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 #else /* #ifdef CONFIG_RCU_BOOST */
49 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
50 * all uses are in dead code. Provide a definition to keep the compiler
51 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
52 * This probably needs to be excluded from -rt builds.
54 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
56 #endif /* #else #ifdef CONFIG_RCU_BOOST */
58 #ifdef CONFIG_RCU_NOCB_CPU
59 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
60 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
61 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
62 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
65 * Check the RCU kernel configuration parameters and print informative
66 * messages about anything out of the ordinary. If you like #ifdef, you
67 * will love this function.
69 static void __init rcu_bootup_announce_oddness(void)
71 if (IS_ENABLED(CONFIG_RCU_TRACE))
72 pr_info("\tRCU debugfs-based tracing is enabled.\n");
73 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
74 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
75 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
78 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
79 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
80 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
81 if (IS_ENABLED(CONFIG_PROVE_RCU))
82 pr_info("\tRCU lockdep checking is enabled.\n");
83 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE))
84 pr_info("\tRCU torture testing starts during boot.\n");
85 if (IS_ENABLED(CONFIG_RCU_CPU_STALL_INFO))
86 pr_info("\tAdditional per-CPU info printed with stalls.\n");
87 if (RCU_NUM_LVLS >= 4)
88 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
89 if (RCU_FANOUT_LEAF != 16)
90 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
92 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
93 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
94 if (nr_cpu_ids != NR_CPUS)
95 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
96 if (IS_ENABLED(CONFIG_RCU_BOOST))
97 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
100 #ifdef CONFIG_PREEMPT_RCU
102 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
103 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
104 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
106 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
107 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
111 * Tell them what RCU they are running.
113 static void __init rcu_bootup_announce(void)
115 pr_info("Preemptible hierarchical RCU implementation.\n");
116 rcu_bootup_announce_oddness();
120 * Record a preemptible-RCU quiescent state for the specified CPU. Note
121 * that this just means that the task currently running on the CPU is
122 * not in a quiescent state. There might be any number of tasks blocked
123 * while in an RCU read-side critical section.
125 * As with the other rcu_*_qs() functions, callers to this function
126 * must disable preemption.
128 static void rcu_preempt_qs(void)
130 if (!__this_cpu_read(rcu_data_p->passed_quiesce)) {
131 trace_rcu_grace_period(TPS("rcu_preempt"),
132 __this_cpu_read(rcu_data_p->gpnum),
134 __this_cpu_write(rcu_data_p->passed_quiesce, 1);
135 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
136 current->rcu_read_unlock_special.b.need_qs = false;
141 * We have entered the scheduler, and the current task might soon be
142 * context-switched away from. If this task is in an RCU read-side
143 * critical section, we will no longer be able to rely on the CPU to
144 * record that fact, so we enqueue the task on the blkd_tasks list.
145 * The task will dequeue itself when it exits the outermost enclosing
146 * RCU read-side critical section. Therefore, the current grace period
147 * cannot be permitted to complete until the blkd_tasks list entries
148 * predating the current grace period drain, in other words, until
149 * rnp->gp_tasks becomes NULL.
151 * Caller must disable preemption.
153 static void rcu_preempt_note_context_switch(void)
155 struct task_struct *t = current;
157 struct rcu_data *rdp;
158 struct rcu_node *rnp;
160 if (t->rcu_read_lock_nesting > 0 &&
161 !t->rcu_read_unlock_special.b.blocked) {
163 /* Possibly blocking in an RCU read-side critical section. */
164 rdp = this_cpu_ptr(rcu_state_p->rda);
166 raw_spin_lock_irqsave(&rnp->lock, flags);
167 smp_mb__after_unlock_lock();
168 t->rcu_read_unlock_special.b.blocked = true;
169 t->rcu_blocked_node = rnp;
172 * If this CPU has already checked in, then this task
173 * will hold up the next grace period rather than the
174 * current grace period. Queue the task accordingly.
175 * If the task is queued for the current grace period
176 * (i.e., this CPU has not yet passed through a quiescent
177 * state for the current grace period), then as long
178 * as that task remains queued, the current grace period
179 * cannot end. Note that there is some uncertainty as
180 * to exactly when the current grace period started.
181 * We take a conservative approach, which can result
182 * in unnecessarily waiting on tasks that started very
183 * slightly after the current grace period began. C'est
186 * But first, note that the current CPU must still be
189 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
190 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
191 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
192 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
193 rnp->gp_tasks = &t->rcu_node_entry;
194 if (IS_ENABLED(CONFIG_RCU_BOOST) &&
195 rnp->boost_tasks != NULL)
196 rnp->boost_tasks = rnp->gp_tasks;
198 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
199 if (rnp->qsmask & rdp->grpmask)
200 rnp->gp_tasks = &t->rcu_node_entry;
202 trace_rcu_preempt_task(rdp->rsp->name,
204 (rnp->qsmask & rdp->grpmask)
207 raw_spin_unlock_irqrestore(&rnp->lock, flags);
208 } else if (t->rcu_read_lock_nesting < 0 &&
209 t->rcu_read_unlock_special.s) {
212 * Complete exit from RCU read-side critical section on
213 * behalf of preempted instance of __rcu_read_unlock().
215 rcu_read_unlock_special(t);
219 * Either we were not in an RCU read-side critical section to
220 * begin with, or we have now recorded that critical section
221 * globally. Either way, we can now note a quiescent state
222 * for this CPU. Again, if we were in an RCU read-side critical
223 * section, and if that critical section was blocking the current
224 * grace period, then the fact that the task has been enqueued
225 * means that we continue to block the current grace period.
231 * Check for preempted RCU readers blocking the current grace period
232 * for the specified rcu_node structure. If the caller needs a reliable
233 * answer, it must hold the rcu_node's ->lock.
235 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
237 return rnp->gp_tasks != NULL;
241 * Advance a ->blkd_tasks-list pointer to the next entry, instead
242 * returning NULL if at the end of the list.
244 static struct list_head *rcu_next_node_entry(struct task_struct *t,
245 struct rcu_node *rnp)
247 struct list_head *np;
249 np = t->rcu_node_entry.next;
250 if (np == &rnp->blkd_tasks)
256 * Return true if the specified rcu_node structure has tasks that were
257 * preempted within an RCU read-side critical section.
259 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
261 return !list_empty(&rnp->blkd_tasks);
265 * Handle special cases during rcu_read_unlock(), such as needing to
266 * notify RCU core processing or task having blocked during the RCU
267 * read-side critical section.
269 void rcu_read_unlock_special(struct task_struct *t)
275 struct list_head *np;
276 bool drop_boost_mutex = false;
277 struct rcu_node *rnp;
278 union rcu_special special;
280 /* NMI handlers cannot block and cannot safely manipulate state. */
284 local_irq_save(flags);
287 * If RCU core is waiting for this CPU to exit critical section,
288 * let it know that we have done so. Because irqs are disabled,
289 * t->rcu_read_unlock_special cannot change.
291 special = t->rcu_read_unlock_special;
292 if (special.b.need_qs) {
294 t->rcu_read_unlock_special.b.need_qs = false;
295 if (!t->rcu_read_unlock_special.s) {
296 local_irq_restore(flags);
301 /* Hardware IRQ handlers cannot block, complain if they get here. */
302 if (in_irq() || in_serving_softirq()) {
303 lockdep_rcu_suspicious(__FILE__, __LINE__,
304 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
305 pr_alert("->rcu_read_unlock_special: %#x (b: %d, nq: %d)\n",
306 t->rcu_read_unlock_special.s,
307 t->rcu_read_unlock_special.b.blocked,
308 t->rcu_read_unlock_special.b.need_qs);
309 local_irq_restore(flags);
313 /* Clean up if blocked during RCU read-side critical section. */
314 if (special.b.blocked) {
315 t->rcu_read_unlock_special.b.blocked = false;
318 * Remove this task from the list it blocked on. The task
319 * now remains queued on the rcu_node corresponding to
320 * the CPU it first blocked on, so the first attempt to
321 * acquire the task's rcu_node's ->lock will succeed.
322 * Keep the loop and add a WARN_ON() out of sheer paranoia.
325 rnp = t->rcu_blocked_node;
326 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
327 smp_mb__after_unlock_lock();
328 if (rnp == t->rcu_blocked_node)
331 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
333 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
334 empty_exp = !rcu_preempted_readers_exp(rnp);
335 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
336 np = rcu_next_node_entry(t, rnp);
337 list_del_init(&t->rcu_node_entry);
338 t->rcu_blocked_node = NULL;
339 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
341 if (&t->rcu_node_entry == rnp->gp_tasks)
343 if (&t->rcu_node_entry == rnp->exp_tasks)
345 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
346 if (&t->rcu_node_entry == rnp->boost_tasks)
347 rnp->boost_tasks = np;
348 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
349 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
353 * If this was the last task on the current list, and if
354 * we aren't waiting on any CPUs, report the quiescent state.
355 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
356 * so we must take a snapshot of the expedited state.
358 empty_exp_now = !rcu_preempted_readers_exp(rnp);
359 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
360 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
367 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
369 raw_spin_unlock_irqrestore(&rnp->lock, flags);
372 /* Unboost if we were boosted. */
373 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
374 rt_mutex_unlock(&rnp->boost_mtx);
377 * If this was the last task on the expedited lists,
378 * then we need to report up the rcu_node hierarchy.
380 if (!empty_exp && empty_exp_now)
381 rcu_report_exp_rnp(rcu_state_p, rnp, true);
383 local_irq_restore(flags);
388 * Dump detailed information for all tasks blocking the current RCU
389 * grace period on the specified rcu_node structure.
391 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
394 struct task_struct *t;
396 raw_spin_lock_irqsave(&rnp->lock, flags);
397 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
398 raw_spin_unlock_irqrestore(&rnp->lock, flags);
401 t = list_entry(rnp->gp_tasks->prev,
402 struct task_struct, rcu_node_entry);
403 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
405 raw_spin_unlock_irqrestore(&rnp->lock, flags);
409 * Dump detailed information for all tasks blocking the current RCU
412 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
414 struct rcu_node *rnp = rcu_get_root(rsp);
416 rcu_print_detail_task_stall_rnp(rnp);
417 rcu_for_each_leaf_node(rsp, rnp)
418 rcu_print_detail_task_stall_rnp(rnp);
421 #ifdef CONFIG_RCU_CPU_STALL_INFO
423 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
425 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
426 rnp->level, rnp->grplo, rnp->grphi);
429 static void rcu_print_task_stall_end(void)
434 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
436 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
440 static void rcu_print_task_stall_end(void)
444 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
447 * Scan the current list of tasks blocked within RCU read-side critical
448 * sections, printing out the tid of each.
450 static int rcu_print_task_stall(struct rcu_node *rnp)
452 struct task_struct *t;
455 if (!rcu_preempt_blocked_readers_cgp(rnp))
457 rcu_print_task_stall_begin(rnp);
458 t = list_entry(rnp->gp_tasks->prev,
459 struct task_struct, rcu_node_entry);
460 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
461 pr_cont(" P%d", t->pid);
464 rcu_print_task_stall_end();
469 * Check that the list of blocked tasks for the newly completed grace
470 * period is in fact empty. It is a serious bug to complete a grace
471 * period that still has RCU readers blocked! This function must be
472 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
473 * must be held by the caller.
475 * Also, if there are blocked tasks on the list, they automatically
476 * block the newly created grace period, so set up ->gp_tasks accordingly.
478 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
480 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
481 if (rcu_preempt_has_tasks(rnp))
482 rnp->gp_tasks = rnp->blkd_tasks.next;
483 WARN_ON_ONCE(rnp->qsmask);
487 * Check for a quiescent state from the current CPU. When a task blocks,
488 * the task is recorded in the corresponding CPU's rcu_node structure,
489 * which is checked elsewhere.
491 * Caller must disable hard irqs.
493 static void rcu_preempt_check_callbacks(void)
495 struct task_struct *t = current;
497 if (t->rcu_read_lock_nesting == 0) {
501 if (t->rcu_read_lock_nesting > 0 &&
502 __this_cpu_read(rcu_data_p->qs_pending) &&
503 !__this_cpu_read(rcu_data_p->passed_quiesce))
504 t->rcu_read_unlock_special.b.need_qs = true;
507 #ifdef CONFIG_RCU_BOOST
509 static void rcu_preempt_do_callbacks(void)
511 rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p));
514 #endif /* #ifdef CONFIG_RCU_BOOST */
517 * Queue a preemptible-RCU callback for invocation after a grace period.
519 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
521 __call_rcu(head, func, rcu_state_p, -1, 0);
523 EXPORT_SYMBOL_GPL(call_rcu);
526 * synchronize_rcu - wait until a grace period has elapsed.
528 * Control will return to the caller some time after a full grace
529 * period has elapsed, in other words after all currently executing RCU
530 * read-side critical sections have completed. Note, however, that
531 * upon return from synchronize_rcu(), the caller might well be executing
532 * concurrently with new RCU read-side critical sections that began while
533 * synchronize_rcu() was waiting. RCU read-side critical sections are
534 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
536 * See the description of synchronize_sched() for more detailed information
537 * on memory ordering guarantees.
539 void synchronize_rcu(void)
541 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
542 !lock_is_held(&rcu_lock_map) &&
543 !lock_is_held(&rcu_sched_lock_map),
544 "Illegal synchronize_rcu() in RCU read-side critical section");
545 if (!rcu_scheduler_active)
547 if (rcu_gp_is_expedited())
548 synchronize_rcu_expedited();
550 wait_rcu_gp(call_rcu);
552 EXPORT_SYMBOL_GPL(synchronize_rcu);
554 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
555 static unsigned long sync_rcu_preempt_exp_count;
556 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
559 * Return non-zero if there are any tasks in RCU read-side critical
560 * sections blocking the current preemptible-RCU expedited grace period.
561 * If there is no preemptible-RCU expedited grace period currently in
562 * progress, returns zero unconditionally.
564 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
566 return rnp->exp_tasks != NULL;
570 * return non-zero if there is no RCU expedited grace period in progress
571 * for the specified rcu_node structure, in other words, if all CPUs and
572 * tasks covered by the specified rcu_node structure have done their bit
573 * for the current expedited grace period. Works only for preemptible
574 * RCU -- other RCU implementation use other means.
576 * Caller must hold sync_rcu_preempt_exp_mutex.
578 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
580 return !rcu_preempted_readers_exp(rnp) &&
581 READ_ONCE(rnp->expmask) == 0;
585 * Report the exit from RCU read-side critical section for the last task
586 * that queued itself during or before the current expedited preemptible-RCU
587 * grace period. This event is reported either to the rcu_node structure on
588 * which the task was queued or to one of that rcu_node structure's ancestors,
589 * recursively up the tree. (Calm down, calm down, we do the recursion
592 * Caller must hold sync_rcu_preempt_exp_mutex.
594 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
600 raw_spin_lock_irqsave(&rnp->lock, flags);
601 smp_mb__after_unlock_lock();
603 if (!sync_rcu_preempt_exp_done(rnp)) {
604 raw_spin_unlock_irqrestore(&rnp->lock, flags);
607 if (rnp->parent == NULL) {
608 raw_spin_unlock_irqrestore(&rnp->lock, flags);
610 smp_mb(); /* EGP done before wake_up(). */
611 wake_up(&sync_rcu_preempt_exp_wq);
616 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
618 raw_spin_lock(&rnp->lock); /* irqs already disabled */
619 smp_mb__after_unlock_lock();
620 rnp->expmask &= ~mask;
625 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
626 * grace period for the specified rcu_node structure, phase 1. If there
627 * are such tasks, set the ->expmask bits up the rcu_node tree and also
628 * set the ->expmask bits on the leaf rcu_node structures to tell phase 2
629 * that work is needed here.
631 * Caller must hold sync_rcu_preempt_exp_mutex.
634 sync_rcu_preempt_exp_init1(struct rcu_state *rsp, struct rcu_node *rnp)
638 struct rcu_node *rnp_up;
640 raw_spin_lock_irqsave(&rnp->lock, flags);
641 smp_mb__after_unlock_lock();
642 WARN_ON_ONCE(rnp->expmask);
643 WARN_ON_ONCE(rnp->exp_tasks);
644 if (!rcu_preempt_has_tasks(rnp)) {
645 /* No blocked tasks, nothing to do. */
646 raw_spin_unlock_irqrestore(&rnp->lock, flags);
649 /* Call for Phase 2 and propagate ->expmask bits up the tree. */
652 while (rnp_up->parent) {
653 mask = rnp_up->grpmask;
654 rnp_up = rnp_up->parent;
655 if (rnp_up->expmask & mask)
657 raw_spin_lock(&rnp_up->lock); /* irqs already off */
658 smp_mb__after_unlock_lock();
659 rnp_up->expmask |= mask;
660 raw_spin_unlock(&rnp_up->lock); /* irqs still off */
662 raw_spin_unlock_irqrestore(&rnp->lock, flags);
666 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
667 * grace period for the specified rcu_node structure, phase 2. If the
668 * leaf rcu_node structure has its ->expmask field set, check for tasks.
669 * If there are some, clear ->expmask and set ->exp_tasks accordingly,
670 * then initiate RCU priority boosting. Otherwise, clear ->expmask and
671 * invoke rcu_report_exp_rnp() to clear out the upper-level ->expmask bits,
672 * enabling rcu_read_unlock_special() to do the bit-clearing.
674 * Caller must hold sync_rcu_preempt_exp_mutex.
677 sync_rcu_preempt_exp_init2(struct rcu_state *rsp, struct rcu_node *rnp)
681 raw_spin_lock_irqsave(&rnp->lock, flags);
682 smp_mb__after_unlock_lock();
684 /* Phase 1 didn't do anything, so Phase 2 doesn't either. */
685 raw_spin_unlock_irqrestore(&rnp->lock, flags);
689 /* Phase 1 is over. */
693 * If there are still blocked tasks, set up ->exp_tasks so that
694 * rcu_read_unlock_special() will wake us and then boost them.
696 if (rcu_preempt_has_tasks(rnp)) {
697 rnp->exp_tasks = rnp->blkd_tasks.next;
698 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
702 /* No longer any blocked tasks, so undo bit setting. */
703 raw_spin_unlock_irqrestore(&rnp->lock, flags);
704 rcu_report_exp_rnp(rsp, rnp, false);
708 * synchronize_rcu_expedited - Brute-force RCU grace period
710 * Wait for an RCU-preempt grace period, but expedite it. The basic
711 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
712 * the ->blkd_tasks lists and wait for this list to drain. This consumes
713 * significant time on all CPUs and is unfriendly to real-time workloads,
714 * so is thus not recommended for any sort of common-case code.
715 * In fact, if you are using synchronize_rcu_expedited() in a loop,
716 * please restructure your code to batch your updates, and then Use a
717 * single synchronize_rcu() instead.
719 void synchronize_rcu_expedited(void)
721 struct rcu_node *rnp;
722 struct rcu_state *rsp = rcu_state_p;
726 smp_mb(); /* Caller's modifications seen first by other CPUs. */
727 snap = READ_ONCE(sync_rcu_preempt_exp_count) + 1;
728 smp_mb(); /* Above access cannot bleed into critical section. */
731 * Acquire lock, falling back to synchronize_rcu() if too many
732 * lock-acquisition failures. Of course, if someone does the
733 * expedited grace period for us, just leave.
735 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
736 if (ULONG_CMP_LT(snap,
737 READ_ONCE(sync_rcu_preempt_exp_count)))
738 goto mb_ret; /* Others did our work for us. */
739 if (trycount++ < 10) {
740 udelay(trycount * num_online_cpus());
742 wait_rcu_gp(call_rcu);
746 if (ULONG_CMP_LT(snap, READ_ONCE(sync_rcu_preempt_exp_count)))
747 goto unlock_mb_ret; /* Others did our work for us. */
749 /* force all RCU readers onto ->blkd_tasks lists. */
750 synchronize_sched_expedited();
753 * Snapshot current state of ->blkd_tasks lists into ->expmask.
754 * Phase 1 sets bits and phase 2 permits rcu_read_unlock_special()
755 * to start clearing them. Doing this in one phase leads to
756 * strange races between setting and clearing bits, so just say "no"!
758 rcu_for_each_leaf_node(rsp, rnp)
759 sync_rcu_preempt_exp_init1(rsp, rnp);
760 rcu_for_each_leaf_node(rsp, rnp)
761 sync_rcu_preempt_exp_init2(rsp, rnp);
763 /* Wait for snapshotted ->blkd_tasks lists to drain. */
764 rnp = rcu_get_root(rsp);
765 wait_event(sync_rcu_preempt_exp_wq,
766 sync_rcu_preempt_exp_done(rnp));
768 /* Clean up and exit. */
769 smp_mb(); /* ensure expedited GP seen before counter increment. */
770 WRITE_ONCE(sync_rcu_preempt_exp_count, sync_rcu_preempt_exp_count + 1);
772 mutex_unlock(&sync_rcu_preempt_exp_mutex);
774 smp_mb(); /* ensure subsequent action seen after grace period. */
776 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
779 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
781 * Note that this primitive does not necessarily wait for an RCU grace period
782 * to complete. For example, if there are no RCU callbacks queued anywhere
783 * in the system, then rcu_barrier() is within its rights to return
784 * immediately, without waiting for anything, much less an RCU grace period.
786 void rcu_barrier(void)
788 _rcu_barrier(rcu_state_p);
790 EXPORT_SYMBOL_GPL(rcu_barrier);
793 * Initialize preemptible RCU's state structures.
795 static void __init __rcu_init_preempt(void)
797 rcu_init_one(rcu_state_p, rcu_data_p);
801 * Check for a task exiting while in a preemptible-RCU read-side
802 * critical section, clean up if so. No need to issue warnings,
803 * as debug_check_no_locks_held() already does this if lockdep
808 struct task_struct *t = current;
810 if (likely(list_empty(¤t->rcu_node_entry)))
812 t->rcu_read_lock_nesting = 1;
814 t->rcu_read_unlock_special.b.blocked = true;
818 #else /* #ifdef CONFIG_PREEMPT_RCU */
820 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
821 static struct rcu_data __percpu *const rcu_data_p = &rcu_sched_data;
824 * Tell them what RCU they are running.
826 static void __init rcu_bootup_announce(void)
828 pr_info("Hierarchical RCU implementation.\n");
829 rcu_bootup_announce_oddness();
833 * Because preemptible RCU does not exist, we never have to check for
834 * CPUs being in quiescent states.
836 static void rcu_preempt_note_context_switch(void)
841 * Because preemptible RCU does not exist, there are never any preempted
844 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
850 * Because there is no preemptible RCU, there can be no readers blocked.
852 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
858 * Because preemptible RCU does not exist, we never have to check for
859 * tasks blocked within RCU read-side critical sections.
861 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
866 * Because preemptible RCU does not exist, we never have to check for
867 * tasks blocked within RCU read-side critical sections.
869 static int rcu_print_task_stall(struct rcu_node *rnp)
875 * Because there is no preemptible RCU, there can be no readers blocked,
876 * so there is no need to check for blocked tasks. So check only for
877 * bogus qsmask values.
879 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
881 WARN_ON_ONCE(rnp->qsmask);
885 * Because preemptible RCU does not exist, it never has any callbacks
888 static void rcu_preempt_check_callbacks(void)
893 * Wait for an rcu-preempt grace period, but make it happen quickly.
894 * But because preemptible RCU does not exist, map to rcu-sched.
896 void synchronize_rcu_expedited(void)
898 synchronize_sched_expedited();
900 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
903 * Because preemptible RCU does not exist, rcu_barrier() is just
904 * another name for rcu_barrier_sched().
906 void rcu_barrier(void)
910 EXPORT_SYMBOL_GPL(rcu_barrier);
913 * Because preemptible RCU does not exist, it need not be initialized.
915 static void __init __rcu_init_preempt(void)
920 * Because preemptible RCU does not exist, tasks cannot possibly exit
921 * while in preemptible RCU read-side critical sections.
927 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
929 #ifdef CONFIG_RCU_BOOST
931 #include "../locking/rtmutex_common.h"
933 #ifdef CONFIG_RCU_TRACE
935 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
937 if (!rcu_preempt_has_tasks(rnp))
938 rnp->n_balk_blkd_tasks++;
939 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
940 rnp->n_balk_exp_gp_tasks++;
941 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
942 rnp->n_balk_boost_tasks++;
943 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
944 rnp->n_balk_notblocked++;
945 else if (rnp->gp_tasks != NULL &&
946 ULONG_CMP_LT(jiffies, rnp->boost_time))
947 rnp->n_balk_notyet++;
952 #else /* #ifdef CONFIG_RCU_TRACE */
954 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
958 #endif /* #else #ifdef CONFIG_RCU_TRACE */
960 static void rcu_wake_cond(struct task_struct *t, int status)
963 * If the thread is yielding, only wake it when this
964 * is invoked from idle
966 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
971 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
972 * or ->boost_tasks, advancing the pointer to the next task in the
975 * Note that irqs must be enabled: boosting the task can block.
976 * Returns 1 if there are more tasks needing to be boosted.
978 static int rcu_boost(struct rcu_node *rnp)
981 struct task_struct *t;
982 struct list_head *tb;
984 if (READ_ONCE(rnp->exp_tasks) == NULL &&
985 READ_ONCE(rnp->boost_tasks) == NULL)
986 return 0; /* Nothing left to boost. */
988 raw_spin_lock_irqsave(&rnp->lock, flags);
989 smp_mb__after_unlock_lock();
992 * Recheck under the lock: all tasks in need of boosting
993 * might exit their RCU read-side critical sections on their own.
995 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
996 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1001 * Preferentially boost tasks blocking expedited grace periods.
1002 * This cannot starve the normal grace periods because a second
1003 * expedited grace period must boost all blocked tasks, including
1004 * those blocking the pre-existing normal grace period.
1006 if (rnp->exp_tasks != NULL) {
1007 tb = rnp->exp_tasks;
1008 rnp->n_exp_boosts++;
1010 tb = rnp->boost_tasks;
1011 rnp->n_normal_boosts++;
1013 rnp->n_tasks_boosted++;
1016 * We boost task t by manufacturing an rt_mutex that appears to
1017 * be held by task t. We leave a pointer to that rt_mutex where
1018 * task t can find it, and task t will release the mutex when it
1019 * exits its outermost RCU read-side critical section. Then
1020 * simply acquiring this artificial rt_mutex will boost task
1021 * t's priority. (Thanks to tglx for suggesting this approach!)
1023 * Note that task t must acquire rnp->lock to remove itself from
1024 * the ->blkd_tasks list, which it will do from exit() if from
1025 * nowhere else. We therefore are guaranteed that task t will
1026 * stay around at least until we drop rnp->lock. Note that
1027 * rnp->lock also resolves races between our priority boosting
1028 * and task t's exiting its outermost RCU read-side critical
1031 t = container_of(tb, struct task_struct, rcu_node_entry);
1032 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1033 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1034 /* Lock only for side effect: boosts task t's priority. */
1035 rt_mutex_lock(&rnp->boost_mtx);
1036 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1038 return READ_ONCE(rnp->exp_tasks) != NULL ||
1039 READ_ONCE(rnp->boost_tasks) != NULL;
1043 * Priority-boosting kthread. One per leaf rcu_node and one for the
1046 static int rcu_boost_kthread(void *arg)
1048 struct rcu_node *rnp = (struct rcu_node *)arg;
1052 trace_rcu_utilization(TPS("Start boost kthread@init"));
1054 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1055 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1056 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1057 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1058 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1059 more2boost = rcu_boost(rnp);
1065 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1066 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1067 schedule_timeout_interruptible(2);
1068 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1073 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1078 * Check to see if it is time to start boosting RCU readers that are
1079 * blocking the current grace period, and, if so, tell the per-rcu_node
1080 * kthread to start boosting them. If there is an expedited grace
1081 * period in progress, it is always time to boost.
1083 * The caller must hold rnp->lock, which this function releases.
1084 * The ->boost_kthread_task is immortal, so we don't need to worry
1085 * about it going away.
1087 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1088 __releases(rnp->lock)
1090 struct task_struct *t;
1092 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1093 rnp->n_balk_exp_gp_tasks++;
1094 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1097 if (rnp->exp_tasks != NULL ||
1098 (rnp->gp_tasks != NULL &&
1099 rnp->boost_tasks == NULL &&
1101 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1102 if (rnp->exp_tasks == NULL)
1103 rnp->boost_tasks = rnp->gp_tasks;
1104 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1105 t = rnp->boost_kthread_task;
1107 rcu_wake_cond(t, rnp->boost_kthread_status);
1109 rcu_initiate_boost_trace(rnp);
1110 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1115 * Wake up the per-CPU kthread to invoke RCU callbacks.
1117 static void invoke_rcu_callbacks_kthread(void)
1119 unsigned long flags;
1121 local_irq_save(flags);
1122 __this_cpu_write(rcu_cpu_has_work, 1);
1123 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1124 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1125 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1126 __this_cpu_read(rcu_cpu_kthread_status));
1128 local_irq_restore(flags);
1132 * Is the current CPU running the RCU-callbacks kthread?
1133 * Caller must have preemption disabled.
1135 static bool rcu_is_callbacks_kthread(void)
1137 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1140 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1143 * Do priority-boost accounting for the start of a new grace period.
1145 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1147 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1151 * Create an RCU-boost kthread for the specified node if one does not
1152 * already exist. We only create this kthread for preemptible RCU.
1153 * Returns zero if all is well, a negated errno otherwise.
1155 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1156 struct rcu_node *rnp)
1158 int rnp_index = rnp - &rsp->node[0];
1159 unsigned long flags;
1160 struct sched_param sp;
1161 struct task_struct *t;
1163 if (rcu_state_p != rsp)
1166 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1170 if (rnp->boost_kthread_task != NULL)
1172 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1173 "rcub/%d", rnp_index);
1176 raw_spin_lock_irqsave(&rnp->lock, flags);
1177 smp_mb__after_unlock_lock();
1178 rnp->boost_kthread_task = t;
1179 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1180 sp.sched_priority = kthread_prio;
1181 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1182 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1186 static void rcu_kthread_do_work(void)
1188 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1189 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1190 rcu_preempt_do_callbacks();
1193 static void rcu_cpu_kthread_setup(unsigned int cpu)
1195 struct sched_param sp;
1197 sp.sched_priority = kthread_prio;
1198 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1201 static void rcu_cpu_kthread_park(unsigned int cpu)
1203 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1206 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1208 return __this_cpu_read(rcu_cpu_has_work);
1212 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1213 * RCU softirq used in flavors and configurations of RCU that do not
1214 * support RCU priority boosting.
1216 static void rcu_cpu_kthread(unsigned int cpu)
1218 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1219 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1222 for (spincnt = 0; spincnt < 10; spincnt++) {
1223 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1225 *statusp = RCU_KTHREAD_RUNNING;
1226 this_cpu_inc(rcu_cpu_kthread_loops);
1227 local_irq_disable();
1232 rcu_kthread_do_work();
1235 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1236 *statusp = RCU_KTHREAD_WAITING;
1240 *statusp = RCU_KTHREAD_YIELDING;
1241 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1242 schedule_timeout_interruptible(2);
1243 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1244 *statusp = RCU_KTHREAD_WAITING;
1248 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1249 * served by the rcu_node in question. The CPU hotplug lock is still
1250 * held, so the value of rnp->qsmaskinit will be stable.
1252 * We don't include outgoingcpu in the affinity set, use -1 if there is
1253 * no outgoing CPU. If there are no CPUs left in the affinity set,
1254 * this function allows the kthread to execute on any CPU.
1256 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1258 struct task_struct *t = rnp->boost_kthread_task;
1259 unsigned long mask = rcu_rnp_online_cpus(rnp);
1265 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1267 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1268 if ((mask & 0x1) && cpu != outgoingcpu)
1269 cpumask_set_cpu(cpu, cm);
1270 if (cpumask_weight(cm) == 0)
1272 set_cpus_allowed_ptr(t, cm);
1273 free_cpumask_var(cm);
1276 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1277 .store = &rcu_cpu_kthread_task,
1278 .thread_should_run = rcu_cpu_kthread_should_run,
1279 .thread_fn = rcu_cpu_kthread,
1280 .thread_comm = "rcuc/%u",
1281 .setup = rcu_cpu_kthread_setup,
1282 .park = rcu_cpu_kthread_park,
1286 * Spawn boost kthreads -- called as soon as the scheduler is running.
1288 static void __init rcu_spawn_boost_kthreads(void)
1290 struct rcu_node *rnp;
1293 for_each_possible_cpu(cpu)
1294 per_cpu(rcu_cpu_has_work, cpu) = 0;
1295 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1296 rcu_for_each_leaf_node(rcu_state_p, rnp)
1297 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1300 static void rcu_prepare_kthreads(int cpu)
1302 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1303 struct rcu_node *rnp = rdp->mynode;
1305 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1306 if (rcu_scheduler_fully_active)
1307 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1310 #else /* #ifdef CONFIG_RCU_BOOST */
1312 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1313 __releases(rnp->lock)
1315 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1318 static void invoke_rcu_callbacks_kthread(void)
1323 static bool rcu_is_callbacks_kthread(void)
1328 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1332 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1336 static void __init rcu_spawn_boost_kthreads(void)
1340 static void rcu_prepare_kthreads(int cpu)
1344 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1346 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1349 * Check to see if any future RCU-related work will need to be done
1350 * by the current CPU, even if none need be done immediately, returning
1351 * 1 if so. This function is part of the RCU implementation; it is -not-
1352 * an exported member of the RCU API.
1354 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1355 * any flavor of RCU.
1357 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1359 *nextevt = KTIME_MAX;
1360 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1361 ? 0 : rcu_cpu_has_callbacks(NULL);
1365 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1368 static void rcu_cleanup_after_idle(void)
1373 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1376 static void rcu_prepare_for_idle(void)
1381 * Don't bother keeping a running count of the number of RCU callbacks
1382 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1384 static void rcu_idle_count_callbacks_posted(void)
1388 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1391 * This code is invoked when a CPU goes idle, at which point we want
1392 * to have the CPU do everything required for RCU so that it can enter
1393 * the energy-efficient dyntick-idle mode. This is handled by a
1394 * state machine implemented by rcu_prepare_for_idle() below.
1396 * The following three proprocessor symbols control this state machine:
1398 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1399 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1400 * is sized to be roughly one RCU grace period. Those energy-efficiency
1401 * benchmarkers who might otherwise be tempted to set this to a large
1402 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1403 * system. And if you are -that- concerned about energy efficiency,
1404 * just power the system down and be done with it!
1405 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1406 * permitted to sleep in dyntick-idle mode with only lazy RCU
1407 * callbacks pending. Setting this too high can OOM your system.
1409 * The values below work well in practice. If future workloads require
1410 * adjustment, they can be converted into kernel config parameters, though
1411 * making the state machine smarter might be a better option.
1413 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1414 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1416 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1417 module_param(rcu_idle_gp_delay, int, 0644);
1418 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1419 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1422 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1423 * only if it has been awhile since the last time we did so. Afterwards,
1424 * if there are any callbacks ready for immediate invocation, return true.
1426 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1428 bool cbs_ready = false;
1429 struct rcu_data *rdp;
1430 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1431 struct rcu_node *rnp;
1432 struct rcu_state *rsp;
1434 /* Exit early if we advanced recently. */
1435 if (jiffies == rdtp->last_advance_all)
1437 rdtp->last_advance_all = jiffies;
1439 for_each_rcu_flavor(rsp) {
1440 rdp = this_cpu_ptr(rsp->rda);
1444 * Don't bother checking unless a grace period has
1445 * completed since we last checked and there are
1446 * callbacks not yet ready to invoke.
1448 if ((rdp->completed != rnp->completed ||
1449 unlikely(READ_ONCE(rdp->gpwrap))) &&
1450 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1451 note_gp_changes(rsp, rdp);
1453 if (cpu_has_callbacks_ready_to_invoke(rdp))
1460 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1461 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1462 * caller to set the timeout based on whether or not there are non-lazy
1465 * The caller must have disabled interrupts.
1467 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1469 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1472 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1473 *nextevt = KTIME_MAX;
1477 /* Snapshot to detect later posting of non-lazy callback. */
1478 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1480 /* If no callbacks, RCU doesn't need the CPU. */
1481 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1482 *nextevt = KTIME_MAX;
1486 /* Attempt to advance callbacks. */
1487 if (rcu_try_advance_all_cbs()) {
1488 /* Some ready to invoke, so initiate later invocation. */
1492 rdtp->last_accelerate = jiffies;
1494 /* Request timer delay depending on laziness, and round. */
1495 if (!rdtp->all_lazy) {
1496 dj = round_up(rcu_idle_gp_delay + jiffies,
1497 rcu_idle_gp_delay) - jiffies;
1499 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1501 *nextevt = basemono + dj * TICK_NSEC;
1506 * Prepare a CPU for idle from an RCU perspective. The first major task
1507 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1508 * The second major task is to check to see if a non-lazy callback has
1509 * arrived at a CPU that previously had only lazy callbacks. The third
1510 * major task is to accelerate (that is, assign grace-period numbers to)
1511 * any recently arrived callbacks.
1513 * The caller must have disabled interrupts.
1515 static void rcu_prepare_for_idle(void)
1518 struct rcu_data *rdp;
1519 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1520 struct rcu_node *rnp;
1521 struct rcu_state *rsp;
1524 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL))
1527 /* Handle nohz enablement switches conservatively. */
1528 tne = READ_ONCE(tick_nohz_active);
1529 if (tne != rdtp->tick_nohz_enabled_snap) {
1530 if (rcu_cpu_has_callbacks(NULL))
1531 invoke_rcu_core(); /* force nohz to see update. */
1532 rdtp->tick_nohz_enabled_snap = tne;
1538 /* If this is a no-CBs CPU, no callbacks, just return. */
1539 if (rcu_is_nocb_cpu(smp_processor_id()))
1543 * If a non-lazy callback arrived at a CPU having only lazy
1544 * callbacks, invoke RCU core for the side-effect of recalculating
1545 * idle duration on re-entry to idle.
1547 if (rdtp->all_lazy &&
1548 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1549 rdtp->all_lazy = false;
1550 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1556 * If we have not yet accelerated this jiffy, accelerate all
1557 * callbacks on this CPU.
1559 if (rdtp->last_accelerate == jiffies)
1561 rdtp->last_accelerate = jiffies;
1562 for_each_rcu_flavor(rsp) {
1563 rdp = this_cpu_ptr(rsp->rda);
1564 if (!*rdp->nxttail[RCU_DONE_TAIL])
1567 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1568 smp_mb__after_unlock_lock();
1569 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1570 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1572 rcu_gp_kthread_wake(rsp);
1577 * Clean up for exit from idle. Attempt to advance callbacks based on
1578 * any grace periods that elapsed while the CPU was idle, and if any
1579 * callbacks are now ready to invoke, initiate invocation.
1581 static void rcu_cleanup_after_idle(void)
1583 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1584 rcu_is_nocb_cpu(smp_processor_id()))
1586 if (rcu_try_advance_all_cbs())
1591 * Keep a running count of the number of non-lazy callbacks posted
1592 * on this CPU. This running counter (which is never decremented) allows
1593 * rcu_prepare_for_idle() to detect when something out of the idle loop
1594 * posts a callback, even if an equal number of callbacks are invoked.
1595 * Of course, callbacks should only be posted from within a trace event
1596 * designed to be called from idle or from within RCU_NONIDLE().
1598 static void rcu_idle_count_callbacks_posted(void)
1600 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1604 * Data for flushing lazy RCU callbacks at OOM time.
1606 static atomic_t oom_callback_count;
1607 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1610 * RCU OOM callback -- decrement the outstanding count and deliver the
1611 * wake-up if we are the last one.
1613 static void rcu_oom_callback(struct rcu_head *rhp)
1615 if (atomic_dec_and_test(&oom_callback_count))
1616 wake_up(&oom_callback_wq);
1620 * Post an rcu_oom_notify callback on the current CPU if it has at
1621 * least one lazy callback. This will unnecessarily post callbacks
1622 * to CPUs that already have a non-lazy callback at the end of their
1623 * callback list, but this is an infrequent operation, so accept some
1624 * extra overhead to keep things simple.
1626 static void rcu_oom_notify_cpu(void *unused)
1628 struct rcu_state *rsp;
1629 struct rcu_data *rdp;
1631 for_each_rcu_flavor(rsp) {
1632 rdp = raw_cpu_ptr(rsp->rda);
1633 if (rdp->qlen_lazy != 0) {
1634 atomic_inc(&oom_callback_count);
1635 rsp->call(&rdp->oom_head, rcu_oom_callback);
1641 * If low on memory, ensure that each CPU has a non-lazy callback.
1642 * This will wake up CPUs that have only lazy callbacks, in turn
1643 * ensuring that they free up the corresponding memory in a timely manner.
1644 * Because an uncertain amount of memory will be freed in some uncertain
1645 * timeframe, we do not claim to have freed anything.
1647 static int rcu_oom_notify(struct notifier_block *self,
1648 unsigned long notused, void *nfreed)
1652 /* Wait for callbacks from earlier instance to complete. */
1653 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1654 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1657 * Prevent premature wakeup: ensure that all increments happen
1658 * before there is a chance of the counter reaching zero.
1660 atomic_set(&oom_callback_count, 1);
1663 for_each_online_cpu(cpu) {
1664 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1665 cond_resched_rcu_qs();
1669 /* Unconditionally decrement: no need to wake ourselves up. */
1670 atomic_dec(&oom_callback_count);
1675 static struct notifier_block rcu_oom_nb = {
1676 .notifier_call = rcu_oom_notify
1679 static int __init rcu_register_oom_notifier(void)
1681 register_oom_notifier(&rcu_oom_nb);
1684 early_initcall(rcu_register_oom_notifier);
1686 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1688 #ifdef CONFIG_RCU_CPU_STALL_INFO
1690 #ifdef CONFIG_RCU_FAST_NO_HZ
1692 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1694 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1695 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1697 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1698 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1700 rdtp->all_lazy ? 'L' : '.',
1701 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1704 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1706 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1711 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1713 /* Initiate the stall-info list. */
1714 static void print_cpu_stall_info_begin(void)
1720 * Print out diagnostic information for the specified stalled CPU.
1722 * If the specified CPU is aware of the current RCU grace period
1723 * (flavor specified by rsp), then print the number of scheduling
1724 * clock interrupts the CPU has taken during the time that it has
1725 * been aware. Otherwise, print the number of RCU grace periods
1726 * that this CPU is ignorant of, for example, "1" if the CPU was
1727 * aware of the previous grace period.
1729 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1731 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1733 char fast_no_hz[72];
1734 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1735 struct rcu_dynticks *rdtp = rdp->dynticks;
1737 unsigned long ticks_value;
1739 if (rsp->gpnum == rdp->gpnum) {
1740 ticks_title = "ticks this GP";
1741 ticks_value = rdp->ticks_this_gp;
1743 ticks_title = "GPs behind";
1744 ticks_value = rsp->gpnum - rdp->gpnum;
1746 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1747 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1748 cpu, ticks_value, ticks_title,
1749 atomic_read(&rdtp->dynticks) & 0xfff,
1750 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1751 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1752 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1756 /* Terminate the stall-info list. */
1757 static void print_cpu_stall_info_end(void)
1762 /* Zero ->ticks_this_gp for all flavors of RCU. */
1763 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1765 rdp->ticks_this_gp = 0;
1766 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1769 /* Increment ->ticks_this_gp for all flavors of RCU. */
1770 static void increment_cpu_stall_ticks(void)
1772 struct rcu_state *rsp;
1774 for_each_rcu_flavor(rsp)
1775 raw_cpu_inc(rsp->rda->ticks_this_gp);
1778 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1780 static void print_cpu_stall_info_begin(void)
1785 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1787 pr_cont(" %d", cpu);
1790 static void print_cpu_stall_info_end(void)
1795 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1799 static void increment_cpu_stall_ticks(void)
1803 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1805 #ifdef CONFIG_RCU_NOCB_CPU
1808 * Offload callback processing from the boot-time-specified set of CPUs
1809 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1810 * kthread created that pulls the callbacks from the corresponding CPU,
1811 * waits for a grace period to elapse, and invokes the callbacks.
1812 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1813 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1814 * has been specified, in which case each kthread actively polls its
1815 * CPU. (Which isn't so great for energy efficiency, but which does
1816 * reduce RCU's overhead on that CPU.)
1818 * This is intended to be used in conjunction with Frederic Weisbecker's
1819 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1820 * running CPU-bound user-mode computations.
1822 * Offloading of callback processing could also in theory be used as
1823 * an energy-efficiency measure because CPUs with no RCU callbacks
1824 * queued are more aggressive about entering dyntick-idle mode.
1828 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1829 static int __init rcu_nocb_setup(char *str)
1831 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1832 have_rcu_nocb_mask = true;
1833 cpulist_parse(str, rcu_nocb_mask);
1836 __setup("rcu_nocbs=", rcu_nocb_setup);
1838 static int __init parse_rcu_nocb_poll(char *arg)
1843 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1846 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1849 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1851 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1855 * Set the root rcu_node structure's ->need_future_gp field
1856 * based on the sum of those of all rcu_node structures. This does
1857 * double-count the root rcu_node structure's requests, but this
1858 * is necessary to handle the possibility of a rcu_nocb_kthread()
1859 * having awakened during the time that the rcu_node structures
1860 * were being updated for the end of the previous grace period.
1862 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1864 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1867 static void rcu_init_one_nocb(struct rcu_node *rnp)
1869 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1870 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1873 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1874 /* Is the specified CPU a no-CBs CPU? */
1875 bool rcu_is_nocb_cpu(int cpu)
1877 if (have_rcu_nocb_mask)
1878 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1881 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1884 * Kick the leader kthread for this NOCB group.
1886 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1888 struct rcu_data *rdp_leader = rdp->nocb_leader;
1890 if (!READ_ONCE(rdp_leader->nocb_kthread))
1892 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1893 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1894 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1895 wake_up(&rdp_leader->nocb_wq);
1900 * Does the specified CPU need an RCU callback for the specified flavor
1903 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1905 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1907 #ifdef CONFIG_PROVE_RCU
1908 struct rcu_head *rhp;
1909 #endif /* #ifdef CONFIG_PROVE_RCU */
1912 * Check count of all no-CBs callbacks awaiting invocation.
1913 * There needs to be a barrier before this function is called,
1914 * but associated with a prior determination that no more
1915 * callbacks would be posted. In the worst case, the first
1916 * barrier in _rcu_barrier() suffices (but the caller cannot
1917 * necessarily rely on this, not a substitute for the caller
1918 * getting the concurrency design right!). There must also be
1919 * a barrier between the following load an posting of a callback
1920 * (if a callback is in fact needed). This is associated with an
1921 * atomic_inc() in the caller.
1923 ret = atomic_long_read(&rdp->nocb_q_count);
1925 #ifdef CONFIG_PROVE_RCU
1926 rhp = READ_ONCE(rdp->nocb_head);
1928 rhp = READ_ONCE(rdp->nocb_gp_head);
1930 rhp = READ_ONCE(rdp->nocb_follower_head);
1932 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1933 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1934 rcu_scheduler_fully_active) {
1935 /* RCU callback enqueued before CPU first came online??? */
1936 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1940 #endif /* #ifdef CONFIG_PROVE_RCU */
1946 * Enqueue the specified string of rcu_head structures onto the specified
1947 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1948 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1949 * counts are supplied by rhcount and rhcount_lazy.
1951 * If warranted, also wake up the kthread servicing this CPUs queues.
1953 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1954 struct rcu_head *rhp,
1955 struct rcu_head **rhtp,
1956 int rhcount, int rhcount_lazy,
1957 unsigned long flags)
1960 struct rcu_head **old_rhpp;
1961 struct task_struct *t;
1963 /* Enqueue the callback on the nocb list and update counts. */
1964 atomic_long_add(rhcount, &rdp->nocb_q_count);
1965 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1966 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1967 WRITE_ONCE(*old_rhpp, rhp);
1968 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1969 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1971 /* If we are not being polled and there is a kthread, awaken it ... */
1972 t = READ_ONCE(rdp->nocb_kthread);
1973 if (rcu_nocb_poll || !t) {
1974 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1975 TPS("WakeNotPoll"));
1978 len = atomic_long_read(&rdp->nocb_q_count);
1979 if (old_rhpp == &rdp->nocb_head) {
1980 if (!irqs_disabled_flags(flags)) {
1981 /* ... if queue was empty ... */
1982 wake_nocb_leader(rdp, false);
1983 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1986 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1987 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1988 TPS("WakeEmptyIsDeferred"));
1990 rdp->qlen_last_fqs_check = 0;
1991 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1992 /* ... or if many callbacks queued. */
1993 if (!irqs_disabled_flags(flags)) {
1994 wake_nocb_leader(rdp, true);
1995 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1998 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1999 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2000 TPS("WakeOvfIsDeferred"));
2002 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2004 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2010 * This is a helper for __call_rcu(), which invokes this when the normal
2011 * callback queue is inoperable. If this is not a no-CBs CPU, this
2012 * function returns failure back to __call_rcu(), which can complain
2015 * Otherwise, this function queues the callback where the corresponding
2016 * "rcuo" kthread can find it.
2018 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2019 bool lazy, unsigned long flags)
2022 if (!rcu_is_nocb_cpu(rdp->cpu))
2024 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2025 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2026 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2027 (unsigned long)rhp->func,
2028 -atomic_long_read(&rdp->nocb_q_count_lazy),
2029 -atomic_long_read(&rdp->nocb_q_count));
2031 trace_rcu_callback(rdp->rsp->name, rhp,
2032 -atomic_long_read(&rdp->nocb_q_count_lazy),
2033 -atomic_long_read(&rdp->nocb_q_count));
2036 * If called from an extended quiescent state with interrupts
2037 * disabled, invoke the RCU core in order to allow the idle-entry
2038 * deferred-wakeup check to function.
2040 if (irqs_disabled_flags(flags) &&
2041 !rcu_is_watching() &&
2042 cpu_online(smp_processor_id()))
2049 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2052 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2053 struct rcu_data *rdp,
2054 unsigned long flags)
2056 long ql = rsp->qlen;
2057 long qll = rsp->qlen_lazy;
2059 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2060 if (!rcu_is_nocb_cpu(smp_processor_id()))
2065 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2066 if (rsp->orphan_donelist != NULL) {
2067 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2068 rsp->orphan_donetail, ql, qll, flags);
2070 rsp->orphan_donelist = NULL;
2071 rsp->orphan_donetail = &rsp->orphan_donelist;
2073 if (rsp->orphan_nxtlist != NULL) {
2074 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2075 rsp->orphan_nxttail, ql, qll, flags);
2077 rsp->orphan_nxtlist = NULL;
2078 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2084 * If necessary, kick off a new grace period, and either way wait
2085 * for a subsequent grace period to complete.
2087 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2091 unsigned long flags;
2093 struct rcu_node *rnp = rdp->mynode;
2095 raw_spin_lock_irqsave(&rnp->lock, flags);
2096 smp_mb__after_unlock_lock();
2097 needwake = rcu_start_future_gp(rnp, rdp, &c);
2098 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2100 rcu_gp_kthread_wake(rdp->rsp);
2103 * Wait for the grace period. Do so interruptibly to avoid messing
2104 * up the load average.
2106 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2108 wait_event_interruptible(
2109 rnp->nocb_gp_wq[c & 0x1],
2110 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
2113 WARN_ON(signal_pending(current));
2114 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2116 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2117 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2121 * Leaders come here to wait for additional callbacks to show up.
2122 * This function does not return until callbacks appear.
2124 static void nocb_leader_wait(struct rcu_data *my_rdp)
2126 bool firsttime = true;
2128 struct rcu_data *rdp;
2129 struct rcu_head **tail;
2133 /* Wait for callbacks to appear. */
2134 if (!rcu_nocb_poll) {
2135 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2136 wait_event_interruptible(my_rdp->nocb_wq,
2137 !READ_ONCE(my_rdp->nocb_leader_sleep));
2138 /* Memory barrier handled by smp_mb() calls below and repoll. */
2139 } else if (firsttime) {
2140 firsttime = false; /* Don't drown trace log with "Poll"! */
2141 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2145 * Each pass through the following loop checks a follower for CBs.
2146 * We are our own first follower. Any CBs found are moved to
2147 * nocb_gp_head, where they await a grace period.
2150 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2151 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2152 if (!rdp->nocb_gp_head)
2153 continue; /* No CBs here, try next follower. */
2155 /* Move callbacks to wait-for-GP list, which is empty. */
2156 WRITE_ONCE(rdp->nocb_head, NULL);
2157 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2162 * If there were no callbacks, sleep a bit, rescan after a
2163 * memory barrier, and go retry.
2165 if (unlikely(!gotcbs)) {
2167 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2169 WARN_ON(signal_pending(current));
2170 schedule_timeout_interruptible(1);
2172 /* Rescan in case we were a victim of memory ordering. */
2173 my_rdp->nocb_leader_sleep = true;
2174 smp_mb(); /* Ensure _sleep true before scan. */
2175 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2176 if (READ_ONCE(rdp->nocb_head)) {
2177 /* Found CB, so short-circuit next wait. */
2178 my_rdp->nocb_leader_sleep = false;
2184 /* Wait for one grace period. */
2185 rcu_nocb_wait_gp(my_rdp);
2188 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2189 * We set it now, but recheck for new callbacks while
2190 * traversing our follower list.
2192 my_rdp->nocb_leader_sleep = true;
2193 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2195 /* Each pass through the following loop wakes a follower, if needed. */
2196 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2197 if (READ_ONCE(rdp->nocb_head))
2198 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2199 if (!rdp->nocb_gp_head)
2200 continue; /* No CBs, so no need to wake follower. */
2202 /* Append callbacks to follower's "done" list. */
2203 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2204 *tail = rdp->nocb_gp_head;
2205 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2206 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2208 * List was empty, wake up the follower.
2209 * Memory barriers supplied by atomic_long_add().
2211 wake_up(&rdp->nocb_wq);
2215 /* If we (the leader) don't have CBs, go wait some more. */
2216 if (!my_rdp->nocb_follower_head)
2221 * Followers come here to wait for additional callbacks to show up.
2222 * This function does not return until callbacks appear.
2224 static void nocb_follower_wait(struct rcu_data *rdp)
2226 bool firsttime = true;
2229 if (!rcu_nocb_poll) {
2230 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2232 wait_event_interruptible(rdp->nocb_wq,
2233 READ_ONCE(rdp->nocb_follower_head));
2234 } else if (firsttime) {
2235 /* Don't drown trace log with "Poll"! */
2237 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2239 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2240 /* ^^^ Ensure CB invocation follows _head test. */
2244 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2246 WARN_ON(signal_pending(current));
2247 schedule_timeout_interruptible(1);
2252 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2253 * callbacks queued by the corresponding no-CBs CPU, however, there is
2254 * an optional leader-follower relationship so that the grace-period
2255 * kthreads don't have to do quite so many wakeups.
2257 static int rcu_nocb_kthread(void *arg)
2260 struct rcu_head *list;
2261 struct rcu_head *next;
2262 struct rcu_head **tail;
2263 struct rcu_data *rdp = arg;
2265 /* Each pass through this loop invokes one batch of callbacks */
2267 /* Wait for callbacks. */
2268 if (rdp->nocb_leader == rdp)
2269 nocb_leader_wait(rdp);
2271 nocb_follower_wait(rdp);
2273 /* Pull the ready-to-invoke callbacks onto local list. */
2274 list = READ_ONCE(rdp->nocb_follower_head);
2276 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2277 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2278 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2280 /* Each pass through the following loop invokes a callback. */
2281 trace_rcu_batch_start(rdp->rsp->name,
2282 atomic_long_read(&rdp->nocb_q_count_lazy),
2283 atomic_long_read(&rdp->nocb_q_count), -1);
2287 /* Wait for enqueuing to complete, if needed. */
2288 while (next == NULL && &list->next != tail) {
2289 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2291 schedule_timeout_interruptible(1);
2292 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2296 debug_rcu_head_unqueue(list);
2298 if (__rcu_reclaim(rdp->rsp->name, list))
2304 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2305 smp_mb__before_atomic(); /* _add after CB invocation. */
2306 atomic_long_add(-c, &rdp->nocb_q_count);
2307 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2308 rdp->n_nocbs_invoked += c;
2313 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2314 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2316 return READ_ONCE(rdp->nocb_defer_wakeup);
2319 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2320 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2324 if (!rcu_nocb_need_deferred_wakeup(rdp))
2326 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2327 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2328 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2329 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2332 void __init rcu_init_nohz(void)
2335 bool need_rcu_nocb_mask = true;
2336 struct rcu_state *rsp;
2338 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2339 need_rcu_nocb_mask = false;
2340 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2342 #if defined(CONFIG_NO_HZ_FULL)
2343 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2344 need_rcu_nocb_mask = true;
2345 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2347 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2348 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2349 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2352 have_rcu_nocb_mask = true;
2354 if (!have_rcu_nocb_mask)
2357 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2358 pr_info("\tOffload RCU callbacks from CPU 0\n");
2359 cpumask_set_cpu(0, rcu_nocb_mask);
2360 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2361 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2362 pr_info("\tOffload RCU callbacks from all CPUs\n");
2363 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2364 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2365 #if defined(CONFIG_NO_HZ_FULL)
2366 if (tick_nohz_full_running)
2367 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2368 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2370 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2371 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2372 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2375 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2376 cpumask_pr_args(rcu_nocb_mask));
2378 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2380 for_each_rcu_flavor(rsp) {
2381 for_each_cpu(cpu, rcu_nocb_mask)
2382 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2383 rcu_organize_nocb_kthreads(rsp);
2387 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2388 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2390 rdp->nocb_tail = &rdp->nocb_head;
2391 init_waitqueue_head(&rdp->nocb_wq);
2392 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2396 * If the specified CPU is a no-CBs CPU that does not already have its
2397 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2398 * brought online out of order, this can require re-organizing the
2399 * leader-follower relationships.
2401 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2403 struct rcu_data *rdp;
2404 struct rcu_data *rdp_last;
2405 struct rcu_data *rdp_old_leader;
2406 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2407 struct task_struct *t;
2410 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2411 * then nothing to do.
2413 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2416 /* If we didn't spawn the leader first, reorganize! */
2417 rdp_old_leader = rdp_spawn->nocb_leader;
2418 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2420 rdp = rdp_old_leader;
2422 rdp->nocb_leader = rdp_spawn;
2423 if (rdp_last && rdp != rdp_spawn)
2424 rdp_last->nocb_next_follower = rdp;
2425 if (rdp == rdp_spawn) {
2426 rdp = rdp->nocb_next_follower;
2429 rdp = rdp->nocb_next_follower;
2430 rdp_last->nocb_next_follower = NULL;
2433 rdp_spawn->nocb_next_follower = rdp_old_leader;
2436 /* Spawn the kthread for this CPU and RCU flavor. */
2437 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2438 "rcuo%c/%d", rsp->abbr, cpu);
2440 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2444 * If the specified CPU is a no-CBs CPU that does not already have its
2445 * rcuo kthreads, spawn them.
2447 static void rcu_spawn_all_nocb_kthreads(int cpu)
2449 struct rcu_state *rsp;
2451 if (rcu_scheduler_fully_active)
2452 for_each_rcu_flavor(rsp)
2453 rcu_spawn_one_nocb_kthread(rsp, cpu);
2457 * Once the scheduler is running, spawn rcuo kthreads for all online
2458 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2459 * non-boot CPUs come online -- if this changes, we will need to add
2460 * some mutual exclusion.
2462 static void __init rcu_spawn_nocb_kthreads(void)
2466 for_each_online_cpu(cpu)
2467 rcu_spawn_all_nocb_kthreads(cpu);
2470 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2471 static int rcu_nocb_leader_stride = -1;
2472 module_param(rcu_nocb_leader_stride, int, 0444);
2475 * Initialize leader-follower relationships for all no-CBs CPU.
2477 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2480 int ls = rcu_nocb_leader_stride;
2481 int nl = 0; /* Next leader. */
2482 struct rcu_data *rdp;
2483 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2484 struct rcu_data *rdp_prev = NULL;
2486 if (!have_rcu_nocb_mask)
2489 ls = int_sqrt(nr_cpu_ids);
2490 rcu_nocb_leader_stride = ls;
2494 * Each pass through this loop sets up one rcu_data structure and
2495 * spawns one rcu_nocb_kthread().
2497 for_each_cpu(cpu, rcu_nocb_mask) {
2498 rdp = per_cpu_ptr(rsp->rda, cpu);
2499 if (rdp->cpu >= nl) {
2500 /* New leader, set up for followers & next leader. */
2501 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2502 rdp->nocb_leader = rdp;
2505 /* Another follower, link to previous leader. */
2506 rdp->nocb_leader = rdp_leader;
2507 rdp_prev->nocb_next_follower = rdp;
2513 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2514 static bool init_nocb_callback_list(struct rcu_data *rdp)
2516 if (!rcu_is_nocb_cpu(rdp->cpu))
2519 /* If there are early-boot callbacks, move them to nocb lists. */
2521 rdp->nocb_head = rdp->nxtlist;
2522 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2523 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2524 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2525 rdp->nxtlist = NULL;
2529 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2533 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2535 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2537 WARN_ON_ONCE(1); /* Should be dead code. */
2541 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2545 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2549 static void rcu_init_one_nocb(struct rcu_node *rnp)
2553 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2554 bool lazy, unsigned long flags)
2559 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2560 struct rcu_data *rdp,
2561 unsigned long flags)
2566 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2570 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2575 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2579 static void rcu_spawn_all_nocb_kthreads(int cpu)
2583 static void __init rcu_spawn_nocb_kthreads(void)
2587 static bool init_nocb_callback_list(struct rcu_data *rdp)
2592 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2595 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2596 * arbitrarily long period of time with the scheduling-clock tick turned
2597 * off. RCU will be paying attention to this CPU because it is in the
2598 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2599 * machine because the scheduling-clock tick has been disabled. Therefore,
2600 * if an adaptive-ticks CPU is failing to respond to the current grace
2601 * period and has not be idle from an RCU perspective, kick it.
2603 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2605 #ifdef CONFIG_NO_HZ_FULL
2606 if (tick_nohz_full_cpu(cpu))
2607 smp_send_reschedule(cpu);
2608 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2612 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2614 static int full_sysidle_state; /* Current system-idle state. */
2615 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2616 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2617 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2618 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2619 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2622 * Invoked to note exit from irq or task transition to idle. Note that
2623 * usermode execution does -not- count as idle here! After all, we want
2624 * to detect full-system idle states, not RCU quiescent states and grace
2625 * periods. The caller must have disabled interrupts.
2627 static void rcu_sysidle_enter(int irq)
2630 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2632 /* If there are no nohz_full= CPUs, no need to track this. */
2633 if (!tick_nohz_full_enabled())
2636 /* Adjust nesting, check for fully idle. */
2638 rdtp->dynticks_idle_nesting--;
2639 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2640 if (rdtp->dynticks_idle_nesting != 0)
2641 return; /* Still not fully idle. */
2643 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2644 DYNTICK_TASK_NEST_VALUE) {
2645 rdtp->dynticks_idle_nesting = 0;
2647 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2648 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2649 return; /* Still not fully idle. */
2653 /* Record start of fully idle period. */
2655 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2656 smp_mb__before_atomic();
2657 atomic_inc(&rdtp->dynticks_idle);
2658 smp_mb__after_atomic();
2659 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2663 * Unconditionally force exit from full system-idle state. This is
2664 * invoked when a normal CPU exits idle, but must be called separately
2665 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2666 * is that the timekeeping CPU is permitted to take scheduling-clock
2667 * interrupts while the system is in system-idle state, and of course
2668 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2669 * interrupt from any other type of interrupt.
2671 void rcu_sysidle_force_exit(void)
2673 int oldstate = READ_ONCE(full_sysidle_state);
2677 * Each pass through the following loop attempts to exit full
2678 * system-idle state. If contention proves to be a problem,
2679 * a trylock-based contention tree could be used here.
2681 while (oldstate > RCU_SYSIDLE_SHORT) {
2682 newoldstate = cmpxchg(&full_sysidle_state,
2683 oldstate, RCU_SYSIDLE_NOT);
2684 if (oldstate == newoldstate &&
2685 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2686 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2687 return; /* We cleared it, done! */
2689 oldstate = newoldstate;
2691 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2695 * Invoked to note entry to irq or task transition from idle. Note that
2696 * usermode execution does -not- count as idle here! The caller must
2697 * have disabled interrupts.
2699 static void rcu_sysidle_exit(int irq)
2701 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2703 /* If there are no nohz_full= CPUs, no need to track this. */
2704 if (!tick_nohz_full_enabled())
2707 /* Adjust nesting, check for already non-idle. */
2709 rdtp->dynticks_idle_nesting++;
2710 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2711 if (rdtp->dynticks_idle_nesting != 1)
2712 return; /* Already non-idle. */
2715 * Allow for irq misnesting. Yes, it really is possible
2716 * to enter an irq handler then never leave it, and maybe
2717 * also vice versa. Handle both possibilities.
2719 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2720 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2721 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2722 return; /* Already non-idle. */
2724 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2728 /* Record end of idle period. */
2729 smp_mb__before_atomic();
2730 atomic_inc(&rdtp->dynticks_idle);
2731 smp_mb__after_atomic();
2732 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2735 * If we are the timekeeping CPU, we are permitted to be non-idle
2736 * during a system-idle state. This must be the case, because
2737 * the timekeeping CPU has to take scheduling-clock interrupts
2738 * during the time that the system is transitioning to full
2739 * system-idle state. This means that the timekeeping CPU must
2740 * invoke rcu_sysidle_force_exit() directly if it does anything
2741 * more than take a scheduling-clock interrupt.
2743 if (smp_processor_id() == tick_do_timer_cpu)
2746 /* Update system-idle state: We are clearly no longer fully idle! */
2747 rcu_sysidle_force_exit();
2751 * Check to see if the current CPU is idle. Note that usermode execution
2752 * does not count as idle. The caller must have disabled interrupts,
2753 * and must be running on tick_do_timer_cpu.
2755 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2756 unsigned long *maxj)
2760 struct rcu_dynticks *rdtp = rdp->dynticks;
2762 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2763 if (!tick_nohz_full_enabled())
2767 * If some other CPU has already reported non-idle, if this is
2768 * not the flavor of RCU that tracks sysidle state, or if this
2769 * is an offline or the timekeeping CPU, nothing to do.
2771 if (!*isidle || rdp->rsp != rcu_state_p ||
2772 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2774 /* Verify affinity of current kthread. */
2775 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2777 /* Pick up current idle and NMI-nesting counter and check. */
2778 cur = atomic_read(&rdtp->dynticks_idle);
2780 *isidle = false; /* We are not idle! */
2783 smp_mb(); /* Read counters before timestamps. */
2785 /* Pick up timestamps. */
2786 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2787 /* If this CPU entered idle more recently, update maxj timestamp. */
2788 if (ULONG_CMP_LT(*maxj, j))
2793 * Is this the flavor of RCU that is handling full-system idle?
2795 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2797 return rsp == rcu_state_p;
2801 * Return a delay in jiffies based on the number of CPUs, rcu_node
2802 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2803 * systems more time to transition to full-idle state in order to
2804 * avoid the cache thrashing that otherwise occur on the state variable.
2805 * Really small systems (less than a couple of tens of CPUs) should
2806 * instead use a single global atomically incremented counter, and later
2807 * versions of this will automatically reconfigure themselves accordingly.
2809 static unsigned long rcu_sysidle_delay(void)
2811 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2813 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2817 * Advance the full-system-idle state. This is invoked when all of
2818 * the non-timekeeping CPUs are idle.
2820 static void rcu_sysidle(unsigned long j)
2822 /* Check the current state. */
2823 switch (READ_ONCE(full_sysidle_state)) {
2824 case RCU_SYSIDLE_NOT:
2826 /* First time all are idle, so note a short idle period. */
2827 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2830 case RCU_SYSIDLE_SHORT:
2833 * Idle for a bit, time to advance to next state?
2834 * cmpxchg failure means race with non-idle, let them win.
2836 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2837 (void)cmpxchg(&full_sysidle_state,
2838 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2841 case RCU_SYSIDLE_LONG:
2844 * Do an additional check pass before advancing to full.
2845 * cmpxchg failure means race with non-idle, let them win.
2847 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2848 (void)cmpxchg(&full_sysidle_state,
2849 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2858 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2859 * back to the beginning.
2861 static void rcu_sysidle_cancel(void)
2864 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2865 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2869 * Update the sysidle state based on the results of a force-quiescent-state
2870 * scan of the CPUs' dyntick-idle state.
2872 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2873 unsigned long maxj, bool gpkt)
2875 if (rsp != rcu_state_p)
2876 return; /* Wrong flavor, ignore. */
2877 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2878 return; /* Running state machine from timekeeping CPU. */
2880 rcu_sysidle(maxj); /* More idle! */
2882 rcu_sysidle_cancel(); /* Idle is over. */
2886 * Wrapper for rcu_sysidle_report() when called from the grace-period
2887 * kthread's context.
2889 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2892 /* If there are no nohz_full= CPUs, no need to track this. */
2893 if (!tick_nohz_full_enabled())
2896 rcu_sysidle_report(rsp, isidle, maxj, true);
2899 /* Callback and function for forcing an RCU grace period. */
2900 struct rcu_sysidle_head {
2905 static void rcu_sysidle_cb(struct rcu_head *rhp)
2907 struct rcu_sysidle_head *rshp;
2910 * The following memory barrier is needed to replace the
2911 * memory barriers that would normally be in the memory
2914 smp_mb(); /* grace period precedes setting inuse. */
2916 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2917 WRITE_ONCE(rshp->inuse, 0);
2921 * Check to see if the system is fully idle, other than the timekeeping CPU.
2922 * The caller must have disabled interrupts. This is not intended to be
2923 * called unless tick_nohz_full_enabled().
2925 bool rcu_sys_is_idle(void)
2927 static struct rcu_sysidle_head rsh;
2928 int rss = READ_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_state_p->rda, cpu);
2950 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2954 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2956 rss = READ_ONCE(full_sysidle_state);
2960 /* If this is the first observation of an idle period, record it. */
2961 if (rss == RCU_SYSIDLE_FULL) {
2962 rss = cmpxchg(&full_sysidle_state,
2963 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2964 return rss == RCU_SYSIDLE_FULL;
2967 smp_mb(); /* ensure rss load happens before later caller actions. */
2969 /* If already fully idle, tell the caller (in case of races). */
2970 if (rss == RCU_SYSIDLE_FULL_NOTED)
2974 * If we aren't there yet, and a grace period is not in flight,
2975 * initiate a grace period. Either way, tell the caller that
2976 * we are not there yet. We use an xchg() rather than an assignment
2977 * to make up for the memory barriers that would otherwise be
2978 * provided by the memory allocator.
2980 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2981 !rcu_gp_in_progress(rcu_state_p) &&
2982 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2983 call_rcu(&rsh.rh, rcu_sysidle_cb);
2988 * Initialize dynticks sysidle state for CPUs coming online.
2990 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2992 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2995 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2997 static void rcu_sysidle_enter(int irq)
3001 static void rcu_sysidle_exit(int irq)
3005 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3006 unsigned long *maxj)
3010 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3015 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3020 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3024 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3027 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3028 * grace-period kthread will do force_quiescent_state() processing?
3029 * The idea is to avoid waking up RCU core processing on such a
3030 * CPU unless the grace period has extended for too long.
3032 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3033 * CONFIG_RCU_NOCB_CPU CPUs.
3035 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3037 #ifdef CONFIG_NO_HZ_FULL
3038 if (tick_nohz_full_cpu(smp_processor_id()) &&
3039 (!rcu_gp_in_progress(rsp) ||
3040 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
3042 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3047 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3050 static void rcu_bind_gp_kthread(void)
3052 int __maybe_unused cpu;
3054 if (!tick_nohz_full_enabled())
3056 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3057 cpu = tick_do_timer_cpu;
3058 if (cpu >= 0 && cpu < nr_cpu_ids)
3059 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3060 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3061 housekeeping_affine(current);
3062 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3065 /* Record the current task on dyntick-idle entry. */
3066 static void rcu_dynticks_task_enter(void)
3068 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3069 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
3070 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3073 /* Record no current task on dyntick-idle exit. */
3074 static void rcu_dynticks_task_exit(void)
3076 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3077 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
3078 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */