2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
31 #include "../time/tick-internal.h"
33 #ifdef CONFIG_RCU_BOOST
35 #include "../locking/rtmutex_common.h"
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
41 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
42 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
43 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
44 DEFINE_PER_CPU(char, rcu_cpu_has_work);
46 #endif /* #ifdef CONFIG_RCU_BOOST */
48 #ifdef CONFIG_RCU_NOCB_CPU
49 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
50 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
51 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
52 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
55 * Check the RCU kernel configuration parameters and print informative
56 * messages about anything out of the ordinary. If you like #ifdef, you
57 * will love this function.
59 static void __init rcu_bootup_announce_oddness(void)
61 if (IS_ENABLED(CONFIG_RCU_TRACE))
62 pr_info("\tRCU debugfs-based tracing is enabled.\n");
63 if ((IS_ENABLED(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) ||
64 (!IS_ENABLED(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32))
65 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
67 if (IS_ENABLED(CONFIG_RCU_FANOUT_EXACT))
68 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
69 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
70 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
71 if (IS_ENABLED(CONFIG_PROVE_RCU))
72 pr_info("\tRCU lockdep checking is enabled.\n");
73 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE))
74 pr_info("\tRCU torture testing starts during boot.\n");
75 if (IS_ENABLED(CONFIG_RCU_CPU_STALL_INFO))
76 pr_info("\tAdditional per-CPU info printed with stalls.\n");
77 if (NUM_RCU_LVL_4 != 0)
78 pr_info("\tFour-level hierarchy is enabled.\n");
79 if (CONFIG_RCU_FANOUT_LEAF != 16)
80 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
81 CONFIG_RCU_FANOUT_LEAF);
82 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
83 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
84 if (nr_cpu_ids != NR_CPUS)
85 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
86 if (IS_ENABLED(CONFIG_RCU_BOOST))
87 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
90 #ifdef CONFIG_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);
96 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
100 * Tell them what RCU they are running.
102 static void __init rcu_bootup_announce(void)
104 pr_info("Preemptible hierarchical RCU implementation.\n");
105 rcu_bootup_announce_oddness();
109 * Record a preemptible-RCU quiescent state for the specified CPU. Note
110 * that this just means that the task currently running on the CPU is
111 * not in a quiescent state. There might be any number of tasks blocked
112 * while in an RCU read-side critical section.
114 * As with the other rcu_*_qs() functions, callers to this function
115 * must disable preemption.
117 static void rcu_preempt_qs(void)
119 if (!__this_cpu_read(rcu_preempt_data.passed_quiesce)) {
120 trace_rcu_grace_period(TPS("rcu_preempt"),
121 __this_cpu_read(rcu_preempt_data.gpnum),
123 __this_cpu_write(rcu_preempt_data.passed_quiesce, 1);
124 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
125 current->rcu_read_unlock_special.b.need_qs = false;
130 * We have entered the scheduler, and the current task might soon be
131 * context-switched away from. If this task is in an RCU read-side
132 * critical section, we will no longer be able to rely on the CPU to
133 * record that fact, so we enqueue the task on the blkd_tasks list.
134 * The task will dequeue itself when it exits the outermost enclosing
135 * RCU read-side critical section. Therefore, the current grace period
136 * cannot be permitted to complete until the blkd_tasks list entries
137 * predating the current grace period drain, in other words, until
138 * rnp->gp_tasks becomes NULL.
140 * Caller must disable preemption.
142 static void rcu_preempt_note_context_switch(void)
144 struct task_struct *t = current;
146 struct rcu_data *rdp;
147 struct rcu_node *rnp;
149 if (t->rcu_read_lock_nesting > 0 &&
150 !t->rcu_read_unlock_special.b.blocked) {
152 /* Possibly blocking in an RCU read-side critical section. */
153 rdp = this_cpu_ptr(rcu_preempt_state.rda);
155 raw_spin_lock_irqsave(&rnp->lock, flags);
156 smp_mb__after_unlock_lock();
157 t->rcu_read_unlock_special.b.blocked = true;
158 t->rcu_blocked_node = rnp;
161 * If this CPU has already checked in, then this task
162 * will hold up the next grace period rather than the
163 * current grace period. Queue the task accordingly.
164 * If the task is queued for the current grace period
165 * (i.e., this CPU has not yet passed through a quiescent
166 * state for the current grace period), then as long
167 * as that task remains queued, the current grace period
168 * cannot end. Note that there is some uncertainty as
169 * to exactly when the current grace period started.
170 * We take a conservative approach, which can result
171 * in unnecessarily waiting on tasks that started very
172 * slightly after the current grace period began. C'est
175 * But first, note that the current CPU must still be
178 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
179 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
180 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
181 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
182 rnp->gp_tasks = &t->rcu_node_entry;
183 #ifdef CONFIG_RCU_BOOST
184 if (rnp->boost_tasks != NULL)
185 rnp->boost_tasks = rnp->gp_tasks;
186 #endif /* #ifdef CONFIG_RCU_BOOST */
188 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
189 if (rnp->qsmask & rdp->grpmask)
190 rnp->gp_tasks = &t->rcu_node_entry;
192 trace_rcu_preempt_task(rdp->rsp->name,
194 (rnp->qsmask & rdp->grpmask)
197 raw_spin_unlock_irqrestore(&rnp->lock, flags);
198 } else if (t->rcu_read_lock_nesting < 0 &&
199 t->rcu_read_unlock_special.s) {
202 * Complete exit from RCU read-side critical section on
203 * behalf of preempted instance of __rcu_read_unlock().
205 rcu_read_unlock_special(t);
209 * Either we were not in an RCU read-side critical section to
210 * begin with, or we have now recorded that critical section
211 * globally. Either way, we can now note a quiescent state
212 * for this CPU. Again, if we were in an RCU read-side critical
213 * section, and if that critical section was blocking the current
214 * grace period, then the fact that the task has been enqueued
215 * means that we continue to block the current grace period.
221 * Check for preempted RCU readers blocking the current grace period
222 * for the specified rcu_node structure. If the caller needs a reliable
223 * answer, it must hold the rcu_node's ->lock.
225 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
227 return rnp->gp_tasks != NULL;
231 * Advance a ->blkd_tasks-list pointer to the next entry, instead
232 * returning NULL if at the end of the list.
234 static struct list_head *rcu_next_node_entry(struct task_struct *t,
235 struct rcu_node *rnp)
237 struct list_head *np;
239 np = t->rcu_node_entry.next;
240 if (np == &rnp->blkd_tasks)
246 * Return true if the specified rcu_node structure has tasks that were
247 * preempted within an RCU read-side critical section.
249 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
251 return !list_empty(&rnp->blkd_tasks);
255 * Handle special cases during rcu_read_unlock(), such as needing to
256 * notify RCU core processing or task having blocked during the RCU
257 * read-side critical section.
259 void rcu_read_unlock_special(struct task_struct *t)
265 struct list_head *np;
266 #ifdef CONFIG_RCU_BOOST
267 bool drop_boost_mutex = false;
268 #endif /* #ifdef CONFIG_RCU_BOOST */
269 struct rcu_node *rnp;
270 union rcu_special special;
272 /* NMI handlers cannot block and cannot safely manipulate state. */
276 local_irq_save(flags);
279 * If RCU core is waiting for this CPU to exit critical section,
280 * let it know that we have done so. Because irqs are disabled,
281 * t->rcu_read_unlock_special cannot change.
283 special = t->rcu_read_unlock_special;
284 if (special.b.need_qs) {
286 t->rcu_read_unlock_special.b.need_qs = false;
287 if (!t->rcu_read_unlock_special.s) {
288 local_irq_restore(flags);
293 /* Hardware IRQ handlers cannot block, complain if they get here. */
294 if (in_irq() || in_serving_softirq()) {
295 lockdep_rcu_suspicious(__FILE__, __LINE__,
296 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
297 pr_alert("->rcu_read_unlock_special: %#x (b: %d, nq: %d)\n",
298 t->rcu_read_unlock_special.s,
299 t->rcu_read_unlock_special.b.blocked,
300 t->rcu_read_unlock_special.b.need_qs);
301 local_irq_restore(flags);
305 /* Clean up if blocked during RCU read-side critical section. */
306 if (special.b.blocked) {
307 t->rcu_read_unlock_special.b.blocked = false;
310 * Remove this task from the list it blocked on. The
311 * task can migrate while we acquire the lock, but at
312 * most one time. So at most two passes through loop.
315 rnp = t->rcu_blocked_node;
316 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
317 smp_mb__after_unlock_lock();
318 if (rnp == t->rcu_blocked_node)
320 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
322 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
323 empty_exp = !rcu_preempted_readers_exp(rnp);
324 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
325 np = rcu_next_node_entry(t, rnp);
326 list_del_init(&t->rcu_node_entry);
327 t->rcu_blocked_node = NULL;
328 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
330 if (&t->rcu_node_entry == rnp->gp_tasks)
332 if (&t->rcu_node_entry == rnp->exp_tasks)
334 #ifdef CONFIG_RCU_BOOST
335 if (&t->rcu_node_entry == rnp->boost_tasks)
336 rnp->boost_tasks = np;
337 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
338 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
339 #endif /* #ifdef CONFIG_RCU_BOOST */
342 * If this was the last task on the current list, and if
343 * we aren't waiting on any CPUs, report the quiescent state.
344 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
345 * so we must take a snapshot of the expedited state.
347 empty_exp_now = !rcu_preempted_readers_exp(rnp);
348 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
349 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
356 rcu_report_unblock_qs_rnp(&rcu_preempt_state,
359 raw_spin_unlock_irqrestore(&rnp->lock, flags);
362 #ifdef CONFIG_RCU_BOOST
363 /* Unboost if we were boosted. */
364 if (drop_boost_mutex)
365 rt_mutex_unlock(&rnp->boost_mtx);
366 #endif /* #ifdef CONFIG_RCU_BOOST */
369 * If this was the last task on the expedited lists,
370 * then we need to report up the rcu_node hierarchy.
372 if (!empty_exp && empty_exp_now)
373 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
375 local_irq_restore(flags);
380 * Dump detailed information for all tasks blocking the current RCU
381 * grace period on the specified rcu_node structure.
383 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
386 struct task_struct *t;
388 raw_spin_lock_irqsave(&rnp->lock, flags);
389 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
390 raw_spin_unlock_irqrestore(&rnp->lock, flags);
393 t = list_entry(rnp->gp_tasks,
394 struct task_struct, rcu_node_entry);
395 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
397 raw_spin_unlock_irqrestore(&rnp->lock, flags);
401 * Dump detailed information for all tasks blocking the current RCU
404 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
406 struct rcu_node *rnp = rcu_get_root(rsp);
408 rcu_print_detail_task_stall_rnp(rnp);
409 rcu_for_each_leaf_node(rsp, rnp)
410 rcu_print_detail_task_stall_rnp(rnp);
413 #ifdef CONFIG_RCU_CPU_STALL_INFO
415 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
417 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
418 rnp->level, rnp->grplo, rnp->grphi);
421 static void rcu_print_task_stall_end(void)
426 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
428 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
432 static void rcu_print_task_stall_end(void)
436 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
439 * Scan the current list of tasks blocked within RCU read-side critical
440 * sections, printing out the tid of each.
442 static int rcu_print_task_stall(struct rcu_node *rnp)
444 struct task_struct *t;
447 if (!rcu_preempt_blocked_readers_cgp(rnp))
449 rcu_print_task_stall_begin(rnp);
450 t = list_entry(rnp->gp_tasks,
451 struct task_struct, rcu_node_entry);
452 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
453 pr_cont(" P%d", t->pid);
456 rcu_print_task_stall_end();
461 * Check that the list of blocked tasks for the newly completed grace
462 * period is in fact empty. It is a serious bug to complete a grace
463 * period that still has RCU readers blocked! This function must be
464 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
465 * must be held by the caller.
467 * Also, if there are blocked tasks on the list, they automatically
468 * block the newly created grace period, so set up ->gp_tasks accordingly.
470 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
472 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
473 if (rcu_preempt_has_tasks(rnp))
474 rnp->gp_tasks = rnp->blkd_tasks.next;
475 WARN_ON_ONCE(rnp->qsmask);
479 * Check for a quiescent state from the current CPU. When a task blocks,
480 * the task is recorded in the corresponding CPU's rcu_node structure,
481 * which is checked elsewhere.
483 * Caller must disable hard irqs.
485 static void rcu_preempt_check_callbacks(void)
487 struct task_struct *t = current;
489 if (t->rcu_read_lock_nesting == 0) {
493 if (t->rcu_read_lock_nesting > 0 &&
494 __this_cpu_read(rcu_preempt_data.qs_pending) &&
495 !__this_cpu_read(rcu_preempt_data.passed_quiesce))
496 t->rcu_read_unlock_special.b.need_qs = true;
499 #ifdef CONFIG_RCU_BOOST
501 static void rcu_preempt_do_callbacks(void)
503 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
506 #endif /* #ifdef CONFIG_RCU_BOOST */
509 * Queue a preemptible-RCU callback for invocation after a grace period.
511 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
513 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
515 EXPORT_SYMBOL_GPL(call_rcu);
518 * synchronize_rcu - wait until a grace period has elapsed.
520 * Control will return to the caller some time after a full grace
521 * period has elapsed, in other words after all currently executing RCU
522 * read-side critical sections have completed. Note, however, that
523 * upon return from synchronize_rcu(), the caller might well be executing
524 * concurrently with new RCU read-side critical sections that began while
525 * synchronize_rcu() was waiting. RCU read-side critical sections are
526 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
528 * See the description of synchronize_sched() for more detailed information
529 * on memory ordering guarantees.
531 void synchronize_rcu(void)
533 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
534 !lock_is_held(&rcu_lock_map) &&
535 !lock_is_held(&rcu_sched_lock_map),
536 "Illegal synchronize_rcu() in RCU read-side critical section");
537 if (!rcu_scheduler_active)
539 if (rcu_gp_is_expedited())
540 synchronize_rcu_expedited();
542 wait_rcu_gp(call_rcu);
544 EXPORT_SYMBOL_GPL(synchronize_rcu);
546 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
547 static unsigned long sync_rcu_preempt_exp_count;
548 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
551 * Return non-zero if there are any tasks in RCU read-side critical
552 * sections blocking the current preemptible-RCU expedited grace period.
553 * If there is no preemptible-RCU expedited grace period currently in
554 * progress, returns zero unconditionally.
556 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
558 return rnp->exp_tasks != NULL;
562 * return non-zero if there is no RCU expedited grace period in progress
563 * for the specified rcu_node structure, in other words, if all CPUs and
564 * tasks covered by the specified rcu_node structure have done their bit
565 * for the current expedited grace period. Works only for preemptible
566 * RCU -- other RCU implementation use other means.
568 * Caller must hold sync_rcu_preempt_exp_mutex.
570 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
572 return !rcu_preempted_readers_exp(rnp) &&
573 READ_ONCE(rnp->expmask) == 0;
577 * Report the exit from RCU read-side critical section for the last task
578 * that queued itself during or before the current expedited preemptible-RCU
579 * grace period. This event is reported either to the rcu_node structure on
580 * which the task was queued or to one of that rcu_node structure's ancestors,
581 * recursively up the tree. (Calm down, calm down, we do the recursion
584 * Caller must hold sync_rcu_preempt_exp_mutex.
586 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
592 raw_spin_lock_irqsave(&rnp->lock, flags);
593 smp_mb__after_unlock_lock();
595 if (!sync_rcu_preempt_exp_done(rnp)) {
596 raw_spin_unlock_irqrestore(&rnp->lock, flags);
599 if (rnp->parent == NULL) {
600 raw_spin_unlock_irqrestore(&rnp->lock, flags);
602 smp_mb(); /* EGP done before wake_up(). */
603 wake_up(&sync_rcu_preempt_exp_wq);
608 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
610 raw_spin_lock(&rnp->lock); /* irqs already disabled */
611 smp_mb__after_unlock_lock();
612 rnp->expmask &= ~mask;
617 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
618 * grace period for the specified rcu_node structure, phase 1. If there
619 * are such tasks, set the ->expmask bits up the rcu_node tree and also
620 * set the ->expmask bits on the leaf rcu_node structures to tell phase 2
621 * that work is needed here.
623 * Caller must hold sync_rcu_preempt_exp_mutex.
626 sync_rcu_preempt_exp_init1(struct rcu_state *rsp, struct rcu_node *rnp)
630 struct rcu_node *rnp_up;
632 raw_spin_lock_irqsave(&rnp->lock, flags);
633 smp_mb__after_unlock_lock();
634 WARN_ON_ONCE(rnp->expmask);
635 WARN_ON_ONCE(rnp->exp_tasks);
636 if (!rcu_preempt_has_tasks(rnp)) {
637 /* No blocked tasks, nothing to do. */
638 raw_spin_unlock_irqrestore(&rnp->lock, flags);
641 /* Call for Phase 2 and propagate ->expmask bits up the tree. */
644 while (rnp_up->parent) {
645 mask = rnp_up->grpmask;
646 rnp_up = rnp_up->parent;
647 if (rnp_up->expmask & mask)
649 raw_spin_lock(&rnp_up->lock); /* irqs already off */
650 smp_mb__after_unlock_lock();
651 rnp_up->expmask |= mask;
652 raw_spin_unlock(&rnp_up->lock); /* irqs still off */
654 raw_spin_unlock_irqrestore(&rnp->lock, flags);
658 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
659 * grace period for the specified rcu_node structure, phase 2. If the
660 * leaf rcu_node structure has its ->expmask field set, check for tasks.
661 * If there are some, clear ->expmask and set ->exp_tasks accordingly,
662 * then initiate RCU priority boosting. Otherwise, clear ->expmask and
663 * invoke rcu_report_exp_rnp() to clear out the upper-level ->expmask bits,
664 * enabling rcu_read_unlock_special() to do the bit-clearing.
666 * Caller must hold sync_rcu_preempt_exp_mutex.
669 sync_rcu_preempt_exp_init2(struct rcu_state *rsp, struct rcu_node *rnp)
673 raw_spin_lock_irqsave(&rnp->lock, flags);
674 smp_mb__after_unlock_lock();
676 /* Phase 1 didn't do anything, so Phase 2 doesn't either. */
677 raw_spin_unlock_irqrestore(&rnp->lock, flags);
681 /* Phase 1 is over. */
685 * If there are still blocked tasks, set up ->exp_tasks so that
686 * rcu_read_unlock_special() will wake us and then boost them.
688 if (rcu_preempt_has_tasks(rnp)) {
689 rnp->exp_tasks = rnp->blkd_tasks.next;
690 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
694 /* No longer any blocked tasks, so undo bit setting. */
695 raw_spin_unlock_irqrestore(&rnp->lock, flags);
696 rcu_report_exp_rnp(rsp, rnp, false);
700 * synchronize_rcu_expedited - Brute-force RCU grace period
702 * Wait for an RCU-preempt grace period, but expedite it. The basic
703 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
704 * the ->blkd_tasks lists and wait for this list to drain. This consumes
705 * significant time on all CPUs and is unfriendly to real-time workloads,
706 * so is thus not recommended for any sort of common-case code.
707 * In fact, if you are using synchronize_rcu_expedited() in a loop,
708 * please restructure your code to batch your updates, and then Use a
709 * single synchronize_rcu() instead.
711 void synchronize_rcu_expedited(void)
713 struct rcu_node *rnp;
714 struct rcu_state *rsp = &rcu_preempt_state;
718 smp_mb(); /* Caller's modifications seen first by other CPUs. */
719 snap = READ_ONCE(sync_rcu_preempt_exp_count) + 1;
720 smp_mb(); /* Above access cannot bleed into critical section. */
723 * Block CPU-hotplug operations. This means that any CPU-hotplug
724 * operation that finds an rcu_node structure with tasks in the
725 * process of being boosted will know that all tasks blocking
726 * this expedited grace period will already be in the process of
727 * being boosted. This simplifies the process of moving tasks
728 * from leaf to root rcu_node structures.
730 if (!try_get_online_cpus()) {
731 /* CPU-hotplug operation in flight, fall back to normal GP. */
732 wait_rcu_gp(call_rcu);
737 * Acquire lock, falling back to synchronize_rcu() if too many
738 * lock-acquisition failures. Of course, if someone does the
739 * expedited grace period for us, just leave.
741 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
742 if (ULONG_CMP_LT(snap,
743 READ_ONCE(sync_rcu_preempt_exp_count))) {
745 goto mb_ret; /* Others did our work for us. */
747 if (trycount++ < 10) {
748 udelay(trycount * num_online_cpus());
751 wait_rcu_gp(call_rcu);
755 if (ULONG_CMP_LT(snap, READ_ONCE(sync_rcu_preempt_exp_count))) {
757 goto unlock_mb_ret; /* Others did our work for us. */
760 /* force all RCU readers onto ->blkd_tasks lists. */
761 synchronize_sched_expedited();
764 * Snapshot current state of ->blkd_tasks lists into ->expmask.
765 * Phase 1 sets bits and phase 2 permits rcu_read_unlock_special()
766 * to start clearing them. Doing this in one phase leads to
767 * strange races between setting and clearing bits, so just say "no"!
769 rcu_for_each_leaf_node(rsp, rnp)
770 sync_rcu_preempt_exp_init1(rsp, rnp);
771 rcu_for_each_leaf_node(rsp, rnp)
772 sync_rcu_preempt_exp_init2(rsp, rnp);
776 /* Wait for snapshotted ->blkd_tasks lists to drain. */
777 rnp = rcu_get_root(rsp);
778 wait_event(sync_rcu_preempt_exp_wq,
779 sync_rcu_preempt_exp_done(rnp));
781 /* Clean up and exit. */
782 smp_mb(); /* ensure expedited GP seen before counter increment. */
783 WRITE_ONCE(sync_rcu_preempt_exp_count, sync_rcu_preempt_exp_count + 1);
785 mutex_unlock(&sync_rcu_preempt_exp_mutex);
787 smp_mb(); /* ensure subsequent action seen after grace period. */
789 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
792 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
794 * Note that this primitive does not necessarily wait for an RCU grace period
795 * to complete. For example, if there are no RCU callbacks queued anywhere
796 * in the system, then rcu_barrier() is within its rights to return
797 * immediately, without waiting for anything, much less an RCU grace period.
799 void rcu_barrier(void)
801 _rcu_barrier(&rcu_preempt_state);
803 EXPORT_SYMBOL_GPL(rcu_barrier);
806 * Initialize preemptible RCU's state structures.
808 static void __init __rcu_init_preempt(void)
810 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
814 * Check for a task exiting while in a preemptible-RCU read-side
815 * critical section, clean up if so. No need to issue warnings,
816 * as debug_check_no_locks_held() already does this if lockdep
821 struct task_struct *t = current;
823 if (likely(list_empty(¤t->rcu_node_entry)))
825 t->rcu_read_lock_nesting = 1;
827 t->rcu_read_unlock_special.b.blocked = true;
831 #else /* #ifdef CONFIG_PREEMPT_RCU */
833 static struct rcu_state *rcu_state_p = &rcu_sched_state;
836 * Tell them what RCU they are running.
838 static void __init rcu_bootup_announce(void)
840 pr_info("Hierarchical RCU implementation.\n");
841 rcu_bootup_announce_oddness();
845 * Because preemptible RCU does not exist, we never have to check for
846 * CPUs being in quiescent states.
848 static void rcu_preempt_note_context_switch(void)
853 * Because preemptible RCU does not exist, there are never any preempted
856 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
862 * Because there is no preemptible RCU, there can be no readers blocked.
864 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
870 * Because preemptible RCU does not exist, we never have to check for
871 * tasks blocked within RCU read-side critical sections.
873 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
878 * Because preemptible RCU does not exist, we never have to check for
879 * tasks blocked within RCU read-side critical sections.
881 static int rcu_print_task_stall(struct rcu_node *rnp)
887 * Because there is no preemptible RCU, there can be no readers blocked,
888 * so there is no need to check for blocked tasks. So check only for
889 * bogus qsmask values.
891 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
893 WARN_ON_ONCE(rnp->qsmask);
897 * Because preemptible RCU does not exist, it never has any callbacks
900 static void rcu_preempt_check_callbacks(void)
905 * Wait for an rcu-preempt grace period, but make it happen quickly.
906 * But because preemptible RCU does not exist, map to rcu-sched.
908 void synchronize_rcu_expedited(void)
910 synchronize_sched_expedited();
912 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
915 * Because preemptible RCU does not exist, rcu_barrier() is just
916 * another name for rcu_barrier_sched().
918 void rcu_barrier(void)
922 EXPORT_SYMBOL_GPL(rcu_barrier);
925 * Because preemptible RCU does not exist, it need not be initialized.
927 static void __init __rcu_init_preempt(void)
932 * Because preemptible RCU does not exist, tasks cannot possibly exit
933 * while in preemptible RCU read-side critical sections.
939 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
941 #ifdef CONFIG_RCU_BOOST
943 #include "../locking/rtmutex_common.h"
945 #ifdef CONFIG_RCU_TRACE
947 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
949 if (!rcu_preempt_has_tasks(rnp))
950 rnp->n_balk_blkd_tasks++;
951 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
952 rnp->n_balk_exp_gp_tasks++;
953 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
954 rnp->n_balk_boost_tasks++;
955 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
956 rnp->n_balk_notblocked++;
957 else if (rnp->gp_tasks != NULL &&
958 ULONG_CMP_LT(jiffies, rnp->boost_time))
959 rnp->n_balk_notyet++;
964 #else /* #ifdef CONFIG_RCU_TRACE */
966 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
970 #endif /* #else #ifdef CONFIG_RCU_TRACE */
972 static void rcu_wake_cond(struct task_struct *t, int status)
975 * If the thread is yielding, only wake it when this
976 * is invoked from idle
978 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
983 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
984 * or ->boost_tasks, advancing the pointer to the next task in the
987 * Note that irqs must be enabled: boosting the task can block.
988 * Returns 1 if there are more tasks needing to be boosted.
990 static int rcu_boost(struct rcu_node *rnp)
993 struct task_struct *t;
994 struct list_head *tb;
996 if (READ_ONCE(rnp->exp_tasks) == NULL &&
997 READ_ONCE(rnp->boost_tasks) == NULL)
998 return 0; /* Nothing left to boost. */
1000 raw_spin_lock_irqsave(&rnp->lock, flags);
1001 smp_mb__after_unlock_lock();
1004 * Recheck under the lock: all tasks in need of boosting
1005 * might exit their RCU read-side critical sections on their own.
1007 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1008 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1013 * Preferentially boost tasks blocking expedited grace periods.
1014 * This cannot starve the normal grace periods because a second
1015 * expedited grace period must boost all blocked tasks, including
1016 * those blocking the pre-existing normal grace period.
1018 if (rnp->exp_tasks != NULL) {
1019 tb = rnp->exp_tasks;
1020 rnp->n_exp_boosts++;
1022 tb = rnp->boost_tasks;
1023 rnp->n_normal_boosts++;
1025 rnp->n_tasks_boosted++;
1028 * We boost task t by manufacturing an rt_mutex that appears to
1029 * be held by task t. We leave a pointer to that rt_mutex where
1030 * task t can find it, and task t will release the mutex when it
1031 * exits its outermost RCU read-side critical section. Then
1032 * simply acquiring this artificial rt_mutex will boost task
1033 * t's priority. (Thanks to tglx for suggesting this approach!)
1035 * Note that task t must acquire rnp->lock to remove itself from
1036 * the ->blkd_tasks list, which it will do from exit() if from
1037 * nowhere else. We therefore are guaranteed that task t will
1038 * stay around at least until we drop rnp->lock. Note that
1039 * rnp->lock also resolves races between our priority boosting
1040 * and task t's exiting its outermost RCU read-side critical
1043 t = container_of(tb, struct task_struct, rcu_node_entry);
1044 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1045 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1046 /* Lock only for side effect: boosts task t's priority. */
1047 rt_mutex_lock(&rnp->boost_mtx);
1048 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1050 return READ_ONCE(rnp->exp_tasks) != NULL ||
1051 READ_ONCE(rnp->boost_tasks) != NULL;
1055 * Priority-boosting kthread. One per leaf rcu_node and one for the
1058 static int rcu_boost_kthread(void *arg)
1060 struct rcu_node *rnp = (struct rcu_node *)arg;
1064 trace_rcu_utilization(TPS("Start boost kthread@init"));
1066 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1067 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1068 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1069 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1070 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1071 more2boost = rcu_boost(rnp);
1077 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1078 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1079 schedule_timeout_interruptible(2);
1080 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1085 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1090 * Check to see if it is time to start boosting RCU readers that are
1091 * blocking the current grace period, and, if so, tell the per-rcu_node
1092 * kthread to start boosting them. If there is an expedited grace
1093 * period in progress, it is always time to boost.
1095 * The caller must hold rnp->lock, which this function releases.
1096 * The ->boost_kthread_task is immortal, so we don't need to worry
1097 * about it going away.
1099 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1100 __releases(rnp->lock)
1102 struct task_struct *t;
1104 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1105 rnp->n_balk_exp_gp_tasks++;
1106 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1109 if (rnp->exp_tasks != NULL ||
1110 (rnp->gp_tasks != NULL &&
1111 rnp->boost_tasks == NULL &&
1113 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1114 if (rnp->exp_tasks == NULL)
1115 rnp->boost_tasks = rnp->gp_tasks;
1116 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1117 t = rnp->boost_kthread_task;
1119 rcu_wake_cond(t, rnp->boost_kthread_status);
1121 rcu_initiate_boost_trace(rnp);
1122 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1127 * Wake up the per-CPU kthread to invoke RCU callbacks.
1129 static void invoke_rcu_callbacks_kthread(void)
1131 unsigned long flags;
1133 local_irq_save(flags);
1134 __this_cpu_write(rcu_cpu_has_work, 1);
1135 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1136 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1137 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1138 __this_cpu_read(rcu_cpu_kthread_status));
1140 local_irq_restore(flags);
1144 * Is the current CPU running the RCU-callbacks kthread?
1145 * Caller must have preemption disabled.
1147 static bool rcu_is_callbacks_kthread(void)
1149 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1152 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1155 * Do priority-boost accounting for the start of a new grace period.
1157 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1159 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1163 * Create an RCU-boost kthread for the specified node if one does not
1164 * already exist. We only create this kthread for preemptible RCU.
1165 * Returns zero if all is well, a negated errno otherwise.
1167 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1168 struct rcu_node *rnp)
1170 int rnp_index = rnp - &rsp->node[0];
1171 unsigned long flags;
1172 struct sched_param sp;
1173 struct task_struct *t;
1175 if (&rcu_preempt_state != rsp)
1178 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1182 if (rnp->boost_kthread_task != NULL)
1184 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1185 "rcub/%d", rnp_index);
1188 raw_spin_lock_irqsave(&rnp->lock, flags);
1189 smp_mb__after_unlock_lock();
1190 rnp->boost_kthread_task = t;
1191 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1192 sp.sched_priority = kthread_prio;
1193 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1194 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1198 static void rcu_kthread_do_work(void)
1200 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1201 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1202 rcu_preempt_do_callbacks();
1205 static void rcu_cpu_kthread_setup(unsigned int cpu)
1207 struct sched_param sp;
1209 sp.sched_priority = kthread_prio;
1210 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1213 static void rcu_cpu_kthread_park(unsigned int cpu)
1215 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1218 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1220 return __this_cpu_read(rcu_cpu_has_work);
1224 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1225 * RCU softirq used in flavors and configurations of RCU that do not
1226 * support RCU priority boosting.
1228 static void rcu_cpu_kthread(unsigned int cpu)
1230 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1231 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1234 for (spincnt = 0; spincnt < 10; spincnt++) {
1235 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1237 *statusp = RCU_KTHREAD_RUNNING;
1238 this_cpu_inc(rcu_cpu_kthread_loops);
1239 local_irq_disable();
1244 rcu_kthread_do_work();
1247 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1248 *statusp = RCU_KTHREAD_WAITING;
1252 *statusp = RCU_KTHREAD_YIELDING;
1253 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1254 schedule_timeout_interruptible(2);
1255 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1256 *statusp = RCU_KTHREAD_WAITING;
1260 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1261 * served by the rcu_node in question. The CPU hotplug lock is still
1262 * held, so the value of rnp->qsmaskinit will be stable.
1264 * We don't include outgoingcpu in the affinity set, use -1 if there is
1265 * no outgoing CPU. If there are no CPUs left in the affinity set,
1266 * this function allows the kthread to execute on any CPU.
1268 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1270 struct task_struct *t = rnp->boost_kthread_task;
1271 unsigned long mask = rcu_rnp_online_cpus(rnp);
1277 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1279 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1280 if ((mask & 0x1) && cpu != outgoingcpu)
1281 cpumask_set_cpu(cpu, cm);
1282 if (cpumask_weight(cm) == 0)
1284 set_cpus_allowed_ptr(t, cm);
1285 free_cpumask_var(cm);
1288 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1289 .store = &rcu_cpu_kthread_task,
1290 .thread_should_run = rcu_cpu_kthread_should_run,
1291 .thread_fn = rcu_cpu_kthread,
1292 .thread_comm = "rcuc/%u",
1293 .setup = rcu_cpu_kthread_setup,
1294 .park = rcu_cpu_kthread_park,
1298 * Spawn boost kthreads -- called as soon as the scheduler is running.
1300 static void __init rcu_spawn_boost_kthreads(void)
1302 struct rcu_node *rnp;
1305 for_each_possible_cpu(cpu)
1306 per_cpu(rcu_cpu_has_work, cpu) = 0;
1307 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1308 rcu_for_each_leaf_node(rcu_state_p, rnp)
1309 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1312 static void rcu_prepare_kthreads(int cpu)
1314 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1315 struct rcu_node *rnp = rdp->mynode;
1317 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1318 if (rcu_scheduler_fully_active)
1319 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1322 #else /* #ifdef CONFIG_RCU_BOOST */
1324 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1325 __releases(rnp->lock)
1327 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1330 static void invoke_rcu_callbacks_kthread(void)
1335 static bool rcu_is_callbacks_kthread(void)
1340 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1344 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1348 static void __init rcu_spawn_boost_kthreads(void)
1352 static void rcu_prepare_kthreads(int cpu)
1356 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1358 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1361 * Check to see if any future RCU-related work will need to be done
1362 * by the current CPU, even if none need be done immediately, returning
1363 * 1 if so. This function is part of the RCU implementation; it is -not-
1364 * an exported member of the RCU API.
1366 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1367 * any flavor of RCU.
1369 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1370 int rcu_needs_cpu(unsigned long *delta_jiffies)
1372 *delta_jiffies = ULONG_MAX;
1373 return rcu_cpu_has_callbacks(NULL);
1375 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1378 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1381 static void rcu_cleanup_after_idle(void)
1386 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1389 static void rcu_prepare_for_idle(void)
1394 * Don't bother keeping a running count of the number of RCU callbacks
1395 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1397 static void rcu_idle_count_callbacks_posted(void)
1401 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1404 * This code is invoked when a CPU goes idle, at which point we want
1405 * to have the CPU do everything required for RCU so that it can enter
1406 * the energy-efficient dyntick-idle mode. This is handled by a
1407 * state machine implemented by rcu_prepare_for_idle() below.
1409 * The following three proprocessor symbols control this state machine:
1411 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1412 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1413 * is sized to be roughly one RCU grace period. Those energy-efficiency
1414 * benchmarkers who might otherwise be tempted to set this to a large
1415 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1416 * system. And if you are -that- concerned about energy efficiency,
1417 * just power the system down and be done with it!
1418 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1419 * permitted to sleep in dyntick-idle mode with only lazy RCU
1420 * callbacks pending. Setting this too high can OOM your system.
1422 * The values below work well in practice. If future workloads require
1423 * adjustment, they can be converted into kernel config parameters, though
1424 * making the state machine smarter might be a better option.
1426 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1427 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1429 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1430 module_param(rcu_idle_gp_delay, int, 0644);
1431 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1432 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1434 extern int tick_nohz_active;
1437 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1438 * only if it has been awhile since the last time we did so. Afterwards,
1439 * if there are any callbacks ready for immediate invocation, return true.
1441 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1443 bool cbs_ready = false;
1444 struct rcu_data *rdp;
1445 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1446 struct rcu_node *rnp;
1447 struct rcu_state *rsp;
1449 /* Exit early if we advanced recently. */
1450 if (jiffies == rdtp->last_advance_all)
1452 rdtp->last_advance_all = jiffies;
1454 for_each_rcu_flavor(rsp) {
1455 rdp = this_cpu_ptr(rsp->rda);
1459 * Don't bother checking unless a grace period has
1460 * completed since we last checked and there are
1461 * callbacks not yet ready to invoke.
1463 if ((rdp->completed != rnp->completed ||
1464 unlikely(READ_ONCE(rdp->gpwrap))) &&
1465 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1466 note_gp_changes(rsp, rdp);
1468 if (cpu_has_callbacks_ready_to_invoke(rdp))
1475 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1476 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1477 * caller to set the timeout based on whether or not there are non-lazy
1480 * The caller must have disabled interrupts.
1482 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1483 int rcu_needs_cpu(unsigned long *dj)
1485 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1487 /* Snapshot to detect later posting of non-lazy callback. */
1488 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1490 /* If no callbacks, RCU doesn't need the CPU. */
1491 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1496 /* Attempt to advance callbacks. */
1497 if (rcu_try_advance_all_cbs()) {
1498 /* Some ready to invoke, so initiate later invocation. */
1502 rdtp->last_accelerate = jiffies;
1504 /* Request timer delay depending on laziness, and round. */
1505 if (!rdtp->all_lazy) {
1506 *dj = round_up(rcu_idle_gp_delay + jiffies,
1507 rcu_idle_gp_delay) - jiffies;
1509 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1513 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1516 * Prepare a CPU for idle from an RCU perspective. The first major task
1517 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1518 * The second major task is to check to see if a non-lazy callback has
1519 * arrived at a CPU that previously had only lazy callbacks. The third
1520 * major task is to accelerate (that is, assign grace-period numbers to)
1521 * any recently arrived callbacks.
1523 * The caller must have disabled interrupts.
1525 static void rcu_prepare_for_idle(void)
1527 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1529 struct rcu_data *rdp;
1530 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1531 struct rcu_node *rnp;
1532 struct rcu_state *rsp;
1535 /* Handle nohz enablement switches conservatively. */
1536 tne = READ_ONCE(tick_nohz_active);
1537 if (tne != rdtp->tick_nohz_enabled_snap) {
1538 if (rcu_cpu_has_callbacks(NULL))
1539 invoke_rcu_core(); /* force nohz to see update. */
1540 rdtp->tick_nohz_enabled_snap = tne;
1546 /* If this is a no-CBs CPU, no callbacks, just return. */
1547 if (rcu_is_nocb_cpu(smp_processor_id()))
1551 * If a non-lazy callback arrived at a CPU having only lazy
1552 * callbacks, invoke RCU core for the side-effect of recalculating
1553 * idle duration on re-entry to idle.
1555 if (rdtp->all_lazy &&
1556 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1557 rdtp->all_lazy = false;
1558 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1564 * If we have not yet accelerated this jiffy, accelerate all
1565 * callbacks on this CPU.
1567 if (rdtp->last_accelerate == jiffies)
1569 rdtp->last_accelerate = jiffies;
1570 for_each_rcu_flavor(rsp) {
1571 rdp = this_cpu_ptr(rsp->rda);
1572 if (!*rdp->nxttail[RCU_DONE_TAIL])
1575 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1576 smp_mb__after_unlock_lock();
1577 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1578 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1580 rcu_gp_kthread_wake(rsp);
1582 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1586 * Clean up for exit from idle. Attempt to advance callbacks based on
1587 * any grace periods that elapsed while the CPU was idle, and if any
1588 * callbacks are now ready to invoke, initiate invocation.
1590 static void rcu_cleanup_after_idle(void)
1592 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1593 if (rcu_is_nocb_cpu(smp_processor_id()))
1595 if (rcu_try_advance_all_cbs())
1597 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1601 * Keep a running count of the number of non-lazy callbacks posted
1602 * on this CPU. This running counter (which is never decremented) allows
1603 * rcu_prepare_for_idle() to detect when something out of the idle loop
1604 * posts a callback, even if an equal number of callbacks are invoked.
1605 * Of course, callbacks should only be posted from within a trace event
1606 * designed to be called from idle or from within RCU_NONIDLE().
1608 static void rcu_idle_count_callbacks_posted(void)
1610 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1614 * Data for flushing lazy RCU callbacks at OOM time.
1616 static atomic_t oom_callback_count;
1617 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1620 * RCU OOM callback -- decrement the outstanding count and deliver the
1621 * wake-up if we are the last one.
1623 static void rcu_oom_callback(struct rcu_head *rhp)
1625 if (atomic_dec_and_test(&oom_callback_count))
1626 wake_up(&oom_callback_wq);
1630 * Post an rcu_oom_notify callback on the current CPU if it has at
1631 * least one lazy callback. This will unnecessarily post callbacks
1632 * to CPUs that already have a non-lazy callback at the end of their
1633 * callback list, but this is an infrequent operation, so accept some
1634 * extra overhead to keep things simple.
1636 static void rcu_oom_notify_cpu(void *unused)
1638 struct rcu_state *rsp;
1639 struct rcu_data *rdp;
1641 for_each_rcu_flavor(rsp) {
1642 rdp = raw_cpu_ptr(rsp->rda);
1643 if (rdp->qlen_lazy != 0) {
1644 atomic_inc(&oom_callback_count);
1645 rsp->call(&rdp->oom_head, rcu_oom_callback);
1651 * If low on memory, ensure that each CPU has a non-lazy callback.
1652 * This will wake up CPUs that have only lazy callbacks, in turn
1653 * ensuring that they free up the corresponding memory in a timely manner.
1654 * Because an uncertain amount of memory will be freed in some uncertain
1655 * timeframe, we do not claim to have freed anything.
1657 static int rcu_oom_notify(struct notifier_block *self,
1658 unsigned long notused, void *nfreed)
1662 /* Wait for callbacks from earlier instance to complete. */
1663 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1664 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1667 * Prevent premature wakeup: ensure that all increments happen
1668 * before there is a chance of the counter reaching zero.
1670 atomic_set(&oom_callback_count, 1);
1673 for_each_online_cpu(cpu) {
1674 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1675 cond_resched_rcu_qs();
1679 /* Unconditionally decrement: no need to wake ourselves up. */
1680 atomic_dec(&oom_callback_count);
1685 static struct notifier_block rcu_oom_nb = {
1686 .notifier_call = rcu_oom_notify
1689 static int __init rcu_register_oom_notifier(void)
1691 register_oom_notifier(&rcu_oom_nb);
1694 early_initcall(rcu_register_oom_notifier);
1696 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1698 #ifdef CONFIG_RCU_CPU_STALL_INFO
1700 #ifdef CONFIG_RCU_FAST_NO_HZ
1702 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1704 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1705 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1707 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1708 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1710 rdtp->all_lazy ? 'L' : '.',
1711 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1714 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1716 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1721 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1723 /* Initiate the stall-info list. */
1724 static void print_cpu_stall_info_begin(void)
1730 * Print out diagnostic information for the specified stalled CPU.
1732 * If the specified CPU is aware of the current RCU grace period
1733 * (flavor specified by rsp), then print the number of scheduling
1734 * clock interrupts the CPU has taken during the time that it has
1735 * been aware. Otherwise, print the number of RCU grace periods
1736 * that this CPU is ignorant of, for example, "1" if the CPU was
1737 * aware of the previous grace period.
1739 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1741 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1743 char fast_no_hz[72];
1744 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1745 struct rcu_dynticks *rdtp = rdp->dynticks;
1747 unsigned long ticks_value;
1749 if (rsp->gpnum == rdp->gpnum) {
1750 ticks_title = "ticks this GP";
1751 ticks_value = rdp->ticks_this_gp;
1753 ticks_title = "GPs behind";
1754 ticks_value = rsp->gpnum - rdp->gpnum;
1756 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1757 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1758 cpu, ticks_value, ticks_title,
1759 atomic_read(&rdtp->dynticks) & 0xfff,
1760 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1761 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1762 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1766 /* Terminate the stall-info list. */
1767 static void print_cpu_stall_info_end(void)
1772 /* Zero ->ticks_this_gp for all flavors of RCU. */
1773 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1775 rdp->ticks_this_gp = 0;
1776 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1779 /* Increment ->ticks_this_gp for all flavors of RCU. */
1780 static void increment_cpu_stall_ticks(void)
1782 struct rcu_state *rsp;
1784 for_each_rcu_flavor(rsp)
1785 raw_cpu_inc(rsp->rda->ticks_this_gp);
1788 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1790 static void print_cpu_stall_info_begin(void)
1795 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1797 pr_cont(" %d", cpu);
1800 static void print_cpu_stall_info_end(void)
1805 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1809 static void increment_cpu_stall_ticks(void)
1813 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1815 #ifdef CONFIG_RCU_NOCB_CPU
1818 * Offload callback processing from the boot-time-specified set of CPUs
1819 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1820 * kthread created that pulls the callbacks from the corresponding CPU,
1821 * waits for a grace period to elapse, and invokes the callbacks.
1822 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1823 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1824 * has been specified, in which case each kthread actively polls its
1825 * CPU. (Which isn't so great for energy efficiency, but which does
1826 * reduce RCU's overhead on that CPU.)
1828 * This is intended to be used in conjunction with Frederic Weisbecker's
1829 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1830 * running CPU-bound user-mode computations.
1832 * Offloading of callback processing could also in theory be used as
1833 * an energy-efficiency measure because CPUs with no RCU callbacks
1834 * queued are more aggressive about entering dyntick-idle mode.
1838 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1839 static int __init rcu_nocb_setup(char *str)
1841 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1842 have_rcu_nocb_mask = true;
1843 cpulist_parse(str, rcu_nocb_mask);
1846 __setup("rcu_nocbs=", rcu_nocb_setup);
1848 static int __init parse_rcu_nocb_poll(char *arg)
1853 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1856 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1859 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1861 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1865 * Set the root rcu_node structure's ->need_future_gp field
1866 * based on the sum of those of all rcu_node structures. This does
1867 * double-count the root rcu_node structure's requests, but this
1868 * is necessary to handle the possibility of a rcu_nocb_kthread()
1869 * having awakened during the time that the rcu_node structures
1870 * were being updated for the end of the previous grace period.
1872 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1874 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1877 static void rcu_init_one_nocb(struct rcu_node *rnp)
1879 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1880 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1883 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1884 /* Is the specified CPU a no-CBs CPU? */
1885 bool rcu_is_nocb_cpu(int cpu)
1887 if (have_rcu_nocb_mask)
1888 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1891 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1894 * Kick the leader kthread for this NOCB group.
1896 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1898 struct rcu_data *rdp_leader = rdp->nocb_leader;
1900 if (!READ_ONCE(rdp_leader->nocb_kthread))
1902 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1903 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1904 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1905 wake_up(&rdp_leader->nocb_wq);
1910 * Does the specified CPU need an RCU callback for the specified flavor
1913 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1915 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1917 #ifdef CONFIG_PROVE_RCU
1918 struct rcu_head *rhp;
1919 #endif /* #ifdef CONFIG_PROVE_RCU */
1922 * Check count of all no-CBs callbacks awaiting invocation.
1923 * There needs to be a barrier before this function is called,
1924 * but associated with a prior determination that no more
1925 * callbacks would be posted. In the worst case, the first
1926 * barrier in _rcu_barrier() suffices (but the caller cannot
1927 * necessarily rely on this, not a substitute for the caller
1928 * getting the concurrency design right!). There must also be
1929 * a barrier between the following load an posting of a callback
1930 * (if a callback is in fact needed). This is associated with an
1931 * atomic_inc() in the caller.
1933 ret = atomic_long_read(&rdp->nocb_q_count);
1935 #ifdef CONFIG_PROVE_RCU
1936 rhp = READ_ONCE(rdp->nocb_head);
1938 rhp = READ_ONCE(rdp->nocb_gp_head);
1940 rhp = READ_ONCE(rdp->nocb_follower_head);
1942 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1943 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1944 rcu_scheduler_fully_active) {
1945 /* RCU callback enqueued before CPU first came online??? */
1946 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1950 #endif /* #ifdef CONFIG_PROVE_RCU */
1956 * Enqueue the specified string of rcu_head structures onto the specified
1957 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1958 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1959 * counts are supplied by rhcount and rhcount_lazy.
1961 * If warranted, also wake up the kthread servicing this CPUs queues.
1963 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1964 struct rcu_head *rhp,
1965 struct rcu_head **rhtp,
1966 int rhcount, int rhcount_lazy,
1967 unsigned long flags)
1970 struct rcu_head **old_rhpp;
1971 struct task_struct *t;
1973 /* Enqueue the callback on the nocb list and update counts. */
1974 atomic_long_add(rhcount, &rdp->nocb_q_count);
1975 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1976 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1977 WRITE_ONCE(*old_rhpp, rhp);
1978 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1979 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1981 /* If we are not being polled and there is a kthread, awaken it ... */
1982 t = READ_ONCE(rdp->nocb_kthread);
1983 if (rcu_nocb_poll || !t) {
1984 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1985 TPS("WakeNotPoll"));
1988 len = atomic_long_read(&rdp->nocb_q_count);
1989 if (old_rhpp == &rdp->nocb_head) {
1990 if (!irqs_disabled_flags(flags)) {
1991 /* ... if queue was empty ... */
1992 wake_nocb_leader(rdp, false);
1993 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1996 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1997 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1998 TPS("WakeEmptyIsDeferred"));
2000 rdp->qlen_last_fqs_check = 0;
2001 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2002 /* ... or if many callbacks queued. */
2003 if (!irqs_disabled_flags(flags)) {
2004 wake_nocb_leader(rdp, true);
2005 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2008 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
2009 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2010 TPS("WakeOvfIsDeferred"));
2012 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2014 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2020 * This is a helper for __call_rcu(), which invokes this when the normal
2021 * callback queue is inoperable. If this is not a no-CBs CPU, this
2022 * function returns failure back to __call_rcu(), which can complain
2025 * Otherwise, this function queues the callback where the corresponding
2026 * "rcuo" kthread can find it.
2028 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2029 bool lazy, unsigned long flags)
2032 if (!rcu_is_nocb_cpu(rdp->cpu))
2034 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2035 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2036 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2037 (unsigned long)rhp->func,
2038 -atomic_long_read(&rdp->nocb_q_count_lazy),
2039 -atomic_long_read(&rdp->nocb_q_count));
2041 trace_rcu_callback(rdp->rsp->name, rhp,
2042 -atomic_long_read(&rdp->nocb_q_count_lazy),
2043 -atomic_long_read(&rdp->nocb_q_count));
2046 * If called from an extended quiescent state with interrupts
2047 * disabled, invoke the RCU core in order to allow the idle-entry
2048 * deferred-wakeup check to function.
2050 if (irqs_disabled_flags(flags) &&
2051 !rcu_is_watching() &&
2052 cpu_online(smp_processor_id()))
2059 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2062 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2063 struct rcu_data *rdp,
2064 unsigned long flags)
2066 long ql = rsp->qlen;
2067 long qll = rsp->qlen_lazy;
2069 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2070 if (!rcu_is_nocb_cpu(smp_processor_id()))
2075 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2076 if (rsp->orphan_donelist != NULL) {
2077 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2078 rsp->orphan_donetail, ql, qll, flags);
2080 rsp->orphan_donelist = NULL;
2081 rsp->orphan_donetail = &rsp->orphan_donelist;
2083 if (rsp->orphan_nxtlist != NULL) {
2084 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2085 rsp->orphan_nxttail, ql, qll, flags);
2087 rsp->orphan_nxtlist = NULL;
2088 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2094 * If necessary, kick off a new grace period, and either way wait
2095 * for a subsequent grace period to complete.
2097 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2101 unsigned long flags;
2103 struct rcu_node *rnp = rdp->mynode;
2105 raw_spin_lock_irqsave(&rnp->lock, flags);
2106 smp_mb__after_unlock_lock();
2107 needwake = rcu_start_future_gp(rnp, rdp, &c);
2108 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2110 rcu_gp_kthread_wake(rdp->rsp);
2113 * Wait for the grace period. Do so interruptibly to avoid messing
2114 * up the load average.
2116 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2118 wait_event_interruptible(
2119 rnp->nocb_gp_wq[c & 0x1],
2120 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
2123 WARN_ON(signal_pending(current));
2124 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2126 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2127 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2131 * Leaders come here to wait for additional callbacks to show up.
2132 * This function does not return until callbacks appear.
2134 static void nocb_leader_wait(struct rcu_data *my_rdp)
2136 bool firsttime = true;
2138 struct rcu_data *rdp;
2139 struct rcu_head **tail;
2143 /* Wait for callbacks to appear. */
2144 if (!rcu_nocb_poll) {
2145 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2146 wait_event_interruptible(my_rdp->nocb_wq,
2147 !READ_ONCE(my_rdp->nocb_leader_sleep));
2148 /* Memory barrier handled by smp_mb() calls below and repoll. */
2149 } else if (firsttime) {
2150 firsttime = false; /* Don't drown trace log with "Poll"! */
2151 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2155 * Each pass through the following loop checks a follower for CBs.
2156 * We are our own first follower. Any CBs found are moved to
2157 * nocb_gp_head, where they await a grace period.
2160 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2161 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2162 if (!rdp->nocb_gp_head)
2163 continue; /* No CBs here, try next follower. */
2165 /* Move callbacks to wait-for-GP list, which is empty. */
2166 WRITE_ONCE(rdp->nocb_head, NULL);
2167 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2172 * If there were no callbacks, sleep a bit, rescan after a
2173 * memory barrier, and go retry.
2175 if (unlikely(!gotcbs)) {
2177 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2179 WARN_ON(signal_pending(current));
2180 schedule_timeout_interruptible(1);
2182 /* Rescan in case we were a victim of memory ordering. */
2183 my_rdp->nocb_leader_sleep = true;
2184 smp_mb(); /* Ensure _sleep true before scan. */
2185 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2186 if (READ_ONCE(rdp->nocb_head)) {
2187 /* Found CB, so short-circuit next wait. */
2188 my_rdp->nocb_leader_sleep = false;
2194 /* Wait for one grace period. */
2195 rcu_nocb_wait_gp(my_rdp);
2198 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2199 * We set it now, but recheck for new callbacks while
2200 * traversing our follower list.
2202 my_rdp->nocb_leader_sleep = true;
2203 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2205 /* Each pass through the following loop wakes a follower, if needed. */
2206 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2207 if (READ_ONCE(rdp->nocb_head))
2208 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2209 if (!rdp->nocb_gp_head)
2210 continue; /* No CBs, so no need to wake follower. */
2212 /* Append callbacks to follower's "done" list. */
2213 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2214 *tail = rdp->nocb_gp_head;
2215 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2216 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2218 * List was empty, wake up the follower.
2219 * Memory barriers supplied by atomic_long_add().
2221 wake_up(&rdp->nocb_wq);
2225 /* If we (the leader) don't have CBs, go wait some more. */
2226 if (!my_rdp->nocb_follower_head)
2231 * Followers come here to wait for additional callbacks to show up.
2232 * This function does not return until callbacks appear.
2234 static void nocb_follower_wait(struct rcu_data *rdp)
2236 bool firsttime = true;
2239 if (!rcu_nocb_poll) {
2240 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2242 wait_event_interruptible(rdp->nocb_wq,
2243 READ_ONCE(rdp->nocb_follower_head));
2244 } else if (firsttime) {
2245 /* Don't drown trace log with "Poll"! */
2247 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2249 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2250 /* ^^^ Ensure CB invocation follows _head test. */
2254 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2256 WARN_ON(signal_pending(current));
2257 schedule_timeout_interruptible(1);
2262 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2263 * callbacks queued by the corresponding no-CBs CPU, however, there is
2264 * an optional leader-follower relationship so that the grace-period
2265 * kthreads don't have to do quite so many wakeups.
2267 static int rcu_nocb_kthread(void *arg)
2270 struct rcu_head *list;
2271 struct rcu_head *next;
2272 struct rcu_head **tail;
2273 struct rcu_data *rdp = arg;
2275 /* Each pass through this loop invokes one batch of callbacks */
2277 /* Wait for callbacks. */
2278 if (rdp->nocb_leader == rdp)
2279 nocb_leader_wait(rdp);
2281 nocb_follower_wait(rdp);
2283 /* Pull the ready-to-invoke callbacks onto local list. */
2284 list = READ_ONCE(rdp->nocb_follower_head);
2286 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2287 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2288 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2290 /* Each pass through the following loop invokes a callback. */
2291 trace_rcu_batch_start(rdp->rsp->name,
2292 atomic_long_read(&rdp->nocb_q_count_lazy),
2293 atomic_long_read(&rdp->nocb_q_count), -1);
2297 /* Wait for enqueuing to complete, if needed. */
2298 while (next == NULL && &list->next != tail) {
2299 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2301 schedule_timeout_interruptible(1);
2302 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2306 debug_rcu_head_unqueue(list);
2308 if (__rcu_reclaim(rdp->rsp->name, list))
2314 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2315 smp_mb__before_atomic(); /* _add after CB invocation. */
2316 atomic_long_add(-c, &rdp->nocb_q_count);
2317 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2318 rdp->n_nocbs_invoked += c;
2323 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2324 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2326 return READ_ONCE(rdp->nocb_defer_wakeup);
2329 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2330 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2334 if (!rcu_nocb_need_deferred_wakeup(rdp))
2336 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2337 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2338 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2339 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2342 void __init rcu_init_nohz(void)
2345 bool need_rcu_nocb_mask = true;
2346 struct rcu_state *rsp;
2348 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2349 need_rcu_nocb_mask = false;
2350 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2352 #if defined(CONFIG_NO_HZ_FULL)
2353 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2354 need_rcu_nocb_mask = true;
2355 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2357 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2358 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2359 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2362 have_rcu_nocb_mask = true;
2364 if (!have_rcu_nocb_mask)
2367 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2368 pr_info("\tOffload RCU callbacks from CPU 0\n");
2369 cpumask_set_cpu(0, rcu_nocb_mask);
2370 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2371 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2372 pr_info("\tOffload RCU callbacks from all CPUs\n");
2373 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2374 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2375 #if defined(CONFIG_NO_HZ_FULL)
2376 if (tick_nohz_full_running)
2377 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2378 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2380 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2381 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2382 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2385 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2386 cpumask_pr_args(rcu_nocb_mask));
2388 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2390 for_each_rcu_flavor(rsp) {
2391 for_each_cpu(cpu, rcu_nocb_mask)
2392 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2393 rcu_organize_nocb_kthreads(rsp);
2397 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2398 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2400 rdp->nocb_tail = &rdp->nocb_head;
2401 init_waitqueue_head(&rdp->nocb_wq);
2402 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2406 * If the specified CPU is a no-CBs CPU that does not already have its
2407 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2408 * brought online out of order, this can require re-organizing the
2409 * leader-follower relationships.
2411 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2413 struct rcu_data *rdp;
2414 struct rcu_data *rdp_last;
2415 struct rcu_data *rdp_old_leader;
2416 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2417 struct task_struct *t;
2420 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2421 * then nothing to do.
2423 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2426 /* If we didn't spawn the leader first, reorganize! */
2427 rdp_old_leader = rdp_spawn->nocb_leader;
2428 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2430 rdp = rdp_old_leader;
2432 rdp->nocb_leader = rdp_spawn;
2433 if (rdp_last && rdp != rdp_spawn)
2434 rdp_last->nocb_next_follower = rdp;
2435 if (rdp == rdp_spawn) {
2436 rdp = rdp->nocb_next_follower;
2439 rdp = rdp->nocb_next_follower;
2440 rdp_last->nocb_next_follower = NULL;
2443 rdp_spawn->nocb_next_follower = rdp_old_leader;
2446 /* Spawn the kthread for this CPU and RCU flavor. */
2447 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2448 "rcuo%c/%d", rsp->abbr, cpu);
2450 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2454 * If the specified CPU is a no-CBs CPU that does not already have its
2455 * rcuo kthreads, spawn them.
2457 static void rcu_spawn_all_nocb_kthreads(int cpu)
2459 struct rcu_state *rsp;
2461 if (rcu_scheduler_fully_active)
2462 for_each_rcu_flavor(rsp)
2463 rcu_spawn_one_nocb_kthread(rsp, cpu);
2467 * Once the scheduler is running, spawn rcuo kthreads for all online
2468 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2469 * non-boot CPUs come online -- if this changes, we will need to add
2470 * some mutual exclusion.
2472 static void __init rcu_spawn_nocb_kthreads(void)
2476 for_each_online_cpu(cpu)
2477 rcu_spawn_all_nocb_kthreads(cpu);
2480 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2481 static int rcu_nocb_leader_stride = -1;
2482 module_param(rcu_nocb_leader_stride, int, 0444);
2485 * Initialize leader-follower relationships for all no-CBs CPU.
2487 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2490 int ls = rcu_nocb_leader_stride;
2491 int nl = 0; /* Next leader. */
2492 struct rcu_data *rdp;
2493 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2494 struct rcu_data *rdp_prev = NULL;
2496 if (!have_rcu_nocb_mask)
2499 ls = int_sqrt(nr_cpu_ids);
2500 rcu_nocb_leader_stride = ls;
2504 * Each pass through this loop sets up one rcu_data structure and
2505 * spawns one rcu_nocb_kthread().
2507 for_each_cpu(cpu, rcu_nocb_mask) {
2508 rdp = per_cpu_ptr(rsp->rda, cpu);
2509 if (rdp->cpu >= nl) {
2510 /* New leader, set up for followers & next leader. */
2511 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2512 rdp->nocb_leader = rdp;
2515 /* Another follower, link to previous leader. */
2516 rdp->nocb_leader = rdp_leader;
2517 rdp_prev->nocb_next_follower = rdp;
2523 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2524 static bool init_nocb_callback_list(struct rcu_data *rdp)
2526 if (!rcu_is_nocb_cpu(rdp->cpu))
2529 /* If there are early-boot callbacks, move them to nocb lists. */
2531 rdp->nocb_head = rdp->nxtlist;
2532 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2533 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2534 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2535 rdp->nxtlist = NULL;
2539 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2543 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2545 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2547 WARN_ON_ONCE(1); /* Should be dead code. */
2551 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2555 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2559 static void rcu_init_one_nocb(struct rcu_node *rnp)
2563 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2564 bool lazy, unsigned long flags)
2569 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2570 struct rcu_data *rdp,
2571 unsigned long flags)
2576 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2580 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2585 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2589 static void rcu_spawn_all_nocb_kthreads(int cpu)
2593 static void __init rcu_spawn_nocb_kthreads(void)
2597 static bool init_nocb_callback_list(struct rcu_data *rdp)
2602 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2605 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2606 * arbitrarily long period of time with the scheduling-clock tick turned
2607 * off. RCU will be paying attention to this CPU because it is in the
2608 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2609 * machine because the scheduling-clock tick has been disabled. Therefore,
2610 * if an adaptive-ticks CPU is failing to respond to the current grace
2611 * period and has not be idle from an RCU perspective, kick it.
2613 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2615 #ifdef CONFIG_NO_HZ_FULL
2616 if (tick_nohz_full_cpu(cpu))
2617 smp_send_reschedule(cpu);
2618 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2622 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2624 static int full_sysidle_state; /* Current system-idle state. */
2625 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2626 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2627 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2628 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2629 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2632 * Invoked to note exit from irq or task transition to idle. Note that
2633 * usermode execution does -not- count as idle here! After all, we want
2634 * to detect full-system idle states, not RCU quiescent states and grace
2635 * periods. The caller must have disabled interrupts.
2637 static void rcu_sysidle_enter(int irq)
2640 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2642 /* If there are no nohz_full= CPUs, no need to track this. */
2643 if (!tick_nohz_full_enabled())
2646 /* Adjust nesting, check for fully idle. */
2648 rdtp->dynticks_idle_nesting--;
2649 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2650 if (rdtp->dynticks_idle_nesting != 0)
2651 return; /* Still not fully idle. */
2653 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2654 DYNTICK_TASK_NEST_VALUE) {
2655 rdtp->dynticks_idle_nesting = 0;
2657 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2658 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2659 return; /* Still not fully idle. */
2663 /* Record start of fully idle period. */
2665 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2666 smp_mb__before_atomic();
2667 atomic_inc(&rdtp->dynticks_idle);
2668 smp_mb__after_atomic();
2669 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2673 * Unconditionally force exit from full system-idle state. This is
2674 * invoked when a normal CPU exits idle, but must be called separately
2675 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2676 * is that the timekeeping CPU is permitted to take scheduling-clock
2677 * interrupts while the system is in system-idle state, and of course
2678 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2679 * interrupt from any other type of interrupt.
2681 void rcu_sysidle_force_exit(void)
2683 int oldstate = READ_ONCE(full_sysidle_state);
2687 * Each pass through the following loop attempts to exit full
2688 * system-idle state. If contention proves to be a problem,
2689 * a trylock-based contention tree could be used here.
2691 while (oldstate > RCU_SYSIDLE_SHORT) {
2692 newoldstate = cmpxchg(&full_sysidle_state,
2693 oldstate, RCU_SYSIDLE_NOT);
2694 if (oldstate == newoldstate &&
2695 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2696 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2697 return; /* We cleared it, done! */
2699 oldstate = newoldstate;
2701 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2705 * Invoked to note entry to irq or task transition from idle. Note that
2706 * usermode execution does -not- count as idle here! The caller must
2707 * have disabled interrupts.
2709 static void rcu_sysidle_exit(int irq)
2711 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2713 /* If there are no nohz_full= CPUs, no need to track this. */
2714 if (!tick_nohz_full_enabled())
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,
2763 * and must be running on tick_do_timer_cpu.
2765 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2766 unsigned long *maxj)
2770 struct rcu_dynticks *rdtp = rdp->dynticks;
2772 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2773 if (!tick_nohz_full_enabled())
2777 * If some other CPU has already reported non-idle, if this is
2778 * not the flavor of RCU that tracks sysidle state, or if this
2779 * is an offline or the timekeeping CPU, nothing to do.
2781 if (!*isidle || rdp->rsp != rcu_state_p ||
2782 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2784 /* Verify affinity of current kthread. */
2785 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2787 /* Pick up current idle and NMI-nesting counter and check. */
2788 cur = atomic_read(&rdtp->dynticks_idle);
2790 *isidle = false; /* We are not idle! */
2793 smp_mb(); /* Read counters before timestamps. */
2795 /* Pick up timestamps. */
2796 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2797 /* If this CPU entered idle more recently, update maxj timestamp. */
2798 if (ULONG_CMP_LT(*maxj, j))
2803 * Is this the flavor of RCU that is handling full-system idle?
2805 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2807 return rsp == rcu_state_p;
2811 * Return a delay in jiffies based on the number of CPUs, rcu_node
2812 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2813 * systems more time to transition to full-idle state in order to
2814 * avoid the cache thrashing that otherwise occur on the state variable.
2815 * Really small systems (less than a couple of tens of CPUs) should
2816 * instead use a single global atomically incremented counter, and later
2817 * versions of this will automatically reconfigure themselves accordingly.
2819 static unsigned long rcu_sysidle_delay(void)
2821 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2823 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2827 * Advance the full-system-idle state. This is invoked when all of
2828 * the non-timekeeping CPUs are idle.
2830 static void rcu_sysidle(unsigned long j)
2832 /* Check the current state. */
2833 switch (READ_ONCE(full_sysidle_state)) {
2834 case RCU_SYSIDLE_NOT:
2836 /* First time all are idle, so note a short idle period. */
2837 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2840 case RCU_SYSIDLE_SHORT:
2843 * Idle for a bit, time to advance to next state?
2844 * cmpxchg failure means race with non-idle, let them win.
2846 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2847 (void)cmpxchg(&full_sysidle_state,
2848 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2851 case RCU_SYSIDLE_LONG:
2854 * Do an additional check pass before advancing to full.
2855 * cmpxchg failure means race with non-idle, let them win.
2857 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2858 (void)cmpxchg(&full_sysidle_state,
2859 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2868 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2869 * back to the beginning.
2871 static void rcu_sysidle_cancel(void)
2874 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2875 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2879 * Update the sysidle state based on the results of a force-quiescent-state
2880 * scan of the CPUs' dyntick-idle state.
2882 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2883 unsigned long maxj, bool gpkt)
2885 if (rsp != rcu_state_p)
2886 return; /* Wrong flavor, ignore. */
2887 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2888 return; /* Running state machine from timekeeping CPU. */
2890 rcu_sysidle(maxj); /* More idle! */
2892 rcu_sysidle_cancel(); /* Idle is over. */
2896 * Wrapper for rcu_sysidle_report() when called from the grace-period
2897 * kthread's context.
2899 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2902 /* If there are no nohz_full= CPUs, no need to track this. */
2903 if (!tick_nohz_full_enabled())
2906 rcu_sysidle_report(rsp, isidle, maxj, true);
2909 /* Callback and function for forcing an RCU grace period. */
2910 struct rcu_sysidle_head {
2915 static void rcu_sysidle_cb(struct rcu_head *rhp)
2917 struct rcu_sysidle_head *rshp;
2920 * The following memory barrier is needed to replace the
2921 * memory barriers that would normally be in the memory
2924 smp_mb(); /* grace period precedes setting inuse. */
2926 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2927 WRITE_ONCE(rshp->inuse, 0);
2931 * Check to see if the system is fully idle, other than the timekeeping CPU.
2932 * The caller must have disabled interrupts. This is not intended to be
2933 * called unless tick_nohz_full_enabled().
2935 bool rcu_sys_is_idle(void)
2937 static struct rcu_sysidle_head rsh;
2938 int rss = READ_ONCE(full_sysidle_state);
2940 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2943 /* Handle small-system case by doing a full scan of CPUs. */
2944 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2945 int oldrss = rss - 1;
2948 * One pass to advance to each state up to _FULL.
2949 * Give up if any pass fails to advance the state.
2951 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2954 unsigned long maxj = jiffies - ULONG_MAX / 4;
2955 struct rcu_data *rdp;
2957 /* Scan all the CPUs looking for nonidle CPUs. */
2958 for_each_possible_cpu(cpu) {
2959 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2960 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2964 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2966 rss = READ_ONCE(full_sysidle_state);
2970 /* If this is the first observation of an idle period, record it. */
2971 if (rss == RCU_SYSIDLE_FULL) {
2972 rss = cmpxchg(&full_sysidle_state,
2973 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2974 return rss == RCU_SYSIDLE_FULL;
2977 smp_mb(); /* ensure rss load happens before later caller actions. */
2979 /* If already fully idle, tell the caller (in case of races). */
2980 if (rss == RCU_SYSIDLE_FULL_NOTED)
2984 * If we aren't there yet, and a grace period is not in flight,
2985 * initiate a grace period. Either way, tell the caller that
2986 * we are not there yet. We use an xchg() rather than an assignment
2987 * to make up for the memory barriers that would otherwise be
2988 * provided by the memory allocator.
2990 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2991 !rcu_gp_in_progress(rcu_state_p) &&
2992 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2993 call_rcu(&rsh.rh, rcu_sysidle_cb);
2998 * Initialize dynticks sysidle state for CPUs coming online.
3000 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3002 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
3005 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3007 static void rcu_sysidle_enter(int irq)
3011 static void rcu_sysidle_exit(int irq)
3015 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3016 unsigned long *maxj)
3020 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3025 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3030 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3034 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3037 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3038 * grace-period kthread will do force_quiescent_state() processing?
3039 * The idea is to avoid waking up RCU core processing on such a
3040 * CPU unless the grace period has extended for too long.
3042 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3043 * CONFIG_RCU_NOCB_CPU CPUs.
3045 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3047 #ifdef CONFIG_NO_HZ_FULL
3048 if (tick_nohz_full_cpu(smp_processor_id()) &&
3049 (!rcu_gp_in_progress(rsp) ||
3050 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
3052 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3057 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3060 static void rcu_bind_gp_kthread(void)
3062 int __maybe_unused cpu;
3064 if (!tick_nohz_full_enabled())
3066 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3067 cpu = tick_do_timer_cpu;
3068 if (cpu >= 0 && cpu < nr_cpu_ids)
3069 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3070 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3071 housekeeping_affine(current);
3072 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3075 /* Record the current task on dyntick-idle entry. */
3076 static void rcu_dynticks_task_enter(void)
3078 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3079 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
3080 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3083 /* Record no current task on dyntick-idle exit. */
3084 static void rcu_dynticks_task_exit(void)
3086 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3087 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
3088 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */