162 smp_mb__after_unlock_lock();
163 t->rcu_read_unlock_special.b.blocked = true;
164 t->rcu_blocked_node = rnp;
167 * If this CPU has already checked in, then this task
168 * will hold up the next grace period rather than the
169 * current grace period. Queue the task accordingly.
170 * If the task is queued for the current grace period
171 * (i.e., this CPU has not yet passed through a quiescent
172 * state for the current grace period), then as long
173 * as that task remains queued, the current grace period
174 * cannot end. Note that there is some uncertainty as
175 * to exactly when the current grace period started.
176 * We take a conservative approach, which can result
177 * in unnecessarily waiting on tasks that started very
178 * slightly after the current grace period began. C'est
181 * But first, note that the current CPU must still be
184 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
185 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
186 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
187 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
188 rnp->gp_tasks = &t->rcu_node_entry;
189 #ifdef CONFIG_RCU_BOOST
190 if (rnp->boost_tasks != NULL)
191 rnp->boost_tasks = rnp->gp_tasks;
192 #endif /* #ifdef CONFIG_RCU_BOOST */
194 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
195 if (rnp->qsmask & rdp->grpmask)
196 rnp->gp_tasks = &t->rcu_node_entry;
198 trace_rcu_preempt_task(rdp->rsp->name,
200 (rnp->qsmask & rdp->grpmask)
203 raw_spin_unlock_irqrestore(&rnp->lock, flags);
204 } else if (t->rcu_read_lock_nesting < 0 &&
205 t->rcu_read_unlock_special.s) {
208 * Complete exit from RCU read-side critical section on
209 * behalf of preempted instance of __rcu_read_unlock().
211 rcu_read_unlock_special(t);
215 * Either we were not in an RCU read-side critical section to
216 * begin with, or we have now recorded that critical section
217 * globally. Either way, we can now note a quiescent state
218 * for this CPU. Again, if we were in an RCU read-side critical
219 * section, and if that critical section was blocking the current
220 * grace period, then the fact that the task has been enqueued
221 * means that we continue to block the current grace period.
227 * Check for preempted RCU readers blocking the current grace period
228 * for the specified rcu_node structure. If the caller needs a reliable
229 * answer, it must hold the rcu_node's ->lock.
231 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
233 return rnp->gp_tasks != NULL;
237 * Record a quiescent state for all tasks that were previously queued
238 * on the specified rcu_node structure and that were blocking the current
239 * RCU grace period. The caller must hold the specified rnp->lock with
240 * irqs disabled, and this lock is released upon return, but irqs remain
243 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
244 __releases(rnp->lock)
247 struct rcu_node *rnp_p;
249 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
250 raw_spin_unlock_irqrestore(&rnp->lock, flags);
251 return; /* Still need more quiescent states! */
257 * Either there is only one rcu_node in the tree,
258 * or tasks were kicked up to root rcu_node due to
259 * CPUs going offline.
261 rcu_report_qs_rsp(&rcu_preempt_state, flags);
265 /* Report up the rest of the hierarchy. */
267 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
268 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
269 smp_mb__after_unlock_lock();
270 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
274 * Advance a ->blkd_tasks-list pointer to the next entry, instead
275 * returning NULL if at the end of the list.
277 static struct list_head *rcu_next_node_entry(struct task_struct *t,
278 struct rcu_node *rnp)
280 struct list_head *np;
282 np = t->rcu_node_entry.next;
283 if (np == &rnp->blkd_tasks)
289 * Return true if the specified rcu_node structure has tasks that were
290 * preempted within an RCU read-side critical section.
292 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
294 return !list_empty(&rnp->blkd_tasks);
298 * Handle special cases during rcu_read_unlock(), such as needing to
299 * notify RCU core processing or task having blocked during the RCU
300 * read-side critical section.
302 void rcu_read_unlock_special(struct task_struct *t)
309 struct list_head *np;
310 #ifdef CONFIG_RCU_BOOST
311 bool drop_boost_mutex = false;
312 #endif /* #ifdef CONFIG_RCU_BOOST */
313 struct rcu_node *rnp;
314 union rcu_special special;
316 /* NMI handlers cannot block and cannot safely manipulate state. */
320 local_irq_save(flags);
323 * If RCU core is waiting for this CPU to exit critical section,
324 * let it know that we have done so. Because irqs are disabled,
325 * t->rcu_read_unlock_special cannot change.
327 special = t->rcu_read_unlock_special;
328 if (special.b.need_qs) {
330 t->rcu_read_unlock_special.b.need_qs = false;
331 if (!t->rcu_read_unlock_special.s) {
332 local_irq_restore(flags);
337 /* Hardware IRQ handlers cannot block, complain if they get here. */
338 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
339 local_irq_restore(flags);
343 /* Clean up if blocked during RCU read-side critical section. */
344 if (special.b.blocked) {
345 t->rcu_read_unlock_special.b.blocked = false;
348 * Remove this task from the list it blocked on. The
349 * task can migrate while we acquire the lock, but at
350 * most one time. So at most two passes through loop.
353 rnp = t->rcu_blocked_node;
354 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
355 smp_mb__after_unlock_lock();
356 if (rnp == t->rcu_blocked_node)
358 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
360 empty = !rcu_preempt_has_tasks(rnp);
361 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
362 empty_exp = !rcu_preempted_readers_exp(rnp);
363 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
364 np = rcu_next_node_entry(t, rnp);
365 list_del_init(&t->rcu_node_entry);
366 t->rcu_blocked_node = NULL;
367 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
369 if (&t->rcu_node_entry == rnp->gp_tasks)
371 if (&t->rcu_node_entry == rnp->exp_tasks)
373 #ifdef CONFIG_RCU_BOOST
374 if (&t->rcu_node_entry == rnp->boost_tasks)
375 rnp->boost_tasks = np;
376 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
377 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
378 #endif /* #ifdef CONFIG_RCU_BOOST */
381 * If this was the last task on the list, go see if we
382 * need to propagate ->qsmaskinit bit clearing up the
385 if (!empty && !rcu_preempt_has_tasks(rnp))
386 rcu_cleanup_dead_rnp(rnp);
389 * If this was the last task on the current list, and if
390 * we aren't waiting on any CPUs, report the quiescent state.
391 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
392 * so we must take a snapshot of the expedited state.
394 empty_exp_now = !rcu_preempted_readers_exp(rnp);
395 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
396 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
403 rcu_report_unblock_qs_rnp(rnp, flags);
405 raw_spin_unlock_irqrestore(&rnp->lock, flags);
408 #ifdef CONFIG_RCU_BOOST
409 /* Unboost if we were boosted. */
410 if (drop_boost_mutex)
411 rt_mutex_unlock(&rnp->boost_mtx);
412 #endif /* #ifdef CONFIG_RCU_BOOST */
415 * If this was the last task on the expedited lists,
416 * then we need to report up the rcu_node hierarchy.
418 if (!empty_exp && empty_exp_now)
419 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
421 local_irq_restore(flags);
426 * Dump detailed information for all tasks blocking the current RCU
427 * grace period on the specified rcu_node structure.
429 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
432 struct task_struct *t;
434 raw_spin_lock_irqsave(&rnp->lock, flags);
435 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
436 raw_spin_unlock_irqrestore(&rnp->lock, flags);
439 t = list_entry(rnp->gp_tasks,
440 struct task_struct, rcu_node_entry);
441 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
443 raw_spin_unlock_irqrestore(&rnp->lock, flags);
447 * Dump detailed information for all tasks blocking the current RCU
450 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
452 struct rcu_node *rnp = rcu_get_root(rsp);
454 rcu_print_detail_task_stall_rnp(rnp);
455 rcu_for_each_leaf_node(rsp, rnp)
456 rcu_print_detail_task_stall_rnp(rnp);
459 #ifdef CONFIG_RCU_CPU_STALL_INFO
461 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
463 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
464 rnp->level, rnp->grplo, rnp->grphi);
467 static void rcu_print_task_stall_end(void)
472 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
474 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
478 static void rcu_print_task_stall_end(void)
482 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
485 * Scan the current list of tasks blocked within RCU read-side critical
486 * sections, printing out the tid of each.
488 static int rcu_print_task_stall(struct rcu_node *rnp)
490 struct task_struct *t;
493 if (!rcu_preempt_blocked_readers_cgp(rnp))
495 rcu_print_task_stall_begin(rnp);
496 t = list_entry(rnp->gp_tasks,
497 struct task_struct, rcu_node_entry);
498 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
499 pr_cont(" P%d", t->pid);
502 rcu_print_task_stall_end();
507 * Check that the list of blocked tasks for the newly completed grace
508 * period is in fact empty. It is a serious bug to complete a grace
509 * period that still has RCU readers blocked! This function must be
510 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
511 * must be held by the caller.
513 * Also, if there are blocked tasks on the list, they automatically
514 * block the newly created grace period, so set up ->gp_tasks accordingly.
516 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
518 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
519 if (rcu_preempt_has_tasks(rnp))
520 rnp->gp_tasks = rnp->blkd_tasks.next;
521 WARN_ON_ONCE(rnp->qsmask);
524 #ifdef CONFIG_HOTPLUG_CPU
526 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
529 * Check for a quiescent state from the current CPU. When a task blocks,
530 * the task is recorded in the corresponding CPU's rcu_node structure,
531 * which is checked elsewhere.
533 * Caller must disable hard irqs.
535 static void rcu_preempt_check_callbacks(void)
537 struct task_struct *t = current;
539 if (t->rcu_read_lock_nesting == 0) {
543 if (t->rcu_read_lock_nesting > 0 &&
544 __this_cpu_read(rcu_preempt_data.qs_pending) &&
545 !__this_cpu_read(rcu_preempt_data.passed_quiesce))
546 t->rcu_read_unlock_special.b.need_qs = true;
549 #ifdef CONFIG_RCU_BOOST
551 static void rcu_preempt_do_callbacks(void)
553 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
556 #endif /* #ifdef CONFIG_RCU_BOOST */
559 * Queue a preemptible-RCU callback for invocation after a grace period.
561 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
563 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
565 EXPORT_SYMBOL_GPL(call_rcu);
568 * synchronize_rcu - wait until a grace period has elapsed.
570 * Control will return to the caller some time after a full grace
571 * period has elapsed, in other words after all currently executing RCU
572 * read-side critical sections have completed. Note, however, that
573 * upon return from synchronize_rcu(), the caller might well be executing
574 * concurrently with new RCU read-side critical sections that began while
575 * synchronize_rcu() was waiting. RCU read-side critical sections are
576 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
578 * See the description of synchronize_sched() for more detailed information
579 * on memory ordering guarantees.
581 void synchronize_rcu(void)
583 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
584 !lock_is_held(&rcu_lock_map) &&
585 !lock_is_held(&rcu_sched_lock_map),
586 "Illegal synchronize_rcu() in RCU read-side critical section");
587 if (!rcu_scheduler_active)
590 synchronize_rcu_expedited();
592 wait_rcu_gp(call_rcu);
594 EXPORT_SYMBOL_GPL(synchronize_rcu);
596 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
597 static unsigned long sync_rcu_preempt_exp_count;
598 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
601 * Return non-zero if there are any tasks in RCU read-side critical
602 * sections blocking the current preemptible-RCU expedited grace period.
603 * If there is no preemptible-RCU expedited grace period currently in
604 * progress, returns zero unconditionally.
606 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
608 return rnp->exp_tasks != NULL;
612 * return non-zero if there is no RCU expedited grace period in progress
613 * for the specified rcu_node structure, in other words, if all CPUs and
614 * tasks covered by the specified rcu_node structure have done their bit
615 * for the current expedited grace period. Works only for preemptible
616 * RCU -- other RCU implementation use other means.
618 * Caller must hold sync_rcu_preempt_exp_mutex.
620 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
622 return !rcu_preempted_readers_exp(rnp) &&
623 ACCESS_ONCE(rnp->expmask) == 0;
627 * Report the exit from RCU read-side critical section for the last task
628 * that queued itself during or before the current expedited preemptible-RCU
629 * grace period. This event is reported either to the rcu_node structure on
630 * which the task was queued or to one of that rcu_node structure's ancestors,
631 * recursively up the tree. (Calm down, calm down, we do the recursion
634 * Most callers will set the "wake" flag, but the task initiating the
635 * expedited grace period need not wake itself.
637 * Caller must hold sync_rcu_preempt_exp_mutex.
639 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
645 raw_spin_lock_irqsave(&rnp->lock, flags);
646 smp_mb__after_unlock_lock();
648 if (!sync_rcu_preempt_exp_done(rnp)) {
649 raw_spin_unlock_irqrestore(&rnp->lock, flags);
652 if (rnp->parent == NULL) {
653 raw_spin_unlock_irqrestore(&rnp->lock, flags);
655 smp_mb(); /* EGP done before wake_up(). */
656 wake_up(&sync_rcu_preempt_exp_wq);
661 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
663 raw_spin_lock(&rnp->lock); /* irqs already disabled */
664 smp_mb__after_unlock_lock();
665 rnp->expmask &= ~mask;
670 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
671 * grace period for the specified rcu_node structure. If there are no such
672 * tasks, report it up the rcu_node hierarchy.
674 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
675 * CPU hotplug operations.
678 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
683 raw_spin_lock_irqsave(&rnp->lock, flags);
684 smp_mb__after_unlock_lock();
685 if (!rcu_preempt_has_tasks(rnp)) {
686 raw_spin_unlock_irqrestore(&rnp->lock, flags);
688 rnp->exp_tasks = rnp->blkd_tasks.next;
689 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
693 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
697 * synchronize_rcu_expedited - Brute-force RCU grace period
699 * Wait for an RCU-preempt grace period, but expedite it. The basic
700 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
701 * the ->blkd_tasks lists and wait for this list to drain. This consumes
702 * significant time on all CPUs and is unfriendly to real-time workloads,
703 * so is thus not recommended for any sort of common-case code.
704 * In fact, if you are using synchronize_rcu_expedited() in a loop,
705 * please restructure your code to batch your updates, and then Use a
706 * single synchronize_rcu() instead.
708 void synchronize_rcu_expedited(void)
711 struct rcu_node *rnp;
712 struct rcu_state *rsp = &rcu_preempt_state;
716 smp_mb(); /* Caller's modifications seen first by other CPUs. */
717 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
718 smp_mb(); /* Above access cannot bleed into critical section. */
721 * Block CPU-hotplug operations. This means that any CPU-hotplug
722 * operation that finds an rcu_node structure with tasks in the
723 * process of being boosted will know that all tasks blocking
724 * this expedited grace period will already be in the process of
725 * being boosted. This simplifies the process of moving tasks
726 * from leaf to root rcu_node structures.
728 if (!try_get_online_cpus()) {
729 /* CPU-hotplug operation in flight, fall back to normal GP. */
730 wait_rcu_gp(call_rcu);
735 * Acquire lock, falling back to synchronize_rcu() if too many
736 * lock-acquisition failures. Of course, if someone does the
737 * expedited grace period for us, just leave.
739 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
740 if (ULONG_CMP_LT(snap,
741 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
743 goto mb_ret; /* Others did our work for us. */
745 if (trycount++ < 10) {
746 udelay(trycount * num_online_cpus());
749 wait_rcu_gp(call_rcu);
753 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
755 goto unlock_mb_ret; /* Others did our work for us. */
758 /* force all RCU readers onto ->blkd_tasks lists. */
759 synchronize_sched_expedited();
761 /* Initialize ->expmask for all non-leaf rcu_node structures. */
762 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
763 raw_spin_lock_irqsave(&rnp->lock, flags);
764 smp_mb__after_unlock_lock();
765 rnp->expmask = rnp->qsmaskinit;
766 raw_spin_unlock_irqrestore(&rnp->lock, flags);
769 /* Snapshot current state of ->blkd_tasks lists. */
770 rcu_for_each_leaf_node(rsp, rnp)
771 sync_rcu_preempt_exp_init(rsp, rnp);
772 if (NUM_RCU_NODES > 1)
773 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
777 /* Wait for snapshotted ->blkd_tasks lists to drain. */
778 rnp = rcu_get_root(rsp);
779 wait_event(sync_rcu_preempt_exp_wq,
780 sync_rcu_preempt_exp_done(rnp));
782 /* Clean up and exit. */
783 smp_mb(); /* ensure expedited GP seen before counter increment. */
784 ACCESS_ONCE(sync_rcu_preempt_exp_count) =
785 sync_rcu_preempt_exp_count + 1;
787 mutex_unlock(&sync_rcu_preempt_exp_mutex);
789 smp_mb(); /* ensure subsequent action seen after grace period. */
791 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
794 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
796 * Note that this primitive does not necessarily wait for an RCU grace period
797 * to complete. For example, if there are no RCU callbacks queued anywhere
798 * in the system, then rcu_barrier() is within its rights to return
799 * immediately, without waiting for anything, much less an RCU grace period.
801 void rcu_barrier(void)
803 _rcu_barrier(&rcu_preempt_state);
805 EXPORT_SYMBOL_GPL(rcu_barrier);
808 * Initialize preemptible RCU's state structures.
810 static void __init __rcu_init_preempt(void)
812 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
816 * Check for a task exiting while in a preemptible-RCU read-side
817 * critical section, clean up if so. No need to issue warnings,
818 * as debug_check_no_locks_held() already does this if lockdep
823 struct task_struct *t = current;
825 if (likely(list_empty(¤t->rcu_node_entry)))
827 t->rcu_read_lock_nesting = 1;
829 t->rcu_read_unlock_special.b.blocked = true;
833 #else /* #ifdef CONFIG_PREEMPT_RCU */
835 static struct rcu_state *rcu_state_p = &rcu_sched_state;
838 * Tell them what RCU they are running.
840 static void __init rcu_bootup_announce(void)
842 pr_info("Hierarchical RCU implementation.\n");
843 rcu_bootup_announce_oddness();
847 * Because preemptible RCU does not exist, we never have to check for
848 * CPUs being in quiescent states.
850 static void rcu_preempt_note_context_switch(void)
855 * Because preemptible RCU does not exist, there are never any preempted
858 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
863 #ifdef CONFIG_HOTPLUG_CPU
866 * Because there is no preemptible RCU, there can be no readers blocked.
868 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
873 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
876 * Because preemptible RCU does not exist, we never have to check for
877 * tasks blocked within RCU read-side critical sections.
879 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
884 * Because preemptible RCU does not exist, we never have to check for
885 * tasks blocked within RCU read-side critical sections.
887 static int rcu_print_task_stall(struct rcu_node *rnp)
893 * Because there is no preemptible RCU, there can be no readers blocked,
894 * so there is no need to check for blocked tasks. So check only for
895 * bogus qsmask values.
897 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
899 WARN_ON_ONCE(rnp->qsmask);
903 * Because preemptible RCU does not exist, it never has any callbacks
906 static void rcu_preempt_check_callbacks(void)
911 * Wait for an rcu-preempt grace period, but make it happen quickly.
912 * But because preemptible RCU does not exist, map to rcu-sched.
914 void synchronize_rcu_expedited(void)
916 synchronize_sched_expedited();
918 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
921 * Because preemptible RCU does not exist, rcu_barrier() is just
922 * another name for rcu_barrier_sched().
924 void rcu_barrier(void)
928 EXPORT_SYMBOL_GPL(rcu_barrier);
931 * Because preemptible RCU does not exist, it need not be initialized.
933 static void __init __rcu_init_preempt(void)
938 * Because preemptible RCU does not exist, tasks cannot possibly exit
939 * while in preemptible RCU read-side critical sections.
945 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
947 #ifdef CONFIG_RCU_BOOST
949 #include "../locking/rtmutex_common.h"
951 #ifdef CONFIG_RCU_TRACE
953 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
955 if (!rcu_preempt_has_tasks(rnp))
956 rnp->n_balk_blkd_tasks++;
957 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
958 rnp->n_balk_exp_gp_tasks++;
959 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
960 rnp->n_balk_boost_tasks++;
961 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
962 rnp->n_balk_notblocked++;
963 else if (rnp->gp_tasks != NULL &&
964 ULONG_CMP_LT(jiffies, rnp->boost_time))
965 rnp->n_balk_notyet++;
970 #else /* #ifdef CONFIG_RCU_TRACE */
972 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
976 #endif /* #else #ifdef CONFIG_RCU_TRACE */
978 static void rcu_wake_cond(struct task_struct *t, int status)
981 * If the thread is yielding, only wake it when this
982 * is invoked from idle
984 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
989 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
990 * or ->boost_tasks, advancing the pointer to the next task in the
993 * Note that irqs must be enabled: boosting the task can block.
994 * Returns 1 if there are more tasks needing to be boosted.
996 static int rcu_boost(struct rcu_node *rnp)
999 struct task_struct *t;
1000 struct list_head *tb;
1002 if (ACCESS_ONCE(rnp->exp_tasks) == NULL &&
1003 ACCESS_ONCE(rnp->boost_tasks) == NULL)
1004 return 0; /* Nothing left to boost. */
1006 raw_spin_lock_irqsave(&rnp->lock, flags);
1007 smp_mb__after_unlock_lock();
1010 * Recheck under the lock: all tasks in need of boosting
1011 * might exit their RCU read-side critical sections on their own.
1013 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1014 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1019 * Preferentially boost tasks blocking expedited grace periods.
1020 * This cannot starve the normal grace periods because a second
1021 * expedited grace period must boost all blocked tasks, including
1022 * those blocking the pre-existing normal grace period.
1024 if (rnp->exp_tasks != NULL) {
1025 tb = rnp->exp_tasks;
1026 rnp->n_exp_boosts++;
1028 tb = rnp->boost_tasks;
1029 rnp->n_normal_boosts++;
1031 rnp->n_tasks_boosted++;
1034 * We boost task t by manufacturing an rt_mutex that appears to
1035 * be held by task t. We leave a pointer to that rt_mutex where
1036 * task t can find it, and task t will release the mutex when it
1037 * exits its outermost RCU read-side critical section. Then
1038 * simply acquiring this artificial rt_mutex will boost task
1039 * t's priority. (Thanks to tglx for suggesting this approach!)
1041 * Note that task t must acquire rnp->lock to remove itself from
1042 * the ->blkd_tasks list, which it will do from exit() if from
1043 * nowhere else. We therefore are guaranteed that task t will
1044 * stay around at least until we drop rnp->lock. Note that
1045 * rnp->lock also resolves races between our priority boosting
1046 * and task t's exiting its outermost RCU read-side critical
1049 t = container_of(tb, struct task_struct, rcu_node_entry);
1050 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1051 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1052 /* Lock only for side effect: boosts task t's priority. */
1053 rt_mutex_lock(&rnp->boost_mtx);
1054 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1056 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1057 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1061 * Priority-boosting kthread. One per leaf rcu_node and one for the
1064 static int rcu_boost_kthread(void *arg)
1066 struct rcu_node *rnp = (struct rcu_node *)arg;
1070 trace_rcu_utilization(TPS("Start boost kthread@init"));
1072 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1073 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1074 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1075 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1076 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1077 more2boost = rcu_boost(rnp);
1083 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1084 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1085 schedule_timeout_interruptible(2);
1086 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1091 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1096 * Check to see if it is time to start boosting RCU readers that are
1097 * blocking the current grace period, and, if so, tell the per-rcu_node
1098 * kthread to start boosting them. If there is an expedited grace
1099 * period in progress, it is always time to boost.
1101 * The caller must hold rnp->lock, which this function releases.
1102 * The ->boost_kthread_task is immortal, so we don't need to worry
1103 * about it going away.
1105 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1106 __releases(rnp->lock)
1108 struct task_struct *t;
1110 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1111 rnp->n_balk_exp_gp_tasks++;
1112 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1115 if (rnp->exp_tasks != NULL ||
1116 (rnp->gp_tasks != NULL &&
1117 rnp->boost_tasks == NULL &&
1119 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1120 if (rnp->exp_tasks == NULL)
1121 rnp->boost_tasks = rnp->gp_tasks;
1122 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1123 t = rnp->boost_kthread_task;
1125 rcu_wake_cond(t, rnp->boost_kthread_status);
1127 rcu_initiate_boost_trace(rnp);
1128 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1133 * Wake up the per-CPU kthread to invoke RCU callbacks.
1135 static void invoke_rcu_callbacks_kthread(void)
1137 unsigned long flags;
1139 local_irq_save(flags);
1140 __this_cpu_write(rcu_cpu_has_work, 1);
1141 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1142 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1143 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1144 __this_cpu_read(rcu_cpu_kthread_status));
1146 local_irq_restore(flags);
1150 * Is the current CPU running the RCU-callbacks kthread?
1151 * Caller must have preemption disabled.
1153 static bool rcu_is_callbacks_kthread(void)
1155 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1158 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1161 * Do priority-boost accounting for the start of a new grace period.
1163 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1165 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1169 * Create an RCU-boost kthread for the specified node if one does not
1170 * already exist. We only create this kthread for preemptible RCU.
1171 * Returns zero if all is well, a negated errno otherwise.
1173 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1174 struct rcu_node *rnp)
1176 int rnp_index = rnp - &rsp->node[0];
1177 unsigned long flags;
1178 struct sched_param sp;
1179 struct task_struct *t;
1181 if (&rcu_preempt_state != rsp)
1184 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1188 if (rnp->boost_kthread_task != NULL)
1190 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1191 "rcub/%d", rnp_index);
1194 raw_spin_lock_irqsave(&rnp->lock, flags);
1195 smp_mb__after_unlock_lock();
1196 rnp->boost_kthread_task = t;
1197 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1198 sp.sched_priority = kthread_prio;
1199 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1200 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1204 static void rcu_kthread_do_work(void)
1206 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1207 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1208 rcu_preempt_do_callbacks();
1211 static void rcu_cpu_kthread_setup(unsigned int cpu)
1213 struct sched_param sp;
1215 sp.sched_priority = kthread_prio;
1216 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1219 static void rcu_cpu_kthread_park(unsigned int cpu)
1221 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1224 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1226 return __this_cpu_read(rcu_cpu_has_work);
1230 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1231 * RCU softirq used in flavors and configurations of RCU that do not
1232 * support RCU priority boosting.
1234 static void rcu_cpu_kthread(unsigned int cpu)
1236 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1237 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1240 for (spincnt = 0; spincnt < 10; spincnt++) {
1241 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1243 *statusp = RCU_KTHREAD_RUNNING;
1244 this_cpu_inc(rcu_cpu_kthread_loops);
1245 local_irq_disable();
1250 rcu_kthread_do_work();
1253 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1254 *statusp = RCU_KTHREAD_WAITING;
1258 *statusp = RCU_KTHREAD_YIELDING;
1259 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1260 schedule_timeout_interruptible(2);
1261 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1262 *statusp = RCU_KTHREAD_WAITING;
1266 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1267 * served by the rcu_node in question. The CPU hotplug lock is still
1268 * held, so the value of rnp->qsmaskinit will be stable.
1270 * We don't include outgoingcpu in the affinity set, use -1 if there is
1271 * no outgoing CPU. If there are no CPUs left in the affinity set,
1272 * this function allows the kthread to execute on any CPU.
1274 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1276 struct task_struct *t = rnp->boost_kthread_task;
1277 unsigned long mask = rnp->qsmaskinit;
1283 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1285 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1286 if ((mask & 0x1) && cpu != outgoingcpu)
1287 cpumask_set_cpu(cpu, cm);
1288 if (cpumask_weight(cm) == 0)
1290 set_cpus_allowed_ptr(t, cm);
1291 free_cpumask_var(cm);
1294 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1295 .store = &rcu_cpu_kthread_task,
1296 .thread_should_run = rcu_cpu_kthread_should_run,
1297 .thread_fn = rcu_cpu_kthread,
1298 .thread_comm = "rcuc/%u",
1299 .setup = rcu_cpu_kthread_setup,
1300 .park = rcu_cpu_kthread_park,
1304 * Spawn boost kthreads -- called as soon as the scheduler is running.
1306 static void __init rcu_spawn_boost_kthreads(void)
1308 struct rcu_node *rnp;
1311 for_each_possible_cpu(cpu)
1312 per_cpu(rcu_cpu_has_work, cpu) = 0;
1313 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1314 rcu_for_each_leaf_node(rcu_state_p, rnp)
1315 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1318 static void rcu_prepare_kthreads(int cpu)
1320 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1321 struct rcu_node *rnp = rdp->mynode;
1323 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1324 if (rcu_scheduler_fully_active)
1325 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1328 #else /* #ifdef CONFIG_RCU_BOOST */
1330 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1331 __releases(rnp->lock)
1333 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1336 static void invoke_rcu_callbacks_kthread(void)
1341 static bool rcu_is_callbacks_kthread(void)
1346 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1350 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1354 static void __init rcu_spawn_boost_kthreads(void)
1358 static void rcu_prepare_kthreads(int cpu)
1362 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1364 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1367 * Check to see if any future RCU-related work will need to be done
1368 * by the current CPU, even if none need be done immediately, returning
1369 * 1 if so. This function is part of the RCU implementation; it is -not-
1370 * an exported member of the RCU API.
1372 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1373 * any flavor of RCU.
1375 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1376 int rcu_needs_cpu(unsigned long *delta_jiffies)
1378 *delta_jiffies = ULONG_MAX;
1379 return rcu_cpu_has_callbacks(NULL);
1381 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1384 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1387 static void rcu_cleanup_after_idle(void)
1392 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1395 static void rcu_prepare_for_idle(void)
1400 * Don't bother keeping a running count of the number of RCU callbacks
1401 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1403 static void rcu_idle_count_callbacks_posted(void)
1407 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1410 * This code is invoked when a CPU goes idle, at which point we want
1411 * to have the CPU do everything required for RCU so that it can enter
1412 * the energy-efficient dyntick-idle mode. This is handled by a
1413 * state machine implemented by rcu_prepare_for_idle() below.
1415 * The following three proprocessor symbols control this state machine:
1417 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1418 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1419 * is sized to be roughly one RCU grace period. Those energy-efficiency
1420 * benchmarkers who might otherwise be tempted to set this to a large
1421 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1422 * system. And if you are -that- concerned about energy efficiency,
1423 * just power the system down and be done with it!
1424 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1425 * permitted to sleep in dyntick-idle mode with only lazy RCU
1426 * callbacks pending. Setting this too high can OOM your system.
1428 * The values below work well in practice. If future workloads require
1429 * adjustment, they can be converted into kernel config parameters, though
1430 * making the state machine smarter might be a better option.
1432 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1433 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1435 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1436 module_param(rcu_idle_gp_delay, int, 0644);
1437 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1438 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1440 extern int tick_nohz_active;
1443 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1444 * only if it has been awhile since the last time we did so. Afterwards,
1445 * if there are any callbacks ready for immediate invocation, return true.
1447 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1449 bool cbs_ready = false;
1450 struct rcu_data *rdp;
1451 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1452 struct rcu_node *rnp;
1453 struct rcu_state *rsp;
1455 /* Exit early if we advanced recently. */
1456 if (jiffies == rdtp->last_advance_all)
1458 rdtp->last_advance_all = jiffies;
1460 for_each_rcu_flavor(rsp) {
1461 rdp = this_cpu_ptr(rsp->rda);
1465 * Don't bother checking unless a grace period has
1466 * completed since we last checked and there are
1467 * callbacks not yet ready to invoke.
1469 if ((rdp->completed != rnp->completed ||
1470 unlikely(ACCESS_ONCE(rdp->gpwrap))) &&
1471 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1472 note_gp_changes(rsp, rdp);
1474 if (cpu_has_callbacks_ready_to_invoke(rdp))
1481 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1482 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1483 * caller to set the timeout based on whether or not there are non-lazy
1486 * The caller must have disabled interrupts.
1488 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1489 int rcu_needs_cpu(unsigned long *dj)
1491 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1493 /* Snapshot to detect later posting of non-lazy callback. */
1494 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1496 /* If no callbacks, RCU doesn't need the CPU. */
1497 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1502 /* Attempt to advance callbacks. */
1503 if (rcu_try_advance_all_cbs()) {
1504 /* Some ready to invoke, so initiate later invocation. */
1508 rdtp->last_accelerate = jiffies;
1510 /* Request timer delay depending on laziness, and round. */
1511 if (!rdtp->all_lazy) {
1512 *dj = round_up(rcu_idle_gp_delay + jiffies,
1513 rcu_idle_gp_delay) - jiffies;
1515 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1519 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1522 * Prepare a CPU for idle from an RCU perspective. The first major task
1523 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1524 * The second major task is to check to see if a non-lazy callback has
1525 * arrived at a CPU that previously had only lazy callbacks. The third
1526 * major task is to accelerate (that is, assign grace-period numbers to)
1527 * any recently arrived callbacks.
1529 * The caller must have disabled interrupts.
1531 static void rcu_prepare_for_idle(void)
1533 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1535 struct rcu_data *rdp;
1536 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1537 struct rcu_node *rnp;
1538 struct rcu_state *rsp;
1541 /* Handle nohz enablement switches conservatively. */
1542 tne = ACCESS_ONCE(tick_nohz_active);
1543 if (tne != rdtp->tick_nohz_enabled_snap) {
1544 if (rcu_cpu_has_callbacks(NULL))
1545 invoke_rcu_core(); /* force nohz to see update. */
1546 rdtp->tick_nohz_enabled_snap = tne;
1552 /* If this is a no-CBs CPU, no callbacks, just return. */
1553 if (rcu_is_nocb_cpu(smp_processor_id()))
1557 * If a non-lazy callback arrived at a CPU having only lazy
1558 * callbacks, invoke RCU core for the side-effect of recalculating
1559 * idle duration on re-entry to idle.
1561 if (rdtp->all_lazy &&
1562 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1563 rdtp->all_lazy = false;
1564 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1570 * If we have not yet accelerated this jiffy, accelerate all
1571 * callbacks on this CPU.
1573 if (rdtp->last_accelerate == jiffies)
1575 rdtp->last_accelerate = jiffies;
1576 for_each_rcu_flavor(rsp) {
1577 rdp = this_cpu_ptr(rsp->rda);
1578 if (!*rdp->nxttail[RCU_DONE_TAIL])
1581 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1582 smp_mb__after_unlock_lock();
1583 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1584 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1586 rcu_gp_kthread_wake(rsp);
1588 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1592 * Clean up for exit from idle. Attempt to advance callbacks based on
1593 * any grace periods that elapsed while the CPU was idle, and if any
1594 * callbacks are now ready to invoke, initiate invocation.
1596 static void rcu_cleanup_after_idle(void)
1598 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1599 if (rcu_is_nocb_cpu(smp_processor_id()))
1601 if (rcu_try_advance_all_cbs())
1603 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1607 * Keep a running count of the number of non-lazy callbacks posted
1608 * on this CPU. This running counter (which is never decremented) allows
1609 * rcu_prepare_for_idle() to detect when something out of the idle loop
1610 * posts a callback, even if an equal number of callbacks are invoked.
1611 * Of course, callbacks should only be posted from within a trace event
1612 * designed to be called from idle or from within RCU_NONIDLE().
1614 static void rcu_idle_count_callbacks_posted(void)
1616 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1620 * Data for flushing lazy RCU callbacks at OOM time.
1622 static atomic_t oom_callback_count;
1623 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1626 * RCU OOM callback -- decrement the outstanding count and deliver the
1627 * wake-up if we are the last one.
1629 static void rcu_oom_callback(struct rcu_head *rhp)
1631 if (atomic_dec_and_test(&oom_callback_count))
1632 wake_up(&oom_callback_wq);
1636 * Post an rcu_oom_notify callback on the current CPU if it has at
1637 * least one lazy callback. This will unnecessarily post callbacks
1638 * to CPUs that already have a non-lazy callback at the end of their
1639 * callback list, but this is an infrequent operation, so accept some
1640 * extra overhead to keep things simple.
1642 static void rcu_oom_notify_cpu(void *unused)
1644 struct rcu_state *rsp;
1645 struct rcu_data *rdp;
1647 for_each_rcu_flavor(rsp) {
1648 rdp = raw_cpu_ptr(rsp->rda);
1649 if (rdp->qlen_lazy != 0) {
1650 atomic_inc(&oom_callback_count);
1651 rsp->call(&rdp->oom_head, rcu_oom_callback);
1657 * If low on memory, ensure that each CPU has a non-lazy callback.
1658 * This will wake up CPUs that have only lazy callbacks, in turn
1659 * ensuring that they free up the corresponding memory in a timely manner.
1660 * Because an uncertain amount of memory will be freed in some uncertain
1661 * timeframe, we do not claim to have freed anything.
1663 static int rcu_oom_notify(struct notifier_block *self,
1664 unsigned long notused, void *nfreed)
1668 /* Wait for callbacks from earlier instance to complete. */
1669 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1670 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1673 * Prevent premature wakeup: ensure that all increments happen
1674 * before there is a chance of the counter reaching zero.
1676 atomic_set(&oom_callback_count, 1);
1679 for_each_online_cpu(cpu) {
1680 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1681 cond_resched_rcu_qs();
1685 /* Unconditionally decrement: no need to wake ourselves up. */
1686 atomic_dec(&oom_callback_count);
1691 static struct notifier_block rcu_oom_nb = {
1692 .notifier_call = rcu_oom_notify
1695 static int __init rcu_register_oom_notifier(void)
1697 register_oom_notifier(&rcu_oom_nb);
1700 early_initcall(rcu_register_oom_notifier);
1702 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1704 #ifdef CONFIG_RCU_CPU_STALL_INFO
1706 #ifdef CONFIG_RCU_FAST_NO_HZ
1708 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1710 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1711 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1713 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1714 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1716 rdtp->all_lazy ? 'L' : '.',
1717 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1720 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1722 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1727 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1729 /* Initiate the stall-info list. */
1730 static void print_cpu_stall_info_begin(void)
1736 * Print out diagnostic information for the specified stalled CPU.
1738 * If the specified CPU is aware of the current RCU grace period
1739 * (flavor specified by rsp), then print the number of scheduling
1740 * clock interrupts the CPU has taken during the time that it has
1741 * been aware. Otherwise, print the number of RCU grace periods
1742 * that this CPU is ignorant of, for example, "1" if the CPU was
1743 * aware of the previous grace period.
1745 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1747 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1749 char fast_no_hz[72];
1750 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1751 struct rcu_dynticks *rdtp = rdp->dynticks;
1753 unsigned long ticks_value;
1755 if (rsp->gpnum == rdp->gpnum) {
1756 ticks_title = "ticks this GP";
1757 ticks_value = rdp->ticks_this_gp;
1759 ticks_title = "GPs behind";
1760 ticks_value = rsp->gpnum - rdp->gpnum;
1762 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1763 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1764 cpu, ticks_value, ticks_title,
1765 atomic_read(&rdtp->dynticks) & 0xfff,
1766 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1767 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1768 ACCESS_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1772 /* Terminate the stall-info list. */
1773 static void print_cpu_stall_info_end(void)
1778 /* Zero ->ticks_this_gp for all flavors of RCU. */
1779 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1781 rdp->ticks_this_gp = 0;
1782 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1785 /* Increment ->ticks_this_gp for all flavors of RCU. */
1786 static void increment_cpu_stall_ticks(void)
1788 struct rcu_state *rsp;
1790 for_each_rcu_flavor(rsp)
1791 raw_cpu_inc(rsp->rda->ticks_this_gp);
1794 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1796 static void print_cpu_stall_info_begin(void)
1801 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1803 pr_cont(" %d", cpu);
1806 static void print_cpu_stall_info_end(void)
1811 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1815 static void increment_cpu_stall_ticks(void)
1819 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1821 #ifdef CONFIG_RCU_NOCB_CPU
1824 * Offload callback processing from the boot-time-specified set of CPUs
1825 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1826 * kthread created that pulls the callbacks from the corresponding CPU,
1827 * waits for a grace period to elapse, and invokes the callbacks.
1828 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1829 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1830 * has been specified, in which case each kthread actively polls its
1831 * CPU. (Which isn't so great for energy efficiency, but which does
1832 * reduce RCU's overhead on that CPU.)
1834 * This is intended to be used in conjunction with Frederic Weisbecker's
1835 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1836 * running CPU-bound user-mode computations.
1838 * Offloading of callback processing could also in theory be used as
1839 * an energy-efficiency measure because CPUs with no RCU callbacks
1840 * queued are more aggressive about entering dyntick-idle mode.
1844 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1845 static int __init rcu_nocb_setup(char *str)
1847 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1848 have_rcu_nocb_mask = true;
1849 cpulist_parse(str, rcu_nocb_mask);
1852 __setup("rcu_nocbs=", rcu_nocb_setup);
1854 static int __init parse_rcu_nocb_poll(char *arg)
1859 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1862 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1865 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1867 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1871 * Set the root rcu_node structure's ->need_future_gp field
1872 * based on the sum of those of all rcu_node structures. This does
1873 * double-count the root rcu_node structure's requests, but this
1874 * is necessary to handle the possibility of a rcu_nocb_kthread()
1875 * having awakened during the time that the rcu_node structures
1876 * were being updated for the end of the previous grace period.
1878 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1880 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1883 static void rcu_init_one_nocb(struct rcu_node *rnp)
1885 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1886 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1889 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1890 /* Is the specified CPU a no-CBs CPU? */
1891 bool rcu_is_nocb_cpu(int cpu)
1893 if (have_rcu_nocb_mask)
1894 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1897 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1900 * Kick the leader kthread for this NOCB group.
1902 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1904 struct rcu_data *rdp_leader = rdp->nocb_leader;
1906 if (!ACCESS_ONCE(rdp_leader->nocb_kthread))
1908 if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1909 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1910 ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false;
1911 wake_up(&rdp_leader->nocb_wq);
1916 * Does the specified CPU need an RCU callback for the specified flavor
1919 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1921 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1923 #ifdef CONFIG_PROVE_RCU
1924 struct rcu_head *rhp;
1925 #endif /* #ifdef CONFIG_PROVE_RCU */
1928 * Check count of all no-CBs callbacks awaiting invocation.
1929 * There needs to be a barrier before this function is called,
1930 * but associated with a prior determination that no more
1931 * callbacks would be posted. In the worst case, the first
1932 * barrier in _rcu_barrier() suffices (but the caller cannot
1933 * necessarily rely on this, not a substitute for the caller
1934 * getting the concurrency design right!). There must also be
1935 * a barrier between the following load an posting of a callback
1936 * (if a callback is in fact needed). This is associated with an
1937 * atomic_inc() in the caller.
1939 ret = atomic_long_read(&rdp->nocb_q_count);
1941 #ifdef CONFIG_PROVE_RCU
1942 rhp = ACCESS_ONCE(rdp->nocb_head);
1944 rhp = ACCESS_ONCE(rdp->nocb_gp_head);
1946 rhp = ACCESS_ONCE(rdp->nocb_follower_head);
1948 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1949 if (!ACCESS_ONCE(rdp->nocb_kthread) && rhp) {
1950 /* RCU callback enqueued before CPU first came online??? */
1951 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1955 #endif /* #ifdef CONFIG_PROVE_RCU */
1961 * Enqueue the specified string of rcu_head structures onto the specified
1962 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1963 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1964 * counts are supplied by rhcount and rhcount_lazy.
1966 * If warranted, also wake up the kthread servicing this CPUs queues.
1968 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1969 struct rcu_head *rhp,
1970 struct rcu_head **rhtp,
1971 int rhcount, int rhcount_lazy,
1972 unsigned long flags)
1975 struct rcu_head **old_rhpp;
1976 struct task_struct *t;
1978 /* Enqueue the callback on the nocb list and update counts. */
1979 atomic_long_add(rhcount, &rdp->nocb_q_count);
1980 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1981 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1982 ACCESS_ONCE(*old_rhpp) = rhp;
1983 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1984 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1986 /* If we are not being polled and there is a kthread, awaken it ... */
1987 t = ACCESS_ONCE(rdp->nocb_kthread);
1988 if (rcu_nocb_poll || !t) {
1989 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1990 TPS("WakeNotPoll"));
1993 len = atomic_long_read(&rdp->nocb_q_count);
1994 if (old_rhpp == &rdp->nocb_head) {
1995 if (!irqs_disabled_flags(flags)) {
1996 /* ... if queue was empty ... */
1997 wake_nocb_leader(rdp, false);
1998 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2001 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
2002 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2003 TPS("WakeEmptyIsDeferred"));
2005 rdp->qlen_last_fqs_check = 0;
2006 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2007 /* ... or if many callbacks queued. */
2008 if (!irqs_disabled_flags(flags)) {
2009 wake_nocb_leader(rdp, true);
2010 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2013 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
2014 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2015 TPS("WakeOvfIsDeferred"));
2017 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2019 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2025 * This is a helper for __call_rcu(), which invokes this when the normal
2026 * callback queue is inoperable. If this is not a no-CBs CPU, this
2027 * function returns failure back to __call_rcu(), which can complain
2030 * Otherwise, this function queues the callback where the corresponding
2031 * "rcuo" kthread can find it.
2033 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2034 bool lazy, unsigned long flags)
2037 if (!rcu_is_nocb_cpu(rdp->cpu))
2039 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2040 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2041 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2042 (unsigned long)rhp->func,
2043 -atomic_long_read(&rdp->nocb_q_count_lazy),
2044 -atomic_long_read(&rdp->nocb_q_count));
2046 trace_rcu_callback(rdp->rsp->name, rhp,
2047 -atomic_long_read(&rdp->nocb_q_count_lazy),
2048 -atomic_long_read(&rdp->nocb_q_count));
2051 * If called from an extended quiescent state with interrupts
2052 * disabled, invoke the RCU core in order to allow the idle-entry
2053 * deferred-wakeup check to function.
2055 if (irqs_disabled_flags(flags) &&
2056 !rcu_is_watching() &&
2057 cpu_online(smp_processor_id()))
2064 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2067 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2068 struct rcu_data *rdp,
2069 unsigned long flags)
2071 long ql = rsp->qlen;
2072 long qll = rsp->qlen_lazy;
2074 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2075 if (!rcu_is_nocb_cpu(smp_processor_id()))
2080 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2081 if (rsp->orphan_donelist != NULL) {
2082 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2083 rsp->orphan_donetail, ql, qll, flags);
2085 rsp->orphan_donelist = NULL;
2086 rsp->orphan_donetail = &rsp->orphan_donelist;
2088 if (rsp->orphan_nxtlist != NULL) {
2089 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2090 rsp->orphan_nxttail, ql, qll, flags);
2092 rsp->orphan_nxtlist = NULL;
2093 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2099 * If necessary, kick off a new grace period, and either way wait
2100 * for a subsequent grace period to complete.
2102 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2106 unsigned long flags;
2108 struct rcu_node *rnp = rdp->mynode;
2110 raw_spin_lock_irqsave(&rnp->lock, flags);
2111 smp_mb__after_unlock_lock();
2112 needwake = rcu_start_future_gp(rnp, rdp, &c);
2113 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2115 rcu_gp_kthread_wake(rdp->rsp);
2118 * Wait for the grace period. Do so interruptibly to avoid messing
2119 * up the load average.
2121 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2123 wait_event_interruptible(
2124 rnp->nocb_gp_wq[c & 0x1],
2125 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2128 WARN_ON(signal_pending(current));
2129 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2131 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2132 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2136 * Leaders come here to wait for additional callbacks to show up.
2137 * This function does not return until callbacks appear.
2139 static void nocb_leader_wait(struct rcu_data *my_rdp)
2141 bool firsttime = true;
2143 struct rcu_data *rdp;
2144 struct rcu_head **tail;
2148 /* Wait for callbacks to appear. */
2149 if (!rcu_nocb_poll) {
2150 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2151 wait_event_interruptible(my_rdp->nocb_wq,
2152 !ACCESS_ONCE(my_rdp->nocb_leader_sleep));
2153 /* Memory barrier handled by smp_mb() calls below and repoll. */
2154 } else if (firsttime) {
2155 firsttime = false; /* Don't drown trace log with "Poll"! */
2156 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2160 * Each pass through the following loop checks a follower for CBs.
2161 * We are our own first follower. Any CBs found are moved to
2162 * nocb_gp_head, where they await a grace period.
2165 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2166 rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head);
2167 if (!rdp->nocb_gp_head)
2168 continue; /* No CBs here, try next follower. */
2170 /* Move callbacks to wait-for-GP list, which is empty. */
2171 ACCESS_ONCE(rdp->nocb_head) = NULL;
2172 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2177 * If there were no callbacks, sleep a bit, rescan after a
2178 * memory barrier, and go retry.
2180 if (unlikely(!gotcbs)) {
2182 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2184 WARN_ON(signal_pending(current));
2185 schedule_timeout_interruptible(1);
2187 /* Rescan in case we were a victim of memory ordering. */
2188 my_rdp->nocb_leader_sleep = true;
2189 smp_mb(); /* Ensure _sleep true before scan. */
2190 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2191 if (ACCESS_ONCE(rdp->nocb_head)) {
2192 /* Found CB, so short-circuit next wait. */
2193 my_rdp->nocb_leader_sleep = false;
2199 /* Wait for one grace period. */
2200 rcu_nocb_wait_gp(my_rdp);
2203 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2204 * We set it now, but recheck for new callbacks while
2205 * traversing our follower list.
2207 my_rdp->nocb_leader_sleep = true;
2208 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2210 /* Each pass through the following loop wakes a follower, if needed. */
2211 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2212 if (ACCESS_ONCE(rdp->nocb_head))
2213 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2214 if (!rdp->nocb_gp_head)
2215 continue; /* No CBs, so no need to wake follower. */
2217 /* Append callbacks to follower's "done" list. */
2218 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2219 *tail = rdp->nocb_gp_head;
2220 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2221 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2223 * List was empty, wake up the follower.
2224 * Memory barriers supplied by atomic_long_add().
2226 wake_up(&rdp->nocb_wq);
2230 /* If we (the leader) don't have CBs, go wait some more. */
2231 if (!my_rdp->nocb_follower_head)
2236 * Followers come here to wait for additional callbacks to show up.
2237 * This function does not return until callbacks appear.
2239 static void nocb_follower_wait(struct rcu_data *rdp)
2241 bool firsttime = true;
2244 if (!rcu_nocb_poll) {
2245 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2247 wait_event_interruptible(rdp->nocb_wq,
2248 ACCESS_ONCE(rdp->nocb_follower_head));
2249 } else if (firsttime) {
2250 /* Don't drown trace log with "Poll"! */
2252 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2254 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2255 /* ^^^ Ensure CB invocation follows _head test. */
2259 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2261 WARN_ON(signal_pending(current));
2262 schedule_timeout_interruptible(1);
2267 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2268 * callbacks queued by the corresponding no-CBs CPU, however, there is
2269 * an optional leader-follower relationship so that the grace-period
2270 * kthreads don't have to do quite so many wakeups.
2272 static int rcu_nocb_kthread(void *arg)
2275 struct rcu_head *list;
2276 struct rcu_head *next;
2277 struct rcu_head **tail;
2278 struct rcu_data *rdp = arg;
2280 /* Each pass through this loop invokes one batch of callbacks */
2282 /* Wait for callbacks. */
2283 if (rdp->nocb_leader == rdp)
2284 nocb_leader_wait(rdp);
2286 nocb_follower_wait(rdp);
2288 /* Pull the ready-to-invoke callbacks onto local list. */
2289 list = ACCESS_ONCE(rdp->nocb_follower_head);
2291 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2292 ACCESS_ONCE(rdp->nocb_follower_head) = NULL;
2293 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2295 /* Each pass through the following loop invokes a callback. */
2296 trace_rcu_batch_start(rdp->rsp->name,
2297 atomic_long_read(&rdp->nocb_q_count_lazy),
2298 atomic_long_read(&rdp->nocb_q_count), -1);
2302 /* Wait for enqueuing to complete, if needed. */
2303 while (next == NULL && &list->next != tail) {
2304 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2306 schedule_timeout_interruptible(1);
2307 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2311 debug_rcu_head_unqueue(list);
2313 if (__rcu_reclaim(rdp->rsp->name, list))
2319 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2320 smp_mb__before_atomic(); /* _add after CB invocation. */
2321 atomic_long_add(-c, &rdp->nocb_q_count);
2322 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2323 rdp->n_nocbs_invoked += c;
2328 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2329 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2331 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2334 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2335 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2339 if (!rcu_nocb_need_deferred_wakeup(rdp))
2341 ndw = ACCESS_ONCE(rdp->nocb_defer_wakeup);
2342 ACCESS_ONCE(rdp->nocb_defer_wakeup) = RCU_NOGP_WAKE_NOT;
2343 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2344 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2347 void __init rcu_init_nohz(void)
2350 bool need_rcu_nocb_mask = true;
2351 struct rcu_state *rsp;
2353 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2354 need_rcu_nocb_mask = false;
2355 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2357 #if defined(CONFIG_NO_HZ_FULL)
2358 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2359 need_rcu_nocb_mask = true;
2360 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2362 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2363 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2364 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2367 have_rcu_nocb_mask = true;
2369 if (!have_rcu_nocb_mask)
2372 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2373 pr_info("\tOffload RCU callbacks from CPU 0\n");
2374 cpumask_set_cpu(0, rcu_nocb_mask);
2375 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2376 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2377 pr_info("\tOffload RCU callbacks from all CPUs\n");
2378 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2379 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2380 #if defined(CONFIG_NO_HZ_FULL)
2381 if (tick_nohz_full_running)
2382 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2383 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2385 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2386 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2387 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2390 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
2391 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
2393 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2395 for_each_rcu_flavor(rsp) {
2396 for_each_cpu(cpu, rcu_nocb_mask) {
2397 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2400 * If there are early callbacks, they will need
2401 * to be moved to the nocb lists.
2403 WARN_ON_ONCE(rdp->nxttail[RCU_NEXT_TAIL] !=
2405 rdp->nxttail[RCU_NEXT_TAIL] != NULL);
2406 init_nocb_callback_list(rdp);
2408 rcu_organize_nocb_kthreads(rsp);
2412 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2413 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2415 rdp->nocb_tail = &rdp->nocb_head;
2416 init_waitqueue_head(&rdp->nocb_wq);
2417 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2421 * If the specified CPU is a no-CBs CPU that does not already have its
2422 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2423 * brought online out of order, this can require re-organizing the
2424 * leader-follower relationships.
2426 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2428 struct rcu_data *rdp;
2429 struct rcu_data *rdp_last;
2430 struct rcu_data *rdp_old_leader;
2431 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2432 struct task_struct *t;
2435 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2436 * then nothing to do.
2438 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2441 /* If we didn't spawn the leader first, reorganize! */
2442 rdp_old_leader = rdp_spawn->nocb_leader;
2443 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2445 rdp = rdp_old_leader;
2447 rdp->nocb_leader = rdp_spawn;
2448 if (rdp_last && rdp != rdp_spawn)
2449 rdp_last->nocb_next_follower = rdp;
2450 if (rdp == rdp_spawn) {
2451 rdp = rdp->nocb_next_follower;
2454 rdp = rdp->nocb_next_follower;
2455 rdp_last->nocb_next_follower = NULL;
2458 rdp_spawn->nocb_next_follower = rdp_old_leader;
2461 /* Spawn the kthread for this CPU and RCU flavor. */
2462 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2463 "rcuo%c/%d", rsp->abbr, cpu);
2465 ACCESS_ONCE(rdp_spawn->nocb_kthread) = t;
2469 * If the specified CPU is a no-CBs CPU that does not already have its
2470 * rcuo kthreads, spawn them.
2472 static void rcu_spawn_all_nocb_kthreads(int cpu)
2474 struct rcu_state *rsp;
2476 if (rcu_scheduler_fully_active)
2477 for_each_rcu_flavor(rsp)
2478 rcu_spawn_one_nocb_kthread(rsp, cpu);
2482 * Once the scheduler is running, spawn rcuo kthreads for all online
2483 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2484 * non-boot CPUs come online -- if this changes, we will need to add
2485 * some mutual exclusion.
2487 static void __init rcu_spawn_nocb_kthreads(void)
2491 for_each_online_cpu(cpu)
2492 rcu_spawn_all_nocb_kthreads(cpu);
2495 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2496 static int rcu_nocb_leader_stride = -1;
2497 module_param(rcu_nocb_leader_stride, int, 0444);
2500 * Initialize leader-follower relationships for all no-CBs CPU.
2502 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2505 int ls = rcu_nocb_leader_stride;
2506 int nl = 0; /* Next leader. */
2507 struct rcu_data *rdp;
2508 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2509 struct rcu_data *rdp_prev = NULL;
2511 if (!have_rcu_nocb_mask)
2514 ls = int_sqrt(nr_cpu_ids);
2515 rcu_nocb_leader_stride = ls;
2519 * Each pass through this loop sets up one rcu_data structure and
2520 * spawns one rcu_nocb_kthread().
2522 for_each_cpu(cpu, rcu_nocb_mask) {
2523 rdp = per_cpu_ptr(rsp->rda, cpu);
2524 if (rdp->cpu >= nl) {
2525 /* New leader, set up for followers & next leader. */
2526 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2527 rdp->nocb_leader = rdp;
2530 /* Another follower, link to previous leader. */
2531 rdp->nocb_leader = rdp_leader;
2532 rdp_prev->nocb_next_follower = rdp;
2538 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2539 static bool init_nocb_callback_list(struct rcu_data *rdp)
2541 if (!rcu_is_nocb_cpu(rdp->cpu))
2544 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2548 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2550 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2552 WARN_ON_ONCE(1); /* Should be dead code. */
2556 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2560 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2564 static void rcu_init_one_nocb(struct rcu_node *rnp)
2568 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2569 bool lazy, unsigned long flags)
2574 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2575 struct rcu_data *rdp,
2576 unsigned long flags)
2581 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2585 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2590 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2594 static void rcu_spawn_all_nocb_kthreads(int cpu)
2598 static void __init rcu_spawn_nocb_kthreads(void)
2602 static bool init_nocb_callback_list(struct rcu_data *rdp)
2607 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2610 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2611 * arbitrarily long period of time with the scheduling-clock tick turned
2612 * off. RCU will be paying attention to this CPU because it is in the
2613 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2614 * machine because the scheduling-clock tick has been disabled. Therefore,
2615 * if an adaptive-ticks CPU is failing to respond to the current grace
2616 * period and has not be idle from an RCU perspective, kick it.
2618 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2620 #ifdef CONFIG_NO_HZ_FULL
2621 if (tick_nohz_full_cpu(cpu))
2622 smp_send_reschedule(cpu);
2623 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2627 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2629 static int full_sysidle_state; /* Current system-idle state. */
2630 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2631 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2632 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2633 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2634 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2637 * Invoked to note exit from irq or task transition to idle. Note that
2638 * usermode execution does -not- count as idle here! After all, we want
2639 * to detect full-system idle states, not RCU quiescent states and grace
2640 * periods. The caller must have disabled interrupts.
2642 static void rcu_sysidle_enter(int irq)
2645 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2647 /* If there are no nohz_full= CPUs, no need to track this. */
2648 if (!tick_nohz_full_enabled())
2651 /* Adjust nesting, check for fully idle. */
2653 rdtp->dynticks_idle_nesting--;
2654 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2655 if (rdtp->dynticks_idle_nesting != 0)
2656 return; /* Still not fully idle. */
2658 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2659 DYNTICK_TASK_NEST_VALUE) {
2660 rdtp->dynticks_idle_nesting = 0;
2662 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2663 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2664 return; /* Still not fully idle. */
2668 /* Record start of fully idle period. */
2670 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2671 smp_mb__before_atomic();
2672 atomic_inc(&rdtp->dynticks_idle);
2673 smp_mb__after_atomic();
2674 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2678 * Unconditionally force exit from full system-idle state. This is
2679 * invoked when a normal CPU exits idle, but must be called separately
2680 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2681 * is that the timekeeping CPU is permitted to take scheduling-clock
2682 * interrupts while the system is in system-idle state, and of course
2683 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2684 * interrupt from any other type of interrupt.
2686 void rcu_sysidle_force_exit(void)
2688 int oldstate = ACCESS_ONCE(full_sysidle_state);
2692 * Each pass through the following loop attempts to exit full
2693 * system-idle state. If contention proves to be a problem,
2694 * a trylock-based contention tree could be used here.
2696 while (oldstate > RCU_SYSIDLE_SHORT) {
2697 newoldstate = cmpxchg(&full_sysidle_state,
2698 oldstate, RCU_SYSIDLE_NOT);
2699 if (oldstate == newoldstate &&
2700 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2701 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2702 return; /* We cleared it, done! */
2704 oldstate = newoldstate;
2706 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2710 * Invoked to note entry to irq or task transition from idle. Note that
2711 * usermode execution does -not- count as idle here! The caller must
2712 * have disabled interrupts.
2714 static void rcu_sysidle_exit(int irq)
2716 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2718 /* If there are no nohz_full= CPUs, no need to track this. */
2719 if (!tick_nohz_full_enabled())
2722 /* Adjust nesting, check for already non-idle. */
2724 rdtp->dynticks_idle_nesting++;
2725 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2726 if (rdtp->dynticks_idle_nesting != 1)
2727 return; /* Already non-idle. */
2730 * Allow for irq misnesting. Yes, it really is possible
2731 * to enter an irq handler then never leave it, and maybe
2732 * also vice versa. Handle both possibilities.
2734 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2735 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2736 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2737 return; /* Already non-idle. */
2739 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2743 /* Record end of idle period. */
2744 smp_mb__before_atomic();
2745 atomic_inc(&rdtp->dynticks_idle);
2746 smp_mb__after_atomic();
2747 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2750 * If we are the timekeeping CPU, we are permitted to be non-idle
2751 * during a system-idle state. This must be the case, because
2752 * the timekeeping CPU has to take scheduling-clock interrupts
2753 * during the time that the system is transitioning to full
2754 * system-idle state. This means that the timekeeping CPU must
2755 * invoke rcu_sysidle_force_exit() directly if it does anything
2756 * more than take a scheduling-clock interrupt.
2758 if (smp_processor_id() == tick_do_timer_cpu)
2761 /* Update system-idle state: We are clearly no longer fully idle! */
2762 rcu_sysidle_force_exit();
2766 * Check to see if the current CPU is idle. Note that usermode execution
2767 * does not count as idle. The caller must have disabled interrupts.
2769 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2770 unsigned long *maxj)
2774 struct rcu_dynticks *rdtp = rdp->dynticks;
2776 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2777 if (!tick_nohz_full_enabled())
2781 * If some other CPU has already reported non-idle, if this is
2782 * not the flavor of RCU that tracks sysidle state, or if this
2783 * is an offline or the timekeeping CPU, nothing to do.
2785 if (!*isidle || rdp->rsp != rcu_state_p ||
2786 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2788 if (rcu_gp_in_progress(rdp->rsp))
2789 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2791 /* Pick up current idle and NMI-nesting counter and check. */
2792 cur = atomic_read(&rdtp->dynticks_idle);
2794 *isidle = false; /* We are not idle! */
2797 smp_mb(); /* Read counters before timestamps. */
2799 /* Pick up timestamps. */
2800 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2801 /* If this CPU entered idle more recently, update maxj timestamp. */
2802 if (ULONG_CMP_LT(*maxj, j))
2807 * Is this the flavor of RCU that is handling full-system idle?
2809 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2811 return rsp == rcu_state_p;
2815 * Return a delay in jiffies based on the number of CPUs, rcu_node
2816 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2817 * systems more time to transition to full-idle state in order to
2818 * avoid the cache thrashing that otherwise occur on the state variable.
2819 * Really small systems (less than a couple of tens of CPUs) should
2820 * instead use a single global atomically incremented counter, and later
2821 * versions of this will automatically reconfigure themselves accordingly.
2823 static unsigned long rcu_sysidle_delay(void)
2825 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2827 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2831 * Advance the full-system-idle state. This is invoked when all of
2832 * the non-timekeeping CPUs are idle.
2834 static void rcu_sysidle(unsigned long j)
2836 /* Check the current state. */
2837 switch (ACCESS_ONCE(full_sysidle_state)) {
2838 case RCU_SYSIDLE_NOT:
2840 /* First time all are idle, so note a short idle period. */
2841 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2844 case RCU_SYSIDLE_SHORT:
2847 * Idle for a bit, time to advance to next state?
2848 * cmpxchg failure means race with non-idle, let them win.
2850 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2851 (void)cmpxchg(&full_sysidle_state,
2852 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2855 case RCU_SYSIDLE_LONG:
2858 * Do an additional check pass before advancing to full.
2859 * cmpxchg failure means race with non-idle, let them win.
2861 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2862 (void)cmpxchg(&full_sysidle_state,
2863 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2872 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2873 * back to the beginning.
2875 static void rcu_sysidle_cancel(void)
2878 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2879 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2883 * Update the sysidle state based on the results of a force-quiescent-state
2884 * scan of the CPUs' dyntick-idle state.
2886 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2887 unsigned long maxj, bool gpkt)
2889 if (rsp != rcu_state_p)
2890 return; /* Wrong flavor, ignore. */
2891 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2892 return; /* Running state machine from timekeeping CPU. */
2894 rcu_sysidle(maxj); /* More idle! */
2896 rcu_sysidle_cancel(); /* Idle is over. */
2900 * Wrapper for rcu_sysidle_report() when called from the grace-period
2901 * kthread's context.
2903 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2906 /* If there are no nohz_full= CPUs, no need to track this. */
2907 if (!tick_nohz_full_enabled())
2910 rcu_sysidle_report(rsp, isidle, maxj, true);
2913 /* Callback and function for forcing an RCU grace period. */
2914 struct rcu_sysidle_head {
2919 static void rcu_sysidle_cb(struct rcu_head *rhp)
2921 struct rcu_sysidle_head *rshp;
2924 * The following memory barrier is needed to replace the
2925 * memory barriers that would normally be in the memory
2928 smp_mb(); /* grace period precedes setting inuse. */
2930 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2931 ACCESS_ONCE(rshp->inuse) = 0;
2935 * Check to see if the system is fully idle, other than the timekeeping CPU.
2936 * The caller must have disabled interrupts. This is not intended to be
2937 * called unless tick_nohz_full_enabled().
2939 bool rcu_sys_is_idle(void)
2941 static struct rcu_sysidle_head rsh;
2942 int rss = ACCESS_ONCE(full_sysidle_state);
2944 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2947 /* Handle small-system case by doing a full scan of CPUs. */
2948 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2949 int oldrss = rss - 1;
2952 * One pass to advance to each state up to _FULL.
2953 * Give up if any pass fails to advance the state.
2955 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2958 unsigned long maxj = jiffies - ULONG_MAX / 4;
2959 struct rcu_data *rdp;
2961 /* Scan all the CPUs looking for nonidle CPUs. */
2962 for_each_possible_cpu(cpu) {
2963 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2964 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2968 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2970 rss = ACCESS_ONCE(full_sysidle_state);
2974 /* If this is the first observation of an idle period, record it. */
2975 if (rss == RCU_SYSIDLE_FULL) {
2976 rss = cmpxchg(&full_sysidle_state,
2977 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2978 return rss == RCU_SYSIDLE_FULL;
2981 smp_mb(); /* ensure rss load happens before later caller actions. */
2983 /* If already fully idle, tell the caller (in case of races). */
2984 if (rss == RCU_SYSIDLE_FULL_NOTED)
2988 * If we aren't there yet, and a grace period is not in flight,
2989 * initiate a grace period. Either way, tell the caller that
2990 * we are not there yet. We use an xchg() rather than an assignment
2991 * to make up for the memory barriers that would otherwise be
2992 * provided by the memory allocator.
2994 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2995 !rcu_gp_in_progress(rcu_state_p) &&
2996 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2997 call_rcu(&rsh.rh, rcu_sysidle_cb);
3002 * Initialize dynticks sysidle state for CPUs coming online.
3004 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3006 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
3009 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3011 static void rcu_sysidle_enter(int irq)
3015 static void rcu_sysidle_exit(int irq)
3019 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3020 unsigned long *maxj)
3024 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3029 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3034 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3038 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3041 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3042 * grace-period kthread will do force_quiescent_state() processing?
3043 * The idea is to avoid waking up RCU core processing on such a
3044 * CPU unless the grace period has extended for too long.
3046 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3047 * CONFIG_RCU_NOCB_CPU CPUs.
3049 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3051 #ifdef CONFIG_NO_HZ_FULL
3052 if (tick_nohz_full_cpu(smp_processor_id()) &&
3053 (!rcu_gp_in_progress(rsp) ||
3054 ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
3056 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3061 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3064 static void rcu_bind_gp_kthread(void)
3066 int __maybe_unused cpu;
3068 if (!tick_nohz_full_enabled())
3070 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3071 cpu = tick_do_timer_cpu;
3072 if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu)
3073 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3074 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3075 if (!is_housekeeping_cpu(raw_smp_processor_id()))
3076 housekeeping_affine(current);
3077 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3080 /* Record the current task on dyntick-idle entry. */
3081 static void rcu_dynticks_task_enter(void)
3083 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3084 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = smp_processor_id();
3085 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3088 /* Record no current task on dyntick-idle exit. */
3089 static void rcu_dynticks_task_exit(void)
3091 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3092 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = -1;
3093 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */