2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
31 #include "../time/tick-internal.h"
33 #ifdef CONFIG_RCU_BOOST
35 #include "../locking/rtmutex_common.h"
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
41 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
42 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
43 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
44 DEFINE_PER_CPU(char, rcu_cpu_has_work);
46 #else /* #ifdef CONFIG_RCU_BOOST */
49 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
50 * all uses are in dead code. Provide a definition to keep the compiler
51 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
52 * This probably needs to be excluded from -rt builds.
54 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
56 #endif /* #else #ifdef CONFIG_RCU_BOOST */
58 #ifdef CONFIG_RCU_NOCB_CPU
59 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
60 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
61 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
62 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
65 * Check the RCU kernel configuration parameters and print informative
66 * messages about anything out of the ordinary. If you like #ifdef, you
67 * will love this function.
69 static void __init rcu_bootup_announce_oddness(void)
71 if (IS_ENABLED(CONFIG_RCU_TRACE))
72 pr_info("\tRCU debugfs-based tracing is enabled.\n");
73 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
74 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
75 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
78 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
79 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
80 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
81 if (IS_ENABLED(CONFIG_PROVE_RCU))
82 pr_info("\tRCU lockdep checking is enabled.\n");
83 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE))
84 pr_info("\tRCU torture testing starts during boot.\n");
85 if (RCU_NUM_LVLS >= 4)
86 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
87 if (RCU_FANOUT_LEAF != 16)
88 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
90 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
91 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
92 if (nr_cpu_ids != NR_CPUS)
93 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
94 if (IS_ENABLED(CONFIG_RCU_BOOST))
95 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
98 #ifdef CONFIG_PREEMPT_RCU
100 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
101 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
102 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
104 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
105 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
109 * Tell them what RCU they are running.
111 static void __init rcu_bootup_announce(void)
113 pr_info("Preemptible hierarchical RCU implementation.\n");
114 rcu_bootup_announce_oddness();
118 * Record a preemptible-RCU quiescent state for the specified CPU. Note
119 * that this just means that the task currently running on the CPU is
120 * not in a quiescent state. There might be any number of tasks blocked
121 * while in an RCU read-side critical section.
123 * As with the other rcu_*_qs() functions, callers to this function
124 * must disable preemption.
126 static void rcu_preempt_qs(void)
128 if (!__this_cpu_read(rcu_data_p->passed_quiesce)) {
129 trace_rcu_grace_period(TPS("rcu_preempt"),
130 __this_cpu_read(rcu_data_p->gpnum),
132 __this_cpu_write(rcu_data_p->passed_quiesce, 1);
133 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
134 current->rcu_read_unlock_special.b.need_qs = false;
139 * We have entered the scheduler, and the current task might soon be
140 * context-switched away from. If this task is in an RCU read-side
141 * critical section, we will no longer be able to rely on the CPU to
142 * record that fact, so we enqueue the task on the blkd_tasks list.
143 * The task will dequeue itself when it exits the outermost enclosing
144 * RCU read-side critical section. Therefore, the current grace period
145 * cannot be permitted to complete until the blkd_tasks list entries
146 * predating the current grace period drain, in other words, until
147 * rnp->gp_tasks becomes NULL.
149 * Caller must disable preemption.
151 static void rcu_preempt_note_context_switch(void)
153 struct task_struct *t = current;
155 struct rcu_data *rdp;
156 struct rcu_node *rnp;
158 if (t->rcu_read_lock_nesting > 0 &&
159 !t->rcu_read_unlock_special.b.blocked) {
161 /* Possibly blocking in an RCU read-side critical section. */
162 rdp = this_cpu_ptr(rcu_state_p->rda);
164 raw_spin_lock_irqsave(&rnp->lock, flags);
165 smp_mb__after_unlock_lock();
166 t->rcu_read_unlock_special.b.blocked = true;
167 t->rcu_blocked_node = rnp;
170 * If this CPU has already checked in, then this task
171 * will hold up the next grace period rather than the
172 * current grace period. Queue the task accordingly.
173 * If the task is queued for the current grace period
174 * (i.e., this CPU has not yet passed through a quiescent
175 * state for the current grace period), then as long
176 * as that task remains queued, the current grace period
177 * cannot end. Note that there is some uncertainty as
178 * to exactly when the current grace period started.
179 * We take a conservative approach, which can result
180 * in unnecessarily waiting on tasks that started very
181 * slightly after the current grace period began. C'est
184 * But first, note that the current CPU must still be
187 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
188 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
189 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
190 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
191 rnp->gp_tasks = &t->rcu_node_entry;
192 if (IS_ENABLED(CONFIG_RCU_BOOST) &&
193 rnp->boost_tasks != NULL)
194 rnp->boost_tasks = rnp->gp_tasks;
196 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
197 if (rnp->qsmask & rdp->grpmask)
198 rnp->gp_tasks = &t->rcu_node_entry;
200 trace_rcu_preempt_task(rdp->rsp->name,
202 (rnp->qsmask & rdp->grpmask)
205 raw_spin_unlock_irqrestore(&rnp->lock, flags);
206 } else if (t->rcu_read_lock_nesting < 0 &&
207 t->rcu_read_unlock_special.s) {
210 * Complete exit from RCU read-side critical section on
211 * behalf of preempted instance of __rcu_read_unlock().
213 rcu_read_unlock_special(t);
217 * Either we were not in an RCU read-side critical section to
218 * begin with, or we have now recorded that critical section
219 * globally. Either way, we can now note a quiescent state
220 * for this CPU. Again, if we were in an RCU read-side critical
221 * section, and if that critical section was blocking the current
222 * grace period, then the fact that the task has been enqueued
223 * means that we continue to block the current grace period.
229 * Check for preempted RCU readers blocking the current grace period
230 * for the specified rcu_node structure. If the caller needs a reliable
231 * answer, it must hold the rcu_node's ->lock.
233 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
235 return rnp->gp_tasks != NULL;
239 * Advance a ->blkd_tasks-list pointer to the next entry, instead
240 * returning NULL if at the end of the list.
242 static struct list_head *rcu_next_node_entry(struct task_struct *t,
243 struct rcu_node *rnp)
245 struct list_head *np;
247 np = t->rcu_node_entry.next;
248 if (np == &rnp->blkd_tasks)
254 * Return true if the specified rcu_node structure has tasks that were
255 * preempted within an RCU read-side critical section.
257 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
259 return !list_empty(&rnp->blkd_tasks);
263 * Handle special cases during rcu_read_unlock(), such as needing to
264 * notify RCU core processing or task having blocked during the RCU
265 * read-side critical section.
267 void rcu_read_unlock_special(struct task_struct *t)
273 struct list_head *np;
274 bool drop_boost_mutex = false;
275 struct rcu_node *rnp;
276 union rcu_special special;
278 /* NMI handlers cannot block and cannot safely manipulate state. */
282 local_irq_save(flags);
285 * If RCU core is waiting for this CPU to exit critical section,
286 * let it know that we have done so. Because irqs are disabled,
287 * t->rcu_read_unlock_special cannot change.
289 special = t->rcu_read_unlock_special;
290 if (special.b.need_qs) {
292 t->rcu_read_unlock_special.b.need_qs = false;
293 if (!t->rcu_read_unlock_special.s) {
294 local_irq_restore(flags);
299 /* Hardware IRQ handlers cannot block, complain if they get here. */
300 if (in_irq() || in_serving_softirq()) {
301 lockdep_rcu_suspicious(__FILE__, __LINE__,
302 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
303 pr_alert("->rcu_read_unlock_special: %#x (b: %d, nq: %d)\n",
304 t->rcu_read_unlock_special.s,
305 t->rcu_read_unlock_special.b.blocked,
306 t->rcu_read_unlock_special.b.need_qs);
307 local_irq_restore(flags);
311 /* Clean up if blocked during RCU read-side critical section. */
312 if (special.b.blocked) {
313 t->rcu_read_unlock_special.b.blocked = false;
316 * Remove this task from the list it blocked on. The task
317 * now remains queued on the rcu_node corresponding to
318 * the CPU it first blocked on, so the first attempt to
319 * acquire the task's rcu_node's ->lock will succeed.
320 * Keep the loop and add a WARN_ON() out of sheer paranoia.
323 rnp = t->rcu_blocked_node;
324 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
325 smp_mb__after_unlock_lock();
326 if (rnp == t->rcu_blocked_node)
329 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
331 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
332 empty_exp = !rcu_preempted_readers_exp(rnp);
333 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
334 np = rcu_next_node_entry(t, rnp);
335 list_del_init(&t->rcu_node_entry);
336 t->rcu_blocked_node = NULL;
337 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
339 if (&t->rcu_node_entry == rnp->gp_tasks)
341 if (&t->rcu_node_entry == rnp->exp_tasks)
343 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
344 if (&t->rcu_node_entry == rnp->boost_tasks)
345 rnp->boost_tasks = np;
346 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
347 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
351 * If this was the last task on the current list, and if
352 * we aren't waiting on any CPUs, report the quiescent state.
353 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
354 * so we must take a snapshot of the expedited state.
356 empty_exp_now = !rcu_preempted_readers_exp(rnp);
357 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
358 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
365 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
367 raw_spin_unlock_irqrestore(&rnp->lock, flags);
370 /* Unboost if we were boosted. */
371 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
372 rt_mutex_unlock(&rnp->boost_mtx);
375 * If this was the last task on the expedited lists,
376 * then we need to report up the rcu_node hierarchy.
378 if (!empty_exp && empty_exp_now)
379 rcu_report_exp_rnp(rcu_state_p, rnp, true);
381 local_irq_restore(flags);
386 * Dump detailed information for all tasks blocking the current RCU
387 * grace period on the specified rcu_node structure.
389 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
392 struct task_struct *t;
394 raw_spin_lock_irqsave(&rnp->lock, flags);
395 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
396 raw_spin_unlock_irqrestore(&rnp->lock, flags);
399 t = list_entry(rnp->gp_tasks->prev,
400 struct task_struct, rcu_node_entry);
401 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
403 raw_spin_unlock_irqrestore(&rnp->lock, flags);
407 * Dump detailed information for all tasks blocking the current RCU
410 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
412 struct rcu_node *rnp = rcu_get_root(rsp);
414 rcu_print_detail_task_stall_rnp(rnp);
415 rcu_for_each_leaf_node(rsp, rnp)
416 rcu_print_detail_task_stall_rnp(rnp);
419 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
421 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
422 rnp->level, rnp->grplo, rnp->grphi);
425 static void rcu_print_task_stall_end(void)
431 * Scan the current list of tasks blocked within RCU read-side critical
432 * sections, printing out the tid of each.
434 static int rcu_print_task_stall(struct rcu_node *rnp)
436 struct task_struct *t;
439 if (!rcu_preempt_blocked_readers_cgp(rnp))
441 rcu_print_task_stall_begin(rnp);
442 t = list_entry(rnp->gp_tasks->prev,
443 struct task_struct, rcu_node_entry);
444 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
445 pr_cont(" P%d", t->pid);
448 rcu_print_task_stall_end();
453 * Check that the list of blocked tasks for the newly completed grace
454 * period is in fact empty. It is a serious bug to complete a grace
455 * period that still has RCU readers blocked! This function must be
456 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
457 * must be held by the caller.
459 * Also, if there are blocked tasks on the list, they automatically
460 * block the newly created grace period, so set up ->gp_tasks accordingly.
462 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
464 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
465 if (rcu_preempt_has_tasks(rnp))
466 rnp->gp_tasks = rnp->blkd_tasks.next;
467 WARN_ON_ONCE(rnp->qsmask);
471 * Check for a quiescent state from the current CPU. When a task blocks,
472 * the task is recorded in the corresponding CPU's rcu_node structure,
473 * which is checked elsewhere.
475 * Caller must disable hard irqs.
477 static void rcu_preempt_check_callbacks(void)
479 struct task_struct *t = current;
481 if (t->rcu_read_lock_nesting == 0) {
485 if (t->rcu_read_lock_nesting > 0 &&
486 __this_cpu_read(rcu_data_p->qs_pending) &&
487 !__this_cpu_read(rcu_data_p->passed_quiesce))
488 t->rcu_read_unlock_special.b.need_qs = true;
491 #ifdef CONFIG_RCU_BOOST
493 static void rcu_preempt_do_callbacks(void)
495 rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p));
498 #endif /* #ifdef CONFIG_RCU_BOOST */
501 * Queue a preemptible-RCU callback for invocation after a grace period.
503 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
505 __call_rcu(head, func, rcu_state_p, -1, 0);
507 EXPORT_SYMBOL_GPL(call_rcu);
510 * synchronize_rcu - wait until a grace period has elapsed.
512 * Control will return to the caller some time after a full grace
513 * period has elapsed, in other words after all currently executing RCU
514 * read-side critical sections have completed. Note, however, that
515 * upon return from synchronize_rcu(), the caller might well be executing
516 * concurrently with new RCU read-side critical sections that began while
517 * synchronize_rcu() was waiting. RCU read-side critical sections are
518 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
520 * See the description of synchronize_sched() for more detailed information
521 * on memory ordering guarantees.
523 void synchronize_rcu(void)
525 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
526 lock_is_held(&rcu_lock_map) ||
527 lock_is_held(&rcu_sched_lock_map),
528 "Illegal synchronize_rcu() in RCU read-side critical section");
529 if (!rcu_scheduler_active)
531 if (rcu_gp_is_expedited())
532 synchronize_rcu_expedited();
534 wait_rcu_gp(call_rcu);
536 EXPORT_SYMBOL_GPL(synchronize_rcu);
539 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
540 * grace period for the specified rcu_node structure, phase 1. If there
541 * are such tasks, set the ->expmask bits up the rcu_node tree and also
542 * set the ->expmask bits on the leaf rcu_node structures to tell phase 2
543 * that work is needed here.
545 * Caller must hold the root rcu_node's exp_funnel_mutex.
548 sync_rcu_preempt_exp_init1(struct rcu_state *rsp, struct rcu_node *rnp)
552 struct rcu_node *rnp_up;
554 raw_spin_lock_irqsave(&rnp->lock, flags);
555 smp_mb__after_unlock_lock();
556 WARN_ON_ONCE(rnp->expmask);
557 WARN_ON_ONCE(rnp->exp_tasks);
558 if (!rcu_preempt_has_tasks(rnp)) {
559 /* No blocked tasks, nothing to do. */
560 raw_spin_unlock_irqrestore(&rnp->lock, flags);
563 /* Call for Phase 2 and propagate ->expmask bits up the tree. */
566 while (rnp_up->parent) {
567 mask = rnp_up->grpmask;
568 rnp_up = rnp_up->parent;
569 if (rnp_up->expmask & mask)
571 raw_spin_lock(&rnp_up->lock); /* irqs already off */
572 smp_mb__after_unlock_lock();
573 rnp_up->expmask |= mask;
574 raw_spin_unlock(&rnp_up->lock); /* irqs still off */
576 raw_spin_unlock_irqrestore(&rnp->lock, flags);
580 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
581 * grace period for the specified rcu_node structure, phase 2. If the
582 * leaf rcu_node structure has its ->expmask field set, check for tasks.
583 * If there are some, clear ->expmask and set ->exp_tasks accordingly,
584 * then initiate RCU priority boosting. Otherwise, clear ->expmask and
585 * invoke rcu_report_exp_rnp() to clear out the upper-level ->expmask bits,
586 * enabling rcu_read_unlock_special() to do the bit-clearing.
588 * Caller must hold the root rcu_node's exp_funnel_mutex.
591 sync_rcu_preempt_exp_init2(struct rcu_state *rsp, struct rcu_node *rnp)
595 raw_spin_lock_irqsave(&rnp->lock, flags);
596 smp_mb__after_unlock_lock();
598 /* Phase 1 didn't do anything, so Phase 2 doesn't either. */
599 raw_spin_unlock_irqrestore(&rnp->lock, flags);
603 /* Phase 1 is over. */
607 * If there are still blocked tasks, set up ->exp_tasks so that
608 * rcu_read_unlock_special() will wake us and then boost them.
610 if (rcu_preempt_has_tasks(rnp)) {
611 rnp->exp_tasks = rnp->blkd_tasks.next;
612 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
616 /* No longer any blocked tasks, so undo bit setting. */
617 raw_spin_unlock_irqrestore(&rnp->lock, flags);
618 rcu_report_exp_rnp(rsp, rnp, false);
622 * synchronize_rcu_expedited - Brute-force RCU grace period
624 * Wait for an RCU-preempt grace period, but expedite it. The basic
625 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
626 * the ->blkd_tasks lists and wait for this list to drain. This consumes
627 * significant time on all CPUs and is unfriendly to real-time workloads,
628 * so is thus not recommended for any sort of common-case code.
629 * In fact, if you are using synchronize_rcu_expedited() in a loop,
630 * please restructure your code to batch your updates, and then Use a
631 * single synchronize_rcu() instead.
633 void synchronize_rcu_expedited(void)
635 struct rcu_node *rnp;
636 struct rcu_node *rnp_unlock;
637 struct rcu_state *rsp = rcu_state_p;
640 s = rcu_exp_gp_seq_snap(rsp);
642 rnp_unlock = exp_funnel_lock(rsp, s);
643 if (rnp_unlock == NULL)
644 return; /* Someone else did our work for us. */
646 rcu_exp_gp_seq_start(rsp);
648 /* force all RCU readers onto ->blkd_tasks lists. */
649 synchronize_sched_expedited();
652 * Snapshot current state of ->blkd_tasks lists into ->expmask.
653 * Phase 1 sets bits and phase 2 permits rcu_read_unlock_special()
654 * to start clearing them. Doing this in one phase leads to
655 * strange races between setting and clearing bits, so just say "no"!
657 rcu_for_each_leaf_node(rsp, rnp)
658 sync_rcu_preempt_exp_init1(rsp, rnp);
659 rcu_for_each_leaf_node(rsp, rnp)
660 sync_rcu_preempt_exp_init2(rsp, rnp);
662 /* Wait for snapshotted ->blkd_tasks lists to drain. */
663 rnp = rcu_get_root(rsp);
664 wait_event(rsp->expedited_wq,
665 sync_rcu_preempt_exp_done(rnp));
667 /* Clean up and exit. */
668 rcu_exp_gp_seq_end(rsp);
669 mutex_unlock(&rnp_unlock->exp_funnel_mutex);
671 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
674 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
676 * Note that this primitive does not necessarily wait for an RCU grace period
677 * to complete. For example, if there are no RCU callbacks queued anywhere
678 * in the system, then rcu_barrier() is within its rights to return
679 * immediately, without waiting for anything, much less an RCU grace period.
681 void rcu_barrier(void)
683 _rcu_barrier(rcu_state_p);
685 EXPORT_SYMBOL_GPL(rcu_barrier);
688 * Initialize preemptible RCU's state structures.
690 static void __init __rcu_init_preempt(void)
692 rcu_init_one(rcu_state_p, rcu_data_p);
696 * Check for a task exiting while in a preemptible-RCU read-side
697 * critical section, clean up if so. No need to issue warnings,
698 * as debug_check_no_locks_held() already does this if lockdep
703 struct task_struct *t = current;
705 if (likely(list_empty(¤t->rcu_node_entry)))
707 t->rcu_read_lock_nesting = 1;
709 t->rcu_read_unlock_special.b.blocked = true;
713 #else /* #ifdef CONFIG_PREEMPT_RCU */
715 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
716 static struct rcu_data __percpu *const rcu_data_p = &rcu_sched_data;
719 * Tell them what RCU they are running.
721 static void __init rcu_bootup_announce(void)
723 pr_info("Hierarchical RCU implementation.\n");
724 rcu_bootup_announce_oddness();
728 * Because preemptible RCU does not exist, we never have to check for
729 * CPUs being in quiescent states.
731 static void rcu_preempt_note_context_switch(void)
736 * Because preemptible RCU does not exist, there are never any preempted
739 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
745 * Because there is no preemptible RCU, there can be no readers blocked.
747 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
753 * Because preemptible RCU does not exist, we never have to check for
754 * tasks blocked within RCU read-side critical sections.
756 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
761 * Because preemptible RCU does not exist, we never have to check for
762 * tasks blocked within RCU read-side critical sections.
764 static int rcu_print_task_stall(struct rcu_node *rnp)
770 * Because there is no preemptible RCU, there can be no readers blocked,
771 * so there is no need to check for blocked tasks. So check only for
772 * bogus qsmask values.
774 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
776 WARN_ON_ONCE(rnp->qsmask);
780 * Because preemptible RCU does not exist, it never has any callbacks
783 static void rcu_preempt_check_callbacks(void)
788 * Wait for an rcu-preempt grace period, but make it happen quickly.
789 * But because preemptible RCU does not exist, map to rcu-sched.
791 void synchronize_rcu_expedited(void)
793 synchronize_sched_expedited();
795 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
798 * Because preemptible RCU does not exist, rcu_barrier() is just
799 * another name for rcu_barrier_sched().
801 void rcu_barrier(void)
805 EXPORT_SYMBOL_GPL(rcu_barrier);
808 * Because preemptible RCU does not exist, it need not be initialized.
810 static void __init __rcu_init_preempt(void)
815 * Because preemptible RCU does not exist, tasks cannot possibly exit
816 * while in preemptible RCU read-side critical sections.
822 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
824 #ifdef CONFIG_RCU_BOOST
826 #include "../locking/rtmutex_common.h"
828 #ifdef CONFIG_RCU_TRACE
830 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
832 if (!rcu_preempt_has_tasks(rnp))
833 rnp->n_balk_blkd_tasks++;
834 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
835 rnp->n_balk_exp_gp_tasks++;
836 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
837 rnp->n_balk_boost_tasks++;
838 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
839 rnp->n_balk_notblocked++;
840 else if (rnp->gp_tasks != NULL &&
841 ULONG_CMP_LT(jiffies, rnp->boost_time))
842 rnp->n_balk_notyet++;
847 #else /* #ifdef CONFIG_RCU_TRACE */
849 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
853 #endif /* #else #ifdef CONFIG_RCU_TRACE */
855 static void rcu_wake_cond(struct task_struct *t, int status)
858 * If the thread is yielding, only wake it when this
859 * is invoked from idle
861 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
866 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
867 * or ->boost_tasks, advancing the pointer to the next task in the
870 * Note that irqs must be enabled: boosting the task can block.
871 * Returns 1 if there are more tasks needing to be boosted.
873 static int rcu_boost(struct rcu_node *rnp)
876 struct task_struct *t;
877 struct list_head *tb;
879 if (READ_ONCE(rnp->exp_tasks) == NULL &&
880 READ_ONCE(rnp->boost_tasks) == NULL)
881 return 0; /* Nothing left to boost. */
883 raw_spin_lock_irqsave(&rnp->lock, flags);
884 smp_mb__after_unlock_lock();
887 * Recheck under the lock: all tasks in need of boosting
888 * might exit their RCU read-side critical sections on their own.
890 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
891 raw_spin_unlock_irqrestore(&rnp->lock, flags);
896 * Preferentially boost tasks blocking expedited grace periods.
897 * This cannot starve the normal grace periods because a second
898 * expedited grace period must boost all blocked tasks, including
899 * those blocking the pre-existing normal grace period.
901 if (rnp->exp_tasks != NULL) {
905 tb = rnp->boost_tasks;
906 rnp->n_normal_boosts++;
908 rnp->n_tasks_boosted++;
911 * We boost task t by manufacturing an rt_mutex that appears to
912 * be held by task t. We leave a pointer to that rt_mutex where
913 * task t can find it, and task t will release the mutex when it
914 * exits its outermost RCU read-side critical section. Then
915 * simply acquiring this artificial rt_mutex will boost task
916 * t's priority. (Thanks to tglx for suggesting this approach!)
918 * Note that task t must acquire rnp->lock to remove itself from
919 * the ->blkd_tasks list, which it will do from exit() if from
920 * nowhere else. We therefore are guaranteed that task t will
921 * stay around at least until we drop rnp->lock. Note that
922 * rnp->lock also resolves races between our priority boosting
923 * and task t's exiting its outermost RCU read-side critical
926 t = container_of(tb, struct task_struct, rcu_node_entry);
927 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
928 raw_spin_unlock_irqrestore(&rnp->lock, flags);
929 /* Lock only for side effect: boosts task t's priority. */
930 rt_mutex_lock(&rnp->boost_mtx);
931 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
933 return READ_ONCE(rnp->exp_tasks) != NULL ||
934 READ_ONCE(rnp->boost_tasks) != NULL;
938 * Priority-boosting kthread, one per leaf rcu_node.
940 static int rcu_boost_kthread(void *arg)
942 struct rcu_node *rnp = (struct rcu_node *)arg;
946 trace_rcu_utilization(TPS("Start boost kthread@init"));
948 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
949 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
950 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
951 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
952 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
953 more2boost = rcu_boost(rnp);
959 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
960 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
961 schedule_timeout_interruptible(2);
962 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
967 trace_rcu_utilization(TPS("End boost kthread@notreached"));
972 * Check to see if it is time to start boosting RCU readers that are
973 * blocking the current grace period, and, if so, tell the per-rcu_node
974 * kthread to start boosting them. If there is an expedited grace
975 * period in progress, it is always time to boost.
977 * The caller must hold rnp->lock, which this function releases.
978 * The ->boost_kthread_task is immortal, so we don't need to worry
979 * about it going away.
981 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
982 __releases(rnp->lock)
984 struct task_struct *t;
986 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
987 rnp->n_balk_exp_gp_tasks++;
988 raw_spin_unlock_irqrestore(&rnp->lock, flags);
991 if (rnp->exp_tasks != NULL ||
992 (rnp->gp_tasks != NULL &&
993 rnp->boost_tasks == NULL &&
995 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
996 if (rnp->exp_tasks == NULL)
997 rnp->boost_tasks = rnp->gp_tasks;
998 raw_spin_unlock_irqrestore(&rnp->lock, flags);
999 t = rnp->boost_kthread_task;
1001 rcu_wake_cond(t, rnp->boost_kthread_status);
1003 rcu_initiate_boost_trace(rnp);
1004 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1009 * Wake up the per-CPU kthread to invoke RCU callbacks.
1011 static void invoke_rcu_callbacks_kthread(void)
1013 unsigned long flags;
1015 local_irq_save(flags);
1016 __this_cpu_write(rcu_cpu_has_work, 1);
1017 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1018 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1019 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1020 __this_cpu_read(rcu_cpu_kthread_status));
1022 local_irq_restore(flags);
1026 * Is the current CPU running the RCU-callbacks kthread?
1027 * Caller must have preemption disabled.
1029 static bool rcu_is_callbacks_kthread(void)
1031 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1034 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1037 * Do priority-boost accounting for the start of a new grace period.
1039 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1041 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1045 * Create an RCU-boost kthread for the specified node if one does not
1046 * already exist. We only create this kthread for preemptible RCU.
1047 * Returns zero if all is well, a negated errno otherwise.
1049 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1050 struct rcu_node *rnp)
1052 int rnp_index = rnp - &rsp->node[0];
1053 unsigned long flags;
1054 struct sched_param sp;
1055 struct task_struct *t;
1057 if (rcu_state_p != rsp)
1060 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1064 if (rnp->boost_kthread_task != NULL)
1066 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1067 "rcub/%d", rnp_index);
1070 raw_spin_lock_irqsave(&rnp->lock, flags);
1071 smp_mb__after_unlock_lock();
1072 rnp->boost_kthread_task = t;
1073 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1074 sp.sched_priority = kthread_prio;
1075 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1076 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1080 static void rcu_kthread_do_work(void)
1082 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1083 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1084 rcu_preempt_do_callbacks();
1087 static void rcu_cpu_kthread_setup(unsigned int cpu)
1089 struct sched_param sp;
1091 sp.sched_priority = kthread_prio;
1092 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1095 static void rcu_cpu_kthread_park(unsigned int cpu)
1097 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1100 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1102 return __this_cpu_read(rcu_cpu_has_work);
1106 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1107 * RCU softirq used in flavors and configurations of RCU that do not
1108 * support RCU priority boosting.
1110 static void rcu_cpu_kthread(unsigned int cpu)
1112 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1113 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1116 for (spincnt = 0; spincnt < 10; spincnt++) {
1117 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1119 *statusp = RCU_KTHREAD_RUNNING;
1120 this_cpu_inc(rcu_cpu_kthread_loops);
1121 local_irq_disable();
1126 rcu_kthread_do_work();
1129 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1130 *statusp = RCU_KTHREAD_WAITING;
1134 *statusp = RCU_KTHREAD_YIELDING;
1135 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1136 schedule_timeout_interruptible(2);
1137 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1138 *statusp = RCU_KTHREAD_WAITING;
1142 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1143 * served by the rcu_node in question. The CPU hotplug lock is still
1144 * held, so the value of rnp->qsmaskinit will be stable.
1146 * We don't include outgoingcpu in the affinity set, use -1 if there is
1147 * no outgoing CPU. If there are no CPUs left in the affinity set,
1148 * this function allows the kthread to execute on any CPU.
1150 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1152 struct task_struct *t = rnp->boost_kthread_task;
1153 unsigned long mask = rcu_rnp_online_cpus(rnp);
1159 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1161 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1162 if ((mask & 0x1) && cpu != outgoingcpu)
1163 cpumask_set_cpu(cpu, cm);
1164 if (cpumask_weight(cm) == 0)
1166 set_cpus_allowed_ptr(t, cm);
1167 free_cpumask_var(cm);
1170 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1171 .store = &rcu_cpu_kthread_task,
1172 .thread_should_run = rcu_cpu_kthread_should_run,
1173 .thread_fn = rcu_cpu_kthread,
1174 .thread_comm = "rcuc/%u",
1175 .setup = rcu_cpu_kthread_setup,
1176 .park = rcu_cpu_kthread_park,
1180 * Spawn boost kthreads -- called as soon as the scheduler is running.
1182 static void __init rcu_spawn_boost_kthreads(void)
1184 struct rcu_node *rnp;
1187 for_each_possible_cpu(cpu)
1188 per_cpu(rcu_cpu_has_work, cpu) = 0;
1189 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1190 rcu_for_each_leaf_node(rcu_state_p, rnp)
1191 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1194 static void rcu_prepare_kthreads(int cpu)
1196 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1197 struct rcu_node *rnp = rdp->mynode;
1199 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1200 if (rcu_scheduler_fully_active)
1201 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1204 #else /* #ifdef CONFIG_RCU_BOOST */
1206 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1207 __releases(rnp->lock)
1209 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1212 static void invoke_rcu_callbacks_kthread(void)
1217 static bool rcu_is_callbacks_kthread(void)
1222 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1226 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1230 static void __init rcu_spawn_boost_kthreads(void)
1234 static void rcu_prepare_kthreads(int cpu)
1238 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1240 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1243 * Check to see if any future RCU-related work will need to be done
1244 * by the current CPU, even if none need be done immediately, returning
1245 * 1 if so. This function is part of the RCU implementation; it is -not-
1246 * an exported member of the RCU API.
1248 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1249 * any flavor of RCU.
1251 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1253 *nextevt = KTIME_MAX;
1254 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1255 ? 0 : rcu_cpu_has_callbacks(NULL);
1259 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1262 static void rcu_cleanup_after_idle(void)
1267 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1270 static void rcu_prepare_for_idle(void)
1275 * Don't bother keeping a running count of the number of RCU callbacks
1276 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1278 static void rcu_idle_count_callbacks_posted(void)
1282 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1285 * This code is invoked when a CPU goes idle, at which point we want
1286 * to have the CPU do everything required for RCU so that it can enter
1287 * the energy-efficient dyntick-idle mode. This is handled by a
1288 * state machine implemented by rcu_prepare_for_idle() below.
1290 * The following three proprocessor symbols control this state machine:
1292 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1293 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1294 * is sized to be roughly one RCU grace period. Those energy-efficiency
1295 * benchmarkers who might otherwise be tempted to set this to a large
1296 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1297 * system. And if you are -that- concerned about energy efficiency,
1298 * just power the system down and be done with it!
1299 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1300 * permitted to sleep in dyntick-idle mode with only lazy RCU
1301 * callbacks pending. Setting this too high can OOM your system.
1303 * The values below work well in practice. If future workloads require
1304 * adjustment, they can be converted into kernel config parameters, though
1305 * making the state machine smarter might be a better option.
1307 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1308 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1310 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1311 module_param(rcu_idle_gp_delay, int, 0644);
1312 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1313 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1316 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1317 * only if it has been awhile since the last time we did so. Afterwards,
1318 * if there are any callbacks ready for immediate invocation, return true.
1320 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1322 bool cbs_ready = false;
1323 struct rcu_data *rdp;
1324 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1325 struct rcu_node *rnp;
1326 struct rcu_state *rsp;
1328 /* Exit early if we advanced recently. */
1329 if (jiffies == rdtp->last_advance_all)
1331 rdtp->last_advance_all = jiffies;
1333 for_each_rcu_flavor(rsp) {
1334 rdp = this_cpu_ptr(rsp->rda);
1338 * Don't bother checking unless a grace period has
1339 * completed since we last checked and there are
1340 * callbacks not yet ready to invoke.
1342 if ((rdp->completed != rnp->completed ||
1343 unlikely(READ_ONCE(rdp->gpwrap))) &&
1344 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1345 note_gp_changes(rsp, rdp);
1347 if (cpu_has_callbacks_ready_to_invoke(rdp))
1354 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1355 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1356 * caller to set the timeout based on whether or not there are non-lazy
1359 * The caller must have disabled interrupts.
1361 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1363 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1366 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1367 *nextevt = KTIME_MAX;
1371 /* Snapshot to detect later posting of non-lazy callback. */
1372 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1374 /* If no callbacks, RCU doesn't need the CPU. */
1375 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1376 *nextevt = KTIME_MAX;
1380 /* Attempt to advance callbacks. */
1381 if (rcu_try_advance_all_cbs()) {
1382 /* Some ready to invoke, so initiate later invocation. */
1386 rdtp->last_accelerate = jiffies;
1388 /* Request timer delay depending on laziness, and round. */
1389 if (!rdtp->all_lazy) {
1390 dj = round_up(rcu_idle_gp_delay + jiffies,
1391 rcu_idle_gp_delay) - jiffies;
1393 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1395 *nextevt = basemono + dj * TICK_NSEC;
1400 * Prepare a CPU for idle from an RCU perspective. The first major task
1401 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1402 * The second major task is to check to see if a non-lazy callback has
1403 * arrived at a CPU that previously had only lazy callbacks. The third
1404 * major task is to accelerate (that is, assign grace-period numbers to)
1405 * any recently arrived callbacks.
1407 * The caller must have disabled interrupts.
1409 static void rcu_prepare_for_idle(void)
1412 struct rcu_data *rdp;
1413 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1414 struct rcu_node *rnp;
1415 struct rcu_state *rsp;
1418 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL))
1421 /* Handle nohz enablement switches conservatively. */
1422 tne = READ_ONCE(tick_nohz_active);
1423 if (tne != rdtp->tick_nohz_enabled_snap) {
1424 if (rcu_cpu_has_callbacks(NULL))
1425 invoke_rcu_core(); /* force nohz to see update. */
1426 rdtp->tick_nohz_enabled_snap = tne;
1432 /* If this is a no-CBs CPU, no callbacks, just return. */
1433 if (rcu_is_nocb_cpu(smp_processor_id()))
1437 * If a non-lazy callback arrived at a CPU having only lazy
1438 * callbacks, invoke RCU core for the side-effect of recalculating
1439 * idle duration on re-entry to idle.
1441 if (rdtp->all_lazy &&
1442 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1443 rdtp->all_lazy = false;
1444 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1450 * If we have not yet accelerated this jiffy, accelerate all
1451 * callbacks on this CPU.
1453 if (rdtp->last_accelerate == jiffies)
1455 rdtp->last_accelerate = jiffies;
1456 for_each_rcu_flavor(rsp) {
1457 rdp = this_cpu_ptr(rsp->rda);
1458 if (!*rdp->nxttail[RCU_DONE_TAIL])
1461 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1462 smp_mb__after_unlock_lock();
1463 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1464 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1466 rcu_gp_kthread_wake(rsp);
1471 * Clean up for exit from idle. Attempt to advance callbacks based on
1472 * any grace periods that elapsed while the CPU was idle, and if any
1473 * callbacks are now ready to invoke, initiate invocation.
1475 static void rcu_cleanup_after_idle(void)
1477 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1478 rcu_is_nocb_cpu(smp_processor_id()))
1480 if (rcu_try_advance_all_cbs())
1485 * Keep a running count of the number of non-lazy callbacks posted
1486 * on this CPU. This running counter (which is never decremented) allows
1487 * rcu_prepare_for_idle() to detect when something out of the idle loop
1488 * posts a callback, even if an equal number of callbacks are invoked.
1489 * Of course, callbacks should only be posted from within a trace event
1490 * designed to be called from idle or from within RCU_NONIDLE().
1492 static void rcu_idle_count_callbacks_posted(void)
1494 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1498 * Data for flushing lazy RCU callbacks at OOM time.
1500 static atomic_t oom_callback_count;
1501 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1504 * RCU OOM callback -- decrement the outstanding count and deliver the
1505 * wake-up if we are the last one.
1507 static void rcu_oom_callback(struct rcu_head *rhp)
1509 if (atomic_dec_and_test(&oom_callback_count))
1510 wake_up(&oom_callback_wq);
1514 * Post an rcu_oom_notify callback on the current CPU if it has at
1515 * least one lazy callback. This will unnecessarily post callbacks
1516 * to CPUs that already have a non-lazy callback at the end of their
1517 * callback list, but this is an infrequent operation, so accept some
1518 * extra overhead to keep things simple.
1520 static void rcu_oom_notify_cpu(void *unused)
1522 struct rcu_state *rsp;
1523 struct rcu_data *rdp;
1525 for_each_rcu_flavor(rsp) {
1526 rdp = raw_cpu_ptr(rsp->rda);
1527 if (rdp->qlen_lazy != 0) {
1528 atomic_inc(&oom_callback_count);
1529 rsp->call(&rdp->oom_head, rcu_oom_callback);
1535 * If low on memory, ensure that each CPU has a non-lazy callback.
1536 * This will wake up CPUs that have only lazy callbacks, in turn
1537 * ensuring that they free up the corresponding memory in a timely manner.
1538 * Because an uncertain amount of memory will be freed in some uncertain
1539 * timeframe, we do not claim to have freed anything.
1541 static int rcu_oom_notify(struct notifier_block *self,
1542 unsigned long notused, void *nfreed)
1546 /* Wait for callbacks from earlier instance to complete. */
1547 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1548 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1551 * Prevent premature wakeup: ensure that all increments happen
1552 * before there is a chance of the counter reaching zero.
1554 atomic_set(&oom_callback_count, 1);
1556 for_each_online_cpu(cpu) {
1557 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1558 cond_resched_rcu_qs();
1561 /* Unconditionally decrement: no need to wake ourselves up. */
1562 atomic_dec(&oom_callback_count);
1567 static struct notifier_block rcu_oom_nb = {
1568 .notifier_call = rcu_oom_notify
1571 static int __init rcu_register_oom_notifier(void)
1573 register_oom_notifier(&rcu_oom_nb);
1576 early_initcall(rcu_register_oom_notifier);
1578 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1580 #ifdef CONFIG_RCU_FAST_NO_HZ
1582 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1584 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1585 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1587 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1588 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1590 rdtp->all_lazy ? 'L' : '.',
1591 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1594 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1596 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1601 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1603 /* Initiate the stall-info list. */
1604 static void print_cpu_stall_info_begin(void)
1610 * Print out diagnostic information for the specified stalled CPU.
1612 * If the specified CPU is aware of the current RCU grace period
1613 * (flavor specified by rsp), then print the number of scheduling
1614 * clock interrupts the CPU has taken during the time that it has
1615 * been aware. Otherwise, print the number of RCU grace periods
1616 * that this CPU is ignorant of, for example, "1" if the CPU was
1617 * aware of the previous grace period.
1619 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1621 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1623 char fast_no_hz[72];
1624 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1625 struct rcu_dynticks *rdtp = rdp->dynticks;
1627 unsigned long ticks_value;
1629 if (rsp->gpnum == rdp->gpnum) {
1630 ticks_title = "ticks this GP";
1631 ticks_value = rdp->ticks_this_gp;
1633 ticks_title = "GPs behind";
1634 ticks_value = rsp->gpnum - rdp->gpnum;
1636 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1637 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1638 cpu, ticks_value, ticks_title,
1639 atomic_read(&rdtp->dynticks) & 0xfff,
1640 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1641 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1642 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1646 /* Terminate the stall-info list. */
1647 static void print_cpu_stall_info_end(void)
1652 /* Zero ->ticks_this_gp for all flavors of RCU. */
1653 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1655 rdp->ticks_this_gp = 0;
1656 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1659 /* Increment ->ticks_this_gp for all flavors of RCU. */
1660 static void increment_cpu_stall_ticks(void)
1662 struct rcu_state *rsp;
1664 for_each_rcu_flavor(rsp)
1665 raw_cpu_inc(rsp->rda->ticks_this_gp);
1668 #ifdef CONFIG_RCU_NOCB_CPU
1671 * Offload callback processing from the boot-time-specified set of CPUs
1672 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1673 * kthread created that pulls the callbacks from the corresponding CPU,
1674 * waits for a grace period to elapse, and invokes the callbacks.
1675 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1676 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1677 * has been specified, in which case each kthread actively polls its
1678 * CPU. (Which isn't so great for energy efficiency, but which does
1679 * reduce RCU's overhead on that CPU.)
1681 * This is intended to be used in conjunction with Frederic Weisbecker's
1682 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1683 * running CPU-bound user-mode computations.
1685 * Offloading of callback processing could also in theory be used as
1686 * an energy-efficiency measure because CPUs with no RCU callbacks
1687 * queued are more aggressive about entering dyntick-idle mode.
1691 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1692 static int __init rcu_nocb_setup(char *str)
1694 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1695 have_rcu_nocb_mask = true;
1696 cpulist_parse(str, rcu_nocb_mask);
1699 __setup("rcu_nocbs=", rcu_nocb_setup);
1701 static int __init parse_rcu_nocb_poll(char *arg)
1706 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1709 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1712 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1714 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1718 * Set the root rcu_node structure's ->need_future_gp field
1719 * based on the sum of those of all rcu_node structures. This does
1720 * double-count the root rcu_node structure's requests, but this
1721 * is necessary to handle the possibility of a rcu_nocb_kthread()
1722 * having awakened during the time that the rcu_node structures
1723 * were being updated for the end of the previous grace period.
1725 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1727 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1730 static void rcu_init_one_nocb(struct rcu_node *rnp)
1732 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1733 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1736 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1737 /* Is the specified CPU a no-CBs CPU? */
1738 bool rcu_is_nocb_cpu(int cpu)
1740 if (have_rcu_nocb_mask)
1741 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1744 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1747 * Kick the leader kthread for this NOCB group.
1749 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1751 struct rcu_data *rdp_leader = rdp->nocb_leader;
1753 if (!READ_ONCE(rdp_leader->nocb_kthread))
1755 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1756 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1757 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1758 wake_up(&rdp_leader->nocb_wq);
1763 * Does the specified CPU need an RCU callback for the specified flavor
1766 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1768 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1770 #ifdef CONFIG_PROVE_RCU
1771 struct rcu_head *rhp;
1772 #endif /* #ifdef CONFIG_PROVE_RCU */
1775 * Check count of all no-CBs callbacks awaiting invocation.
1776 * There needs to be a barrier before this function is called,
1777 * but associated with a prior determination that no more
1778 * callbacks would be posted. In the worst case, the first
1779 * barrier in _rcu_barrier() suffices (but the caller cannot
1780 * necessarily rely on this, not a substitute for the caller
1781 * getting the concurrency design right!). There must also be
1782 * a barrier between the following load an posting of a callback
1783 * (if a callback is in fact needed). This is associated with an
1784 * atomic_inc() in the caller.
1786 ret = atomic_long_read(&rdp->nocb_q_count);
1788 #ifdef CONFIG_PROVE_RCU
1789 rhp = READ_ONCE(rdp->nocb_head);
1791 rhp = READ_ONCE(rdp->nocb_gp_head);
1793 rhp = READ_ONCE(rdp->nocb_follower_head);
1795 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1796 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1797 rcu_scheduler_fully_active) {
1798 /* RCU callback enqueued before CPU first came online??? */
1799 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1803 #endif /* #ifdef CONFIG_PROVE_RCU */
1809 * Enqueue the specified string of rcu_head structures onto the specified
1810 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1811 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1812 * counts are supplied by rhcount and rhcount_lazy.
1814 * If warranted, also wake up the kthread servicing this CPUs queues.
1816 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1817 struct rcu_head *rhp,
1818 struct rcu_head **rhtp,
1819 int rhcount, int rhcount_lazy,
1820 unsigned long flags)
1823 struct rcu_head **old_rhpp;
1824 struct task_struct *t;
1826 /* Enqueue the callback on the nocb list and update counts. */
1827 atomic_long_add(rhcount, &rdp->nocb_q_count);
1828 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1829 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1830 WRITE_ONCE(*old_rhpp, rhp);
1831 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1832 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1834 /* If we are not being polled and there is a kthread, awaken it ... */
1835 t = READ_ONCE(rdp->nocb_kthread);
1836 if (rcu_nocb_poll || !t) {
1837 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1838 TPS("WakeNotPoll"));
1841 len = atomic_long_read(&rdp->nocb_q_count);
1842 if (old_rhpp == &rdp->nocb_head) {
1843 if (!irqs_disabled_flags(flags)) {
1844 /* ... if queue was empty ... */
1845 wake_nocb_leader(rdp, false);
1846 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1849 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1850 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1851 TPS("WakeEmptyIsDeferred"));
1853 rdp->qlen_last_fqs_check = 0;
1854 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1855 /* ... or if many callbacks queued. */
1856 if (!irqs_disabled_flags(flags)) {
1857 wake_nocb_leader(rdp, true);
1858 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1861 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1862 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1863 TPS("WakeOvfIsDeferred"));
1865 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1867 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1873 * This is a helper for __call_rcu(), which invokes this when the normal
1874 * callback queue is inoperable. If this is not a no-CBs CPU, this
1875 * function returns failure back to __call_rcu(), which can complain
1878 * Otherwise, this function queues the callback where the corresponding
1879 * "rcuo" kthread can find it.
1881 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1882 bool lazy, unsigned long flags)
1885 if (!rcu_is_nocb_cpu(rdp->cpu))
1887 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1888 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1889 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
1890 (unsigned long)rhp->func,
1891 -atomic_long_read(&rdp->nocb_q_count_lazy),
1892 -atomic_long_read(&rdp->nocb_q_count));
1894 trace_rcu_callback(rdp->rsp->name, rhp,
1895 -atomic_long_read(&rdp->nocb_q_count_lazy),
1896 -atomic_long_read(&rdp->nocb_q_count));
1899 * If called from an extended quiescent state with interrupts
1900 * disabled, invoke the RCU core in order to allow the idle-entry
1901 * deferred-wakeup check to function.
1903 if (irqs_disabled_flags(flags) &&
1904 !rcu_is_watching() &&
1905 cpu_online(smp_processor_id()))
1912 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
1915 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
1916 struct rcu_data *rdp,
1917 unsigned long flags)
1919 long ql = rsp->qlen;
1920 long qll = rsp->qlen_lazy;
1922 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
1923 if (!rcu_is_nocb_cpu(smp_processor_id()))
1928 /* First, enqueue the donelist, if any. This preserves CB ordering. */
1929 if (rsp->orphan_donelist != NULL) {
1930 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
1931 rsp->orphan_donetail, ql, qll, flags);
1933 rsp->orphan_donelist = NULL;
1934 rsp->orphan_donetail = &rsp->orphan_donelist;
1936 if (rsp->orphan_nxtlist != NULL) {
1937 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
1938 rsp->orphan_nxttail, ql, qll, flags);
1940 rsp->orphan_nxtlist = NULL;
1941 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1947 * If necessary, kick off a new grace period, and either way wait
1948 * for a subsequent grace period to complete.
1950 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
1954 unsigned long flags;
1956 struct rcu_node *rnp = rdp->mynode;
1958 raw_spin_lock_irqsave(&rnp->lock, flags);
1959 smp_mb__after_unlock_lock();
1960 needwake = rcu_start_future_gp(rnp, rdp, &c);
1961 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1963 rcu_gp_kthread_wake(rdp->rsp);
1966 * Wait for the grace period. Do so interruptibly to avoid messing
1967 * up the load average.
1969 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
1971 wait_event_interruptible(
1972 rnp->nocb_gp_wq[c & 0x1],
1973 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
1976 WARN_ON(signal_pending(current));
1977 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
1979 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
1980 smp_mb(); /* Ensure that CB invocation happens after GP end. */
1984 * Leaders come here to wait for additional callbacks to show up.
1985 * This function does not return until callbacks appear.
1987 static void nocb_leader_wait(struct rcu_data *my_rdp)
1989 bool firsttime = true;
1991 struct rcu_data *rdp;
1992 struct rcu_head **tail;
1996 /* Wait for callbacks to appear. */
1997 if (!rcu_nocb_poll) {
1998 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
1999 wait_event_interruptible(my_rdp->nocb_wq,
2000 !READ_ONCE(my_rdp->nocb_leader_sleep));
2001 /* Memory barrier handled by smp_mb() calls below and repoll. */
2002 } else if (firsttime) {
2003 firsttime = false; /* Don't drown trace log with "Poll"! */
2004 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2008 * Each pass through the following loop checks a follower for CBs.
2009 * We are our own first follower. Any CBs found are moved to
2010 * nocb_gp_head, where they await a grace period.
2013 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2014 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2015 if (!rdp->nocb_gp_head)
2016 continue; /* No CBs here, try next follower. */
2018 /* Move callbacks to wait-for-GP list, which is empty. */
2019 WRITE_ONCE(rdp->nocb_head, NULL);
2020 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2025 * If there were no callbacks, sleep a bit, rescan after a
2026 * memory barrier, and go retry.
2028 if (unlikely(!gotcbs)) {
2030 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2032 WARN_ON(signal_pending(current));
2033 schedule_timeout_interruptible(1);
2035 /* Rescan in case we were a victim of memory ordering. */
2036 my_rdp->nocb_leader_sleep = true;
2037 smp_mb(); /* Ensure _sleep true before scan. */
2038 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2039 if (READ_ONCE(rdp->nocb_head)) {
2040 /* Found CB, so short-circuit next wait. */
2041 my_rdp->nocb_leader_sleep = false;
2047 /* Wait for one grace period. */
2048 rcu_nocb_wait_gp(my_rdp);
2051 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2052 * We set it now, but recheck for new callbacks while
2053 * traversing our follower list.
2055 my_rdp->nocb_leader_sleep = true;
2056 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2058 /* Each pass through the following loop wakes a follower, if needed. */
2059 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2060 if (READ_ONCE(rdp->nocb_head))
2061 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2062 if (!rdp->nocb_gp_head)
2063 continue; /* No CBs, so no need to wake follower. */
2065 /* Append callbacks to follower's "done" list. */
2066 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2067 *tail = rdp->nocb_gp_head;
2068 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2069 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2071 * List was empty, wake up the follower.
2072 * Memory barriers supplied by atomic_long_add().
2074 wake_up(&rdp->nocb_wq);
2078 /* If we (the leader) don't have CBs, go wait some more. */
2079 if (!my_rdp->nocb_follower_head)
2084 * Followers come here to wait for additional callbacks to show up.
2085 * This function does not return until callbacks appear.
2087 static void nocb_follower_wait(struct rcu_data *rdp)
2089 bool firsttime = true;
2092 if (!rcu_nocb_poll) {
2093 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2095 wait_event_interruptible(rdp->nocb_wq,
2096 READ_ONCE(rdp->nocb_follower_head));
2097 } else if (firsttime) {
2098 /* Don't drown trace log with "Poll"! */
2100 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2102 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2103 /* ^^^ Ensure CB invocation follows _head test. */
2107 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2109 WARN_ON(signal_pending(current));
2110 schedule_timeout_interruptible(1);
2115 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2116 * callbacks queued by the corresponding no-CBs CPU, however, there is
2117 * an optional leader-follower relationship so that the grace-period
2118 * kthreads don't have to do quite so many wakeups.
2120 static int rcu_nocb_kthread(void *arg)
2123 struct rcu_head *list;
2124 struct rcu_head *next;
2125 struct rcu_head **tail;
2126 struct rcu_data *rdp = arg;
2128 /* Each pass through this loop invokes one batch of callbacks */
2130 /* Wait for callbacks. */
2131 if (rdp->nocb_leader == rdp)
2132 nocb_leader_wait(rdp);
2134 nocb_follower_wait(rdp);
2136 /* Pull the ready-to-invoke callbacks onto local list. */
2137 list = READ_ONCE(rdp->nocb_follower_head);
2139 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2140 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2141 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2143 /* Each pass through the following loop invokes a callback. */
2144 trace_rcu_batch_start(rdp->rsp->name,
2145 atomic_long_read(&rdp->nocb_q_count_lazy),
2146 atomic_long_read(&rdp->nocb_q_count), -1);
2150 /* Wait for enqueuing to complete, if needed. */
2151 while (next == NULL && &list->next != tail) {
2152 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2154 schedule_timeout_interruptible(1);
2155 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2159 debug_rcu_head_unqueue(list);
2161 if (__rcu_reclaim(rdp->rsp->name, list))
2167 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2168 smp_mb__before_atomic(); /* _add after CB invocation. */
2169 atomic_long_add(-c, &rdp->nocb_q_count);
2170 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2171 rdp->n_nocbs_invoked += c;
2176 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2177 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2179 return READ_ONCE(rdp->nocb_defer_wakeup);
2182 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2183 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2187 if (!rcu_nocb_need_deferred_wakeup(rdp))
2189 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2190 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2191 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2192 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2195 void __init rcu_init_nohz(void)
2198 bool need_rcu_nocb_mask = true;
2199 struct rcu_state *rsp;
2201 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2202 need_rcu_nocb_mask = false;
2203 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2205 #if defined(CONFIG_NO_HZ_FULL)
2206 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2207 need_rcu_nocb_mask = true;
2208 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2210 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2211 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2212 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2215 have_rcu_nocb_mask = true;
2217 if (!have_rcu_nocb_mask)
2220 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2221 pr_info("\tOffload RCU callbacks from CPU 0\n");
2222 cpumask_set_cpu(0, rcu_nocb_mask);
2223 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2224 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2225 pr_info("\tOffload RCU callbacks from all CPUs\n");
2226 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2227 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2228 #if defined(CONFIG_NO_HZ_FULL)
2229 if (tick_nohz_full_running)
2230 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2231 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2233 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2234 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2235 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2238 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2239 cpumask_pr_args(rcu_nocb_mask));
2241 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2243 for_each_rcu_flavor(rsp) {
2244 for_each_cpu(cpu, rcu_nocb_mask)
2245 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2246 rcu_organize_nocb_kthreads(rsp);
2250 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2251 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2253 rdp->nocb_tail = &rdp->nocb_head;
2254 init_waitqueue_head(&rdp->nocb_wq);
2255 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2259 * If the specified CPU is a no-CBs CPU that does not already have its
2260 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2261 * brought online out of order, this can require re-organizing the
2262 * leader-follower relationships.
2264 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2266 struct rcu_data *rdp;
2267 struct rcu_data *rdp_last;
2268 struct rcu_data *rdp_old_leader;
2269 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2270 struct task_struct *t;
2273 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2274 * then nothing to do.
2276 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2279 /* If we didn't spawn the leader first, reorganize! */
2280 rdp_old_leader = rdp_spawn->nocb_leader;
2281 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2283 rdp = rdp_old_leader;
2285 rdp->nocb_leader = rdp_spawn;
2286 if (rdp_last && rdp != rdp_spawn)
2287 rdp_last->nocb_next_follower = rdp;
2288 if (rdp == rdp_spawn) {
2289 rdp = rdp->nocb_next_follower;
2292 rdp = rdp->nocb_next_follower;
2293 rdp_last->nocb_next_follower = NULL;
2296 rdp_spawn->nocb_next_follower = rdp_old_leader;
2299 /* Spawn the kthread for this CPU and RCU flavor. */
2300 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2301 "rcuo%c/%d", rsp->abbr, cpu);
2303 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2307 * If the specified CPU is a no-CBs CPU that does not already have its
2308 * rcuo kthreads, spawn them.
2310 static void rcu_spawn_all_nocb_kthreads(int cpu)
2312 struct rcu_state *rsp;
2314 if (rcu_scheduler_fully_active)
2315 for_each_rcu_flavor(rsp)
2316 rcu_spawn_one_nocb_kthread(rsp, cpu);
2320 * Once the scheduler is running, spawn rcuo kthreads for all online
2321 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2322 * non-boot CPUs come online -- if this changes, we will need to add
2323 * some mutual exclusion.
2325 static void __init rcu_spawn_nocb_kthreads(void)
2329 for_each_online_cpu(cpu)
2330 rcu_spawn_all_nocb_kthreads(cpu);
2333 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2334 static int rcu_nocb_leader_stride = -1;
2335 module_param(rcu_nocb_leader_stride, int, 0444);
2338 * Initialize leader-follower relationships for all no-CBs CPU.
2340 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2343 int ls = rcu_nocb_leader_stride;
2344 int nl = 0; /* Next leader. */
2345 struct rcu_data *rdp;
2346 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2347 struct rcu_data *rdp_prev = NULL;
2349 if (!have_rcu_nocb_mask)
2352 ls = int_sqrt(nr_cpu_ids);
2353 rcu_nocb_leader_stride = ls;
2357 * Each pass through this loop sets up one rcu_data structure and
2358 * spawns one rcu_nocb_kthread().
2360 for_each_cpu(cpu, rcu_nocb_mask) {
2361 rdp = per_cpu_ptr(rsp->rda, cpu);
2362 if (rdp->cpu >= nl) {
2363 /* New leader, set up for followers & next leader. */
2364 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2365 rdp->nocb_leader = rdp;
2368 /* Another follower, link to previous leader. */
2369 rdp->nocb_leader = rdp_leader;
2370 rdp_prev->nocb_next_follower = rdp;
2376 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2377 static bool init_nocb_callback_list(struct rcu_data *rdp)
2379 if (!rcu_is_nocb_cpu(rdp->cpu))
2382 /* If there are early-boot callbacks, move them to nocb lists. */
2384 rdp->nocb_head = rdp->nxtlist;
2385 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2386 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2387 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2388 rdp->nxtlist = NULL;
2392 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2396 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2398 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2400 WARN_ON_ONCE(1); /* Should be dead code. */
2404 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2408 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2412 static void rcu_init_one_nocb(struct rcu_node *rnp)
2416 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2417 bool lazy, unsigned long flags)
2422 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2423 struct rcu_data *rdp,
2424 unsigned long flags)
2429 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2433 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2438 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2442 static void rcu_spawn_all_nocb_kthreads(int cpu)
2446 static void __init rcu_spawn_nocb_kthreads(void)
2450 static bool init_nocb_callback_list(struct rcu_data *rdp)
2455 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2458 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2459 * arbitrarily long period of time with the scheduling-clock tick turned
2460 * off. RCU will be paying attention to this CPU because it is in the
2461 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2462 * machine because the scheduling-clock tick has been disabled. Therefore,
2463 * if an adaptive-ticks CPU is failing to respond to the current grace
2464 * period and has not be idle from an RCU perspective, kick it.
2466 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2468 #ifdef CONFIG_NO_HZ_FULL
2469 if (tick_nohz_full_cpu(cpu))
2470 smp_send_reschedule(cpu);
2471 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2475 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2477 static int full_sysidle_state; /* Current system-idle state. */
2478 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2479 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2480 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2481 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2482 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2485 * Invoked to note exit from irq or task transition to idle. Note that
2486 * usermode execution does -not- count as idle here! After all, we want
2487 * to detect full-system idle states, not RCU quiescent states and grace
2488 * periods. The caller must have disabled interrupts.
2490 static void rcu_sysidle_enter(int irq)
2493 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2495 /* If there are no nohz_full= CPUs, no need to track this. */
2496 if (!tick_nohz_full_enabled())
2499 /* Adjust nesting, check for fully idle. */
2501 rdtp->dynticks_idle_nesting--;
2502 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2503 if (rdtp->dynticks_idle_nesting != 0)
2504 return; /* Still not fully idle. */
2506 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2507 DYNTICK_TASK_NEST_VALUE) {
2508 rdtp->dynticks_idle_nesting = 0;
2510 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2511 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2512 return; /* Still not fully idle. */
2516 /* Record start of fully idle period. */
2518 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2519 smp_mb__before_atomic();
2520 atomic_inc(&rdtp->dynticks_idle);
2521 smp_mb__after_atomic();
2522 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2526 * Unconditionally force exit from full system-idle state. This is
2527 * invoked when a normal CPU exits idle, but must be called separately
2528 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2529 * is that the timekeeping CPU is permitted to take scheduling-clock
2530 * interrupts while the system is in system-idle state, and of course
2531 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2532 * interrupt from any other type of interrupt.
2534 void rcu_sysidle_force_exit(void)
2536 int oldstate = READ_ONCE(full_sysidle_state);
2540 * Each pass through the following loop attempts to exit full
2541 * system-idle state. If contention proves to be a problem,
2542 * a trylock-based contention tree could be used here.
2544 while (oldstate > RCU_SYSIDLE_SHORT) {
2545 newoldstate = cmpxchg(&full_sysidle_state,
2546 oldstate, RCU_SYSIDLE_NOT);
2547 if (oldstate == newoldstate &&
2548 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2549 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2550 return; /* We cleared it, done! */
2552 oldstate = newoldstate;
2554 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2558 * Invoked to note entry to irq or task transition from idle. Note that
2559 * usermode execution does -not- count as idle here! The caller must
2560 * have disabled interrupts.
2562 static void rcu_sysidle_exit(int irq)
2564 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2566 /* If there are no nohz_full= CPUs, no need to track this. */
2567 if (!tick_nohz_full_enabled())
2570 /* Adjust nesting, check for already non-idle. */
2572 rdtp->dynticks_idle_nesting++;
2573 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2574 if (rdtp->dynticks_idle_nesting != 1)
2575 return; /* Already non-idle. */
2578 * Allow for irq misnesting. Yes, it really is possible
2579 * to enter an irq handler then never leave it, and maybe
2580 * also vice versa. Handle both possibilities.
2582 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2583 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2584 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2585 return; /* Already non-idle. */
2587 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2591 /* Record end of idle period. */
2592 smp_mb__before_atomic();
2593 atomic_inc(&rdtp->dynticks_idle);
2594 smp_mb__after_atomic();
2595 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2598 * If we are the timekeeping CPU, we are permitted to be non-idle
2599 * during a system-idle state. This must be the case, because
2600 * the timekeeping CPU has to take scheduling-clock interrupts
2601 * during the time that the system is transitioning to full
2602 * system-idle state. This means that the timekeeping CPU must
2603 * invoke rcu_sysidle_force_exit() directly if it does anything
2604 * more than take a scheduling-clock interrupt.
2606 if (smp_processor_id() == tick_do_timer_cpu)
2609 /* Update system-idle state: We are clearly no longer fully idle! */
2610 rcu_sysidle_force_exit();
2614 * Check to see if the current CPU is idle. Note that usermode execution
2615 * does not count as idle. The caller must have disabled interrupts,
2616 * and must be running on tick_do_timer_cpu.
2618 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2619 unsigned long *maxj)
2623 struct rcu_dynticks *rdtp = rdp->dynticks;
2625 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2626 if (!tick_nohz_full_enabled())
2630 * If some other CPU has already reported non-idle, if this is
2631 * not the flavor of RCU that tracks sysidle state, or if this
2632 * is an offline or the timekeeping CPU, nothing to do.
2634 if (!*isidle || rdp->rsp != rcu_state_p ||
2635 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2637 /* Verify affinity of current kthread. */
2638 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2640 /* Pick up current idle and NMI-nesting counter and check. */
2641 cur = atomic_read(&rdtp->dynticks_idle);
2643 *isidle = false; /* We are not idle! */
2646 smp_mb(); /* Read counters before timestamps. */
2648 /* Pick up timestamps. */
2649 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2650 /* If this CPU entered idle more recently, update maxj timestamp. */
2651 if (ULONG_CMP_LT(*maxj, j))
2656 * Is this the flavor of RCU that is handling full-system idle?
2658 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2660 return rsp == rcu_state_p;
2664 * Return a delay in jiffies based on the number of CPUs, rcu_node
2665 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2666 * systems more time to transition to full-idle state in order to
2667 * avoid the cache thrashing that otherwise occur on the state variable.
2668 * Really small systems (less than a couple of tens of CPUs) should
2669 * instead use a single global atomically incremented counter, and later
2670 * versions of this will automatically reconfigure themselves accordingly.
2672 static unsigned long rcu_sysidle_delay(void)
2674 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2676 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2680 * Advance the full-system-idle state. This is invoked when all of
2681 * the non-timekeeping CPUs are idle.
2683 static void rcu_sysidle(unsigned long j)
2685 /* Check the current state. */
2686 switch (READ_ONCE(full_sysidle_state)) {
2687 case RCU_SYSIDLE_NOT:
2689 /* First time all are idle, so note a short idle period. */
2690 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2693 case RCU_SYSIDLE_SHORT:
2696 * Idle for a bit, time to advance to next state?
2697 * cmpxchg failure means race with non-idle, let them win.
2699 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2700 (void)cmpxchg(&full_sysidle_state,
2701 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2704 case RCU_SYSIDLE_LONG:
2707 * Do an additional check pass before advancing to full.
2708 * cmpxchg failure means race with non-idle, let them win.
2710 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2711 (void)cmpxchg(&full_sysidle_state,
2712 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2721 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2722 * back to the beginning.
2724 static void rcu_sysidle_cancel(void)
2727 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2728 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2732 * Update the sysidle state based on the results of a force-quiescent-state
2733 * scan of the CPUs' dyntick-idle state.
2735 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2736 unsigned long maxj, bool gpkt)
2738 if (rsp != rcu_state_p)
2739 return; /* Wrong flavor, ignore. */
2740 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2741 return; /* Running state machine from timekeeping CPU. */
2743 rcu_sysidle(maxj); /* More idle! */
2745 rcu_sysidle_cancel(); /* Idle is over. */
2749 * Wrapper for rcu_sysidle_report() when called from the grace-period
2750 * kthread's context.
2752 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2755 /* If there are no nohz_full= CPUs, no need to track this. */
2756 if (!tick_nohz_full_enabled())
2759 rcu_sysidle_report(rsp, isidle, maxj, true);
2762 /* Callback and function for forcing an RCU grace period. */
2763 struct rcu_sysidle_head {
2768 static void rcu_sysidle_cb(struct rcu_head *rhp)
2770 struct rcu_sysidle_head *rshp;
2773 * The following memory barrier is needed to replace the
2774 * memory barriers that would normally be in the memory
2777 smp_mb(); /* grace period precedes setting inuse. */
2779 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2780 WRITE_ONCE(rshp->inuse, 0);
2784 * Check to see if the system is fully idle, other than the timekeeping CPU.
2785 * The caller must have disabled interrupts. This is not intended to be
2786 * called unless tick_nohz_full_enabled().
2788 bool rcu_sys_is_idle(void)
2790 static struct rcu_sysidle_head rsh;
2791 int rss = READ_ONCE(full_sysidle_state);
2793 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2796 /* Handle small-system case by doing a full scan of CPUs. */
2797 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2798 int oldrss = rss - 1;
2801 * One pass to advance to each state up to _FULL.
2802 * Give up if any pass fails to advance the state.
2804 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2807 unsigned long maxj = jiffies - ULONG_MAX / 4;
2808 struct rcu_data *rdp;
2810 /* Scan all the CPUs looking for nonidle CPUs. */
2811 for_each_possible_cpu(cpu) {
2812 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2813 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2817 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2819 rss = READ_ONCE(full_sysidle_state);
2823 /* If this is the first observation of an idle period, record it. */
2824 if (rss == RCU_SYSIDLE_FULL) {
2825 rss = cmpxchg(&full_sysidle_state,
2826 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2827 return rss == RCU_SYSIDLE_FULL;
2830 smp_mb(); /* ensure rss load happens before later caller actions. */
2832 /* If already fully idle, tell the caller (in case of races). */
2833 if (rss == RCU_SYSIDLE_FULL_NOTED)
2837 * If we aren't there yet, and a grace period is not in flight,
2838 * initiate a grace period. Either way, tell the caller that
2839 * we are not there yet. We use an xchg() rather than an assignment
2840 * to make up for the memory barriers that would otherwise be
2841 * provided by the memory allocator.
2843 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2844 !rcu_gp_in_progress(rcu_state_p) &&
2845 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2846 call_rcu(&rsh.rh, rcu_sysidle_cb);
2851 * Initialize dynticks sysidle state for CPUs coming online.
2853 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2855 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2858 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2860 static void rcu_sysidle_enter(int irq)
2864 static void rcu_sysidle_exit(int irq)
2868 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2869 unsigned long *maxj)
2873 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2878 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2883 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2887 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2890 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2891 * grace-period kthread will do force_quiescent_state() processing?
2892 * The idea is to avoid waking up RCU core processing on such a
2893 * CPU unless the grace period has extended for too long.
2895 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2896 * CONFIG_RCU_NOCB_CPU CPUs.
2898 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2900 #ifdef CONFIG_NO_HZ_FULL
2901 if (tick_nohz_full_cpu(smp_processor_id()) &&
2902 (!rcu_gp_in_progress(rsp) ||
2903 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
2905 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2910 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2913 static void rcu_bind_gp_kthread(void)
2915 int __maybe_unused cpu;
2917 if (!tick_nohz_full_enabled())
2919 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2920 cpu = tick_do_timer_cpu;
2921 if (cpu >= 0 && cpu < nr_cpu_ids)
2922 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2923 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2924 housekeeping_affine(current);
2925 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2928 /* Record the current task on dyntick-idle entry. */
2929 static void rcu_dynticks_task_enter(void)
2931 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2932 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2933 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2936 /* Record no current task on dyntick-idle exit. */
2937 static void rcu_dynticks_task_exit(void)
2939 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2940 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2941 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */