2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
31 #include "../time/tick-internal.h"
33 #ifdef CONFIG_RCU_BOOST
35 #include "../locking/rtmutex_common.h"
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
41 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
42 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
43 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
44 DEFINE_PER_CPU(char, rcu_cpu_has_work);
46 #endif /* #ifdef CONFIG_RCU_BOOST */
48 #ifdef CONFIG_RCU_NOCB_CPU
49 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
50 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
51 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
52 static char __initdata nocb_buf[NR_CPUS * 5];
53 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
56 * Check the RCU kernel configuration parameters and print informative
57 * messages about anything out of the ordinary. If you like #ifdef, you
58 * will love this function.
60 static void __init rcu_bootup_announce_oddness(void)
62 #ifdef CONFIG_RCU_TRACE
63 pr_info("\tRCU debugfs-based tracing is enabled.\n");
65 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
66 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
69 #ifdef CONFIG_RCU_FANOUT_EXACT
70 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
72 #ifdef CONFIG_RCU_FAST_NO_HZ
73 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
75 #ifdef CONFIG_PROVE_RCU
76 pr_info("\tRCU lockdep checking is enabled.\n");
78 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
79 pr_info("\tRCU torture testing starts during boot.\n");
81 #if defined(CONFIG_RCU_CPU_STALL_INFO)
82 pr_info("\tAdditional per-CPU info printed with stalls.\n");
84 #if NUM_RCU_LVL_4 != 0
85 pr_info("\tFour-level hierarchy is enabled.\n");
87 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
88 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
89 if (nr_cpu_ids != NR_CPUS)
90 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
91 #ifdef CONFIG_RCU_BOOST
92 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
96 #ifdef CONFIG_PREEMPT_RCU
98 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
99 static struct rcu_state *rcu_state_p = &rcu_preempt_state;
101 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
104 * Tell them what RCU they are running.
106 static void __init rcu_bootup_announce(void)
108 pr_info("Preemptible hierarchical RCU implementation.\n");
109 rcu_bootup_announce_oddness();
113 * Return the number of RCU-preempt batches processed thus far
114 * for debug and statistics.
116 static long rcu_batches_completed_preempt(void)
118 return rcu_preempt_state.completed;
120 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
123 * Return the number of RCU batches processed thus far for debug & stats.
125 long rcu_batches_completed(void)
127 return rcu_batches_completed_preempt();
129 EXPORT_SYMBOL_GPL(rcu_batches_completed);
132 * Record a preemptible-RCU quiescent state for the specified CPU. Note
133 * that this just means that the task currently running on the CPU is
134 * not in a quiescent state. There might be any number of tasks blocked
135 * while in an RCU read-side critical section.
137 * As with the other rcu_*_qs() functions, callers to this function
138 * must disable preemption.
140 static void rcu_preempt_qs(void)
142 if (!__this_cpu_read(rcu_preempt_data.passed_quiesce)) {
143 trace_rcu_grace_period(TPS("rcu_preempt"),
144 __this_cpu_read(rcu_preempt_data.gpnum),
146 __this_cpu_write(rcu_preempt_data.passed_quiesce, 1);
147 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
148 current->rcu_read_unlock_special.b.need_qs = false;
153 * We have entered the scheduler, and the current task might soon be
154 * context-switched away from. If this task is in an RCU read-side
155 * critical section, we will no longer be able to rely on the CPU to
156 * record that fact, so we enqueue the task on the blkd_tasks list.
157 * The task will dequeue itself when it exits the outermost enclosing
158 * RCU read-side critical section. Therefore, the current grace period
159 * cannot be permitted to complete until the blkd_tasks list entries
160 * predating the current grace period drain, in other words, until
161 * rnp->gp_tasks becomes NULL.
163 * Caller must disable preemption.
165 static void rcu_preempt_note_context_switch(void)
167 struct task_struct *t = current;
169 struct rcu_data *rdp;
170 struct rcu_node *rnp;
172 if (t->rcu_read_lock_nesting > 0 &&
173 !t->rcu_read_unlock_special.b.blocked) {
175 /* Possibly blocking in an RCU read-side critical section. */
176 rdp = this_cpu_ptr(rcu_preempt_state.rda);
178 raw_spin_lock_irqsave(&rnp->lock, flags);
179 smp_mb__after_unlock_lock();
180 t->rcu_read_unlock_special.b.blocked = true;
181 t->rcu_blocked_node = rnp;
184 * If this CPU has already checked in, then this task
185 * will hold up the next grace period rather than the
186 * current grace period. Queue the task accordingly.
187 * If the task is queued for the current grace period
188 * (i.e., this CPU has not yet passed through a quiescent
189 * state for the current grace period), then as long
190 * as that task remains queued, the current grace period
191 * cannot end. Note that there is some uncertainty as
192 * to exactly when the current grace period started.
193 * We take a conservative approach, which can result
194 * in unnecessarily waiting on tasks that started very
195 * slightly after the current grace period began. C'est
198 * But first, note that the current CPU must still be
201 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
202 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
203 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
204 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
205 rnp->gp_tasks = &t->rcu_node_entry;
206 #ifdef CONFIG_RCU_BOOST
207 if (rnp->boost_tasks != NULL)
208 rnp->boost_tasks = rnp->gp_tasks;
209 #endif /* #ifdef CONFIG_RCU_BOOST */
211 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
212 if (rnp->qsmask & rdp->grpmask)
213 rnp->gp_tasks = &t->rcu_node_entry;
215 trace_rcu_preempt_task(rdp->rsp->name,
217 (rnp->qsmask & rdp->grpmask)
220 raw_spin_unlock_irqrestore(&rnp->lock, flags);
221 } else if (t->rcu_read_lock_nesting < 0 &&
222 t->rcu_read_unlock_special.s) {
225 * Complete exit from RCU read-side critical section on
226 * behalf of preempted instance of __rcu_read_unlock().
228 rcu_read_unlock_special(t);
232 * Either we were not in an RCU read-side critical section to
233 * begin with, or we have now recorded that critical section
234 * globally. Either way, we can now note a quiescent state
235 * for this CPU. Again, if we were in an RCU read-side critical
236 * section, and if that critical section was blocking the current
237 * grace period, then the fact that the task has been enqueued
238 * means that we continue to block the current grace period.
244 * Check for preempted RCU readers blocking the current grace period
245 * for the specified rcu_node structure. If the caller needs a reliable
246 * answer, it must hold the rcu_node's ->lock.
248 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
250 return rnp->gp_tasks != NULL;
254 * Record a quiescent state for all tasks that were previously queued
255 * on the specified rcu_node structure and that were blocking the current
256 * RCU grace period. The caller must hold the specified rnp->lock with
257 * irqs disabled, and this lock is released upon return, but irqs remain
260 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
261 __releases(rnp->lock)
264 struct rcu_node *rnp_p;
266 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
267 raw_spin_unlock_irqrestore(&rnp->lock, flags);
268 return; /* Still need more quiescent states! */
274 * Either there is only one rcu_node in the tree,
275 * or tasks were kicked up to root rcu_node due to
276 * CPUs going offline.
278 rcu_report_qs_rsp(&rcu_preempt_state, flags);
282 /* Report up the rest of the hierarchy. */
284 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
285 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
286 smp_mb__after_unlock_lock();
287 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
291 * Advance a ->blkd_tasks-list pointer to the next entry, instead
292 * returning NULL if at the end of the list.
294 static struct list_head *rcu_next_node_entry(struct task_struct *t,
295 struct rcu_node *rnp)
297 struct list_head *np;
299 np = t->rcu_node_entry.next;
300 if (np == &rnp->blkd_tasks)
306 * Handle special cases during rcu_read_unlock(), such as needing to
307 * notify RCU core processing or task having blocked during the RCU
308 * read-side critical section.
310 void rcu_read_unlock_special(struct task_struct *t)
316 struct list_head *np;
317 #ifdef CONFIG_RCU_BOOST
318 bool drop_boost_mutex = false;
319 #endif /* #ifdef CONFIG_RCU_BOOST */
320 struct rcu_node *rnp;
321 union rcu_special special;
323 /* NMI handlers cannot block and cannot safely manipulate state. */
327 local_irq_save(flags);
330 * If RCU core is waiting for this CPU to exit critical section,
331 * let it know that we have done so. Because irqs are disabled,
332 * t->rcu_read_unlock_special cannot change.
334 special = t->rcu_read_unlock_special;
335 if (special.b.need_qs) {
337 if (!t->rcu_read_unlock_special.s) {
338 local_irq_restore(flags);
343 /* Hardware IRQ handlers cannot block, complain if they get here. */
344 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
345 local_irq_restore(flags);
349 /* Clean up if blocked during RCU read-side critical section. */
350 if (special.b.blocked) {
351 t->rcu_read_unlock_special.b.blocked = false;
354 * Remove this task from the list it blocked on. The
355 * task can migrate while we acquire the lock, but at
356 * most one time. So at most two passes through loop.
359 rnp = t->rcu_blocked_node;
360 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
361 smp_mb__after_unlock_lock();
362 if (rnp == t->rcu_blocked_node)
364 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
366 empty = !rcu_preempt_blocked_readers_cgp(rnp);
367 empty_exp = !rcu_preempted_readers_exp(rnp);
368 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
369 np = rcu_next_node_entry(t, rnp);
370 list_del_init(&t->rcu_node_entry);
371 t->rcu_blocked_node = NULL;
372 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
374 if (&t->rcu_node_entry == rnp->gp_tasks)
376 if (&t->rcu_node_entry == rnp->exp_tasks)
378 #ifdef CONFIG_RCU_BOOST
379 if (&t->rcu_node_entry == rnp->boost_tasks)
380 rnp->boost_tasks = np;
381 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
382 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
383 #endif /* #ifdef CONFIG_RCU_BOOST */
386 * If this was the last task on the current list, and if
387 * we aren't waiting on any CPUs, report the quiescent state.
388 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
389 * so we must take a snapshot of the expedited state.
391 empty_exp_now = !rcu_preempted_readers_exp(rnp);
392 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
393 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
400 rcu_report_unblock_qs_rnp(rnp, flags);
402 raw_spin_unlock_irqrestore(&rnp->lock, flags);
405 #ifdef CONFIG_RCU_BOOST
406 /* Unboost if we were boosted. */
407 if (drop_boost_mutex) {
408 rt_mutex_unlock(&rnp->boost_mtx);
409 complete(&rnp->boost_completion);
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 (!list_empty(&rnp->blkd_tasks))
519 rnp->gp_tasks = rnp->blkd_tasks.next;
520 WARN_ON_ONCE(rnp->qsmask);
523 #ifdef CONFIG_HOTPLUG_CPU
526 * Handle tasklist migration for case in which all CPUs covered by the
527 * specified rcu_node have gone offline. Move them up to the root
528 * rcu_node. The reason for not just moving them to the immediate
529 * parent is to remove the need for rcu_read_unlock_special() to
530 * make more than two attempts to acquire the target rcu_node's lock.
531 * Returns true if there were tasks blocking the current RCU grace
534 * Returns 1 if there was previously a task blocking the current grace
535 * period on the specified rcu_node structure.
537 * The caller must hold rnp->lock with irqs disabled.
539 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
540 struct rcu_node *rnp,
541 struct rcu_data *rdp)
543 struct list_head *lp;
544 struct list_head *lp_root;
546 struct rcu_node *rnp_root = rcu_get_root(rsp);
547 struct task_struct *t;
549 if (rnp == rnp_root) {
550 WARN_ONCE(1, "Last CPU thought to be offlined?");
551 return 0; /* Shouldn't happen: at least one CPU online. */
554 /* If we are on an internal node, complain bitterly. */
555 WARN_ON_ONCE(rnp != rdp->mynode);
558 * Move tasks up to root rcu_node. Don't try to get fancy for
559 * this corner-case operation -- just put this node's tasks
560 * at the head of the root node's list, and update the root node's
561 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
562 * if non-NULL. This might result in waiting for more tasks than
563 * absolutely necessary, but this is a good performance/complexity
566 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
567 retval |= RCU_OFL_TASKS_NORM_GP;
568 if (rcu_preempted_readers_exp(rnp))
569 retval |= RCU_OFL_TASKS_EXP_GP;
570 lp = &rnp->blkd_tasks;
571 lp_root = &rnp_root->blkd_tasks;
572 while (!list_empty(lp)) {
573 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
574 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
575 smp_mb__after_unlock_lock();
576 list_del(&t->rcu_node_entry);
577 t->rcu_blocked_node = rnp_root;
578 list_add(&t->rcu_node_entry, lp_root);
579 if (&t->rcu_node_entry == rnp->gp_tasks)
580 rnp_root->gp_tasks = rnp->gp_tasks;
581 if (&t->rcu_node_entry == rnp->exp_tasks)
582 rnp_root->exp_tasks = rnp->exp_tasks;
583 #ifdef CONFIG_RCU_BOOST
584 if (&t->rcu_node_entry == rnp->boost_tasks)
585 rnp_root->boost_tasks = rnp->boost_tasks;
586 #endif /* #ifdef CONFIG_RCU_BOOST */
587 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
590 rnp->gp_tasks = NULL;
591 rnp->exp_tasks = NULL;
592 #ifdef CONFIG_RCU_BOOST
593 rnp->boost_tasks = NULL;
595 * In case root is being boosted and leaf was not. Make sure
596 * that we boost the tasks blocking the current grace period
599 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
600 smp_mb__after_unlock_lock();
601 if (rnp_root->boost_tasks != NULL &&
602 rnp_root->boost_tasks != rnp_root->gp_tasks &&
603 rnp_root->boost_tasks != rnp_root->exp_tasks)
604 rnp_root->boost_tasks = rnp_root->gp_tasks;
605 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
606 #endif /* #ifdef CONFIG_RCU_BOOST */
611 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
614 * Check for a quiescent state from the current CPU. When a task blocks,
615 * the task is recorded in the corresponding CPU's rcu_node structure,
616 * which is checked elsewhere.
618 * Caller must disable hard irqs.
620 static void rcu_preempt_check_callbacks(void)
622 struct task_struct *t = current;
624 if (t->rcu_read_lock_nesting == 0) {
628 if (t->rcu_read_lock_nesting > 0 &&
629 __this_cpu_read(rcu_preempt_data.qs_pending) &&
630 !__this_cpu_read(rcu_preempt_data.passed_quiesce))
631 t->rcu_read_unlock_special.b.need_qs = true;
634 #ifdef CONFIG_RCU_BOOST
636 static void rcu_preempt_do_callbacks(void)
638 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
641 #endif /* #ifdef CONFIG_RCU_BOOST */
644 * Queue a preemptible-RCU callback for invocation after a grace period.
646 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
648 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
650 EXPORT_SYMBOL_GPL(call_rcu);
653 * synchronize_rcu - wait until a grace period has elapsed.
655 * Control will return to the caller some time after a full grace
656 * period has elapsed, in other words after all currently executing RCU
657 * read-side critical sections have completed. Note, however, that
658 * upon return from synchronize_rcu(), the caller might well be executing
659 * concurrently with new RCU read-side critical sections that began while
660 * synchronize_rcu() was waiting. RCU read-side critical sections are
661 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
663 * See the description of synchronize_sched() for more detailed information
664 * on memory ordering guarantees.
666 void synchronize_rcu(void)
668 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
669 !lock_is_held(&rcu_lock_map) &&
670 !lock_is_held(&rcu_sched_lock_map),
671 "Illegal synchronize_rcu() in RCU read-side critical section");
672 if (!rcu_scheduler_active)
675 synchronize_rcu_expedited();
677 wait_rcu_gp(call_rcu);
679 EXPORT_SYMBOL_GPL(synchronize_rcu);
681 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
682 static unsigned long sync_rcu_preempt_exp_count;
683 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
686 * Return non-zero if there are any tasks in RCU read-side critical
687 * sections blocking the current preemptible-RCU expedited grace period.
688 * If there is no preemptible-RCU expedited grace period currently in
689 * progress, returns zero unconditionally.
691 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
693 return rnp->exp_tasks != NULL;
697 * return non-zero if there is no RCU expedited grace period in progress
698 * for the specified rcu_node structure, in other words, if all CPUs and
699 * tasks covered by the specified rcu_node structure have done their bit
700 * for the current expedited grace period. Works only for preemptible
701 * RCU -- other RCU implementation use other means.
703 * Caller must hold sync_rcu_preempt_exp_mutex.
705 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
707 return !rcu_preempted_readers_exp(rnp) &&
708 ACCESS_ONCE(rnp->expmask) == 0;
712 * Report the exit from RCU read-side critical section for the last task
713 * that queued itself during or before the current expedited preemptible-RCU
714 * grace period. This event is reported either to the rcu_node structure on
715 * which the task was queued or to one of that rcu_node structure's ancestors,
716 * recursively up the tree. (Calm down, calm down, we do the recursion
719 * Most callers will set the "wake" flag, but the task initiating the
720 * expedited grace period need not wake itself.
722 * Caller must hold sync_rcu_preempt_exp_mutex.
724 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
730 raw_spin_lock_irqsave(&rnp->lock, flags);
731 smp_mb__after_unlock_lock();
733 if (!sync_rcu_preempt_exp_done(rnp)) {
734 raw_spin_unlock_irqrestore(&rnp->lock, flags);
737 if (rnp->parent == NULL) {
738 raw_spin_unlock_irqrestore(&rnp->lock, flags);
740 smp_mb(); /* EGP done before wake_up(). */
741 wake_up(&sync_rcu_preempt_exp_wq);
746 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
748 raw_spin_lock(&rnp->lock); /* irqs already disabled */
749 smp_mb__after_unlock_lock();
750 rnp->expmask &= ~mask;
755 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
756 * grace period for the specified rcu_node structure. If there are no such
757 * tasks, report it up the rcu_node hierarchy.
759 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
760 * CPU hotplug operations.
763 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
768 raw_spin_lock_irqsave(&rnp->lock, flags);
769 smp_mb__after_unlock_lock();
770 if (list_empty(&rnp->blkd_tasks)) {
771 raw_spin_unlock_irqrestore(&rnp->lock, flags);
773 rnp->exp_tasks = rnp->blkd_tasks.next;
774 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
778 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
782 * synchronize_rcu_expedited - Brute-force RCU grace period
784 * Wait for an RCU-preempt grace period, but expedite it. The basic
785 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
786 * the ->blkd_tasks lists and wait for this list to drain. This consumes
787 * significant time on all CPUs and is unfriendly to real-time workloads,
788 * so is thus not recommended for any sort of common-case code.
789 * In fact, if you are using synchronize_rcu_expedited() in a loop,
790 * please restructure your code to batch your updates, and then Use a
791 * single synchronize_rcu() instead.
793 void synchronize_rcu_expedited(void)
796 struct rcu_node *rnp;
797 struct rcu_state *rsp = &rcu_preempt_state;
801 smp_mb(); /* Caller's modifications seen first by other CPUs. */
802 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
803 smp_mb(); /* Above access cannot bleed into critical section. */
806 * Block CPU-hotplug operations. This means that any CPU-hotplug
807 * operation that finds an rcu_node structure with tasks in the
808 * process of being boosted will know that all tasks blocking
809 * this expedited grace period will already be in the process of
810 * being boosted. This simplifies the process of moving tasks
811 * from leaf to root rcu_node structures.
813 if (!try_get_online_cpus()) {
814 /* CPU-hotplug operation in flight, fall back to normal GP. */
815 wait_rcu_gp(call_rcu);
820 * Acquire lock, falling back to synchronize_rcu() if too many
821 * lock-acquisition failures. Of course, if someone does the
822 * expedited grace period for us, just leave.
824 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
825 if (ULONG_CMP_LT(snap,
826 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
828 goto mb_ret; /* Others did our work for us. */
830 if (trycount++ < 10) {
831 udelay(trycount * num_online_cpus());
834 wait_rcu_gp(call_rcu);
838 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
840 goto unlock_mb_ret; /* Others did our work for us. */
843 /* force all RCU readers onto ->blkd_tasks lists. */
844 synchronize_sched_expedited();
846 /* Initialize ->expmask for all non-leaf rcu_node structures. */
847 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
848 raw_spin_lock_irqsave(&rnp->lock, flags);
849 smp_mb__after_unlock_lock();
850 rnp->expmask = rnp->qsmaskinit;
851 raw_spin_unlock_irqrestore(&rnp->lock, flags);
854 /* Snapshot current state of ->blkd_tasks lists. */
855 rcu_for_each_leaf_node(rsp, rnp)
856 sync_rcu_preempt_exp_init(rsp, rnp);
857 if (NUM_RCU_NODES > 1)
858 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
862 /* Wait for snapshotted ->blkd_tasks lists to drain. */
863 rnp = rcu_get_root(rsp);
864 wait_event(sync_rcu_preempt_exp_wq,
865 sync_rcu_preempt_exp_done(rnp));
867 /* Clean up and exit. */
868 smp_mb(); /* ensure expedited GP seen before counter increment. */
869 ACCESS_ONCE(sync_rcu_preempt_exp_count) =
870 sync_rcu_preempt_exp_count + 1;
872 mutex_unlock(&sync_rcu_preempt_exp_mutex);
874 smp_mb(); /* ensure subsequent action seen after grace period. */
876 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
879 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
881 * Note that this primitive does not necessarily wait for an RCU grace period
882 * to complete. For example, if there are no RCU callbacks queued anywhere
883 * in the system, then rcu_barrier() is within its rights to return
884 * immediately, without waiting for anything, much less an RCU grace period.
886 void rcu_barrier(void)
888 _rcu_barrier(&rcu_preempt_state);
890 EXPORT_SYMBOL_GPL(rcu_barrier);
893 * Initialize preemptible RCU's state structures.
895 static void __init __rcu_init_preempt(void)
897 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
901 * Check for a task exiting while in a preemptible-RCU read-side
902 * critical section, clean up if so. No need to issue warnings,
903 * as debug_check_no_locks_held() already does this if lockdep
908 struct task_struct *t = current;
910 if (likely(list_empty(¤t->rcu_node_entry)))
912 t->rcu_read_lock_nesting = 1;
914 t->rcu_read_unlock_special.b.blocked = true;
918 #else /* #ifdef CONFIG_PREEMPT_RCU */
920 static struct rcu_state *rcu_state_p = &rcu_sched_state;
923 * Tell them what RCU they are running.
925 static void __init rcu_bootup_announce(void)
927 pr_info("Hierarchical RCU implementation.\n");
928 rcu_bootup_announce_oddness();
932 * Return the number of RCU batches processed thus far for debug & stats.
934 long rcu_batches_completed(void)
936 return rcu_batches_completed_sched();
938 EXPORT_SYMBOL_GPL(rcu_batches_completed);
941 * Because preemptible RCU does not exist, we never have to check for
942 * CPUs being in quiescent states.
944 static void rcu_preempt_note_context_switch(void)
949 * Because preemptible RCU does not exist, there are never any preempted
952 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
957 #ifdef CONFIG_HOTPLUG_CPU
959 /* Because preemptible RCU does not exist, no quieting of tasks. */
960 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
961 __releases(rnp->lock)
963 raw_spin_unlock_irqrestore(&rnp->lock, flags);
966 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
969 * Because preemptible RCU does not exist, we never have to check for
970 * tasks blocked within RCU read-side critical sections.
972 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
977 * Because preemptible RCU does not exist, we never have to check for
978 * tasks blocked within RCU read-side critical sections.
980 static int rcu_print_task_stall(struct rcu_node *rnp)
986 * Because there is no preemptible RCU, there can be no readers blocked,
987 * so there is no need to check for blocked tasks. So check only for
988 * bogus qsmask values.
990 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
992 WARN_ON_ONCE(rnp->qsmask);
995 #ifdef CONFIG_HOTPLUG_CPU
998 * Because preemptible RCU does not exist, it never needs to migrate
999 * tasks that were blocked within RCU read-side critical sections, and
1000 * such non-existent tasks cannot possibly have been blocking the current
1003 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1004 struct rcu_node *rnp,
1005 struct rcu_data *rdp)
1010 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1013 * Because preemptible RCU does not exist, it never has any callbacks
1016 static void rcu_preempt_check_callbacks(void)
1021 * Wait for an rcu-preempt grace period, but make it happen quickly.
1022 * But because preemptible RCU does not exist, map to rcu-sched.
1024 void synchronize_rcu_expedited(void)
1026 synchronize_sched_expedited();
1028 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1030 #ifdef CONFIG_HOTPLUG_CPU
1033 * Because preemptible RCU does not exist, there is never any need to
1034 * report on tasks preempted in RCU read-side critical sections during
1035 * expedited RCU grace periods.
1037 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1042 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1045 * Because preemptible RCU does not exist, rcu_barrier() is just
1046 * another name for rcu_barrier_sched().
1048 void rcu_barrier(void)
1050 rcu_barrier_sched();
1052 EXPORT_SYMBOL_GPL(rcu_barrier);
1055 * Because preemptible RCU does not exist, it need not be initialized.
1057 static void __init __rcu_init_preempt(void)
1062 * Because preemptible RCU does not exist, tasks cannot possibly exit
1063 * while in preemptible RCU read-side critical sections.
1069 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
1071 #ifdef CONFIG_RCU_BOOST
1073 #include "../locking/rtmutex_common.h"
1075 #ifdef CONFIG_RCU_TRACE
1077 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1079 if (list_empty(&rnp->blkd_tasks))
1080 rnp->n_balk_blkd_tasks++;
1081 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1082 rnp->n_balk_exp_gp_tasks++;
1083 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1084 rnp->n_balk_boost_tasks++;
1085 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1086 rnp->n_balk_notblocked++;
1087 else if (rnp->gp_tasks != NULL &&
1088 ULONG_CMP_LT(jiffies, rnp->boost_time))
1089 rnp->n_balk_notyet++;
1094 #else /* #ifdef CONFIG_RCU_TRACE */
1096 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1100 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1102 static void rcu_wake_cond(struct task_struct *t, int status)
1105 * If the thread is yielding, only wake it when this
1106 * is invoked from idle
1108 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1113 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1114 * or ->boost_tasks, advancing the pointer to the next task in the
1115 * ->blkd_tasks list.
1117 * Note that irqs must be enabled: boosting the task can block.
1118 * Returns 1 if there are more tasks needing to be boosted.
1120 static int rcu_boost(struct rcu_node *rnp)
1122 unsigned long flags;
1123 struct task_struct *t;
1124 struct list_head *tb;
1126 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1127 return 0; /* Nothing left to boost. */
1129 raw_spin_lock_irqsave(&rnp->lock, flags);
1130 smp_mb__after_unlock_lock();
1133 * Recheck under the lock: all tasks in need of boosting
1134 * might exit their RCU read-side critical sections on their own.
1136 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1137 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1142 * Preferentially boost tasks blocking expedited grace periods.
1143 * This cannot starve the normal grace periods because a second
1144 * expedited grace period must boost all blocked tasks, including
1145 * those blocking the pre-existing normal grace period.
1147 if (rnp->exp_tasks != NULL) {
1148 tb = rnp->exp_tasks;
1149 rnp->n_exp_boosts++;
1151 tb = rnp->boost_tasks;
1152 rnp->n_normal_boosts++;
1154 rnp->n_tasks_boosted++;
1157 * We boost task t by manufacturing an rt_mutex that appears to
1158 * be held by task t. We leave a pointer to that rt_mutex where
1159 * task t can find it, and task t will release the mutex when it
1160 * exits its outermost RCU read-side critical section. Then
1161 * simply acquiring this artificial rt_mutex will boost task
1162 * t's priority. (Thanks to tglx for suggesting this approach!)
1164 * Note that task t must acquire rnp->lock to remove itself from
1165 * the ->blkd_tasks list, which it will do from exit() if from
1166 * nowhere else. We therefore are guaranteed that task t will
1167 * stay around at least until we drop rnp->lock. Note that
1168 * rnp->lock also resolves races between our priority boosting
1169 * and task t's exiting its outermost RCU read-side critical
1172 t = container_of(tb, struct task_struct, rcu_node_entry);
1173 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1174 init_completion(&rnp->boost_completion);
1175 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1176 /* Lock only for side effect: boosts task t's priority. */
1177 rt_mutex_lock(&rnp->boost_mtx);
1178 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1180 /* Wait for boostee to be done w/boost_mtx before reinitializing. */
1181 wait_for_completion(&rnp->boost_completion);
1183 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1184 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1188 * Priority-boosting kthread. One per leaf rcu_node and one for the
1191 static int rcu_boost_kthread(void *arg)
1193 struct rcu_node *rnp = (struct rcu_node *)arg;
1197 trace_rcu_utilization(TPS("Start boost kthread@init"));
1199 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1200 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1201 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1202 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1203 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1204 more2boost = rcu_boost(rnp);
1210 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1211 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1212 schedule_timeout_interruptible(2);
1213 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1218 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1223 * Check to see if it is time to start boosting RCU readers that are
1224 * blocking the current grace period, and, if so, tell the per-rcu_node
1225 * kthread to start boosting them. If there is an expedited grace
1226 * period in progress, it is always time to boost.
1228 * The caller must hold rnp->lock, which this function releases.
1229 * The ->boost_kthread_task is immortal, so we don't need to worry
1230 * about it going away.
1232 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1233 __releases(rnp->lock)
1235 struct task_struct *t;
1237 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1238 rnp->n_balk_exp_gp_tasks++;
1239 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1242 if (rnp->exp_tasks != NULL ||
1243 (rnp->gp_tasks != NULL &&
1244 rnp->boost_tasks == NULL &&
1246 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1247 if (rnp->exp_tasks == NULL)
1248 rnp->boost_tasks = rnp->gp_tasks;
1249 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1250 t = rnp->boost_kthread_task;
1252 rcu_wake_cond(t, rnp->boost_kthread_status);
1254 rcu_initiate_boost_trace(rnp);
1255 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1260 * Wake up the per-CPU kthread to invoke RCU callbacks.
1262 static void invoke_rcu_callbacks_kthread(void)
1264 unsigned long flags;
1266 local_irq_save(flags);
1267 __this_cpu_write(rcu_cpu_has_work, 1);
1268 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1269 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1270 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1271 __this_cpu_read(rcu_cpu_kthread_status));
1273 local_irq_restore(flags);
1277 * Is the current CPU running the RCU-callbacks kthread?
1278 * Caller must have preemption disabled.
1280 static bool rcu_is_callbacks_kthread(void)
1282 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1285 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1288 * Do priority-boost accounting for the start of a new grace period.
1290 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1292 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1296 * Create an RCU-boost kthread for the specified node if one does not
1297 * already exist. We only create this kthread for preemptible RCU.
1298 * Returns zero if all is well, a negated errno otherwise.
1300 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1301 struct rcu_node *rnp)
1303 int rnp_index = rnp - &rsp->node[0];
1304 unsigned long flags;
1305 struct sched_param sp;
1306 struct task_struct *t;
1308 if (&rcu_preempt_state != rsp)
1311 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1315 if (rnp->boost_kthread_task != NULL)
1317 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1318 "rcub/%d", rnp_index);
1321 raw_spin_lock_irqsave(&rnp->lock, flags);
1322 smp_mb__after_unlock_lock();
1323 rnp->boost_kthread_task = t;
1324 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1325 sp.sched_priority = kthread_prio;
1326 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1327 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1331 static void rcu_kthread_do_work(void)
1333 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1334 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1335 rcu_preempt_do_callbacks();
1338 static void rcu_cpu_kthread_setup(unsigned int cpu)
1340 struct sched_param sp;
1342 sp.sched_priority = kthread_prio;
1343 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1346 static void rcu_cpu_kthread_park(unsigned int cpu)
1348 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1351 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1353 return __this_cpu_read(rcu_cpu_has_work);
1357 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1358 * RCU softirq used in flavors and configurations of RCU that do not
1359 * support RCU priority boosting.
1361 static void rcu_cpu_kthread(unsigned int cpu)
1363 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1364 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1367 for (spincnt = 0; spincnt < 10; spincnt++) {
1368 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1370 *statusp = RCU_KTHREAD_RUNNING;
1371 this_cpu_inc(rcu_cpu_kthread_loops);
1372 local_irq_disable();
1377 rcu_kthread_do_work();
1380 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1381 *statusp = RCU_KTHREAD_WAITING;
1385 *statusp = RCU_KTHREAD_YIELDING;
1386 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1387 schedule_timeout_interruptible(2);
1388 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1389 *statusp = RCU_KTHREAD_WAITING;
1393 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1394 * served by the rcu_node in question. The CPU hotplug lock is still
1395 * held, so the value of rnp->qsmaskinit will be stable.
1397 * We don't include outgoingcpu in the affinity set, use -1 if there is
1398 * no outgoing CPU. If there are no CPUs left in the affinity set,
1399 * this function allows the kthread to execute on any CPU.
1401 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1403 struct task_struct *t = rnp->boost_kthread_task;
1404 unsigned long mask = rnp->qsmaskinit;
1410 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1412 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1413 if ((mask & 0x1) && cpu != outgoingcpu)
1414 cpumask_set_cpu(cpu, cm);
1415 if (cpumask_weight(cm) == 0) {
1417 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1418 cpumask_clear_cpu(cpu, cm);
1419 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1421 set_cpus_allowed_ptr(t, cm);
1422 free_cpumask_var(cm);
1425 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1426 .store = &rcu_cpu_kthread_task,
1427 .thread_should_run = rcu_cpu_kthread_should_run,
1428 .thread_fn = rcu_cpu_kthread,
1429 .thread_comm = "rcuc/%u",
1430 .setup = rcu_cpu_kthread_setup,
1431 .park = rcu_cpu_kthread_park,
1435 * Spawn boost kthreads -- called as soon as the scheduler is running.
1437 static void __init rcu_spawn_boost_kthreads(void)
1439 struct rcu_node *rnp;
1442 for_each_possible_cpu(cpu)
1443 per_cpu(rcu_cpu_has_work, cpu) = 0;
1444 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1445 rnp = rcu_get_root(rcu_state_p);
1446 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1447 if (NUM_RCU_NODES > 1) {
1448 rcu_for_each_leaf_node(rcu_state_p, rnp)
1449 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1453 static void rcu_prepare_kthreads(int cpu)
1455 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1456 struct rcu_node *rnp = rdp->mynode;
1458 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1459 if (rcu_scheduler_fully_active)
1460 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1463 #else /* #ifdef CONFIG_RCU_BOOST */
1465 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1466 __releases(rnp->lock)
1468 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1471 static void invoke_rcu_callbacks_kthread(void)
1476 static bool rcu_is_callbacks_kthread(void)
1481 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1485 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1489 static void __init rcu_spawn_boost_kthreads(void)
1493 static void rcu_prepare_kthreads(int cpu)
1497 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1499 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1502 * Check to see if any future RCU-related work will need to be done
1503 * by the current CPU, even if none need be done immediately, returning
1504 * 1 if so. This function is part of the RCU implementation; it is -not-
1505 * an exported member of the RCU API.
1507 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1508 * any flavor of RCU.
1510 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1511 int rcu_needs_cpu(unsigned long *delta_jiffies)
1513 *delta_jiffies = ULONG_MAX;
1514 return rcu_cpu_has_callbacks(NULL);
1516 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1519 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1522 static void rcu_cleanup_after_idle(void)
1527 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1530 static void rcu_prepare_for_idle(void)
1535 * Don't bother keeping a running count of the number of RCU callbacks
1536 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1538 static void rcu_idle_count_callbacks_posted(void)
1542 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1545 * This code is invoked when a CPU goes idle, at which point we want
1546 * to have the CPU do everything required for RCU so that it can enter
1547 * the energy-efficient dyntick-idle mode. This is handled by a
1548 * state machine implemented by rcu_prepare_for_idle() below.
1550 * The following three proprocessor symbols control this state machine:
1552 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1553 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1554 * is sized to be roughly one RCU grace period. Those energy-efficiency
1555 * benchmarkers who might otherwise be tempted to set this to a large
1556 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1557 * system. And if you are -that- concerned about energy efficiency,
1558 * just power the system down and be done with it!
1559 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1560 * permitted to sleep in dyntick-idle mode with only lazy RCU
1561 * callbacks pending. Setting this too high can OOM your system.
1563 * The values below work well in practice. If future workloads require
1564 * adjustment, they can be converted into kernel config parameters, though
1565 * making the state machine smarter might be a better option.
1567 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1568 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1570 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1571 module_param(rcu_idle_gp_delay, int, 0644);
1572 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1573 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1575 extern int tick_nohz_active;
1578 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1579 * only if it has been awhile since the last time we did so. Afterwards,
1580 * if there are any callbacks ready for immediate invocation, return true.
1582 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1584 bool cbs_ready = false;
1585 struct rcu_data *rdp;
1586 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1587 struct rcu_node *rnp;
1588 struct rcu_state *rsp;
1590 /* Exit early if we advanced recently. */
1591 if (jiffies == rdtp->last_advance_all)
1593 rdtp->last_advance_all = jiffies;
1595 for_each_rcu_flavor(rsp) {
1596 rdp = this_cpu_ptr(rsp->rda);
1600 * Don't bother checking unless a grace period has
1601 * completed since we last checked and there are
1602 * callbacks not yet ready to invoke.
1604 if ((rdp->completed != rnp->completed ||
1605 unlikely(ACCESS_ONCE(rdp->gpwrap))) &&
1606 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1607 note_gp_changes(rsp, rdp);
1609 if (cpu_has_callbacks_ready_to_invoke(rdp))
1616 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1617 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1618 * caller to set the timeout based on whether or not there are non-lazy
1621 * The caller must have disabled interrupts.
1623 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1624 int rcu_needs_cpu(unsigned long *dj)
1626 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1628 /* Snapshot to detect later posting of non-lazy callback. */
1629 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1631 /* If no callbacks, RCU doesn't need the CPU. */
1632 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1637 /* Attempt to advance callbacks. */
1638 if (rcu_try_advance_all_cbs()) {
1639 /* Some ready to invoke, so initiate later invocation. */
1643 rdtp->last_accelerate = jiffies;
1645 /* Request timer delay depending on laziness, and round. */
1646 if (!rdtp->all_lazy) {
1647 *dj = round_up(rcu_idle_gp_delay + jiffies,
1648 rcu_idle_gp_delay) - jiffies;
1650 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1654 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1657 * Prepare a CPU for idle from an RCU perspective. The first major task
1658 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1659 * The second major task is to check to see if a non-lazy callback has
1660 * arrived at a CPU that previously had only lazy callbacks. The third
1661 * major task is to accelerate (that is, assign grace-period numbers to)
1662 * any recently arrived callbacks.
1664 * The caller must have disabled interrupts.
1666 static void rcu_prepare_for_idle(void)
1668 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1670 struct rcu_data *rdp;
1671 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1672 struct rcu_node *rnp;
1673 struct rcu_state *rsp;
1676 /* Handle nohz enablement switches conservatively. */
1677 tne = ACCESS_ONCE(tick_nohz_active);
1678 if (tne != rdtp->tick_nohz_enabled_snap) {
1679 if (rcu_cpu_has_callbacks(NULL))
1680 invoke_rcu_core(); /* force nohz to see update. */
1681 rdtp->tick_nohz_enabled_snap = tne;
1687 /* If this is a no-CBs CPU, no callbacks, just return. */
1688 if (rcu_is_nocb_cpu(smp_processor_id()))
1692 * If a non-lazy callback arrived at a CPU having only lazy
1693 * callbacks, invoke RCU core for the side-effect of recalculating
1694 * idle duration on re-entry to idle.
1696 if (rdtp->all_lazy &&
1697 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1698 rdtp->all_lazy = false;
1699 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1705 * If we have not yet accelerated this jiffy, accelerate all
1706 * callbacks on this CPU.
1708 if (rdtp->last_accelerate == jiffies)
1710 rdtp->last_accelerate = jiffies;
1711 for_each_rcu_flavor(rsp) {
1712 rdp = this_cpu_ptr(rsp->rda);
1713 if (!*rdp->nxttail[RCU_DONE_TAIL])
1716 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1717 smp_mb__after_unlock_lock();
1718 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1719 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1721 rcu_gp_kthread_wake(rsp);
1723 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1727 * Clean up for exit from idle. Attempt to advance callbacks based on
1728 * any grace periods that elapsed while the CPU was idle, and if any
1729 * callbacks are now ready to invoke, initiate invocation.
1731 static void rcu_cleanup_after_idle(void)
1733 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1734 if (rcu_is_nocb_cpu(smp_processor_id()))
1736 if (rcu_try_advance_all_cbs())
1738 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1742 * Keep a running count of the number of non-lazy callbacks posted
1743 * on this CPU. This running counter (which is never decremented) allows
1744 * rcu_prepare_for_idle() to detect when something out of the idle loop
1745 * posts a callback, even if an equal number of callbacks are invoked.
1746 * Of course, callbacks should only be posted from within a trace event
1747 * designed to be called from idle or from within RCU_NONIDLE().
1749 static void rcu_idle_count_callbacks_posted(void)
1751 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1755 * Data for flushing lazy RCU callbacks at OOM time.
1757 static atomic_t oom_callback_count;
1758 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1761 * RCU OOM callback -- decrement the outstanding count and deliver the
1762 * wake-up if we are the last one.
1764 static void rcu_oom_callback(struct rcu_head *rhp)
1766 if (atomic_dec_and_test(&oom_callback_count))
1767 wake_up(&oom_callback_wq);
1771 * Post an rcu_oom_notify callback on the current CPU if it has at
1772 * least one lazy callback. This will unnecessarily post callbacks
1773 * to CPUs that already have a non-lazy callback at the end of their
1774 * callback list, but this is an infrequent operation, so accept some
1775 * extra overhead to keep things simple.
1777 static void rcu_oom_notify_cpu(void *unused)
1779 struct rcu_state *rsp;
1780 struct rcu_data *rdp;
1782 for_each_rcu_flavor(rsp) {
1783 rdp = raw_cpu_ptr(rsp->rda);
1784 if (rdp->qlen_lazy != 0) {
1785 atomic_inc(&oom_callback_count);
1786 rsp->call(&rdp->oom_head, rcu_oom_callback);
1792 * If low on memory, ensure that each CPU has a non-lazy callback.
1793 * This will wake up CPUs that have only lazy callbacks, in turn
1794 * ensuring that they free up the corresponding memory in a timely manner.
1795 * Because an uncertain amount of memory will be freed in some uncertain
1796 * timeframe, we do not claim to have freed anything.
1798 static int rcu_oom_notify(struct notifier_block *self,
1799 unsigned long notused, void *nfreed)
1803 /* Wait for callbacks from earlier instance to complete. */
1804 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1805 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1808 * Prevent premature wakeup: ensure that all increments happen
1809 * before there is a chance of the counter reaching zero.
1811 atomic_set(&oom_callback_count, 1);
1814 for_each_online_cpu(cpu) {
1815 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1816 cond_resched_rcu_qs();
1820 /* Unconditionally decrement: no need to wake ourselves up. */
1821 atomic_dec(&oom_callback_count);
1826 static struct notifier_block rcu_oom_nb = {
1827 .notifier_call = rcu_oom_notify
1830 static int __init rcu_register_oom_notifier(void)
1832 register_oom_notifier(&rcu_oom_nb);
1835 early_initcall(rcu_register_oom_notifier);
1837 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1839 #ifdef CONFIG_RCU_CPU_STALL_INFO
1841 #ifdef CONFIG_RCU_FAST_NO_HZ
1843 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1845 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1846 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1848 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1849 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1851 rdtp->all_lazy ? 'L' : '.',
1852 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1855 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1857 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1862 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1864 /* Initiate the stall-info list. */
1865 static void print_cpu_stall_info_begin(void)
1871 * Print out diagnostic information for the specified stalled CPU.
1873 * If the specified CPU is aware of the current RCU grace period
1874 * (flavor specified by rsp), then print the number of scheduling
1875 * clock interrupts the CPU has taken during the time that it has
1876 * been aware. Otherwise, print the number of RCU grace periods
1877 * that this CPU is ignorant of, for example, "1" if the CPU was
1878 * aware of the previous grace period.
1880 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1882 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1884 char fast_no_hz[72];
1885 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1886 struct rcu_dynticks *rdtp = rdp->dynticks;
1888 unsigned long ticks_value;
1890 if (rsp->gpnum == rdp->gpnum) {
1891 ticks_title = "ticks this GP";
1892 ticks_value = rdp->ticks_this_gp;
1894 ticks_title = "GPs behind";
1895 ticks_value = rsp->gpnum - rdp->gpnum;
1897 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1898 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1899 cpu, ticks_value, ticks_title,
1900 atomic_read(&rdtp->dynticks) & 0xfff,
1901 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1902 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1903 ACCESS_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1907 /* Terminate the stall-info list. */
1908 static void print_cpu_stall_info_end(void)
1913 /* Zero ->ticks_this_gp for all flavors of RCU. */
1914 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1916 rdp->ticks_this_gp = 0;
1917 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1920 /* Increment ->ticks_this_gp for all flavors of RCU. */
1921 static void increment_cpu_stall_ticks(void)
1923 struct rcu_state *rsp;
1925 for_each_rcu_flavor(rsp)
1926 raw_cpu_inc(rsp->rda->ticks_this_gp);
1929 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1931 static void print_cpu_stall_info_begin(void)
1936 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1938 pr_cont(" %d", cpu);
1941 static void print_cpu_stall_info_end(void)
1946 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1950 static void increment_cpu_stall_ticks(void)
1954 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1956 #ifdef CONFIG_RCU_NOCB_CPU
1959 * Offload callback processing from the boot-time-specified set of CPUs
1960 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1961 * kthread created that pulls the callbacks from the corresponding CPU,
1962 * waits for a grace period to elapse, and invokes the callbacks.
1963 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1964 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1965 * has been specified, in which case each kthread actively polls its
1966 * CPU. (Which isn't so great for energy efficiency, but which does
1967 * reduce RCU's overhead on that CPU.)
1969 * This is intended to be used in conjunction with Frederic Weisbecker's
1970 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1971 * running CPU-bound user-mode computations.
1973 * Offloading of callback processing could also in theory be used as
1974 * an energy-efficiency measure because CPUs with no RCU callbacks
1975 * queued are more aggressive about entering dyntick-idle mode.
1979 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1980 static int __init rcu_nocb_setup(char *str)
1982 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1983 have_rcu_nocb_mask = true;
1984 cpulist_parse(str, rcu_nocb_mask);
1987 __setup("rcu_nocbs=", rcu_nocb_setup);
1989 static int __init parse_rcu_nocb_poll(char *arg)
1994 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1997 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2000 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2002 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2006 * Set the root rcu_node structure's ->need_future_gp field
2007 * based on the sum of those of all rcu_node structures. This does
2008 * double-count the root rcu_node structure's requests, but this
2009 * is necessary to handle the possibility of a rcu_nocb_kthread()
2010 * having awakened during the time that the rcu_node structures
2011 * were being updated for the end of the previous grace period.
2013 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2015 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2018 static void rcu_init_one_nocb(struct rcu_node *rnp)
2020 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2021 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2024 #ifndef CONFIG_RCU_NOCB_CPU_ALL
2025 /* Is the specified CPU a no-CBs CPU? */
2026 bool rcu_is_nocb_cpu(int cpu)
2028 if (have_rcu_nocb_mask)
2029 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2032 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
2035 * Kick the leader kthread for this NOCB group.
2037 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
2039 struct rcu_data *rdp_leader = rdp->nocb_leader;
2041 if (!ACCESS_ONCE(rdp_leader->nocb_kthread))
2043 if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) {
2044 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
2045 ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false;
2046 wake_up(&rdp_leader->nocb_wq);
2051 * Does the specified CPU need an RCU callback for the specified flavor
2054 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2056 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2057 struct rcu_head *rhp;
2059 /* No-CBs CPUs might have callbacks on any of three lists. */
2060 rhp = ACCESS_ONCE(rdp->nocb_head);
2062 rhp = ACCESS_ONCE(rdp->nocb_gp_head);
2064 rhp = ACCESS_ONCE(rdp->nocb_follower_head);
2066 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
2067 if (!ACCESS_ONCE(rdp->nocb_kthread) && rhp) {
2068 /* RCU callback enqueued before CPU first came online??? */
2069 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
2078 * Enqueue the specified string of rcu_head structures onto the specified
2079 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2080 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2081 * counts are supplied by rhcount and rhcount_lazy.
2083 * If warranted, also wake up the kthread servicing this CPUs queues.
2085 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2086 struct rcu_head *rhp,
2087 struct rcu_head **rhtp,
2088 int rhcount, int rhcount_lazy,
2089 unsigned long flags)
2092 struct rcu_head **old_rhpp;
2093 struct task_struct *t;
2095 /* Enqueue the callback on the nocb list and update counts. */
2096 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2097 ACCESS_ONCE(*old_rhpp) = rhp;
2098 atomic_long_add(rhcount, &rdp->nocb_q_count);
2099 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2100 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
2102 /* If we are not being polled and there is a kthread, awaken it ... */
2103 t = ACCESS_ONCE(rdp->nocb_kthread);
2104 if (rcu_nocb_poll || !t) {
2105 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2106 TPS("WakeNotPoll"));
2109 len = atomic_long_read(&rdp->nocb_q_count);
2110 if (old_rhpp == &rdp->nocb_head) {
2111 if (!irqs_disabled_flags(flags)) {
2112 /* ... if queue was empty ... */
2113 wake_nocb_leader(rdp, false);
2114 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2117 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
2118 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2119 TPS("WakeEmptyIsDeferred"));
2121 rdp->qlen_last_fqs_check = 0;
2122 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2123 /* ... or if many callbacks queued. */
2124 if (!irqs_disabled_flags(flags)) {
2125 wake_nocb_leader(rdp, true);
2126 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2129 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
2130 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2131 TPS("WakeOvfIsDeferred"));
2133 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2135 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2141 * This is a helper for __call_rcu(), which invokes this when the normal
2142 * callback queue is inoperable. If this is not a no-CBs CPU, this
2143 * function returns failure back to __call_rcu(), which can complain
2146 * Otherwise, this function queues the callback where the corresponding
2147 * "rcuo" kthread can find it.
2149 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2150 bool lazy, unsigned long flags)
2153 if (!rcu_is_nocb_cpu(rdp->cpu))
2155 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2156 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2157 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2158 (unsigned long)rhp->func,
2159 -atomic_long_read(&rdp->nocb_q_count_lazy),
2160 -atomic_long_read(&rdp->nocb_q_count));
2162 trace_rcu_callback(rdp->rsp->name, rhp,
2163 -atomic_long_read(&rdp->nocb_q_count_lazy),
2164 -atomic_long_read(&rdp->nocb_q_count));
2167 * If called from an extended quiescent state with interrupts
2168 * disabled, invoke the RCU core in order to allow the idle-entry
2169 * deferred-wakeup check to function.
2171 if (irqs_disabled_flags(flags) &&
2172 !rcu_is_watching() &&
2173 cpu_online(smp_processor_id()))
2180 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2183 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2184 struct rcu_data *rdp,
2185 unsigned long flags)
2187 long ql = rsp->qlen;
2188 long qll = rsp->qlen_lazy;
2190 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2191 if (!rcu_is_nocb_cpu(smp_processor_id()))
2196 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2197 if (rsp->orphan_donelist != NULL) {
2198 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2199 rsp->orphan_donetail, ql, qll, flags);
2201 rsp->orphan_donelist = NULL;
2202 rsp->orphan_donetail = &rsp->orphan_donelist;
2204 if (rsp->orphan_nxtlist != NULL) {
2205 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2206 rsp->orphan_nxttail, ql, qll, flags);
2208 rsp->orphan_nxtlist = NULL;
2209 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2215 * If necessary, kick off a new grace period, and either way wait
2216 * for a subsequent grace period to complete.
2218 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2222 unsigned long flags;
2224 struct rcu_node *rnp = rdp->mynode;
2226 raw_spin_lock_irqsave(&rnp->lock, flags);
2227 smp_mb__after_unlock_lock();
2228 needwake = rcu_start_future_gp(rnp, rdp, &c);
2229 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2231 rcu_gp_kthread_wake(rdp->rsp);
2234 * Wait for the grace period. Do so interruptibly to avoid messing
2235 * up the load average.
2237 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2239 wait_event_interruptible(
2240 rnp->nocb_gp_wq[c & 0x1],
2241 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2244 WARN_ON(signal_pending(current));
2245 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2247 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2248 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2252 * Leaders come here to wait for additional callbacks to show up.
2253 * This function does not return until callbacks appear.
2255 static void nocb_leader_wait(struct rcu_data *my_rdp)
2257 bool firsttime = true;
2259 struct rcu_data *rdp;
2260 struct rcu_head **tail;
2264 /* Wait for callbacks to appear. */
2265 if (!rcu_nocb_poll) {
2266 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2267 wait_event_interruptible(my_rdp->nocb_wq,
2268 !ACCESS_ONCE(my_rdp->nocb_leader_sleep));
2269 /* Memory barrier handled by smp_mb() calls below and repoll. */
2270 } else if (firsttime) {
2271 firsttime = false; /* Don't drown trace log with "Poll"! */
2272 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2276 * Each pass through the following loop checks a follower for CBs.
2277 * We are our own first follower. Any CBs found are moved to
2278 * nocb_gp_head, where they await a grace period.
2281 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2282 rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head);
2283 if (!rdp->nocb_gp_head)
2284 continue; /* No CBs here, try next follower. */
2286 /* Move callbacks to wait-for-GP list, which is empty. */
2287 ACCESS_ONCE(rdp->nocb_head) = NULL;
2288 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2289 rdp->nocb_gp_count = atomic_long_xchg(&rdp->nocb_q_count, 0);
2290 rdp->nocb_gp_count_lazy =
2291 atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2296 * If there were no callbacks, sleep a bit, rescan after a
2297 * memory barrier, and go retry.
2299 if (unlikely(!gotcbs)) {
2301 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2303 WARN_ON(signal_pending(current));
2304 schedule_timeout_interruptible(1);
2306 /* Rescan in case we were a victim of memory ordering. */
2307 my_rdp->nocb_leader_sleep = true;
2308 smp_mb(); /* Ensure _sleep true before scan. */
2309 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2310 if (ACCESS_ONCE(rdp->nocb_head)) {
2311 /* Found CB, so short-circuit next wait. */
2312 my_rdp->nocb_leader_sleep = false;
2318 /* Wait for one grace period. */
2319 rcu_nocb_wait_gp(my_rdp);
2322 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2323 * We set it now, but recheck for new callbacks while
2324 * traversing our follower list.
2326 my_rdp->nocb_leader_sleep = true;
2327 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2329 /* Each pass through the following loop wakes a follower, if needed. */
2330 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2331 if (ACCESS_ONCE(rdp->nocb_head))
2332 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2333 if (!rdp->nocb_gp_head)
2334 continue; /* No CBs, so no need to wake follower. */
2336 /* Append callbacks to follower's "done" list. */
2337 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2338 *tail = rdp->nocb_gp_head;
2339 atomic_long_add(rdp->nocb_gp_count, &rdp->nocb_follower_count);
2340 atomic_long_add(rdp->nocb_gp_count_lazy,
2341 &rdp->nocb_follower_count_lazy);
2342 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2343 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2345 * List was empty, wake up the follower.
2346 * Memory barriers supplied by atomic_long_add().
2348 wake_up(&rdp->nocb_wq);
2352 /* If we (the leader) don't have CBs, go wait some more. */
2353 if (!my_rdp->nocb_follower_head)
2358 * Followers come here to wait for additional callbacks to show up.
2359 * This function does not return until callbacks appear.
2361 static void nocb_follower_wait(struct rcu_data *rdp)
2363 bool firsttime = true;
2366 if (!rcu_nocb_poll) {
2367 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2369 wait_event_interruptible(rdp->nocb_wq,
2370 ACCESS_ONCE(rdp->nocb_follower_head));
2371 } else if (firsttime) {
2372 /* Don't drown trace log with "Poll"! */
2374 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2376 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2377 /* ^^^ Ensure CB invocation follows _head test. */
2381 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2383 WARN_ON(signal_pending(current));
2384 schedule_timeout_interruptible(1);
2389 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2390 * callbacks queued by the corresponding no-CBs CPU, however, there is
2391 * an optional leader-follower relationship so that the grace-period
2392 * kthreads don't have to do quite so many wakeups.
2394 static int rcu_nocb_kthread(void *arg)
2397 struct rcu_head *list;
2398 struct rcu_head *next;
2399 struct rcu_head **tail;
2400 struct rcu_data *rdp = arg;
2402 /* Each pass through this loop invokes one batch of callbacks */
2404 /* Wait for callbacks. */
2405 if (rdp->nocb_leader == rdp)
2406 nocb_leader_wait(rdp);
2408 nocb_follower_wait(rdp);
2410 /* Pull the ready-to-invoke callbacks onto local list. */
2411 list = ACCESS_ONCE(rdp->nocb_follower_head);
2413 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2414 ACCESS_ONCE(rdp->nocb_follower_head) = NULL;
2415 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2416 c = atomic_long_xchg(&rdp->nocb_follower_count, 0);
2417 cl = atomic_long_xchg(&rdp->nocb_follower_count_lazy, 0);
2418 rdp->nocb_p_count += c;
2419 rdp->nocb_p_count_lazy += cl;
2421 /* Each pass through the following loop invokes a callback. */
2422 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2426 /* Wait for enqueuing to complete, if needed. */
2427 while (next == NULL && &list->next != tail) {
2428 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2430 schedule_timeout_interruptible(1);
2431 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2435 debug_rcu_head_unqueue(list);
2437 if (__rcu_reclaim(rdp->rsp->name, list))
2443 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2444 ACCESS_ONCE(rdp->nocb_p_count) = rdp->nocb_p_count - c;
2445 ACCESS_ONCE(rdp->nocb_p_count_lazy) =
2446 rdp->nocb_p_count_lazy - cl;
2447 rdp->n_nocbs_invoked += c;
2452 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2453 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2455 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2458 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2459 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2463 if (!rcu_nocb_need_deferred_wakeup(rdp))
2465 ndw = ACCESS_ONCE(rdp->nocb_defer_wakeup);
2466 ACCESS_ONCE(rdp->nocb_defer_wakeup) = RCU_NOGP_WAKE_NOT;
2467 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2468 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2471 void __init rcu_init_nohz(void)
2474 bool need_rcu_nocb_mask = true;
2475 struct rcu_state *rsp;
2477 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2478 need_rcu_nocb_mask = false;
2479 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2481 #if defined(CONFIG_NO_HZ_FULL)
2482 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2483 need_rcu_nocb_mask = true;
2484 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2486 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2487 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2488 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2491 have_rcu_nocb_mask = true;
2493 if (!have_rcu_nocb_mask)
2496 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2497 pr_info("\tOffload RCU callbacks from CPU 0\n");
2498 cpumask_set_cpu(0, rcu_nocb_mask);
2499 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2500 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2501 pr_info("\tOffload RCU callbacks from all CPUs\n");
2502 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2503 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2504 #if defined(CONFIG_NO_HZ_FULL)
2505 if (tick_nohz_full_running)
2506 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2507 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2509 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2510 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2511 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2514 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
2515 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
2517 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2519 for_each_rcu_flavor(rsp) {
2520 for_each_cpu(cpu, rcu_nocb_mask) {
2521 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2524 * If there are early callbacks, they will need
2525 * to be moved to the nocb lists.
2527 WARN_ON_ONCE(rdp->nxttail[RCU_NEXT_TAIL] !=
2529 rdp->nxttail[RCU_NEXT_TAIL] != NULL);
2530 init_nocb_callback_list(rdp);
2532 rcu_organize_nocb_kthreads(rsp);
2536 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2537 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2539 rdp->nocb_tail = &rdp->nocb_head;
2540 init_waitqueue_head(&rdp->nocb_wq);
2541 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2545 * If the specified CPU is a no-CBs CPU that does not already have its
2546 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2547 * brought online out of order, this can require re-organizing the
2548 * leader-follower relationships.
2550 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2552 struct rcu_data *rdp;
2553 struct rcu_data *rdp_last;
2554 struct rcu_data *rdp_old_leader;
2555 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2556 struct task_struct *t;
2559 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2560 * then nothing to do.
2562 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2565 /* If we didn't spawn the leader first, reorganize! */
2566 rdp_old_leader = rdp_spawn->nocb_leader;
2567 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2569 rdp = rdp_old_leader;
2571 rdp->nocb_leader = rdp_spawn;
2572 if (rdp_last && rdp != rdp_spawn)
2573 rdp_last->nocb_next_follower = rdp;
2574 if (rdp == rdp_spawn) {
2575 rdp = rdp->nocb_next_follower;
2578 rdp = rdp->nocb_next_follower;
2579 rdp_last->nocb_next_follower = NULL;
2582 rdp_spawn->nocb_next_follower = rdp_old_leader;
2585 /* Spawn the kthread for this CPU and RCU flavor. */
2586 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2587 "rcuo%c/%d", rsp->abbr, cpu);
2589 ACCESS_ONCE(rdp_spawn->nocb_kthread) = t;
2593 * If the specified CPU is a no-CBs CPU that does not already have its
2594 * rcuo kthreads, spawn them.
2596 static void rcu_spawn_all_nocb_kthreads(int cpu)
2598 struct rcu_state *rsp;
2600 if (rcu_scheduler_fully_active)
2601 for_each_rcu_flavor(rsp)
2602 rcu_spawn_one_nocb_kthread(rsp, cpu);
2606 * Once the scheduler is running, spawn rcuo kthreads for all online
2607 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2608 * non-boot CPUs come online -- if this changes, we will need to add
2609 * some mutual exclusion.
2611 static void __init rcu_spawn_nocb_kthreads(void)
2615 for_each_online_cpu(cpu)
2616 rcu_spawn_all_nocb_kthreads(cpu);
2619 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2620 static int rcu_nocb_leader_stride = -1;
2621 module_param(rcu_nocb_leader_stride, int, 0444);
2624 * Initialize leader-follower relationships for all no-CBs CPU.
2626 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2629 int ls = rcu_nocb_leader_stride;
2630 int nl = 0; /* Next leader. */
2631 struct rcu_data *rdp;
2632 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2633 struct rcu_data *rdp_prev = NULL;
2635 if (!have_rcu_nocb_mask)
2638 ls = int_sqrt(nr_cpu_ids);
2639 rcu_nocb_leader_stride = ls;
2643 * Each pass through this loop sets up one rcu_data structure and
2644 * spawns one rcu_nocb_kthread().
2646 for_each_cpu(cpu, rcu_nocb_mask) {
2647 rdp = per_cpu_ptr(rsp->rda, cpu);
2648 if (rdp->cpu >= nl) {
2649 /* New leader, set up for followers & next leader. */
2650 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2651 rdp->nocb_leader = rdp;
2654 /* Another follower, link to previous leader. */
2655 rdp->nocb_leader = rdp_leader;
2656 rdp_prev->nocb_next_follower = rdp;
2662 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2663 static bool init_nocb_callback_list(struct rcu_data *rdp)
2665 if (!rcu_is_nocb_cpu(rdp->cpu))
2668 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2672 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2674 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2676 WARN_ON_ONCE(1); /* Should be dead code. */
2680 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2684 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2688 static void rcu_init_one_nocb(struct rcu_node *rnp)
2692 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2693 bool lazy, unsigned long flags)
2698 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2699 struct rcu_data *rdp,
2700 unsigned long flags)
2705 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2709 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2714 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2718 static void rcu_spawn_all_nocb_kthreads(int cpu)
2722 static void __init rcu_spawn_nocb_kthreads(void)
2726 static bool init_nocb_callback_list(struct rcu_data *rdp)
2731 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2734 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2735 * arbitrarily long period of time with the scheduling-clock tick turned
2736 * off. RCU will be paying attention to this CPU because it is in the
2737 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2738 * machine because the scheduling-clock tick has been disabled. Therefore,
2739 * if an adaptive-ticks CPU is failing to respond to the current grace
2740 * period and has not be idle from an RCU perspective, kick it.
2742 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2744 #ifdef CONFIG_NO_HZ_FULL
2745 if (tick_nohz_full_cpu(cpu))
2746 smp_send_reschedule(cpu);
2747 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2751 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2753 static int full_sysidle_state; /* Current system-idle state. */
2754 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2755 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2756 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2757 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2758 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2761 * Invoked to note exit from irq or task transition to idle. Note that
2762 * usermode execution does -not- count as idle here! After all, we want
2763 * to detect full-system idle states, not RCU quiescent states and grace
2764 * periods. The caller must have disabled interrupts.
2766 static void rcu_sysidle_enter(int irq)
2769 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2771 /* If there are no nohz_full= CPUs, no need to track this. */
2772 if (!tick_nohz_full_enabled())
2775 /* Adjust nesting, check for fully idle. */
2777 rdtp->dynticks_idle_nesting--;
2778 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2779 if (rdtp->dynticks_idle_nesting != 0)
2780 return; /* Still not fully idle. */
2782 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2783 DYNTICK_TASK_NEST_VALUE) {
2784 rdtp->dynticks_idle_nesting = 0;
2786 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2787 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2788 return; /* Still not fully idle. */
2792 /* Record start of fully idle period. */
2794 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2795 smp_mb__before_atomic();
2796 atomic_inc(&rdtp->dynticks_idle);
2797 smp_mb__after_atomic();
2798 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2802 * Unconditionally force exit from full system-idle state. This is
2803 * invoked when a normal CPU exits idle, but must be called separately
2804 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2805 * is that the timekeeping CPU is permitted to take scheduling-clock
2806 * interrupts while the system is in system-idle state, and of course
2807 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2808 * interrupt from any other type of interrupt.
2810 void rcu_sysidle_force_exit(void)
2812 int oldstate = ACCESS_ONCE(full_sysidle_state);
2816 * Each pass through the following loop attempts to exit full
2817 * system-idle state. If contention proves to be a problem,
2818 * a trylock-based contention tree could be used here.
2820 while (oldstate > RCU_SYSIDLE_SHORT) {
2821 newoldstate = cmpxchg(&full_sysidle_state,
2822 oldstate, RCU_SYSIDLE_NOT);
2823 if (oldstate == newoldstate &&
2824 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2825 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2826 return; /* We cleared it, done! */
2828 oldstate = newoldstate;
2830 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2834 * Invoked to note entry to irq or task transition from idle. Note that
2835 * usermode execution does -not- count as idle here! The caller must
2836 * have disabled interrupts.
2838 static void rcu_sysidle_exit(int irq)
2840 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2842 /* If there are no nohz_full= CPUs, no need to track this. */
2843 if (!tick_nohz_full_enabled())
2846 /* Adjust nesting, check for already non-idle. */
2848 rdtp->dynticks_idle_nesting++;
2849 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2850 if (rdtp->dynticks_idle_nesting != 1)
2851 return; /* Already non-idle. */
2854 * Allow for irq misnesting. Yes, it really is possible
2855 * to enter an irq handler then never leave it, and maybe
2856 * also vice versa. Handle both possibilities.
2858 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2859 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2860 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2861 return; /* Already non-idle. */
2863 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2867 /* Record end of idle period. */
2868 smp_mb__before_atomic();
2869 atomic_inc(&rdtp->dynticks_idle);
2870 smp_mb__after_atomic();
2871 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2874 * If we are the timekeeping CPU, we are permitted to be non-idle
2875 * during a system-idle state. This must be the case, because
2876 * the timekeeping CPU has to take scheduling-clock interrupts
2877 * during the time that the system is transitioning to full
2878 * system-idle state. This means that the timekeeping CPU must
2879 * invoke rcu_sysidle_force_exit() directly if it does anything
2880 * more than take a scheduling-clock interrupt.
2882 if (smp_processor_id() == tick_do_timer_cpu)
2885 /* Update system-idle state: We are clearly no longer fully idle! */
2886 rcu_sysidle_force_exit();
2890 * Check to see if the current CPU is idle. Note that usermode execution
2891 * does not count as idle. The caller must have disabled interrupts.
2893 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2894 unsigned long *maxj)
2898 struct rcu_dynticks *rdtp = rdp->dynticks;
2900 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2901 if (!tick_nohz_full_enabled())
2905 * If some other CPU has already reported non-idle, if this is
2906 * not the flavor of RCU that tracks sysidle state, or if this
2907 * is an offline or the timekeeping CPU, nothing to do.
2909 if (!*isidle || rdp->rsp != rcu_state_p ||
2910 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2912 if (rcu_gp_in_progress(rdp->rsp))
2913 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2915 /* Pick up current idle and NMI-nesting counter and check. */
2916 cur = atomic_read(&rdtp->dynticks_idle);
2918 *isidle = false; /* We are not idle! */
2921 smp_mb(); /* Read counters before timestamps. */
2923 /* Pick up timestamps. */
2924 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2925 /* If this CPU entered idle more recently, update maxj timestamp. */
2926 if (ULONG_CMP_LT(*maxj, j))
2931 * Is this the flavor of RCU that is handling full-system idle?
2933 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2935 return rsp == rcu_state_p;
2939 * Return a delay in jiffies based on the number of CPUs, rcu_node
2940 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2941 * systems more time to transition to full-idle state in order to
2942 * avoid the cache thrashing that otherwise occur on the state variable.
2943 * Really small systems (less than a couple of tens of CPUs) should
2944 * instead use a single global atomically incremented counter, and later
2945 * versions of this will automatically reconfigure themselves accordingly.
2947 static unsigned long rcu_sysidle_delay(void)
2949 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2951 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2955 * Advance the full-system-idle state. This is invoked when all of
2956 * the non-timekeeping CPUs are idle.
2958 static void rcu_sysidle(unsigned long j)
2960 /* Check the current state. */
2961 switch (ACCESS_ONCE(full_sysidle_state)) {
2962 case RCU_SYSIDLE_NOT:
2964 /* First time all are idle, so note a short idle period. */
2965 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2968 case RCU_SYSIDLE_SHORT:
2971 * Idle for a bit, time to advance to next state?
2972 * cmpxchg failure means race with non-idle, let them win.
2974 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2975 (void)cmpxchg(&full_sysidle_state,
2976 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2979 case RCU_SYSIDLE_LONG:
2982 * Do an additional check pass before advancing to full.
2983 * cmpxchg failure means race with non-idle, let them win.
2985 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2986 (void)cmpxchg(&full_sysidle_state,
2987 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2996 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2997 * back to the beginning.
2999 static void rcu_sysidle_cancel(void)
3002 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
3003 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
3007 * Update the sysidle state based on the results of a force-quiescent-state
3008 * scan of the CPUs' dyntick-idle state.
3010 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
3011 unsigned long maxj, bool gpkt)
3013 if (rsp != rcu_state_p)
3014 return; /* Wrong flavor, ignore. */
3015 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
3016 return; /* Running state machine from timekeeping CPU. */
3018 rcu_sysidle(maxj); /* More idle! */
3020 rcu_sysidle_cancel(); /* Idle is over. */
3024 * Wrapper for rcu_sysidle_report() when called from the grace-period
3025 * kthread's context.
3027 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3030 /* If there are no nohz_full= CPUs, no need to track this. */
3031 if (!tick_nohz_full_enabled())
3034 rcu_sysidle_report(rsp, isidle, maxj, true);
3037 /* Callback and function for forcing an RCU grace period. */
3038 struct rcu_sysidle_head {
3043 static void rcu_sysidle_cb(struct rcu_head *rhp)
3045 struct rcu_sysidle_head *rshp;
3048 * The following memory barrier is needed to replace the
3049 * memory barriers that would normally be in the memory
3052 smp_mb(); /* grace period precedes setting inuse. */
3054 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
3055 ACCESS_ONCE(rshp->inuse) = 0;
3059 * Check to see if the system is fully idle, other than the timekeeping CPU.
3060 * The caller must have disabled interrupts. This is not intended to be
3061 * called unless tick_nohz_full_enabled().
3063 bool rcu_sys_is_idle(void)
3065 static struct rcu_sysidle_head rsh;
3066 int rss = ACCESS_ONCE(full_sysidle_state);
3068 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
3071 /* Handle small-system case by doing a full scan of CPUs. */
3072 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
3073 int oldrss = rss - 1;
3076 * One pass to advance to each state up to _FULL.
3077 * Give up if any pass fails to advance the state.
3079 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
3082 unsigned long maxj = jiffies - ULONG_MAX / 4;
3083 struct rcu_data *rdp;
3085 /* Scan all the CPUs looking for nonidle CPUs. */
3086 for_each_possible_cpu(cpu) {
3087 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3088 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
3092 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
3094 rss = ACCESS_ONCE(full_sysidle_state);
3098 /* If this is the first observation of an idle period, record it. */
3099 if (rss == RCU_SYSIDLE_FULL) {
3100 rss = cmpxchg(&full_sysidle_state,
3101 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
3102 return rss == RCU_SYSIDLE_FULL;
3105 smp_mb(); /* ensure rss load happens before later caller actions. */
3107 /* If already fully idle, tell the caller (in case of races). */
3108 if (rss == RCU_SYSIDLE_FULL_NOTED)
3112 * If we aren't there yet, and a grace period is not in flight,
3113 * initiate a grace period. Either way, tell the caller that
3114 * we are not there yet. We use an xchg() rather than an assignment
3115 * to make up for the memory barriers that would otherwise be
3116 * provided by the memory allocator.
3118 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
3119 !rcu_gp_in_progress(rcu_state_p) &&
3120 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
3121 call_rcu(&rsh.rh, rcu_sysidle_cb);
3126 * Initialize dynticks sysidle state for CPUs coming online.
3128 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3130 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
3133 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3135 static void rcu_sysidle_enter(int irq)
3139 static void rcu_sysidle_exit(int irq)
3143 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3144 unsigned long *maxj)
3148 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3153 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3158 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3162 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3165 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3166 * grace-period kthread will do force_quiescent_state() processing?
3167 * The idea is to avoid waking up RCU core processing on such a
3168 * CPU unless the grace period has extended for too long.
3170 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3171 * CONFIG_RCU_NOCB_CPU CPUs.
3173 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3175 #ifdef CONFIG_NO_HZ_FULL
3176 if (tick_nohz_full_cpu(smp_processor_id()) &&
3177 (!rcu_gp_in_progress(rsp) ||
3178 ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
3180 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3185 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3188 static void rcu_bind_gp_kthread(void)
3190 int __maybe_unused cpu;
3192 if (!tick_nohz_full_enabled())
3194 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3195 cpu = tick_do_timer_cpu;
3196 if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu)
3197 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3198 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3199 if (!is_housekeeping_cpu(raw_smp_processor_id()))
3200 housekeeping_affine(current);
3201 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3204 /* Record the current task on dyntick-idle entry. */
3205 static void rcu_dynticks_task_enter(void)
3207 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3208 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = smp_processor_id();
3209 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3212 /* Record no current task on dyntick-idle exit. */
3213 static void rcu_dynticks_task_exit(void)
3215 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3216 ACCESS_ONCE(current->rcu_tasks_idle_cpu) = -1;
3217 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */