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);
90 #ifdef CONFIG_TREE_PREEMPT_RCU
92 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
93 static struct rcu_state *rcu_state_p = &rcu_preempt_state;
95 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
98 * Tell them what RCU they are running.
100 static void __init rcu_bootup_announce(void)
102 pr_info("Preemptible hierarchical RCU implementation.\n");
103 rcu_bootup_announce_oddness();
107 * Return the number of RCU-preempt batches processed thus far
108 * for debug and statistics.
110 long rcu_batches_completed_preempt(void)
112 return rcu_preempt_state.completed;
114 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
117 * Return the number of RCU batches processed thus far for debug & stats.
119 long rcu_batches_completed(void)
121 return rcu_batches_completed_preempt();
123 EXPORT_SYMBOL_GPL(rcu_batches_completed);
126 * Record a preemptible-RCU quiescent state for the specified CPU. Note
127 * that this just means that the task currently running on the CPU is
128 * not in a quiescent state. There might be any number of tasks blocked
129 * while in an RCU read-side critical section.
131 * Unlike the other rcu_*_qs() functions, callers to this function
132 * must disable irqs in order to protect the assignment to
133 * ->rcu_read_unlock_special.
135 static void rcu_preempt_qs(int cpu)
137 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
139 if (rdp->passed_quiesce == 0)
140 trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs"));
141 rdp->passed_quiesce = 1;
142 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
146 * We have entered the scheduler, and the current task might soon be
147 * context-switched away from. If this task is in an RCU read-side
148 * critical section, we will no longer be able to rely on the CPU to
149 * record that fact, so we enqueue the task on the blkd_tasks list.
150 * The task will dequeue itself when it exits the outermost enclosing
151 * RCU read-side critical section. Therefore, the current grace period
152 * cannot be permitted to complete until the blkd_tasks list entries
153 * predating the current grace period drain, in other words, until
154 * rnp->gp_tasks becomes NULL.
156 * Caller must disable preemption.
158 static void rcu_preempt_note_context_switch(int cpu)
160 struct task_struct *t = current;
162 struct rcu_data *rdp;
163 struct rcu_node *rnp;
165 if (t->rcu_read_lock_nesting > 0 &&
166 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
168 /* Possibly blocking in an RCU read-side critical section. */
169 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
171 raw_spin_lock_irqsave(&rnp->lock, flags);
172 smp_mb__after_unlock_lock();
173 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
174 t->rcu_blocked_node = rnp;
177 * If this CPU has already checked in, then this task
178 * will hold up the next grace period rather than the
179 * current grace period. Queue the task accordingly.
180 * If the task is queued for the current grace period
181 * (i.e., this CPU has not yet passed through a quiescent
182 * state for the current grace period), then as long
183 * as that task remains queued, the current grace period
184 * cannot end. Note that there is some uncertainty as
185 * to exactly when the current grace period started.
186 * We take a conservative approach, which can result
187 * in unnecessarily waiting on tasks that started very
188 * slightly after the current grace period began. C'est
191 * But first, note that the current CPU must still be
194 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
195 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
196 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
197 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
198 rnp->gp_tasks = &t->rcu_node_entry;
199 #ifdef CONFIG_RCU_BOOST
200 if (rnp->boost_tasks != NULL)
201 rnp->boost_tasks = rnp->gp_tasks;
202 #endif /* #ifdef CONFIG_RCU_BOOST */
204 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
205 if (rnp->qsmask & rdp->grpmask)
206 rnp->gp_tasks = &t->rcu_node_entry;
208 trace_rcu_preempt_task(rdp->rsp->name,
210 (rnp->qsmask & rdp->grpmask)
213 raw_spin_unlock_irqrestore(&rnp->lock, flags);
214 } else if (t->rcu_read_lock_nesting < 0 &&
215 t->rcu_read_unlock_special) {
218 * Complete exit from RCU read-side critical section on
219 * behalf of preempted instance of __rcu_read_unlock().
221 rcu_read_unlock_special(t);
225 * Either we were not in an RCU read-side critical section to
226 * begin with, or we have now recorded that critical section
227 * globally. Either way, we can now note a quiescent state
228 * for this CPU. Again, if we were in an RCU read-side critical
229 * section, and if that critical section was blocking the current
230 * grace period, then the fact that the task has been enqueued
231 * means that we continue to block the current grace period.
233 local_irq_save(flags);
235 local_irq_restore(flags);
239 * Check for preempted RCU readers blocking the current grace period
240 * for the specified rcu_node structure. If the caller needs a reliable
241 * answer, it must hold the rcu_node's ->lock.
243 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
245 return rnp->gp_tasks != NULL;
249 * Record a quiescent state for all tasks that were previously queued
250 * on the specified rcu_node structure and that were blocking the current
251 * RCU grace period. The caller must hold the specified rnp->lock with
252 * irqs disabled, and this lock is released upon return, but irqs remain
255 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
256 __releases(rnp->lock)
259 struct rcu_node *rnp_p;
261 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
262 raw_spin_unlock_irqrestore(&rnp->lock, flags);
263 return; /* Still need more quiescent states! */
269 * Either there is only one rcu_node in the tree,
270 * or tasks were kicked up to root rcu_node due to
271 * CPUs going offline.
273 rcu_report_qs_rsp(&rcu_preempt_state, flags);
277 /* Report up the rest of the hierarchy. */
279 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
280 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
281 smp_mb__after_unlock_lock();
282 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
286 * Advance a ->blkd_tasks-list pointer to the next entry, instead
287 * returning NULL if at the end of the list.
289 static struct list_head *rcu_next_node_entry(struct task_struct *t,
290 struct rcu_node *rnp)
292 struct list_head *np;
294 np = t->rcu_node_entry.next;
295 if (np == &rnp->blkd_tasks)
301 * Handle special cases during rcu_read_unlock(), such as needing to
302 * notify RCU core processing or task having blocked during the RCU
303 * read-side critical section.
305 void rcu_read_unlock_special(struct task_struct *t)
311 struct list_head *np;
312 #ifdef CONFIG_RCU_BOOST
313 bool drop_boost_mutex = false;
314 #endif /* #ifdef CONFIG_RCU_BOOST */
315 struct rcu_node *rnp;
318 /* NMI handlers cannot block and cannot safely manipulate state. */
322 local_irq_save(flags);
325 * If RCU core is waiting for this CPU to exit critical section,
326 * let it know that we have done so.
328 special = t->rcu_read_unlock_special;
329 if (special & RCU_READ_UNLOCK_NEED_QS) {
330 rcu_preempt_qs(smp_processor_id());
331 if (!t->rcu_read_unlock_special) {
332 local_irq_restore(flags);
337 /* Hardware IRQ handlers cannot block, complain if they get here. */
338 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
339 local_irq_restore(flags);
343 /* Clean up if blocked during RCU read-side critical section. */
344 if (special & RCU_READ_UNLOCK_BLOCKED) {
345 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
348 * Remove this task from the list it blocked on. The
349 * task can migrate while we acquire the lock, but at
350 * most one time. So at most two passes through loop.
353 rnp = t->rcu_blocked_node;
354 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
355 smp_mb__after_unlock_lock();
356 if (rnp == t->rcu_blocked_node)
358 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
360 empty = !rcu_preempt_blocked_readers_cgp(rnp);
361 empty_exp = !rcu_preempted_readers_exp(rnp);
362 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
363 np = rcu_next_node_entry(t, rnp);
364 list_del_init(&t->rcu_node_entry);
365 t->rcu_blocked_node = NULL;
366 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
368 if (&t->rcu_node_entry == rnp->gp_tasks)
370 if (&t->rcu_node_entry == rnp->exp_tasks)
372 #ifdef CONFIG_RCU_BOOST
373 if (&t->rcu_node_entry == rnp->boost_tasks)
374 rnp->boost_tasks = np;
375 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
376 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
377 #endif /* #ifdef CONFIG_RCU_BOOST */
380 * If this was the last task on the current list, and if
381 * we aren't waiting on any CPUs, report the quiescent state.
382 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
383 * so we must take a snapshot of the expedited state.
385 empty_exp_now = !rcu_preempted_readers_exp(rnp);
386 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
387 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
394 rcu_report_unblock_qs_rnp(rnp, flags);
396 raw_spin_unlock_irqrestore(&rnp->lock, flags);
399 #ifdef CONFIG_RCU_BOOST
400 /* Unboost if we were boosted. */
401 if (drop_boost_mutex) {
402 rt_mutex_unlock(&rnp->boost_mtx);
403 complete(&rnp->boost_completion);
405 #endif /* #ifdef CONFIG_RCU_BOOST */
408 * If this was the last task on the expedited lists,
409 * then we need to report up the rcu_node hierarchy.
411 if (!empty_exp && empty_exp_now)
412 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
414 local_irq_restore(flags);
418 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
421 * Dump detailed information for all tasks blocking the current RCU
422 * grace period on the specified rcu_node structure.
424 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
427 struct task_struct *t;
429 raw_spin_lock_irqsave(&rnp->lock, flags);
430 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
431 raw_spin_unlock_irqrestore(&rnp->lock, flags);
434 t = list_entry(rnp->gp_tasks,
435 struct task_struct, rcu_node_entry);
436 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
438 raw_spin_unlock_irqrestore(&rnp->lock, flags);
442 * Dump detailed information for all tasks blocking the current RCU
445 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
447 struct rcu_node *rnp = rcu_get_root(rsp);
449 rcu_print_detail_task_stall_rnp(rnp);
450 rcu_for_each_leaf_node(rsp, rnp)
451 rcu_print_detail_task_stall_rnp(rnp);
454 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
456 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
460 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
462 #ifdef CONFIG_RCU_CPU_STALL_INFO
464 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
466 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
467 rnp->level, rnp->grplo, rnp->grphi);
470 static void rcu_print_task_stall_end(void)
475 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
477 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
481 static void rcu_print_task_stall_end(void)
485 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
488 * Scan the current list of tasks blocked within RCU read-side critical
489 * sections, printing out the tid of each.
491 static int rcu_print_task_stall(struct rcu_node *rnp)
493 struct task_struct *t;
496 if (!rcu_preempt_blocked_readers_cgp(rnp))
498 rcu_print_task_stall_begin(rnp);
499 t = list_entry(rnp->gp_tasks,
500 struct task_struct, rcu_node_entry);
501 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
502 pr_cont(" P%d", t->pid);
505 rcu_print_task_stall_end();
510 * Check that the list of blocked tasks for the newly completed grace
511 * period is in fact empty. It is a serious bug to complete a grace
512 * period that still has RCU readers blocked! This function must be
513 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
514 * must be held by the caller.
516 * Also, if there are blocked tasks on the list, they automatically
517 * block the newly created grace period, so set up ->gp_tasks accordingly.
519 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
521 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
522 if (!list_empty(&rnp->blkd_tasks))
523 rnp->gp_tasks = rnp->blkd_tasks.next;
524 WARN_ON_ONCE(rnp->qsmask);
527 #ifdef CONFIG_HOTPLUG_CPU
530 * Handle tasklist migration for case in which all CPUs covered by the
531 * specified rcu_node have gone offline. Move them up to the root
532 * rcu_node. The reason for not just moving them to the immediate
533 * parent is to remove the need for rcu_read_unlock_special() to
534 * make more than two attempts to acquire the target rcu_node's lock.
535 * Returns true if there were tasks blocking the current RCU grace
538 * Returns 1 if there was previously a task blocking the current grace
539 * period on the specified rcu_node structure.
541 * The caller must hold rnp->lock with irqs disabled.
543 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
544 struct rcu_node *rnp,
545 struct rcu_data *rdp)
547 struct list_head *lp;
548 struct list_head *lp_root;
550 struct rcu_node *rnp_root = rcu_get_root(rsp);
551 struct task_struct *t;
553 if (rnp == rnp_root) {
554 WARN_ONCE(1, "Last CPU thought to be offlined?");
555 return 0; /* Shouldn't happen: at least one CPU online. */
558 /* If we are on an internal node, complain bitterly. */
559 WARN_ON_ONCE(rnp != rdp->mynode);
562 * Move tasks up to root rcu_node. Don't try to get fancy for
563 * this corner-case operation -- just put this node's tasks
564 * at the head of the root node's list, and update the root node's
565 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
566 * if non-NULL. This might result in waiting for more tasks than
567 * absolutely necessary, but this is a good performance/complexity
570 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
571 retval |= RCU_OFL_TASKS_NORM_GP;
572 if (rcu_preempted_readers_exp(rnp))
573 retval |= RCU_OFL_TASKS_EXP_GP;
574 lp = &rnp->blkd_tasks;
575 lp_root = &rnp_root->blkd_tasks;
576 while (!list_empty(lp)) {
577 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
578 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
579 smp_mb__after_unlock_lock();
580 list_del(&t->rcu_node_entry);
581 t->rcu_blocked_node = rnp_root;
582 list_add(&t->rcu_node_entry, lp_root);
583 if (&t->rcu_node_entry == rnp->gp_tasks)
584 rnp_root->gp_tasks = rnp->gp_tasks;
585 if (&t->rcu_node_entry == rnp->exp_tasks)
586 rnp_root->exp_tasks = rnp->exp_tasks;
587 #ifdef CONFIG_RCU_BOOST
588 if (&t->rcu_node_entry == rnp->boost_tasks)
589 rnp_root->boost_tasks = rnp->boost_tasks;
590 #endif /* #ifdef CONFIG_RCU_BOOST */
591 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
594 rnp->gp_tasks = NULL;
595 rnp->exp_tasks = NULL;
596 #ifdef CONFIG_RCU_BOOST
597 rnp->boost_tasks = NULL;
599 * In case root is being boosted and leaf was not. Make sure
600 * that we boost the tasks blocking the current grace period
603 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
604 smp_mb__after_unlock_lock();
605 if (rnp_root->boost_tasks != NULL &&
606 rnp_root->boost_tasks != rnp_root->gp_tasks &&
607 rnp_root->boost_tasks != rnp_root->exp_tasks)
608 rnp_root->boost_tasks = rnp_root->gp_tasks;
609 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
610 #endif /* #ifdef CONFIG_RCU_BOOST */
615 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
618 * Check for a quiescent state from the current CPU. When a task blocks,
619 * the task is recorded in the corresponding CPU's rcu_node structure,
620 * which is checked elsewhere.
622 * Caller must disable hard irqs.
624 static void rcu_preempt_check_callbacks(int cpu)
626 struct task_struct *t = current;
628 if (t->rcu_read_lock_nesting == 0) {
632 if (t->rcu_read_lock_nesting > 0 &&
633 per_cpu(rcu_preempt_data, cpu).qs_pending)
634 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
637 #ifdef CONFIG_RCU_BOOST
639 static void rcu_preempt_do_callbacks(void)
641 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
644 #endif /* #ifdef CONFIG_RCU_BOOST */
647 * Queue a preemptible-RCU callback for invocation after a grace period.
649 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
651 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
653 EXPORT_SYMBOL_GPL(call_rcu);
656 * synchronize_rcu - wait until a grace period has elapsed.
658 * Control will return to the caller some time after a full grace
659 * period has elapsed, in other words after all currently executing RCU
660 * read-side critical sections have completed. Note, however, that
661 * upon return from synchronize_rcu(), the caller might well be executing
662 * concurrently with new RCU read-side critical sections that began while
663 * synchronize_rcu() was waiting. RCU read-side critical sections are
664 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
666 * See the description of synchronize_sched() for more detailed information
667 * on memory ordering guarantees.
669 void synchronize_rcu(void)
671 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
672 !lock_is_held(&rcu_lock_map) &&
673 !lock_is_held(&rcu_sched_lock_map),
674 "Illegal synchronize_rcu() in RCU read-side critical section");
675 if (!rcu_scheduler_active)
678 synchronize_rcu_expedited();
680 wait_rcu_gp(call_rcu);
682 EXPORT_SYMBOL_GPL(synchronize_rcu);
684 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
685 static unsigned long sync_rcu_preempt_exp_count;
686 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
689 * Return non-zero if there are any tasks in RCU read-side critical
690 * sections blocking the current preemptible-RCU expedited grace period.
691 * If there is no preemptible-RCU expedited grace period currently in
692 * progress, returns zero unconditionally.
694 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
696 return rnp->exp_tasks != NULL;
700 * return non-zero if there is no RCU expedited grace period in progress
701 * for the specified rcu_node structure, in other words, if all CPUs and
702 * tasks covered by the specified rcu_node structure have done their bit
703 * for the current expedited grace period. Works only for preemptible
704 * RCU -- other RCU implementation use other means.
706 * Caller must hold sync_rcu_preempt_exp_mutex.
708 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
710 return !rcu_preempted_readers_exp(rnp) &&
711 ACCESS_ONCE(rnp->expmask) == 0;
715 * Report the exit from RCU read-side critical section for the last task
716 * that queued itself during or before the current expedited preemptible-RCU
717 * grace period. This event is reported either to the rcu_node structure on
718 * which the task was queued or to one of that rcu_node structure's ancestors,
719 * recursively up the tree. (Calm down, calm down, we do the recursion
722 * Most callers will set the "wake" flag, but the task initiating the
723 * expedited grace period need not wake itself.
725 * Caller must hold sync_rcu_preempt_exp_mutex.
727 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
733 raw_spin_lock_irqsave(&rnp->lock, flags);
734 smp_mb__after_unlock_lock();
736 if (!sync_rcu_preempt_exp_done(rnp)) {
737 raw_spin_unlock_irqrestore(&rnp->lock, flags);
740 if (rnp->parent == NULL) {
741 raw_spin_unlock_irqrestore(&rnp->lock, flags);
743 smp_mb(); /* EGP done before wake_up(). */
744 wake_up(&sync_rcu_preempt_exp_wq);
749 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
751 raw_spin_lock(&rnp->lock); /* irqs already disabled */
752 smp_mb__after_unlock_lock();
753 rnp->expmask &= ~mask;
758 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
759 * grace period for the specified rcu_node structure. If there are no such
760 * tasks, report it up the rcu_node hierarchy.
762 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
763 * CPU hotplug operations.
766 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
771 raw_spin_lock_irqsave(&rnp->lock, flags);
772 smp_mb__after_unlock_lock();
773 if (list_empty(&rnp->blkd_tasks)) {
774 raw_spin_unlock_irqrestore(&rnp->lock, flags);
776 rnp->exp_tasks = rnp->blkd_tasks.next;
777 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
781 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
785 * synchronize_rcu_expedited - Brute-force RCU grace period
787 * Wait for an RCU-preempt grace period, but expedite it. The basic
788 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
789 * the ->blkd_tasks lists and wait for this list to drain. This consumes
790 * significant time on all CPUs and is unfriendly to real-time workloads,
791 * so is thus not recommended for any sort of common-case code.
792 * In fact, if you are using synchronize_rcu_expedited() in a loop,
793 * please restructure your code to batch your updates, and then Use a
794 * single synchronize_rcu() instead.
796 * Note that it is illegal to call this function while holding any lock
797 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
798 * to call this function from a CPU-hotplug notifier. Failing to observe
799 * these restriction will result in deadlock.
801 void synchronize_rcu_expedited(void)
804 struct rcu_node *rnp;
805 struct rcu_state *rsp = &rcu_preempt_state;
809 smp_mb(); /* Caller's modifications seen first by other CPUs. */
810 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
811 smp_mb(); /* Above access cannot bleed into critical section. */
814 * Block CPU-hotplug operations. This means that any CPU-hotplug
815 * operation that finds an rcu_node structure with tasks in the
816 * process of being boosted will know that all tasks blocking
817 * this expedited grace period will already be in the process of
818 * being boosted. This simplifies the process of moving tasks
819 * from leaf to root rcu_node structures.
824 * Acquire lock, falling back to synchronize_rcu() if too many
825 * lock-acquisition failures. Of course, if someone does the
826 * expedited grace period for us, just leave.
828 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
829 if (ULONG_CMP_LT(snap,
830 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
832 goto mb_ret; /* Others did our work for us. */
834 if (trycount++ < 10) {
835 udelay(trycount * num_online_cpus());
838 wait_rcu_gp(call_rcu);
842 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
844 goto unlock_mb_ret; /* Others did our work for us. */
847 /* force all RCU readers onto ->blkd_tasks lists. */
848 synchronize_sched_expedited();
850 /* Initialize ->expmask for all non-leaf rcu_node structures. */
851 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
852 raw_spin_lock_irqsave(&rnp->lock, flags);
853 smp_mb__after_unlock_lock();
854 rnp->expmask = rnp->qsmaskinit;
855 raw_spin_unlock_irqrestore(&rnp->lock, flags);
858 /* Snapshot current state of ->blkd_tasks lists. */
859 rcu_for_each_leaf_node(rsp, rnp)
860 sync_rcu_preempt_exp_init(rsp, rnp);
861 if (NUM_RCU_NODES > 1)
862 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
866 /* Wait for snapshotted ->blkd_tasks lists to drain. */
867 rnp = rcu_get_root(rsp);
868 wait_event(sync_rcu_preempt_exp_wq,
869 sync_rcu_preempt_exp_done(rnp));
871 /* Clean up and exit. */
872 smp_mb(); /* ensure expedited GP seen before counter increment. */
873 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
875 mutex_unlock(&sync_rcu_preempt_exp_mutex);
877 smp_mb(); /* ensure subsequent action seen after grace period. */
879 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
882 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
884 * Note that this primitive does not necessarily wait for an RCU grace period
885 * to complete. For example, if there are no RCU callbacks queued anywhere
886 * in the system, then rcu_barrier() is within its rights to return
887 * immediately, without waiting for anything, much less an RCU grace period.
889 void rcu_barrier(void)
891 _rcu_barrier(&rcu_preempt_state);
893 EXPORT_SYMBOL_GPL(rcu_barrier);
896 * Initialize preemptible RCU's state structures.
898 static void __init __rcu_init_preempt(void)
900 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
904 * Check for a task exiting while in a preemptible-RCU read-side
905 * critical section, clean up if so. No need to issue warnings,
906 * as debug_check_no_locks_held() already does this if lockdep
911 struct task_struct *t = current;
913 if (likely(list_empty(¤t->rcu_node_entry)))
915 t->rcu_read_lock_nesting = 1;
917 t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED;
921 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
923 static struct rcu_state *rcu_state_p = &rcu_sched_state;
926 * Tell them what RCU they are running.
928 static void __init rcu_bootup_announce(void)
930 pr_info("Hierarchical RCU implementation.\n");
931 rcu_bootup_announce_oddness();
935 * Return the number of RCU batches processed thus far for debug & stats.
937 long rcu_batches_completed(void)
939 return rcu_batches_completed_sched();
941 EXPORT_SYMBOL_GPL(rcu_batches_completed);
944 * Because preemptible RCU does not exist, we never have to check for
945 * CPUs being in quiescent states.
947 static void rcu_preempt_note_context_switch(int cpu)
952 * Because preemptible RCU does not exist, there are never any preempted
955 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
960 #ifdef CONFIG_HOTPLUG_CPU
962 /* Because preemptible RCU does not exist, no quieting of tasks. */
963 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
964 __releases(rnp->lock)
966 raw_spin_unlock_irqrestore(&rnp->lock, flags);
969 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
972 * Because preemptible RCU does not exist, we never have to check for
973 * tasks blocked within RCU read-side critical sections.
975 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
980 * Because preemptible RCU does not exist, we never have to check for
981 * tasks blocked within RCU read-side critical sections.
983 static int rcu_print_task_stall(struct rcu_node *rnp)
989 * Because there is no preemptible RCU, there can be no readers blocked,
990 * so there is no need to check for blocked tasks. So check only for
991 * bogus qsmask values.
993 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
995 WARN_ON_ONCE(rnp->qsmask);
998 #ifdef CONFIG_HOTPLUG_CPU
1001 * Because preemptible RCU does not exist, it never needs to migrate
1002 * tasks that were blocked within RCU read-side critical sections, and
1003 * such non-existent tasks cannot possibly have been blocking the current
1006 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1007 struct rcu_node *rnp,
1008 struct rcu_data *rdp)
1013 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1016 * Because preemptible RCU does not exist, it never has any callbacks
1019 static void rcu_preempt_check_callbacks(int cpu)
1024 * Wait for an rcu-preempt grace period, but make it happen quickly.
1025 * But because preemptible RCU does not exist, map to rcu-sched.
1027 void synchronize_rcu_expedited(void)
1029 synchronize_sched_expedited();
1031 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1033 #ifdef CONFIG_HOTPLUG_CPU
1036 * Because preemptible RCU does not exist, there is never any need to
1037 * report on tasks preempted in RCU read-side critical sections during
1038 * expedited RCU grace periods.
1040 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1045 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1048 * Because preemptible RCU does not exist, rcu_barrier() is just
1049 * another name for rcu_barrier_sched().
1051 void rcu_barrier(void)
1053 rcu_barrier_sched();
1055 EXPORT_SYMBOL_GPL(rcu_barrier);
1058 * Because preemptible RCU does not exist, it need not be initialized.
1060 static void __init __rcu_init_preempt(void)
1065 * Because preemptible RCU does not exist, tasks cannot possibly exit
1066 * while in preemptible RCU read-side critical sections.
1072 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1074 #ifdef CONFIG_RCU_BOOST
1076 #include "../locking/rtmutex_common.h"
1078 #ifdef CONFIG_RCU_TRACE
1080 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1082 if (list_empty(&rnp->blkd_tasks))
1083 rnp->n_balk_blkd_tasks++;
1084 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1085 rnp->n_balk_exp_gp_tasks++;
1086 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1087 rnp->n_balk_boost_tasks++;
1088 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1089 rnp->n_balk_notblocked++;
1090 else if (rnp->gp_tasks != NULL &&
1091 ULONG_CMP_LT(jiffies, rnp->boost_time))
1092 rnp->n_balk_notyet++;
1097 #else /* #ifdef CONFIG_RCU_TRACE */
1099 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1103 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1105 static void rcu_wake_cond(struct task_struct *t, int status)
1108 * If the thread is yielding, only wake it when this
1109 * is invoked from idle
1111 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1116 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1117 * or ->boost_tasks, advancing the pointer to the next task in the
1118 * ->blkd_tasks list.
1120 * Note that irqs must be enabled: boosting the task can block.
1121 * Returns 1 if there are more tasks needing to be boosted.
1123 static int rcu_boost(struct rcu_node *rnp)
1125 unsigned long flags;
1126 struct task_struct *t;
1127 struct list_head *tb;
1129 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1130 return 0; /* Nothing left to boost. */
1132 raw_spin_lock_irqsave(&rnp->lock, flags);
1133 smp_mb__after_unlock_lock();
1136 * Recheck under the lock: all tasks in need of boosting
1137 * might exit their RCU read-side critical sections on their own.
1139 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1140 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1145 * Preferentially boost tasks blocking expedited grace periods.
1146 * This cannot starve the normal grace periods because a second
1147 * expedited grace period must boost all blocked tasks, including
1148 * those blocking the pre-existing normal grace period.
1150 if (rnp->exp_tasks != NULL) {
1151 tb = rnp->exp_tasks;
1152 rnp->n_exp_boosts++;
1154 tb = rnp->boost_tasks;
1155 rnp->n_normal_boosts++;
1157 rnp->n_tasks_boosted++;
1160 * We boost task t by manufacturing an rt_mutex that appears to
1161 * be held by task t. We leave a pointer to that rt_mutex where
1162 * task t can find it, and task t will release the mutex when it
1163 * exits its outermost RCU read-side critical section. Then
1164 * simply acquiring this artificial rt_mutex will boost task
1165 * t's priority. (Thanks to tglx for suggesting this approach!)
1167 * Note that task t must acquire rnp->lock to remove itself from
1168 * the ->blkd_tasks list, which it will do from exit() if from
1169 * nowhere else. We therefore are guaranteed that task t will
1170 * stay around at least until we drop rnp->lock. Note that
1171 * rnp->lock also resolves races between our priority boosting
1172 * and task t's exiting its outermost RCU read-side critical
1175 t = container_of(tb, struct task_struct, rcu_node_entry);
1176 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1177 init_completion(&rnp->boost_completion);
1178 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1179 /* Lock only for side effect: boosts task t's priority. */
1180 rt_mutex_lock(&rnp->boost_mtx);
1181 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1183 /* Wait for boostee to be done w/boost_mtx before reinitializing. */
1184 wait_for_completion(&rnp->boost_completion);
1186 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1187 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1191 * Priority-boosting kthread. One per leaf rcu_node and one for the
1194 static int rcu_boost_kthread(void *arg)
1196 struct rcu_node *rnp = (struct rcu_node *)arg;
1200 trace_rcu_utilization(TPS("Start boost kthread@init"));
1202 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1203 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1204 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1205 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1206 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1207 more2boost = rcu_boost(rnp);
1213 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1214 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1215 schedule_timeout_interruptible(2);
1216 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1221 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1226 * Check to see if it is time to start boosting RCU readers that are
1227 * blocking the current grace period, and, if so, tell the per-rcu_node
1228 * kthread to start boosting them. If there is an expedited grace
1229 * period in progress, it is always time to boost.
1231 * The caller must hold rnp->lock, which this function releases.
1232 * The ->boost_kthread_task is immortal, so we don't need to worry
1233 * about it going away.
1235 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1236 __releases(rnp->lock)
1238 struct task_struct *t;
1240 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1241 rnp->n_balk_exp_gp_tasks++;
1242 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1245 if (rnp->exp_tasks != NULL ||
1246 (rnp->gp_tasks != NULL &&
1247 rnp->boost_tasks == NULL &&
1249 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1250 if (rnp->exp_tasks == NULL)
1251 rnp->boost_tasks = rnp->gp_tasks;
1252 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1253 t = rnp->boost_kthread_task;
1255 rcu_wake_cond(t, rnp->boost_kthread_status);
1257 rcu_initiate_boost_trace(rnp);
1258 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1263 * Wake up the per-CPU kthread to invoke RCU callbacks.
1265 static void invoke_rcu_callbacks_kthread(void)
1267 unsigned long flags;
1269 local_irq_save(flags);
1270 __this_cpu_write(rcu_cpu_has_work, 1);
1271 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1272 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1273 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1274 __this_cpu_read(rcu_cpu_kthread_status));
1276 local_irq_restore(flags);
1280 * Is the current CPU running the RCU-callbacks kthread?
1281 * Caller must have preemption disabled.
1283 static bool rcu_is_callbacks_kthread(void)
1285 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1288 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1291 * Do priority-boost accounting for the start of a new grace period.
1293 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1295 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1299 * Create an RCU-boost kthread for the specified node if one does not
1300 * already exist. We only create this kthread for preemptible RCU.
1301 * Returns zero if all is well, a negated errno otherwise.
1303 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1304 struct rcu_node *rnp)
1306 int rnp_index = rnp - &rsp->node[0];
1307 unsigned long flags;
1308 struct sched_param sp;
1309 struct task_struct *t;
1311 if (&rcu_preempt_state != rsp)
1314 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1318 if (rnp->boost_kthread_task != NULL)
1320 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1321 "rcub/%d", rnp_index);
1324 raw_spin_lock_irqsave(&rnp->lock, flags);
1325 smp_mb__after_unlock_lock();
1326 rnp->boost_kthread_task = t;
1327 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1328 sp.sched_priority = RCU_BOOST_PRIO;
1329 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1330 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1334 static void rcu_kthread_do_work(void)
1336 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1337 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1338 rcu_preempt_do_callbacks();
1341 static void rcu_cpu_kthread_setup(unsigned int cpu)
1343 struct sched_param sp;
1345 sp.sched_priority = RCU_KTHREAD_PRIO;
1346 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1349 static void rcu_cpu_kthread_park(unsigned int cpu)
1351 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1354 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1356 return __this_cpu_read(rcu_cpu_has_work);
1360 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1361 * RCU softirq used in flavors and configurations of RCU that do not
1362 * support RCU priority boosting.
1364 static void rcu_cpu_kthread(unsigned int cpu)
1366 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1367 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1370 for (spincnt = 0; spincnt < 10; spincnt++) {
1371 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1373 *statusp = RCU_KTHREAD_RUNNING;
1374 this_cpu_inc(rcu_cpu_kthread_loops);
1375 local_irq_disable();
1380 rcu_kthread_do_work();
1383 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1384 *statusp = RCU_KTHREAD_WAITING;
1388 *statusp = RCU_KTHREAD_YIELDING;
1389 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1390 schedule_timeout_interruptible(2);
1391 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1392 *statusp = RCU_KTHREAD_WAITING;
1396 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1397 * served by the rcu_node in question. The CPU hotplug lock is still
1398 * held, so the value of rnp->qsmaskinit will be stable.
1400 * We don't include outgoingcpu in the affinity set, use -1 if there is
1401 * no outgoing CPU. If there are no CPUs left in the affinity set,
1402 * this function allows the kthread to execute on any CPU.
1404 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1406 struct task_struct *t = rnp->boost_kthread_task;
1407 unsigned long mask = rnp->qsmaskinit;
1413 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1415 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1416 if ((mask & 0x1) && cpu != outgoingcpu)
1417 cpumask_set_cpu(cpu, cm);
1418 if (cpumask_weight(cm) == 0) {
1420 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1421 cpumask_clear_cpu(cpu, cm);
1422 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1424 set_cpus_allowed_ptr(t, cm);
1425 free_cpumask_var(cm);
1428 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1429 .store = &rcu_cpu_kthread_task,
1430 .thread_should_run = rcu_cpu_kthread_should_run,
1431 .thread_fn = rcu_cpu_kthread,
1432 .thread_comm = "rcuc/%u",
1433 .setup = rcu_cpu_kthread_setup,
1434 .park = rcu_cpu_kthread_park,
1438 * Spawn boost kthreads -- called as soon as the scheduler is running.
1440 static void __init rcu_spawn_boost_kthreads(void)
1442 struct rcu_node *rnp;
1445 for_each_possible_cpu(cpu)
1446 per_cpu(rcu_cpu_has_work, cpu) = 0;
1447 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1448 rnp = rcu_get_root(rcu_state_p);
1449 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1450 if (NUM_RCU_NODES > 1) {
1451 rcu_for_each_leaf_node(rcu_state_p, rnp)
1452 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1456 static void rcu_prepare_kthreads(int cpu)
1458 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1459 struct rcu_node *rnp = rdp->mynode;
1461 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1462 if (rcu_scheduler_fully_active)
1463 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1466 #else /* #ifdef CONFIG_RCU_BOOST */
1468 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1469 __releases(rnp->lock)
1471 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1474 static void invoke_rcu_callbacks_kthread(void)
1479 static bool rcu_is_callbacks_kthread(void)
1484 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1488 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1492 static void __init rcu_spawn_boost_kthreads(void)
1496 static void rcu_prepare_kthreads(int cpu)
1500 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1502 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1505 * Check to see if any future RCU-related work will need to be done
1506 * by the current CPU, even if none need be done immediately, returning
1507 * 1 if so. This function is part of the RCU implementation; it is -not-
1508 * an exported member of the RCU API.
1510 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1511 * any flavor of RCU.
1513 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1514 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1516 *delta_jiffies = ULONG_MAX;
1517 return rcu_cpu_has_callbacks(cpu, NULL);
1519 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1522 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1525 static void rcu_cleanup_after_idle(int cpu)
1530 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1533 static void rcu_prepare_for_idle(int cpu)
1538 * Don't bother keeping a running count of the number of RCU callbacks
1539 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1541 static void rcu_idle_count_callbacks_posted(void)
1545 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1548 * This code is invoked when a CPU goes idle, at which point we want
1549 * to have the CPU do everything required for RCU so that it can enter
1550 * the energy-efficient dyntick-idle mode. This is handled by a
1551 * state machine implemented by rcu_prepare_for_idle() below.
1553 * The following three proprocessor symbols control this state machine:
1555 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1556 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1557 * is sized to be roughly one RCU grace period. Those energy-efficiency
1558 * benchmarkers who might otherwise be tempted to set this to a large
1559 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1560 * system. And if you are -that- concerned about energy efficiency,
1561 * just power the system down and be done with it!
1562 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1563 * permitted to sleep in dyntick-idle mode with only lazy RCU
1564 * callbacks pending. Setting this too high can OOM your system.
1566 * The values below work well in practice. If future workloads require
1567 * adjustment, they can be converted into kernel config parameters, though
1568 * making the state machine smarter might be a better option.
1570 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1571 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1573 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1574 module_param(rcu_idle_gp_delay, int, 0644);
1575 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1576 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1578 extern int tick_nohz_active;
1581 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1582 * only if it has been awhile since the last time we did so. Afterwards,
1583 * if there are any callbacks ready for immediate invocation, return true.
1585 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1587 bool cbs_ready = false;
1588 struct rcu_data *rdp;
1589 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1590 struct rcu_node *rnp;
1591 struct rcu_state *rsp;
1593 /* Exit early if we advanced recently. */
1594 if (jiffies == rdtp->last_advance_all)
1596 rdtp->last_advance_all = jiffies;
1598 for_each_rcu_flavor(rsp) {
1599 rdp = this_cpu_ptr(rsp->rda);
1603 * Don't bother checking unless a grace period has
1604 * completed since we last checked and there are
1605 * callbacks not yet ready to invoke.
1607 if (rdp->completed != rnp->completed &&
1608 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1609 note_gp_changes(rsp, rdp);
1611 if (cpu_has_callbacks_ready_to_invoke(rdp))
1618 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1619 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1620 * caller to set the timeout based on whether or not there are non-lazy
1623 * The caller must have disabled interrupts.
1625 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1626 int rcu_needs_cpu(int cpu, unsigned long *dj)
1628 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1630 /* Snapshot to detect later posting of non-lazy callback. */
1631 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1633 /* If no callbacks, RCU doesn't need the CPU. */
1634 if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1639 /* Attempt to advance callbacks. */
1640 if (rcu_try_advance_all_cbs()) {
1641 /* Some ready to invoke, so initiate later invocation. */
1645 rdtp->last_accelerate = jiffies;
1647 /* Request timer delay depending on laziness, and round. */
1648 if (!rdtp->all_lazy) {
1649 *dj = round_up(rcu_idle_gp_delay + jiffies,
1650 rcu_idle_gp_delay) - jiffies;
1652 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1656 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1659 * Prepare a CPU for idle from an RCU perspective. The first major task
1660 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1661 * The second major task is to check to see if a non-lazy callback has
1662 * arrived at a CPU that previously had only lazy callbacks. The third
1663 * major task is to accelerate (that is, assign grace-period numbers to)
1664 * any recently arrived callbacks.
1666 * The caller must have disabled interrupts.
1668 static void rcu_prepare_for_idle(int cpu)
1670 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1672 struct rcu_data *rdp;
1673 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1674 struct rcu_node *rnp;
1675 struct rcu_state *rsp;
1678 /* Handle nohz enablement switches conservatively. */
1679 tne = ACCESS_ONCE(tick_nohz_active);
1680 if (tne != rdtp->tick_nohz_enabled_snap) {
1681 if (rcu_cpu_has_callbacks(cpu, NULL))
1682 invoke_rcu_core(); /* force nohz to see update. */
1683 rdtp->tick_nohz_enabled_snap = tne;
1689 /* If this is a no-CBs CPU, no callbacks, just return. */
1690 if (rcu_is_nocb_cpu(cpu))
1694 * If a non-lazy callback arrived at a CPU having only lazy
1695 * callbacks, invoke RCU core for the side-effect of recalculating
1696 * idle duration on re-entry to idle.
1698 if (rdtp->all_lazy &&
1699 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1700 rdtp->all_lazy = false;
1701 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1707 * If we have not yet accelerated this jiffy, accelerate all
1708 * callbacks on this CPU.
1710 if (rdtp->last_accelerate == jiffies)
1712 rdtp->last_accelerate = jiffies;
1713 for_each_rcu_flavor(rsp) {
1714 rdp = per_cpu_ptr(rsp->rda, cpu);
1715 if (!*rdp->nxttail[RCU_DONE_TAIL])
1718 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1719 smp_mb__after_unlock_lock();
1720 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1721 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1723 rcu_gp_kthread_wake(rsp);
1725 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1729 * Clean up for exit from idle. Attempt to advance callbacks based on
1730 * any grace periods that elapsed while the CPU was idle, and if any
1731 * callbacks are now ready to invoke, initiate invocation.
1733 static void rcu_cleanup_after_idle(int cpu)
1735 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1736 if (rcu_is_nocb_cpu(cpu))
1738 if (rcu_try_advance_all_cbs())
1740 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1744 * Keep a running count of the number of non-lazy callbacks posted
1745 * on this CPU. This running counter (which is never decremented) allows
1746 * rcu_prepare_for_idle() to detect when something out of the idle loop
1747 * posts a callback, even if an equal number of callbacks are invoked.
1748 * Of course, callbacks should only be posted from within a trace event
1749 * designed to be called from idle or from within RCU_NONIDLE().
1751 static void rcu_idle_count_callbacks_posted(void)
1753 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1757 * Data for flushing lazy RCU callbacks at OOM time.
1759 static atomic_t oom_callback_count;
1760 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1763 * RCU OOM callback -- decrement the outstanding count and deliver the
1764 * wake-up if we are the last one.
1766 static void rcu_oom_callback(struct rcu_head *rhp)
1768 if (atomic_dec_and_test(&oom_callback_count))
1769 wake_up(&oom_callback_wq);
1773 * Post an rcu_oom_notify callback on the current CPU if it has at
1774 * least one lazy callback. This will unnecessarily post callbacks
1775 * to CPUs that already have a non-lazy callback at the end of their
1776 * callback list, but this is an infrequent operation, so accept some
1777 * extra overhead to keep things simple.
1779 static void rcu_oom_notify_cpu(void *unused)
1781 struct rcu_state *rsp;
1782 struct rcu_data *rdp;
1784 for_each_rcu_flavor(rsp) {
1785 rdp = raw_cpu_ptr(rsp->rda);
1786 if (rdp->qlen_lazy != 0) {
1787 atomic_inc(&oom_callback_count);
1788 rsp->call(&rdp->oom_head, rcu_oom_callback);
1794 * If low on memory, ensure that each CPU has a non-lazy callback.
1795 * This will wake up CPUs that have only lazy callbacks, in turn
1796 * ensuring that they free up the corresponding memory in a timely manner.
1797 * Because an uncertain amount of memory will be freed in some uncertain
1798 * timeframe, we do not claim to have freed anything.
1800 static int rcu_oom_notify(struct notifier_block *self,
1801 unsigned long notused, void *nfreed)
1805 /* Wait for callbacks from earlier instance to complete. */
1806 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1807 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1810 * Prevent premature wakeup: ensure that all increments happen
1811 * before there is a chance of the counter reaching zero.
1813 atomic_set(&oom_callback_count, 1);
1816 for_each_online_cpu(cpu) {
1817 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1822 /* Unconditionally decrement: no need to wake ourselves up. */
1823 atomic_dec(&oom_callback_count);
1828 static struct notifier_block rcu_oom_nb = {
1829 .notifier_call = rcu_oom_notify
1832 static int __init rcu_register_oom_notifier(void)
1834 register_oom_notifier(&rcu_oom_nb);
1837 early_initcall(rcu_register_oom_notifier);
1839 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1841 #ifdef CONFIG_RCU_CPU_STALL_INFO
1843 #ifdef CONFIG_RCU_FAST_NO_HZ
1845 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1847 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1848 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1850 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1851 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1853 rdtp->all_lazy ? 'L' : '.',
1854 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1857 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1859 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1864 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1866 /* Initiate the stall-info list. */
1867 static void print_cpu_stall_info_begin(void)
1873 * Print out diagnostic information for the specified stalled CPU.
1875 * If the specified CPU is aware of the current RCU grace period
1876 * (flavor specified by rsp), then print the number of scheduling
1877 * clock interrupts the CPU has taken during the time that it has
1878 * been aware. Otherwise, print the number of RCU grace periods
1879 * that this CPU is ignorant of, for example, "1" if the CPU was
1880 * aware of the previous grace period.
1882 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1884 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1886 char fast_no_hz[72];
1887 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1888 struct rcu_dynticks *rdtp = rdp->dynticks;
1890 unsigned long ticks_value;
1892 if (rsp->gpnum == rdp->gpnum) {
1893 ticks_title = "ticks this GP";
1894 ticks_value = rdp->ticks_this_gp;
1896 ticks_title = "GPs behind";
1897 ticks_value = rsp->gpnum - rdp->gpnum;
1899 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1900 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1901 cpu, ticks_value, ticks_title,
1902 atomic_read(&rdtp->dynticks) & 0xfff,
1903 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1904 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1908 /* Terminate the stall-info list. */
1909 static void print_cpu_stall_info_end(void)
1914 /* Zero ->ticks_this_gp for all flavors of RCU. */
1915 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1917 rdp->ticks_this_gp = 0;
1918 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1921 /* Increment ->ticks_this_gp for all flavors of RCU. */
1922 static void increment_cpu_stall_ticks(void)
1924 struct rcu_state *rsp;
1926 for_each_rcu_flavor(rsp)
1927 raw_cpu_inc(rsp->rda->ticks_this_gp);
1930 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1932 static void print_cpu_stall_info_begin(void)
1937 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1939 pr_cont(" %d", cpu);
1942 static void print_cpu_stall_info_end(void)
1947 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1951 static void increment_cpu_stall_ticks(void)
1955 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1957 #ifdef CONFIG_RCU_NOCB_CPU
1960 * Offload callback processing from the boot-time-specified set of CPUs
1961 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1962 * kthread created that pulls the callbacks from the corresponding CPU,
1963 * waits for a grace period to elapse, and invokes the callbacks.
1964 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1965 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1966 * has been specified, in which case each kthread actively polls its
1967 * CPU. (Which isn't so great for energy efficiency, but which does
1968 * reduce RCU's overhead on that CPU.)
1970 * This is intended to be used in conjunction with Frederic Weisbecker's
1971 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1972 * running CPU-bound user-mode computations.
1974 * Offloading of callback processing could also in theory be used as
1975 * an energy-efficiency measure because CPUs with no RCU callbacks
1976 * queued are more aggressive about entering dyntick-idle mode.
1980 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1981 static int __init rcu_nocb_setup(char *str)
1983 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1984 have_rcu_nocb_mask = true;
1985 cpulist_parse(str, rcu_nocb_mask);
1988 __setup("rcu_nocbs=", rcu_nocb_setup);
1990 static int __init parse_rcu_nocb_poll(char *arg)
1995 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1998 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2001 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2003 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2007 * Set the root rcu_node structure's ->need_future_gp field
2008 * based on the sum of those of all rcu_node structures. This does
2009 * double-count the root rcu_node structure's requests, but this
2010 * is necessary to handle the possibility of a rcu_nocb_kthread()
2011 * having awakened during the time that the rcu_node structures
2012 * were being updated for the end of the previous grace period.
2014 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2016 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2019 static void rcu_init_one_nocb(struct rcu_node *rnp)
2021 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2022 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2025 #ifndef CONFIG_RCU_NOCB_CPU_ALL
2026 /* Is the specified CPU a no-CBs CPU? */
2027 bool rcu_is_nocb_cpu(int cpu)
2029 if (have_rcu_nocb_mask)
2030 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2033 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
2036 * Kick the leader kthread for this NOCB group.
2038 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
2040 struct rcu_data *rdp_leader = rdp->nocb_leader;
2042 if (!ACCESS_ONCE(rdp_leader->nocb_kthread))
2044 if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) {
2045 /* Prior xchg orders against prior callback enqueue. */
2046 ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false;
2047 wake_up(&rdp_leader->nocb_wq);
2052 * Enqueue the specified string of rcu_head structures onto the specified
2053 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2054 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2055 * counts are supplied by rhcount and rhcount_lazy.
2057 * If warranted, also wake up the kthread servicing this CPUs queues.
2059 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2060 struct rcu_head *rhp,
2061 struct rcu_head **rhtp,
2062 int rhcount, int rhcount_lazy,
2063 unsigned long flags)
2066 struct rcu_head **old_rhpp;
2067 struct task_struct *t;
2069 /* Enqueue the callback on the nocb list and update counts. */
2070 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2071 ACCESS_ONCE(*old_rhpp) = rhp;
2072 atomic_long_add(rhcount, &rdp->nocb_q_count);
2073 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2075 /* If we are not being polled and there is a kthread, awaken it ... */
2076 t = ACCESS_ONCE(rdp->nocb_kthread);
2077 if (rcu_nocb_poll || !t) {
2078 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2079 TPS("WakeNotPoll"));
2082 len = atomic_long_read(&rdp->nocb_q_count);
2083 if (old_rhpp == &rdp->nocb_head) {
2084 if (!irqs_disabled_flags(flags)) {
2085 /* ... if queue was empty ... */
2086 wake_nocb_leader(rdp, false);
2087 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2090 rdp->nocb_defer_wakeup = true;
2091 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2092 TPS("WakeEmptyIsDeferred"));
2094 rdp->qlen_last_fqs_check = 0;
2095 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2096 /* ... or if many callbacks queued. */
2097 wake_nocb_leader(rdp, true);
2098 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2099 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf"));
2101 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2107 * This is a helper for __call_rcu(), which invokes this when the normal
2108 * callback queue is inoperable. If this is not a no-CBs CPU, this
2109 * function returns failure back to __call_rcu(), which can complain
2112 * Otherwise, this function queues the callback where the corresponding
2113 * "rcuo" kthread can find it.
2115 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2116 bool lazy, unsigned long flags)
2119 if (!rcu_is_nocb_cpu(rdp->cpu))
2121 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2122 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2123 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2124 (unsigned long)rhp->func,
2125 -atomic_long_read(&rdp->nocb_q_count_lazy),
2126 -atomic_long_read(&rdp->nocb_q_count));
2128 trace_rcu_callback(rdp->rsp->name, rhp,
2129 -atomic_long_read(&rdp->nocb_q_count_lazy),
2130 -atomic_long_read(&rdp->nocb_q_count));
2135 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2138 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2139 struct rcu_data *rdp,
2140 unsigned long flags)
2142 long ql = rsp->qlen;
2143 long qll = rsp->qlen_lazy;
2145 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2146 if (!rcu_is_nocb_cpu(smp_processor_id()))
2151 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2152 if (rsp->orphan_donelist != NULL) {
2153 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2154 rsp->orphan_donetail, ql, qll, flags);
2156 rsp->orphan_donelist = NULL;
2157 rsp->orphan_donetail = &rsp->orphan_donelist;
2159 if (rsp->orphan_nxtlist != NULL) {
2160 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2161 rsp->orphan_nxttail, ql, qll, flags);
2163 rsp->orphan_nxtlist = NULL;
2164 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2170 * If necessary, kick off a new grace period, and either way wait
2171 * for a subsequent grace period to complete.
2173 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2177 unsigned long flags;
2179 struct rcu_node *rnp = rdp->mynode;
2181 raw_spin_lock_irqsave(&rnp->lock, flags);
2182 smp_mb__after_unlock_lock();
2183 needwake = rcu_start_future_gp(rnp, rdp, &c);
2184 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2186 rcu_gp_kthread_wake(rdp->rsp);
2189 * Wait for the grace period. Do so interruptibly to avoid messing
2190 * up the load average.
2192 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2194 wait_event_interruptible(
2195 rnp->nocb_gp_wq[c & 0x1],
2196 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2199 flush_signals(current);
2200 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2202 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2203 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2207 * Leaders come here to wait for additional callbacks to show up.
2208 * This function does not return until callbacks appear.
2210 static void nocb_leader_wait(struct rcu_data *my_rdp)
2212 bool firsttime = true;
2214 struct rcu_data *rdp;
2215 struct rcu_head **tail;
2219 /* Wait for callbacks to appear. */
2220 if (!rcu_nocb_poll) {
2221 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2222 wait_event_interruptible(my_rdp->nocb_wq,
2223 !ACCESS_ONCE(my_rdp->nocb_leader_sleep));
2224 /* Memory barrier handled by smp_mb() calls below and repoll. */
2225 } else if (firsttime) {
2226 firsttime = false; /* Don't drown trace log with "Poll"! */
2227 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2231 * Each pass through the following loop checks a follower for CBs.
2232 * We are our own first follower. Any CBs found are moved to
2233 * nocb_gp_head, where they await a grace period.
2236 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2237 rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head);
2238 if (!rdp->nocb_gp_head)
2239 continue; /* No CBs here, try next follower. */
2241 /* Move callbacks to wait-for-GP list, which is empty. */
2242 ACCESS_ONCE(rdp->nocb_head) = NULL;
2243 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2244 rdp->nocb_gp_count = atomic_long_xchg(&rdp->nocb_q_count, 0);
2245 rdp->nocb_gp_count_lazy =
2246 atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2251 * If there were no callbacks, sleep a bit, rescan after a
2252 * memory barrier, and go retry.
2254 if (unlikely(!gotcbs)) {
2256 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2258 flush_signals(current);
2259 schedule_timeout_interruptible(1);
2261 /* Rescan in case we were a victim of memory ordering. */
2262 my_rdp->nocb_leader_sleep = true;
2263 smp_mb(); /* Ensure _sleep true before scan. */
2264 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2265 if (ACCESS_ONCE(rdp->nocb_head)) {
2266 /* Found CB, so short-circuit next wait. */
2267 my_rdp->nocb_leader_sleep = false;
2273 /* Wait for one grace period. */
2274 rcu_nocb_wait_gp(my_rdp);
2277 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2278 * We set it now, but recheck for new callbacks while
2279 * traversing our follower list.
2281 my_rdp->nocb_leader_sleep = true;
2282 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2284 /* Each pass through the following loop wakes a follower, if needed. */
2285 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2286 if (ACCESS_ONCE(rdp->nocb_head))
2287 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2288 if (!rdp->nocb_gp_head)
2289 continue; /* No CBs, so no need to wake follower. */
2291 /* Append callbacks to follower's "done" list. */
2292 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2293 *tail = rdp->nocb_gp_head;
2294 atomic_long_add(rdp->nocb_gp_count, &rdp->nocb_follower_count);
2295 atomic_long_add(rdp->nocb_gp_count_lazy,
2296 &rdp->nocb_follower_count_lazy);
2297 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2299 * List was empty, wake up the follower.
2300 * Memory barriers supplied by atomic_long_add().
2302 wake_up(&rdp->nocb_wq);
2306 /* If we (the leader) don't have CBs, go wait some more. */
2307 if (!my_rdp->nocb_follower_head)
2312 * Followers come here to wait for additional callbacks to show up.
2313 * This function does not return until callbacks appear.
2315 static void nocb_follower_wait(struct rcu_data *rdp)
2317 bool firsttime = true;
2320 if (!rcu_nocb_poll) {
2321 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2323 wait_event_interruptible(rdp->nocb_wq,
2324 ACCESS_ONCE(rdp->nocb_follower_head));
2325 } else if (firsttime) {
2326 /* Don't drown trace log with "Poll"! */
2328 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2330 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2331 /* ^^^ Ensure CB invocation follows _head test. */
2335 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2337 flush_signals(current);
2338 schedule_timeout_interruptible(1);
2343 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2344 * callbacks queued by the corresponding no-CBs CPU, however, there is
2345 * an optional leader-follower relationship so that the grace-period
2346 * kthreads don't have to do quite so many wakeups.
2348 static int rcu_nocb_kthread(void *arg)
2351 struct rcu_head *list;
2352 struct rcu_head *next;
2353 struct rcu_head **tail;
2354 struct rcu_data *rdp = arg;
2356 /* Each pass through this loop invokes one batch of callbacks */
2358 /* Wait for callbacks. */
2359 if (rdp->nocb_leader == rdp)
2360 nocb_leader_wait(rdp);
2362 nocb_follower_wait(rdp);
2364 /* Pull the ready-to-invoke callbacks onto local list. */
2365 list = ACCESS_ONCE(rdp->nocb_follower_head);
2367 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2368 ACCESS_ONCE(rdp->nocb_follower_head) = NULL;
2369 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2370 c = atomic_long_xchg(&rdp->nocb_follower_count, 0);
2371 cl = atomic_long_xchg(&rdp->nocb_follower_count_lazy, 0);
2372 rdp->nocb_p_count += c;
2373 rdp->nocb_p_count_lazy += cl;
2375 /* Each pass through the following loop invokes a callback. */
2376 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2380 /* Wait for enqueuing to complete, if needed. */
2381 while (next == NULL && &list->next != tail) {
2382 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2384 schedule_timeout_interruptible(1);
2385 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2389 debug_rcu_head_unqueue(list);
2391 if (__rcu_reclaim(rdp->rsp->name, list))
2397 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2398 ACCESS_ONCE(rdp->nocb_p_count) -= c;
2399 ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2400 rdp->n_nocbs_invoked += c;
2405 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2406 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2408 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2411 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2412 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2414 if (!rcu_nocb_need_deferred_wakeup(rdp))
2416 ACCESS_ONCE(rdp->nocb_defer_wakeup) = false;
2417 wake_nocb_leader(rdp, false);
2418 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWakeEmpty"));
2421 void __init rcu_init_nohz(void)
2424 bool need_rcu_nocb_mask = true;
2425 struct rcu_state *rsp;
2427 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2428 need_rcu_nocb_mask = false;
2429 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2431 #if defined(CONFIG_NO_HZ_FULL)
2432 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2433 need_rcu_nocb_mask = true;
2434 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2436 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2437 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2438 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2441 have_rcu_nocb_mask = true;
2443 if (!have_rcu_nocb_mask)
2446 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2447 pr_info("\tOffload RCU callbacks from CPU 0\n");
2448 cpumask_set_cpu(0, rcu_nocb_mask);
2449 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2450 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2451 pr_info("\tOffload RCU callbacks from all CPUs\n");
2452 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2453 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2454 #if defined(CONFIG_NO_HZ_FULL)
2455 if (tick_nohz_full_running)
2456 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2457 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2459 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2460 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2461 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2464 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
2465 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
2467 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2469 for_each_rcu_flavor(rsp) {
2470 for_each_cpu(cpu, rcu_nocb_mask) {
2471 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2474 * If there are early callbacks, they will need
2475 * to be moved to the nocb lists.
2477 WARN_ON_ONCE(rdp->nxttail[RCU_NEXT_TAIL] !=
2479 rdp->nxttail[RCU_NEXT_TAIL] != NULL);
2480 init_nocb_callback_list(rdp);
2485 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2486 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2488 rdp->nocb_tail = &rdp->nocb_head;
2489 init_waitqueue_head(&rdp->nocb_wq);
2490 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2493 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2494 static int rcu_nocb_leader_stride = -1;
2495 module_param(rcu_nocb_leader_stride, int, 0444);
2498 * Create a kthread for each RCU flavor for each no-CBs CPU.
2499 * Also initialize leader-follower relationships.
2501 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2504 int ls = rcu_nocb_leader_stride;
2505 int nl = 0; /* Next leader. */
2506 struct rcu_data *rdp;
2507 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2508 struct rcu_data *rdp_prev = NULL;
2509 struct task_struct *t;
2511 if (rcu_nocb_mask == NULL)
2514 ls = int_sqrt(nr_cpu_ids);
2515 rcu_nocb_leader_stride = ls;
2519 * Each pass through this loop sets up one rcu_data structure and
2520 * spawns one rcu_nocb_kthread().
2522 for_each_cpu(cpu, rcu_nocb_mask) {
2523 rdp = per_cpu_ptr(rsp->rda, cpu);
2524 if (rdp->cpu >= nl) {
2525 /* New leader, set up for followers & next leader. */
2526 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2527 rdp->nocb_leader = rdp;
2530 /* Another follower, link to previous leader. */
2531 rdp->nocb_leader = rdp_leader;
2532 rdp_prev->nocb_next_follower = rdp;
2536 /* Spawn the kthread for this CPU. */
2537 t = kthread_run(rcu_nocb_kthread, rdp,
2538 "rcuo%c/%d", rsp->abbr, cpu);
2540 ACCESS_ONCE(rdp->nocb_kthread) = t;
2544 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2545 static bool init_nocb_callback_list(struct rcu_data *rdp)
2547 if (rcu_nocb_mask == NULL ||
2548 !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2550 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2554 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2556 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2560 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2564 static void rcu_init_one_nocb(struct rcu_node *rnp)
2568 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2569 bool lazy, unsigned long flags)
2574 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2575 struct rcu_data *rdp,
2576 unsigned long flags)
2581 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2585 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2590 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2594 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2598 static bool init_nocb_callback_list(struct rcu_data *rdp)
2603 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2606 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2607 * arbitrarily long period of time with the scheduling-clock tick turned
2608 * off. RCU will be paying attention to this CPU because it is in the
2609 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2610 * machine because the scheduling-clock tick has been disabled. Therefore,
2611 * if an adaptive-ticks CPU is failing to respond to the current grace
2612 * period and has not be idle from an RCU perspective, kick it.
2614 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2616 #ifdef CONFIG_NO_HZ_FULL
2617 if (tick_nohz_full_cpu(cpu))
2618 smp_send_reschedule(cpu);
2619 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2623 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2626 * Define RCU flavor that holds sysidle state. This needs to be the
2627 * most active flavor of RCU.
2629 #ifdef CONFIG_PREEMPT_RCU
2630 static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state;
2631 #else /* #ifdef CONFIG_PREEMPT_RCU */
2632 static struct rcu_state *rcu_sysidle_state = &rcu_sched_state;
2633 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2635 static int full_sysidle_state; /* Current system-idle state. */
2636 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2637 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2638 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2639 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2640 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2643 * Invoked to note exit from irq or task transition to idle. Note that
2644 * usermode execution does -not- count as idle here! After all, we want
2645 * to detect full-system idle states, not RCU quiescent states and grace
2646 * periods. The caller must have disabled interrupts.
2648 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2652 /* Adjust nesting, check for fully idle. */
2654 rdtp->dynticks_idle_nesting--;
2655 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2656 if (rdtp->dynticks_idle_nesting != 0)
2657 return; /* Still not fully idle. */
2659 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2660 DYNTICK_TASK_NEST_VALUE) {
2661 rdtp->dynticks_idle_nesting = 0;
2663 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2664 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2665 return; /* Still not fully idle. */
2669 /* Record start of fully idle period. */
2671 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2672 smp_mb__before_atomic();
2673 atomic_inc(&rdtp->dynticks_idle);
2674 smp_mb__after_atomic();
2675 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2679 * Unconditionally force exit from full system-idle state. This is
2680 * invoked when a normal CPU exits idle, but must be called separately
2681 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2682 * is that the timekeeping CPU is permitted to take scheduling-clock
2683 * interrupts while the system is in system-idle state, and of course
2684 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2685 * interrupt from any other type of interrupt.
2687 void rcu_sysidle_force_exit(void)
2689 int oldstate = ACCESS_ONCE(full_sysidle_state);
2693 * Each pass through the following loop attempts to exit full
2694 * system-idle state. If contention proves to be a problem,
2695 * a trylock-based contention tree could be used here.
2697 while (oldstate > RCU_SYSIDLE_SHORT) {
2698 newoldstate = cmpxchg(&full_sysidle_state,
2699 oldstate, RCU_SYSIDLE_NOT);
2700 if (oldstate == newoldstate &&
2701 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2702 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2703 return; /* We cleared it, done! */
2705 oldstate = newoldstate;
2707 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2711 * Invoked to note entry to irq or task transition from idle. Note that
2712 * usermode execution does -not- count as idle here! The caller must
2713 * have disabled interrupts.
2715 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2717 /* Adjust nesting, check for already non-idle. */
2719 rdtp->dynticks_idle_nesting++;
2720 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2721 if (rdtp->dynticks_idle_nesting != 1)
2722 return; /* Already non-idle. */
2725 * Allow for irq misnesting. Yes, it really is possible
2726 * to enter an irq handler then never leave it, and maybe
2727 * also vice versa. Handle both possibilities.
2729 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2730 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2731 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2732 return; /* Already non-idle. */
2734 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2738 /* Record end of idle period. */
2739 smp_mb__before_atomic();
2740 atomic_inc(&rdtp->dynticks_idle);
2741 smp_mb__after_atomic();
2742 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2745 * If we are the timekeeping CPU, we are permitted to be non-idle
2746 * during a system-idle state. This must be the case, because
2747 * the timekeeping CPU has to take scheduling-clock interrupts
2748 * during the time that the system is transitioning to full
2749 * system-idle state. This means that the timekeeping CPU must
2750 * invoke rcu_sysidle_force_exit() directly if it does anything
2751 * more than take a scheduling-clock interrupt.
2753 if (smp_processor_id() == tick_do_timer_cpu)
2756 /* Update system-idle state: We are clearly no longer fully idle! */
2757 rcu_sysidle_force_exit();
2761 * Check to see if the current CPU is idle. Note that usermode execution
2762 * does not count as idle. The caller must have disabled interrupts.
2764 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2765 unsigned long *maxj)
2769 struct rcu_dynticks *rdtp = rdp->dynticks;
2772 * If some other CPU has already reported non-idle, if this is
2773 * not the flavor of RCU that tracks sysidle state, or if this
2774 * is an offline or the timekeeping CPU, nothing to do.
2776 if (!*isidle || rdp->rsp != rcu_sysidle_state ||
2777 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2779 if (rcu_gp_in_progress(rdp->rsp))
2780 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2782 /* Pick up current idle and NMI-nesting counter and check. */
2783 cur = atomic_read(&rdtp->dynticks_idle);
2785 *isidle = false; /* We are not idle! */
2788 smp_mb(); /* Read counters before timestamps. */
2790 /* Pick up timestamps. */
2791 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2792 /* If this CPU entered idle more recently, update maxj timestamp. */
2793 if (ULONG_CMP_LT(*maxj, j))
2798 * Is this the flavor of RCU that is handling full-system idle?
2800 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2802 return rsp == rcu_sysidle_state;
2806 * Return a delay in jiffies based on the number of CPUs, rcu_node
2807 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2808 * systems more time to transition to full-idle state in order to
2809 * avoid the cache thrashing that otherwise occur on the state variable.
2810 * Really small systems (less than a couple of tens of CPUs) should
2811 * instead use a single global atomically incremented counter, and later
2812 * versions of this will automatically reconfigure themselves accordingly.
2814 static unsigned long rcu_sysidle_delay(void)
2816 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2818 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2822 * Advance the full-system-idle state. This is invoked when all of
2823 * the non-timekeeping CPUs are idle.
2825 static void rcu_sysidle(unsigned long j)
2827 /* Check the current state. */
2828 switch (ACCESS_ONCE(full_sysidle_state)) {
2829 case RCU_SYSIDLE_NOT:
2831 /* First time all are idle, so note a short idle period. */
2832 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2835 case RCU_SYSIDLE_SHORT:
2838 * Idle for a bit, time to advance to next state?
2839 * cmpxchg failure means race with non-idle, let them win.
2841 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2842 (void)cmpxchg(&full_sysidle_state,
2843 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2846 case RCU_SYSIDLE_LONG:
2849 * Do an additional check pass before advancing to full.
2850 * cmpxchg failure means race with non-idle, let them win.
2852 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2853 (void)cmpxchg(&full_sysidle_state,
2854 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2863 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2864 * back to the beginning.
2866 static void rcu_sysidle_cancel(void)
2869 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2870 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2874 * Update the sysidle state based on the results of a force-quiescent-state
2875 * scan of the CPUs' dyntick-idle state.
2877 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2878 unsigned long maxj, bool gpkt)
2880 if (rsp != rcu_sysidle_state)
2881 return; /* Wrong flavor, ignore. */
2882 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2883 return; /* Running state machine from timekeeping CPU. */
2885 rcu_sysidle(maxj); /* More idle! */
2887 rcu_sysidle_cancel(); /* Idle is over. */
2891 * Wrapper for rcu_sysidle_report() when called from the grace-period
2892 * kthread's context.
2894 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2897 rcu_sysidle_report(rsp, isidle, maxj, true);
2900 /* Callback and function for forcing an RCU grace period. */
2901 struct rcu_sysidle_head {
2906 static void rcu_sysidle_cb(struct rcu_head *rhp)
2908 struct rcu_sysidle_head *rshp;
2911 * The following memory barrier is needed to replace the
2912 * memory barriers that would normally be in the memory
2915 smp_mb(); /* grace period precedes setting inuse. */
2917 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2918 ACCESS_ONCE(rshp->inuse) = 0;
2922 * Check to see if the system is fully idle, other than the timekeeping CPU.
2923 * The caller must have disabled interrupts.
2925 bool rcu_sys_is_idle(void)
2927 static struct rcu_sysidle_head rsh;
2928 int rss = ACCESS_ONCE(full_sysidle_state);
2930 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2933 /* Handle small-system case by doing a full scan of CPUs. */
2934 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2935 int oldrss = rss - 1;
2938 * One pass to advance to each state up to _FULL.
2939 * Give up if any pass fails to advance the state.
2941 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2944 unsigned long maxj = jiffies - ULONG_MAX / 4;
2945 struct rcu_data *rdp;
2947 /* Scan all the CPUs looking for nonidle CPUs. */
2948 for_each_possible_cpu(cpu) {
2949 rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu);
2950 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2954 rcu_sysidle_report(rcu_sysidle_state,
2955 isidle, maxj, false);
2957 rss = ACCESS_ONCE(full_sysidle_state);
2961 /* If this is the first observation of an idle period, record it. */
2962 if (rss == RCU_SYSIDLE_FULL) {
2963 rss = cmpxchg(&full_sysidle_state,
2964 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2965 return rss == RCU_SYSIDLE_FULL;
2968 smp_mb(); /* ensure rss load happens before later caller actions. */
2970 /* If already fully idle, tell the caller (in case of races). */
2971 if (rss == RCU_SYSIDLE_FULL_NOTED)
2975 * If we aren't there yet, and a grace period is not in flight,
2976 * initiate a grace period. Either way, tell the caller that
2977 * we are not there yet. We use an xchg() rather than an assignment
2978 * to make up for the memory barriers that would otherwise be
2979 * provided by the memory allocator.
2981 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2982 !rcu_gp_in_progress(rcu_sysidle_state) &&
2983 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2984 call_rcu(&rsh.rh, rcu_sysidle_cb);
2989 * Initialize dynticks sysidle state for CPUs coming online.
2991 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2993 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2996 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2998 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
3002 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
3006 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3007 unsigned long *maxj)
3011 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3016 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3021 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3025 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3028 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3029 * grace-period kthread will do force_quiescent_state() processing?
3030 * The idea is to avoid waking up RCU core processing on such a
3031 * CPU unless the grace period has extended for too long.
3033 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3034 * CONFIG_RCU_NOCB_CPU CPUs.
3036 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3038 #ifdef CONFIG_NO_HZ_FULL
3039 if (tick_nohz_full_cpu(smp_processor_id()) &&
3040 (!rcu_gp_in_progress(rsp) ||
3041 ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
3043 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3048 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3051 static void rcu_bind_gp_kthread(void)
3053 int __maybe_unused cpu;
3055 if (!tick_nohz_full_enabled())
3057 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3058 cpu = tick_do_timer_cpu;
3059 if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu)
3060 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3061 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3062 if (!is_housekeeping_cpu(raw_smp_processor_id()))
3063 housekeeping_affine(current);
3064 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */