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"
37 /* rcuc/rcub kthread realtime priority */
38 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
39 module_param(kthread_prio, int, 0644);
42 * Control variables for per-CPU and per-rcu_node kthreads. These
43 * handle all flavors of RCU.
45 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
46 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
47 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
48 DEFINE_PER_CPU(char, rcu_cpu_has_work);
50 #endif /* #ifdef CONFIG_RCU_BOOST */
52 #ifdef CONFIG_RCU_NOCB_CPU
53 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
54 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
55 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
56 static char __initdata nocb_buf[NR_CPUS * 5];
57 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
60 * Check the RCU kernel configuration parameters and print informative
61 * messages about anything out of the ordinary. If you like #ifdef, you
62 * will love this function.
64 static void __init rcu_bootup_announce_oddness(void)
66 #ifdef CONFIG_RCU_TRACE
67 pr_info("\tRCU debugfs-based tracing is enabled.\n");
69 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
70 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
73 #ifdef CONFIG_RCU_FANOUT_EXACT
74 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
76 #ifdef CONFIG_RCU_FAST_NO_HZ
77 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
79 #ifdef CONFIG_PROVE_RCU
80 pr_info("\tRCU lockdep checking is enabled.\n");
82 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
83 pr_info("\tRCU torture testing starts during boot.\n");
85 #if defined(CONFIG_RCU_CPU_STALL_INFO)
86 pr_info("\tAdditional per-CPU info printed with stalls.\n");
88 #if NUM_RCU_LVL_4 != 0
89 pr_info("\tFour-level hierarchy is enabled.\n");
91 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
92 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
93 if (nr_cpu_ids != NR_CPUS)
94 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
95 #ifdef CONFIG_RCU_BOOST
96 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
100 #ifdef CONFIG_PREEMPT_RCU
102 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
103 static struct rcu_state *rcu_state_p = &rcu_preempt_state;
105 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
106 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
110 * Tell them what RCU they are running.
112 static void __init rcu_bootup_announce(void)
114 pr_info("Preemptible hierarchical RCU implementation.\n");
115 rcu_bootup_announce_oddness();
119 * Return the number of RCU-preempt batches processed thus far
120 * for debug and statistics.
122 static long rcu_batches_completed_preempt(void)
124 return rcu_preempt_state.completed;
126 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
129 * Return the number of RCU batches processed thus far for debug & stats.
131 long rcu_batches_completed(void)
133 return rcu_batches_completed_preempt();
135 EXPORT_SYMBOL_GPL(rcu_batches_completed);
138 * Record a preemptible-RCU quiescent state for the specified CPU. Note
139 * that this just means that the task currently running on the CPU is
140 * not in a quiescent state. There might be any number of tasks blocked
141 * while in an RCU read-side critical section.
143 * As with the other rcu_*_qs() functions, callers to this function
144 * must disable preemption.
146 static void rcu_preempt_qs(void)
148 if (!__this_cpu_read(rcu_preempt_data.passed_quiesce)) {
149 trace_rcu_grace_period(TPS("rcu_preempt"),
150 __this_cpu_read(rcu_preempt_data.gpnum),
152 __this_cpu_write(rcu_preempt_data.passed_quiesce, 1);
153 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
154 current->rcu_read_unlock_special.b.need_qs = false;
159 * We have entered the scheduler, and the current task might soon be
160 * context-switched away from. If this task is in an RCU read-side
161 * critical section, we will no longer be able to rely on the CPU to
162 * record that fact, so we enqueue the task on the blkd_tasks list.
163 * The task will dequeue itself when it exits the outermost enclosing
164 * RCU read-side critical section. Therefore, the current grace period
165 * cannot be permitted to complete until the blkd_tasks list entries
166 * predating the current grace period drain, in other words, until
167 * rnp->gp_tasks becomes NULL.
169 * Caller must disable preemption.
171 static void rcu_preempt_note_context_switch(void)
173 struct task_struct *t = current;
175 struct rcu_data *rdp;
176 struct rcu_node *rnp;
178 if (t->rcu_read_lock_nesting > 0 &&
179 !t->rcu_read_unlock_special.b.blocked) {
181 /* Possibly blocking in an RCU read-side critical section. */
182 rdp = this_cpu_ptr(rcu_preempt_state.rda);
184 raw_spin_lock_irqsave(&rnp->lock, flags);
185 smp_mb__after_unlock_lock();
186 t->rcu_read_unlock_special.b.blocked = true;
187 t->rcu_blocked_node = rnp;
190 * If this CPU has already checked in, then this task
191 * will hold up the next grace period rather than the
192 * current grace period. Queue the task accordingly.
193 * If the task is queued for the current grace period
194 * (i.e., this CPU has not yet passed through a quiescent
195 * state for the current grace period), then as long
196 * as that task remains queued, the current grace period
197 * cannot end. Note that there is some uncertainty as
198 * to exactly when the current grace period started.
199 * We take a conservative approach, which can result
200 * in unnecessarily waiting on tasks that started very
201 * slightly after the current grace period began. C'est
204 * But first, note that the current CPU must still be
207 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
208 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
209 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
210 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
211 rnp->gp_tasks = &t->rcu_node_entry;
212 #ifdef CONFIG_RCU_BOOST
213 if (rnp->boost_tasks != NULL)
214 rnp->boost_tasks = rnp->gp_tasks;
215 #endif /* #ifdef CONFIG_RCU_BOOST */
217 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
218 if (rnp->qsmask & rdp->grpmask)
219 rnp->gp_tasks = &t->rcu_node_entry;
221 trace_rcu_preempt_task(rdp->rsp->name,
223 (rnp->qsmask & rdp->grpmask)
226 raw_spin_unlock_irqrestore(&rnp->lock, flags);
227 } else if (t->rcu_read_lock_nesting < 0 &&
228 t->rcu_read_unlock_special.s) {
231 * Complete exit from RCU read-side critical section on
232 * behalf of preempted instance of __rcu_read_unlock().
234 rcu_read_unlock_special(t);
238 * Either we were not in an RCU read-side critical section to
239 * begin with, or we have now recorded that critical section
240 * globally. Either way, we can now note a quiescent state
241 * for this CPU. Again, if we were in an RCU read-side critical
242 * section, and if that critical section was blocking the current
243 * grace period, then the fact that the task has been enqueued
244 * means that we continue to block the current grace period.
250 * Check for preempted RCU readers blocking the current grace period
251 * for the specified rcu_node structure. If the caller needs a reliable
252 * answer, it must hold the rcu_node's ->lock.
254 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
256 return rnp->gp_tasks != NULL;
260 * Record a quiescent state for all tasks that were previously queued
261 * on the specified rcu_node structure and that were blocking the current
262 * RCU grace period. The caller must hold the specified rnp->lock with
263 * irqs disabled, and this lock is released upon return, but irqs remain
266 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
267 __releases(rnp->lock)
270 struct rcu_node *rnp_p;
272 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
273 raw_spin_unlock_irqrestore(&rnp->lock, flags);
274 return; /* Still need more quiescent states! */
280 * Either there is only one rcu_node in the tree,
281 * or tasks were kicked up to root rcu_node due to
282 * CPUs going offline.
284 rcu_report_qs_rsp(&rcu_preempt_state, flags);
288 /* Report up the rest of the hierarchy. */
290 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
291 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
292 smp_mb__after_unlock_lock();
293 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
297 * Advance a ->blkd_tasks-list pointer to the next entry, instead
298 * returning NULL if at the end of the list.
300 static struct list_head *rcu_next_node_entry(struct task_struct *t,
301 struct rcu_node *rnp)
303 struct list_head *np;
305 np = t->rcu_node_entry.next;
306 if (np == &rnp->blkd_tasks)
312 * Return true if the specified rcu_node structure has tasks that were
313 * preempted within an RCU read-side critical section.
315 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
317 return !list_empty(&rnp->blkd_tasks);
321 * Handle special cases during rcu_read_unlock(), such as needing to
322 * notify RCU core processing or task having blocked during the RCU
323 * read-side critical section.
325 void rcu_read_unlock_special(struct task_struct *t)
332 struct list_head *np;
333 #ifdef CONFIG_RCU_BOOST
334 bool drop_boost_mutex = false;
335 #endif /* #ifdef CONFIG_RCU_BOOST */
336 struct rcu_node *rnp;
337 union rcu_special special;
339 /* NMI handlers cannot block and cannot safely manipulate state. */
343 local_irq_save(flags);
346 * If RCU core is waiting for this CPU to exit critical section,
347 * let it know that we have done so. Because irqs are disabled,
348 * t->rcu_read_unlock_special cannot change.
350 special = t->rcu_read_unlock_special;
351 if (special.b.need_qs) {
353 if (!t->rcu_read_unlock_special.s) {
354 local_irq_restore(flags);
359 /* Hardware IRQ handlers cannot block, complain if they get here. */
360 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
361 local_irq_restore(flags);
365 /* Clean up if blocked during RCU read-side critical section. */
366 if (special.b.blocked) {
367 t->rcu_read_unlock_special.b.blocked = false;
370 * Remove this task from the list it blocked on. The
371 * task can migrate while we acquire the lock, but at
372 * most one time. So at most two passes through loop.
375 rnp = t->rcu_blocked_node;
376 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
377 smp_mb__after_unlock_lock();
378 if (rnp == t->rcu_blocked_node)
380 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
382 empty = !rcu_preempt_has_tasks(rnp);
383 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
384 empty_exp = !rcu_preempted_readers_exp(rnp);
385 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
386 np = rcu_next_node_entry(t, rnp);
387 list_del_init(&t->rcu_node_entry);
388 t->rcu_blocked_node = NULL;
389 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
391 if (&t->rcu_node_entry == rnp->gp_tasks)
393 if (&t->rcu_node_entry == rnp->exp_tasks)
395 #ifdef CONFIG_RCU_BOOST
396 if (&t->rcu_node_entry == rnp->boost_tasks)
397 rnp->boost_tasks = np;
398 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
399 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
400 #endif /* #ifdef CONFIG_RCU_BOOST */
403 * If this was the last task on the list, go see if we
404 * need to propagate ->qsmaskinit bit clearing up the
407 if (!empty && !rcu_preempt_has_tasks(rnp))
408 rcu_cleanup_dead_rnp(rnp);
411 * If this was the last task on the current list, and if
412 * we aren't waiting on any CPUs, report the quiescent state.
413 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
414 * so we must take a snapshot of the expedited state.
416 empty_exp_now = !rcu_preempted_readers_exp(rnp);
417 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
418 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
425 rcu_report_unblock_qs_rnp(rnp, flags);
427 raw_spin_unlock_irqrestore(&rnp->lock, flags);
430 #ifdef CONFIG_RCU_BOOST
431 /* Unboost if we were boosted. */
432 if (drop_boost_mutex) {
433 rt_mutex_unlock(&rnp->boost_mtx);
434 complete(&rnp->boost_completion);
436 #endif /* #ifdef CONFIG_RCU_BOOST */
439 * If this was the last task on the expedited lists,
440 * then we need to report up the rcu_node hierarchy.
442 if (!empty_exp && empty_exp_now)
443 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
445 local_irq_restore(flags);
450 * Dump detailed information for all tasks blocking the current RCU
451 * grace period on the specified rcu_node structure.
453 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
456 struct task_struct *t;
458 raw_spin_lock_irqsave(&rnp->lock, flags);
459 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
460 raw_spin_unlock_irqrestore(&rnp->lock, flags);
463 t = list_entry(rnp->gp_tasks,
464 struct task_struct, rcu_node_entry);
465 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
467 raw_spin_unlock_irqrestore(&rnp->lock, flags);
471 * Dump detailed information for all tasks blocking the current RCU
474 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
476 struct rcu_node *rnp = rcu_get_root(rsp);
478 rcu_print_detail_task_stall_rnp(rnp);
479 rcu_for_each_leaf_node(rsp, rnp)
480 rcu_print_detail_task_stall_rnp(rnp);
483 #ifdef CONFIG_RCU_CPU_STALL_INFO
485 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
487 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
488 rnp->level, rnp->grplo, rnp->grphi);
491 static void rcu_print_task_stall_end(void)
496 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
498 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
502 static void rcu_print_task_stall_end(void)
506 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
509 * Scan the current list of tasks blocked within RCU read-side critical
510 * sections, printing out the tid of each.
512 static int rcu_print_task_stall(struct rcu_node *rnp)
514 struct task_struct *t;
517 if (!rcu_preempt_blocked_readers_cgp(rnp))
519 rcu_print_task_stall_begin(rnp);
520 t = list_entry(rnp->gp_tasks,
521 struct task_struct, rcu_node_entry);
522 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
523 pr_cont(" P%d", t->pid);
526 rcu_print_task_stall_end();
531 * Check that the list of blocked tasks for the newly completed grace
532 * period is in fact empty. It is a serious bug to complete a grace
533 * period that still has RCU readers blocked! This function must be
534 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
535 * must be held by the caller.
537 * Also, if there are blocked tasks on the list, they automatically
538 * block the newly created grace period, so set up ->gp_tasks accordingly.
540 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
542 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
543 if (rcu_preempt_has_tasks(rnp))
544 rnp->gp_tasks = rnp->blkd_tasks.next;
545 WARN_ON_ONCE(rnp->qsmask);
548 #ifdef CONFIG_HOTPLUG_CPU
550 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
553 * Check for a quiescent state from the current CPU. When a task blocks,
554 * the task is recorded in the corresponding CPU's rcu_node structure,
555 * which is checked elsewhere.
557 * Caller must disable hard irqs.
559 static void rcu_preempt_check_callbacks(void)
561 struct task_struct *t = current;
563 if (t->rcu_read_lock_nesting == 0) {
567 if (t->rcu_read_lock_nesting > 0 &&
568 __this_cpu_read(rcu_preempt_data.qs_pending) &&
569 !__this_cpu_read(rcu_preempt_data.passed_quiesce))
570 t->rcu_read_unlock_special.b.need_qs = true;
573 #ifdef CONFIG_RCU_BOOST
575 static void rcu_preempt_do_callbacks(void)
577 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
580 #endif /* #ifdef CONFIG_RCU_BOOST */
583 * Queue a preemptible-RCU callback for invocation after a grace period.
585 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
587 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
589 EXPORT_SYMBOL_GPL(call_rcu);
592 * synchronize_rcu - wait until a grace period has elapsed.
594 * Control will return to the caller some time after a full grace
595 * period has elapsed, in other words after all currently executing RCU
596 * read-side critical sections have completed. Note, however, that
597 * upon return from synchronize_rcu(), the caller might well be executing
598 * concurrently with new RCU read-side critical sections that began while
599 * synchronize_rcu() was waiting. RCU read-side critical sections are
600 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
602 * See the description of synchronize_sched() for more detailed information
603 * on memory ordering guarantees.
605 void synchronize_rcu(void)
607 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
608 !lock_is_held(&rcu_lock_map) &&
609 !lock_is_held(&rcu_sched_lock_map),
610 "Illegal synchronize_rcu() in RCU read-side critical section");
611 if (!rcu_scheduler_active)
614 synchronize_rcu_expedited();
616 wait_rcu_gp(call_rcu);
618 EXPORT_SYMBOL_GPL(synchronize_rcu);
620 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
621 static unsigned long sync_rcu_preempt_exp_count;
622 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
625 * Return non-zero if there are any tasks in RCU read-side critical
626 * sections blocking the current preemptible-RCU expedited grace period.
627 * If there is no preemptible-RCU expedited grace period currently in
628 * progress, returns zero unconditionally.
630 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
632 return rnp->exp_tasks != NULL;
636 * return non-zero if there is no RCU expedited grace period in progress
637 * for the specified rcu_node structure, in other words, if all CPUs and
638 * tasks covered by the specified rcu_node structure have done their bit
639 * for the current expedited grace period. Works only for preemptible
640 * RCU -- other RCU implementation use other means.
642 * Caller must hold sync_rcu_preempt_exp_mutex.
644 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
646 return !rcu_preempted_readers_exp(rnp) &&
647 ACCESS_ONCE(rnp->expmask) == 0;
651 * Report the exit from RCU read-side critical section for the last task
652 * that queued itself during or before the current expedited preemptible-RCU
653 * grace period. This event is reported either to the rcu_node structure on
654 * which the task was queued or to one of that rcu_node structure's ancestors,
655 * recursively up the tree. (Calm down, calm down, we do the recursion
658 * Most callers will set the "wake" flag, but the task initiating the
659 * expedited grace period need not wake itself.
661 * Caller must hold sync_rcu_preempt_exp_mutex.
663 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
669 raw_spin_lock_irqsave(&rnp->lock, flags);
670 smp_mb__after_unlock_lock();
672 if (!sync_rcu_preempt_exp_done(rnp)) {
673 raw_spin_unlock_irqrestore(&rnp->lock, flags);
676 if (rnp->parent == NULL) {
677 raw_spin_unlock_irqrestore(&rnp->lock, flags);
679 smp_mb(); /* EGP done before wake_up(). */
680 wake_up(&sync_rcu_preempt_exp_wq);
685 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
687 raw_spin_lock(&rnp->lock); /* irqs already disabled */
688 smp_mb__after_unlock_lock();
689 rnp->expmask &= ~mask;
694 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
695 * grace period for the specified rcu_node structure. If there are no such
696 * tasks, report it up the rcu_node hierarchy.
698 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
699 * CPU hotplug operations.
702 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
707 raw_spin_lock_irqsave(&rnp->lock, flags);
708 smp_mb__after_unlock_lock();
709 if (!rcu_preempt_has_tasks(rnp)) {
710 raw_spin_unlock_irqrestore(&rnp->lock, flags);
712 rnp->exp_tasks = rnp->blkd_tasks.next;
713 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
717 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
721 * synchronize_rcu_expedited - Brute-force RCU grace period
723 * Wait for an RCU-preempt grace period, but expedite it. The basic
724 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
725 * the ->blkd_tasks lists and wait for this list to drain. This consumes
726 * significant time on all CPUs and is unfriendly to real-time workloads,
727 * so is thus not recommended for any sort of common-case code.
728 * In fact, if you are using synchronize_rcu_expedited() in a loop,
729 * please restructure your code to batch your updates, and then Use a
730 * single synchronize_rcu() instead.
732 void synchronize_rcu_expedited(void)
735 struct rcu_node *rnp;
736 struct rcu_state *rsp = &rcu_preempt_state;
740 smp_mb(); /* Caller's modifications seen first by other CPUs. */
741 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
742 smp_mb(); /* Above access cannot bleed into critical section. */
745 * Block CPU-hotplug operations. This means that any CPU-hotplug
746 * operation that finds an rcu_node structure with tasks in the
747 * process of being boosted will know that all tasks blocking
748 * this expedited grace period will already be in the process of
749 * being boosted. This simplifies the process of moving tasks
750 * from leaf to root rcu_node structures.
752 if (!try_get_online_cpus()) {
753 /* CPU-hotplug operation in flight, fall back to normal GP. */
754 wait_rcu_gp(call_rcu);
759 * Acquire lock, falling back to synchronize_rcu() if too many
760 * lock-acquisition failures. Of course, if someone does the
761 * expedited grace period for us, just leave.
763 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
764 if (ULONG_CMP_LT(snap,
765 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
767 goto mb_ret; /* Others did our work for us. */
769 if (trycount++ < 10) {
770 udelay(trycount * num_online_cpus());
773 wait_rcu_gp(call_rcu);
777 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
779 goto unlock_mb_ret; /* Others did our work for us. */
782 /* force all RCU readers onto ->blkd_tasks lists. */
783 synchronize_sched_expedited();
785 /* Initialize ->expmask for all non-leaf rcu_node structures. */
786 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
787 raw_spin_lock_irqsave(&rnp->lock, flags);
788 smp_mb__after_unlock_lock();
789 rnp->expmask = rnp->qsmaskinit;
790 raw_spin_unlock_irqrestore(&rnp->lock, flags);
793 /* Snapshot current state of ->blkd_tasks lists. */
794 rcu_for_each_leaf_node(rsp, rnp)
795 sync_rcu_preempt_exp_init(rsp, rnp);
796 if (NUM_RCU_NODES > 1)
797 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
801 /* Wait for snapshotted ->blkd_tasks lists to drain. */
802 rnp = rcu_get_root(rsp);
803 wait_event(sync_rcu_preempt_exp_wq,
804 sync_rcu_preempt_exp_done(rnp));
806 /* Clean up and exit. */
807 smp_mb(); /* ensure expedited GP seen before counter increment. */
808 ACCESS_ONCE(sync_rcu_preempt_exp_count) =
809 sync_rcu_preempt_exp_count + 1;
811 mutex_unlock(&sync_rcu_preempt_exp_mutex);
813 smp_mb(); /* ensure subsequent action seen after grace period. */
815 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
818 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
820 * Note that this primitive does not necessarily wait for an RCU grace period
821 * to complete. For example, if there are no RCU callbacks queued anywhere
822 * in the system, then rcu_barrier() is within its rights to return
823 * immediately, without waiting for anything, much less an RCU grace period.
825 void rcu_barrier(void)
827 _rcu_barrier(&rcu_preempt_state);
829 EXPORT_SYMBOL_GPL(rcu_barrier);
832 * Initialize preemptible RCU's state structures.
834 static void __init __rcu_init_preempt(void)
836 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
840 * Check for a task exiting while in a preemptible-RCU read-side
841 * critical section, clean up if so. No need to issue warnings,
842 * as debug_check_no_locks_held() already does this if lockdep
847 struct task_struct *t = current;
849 if (likely(list_empty(¤t->rcu_node_entry)))
851 t->rcu_read_lock_nesting = 1;
853 t->rcu_read_unlock_special.b.blocked = true;
857 #else /* #ifdef CONFIG_PREEMPT_RCU */
859 static struct rcu_state *rcu_state_p = &rcu_sched_state;
862 * Tell them what RCU they are running.
864 static void __init rcu_bootup_announce(void)
866 pr_info("Hierarchical RCU implementation.\n");
867 rcu_bootup_announce_oddness();
871 * Return the number of RCU batches processed thus far for debug & stats.
873 long rcu_batches_completed(void)
875 return rcu_batches_completed_sched();
877 EXPORT_SYMBOL_GPL(rcu_batches_completed);
880 * Because preemptible RCU does not exist, we never have to check for
881 * CPUs being in quiescent states.
883 static void rcu_preempt_note_context_switch(void)
888 * Because preemptible RCU does not exist, there are never any preempted
891 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
896 #ifdef CONFIG_HOTPLUG_CPU
899 * Because there is no preemptible RCU, there can be no readers blocked.
901 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
906 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
909 * Because preemptible RCU does not exist, we never have to check for
910 * tasks blocked within RCU read-side critical sections.
912 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
917 * Because preemptible RCU does not exist, we never have to check for
918 * tasks blocked within RCU read-side critical sections.
920 static int rcu_print_task_stall(struct rcu_node *rnp)
926 * Because there is no preemptible RCU, there can be no readers blocked,
927 * so there is no need to check for blocked tasks. So check only for
928 * bogus qsmask values.
930 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
932 WARN_ON_ONCE(rnp->qsmask);
936 * Because preemptible RCU does not exist, it never has any callbacks
939 static void rcu_preempt_check_callbacks(void)
944 * Wait for an rcu-preempt grace period, but make it happen quickly.
945 * But because preemptible RCU does not exist, map to rcu-sched.
947 void synchronize_rcu_expedited(void)
949 synchronize_sched_expedited();
951 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
954 * Because preemptible RCU does not exist, rcu_barrier() is just
955 * another name for rcu_barrier_sched().
957 void rcu_barrier(void)
961 EXPORT_SYMBOL_GPL(rcu_barrier);
964 * Because preemptible RCU does not exist, it need not be initialized.
966 static void __init __rcu_init_preempt(void)
971 * Because preemptible RCU does not exist, tasks cannot possibly exit
972 * while in preemptible RCU read-side critical sections.
978 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
980 #ifdef CONFIG_RCU_BOOST
982 #include "../locking/rtmutex_common.h"
984 #ifdef CONFIG_RCU_TRACE
986 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
988 if (!rcu_preempt_has_tasks(rnp))
989 rnp->n_balk_blkd_tasks++;
990 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
991 rnp->n_balk_exp_gp_tasks++;
992 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
993 rnp->n_balk_boost_tasks++;
994 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
995 rnp->n_balk_notblocked++;
996 else if (rnp->gp_tasks != NULL &&
997 ULONG_CMP_LT(jiffies, rnp->boost_time))
998 rnp->n_balk_notyet++;
1003 #else /* #ifdef CONFIG_RCU_TRACE */
1005 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1009 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1011 static void rcu_wake_cond(struct task_struct *t, int status)
1014 * If the thread is yielding, only wake it when this
1015 * is invoked from idle
1017 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1022 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1023 * or ->boost_tasks, advancing the pointer to the next task in the
1024 * ->blkd_tasks list.
1026 * Note that irqs must be enabled: boosting the task can block.
1027 * Returns 1 if there are more tasks needing to be boosted.
1029 static int rcu_boost(struct rcu_node *rnp)
1031 unsigned long flags;
1032 struct task_struct *t;
1033 struct list_head *tb;
1035 if (ACCESS_ONCE(rnp->exp_tasks) == NULL &&
1036 ACCESS_ONCE(rnp->boost_tasks) == NULL)
1037 return 0; /* Nothing left to boost. */
1039 raw_spin_lock_irqsave(&rnp->lock, flags);
1040 smp_mb__after_unlock_lock();
1043 * Recheck under the lock: all tasks in need of boosting
1044 * might exit their RCU read-side critical sections on their own.
1046 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1047 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1052 * Preferentially boost tasks blocking expedited grace periods.
1053 * This cannot starve the normal grace periods because a second
1054 * expedited grace period must boost all blocked tasks, including
1055 * those blocking the pre-existing normal grace period.
1057 if (rnp->exp_tasks != NULL) {
1058 tb = rnp->exp_tasks;
1059 rnp->n_exp_boosts++;
1061 tb = rnp->boost_tasks;
1062 rnp->n_normal_boosts++;
1064 rnp->n_tasks_boosted++;
1067 * We boost task t by manufacturing an rt_mutex that appears to
1068 * be held by task t. We leave a pointer to that rt_mutex where
1069 * task t can find it, and task t will release the mutex when it
1070 * exits its outermost RCU read-side critical section. Then
1071 * simply acquiring this artificial rt_mutex will boost task
1072 * t's priority. (Thanks to tglx for suggesting this approach!)
1074 * Note that task t must acquire rnp->lock to remove itself from
1075 * the ->blkd_tasks list, which it will do from exit() if from
1076 * nowhere else. We therefore are guaranteed that task t will
1077 * stay around at least until we drop rnp->lock. Note that
1078 * rnp->lock also resolves races between our priority boosting
1079 * and task t's exiting its outermost RCU read-side critical
1082 t = container_of(tb, struct task_struct, rcu_node_entry);
1083 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1084 init_completion(&rnp->boost_completion);
1085 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1086 /* Lock only for side effect: boosts task t's priority. */
1087 rt_mutex_lock(&rnp->boost_mtx);
1088 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1090 /* Wait for boostee to be done w/boost_mtx before reinitializing. */
1091 wait_for_completion(&rnp->boost_completion);
1093 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1094 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1098 * Priority-boosting kthread. One per leaf rcu_node and one for the
1101 static int rcu_boost_kthread(void *arg)
1103 struct rcu_node *rnp = (struct rcu_node *)arg;
1107 trace_rcu_utilization(TPS("Start boost kthread@init"));
1109 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1110 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1111 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1112 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1113 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1114 more2boost = rcu_boost(rnp);
1120 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1121 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1122 schedule_timeout_interruptible(2);
1123 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1128 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1133 * Check to see if it is time to start boosting RCU readers that are
1134 * blocking the current grace period, and, if so, tell the per-rcu_node
1135 * kthread to start boosting them. If there is an expedited grace
1136 * period in progress, it is always time to boost.
1138 * The caller must hold rnp->lock, which this function releases.
1139 * The ->boost_kthread_task is immortal, so we don't need to worry
1140 * about it going away.
1142 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1143 __releases(rnp->lock)
1145 struct task_struct *t;
1147 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1148 rnp->n_balk_exp_gp_tasks++;
1149 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1152 if (rnp->exp_tasks != NULL ||
1153 (rnp->gp_tasks != NULL &&
1154 rnp->boost_tasks == NULL &&
1156 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1157 if (rnp->exp_tasks == NULL)
1158 rnp->boost_tasks = rnp->gp_tasks;
1159 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1160 t = rnp->boost_kthread_task;
1162 rcu_wake_cond(t, rnp->boost_kthread_status);
1164 rcu_initiate_boost_trace(rnp);
1165 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1170 * Wake up the per-CPU kthread to invoke RCU callbacks.
1172 static void invoke_rcu_callbacks_kthread(void)
1174 unsigned long flags;
1176 local_irq_save(flags);
1177 __this_cpu_write(rcu_cpu_has_work, 1);
1178 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1179 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1180 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1181 __this_cpu_read(rcu_cpu_kthread_status));
1183 local_irq_restore(flags);
1187 * Is the current CPU running the RCU-callbacks kthread?
1188 * Caller must have preemption disabled.
1190 static bool rcu_is_callbacks_kthread(void)
1192 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1195 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1198 * Do priority-boost accounting for the start of a new grace period.
1200 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1202 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1206 * Create an RCU-boost kthread for the specified node if one does not
1207 * already exist. We only create this kthread for preemptible RCU.
1208 * Returns zero if all is well, a negated errno otherwise.
1210 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1211 struct rcu_node *rnp)
1213 int rnp_index = rnp - &rsp->node[0];
1214 unsigned long flags;
1215 struct sched_param sp;
1216 struct task_struct *t;
1218 if (&rcu_preempt_state != rsp)
1221 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1225 if (rnp->boost_kthread_task != NULL)
1227 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1228 "rcub/%d", rnp_index);
1231 raw_spin_lock_irqsave(&rnp->lock, flags);
1232 smp_mb__after_unlock_lock();
1233 rnp->boost_kthread_task = t;
1234 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1235 sp.sched_priority = kthread_prio;
1236 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1237 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1241 static void rcu_kthread_do_work(void)
1243 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1244 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1245 rcu_preempt_do_callbacks();
1248 static void rcu_cpu_kthread_setup(unsigned int cpu)
1250 struct sched_param sp;
1252 sp.sched_priority = kthread_prio;
1253 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1256 static void rcu_cpu_kthread_park(unsigned int cpu)
1258 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1261 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1263 return __this_cpu_read(rcu_cpu_has_work);
1267 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1268 * RCU softirq used in flavors and configurations of RCU that do not
1269 * support RCU priority boosting.
1271 static void rcu_cpu_kthread(unsigned int cpu)
1273 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1274 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1277 for (spincnt = 0; spincnt < 10; spincnt++) {
1278 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1280 *statusp = RCU_KTHREAD_RUNNING;
1281 this_cpu_inc(rcu_cpu_kthread_loops);
1282 local_irq_disable();
1287 rcu_kthread_do_work();
1290 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1291 *statusp = RCU_KTHREAD_WAITING;
1295 *statusp = RCU_KTHREAD_YIELDING;
1296 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1297 schedule_timeout_interruptible(2);
1298 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1299 *statusp = RCU_KTHREAD_WAITING;
1303 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1304 * served by the rcu_node in question. The CPU hotplug lock is still
1305 * held, so the value of rnp->qsmaskinit will be stable.
1307 * We don't include outgoingcpu in the affinity set, use -1 if there is
1308 * no outgoing CPU. If there are no CPUs left in the affinity set,
1309 * this function allows the kthread to execute on any CPU.
1311 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1313 struct task_struct *t = rnp->boost_kthread_task;
1314 unsigned long mask = rnp->qsmaskinit;
1320 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1322 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1323 if ((mask & 0x1) && cpu != outgoingcpu)
1324 cpumask_set_cpu(cpu, cm);
1325 if (cpumask_weight(cm) == 0) {
1327 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1328 cpumask_clear_cpu(cpu, cm);
1329 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1331 set_cpus_allowed_ptr(t, cm);
1332 free_cpumask_var(cm);
1335 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1336 .store = &rcu_cpu_kthread_task,
1337 .thread_should_run = rcu_cpu_kthread_should_run,
1338 .thread_fn = rcu_cpu_kthread,
1339 .thread_comm = "rcuc/%u",
1340 .setup = rcu_cpu_kthread_setup,
1341 .park = rcu_cpu_kthread_park,
1345 * Spawn boost kthreads -- called as soon as the scheduler is running.
1347 static void __init rcu_spawn_boost_kthreads(void)
1349 struct rcu_node *rnp;
1352 for_each_possible_cpu(cpu)
1353 per_cpu(rcu_cpu_has_work, cpu) = 0;
1354 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1355 rcu_for_each_leaf_node(rcu_state_p, rnp)
1356 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1359 static void rcu_prepare_kthreads(int cpu)
1361 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1362 struct rcu_node *rnp = rdp->mynode;
1364 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1365 if (rcu_scheduler_fully_active)
1366 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1369 #else /* #ifdef CONFIG_RCU_BOOST */
1371 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1372 __releases(rnp->lock)
1374 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1377 static void invoke_rcu_callbacks_kthread(void)
1382 static bool rcu_is_callbacks_kthread(void)
1387 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1391 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1395 static void __init rcu_spawn_boost_kthreads(void)
1399 static void rcu_prepare_kthreads(int cpu)
1403 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1405 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1408 * Check to see if any future RCU-related work will need to be done
1409 * by the current CPU, even if none need be done immediately, returning
1410 * 1 if so. This function is part of the RCU implementation; it is -not-
1411 * an exported member of the RCU API.
1413 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1414 * any flavor of RCU.
1416 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1417 int rcu_needs_cpu(unsigned long *delta_jiffies)
1419 *delta_jiffies = ULONG_MAX;
1420 return rcu_cpu_has_callbacks(NULL);
1422 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1425 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1428 static void rcu_cleanup_after_idle(void)
1433 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1436 static void rcu_prepare_for_idle(void)
1441 * Don't bother keeping a running count of the number of RCU callbacks
1442 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1444 static void rcu_idle_count_callbacks_posted(void)
1448 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1451 * This code is invoked when a CPU goes idle, at which point we want
1452 * to have the CPU do everything required for RCU so that it can enter
1453 * the energy-efficient dyntick-idle mode. This is handled by a
1454 * state machine implemented by rcu_prepare_for_idle() below.
1456 * The following three proprocessor symbols control this state machine:
1458 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1459 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1460 * is sized to be roughly one RCU grace period. Those energy-efficiency
1461 * benchmarkers who might otherwise be tempted to set this to a large
1462 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1463 * system. And if you are -that- concerned about energy efficiency,
1464 * just power the system down and be done with it!
1465 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1466 * permitted to sleep in dyntick-idle mode with only lazy RCU
1467 * callbacks pending. Setting this too high can OOM your system.
1469 * The values below work well in practice. If future workloads require
1470 * adjustment, they can be converted into kernel config parameters, though
1471 * making the state machine smarter might be a better option.
1473 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1474 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1476 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1477 module_param(rcu_idle_gp_delay, int, 0644);
1478 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1479 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1481 extern int tick_nohz_active;
1484 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1485 * only if it has been awhile since the last time we did so. Afterwards,
1486 * if there are any callbacks ready for immediate invocation, return true.
1488 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1490 bool cbs_ready = false;
1491 struct rcu_data *rdp;
1492 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1493 struct rcu_node *rnp;
1494 struct rcu_state *rsp;
1496 /* Exit early if we advanced recently. */
1497 if (jiffies == rdtp->last_advance_all)
1499 rdtp->last_advance_all = jiffies;
1501 for_each_rcu_flavor(rsp) {
1502 rdp = this_cpu_ptr(rsp->rda);
1506 * Don't bother checking unless a grace period has
1507 * completed since we last checked and there are
1508 * callbacks not yet ready to invoke.
1510 if (rdp->completed != rnp->completed &&
1511 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1512 note_gp_changes(rsp, rdp);
1514 if (cpu_has_callbacks_ready_to_invoke(rdp))
1521 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1522 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1523 * caller to set the timeout based on whether or not there are non-lazy
1526 * The caller must have disabled interrupts.
1528 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1529 int rcu_needs_cpu(unsigned long *dj)
1531 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1533 /* Snapshot to detect later posting of non-lazy callback. */
1534 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1536 /* If no callbacks, RCU doesn't need the CPU. */
1537 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1542 /* Attempt to advance callbacks. */
1543 if (rcu_try_advance_all_cbs()) {
1544 /* Some ready to invoke, so initiate later invocation. */
1548 rdtp->last_accelerate = jiffies;
1550 /* Request timer delay depending on laziness, and round. */
1551 if (!rdtp->all_lazy) {
1552 *dj = round_up(rcu_idle_gp_delay + jiffies,
1553 rcu_idle_gp_delay) - jiffies;
1555 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1559 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1562 * Prepare a CPU for idle from an RCU perspective. The first major task
1563 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1564 * The second major task is to check to see if a non-lazy callback has
1565 * arrived at a CPU that previously had only lazy callbacks. The third
1566 * major task is to accelerate (that is, assign grace-period numbers to)
1567 * any recently arrived callbacks.
1569 * The caller must have disabled interrupts.
1571 static void rcu_prepare_for_idle(void)
1573 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1575 struct rcu_data *rdp;
1576 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1577 struct rcu_node *rnp;
1578 struct rcu_state *rsp;
1581 /* Handle nohz enablement switches conservatively. */
1582 tne = ACCESS_ONCE(tick_nohz_active);
1583 if (tne != rdtp->tick_nohz_enabled_snap) {
1584 if (rcu_cpu_has_callbacks(NULL))
1585 invoke_rcu_core(); /* force nohz to see update. */
1586 rdtp->tick_nohz_enabled_snap = tne;
1592 /* If this is a no-CBs CPU, no callbacks, just return. */
1593 if (rcu_is_nocb_cpu(smp_processor_id()))
1597 * If a non-lazy callback arrived at a CPU having only lazy
1598 * callbacks, invoke RCU core for the side-effect of recalculating
1599 * idle duration on re-entry to idle.
1601 if (rdtp->all_lazy &&
1602 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1603 rdtp->all_lazy = false;
1604 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1610 * If we have not yet accelerated this jiffy, accelerate all
1611 * callbacks on this CPU.
1613 if (rdtp->last_accelerate == jiffies)
1615 rdtp->last_accelerate = jiffies;
1616 for_each_rcu_flavor(rsp) {
1617 rdp = this_cpu_ptr(rsp->rda);
1618 if (!*rdp->nxttail[RCU_DONE_TAIL])
1621 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1622 smp_mb__after_unlock_lock();
1623 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1624 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1626 rcu_gp_kthread_wake(rsp);
1628 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1632 * Clean up for exit from idle. Attempt to advance callbacks based on
1633 * any grace periods that elapsed while the CPU was idle, and if any
1634 * callbacks are now ready to invoke, initiate invocation.
1636 static void rcu_cleanup_after_idle(void)
1638 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1639 if (rcu_is_nocb_cpu(smp_processor_id()))
1641 if (rcu_try_advance_all_cbs())
1643 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1647 * Keep a running count of the number of non-lazy callbacks posted
1648 * on this CPU. This running counter (which is never decremented) allows
1649 * rcu_prepare_for_idle() to detect when something out of the idle loop
1650 * posts a callback, even if an equal number of callbacks are invoked.
1651 * Of course, callbacks should only be posted from within a trace event
1652 * designed to be called from idle or from within RCU_NONIDLE().
1654 static void rcu_idle_count_callbacks_posted(void)
1656 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1660 * Data for flushing lazy RCU callbacks at OOM time.
1662 static atomic_t oom_callback_count;
1663 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1666 * RCU OOM callback -- decrement the outstanding count and deliver the
1667 * wake-up if we are the last one.
1669 static void rcu_oom_callback(struct rcu_head *rhp)
1671 if (atomic_dec_and_test(&oom_callback_count))
1672 wake_up(&oom_callback_wq);
1676 * Post an rcu_oom_notify callback on the current CPU if it has at
1677 * least one lazy callback. This will unnecessarily post callbacks
1678 * to CPUs that already have a non-lazy callback at the end of their
1679 * callback list, but this is an infrequent operation, so accept some
1680 * extra overhead to keep things simple.
1682 static void rcu_oom_notify_cpu(void *unused)
1684 struct rcu_state *rsp;
1685 struct rcu_data *rdp;
1687 for_each_rcu_flavor(rsp) {
1688 rdp = raw_cpu_ptr(rsp->rda);
1689 if (rdp->qlen_lazy != 0) {
1690 atomic_inc(&oom_callback_count);
1691 rsp->call(&rdp->oom_head, rcu_oom_callback);
1697 * If low on memory, ensure that each CPU has a non-lazy callback.
1698 * This will wake up CPUs that have only lazy callbacks, in turn
1699 * ensuring that they free up the corresponding memory in a timely manner.
1700 * Because an uncertain amount of memory will be freed in some uncertain
1701 * timeframe, we do not claim to have freed anything.
1703 static int rcu_oom_notify(struct notifier_block *self,
1704 unsigned long notused, void *nfreed)
1708 /* Wait for callbacks from earlier instance to complete. */
1709 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1710 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1713 * Prevent premature wakeup: ensure that all increments happen
1714 * before there is a chance of the counter reaching zero.
1716 atomic_set(&oom_callback_count, 1);
1719 for_each_online_cpu(cpu) {
1720 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1721 cond_resched_rcu_qs();
1725 /* Unconditionally decrement: no need to wake ourselves up. */
1726 atomic_dec(&oom_callback_count);
1731 static struct notifier_block rcu_oom_nb = {
1732 .notifier_call = rcu_oom_notify
1735 static int __init rcu_register_oom_notifier(void)
1737 register_oom_notifier(&rcu_oom_nb);
1740 early_initcall(rcu_register_oom_notifier);
1742 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1744 #ifdef CONFIG_RCU_CPU_STALL_INFO
1746 #ifdef CONFIG_RCU_FAST_NO_HZ
1748 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1750 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1751 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1753 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1754 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1756 rdtp->all_lazy ? 'L' : '.',
1757 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1760 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1762 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1767 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1769 /* Initiate the stall-info list. */
1770 static void print_cpu_stall_info_begin(void)
1776 * Print out diagnostic information for the specified stalled CPU.
1778 * If the specified CPU is aware of the current RCU grace period
1779 * (flavor specified by rsp), then print the number of scheduling
1780 * clock interrupts the CPU has taken during the time that it has
1781 * been aware. Otherwise, print the number of RCU grace periods
1782 * that this CPU is ignorant of, for example, "1" if the CPU was
1783 * aware of the previous grace period.
1785 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1787 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1789 char fast_no_hz[72];
1790 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1791 struct rcu_dynticks *rdtp = rdp->dynticks;
1793 unsigned long ticks_value;
1795 if (rsp->gpnum == rdp->gpnum) {
1796 ticks_title = "ticks this GP";
1797 ticks_value = rdp->ticks_this_gp;
1799 ticks_title = "GPs behind";
1800 ticks_value = rsp->gpnum - rdp->gpnum;
1802 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1803 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1804 cpu, ticks_value, ticks_title,
1805 atomic_read(&rdtp->dynticks) & 0xfff,
1806 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1807 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1811 /* Terminate the stall-info list. */
1812 static void print_cpu_stall_info_end(void)
1817 /* Zero ->ticks_this_gp for all flavors of RCU. */
1818 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1820 rdp->ticks_this_gp = 0;
1821 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1824 /* Increment ->ticks_this_gp for all flavors of RCU. */
1825 static void increment_cpu_stall_ticks(void)
1827 struct rcu_state *rsp;
1829 for_each_rcu_flavor(rsp)
1830 raw_cpu_inc(rsp->rda->ticks_this_gp);
1833 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1835 static void print_cpu_stall_info_begin(void)
1840 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1842 pr_cont(" %d", cpu);
1845 static void print_cpu_stall_info_end(void)
1850 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1854 static void increment_cpu_stall_ticks(void)
1858 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1860 #ifdef CONFIG_RCU_NOCB_CPU
1863 * Offload callback processing from the boot-time-specified set of CPUs
1864 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1865 * kthread created that pulls the callbacks from the corresponding CPU,
1866 * waits for a grace period to elapse, and invokes the callbacks.
1867 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1868 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1869 * has been specified, in which case each kthread actively polls its
1870 * CPU. (Which isn't so great for energy efficiency, but which does
1871 * reduce RCU's overhead on that CPU.)
1873 * This is intended to be used in conjunction with Frederic Weisbecker's
1874 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1875 * running CPU-bound user-mode computations.
1877 * Offloading of callback processing could also in theory be used as
1878 * an energy-efficiency measure because CPUs with no RCU callbacks
1879 * queued are more aggressive about entering dyntick-idle mode.
1883 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1884 static int __init rcu_nocb_setup(char *str)
1886 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1887 have_rcu_nocb_mask = true;
1888 cpulist_parse(str, rcu_nocb_mask);
1891 __setup("rcu_nocbs=", rcu_nocb_setup);
1893 static int __init parse_rcu_nocb_poll(char *arg)
1898 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1901 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1904 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1906 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1910 * Set the root rcu_node structure's ->need_future_gp field
1911 * based on the sum of those of all rcu_node structures. This does
1912 * double-count the root rcu_node structure's requests, but this
1913 * is necessary to handle the possibility of a rcu_nocb_kthread()
1914 * having awakened during the time that the rcu_node structures
1915 * were being updated for the end of the previous grace period.
1917 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1919 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1922 static void rcu_init_one_nocb(struct rcu_node *rnp)
1924 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1925 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1928 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1929 /* Is the specified CPU a no-CBs CPU? */
1930 bool rcu_is_nocb_cpu(int cpu)
1932 if (have_rcu_nocb_mask)
1933 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1936 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1939 * Kick the leader kthread for this NOCB group.
1941 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1943 struct rcu_data *rdp_leader = rdp->nocb_leader;
1945 if (!ACCESS_ONCE(rdp_leader->nocb_kthread))
1947 if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1948 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1949 ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false;
1950 wake_up(&rdp_leader->nocb_wq);
1955 * Does the specified CPU need an RCU callback for the specified flavor
1958 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1960 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1961 struct rcu_head *rhp;
1963 /* No-CBs CPUs might have callbacks on any of three lists. */
1964 rhp = ACCESS_ONCE(rdp->nocb_head);
1966 rhp = ACCESS_ONCE(rdp->nocb_gp_head);
1968 rhp = ACCESS_ONCE(rdp->nocb_follower_head);
1970 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1971 if (!ACCESS_ONCE(rdp->nocb_kthread) && rhp) {
1972 /* RCU callback enqueued before CPU first came online??? */
1973 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1982 * Enqueue the specified string of rcu_head structures onto the specified
1983 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1984 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1985 * counts are supplied by rhcount and rhcount_lazy.
1987 * If warranted, also wake up the kthread servicing this CPUs queues.
1989 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1990 struct rcu_head *rhp,
1991 struct rcu_head **rhtp,
1992 int rhcount, int rhcount_lazy,
1993 unsigned long flags)
1996 struct rcu_head **old_rhpp;
1997 struct task_struct *t;
1999 /* Enqueue the callback on the nocb list and update counts. */
2000 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2001 ACCESS_ONCE(*old_rhpp) = rhp;
2002 atomic_long_add(rhcount, &rdp->nocb_q_count);
2003 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2004 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
2006 /* If we are not being polled and there is a kthread, awaken it ... */
2007 t = ACCESS_ONCE(rdp->nocb_kthread);
2008 if (rcu_nocb_poll || !t) {
2009 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2010 TPS("WakeNotPoll"));
2013 len = atomic_long_read(&rdp->nocb_q_count);
2014 if (old_rhpp == &rdp->nocb_head) {
2015 if (!irqs_disabled_flags(flags)) {
2016 /* ... if queue was empty ... */
2017 wake_nocb_leader(rdp, false);
2018 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2021 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
2022 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2023 TPS("WakeEmptyIsDeferred"));
2025 rdp->qlen_last_fqs_check = 0;
2026 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2027 /* ... or if many callbacks queued. */
2028 if (!irqs_disabled_flags(flags)) {
2029 wake_nocb_leader(rdp, true);
2030 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2033 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
2034 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2035 TPS("WakeOvfIsDeferred"));
2037 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2039 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2045 * This is a helper for __call_rcu(), which invokes this when the normal
2046 * callback queue is inoperable. If this is not a no-CBs CPU, this
2047 * function returns failure back to __call_rcu(), which can complain
2050 * Otherwise, this function queues the callback where the corresponding
2051 * "rcuo" kthread can find it.
2053 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2054 bool lazy, unsigned long flags)
2057 if (!rcu_is_nocb_cpu(rdp->cpu))
2059 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2060 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2061 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2062 (unsigned long)rhp->func,
2063 -atomic_long_read(&rdp->nocb_q_count_lazy),
2064 -atomic_long_read(&rdp->nocb_q_count));
2066 trace_rcu_callback(rdp->rsp->name, rhp,
2067 -atomic_long_read(&rdp->nocb_q_count_lazy),
2068 -atomic_long_read(&rdp->nocb_q_count));
2071 * If called from an extended quiescent state with interrupts
2072 * disabled, invoke the RCU core in order to allow the idle-entry
2073 * deferred-wakeup check to function.
2075 if (irqs_disabled_flags(flags) &&
2076 !rcu_is_watching() &&
2077 cpu_online(smp_processor_id()))
2084 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2087 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2088 struct rcu_data *rdp,
2089 unsigned long flags)
2091 long ql = rsp->qlen;
2092 long qll = rsp->qlen_lazy;
2094 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2095 if (!rcu_is_nocb_cpu(smp_processor_id()))
2100 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2101 if (rsp->orphan_donelist != NULL) {
2102 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2103 rsp->orphan_donetail, ql, qll, flags);
2105 rsp->orphan_donelist = NULL;
2106 rsp->orphan_donetail = &rsp->orphan_donelist;
2108 if (rsp->orphan_nxtlist != NULL) {
2109 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2110 rsp->orphan_nxttail, ql, qll, flags);
2112 rsp->orphan_nxtlist = NULL;
2113 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2119 * If necessary, kick off a new grace period, and either way wait
2120 * for a subsequent grace period to complete.
2122 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2126 unsigned long flags;
2128 struct rcu_node *rnp = rdp->mynode;
2130 raw_spin_lock_irqsave(&rnp->lock, flags);
2131 smp_mb__after_unlock_lock();
2132 needwake = rcu_start_future_gp(rnp, rdp, &c);
2133 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2135 rcu_gp_kthread_wake(rdp->rsp);
2138 * Wait for the grace period. Do so interruptibly to avoid messing
2139 * up the load average.
2141 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2143 wait_event_interruptible(
2144 rnp->nocb_gp_wq[c & 0x1],
2145 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2148 WARN_ON(signal_pending(current));
2149 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2151 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2152 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2156 * Leaders come here to wait for additional callbacks to show up.
2157 * This function does not return until callbacks appear.
2159 static void nocb_leader_wait(struct rcu_data *my_rdp)
2161 bool firsttime = true;
2163 struct rcu_data *rdp;
2164 struct rcu_head **tail;
2168 /* Wait for callbacks to appear. */
2169 if (!rcu_nocb_poll) {
2170 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2171 wait_event_interruptible(my_rdp->nocb_wq,
2172 !ACCESS_ONCE(my_rdp->nocb_leader_sleep));
2173 /* Memory barrier handled by smp_mb() calls below and repoll. */
2174 } else if (firsttime) {
2175 firsttime = false; /* Don't drown trace log with "Poll"! */
2176 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2180 * Each pass through the following loop checks a follower for CBs.
2181 * We are our own first follower. Any CBs found are moved to
2182 * nocb_gp_head, where they await a grace period.
2185 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2186 rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head);
2187 if (!rdp->nocb_gp_head)
2188 continue; /* No CBs here, try next follower. */
2190 /* Move callbacks to wait-for-GP list, which is empty. */
2191 ACCESS_ONCE(rdp->nocb_head) = NULL;
2192 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2193 rdp->nocb_gp_count = atomic_long_xchg(&rdp->nocb_q_count, 0);
2194 rdp->nocb_gp_count_lazy =
2195 atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2200 * If there were no callbacks, sleep a bit, rescan after a
2201 * memory barrier, and go retry.
2203 if (unlikely(!gotcbs)) {
2205 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2207 WARN_ON(signal_pending(current));
2208 schedule_timeout_interruptible(1);
2210 /* Rescan in case we were a victim of memory ordering. */
2211 my_rdp->nocb_leader_sleep = true;
2212 smp_mb(); /* Ensure _sleep true before scan. */
2213 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2214 if (ACCESS_ONCE(rdp->nocb_head)) {
2215 /* Found CB, so short-circuit next wait. */
2216 my_rdp->nocb_leader_sleep = false;
2222 /* Wait for one grace period. */
2223 rcu_nocb_wait_gp(my_rdp);
2226 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2227 * We set it now, but recheck for new callbacks while
2228 * traversing our follower list.
2230 my_rdp->nocb_leader_sleep = true;
2231 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2233 /* Each pass through the following loop wakes a follower, if needed. */
2234 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2235 if (ACCESS_ONCE(rdp->nocb_head))
2236 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2237 if (!rdp->nocb_gp_head)
2238 continue; /* No CBs, so no need to wake follower. */
2240 /* Append callbacks to follower's "done" list. */
2241 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2242 *tail = rdp->nocb_gp_head;
2243 atomic_long_add(rdp->nocb_gp_count, &rdp->nocb_follower_count);
2244 atomic_long_add(rdp->nocb_gp_count_lazy,
2245 &rdp->nocb_follower_count_lazy);
2246 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2247 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2249 * List was empty, wake up the follower.
2250 * Memory barriers supplied by atomic_long_add().
2252 wake_up(&rdp->nocb_wq);
2256 /* If we (the leader) don't have CBs, go wait some more. */
2257 if (!my_rdp->nocb_follower_head)
2262 * Followers come here to wait for additional callbacks to show up.
2263 * This function does not return until callbacks appear.
2265 static void nocb_follower_wait(struct rcu_data *rdp)
2267 bool firsttime = true;
2270 if (!rcu_nocb_poll) {
2271 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2273 wait_event_interruptible(rdp->nocb_wq,
2274 ACCESS_ONCE(rdp->nocb_follower_head));
2275 } else if (firsttime) {
2276 /* Don't drown trace log with "Poll"! */
2278 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2280 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2281 /* ^^^ Ensure CB invocation follows _head test. */
2285 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2287 WARN_ON(signal_pending(current));
2288 schedule_timeout_interruptible(1);
2293 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2294 * callbacks queued by the corresponding no-CBs CPU, however, there is
2295 * an optional leader-follower relationship so that the grace-period
2296 * kthreads don't have to do quite so many wakeups.
2298 static int rcu_nocb_kthread(void *arg)
2301 struct rcu_head *list;
2302 struct rcu_head *next;
2303 struct rcu_head **tail;
2304 struct rcu_data *rdp = arg;
2306 /* Each pass through this loop invokes one batch of callbacks */
2308 /* Wait for callbacks. */
2309 if (rdp->nocb_leader == rdp)
2310 nocb_leader_wait(rdp);
2312 nocb_follower_wait(rdp);
2314 /* Pull the ready-to-invoke callbacks onto local list. */
2315 list = ACCESS_ONCE(rdp->nocb_follower_head);
2317 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2318 ACCESS_ONCE(rdp->nocb_follower_head) = NULL;
2319 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2320 c = atomic_long_xchg(&rdp->nocb_follower_count, 0);
2321 cl = atomic_long_xchg(&rdp->nocb_follower_count_lazy, 0);
2322 rdp->nocb_p_count += c;
2323 rdp->nocb_p_count_lazy += cl;
2325 /* Each pass through the following loop invokes a callback. */
2326 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2330 /* Wait for enqueuing to complete, if needed. */
2331 while (next == NULL && &list->next != tail) {
2332 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2334 schedule_timeout_interruptible(1);
2335 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2339 debug_rcu_head_unqueue(list);
2341 if (__rcu_reclaim(rdp->rsp->name, list))
2347 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2348 ACCESS_ONCE(rdp->nocb_p_count) = rdp->nocb_p_count - c;
2349 ACCESS_ONCE(rdp->nocb_p_count_lazy) =
2350 rdp->nocb_p_count_lazy - cl;
2351 rdp->n_nocbs_invoked += c;
2356 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2357 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2359 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2362 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2363 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2367 if (!rcu_nocb_need_deferred_wakeup(rdp))
2369 ndw = ACCESS_ONCE(rdp->nocb_defer_wakeup);
2370 ACCESS_ONCE(rdp->nocb_defer_wakeup) = RCU_NOGP_WAKE_NOT;
2371 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2372 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2375 void __init rcu_init_nohz(void)
2378 bool need_rcu_nocb_mask = true;
2379 struct rcu_state *rsp;
2381 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2382 need_rcu_nocb_mask = false;
2383 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2385 #if defined(CONFIG_NO_HZ_FULL)
2386 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2387 need_rcu_nocb_mask = true;
2388 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2390 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2391 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2392 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2395 have_rcu_nocb_mask = true;
2397 if (!have_rcu_nocb_mask)
2400 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2401 pr_info("\tOffload RCU callbacks from CPU 0\n");
2402 cpumask_set_cpu(0, rcu_nocb_mask);
2403 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2404 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2405 pr_info("\tOffload RCU callbacks from all CPUs\n");
2406 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2407 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2408 #if defined(CONFIG_NO_HZ_FULL)
2409 if (tick_nohz_full_running)
2410 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2411 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2413 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2414 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2415 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2418 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
2419 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
2421 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2423 for_each_rcu_flavor(rsp) {
2424 for_each_cpu(cpu, rcu_nocb_mask) {
2425 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2428 * If there are early callbacks, they will need
2429 * to be moved to the nocb lists.
2431 WARN_ON_ONCE(rdp->nxttail[RCU_NEXT_TAIL] !=
2433 rdp->nxttail[RCU_NEXT_TAIL] != NULL);
2434 init_nocb_callback_list(rdp);
2436 rcu_organize_nocb_kthreads(rsp);
2440 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2441 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2443 rdp->nocb_tail = &rdp->nocb_head;
2444 init_waitqueue_head(&rdp->nocb_wq);
2445 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2449 * If the specified CPU is a no-CBs CPU that does not already have its
2450 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2451 * brought online out of order, this can require re-organizing the
2452 * leader-follower relationships.
2454 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2456 struct rcu_data *rdp;
2457 struct rcu_data *rdp_last;
2458 struct rcu_data *rdp_old_leader;
2459 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2460 struct task_struct *t;
2463 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2464 * then nothing to do.
2466 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2469 /* If we didn't spawn the leader first, reorganize! */
2470 rdp_old_leader = rdp_spawn->nocb_leader;
2471 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2473 rdp = rdp_old_leader;
2475 rdp->nocb_leader = rdp_spawn;
2476 if (rdp_last && rdp != rdp_spawn)
2477 rdp_last->nocb_next_follower = rdp;
2478 if (rdp == rdp_spawn) {
2479 rdp = rdp->nocb_next_follower;
2482 rdp = rdp->nocb_next_follower;
2483 rdp_last->nocb_next_follower = NULL;
2486 rdp_spawn->nocb_next_follower = rdp_old_leader;
2489 /* Spawn the kthread for this CPU and RCU flavor. */
2490 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2491 "rcuo%c/%d", rsp->abbr, cpu);
2493 ACCESS_ONCE(rdp_spawn->nocb_kthread) = t;
2497 * If the specified CPU is a no-CBs CPU that does not already have its
2498 * rcuo kthreads, spawn them.
2500 static void rcu_spawn_all_nocb_kthreads(int cpu)
2502 struct rcu_state *rsp;
2504 if (rcu_scheduler_fully_active)
2505 for_each_rcu_flavor(rsp)
2506 rcu_spawn_one_nocb_kthread(rsp, cpu);
2510 * Once the scheduler is running, spawn rcuo kthreads for all online
2511 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2512 * non-boot CPUs come online -- if this changes, we will need to add
2513 * some mutual exclusion.
2515 static void __init rcu_spawn_nocb_kthreads(void)
2519 for_each_online_cpu(cpu)
2520 rcu_spawn_all_nocb_kthreads(cpu);
2523 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2524 static int rcu_nocb_leader_stride = -1;
2525 module_param(rcu_nocb_leader_stride, int, 0444);
2528 * Initialize leader-follower relationships for all no-CBs CPU.
2530 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2533 int ls = rcu_nocb_leader_stride;
2534 int nl = 0; /* Next leader. */
2535 struct rcu_data *rdp;
2536 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2537 struct rcu_data *rdp_prev = NULL;
2539 if (!have_rcu_nocb_mask)
2542 ls = int_sqrt(nr_cpu_ids);
2543 rcu_nocb_leader_stride = ls;
2547 * Each pass through this loop sets up one rcu_data structure and
2548 * spawns one rcu_nocb_kthread().
2550 for_each_cpu(cpu, rcu_nocb_mask) {
2551 rdp = per_cpu_ptr(rsp->rda, cpu);
2552 if (rdp->cpu >= nl) {
2553 /* New leader, set up for followers & next leader. */
2554 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2555 rdp->nocb_leader = rdp;
2558 /* Another follower, link to previous leader. */
2559 rdp->nocb_leader = rdp_leader;
2560 rdp_prev->nocb_next_follower = rdp;
2566 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2567 static bool init_nocb_callback_list(struct rcu_data *rdp)
2569 if (!rcu_is_nocb_cpu(rdp->cpu))
2572 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2576 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2578 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2580 WARN_ON_ONCE(1); /* Should be dead code. */
2584 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2588 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2592 static void rcu_init_one_nocb(struct rcu_node *rnp)
2596 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2597 bool lazy, unsigned long flags)
2602 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2603 struct rcu_data *rdp,
2604 unsigned long flags)
2609 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2613 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2618 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2622 static void rcu_spawn_all_nocb_kthreads(int cpu)
2626 static void __init rcu_spawn_nocb_kthreads(void)
2630 static bool init_nocb_callback_list(struct rcu_data *rdp)
2635 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2638 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2639 * arbitrarily long period of time with the scheduling-clock tick turned
2640 * off. RCU will be paying attention to this CPU because it is in the
2641 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2642 * machine because the scheduling-clock tick has been disabled. Therefore,
2643 * if an adaptive-ticks CPU is failing to respond to the current grace
2644 * period and has not be idle from an RCU perspective, kick it.
2646 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2648 #ifdef CONFIG_NO_HZ_FULL
2649 if (tick_nohz_full_cpu(cpu))
2650 smp_send_reschedule(cpu);
2651 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2655 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2657 static int full_sysidle_state; /* Current system-idle state. */
2658 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2659 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2660 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2661 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2662 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2665 * Invoked to note exit from irq or task transition to idle. Note that
2666 * usermode execution does -not- count as idle here! After all, we want
2667 * to detect full-system idle states, not RCU quiescent states and grace
2668 * periods. The caller must have disabled interrupts.
2670 static void rcu_sysidle_enter(int irq)
2673 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2675 /* If there are no nohz_full= CPUs, no need to track this. */
2676 if (!tick_nohz_full_enabled())
2679 /* Adjust nesting, check for fully idle. */
2681 rdtp->dynticks_idle_nesting--;
2682 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2683 if (rdtp->dynticks_idle_nesting != 0)
2684 return; /* Still not fully idle. */
2686 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2687 DYNTICK_TASK_NEST_VALUE) {
2688 rdtp->dynticks_idle_nesting = 0;
2690 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2691 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2692 return; /* Still not fully idle. */
2696 /* Record start of fully idle period. */
2698 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2699 smp_mb__before_atomic();
2700 atomic_inc(&rdtp->dynticks_idle);
2701 smp_mb__after_atomic();
2702 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2706 * Unconditionally force exit from full system-idle state. This is
2707 * invoked when a normal CPU exits idle, but must be called separately
2708 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2709 * is that the timekeeping CPU is permitted to take scheduling-clock
2710 * interrupts while the system is in system-idle state, and of course
2711 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2712 * interrupt from any other type of interrupt.
2714 void rcu_sysidle_force_exit(void)
2716 int oldstate = ACCESS_ONCE(full_sysidle_state);
2720 * Each pass through the following loop attempts to exit full
2721 * system-idle state. If contention proves to be a problem,
2722 * a trylock-based contention tree could be used here.
2724 while (oldstate > RCU_SYSIDLE_SHORT) {
2725 newoldstate = cmpxchg(&full_sysidle_state,
2726 oldstate, RCU_SYSIDLE_NOT);
2727 if (oldstate == newoldstate &&
2728 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2729 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2730 return; /* We cleared it, done! */
2732 oldstate = newoldstate;
2734 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2738 * Invoked to note entry to irq or task transition from idle. Note that
2739 * usermode execution does -not- count as idle here! The caller must
2740 * have disabled interrupts.
2742 static void rcu_sysidle_exit(int irq)
2744 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2746 /* If there are no nohz_full= CPUs, no need to track this. */
2747 if (!tick_nohz_full_enabled())
2750 /* Adjust nesting, check for already non-idle. */
2752 rdtp->dynticks_idle_nesting++;
2753 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2754 if (rdtp->dynticks_idle_nesting != 1)
2755 return; /* Already non-idle. */
2758 * Allow for irq misnesting. Yes, it really is possible
2759 * to enter an irq handler then never leave it, and maybe
2760 * also vice versa. Handle both possibilities.
2762 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2763 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2764 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2765 return; /* Already non-idle. */
2767 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2771 /* Record end of idle period. */
2772 smp_mb__before_atomic();
2773 atomic_inc(&rdtp->dynticks_idle);
2774 smp_mb__after_atomic();
2775 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2778 * If we are the timekeeping CPU, we are permitted to be non-idle
2779 * during a system-idle state. This must be the case, because
2780 * the timekeeping CPU has to take scheduling-clock interrupts
2781 * during the time that the system is transitioning to full
2782 * system-idle state. This means that the timekeeping CPU must
2783 * invoke rcu_sysidle_force_exit() directly if it does anything
2784 * more than take a scheduling-clock interrupt.
2786 if (smp_processor_id() == tick_do_timer_cpu)
2789 /* Update system-idle state: We are clearly no longer fully idle! */
2790 rcu_sysidle_force_exit();
2794 * Check to see if the current CPU is idle. Note that usermode execution
2795 * does not count as idle. The caller must have disabled interrupts.
2797 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2798 unsigned long *maxj)
2802 struct rcu_dynticks *rdtp = rdp->dynticks;
2804 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2805 if (!tick_nohz_full_enabled())
2809 * If some other CPU has already reported non-idle, if this is
2810 * not the flavor of RCU that tracks sysidle state, or if this
2811 * is an offline or the timekeeping CPU, nothing to do.
2813 if (!*isidle || rdp->rsp != rcu_state_p ||
2814 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2816 if (rcu_gp_in_progress(rdp->rsp))
2817 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2819 /* Pick up current idle and NMI-nesting counter and check. */
2820 cur = atomic_read(&rdtp->dynticks_idle);
2822 *isidle = false; /* We are not idle! */
2825 smp_mb(); /* Read counters before timestamps. */
2827 /* Pick up timestamps. */
2828 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2829 /* If this CPU entered idle more recently, update maxj timestamp. */
2830 if (ULONG_CMP_LT(*maxj, j))
2835 * Is this the flavor of RCU that is handling full-system idle?
2837 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2839 return rsp == rcu_state_p;
2843 * Return a delay in jiffies based on the number of CPUs, rcu_node
2844 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2845 * systems more time to transition to full-idle state in order to
2846 * avoid the cache thrashing that otherwise occur on the state variable.
2847 * Really small systems (less than a couple of tens of CPUs) should
2848 * instead use a single global atomically incremented counter, and later
2849 * versions of this will automatically reconfigure themselves accordingly.
2851 static unsigned long rcu_sysidle_delay(void)
2853 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2855 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2859 * Advance the full-system-idle state. This is invoked when all of
2860 * the non-timekeeping CPUs are idle.
2862 static void rcu_sysidle(unsigned long j)
2864 /* Check the current state. */
2865 switch (ACCESS_ONCE(full_sysidle_state)) {
2866 case RCU_SYSIDLE_NOT:
2868 /* First time all are idle, so note a short idle period. */
2869 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2872 case RCU_SYSIDLE_SHORT:
2875 * Idle for a bit, time to advance to next state?
2876 * cmpxchg failure means race with non-idle, let them win.
2878 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2879 (void)cmpxchg(&full_sysidle_state,
2880 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2883 case RCU_SYSIDLE_LONG:
2886 * Do an additional check pass before advancing to full.
2887 * cmpxchg failure means race with non-idle, let them win.
2889 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2890 (void)cmpxchg(&full_sysidle_state,
2891 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2900 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2901 * back to the beginning.
2903 static void rcu_sysidle_cancel(void)
2906 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2907 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2911 * Update the sysidle state based on the results of a force-quiescent-state
2912 * scan of the CPUs' dyntick-idle state.
2914 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2915 unsigned long maxj, bool gpkt)
2917 if (rsp != rcu_state_p)
2918 return; /* Wrong flavor, ignore. */
2919 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2920 return; /* Running state machine from timekeeping CPU. */
2922 rcu_sysidle(maxj); /* More idle! */
2924 rcu_sysidle_cancel(); /* Idle is over. */
2928 * Wrapper for rcu_sysidle_report() when called from the grace-period
2929 * kthread's context.
2931 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2934 /* If there are no nohz_full= CPUs, no need to track this. */
2935 if (!tick_nohz_full_enabled())
2938 rcu_sysidle_report(rsp, isidle, maxj, true);
2941 /* Callback and function for forcing an RCU grace period. */
2942 struct rcu_sysidle_head {
2947 static void rcu_sysidle_cb(struct rcu_head *rhp)
2949 struct rcu_sysidle_head *rshp;
2952 * The following memory barrier is needed to replace the
2953 * memory barriers that would normally be in the memory
2956 smp_mb(); /* grace period precedes setting inuse. */
2958 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2959 ACCESS_ONCE(rshp->inuse) = 0;
2963 * Check to see if the system is fully idle, other than the timekeeping CPU.
2964 * The caller must have disabled interrupts. This is not intended to be
2965 * called unless tick_nohz_full_enabled().
2967 bool rcu_sys_is_idle(void)
2969 static struct rcu_sysidle_head rsh;
2970 int rss = ACCESS_ONCE(full_sysidle_state);
2972 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2975 /* Handle small-system case by doing a full scan of CPUs. */
2976 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2977 int oldrss = rss - 1;
2980 * One pass to advance to each state up to _FULL.
2981 * Give up if any pass fails to advance the state.
2983 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2986 unsigned long maxj = jiffies - ULONG_MAX / 4;
2987 struct rcu_data *rdp;
2989 /* Scan all the CPUs looking for nonidle CPUs. */
2990 for_each_possible_cpu(cpu) {
2991 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2992 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2996 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2998 rss = ACCESS_ONCE(full_sysidle_state);
3002 /* If this is the first observation of an idle period, record it. */
3003 if (rss == RCU_SYSIDLE_FULL) {
3004 rss = cmpxchg(&full_sysidle_state,
3005 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
3006 return rss == RCU_SYSIDLE_FULL;
3009 smp_mb(); /* ensure rss load happens before later caller actions. */
3011 /* If already fully idle, tell the caller (in case of races). */
3012 if (rss == RCU_SYSIDLE_FULL_NOTED)
3016 * If we aren't there yet, and a grace period is not in flight,
3017 * initiate a grace period. Either way, tell the caller that
3018 * we are not there yet. We use an xchg() rather than an assignment
3019 * to make up for the memory barriers that would otherwise be
3020 * provided by the memory allocator.
3022 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
3023 !rcu_gp_in_progress(rcu_state_p) &&
3024 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
3025 call_rcu(&rsh.rh, rcu_sysidle_cb);
3030 * Initialize dynticks sysidle state for CPUs coming online.
3032 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3034 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
3037 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3039 static void rcu_sysidle_enter(int irq)
3043 static void rcu_sysidle_exit(int irq)
3047 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3048 unsigned long *maxj)
3052 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3057 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3062 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3066 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3069 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3070 * grace-period kthread will do force_quiescent_state() processing?
3071 * The idea is to avoid waking up RCU core processing on such a
3072 * CPU unless the grace period has extended for too long.
3074 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3075 * CONFIG_RCU_NOCB_CPU CPUs.
3077 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3079 #ifdef CONFIG_NO_HZ_FULL
3080 if (tick_nohz_full_cpu(smp_processor_id()) &&
3081 (!rcu_gp_in_progress(rsp) ||
3082 ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
3084 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3089 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3092 static void rcu_bind_gp_kthread(void)
3094 int __maybe_unused cpu;
3096 if (!tick_nohz_full_enabled())
3098 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3099 cpu = tick_do_timer_cpu;
3100 if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu)
3101 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3102 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3103 if (!is_housekeeping_cpu(raw_smp_processor_id()))
3104 housekeeping_affine(current);
3105 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3108 /* Record the current task on dyntick-idle entry. */
3109 static void rcu_dynticks_task_enter(void)
3111 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3112 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = smp_processor_id();
3113 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3116 /* Record no current task on dyntick-idle exit. */
3117 static void rcu_dynticks_task_exit(void)
3119 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3120 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = -1;
3121 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */