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 #define RCU_KTHREAD_PRIO 1
35 #ifdef CONFIG_RCU_BOOST
36 #include "../locking/rtmutex_common.h"
37 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
39 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
42 #ifdef CONFIG_RCU_NOCB_CPU
43 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
44 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
45 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
46 static char __initdata nocb_buf[NR_CPUS * 5];
47 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
50 * Check the RCU kernel configuration parameters and print informative
51 * messages about anything out of the ordinary. If you like #ifdef, you
52 * will love this function.
54 static void __init rcu_bootup_announce_oddness(void)
56 #ifdef CONFIG_RCU_TRACE
57 pr_info("\tRCU debugfs-based tracing is enabled.\n");
59 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
60 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
63 #ifdef CONFIG_RCU_FANOUT_EXACT
64 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
66 #ifdef CONFIG_RCU_FAST_NO_HZ
67 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
69 #ifdef CONFIG_PROVE_RCU
70 pr_info("\tRCU lockdep checking is enabled.\n");
72 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
73 pr_info("\tRCU torture testing starts during boot.\n");
75 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
76 pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
78 #if defined(CONFIG_RCU_CPU_STALL_INFO)
79 pr_info("\tAdditional per-CPU info printed with stalls.\n");
81 #if NUM_RCU_LVL_4 != 0
82 pr_info("\tFour-level hierarchy is enabled.\n");
84 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
85 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
86 if (nr_cpu_ids != NR_CPUS)
87 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
88 #ifdef CONFIG_RCU_NOCB_CPU
89 #ifndef CONFIG_RCU_NOCB_CPU_NONE
90 if (!have_rcu_nocb_mask) {
91 zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL);
92 have_rcu_nocb_mask = true;
94 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
95 pr_info("\tOffload RCU callbacks from CPU 0\n");
96 cpumask_set_cpu(0, rcu_nocb_mask);
97 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
98 #ifdef CONFIG_RCU_NOCB_CPU_ALL
99 pr_info("\tOffload RCU callbacks from all CPUs\n");
100 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
101 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
102 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
103 if (have_rcu_nocb_mask) {
104 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
105 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
106 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
109 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
110 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
112 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
114 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
117 #ifdef CONFIG_TREE_PREEMPT_RCU
119 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
120 static struct rcu_state *rcu_state_p = &rcu_preempt_state;
122 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
125 * Tell them what RCU they are running.
127 static void __init rcu_bootup_announce(void)
129 pr_info("Preemptible hierarchical RCU implementation.\n");
130 rcu_bootup_announce_oddness();
134 * Return the number of RCU-preempt batches processed thus far
135 * for debug and statistics.
137 long rcu_batches_completed_preempt(void)
139 return rcu_preempt_state.completed;
141 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
144 * Return the number of RCU batches processed thus far for debug & stats.
146 long rcu_batches_completed(void)
148 return rcu_batches_completed_preempt();
150 EXPORT_SYMBOL_GPL(rcu_batches_completed);
153 * Record a preemptible-RCU quiescent state for the specified CPU. Note
154 * that this just means that the task currently running on the CPU is
155 * not in a quiescent state. There might be any number of tasks blocked
156 * while in an RCU read-side critical section.
158 * Unlike the other rcu_*_qs() functions, callers to this function
159 * must disable irqs in order to protect the assignment to
160 * ->rcu_read_unlock_special.
162 static void rcu_preempt_qs(int cpu)
164 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
166 if (rdp->passed_quiesce == 0)
167 trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs"));
168 rdp->passed_quiesce = 1;
169 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
173 * We have entered the scheduler, and the current task might soon be
174 * context-switched away from. If this task is in an RCU read-side
175 * critical section, we will no longer be able to rely on the CPU to
176 * record that fact, so we enqueue the task on the blkd_tasks list.
177 * The task will dequeue itself when it exits the outermost enclosing
178 * RCU read-side critical section. Therefore, the current grace period
179 * cannot be permitted to complete until the blkd_tasks list entries
180 * predating the current grace period drain, in other words, until
181 * rnp->gp_tasks becomes NULL.
183 * Caller must disable preemption.
185 static void rcu_preempt_note_context_switch(int cpu)
187 struct task_struct *t = current;
189 struct rcu_data *rdp;
190 struct rcu_node *rnp;
192 if (t->rcu_read_lock_nesting > 0 &&
193 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
195 /* Possibly blocking in an RCU read-side critical section. */
196 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
198 raw_spin_lock_irqsave(&rnp->lock, flags);
199 smp_mb__after_unlock_lock();
200 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
201 t->rcu_blocked_node = rnp;
204 * If this CPU has already checked in, then this task
205 * will hold up the next grace period rather than the
206 * current grace period. Queue the task accordingly.
207 * If the task is queued for the current grace period
208 * (i.e., this CPU has not yet passed through a quiescent
209 * state for the current grace period), then as long
210 * as that task remains queued, the current grace period
211 * cannot end. Note that there is some uncertainty as
212 * to exactly when the current grace period started.
213 * We take a conservative approach, which can result
214 * in unnecessarily waiting on tasks that started very
215 * slightly after the current grace period began. C'est
218 * But first, note that the current CPU must still be
221 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
222 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
223 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
224 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
225 rnp->gp_tasks = &t->rcu_node_entry;
226 #ifdef CONFIG_RCU_BOOST
227 if (rnp->boost_tasks != NULL)
228 rnp->boost_tasks = rnp->gp_tasks;
229 #endif /* #ifdef CONFIG_RCU_BOOST */
231 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
232 if (rnp->qsmask & rdp->grpmask)
233 rnp->gp_tasks = &t->rcu_node_entry;
235 trace_rcu_preempt_task(rdp->rsp->name,
237 (rnp->qsmask & rdp->grpmask)
240 raw_spin_unlock_irqrestore(&rnp->lock, flags);
241 } else if (t->rcu_read_lock_nesting < 0 &&
242 t->rcu_read_unlock_special) {
245 * Complete exit from RCU read-side critical section on
246 * behalf of preempted instance of __rcu_read_unlock().
248 rcu_read_unlock_special(t);
252 * Either we were not in an RCU read-side critical section to
253 * begin with, or we have now recorded that critical section
254 * globally. Either way, we can now note a quiescent state
255 * for this CPU. Again, if we were in an RCU read-side critical
256 * section, and if that critical section was blocking the current
257 * grace period, then the fact that the task has been enqueued
258 * means that we continue to block the current grace period.
260 local_irq_save(flags);
262 local_irq_restore(flags);
266 * Check for preempted RCU readers blocking the current grace period
267 * for the specified rcu_node structure. If the caller needs a reliable
268 * answer, it must hold the rcu_node's ->lock.
270 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
272 return rnp->gp_tasks != NULL;
276 * Record a quiescent state for all tasks that were previously queued
277 * on the specified rcu_node structure and that were blocking the current
278 * RCU grace period. The caller must hold the specified rnp->lock with
279 * irqs disabled, and this lock is released upon return, but irqs remain
282 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
283 __releases(rnp->lock)
286 struct rcu_node *rnp_p;
288 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
289 raw_spin_unlock_irqrestore(&rnp->lock, flags);
290 return; /* Still need more quiescent states! */
296 * Either there is only one rcu_node in the tree,
297 * or tasks were kicked up to root rcu_node due to
298 * CPUs going offline.
300 rcu_report_qs_rsp(&rcu_preempt_state, flags);
304 /* Report up the rest of the hierarchy. */
306 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
307 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
308 smp_mb__after_unlock_lock();
309 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
313 * Advance a ->blkd_tasks-list pointer to the next entry, instead
314 * returning NULL if at the end of the list.
316 static struct list_head *rcu_next_node_entry(struct task_struct *t,
317 struct rcu_node *rnp)
319 struct list_head *np;
321 np = t->rcu_node_entry.next;
322 if (np == &rnp->blkd_tasks)
328 * Handle special cases during rcu_read_unlock(), such as needing to
329 * notify RCU core processing or task having blocked during the RCU
330 * read-side critical section.
332 void rcu_read_unlock_special(struct task_struct *t)
338 struct list_head *np;
339 #ifdef CONFIG_RCU_BOOST
340 bool drop_boost_mutex = false;
341 #endif /* #ifdef CONFIG_RCU_BOOST */
342 struct rcu_node *rnp;
345 /* NMI handlers cannot block and cannot safely manipulate state. */
349 local_irq_save(flags);
352 * If RCU core is waiting for this CPU to exit critical section,
353 * let it know that we have done so.
355 special = t->rcu_read_unlock_special;
356 if (special & RCU_READ_UNLOCK_NEED_QS) {
357 rcu_preempt_qs(smp_processor_id());
358 if (!t->rcu_read_unlock_special) {
359 local_irq_restore(flags);
364 /* Hardware IRQ handlers cannot block, complain if they get here. */
365 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
366 local_irq_restore(flags);
370 /* Clean up if blocked during RCU read-side critical section. */
371 if (special & RCU_READ_UNLOCK_BLOCKED) {
372 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
375 * Remove this task from the list it blocked on. The
376 * task can migrate while we acquire the lock, but at
377 * most one time. So at most two passes through loop.
380 rnp = t->rcu_blocked_node;
381 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
382 smp_mb__after_unlock_lock();
383 if (rnp == t->rcu_blocked_node)
385 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
387 empty = !rcu_preempt_blocked_readers_cgp(rnp);
388 empty_exp = !rcu_preempted_readers_exp(rnp);
389 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
390 np = rcu_next_node_entry(t, rnp);
391 list_del_init(&t->rcu_node_entry);
392 t->rcu_blocked_node = NULL;
393 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
395 if (&t->rcu_node_entry == rnp->gp_tasks)
397 if (&t->rcu_node_entry == rnp->exp_tasks)
399 #ifdef CONFIG_RCU_BOOST
400 if (&t->rcu_node_entry == rnp->boost_tasks)
401 rnp->boost_tasks = np;
402 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
403 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
404 #endif /* #ifdef CONFIG_RCU_BOOST */
407 * If this was the last task on the current list, and if
408 * we aren't waiting on any CPUs, report the quiescent state.
409 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
410 * so we must take a snapshot of the expedited state.
412 empty_exp_now = !rcu_preempted_readers_exp(rnp);
413 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
414 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
421 rcu_report_unblock_qs_rnp(rnp, flags);
423 raw_spin_unlock_irqrestore(&rnp->lock, flags);
426 #ifdef CONFIG_RCU_BOOST
427 /* Unboost if we were boosted. */
428 if (drop_boost_mutex) {
429 rt_mutex_unlock(&rnp->boost_mtx);
430 complete(&rnp->boost_completion);
432 #endif /* #ifdef CONFIG_RCU_BOOST */
435 * If this was the last task on the expedited lists,
436 * then we need to report up the rcu_node hierarchy.
438 if (!empty_exp && empty_exp_now)
439 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
441 local_irq_restore(flags);
445 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
448 * Dump detailed information for all tasks blocking the current RCU
449 * grace period on the specified rcu_node structure.
451 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
454 struct task_struct *t;
456 raw_spin_lock_irqsave(&rnp->lock, flags);
457 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
458 raw_spin_unlock_irqrestore(&rnp->lock, flags);
461 t = list_entry(rnp->gp_tasks,
462 struct task_struct, rcu_node_entry);
463 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
465 raw_spin_unlock_irqrestore(&rnp->lock, flags);
469 * Dump detailed information for all tasks blocking the current RCU
472 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
474 struct rcu_node *rnp = rcu_get_root(rsp);
476 rcu_print_detail_task_stall_rnp(rnp);
477 rcu_for_each_leaf_node(rsp, rnp)
478 rcu_print_detail_task_stall_rnp(rnp);
481 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
483 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
487 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
489 #ifdef CONFIG_RCU_CPU_STALL_INFO
491 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
493 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
494 rnp->level, rnp->grplo, rnp->grphi);
497 static void rcu_print_task_stall_end(void)
502 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
504 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
508 static void rcu_print_task_stall_end(void)
512 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
515 * Scan the current list of tasks blocked within RCU read-side critical
516 * sections, printing out the tid of each.
518 static int rcu_print_task_stall(struct rcu_node *rnp)
520 struct task_struct *t;
523 if (!rcu_preempt_blocked_readers_cgp(rnp))
525 rcu_print_task_stall_begin(rnp);
526 t = list_entry(rnp->gp_tasks,
527 struct task_struct, rcu_node_entry);
528 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
529 pr_cont(" P%d", t->pid);
532 rcu_print_task_stall_end();
537 * Check that the list of blocked tasks for the newly completed grace
538 * period is in fact empty. It is a serious bug to complete a grace
539 * period that still has RCU readers blocked! This function must be
540 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
541 * must be held by the caller.
543 * Also, if there are blocked tasks on the list, they automatically
544 * block the newly created grace period, so set up ->gp_tasks accordingly.
546 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
548 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
549 if (!list_empty(&rnp->blkd_tasks))
550 rnp->gp_tasks = rnp->blkd_tasks.next;
551 WARN_ON_ONCE(rnp->qsmask);
554 #ifdef CONFIG_HOTPLUG_CPU
557 * Handle tasklist migration for case in which all CPUs covered by the
558 * specified rcu_node have gone offline. Move them up to the root
559 * rcu_node. The reason for not just moving them to the immediate
560 * parent is to remove the need for rcu_read_unlock_special() to
561 * make more than two attempts to acquire the target rcu_node's lock.
562 * Returns true if there were tasks blocking the current RCU grace
565 * Returns 1 if there was previously a task blocking the current grace
566 * period on the specified rcu_node structure.
568 * The caller must hold rnp->lock with irqs disabled.
570 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
571 struct rcu_node *rnp,
572 struct rcu_data *rdp)
574 struct list_head *lp;
575 struct list_head *lp_root;
577 struct rcu_node *rnp_root = rcu_get_root(rsp);
578 struct task_struct *t;
580 if (rnp == rnp_root) {
581 WARN_ONCE(1, "Last CPU thought to be offlined?");
582 return 0; /* Shouldn't happen: at least one CPU online. */
585 /* If we are on an internal node, complain bitterly. */
586 WARN_ON_ONCE(rnp != rdp->mynode);
589 * Move tasks up to root rcu_node. Don't try to get fancy for
590 * this corner-case operation -- just put this node's tasks
591 * at the head of the root node's list, and update the root node's
592 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
593 * if non-NULL. This might result in waiting for more tasks than
594 * absolutely necessary, but this is a good performance/complexity
597 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
598 retval |= RCU_OFL_TASKS_NORM_GP;
599 if (rcu_preempted_readers_exp(rnp))
600 retval |= RCU_OFL_TASKS_EXP_GP;
601 lp = &rnp->blkd_tasks;
602 lp_root = &rnp_root->blkd_tasks;
603 while (!list_empty(lp)) {
604 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
605 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
606 smp_mb__after_unlock_lock();
607 list_del(&t->rcu_node_entry);
608 t->rcu_blocked_node = rnp_root;
609 list_add(&t->rcu_node_entry, lp_root);
610 if (&t->rcu_node_entry == rnp->gp_tasks)
611 rnp_root->gp_tasks = rnp->gp_tasks;
612 if (&t->rcu_node_entry == rnp->exp_tasks)
613 rnp_root->exp_tasks = rnp->exp_tasks;
614 #ifdef CONFIG_RCU_BOOST
615 if (&t->rcu_node_entry == rnp->boost_tasks)
616 rnp_root->boost_tasks = rnp->boost_tasks;
617 #endif /* #ifdef CONFIG_RCU_BOOST */
618 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
621 rnp->gp_tasks = NULL;
622 rnp->exp_tasks = NULL;
623 #ifdef CONFIG_RCU_BOOST
624 rnp->boost_tasks = NULL;
626 * In case root is being boosted and leaf was not. Make sure
627 * that we boost the tasks blocking the current grace period
630 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
631 smp_mb__after_unlock_lock();
632 if (rnp_root->boost_tasks != NULL &&
633 rnp_root->boost_tasks != rnp_root->gp_tasks &&
634 rnp_root->boost_tasks != rnp_root->exp_tasks)
635 rnp_root->boost_tasks = rnp_root->gp_tasks;
636 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
637 #endif /* #ifdef CONFIG_RCU_BOOST */
642 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
645 * Check for a quiescent state from the current CPU. When a task blocks,
646 * the task is recorded in the corresponding CPU's rcu_node structure,
647 * which is checked elsewhere.
649 * Caller must disable hard irqs.
651 static void rcu_preempt_check_callbacks(int cpu)
653 struct task_struct *t = current;
655 if (t->rcu_read_lock_nesting == 0) {
659 if (t->rcu_read_lock_nesting > 0 &&
660 per_cpu(rcu_preempt_data, cpu).qs_pending)
661 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
664 #ifdef CONFIG_RCU_BOOST
666 static void rcu_preempt_do_callbacks(void)
668 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
671 #endif /* #ifdef CONFIG_RCU_BOOST */
674 * Queue a preemptible-RCU callback for invocation after a grace period.
676 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
678 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
680 EXPORT_SYMBOL_GPL(call_rcu);
683 * synchronize_rcu - wait until a grace period has elapsed.
685 * Control will return to the caller some time after a full grace
686 * period has elapsed, in other words after all currently executing RCU
687 * read-side critical sections have completed. Note, however, that
688 * upon return from synchronize_rcu(), the caller might well be executing
689 * concurrently with new RCU read-side critical sections that began while
690 * synchronize_rcu() was waiting. RCU read-side critical sections are
691 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
693 * See the description of synchronize_sched() for more detailed information
694 * on memory ordering guarantees.
696 void synchronize_rcu(void)
698 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
699 !lock_is_held(&rcu_lock_map) &&
700 !lock_is_held(&rcu_sched_lock_map),
701 "Illegal synchronize_rcu() in RCU read-side critical section");
702 if (!rcu_scheduler_active)
705 synchronize_rcu_expedited();
707 wait_rcu_gp(call_rcu);
709 EXPORT_SYMBOL_GPL(synchronize_rcu);
711 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
712 static unsigned long sync_rcu_preempt_exp_count;
713 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
716 * Return non-zero if there are any tasks in RCU read-side critical
717 * sections blocking the current preemptible-RCU expedited grace period.
718 * If there is no preemptible-RCU expedited grace period currently in
719 * progress, returns zero unconditionally.
721 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
723 return rnp->exp_tasks != NULL;
727 * return non-zero if there is no RCU expedited grace period in progress
728 * for the specified rcu_node structure, in other words, if all CPUs and
729 * tasks covered by the specified rcu_node structure have done their bit
730 * for the current expedited grace period. Works only for preemptible
731 * RCU -- other RCU implementation use other means.
733 * Caller must hold sync_rcu_preempt_exp_mutex.
735 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
737 return !rcu_preempted_readers_exp(rnp) &&
738 ACCESS_ONCE(rnp->expmask) == 0;
742 * Report the exit from RCU read-side critical section for the last task
743 * that queued itself during or before the current expedited preemptible-RCU
744 * grace period. This event is reported either to the rcu_node structure on
745 * which the task was queued or to one of that rcu_node structure's ancestors,
746 * recursively up the tree. (Calm down, calm down, we do the recursion
749 * Most callers will set the "wake" flag, but the task initiating the
750 * expedited grace period need not wake itself.
752 * Caller must hold sync_rcu_preempt_exp_mutex.
754 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
760 raw_spin_lock_irqsave(&rnp->lock, flags);
761 smp_mb__after_unlock_lock();
763 if (!sync_rcu_preempt_exp_done(rnp)) {
764 raw_spin_unlock_irqrestore(&rnp->lock, flags);
767 if (rnp->parent == NULL) {
768 raw_spin_unlock_irqrestore(&rnp->lock, flags);
770 smp_mb(); /* EGP done before wake_up(). */
771 wake_up(&sync_rcu_preempt_exp_wq);
776 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
778 raw_spin_lock(&rnp->lock); /* irqs already disabled */
779 smp_mb__after_unlock_lock();
780 rnp->expmask &= ~mask;
785 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
786 * grace period for the specified rcu_node structure. If there are no such
787 * tasks, report it up the rcu_node hierarchy.
789 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
790 * CPU hotplug operations.
793 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
798 raw_spin_lock_irqsave(&rnp->lock, flags);
799 smp_mb__after_unlock_lock();
800 if (list_empty(&rnp->blkd_tasks)) {
801 raw_spin_unlock_irqrestore(&rnp->lock, flags);
803 rnp->exp_tasks = rnp->blkd_tasks.next;
804 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
808 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
812 * synchronize_rcu_expedited - Brute-force RCU grace period
814 * Wait for an RCU-preempt grace period, but expedite it. The basic
815 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
816 * the ->blkd_tasks lists and wait for this list to drain. This consumes
817 * significant time on all CPUs and is unfriendly to real-time workloads,
818 * so is thus not recommended for any sort of common-case code.
819 * In fact, if you are using synchronize_rcu_expedited() in a loop,
820 * please restructure your code to batch your updates, and then Use a
821 * single synchronize_rcu() instead.
823 * Note that it is illegal to call this function while holding any lock
824 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
825 * to call this function from a CPU-hotplug notifier. Failing to observe
826 * these restriction will result in deadlock.
828 void synchronize_rcu_expedited(void)
831 struct rcu_node *rnp;
832 struct rcu_state *rsp = &rcu_preempt_state;
836 smp_mb(); /* Caller's modifications seen first by other CPUs. */
837 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
838 smp_mb(); /* Above access cannot bleed into critical section. */
841 * Block CPU-hotplug operations. This means that any CPU-hotplug
842 * operation that finds an rcu_node structure with tasks in the
843 * process of being boosted will know that all tasks blocking
844 * this expedited grace period will already be in the process of
845 * being boosted. This simplifies the process of moving tasks
846 * from leaf to root rcu_node structures.
851 * Acquire lock, falling back to synchronize_rcu() if too many
852 * lock-acquisition failures. Of course, if someone does the
853 * expedited grace period for us, just leave.
855 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
856 if (ULONG_CMP_LT(snap,
857 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
859 goto mb_ret; /* Others did our work for us. */
861 if (trycount++ < 10) {
862 udelay(trycount * num_online_cpus());
865 wait_rcu_gp(call_rcu);
869 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
871 goto unlock_mb_ret; /* Others did our work for us. */
874 /* force all RCU readers onto ->blkd_tasks lists. */
875 synchronize_sched_expedited();
877 /* Initialize ->expmask for all non-leaf rcu_node structures. */
878 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
879 raw_spin_lock_irqsave(&rnp->lock, flags);
880 smp_mb__after_unlock_lock();
881 rnp->expmask = rnp->qsmaskinit;
882 raw_spin_unlock_irqrestore(&rnp->lock, flags);
885 /* Snapshot current state of ->blkd_tasks lists. */
886 rcu_for_each_leaf_node(rsp, rnp)
887 sync_rcu_preempt_exp_init(rsp, rnp);
888 if (NUM_RCU_NODES > 1)
889 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
893 /* Wait for snapshotted ->blkd_tasks lists to drain. */
894 rnp = rcu_get_root(rsp);
895 wait_event(sync_rcu_preempt_exp_wq,
896 sync_rcu_preempt_exp_done(rnp));
898 /* Clean up and exit. */
899 smp_mb(); /* ensure expedited GP seen before counter increment. */
900 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
902 mutex_unlock(&sync_rcu_preempt_exp_mutex);
904 smp_mb(); /* ensure subsequent action seen after grace period. */
906 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
909 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
911 * Note that this primitive does not necessarily wait for an RCU grace period
912 * to complete. For example, if there are no RCU callbacks queued anywhere
913 * in the system, then rcu_barrier() is within its rights to return
914 * immediately, without waiting for anything, much less an RCU grace period.
916 void rcu_barrier(void)
918 _rcu_barrier(&rcu_preempt_state);
920 EXPORT_SYMBOL_GPL(rcu_barrier);
923 * Initialize preemptible RCU's state structures.
925 static void __init __rcu_init_preempt(void)
927 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
931 * Check for a task exiting while in a preemptible-RCU read-side
932 * critical section, clean up if so. No need to issue warnings,
933 * as debug_check_no_locks_held() already does this if lockdep
938 struct task_struct *t = current;
940 if (likely(list_empty(¤t->rcu_node_entry)))
942 t->rcu_read_lock_nesting = 1;
944 t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED;
948 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
950 static struct rcu_state *rcu_state_p = &rcu_sched_state;
953 * Tell them what RCU they are running.
955 static void __init rcu_bootup_announce(void)
957 pr_info("Hierarchical RCU implementation.\n");
958 rcu_bootup_announce_oddness();
962 * Return the number of RCU batches processed thus far for debug & stats.
964 long rcu_batches_completed(void)
966 return rcu_batches_completed_sched();
968 EXPORT_SYMBOL_GPL(rcu_batches_completed);
971 * Because preemptible RCU does not exist, we never have to check for
972 * CPUs being in quiescent states.
974 static void rcu_preempt_note_context_switch(int cpu)
979 * Because preemptible RCU does not exist, there are never any preempted
982 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
987 #ifdef CONFIG_HOTPLUG_CPU
989 /* Because preemptible RCU does not exist, no quieting of tasks. */
990 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
991 __releases(rnp->lock)
993 raw_spin_unlock_irqrestore(&rnp->lock, flags);
996 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
999 * Because preemptible RCU does not exist, we never have to check for
1000 * tasks blocked within RCU read-side critical sections.
1002 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1007 * Because preemptible RCU does not exist, we never have to check for
1008 * tasks blocked within RCU read-side critical sections.
1010 static int rcu_print_task_stall(struct rcu_node *rnp)
1016 * Because there is no preemptible RCU, there can be no readers blocked,
1017 * so there is no need to check for blocked tasks. So check only for
1018 * bogus qsmask values.
1020 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1022 WARN_ON_ONCE(rnp->qsmask);
1025 #ifdef CONFIG_HOTPLUG_CPU
1028 * Because preemptible RCU does not exist, it never needs to migrate
1029 * tasks that were blocked within RCU read-side critical sections, and
1030 * such non-existent tasks cannot possibly have been blocking the current
1033 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1034 struct rcu_node *rnp,
1035 struct rcu_data *rdp)
1040 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1043 * Because preemptible RCU does not exist, it never has any callbacks
1046 static void rcu_preempt_check_callbacks(int cpu)
1051 * Wait for an rcu-preempt grace period, but make it happen quickly.
1052 * But because preemptible RCU does not exist, map to rcu-sched.
1054 void synchronize_rcu_expedited(void)
1056 synchronize_sched_expedited();
1058 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1060 #ifdef CONFIG_HOTPLUG_CPU
1063 * Because preemptible RCU does not exist, there is never any need to
1064 * report on tasks preempted in RCU read-side critical sections during
1065 * expedited RCU grace periods.
1067 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1072 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1075 * Because preemptible RCU does not exist, rcu_barrier() is just
1076 * another name for rcu_barrier_sched().
1078 void rcu_barrier(void)
1080 rcu_barrier_sched();
1082 EXPORT_SYMBOL_GPL(rcu_barrier);
1085 * Because preemptible RCU does not exist, it need not be initialized.
1087 static void __init __rcu_init_preempt(void)
1092 * Because preemptible RCU does not exist, tasks cannot possibly exit
1093 * while in preemptible RCU read-side critical sections.
1099 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1101 #ifdef CONFIG_RCU_BOOST
1103 #include "../locking/rtmutex_common.h"
1105 #ifdef CONFIG_RCU_TRACE
1107 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1109 if (list_empty(&rnp->blkd_tasks))
1110 rnp->n_balk_blkd_tasks++;
1111 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1112 rnp->n_balk_exp_gp_tasks++;
1113 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1114 rnp->n_balk_boost_tasks++;
1115 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1116 rnp->n_balk_notblocked++;
1117 else if (rnp->gp_tasks != NULL &&
1118 ULONG_CMP_LT(jiffies, rnp->boost_time))
1119 rnp->n_balk_notyet++;
1124 #else /* #ifdef CONFIG_RCU_TRACE */
1126 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1130 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1132 static void rcu_wake_cond(struct task_struct *t, int status)
1135 * If the thread is yielding, only wake it when this
1136 * is invoked from idle
1138 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1143 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1144 * or ->boost_tasks, advancing the pointer to the next task in the
1145 * ->blkd_tasks list.
1147 * Note that irqs must be enabled: boosting the task can block.
1148 * Returns 1 if there are more tasks needing to be boosted.
1150 static int rcu_boost(struct rcu_node *rnp)
1152 unsigned long flags;
1153 struct task_struct *t;
1154 struct list_head *tb;
1156 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1157 return 0; /* Nothing left to boost. */
1159 raw_spin_lock_irqsave(&rnp->lock, flags);
1160 smp_mb__after_unlock_lock();
1163 * Recheck under the lock: all tasks in need of boosting
1164 * might exit their RCU read-side critical sections on their own.
1166 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1167 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1172 * Preferentially boost tasks blocking expedited grace periods.
1173 * This cannot starve the normal grace periods because a second
1174 * expedited grace period must boost all blocked tasks, including
1175 * those blocking the pre-existing normal grace period.
1177 if (rnp->exp_tasks != NULL) {
1178 tb = rnp->exp_tasks;
1179 rnp->n_exp_boosts++;
1181 tb = rnp->boost_tasks;
1182 rnp->n_normal_boosts++;
1184 rnp->n_tasks_boosted++;
1187 * We boost task t by manufacturing an rt_mutex that appears to
1188 * be held by task t. We leave a pointer to that rt_mutex where
1189 * task t can find it, and task t will release the mutex when it
1190 * exits its outermost RCU read-side critical section. Then
1191 * simply acquiring this artificial rt_mutex will boost task
1192 * t's priority. (Thanks to tglx for suggesting this approach!)
1194 * Note that task t must acquire rnp->lock to remove itself from
1195 * the ->blkd_tasks list, which it will do from exit() if from
1196 * nowhere else. We therefore are guaranteed that task t will
1197 * stay around at least until we drop rnp->lock. Note that
1198 * rnp->lock also resolves races between our priority boosting
1199 * and task t's exiting its outermost RCU read-side critical
1202 t = container_of(tb, struct task_struct, rcu_node_entry);
1203 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1204 init_completion(&rnp->boost_completion);
1205 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1206 /* Lock only for side effect: boosts task t's priority. */
1207 rt_mutex_lock(&rnp->boost_mtx);
1208 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1210 /* Wait for boostee to be done w/boost_mtx before reinitializing. */
1211 wait_for_completion(&rnp->boost_completion);
1213 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1214 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1218 * Priority-boosting kthread. One per leaf rcu_node and one for the
1221 static int rcu_boost_kthread(void *arg)
1223 struct rcu_node *rnp = (struct rcu_node *)arg;
1227 trace_rcu_utilization(TPS("Start boost kthread@init"));
1229 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1230 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1231 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1232 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1233 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1234 more2boost = rcu_boost(rnp);
1240 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1241 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1242 schedule_timeout_interruptible(2);
1243 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1248 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1253 * Check to see if it is time to start boosting RCU readers that are
1254 * blocking the current grace period, and, if so, tell the per-rcu_node
1255 * kthread to start boosting them. If there is an expedited grace
1256 * period in progress, it is always time to boost.
1258 * The caller must hold rnp->lock, which this function releases.
1259 * The ->boost_kthread_task is immortal, so we don't need to worry
1260 * about it going away.
1262 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1263 __releases(rnp->lock)
1265 struct task_struct *t;
1267 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1268 rnp->n_balk_exp_gp_tasks++;
1269 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1272 if (rnp->exp_tasks != NULL ||
1273 (rnp->gp_tasks != NULL &&
1274 rnp->boost_tasks == NULL &&
1276 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1277 if (rnp->exp_tasks == NULL)
1278 rnp->boost_tasks = rnp->gp_tasks;
1279 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1280 t = rnp->boost_kthread_task;
1282 rcu_wake_cond(t, rnp->boost_kthread_status);
1284 rcu_initiate_boost_trace(rnp);
1285 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1290 * Wake up the per-CPU kthread to invoke RCU callbacks.
1292 static void invoke_rcu_callbacks_kthread(void)
1294 unsigned long flags;
1296 local_irq_save(flags);
1297 __this_cpu_write(rcu_cpu_has_work, 1);
1298 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1299 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1300 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1301 __this_cpu_read(rcu_cpu_kthread_status));
1303 local_irq_restore(flags);
1307 * Is the current CPU running the RCU-callbacks kthread?
1308 * Caller must have preemption disabled.
1310 static bool rcu_is_callbacks_kthread(void)
1312 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1315 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1318 * Do priority-boost accounting for the start of a new grace period.
1320 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1322 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1326 * Create an RCU-boost kthread for the specified node if one does not
1327 * already exist. We only create this kthread for preemptible RCU.
1328 * Returns zero if all is well, a negated errno otherwise.
1330 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1331 struct rcu_node *rnp)
1333 int rnp_index = rnp - &rsp->node[0];
1334 unsigned long flags;
1335 struct sched_param sp;
1336 struct task_struct *t;
1338 if (&rcu_preempt_state != rsp)
1341 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1345 if (rnp->boost_kthread_task != NULL)
1347 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1348 "rcub/%d", rnp_index);
1351 raw_spin_lock_irqsave(&rnp->lock, flags);
1352 smp_mb__after_unlock_lock();
1353 rnp->boost_kthread_task = t;
1354 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1355 sp.sched_priority = RCU_BOOST_PRIO;
1356 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1357 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1361 static void rcu_kthread_do_work(void)
1363 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1364 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1365 rcu_preempt_do_callbacks();
1368 static void rcu_cpu_kthread_setup(unsigned int cpu)
1370 struct sched_param sp;
1372 sp.sched_priority = RCU_KTHREAD_PRIO;
1373 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1376 static void rcu_cpu_kthread_park(unsigned int cpu)
1378 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1381 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1383 return __this_cpu_read(rcu_cpu_has_work);
1387 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1388 * RCU softirq used in flavors and configurations of RCU that do not
1389 * support RCU priority boosting.
1391 static void rcu_cpu_kthread(unsigned int cpu)
1393 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1394 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1397 for (spincnt = 0; spincnt < 10; spincnt++) {
1398 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1400 *statusp = RCU_KTHREAD_RUNNING;
1401 this_cpu_inc(rcu_cpu_kthread_loops);
1402 local_irq_disable();
1407 rcu_kthread_do_work();
1410 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1411 *statusp = RCU_KTHREAD_WAITING;
1415 *statusp = RCU_KTHREAD_YIELDING;
1416 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1417 schedule_timeout_interruptible(2);
1418 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1419 *statusp = RCU_KTHREAD_WAITING;
1423 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1424 * served by the rcu_node in question. The CPU hotplug lock is still
1425 * held, so the value of rnp->qsmaskinit will be stable.
1427 * We don't include outgoingcpu in the affinity set, use -1 if there is
1428 * no outgoing CPU. If there are no CPUs left in the affinity set,
1429 * this function allows the kthread to execute on any CPU.
1431 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1433 struct task_struct *t = rnp->boost_kthread_task;
1434 unsigned long mask = rnp->qsmaskinit;
1440 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1442 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1443 if ((mask & 0x1) && cpu != outgoingcpu)
1444 cpumask_set_cpu(cpu, cm);
1445 if (cpumask_weight(cm) == 0) {
1447 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1448 cpumask_clear_cpu(cpu, cm);
1449 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1451 set_cpus_allowed_ptr(t, cm);
1452 free_cpumask_var(cm);
1455 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1456 .store = &rcu_cpu_kthread_task,
1457 .thread_should_run = rcu_cpu_kthread_should_run,
1458 .thread_fn = rcu_cpu_kthread,
1459 .thread_comm = "rcuc/%u",
1460 .setup = rcu_cpu_kthread_setup,
1461 .park = rcu_cpu_kthread_park,
1465 * Spawn all kthreads -- called as soon as the scheduler is running.
1467 static int __init rcu_spawn_kthreads(void)
1469 struct rcu_node *rnp;
1472 rcu_scheduler_fully_active = 1;
1473 for_each_possible_cpu(cpu)
1474 per_cpu(rcu_cpu_has_work, cpu) = 0;
1475 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1476 rnp = rcu_get_root(rcu_state_p);
1477 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1478 if (NUM_RCU_NODES > 1) {
1479 rcu_for_each_leaf_node(rcu_state_p, rnp)
1480 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1484 early_initcall(rcu_spawn_kthreads);
1486 static void rcu_prepare_kthreads(int cpu)
1488 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1489 struct rcu_node *rnp = rdp->mynode;
1491 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1492 if (rcu_scheduler_fully_active)
1493 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1496 #else /* #ifdef CONFIG_RCU_BOOST */
1498 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1499 __releases(rnp->lock)
1501 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1504 static void invoke_rcu_callbacks_kthread(void)
1509 static bool rcu_is_callbacks_kthread(void)
1514 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1518 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1522 static int __init rcu_scheduler_really_started(void)
1524 rcu_scheduler_fully_active = 1;
1527 early_initcall(rcu_scheduler_really_started);
1529 static void rcu_prepare_kthreads(int cpu)
1533 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1535 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1538 * Check to see if any future RCU-related work will need to be done
1539 * by the current CPU, even if none need be done immediately, returning
1540 * 1 if so. This function is part of the RCU implementation; it is -not-
1541 * an exported member of the RCU API.
1543 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1544 * any flavor of RCU.
1546 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1547 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1549 *delta_jiffies = ULONG_MAX;
1550 return rcu_cpu_has_callbacks(cpu, NULL);
1552 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1555 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1558 static void rcu_cleanup_after_idle(int cpu)
1563 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1566 static void rcu_prepare_for_idle(int cpu)
1571 * Don't bother keeping a running count of the number of RCU callbacks
1572 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1574 static void rcu_idle_count_callbacks_posted(void)
1578 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1581 * This code is invoked when a CPU goes idle, at which point we want
1582 * to have the CPU do everything required for RCU so that it can enter
1583 * the energy-efficient dyntick-idle mode. This is handled by a
1584 * state machine implemented by rcu_prepare_for_idle() below.
1586 * The following three proprocessor symbols control this state machine:
1588 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1589 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1590 * is sized to be roughly one RCU grace period. Those energy-efficiency
1591 * benchmarkers who might otherwise be tempted to set this to a large
1592 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1593 * system. And if you are -that- concerned about energy efficiency,
1594 * just power the system down and be done with it!
1595 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1596 * permitted to sleep in dyntick-idle mode with only lazy RCU
1597 * callbacks pending. Setting this too high can OOM your system.
1599 * The values below work well in practice. If future workloads require
1600 * adjustment, they can be converted into kernel config parameters, though
1601 * making the state machine smarter might be a better option.
1603 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1604 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1606 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1607 module_param(rcu_idle_gp_delay, int, 0644);
1608 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1609 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1611 extern int tick_nohz_active;
1614 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1615 * only if it has been awhile since the last time we did so. Afterwards,
1616 * if there are any callbacks ready for immediate invocation, return true.
1618 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1620 bool cbs_ready = false;
1621 struct rcu_data *rdp;
1622 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1623 struct rcu_node *rnp;
1624 struct rcu_state *rsp;
1626 /* Exit early if we advanced recently. */
1627 if (jiffies == rdtp->last_advance_all)
1629 rdtp->last_advance_all = jiffies;
1631 for_each_rcu_flavor(rsp) {
1632 rdp = this_cpu_ptr(rsp->rda);
1636 * Don't bother checking unless a grace period has
1637 * completed since we last checked and there are
1638 * callbacks not yet ready to invoke.
1640 if (rdp->completed != rnp->completed &&
1641 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1642 note_gp_changes(rsp, rdp);
1644 if (cpu_has_callbacks_ready_to_invoke(rdp))
1651 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1652 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1653 * caller to set the timeout based on whether or not there are non-lazy
1656 * The caller must have disabled interrupts.
1658 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1659 int rcu_needs_cpu(int cpu, unsigned long *dj)
1661 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1663 /* Snapshot to detect later posting of non-lazy callback. */
1664 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1666 /* If no callbacks, RCU doesn't need the CPU. */
1667 if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1672 /* Attempt to advance callbacks. */
1673 if (rcu_try_advance_all_cbs()) {
1674 /* Some ready to invoke, so initiate later invocation. */
1678 rdtp->last_accelerate = jiffies;
1680 /* Request timer delay depending on laziness, and round. */
1681 if (!rdtp->all_lazy) {
1682 *dj = round_up(rcu_idle_gp_delay + jiffies,
1683 rcu_idle_gp_delay) - jiffies;
1685 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1689 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1692 * Prepare a CPU for idle from an RCU perspective. The first major task
1693 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1694 * The second major task is to check to see if a non-lazy callback has
1695 * arrived at a CPU that previously had only lazy callbacks. The third
1696 * major task is to accelerate (that is, assign grace-period numbers to)
1697 * any recently arrived callbacks.
1699 * The caller must have disabled interrupts.
1701 static void rcu_prepare_for_idle(int cpu)
1703 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1705 struct rcu_data *rdp;
1706 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1707 struct rcu_node *rnp;
1708 struct rcu_state *rsp;
1711 /* Handle nohz enablement switches conservatively. */
1712 tne = ACCESS_ONCE(tick_nohz_active);
1713 if (tne != rdtp->tick_nohz_enabled_snap) {
1714 if (rcu_cpu_has_callbacks(cpu, NULL))
1715 invoke_rcu_core(); /* force nohz to see update. */
1716 rdtp->tick_nohz_enabled_snap = tne;
1722 /* If this is a no-CBs CPU, no callbacks, just return. */
1723 if (rcu_is_nocb_cpu(cpu))
1727 * If a non-lazy callback arrived at a CPU having only lazy
1728 * callbacks, invoke RCU core for the side-effect of recalculating
1729 * idle duration on re-entry to idle.
1731 if (rdtp->all_lazy &&
1732 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1733 rdtp->all_lazy = false;
1734 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1740 * If we have not yet accelerated this jiffy, accelerate all
1741 * callbacks on this CPU.
1743 if (rdtp->last_accelerate == jiffies)
1745 rdtp->last_accelerate = jiffies;
1746 for_each_rcu_flavor(rsp) {
1747 rdp = per_cpu_ptr(rsp->rda, cpu);
1748 if (!*rdp->nxttail[RCU_DONE_TAIL])
1751 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1752 smp_mb__after_unlock_lock();
1753 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1754 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1756 rcu_gp_kthread_wake(rsp);
1758 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1762 * Clean up for exit from idle. Attempt to advance callbacks based on
1763 * any grace periods that elapsed while the CPU was idle, and if any
1764 * callbacks are now ready to invoke, initiate invocation.
1766 static void rcu_cleanup_after_idle(int cpu)
1768 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1769 if (rcu_is_nocb_cpu(cpu))
1771 if (rcu_try_advance_all_cbs())
1773 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1777 * Keep a running count of the number of non-lazy callbacks posted
1778 * on this CPU. This running counter (which is never decremented) allows
1779 * rcu_prepare_for_idle() to detect when something out of the idle loop
1780 * posts a callback, even if an equal number of callbacks are invoked.
1781 * Of course, callbacks should only be posted from within a trace event
1782 * designed to be called from idle or from within RCU_NONIDLE().
1784 static void rcu_idle_count_callbacks_posted(void)
1786 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1790 * Data for flushing lazy RCU callbacks at OOM time.
1792 static atomic_t oom_callback_count;
1793 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1796 * RCU OOM callback -- decrement the outstanding count and deliver the
1797 * wake-up if we are the last one.
1799 static void rcu_oom_callback(struct rcu_head *rhp)
1801 if (atomic_dec_and_test(&oom_callback_count))
1802 wake_up(&oom_callback_wq);
1806 * Post an rcu_oom_notify callback on the current CPU if it has at
1807 * least one lazy callback. This will unnecessarily post callbacks
1808 * to CPUs that already have a non-lazy callback at the end of their
1809 * callback list, but this is an infrequent operation, so accept some
1810 * extra overhead to keep things simple.
1812 static void rcu_oom_notify_cpu(void *unused)
1814 struct rcu_state *rsp;
1815 struct rcu_data *rdp;
1817 for_each_rcu_flavor(rsp) {
1818 rdp = raw_cpu_ptr(rsp->rda);
1819 if (rdp->qlen_lazy != 0) {
1820 atomic_inc(&oom_callback_count);
1821 rsp->call(&rdp->oom_head, rcu_oom_callback);
1827 * If low on memory, ensure that each CPU has a non-lazy callback.
1828 * This will wake up CPUs that have only lazy callbacks, in turn
1829 * ensuring that they free up the corresponding memory in a timely manner.
1830 * Because an uncertain amount of memory will be freed in some uncertain
1831 * timeframe, we do not claim to have freed anything.
1833 static int rcu_oom_notify(struct notifier_block *self,
1834 unsigned long notused, void *nfreed)
1838 /* Wait for callbacks from earlier instance to complete. */
1839 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1840 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1843 * Prevent premature wakeup: ensure that all increments happen
1844 * before there is a chance of the counter reaching zero.
1846 atomic_set(&oom_callback_count, 1);
1849 for_each_online_cpu(cpu) {
1850 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1855 /* Unconditionally decrement: no need to wake ourselves up. */
1856 atomic_dec(&oom_callback_count);
1861 static struct notifier_block rcu_oom_nb = {
1862 .notifier_call = rcu_oom_notify
1865 static int __init rcu_register_oom_notifier(void)
1867 register_oom_notifier(&rcu_oom_nb);
1870 early_initcall(rcu_register_oom_notifier);
1872 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1874 #ifdef CONFIG_RCU_CPU_STALL_INFO
1876 #ifdef CONFIG_RCU_FAST_NO_HZ
1878 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1880 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1881 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1883 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1884 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1886 rdtp->all_lazy ? 'L' : '.',
1887 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1890 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1892 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1897 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1899 /* Initiate the stall-info list. */
1900 static void print_cpu_stall_info_begin(void)
1906 * Print out diagnostic information for the specified stalled CPU.
1908 * If the specified CPU is aware of the current RCU grace period
1909 * (flavor specified by rsp), then print the number of scheduling
1910 * clock interrupts the CPU has taken during the time that it has
1911 * been aware. Otherwise, print the number of RCU grace periods
1912 * that this CPU is ignorant of, for example, "1" if the CPU was
1913 * aware of the previous grace period.
1915 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1917 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1919 char fast_no_hz[72];
1920 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1921 struct rcu_dynticks *rdtp = rdp->dynticks;
1923 unsigned long ticks_value;
1925 if (rsp->gpnum == rdp->gpnum) {
1926 ticks_title = "ticks this GP";
1927 ticks_value = rdp->ticks_this_gp;
1929 ticks_title = "GPs behind";
1930 ticks_value = rsp->gpnum - rdp->gpnum;
1932 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1933 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1934 cpu, ticks_value, ticks_title,
1935 atomic_read(&rdtp->dynticks) & 0xfff,
1936 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1937 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1941 /* Terminate the stall-info list. */
1942 static void print_cpu_stall_info_end(void)
1947 /* Zero ->ticks_this_gp for all flavors of RCU. */
1948 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1950 rdp->ticks_this_gp = 0;
1951 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1954 /* Increment ->ticks_this_gp for all flavors of RCU. */
1955 static void increment_cpu_stall_ticks(void)
1957 struct rcu_state *rsp;
1959 for_each_rcu_flavor(rsp)
1960 raw_cpu_inc(rsp->rda->ticks_this_gp);
1963 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1965 static void print_cpu_stall_info_begin(void)
1970 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1972 pr_cont(" %d", cpu);
1975 static void print_cpu_stall_info_end(void)
1980 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1984 static void increment_cpu_stall_ticks(void)
1988 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1990 #ifdef CONFIG_RCU_NOCB_CPU
1993 * Offload callback processing from the boot-time-specified set of CPUs
1994 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1995 * kthread created that pulls the callbacks from the corresponding CPU,
1996 * waits for a grace period to elapse, and invokes the callbacks.
1997 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1998 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1999 * has been specified, in which case each kthread actively polls its
2000 * CPU. (Which isn't so great for energy efficiency, but which does
2001 * reduce RCU's overhead on that CPU.)
2003 * This is intended to be used in conjunction with Frederic Weisbecker's
2004 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2005 * running CPU-bound user-mode computations.
2007 * Offloading of callback processing could also in theory be used as
2008 * an energy-efficiency measure because CPUs with no RCU callbacks
2009 * queued are more aggressive about entering dyntick-idle mode.
2013 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2014 static int __init rcu_nocb_setup(char *str)
2016 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
2017 have_rcu_nocb_mask = true;
2018 cpulist_parse(str, rcu_nocb_mask);
2021 __setup("rcu_nocbs=", rcu_nocb_setup);
2023 static int __init parse_rcu_nocb_poll(char *arg)
2028 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2031 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2034 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2036 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2040 * Set the root rcu_node structure's ->need_future_gp field
2041 * based on the sum of those of all rcu_node structures. This does
2042 * double-count the root rcu_node structure's requests, but this
2043 * is necessary to handle the possibility of a rcu_nocb_kthread()
2044 * having awakened during the time that the rcu_node structures
2045 * were being updated for the end of the previous grace period.
2047 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2049 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2052 static void rcu_init_one_nocb(struct rcu_node *rnp)
2054 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2055 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2058 #ifndef CONFIG_RCU_NOCB_CPU_ALL
2059 /* Is the specified CPU a no-CBs CPU? */
2060 bool rcu_is_nocb_cpu(int cpu)
2062 if (have_rcu_nocb_mask)
2063 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2066 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
2069 * Enqueue the specified string of rcu_head structures onto the specified
2070 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2071 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2072 * counts are supplied by rhcount and rhcount_lazy.
2074 * If warranted, also wake up the kthread servicing this CPUs queues.
2076 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2077 struct rcu_head *rhp,
2078 struct rcu_head **rhtp,
2079 int rhcount, int rhcount_lazy,
2080 unsigned long flags)
2083 struct rcu_head **old_rhpp;
2084 struct task_struct *t;
2086 /* Enqueue the callback on the nocb list and update counts. */
2087 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2088 ACCESS_ONCE(*old_rhpp) = rhp;
2089 atomic_long_add(rhcount, &rdp->nocb_q_count);
2090 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2092 /* If we are not being polled and there is a kthread, awaken it ... */
2093 t = ACCESS_ONCE(rdp->nocb_kthread);
2094 if (rcu_nocb_poll || !t) {
2095 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2096 TPS("WakeNotPoll"));
2099 len = atomic_long_read(&rdp->nocb_q_count);
2100 if (old_rhpp == &rdp->nocb_head) {
2101 if (!irqs_disabled_flags(flags)) {
2102 wake_up(&rdp->nocb_wq); /* ... if queue was empty ... */
2103 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2106 rdp->nocb_defer_wakeup = true;
2107 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2108 TPS("WakeEmptyIsDeferred"));
2110 rdp->qlen_last_fqs_check = 0;
2111 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2112 wake_up_process(t); /* ... or if many callbacks queued. */
2113 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2114 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf"));
2116 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2122 * This is a helper for __call_rcu(), which invokes this when the normal
2123 * callback queue is inoperable. If this is not a no-CBs CPU, this
2124 * function returns failure back to __call_rcu(), which can complain
2127 * Otherwise, this function queues the callback where the corresponding
2128 * "rcuo" kthread can find it.
2130 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2131 bool lazy, unsigned long flags)
2134 if (!rcu_is_nocb_cpu(rdp->cpu))
2136 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2137 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2138 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2139 (unsigned long)rhp->func,
2140 -atomic_long_read(&rdp->nocb_q_count_lazy),
2141 -atomic_long_read(&rdp->nocb_q_count));
2143 trace_rcu_callback(rdp->rsp->name, rhp,
2144 -atomic_long_read(&rdp->nocb_q_count_lazy),
2145 -atomic_long_read(&rdp->nocb_q_count));
2150 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2153 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2154 struct rcu_data *rdp,
2155 unsigned long flags)
2157 long ql = rsp->qlen;
2158 long qll = rsp->qlen_lazy;
2160 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2161 if (!rcu_is_nocb_cpu(smp_processor_id()))
2166 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2167 if (rsp->orphan_donelist != NULL) {
2168 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2169 rsp->orphan_donetail, ql, qll, flags);
2171 rsp->orphan_donelist = NULL;
2172 rsp->orphan_donetail = &rsp->orphan_donelist;
2174 if (rsp->orphan_nxtlist != NULL) {
2175 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2176 rsp->orphan_nxttail, ql, qll, flags);
2178 rsp->orphan_nxtlist = NULL;
2179 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2185 * If necessary, kick off a new grace period, and either way wait
2186 * for a subsequent grace period to complete.
2188 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2192 unsigned long flags;
2194 struct rcu_node *rnp = rdp->mynode;
2196 raw_spin_lock_irqsave(&rnp->lock, flags);
2197 smp_mb__after_unlock_lock();
2198 needwake = rcu_start_future_gp(rnp, rdp, &c);
2199 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2201 rcu_gp_kthread_wake(rdp->rsp);
2204 * Wait for the grace period. Do so interruptibly to avoid messing
2205 * up the load average.
2207 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2209 wait_event_interruptible(
2210 rnp->nocb_gp_wq[c & 0x1],
2211 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2214 flush_signals(current);
2215 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2217 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2218 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2222 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2223 * callbacks queued by the corresponding no-CBs CPU.
2225 static int rcu_nocb_kthread(void *arg)
2229 struct rcu_head *list;
2230 struct rcu_head *next;
2231 struct rcu_head **tail;
2232 struct rcu_data *rdp = arg;
2234 /* Each pass through this loop invokes one batch of callbacks */
2236 /* If not polling, wait for next batch of callbacks. */
2237 if (!rcu_nocb_poll) {
2238 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2240 wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head);
2241 /* Memory barrier provide by xchg() below. */
2242 } else if (firsttime) {
2244 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2247 list = ACCESS_ONCE(rdp->nocb_head);
2250 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2252 schedule_timeout_interruptible(1);
2253 flush_signals(current);
2257 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2258 TPS("WokeNonEmpty"));
2261 * Extract queued callbacks, update counts, and wait
2262 * for a grace period to elapse.
2264 ACCESS_ONCE(rdp->nocb_head) = NULL;
2265 tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2266 c = atomic_long_xchg(&rdp->nocb_q_count, 0);
2267 cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2268 ACCESS_ONCE(rdp->nocb_p_count) += c;
2269 ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl;
2270 rcu_nocb_wait_gp(rdp);
2272 /* Each pass through the following loop invokes a callback. */
2273 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2277 /* Wait for enqueuing to complete, if needed. */
2278 while (next == NULL && &list->next != tail) {
2279 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2281 schedule_timeout_interruptible(1);
2282 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2286 debug_rcu_head_unqueue(list);
2288 if (__rcu_reclaim(rdp->rsp->name, list))
2294 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2295 ACCESS_ONCE(rdp->nocb_p_count) -= c;
2296 ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2297 rdp->n_nocbs_invoked += c;
2302 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2303 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2305 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2308 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2309 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2311 if (!rcu_nocb_need_deferred_wakeup(rdp))
2313 ACCESS_ONCE(rdp->nocb_defer_wakeup) = false;
2314 wake_up(&rdp->nocb_wq);
2315 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWakeEmpty"));
2318 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2319 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2321 rdp->nocb_tail = &rdp->nocb_head;
2322 init_waitqueue_head(&rdp->nocb_wq);
2325 /* Create a kthread for each RCU flavor for each no-CBs CPU. */
2326 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2329 struct rcu_data *rdp;
2330 struct task_struct *t;
2332 if (rcu_nocb_mask == NULL)
2334 for_each_cpu(cpu, rcu_nocb_mask) {
2335 rdp = per_cpu_ptr(rsp->rda, cpu);
2336 t = kthread_run(rcu_nocb_kthread, rdp,
2337 "rcuo%c/%d", rsp->abbr, cpu);
2339 ACCESS_ONCE(rdp->nocb_kthread) = t;
2343 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2344 static bool init_nocb_callback_list(struct rcu_data *rdp)
2346 if (rcu_nocb_mask == NULL ||
2347 !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2349 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2353 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2355 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2359 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2363 static void rcu_init_one_nocb(struct rcu_node *rnp)
2367 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2368 bool lazy, unsigned long flags)
2373 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2374 struct rcu_data *rdp,
2375 unsigned long flags)
2380 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2384 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2389 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2393 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2397 static bool init_nocb_callback_list(struct rcu_data *rdp)
2402 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2405 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2406 * arbitrarily long period of time with the scheduling-clock tick turned
2407 * off. RCU will be paying attention to this CPU because it is in the
2408 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2409 * machine because the scheduling-clock tick has been disabled. Therefore,
2410 * if an adaptive-ticks CPU is failing to respond to the current grace
2411 * period and has not be idle from an RCU perspective, kick it.
2413 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2415 #ifdef CONFIG_NO_HZ_FULL
2416 if (tick_nohz_full_cpu(cpu))
2417 smp_send_reschedule(cpu);
2418 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2422 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2425 * Define RCU flavor that holds sysidle state. This needs to be the
2426 * most active flavor of RCU.
2428 #ifdef CONFIG_PREEMPT_RCU
2429 static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state;
2430 #else /* #ifdef CONFIG_PREEMPT_RCU */
2431 static struct rcu_state *rcu_sysidle_state = &rcu_sched_state;
2432 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2434 static int full_sysidle_state; /* Current system-idle state. */
2435 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2436 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2437 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2438 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2439 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2442 * Invoked to note exit from irq or task transition to idle. Note that
2443 * usermode execution does -not- count as idle here! After all, we want
2444 * to detect full-system idle states, not RCU quiescent states and grace
2445 * periods. The caller must have disabled interrupts.
2447 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2451 /* Adjust nesting, check for fully idle. */
2453 rdtp->dynticks_idle_nesting--;
2454 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2455 if (rdtp->dynticks_idle_nesting != 0)
2456 return; /* Still not fully idle. */
2458 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2459 DYNTICK_TASK_NEST_VALUE) {
2460 rdtp->dynticks_idle_nesting = 0;
2462 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2463 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2464 return; /* Still not fully idle. */
2468 /* Record start of fully idle period. */
2470 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2471 smp_mb__before_atomic();
2472 atomic_inc(&rdtp->dynticks_idle);
2473 smp_mb__after_atomic();
2474 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2478 * Unconditionally force exit from full system-idle state. This is
2479 * invoked when a normal CPU exits idle, but must be called separately
2480 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2481 * is that the timekeeping CPU is permitted to take scheduling-clock
2482 * interrupts while the system is in system-idle state, and of course
2483 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2484 * interrupt from any other type of interrupt.
2486 void rcu_sysidle_force_exit(void)
2488 int oldstate = ACCESS_ONCE(full_sysidle_state);
2492 * Each pass through the following loop attempts to exit full
2493 * system-idle state. If contention proves to be a problem,
2494 * a trylock-based contention tree could be used here.
2496 while (oldstate > RCU_SYSIDLE_SHORT) {
2497 newoldstate = cmpxchg(&full_sysidle_state,
2498 oldstate, RCU_SYSIDLE_NOT);
2499 if (oldstate == newoldstate &&
2500 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2501 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2502 return; /* We cleared it, done! */
2504 oldstate = newoldstate;
2506 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2510 * Invoked to note entry to irq or task transition from idle. Note that
2511 * usermode execution does -not- count as idle here! The caller must
2512 * have disabled interrupts.
2514 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2516 /* Adjust nesting, check for already non-idle. */
2518 rdtp->dynticks_idle_nesting++;
2519 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2520 if (rdtp->dynticks_idle_nesting != 1)
2521 return; /* Already non-idle. */
2524 * Allow for irq misnesting. Yes, it really is possible
2525 * to enter an irq handler then never leave it, and maybe
2526 * also vice versa. Handle both possibilities.
2528 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2529 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2530 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2531 return; /* Already non-idle. */
2533 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2537 /* Record end of idle period. */
2538 smp_mb__before_atomic();
2539 atomic_inc(&rdtp->dynticks_idle);
2540 smp_mb__after_atomic();
2541 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2544 * If we are the timekeeping CPU, we are permitted to be non-idle
2545 * during a system-idle state. This must be the case, because
2546 * the timekeeping CPU has to take scheduling-clock interrupts
2547 * during the time that the system is transitioning to full
2548 * system-idle state. This means that the timekeeping CPU must
2549 * invoke rcu_sysidle_force_exit() directly if it does anything
2550 * more than take a scheduling-clock interrupt.
2552 if (smp_processor_id() == tick_do_timer_cpu)
2555 /* Update system-idle state: We are clearly no longer fully idle! */
2556 rcu_sysidle_force_exit();
2560 * Check to see if the current CPU is idle. Note that usermode execution
2561 * does not count as idle. The caller must have disabled interrupts.
2563 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2564 unsigned long *maxj)
2568 struct rcu_dynticks *rdtp = rdp->dynticks;
2571 * If some other CPU has already reported non-idle, if this is
2572 * not the flavor of RCU that tracks sysidle state, or if this
2573 * is an offline or the timekeeping CPU, nothing to do.
2575 if (!*isidle || rdp->rsp != rcu_sysidle_state ||
2576 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2578 if (rcu_gp_in_progress(rdp->rsp))
2579 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2581 /* Pick up current idle and NMI-nesting counter and check. */
2582 cur = atomic_read(&rdtp->dynticks_idle);
2584 *isidle = false; /* We are not idle! */
2587 smp_mb(); /* Read counters before timestamps. */
2589 /* Pick up timestamps. */
2590 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2591 /* If this CPU entered idle more recently, update maxj timestamp. */
2592 if (ULONG_CMP_LT(*maxj, j))
2597 * Is this the flavor of RCU that is handling full-system idle?
2599 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2601 return rsp == rcu_sysidle_state;
2605 * Return a delay in jiffies based on the number of CPUs, rcu_node
2606 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2607 * systems more time to transition to full-idle state in order to
2608 * avoid the cache thrashing that otherwise occur on the state variable.
2609 * Really small systems (less than a couple of tens of CPUs) should
2610 * instead use a single global atomically incremented counter, and later
2611 * versions of this will automatically reconfigure themselves accordingly.
2613 static unsigned long rcu_sysidle_delay(void)
2615 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2617 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2621 * Advance the full-system-idle state. This is invoked when all of
2622 * the non-timekeeping CPUs are idle.
2624 static void rcu_sysidle(unsigned long j)
2626 /* Check the current state. */
2627 switch (ACCESS_ONCE(full_sysidle_state)) {
2628 case RCU_SYSIDLE_NOT:
2630 /* First time all are idle, so note a short idle period. */
2631 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2634 case RCU_SYSIDLE_SHORT:
2637 * Idle for a bit, time to advance to next state?
2638 * cmpxchg failure means race with non-idle, let them win.
2640 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2641 (void)cmpxchg(&full_sysidle_state,
2642 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2645 case RCU_SYSIDLE_LONG:
2648 * Do an additional check pass before advancing to full.
2649 * cmpxchg failure means race with non-idle, let them win.
2651 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2652 (void)cmpxchg(&full_sysidle_state,
2653 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2662 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2663 * back to the beginning.
2665 static void rcu_sysidle_cancel(void)
2668 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2669 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2673 * Update the sysidle state based on the results of a force-quiescent-state
2674 * scan of the CPUs' dyntick-idle state.
2676 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2677 unsigned long maxj, bool gpkt)
2679 if (rsp != rcu_sysidle_state)
2680 return; /* Wrong flavor, ignore. */
2681 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2682 return; /* Running state machine from timekeeping CPU. */
2684 rcu_sysidle(maxj); /* More idle! */
2686 rcu_sysidle_cancel(); /* Idle is over. */
2690 * Wrapper for rcu_sysidle_report() when called from the grace-period
2691 * kthread's context.
2693 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2696 rcu_sysidle_report(rsp, isidle, maxj, true);
2699 /* Callback and function for forcing an RCU grace period. */
2700 struct rcu_sysidle_head {
2705 static void rcu_sysidle_cb(struct rcu_head *rhp)
2707 struct rcu_sysidle_head *rshp;
2710 * The following memory barrier is needed to replace the
2711 * memory barriers that would normally be in the memory
2714 smp_mb(); /* grace period precedes setting inuse. */
2716 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2717 ACCESS_ONCE(rshp->inuse) = 0;
2721 * Check to see if the system is fully idle, other than the timekeeping CPU.
2722 * The caller must have disabled interrupts.
2724 bool rcu_sys_is_idle(void)
2726 static struct rcu_sysidle_head rsh;
2727 int rss = ACCESS_ONCE(full_sysidle_state);
2729 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2732 /* Handle small-system case by doing a full scan of CPUs. */
2733 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2734 int oldrss = rss - 1;
2737 * One pass to advance to each state up to _FULL.
2738 * Give up if any pass fails to advance the state.
2740 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2743 unsigned long maxj = jiffies - ULONG_MAX / 4;
2744 struct rcu_data *rdp;
2746 /* Scan all the CPUs looking for nonidle CPUs. */
2747 for_each_possible_cpu(cpu) {
2748 rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu);
2749 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2753 rcu_sysidle_report(rcu_sysidle_state,
2754 isidle, maxj, false);
2756 rss = ACCESS_ONCE(full_sysidle_state);
2760 /* If this is the first observation of an idle period, record it. */
2761 if (rss == RCU_SYSIDLE_FULL) {
2762 rss = cmpxchg(&full_sysidle_state,
2763 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2764 return rss == RCU_SYSIDLE_FULL;
2767 smp_mb(); /* ensure rss load happens before later caller actions. */
2769 /* If already fully idle, tell the caller (in case of races). */
2770 if (rss == RCU_SYSIDLE_FULL_NOTED)
2774 * If we aren't there yet, and a grace period is not in flight,
2775 * initiate a grace period. Either way, tell the caller that
2776 * we are not there yet. We use an xchg() rather than an assignment
2777 * to make up for the memory barriers that would otherwise be
2778 * provided by the memory allocator.
2780 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2781 !rcu_gp_in_progress(rcu_sysidle_state) &&
2782 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2783 call_rcu(&rsh.rh, rcu_sysidle_cb);
2788 * Initialize dynticks sysidle state for CPUs coming online.
2790 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2792 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2795 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2797 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2801 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2805 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2806 unsigned long *maxj)
2810 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2815 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2820 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2824 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2827 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2828 * grace-period kthread will do force_quiescent_state() processing?
2829 * The idea is to avoid waking up RCU core processing on such a
2830 * CPU unless the grace period has extended for too long.
2832 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2833 * CONFIG_RCU_NOCB_CPU CPUs.
2835 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2837 #ifdef CONFIG_NO_HZ_FULL
2838 if (tick_nohz_full_cpu(smp_processor_id()) &&
2839 (!rcu_gp_in_progress(rsp) ||
2840 ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
2842 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2847 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2850 static void rcu_bind_gp_kthread(void)
2852 int __maybe_unused cpu;
2854 if (!tick_nohz_full_enabled())
2856 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2857 cpu = tick_do_timer_cpu;
2858 if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu)
2859 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2860 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2861 if (!is_housekeeping_cpu(raw_smp_processor_id()))
2862 housekeeping_affine(current);
2863 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */