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 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
55 * Check the RCU kernel configuration parameters and print informative
56 * messages about anything out of the ordinary. If you like #ifdef, you
57 * will love this function.
59 static void __init rcu_bootup_announce_oddness(void)
61 #ifdef CONFIG_RCU_TRACE
62 pr_info("\tRCU debugfs-based tracing is enabled.\n");
64 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
65 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
68 #ifdef CONFIG_RCU_FANOUT_EXACT
69 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
71 #ifdef CONFIG_RCU_FAST_NO_HZ
72 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
74 #ifdef CONFIG_PROVE_RCU
75 pr_info("\tRCU lockdep checking is enabled.\n");
77 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
78 pr_info("\tRCU torture testing starts during boot.\n");
80 #if defined(CONFIG_RCU_CPU_STALL_INFO)
81 pr_info("\tAdditional per-CPU info printed with stalls.\n");
83 #if NUM_RCU_LVL_4 != 0
84 pr_info("\tFour-level hierarchy is enabled.\n");
86 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
87 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
88 if (nr_cpu_ids != NR_CPUS)
89 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
90 #ifdef CONFIG_RCU_BOOST
91 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
95 #ifdef CONFIG_PREEMPT_RCU
97 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
98 static struct rcu_state *rcu_state_p = &rcu_preempt_state;
100 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
101 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
105 * Tell them what RCU they are running.
107 static void __init rcu_bootup_announce(void)
109 pr_info("Preemptible hierarchical RCU implementation.\n");
110 rcu_bootup_announce_oddness();
114 * Record a preemptible-RCU quiescent state for the specified CPU. Note
115 * that this just means that the task currently running on the CPU is
116 * not in a quiescent state. There might be any number of tasks blocked
117 * while in an RCU read-side critical section.
119 * As with the other rcu_*_qs() functions, callers to this function
120 * must disable preemption.
122 static void rcu_preempt_qs(void)
124 if (!__this_cpu_read(rcu_preempt_data.passed_quiesce)) {
125 trace_rcu_grace_period(TPS("rcu_preempt"),
126 __this_cpu_read(rcu_preempt_data.gpnum),
128 __this_cpu_write(rcu_preempt_data.passed_quiesce, 1);
129 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
130 current->rcu_read_unlock_special.b.need_qs = false;
135 * We have entered the scheduler, and the current task might soon be
136 * context-switched away from. If this task is in an RCU read-side
137 * critical section, we will no longer be able to rely on the CPU to
138 * record that fact, so we enqueue the task on the blkd_tasks list.
139 * The task will dequeue itself when it exits the outermost enclosing
140 * RCU read-side critical section. Therefore, the current grace period
141 * cannot be permitted to complete until the blkd_tasks list entries
142 * predating the current grace period drain, in other words, until
143 * rnp->gp_tasks becomes NULL.
145 * Caller must disable preemption.
147 static void rcu_preempt_note_context_switch(void)
149 struct task_struct *t = current;
151 struct rcu_data *rdp;
152 struct rcu_node *rnp;
154 if (t->rcu_read_lock_nesting > 0 &&
155 !t->rcu_read_unlock_special.b.blocked) {
157 /* Possibly blocking in an RCU read-side critical section. */
158 rdp = this_cpu_ptr(rcu_preempt_state.rda);
160 raw_spin_lock_irqsave(&rnp->lock, flags);
161 smp_mb__after_unlock_lock();
162 t->rcu_read_unlock_special.b.blocked = true;
163 t->rcu_blocked_node = rnp;
166 * If this CPU has already checked in, then this task
167 * will hold up the next grace period rather than the
168 * current grace period. Queue the task accordingly.
169 * If the task is queued for the current grace period
170 * (i.e., this CPU has not yet passed through a quiescent
171 * state for the current grace period), then as long
172 * as that task remains queued, the current grace period
173 * cannot end. Note that there is some uncertainty as
174 * to exactly when the current grace period started.
175 * We take a conservative approach, which can result
176 * in unnecessarily waiting on tasks that started very
177 * slightly after the current grace period began. C'est
180 * But first, note that the current CPU must still be
183 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
184 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
185 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
186 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
187 rnp->gp_tasks = &t->rcu_node_entry;
188 #ifdef CONFIG_RCU_BOOST
189 if (rnp->boost_tasks != NULL)
190 rnp->boost_tasks = rnp->gp_tasks;
191 #endif /* #ifdef CONFIG_RCU_BOOST */
193 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
194 if (rnp->qsmask & rdp->grpmask)
195 rnp->gp_tasks = &t->rcu_node_entry;
197 trace_rcu_preempt_task(rdp->rsp->name,
199 (rnp->qsmask & rdp->grpmask)
202 raw_spin_unlock_irqrestore(&rnp->lock, flags);
203 } else if (t->rcu_read_lock_nesting < 0 &&
204 t->rcu_read_unlock_special.s) {
207 * Complete exit from RCU read-side critical section on
208 * behalf of preempted instance of __rcu_read_unlock().
210 rcu_read_unlock_special(t);
214 * Either we were not in an RCU read-side critical section to
215 * begin with, or we have now recorded that critical section
216 * globally. Either way, we can now note a quiescent state
217 * for this CPU. Again, if we were in an RCU read-side critical
218 * section, and if that critical section was blocking the current
219 * grace period, then the fact that the task has been enqueued
220 * means that we continue to block the current grace period.
226 * Check for preempted RCU readers blocking the current grace period
227 * for the specified rcu_node structure. If the caller needs a reliable
228 * answer, it must hold the rcu_node's ->lock.
230 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
232 return rnp->gp_tasks != NULL;
236 * Record a quiescent state for all tasks that were previously queued
237 * on the specified rcu_node structure and that were blocking the current
238 * RCU grace period. The caller must hold the specified rnp->lock with
239 * irqs disabled, and this lock is released upon return, but irqs remain
242 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
243 __releases(rnp->lock)
246 struct rcu_node *rnp_p;
248 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
249 raw_spin_unlock_irqrestore(&rnp->lock, flags);
250 return; /* Still need more quiescent states! */
256 * Either there is only one rcu_node in the tree,
257 * or tasks were kicked up to root rcu_node due to
258 * CPUs going offline.
260 rcu_report_qs_rsp(&rcu_preempt_state, flags);
264 /* Report up the rest of the hierarchy. */
266 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
267 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
268 smp_mb__after_unlock_lock();
269 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
273 * Advance a ->blkd_tasks-list pointer to the next entry, instead
274 * returning NULL if at the end of the list.
276 static struct list_head *rcu_next_node_entry(struct task_struct *t,
277 struct rcu_node *rnp)
279 struct list_head *np;
281 np = t->rcu_node_entry.next;
282 if (np == &rnp->blkd_tasks)
288 * Return true if the specified rcu_node structure has tasks that were
289 * preempted within an RCU read-side critical section.
291 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
293 return !list_empty(&rnp->blkd_tasks);
297 * Handle special cases during rcu_read_unlock(), such as needing to
298 * notify RCU core processing or task having blocked during the RCU
299 * read-side critical section.
301 void rcu_read_unlock_special(struct task_struct *t)
308 struct list_head *np;
309 #ifdef CONFIG_RCU_BOOST
310 bool drop_boost_mutex = false;
311 #endif /* #ifdef CONFIG_RCU_BOOST */
312 struct rcu_node *rnp;
313 union rcu_special special;
315 /* NMI handlers cannot block and cannot safely manipulate state. */
319 local_irq_save(flags);
322 * If RCU core is waiting for this CPU to exit critical section,
323 * let it know that we have done so. Because irqs are disabled,
324 * t->rcu_read_unlock_special cannot change.
326 special = t->rcu_read_unlock_special;
327 if (special.b.need_qs) {
329 t->rcu_read_unlock_special.b.need_qs = false;
330 if (!t->rcu_read_unlock_special.s) {
331 local_irq_restore(flags);
336 /* Hardware IRQ handlers cannot block, complain if they get here. */
337 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
338 local_irq_restore(flags);
342 /* Clean up if blocked during RCU read-side critical section. */
343 if (special.b.blocked) {
344 t->rcu_read_unlock_special.b.blocked = false;
347 * Remove this task from the list it blocked on. The
348 * task can migrate while we acquire the lock, but at
349 * most one time. So at most two passes through loop.
352 rnp = t->rcu_blocked_node;
353 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
354 smp_mb__after_unlock_lock();
355 if (rnp == t->rcu_blocked_node)
357 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
359 empty = !rcu_preempt_has_tasks(rnp);
360 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
361 empty_exp = !rcu_preempted_readers_exp(rnp);
362 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
363 np = rcu_next_node_entry(t, rnp);
364 list_del_init(&t->rcu_node_entry);
365 t->rcu_blocked_node = NULL;
366 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
368 if (&t->rcu_node_entry == rnp->gp_tasks)
370 if (&t->rcu_node_entry == rnp->exp_tasks)
372 #ifdef CONFIG_RCU_BOOST
373 if (&t->rcu_node_entry == rnp->boost_tasks)
374 rnp->boost_tasks = np;
375 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
376 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
377 #endif /* #ifdef CONFIG_RCU_BOOST */
380 * If this was the last task on the list, go see if we
381 * need to propagate ->qsmaskinit bit clearing up the
384 if (!empty && !rcu_preempt_has_tasks(rnp))
385 rcu_cleanup_dead_rnp(rnp);
388 * If this was the last task on the current list, and if
389 * we aren't waiting on any CPUs, report the quiescent state.
390 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
391 * so we must take a snapshot of the expedited state.
393 empty_exp_now = !rcu_preempted_readers_exp(rnp);
394 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
395 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
402 rcu_report_unblock_qs_rnp(rnp, flags);
404 raw_spin_unlock_irqrestore(&rnp->lock, flags);
407 #ifdef CONFIG_RCU_BOOST
408 /* Unboost if we were boosted. */
409 if (drop_boost_mutex)
410 rt_mutex_unlock(&rnp->boost_mtx);
411 #endif /* #ifdef CONFIG_RCU_BOOST */
414 * If this was the last task on the expedited lists,
415 * then we need to report up the rcu_node hierarchy.
417 if (!empty_exp && empty_exp_now)
418 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
420 local_irq_restore(flags);
425 * Dump detailed information for all tasks blocking the current RCU
426 * grace period on the specified rcu_node structure.
428 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
431 struct task_struct *t;
433 raw_spin_lock_irqsave(&rnp->lock, flags);
434 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
435 raw_spin_unlock_irqrestore(&rnp->lock, flags);
438 t = list_entry(rnp->gp_tasks,
439 struct task_struct, rcu_node_entry);
440 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
442 raw_spin_unlock_irqrestore(&rnp->lock, flags);
446 * Dump detailed information for all tasks blocking the current RCU
449 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
451 struct rcu_node *rnp = rcu_get_root(rsp);
453 rcu_print_detail_task_stall_rnp(rnp);
454 rcu_for_each_leaf_node(rsp, rnp)
455 rcu_print_detail_task_stall_rnp(rnp);
458 #ifdef CONFIG_RCU_CPU_STALL_INFO
460 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
462 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
463 rnp->level, rnp->grplo, rnp->grphi);
466 static void rcu_print_task_stall_end(void)
471 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
473 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
477 static void rcu_print_task_stall_end(void)
481 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
484 * Scan the current list of tasks blocked within RCU read-side critical
485 * sections, printing out the tid of each.
487 static int rcu_print_task_stall(struct rcu_node *rnp)
489 struct task_struct *t;
492 if (!rcu_preempt_blocked_readers_cgp(rnp))
494 rcu_print_task_stall_begin(rnp);
495 t = list_entry(rnp->gp_tasks,
496 struct task_struct, rcu_node_entry);
497 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
498 pr_cont(" P%d", t->pid);
501 rcu_print_task_stall_end();
506 * Check that the list of blocked tasks for the newly completed grace
507 * period is in fact empty. It is a serious bug to complete a grace
508 * period that still has RCU readers blocked! This function must be
509 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
510 * must be held by the caller.
512 * Also, if there are blocked tasks on the list, they automatically
513 * block the newly created grace period, so set up ->gp_tasks accordingly.
515 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
517 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
518 if (rcu_preempt_has_tasks(rnp))
519 rnp->gp_tasks = rnp->blkd_tasks.next;
520 WARN_ON_ONCE(rnp->qsmask);
524 * Check for a quiescent state from the current CPU. When a task blocks,
525 * the task is recorded in the corresponding CPU's rcu_node structure,
526 * which is checked elsewhere.
528 * Caller must disable hard irqs.
530 static void rcu_preempt_check_callbacks(void)
532 struct task_struct *t = current;
534 if (t->rcu_read_lock_nesting == 0) {
538 if (t->rcu_read_lock_nesting > 0 &&
539 __this_cpu_read(rcu_preempt_data.qs_pending) &&
540 !__this_cpu_read(rcu_preempt_data.passed_quiesce))
541 t->rcu_read_unlock_special.b.need_qs = true;
544 #ifdef CONFIG_RCU_BOOST
546 static void rcu_preempt_do_callbacks(void)
548 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
551 #endif /* #ifdef CONFIG_RCU_BOOST */
554 * Queue a preemptible-RCU callback for invocation after a grace period.
556 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
558 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
560 EXPORT_SYMBOL_GPL(call_rcu);
563 * synchronize_rcu - wait until a grace period has elapsed.
565 * Control will return to the caller some time after a full grace
566 * period has elapsed, in other words after all currently executing RCU
567 * read-side critical sections have completed. Note, however, that
568 * upon return from synchronize_rcu(), the caller might well be executing
569 * concurrently with new RCU read-side critical sections that began while
570 * synchronize_rcu() was waiting. RCU read-side critical sections are
571 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
573 * See the description of synchronize_sched() for more detailed information
574 * on memory ordering guarantees.
576 void synchronize_rcu(void)
578 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
579 !lock_is_held(&rcu_lock_map) &&
580 !lock_is_held(&rcu_sched_lock_map),
581 "Illegal synchronize_rcu() in RCU read-side critical section");
582 if (!rcu_scheduler_active)
585 synchronize_rcu_expedited();
587 wait_rcu_gp(call_rcu);
589 EXPORT_SYMBOL_GPL(synchronize_rcu);
591 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
592 static unsigned long sync_rcu_preempt_exp_count;
593 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
596 * Return non-zero if there are any tasks in RCU read-side critical
597 * sections blocking the current preemptible-RCU expedited grace period.
598 * If there is no preemptible-RCU expedited grace period currently in
599 * progress, returns zero unconditionally.
601 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
603 return rnp->exp_tasks != NULL;
607 * return non-zero if there is no RCU expedited grace period in progress
608 * for the specified rcu_node structure, in other words, if all CPUs and
609 * tasks covered by the specified rcu_node structure have done their bit
610 * for the current expedited grace period. Works only for preemptible
611 * RCU -- other RCU implementation use other means.
613 * Caller must hold sync_rcu_preempt_exp_mutex.
615 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
617 return !rcu_preempted_readers_exp(rnp) &&
618 ACCESS_ONCE(rnp->expmask) == 0;
622 * Report the exit from RCU read-side critical section for the last task
623 * that queued itself during or before the current expedited preemptible-RCU
624 * grace period. This event is reported either to the rcu_node structure on
625 * which the task was queued or to one of that rcu_node structure's ancestors,
626 * recursively up the tree. (Calm down, calm down, we do the recursion
629 * Caller must hold sync_rcu_preempt_exp_mutex.
631 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
637 raw_spin_lock_irqsave(&rnp->lock, flags);
638 smp_mb__after_unlock_lock();
640 if (!sync_rcu_preempt_exp_done(rnp)) {
641 raw_spin_unlock_irqrestore(&rnp->lock, flags);
644 if (rnp->parent == NULL) {
645 raw_spin_unlock_irqrestore(&rnp->lock, flags);
647 smp_mb(); /* EGP done before wake_up(). */
648 wake_up(&sync_rcu_preempt_exp_wq);
653 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
655 raw_spin_lock(&rnp->lock); /* irqs already disabled */
656 smp_mb__after_unlock_lock();
657 rnp->expmask &= ~mask;
662 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
663 * grace period for the specified rcu_node structure, phase 1. If there
664 * are such tasks, set the ->expmask bits up the rcu_node tree and also
665 * set the ->expmask bits on the leaf rcu_node structures to tell phase 2
666 * that work is needed here.
668 * Caller must hold sync_rcu_preempt_exp_mutex.
671 sync_rcu_preempt_exp_init1(struct rcu_state *rsp, struct rcu_node *rnp)
675 struct rcu_node *rnp_up;
677 raw_spin_lock_irqsave(&rnp->lock, flags);
678 smp_mb__after_unlock_lock();
679 WARN_ON_ONCE(rnp->expmask);
680 WARN_ON_ONCE(rnp->exp_tasks);
681 if (!rcu_preempt_has_tasks(rnp)) {
682 /* No blocked tasks, nothing to do. */
683 raw_spin_unlock_irqrestore(&rnp->lock, flags);
686 /* Call for Phase 2 and propagate ->expmask bits up the tree. */
689 while (rnp_up->parent) {
690 mask = rnp_up->grpmask;
691 rnp_up = rnp_up->parent;
692 if (rnp_up->expmask & mask)
694 raw_spin_lock(&rnp_up->lock); /* irqs already off */
695 smp_mb__after_unlock_lock();
696 rnp_up->expmask |= mask;
697 raw_spin_unlock(&rnp_up->lock); /* irqs still off */
699 raw_spin_unlock_irqrestore(&rnp->lock, flags);
703 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
704 * grace period for the specified rcu_node structure, phase 2. If the
705 * leaf rcu_node structure has its ->expmask field set, check for tasks.
706 * If there are some, clear ->expmask and set ->exp_tasks accordingly,
707 * then initiate RCU priority boosting. Otherwise, clear ->expmask and
708 * invoke rcu_report_exp_rnp() to clear out the upper-level ->expmask bits,
709 * enabling rcu_read_unlock_special() to do the bit-clearing.
711 * Caller must hold sync_rcu_preempt_exp_mutex.
714 sync_rcu_preempt_exp_init2(struct rcu_state *rsp, struct rcu_node *rnp)
718 raw_spin_lock_irqsave(&rnp->lock, flags);
719 smp_mb__after_unlock_lock();
721 /* Phase 1 didn't do anything, so Phase 2 doesn't either. */
722 raw_spin_unlock_irqrestore(&rnp->lock, flags);
726 /* Phase 1 is over. */
730 * If there are still blocked tasks, set up ->exp_tasks so that
731 * rcu_read_unlock_special() will wake us and then boost them.
733 if (rcu_preempt_has_tasks(rnp)) {
734 rnp->exp_tasks = rnp->blkd_tasks.next;
735 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
739 /* No longer any blocked tasks, so undo bit setting. */
740 raw_spin_unlock_irqrestore(&rnp->lock, flags);
741 rcu_report_exp_rnp(rsp, rnp, false);
745 * synchronize_rcu_expedited - Brute-force RCU grace period
747 * Wait for an RCU-preempt grace period, but expedite it. The basic
748 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
749 * the ->blkd_tasks lists and wait for this list to drain. This consumes
750 * significant time on all CPUs and is unfriendly to real-time workloads,
751 * so is thus not recommended for any sort of common-case code.
752 * In fact, if you are using synchronize_rcu_expedited() in a loop,
753 * please restructure your code to batch your updates, and then Use a
754 * single synchronize_rcu() instead.
756 void synchronize_rcu_expedited(void)
758 struct rcu_node *rnp;
759 struct rcu_state *rsp = &rcu_preempt_state;
763 smp_mb(); /* Caller's modifications seen first by other CPUs. */
764 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
765 smp_mb(); /* Above access cannot bleed into critical section. */
768 * Block CPU-hotplug operations. This means that any CPU-hotplug
769 * operation that finds an rcu_node structure with tasks in the
770 * process of being boosted will know that all tasks blocking
771 * this expedited grace period will already be in the process of
772 * being boosted. This simplifies the process of moving tasks
773 * from leaf to root rcu_node structures.
775 if (!try_get_online_cpus()) {
776 /* CPU-hotplug operation in flight, fall back to normal GP. */
777 wait_rcu_gp(call_rcu);
782 * Acquire lock, falling back to synchronize_rcu() if too many
783 * lock-acquisition failures. Of course, if someone does the
784 * expedited grace period for us, just leave.
786 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
787 if (ULONG_CMP_LT(snap,
788 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
790 goto mb_ret; /* Others did our work for us. */
792 if (trycount++ < 10) {
793 udelay(trycount * num_online_cpus());
796 wait_rcu_gp(call_rcu);
800 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
802 goto unlock_mb_ret; /* Others did our work for us. */
805 /* force all RCU readers onto ->blkd_tasks lists. */
806 synchronize_sched_expedited();
809 * Snapshot current state of ->blkd_tasks lists into ->expmask.
810 * Phase 1 sets bits and phase 2 permits rcu_read_unlock_special()
811 * to start clearing them. Doing this in one phase leads to
812 * strange races between setting and clearing bits, so just say "no"!
814 rcu_for_each_leaf_node(rsp, rnp)
815 sync_rcu_preempt_exp_init1(rsp, rnp);
816 rcu_for_each_leaf_node(rsp, rnp)
817 sync_rcu_preempt_exp_init2(rsp, rnp);
821 /* Wait for snapshotted ->blkd_tasks lists to drain. */
822 rnp = rcu_get_root(rsp);
823 wait_event(sync_rcu_preempt_exp_wq,
824 sync_rcu_preempt_exp_done(rnp));
826 /* Clean up and exit. */
827 smp_mb(); /* ensure expedited GP seen before counter increment. */
828 ACCESS_ONCE(sync_rcu_preempt_exp_count) =
829 sync_rcu_preempt_exp_count + 1;
831 mutex_unlock(&sync_rcu_preempt_exp_mutex);
833 smp_mb(); /* ensure subsequent action seen after grace period. */
835 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
838 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
840 * Note that this primitive does not necessarily wait for an RCU grace period
841 * to complete. For example, if there are no RCU callbacks queued anywhere
842 * in the system, then rcu_barrier() is within its rights to return
843 * immediately, without waiting for anything, much less an RCU grace period.
845 void rcu_barrier(void)
847 _rcu_barrier(&rcu_preempt_state);
849 EXPORT_SYMBOL_GPL(rcu_barrier);
852 * Initialize preemptible RCU's state structures.
854 static void __init __rcu_init_preempt(void)
856 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
860 * Check for a task exiting while in a preemptible-RCU read-side
861 * critical section, clean up if so. No need to issue warnings,
862 * as debug_check_no_locks_held() already does this if lockdep
867 struct task_struct *t = current;
869 if (likely(list_empty(¤t->rcu_node_entry)))
871 t->rcu_read_lock_nesting = 1;
873 t->rcu_read_unlock_special.b.blocked = true;
877 #else /* #ifdef CONFIG_PREEMPT_RCU */
879 static struct rcu_state *rcu_state_p = &rcu_sched_state;
882 * Tell them what RCU they are running.
884 static void __init rcu_bootup_announce(void)
886 pr_info("Hierarchical RCU implementation.\n");
887 rcu_bootup_announce_oddness();
891 * Because preemptible RCU does not exist, we never have to check for
892 * CPUs being in quiescent states.
894 static void rcu_preempt_note_context_switch(void)
899 * Because preemptible RCU does not exist, there are never any preempted
902 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
907 #ifdef CONFIG_HOTPLUG_CPU
910 * Because there is no preemptible RCU, there can be no readers blocked.
912 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
917 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
920 * Because preemptible RCU does not exist, we never have to check for
921 * tasks blocked within RCU read-side critical sections.
923 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
928 * Because preemptible RCU does not exist, we never have to check for
929 * tasks blocked within RCU read-side critical sections.
931 static int rcu_print_task_stall(struct rcu_node *rnp)
937 * Because there is no preemptible RCU, there can be no readers blocked,
938 * so there is no need to check for blocked tasks. So check only for
939 * bogus qsmask values.
941 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
943 WARN_ON_ONCE(rnp->qsmask);
947 * Because preemptible RCU does not exist, it never has any callbacks
950 static void rcu_preempt_check_callbacks(void)
955 * Wait for an rcu-preempt grace period, but make it happen quickly.
956 * But because preemptible RCU does not exist, map to rcu-sched.
958 void synchronize_rcu_expedited(void)
960 synchronize_sched_expedited();
962 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
965 * Because preemptible RCU does not exist, rcu_barrier() is just
966 * another name for rcu_barrier_sched().
968 void rcu_barrier(void)
972 EXPORT_SYMBOL_GPL(rcu_barrier);
975 * Because preemptible RCU does not exist, it need not be initialized.
977 static void __init __rcu_init_preempt(void)
982 * Because preemptible RCU does not exist, tasks cannot possibly exit
983 * while in preemptible RCU read-side critical sections.
989 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
991 #ifdef CONFIG_RCU_BOOST
993 #include "../locking/rtmutex_common.h"
995 #ifdef CONFIG_RCU_TRACE
997 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
999 if (!rcu_preempt_has_tasks(rnp))
1000 rnp->n_balk_blkd_tasks++;
1001 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1002 rnp->n_balk_exp_gp_tasks++;
1003 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1004 rnp->n_balk_boost_tasks++;
1005 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1006 rnp->n_balk_notblocked++;
1007 else if (rnp->gp_tasks != NULL &&
1008 ULONG_CMP_LT(jiffies, rnp->boost_time))
1009 rnp->n_balk_notyet++;
1014 #else /* #ifdef CONFIG_RCU_TRACE */
1016 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1020 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1022 static void rcu_wake_cond(struct task_struct *t, int status)
1025 * If the thread is yielding, only wake it when this
1026 * is invoked from idle
1028 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1033 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1034 * or ->boost_tasks, advancing the pointer to the next task in the
1035 * ->blkd_tasks list.
1037 * Note that irqs must be enabled: boosting the task can block.
1038 * Returns 1 if there are more tasks needing to be boosted.
1040 static int rcu_boost(struct rcu_node *rnp)
1042 unsigned long flags;
1043 struct task_struct *t;
1044 struct list_head *tb;
1046 if (ACCESS_ONCE(rnp->exp_tasks) == NULL &&
1047 ACCESS_ONCE(rnp->boost_tasks) == NULL)
1048 return 0; /* Nothing left to boost. */
1050 raw_spin_lock_irqsave(&rnp->lock, flags);
1051 smp_mb__after_unlock_lock();
1054 * Recheck under the lock: all tasks in need of boosting
1055 * might exit their RCU read-side critical sections on their own.
1057 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1058 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1063 * Preferentially boost tasks blocking expedited grace periods.
1064 * This cannot starve the normal grace periods because a second
1065 * expedited grace period must boost all blocked tasks, including
1066 * those blocking the pre-existing normal grace period.
1068 if (rnp->exp_tasks != NULL) {
1069 tb = rnp->exp_tasks;
1070 rnp->n_exp_boosts++;
1072 tb = rnp->boost_tasks;
1073 rnp->n_normal_boosts++;
1075 rnp->n_tasks_boosted++;
1078 * We boost task t by manufacturing an rt_mutex that appears to
1079 * be held by task t. We leave a pointer to that rt_mutex where
1080 * task t can find it, and task t will release the mutex when it
1081 * exits its outermost RCU read-side critical section. Then
1082 * simply acquiring this artificial rt_mutex will boost task
1083 * t's priority. (Thanks to tglx for suggesting this approach!)
1085 * Note that task t must acquire rnp->lock to remove itself from
1086 * the ->blkd_tasks list, which it will do from exit() if from
1087 * nowhere else. We therefore are guaranteed that task t will
1088 * stay around at least until we drop rnp->lock. Note that
1089 * rnp->lock also resolves races between our priority boosting
1090 * and task t's exiting its outermost RCU read-side critical
1093 t = container_of(tb, struct task_struct, rcu_node_entry);
1094 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1095 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1096 /* Lock only for side effect: boosts task t's priority. */
1097 rt_mutex_lock(&rnp->boost_mtx);
1098 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1100 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1101 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1105 * Priority-boosting kthread. One per leaf rcu_node and one for the
1108 static int rcu_boost_kthread(void *arg)
1110 struct rcu_node *rnp = (struct rcu_node *)arg;
1114 trace_rcu_utilization(TPS("Start boost kthread@init"));
1116 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1117 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1118 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1119 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1120 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1121 more2boost = rcu_boost(rnp);
1127 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1128 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1129 schedule_timeout_interruptible(2);
1130 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1135 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1140 * Check to see if it is time to start boosting RCU readers that are
1141 * blocking the current grace period, and, if so, tell the per-rcu_node
1142 * kthread to start boosting them. If there is an expedited grace
1143 * period in progress, it is always time to boost.
1145 * The caller must hold rnp->lock, which this function releases.
1146 * The ->boost_kthread_task is immortal, so we don't need to worry
1147 * about it going away.
1149 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1150 __releases(rnp->lock)
1152 struct task_struct *t;
1154 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1155 rnp->n_balk_exp_gp_tasks++;
1156 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1159 if (rnp->exp_tasks != NULL ||
1160 (rnp->gp_tasks != NULL &&
1161 rnp->boost_tasks == NULL &&
1163 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1164 if (rnp->exp_tasks == NULL)
1165 rnp->boost_tasks = rnp->gp_tasks;
1166 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1167 t = rnp->boost_kthread_task;
1169 rcu_wake_cond(t, rnp->boost_kthread_status);
1171 rcu_initiate_boost_trace(rnp);
1172 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1177 * Wake up the per-CPU kthread to invoke RCU callbacks.
1179 static void invoke_rcu_callbacks_kthread(void)
1181 unsigned long flags;
1183 local_irq_save(flags);
1184 __this_cpu_write(rcu_cpu_has_work, 1);
1185 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1186 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1187 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1188 __this_cpu_read(rcu_cpu_kthread_status));
1190 local_irq_restore(flags);
1194 * Is the current CPU running the RCU-callbacks kthread?
1195 * Caller must have preemption disabled.
1197 static bool rcu_is_callbacks_kthread(void)
1199 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1202 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1205 * Do priority-boost accounting for the start of a new grace period.
1207 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1209 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1213 * Create an RCU-boost kthread for the specified node if one does not
1214 * already exist. We only create this kthread for preemptible RCU.
1215 * Returns zero if all is well, a negated errno otherwise.
1217 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1218 struct rcu_node *rnp)
1220 int rnp_index = rnp - &rsp->node[0];
1221 unsigned long flags;
1222 struct sched_param sp;
1223 struct task_struct *t;
1225 if (&rcu_preempt_state != rsp)
1228 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1232 if (rnp->boost_kthread_task != NULL)
1234 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1235 "rcub/%d", rnp_index);
1238 raw_spin_lock_irqsave(&rnp->lock, flags);
1239 smp_mb__after_unlock_lock();
1240 rnp->boost_kthread_task = t;
1241 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1242 sp.sched_priority = kthread_prio;
1243 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1244 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1248 static void rcu_kthread_do_work(void)
1250 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1251 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1252 rcu_preempt_do_callbacks();
1255 static void rcu_cpu_kthread_setup(unsigned int cpu)
1257 struct sched_param sp;
1259 sp.sched_priority = kthread_prio;
1260 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1263 static void rcu_cpu_kthread_park(unsigned int cpu)
1265 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1268 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1270 return __this_cpu_read(rcu_cpu_has_work);
1274 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1275 * RCU softirq used in flavors and configurations of RCU that do not
1276 * support RCU priority boosting.
1278 static void rcu_cpu_kthread(unsigned int cpu)
1280 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1281 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1284 for (spincnt = 0; spincnt < 10; spincnt++) {
1285 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1287 *statusp = RCU_KTHREAD_RUNNING;
1288 this_cpu_inc(rcu_cpu_kthread_loops);
1289 local_irq_disable();
1294 rcu_kthread_do_work();
1297 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1298 *statusp = RCU_KTHREAD_WAITING;
1302 *statusp = RCU_KTHREAD_YIELDING;
1303 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1304 schedule_timeout_interruptible(2);
1305 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1306 *statusp = RCU_KTHREAD_WAITING;
1310 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1311 * served by the rcu_node in question. The CPU hotplug lock is still
1312 * held, so the value of rnp->qsmaskinit will be stable.
1314 * We don't include outgoingcpu in the affinity set, use -1 if there is
1315 * no outgoing CPU. If there are no CPUs left in the affinity set,
1316 * this function allows the kthread to execute on any CPU.
1318 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1320 struct task_struct *t = rnp->boost_kthread_task;
1321 unsigned long mask = rnp->qsmaskinit;
1327 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1329 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1330 if ((mask & 0x1) && cpu != outgoingcpu)
1331 cpumask_set_cpu(cpu, cm);
1332 if (cpumask_weight(cm) == 0)
1334 set_cpus_allowed_ptr(t, cm);
1335 free_cpumask_var(cm);
1338 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1339 .store = &rcu_cpu_kthread_task,
1340 .thread_should_run = rcu_cpu_kthread_should_run,
1341 .thread_fn = rcu_cpu_kthread,
1342 .thread_comm = "rcuc/%u",
1343 .setup = rcu_cpu_kthread_setup,
1344 .park = rcu_cpu_kthread_park,
1348 * Spawn boost kthreads -- called as soon as the scheduler is running.
1350 static void __init rcu_spawn_boost_kthreads(void)
1352 struct rcu_node *rnp;
1355 for_each_possible_cpu(cpu)
1356 per_cpu(rcu_cpu_has_work, cpu) = 0;
1357 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1358 rcu_for_each_leaf_node(rcu_state_p, rnp)
1359 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1362 static void rcu_prepare_kthreads(int cpu)
1364 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1365 struct rcu_node *rnp = rdp->mynode;
1367 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1368 if (rcu_scheduler_fully_active)
1369 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1372 #else /* #ifdef CONFIG_RCU_BOOST */
1374 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1375 __releases(rnp->lock)
1377 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1380 static void invoke_rcu_callbacks_kthread(void)
1385 static bool rcu_is_callbacks_kthread(void)
1390 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1394 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1398 static void __init rcu_spawn_boost_kthreads(void)
1402 static void rcu_prepare_kthreads(int cpu)
1406 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1408 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1411 * Check to see if any future RCU-related work will need to be done
1412 * by the current CPU, even if none need be done immediately, returning
1413 * 1 if so. This function is part of the RCU implementation; it is -not-
1414 * an exported member of the RCU API.
1416 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1417 * any flavor of RCU.
1419 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1420 int rcu_needs_cpu(unsigned long *delta_jiffies)
1422 *delta_jiffies = ULONG_MAX;
1423 return rcu_cpu_has_callbacks(NULL);
1425 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1428 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1431 static void rcu_cleanup_after_idle(void)
1436 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1439 static void rcu_prepare_for_idle(void)
1444 * Don't bother keeping a running count of the number of RCU callbacks
1445 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1447 static void rcu_idle_count_callbacks_posted(void)
1451 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1454 * This code is invoked when a CPU goes idle, at which point we want
1455 * to have the CPU do everything required for RCU so that it can enter
1456 * the energy-efficient dyntick-idle mode. This is handled by a
1457 * state machine implemented by rcu_prepare_for_idle() below.
1459 * The following three proprocessor symbols control this state machine:
1461 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1462 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1463 * is sized to be roughly one RCU grace period. Those energy-efficiency
1464 * benchmarkers who might otherwise be tempted to set this to a large
1465 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1466 * system. And if you are -that- concerned about energy efficiency,
1467 * just power the system down and be done with it!
1468 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1469 * permitted to sleep in dyntick-idle mode with only lazy RCU
1470 * callbacks pending. Setting this too high can OOM your system.
1472 * The values below work well in practice. If future workloads require
1473 * adjustment, they can be converted into kernel config parameters, though
1474 * making the state machine smarter might be a better option.
1476 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1477 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1479 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1480 module_param(rcu_idle_gp_delay, int, 0644);
1481 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1482 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1484 extern int tick_nohz_active;
1487 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1488 * only if it has been awhile since the last time we did so. Afterwards,
1489 * if there are any callbacks ready for immediate invocation, return true.
1491 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1493 bool cbs_ready = false;
1494 struct rcu_data *rdp;
1495 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1496 struct rcu_node *rnp;
1497 struct rcu_state *rsp;
1499 /* Exit early if we advanced recently. */
1500 if (jiffies == rdtp->last_advance_all)
1502 rdtp->last_advance_all = jiffies;
1504 for_each_rcu_flavor(rsp) {
1505 rdp = this_cpu_ptr(rsp->rda);
1509 * Don't bother checking unless a grace period has
1510 * completed since we last checked and there are
1511 * callbacks not yet ready to invoke.
1513 if ((rdp->completed != rnp->completed ||
1514 unlikely(ACCESS_ONCE(rdp->gpwrap))) &&
1515 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1516 note_gp_changes(rsp, rdp);
1518 if (cpu_has_callbacks_ready_to_invoke(rdp))
1525 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1526 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1527 * caller to set the timeout based on whether or not there are non-lazy
1530 * The caller must have disabled interrupts.
1532 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1533 int rcu_needs_cpu(unsigned long *dj)
1535 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1537 /* Snapshot to detect later posting of non-lazy callback. */
1538 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1540 /* If no callbacks, RCU doesn't need the CPU. */
1541 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1546 /* Attempt to advance callbacks. */
1547 if (rcu_try_advance_all_cbs()) {
1548 /* Some ready to invoke, so initiate later invocation. */
1552 rdtp->last_accelerate = jiffies;
1554 /* Request timer delay depending on laziness, and round. */
1555 if (!rdtp->all_lazy) {
1556 *dj = round_up(rcu_idle_gp_delay + jiffies,
1557 rcu_idle_gp_delay) - jiffies;
1559 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1563 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1566 * Prepare a CPU for idle from an RCU perspective. The first major task
1567 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1568 * The second major task is to check to see if a non-lazy callback has
1569 * arrived at a CPU that previously had only lazy callbacks. The third
1570 * major task is to accelerate (that is, assign grace-period numbers to)
1571 * any recently arrived callbacks.
1573 * The caller must have disabled interrupts.
1575 static void rcu_prepare_for_idle(void)
1577 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1579 struct rcu_data *rdp;
1580 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1581 struct rcu_node *rnp;
1582 struct rcu_state *rsp;
1585 /* Handle nohz enablement switches conservatively. */
1586 tne = ACCESS_ONCE(tick_nohz_active);
1587 if (tne != rdtp->tick_nohz_enabled_snap) {
1588 if (rcu_cpu_has_callbacks(NULL))
1589 invoke_rcu_core(); /* force nohz to see update. */
1590 rdtp->tick_nohz_enabled_snap = tne;
1596 /* If this is a no-CBs CPU, no callbacks, just return. */
1597 if (rcu_is_nocb_cpu(smp_processor_id()))
1601 * If a non-lazy callback arrived at a CPU having only lazy
1602 * callbacks, invoke RCU core for the side-effect of recalculating
1603 * idle duration on re-entry to idle.
1605 if (rdtp->all_lazy &&
1606 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1607 rdtp->all_lazy = false;
1608 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1614 * If we have not yet accelerated this jiffy, accelerate all
1615 * callbacks on this CPU.
1617 if (rdtp->last_accelerate == jiffies)
1619 rdtp->last_accelerate = jiffies;
1620 for_each_rcu_flavor(rsp) {
1621 rdp = this_cpu_ptr(rsp->rda);
1622 if (!*rdp->nxttail[RCU_DONE_TAIL])
1625 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1626 smp_mb__after_unlock_lock();
1627 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1628 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1630 rcu_gp_kthread_wake(rsp);
1632 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1636 * Clean up for exit from idle. Attempt to advance callbacks based on
1637 * any grace periods that elapsed while the CPU was idle, and if any
1638 * callbacks are now ready to invoke, initiate invocation.
1640 static void rcu_cleanup_after_idle(void)
1642 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1643 if (rcu_is_nocb_cpu(smp_processor_id()))
1645 if (rcu_try_advance_all_cbs())
1647 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1651 * Keep a running count of the number of non-lazy callbacks posted
1652 * on this CPU. This running counter (which is never decremented) allows
1653 * rcu_prepare_for_idle() to detect when something out of the idle loop
1654 * posts a callback, even if an equal number of callbacks are invoked.
1655 * Of course, callbacks should only be posted from within a trace event
1656 * designed to be called from idle or from within RCU_NONIDLE().
1658 static void rcu_idle_count_callbacks_posted(void)
1660 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1664 * Data for flushing lazy RCU callbacks at OOM time.
1666 static atomic_t oom_callback_count;
1667 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1670 * RCU OOM callback -- decrement the outstanding count and deliver the
1671 * wake-up if we are the last one.
1673 static void rcu_oom_callback(struct rcu_head *rhp)
1675 if (atomic_dec_and_test(&oom_callback_count))
1676 wake_up(&oom_callback_wq);
1680 * Post an rcu_oom_notify callback on the current CPU if it has at
1681 * least one lazy callback. This will unnecessarily post callbacks
1682 * to CPUs that already have a non-lazy callback at the end of their
1683 * callback list, but this is an infrequent operation, so accept some
1684 * extra overhead to keep things simple.
1686 static void rcu_oom_notify_cpu(void *unused)
1688 struct rcu_state *rsp;
1689 struct rcu_data *rdp;
1691 for_each_rcu_flavor(rsp) {
1692 rdp = raw_cpu_ptr(rsp->rda);
1693 if (rdp->qlen_lazy != 0) {
1694 atomic_inc(&oom_callback_count);
1695 rsp->call(&rdp->oom_head, rcu_oom_callback);
1701 * If low on memory, ensure that each CPU has a non-lazy callback.
1702 * This will wake up CPUs that have only lazy callbacks, in turn
1703 * ensuring that they free up the corresponding memory in a timely manner.
1704 * Because an uncertain amount of memory will be freed in some uncertain
1705 * timeframe, we do not claim to have freed anything.
1707 static int rcu_oom_notify(struct notifier_block *self,
1708 unsigned long notused, void *nfreed)
1712 /* Wait for callbacks from earlier instance to complete. */
1713 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1714 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1717 * Prevent premature wakeup: ensure that all increments happen
1718 * before there is a chance of the counter reaching zero.
1720 atomic_set(&oom_callback_count, 1);
1723 for_each_online_cpu(cpu) {
1724 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1725 cond_resched_rcu_qs();
1729 /* Unconditionally decrement: no need to wake ourselves up. */
1730 atomic_dec(&oom_callback_count);
1735 static struct notifier_block rcu_oom_nb = {
1736 .notifier_call = rcu_oom_notify
1739 static int __init rcu_register_oom_notifier(void)
1741 register_oom_notifier(&rcu_oom_nb);
1744 early_initcall(rcu_register_oom_notifier);
1746 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1748 #ifdef CONFIG_RCU_CPU_STALL_INFO
1750 #ifdef CONFIG_RCU_FAST_NO_HZ
1752 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1754 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1755 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1757 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1758 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1760 rdtp->all_lazy ? 'L' : '.',
1761 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1764 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1766 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1771 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1773 /* Initiate the stall-info list. */
1774 static void print_cpu_stall_info_begin(void)
1780 * Print out diagnostic information for the specified stalled CPU.
1782 * If the specified CPU is aware of the current RCU grace period
1783 * (flavor specified by rsp), then print the number of scheduling
1784 * clock interrupts the CPU has taken during the time that it has
1785 * been aware. Otherwise, print the number of RCU grace periods
1786 * that this CPU is ignorant of, for example, "1" if the CPU was
1787 * aware of the previous grace period.
1789 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1791 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1793 char fast_no_hz[72];
1794 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1795 struct rcu_dynticks *rdtp = rdp->dynticks;
1797 unsigned long ticks_value;
1799 if (rsp->gpnum == rdp->gpnum) {
1800 ticks_title = "ticks this GP";
1801 ticks_value = rdp->ticks_this_gp;
1803 ticks_title = "GPs behind";
1804 ticks_value = rsp->gpnum - rdp->gpnum;
1806 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1807 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1808 cpu, ticks_value, ticks_title,
1809 atomic_read(&rdtp->dynticks) & 0xfff,
1810 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1811 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1812 ACCESS_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1816 /* Terminate the stall-info list. */
1817 static void print_cpu_stall_info_end(void)
1822 /* Zero ->ticks_this_gp for all flavors of RCU. */
1823 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1825 rdp->ticks_this_gp = 0;
1826 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1829 /* Increment ->ticks_this_gp for all flavors of RCU. */
1830 static void increment_cpu_stall_ticks(void)
1832 struct rcu_state *rsp;
1834 for_each_rcu_flavor(rsp)
1835 raw_cpu_inc(rsp->rda->ticks_this_gp);
1838 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1840 static void print_cpu_stall_info_begin(void)
1845 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1847 pr_cont(" %d", cpu);
1850 static void print_cpu_stall_info_end(void)
1855 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1859 static void increment_cpu_stall_ticks(void)
1863 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1865 #ifdef CONFIG_RCU_NOCB_CPU
1868 * Offload callback processing from the boot-time-specified set of CPUs
1869 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1870 * kthread created that pulls the callbacks from the corresponding CPU,
1871 * waits for a grace period to elapse, and invokes the callbacks.
1872 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1873 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1874 * has been specified, in which case each kthread actively polls its
1875 * CPU. (Which isn't so great for energy efficiency, but which does
1876 * reduce RCU's overhead on that CPU.)
1878 * This is intended to be used in conjunction with Frederic Weisbecker's
1879 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1880 * running CPU-bound user-mode computations.
1882 * Offloading of callback processing could also in theory be used as
1883 * an energy-efficiency measure because CPUs with no RCU callbacks
1884 * queued are more aggressive about entering dyntick-idle mode.
1888 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1889 static int __init rcu_nocb_setup(char *str)
1891 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1892 have_rcu_nocb_mask = true;
1893 cpulist_parse(str, rcu_nocb_mask);
1896 __setup("rcu_nocbs=", rcu_nocb_setup);
1898 static int __init parse_rcu_nocb_poll(char *arg)
1903 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1906 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1909 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1911 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1915 * Set the root rcu_node structure's ->need_future_gp field
1916 * based on the sum of those of all rcu_node structures. This does
1917 * double-count the root rcu_node structure's requests, but this
1918 * is necessary to handle the possibility of a rcu_nocb_kthread()
1919 * having awakened during the time that the rcu_node structures
1920 * were being updated for the end of the previous grace period.
1922 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1924 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1927 static void rcu_init_one_nocb(struct rcu_node *rnp)
1929 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1930 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1933 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1934 /* Is the specified CPU a no-CBs CPU? */
1935 bool rcu_is_nocb_cpu(int cpu)
1937 if (have_rcu_nocb_mask)
1938 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1941 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1944 * Kick the leader kthread for this NOCB group.
1946 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1948 struct rcu_data *rdp_leader = rdp->nocb_leader;
1950 if (!ACCESS_ONCE(rdp_leader->nocb_kthread))
1952 if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1953 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1954 ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false;
1955 wake_up(&rdp_leader->nocb_wq);
1960 * Does the specified CPU need an RCU callback for the specified flavor
1963 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1965 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1967 #ifdef CONFIG_PROVE_RCU
1968 struct rcu_head *rhp;
1969 #endif /* #ifdef CONFIG_PROVE_RCU */
1972 * Check count of all no-CBs callbacks awaiting invocation.
1973 * There needs to be a barrier before this function is called,
1974 * but associated with a prior determination that no more
1975 * callbacks would be posted. In the worst case, the first
1976 * barrier in _rcu_barrier() suffices (but the caller cannot
1977 * necessarily rely on this, not a substitute for the caller
1978 * getting the concurrency design right!). There must also be
1979 * a barrier between the following load an posting of a callback
1980 * (if a callback is in fact needed). This is associated with an
1981 * atomic_inc() in the caller.
1983 ret = atomic_long_read(&rdp->nocb_q_count);
1985 #ifdef CONFIG_PROVE_RCU
1986 rhp = ACCESS_ONCE(rdp->nocb_head);
1988 rhp = ACCESS_ONCE(rdp->nocb_gp_head);
1990 rhp = ACCESS_ONCE(rdp->nocb_follower_head);
1992 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1993 if (!ACCESS_ONCE(rdp->nocb_kthread) && rhp) {
1994 /* RCU callback enqueued before CPU first came online??? */
1995 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1999 #endif /* #ifdef CONFIG_PROVE_RCU */
2005 * Enqueue the specified string of rcu_head structures onto the specified
2006 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2007 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2008 * counts are supplied by rhcount and rhcount_lazy.
2010 * If warranted, also wake up the kthread servicing this CPUs queues.
2012 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2013 struct rcu_head *rhp,
2014 struct rcu_head **rhtp,
2015 int rhcount, int rhcount_lazy,
2016 unsigned long flags)
2019 struct rcu_head **old_rhpp;
2020 struct task_struct *t;
2022 /* Enqueue the callback on the nocb list and update counts. */
2023 atomic_long_add(rhcount, &rdp->nocb_q_count);
2024 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
2025 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2026 ACCESS_ONCE(*old_rhpp) = rhp;
2027 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2028 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
2030 /* If we are not being polled and there is a kthread, awaken it ... */
2031 t = ACCESS_ONCE(rdp->nocb_kthread);
2032 if (rcu_nocb_poll || !t) {
2033 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2034 TPS("WakeNotPoll"));
2037 len = atomic_long_read(&rdp->nocb_q_count);
2038 if (old_rhpp == &rdp->nocb_head) {
2039 if (!irqs_disabled_flags(flags)) {
2040 /* ... if queue was empty ... */
2041 wake_nocb_leader(rdp, false);
2042 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2045 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
2046 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2047 TPS("WakeEmptyIsDeferred"));
2049 rdp->qlen_last_fqs_check = 0;
2050 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2051 /* ... or if many callbacks queued. */
2052 if (!irqs_disabled_flags(flags)) {
2053 wake_nocb_leader(rdp, true);
2054 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2057 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
2058 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2059 TPS("WakeOvfIsDeferred"));
2061 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2063 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2069 * This is a helper for __call_rcu(), which invokes this when the normal
2070 * callback queue is inoperable. If this is not a no-CBs CPU, this
2071 * function returns failure back to __call_rcu(), which can complain
2074 * Otherwise, this function queues the callback where the corresponding
2075 * "rcuo" kthread can find it.
2077 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2078 bool lazy, unsigned long flags)
2081 if (!rcu_is_nocb_cpu(rdp->cpu))
2083 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2084 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2085 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2086 (unsigned long)rhp->func,
2087 -atomic_long_read(&rdp->nocb_q_count_lazy),
2088 -atomic_long_read(&rdp->nocb_q_count));
2090 trace_rcu_callback(rdp->rsp->name, rhp,
2091 -atomic_long_read(&rdp->nocb_q_count_lazy),
2092 -atomic_long_read(&rdp->nocb_q_count));
2095 * If called from an extended quiescent state with interrupts
2096 * disabled, invoke the RCU core in order to allow the idle-entry
2097 * deferred-wakeup check to function.
2099 if (irqs_disabled_flags(flags) &&
2100 !rcu_is_watching() &&
2101 cpu_online(smp_processor_id()))
2108 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2111 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2112 struct rcu_data *rdp,
2113 unsigned long flags)
2115 long ql = rsp->qlen;
2116 long qll = rsp->qlen_lazy;
2118 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2119 if (!rcu_is_nocb_cpu(smp_processor_id()))
2124 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2125 if (rsp->orphan_donelist != NULL) {
2126 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2127 rsp->orphan_donetail, ql, qll, flags);
2129 rsp->orphan_donelist = NULL;
2130 rsp->orphan_donetail = &rsp->orphan_donelist;
2132 if (rsp->orphan_nxtlist != NULL) {
2133 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2134 rsp->orphan_nxttail, ql, qll, flags);
2136 rsp->orphan_nxtlist = NULL;
2137 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2143 * If necessary, kick off a new grace period, and either way wait
2144 * for a subsequent grace period to complete.
2146 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2150 unsigned long flags;
2152 struct rcu_node *rnp = rdp->mynode;
2154 raw_spin_lock_irqsave(&rnp->lock, flags);
2155 smp_mb__after_unlock_lock();
2156 needwake = rcu_start_future_gp(rnp, rdp, &c);
2157 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2159 rcu_gp_kthread_wake(rdp->rsp);
2162 * Wait for the grace period. Do so interruptibly to avoid messing
2163 * up the load average.
2165 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2167 wait_event_interruptible(
2168 rnp->nocb_gp_wq[c & 0x1],
2169 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2172 WARN_ON(signal_pending(current));
2173 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2175 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2176 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2180 * Leaders come here to wait for additional callbacks to show up.
2181 * This function does not return until callbacks appear.
2183 static void nocb_leader_wait(struct rcu_data *my_rdp)
2185 bool firsttime = true;
2187 struct rcu_data *rdp;
2188 struct rcu_head **tail;
2192 /* Wait for callbacks to appear. */
2193 if (!rcu_nocb_poll) {
2194 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2195 wait_event_interruptible(my_rdp->nocb_wq,
2196 !ACCESS_ONCE(my_rdp->nocb_leader_sleep));
2197 /* Memory barrier handled by smp_mb() calls below and repoll. */
2198 } else if (firsttime) {
2199 firsttime = false; /* Don't drown trace log with "Poll"! */
2200 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2204 * Each pass through the following loop checks a follower for CBs.
2205 * We are our own first follower. Any CBs found are moved to
2206 * nocb_gp_head, where they await a grace period.
2209 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2210 rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head);
2211 if (!rdp->nocb_gp_head)
2212 continue; /* No CBs here, try next follower. */
2214 /* Move callbacks to wait-for-GP list, which is empty. */
2215 ACCESS_ONCE(rdp->nocb_head) = NULL;
2216 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2221 * If there were no callbacks, sleep a bit, rescan after a
2222 * memory barrier, and go retry.
2224 if (unlikely(!gotcbs)) {
2226 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2228 WARN_ON(signal_pending(current));
2229 schedule_timeout_interruptible(1);
2231 /* Rescan in case we were a victim of memory ordering. */
2232 my_rdp->nocb_leader_sleep = true;
2233 smp_mb(); /* Ensure _sleep true before scan. */
2234 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2235 if (ACCESS_ONCE(rdp->nocb_head)) {
2236 /* Found CB, so short-circuit next wait. */
2237 my_rdp->nocb_leader_sleep = false;
2243 /* Wait for one grace period. */
2244 rcu_nocb_wait_gp(my_rdp);
2247 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2248 * We set it now, but recheck for new callbacks while
2249 * traversing our follower list.
2251 my_rdp->nocb_leader_sleep = true;
2252 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2254 /* Each pass through the following loop wakes a follower, if needed. */
2255 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2256 if (ACCESS_ONCE(rdp->nocb_head))
2257 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2258 if (!rdp->nocb_gp_head)
2259 continue; /* No CBs, so no need to wake follower. */
2261 /* Append callbacks to follower's "done" list. */
2262 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2263 *tail = rdp->nocb_gp_head;
2264 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2265 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2267 * List was empty, wake up the follower.
2268 * Memory barriers supplied by atomic_long_add().
2270 wake_up(&rdp->nocb_wq);
2274 /* If we (the leader) don't have CBs, go wait some more. */
2275 if (!my_rdp->nocb_follower_head)
2280 * Followers come here to wait for additional callbacks to show up.
2281 * This function does not return until callbacks appear.
2283 static void nocb_follower_wait(struct rcu_data *rdp)
2285 bool firsttime = true;
2288 if (!rcu_nocb_poll) {
2289 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2291 wait_event_interruptible(rdp->nocb_wq,
2292 ACCESS_ONCE(rdp->nocb_follower_head));
2293 } else if (firsttime) {
2294 /* Don't drown trace log with "Poll"! */
2296 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2298 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2299 /* ^^^ Ensure CB invocation follows _head test. */
2303 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2305 WARN_ON(signal_pending(current));
2306 schedule_timeout_interruptible(1);
2311 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2312 * callbacks queued by the corresponding no-CBs CPU, however, there is
2313 * an optional leader-follower relationship so that the grace-period
2314 * kthreads don't have to do quite so many wakeups.
2316 static int rcu_nocb_kthread(void *arg)
2319 struct rcu_head *list;
2320 struct rcu_head *next;
2321 struct rcu_head **tail;
2322 struct rcu_data *rdp = arg;
2324 /* Each pass through this loop invokes one batch of callbacks */
2326 /* Wait for callbacks. */
2327 if (rdp->nocb_leader == rdp)
2328 nocb_leader_wait(rdp);
2330 nocb_follower_wait(rdp);
2332 /* Pull the ready-to-invoke callbacks onto local list. */
2333 list = ACCESS_ONCE(rdp->nocb_follower_head);
2335 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2336 ACCESS_ONCE(rdp->nocb_follower_head) = NULL;
2337 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2339 /* Each pass through the following loop invokes a callback. */
2340 trace_rcu_batch_start(rdp->rsp->name,
2341 atomic_long_read(&rdp->nocb_q_count_lazy),
2342 atomic_long_read(&rdp->nocb_q_count), -1);
2346 /* Wait for enqueuing to complete, if needed. */
2347 while (next == NULL && &list->next != tail) {
2348 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2350 schedule_timeout_interruptible(1);
2351 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2355 debug_rcu_head_unqueue(list);
2357 if (__rcu_reclaim(rdp->rsp->name, list))
2363 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2364 smp_mb__before_atomic(); /* _add after CB invocation. */
2365 atomic_long_add(-c, &rdp->nocb_q_count);
2366 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2367 rdp->n_nocbs_invoked += c;
2372 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2373 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2375 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2378 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2379 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2383 if (!rcu_nocb_need_deferred_wakeup(rdp))
2385 ndw = ACCESS_ONCE(rdp->nocb_defer_wakeup);
2386 ACCESS_ONCE(rdp->nocb_defer_wakeup) = RCU_NOGP_WAKE_NOT;
2387 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2388 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2391 void __init rcu_init_nohz(void)
2394 bool need_rcu_nocb_mask = true;
2395 struct rcu_state *rsp;
2397 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2398 need_rcu_nocb_mask = false;
2399 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2401 #if defined(CONFIG_NO_HZ_FULL)
2402 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2403 need_rcu_nocb_mask = true;
2404 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2406 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2407 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2408 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2411 have_rcu_nocb_mask = true;
2413 if (!have_rcu_nocb_mask)
2416 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2417 pr_info("\tOffload RCU callbacks from CPU 0\n");
2418 cpumask_set_cpu(0, rcu_nocb_mask);
2419 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2420 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2421 pr_info("\tOffload RCU callbacks from all CPUs\n");
2422 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2423 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2424 #if defined(CONFIG_NO_HZ_FULL)
2425 if (tick_nohz_full_running)
2426 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2427 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2429 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2430 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2431 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2434 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2435 cpumask_pr_args(rcu_nocb_mask));
2437 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2439 for_each_rcu_flavor(rsp) {
2440 for_each_cpu(cpu, rcu_nocb_mask) {
2441 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2444 * If there are early callbacks, they will need
2445 * to be moved to the nocb lists.
2447 WARN_ON_ONCE(rdp->nxttail[RCU_NEXT_TAIL] !=
2449 rdp->nxttail[RCU_NEXT_TAIL] != NULL);
2450 init_nocb_callback_list(rdp);
2452 rcu_organize_nocb_kthreads(rsp);
2456 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2457 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2459 rdp->nocb_tail = &rdp->nocb_head;
2460 init_waitqueue_head(&rdp->nocb_wq);
2461 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2465 * If the specified CPU is a no-CBs CPU that does not already have its
2466 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2467 * brought online out of order, this can require re-organizing the
2468 * leader-follower relationships.
2470 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2472 struct rcu_data *rdp;
2473 struct rcu_data *rdp_last;
2474 struct rcu_data *rdp_old_leader;
2475 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2476 struct task_struct *t;
2479 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2480 * then nothing to do.
2482 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2485 /* If we didn't spawn the leader first, reorganize! */
2486 rdp_old_leader = rdp_spawn->nocb_leader;
2487 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2489 rdp = rdp_old_leader;
2491 rdp->nocb_leader = rdp_spawn;
2492 if (rdp_last && rdp != rdp_spawn)
2493 rdp_last->nocb_next_follower = rdp;
2494 if (rdp == rdp_spawn) {
2495 rdp = rdp->nocb_next_follower;
2498 rdp = rdp->nocb_next_follower;
2499 rdp_last->nocb_next_follower = NULL;
2502 rdp_spawn->nocb_next_follower = rdp_old_leader;
2505 /* Spawn the kthread for this CPU and RCU flavor. */
2506 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2507 "rcuo%c/%d", rsp->abbr, cpu);
2509 ACCESS_ONCE(rdp_spawn->nocb_kthread) = t;
2513 * If the specified CPU is a no-CBs CPU that does not already have its
2514 * rcuo kthreads, spawn them.
2516 static void rcu_spawn_all_nocb_kthreads(int cpu)
2518 struct rcu_state *rsp;
2520 if (rcu_scheduler_fully_active)
2521 for_each_rcu_flavor(rsp)
2522 rcu_spawn_one_nocb_kthread(rsp, cpu);
2526 * Once the scheduler is running, spawn rcuo kthreads for all online
2527 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2528 * non-boot CPUs come online -- if this changes, we will need to add
2529 * some mutual exclusion.
2531 static void __init rcu_spawn_nocb_kthreads(void)
2535 for_each_online_cpu(cpu)
2536 rcu_spawn_all_nocb_kthreads(cpu);
2539 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2540 static int rcu_nocb_leader_stride = -1;
2541 module_param(rcu_nocb_leader_stride, int, 0444);
2544 * Initialize leader-follower relationships for all no-CBs CPU.
2546 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2549 int ls = rcu_nocb_leader_stride;
2550 int nl = 0; /* Next leader. */
2551 struct rcu_data *rdp;
2552 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2553 struct rcu_data *rdp_prev = NULL;
2555 if (!have_rcu_nocb_mask)
2558 ls = int_sqrt(nr_cpu_ids);
2559 rcu_nocb_leader_stride = ls;
2563 * Each pass through this loop sets up one rcu_data structure and
2564 * spawns one rcu_nocb_kthread().
2566 for_each_cpu(cpu, rcu_nocb_mask) {
2567 rdp = per_cpu_ptr(rsp->rda, cpu);
2568 if (rdp->cpu >= nl) {
2569 /* New leader, set up for followers & next leader. */
2570 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2571 rdp->nocb_leader = rdp;
2574 /* Another follower, link to previous leader. */
2575 rdp->nocb_leader = rdp_leader;
2576 rdp_prev->nocb_next_follower = rdp;
2582 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2583 static bool init_nocb_callback_list(struct rcu_data *rdp)
2585 if (!rcu_is_nocb_cpu(rdp->cpu))
2588 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2592 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2594 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2596 WARN_ON_ONCE(1); /* Should be dead code. */
2600 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2604 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2608 static void rcu_init_one_nocb(struct rcu_node *rnp)
2612 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2613 bool lazy, unsigned long flags)
2618 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2619 struct rcu_data *rdp,
2620 unsigned long flags)
2625 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2629 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2634 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2638 static void rcu_spawn_all_nocb_kthreads(int cpu)
2642 static void __init rcu_spawn_nocb_kthreads(void)
2646 static bool init_nocb_callback_list(struct rcu_data *rdp)
2651 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2654 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2655 * arbitrarily long period of time with the scheduling-clock tick turned
2656 * off. RCU will be paying attention to this CPU because it is in the
2657 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2658 * machine because the scheduling-clock tick has been disabled. Therefore,
2659 * if an adaptive-ticks CPU is failing to respond to the current grace
2660 * period and has not be idle from an RCU perspective, kick it.
2662 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2664 #ifdef CONFIG_NO_HZ_FULL
2665 if (tick_nohz_full_cpu(cpu))
2666 smp_send_reschedule(cpu);
2667 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2671 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2673 static int full_sysidle_state; /* Current system-idle state. */
2674 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2675 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2676 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2677 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2678 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2681 * Invoked to note exit from irq or task transition to idle. Note that
2682 * usermode execution does -not- count as idle here! After all, we want
2683 * to detect full-system idle states, not RCU quiescent states and grace
2684 * periods. The caller must have disabled interrupts.
2686 static void rcu_sysidle_enter(int irq)
2689 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2691 /* If there are no nohz_full= CPUs, no need to track this. */
2692 if (!tick_nohz_full_enabled())
2695 /* Adjust nesting, check for fully idle. */
2697 rdtp->dynticks_idle_nesting--;
2698 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2699 if (rdtp->dynticks_idle_nesting != 0)
2700 return; /* Still not fully idle. */
2702 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2703 DYNTICK_TASK_NEST_VALUE) {
2704 rdtp->dynticks_idle_nesting = 0;
2706 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2707 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2708 return; /* Still not fully idle. */
2712 /* Record start of fully idle period. */
2714 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2715 smp_mb__before_atomic();
2716 atomic_inc(&rdtp->dynticks_idle);
2717 smp_mb__after_atomic();
2718 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2722 * Unconditionally force exit from full system-idle state. This is
2723 * invoked when a normal CPU exits idle, but must be called separately
2724 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2725 * is that the timekeeping CPU is permitted to take scheduling-clock
2726 * interrupts while the system is in system-idle state, and of course
2727 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2728 * interrupt from any other type of interrupt.
2730 void rcu_sysidle_force_exit(void)
2732 int oldstate = ACCESS_ONCE(full_sysidle_state);
2736 * Each pass through the following loop attempts to exit full
2737 * system-idle state. If contention proves to be a problem,
2738 * a trylock-based contention tree could be used here.
2740 while (oldstate > RCU_SYSIDLE_SHORT) {
2741 newoldstate = cmpxchg(&full_sysidle_state,
2742 oldstate, RCU_SYSIDLE_NOT);
2743 if (oldstate == newoldstate &&
2744 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2745 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2746 return; /* We cleared it, done! */
2748 oldstate = newoldstate;
2750 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2754 * Invoked to note entry to irq or task transition from idle. Note that
2755 * usermode execution does -not- count as idle here! The caller must
2756 * have disabled interrupts.
2758 static void rcu_sysidle_exit(int irq)
2760 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2762 /* If there are no nohz_full= CPUs, no need to track this. */
2763 if (!tick_nohz_full_enabled())
2766 /* Adjust nesting, check for already non-idle. */
2768 rdtp->dynticks_idle_nesting++;
2769 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2770 if (rdtp->dynticks_idle_nesting != 1)
2771 return; /* Already non-idle. */
2774 * Allow for irq misnesting. Yes, it really is possible
2775 * to enter an irq handler then never leave it, and maybe
2776 * also vice versa. Handle both possibilities.
2778 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2779 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2780 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2781 return; /* Already non-idle. */
2783 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2787 /* Record end of idle period. */
2788 smp_mb__before_atomic();
2789 atomic_inc(&rdtp->dynticks_idle);
2790 smp_mb__after_atomic();
2791 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2794 * If we are the timekeeping CPU, we are permitted to be non-idle
2795 * during a system-idle state. This must be the case, because
2796 * the timekeeping CPU has to take scheduling-clock interrupts
2797 * during the time that the system is transitioning to full
2798 * system-idle state. This means that the timekeeping CPU must
2799 * invoke rcu_sysidle_force_exit() directly if it does anything
2800 * more than take a scheduling-clock interrupt.
2802 if (smp_processor_id() == tick_do_timer_cpu)
2805 /* Update system-idle state: We are clearly no longer fully idle! */
2806 rcu_sysidle_force_exit();
2810 * Check to see if the current CPU is idle. Note that usermode execution
2811 * does not count as idle. The caller must have disabled interrupts.
2813 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2814 unsigned long *maxj)
2818 struct rcu_dynticks *rdtp = rdp->dynticks;
2820 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2821 if (!tick_nohz_full_enabled())
2825 * If some other CPU has already reported non-idle, if this is
2826 * not the flavor of RCU that tracks sysidle state, or if this
2827 * is an offline or the timekeeping CPU, nothing to do.
2829 if (!*isidle || rdp->rsp != rcu_state_p ||
2830 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2832 if (rcu_gp_in_progress(rdp->rsp))
2833 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2835 /* Pick up current idle and NMI-nesting counter and check. */
2836 cur = atomic_read(&rdtp->dynticks_idle);
2838 *isidle = false; /* We are not idle! */
2841 smp_mb(); /* Read counters before timestamps. */
2843 /* Pick up timestamps. */
2844 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2845 /* If this CPU entered idle more recently, update maxj timestamp. */
2846 if (ULONG_CMP_LT(*maxj, j))
2851 * Is this the flavor of RCU that is handling full-system idle?
2853 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2855 return rsp == rcu_state_p;
2859 * Return a delay in jiffies based on the number of CPUs, rcu_node
2860 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2861 * systems more time to transition to full-idle state in order to
2862 * avoid the cache thrashing that otherwise occur on the state variable.
2863 * Really small systems (less than a couple of tens of CPUs) should
2864 * instead use a single global atomically incremented counter, and later
2865 * versions of this will automatically reconfigure themselves accordingly.
2867 static unsigned long rcu_sysidle_delay(void)
2869 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2871 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2875 * Advance the full-system-idle state. This is invoked when all of
2876 * the non-timekeeping CPUs are idle.
2878 static void rcu_sysidle(unsigned long j)
2880 /* Check the current state. */
2881 switch (ACCESS_ONCE(full_sysidle_state)) {
2882 case RCU_SYSIDLE_NOT:
2884 /* First time all are idle, so note a short idle period. */
2885 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2888 case RCU_SYSIDLE_SHORT:
2891 * Idle for a bit, time to advance to next state?
2892 * cmpxchg failure means race with non-idle, let them win.
2894 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2895 (void)cmpxchg(&full_sysidle_state,
2896 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2899 case RCU_SYSIDLE_LONG:
2902 * Do an additional check pass before advancing to full.
2903 * cmpxchg failure means race with non-idle, let them win.
2905 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2906 (void)cmpxchg(&full_sysidle_state,
2907 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2916 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2917 * back to the beginning.
2919 static void rcu_sysidle_cancel(void)
2922 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2923 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2927 * Update the sysidle state based on the results of a force-quiescent-state
2928 * scan of the CPUs' dyntick-idle state.
2930 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2931 unsigned long maxj, bool gpkt)
2933 if (rsp != rcu_state_p)
2934 return; /* Wrong flavor, ignore. */
2935 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2936 return; /* Running state machine from timekeeping CPU. */
2938 rcu_sysidle(maxj); /* More idle! */
2940 rcu_sysidle_cancel(); /* Idle is over. */
2944 * Wrapper for rcu_sysidle_report() when called from the grace-period
2945 * kthread's context.
2947 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2950 /* If there are no nohz_full= CPUs, no need to track this. */
2951 if (!tick_nohz_full_enabled())
2954 rcu_sysidle_report(rsp, isidle, maxj, true);
2957 /* Callback and function for forcing an RCU grace period. */
2958 struct rcu_sysidle_head {
2963 static void rcu_sysidle_cb(struct rcu_head *rhp)
2965 struct rcu_sysidle_head *rshp;
2968 * The following memory barrier is needed to replace the
2969 * memory barriers that would normally be in the memory
2972 smp_mb(); /* grace period precedes setting inuse. */
2974 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2975 ACCESS_ONCE(rshp->inuse) = 0;
2979 * Check to see if the system is fully idle, other than the timekeeping CPU.
2980 * The caller must have disabled interrupts. This is not intended to be
2981 * called unless tick_nohz_full_enabled().
2983 bool rcu_sys_is_idle(void)
2985 static struct rcu_sysidle_head rsh;
2986 int rss = ACCESS_ONCE(full_sysidle_state);
2988 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2991 /* Handle small-system case by doing a full scan of CPUs. */
2992 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2993 int oldrss = rss - 1;
2996 * One pass to advance to each state up to _FULL.
2997 * Give up if any pass fails to advance the state.
2999 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
3002 unsigned long maxj = jiffies - ULONG_MAX / 4;
3003 struct rcu_data *rdp;
3005 /* Scan all the CPUs looking for nonidle CPUs. */
3006 for_each_possible_cpu(cpu) {
3007 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3008 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
3012 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
3014 rss = ACCESS_ONCE(full_sysidle_state);
3018 /* If this is the first observation of an idle period, record it. */
3019 if (rss == RCU_SYSIDLE_FULL) {
3020 rss = cmpxchg(&full_sysidle_state,
3021 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
3022 return rss == RCU_SYSIDLE_FULL;
3025 smp_mb(); /* ensure rss load happens before later caller actions. */
3027 /* If already fully idle, tell the caller (in case of races). */
3028 if (rss == RCU_SYSIDLE_FULL_NOTED)
3032 * If we aren't there yet, and a grace period is not in flight,
3033 * initiate a grace period. Either way, tell the caller that
3034 * we are not there yet. We use an xchg() rather than an assignment
3035 * to make up for the memory barriers that would otherwise be
3036 * provided by the memory allocator.
3038 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
3039 !rcu_gp_in_progress(rcu_state_p) &&
3040 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
3041 call_rcu(&rsh.rh, rcu_sysidle_cb);
3046 * Initialize dynticks sysidle state for CPUs coming online.
3048 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3050 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
3053 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3055 static void rcu_sysidle_enter(int irq)
3059 static void rcu_sysidle_exit(int irq)
3063 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3064 unsigned long *maxj)
3068 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3073 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3078 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3082 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3085 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3086 * grace-period kthread will do force_quiescent_state() processing?
3087 * The idea is to avoid waking up RCU core processing on such a
3088 * CPU unless the grace period has extended for too long.
3090 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3091 * CONFIG_RCU_NOCB_CPU CPUs.
3093 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3095 #ifdef CONFIG_NO_HZ_FULL
3096 if (tick_nohz_full_cpu(smp_processor_id()) &&
3097 (!rcu_gp_in_progress(rsp) ||
3098 ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
3100 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3105 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3108 static void rcu_bind_gp_kthread(void)
3110 int __maybe_unused cpu;
3112 if (!tick_nohz_full_enabled())
3114 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3115 cpu = tick_do_timer_cpu;
3116 if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu)
3117 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3118 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3119 if (!is_housekeeping_cpu(raw_smp_processor_id()))
3120 housekeeping_affine(current);
3121 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3124 /* Record the current task on dyntick-idle entry. */
3125 static void rcu_dynticks_task_enter(void)
3127 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3128 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = smp_processor_id();
3129 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3132 /* Record no current task on dyntick-idle exit. */
3133 static void rcu_dynticks_task_exit(void)
3135 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3136 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = -1;
3137 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */