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 #ifdef CONFIG_RCU_BOOST
29 #include "../locking/rtmutex_common.h"
31 #else /* #ifdef CONFIG_RCU_BOOST */
34 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
35 * all uses are in dead code. Provide a definition to keep the compiler
36 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
37 * This probably needs to be excluded from -rt builds.
39 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
41 #endif /* #else #ifdef CONFIG_RCU_BOOST */
44 * Control variables for per-CPU and per-rcu_node kthreads. These
45 * handle all flavors of RCU.
47 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
48 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
49 DEFINE_PER_CPU(char, rcu_cpu_has_work);
51 #ifdef CONFIG_RCU_NOCB_CPU
52 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
53 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
54 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
55 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
58 * Check the RCU kernel configuration parameters and print informative
59 * messages about anything out of the ordinary.
61 static void __init rcu_bootup_announce_oddness(void)
63 if (IS_ENABLED(CONFIG_RCU_TRACE))
64 pr_info("\tRCU debugfs-based tracing is enabled.\n");
65 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
66 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
67 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
70 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
71 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
72 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
73 if (IS_ENABLED(CONFIG_PROVE_RCU))
74 pr_info("\tRCU lockdep checking is enabled.\n");
75 if (RCU_NUM_LVLS >= 4)
76 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
77 if (RCU_FANOUT_LEAF != 16)
78 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
80 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
81 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
82 if (nr_cpu_ids != NR_CPUS)
83 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
84 if (IS_ENABLED(CONFIG_RCU_BOOST))
85 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
88 #ifdef CONFIG_PREEMPT_RCU
90 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
91 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
92 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
94 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
98 * Tell them what RCU they are running.
100 static void __init rcu_bootup_announce(void)
102 pr_info("Preemptible hierarchical RCU implementation.\n");
103 rcu_bootup_announce_oddness();
106 /* Flags for rcu_preempt_ctxt_queue() decision table. */
107 #define RCU_GP_TASKS 0x8
108 #define RCU_EXP_TASKS 0x4
109 #define RCU_GP_BLKD 0x2
110 #define RCU_EXP_BLKD 0x1
113 * Queues a task preempted within an RCU-preempt read-side critical
114 * section into the appropriate location within the ->blkd_tasks list,
115 * depending on the states of any ongoing normal and expedited grace
116 * periods. The ->gp_tasks pointer indicates which element the normal
117 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
118 * indicates which element the expedited grace period is waiting on (again,
119 * NULL if none). If a grace period is waiting on a given element in the
120 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
121 * adding a task to the tail of the list blocks any grace period that is
122 * already waiting on one of the elements. In contrast, adding a task
123 * to the head of the list won't block any grace period that is already
124 * waiting on one of the elements.
126 * This queuing is imprecise, and can sometimes make an ongoing grace
127 * period wait for a task that is not strictly speaking blocking it.
128 * Given the choice, we needlessly block a normal grace period rather than
129 * blocking an expedited grace period.
131 * Note that an endless sequence of expedited grace periods still cannot
132 * indefinitely postpone a normal grace period. Eventually, all of the
133 * fixed number of preempted tasks blocking the normal grace period that are
134 * not also blocking the expedited grace period will resume and complete
135 * their RCU read-side critical sections. At that point, the ->gp_tasks
136 * pointer will equal the ->exp_tasks pointer, at which point the end of
137 * the corresponding expedited grace period will also be the end of the
138 * normal grace period.
140 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
141 __releases(rnp->lock) /* But leaves rrupts disabled. */
143 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
144 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
145 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
146 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
147 struct task_struct *t = current;
150 * Decide where to queue the newly blocked task. In theory,
151 * this could be an if-statement. In practice, when I tried
152 * that, it was quite messy.
154 switch (blkd_state) {
157 case RCU_EXP_TASKS + RCU_GP_BLKD:
159 case RCU_GP_TASKS + RCU_EXP_TASKS:
162 * Blocking neither GP, or first task blocking the normal
163 * GP but not blocking the already-waiting expedited GP.
164 * Queue at the head of the list to avoid unnecessarily
165 * blocking the already-waiting GPs.
167 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
172 case RCU_GP_BLKD + RCU_EXP_BLKD:
173 case RCU_GP_TASKS + RCU_EXP_BLKD:
174 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
175 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
178 * First task arriving that blocks either GP, or first task
179 * arriving that blocks the expedited GP (with the normal
180 * GP already waiting), or a task arriving that blocks
181 * both GPs with both GPs already waiting. Queue at the
182 * tail of the list to avoid any GP waiting on any of the
183 * already queued tasks that are not blocking it.
185 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
188 case RCU_EXP_TASKS + RCU_EXP_BLKD:
189 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
190 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
193 * Second or subsequent task blocking the expedited GP.
194 * The task either does not block the normal GP, or is the
195 * first task blocking the normal GP. Queue just after
196 * the first task blocking the expedited GP.
198 list_add(&t->rcu_node_entry, rnp->exp_tasks);
201 case RCU_GP_TASKS + RCU_GP_BLKD:
202 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
205 * Second or subsequent task blocking the normal GP.
206 * The task does not block the expedited GP. Queue just
207 * after the first task blocking the normal GP.
209 list_add(&t->rcu_node_entry, rnp->gp_tasks);
214 /* Yet another exercise in excessive paranoia. */
220 * We have now queued the task. If it was the first one to
221 * block either grace period, update the ->gp_tasks and/or
222 * ->exp_tasks pointers, respectively, to reference the newly
225 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD))
226 rnp->gp_tasks = &t->rcu_node_entry;
227 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
228 rnp->exp_tasks = &t->rcu_node_entry;
229 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
232 * Report the quiescent state for the expedited GP. This expedited
233 * GP should not be able to end until we report, so there should be
234 * no need to check for a subsequent expedited GP. (Though we are
235 * still in a quiescent state in any case.)
237 if (blkd_state & RCU_EXP_BLKD &&
238 t->rcu_read_unlock_special.b.exp_need_qs) {
239 t->rcu_read_unlock_special.b.exp_need_qs = false;
240 rcu_report_exp_rdp(rdp->rsp, rdp, true);
242 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
247 * Record a preemptible-RCU quiescent state for the specified CPU. Note
248 * that this just means that the task currently running on the CPU is
249 * not in a quiescent state. There might be any number of tasks blocked
250 * while in an RCU read-side critical section.
252 * As with the other rcu_*_qs() functions, callers to this function
253 * must disable preemption.
255 static void rcu_preempt_qs(void)
257 if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
258 trace_rcu_grace_period(TPS("rcu_preempt"),
259 __this_cpu_read(rcu_data_p->gpnum),
261 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
262 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
263 current->rcu_read_unlock_special.b.need_qs = false;
268 * We have entered the scheduler, and the current task might soon be
269 * context-switched away from. If this task is in an RCU read-side
270 * critical section, we will no longer be able to rely on the CPU to
271 * record that fact, so we enqueue the task on the blkd_tasks list.
272 * The task will dequeue itself when it exits the outermost enclosing
273 * RCU read-side critical section. Therefore, the current grace period
274 * cannot be permitted to complete until the blkd_tasks list entries
275 * predating the current grace period drain, in other words, until
276 * rnp->gp_tasks becomes NULL.
278 * Caller must disable interrupts.
280 static void rcu_preempt_note_context_switch(void)
282 struct task_struct *t = current;
283 struct rcu_data *rdp;
284 struct rcu_node *rnp;
286 if (t->rcu_read_lock_nesting > 0 &&
287 !t->rcu_read_unlock_special.b.blocked) {
289 /* Possibly blocking in an RCU read-side critical section. */
290 rdp = this_cpu_ptr(rcu_state_p->rda);
292 raw_spin_lock_rcu_node(rnp);
293 t->rcu_read_unlock_special.b.blocked = true;
294 t->rcu_blocked_node = rnp;
297 * Verify the CPU's sanity, trace the preemption, and
298 * then queue the task as required based on the states
299 * of any ongoing and expedited grace periods.
301 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
302 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
303 trace_rcu_preempt_task(rdp->rsp->name,
305 (rnp->qsmask & rdp->grpmask)
308 rcu_preempt_ctxt_queue(rnp, rdp);
309 } else if (t->rcu_read_lock_nesting < 0 &&
310 t->rcu_read_unlock_special.s) {
313 * Complete exit from RCU read-side critical section on
314 * behalf of preempted instance of __rcu_read_unlock().
316 rcu_read_unlock_special(t);
320 * Either we were not in an RCU read-side critical section to
321 * begin with, or we have now recorded that critical section
322 * globally. Either way, we can now note a quiescent state
323 * for this CPU. Again, if we were in an RCU read-side critical
324 * section, and if that critical section was blocking the current
325 * grace period, then the fact that the task has been enqueued
326 * means that we continue to block the current grace period.
332 * Check for preempted RCU readers blocking the current grace period
333 * for the specified rcu_node structure. If the caller needs a reliable
334 * answer, it must hold the rcu_node's ->lock.
336 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
338 return rnp->gp_tasks != NULL;
342 * Advance a ->blkd_tasks-list pointer to the next entry, instead
343 * returning NULL if at the end of the list.
345 static struct list_head *rcu_next_node_entry(struct task_struct *t,
346 struct rcu_node *rnp)
348 struct list_head *np;
350 np = t->rcu_node_entry.next;
351 if (np == &rnp->blkd_tasks)
357 * Return true if the specified rcu_node structure has tasks that were
358 * preempted within an RCU read-side critical section.
360 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
362 return !list_empty(&rnp->blkd_tasks);
366 * Handle special cases during rcu_read_unlock(), such as needing to
367 * notify RCU core processing or task having blocked during the RCU
368 * read-side critical section.
370 void rcu_read_unlock_special(struct task_struct *t)
376 struct list_head *np;
377 bool drop_boost_mutex = false;
378 struct rcu_data *rdp;
379 struct rcu_node *rnp;
380 union rcu_special special;
382 /* NMI handlers cannot block and cannot safely manipulate state. */
386 local_irq_save(flags);
389 * If RCU core is waiting for this CPU to exit its critical section,
390 * report the fact that it has exited. Because irqs are disabled,
391 * t->rcu_read_unlock_special cannot change.
393 special = t->rcu_read_unlock_special;
394 if (special.b.need_qs) {
396 t->rcu_read_unlock_special.b.need_qs = false;
397 if (!t->rcu_read_unlock_special.s) {
398 local_irq_restore(flags);
404 * Respond to a request for an expedited grace period, but only if
405 * we were not preempted, meaning that we were running on the same
406 * CPU throughout. If we were preempted, the exp_need_qs flag
407 * would have been cleared at the time of the first preemption,
408 * and the quiescent state would be reported when we were dequeued.
410 if (special.b.exp_need_qs) {
411 WARN_ON_ONCE(special.b.blocked);
412 t->rcu_read_unlock_special.b.exp_need_qs = false;
413 rdp = this_cpu_ptr(rcu_state_p->rda);
414 rcu_report_exp_rdp(rcu_state_p, rdp, true);
415 if (!t->rcu_read_unlock_special.s) {
416 local_irq_restore(flags);
421 /* Hardware IRQ handlers cannot block, complain if they get here. */
422 if (preempt_count() & (HARDIRQ_MASK | SOFTIRQ_OFFSET)) {
423 lockdep_rcu_suspicious(__FILE__, __LINE__,
424 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
425 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
426 t->rcu_read_unlock_special.s,
427 t->rcu_read_unlock_special.b.blocked,
428 t->rcu_read_unlock_special.b.exp_need_qs,
429 t->rcu_read_unlock_special.b.need_qs);
430 local_irq_restore(flags);
434 /* Clean up if blocked during RCU read-side critical section. */
435 if (special.b.blocked) {
436 t->rcu_read_unlock_special.b.blocked = false;
439 * Remove this task from the list it blocked on. The task
440 * now remains queued on the rcu_node corresponding to the
441 * CPU it first blocked on, so there is no longer any need
442 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
444 rnp = t->rcu_blocked_node;
445 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
446 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
447 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
448 empty_exp = sync_rcu_preempt_exp_done(rnp);
449 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
450 np = rcu_next_node_entry(t, rnp);
451 list_del_init(&t->rcu_node_entry);
452 t->rcu_blocked_node = NULL;
453 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
455 if (&t->rcu_node_entry == rnp->gp_tasks)
457 if (&t->rcu_node_entry == rnp->exp_tasks)
459 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
460 if (&t->rcu_node_entry == rnp->boost_tasks)
461 rnp->boost_tasks = np;
462 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
463 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
467 * If this was the last task on the current list, and if
468 * we aren't waiting on any CPUs, report the quiescent state.
469 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
470 * so we must take a snapshot of the expedited state.
472 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
473 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
474 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
481 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
483 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
486 /* Unboost if we were boosted. */
487 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
488 rt_mutex_unlock(&rnp->boost_mtx);
491 * If this was the last task on the expedited lists,
492 * then we need to report up the rcu_node hierarchy.
494 if (!empty_exp && empty_exp_now)
495 rcu_report_exp_rnp(rcu_state_p, rnp, true);
497 local_irq_restore(flags);
502 * Dump detailed information for all tasks blocking the current RCU
503 * grace period on the specified rcu_node structure.
505 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
508 struct task_struct *t;
510 raw_spin_lock_irqsave_rcu_node(rnp, flags);
511 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
512 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
515 t = list_entry(rnp->gp_tasks->prev,
516 struct task_struct, rcu_node_entry);
517 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
519 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
523 * Dump detailed information for all tasks blocking the current RCU
526 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
528 struct rcu_node *rnp = rcu_get_root(rsp);
530 rcu_print_detail_task_stall_rnp(rnp);
531 rcu_for_each_leaf_node(rsp, rnp)
532 rcu_print_detail_task_stall_rnp(rnp);
535 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
537 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
538 rnp->level, rnp->grplo, rnp->grphi);
541 static void rcu_print_task_stall_end(void)
547 * Scan the current list of tasks blocked within RCU read-side critical
548 * sections, printing out the tid of each.
550 static int rcu_print_task_stall(struct rcu_node *rnp)
552 struct task_struct *t;
555 if (!rcu_preempt_blocked_readers_cgp(rnp))
557 rcu_print_task_stall_begin(rnp);
558 t = list_entry(rnp->gp_tasks->prev,
559 struct task_struct, rcu_node_entry);
560 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
561 pr_cont(" P%d", t->pid);
564 rcu_print_task_stall_end();
569 * Scan the current list of tasks blocked within RCU read-side critical
570 * sections, printing out the tid of each that is blocking the current
571 * expedited grace period.
573 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
575 struct task_struct *t;
580 t = list_entry(rnp->exp_tasks->prev,
581 struct task_struct, rcu_node_entry);
582 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
583 pr_cont(" P%d", t->pid);
590 * Check that the list of blocked tasks for the newly completed grace
591 * period is in fact empty. It is a serious bug to complete a grace
592 * period that still has RCU readers blocked! This function must be
593 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
594 * must be held by the caller.
596 * Also, if there are blocked tasks on the list, they automatically
597 * block the newly created grace period, so set up ->gp_tasks accordingly.
599 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
601 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
602 if (rcu_preempt_has_tasks(rnp))
603 rnp->gp_tasks = rnp->blkd_tasks.next;
604 WARN_ON_ONCE(rnp->qsmask);
608 * Check for a quiescent state from the current CPU. When a task blocks,
609 * the task is recorded in the corresponding CPU's rcu_node structure,
610 * which is checked elsewhere.
612 * Caller must disable hard irqs.
614 static void rcu_preempt_check_callbacks(void)
616 struct task_struct *t = current;
618 if (t->rcu_read_lock_nesting == 0) {
622 if (t->rcu_read_lock_nesting > 0 &&
623 __this_cpu_read(rcu_data_p->core_needs_qs) &&
624 __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm))
625 t->rcu_read_unlock_special.b.need_qs = true;
629 * Queue a preemptible-RCU callback for invocation after a grace period.
631 void call_rcu(struct rcu_head *head, rcu_callback_t func)
633 __call_rcu(head, func, rcu_state_p, -1, 0);
635 EXPORT_SYMBOL_GPL(call_rcu);
638 * synchronize_rcu - wait until a grace period has elapsed.
640 * Control will return to the caller some time after a full grace
641 * period has elapsed, in other words after all currently executing RCU
642 * read-side critical sections have completed. Note, however, that
643 * upon return from synchronize_rcu(), the caller might well be executing
644 * concurrently with new RCU read-side critical sections that began while
645 * synchronize_rcu() was waiting. RCU read-side critical sections are
646 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
648 * See the description of synchronize_sched() for more detailed information
649 * on memory ordering guarantees.
651 void synchronize_rcu(void)
653 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
654 lock_is_held(&rcu_lock_map) ||
655 lock_is_held(&rcu_sched_lock_map),
656 "Illegal synchronize_rcu() in RCU read-side critical section");
657 if (!rcu_scheduler_active)
659 if (rcu_gp_is_expedited())
660 synchronize_rcu_expedited();
662 wait_rcu_gp(call_rcu);
664 EXPORT_SYMBOL_GPL(synchronize_rcu);
667 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
669 * Note that this primitive does not necessarily wait for an RCU grace period
670 * to complete. For example, if there are no RCU callbacks queued anywhere
671 * in the system, then rcu_barrier() is within its rights to return
672 * immediately, without waiting for anything, much less an RCU grace period.
674 void rcu_barrier(void)
676 _rcu_barrier(rcu_state_p);
678 EXPORT_SYMBOL_GPL(rcu_barrier);
681 * Initialize preemptible RCU's state structures.
683 static void __init __rcu_init_preempt(void)
685 rcu_init_one(rcu_state_p);
689 * Check for a task exiting while in a preemptible-RCU read-side
690 * critical section, clean up if so. No need to issue warnings,
691 * as debug_check_no_locks_held() already does this if lockdep
696 struct task_struct *t = current;
698 if (likely(list_empty(¤t->rcu_node_entry)))
700 t->rcu_read_lock_nesting = 1;
702 t->rcu_read_unlock_special.b.blocked = true;
706 #else /* #ifdef CONFIG_PREEMPT_RCU */
708 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
711 * Tell them what RCU they are running.
713 static void __init rcu_bootup_announce(void)
715 pr_info("Hierarchical RCU implementation.\n");
716 rcu_bootup_announce_oddness();
720 * Because preemptible RCU does not exist, we never have to check for
721 * CPUs being in quiescent states.
723 static void rcu_preempt_note_context_switch(void)
728 * Because preemptible RCU does not exist, there are never any preempted
731 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
737 * Because there is no preemptible RCU, there can be no readers blocked.
739 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
745 * Because preemptible RCU does not exist, we never have to check for
746 * tasks blocked within RCU read-side critical sections.
748 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
753 * Because preemptible RCU does not exist, we never have to check for
754 * tasks blocked within RCU read-side critical sections.
756 static int rcu_print_task_stall(struct rcu_node *rnp)
762 * Because preemptible RCU does not exist, we never have to check for
763 * tasks blocked within RCU read-side critical sections that are
764 * blocking the current expedited grace period.
766 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
772 * Because there is no preemptible RCU, there can be no readers blocked,
773 * so there is no need to check for blocked tasks. So check only for
774 * bogus qsmask values.
776 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
778 WARN_ON_ONCE(rnp->qsmask);
782 * Because preemptible RCU does not exist, it never has any callbacks
785 static void rcu_preempt_check_callbacks(void)
790 * Because preemptible RCU does not exist, rcu_barrier() is just
791 * another name for rcu_barrier_sched().
793 void rcu_barrier(void)
797 EXPORT_SYMBOL_GPL(rcu_barrier);
800 * Because preemptible RCU does not exist, it need not be initialized.
802 static void __init __rcu_init_preempt(void)
807 * Because preemptible RCU does not exist, tasks cannot possibly exit
808 * while in preemptible RCU read-side critical sections.
814 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
817 * If boosting, set rcuc kthreads to realtime priority.
819 static void rcu_cpu_kthread_setup(unsigned int cpu)
821 #ifdef CONFIG_RCU_BOOST
822 struct sched_param sp;
824 sp.sched_priority = kthread_prio;
825 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
826 #endif /* #ifdef CONFIG_RCU_BOOST */
829 #ifdef CONFIG_RCU_BOOST
831 #include "../locking/rtmutex_common.h"
833 #ifdef CONFIG_RCU_TRACE
835 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
837 if (!rcu_preempt_has_tasks(rnp))
838 rnp->n_balk_blkd_tasks++;
839 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
840 rnp->n_balk_exp_gp_tasks++;
841 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
842 rnp->n_balk_boost_tasks++;
843 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
844 rnp->n_balk_notblocked++;
845 else if (rnp->gp_tasks != NULL &&
846 ULONG_CMP_LT(jiffies, rnp->boost_time))
847 rnp->n_balk_notyet++;
852 #else /* #ifdef CONFIG_RCU_TRACE */
854 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
858 #endif /* #else #ifdef CONFIG_RCU_TRACE */
861 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
862 * or ->boost_tasks, advancing the pointer to the next task in the
865 * Note that irqs must be enabled: boosting the task can block.
866 * Returns 1 if there are more tasks needing to be boosted.
868 static int rcu_boost(struct rcu_node *rnp)
871 struct task_struct *t;
872 struct list_head *tb;
874 if (READ_ONCE(rnp->exp_tasks) == NULL &&
875 READ_ONCE(rnp->boost_tasks) == NULL)
876 return 0; /* Nothing left to boost. */
878 raw_spin_lock_irqsave_rcu_node(rnp, flags);
881 * Recheck under the lock: all tasks in need of boosting
882 * might exit their RCU read-side critical sections on their own.
884 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
885 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
890 * Preferentially boost tasks blocking expedited grace periods.
891 * This cannot starve the normal grace periods because a second
892 * expedited grace period must boost all blocked tasks, including
893 * those blocking the pre-existing normal grace period.
895 if (rnp->exp_tasks != NULL) {
899 tb = rnp->boost_tasks;
900 rnp->n_normal_boosts++;
902 rnp->n_tasks_boosted++;
905 * We boost task t by manufacturing an rt_mutex that appears to
906 * be held by task t. We leave a pointer to that rt_mutex where
907 * task t can find it, and task t will release the mutex when it
908 * exits its outermost RCU read-side critical section. Then
909 * simply acquiring this artificial rt_mutex will boost task
910 * t's priority. (Thanks to tglx for suggesting this approach!)
912 * Note that task t must acquire rnp->lock to remove itself from
913 * the ->blkd_tasks list, which it will do from exit() if from
914 * nowhere else. We therefore are guaranteed that task t will
915 * stay around at least until we drop rnp->lock. Note that
916 * rnp->lock also resolves races between our priority boosting
917 * and task t's exiting its outermost RCU read-side critical
920 t = container_of(tb, struct task_struct, rcu_node_entry);
921 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
922 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
923 /* Lock only for side effect: boosts task t's priority. */
924 rt_mutex_lock(&rnp->boost_mtx);
925 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
927 return READ_ONCE(rnp->exp_tasks) != NULL ||
928 READ_ONCE(rnp->boost_tasks) != NULL;
932 * Priority-boosting kthread, one per leaf rcu_node.
934 static int rcu_boost_kthread(void *arg)
936 struct rcu_node *rnp = (struct rcu_node *)arg;
940 trace_rcu_utilization(TPS("Start boost kthread@init"));
942 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
943 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
944 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
945 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
946 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
947 more2boost = rcu_boost(rnp);
953 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
954 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
955 schedule_timeout_interruptible(2);
956 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
961 trace_rcu_utilization(TPS("End boost kthread@notreached"));
966 * Check to see if it is time to start boosting RCU readers that are
967 * blocking the current grace period, and, if so, tell the per-rcu_node
968 * kthread to start boosting them. If there is an expedited grace
969 * period in progress, it is always time to boost.
971 * The caller must hold rnp->lock, which this function releases.
972 * The ->boost_kthread_task is immortal, so we don't need to worry
973 * about it going away.
975 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
976 __releases(rnp->lock)
978 struct task_struct *t;
980 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
981 rnp->n_balk_exp_gp_tasks++;
982 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
985 if (rnp->exp_tasks != NULL ||
986 (rnp->gp_tasks != NULL &&
987 rnp->boost_tasks == NULL &&
989 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
990 if (rnp->exp_tasks == NULL)
991 rnp->boost_tasks = rnp->gp_tasks;
992 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
993 t = rnp->boost_kthread_task;
995 rcu_wake_cond(t, rnp->boost_kthread_status);
997 rcu_initiate_boost_trace(rnp);
998 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1003 * Is the current CPU running the RCU-callbacks kthread?
1004 * Caller must have preemption disabled.
1006 static bool rcu_is_callbacks_kthread(void)
1008 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1011 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1014 * Do priority-boost accounting for the start of a new grace period.
1016 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1018 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1022 * Create an RCU-boost kthread for the specified node if one does not
1023 * already exist. We only create this kthread for preemptible RCU.
1024 * Returns zero if all is well, a negated errno otherwise.
1026 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1027 struct rcu_node *rnp)
1029 int rnp_index = rnp - &rsp->node[0];
1030 unsigned long flags;
1031 struct sched_param sp;
1032 struct task_struct *t;
1034 if (rcu_state_p != rsp)
1037 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1041 if (rnp->boost_kthread_task != NULL)
1043 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1044 "rcub/%d", rnp_index);
1047 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1048 rnp->boost_kthread_task = t;
1049 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1050 sp.sched_priority = kthread_prio;
1051 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1052 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1057 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1058 * served by the rcu_node in question. The CPU hotplug lock is still
1059 * held, so the value of rnp->qsmaskinit will be stable.
1061 * We don't include outgoingcpu in the affinity set, use -1 if there is
1062 * no outgoing CPU. If there are no CPUs left in the affinity set,
1063 * this function allows the kthread to execute on any CPU.
1065 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1067 struct task_struct *t = rnp->boost_kthread_task;
1068 unsigned long mask = rcu_rnp_online_cpus(rnp);
1074 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1076 for_each_leaf_node_possible_cpu(rnp, cpu)
1077 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1079 cpumask_set_cpu(cpu, cm);
1080 if (cpumask_weight(cm) == 0)
1082 set_cpus_allowed_ptr(t, cm);
1083 free_cpumask_var(cm);
1087 * Spawn boost kthreads -- called as soon as the scheduler is running.
1089 static void __init rcu_spawn_boost_kthreads(void)
1091 struct rcu_node *rnp;
1092 rcu_for_each_leaf_node(rcu_state_p, rnp)
1093 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1096 static void rcu_prepare_kthreads(int cpu)
1098 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1099 struct rcu_node *rnp = rdp->mynode;
1101 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1102 if (rcu_scheduler_fully_active)
1103 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1106 #else /* #ifdef CONFIG_RCU_BOOST */
1108 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1109 __releases(rnp->lock)
1111 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1114 static bool rcu_is_callbacks_kthread(void)
1119 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1123 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1127 static void __init rcu_spawn_boost_kthreads(void)
1131 static void rcu_prepare_kthreads(int cpu)
1135 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1137 #if !defined(CONFIG_RCU_FAST_NO_HZ) || defined(CONFIG_PREEMPT_RT_FULL)
1140 * Check to see if any future RCU-related work will need to be done
1141 * by the current CPU, even if none need be done immediately, returning
1142 * 1 if so. This function is part of the RCU implementation; it is -not-
1143 * an exported member of the RCU API.
1145 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1146 * any flavor of RCU.
1148 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1150 *nextevt = KTIME_MAX;
1151 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1152 ? 0 : rcu_cpu_has_callbacks(NULL);
1154 #endif /* !defined(CONFIG_RCU_FAST_NO_HZ) || defined(CONFIG_PREEMPT_RT_FULL) */
1156 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1158 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1161 static void rcu_cleanup_after_idle(void)
1166 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1169 static void rcu_prepare_for_idle(void)
1174 * Don't bother keeping a running count of the number of RCU callbacks
1175 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1177 static void rcu_idle_count_callbacks_posted(void)
1181 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1184 * This code is invoked when a CPU goes idle, at which point we want
1185 * to have the CPU do everything required for RCU so that it can enter
1186 * the energy-efficient dyntick-idle mode. This is handled by a
1187 * state machine implemented by rcu_prepare_for_idle() below.
1189 * The following three proprocessor symbols control this state machine:
1191 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1192 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1193 * is sized to be roughly one RCU grace period. Those energy-efficiency
1194 * benchmarkers who might otherwise be tempted to set this to a large
1195 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1196 * system. And if you are -that- concerned about energy efficiency,
1197 * just power the system down and be done with it!
1198 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1199 * permitted to sleep in dyntick-idle mode with only lazy RCU
1200 * callbacks pending. Setting this too high can OOM your system.
1202 * The values below work well in practice. If future workloads require
1203 * adjustment, they can be converted into kernel config parameters, though
1204 * making the state machine smarter might be a better option.
1206 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1207 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1209 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1210 module_param(rcu_idle_gp_delay, int, 0644);
1211 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1212 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1215 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1216 * only if it has been awhile since the last time we did so. Afterwards,
1217 * if there are any callbacks ready for immediate invocation, return true.
1219 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1221 bool cbs_ready = false;
1222 struct rcu_data *rdp;
1223 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1224 struct rcu_node *rnp;
1225 struct rcu_state *rsp;
1227 /* Exit early if we advanced recently. */
1228 if (jiffies == rdtp->last_advance_all)
1230 rdtp->last_advance_all = jiffies;
1232 for_each_rcu_flavor(rsp) {
1233 rdp = this_cpu_ptr(rsp->rda);
1237 * Don't bother checking unless a grace period has
1238 * completed since we last checked and there are
1239 * callbacks not yet ready to invoke.
1241 if ((rdp->completed != rnp->completed ||
1242 unlikely(READ_ONCE(rdp->gpwrap))) &&
1243 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1244 note_gp_changes(rsp, rdp);
1246 if (cpu_has_callbacks_ready_to_invoke(rdp))
1252 #ifndef CONFIG_PREEMPT_RT_FULL
1255 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1256 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1257 * caller to set the timeout based on whether or not there are non-lazy
1260 * The caller must have disabled interrupts.
1262 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1264 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1267 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1268 *nextevt = KTIME_MAX;
1272 /* Snapshot to detect later posting of non-lazy callback. */
1273 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1275 /* If no callbacks, RCU doesn't need the CPU. */
1276 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1277 *nextevt = KTIME_MAX;
1281 /* Attempt to advance callbacks. */
1282 if (rcu_try_advance_all_cbs()) {
1283 /* Some ready to invoke, so initiate later invocation. */
1287 rdtp->last_accelerate = jiffies;
1289 /* Request timer delay depending on laziness, and round. */
1290 if (!rdtp->all_lazy) {
1291 dj = round_up(rcu_idle_gp_delay + jiffies,
1292 rcu_idle_gp_delay) - jiffies;
1294 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1296 *nextevt = basemono + dj * TICK_NSEC;
1299 #endif /* #ifndef CONFIG_PREEMPT_RT_FULL */
1302 * Prepare a CPU for idle from an RCU perspective. The first major task
1303 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1304 * The second major task is to check to see if a non-lazy callback has
1305 * arrived at a CPU that previously had only lazy callbacks. The third
1306 * major task is to accelerate (that is, assign grace-period numbers to)
1307 * any recently arrived callbacks.
1309 * The caller must have disabled interrupts.
1311 static void rcu_prepare_for_idle(void)
1314 struct rcu_data *rdp;
1315 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1316 struct rcu_node *rnp;
1317 struct rcu_state *rsp;
1320 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1321 rcu_is_nocb_cpu(smp_processor_id()))
1324 /* Handle nohz enablement switches conservatively. */
1325 tne = READ_ONCE(tick_nohz_active);
1326 if (tne != rdtp->tick_nohz_enabled_snap) {
1327 if (rcu_cpu_has_callbacks(NULL))
1328 invoke_rcu_core(); /* force nohz to see update. */
1329 rdtp->tick_nohz_enabled_snap = tne;
1336 * If a non-lazy callback arrived at a CPU having only lazy
1337 * callbacks, invoke RCU core for the side-effect of recalculating
1338 * idle duration on re-entry to idle.
1340 if (rdtp->all_lazy &&
1341 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1342 rdtp->all_lazy = false;
1343 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1349 * If we have not yet accelerated this jiffy, accelerate all
1350 * callbacks on this CPU.
1352 if (rdtp->last_accelerate == jiffies)
1354 rdtp->last_accelerate = jiffies;
1355 for_each_rcu_flavor(rsp) {
1356 rdp = this_cpu_ptr(rsp->rda);
1357 if (!*rdp->nxttail[RCU_DONE_TAIL])
1360 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1361 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1362 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1364 rcu_gp_kthread_wake(rsp);
1369 * Clean up for exit from idle. Attempt to advance callbacks based on
1370 * any grace periods that elapsed while the CPU was idle, and if any
1371 * callbacks are now ready to invoke, initiate invocation.
1373 static void rcu_cleanup_after_idle(void)
1375 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1376 rcu_is_nocb_cpu(smp_processor_id()))
1378 if (rcu_try_advance_all_cbs())
1383 * Keep a running count of the number of non-lazy callbacks posted
1384 * on this CPU. This running counter (which is never decremented) allows
1385 * rcu_prepare_for_idle() to detect when something out of the idle loop
1386 * posts a callback, even if an equal number of callbacks are invoked.
1387 * Of course, callbacks should only be posted from within a trace event
1388 * designed to be called from idle or from within RCU_NONIDLE().
1390 static void rcu_idle_count_callbacks_posted(void)
1392 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1396 * Data for flushing lazy RCU callbacks at OOM time.
1398 static atomic_t oom_callback_count;
1399 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1402 * RCU OOM callback -- decrement the outstanding count and deliver the
1403 * wake-up if we are the last one.
1405 static void rcu_oom_callback(struct rcu_head *rhp)
1407 if (atomic_dec_and_test(&oom_callback_count))
1408 wake_up(&oom_callback_wq);
1412 * Post an rcu_oom_notify callback on the current CPU if it has at
1413 * least one lazy callback. This will unnecessarily post callbacks
1414 * to CPUs that already have a non-lazy callback at the end of their
1415 * callback list, but this is an infrequent operation, so accept some
1416 * extra overhead to keep things simple.
1418 static void rcu_oom_notify_cpu(void *unused)
1420 struct rcu_state *rsp;
1421 struct rcu_data *rdp;
1423 for_each_rcu_flavor(rsp) {
1424 rdp = raw_cpu_ptr(rsp->rda);
1425 if (rdp->qlen_lazy != 0) {
1426 atomic_inc(&oom_callback_count);
1427 rsp->call(&rdp->oom_head, rcu_oom_callback);
1433 * If low on memory, ensure that each CPU has a non-lazy callback.
1434 * This will wake up CPUs that have only lazy callbacks, in turn
1435 * ensuring that they free up the corresponding memory in a timely manner.
1436 * Because an uncertain amount of memory will be freed in some uncertain
1437 * timeframe, we do not claim to have freed anything.
1439 static int rcu_oom_notify(struct notifier_block *self,
1440 unsigned long notused, void *nfreed)
1444 /* Wait for callbacks from earlier instance to complete. */
1445 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1446 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1449 * Prevent premature wakeup: ensure that all increments happen
1450 * before there is a chance of the counter reaching zero.
1452 atomic_set(&oom_callback_count, 1);
1454 for_each_online_cpu(cpu) {
1455 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1456 cond_resched_rcu_qs();
1459 /* Unconditionally decrement: no need to wake ourselves up. */
1460 atomic_dec(&oom_callback_count);
1465 static struct notifier_block rcu_oom_nb = {
1466 .notifier_call = rcu_oom_notify
1469 static int __init rcu_register_oom_notifier(void)
1471 register_oom_notifier(&rcu_oom_nb);
1474 early_initcall(rcu_register_oom_notifier);
1476 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1478 #ifdef CONFIG_RCU_FAST_NO_HZ
1480 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1482 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1483 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1485 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1486 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1488 rdtp->all_lazy ? 'L' : '.',
1489 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1492 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1494 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1499 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1501 /* Initiate the stall-info list. */
1502 static void print_cpu_stall_info_begin(void)
1508 * Print out diagnostic information for the specified stalled CPU.
1510 * If the specified CPU is aware of the current RCU grace period
1511 * (flavor specified by rsp), then print the number of scheduling
1512 * clock interrupts the CPU has taken during the time that it has
1513 * been aware. Otherwise, print the number of RCU grace periods
1514 * that this CPU is ignorant of, for example, "1" if the CPU was
1515 * aware of the previous grace period.
1517 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1519 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1521 char fast_no_hz[72];
1522 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1523 struct rcu_dynticks *rdtp = rdp->dynticks;
1525 unsigned long ticks_value;
1527 if (rsp->gpnum == rdp->gpnum) {
1528 ticks_title = "ticks this GP";
1529 ticks_value = rdp->ticks_this_gp;
1531 ticks_title = "GPs behind";
1532 ticks_value = rsp->gpnum - rdp->gpnum;
1534 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1535 pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1537 "O."[!!cpu_online(cpu)],
1538 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1539 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1540 ticks_value, ticks_title,
1541 atomic_read(&rdtp->dynticks) & 0xfff,
1542 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1543 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1544 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1548 /* Terminate the stall-info list. */
1549 static void print_cpu_stall_info_end(void)
1554 /* Zero ->ticks_this_gp for all flavors of RCU. */
1555 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1557 rdp->ticks_this_gp = 0;
1558 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1561 /* Increment ->ticks_this_gp for all flavors of RCU. */
1562 static void increment_cpu_stall_ticks(void)
1564 struct rcu_state *rsp;
1566 for_each_rcu_flavor(rsp)
1567 raw_cpu_inc(rsp->rda->ticks_this_gp);
1570 #ifdef CONFIG_RCU_NOCB_CPU
1573 * Offload callback processing from the boot-time-specified set of CPUs
1574 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1575 * kthread created that pulls the callbacks from the corresponding CPU,
1576 * waits for a grace period to elapse, and invokes the callbacks.
1577 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1578 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1579 * has been specified, in which case each kthread actively polls its
1580 * CPU. (Which isn't so great for energy efficiency, but which does
1581 * reduce RCU's overhead on that CPU.)
1583 * This is intended to be used in conjunction with Frederic Weisbecker's
1584 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1585 * running CPU-bound user-mode computations.
1587 * Offloading of callback processing could also in theory be used as
1588 * an energy-efficiency measure because CPUs with no RCU callbacks
1589 * queued are more aggressive about entering dyntick-idle mode.
1593 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1594 static int __init rcu_nocb_setup(char *str)
1596 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1597 have_rcu_nocb_mask = true;
1598 cpulist_parse(str, rcu_nocb_mask);
1601 __setup("rcu_nocbs=", rcu_nocb_setup);
1603 static int __init parse_rcu_nocb_poll(char *arg)
1608 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1611 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1614 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1620 * Set the root rcu_node structure's ->need_future_gp field
1621 * based on the sum of those of all rcu_node structures. This does
1622 * double-count the root rcu_node structure's requests, but this
1623 * is necessary to handle the possibility of a rcu_nocb_kthread()
1624 * having awakened during the time that the rcu_node structures
1625 * were being updated for the end of the previous grace period.
1627 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1629 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1632 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1634 return &rnp->nocb_gp_wq[rnp->completed & 0x1];
1637 static void rcu_init_one_nocb(struct rcu_node *rnp)
1639 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1640 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1643 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1644 /* Is the specified CPU a no-CBs CPU? */
1645 bool rcu_is_nocb_cpu(int cpu)
1647 if (have_rcu_nocb_mask)
1648 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1651 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1654 * Kick the leader kthread for this NOCB group.
1656 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1658 struct rcu_data *rdp_leader = rdp->nocb_leader;
1660 if (!READ_ONCE(rdp_leader->nocb_kthread))
1662 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1663 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1664 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1665 swake_up(&rdp_leader->nocb_wq);
1670 * Does the specified CPU need an RCU callback for the specified flavor
1673 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1675 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1677 #ifdef CONFIG_PROVE_RCU
1678 struct rcu_head *rhp;
1679 #endif /* #ifdef CONFIG_PROVE_RCU */
1682 * Check count of all no-CBs callbacks awaiting invocation.
1683 * There needs to be a barrier before this function is called,
1684 * but associated with a prior determination that no more
1685 * callbacks would be posted. In the worst case, the first
1686 * barrier in _rcu_barrier() suffices (but the caller cannot
1687 * necessarily rely on this, not a substitute for the caller
1688 * getting the concurrency design right!). There must also be
1689 * a barrier between the following load an posting of a callback
1690 * (if a callback is in fact needed). This is associated with an
1691 * atomic_inc() in the caller.
1693 ret = atomic_long_read(&rdp->nocb_q_count);
1695 #ifdef CONFIG_PROVE_RCU
1696 rhp = READ_ONCE(rdp->nocb_head);
1698 rhp = READ_ONCE(rdp->nocb_gp_head);
1700 rhp = READ_ONCE(rdp->nocb_follower_head);
1702 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1703 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1704 rcu_scheduler_fully_active) {
1705 /* RCU callback enqueued before CPU first came online??? */
1706 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1710 #endif /* #ifdef CONFIG_PROVE_RCU */
1716 * Enqueue the specified string of rcu_head structures onto the specified
1717 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1718 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1719 * counts are supplied by rhcount and rhcount_lazy.
1721 * If warranted, also wake up the kthread servicing this CPUs queues.
1723 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1724 struct rcu_head *rhp,
1725 struct rcu_head **rhtp,
1726 int rhcount, int rhcount_lazy,
1727 unsigned long flags)
1730 struct rcu_head **old_rhpp;
1731 struct task_struct *t;
1733 /* Enqueue the callback on the nocb list and update counts. */
1734 atomic_long_add(rhcount, &rdp->nocb_q_count);
1735 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1736 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1737 WRITE_ONCE(*old_rhpp, rhp);
1738 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1739 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1741 /* If we are not being polled and there is a kthread, awaken it ... */
1742 t = READ_ONCE(rdp->nocb_kthread);
1743 if (rcu_nocb_poll || !t) {
1744 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1745 TPS("WakeNotPoll"));
1748 len = atomic_long_read(&rdp->nocb_q_count);
1749 if (old_rhpp == &rdp->nocb_head) {
1750 if (!irqs_disabled_flags(flags)) {
1751 /* ... if queue was empty ... */
1752 wake_nocb_leader(rdp, false);
1753 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1756 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1757 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1758 TPS("WakeEmptyIsDeferred"));
1760 rdp->qlen_last_fqs_check = 0;
1761 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1762 /* ... or if many callbacks queued. */
1763 if (!irqs_disabled_flags(flags)) {
1764 wake_nocb_leader(rdp, true);
1765 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1768 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1769 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1770 TPS("WakeOvfIsDeferred"));
1772 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1774 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1780 * This is a helper for __call_rcu(), which invokes this when the normal
1781 * callback queue is inoperable. If this is not a no-CBs CPU, this
1782 * function returns failure back to __call_rcu(), which can complain
1785 * Otherwise, this function queues the callback where the corresponding
1786 * "rcuo" kthread can find it.
1788 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1789 bool lazy, unsigned long flags)
1792 if (!rcu_is_nocb_cpu(rdp->cpu))
1794 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1795 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1796 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
1797 (unsigned long)rhp->func,
1798 -atomic_long_read(&rdp->nocb_q_count_lazy),
1799 -atomic_long_read(&rdp->nocb_q_count));
1801 trace_rcu_callback(rdp->rsp->name, rhp,
1802 -atomic_long_read(&rdp->nocb_q_count_lazy),
1803 -atomic_long_read(&rdp->nocb_q_count));
1806 * If called from an extended quiescent state with interrupts
1807 * disabled, invoke the RCU core in order to allow the idle-entry
1808 * deferred-wakeup check to function.
1810 if (irqs_disabled_flags(flags) &&
1811 !rcu_is_watching() &&
1812 cpu_online(smp_processor_id()))
1819 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
1822 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
1823 struct rcu_data *rdp,
1824 unsigned long flags)
1826 long ql = rsp->qlen;
1827 long qll = rsp->qlen_lazy;
1829 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
1830 if (!rcu_is_nocb_cpu(smp_processor_id()))
1835 /* First, enqueue the donelist, if any. This preserves CB ordering. */
1836 if (rsp->orphan_donelist != NULL) {
1837 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
1838 rsp->orphan_donetail, ql, qll, flags);
1840 rsp->orphan_donelist = NULL;
1841 rsp->orphan_donetail = &rsp->orphan_donelist;
1843 if (rsp->orphan_nxtlist != NULL) {
1844 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
1845 rsp->orphan_nxttail, ql, qll, flags);
1847 rsp->orphan_nxtlist = NULL;
1848 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1854 * If necessary, kick off a new grace period, and either way wait
1855 * for a subsequent grace period to complete.
1857 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
1861 unsigned long flags;
1863 struct rcu_node *rnp = rdp->mynode;
1865 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1866 needwake = rcu_start_future_gp(rnp, rdp, &c);
1867 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1869 rcu_gp_kthread_wake(rdp->rsp);
1872 * Wait for the grace period. Do so interruptibly to avoid messing
1873 * up the load average.
1875 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
1877 swait_event_interruptible(
1878 rnp->nocb_gp_wq[c & 0x1],
1879 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
1882 WARN_ON(signal_pending(current));
1883 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
1885 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
1886 smp_mb(); /* Ensure that CB invocation happens after GP end. */
1890 * Leaders come here to wait for additional callbacks to show up.
1891 * This function does not return until callbacks appear.
1893 static void nocb_leader_wait(struct rcu_data *my_rdp)
1895 bool firsttime = true;
1897 struct rcu_data *rdp;
1898 struct rcu_head **tail;
1902 /* Wait for callbacks to appear. */
1903 if (!rcu_nocb_poll) {
1904 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
1905 swait_event_interruptible(my_rdp->nocb_wq,
1906 !READ_ONCE(my_rdp->nocb_leader_sleep));
1907 /* Memory barrier handled by smp_mb() calls below and repoll. */
1908 } else if (firsttime) {
1909 firsttime = false; /* Don't drown trace log with "Poll"! */
1910 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
1914 * Each pass through the following loop checks a follower for CBs.
1915 * We are our own first follower. Any CBs found are moved to
1916 * nocb_gp_head, where they await a grace period.
1919 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
1920 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
1921 if (!rdp->nocb_gp_head)
1922 continue; /* No CBs here, try next follower. */
1924 /* Move callbacks to wait-for-GP list, which is empty. */
1925 WRITE_ONCE(rdp->nocb_head, NULL);
1926 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
1931 * If there were no callbacks, sleep a bit, rescan after a
1932 * memory barrier, and go retry.
1934 if (unlikely(!gotcbs)) {
1936 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
1938 WARN_ON(signal_pending(current));
1939 schedule_timeout_interruptible(1);
1941 /* Rescan in case we were a victim of memory ordering. */
1942 my_rdp->nocb_leader_sleep = true;
1943 smp_mb(); /* Ensure _sleep true before scan. */
1944 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
1945 if (READ_ONCE(rdp->nocb_head)) {
1946 /* Found CB, so short-circuit next wait. */
1947 my_rdp->nocb_leader_sleep = false;
1953 /* Wait for one grace period. */
1954 rcu_nocb_wait_gp(my_rdp);
1957 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
1958 * We set it now, but recheck for new callbacks while
1959 * traversing our follower list.
1961 my_rdp->nocb_leader_sleep = true;
1962 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
1964 /* Each pass through the following loop wakes a follower, if needed. */
1965 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
1966 if (READ_ONCE(rdp->nocb_head))
1967 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
1968 if (!rdp->nocb_gp_head)
1969 continue; /* No CBs, so no need to wake follower. */
1971 /* Append callbacks to follower's "done" list. */
1972 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
1973 *tail = rdp->nocb_gp_head;
1974 smp_mb__after_atomic(); /* Store *tail before wakeup. */
1975 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
1977 * List was empty, wake up the follower.
1978 * Memory barriers supplied by atomic_long_add().
1980 swake_up(&rdp->nocb_wq);
1984 /* If we (the leader) don't have CBs, go wait some more. */
1985 if (!my_rdp->nocb_follower_head)
1990 * Followers come here to wait for additional callbacks to show up.
1991 * This function does not return until callbacks appear.
1993 static void nocb_follower_wait(struct rcu_data *rdp)
1995 bool firsttime = true;
1998 if (!rcu_nocb_poll) {
1999 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2001 swait_event_interruptible(rdp->nocb_wq,
2002 READ_ONCE(rdp->nocb_follower_head));
2003 } else if (firsttime) {
2004 /* Don't drown trace log with "Poll"! */
2006 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2008 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2009 /* ^^^ Ensure CB invocation follows _head test. */
2013 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2015 WARN_ON(signal_pending(current));
2016 schedule_timeout_interruptible(1);
2021 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2022 * callbacks queued by the corresponding no-CBs CPU, however, there is
2023 * an optional leader-follower relationship so that the grace-period
2024 * kthreads don't have to do quite so many wakeups.
2026 static int rcu_nocb_kthread(void *arg)
2029 struct rcu_head *list;
2030 struct rcu_head *next;
2031 struct rcu_head **tail;
2032 struct rcu_data *rdp = arg;
2034 /* Each pass through this loop invokes one batch of callbacks */
2036 /* Wait for callbacks. */
2037 if (rdp->nocb_leader == rdp)
2038 nocb_leader_wait(rdp);
2040 nocb_follower_wait(rdp);
2042 /* Pull the ready-to-invoke callbacks onto local list. */
2043 list = READ_ONCE(rdp->nocb_follower_head);
2045 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2046 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2047 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2049 /* Each pass through the following loop invokes a callback. */
2050 trace_rcu_batch_start(rdp->rsp->name,
2051 atomic_long_read(&rdp->nocb_q_count_lazy),
2052 atomic_long_read(&rdp->nocb_q_count), -1);
2056 /* Wait for enqueuing to complete, if needed. */
2057 while (next == NULL && &list->next != tail) {
2058 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2060 schedule_timeout_interruptible(1);
2061 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2065 debug_rcu_head_unqueue(list);
2067 if (__rcu_reclaim(rdp->rsp->name, list))
2071 cond_resched_rcu_qs();
2074 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2075 smp_mb__before_atomic(); /* _add after CB invocation. */
2076 atomic_long_add(-c, &rdp->nocb_q_count);
2077 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2078 rdp->n_nocbs_invoked += c;
2083 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2084 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2086 return READ_ONCE(rdp->nocb_defer_wakeup);
2089 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2090 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2094 if (!rcu_nocb_need_deferred_wakeup(rdp))
2096 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2097 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2098 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2099 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2102 void __init rcu_init_nohz(void)
2105 bool need_rcu_nocb_mask = true;
2106 struct rcu_state *rsp;
2108 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2109 need_rcu_nocb_mask = false;
2110 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2112 #if defined(CONFIG_NO_HZ_FULL)
2113 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2114 need_rcu_nocb_mask = true;
2115 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2117 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2118 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2119 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2122 have_rcu_nocb_mask = true;
2124 if (!have_rcu_nocb_mask)
2127 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2128 pr_info("\tOffload RCU callbacks from CPU 0\n");
2129 cpumask_set_cpu(0, rcu_nocb_mask);
2130 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2131 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2132 pr_info("\tOffload RCU callbacks from all CPUs\n");
2133 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2134 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2135 #if defined(CONFIG_NO_HZ_FULL)
2136 if (tick_nohz_full_running)
2137 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2138 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2140 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2141 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2142 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2145 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2146 cpumask_pr_args(rcu_nocb_mask));
2148 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2150 for_each_rcu_flavor(rsp) {
2151 for_each_cpu(cpu, rcu_nocb_mask)
2152 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2153 rcu_organize_nocb_kthreads(rsp);
2157 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2158 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2160 rdp->nocb_tail = &rdp->nocb_head;
2161 init_swait_queue_head(&rdp->nocb_wq);
2162 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2166 * If the specified CPU is a no-CBs CPU that does not already have its
2167 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2168 * brought online out of order, this can require re-organizing the
2169 * leader-follower relationships.
2171 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2173 struct rcu_data *rdp;
2174 struct rcu_data *rdp_last;
2175 struct rcu_data *rdp_old_leader;
2176 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2177 struct task_struct *t;
2180 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2181 * then nothing to do.
2183 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2186 /* If we didn't spawn the leader first, reorganize! */
2187 rdp_old_leader = rdp_spawn->nocb_leader;
2188 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2190 rdp = rdp_old_leader;
2192 rdp->nocb_leader = rdp_spawn;
2193 if (rdp_last && rdp != rdp_spawn)
2194 rdp_last->nocb_next_follower = rdp;
2195 if (rdp == rdp_spawn) {
2196 rdp = rdp->nocb_next_follower;
2199 rdp = rdp->nocb_next_follower;
2200 rdp_last->nocb_next_follower = NULL;
2203 rdp_spawn->nocb_next_follower = rdp_old_leader;
2206 /* Spawn the kthread for this CPU and RCU flavor. */
2207 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2208 "rcuo%c/%d", rsp->abbr, cpu);
2210 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2214 * If the specified CPU is a no-CBs CPU that does not already have its
2215 * rcuo kthreads, spawn them.
2217 static void rcu_spawn_all_nocb_kthreads(int cpu)
2219 struct rcu_state *rsp;
2221 if (rcu_scheduler_fully_active)
2222 for_each_rcu_flavor(rsp)
2223 rcu_spawn_one_nocb_kthread(rsp, cpu);
2227 * Once the scheduler is running, spawn rcuo kthreads for all online
2228 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2229 * non-boot CPUs come online -- if this changes, we will need to add
2230 * some mutual exclusion.
2232 static void __init rcu_spawn_nocb_kthreads(void)
2236 for_each_online_cpu(cpu)
2237 rcu_spawn_all_nocb_kthreads(cpu);
2240 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2241 static int rcu_nocb_leader_stride = -1;
2242 module_param(rcu_nocb_leader_stride, int, 0444);
2245 * Initialize leader-follower relationships for all no-CBs CPU.
2247 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2250 int ls = rcu_nocb_leader_stride;
2251 int nl = 0; /* Next leader. */
2252 struct rcu_data *rdp;
2253 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2254 struct rcu_data *rdp_prev = NULL;
2256 if (!have_rcu_nocb_mask)
2259 ls = int_sqrt(nr_cpu_ids);
2260 rcu_nocb_leader_stride = ls;
2264 * Each pass through this loop sets up one rcu_data structure and
2265 * spawns one rcu_nocb_kthread().
2267 for_each_cpu(cpu, rcu_nocb_mask) {
2268 rdp = per_cpu_ptr(rsp->rda, cpu);
2269 if (rdp->cpu >= nl) {
2270 /* New leader, set up for followers & next leader. */
2271 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2272 rdp->nocb_leader = rdp;
2275 /* Another follower, link to previous leader. */
2276 rdp->nocb_leader = rdp_leader;
2277 rdp_prev->nocb_next_follower = rdp;
2283 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2284 static bool init_nocb_callback_list(struct rcu_data *rdp)
2286 if (!rcu_is_nocb_cpu(rdp->cpu))
2289 /* If there are early-boot callbacks, move them to nocb lists. */
2291 rdp->nocb_head = rdp->nxtlist;
2292 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2293 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2294 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2295 rdp->nxtlist = NULL;
2299 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2303 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2305 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2307 WARN_ON_ONCE(1); /* Should be dead code. */
2311 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2315 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2319 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2324 static void rcu_init_one_nocb(struct rcu_node *rnp)
2328 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2329 bool lazy, unsigned long flags)
2334 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2335 struct rcu_data *rdp,
2336 unsigned long flags)
2341 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2345 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2350 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2354 static void rcu_spawn_all_nocb_kthreads(int cpu)
2358 static void __init rcu_spawn_nocb_kthreads(void)
2362 static bool init_nocb_callback_list(struct rcu_data *rdp)
2367 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2370 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2371 * arbitrarily long period of time with the scheduling-clock tick turned
2372 * off. RCU will be paying attention to this CPU because it is in the
2373 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2374 * machine because the scheduling-clock tick has been disabled. Therefore,
2375 * if an adaptive-ticks CPU is failing to respond to the current grace
2376 * period and has not be idle from an RCU perspective, kick it.
2378 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2380 #ifdef CONFIG_NO_HZ_FULL
2381 if (tick_nohz_full_cpu(cpu))
2382 smp_send_reschedule(cpu);
2383 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2387 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2389 static int full_sysidle_state; /* Current system-idle state. */
2390 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2391 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2392 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2393 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2394 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2397 * Invoked to note exit from irq or task transition to idle. Note that
2398 * usermode execution does -not- count as idle here! After all, we want
2399 * to detect full-system idle states, not RCU quiescent states and grace
2400 * periods. The caller must have disabled interrupts.
2402 static void rcu_sysidle_enter(int irq)
2405 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2407 /* If there are no nohz_full= CPUs, no need to track this. */
2408 if (!tick_nohz_full_enabled())
2411 /* Adjust nesting, check for fully idle. */
2413 rdtp->dynticks_idle_nesting--;
2414 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2415 if (rdtp->dynticks_idle_nesting != 0)
2416 return; /* Still not fully idle. */
2418 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2419 DYNTICK_TASK_NEST_VALUE) {
2420 rdtp->dynticks_idle_nesting = 0;
2422 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2423 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2424 return; /* Still not fully idle. */
2428 /* Record start of fully idle period. */
2430 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2431 smp_mb__before_atomic();
2432 atomic_inc(&rdtp->dynticks_idle);
2433 smp_mb__after_atomic();
2434 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2438 * Unconditionally force exit from full system-idle state. This is
2439 * invoked when a normal CPU exits idle, but must be called separately
2440 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2441 * is that the timekeeping CPU is permitted to take scheduling-clock
2442 * interrupts while the system is in system-idle state, and of course
2443 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2444 * interrupt from any other type of interrupt.
2446 void rcu_sysidle_force_exit(void)
2448 int oldstate = READ_ONCE(full_sysidle_state);
2452 * Each pass through the following loop attempts to exit full
2453 * system-idle state. If contention proves to be a problem,
2454 * a trylock-based contention tree could be used here.
2456 while (oldstate > RCU_SYSIDLE_SHORT) {
2457 newoldstate = cmpxchg(&full_sysidle_state,
2458 oldstate, RCU_SYSIDLE_NOT);
2459 if (oldstate == newoldstate &&
2460 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2461 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2462 return; /* We cleared it, done! */
2464 oldstate = newoldstate;
2466 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2470 * Invoked to note entry to irq or task transition from idle. Note that
2471 * usermode execution does -not- count as idle here! The caller must
2472 * have disabled interrupts.
2474 static void rcu_sysidle_exit(int irq)
2476 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2478 /* If there are no nohz_full= CPUs, no need to track this. */
2479 if (!tick_nohz_full_enabled())
2482 /* Adjust nesting, check for already non-idle. */
2484 rdtp->dynticks_idle_nesting++;
2485 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2486 if (rdtp->dynticks_idle_nesting != 1)
2487 return; /* Already non-idle. */
2490 * Allow for irq misnesting. Yes, it really is possible
2491 * to enter an irq handler then never leave it, and maybe
2492 * also vice versa. Handle both possibilities.
2494 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2495 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2496 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2497 return; /* Already non-idle. */
2499 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2503 /* Record end of idle period. */
2504 smp_mb__before_atomic();
2505 atomic_inc(&rdtp->dynticks_idle);
2506 smp_mb__after_atomic();
2507 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2510 * If we are the timekeeping CPU, we are permitted to be non-idle
2511 * during a system-idle state. This must be the case, because
2512 * the timekeeping CPU has to take scheduling-clock interrupts
2513 * during the time that the system is transitioning to full
2514 * system-idle state. This means that the timekeeping CPU must
2515 * invoke rcu_sysidle_force_exit() directly if it does anything
2516 * more than take a scheduling-clock interrupt.
2518 if (smp_processor_id() == tick_do_timer_cpu)
2521 /* Update system-idle state: We are clearly no longer fully idle! */
2522 rcu_sysidle_force_exit();
2526 * Check to see if the current CPU is idle. Note that usermode execution
2527 * does not count as idle. The caller must have disabled interrupts,
2528 * and must be running on tick_do_timer_cpu.
2530 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2531 unsigned long *maxj)
2535 struct rcu_dynticks *rdtp = rdp->dynticks;
2537 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2538 if (!tick_nohz_full_enabled())
2542 * If some other CPU has already reported non-idle, if this is
2543 * not the flavor of RCU that tracks sysidle state, or if this
2544 * is an offline or the timekeeping CPU, nothing to do.
2546 if (!*isidle || rdp->rsp != rcu_state_p ||
2547 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2549 /* Verify affinity of current kthread. */
2550 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2552 /* Pick up current idle and NMI-nesting counter and check. */
2553 cur = atomic_read(&rdtp->dynticks_idle);
2555 *isidle = false; /* We are not idle! */
2558 smp_mb(); /* Read counters before timestamps. */
2560 /* Pick up timestamps. */
2561 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2562 /* If this CPU entered idle more recently, update maxj timestamp. */
2563 if (ULONG_CMP_LT(*maxj, j))
2568 * Is this the flavor of RCU that is handling full-system idle?
2570 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2572 return rsp == rcu_state_p;
2576 * Return a delay in jiffies based on the number of CPUs, rcu_node
2577 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2578 * systems more time to transition to full-idle state in order to
2579 * avoid the cache thrashing that otherwise occur on the state variable.
2580 * Really small systems (less than a couple of tens of CPUs) should
2581 * instead use a single global atomically incremented counter, and later
2582 * versions of this will automatically reconfigure themselves accordingly.
2584 static unsigned long rcu_sysidle_delay(void)
2586 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2588 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2592 * Advance the full-system-idle state. This is invoked when all of
2593 * the non-timekeeping CPUs are idle.
2595 static void rcu_sysidle(unsigned long j)
2597 /* Check the current state. */
2598 switch (READ_ONCE(full_sysidle_state)) {
2599 case RCU_SYSIDLE_NOT:
2601 /* First time all are idle, so note a short idle period. */
2602 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2605 case RCU_SYSIDLE_SHORT:
2608 * Idle for a bit, time to advance to next state?
2609 * cmpxchg failure means race with non-idle, let them win.
2611 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2612 (void)cmpxchg(&full_sysidle_state,
2613 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2616 case RCU_SYSIDLE_LONG:
2619 * Do an additional check pass before advancing to full.
2620 * cmpxchg failure means race with non-idle, let them win.
2622 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2623 (void)cmpxchg(&full_sysidle_state,
2624 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2633 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2634 * back to the beginning.
2636 static void rcu_sysidle_cancel(void)
2639 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2640 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2644 * Update the sysidle state based on the results of a force-quiescent-state
2645 * scan of the CPUs' dyntick-idle state.
2647 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2648 unsigned long maxj, bool gpkt)
2650 if (rsp != rcu_state_p)
2651 return; /* Wrong flavor, ignore. */
2652 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2653 return; /* Running state machine from timekeeping CPU. */
2655 rcu_sysidle(maxj); /* More idle! */
2657 rcu_sysidle_cancel(); /* Idle is over. */
2661 * Wrapper for rcu_sysidle_report() when called from the grace-period
2662 * kthread's context.
2664 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2667 /* If there are no nohz_full= CPUs, no need to track this. */
2668 if (!tick_nohz_full_enabled())
2671 rcu_sysidle_report(rsp, isidle, maxj, true);
2674 /* Callback and function for forcing an RCU grace period. */
2675 struct rcu_sysidle_head {
2680 static void rcu_sysidle_cb(struct rcu_head *rhp)
2682 struct rcu_sysidle_head *rshp;
2685 * The following memory barrier is needed to replace the
2686 * memory barriers that would normally be in the memory
2689 smp_mb(); /* grace period precedes setting inuse. */
2691 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2692 WRITE_ONCE(rshp->inuse, 0);
2696 * Check to see if the system is fully idle, other than the timekeeping CPU.
2697 * The caller must have disabled interrupts. This is not intended to be
2698 * called unless tick_nohz_full_enabled().
2700 bool rcu_sys_is_idle(void)
2702 static struct rcu_sysidle_head rsh;
2703 int rss = READ_ONCE(full_sysidle_state);
2705 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2708 /* Handle small-system case by doing a full scan of CPUs. */
2709 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2710 int oldrss = rss - 1;
2713 * One pass to advance to each state up to _FULL.
2714 * Give up if any pass fails to advance the state.
2716 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2719 unsigned long maxj = jiffies - ULONG_MAX / 4;
2720 struct rcu_data *rdp;
2722 /* Scan all the CPUs looking for nonidle CPUs. */
2723 for_each_possible_cpu(cpu) {
2724 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2725 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2729 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2731 rss = READ_ONCE(full_sysidle_state);
2735 /* If this is the first observation of an idle period, record it. */
2736 if (rss == RCU_SYSIDLE_FULL) {
2737 rss = cmpxchg(&full_sysidle_state,
2738 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2739 return rss == RCU_SYSIDLE_FULL;
2742 smp_mb(); /* ensure rss load happens before later caller actions. */
2744 /* If already fully idle, tell the caller (in case of races). */
2745 if (rss == RCU_SYSIDLE_FULL_NOTED)
2749 * If we aren't there yet, and a grace period is not in flight,
2750 * initiate a grace period. Either way, tell the caller that
2751 * we are not there yet. We use an xchg() rather than an assignment
2752 * to make up for the memory barriers that would otherwise be
2753 * provided by the memory allocator.
2755 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2756 !rcu_gp_in_progress(rcu_state_p) &&
2757 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2758 call_rcu(&rsh.rh, rcu_sysidle_cb);
2763 * Initialize dynticks sysidle state for CPUs coming online.
2765 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2767 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2770 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2772 static void rcu_sysidle_enter(int irq)
2776 static void rcu_sysidle_exit(int irq)
2780 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2781 unsigned long *maxj)
2785 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2790 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2795 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2799 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2802 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2803 * grace-period kthread will do force_quiescent_state() processing?
2804 * The idea is to avoid waking up RCU core processing on such a
2805 * CPU unless the grace period has extended for too long.
2807 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2808 * CONFIG_RCU_NOCB_CPU CPUs.
2810 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2812 #ifdef CONFIG_NO_HZ_FULL
2813 if (tick_nohz_full_cpu(smp_processor_id()) &&
2814 (!rcu_gp_in_progress(rsp) ||
2815 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
2817 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2822 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2825 static void rcu_bind_gp_kthread(void)
2827 int __maybe_unused cpu;
2829 if (!tick_nohz_full_enabled())
2831 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2832 cpu = tick_do_timer_cpu;
2833 if (cpu >= 0 && cpu < nr_cpu_ids)
2834 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2835 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2836 housekeeping_affine(current);
2837 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2840 /* Record the current task on dyntick-idle entry. */
2841 static void rcu_dynticks_task_enter(void)
2843 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2844 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2845 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2848 /* Record no current task on dyntick-idle exit. */
2849 static void rcu_dynticks_task_exit(void)
2851 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2852 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2853 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */