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. If you like #ifdef, you
60 * will love this function.
62 static void __init rcu_bootup_announce_oddness(void)
64 if (IS_ENABLED(CONFIG_RCU_TRACE))
65 pr_info("\tRCU debugfs-based tracing is enabled.\n");
66 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
67 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
68 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
71 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
72 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
73 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
74 if (IS_ENABLED(CONFIG_PROVE_RCU))
75 pr_info("\tRCU lockdep checking is enabled.\n");
76 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE))
77 pr_info("\tRCU torture testing starts during boot.\n");
78 if (RCU_NUM_LVLS >= 4)
79 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
80 if (RCU_FANOUT_LEAF != 16)
81 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
83 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
84 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
85 if (nr_cpu_ids != NR_CPUS)
86 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
87 if (IS_ENABLED(CONFIG_RCU_BOOST))
88 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
91 #ifdef CONFIG_PREEMPT_RCU
93 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
94 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
95 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
97 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
101 * Tell them what RCU they are running.
103 static void __init rcu_bootup_announce(void)
105 pr_info("Preemptible hierarchical RCU implementation.\n");
106 rcu_bootup_announce_oddness();
109 /* Flags for rcu_preempt_ctxt_queue() decision table. */
110 #define RCU_GP_TASKS 0x8
111 #define RCU_EXP_TASKS 0x4
112 #define RCU_GP_BLKD 0x2
113 #define RCU_EXP_BLKD 0x1
116 * Queues a task preempted within an RCU-preempt read-side critical
117 * section into the appropriate location within the ->blkd_tasks list,
118 * depending on the states of any ongoing normal and expedited grace
119 * periods. The ->gp_tasks pointer indicates which element the normal
120 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
121 * indicates which element the expedited grace period is waiting on (again,
122 * NULL if none). If a grace period is waiting on a given element in the
123 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
124 * adding a task to the tail of the list blocks any grace period that is
125 * already waiting on one of the elements. In contrast, adding a task
126 * to the head of the list won't block any grace period that is already
127 * waiting on one of the elements.
129 * This queuing is imprecise, and can sometimes make an ongoing grace
130 * period wait for a task that is not strictly speaking blocking it.
131 * Given the choice, we needlessly block a normal grace period rather than
132 * blocking an expedited grace period.
134 * Note that an endless sequence of expedited grace periods still cannot
135 * indefinitely postpone a normal grace period. Eventually, all of the
136 * fixed number of preempted tasks blocking the normal grace period that are
137 * not also blocking the expedited grace period will resume and complete
138 * their RCU read-side critical sections. At that point, the ->gp_tasks
139 * pointer will equal the ->exp_tasks pointer, at which point the end of
140 * the corresponding expedited grace period will also be the end of the
141 * normal grace period.
143 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp,
144 unsigned long flags) __releases(rnp->lock)
146 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
147 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
148 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
149 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
150 struct task_struct *t = current;
153 * Decide where to queue the newly blocked task. In theory,
154 * this could be an if-statement. In practice, when I tried
155 * that, it was quite messy.
157 switch (blkd_state) {
160 case RCU_EXP_TASKS + RCU_GP_BLKD:
162 case RCU_GP_TASKS + RCU_EXP_TASKS:
165 * Blocking neither GP, or first task blocking the normal
166 * GP but not blocking the already-waiting expedited GP.
167 * Queue at the head of the list to avoid unnecessarily
168 * blocking the already-waiting GPs.
170 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
175 case RCU_GP_BLKD + RCU_EXP_BLKD:
176 case RCU_GP_TASKS + RCU_EXP_BLKD:
177 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
178 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
181 * First task arriving that blocks either GP, or first task
182 * arriving that blocks the expedited GP (with the normal
183 * GP already waiting), or a task arriving that blocks
184 * both GPs with both GPs already waiting. Queue at the
185 * tail of the list to avoid any GP waiting on any of the
186 * already queued tasks that are not blocking it.
188 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
191 case RCU_EXP_TASKS + RCU_EXP_BLKD:
192 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
193 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
196 * Second or subsequent task blocking the expedited GP.
197 * The task either does not block the normal GP, or is the
198 * first task blocking the normal GP. Queue just after
199 * the first task blocking the expedited GP.
201 list_add(&t->rcu_node_entry, rnp->exp_tasks);
204 case RCU_GP_TASKS + RCU_GP_BLKD:
205 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
208 * Second or subsequent task blocking the normal GP.
209 * The task does not block the expedited GP. Queue just
210 * after the first task blocking the normal GP.
212 list_add(&t->rcu_node_entry, rnp->gp_tasks);
217 /* Yet another exercise in excessive paranoia. */
223 * We have now queued the task. If it was the first one to
224 * block either grace period, update the ->gp_tasks and/or
225 * ->exp_tasks pointers, respectively, to reference the newly
228 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD))
229 rnp->gp_tasks = &t->rcu_node_entry;
230 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
231 rnp->exp_tasks = &t->rcu_node_entry;
232 raw_spin_unlock(&rnp->lock);
235 * Report the quiescent state for the expedited GP. This expedited
236 * GP should not be able to end until we report, so there should be
237 * no need to check for a subsequent expedited GP. (Though we are
238 * still in a quiescent state in any case.)
240 if (blkd_state & RCU_EXP_BLKD &&
241 t->rcu_read_unlock_special.b.exp_need_qs) {
242 t->rcu_read_unlock_special.b.exp_need_qs = false;
243 rcu_report_exp_rdp(rdp->rsp, rdp, true);
245 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
247 local_irq_restore(flags);
251 * Record a preemptible-RCU quiescent state for the specified CPU. Note
252 * that this just means that the task currently running on the CPU is
253 * not in a quiescent state. There might be any number of tasks blocked
254 * while in an RCU read-side critical section.
256 * As with the other rcu_*_qs() functions, callers to this function
257 * must disable preemption.
259 static void rcu_preempt_qs(void)
261 if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
262 trace_rcu_grace_period(TPS("rcu_preempt"),
263 __this_cpu_read(rcu_data_p->gpnum),
265 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
266 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
267 current->rcu_read_unlock_special.b.need_qs = false;
272 * We have entered the scheduler, and the current task might soon be
273 * context-switched away from. If this task is in an RCU read-side
274 * critical section, we will no longer be able to rely on the CPU to
275 * record that fact, so we enqueue the task on the blkd_tasks list.
276 * The task will dequeue itself when it exits the outermost enclosing
277 * RCU read-side critical section. Therefore, the current grace period
278 * cannot be permitted to complete until the blkd_tasks list entries
279 * predating the current grace period drain, in other words, until
280 * rnp->gp_tasks becomes NULL.
282 * Caller must disable preemption.
284 static void rcu_preempt_note_context_switch(void)
286 struct task_struct *t = current;
288 struct rcu_data *rdp;
289 struct rcu_node *rnp;
291 if (t->rcu_read_lock_nesting > 0 &&
292 !t->rcu_read_unlock_special.b.blocked) {
294 /* Possibly blocking in an RCU read-side critical section. */
295 rdp = this_cpu_ptr(rcu_state_p->rda);
297 raw_spin_lock_irqsave(&rnp->lock, flags);
298 smp_mb__after_unlock_lock();
299 t->rcu_read_unlock_special.b.blocked = true;
300 t->rcu_blocked_node = rnp;
303 * Verify the CPU's sanity, trace the preemption, and
304 * then queue the task as required based on the states
305 * of any ongoing and expedited grace periods.
307 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
308 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
309 trace_rcu_preempt_task(rdp->rsp->name,
311 (rnp->qsmask & rdp->grpmask)
314 rcu_preempt_ctxt_queue(rnp, rdp, flags);
315 } else if (t->rcu_read_lock_nesting < 0 &&
316 t->rcu_read_unlock_special.s) {
319 * Complete exit from RCU read-side critical section on
320 * behalf of preempted instance of __rcu_read_unlock().
322 rcu_read_unlock_special(t);
326 * Either we were not in an RCU read-side critical section to
327 * begin with, or we have now recorded that critical section
328 * globally. Either way, we can now note a quiescent state
329 * for this CPU. Again, if we were in an RCU read-side critical
330 * section, and if that critical section was blocking the current
331 * grace period, then the fact that the task has been enqueued
332 * means that we continue to block the current grace period.
338 * Check for preempted RCU readers blocking the current grace period
339 * for the specified rcu_node structure. If the caller needs a reliable
340 * answer, it must hold the rcu_node's ->lock.
342 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
344 return rnp->gp_tasks != NULL;
348 * Advance a ->blkd_tasks-list pointer to the next entry, instead
349 * returning NULL if at the end of the list.
351 static struct list_head *rcu_next_node_entry(struct task_struct *t,
352 struct rcu_node *rnp)
354 struct list_head *np;
356 np = t->rcu_node_entry.next;
357 if (np == &rnp->blkd_tasks)
363 * Return true if the specified rcu_node structure has tasks that were
364 * preempted within an RCU read-side critical section.
366 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
368 return !list_empty(&rnp->blkd_tasks);
372 * Handle special cases during rcu_read_unlock(), such as needing to
373 * notify RCU core processing or task having blocked during the RCU
374 * read-side critical section.
376 void rcu_read_unlock_special(struct task_struct *t)
382 struct list_head *np;
383 bool drop_boost_mutex = false;
384 struct rcu_data *rdp;
385 struct rcu_node *rnp;
386 union rcu_special special;
388 /* NMI handlers cannot block and cannot safely manipulate state. */
392 local_irq_save(flags);
395 * If RCU core is waiting for this CPU to exit its critical section,
396 * report the fact that it has exited. Because irqs are disabled,
397 * t->rcu_read_unlock_special cannot change.
399 special = t->rcu_read_unlock_special;
400 if (special.b.need_qs) {
402 t->rcu_read_unlock_special.b.need_qs = false;
403 if (!t->rcu_read_unlock_special.s) {
404 local_irq_restore(flags);
410 * Respond to a request for an expedited grace period, but only if
411 * we were not preempted, meaning that we were running on the same
412 * CPU throughout. If we were preempted, the exp_need_qs flag
413 * would have been cleared at the time of the first preemption,
414 * and the quiescent state would be reported when we were dequeued.
416 if (special.b.exp_need_qs) {
417 WARN_ON_ONCE(special.b.blocked);
418 t->rcu_read_unlock_special.b.exp_need_qs = false;
419 rdp = this_cpu_ptr(rcu_state_p->rda);
420 rcu_report_exp_rdp(rcu_state_p, rdp, true);
421 if (!t->rcu_read_unlock_special.s) {
422 local_irq_restore(flags);
427 /* Hardware IRQ handlers cannot block, complain if they get here. */
428 if (preempt_count() & (HARDIRQ_MASK | SOFTIRQ_OFFSET)) {
429 lockdep_rcu_suspicious(__FILE__, __LINE__,
430 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
431 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
432 t->rcu_read_unlock_special.s,
433 t->rcu_read_unlock_special.b.blocked,
434 t->rcu_read_unlock_special.b.exp_need_qs,
435 t->rcu_read_unlock_special.b.need_qs);
436 local_irq_restore(flags);
440 /* Clean up if blocked during RCU read-side critical section. */
441 if (special.b.blocked) {
442 t->rcu_read_unlock_special.b.blocked = false;
445 * Remove this task from the list it blocked on. The task
446 * now remains queued on the rcu_node corresponding to
447 * the CPU it first blocked on, so the first attempt to
448 * acquire the task's rcu_node's ->lock will succeed.
449 * Keep the loop and add a WARN_ON() out of sheer paranoia.
452 rnp = t->rcu_blocked_node;
453 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
454 smp_mb__after_unlock_lock();
455 if (rnp == t->rcu_blocked_node)
458 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
460 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
461 empty_exp = sync_rcu_preempt_exp_done(rnp);
462 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
463 np = rcu_next_node_entry(t, rnp);
464 list_del_init(&t->rcu_node_entry);
465 t->rcu_blocked_node = NULL;
466 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
468 if (&t->rcu_node_entry == rnp->gp_tasks)
470 if (&t->rcu_node_entry == rnp->exp_tasks)
472 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
473 if (&t->rcu_node_entry == rnp->boost_tasks)
474 rnp->boost_tasks = np;
475 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
476 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
480 * If this was the last task on the current list, and if
481 * we aren't waiting on any CPUs, report the quiescent state.
482 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
483 * so we must take a snapshot of the expedited state.
485 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
486 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
487 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
494 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
496 raw_spin_unlock_irqrestore(&rnp->lock, flags);
499 /* Unboost if we were boosted. */
500 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
501 rt_mutex_unlock(&rnp->boost_mtx);
504 * If this was the last task on the expedited lists,
505 * then we need to report up the rcu_node hierarchy.
507 if (!empty_exp && empty_exp_now)
508 rcu_report_exp_rnp(rcu_state_p, rnp, true);
510 local_irq_restore(flags);
515 * Dump detailed information for all tasks blocking the current RCU
516 * grace period on the specified rcu_node structure.
518 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
521 struct task_struct *t;
523 raw_spin_lock_irqsave(&rnp->lock, flags);
524 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
525 raw_spin_unlock_irqrestore(&rnp->lock, flags);
528 t = list_entry(rnp->gp_tasks->prev,
529 struct task_struct, rcu_node_entry);
530 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
532 raw_spin_unlock_irqrestore(&rnp->lock, flags);
536 * Dump detailed information for all tasks blocking the current RCU
539 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
541 struct rcu_node *rnp = rcu_get_root(rsp);
543 rcu_print_detail_task_stall_rnp(rnp);
544 rcu_for_each_leaf_node(rsp, rnp)
545 rcu_print_detail_task_stall_rnp(rnp);
548 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
550 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
551 rnp->level, rnp->grplo, rnp->grphi);
554 static void rcu_print_task_stall_end(void)
560 * Scan the current list of tasks blocked within RCU read-side critical
561 * sections, printing out the tid of each.
563 static int rcu_print_task_stall(struct rcu_node *rnp)
565 struct task_struct *t;
568 if (!rcu_preempt_blocked_readers_cgp(rnp))
570 rcu_print_task_stall_begin(rnp);
571 t = list_entry(rnp->gp_tasks->prev,
572 struct task_struct, rcu_node_entry);
573 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
574 pr_cont(" P%d", t->pid);
577 rcu_print_task_stall_end();
582 * Scan the current list of tasks blocked within RCU read-side critical
583 * sections, printing out the tid of each that is blocking the current
584 * expedited grace period.
586 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
588 struct task_struct *t;
593 t = list_entry(rnp->exp_tasks->prev,
594 struct task_struct, rcu_node_entry);
595 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
596 pr_cont(" P%d", t->pid);
603 * Check that the list of blocked tasks for the newly completed grace
604 * period is in fact empty. It is a serious bug to complete a grace
605 * period that still has RCU readers blocked! This function must be
606 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
607 * must be held by the caller.
609 * Also, if there are blocked tasks on the list, they automatically
610 * block the newly created grace period, so set up ->gp_tasks accordingly.
612 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
614 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
615 if (rcu_preempt_has_tasks(rnp))
616 rnp->gp_tasks = rnp->blkd_tasks.next;
617 WARN_ON_ONCE(rnp->qsmask);
621 * Check for a quiescent state from the current CPU. When a task blocks,
622 * the task is recorded in the corresponding CPU's rcu_node structure,
623 * which is checked elsewhere.
625 * Caller must disable hard irqs.
627 static void rcu_preempt_check_callbacks(void)
629 struct task_struct *t = current;
631 if (t->rcu_read_lock_nesting == 0) {
635 if (t->rcu_read_lock_nesting > 0 &&
636 __this_cpu_read(rcu_data_p->core_needs_qs) &&
637 __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm))
638 t->rcu_read_unlock_special.b.need_qs = true;
642 * Queue a preemptible-RCU callback for invocation after a grace period.
644 void call_rcu(struct rcu_head *head, rcu_callback_t func)
646 __call_rcu(head, func, rcu_state_p, -1, 0);
648 EXPORT_SYMBOL_GPL(call_rcu);
651 * synchronize_rcu - wait until a grace period has elapsed.
653 * Control will return to the caller some time after a full grace
654 * period has elapsed, in other words after all currently executing RCU
655 * read-side critical sections have completed. Note, however, that
656 * upon return from synchronize_rcu(), the caller might well be executing
657 * concurrently with new RCU read-side critical sections that began while
658 * synchronize_rcu() was waiting. RCU read-side critical sections are
659 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
661 * See the description of synchronize_sched() for more detailed information
662 * on memory ordering guarantees.
664 void synchronize_rcu(void)
666 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
667 lock_is_held(&rcu_lock_map) ||
668 lock_is_held(&rcu_sched_lock_map),
669 "Illegal synchronize_rcu() in RCU read-side critical section");
670 if (!rcu_scheduler_active)
672 if (rcu_gp_is_expedited())
673 synchronize_rcu_expedited();
675 wait_rcu_gp(call_rcu);
677 EXPORT_SYMBOL_GPL(synchronize_rcu);
680 * Remote handler for smp_call_function_single(). If there is an
681 * RCU read-side critical section in effect, request that the
682 * next rcu_read_unlock() record the quiescent state up the
683 * ->expmask fields in the rcu_node tree. Otherwise, immediately
684 * report the quiescent state.
686 static void sync_rcu_exp_handler(void *info)
688 struct rcu_data *rdp;
689 struct rcu_state *rsp = info;
690 struct task_struct *t = current;
693 * Within an RCU read-side critical section, request that the next
694 * rcu_read_unlock() report. Unless this RCU read-side critical
695 * section has already blocked, in which case it is already set
696 * up for the expedited grace period to wait on it.
698 if (t->rcu_read_lock_nesting > 0 &&
699 !t->rcu_read_unlock_special.b.blocked) {
700 t->rcu_read_unlock_special.b.exp_need_qs = true;
705 * We are either exiting an RCU read-side critical section (negative
706 * values of t->rcu_read_lock_nesting) or are not in one at all
707 * (zero value of t->rcu_read_lock_nesting). Or we are in an RCU
708 * read-side critical section that blocked before this expedited
709 * grace period started. Either way, we can immediately report
710 * the quiescent state.
712 rdp = this_cpu_ptr(rsp->rda);
713 rcu_report_exp_rdp(rsp, rdp, true);
717 * synchronize_rcu_expedited - Brute-force RCU grace period
719 * Wait for an RCU-preempt grace period, but expedite it. The basic
720 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
721 * the ->blkd_tasks lists and wait for this list to drain. This consumes
722 * significant time on all CPUs and is unfriendly to real-time workloads,
723 * so is thus not recommended for any sort of common-case code.
724 * In fact, if you are using synchronize_rcu_expedited() in a loop,
725 * please restructure your code to batch your updates, and then Use a
726 * single synchronize_rcu() instead.
728 void synchronize_rcu_expedited(void)
730 struct rcu_node *rnp;
731 struct rcu_node *rnp_unlock;
732 struct rcu_state *rsp = rcu_state_p;
735 s = rcu_exp_gp_seq_snap(rsp);
737 rnp_unlock = exp_funnel_lock(rsp, s);
738 if (rnp_unlock == NULL)
739 return; /* Someone else did our work for us. */
741 rcu_exp_gp_seq_start(rsp);
743 /* Initialize the rcu_node tree in preparation for the wait. */
744 sync_rcu_exp_select_cpus(rsp, sync_rcu_exp_handler);
746 /* Wait for snapshotted ->blkd_tasks lists to drain. */
747 rnp = rcu_get_root(rsp);
748 synchronize_sched_expedited_wait(rsp);
750 /* Clean up and exit. */
751 rcu_exp_gp_seq_end(rsp);
752 mutex_unlock(&rnp_unlock->exp_funnel_mutex);
754 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
757 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
759 * Note that this primitive does not necessarily wait for an RCU grace period
760 * to complete. For example, if there are no RCU callbacks queued anywhere
761 * in the system, then rcu_barrier() is within its rights to return
762 * immediately, without waiting for anything, much less an RCU grace period.
764 void rcu_barrier(void)
766 _rcu_barrier(rcu_state_p);
768 EXPORT_SYMBOL_GPL(rcu_barrier);
771 * Initialize preemptible RCU's state structures.
773 static void __init __rcu_init_preempt(void)
775 rcu_init_one(rcu_state_p, rcu_data_p);
779 * Check for a task exiting while in a preemptible-RCU read-side
780 * critical section, clean up if so. No need to issue warnings,
781 * as debug_check_no_locks_held() already does this if lockdep
786 struct task_struct *t = current;
788 if (likely(list_empty(¤t->rcu_node_entry)))
790 t->rcu_read_lock_nesting = 1;
792 t->rcu_read_unlock_special.b.blocked = true;
796 #else /* #ifdef CONFIG_PREEMPT_RCU */
798 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
799 static struct rcu_data __percpu *const rcu_data_p = &rcu_sched_data;
802 * Tell them what RCU they are running.
804 static void __init rcu_bootup_announce(void)
806 pr_info("Hierarchical RCU implementation.\n");
807 rcu_bootup_announce_oddness();
811 * Because preemptible RCU does not exist, we never have to check for
812 * CPUs being in quiescent states.
814 static void rcu_preempt_note_context_switch(void)
819 * Because preemptible RCU does not exist, there are never any preempted
822 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
828 * Because there is no preemptible RCU, there can be no readers blocked.
830 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
836 * Because preemptible RCU does not exist, we never have to check for
837 * tasks blocked within RCU read-side critical sections.
839 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
844 * Because preemptible RCU does not exist, we never have to check for
845 * tasks blocked within RCU read-side critical sections.
847 static int rcu_print_task_stall(struct rcu_node *rnp)
853 * Because preemptible RCU does not exist, we never have to check for
854 * tasks blocked within RCU read-side critical sections that are
855 * blocking the current expedited grace period.
857 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
863 * Because there is no preemptible RCU, there can be no readers blocked,
864 * so there is no need to check for blocked tasks. So check only for
865 * bogus qsmask values.
867 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
869 WARN_ON_ONCE(rnp->qsmask);
873 * Because preemptible RCU does not exist, it never has any callbacks
876 static void rcu_preempt_check_callbacks(void)
881 * Wait for an rcu-preempt grace period, but make it happen quickly.
882 * But because preemptible RCU does not exist, map to rcu-sched.
884 void synchronize_rcu_expedited(void)
886 synchronize_sched_expedited();
888 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
891 * Because preemptible RCU does not exist, rcu_barrier() is just
892 * another name for rcu_barrier_sched().
894 void rcu_barrier(void)
898 EXPORT_SYMBOL_GPL(rcu_barrier);
901 * Because preemptible RCU does not exist, it need not be initialized.
903 static void __init __rcu_init_preempt(void)
908 * Because preemptible RCU does not exist, tasks cannot possibly exit
909 * while in preemptible RCU read-side critical sections.
915 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
918 * If boosting, set rcuc kthreads to realtime priority.
920 static void rcu_cpu_kthread_setup(unsigned int cpu)
922 #ifdef CONFIG_RCU_BOOST
923 struct sched_param sp;
925 sp.sched_priority = kthread_prio;
926 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
927 #endif /* #ifdef CONFIG_RCU_BOOST */
930 #ifdef CONFIG_RCU_BOOST
932 #include "../locking/rtmutex_common.h"
934 #ifdef CONFIG_RCU_TRACE
936 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
938 if (!rcu_preempt_has_tasks(rnp))
939 rnp->n_balk_blkd_tasks++;
940 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
941 rnp->n_balk_exp_gp_tasks++;
942 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
943 rnp->n_balk_boost_tasks++;
944 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
945 rnp->n_balk_notblocked++;
946 else if (rnp->gp_tasks != NULL &&
947 ULONG_CMP_LT(jiffies, rnp->boost_time))
948 rnp->n_balk_notyet++;
953 #else /* #ifdef CONFIG_RCU_TRACE */
955 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
959 #endif /* #else #ifdef CONFIG_RCU_TRACE */
962 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
963 * or ->boost_tasks, advancing the pointer to the next task in the
966 * Note that irqs must be enabled: boosting the task can block.
967 * Returns 1 if there are more tasks needing to be boosted.
969 static int rcu_boost(struct rcu_node *rnp)
972 struct task_struct *t;
973 struct list_head *tb;
975 if (READ_ONCE(rnp->exp_tasks) == NULL &&
976 READ_ONCE(rnp->boost_tasks) == NULL)
977 return 0; /* Nothing left to boost. */
979 raw_spin_lock_irqsave(&rnp->lock, flags);
980 smp_mb__after_unlock_lock();
983 * Recheck under the lock: all tasks in need of boosting
984 * might exit their RCU read-side critical sections on their own.
986 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
987 raw_spin_unlock_irqrestore(&rnp->lock, flags);
992 * Preferentially boost tasks blocking expedited grace periods.
993 * This cannot starve the normal grace periods because a second
994 * expedited grace period must boost all blocked tasks, including
995 * those blocking the pre-existing normal grace period.
997 if (rnp->exp_tasks != NULL) {
1001 tb = rnp->boost_tasks;
1002 rnp->n_normal_boosts++;
1004 rnp->n_tasks_boosted++;
1007 * We boost task t by manufacturing an rt_mutex that appears to
1008 * be held by task t. We leave a pointer to that rt_mutex where
1009 * task t can find it, and task t will release the mutex when it
1010 * exits its outermost RCU read-side critical section. Then
1011 * simply acquiring this artificial rt_mutex will boost task
1012 * t's priority. (Thanks to tglx for suggesting this approach!)
1014 * Note that task t must acquire rnp->lock to remove itself from
1015 * the ->blkd_tasks list, which it will do from exit() if from
1016 * nowhere else. We therefore are guaranteed that task t will
1017 * stay around at least until we drop rnp->lock. Note that
1018 * rnp->lock also resolves races between our priority boosting
1019 * and task t's exiting its outermost RCU read-side critical
1022 t = container_of(tb, struct task_struct, rcu_node_entry);
1023 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1024 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1025 /* Lock only for side effect: boosts task t's priority. */
1026 rt_mutex_lock(&rnp->boost_mtx);
1027 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1029 return READ_ONCE(rnp->exp_tasks) != NULL ||
1030 READ_ONCE(rnp->boost_tasks) != NULL;
1034 * Priority-boosting kthread, one per leaf rcu_node.
1036 static int rcu_boost_kthread(void *arg)
1038 struct rcu_node *rnp = (struct rcu_node *)arg;
1042 trace_rcu_utilization(TPS("Start boost kthread@init"));
1044 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1045 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1046 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1047 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1048 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1049 more2boost = rcu_boost(rnp);
1055 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1056 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1057 schedule_timeout_interruptible(2);
1058 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1063 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1068 * Check to see if it is time to start boosting RCU readers that are
1069 * blocking the current grace period, and, if so, tell the per-rcu_node
1070 * kthread to start boosting them. If there is an expedited grace
1071 * period in progress, it is always time to boost.
1073 * The caller must hold rnp->lock, which this function releases.
1074 * The ->boost_kthread_task is immortal, so we don't need to worry
1075 * about it going away.
1077 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1078 __releases(rnp->lock)
1080 struct task_struct *t;
1082 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1083 rnp->n_balk_exp_gp_tasks++;
1084 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1087 if (rnp->exp_tasks != NULL ||
1088 (rnp->gp_tasks != NULL &&
1089 rnp->boost_tasks == NULL &&
1091 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1092 if (rnp->exp_tasks == NULL)
1093 rnp->boost_tasks = rnp->gp_tasks;
1094 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1095 t = rnp->boost_kthread_task;
1097 rcu_wake_cond(t, rnp->boost_kthread_status);
1099 rcu_initiate_boost_trace(rnp);
1100 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1105 * Is the current CPU running the RCU-callbacks kthread?
1106 * Caller must have preemption disabled.
1108 static bool rcu_is_callbacks_kthread(void)
1110 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1113 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1116 * Do priority-boost accounting for the start of a new grace period.
1118 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1120 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1124 * Create an RCU-boost kthread for the specified node if one does not
1125 * already exist. We only create this kthread for preemptible RCU.
1126 * Returns zero if all is well, a negated errno otherwise.
1128 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1129 struct rcu_node *rnp)
1131 int rnp_index = rnp - &rsp->node[0];
1132 unsigned long flags;
1133 struct sched_param sp;
1134 struct task_struct *t;
1136 if (rcu_state_p != rsp)
1139 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1143 if (rnp->boost_kthread_task != NULL)
1145 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1146 "rcub/%d", rnp_index);
1149 raw_spin_lock_irqsave(&rnp->lock, flags);
1150 smp_mb__after_unlock_lock();
1151 rnp->boost_kthread_task = t;
1152 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1153 sp.sched_priority = kthread_prio;
1154 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1155 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1160 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1161 * served by the rcu_node in question. The CPU hotplug lock is still
1162 * held, so the value of rnp->qsmaskinit will be stable.
1164 * We don't include outgoingcpu in the affinity set, use -1 if there is
1165 * no outgoing CPU. If there are no CPUs left in the affinity set,
1166 * this function allows the kthread to execute on any CPU.
1168 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1170 struct task_struct *t = rnp->boost_kthread_task;
1171 unsigned long mask = rcu_rnp_online_cpus(rnp);
1177 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1179 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1180 if ((mask & 0x1) && cpu != outgoingcpu)
1181 cpumask_set_cpu(cpu, cm);
1182 if (cpumask_weight(cm) == 0)
1184 set_cpus_allowed_ptr(t, cm);
1185 free_cpumask_var(cm);
1189 * Spawn boost kthreads -- called as soon as the scheduler is running.
1191 static void __init rcu_spawn_boost_kthreads(void)
1193 struct rcu_node *rnp;
1194 rcu_for_each_leaf_node(rcu_state_p, rnp)
1195 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1198 static void rcu_prepare_kthreads(int cpu)
1200 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1201 struct rcu_node *rnp = rdp->mynode;
1203 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1204 if (rcu_scheduler_fully_active)
1205 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1208 #else /* #ifdef CONFIG_RCU_BOOST */
1210 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1211 __releases(rnp->lock)
1213 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1216 static bool rcu_is_callbacks_kthread(void)
1221 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1225 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1229 static void __init rcu_spawn_boost_kthreads(void)
1233 static void rcu_prepare_kthreads(int cpu)
1237 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1239 #if !defined(CONFIG_RCU_FAST_NO_HZ) || defined(CONFIG_PREEMPT_RT_FULL)
1242 * Check to see if any future RCU-related work will need to be done
1243 * by the current CPU, even if none need be done immediately, returning
1244 * 1 if so. This function is part of the RCU implementation; it is -not-
1245 * an exported member of the RCU API.
1247 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1248 * any flavor of RCU.
1250 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1252 *nextevt = KTIME_MAX;
1253 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1254 ? 0 : rcu_cpu_has_callbacks(NULL);
1256 #endif /* !defined(CONFIG_RCU_FAST_NO_HZ) || defined(CONFIG_PREEMPT_RT_FULL) */
1258 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1260 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1263 static void rcu_cleanup_after_idle(void)
1268 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1271 static void rcu_prepare_for_idle(void)
1276 * Don't bother keeping a running count of the number of RCU callbacks
1277 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1279 static void rcu_idle_count_callbacks_posted(void)
1283 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1286 * This code is invoked when a CPU goes idle, at which point we want
1287 * to have the CPU do everything required for RCU so that it can enter
1288 * the energy-efficient dyntick-idle mode. This is handled by a
1289 * state machine implemented by rcu_prepare_for_idle() below.
1291 * The following three proprocessor symbols control this state machine:
1293 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1294 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1295 * is sized to be roughly one RCU grace period. Those energy-efficiency
1296 * benchmarkers who might otherwise be tempted to set this to a large
1297 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1298 * system. And if you are -that- concerned about energy efficiency,
1299 * just power the system down and be done with it!
1300 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1301 * permitted to sleep in dyntick-idle mode with only lazy RCU
1302 * callbacks pending. Setting this too high can OOM your system.
1304 * The values below work well in practice. If future workloads require
1305 * adjustment, they can be converted into kernel config parameters, though
1306 * making the state machine smarter might be a better option.
1308 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1309 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1311 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1312 module_param(rcu_idle_gp_delay, int, 0644);
1313 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1314 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1317 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1318 * only if it has been awhile since the last time we did so. Afterwards,
1319 * if there are any callbacks ready for immediate invocation, return true.
1321 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1323 bool cbs_ready = false;
1324 struct rcu_data *rdp;
1325 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1326 struct rcu_node *rnp;
1327 struct rcu_state *rsp;
1329 /* Exit early if we advanced recently. */
1330 if (jiffies == rdtp->last_advance_all)
1332 rdtp->last_advance_all = jiffies;
1334 for_each_rcu_flavor(rsp) {
1335 rdp = this_cpu_ptr(rsp->rda);
1339 * Don't bother checking unless a grace period has
1340 * completed since we last checked and there are
1341 * callbacks not yet ready to invoke.
1343 if ((rdp->completed != rnp->completed ||
1344 unlikely(READ_ONCE(rdp->gpwrap))) &&
1345 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1346 note_gp_changes(rsp, rdp);
1348 if (cpu_has_callbacks_ready_to_invoke(rdp))
1354 #ifndef CONFIG_PREEMPT_RT_FULL
1357 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1358 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1359 * caller to set the timeout based on whether or not there are non-lazy
1362 * The caller must have disabled interrupts.
1364 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1366 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1369 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1370 *nextevt = KTIME_MAX;
1374 /* Snapshot to detect later posting of non-lazy callback. */
1375 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1377 /* If no callbacks, RCU doesn't need the CPU. */
1378 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1379 *nextevt = KTIME_MAX;
1383 /* Attempt to advance callbacks. */
1384 if (rcu_try_advance_all_cbs()) {
1385 /* Some ready to invoke, so initiate later invocation. */
1389 rdtp->last_accelerate = jiffies;
1391 /* Request timer delay depending on laziness, and round. */
1392 if (!rdtp->all_lazy) {
1393 dj = round_up(rcu_idle_gp_delay + jiffies,
1394 rcu_idle_gp_delay) - jiffies;
1396 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1398 *nextevt = basemono + dj * TICK_NSEC;
1401 #endif /* #ifndef CONFIG_PREEMPT_RT_FULL */
1404 * Prepare a CPU for idle from an RCU perspective. The first major task
1405 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1406 * The second major task is to check to see if a non-lazy callback has
1407 * arrived at a CPU that previously had only lazy callbacks. The third
1408 * major task is to accelerate (that is, assign grace-period numbers to)
1409 * any recently arrived callbacks.
1411 * The caller must have disabled interrupts.
1413 static void rcu_prepare_for_idle(void)
1416 struct rcu_data *rdp;
1417 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1418 struct rcu_node *rnp;
1419 struct rcu_state *rsp;
1422 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL))
1425 /* Handle nohz enablement switches conservatively. */
1426 tne = READ_ONCE(tick_nohz_active);
1427 if (tne != rdtp->tick_nohz_enabled_snap) {
1428 if (rcu_cpu_has_callbacks(NULL))
1429 invoke_rcu_core(); /* force nohz to see update. */
1430 rdtp->tick_nohz_enabled_snap = tne;
1436 /* If this is a no-CBs CPU, no callbacks, just return. */
1437 if (rcu_is_nocb_cpu(smp_processor_id()))
1441 * If a non-lazy callback arrived at a CPU having only lazy
1442 * callbacks, invoke RCU core for the side-effect of recalculating
1443 * idle duration on re-entry to idle.
1445 if (rdtp->all_lazy &&
1446 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1447 rdtp->all_lazy = false;
1448 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1454 * If we have not yet accelerated this jiffy, accelerate all
1455 * callbacks on this CPU.
1457 if (rdtp->last_accelerate == jiffies)
1459 rdtp->last_accelerate = jiffies;
1460 for_each_rcu_flavor(rsp) {
1461 rdp = this_cpu_ptr(rsp->rda);
1462 if (!*rdp->nxttail[RCU_DONE_TAIL])
1465 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1466 smp_mb__after_unlock_lock();
1467 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1468 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1470 rcu_gp_kthread_wake(rsp);
1475 * Clean up for exit from idle. Attempt to advance callbacks based on
1476 * any grace periods that elapsed while the CPU was idle, and if any
1477 * callbacks are now ready to invoke, initiate invocation.
1479 static void rcu_cleanup_after_idle(void)
1481 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1482 rcu_is_nocb_cpu(smp_processor_id()))
1484 if (rcu_try_advance_all_cbs())
1489 * Keep a running count of the number of non-lazy callbacks posted
1490 * on this CPU. This running counter (which is never decremented) allows
1491 * rcu_prepare_for_idle() to detect when something out of the idle loop
1492 * posts a callback, even if an equal number of callbacks are invoked.
1493 * Of course, callbacks should only be posted from within a trace event
1494 * designed to be called from idle or from within RCU_NONIDLE().
1496 static void rcu_idle_count_callbacks_posted(void)
1498 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1502 * Data for flushing lazy RCU callbacks at OOM time.
1504 static atomic_t oom_callback_count;
1505 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1508 * RCU OOM callback -- decrement the outstanding count and deliver the
1509 * wake-up if we are the last one.
1511 static void rcu_oom_callback(struct rcu_head *rhp)
1513 if (atomic_dec_and_test(&oom_callback_count))
1514 wake_up(&oom_callback_wq);
1518 * Post an rcu_oom_notify callback on the current CPU if it has at
1519 * least one lazy callback. This will unnecessarily post callbacks
1520 * to CPUs that already have a non-lazy callback at the end of their
1521 * callback list, but this is an infrequent operation, so accept some
1522 * extra overhead to keep things simple.
1524 static void rcu_oom_notify_cpu(void *unused)
1526 struct rcu_state *rsp;
1527 struct rcu_data *rdp;
1529 for_each_rcu_flavor(rsp) {
1530 rdp = raw_cpu_ptr(rsp->rda);
1531 if (rdp->qlen_lazy != 0) {
1532 atomic_inc(&oom_callback_count);
1533 rsp->call(&rdp->oom_head, rcu_oom_callback);
1539 * If low on memory, ensure that each CPU has a non-lazy callback.
1540 * This will wake up CPUs that have only lazy callbacks, in turn
1541 * ensuring that they free up the corresponding memory in a timely manner.
1542 * Because an uncertain amount of memory will be freed in some uncertain
1543 * timeframe, we do not claim to have freed anything.
1545 static int rcu_oom_notify(struct notifier_block *self,
1546 unsigned long notused, void *nfreed)
1550 /* Wait for callbacks from earlier instance to complete. */
1551 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1552 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1555 * Prevent premature wakeup: ensure that all increments happen
1556 * before there is a chance of the counter reaching zero.
1558 atomic_set(&oom_callback_count, 1);
1560 for_each_online_cpu(cpu) {
1561 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1562 cond_resched_rcu_qs();
1565 /* Unconditionally decrement: no need to wake ourselves up. */
1566 atomic_dec(&oom_callback_count);
1571 static struct notifier_block rcu_oom_nb = {
1572 .notifier_call = rcu_oom_notify
1575 static int __init rcu_register_oom_notifier(void)
1577 register_oom_notifier(&rcu_oom_nb);
1580 early_initcall(rcu_register_oom_notifier);
1582 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1584 #ifdef CONFIG_RCU_FAST_NO_HZ
1586 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1588 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1589 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1591 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1592 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1594 rdtp->all_lazy ? 'L' : '.',
1595 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1598 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1600 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1605 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1607 /* Initiate the stall-info list. */
1608 static void print_cpu_stall_info_begin(void)
1614 * Print out diagnostic information for the specified stalled CPU.
1616 * If the specified CPU is aware of the current RCU grace period
1617 * (flavor specified by rsp), then print the number of scheduling
1618 * clock interrupts the CPU has taken during the time that it has
1619 * been aware. Otherwise, print the number of RCU grace periods
1620 * that this CPU is ignorant of, for example, "1" if the CPU was
1621 * aware of the previous grace period.
1623 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1625 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1627 char fast_no_hz[72];
1628 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1629 struct rcu_dynticks *rdtp = rdp->dynticks;
1631 unsigned long ticks_value;
1633 if (rsp->gpnum == rdp->gpnum) {
1634 ticks_title = "ticks this GP";
1635 ticks_value = rdp->ticks_this_gp;
1637 ticks_title = "GPs behind";
1638 ticks_value = rsp->gpnum - rdp->gpnum;
1640 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1641 pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1643 "O."[!!cpu_online(cpu)],
1644 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1645 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1646 ticks_value, ticks_title,
1647 atomic_read(&rdtp->dynticks) & 0xfff,
1648 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1649 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1650 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1654 /* Terminate the stall-info list. */
1655 static void print_cpu_stall_info_end(void)
1660 /* Zero ->ticks_this_gp for all flavors of RCU. */
1661 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1663 rdp->ticks_this_gp = 0;
1664 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1667 /* Increment ->ticks_this_gp for all flavors of RCU. */
1668 static void increment_cpu_stall_ticks(void)
1670 struct rcu_state *rsp;
1672 for_each_rcu_flavor(rsp)
1673 raw_cpu_inc(rsp->rda->ticks_this_gp);
1676 #ifdef CONFIG_RCU_NOCB_CPU
1679 * Offload callback processing from the boot-time-specified set of CPUs
1680 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1681 * kthread created that pulls the callbacks from the corresponding CPU,
1682 * waits for a grace period to elapse, and invokes the callbacks.
1683 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1684 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1685 * has been specified, in which case each kthread actively polls its
1686 * CPU. (Which isn't so great for energy efficiency, but which does
1687 * reduce RCU's overhead on that CPU.)
1689 * This is intended to be used in conjunction with Frederic Weisbecker's
1690 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1691 * running CPU-bound user-mode computations.
1693 * Offloading of callback processing could also in theory be used as
1694 * an energy-efficiency measure because CPUs with no RCU callbacks
1695 * queued are more aggressive about entering dyntick-idle mode.
1699 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1700 static int __init rcu_nocb_setup(char *str)
1702 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1703 have_rcu_nocb_mask = true;
1704 cpulist_parse(str, rcu_nocb_mask);
1707 __setup("rcu_nocbs=", rcu_nocb_setup);
1709 static int __init parse_rcu_nocb_poll(char *arg)
1714 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1717 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1720 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1726 * Set the root rcu_node structure's ->need_future_gp field
1727 * based on the sum of those of all rcu_node structures. This does
1728 * double-count the root rcu_node structure's requests, but this
1729 * is necessary to handle the possibility of a rcu_nocb_kthread()
1730 * having awakened during the time that the rcu_node structures
1731 * were being updated for the end of the previous grace period.
1733 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1735 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1738 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1740 return &rnp->nocb_gp_wq[rnp->completed & 0x1];
1743 static void rcu_init_one_nocb(struct rcu_node *rnp)
1745 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1746 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1749 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1750 /* Is the specified CPU a no-CBs CPU? */
1751 bool rcu_is_nocb_cpu(int cpu)
1753 if (have_rcu_nocb_mask)
1754 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1757 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1760 * Kick the leader kthread for this NOCB group.
1762 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1764 struct rcu_data *rdp_leader = rdp->nocb_leader;
1766 if (!READ_ONCE(rdp_leader->nocb_kthread))
1768 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1769 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1770 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1771 swake_up(&rdp_leader->nocb_wq);
1776 * Does the specified CPU need an RCU callback for the specified flavor
1779 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1781 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1783 #ifdef CONFIG_PROVE_RCU
1784 struct rcu_head *rhp;
1785 #endif /* #ifdef CONFIG_PROVE_RCU */
1788 * Check count of all no-CBs callbacks awaiting invocation.
1789 * There needs to be a barrier before this function is called,
1790 * but associated with a prior determination that no more
1791 * callbacks would be posted. In the worst case, the first
1792 * barrier in _rcu_barrier() suffices (but the caller cannot
1793 * necessarily rely on this, not a substitute for the caller
1794 * getting the concurrency design right!). There must also be
1795 * a barrier between the following load an posting of a callback
1796 * (if a callback is in fact needed). This is associated with an
1797 * atomic_inc() in the caller.
1799 ret = atomic_long_read(&rdp->nocb_q_count);
1801 #ifdef CONFIG_PROVE_RCU
1802 rhp = READ_ONCE(rdp->nocb_head);
1804 rhp = READ_ONCE(rdp->nocb_gp_head);
1806 rhp = READ_ONCE(rdp->nocb_follower_head);
1808 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1809 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1810 rcu_scheduler_fully_active) {
1811 /* RCU callback enqueued before CPU first came online??? */
1812 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1816 #endif /* #ifdef CONFIG_PROVE_RCU */
1822 * Enqueue the specified string of rcu_head structures onto the specified
1823 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1824 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1825 * counts are supplied by rhcount and rhcount_lazy.
1827 * If warranted, also wake up the kthread servicing this CPUs queues.
1829 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1830 struct rcu_head *rhp,
1831 struct rcu_head **rhtp,
1832 int rhcount, int rhcount_lazy,
1833 unsigned long flags)
1836 struct rcu_head **old_rhpp;
1837 struct task_struct *t;
1839 /* Enqueue the callback on the nocb list and update counts. */
1840 atomic_long_add(rhcount, &rdp->nocb_q_count);
1841 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1842 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1843 WRITE_ONCE(*old_rhpp, rhp);
1844 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1845 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1847 /* If we are not being polled and there is a kthread, awaken it ... */
1848 t = READ_ONCE(rdp->nocb_kthread);
1849 if (rcu_nocb_poll || !t) {
1850 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1851 TPS("WakeNotPoll"));
1854 len = atomic_long_read(&rdp->nocb_q_count);
1855 if (old_rhpp == &rdp->nocb_head) {
1856 if (!irqs_disabled_flags(flags)) {
1857 /* ... if queue was empty ... */
1858 wake_nocb_leader(rdp, false);
1859 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1862 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1863 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1864 TPS("WakeEmptyIsDeferred"));
1866 rdp->qlen_last_fqs_check = 0;
1867 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1868 /* ... or if many callbacks queued. */
1869 if (!irqs_disabled_flags(flags)) {
1870 wake_nocb_leader(rdp, true);
1871 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1874 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1875 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1876 TPS("WakeOvfIsDeferred"));
1878 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1880 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1886 * This is a helper for __call_rcu(), which invokes this when the normal
1887 * callback queue is inoperable. If this is not a no-CBs CPU, this
1888 * function returns failure back to __call_rcu(), which can complain
1891 * Otherwise, this function queues the callback where the corresponding
1892 * "rcuo" kthread can find it.
1894 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1895 bool lazy, unsigned long flags)
1898 if (!rcu_is_nocb_cpu(rdp->cpu))
1900 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1901 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1902 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
1903 (unsigned long)rhp->func,
1904 -atomic_long_read(&rdp->nocb_q_count_lazy),
1905 -atomic_long_read(&rdp->nocb_q_count));
1907 trace_rcu_callback(rdp->rsp->name, rhp,
1908 -atomic_long_read(&rdp->nocb_q_count_lazy),
1909 -atomic_long_read(&rdp->nocb_q_count));
1912 * If called from an extended quiescent state with interrupts
1913 * disabled, invoke the RCU core in order to allow the idle-entry
1914 * deferred-wakeup check to function.
1916 if (irqs_disabled_flags(flags) &&
1917 !rcu_is_watching() &&
1918 cpu_online(smp_processor_id()))
1925 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
1928 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
1929 struct rcu_data *rdp,
1930 unsigned long flags)
1932 long ql = rsp->qlen;
1933 long qll = rsp->qlen_lazy;
1935 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
1936 if (!rcu_is_nocb_cpu(smp_processor_id()))
1941 /* First, enqueue the donelist, if any. This preserves CB ordering. */
1942 if (rsp->orphan_donelist != NULL) {
1943 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
1944 rsp->orphan_donetail, ql, qll, flags);
1946 rsp->orphan_donelist = NULL;
1947 rsp->orphan_donetail = &rsp->orphan_donelist;
1949 if (rsp->orphan_nxtlist != NULL) {
1950 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
1951 rsp->orphan_nxttail, ql, qll, flags);
1953 rsp->orphan_nxtlist = NULL;
1954 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1960 * If necessary, kick off a new grace period, and either way wait
1961 * for a subsequent grace period to complete.
1963 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
1967 unsigned long flags;
1969 struct rcu_node *rnp = rdp->mynode;
1971 raw_spin_lock_irqsave(&rnp->lock, flags);
1972 smp_mb__after_unlock_lock();
1973 needwake = rcu_start_future_gp(rnp, rdp, &c);
1974 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1976 rcu_gp_kthread_wake(rdp->rsp);
1979 * Wait for the grace period. Do so interruptibly to avoid messing
1980 * up the load average.
1982 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
1984 swait_event_interruptible(
1985 rnp->nocb_gp_wq[c & 0x1],
1986 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
1989 WARN_ON(signal_pending(current));
1990 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
1992 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
1993 smp_mb(); /* Ensure that CB invocation happens after GP end. */
1997 * Leaders come here to wait for additional callbacks to show up.
1998 * This function does not return until callbacks appear.
2000 static void nocb_leader_wait(struct rcu_data *my_rdp)
2002 bool firsttime = true;
2004 struct rcu_data *rdp;
2005 struct rcu_head **tail;
2009 /* Wait for callbacks to appear. */
2010 if (!rcu_nocb_poll) {
2011 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2012 swait_event_interruptible(my_rdp->nocb_wq,
2013 !READ_ONCE(my_rdp->nocb_leader_sleep));
2014 /* Memory barrier handled by smp_mb() calls below and repoll. */
2015 } else if (firsttime) {
2016 firsttime = false; /* Don't drown trace log with "Poll"! */
2017 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2021 * Each pass through the following loop checks a follower for CBs.
2022 * We are our own first follower. Any CBs found are moved to
2023 * nocb_gp_head, where they await a grace period.
2026 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2027 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2028 if (!rdp->nocb_gp_head)
2029 continue; /* No CBs here, try next follower. */
2031 /* Move callbacks to wait-for-GP list, which is empty. */
2032 WRITE_ONCE(rdp->nocb_head, NULL);
2033 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2038 * If there were no callbacks, sleep a bit, rescan after a
2039 * memory barrier, and go retry.
2041 if (unlikely(!gotcbs)) {
2043 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2045 WARN_ON(signal_pending(current));
2046 schedule_timeout_interruptible(1);
2048 /* Rescan in case we were a victim of memory ordering. */
2049 my_rdp->nocb_leader_sleep = true;
2050 smp_mb(); /* Ensure _sleep true before scan. */
2051 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2052 if (READ_ONCE(rdp->nocb_head)) {
2053 /* Found CB, so short-circuit next wait. */
2054 my_rdp->nocb_leader_sleep = false;
2060 /* Wait for one grace period. */
2061 rcu_nocb_wait_gp(my_rdp);
2064 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2065 * We set it now, but recheck for new callbacks while
2066 * traversing our follower list.
2068 my_rdp->nocb_leader_sleep = true;
2069 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2071 /* Each pass through the following loop wakes a follower, if needed. */
2072 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2073 if (READ_ONCE(rdp->nocb_head))
2074 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2075 if (!rdp->nocb_gp_head)
2076 continue; /* No CBs, so no need to wake follower. */
2078 /* Append callbacks to follower's "done" list. */
2079 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2080 *tail = rdp->nocb_gp_head;
2081 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2082 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2084 * List was empty, wake up the follower.
2085 * Memory barriers supplied by atomic_long_add().
2087 swake_up(&rdp->nocb_wq);
2091 /* If we (the leader) don't have CBs, go wait some more. */
2092 if (!my_rdp->nocb_follower_head)
2097 * Followers come here to wait for additional callbacks to show up.
2098 * This function does not return until callbacks appear.
2100 static void nocb_follower_wait(struct rcu_data *rdp)
2102 bool firsttime = true;
2105 if (!rcu_nocb_poll) {
2106 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2108 swait_event_interruptible(rdp->nocb_wq,
2109 READ_ONCE(rdp->nocb_follower_head));
2110 } else if (firsttime) {
2111 /* Don't drown trace log with "Poll"! */
2113 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2115 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2116 /* ^^^ Ensure CB invocation follows _head test. */
2120 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2122 WARN_ON(signal_pending(current));
2123 schedule_timeout_interruptible(1);
2128 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2129 * callbacks queued by the corresponding no-CBs CPU, however, there is
2130 * an optional leader-follower relationship so that the grace-period
2131 * kthreads don't have to do quite so many wakeups.
2133 static int rcu_nocb_kthread(void *arg)
2136 struct rcu_head *list;
2137 struct rcu_head *next;
2138 struct rcu_head **tail;
2139 struct rcu_data *rdp = arg;
2141 /* Each pass through this loop invokes one batch of callbacks */
2143 /* Wait for callbacks. */
2144 if (rdp->nocb_leader == rdp)
2145 nocb_leader_wait(rdp);
2147 nocb_follower_wait(rdp);
2149 /* Pull the ready-to-invoke callbacks onto local list. */
2150 list = READ_ONCE(rdp->nocb_follower_head);
2152 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2153 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2154 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2156 /* Each pass through the following loop invokes a callback. */
2157 trace_rcu_batch_start(rdp->rsp->name,
2158 atomic_long_read(&rdp->nocb_q_count_lazy),
2159 atomic_long_read(&rdp->nocb_q_count), -1);
2163 /* Wait for enqueuing to complete, if needed. */
2164 while (next == NULL && &list->next != tail) {
2165 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2167 schedule_timeout_interruptible(1);
2168 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2172 debug_rcu_head_unqueue(list);
2174 if (__rcu_reclaim(rdp->rsp->name, list))
2178 cond_resched_rcu_qs();
2181 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2182 smp_mb__before_atomic(); /* _add after CB invocation. */
2183 atomic_long_add(-c, &rdp->nocb_q_count);
2184 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2185 rdp->n_nocbs_invoked += c;
2190 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2191 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2193 return READ_ONCE(rdp->nocb_defer_wakeup);
2196 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2197 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2201 if (!rcu_nocb_need_deferred_wakeup(rdp))
2203 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2204 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2205 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2206 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2209 void __init rcu_init_nohz(void)
2212 bool need_rcu_nocb_mask = true;
2213 struct rcu_state *rsp;
2215 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2216 need_rcu_nocb_mask = false;
2217 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2219 #if defined(CONFIG_NO_HZ_FULL)
2220 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2221 need_rcu_nocb_mask = true;
2222 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2224 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2225 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2226 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2229 have_rcu_nocb_mask = true;
2231 if (!have_rcu_nocb_mask)
2234 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2235 pr_info("\tOffload RCU callbacks from CPU 0\n");
2236 cpumask_set_cpu(0, rcu_nocb_mask);
2237 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2238 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2239 pr_info("\tOffload RCU callbacks from all CPUs\n");
2240 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2241 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2242 #if defined(CONFIG_NO_HZ_FULL)
2243 if (tick_nohz_full_running)
2244 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2245 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2247 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2248 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2249 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2252 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2253 cpumask_pr_args(rcu_nocb_mask));
2255 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2257 for_each_rcu_flavor(rsp) {
2258 for_each_cpu(cpu, rcu_nocb_mask)
2259 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2260 rcu_organize_nocb_kthreads(rsp);
2264 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2265 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2267 rdp->nocb_tail = &rdp->nocb_head;
2268 init_swait_queue_head(&rdp->nocb_wq);
2269 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2273 * If the specified CPU is a no-CBs CPU that does not already have its
2274 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2275 * brought online out of order, this can require re-organizing the
2276 * leader-follower relationships.
2278 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2280 struct rcu_data *rdp;
2281 struct rcu_data *rdp_last;
2282 struct rcu_data *rdp_old_leader;
2283 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2284 struct task_struct *t;
2287 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2288 * then nothing to do.
2290 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2293 /* If we didn't spawn the leader first, reorganize! */
2294 rdp_old_leader = rdp_spawn->nocb_leader;
2295 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2297 rdp = rdp_old_leader;
2299 rdp->nocb_leader = rdp_spawn;
2300 if (rdp_last && rdp != rdp_spawn)
2301 rdp_last->nocb_next_follower = rdp;
2302 if (rdp == rdp_spawn) {
2303 rdp = rdp->nocb_next_follower;
2306 rdp = rdp->nocb_next_follower;
2307 rdp_last->nocb_next_follower = NULL;
2310 rdp_spawn->nocb_next_follower = rdp_old_leader;
2313 /* Spawn the kthread for this CPU and RCU flavor. */
2314 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2315 "rcuo%c/%d", rsp->abbr, cpu);
2317 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2321 * If the specified CPU is a no-CBs CPU that does not already have its
2322 * rcuo kthreads, spawn them.
2324 static void rcu_spawn_all_nocb_kthreads(int cpu)
2326 struct rcu_state *rsp;
2328 if (rcu_scheduler_fully_active)
2329 for_each_rcu_flavor(rsp)
2330 rcu_spawn_one_nocb_kthread(rsp, cpu);
2334 * Once the scheduler is running, spawn rcuo kthreads for all online
2335 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2336 * non-boot CPUs come online -- if this changes, we will need to add
2337 * some mutual exclusion.
2339 static void __init rcu_spawn_nocb_kthreads(void)
2343 for_each_online_cpu(cpu)
2344 rcu_spawn_all_nocb_kthreads(cpu);
2347 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2348 static int rcu_nocb_leader_stride = -1;
2349 module_param(rcu_nocb_leader_stride, int, 0444);
2352 * Initialize leader-follower relationships for all no-CBs CPU.
2354 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2357 int ls = rcu_nocb_leader_stride;
2358 int nl = 0; /* Next leader. */
2359 struct rcu_data *rdp;
2360 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2361 struct rcu_data *rdp_prev = NULL;
2363 if (!have_rcu_nocb_mask)
2366 ls = int_sqrt(nr_cpu_ids);
2367 rcu_nocb_leader_stride = ls;
2371 * Each pass through this loop sets up one rcu_data structure and
2372 * spawns one rcu_nocb_kthread().
2374 for_each_cpu(cpu, rcu_nocb_mask) {
2375 rdp = per_cpu_ptr(rsp->rda, cpu);
2376 if (rdp->cpu >= nl) {
2377 /* New leader, set up for followers & next leader. */
2378 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2379 rdp->nocb_leader = rdp;
2382 /* Another follower, link to previous leader. */
2383 rdp->nocb_leader = rdp_leader;
2384 rdp_prev->nocb_next_follower = rdp;
2390 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2391 static bool init_nocb_callback_list(struct rcu_data *rdp)
2393 if (!rcu_is_nocb_cpu(rdp->cpu))
2396 /* If there are early-boot callbacks, move them to nocb lists. */
2398 rdp->nocb_head = rdp->nxtlist;
2399 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2400 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2401 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2402 rdp->nxtlist = NULL;
2406 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2410 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2412 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2414 WARN_ON_ONCE(1); /* Should be dead code. */
2418 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2422 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2426 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2431 static void rcu_init_one_nocb(struct rcu_node *rnp)
2435 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2436 bool lazy, unsigned long flags)
2441 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2442 struct rcu_data *rdp,
2443 unsigned long flags)
2448 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2452 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2457 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2461 static void rcu_spawn_all_nocb_kthreads(int cpu)
2465 static void __init rcu_spawn_nocb_kthreads(void)
2469 static bool init_nocb_callback_list(struct rcu_data *rdp)
2474 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2477 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2478 * arbitrarily long period of time with the scheduling-clock tick turned
2479 * off. RCU will be paying attention to this CPU because it is in the
2480 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2481 * machine because the scheduling-clock tick has been disabled. Therefore,
2482 * if an adaptive-ticks CPU is failing to respond to the current grace
2483 * period and has not be idle from an RCU perspective, kick it.
2485 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2487 #ifdef CONFIG_NO_HZ_FULL
2488 if (tick_nohz_full_cpu(cpu))
2489 smp_send_reschedule(cpu);
2490 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2494 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2496 static int full_sysidle_state; /* Current system-idle state. */
2497 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2498 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2499 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2500 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2501 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2504 * Invoked to note exit from irq or task transition to idle. Note that
2505 * usermode execution does -not- count as idle here! After all, we want
2506 * to detect full-system idle states, not RCU quiescent states and grace
2507 * periods. The caller must have disabled interrupts.
2509 static void rcu_sysidle_enter(int irq)
2512 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2514 /* If there are no nohz_full= CPUs, no need to track this. */
2515 if (!tick_nohz_full_enabled())
2518 /* Adjust nesting, check for fully idle. */
2520 rdtp->dynticks_idle_nesting--;
2521 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2522 if (rdtp->dynticks_idle_nesting != 0)
2523 return; /* Still not fully idle. */
2525 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2526 DYNTICK_TASK_NEST_VALUE) {
2527 rdtp->dynticks_idle_nesting = 0;
2529 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2530 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2531 return; /* Still not fully idle. */
2535 /* Record start of fully idle period. */
2537 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2538 smp_mb__before_atomic();
2539 atomic_inc(&rdtp->dynticks_idle);
2540 smp_mb__after_atomic();
2541 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2545 * Unconditionally force exit from full system-idle state. This is
2546 * invoked when a normal CPU exits idle, but must be called separately
2547 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2548 * is that the timekeeping CPU is permitted to take scheduling-clock
2549 * interrupts while the system is in system-idle state, and of course
2550 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2551 * interrupt from any other type of interrupt.
2553 void rcu_sysidle_force_exit(void)
2555 int oldstate = READ_ONCE(full_sysidle_state);
2559 * Each pass through the following loop attempts to exit full
2560 * system-idle state. If contention proves to be a problem,
2561 * a trylock-based contention tree could be used here.
2563 while (oldstate > RCU_SYSIDLE_SHORT) {
2564 newoldstate = cmpxchg(&full_sysidle_state,
2565 oldstate, RCU_SYSIDLE_NOT);
2566 if (oldstate == newoldstate &&
2567 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2568 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2569 return; /* We cleared it, done! */
2571 oldstate = newoldstate;
2573 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2577 * Invoked to note entry to irq or task transition from idle. Note that
2578 * usermode execution does -not- count as idle here! The caller must
2579 * have disabled interrupts.
2581 static void rcu_sysidle_exit(int irq)
2583 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2585 /* If there are no nohz_full= CPUs, no need to track this. */
2586 if (!tick_nohz_full_enabled())
2589 /* Adjust nesting, check for already non-idle. */
2591 rdtp->dynticks_idle_nesting++;
2592 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2593 if (rdtp->dynticks_idle_nesting != 1)
2594 return; /* Already non-idle. */
2597 * Allow for irq misnesting. Yes, it really is possible
2598 * to enter an irq handler then never leave it, and maybe
2599 * also vice versa. Handle both possibilities.
2601 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2602 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2603 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2604 return; /* Already non-idle. */
2606 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2610 /* Record end of idle period. */
2611 smp_mb__before_atomic();
2612 atomic_inc(&rdtp->dynticks_idle);
2613 smp_mb__after_atomic();
2614 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2617 * If we are the timekeeping CPU, we are permitted to be non-idle
2618 * during a system-idle state. This must be the case, because
2619 * the timekeeping CPU has to take scheduling-clock interrupts
2620 * during the time that the system is transitioning to full
2621 * system-idle state. This means that the timekeeping CPU must
2622 * invoke rcu_sysidle_force_exit() directly if it does anything
2623 * more than take a scheduling-clock interrupt.
2625 if (smp_processor_id() == tick_do_timer_cpu)
2628 /* Update system-idle state: We are clearly no longer fully idle! */
2629 rcu_sysidle_force_exit();
2633 * Check to see if the current CPU is idle. Note that usermode execution
2634 * does not count as idle. The caller must have disabled interrupts,
2635 * and must be running on tick_do_timer_cpu.
2637 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2638 unsigned long *maxj)
2642 struct rcu_dynticks *rdtp = rdp->dynticks;
2644 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2645 if (!tick_nohz_full_enabled())
2649 * If some other CPU has already reported non-idle, if this is
2650 * not the flavor of RCU that tracks sysidle state, or if this
2651 * is an offline or the timekeeping CPU, nothing to do.
2653 if (!*isidle || rdp->rsp != rcu_state_p ||
2654 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2656 /* Verify affinity of current kthread. */
2657 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2659 /* Pick up current idle and NMI-nesting counter and check. */
2660 cur = atomic_read(&rdtp->dynticks_idle);
2662 *isidle = false; /* We are not idle! */
2665 smp_mb(); /* Read counters before timestamps. */
2667 /* Pick up timestamps. */
2668 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2669 /* If this CPU entered idle more recently, update maxj timestamp. */
2670 if (ULONG_CMP_LT(*maxj, j))
2675 * Is this the flavor of RCU that is handling full-system idle?
2677 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2679 return rsp == rcu_state_p;
2683 * Return a delay in jiffies based on the number of CPUs, rcu_node
2684 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2685 * systems more time to transition to full-idle state in order to
2686 * avoid the cache thrashing that otherwise occur on the state variable.
2687 * Really small systems (less than a couple of tens of CPUs) should
2688 * instead use a single global atomically incremented counter, and later
2689 * versions of this will automatically reconfigure themselves accordingly.
2691 static unsigned long rcu_sysidle_delay(void)
2693 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2695 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2699 * Advance the full-system-idle state. This is invoked when all of
2700 * the non-timekeeping CPUs are idle.
2702 static void rcu_sysidle(unsigned long j)
2704 /* Check the current state. */
2705 switch (READ_ONCE(full_sysidle_state)) {
2706 case RCU_SYSIDLE_NOT:
2708 /* First time all are idle, so note a short idle period. */
2709 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2712 case RCU_SYSIDLE_SHORT:
2715 * Idle for a bit, time to advance to next state?
2716 * cmpxchg failure means race with non-idle, let them win.
2718 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2719 (void)cmpxchg(&full_sysidle_state,
2720 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2723 case RCU_SYSIDLE_LONG:
2726 * Do an additional check pass before advancing to full.
2727 * cmpxchg failure means race with non-idle, let them win.
2729 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2730 (void)cmpxchg(&full_sysidle_state,
2731 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2740 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2741 * back to the beginning.
2743 static void rcu_sysidle_cancel(void)
2746 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2747 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2751 * Update the sysidle state based on the results of a force-quiescent-state
2752 * scan of the CPUs' dyntick-idle state.
2754 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2755 unsigned long maxj, bool gpkt)
2757 if (rsp != rcu_state_p)
2758 return; /* Wrong flavor, ignore. */
2759 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2760 return; /* Running state machine from timekeeping CPU. */
2762 rcu_sysidle(maxj); /* More idle! */
2764 rcu_sysidle_cancel(); /* Idle is over. */
2768 * Wrapper for rcu_sysidle_report() when called from the grace-period
2769 * kthread's context.
2771 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2774 /* If there are no nohz_full= CPUs, no need to track this. */
2775 if (!tick_nohz_full_enabled())
2778 rcu_sysidle_report(rsp, isidle, maxj, true);
2781 /* Callback and function for forcing an RCU grace period. */
2782 struct rcu_sysidle_head {
2787 static void rcu_sysidle_cb(struct rcu_head *rhp)
2789 struct rcu_sysidle_head *rshp;
2792 * The following memory barrier is needed to replace the
2793 * memory barriers that would normally be in the memory
2796 smp_mb(); /* grace period precedes setting inuse. */
2798 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2799 WRITE_ONCE(rshp->inuse, 0);
2803 * Check to see if the system is fully idle, other than the timekeeping CPU.
2804 * The caller must have disabled interrupts. This is not intended to be
2805 * called unless tick_nohz_full_enabled().
2807 bool rcu_sys_is_idle(void)
2809 static struct rcu_sysidle_head rsh;
2810 int rss = READ_ONCE(full_sysidle_state);
2812 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2815 /* Handle small-system case by doing a full scan of CPUs. */
2816 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2817 int oldrss = rss - 1;
2820 * One pass to advance to each state up to _FULL.
2821 * Give up if any pass fails to advance the state.
2823 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2826 unsigned long maxj = jiffies - ULONG_MAX / 4;
2827 struct rcu_data *rdp;
2829 /* Scan all the CPUs looking for nonidle CPUs. */
2830 for_each_possible_cpu(cpu) {
2831 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2832 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2836 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2838 rss = READ_ONCE(full_sysidle_state);
2842 /* If this is the first observation of an idle period, record it. */
2843 if (rss == RCU_SYSIDLE_FULL) {
2844 rss = cmpxchg(&full_sysidle_state,
2845 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2846 return rss == RCU_SYSIDLE_FULL;
2849 smp_mb(); /* ensure rss load happens before later caller actions. */
2851 /* If already fully idle, tell the caller (in case of races). */
2852 if (rss == RCU_SYSIDLE_FULL_NOTED)
2856 * If we aren't there yet, and a grace period is not in flight,
2857 * initiate a grace period. Either way, tell the caller that
2858 * we are not there yet. We use an xchg() rather than an assignment
2859 * to make up for the memory barriers that would otherwise be
2860 * provided by the memory allocator.
2862 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2863 !rcu_gp_in_progress(rcu_state_p) &&
2864 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2865 call_rcu(&rsh.rh, rcu_sysidle_cb);
2870 * Initialize dynticks sysidle state for CPUs coming online.
2872 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2874 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2877 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2879 static void rcu_sysidle_enter(int irq)
2883 static void rcu_sysidle_exit(int irq)
2887 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2888 unsigned long *maxj)
2892 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2897 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2902 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2906 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2909 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2910 * grace-period kthread will do force_quiescent_state() processing?
2911 * The idea is to avoid waking up RCU core processing on such a
2912 * CPU unless the grace period has extended for too long.
2914 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2915 * CONFIG_RCU_NOCB_CPU CPUs.
2917 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2919 #ifdef CONFIG_NO_HZ_FULL
2920 if (tick_nohz_full_cpu(smp_processor_id()) &&
2921 (!rcu_gp_in_progress(rsp) ||
2922 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
2924 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2929 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2932 static void rcu_bind_gp_kthread(void)
2934 int __maybe_unused cpu;
2936 if (!tick_nohz_full_enabled())
2938 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2939 cpu = tick_do_timer_cpu;
2940 if (cpu >= 0 && cpu < nr_cpu_ids)
2941 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2942 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2943 housekeeping_affine(current);
2944 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2947 /* Record the current task on dyntick-idle entry. */
2948 static void rcu_dynticks_task_enter(void)
2950 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2951 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2952 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2955 /* Record no current task on dyntick-idle exit. */
2956 static void rcu_dynticks_task_exit(void)
2958 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2959 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2960 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */