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.
68 static void __init rcu_bootup_announce_oddness(void)
70 if (IS_ENABLED(CONFIG_RCU_TRACE))
71 pr_info("\tRCU debugfs-based tracing is enabled.\n");
72 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
73 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
74 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
77 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
78 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
79 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
80 if (IS_ENABLED(CONFIG_PROVE_RCU))
81 pr_info("\tRCU lockdep checking is enabled.\n");
82 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE))
83 pr_info("\tRCU torture testing starts during boot.\n");
84 if (RCU_NUM_LVLS >= 4)
85 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
86 if (RCU_FANOUT_LEAF != 16)
87 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
89 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
90 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
91 if (nr_cpu_ids != NR_CPUS)
92 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
93 if (IS_ENABLED(CONFIG_RCU_BOOST))
94 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
97 #ifdef CONFIG_PREEMPT_RCU
99 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
100 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
101 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
103 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
107 * Tell them what RCU they are running.
109 static void __init rcu_bootup_announce(void)
111 pr_info("Preemptible hierarchical RCU implementation.\n");
112 rcu_bootup_announce_oddness();
115 /* Flags for rcu_preempt_ctxt_queue() decision table. */
116 #define RCU_GP_TASKS 0x8
117 #define RCU_EXP_TASKS 0x4
118 #define RCU_GP_BLKD 0x2
119 #define RCU_EXP_BLKD 0x1
122 * Queues a task preempted within an RCU-preempt read-side critical
123 * section into the appropriate location within the ->blkd_tasks list,
124 * depending on the states of any ongoing normal and expedited grace
125 * periods. The ->gp_tasks pointer indicates which element the normal
126 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
127 * indicates which element the expedited grace period is waiting on (again,
128 * NULL if none). If a grace period is waiting on a given element in the
129 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
130 * adding a task to the tail of the list blocks any grace period that is
131 * already waiting on one of the elements. In contrast, adding a task
132 * to the head of the list won't block any grace period that is already
133 * waiting on one of the elements.
135 * This queuing is imprecise, and can sometimes make an ongoing grace
136 * period wait for a task that is not strictly speaking blocking it.
137 * Given the choice, we needlessly block a normal grace period rather than
138 * blocking an expedited grace period.
140 * Note that an endless sequence of expedited grace periods still cannot
141 * indefinitely postpone a normal grace period. Eventually, all of the
142 * fixed number of preempted tasks blocking the normal grace period that are
143 * not also blocking the expedited grace period will resume and complete
144 * their RCU read-side critical sections. At that point, the ->gp_tasks
145 * pointer will equal the ->exp_tasks pointer, at which point the end of
146 * the corresponding expedited grace period will also be the end of the
147 * normal grace period.
149 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
150 __releases(rnp->lock) /* But leaves rrupts disabled. */
152 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
153 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
154 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
155 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
156 struct task_struct *t = current;
159 * Decide where to queue the newly blocked task. In theory,
160 * this could be an if-statement. In practice, when I tried
161 * that, it was quite messy.
163 switch (blkd_state) {
166 case RCU_EXP_TASKS + RCU_GP_BLKD:
168 case RCU_GP_TASKS + RCU_EXP_TASKS:
171 * Blocking neither GP, or first task blocking the normal
172 * GP but not blocking the already-waiting expedited GP.
173 * Queue at the head of the list to avoid unnecessarily
174 * blocking the already-waiting GPs.
176 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
181 case RCU_GP_BLKD + RCU_EXP_BLKD:
182 case RCU_GP_TASKS + RCU_EXP_BLKD:
183 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
184 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
187 * First task arriving that blocks either GP, or first task
188 * arriving that blocks the expedited GP (with the normal
189 * GP already waiting), or a task arriving that blocks
190 * both GPs with both GPs already waiting. Queue at the
191 * tail of the list to avoid any GP waiting on any of the
192 * already queued tasks that are not blocking it.
194 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
197 case RCU_EXP_TASKS + RCU_EXP_BLKD:
198 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
199 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
202 * Second or subsequent task blocking the expedited GP.
203 * The task either does not block the normal GP, or is the
204 * first task blocking the normal GP. Queue just after
205 * the first task blocking the expedited GP.
207 list_add(&t->rcu_node_entry, rnp->exp_tasks);
210 case RCU_GP_TASKS + RCU_GP_BLKD:
211 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
214 * Second or subsequent task blocking the normal GP.
215 * The task does not block the expedited GP. Queue just
216 * after the first task blocking the normal GP.
218 list_add(&t->rcu_node_entry, rnp->gp_tasks);
223 /* Yet another exercise in excessive paranoia. */
229 * We have now queued the task. If it was the first one to
230 * block either grace period, update the ->gp_tasks and/or
231 * ->exp_tasks pointers, respectively, to reference the newly
234 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD))
235 rnp->gp_tasks = &t->rcu_node_entry;
236 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
237 rnp->exp_tasks = &t->rcu_node_entry;
238 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
241 * Report the quiescent state for the expedited GP. This expedited
242 * GP should not be able to end until we report, so there should be
243 * no need to check for a subsequent expedited GP. (Though we are
244 * still in a quiescent state in any case.)
246 if (blkd_state & RCU_EXP_BLKD &&
247 t->rcu_read_unlock_special.b.exp_need_qs) {
248 t->rcu_read_unlock_special.b.exp_need_qs = false;
249 rcu_report_exp_rdp(rdp->rsp, rdp, true);
251 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
256 * Record a preemptible-RCU quiescent state for the specified CPU. Note
257 * that this just means that the task currently running on the CPU is
258 * not in a quiescent state. There might be any number of tasks blocked
259 * while in an RCU read-side critical section.
261 * As with the other rcu_*_qs() functions, callers to this function
262 * must disable preemption.
264 static void rcu_preempt_qs(void)
266 if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
267 trace_rcu_grace_period(TPS("rcu_preempt"),
268 __this_cpu_read(rcu_data_p->gpnum),
270 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
271 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
272 current->rcu_read_unlock_special.b.need_qs = false;
277 * We have entered the scheduler, and the current task might soon be
278 * context-switched away from. If this task is in an RCU read-side
279 * critical section, we will no longer be able to rely on the CPU to
280 * record that fact, so we enqueue the task on the blkd_tasks list.
281 * The task will dequeue itself when it exits the outermost enclosing
282 * RCU read-side critical section. Therefore, the current grace period
283 * cannot be permitted to complete until the blkd_tasks list entries
284 * predating the current grace period drain, in other words, until
285 * rnp->gp_tasks becomes NULL.
287 * Caller must disable interrupts.
289 static void rcu_preempt_note_context_switch(void)
291 struct task_struct *t = current;
292 struct rcu_data *rdp;
293 struct rcu_node *rnp;
295 if (t->rcu_read_lock_nesting > 0 &&
296 !t->rcu_read_unlock_special.b.blocked) {
298 /* Possibly blocking in an RCU read-side critical section. */
299 rdp = this_cpu_ptr(rcu_state_p->rda);
301 raw_spin_lock_rcu_node(rnp);
302 t->rcu_read_unlock_special.b.blocked = true;
303 t->rcu_blocked_node = rnp;
306 * Verify the CPU's sanity, trace the preemption, and
307 * then queue the task as required based on the states
308 * of any ongoing and expedited grace periods.
310 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
311 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
312 trace_rcu_preempt_task(rdp->rsp->name,
314 (rnp->qsmask & rdp->grpmask)
317 rcu_preempt_ctxt_queue(rnp, rdp);
318 } else if (t->rcu_read_lock_nesting < 0 &&
319 t->rcu_read_unlock_special.s) {
322 * Complete exit from RCU read-side critical section on
323 * behalf of preempted instance of __rcu_read_unlock().
325 rcu_read_unlock_special(t);
329 * Either we were not in an RCU read-side critical section to
330 * begin with, or we have now recorded that critical section
331 * globally. Either way, we can now note a quiescent state
332 * for this CPU. Again, if we were in an RCU read-side critical
333 * section, and if that critical section was blocking the current
334 * grace period, then the fact that the task has been enqueued
335 * means that we continue to block the current grace period.
341 * Check for preempted RCU readers blocking the current grace period
342 * for the specified rcu_node structure. If the caller needs a reliable
343 * answer, it must hold the rcu_node's ->lock.
345 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
347 return rnp->gp_tasks != NULL;
351 * Advance a ->blkd_tasks-list pointer to the next entry, instead
352 * returning NULL if at the end of the list.
354 static struct list_head *rcu_next_node_entry(struct task_struct *t,
355 struct rcu_node *rnp)
357 struct list_head *np;
359 np = t->rcu_node_entry.next;
360 if (np == &rnp->blkd_tasks)
366 * Return true if the specified rcu_node structure has tasks that were
367 * preempted within an RCU read-side critical section.
369 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
371 return !list_empty(&rnp->blkd_tasks);
375 * Handle special cases during rcu_read_unlock(), such as needing to
376 * notify RCU core processing or task having blocked during the RCU
377 * read-side critical section.
379 void rcu_read_unlock_special(struct task_struct *t)
385 struct list_head *np;
386 bool drop_boost_mutex = false;
387 struct rcu_data *rdp;
388 struct rcu_node *rnp;
389 union rcu_special special;
391 /* NMI handlers cannot block and cannot safely manipulate state. */
395 local_irq_save(flags);
398 * If RCU core is waiting for this CPU to exit its critical section,
399 * report the fact that it has exited. Because irqs are disabled,
400 * t->rcu_read_unlock_special cannot change.
402 special = t->rcu_read_unlock_special;
403 if (special.b.need_qs) {
405 t->rcu_read_unlock_special.b.need_qs = false;
406 if (!t->rcu_read_unlock_special.s) {
407 local_irq_restore(flags);
413 * Respond to a request for an expedited grace period, but only if
414 * we were not preempted, meaning that we were running on the same
415 * CPU throughout. If we were preempted, the exp_need_qs flag
416 * would have been cleared at the time of the first preemption,
417 * and the quiescent state would be reported when we were dequeued.
419 if (special.b.exp_need_qs) {
420 WARN_ON_ONCE(special.b.blocked);
421 t->rcu_read_unlock_special.b.exp_need_qs = false;
422 rdp = this_cpu_ptr(rcu_state_p->rda);
423 rcu_report_exp_rdp(rcu_state_p, rdp, true);
424 if (!t->rcu_read_unlock_special.s) {
425 local_irq_restore(flags);
430 /* Hardware IRQ handlers cannot block, complain if they get here. */
431 if (in_irq() || in_serving_softirq()) {
432 lockdep_rcu_suspicious(__FILE__, __LINE__,
433 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
434 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
435 t->rcu_read_unlock_special.s,
436 t->rcu_read_unlock_special.b.blocked,
437 t->rcu_read_unlock_special.b.exp_need_qs,
438 t->rcu_read_unlock_special.b.need_qs);
439 local_irq_restore(flags);
443 /* Clean up if blocked during RCU read-side critical section. */
444 if (special.b.blocked) {
445 t->rcu_read_unlock_special.b.blocked = false;
448 * Remove this task from the list it blocked on. The task
449 * now remains queued on the rcu_node corresponding to the
450 * CPU it first blocked on, so there is no longer any need
451 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
453 rnp = t->rcu_blocked_node;
454 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
455 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
456 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
457 empty_exp = sync_rcu_preempt_exp_done(rnp);
458 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
459 np = rcu_next_node_entry(t, rnp);
460 list_del_init(&t->rcu_node_entry);
461 t->rcu_blocked_node = NULL;
462 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
464 if (&t->rcu_node_entry == rnp->gp_tasks)
466 if (&t->rcu_node_entry == rnp->exp_tasks)
468 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
469 if (&t->rcu_node_entry == rnp->boost_tasks)
470 rnp->boost_tasks = np;
471 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
472 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
476 * If this was the last task on the current list, and if
477 * we aren't waiting on any CPUs, report the quiescent state.
478 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
479 * so we must take a snapshot of the expedited state.
481 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
482 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
483 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
490 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
492 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
495 /* Unboost if we were boosted. */
496 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
497 rt_mutex_unlock(&rnp->boost_mtx);
500 * If this was the last task on the expedited lists,
501 * then we need to report up the rcu_node hierarchy.
503 if (!empty_exp && empty_exp_now)
504 rcu_report_exp_rnp(rcu_state_p, rnp, true);
506 local_irq_restore(flags);
511 * Dump detailed information for all tasks blocking the current RCU
512 * grace period on the specified rcu_node structure.
514 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
517 struct task_struct *t;
519 raw_spin_lock_irqsave_rcu_node(rnp, flags);
520 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
521 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
524 t = list_entry(rnp->gp_tasks->prev,
525 struct task_struct, rcu_node_entry);
526 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
528 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
532 * Dump detailed information for all tasks blocking the current RCU
535 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
537 struct rcu_node *rnp = rcu_get_root(rsp);
539 rcu_print_detail_task_stall_rnp(rnp);
540 rcu_for_each_leaf_node(rsp, rnp)
541 rcu_print_detail_task_stall_rnp(rnp);
544 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
546 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
547 rnp->level, rnp->grplo, rnp->grphi);
550 static void rcu_print_task_stall_end(void)
556 * Scan the current list of tasks blocked within RCU read-side critical
557 * sections, printing out the tid of each.
559 static int rcu_print_task_stall(struct rcu_node *rnp)
561 struct task_struct *t;
564 if (!rcu_preempt_blocked_readers_cgp(rnp))
566 rcu_print_task_stall_begin(rnp);
567 t = list_entry(rnp->gp_tasks->prev,
568 struct task_struct, rcu_node_entry);
569 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
570 pr_cont(" P%d", t->pid);
573 rcu_print_task_stall_end();
578 * Scan the current list of tasks blocked within RCU read-side critical
579 * sections, printing out the tid of each that is blocking the current
580 * expedited grace period.
582 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
584 struct task_struct *t;
589 t = list_entry(rnp->exp_tasks->prev,
590 struct task_struct, rcu_node_entry);
591 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
592 pr_cont(" P%d", t->pid);
599 * Check that the list of blocked tasks for the newly completed grace
600 * period is in fact empty. It is a serious bug to complete a grace
601 * period that still has RCU readers blocked! This function must be
602 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
603 * must be held by the caller.
605 * Also, if there are blocked tasks on the list, they automatically
606 * block the newly created grace period, so set up ->gp_tasks accordingly.
608 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
610 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
611 if (rcu_preempt_has_tasks(rnp))
612 rnp->gp_tasks = rnp->blkd_tasks.next;
613 WARN_ON_ONCE(rnp->qsmask);
617 * Check for a quiescent state from the current CPU. When a task blocks,
618 * the task is recorded in the corresponding CPU's rcu_node structure,
619 * which is checked elsewhere.
621 * Caller must disable hard irqs.
623 static void rcu_preempt_check_callbacks(void)
625 struct task_struct *t = current;
627 if (t->rcu_read_lock_nesting == 0) {
631 if (t->rcu_read_lock_nesting > 0 &&
632 __this_cpu_read(rcu_data_p->core_needs_qs) &&
633 __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm))
634 t->rcu_read_unlock_special.b.need_qs = true;
637 #ifdef CONFIG_RCU_BOOST
639 static void rcu_preempt_do_callbacks(void)
641 rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p));
644 #endif /* #ifdef CONFIG_RCU_BOOST */
647 * Queue a preemptible-RCU callback for invocation after a grace period.
649 void call_rcu(struct rcu_head *head, rcu_callback_t func)
651 __call_rcu(head, func, rcu_state_p, -1, 0);
653 EXPORT_SYMBOL_GPL(call_rcu);
656 * synchronize_rcu - wait until a grace period has elapsed.
658 * Control will return to the caller some time after a full grace
659 * period has elapsed, in other words after all currently executing RCU
660 * read-side critical sections have completed. Note, however, that
661 * upon return from synchronize_rcu(), the caller might well be executing
662 * concurrently with new RCU read-side critical sections that began while
663 * synchronize_rcu() was waiting. RCU read-side critical sections are
664 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
666 * See the description of synchronize_sched() for more detailed information
667 * on memory ordering guarantees.
669 void synchronize_rcu(void)
671 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
672 lock_is_held(&rcu_lock_map) ||
673 lock_is_held(&rcu_sched_lock_map),
674 "Illegal synchronize_rcu() in RCU read-side critical section");
675 if (!rcu_scheduler_active)
677 if (rcu_gp_is_expedited())
678 synchronize_rcu_expedited();
680 wait_rcu_gp(call_rcu);
682 EXPORT_SYMBOL_GPL(synchronize_rcu);
685 * Remote handler for smp_call_function_single(). If there is an
686 * RCU read-side critical section in effect, request that the
687 * next rcu_read_unlock() record the quiescent state up the
688 * ->expmask fields in the rcu_node tree. Otherwise, immediately
689 * report the quiescent state.
691 static void sync_rcu_exp_handler(void *info)
693 struct rcu_data *rdp;
694 struct rcu_state *rsp = info;
695 struct task_struct *t = current;
698 * Within an RCU read-side critical section, request that the next
699 * rcu_read_unlock() report. Unless this RCU read-side critical
700 * section has already blocked, in which case it is already set
701 * up for the expedited grace period to wait on it.
703 if (t->rcu_read_lock_nesting > 0 &&
704 !t->rcu_read_unlock_special.b.blocked) {
705 t->rcu_read_unlock_special.b.exp_need_qs = true;
710 * We are either exiting an RCU read-side critical section (negative
711 * values of t->rcu_read_lock_nesting) or are not in one at all
712 * (zero value of t->rcu_read_lock_nesting). Or we are in an RCU
713 * read-side critical section that blocked before this expedited
714 * grace period started. Either way, we can immediately report
715 * the quiescent state.
717 rdp = this_cpu_ptr(rsp->rda);
718 rcu_report_exp_rdp(rsp, rdp, true);
722 * synchronize_rcu_expedited - Brute-force RCU grace period
724 * Wait for an RCU-preempt grace period, but expedite it. The basic
725 * idea is to IPI all non-idle non-nohz online CPUs. The IPI handler
726 * checks whether the CPU is in an RCU-preempt critical section, and
727 * if so, it sets a flag that causes the outermost rcu_read_unlock()
728 * to report the quiescent state. On the other hand, if the CPU is
729 * not in an RCU read-side critical section, the IPI handler reports
730 * the quiescent state immediately.
732 * Although this is a greate improvement over previous expedited
733 * implementations, it is still unfriendly to real-time workloads, so is
734 * thus not recommended for any sort of common-case code. In fact, if
735 * you are using synchronize_rcu_expedited() in a loop, please restructure
736 * your code to batch your updates, and then Use a single synchronize_rcu()
739 void synchronize_rcu_expedited(void)
741 struct rcu_node *rnp;
742 struct rcu_node *rnp_unlock;
743 struct rcu_state *rsp = rcu_state_p;
746 /* If expedited grace periods are prohibited, fall back to normal. */
747 if (rcu_gp_is_normal()) {
748 wait_rcu_gp(call_rcu);
752 s = rcu_exp_gp_seq_snap(rsp);
753 trace_rcu_exp_grace_period(rsp->name, s, TPS("snap"));
755 rnp_unlock = exp_funnel_lock(rsp, s);
756 if (rnp_unlock == NULL)
757 return; /* Someone else did our work for us. */
759 rcu_exp_gp_seq_start(rsp);
760 trace_rcu_exp_grace_period(rsp->name, s, TPS("start"));
762 /* Initialize the rcu_node tree in preparation for the wait. */
763 sync_rcu_exp_select_cpus(rsp, sync_rcu_exp_handler);
765 /* Wait for snapshotted ->blkd_tasks lists to drain. */
766 rnp = rcu_get_root(rsp);
767 synchronize_sched_expedited_wait(rsp);
769 /* Clean up and exit. */
770 rcu_exp_gp_seq_end(rsp);
771 trace_rcu_exp_grace_period(rsp->name, s, TPS("end"));
772 mutex_unlock(&rnp_unlock->exp_funnel_mutex);
773 trace_rcu_exp_funnel_lock(rsp->name, rnp_unlock->level,
774 rnp_unlock->grplo, rnp_unlock->grphi,
777 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
780 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
782 * Note that this primitive does not necessarily wait for an RCU grace period
783 * to complete. For example, if there are no RCU callbacks queued anywhere
784 * in the system, then rcu_barrier() is within its rights to return
785 * immediately, without waiting for anything, much less an RCU grace period.
787 void rcu_barrier(void)
789 _rcu_barrier(rcu_state_p);
791 EXPORT_SYMBOL_GPL(rcu_barrier);
794 * Initialize preemptible RCU's state structures.
796 static void __init __rcu_init_preempt(void)
798 rcu_init_one(rcu_state_p);
802 * Check for a task exiting while in a preemptible-RCU read-side
803 * critical section, clean up if so. No need to issue warnings,
804 * as debug_check_no_locks_held() already does this if lockdep
809 struct task_struct *t = current;
811 if (likely(list_empty(¤t->rcu_node_entry)))
813 t->rcu_read_lock_nesting = 1;
815 t->rcu_read_unlock_special.b.blocked = true;
819 #else /* #ifdef CONFIG_PREEMPT_RCU */
821 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
824 * Tell them what RCU they are running.
826 static void __init rcu_bootup_announce(void)
828 pr_info("Hierarchical RCU implementation.\n");
829 rcu_bootup_announce_oddness();
833 * Because preemptible RCU does not exist, we never have to check for
834 * CPUs being in quiescent states.
836 static void rcu_preempt_note_context_switch(void)
841 * Because preemptible RCU does not exist, there are never any preempted
844 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
850 * Because there is no preemptible RCU, there can be no readers blocked.
852 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
858 * Because preemptible RCU does not exist, we never have to check for
859 * tasks blocked within RCU read-side critical sections.
861 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
866 * Because preemptible RCU does not exist, we never have to check for
867 * tasks blocked within RCU read-side critical sections.
869 static int rcu_print_task_stall(struct rcu_node *rnp)
875 * Because preemptible RCU does not exist, we never have to check for
876 * tasks blocked within RCU read-side critical sections that are
877 * blocking the current expedited grace period.
879 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
885 * Because there is no preemptible RCU, there can be no readers blocked,
886 * so there is no need to check for blocked tasks. So check only for
887 * bogus qsmask values.
889 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
891 WARN_ON_ONCE(rnp->qsmask);
895 * Because preemptible RCU does not exist, it never has any callbacks
898 static void rcu_preempt_check_callbacks(void)
903 * Wait for an rcu-preempt grace period, but make it happen quickly.
904 * But because preemptible RCU does not exist, map to rcu-sched.
906 void synchronize_rcu_expedited(void)
908 synchronize_sched_expedited();
910 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
913 * Because preemptible RCU does not exist, rcu_barrier() is just
914 * another name for rcu_barrier_sched().
916 void rcu_barrier(void)
920 EXPORT_SYMBOL_GPL(rcu_barrier);
923 * Because preemptible RCU does not exist, it need not be initialized.
925 static void __init __rcu_init_preempt(void)
930 * Because preemptible RCU does not exist, tasks cannot possibly exit
931 * while in preemptible RCU read-side critical sections.
937 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
939 #ifdef CONFIG_RCU_BOOST
941 #include "../locking/rtmutex_common.h"
943 #ifdef CONFIG_RCU_TRACE
945 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
947 if (!rcu_preempt_has_tasks(rnp))
948 rnp->n_balk_blkd_tasks++;
949 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
950 rnp->n_balk_exp_gp_tasks++;
951 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
952 rnp->n_balk_boost_tasks++;
953 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
954 rnp->n_balk_notblocked++;
955 else if (rnp->gp_tasks != NULL &&
956 ULONG_CMP_LT(jiffies, rnp->boost_time))
957 rnp->n_balk_notyet++;
962 #else /* #ifdef CONFIG_RCU_TRACE */
964 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
968 #endif /* #else #ifdef CONFIG_RCU_TRACE */
970 static void rcu_wake_cond(struct task_struct *t, int status)
973 * If the thread is yielding, only wake it when this
974 * is invoked from idle
976 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
981 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
982 * or ->boost_tasks, advancing the pointer to the next task in the
985 * Note that irqs must be enabled: boosting the task can block.
986 * Returns 1 if there are more tasks needing to be boosted.
988 static int rcu_boost(struct rcu_node *rnp)
991 struct task_struct *t;
992 struct list_head *tb;
994 if (READ_ONCE(rnp->exp_tasks) == NULL &&
995 READ_ONCE(rnp->boost_tasks) == NULL)
996 return 0; /* Nothing left to boost. */
998 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1001 * Recheck under the lock: all tasks in need of boosting
1002 * might exit their RCU read-side critical sections on their own.
1004 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1005 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1010 * Preferentially boost tasks blocking expedited grace periods.
1011 * This cannot starve the normal grace periods because a second
1012 * expedited grace period must boost all blocked tasks, including
1013 * those blocking the pre-existing normal grace period.
1015 if (rnp->exp_tasks != NULL) {
1016 tb = rnp->exp_tasks;
1017 rnp->n_exp_boosts++;
1019 tb = rnp->boost_tasks;
1020 rnp->n_normal_boosts++;
1022 rnp->n_tasks_boosted++;
1025 * We boost task t by manufacturing an rt_mutex that appears to
1026 * be held by task t. We leave a pointer to that rt_mutex where
1027 * task t can find it, and task t will release the mutex when it
1028 * exits its outermost RCU read-side critical section. Then
1029 * simply acquiring this artificial rt_mutex will boost task
1030 * t's priority. (Thanks to tglx for suggesting this approach!)
1032 * Note that task t must acquire rnp->lock to remove itself from
1033 * the ->blkd_tasks list, which it will do from exit() if from
1034 * nowhere else. We therefore are guaranteed that task t will
1035 * stay around at least until we drop rnp->lock. Note that
1036 * rnp->lock also resolves races between our priority boosting
1037 * and task t's exiting its outermost RCU read-side critical
1040 t = container_of(tb, struct task_struct, rcu_node_entry);
1041 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1042 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1043 /* Lock only for side effect: boosts task t's priority. */
1044 rt_mutex_lock(&rnp->boost_mtx);
1045 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1047 return READ_ONCE(rnp->exp_tasks) != NULL ||
1048 READ_ONCE(rnp->boost_tasks) != NULL;
1052 * Priority-boosting kthread, one per leaf rcu_node.
1054 static int rcu_boost_kthread(void *arg)
1056 struct rcu_node *rnp = (struct rcu_node *)arg;
1060 trace_rcu_utilization(TPS("Start boost kthread@init"));
1062 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1063 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1064 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1065 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1066 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1067 more2boost = rcu_boost(rnp);
1073 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1074 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1075 schedule_timeout_interruptible(2);
1076 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1081 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1086 * Check to see if it is time to start boosting RCU readers that are
1087 * blocking the current grace period, and, if so, tell the per-rcu_node
1088 * kthread to start boosting them. If there is an expedited grace
1089 * period in progress, it is always time to boost.
1091 * The caller must hold rnp->lock, which this function releases.
1092 * The ->boost_kthread_task is immortal, so we don't need to worry
1093 * about it going away.
1095 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1096 __releases(rnp->lock)
1098 struct task_struct *t;
1100 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1101 rnp->n_balk_exp_gp_tasks++;
1102 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1105 if (rnp->exp_tasks != NULL ||
1106 (rnp->gp_tasks != NULL &&
1107 rnp->boost_tasks == NULL &&
1109 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1110 if (rnp->exp_tasks == NULL)
1111 rnp->boost_tasks = rnp->gp_tasks;
1112 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1113 t = rnp->boost_kthread_task;
1115 rcu_wake_cond(t, rnp->boost_kthread_status);
1117 rcu_initiate_boost_trace(rnp);
1118 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1123 * Wake up the per-CPU kthread to invoke RCU callbacks.
1125 static void invoke_rcu_callbacks_kthread(void)
1127 unsigned long flags;
1129 local_irq_save(flags);
1130 __this_cpu_write(rcu_cpu_has_work, 1);
1131 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1132 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1133 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1134 __this_cpu_read(rcu_cpu_kthread_status));
1136 local_irq_restore(flags);
1140 * Is the current CPU running the RCU-callbacks kthread?
1141 * Caller must have preemption disabled.
1143 static bool rcu_is_callbacks_kthread(void)
1145 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1148 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1151 * Do priority-boost accounting for the start of a new grace period.
1153 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1155 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1159 * Create an RCU-boost kthread for the specified node if one does not
1160 * already exist. We only create this kthread for preemptible RCU.
1161 * Returns zero if all is well, a negated errno otherwise.
1163 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1164 struct rcu_node *rnp)
1166 int rnp_index = rnp - &rsp->node[0];
1167 unsigned long flags;
1168 struct sched_param sp;
1169 struct task_struct *t;
1171 if (rcu_state_p != rsp)
1174 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1178 if (rnp->boost_kthread_task != NULL)
1180 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1181 "rcub/%d", rnp_index);
1184 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1185 rnp->boost_kthread_task = t;
1186 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1187 sp.sched_priority = kthread_prio;
1188 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1189 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1193 static void rcu_kthread_do_work(void)
1195 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1196 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1197 rcu_preempt_do_callbacks();
1200 static void rcu_cpu_kthread_setup(unsigned int cpu)
1202 struct sched_param sp;
1204 sp.sched_priority = kthread_prio;
1205 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1208 static void rcu_cpu_kthread_park(unsigned int cpu)
1210 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1213 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1215 return __this_cpu_read(rcu_cpu_has_work);
1219 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1220 * RCU softirq used in flavors and configurations of RCU that do not
1221 * support RCU priority boosting.
1223 static void rcu_cpu_kthread(unsigned int cpu)
1225 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1226 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1229 for (spincnt = 0; spincnt < 10; spincnt++) {
1230 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1232 *statusp = RCU_KTHREAD_RUNNING;
1233 this_cpu_inc(rcu_cpu_kthread_loops);
1234 local_irq_disable();
1239 rcu_kthread_do_work();
1242 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1243 *statusp = RCU_KTHREAD_WAITING;
1247 *statusp = RCU_KTHREAD_YIELDING;
1248 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1249 schedule_timeout_interruptible(2);
1250 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1251 *statusp = RCU_KTHREAD_WAITING;
1255 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1256 * served by the rcu_node in question. The CPU hotplug lock is still
1257 * held, so the value of rnp->qsmaskinit will be stable.
1259 * We don't include outgoingcpu in the affinity set, use -1 if there is
1260 * no outgoing CPU. If there are no CPUs left in the affinity set,
1261 * this function allows the kthread to execute on any CPU.
1263 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1265 struct task_struct *t = rnp->boost_kthread_task;
1266 unsigned long mask = rcu_rnp_online_cpus(rnp);
1272 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1274 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1275 if ((mask & 0x1) && cpu != outgoingcpu)
1276 cpumask_set_cpu(cpu, cm);
1277 if (cpumask_weight(cm) == 0)
1279 set_cpus_allowed_ptr(t, cm);
1280 free_cpumask_var(cm);
1283 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1284 .store = &rcu_cpu_kthread_task,
1285 .thread_should_run = rcu_cpu_kthread_should_run,
1286 .thread_fn = rcu_cpu_kthread,
1287 .thread_comm = "rcuc/%u",
1288 .setup = rcu_cpu_kthread_setup,
1289 .park = rcu_cpu_kthread_park,
1293 * Spawn boost kthreads -- called as soon as the scheduler is running.
1295 static void __init rcu_spawn_boost_kthreads(void)
1297 struct rcu_node *rnp;
1300 for_each_possible_cpu(cpu)
1301 per_cpu(rcu_cpu_has_work, cpu) = 0;
1302 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1303 rcu_for_each_leaf_node(rcu_state_p, rnp)
1304 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1307 static void rcu_prepare_kthreads(int cpu)
1309 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1310 struct rcu_node *rnp = rdp->mynode;
1312 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1313 if (rcu_scheduler_fully_active)
1314 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1317 #else /* #ifdef CONFIG_RCU_BOOST */
1319 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1320 __releases(rnp->lock)
1322 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1325 static void invoke_rcu_callbacks_kthread(void)
1330 static bool rcu_is_callbacks_kthread(void)
1335 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1339 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1343 static void __init rcu_spawn_boost_kthreads(void)
1347 static void rcu_prepare_kthreads(int cpu)
1351 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1353 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1356 * Check to see if any future RCU-related work will need to be done
1357 * by the current CPU, even if none need be done immediately, returning
1358 * 1 if so. This function is part of the RCU implementation; it is -not-
1359 * an exported member of the RCU API.
1361 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1362 * any flavor of RCU.
1364 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1366 *nextevt = KTIME_MAX;
1367 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1368 ? 0 : rcu_cpu_has_callbacks(NULL);
1372 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1375 static void rcu_cleanup_after_idle(void)
1380 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1383 static void rcu_prepare_for_idle(void)
1388 * Don't bother keeping a running count of the number of RCU callbacks
1389 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1391 static void rcu_idle_count_callbacks_posted(void)
1395 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1398 * This code is invoked when a CPU goes idle, at which point we want
1399 * to have the CPU do everything required for RCU so that it can enter
1400 * the energy-efficient dyntick-idle mode. This is handled by a
1401 * state machine implemented by rcu_prepare_for_idle() below.
1403 * The following three proprocessor symbols control this state machine:
1405 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1406 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1407 * is sized to be roughly one RCU grace period. Those energy-efficiency
1408 * benchmarkers who might otherwise be tempted to set this to a large
1409 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1410 * system. And if you are -that- concerned about energy efficiency,
1411 * just power the system down and be done with it!
1412 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1413 * permitted to sleep in dyntick-idle mode with only lazy RCU
1414 * callbacks pending. Setting this too high can OOM your system.
1416 * The values below work well in practice. If future workloads require
1417 * adjustment, they can be converted into kernel config parameters, though
1418 * making the state machine smarter might be a better option.
1420 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1421 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1423 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1424 module_param(rcu_idle_gp_delay, int, 0644);
1425 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1426 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1429 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1430 * only if it has been awhile since the last time we did so. Afterwards,
1431 * if there are any callbacks ready for immediate invocation, return true.
1433 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1435 bool cbs_ready = false;
1436 struct rcu_data *rdp;
1437 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1438 struct rcu_node *rnp;
1439 struct rcu_state *rsp;
1441 /* Exit early if we advanced recently. */
1442 if (jiffies == rdtp->last_advance_all)
1444 rdtp->last_advance_all = jiffies;
1446 for_each_rcu_flavor(rsp) {
1447 rdp = this_cpu_ptr(rsp->rda);
1451 * Don't bother checking unless a grace period has
1452 * completed since we last checked and there are
1453 * callbacks not yet ready to invoke.
1455 if ((rdp->completed != rnp->completed ||
1456 unlikely(READ_ONCE(rdp->gpwrap))) &&
1457 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1458 note_gp_changes(rsp, rdp);
1460 if (cpu_has_callbacks_ready_to_invoke(rdp))
1467 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1468 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1469 * caller to set the timeout based on whether or not there are non-lazy
1472 * The caller must have disabled interrupts.
1474 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1476 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1479 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1480 *nextevt = KTIME_MAX;
1484 /* Snapshot to detect later posting of non-lazy callback. */
1485 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1487 /* If no callbacks, RCU doesn't need the CPU. */
1488 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1489 *nextevt = KTIME_MAX;
1493 /* Attempt to advance callbacks. */
1494 if (rcu_try_advance_all_cbs()) {
1495 /* Some ready to invoke, so initiate later invocation. */
1499 rdtp->last_accelerate = jiffies;
1501 /* Request timer delay depending on laziness, and round. */
1502 if (!rdtp->all_lazy) {
1503 dj = round_up(rcu_idle_gp_delay + jiffies,
1504 rcu_idle_gp_delay) - jiffies;
1506 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1508 *nextevt = basemono + dj * TICK_NSEC;
1513 * Prepare a CPU for idle from an RCU perspective. The first major task
1514 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1515 * The second major task is to check to see if a non-lazy callback has
1516 * arrived at a CPU that previously had only lazy callbacks. The third
1517 * major task is to accelerate (that is, assign grace-period numbers to)
1518 * any recently arrived callbacks.
1520 * The caller must have disabled interrupts.
1522 static void rcu_prepare_for_idle(void)
1525 struct rcu_data *rdp;
1526 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1527 struct rcu_node *rnp;
1528 struct rcu_state *rsp;
1531 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1532 rcu_is_nocb_cpu(smp_processor_id()))
1535 /* Handle nohz enablement switches conservatively. */
1536 tne = READ_ONCE(tick_nohz_active);
1537 if (tne != rdtp->tick_nohz_enabled_snap) {
1538 if (rcu_cpu_has_callbacks(NULL))
1539 invoke_rcu_core(); /* force nohz to see update. */
1540 rdtp->tick_nohz_enabled_snap = tne;
1547 * If a non-lazy callback arrived at a CPU having only lazy
1548 * callbacks, invoke RCU core for the side-effect of recalculating
1549 * idle duration on re-entry to idle.
1551 if (rdtp->all_lazy &&
1552 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1553 rdtp->all_lazy = false;
1554 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1560 * If we have not yet accelerated this jiffy, accelerate all
1561 * callbacks on this CPU.
1563 if (rdtp->last_accelerate == jiffies)
1565 rdtp->last_accelerate = jiffies;
1566 for_each_rcu_flavor(rsp) {
1567 rdp = this_cpu_ptr(rsp->rda);
1568 if (!*rdp->nxttail[RCU_DONE_TAIL])
1571 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1572 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1573 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1575 rcu_gp_kthread_wake(rsp);
1580 * Clean up for exit from idle. Attempt to advance callbacks based on
1581 * any grace periods that elapsed while the CPU was idle, and if any
1582 * callbacks are now ready to invoke, initiate invocation.
1584 static void rcu_cleanup_after_idle(void)
1586 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1587 rcu_is_nocb_cpu(smp_processor_id()))
1589 if (rcu_try_advance_all_cbs())
1594 * Keep a running count of the number of non-lazy callbacks posted
1595 * on this CPU. This running counter (which is never decremented) allows
1596 * rcu_prepare_for_idle() to detect when something out of the idle loop
1597 * posts a callback, even if an equal number of callbacks are invoked.
1598 * Of course, callbacks should only be posted from within a trace event
1599 * designed to be called from idle or from within RCU_NONIDLE().
1601 static void rcu_idle_count_callbacks_posted(void)
1603 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1607 * Data for flushing lazy RCU callbacks at OOM time.
1609 static atomic_t oom_callback_count;
1610 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1613 * RCU OOM callback -- decrement the outstanding count and deliver the
1614 * wake-up if we are the last one.
1616 static void rcu_oom_callback(struct rcu_head *rhp)
1618 if (atomic_dec_and_test(&oom_callback_count))
1619 wake_up(&oom_callback_wq);
1623 * Post an rcu_oom_notify callback on the current CPU if it has at
1624 * least one lazy callback. This will unnecessarily post callbacks
1625 * to CPUs that already have a non-lazy callback at the end of their
1626 * callback list, but this is an infrequent operation, so accept some
1627 * extra overhead to keep things simple.
1629 static void rcu_oom_notify_cpu(void *unused)
1631 struct rcu_state *rsp;
1632 struct rcu_data *rdp;
1634 for_each_rcu_flavor(rsp) {
1635 rdp = raw_cpu_ptr(rsp->rda);
1636 if (rdp->qlen_lazy != 0) {
1637 atomic_inc(&oom_callback_count);
1638 rsp->call(&rdp->oom_head, rcu_oom_callback);
1644 * If low on memory, ensure that each CPU has a non-lazy callback.
1645 * This will wake up CPUs that have only lazy callbacks, in turn
1646 * ensuring that they free up the corresponding memory in a timely manner.
1647 * Because an uncertain amount of memory will be freed in some uncertain
1648 * timeframe, we do not claim to have freed anything.
1650 static int rcu_oom_notify(struct notifier_block *self,
1651 unsigned long notused, void *nfreed)
1655 /* Wait for callbacks from earlier instance to complete. */
1656 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1657 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1660 * Prevent premature wakeup: ensure that all increments happen
1661 * before there is a chance of the counter reaching zero.
1663 atomic_set(&oom_callback_count, 1);
1665 for_each_online_cpu(cpu) {
1666 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1667 cond_resched_rcu_qs();
1670 /* Unconditionally decrement: no need to wake ourselves up. */
1671 atomic_dec(&oom_callback_count);
1676 static struct notifier_block rcu_oom_nb = {
1677 .notifier_call = rcu_oom_notify
1680 static int __init rcu_register_oom_notifier(void)
1682 register_oom_notifier(&rcu_oom_nb);
1685 early_initcall(rcu_register_oom_notifier);
1687 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1689 #ifdef CONFIG_RCU_FAST_NO_HZ
1691 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1693 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1694 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1696 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1697 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1699 rdtp->all_lazy ? 'L' : '.',
1700 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1703 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1705 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1710 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1712 /* Initiate the stall-info list. */
1713 static void print_cpu_stall_info_begin(void)
1719 * Print out diagnostic information for the specified stalled CPU.
1721 * If the specified CPU is aware of the current RCU grace period
1722 * (flavor specified by rsp), then print the number of scheduling
1723 * clock interrupts the CPU has taken during the time that it has
1724 * been aware. Otherwise, print the number of RCU grace periods
1725 * that this CPU is ignorant of, for example, "1" if the CPU was
1726 * aware of the previous grace period.
1728 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1730 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1732 char fast_no_hz[72];
1733 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1734 struct rcu_dynticks *rdtp = rdp->dynticks;
1736 unsigned long ticks_value;
1738 if (rsp->gpnum == rdp->gpnum) {
1739 ticks_title = "ticks this GP";
1740 ticks_value = rdp->ticks_this_gp;
1742 ticks_title = "GPs behind";
1743 ticks_value = rsp->gpnum - rdp->gpnum;
1745 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1746 pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1748 "O."[!!cpu_online(cpu)],
1749 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1750 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1751 ticks_value, ticks_title,
1752 atomic_read(&rdtp->dynticks) & 0xfff,
1753 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1754 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1755 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1759 /* Terminate the stall-info list. */
1760 static void print_cpu_stall_info_end(void)
1765 /* Zero ->ticks_this_gp for all flavors of RCU. */
1766 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1768 rdp->ticks_this_gp = 0;
1769 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1772 /* Increment ->ticks_this_gp for all flavors of RCU. */
1773 static void increment_cpu_stall_ticks(void)
1775 struct rcu_state *rsp;
1777 for_each_rcu_flavor(rsp)
1778 raw_cpu_inc(rsp->rda->ticks_this_gp);
1781 #ifdef CONFIG_RCU_NOCB_CPU
1784 * Offload callback processing from the boot-time-specified set of CPUs
1785 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1786 * kthread created that pulls the callbacks from the corresponding CPU,
1787 * waits for a grace period to elapse, and invokes the callbacks.
1788 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1789 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1790 * has been specified, in which case each kthread actively polls its
1791 * CPU. (Which isn't so great for energy efficiency, but which does
1792 * reduce RCU's overhead on that CPU.)
1794 * This is intended to be used in conjunction with Frederic Weisbecker's
1795 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1796 * running CPU-bound user-mode computations.
1798 * Offloading of callback processing could also in theory be used as
1799 * an energy-efficiency measure because CPUs with no RCU callbacks
1800 * queued are more aggressive about entering dyntick-idle mode.
1804 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1805 static int __init rcu_nocb_setup(char *str)
1807 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1808 have_rcu_nocb_mask = true;
1809 cpulist_parse(str, rcu_nocb_mask);
1812 __setup("rcu_nocbs=", rcu_nocb_setup);
1814 static int __init parse_rcu_nocb_poll(char *arg)
1819 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1822 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1825 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1827 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1831 * Set the root rcu_node structure's ->need_future_gp field
1832 * based on the sum of those of all rcu_node structures. This does
1833 * double-count the root rcu_node structure's requests, but this
1834 * is necessary to handle the possibility of a rcu_nocb_kthread()
1835 * having awakened during the time that the rcu_node structures
1836 * were being updated for the end of the previous grace period.
1838 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1840 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1843 static void rcu_init_one_nocb(struct rcu_node *rnp)
1845 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1846 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1849 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1850 /* Is the specified CPU a no-CBs CPU? */
1851 bool rcu_is_nocb_cpu(int cpu)
1853 if (have_rcu_nocb_mask)
1854 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1857 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1860 * Kick the leader kthread for this NOCB group.
1862 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1864 struct rcu_data *rdp_leader = rdp->nocb_leader;
1866 if (!READ_ONCE(rdp_leader->nocb_kthread))
1868 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1869 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1870 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1871 wake_up(&rdp_leader->nocb_wq);
1876 * Does the specified CPU need an RCU callback for the specified flavor
1879 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1881 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1883 #ifdef CONFIG_PROVE_RCU
1884 struct rcu_head *rhp;
1885 #endif /* #ifdef CONFIG_PROVE_RCU */
1888 * Check count of all no-CBs callbacks awaiting invocation.
1889 * There needs to be a barrier before this function is called,
1890 * but associated with a prior determination that no more
1891 * callbacks would be posted. In the worst case, the first
1892 * barrier in _rcu_barrier() suffices (but the caller cannot
1893 * necessarily rely on this, not a substitute for the caller
1894 * getting the concurrency design right!). There must also be
1895 * a barrier between the following load an posting of a callback
1896 * (if a callback is in fact needed). This is associated with an
1897 * atomic_inc() in the caller.
1899 ret = atomic_long_read(&rdp->nocb_q_count);
1901 #ifdef CONFIG_PROVE_RCU
1902 rhp = READ_ONCE(rdp->nocb_head);
1904 rhp = READ_ONCE(rdp->nocb_gp_head);
1906 rhp = READ_ONCE(rdp->nocb_follower_head);
1908 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1909 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1910 rcu_scheduler_fully_active) {
1911 /* RCU callback enqueued before CPU first came online??? */
1912 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1916 #endif /* #ifdef CONFIG_PROVE_RCU */
1922 * Enqueue the specified string of rcu_head structures onto the specified
1923 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1924 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1925 * counts are supplied by rhcount and rhcount_lazy.
1927 * If warranted, also wake up the kthread servicing this CPUs queues.
1929 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1930 struct rcu_head *rhp,
1931 struct rcu_head **rhtp,
1932 int rhcount, int rhcount_lazy,
1933 unsigned long flags)
1936 struct rcu_head **old_rhpp;
1937 struct task_struct *t;
1939 /* Enqueue the callback on the nocb list and update counts. */
1940 atomic_long_add(rhcount, &rdp->nocb_q_count);
1941 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1942 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1943 WRITE_ONCE(*old_rhpp, rhp);
1944 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1945 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1947 /* If we are not being polled and there is a kthread, awaken it ... */
1948 t = READ_ONCE(rdp->nocb_kthread);
1949 if (rcu_nocb_poll || !t) {
1950 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1951 TPS("WakeNotPoll"));
1954 len = atomic_long_read(&rdp->nocb_q_count);
1955 if (old_rhpp == &rdp->nocb_head) {
1956 if (!irqs_disabled_flags(flags)) {
1957 /* ... if queue was empty ... */
1958 wake_nocb_leader(rdp, false);
1959 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1962 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1963 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1964 TPS("WakeEmptyIsDeferred"));
1966 rdp->qlen_last_fqs_check = 0;
1967 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1968 /* ... or if many callbacks queued. */
1969 if (!irqs_disabled_flags(flags)) {
1970 wake_nocb_leader(rdp, true);
1971 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1974 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1975 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1976 TPS("WakeOvfIsDeferred"));
1978 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1980 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1986 * This is a helper for __call_rcu(), which invokes this when the normal
1987 * callback queue is inoperable. If this is not a no-CBs CPU, this
1988 * function returns failure back to __call_rcu(), which can complain
1991 * Otherwise, this function queues the callback where the corresponding
1992 * "rcuo" kthread can find it.
1994 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1995 bool lazy, unsigned long flags)
1998 if (!rcu_is_nocb_cpu(rdp->cpu))
2000 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2001 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2002 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2003 (unsigned long)rhp->func,
2004 -atomic_long_read(&rdp->nocb_q_count_lazy),
2005 -atomic_long_read(&rdp->nocb_q_count));
2007 trace_rcu_callback(rdp->rsp->name, rhp,
2008 -atomic_long_read(&rdp->nocb_q_count_lazy),
2009 -atomic_long_read(&rdp->nocb_q_count));
2012 * If called from an extended quiescent state with interrupts
2013 * disabled, invoke the RCU core in order to allow the idle-entry
2014 * deferred-wakeup check to function.
2016 if (irqs_disabled_flags(flags) &&
2017 !rcu_is_watching() &&
2018 cpu_online(smp_processor_id()))
2025 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2028 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2029 struct rcu_data *rdp,
2030 unsigned long flags)
2032 long ql = rsp->qlen;
2033 long qll = rsp->qlen_lazy;
2035 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2036 if (!rcu_is_nocb_cpu(smp_processor_id()))
2041 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2042 if (rsp->orphan_donelist != NULL) {
2043 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2044 rsp->orphan_donetail, ql, qll, flags);
2046 rsp->orphan_donelist = NULL;
2047 rsp->orphan_donetail = &rsp->orphan_donelist;
2049 if (rsp->orphan_nxtlist != NULL) {
2050 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2051 rsp->orphan_nxttail, ql, qll, flags);
2053 rsp->orphan_nxtlist = NULL;
2054 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2060 * If necessary, kick off a new grace period, and either way wait
2061 * for a subsequent grace period to complete.
2063 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2067 unsigned long flags;
2069 struct rcu_node *rnp = rdp->mynode;
2071 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2072 needwake = rcu_start_future_gp(rnp, rdp, &c);
2073 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2075 rcu_gp_kthread_wake(rdp->rsp);
2078 * Wait for the grace period. Do so interruptibly to avoid messing
2079 * up the load average.
2081 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2083 wait_event_interruptible(
2084 rnp->nocb_gp_wq[c & 0x1],
2085 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
2088 WARN_ON(signal_pending(current));
2089 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2091 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2092 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2096 * Leaders come here to wait for additional callbacks to show up.
2097 * This function does not return until callbacks appear.
2099 static void nocb_leader_wait(struct rcu_data *my_rdp)
2101 bool firsttime = true;
2103 struct rcu_data *rdp;
2104 struct rcu_head **tail;
2108 /* Wait for callbacks to appear. */
2109 if (!rcu_nocb_poll) {
2110 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2111 wait_event_interruptible(my_rdp->nocb_wq,
2112 !READ_ONCE(my_rdp->nocb_leader_sleep));
2113 /* Memory barrier handled by smp_mb() calls below and repoll. */
2114 } else if (firsttime) {
2115 firsttime = false; /* Don't drown trace log with "Poll"! */
2116 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2120 * Each pass through the following loop checks a follower for CBs.
2121 * We are our own first follower. Any CBs found are moved to
2122 * nocb_gp_head, where they await a grace period.
2125 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2126 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2127 if (!rdp->nocb_gp_head)
2128 continue; /* No CBs here, try next follower. */
2130 /* Move callbacks to wait-for-GP list, which is empty. */
2131 WRITE_ONCE(rdp->nocb_head, NULL);
2132 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2137 * If there were no callbacks, sleep a bit, rescan after a
2138 * memory barrier, and go retry.
2140 if (unlikely(!gotcbs)) {
2142 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2144 WARN_ON(signal_pending(current));
2145 schedule_timeout_interruptible(1);
2147 /* Rescan in case we were a victim of memory ordering. */
2148 my_rdp->nocb_leader_sleep = true;
2149 smp_mb(); /* Ensure _sleep true before scan. */
2150 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2151 if (READ_ONCE(rdp->nocb_head)) {
2152 /* Found CB, so short-circuit next wait. */
2153 my_rdp->nocb_leader_sleep = false;
2159 /* Wait for one grace period. */
2160 rcu_nocb_wait_gp(my_rdp);
2163 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2164 * We set it now, but recheck for new callbacks while
2165 * traversing our follower list.
2167 my_rdp->nocb_leader_sleep = true;
2168 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2170 /* Each pass through the following loop wakes a follower, if needed. */
2171 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2172 if (READ_ONCE(rdp->nocb_head))
2173 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2174 if (!rdp->nocb_gp_head)
2175 continue; /* No CBs, so no need to wake follower. */
2177 /* Append callbacks to follower's "done" list. */
2178 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2179 *tail = rdp->nocb_gp_head;
2180 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2181 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2183 * List was empty, wake up the follower.
2184 * Memory barriers supplied by atomic_long_add().
2186 wake_up(&rdp->nocb_wq);
2190 /* If we (the leader) don't have CBs, go wait some more. */
2191 if (!my_rdp->nocb_follower_head)
2196 * Followers come here to wait for additional callbacks to show up.
2197 * This function does not return until callbacks appear.
2199 static void nocb_follower_wait(struct rcu_data *rdp)
2201 bool firsttime = true;
2204 if (!rcu_nocb_poll) {
2205 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2207 wait_event_interruptible(rdp->nocb_wq,
2208 READ_ONCE(rdp->nocb_follower_head));
2209 } else if (firsttime) {
2210 /* Don't drown trace log with "Poll"! */
2212 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2214 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2215 /* ^^^ Ensure CB invocation follows _head test. */
2219 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2221 WARN_ON(signal_pending(current));
2222 schedule_timeout_interruptible(1);
2227 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2228 * callbacks queued by the corresponding no-CBs CPU, however, there is
2229 * an optional leader-follower relationship so that the grace-period
2230 * kthreads don't have to do quite so many wakeups.
2232 static int rcu_nocb_kthread(void *arg)
2235 struct rcu_head *list;
2236 struct rcu_head *next;
2237 struct rcu_head **tail;
2238 struct rcu_data *rdp = arg;
2240 /* Each pass through this loop invokes one batch of callbacks */
2242 /* Wait for callbacks. */
2243 if (rdp->nocb_leader == rdp)
2244 nocb_leader_wait(rdp);
2246 nocb_follower_wait(rdp);
2248 /* Pull the ready-to-invoke callbacks onto local list. */
2249 list = READ_ONCE(rdp->nocb_follower_head);
2251 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2252 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2253 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2255 /* Each pass through the following loop invokes a callback. */
2256 trace_rcu_batch_start(rdp->rsp->name,
2257 atomic_long_read(&rdp->nocb_q_count_lazy),
2258 atomic_long_read(&rdp->nocb_q_count), -1);
2262 /* Wait for enqueuing to complete, if needed. */
2263 while (next == NULL && &list->next != tail) {
2264 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2266 schedule_timeout_interruptible(1);
2267 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2271 debug_rcu_head_unqueue(list);
2273 if (__rcu_reclaim(rdp->rsp->name, list))
2279 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2280 smp_mb__before_atomic(); /* _add after CB invocation. */
2281 atomic_long_add(-c, &rdp->nocb_q_count);
2282 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2283 rdp->n_nocbs_invoked += c;
2288 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2289 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2291 return READ_ONCE(rdp->nocb_defer_wakeup);
2294 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2295 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2299 if (!rcu_nocb_need_deferred_wakeup(rdp))
2301 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2302 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2303 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2304 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2307 void __init rcu_init_nohz(void)
2310 bool need_rcu_nocb_mask = true;
2311 struct rcu_state *rsp;
2313 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2314 need_rcu_nocb_mask = false;
2315 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2317 #if defined(CONFIG_NO_HZ_FULL)
2318 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2319 need_rcu_nocb_mask = true;
2320 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2322 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2323 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2324 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2327 have_rcu_nocb_mask = true;
2329 if (!have_rcu_nocb_mask)
2332 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2333 pr_info("\tOffload RCU callbacks from CPU 0\n");
2334 cpumask_set_cpu(0, rcu_nocb_mask);
2335 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2336 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2337 pr_info("\tOffload RCU callbacks from all CPUs\n");
2338 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2339 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2340 #if defined(CONFIG_NO_HZ_FULL)
2341 if (tick_nohz_full_running)
2342 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2343 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2345 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2346 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2347 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2350 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2351 cpumask_pr_args(rcu_nocb_mask));
2353 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2355 for_each_rcu_flavor(rsp) {
2356 for_each_cpu(cpu, rcu_nocb_mask)
2357 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2358 rcu_organize_nocb_kthreads(rsp);
2362 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2363 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2365 rdp->nocb_tail = &rdp->nocb_head;
2366 init_waitqueue_head(&rdp->nocb_wq);
2367 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2371 * If the specified CPU is a no-CBs CPU that does not already have its
2372 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2373 * brought online out of order, this can require re-organizing the
2374 * leader-follower relationships.
2376 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2378 struct rcu_data *rdp;
2379 struct rcu_data *rdp_last;
2380 struct rcu_data *rdp_old_leader;
2381 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2382 struct task_struct *t;
2385 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2386 * then nothing to do.
2388 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2391 /* If we didn't spawn the leader first, reorganize! */
2392 rdp_old_leader = rdp_spawn->nocb_leader;
2393 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2395 rdp = rdp_old_leader;
2397 rdp->nocb_leader = rdp_spawn;
2398 if (rdp_last && rdp != rdp_spawn)
2399 rdp_last->nocb_next_follower = rdp;
2400 if (rdp == rdp_spawn) {
2401 rdp = rdp->nocb_next_follower;
2404 rdp = rdp->nocb_next_follower;
2405 rdp_last->nocb_next_follower = NULL;
2408 rdp_spawn->nocb_next_follower = rdp_old_leader;
2411 /* Spawn the kthread for this CPU and RCU flavor. */
2412 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2413 "rcuo%c/%d", rsp->abbr, cpu);
2415 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2419 * If the specified CPU is a no-CBs CPU that does not already have its
2420 * rcuo kthreads, spawn them.
2422 static void rcu_spawn_all_nocb_kthreads(int cpu)
2424 struct rcu_state *rsp;
2426 if (rcu_scheduler_fully_active)
2427 for_each_rcu_flavor(rsp)
2428 rcu_spawn_one_nocb_kthread(rsp, cpu);
2432 * Once the scheduler is running, spawn rcuo kthreads for all online
2433 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2434 * non-boot CPUs come online -- if this changes, we will need to add
2435 * some mutual exclusion.
2437 static void __init rcu_spawn_nocb_kthreads(void)
2441 for_each_online_cpu(cpu)
2442 rcu_spawn_all_nocb_kthreads(cpu);
2445 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2446 static int rcu_nocb_leader_stride = -1;
2447 module_param(rcu_nocb_leader_stride, int, 0444);
2450 * Initialize leader-follower relationships for all no-CBs CPU.
2452 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2455 int ls = rcu_nocb_leader_stride;
2456 int nl = 0; /* Next leader. */
2457 struct rcu_data *rdp;
2458 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2459 struct rcu_data *rdp_prev = NULL;
2461 if (!have_rcu_nocb_mask)
2464 ls = int_sqrt(nr_cpu_ids);
2465 rcu_nocb_leader_stride = ls;
2469 * Each pass through this loop sets up one rcu_data structure and
2470 * spawns one rcu_nocb_kthread().
2472 for_each_cpu(cpu, rcu_nocb_mask) {
2473 rdp = per_cpu_ptr(rsp->rda, cpu);
2474 if (rdp->cpu >= nl) {
2475 /* New leader, set up for followers & next leader. */
2476 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2477 rdp->nocb_leader = rdp;
2480 /* Another follower, link to previous leader. */
2481 rdp->nocb_leader = rdp_leader;
2482 rdp_prev->nocb_next_follower = rdp;
2488 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2489 static bool init_nocb_callback_list(struct rcu_data *rdp)
2491 if (!rcu_is_nocb_cpu(rdp->cpu))
2494 /* If there are early-boot callbacks, move them to nocb lists. */
2496 rdp->nocb_head = rdp->nxtlist;
2497 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2498 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2499 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2500 rdp->nxtlist = NULL;
2504 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2508 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2510 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2512 WARN_ON_ONCE(1); /* Should be dead code. */
2516 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2520 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2524 static void rcu_init_one_nocb(struct rcu_node *rnp)
2528 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2529 bool lazy, unsigned long flags)
2534 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2535 struct rcu_data *rdp,
2536 unsigned long flags)
2541 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2545 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2550 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2554 static void rcu_spawn_all_nocb_kthreads(int cpu)
2558 static void __init rcu_spawn_nocb_kthreads(void)
2562 static bool init_nocb_callback_list(struct rcu_data *rdp)
2567 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2570 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2571 * arbitrarily long period of time with the scheduling-clock tick turned
2572 * off. RCU will be paying attention to this CPU because it is in the
2573 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2574 * machine because the scheduling-clock tick has been disabled. Therefore,
2575 * if an adaptive-ticks CPU is failing to respond to the current grace
2576 * period and has not be idle from an RCU perspective, kick it.
2578 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2580 #ifdef CONFIG_NO_HZ_FULL
2581 if (tick_nohz_full_cpu(cpu))
2582 smp_send_reschedule(cpu);
2583 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2587 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2589 static int full_sysidle_state; /* Current system-idle state. */
2590 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2591 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2592 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2593 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2594 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2597 * Invoked to note exit from irq or task transition to idle. Note that
2598 * usermode execution does -not- count as idle here! After all, we want
2599 * to detect full-system idle states, not RCU quiescent states and grace
2600 * periods. The caller must have disabled interrupts.
2602 static void rcu_sysidle_enter(int irq)
2605 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2607 /* If there are no nohz_full= CPUs, no need to track this. */
2608 if (!tick_nohz_full_enabled())
2611 /* Adjust nesting, check for fully idle. */
2613 rdtp->dynticks_idle_nesting--;
2614 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2615 if (rdtp->dynticks_idle_nesting != 0)
2616 return; /* Still not fully idle. */
2618 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2619 DYNTICK_TASK_NEST_VALUE) {
2620 rdtp->dynticks_idle_nesting = 0;
2622 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2623 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2624 return; /* Still not fully idle. */
2628 /* Record start of fully idle period. */
2630 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2631 smp_mb__before_atomic();
2632 atomic_inc(&rdtp->dynticks_idle);
2633 smp_mb__after_atomic();
2634 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2638 * Unconditionally force exit from full system-idle state. This is
2639 * invoked when a normal CPU exits idle, but must be called separately
2640 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2641 * is that the timekeeping CPU is permitted to take scheduling-clock
2642 * interrupts while the system is in system-idle state, and of course
2643 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2644 * interrupt from any other type of interrupt.
2646 void rcu_sysidle_force_exit(void)
2648 int oldstate = READ_ONCE(full_sysidle_state);
2652 * Each pass through the following loop attempts to exit full
2653 * system-idle state. If contention proves to be a problem,
2654 * a trylock-based contention tree could be used here.
2656 while (oldstate > RCU_SYSIDLE_SHORT) {
2657 newoldstate = cmpxchg(&full_sysidle_state,
2658 oldstate, RCU_SYSIDLE_NOT);
2659 if (oldstate == newoldstate &&
2660 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2661 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2662 return; /* We cleared it, done! */
2664 oldstate = newoldstate;
2666 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2670 * Invoked to note entry to irq or task transition from idle. Note that
2671 * usermode execution does -not- count as idle here! The caller must
2672 * have disabled interrupts.
2674 static void rcu_sysidle_exit(int irq)
2676 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2678 /* If there are no nohz_full= CPUs, no need to track this. */
2679 if (!tick_nohz_full_enabled())
2682 /* Adjust nesting, check for already non-idle. */
2684 rdtp->dynticks_idle_nesting++;
2685 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2686 if (rdtp->dynticks_idle_nesting != 1)
2687 return; /* Already non-idle. */
2690 * Allow for irq misnesting. Yes, it really is possible
2691 * to enter an irq handler then never leave it, and maybe
2692 * also vice versa. Handle both possibilities.
2694 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2695 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2696 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2697 return; /* Already non-idle. */
2699 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2703 /* Record end of idle period. */
2704 smp_mb__before_atomic();
2705 atomic_inc(&rdtp->dynticks_idle);
2706 smp_mb__after_atomic();
2707 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2710 * If we are the timekeeping CPU, we are permitted to be non-idle
2711 * during a system-idle state. This must be the case, because
2712 * the timekeeping CPU has to take scheduling-clock interrupts
2713 * during the time that the system is transitioning to full
2714 * system-idle state. This means that the timekeeping CPU must
2715 * invoke rcu_sysidle_force_exit() directly if it does anything
2716 * more than take a scheduling-clock interrupt.
2718 if (smp_processor_id() == tick_do_timer_cpu)
2721 /* Update system-idle state: We are clearly no longer fully idle! */
2722 rcu_sysidle_force_exit();
2726 * Check to see if the current CPU is idle. Note that usermode execution
2727 * does not count as idle. The caller must have disabled interrupts,
2728 * and must be running on tick_do_timer_cpu.
2730 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2731 unsigned long *maxj)
2735 struct rcu_dynticks *rdtp = rdp->dynticks;
2737 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2738 if (!tick_nohz_full_enabled())
2742 * If some other CPU has already reported non-idle, if this is
2743 * not the flavor of RCU that tracks sysidle state, or if this
2744 * is an offline or the timekeeping CPU, nothing to do.
2746 if (!*isidle || rdp->rsp != rcu_state_p ||
2747 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2749 /* Verify affinity of current kthread. */
2750 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2752 /* Pick up current idle and NMI-nesting counter and check. */
2753 cur = atomic_read(&rdtp->dynticks_idle);
2755 *isidle = false; /* We are not idle! */
2758 smp_mb(); /* Read counters before timestamps. */
2760 /* Pick up timestamps. */
2761 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2762 /* If this CPU entered idle more recently, update maxj timestamp. */
2763 if (ULONG_CMP_LT(*maxj, j))
2768 * Is this the flavor of RCU that is handling full-system idle?
2770 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2772 return rsp == rcu_state_p;
2776 * Return a delay in jiffies based on the number of CPUs, rcu_node
2777 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2778 * systems more time to transition to full-idle state in order to
2779 * avoid the cache thrashing that otherwise occur on the state variable.
2780 * Really small systems (less than a couple of tens of CPUs) should
2781 * instead use a single global atomically incremented counter, and later
2782 * versions of this will automatically reconfigure themselves accordingly.
2784 static unsigned long rcu_sysidle_delay(void)
2786 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2788 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2792 * Advance the full-system-idle state. This is invoked when all of
2793 * the non-timekeeping CPUs are idle.
2795 static void rcu_sysidle(unsigned long j)
2797 /* Check the current state. */
2798 switch (READ_ONCE(full_sysidle_state)) {
2799 case RCU_SYSIDLE_NOT:
2801 /* First time all are idle, so note a short idle period. */
2802 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2805 case RCU_SYSIDLE_SHORT:
2808 * Idle for a bit, time to advance to next state?
2809 * cmpxchg failure means race with non-idle, let them win.
2811 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2812 (void)cmpxchg(&full_sysidle_state,
2813 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2816 case RCU_SYSIDLE_LONG:
2819 * Do an additional check pass before advancing to full.
2820 * cmpxchg failure means race with non-idle, let them win.
2822 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2823 (void)cmpxchg(&full_sysidle_state,
2824 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2833 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2834 * back to the beginning.
2836 static void rcu_sysidle_cancel(void)
2839 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2840 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2844 * Update the sysidle state based on the results of a force-quiescent-state
2845 * scan of the CPUs' dyntick-idle state.
2847 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2848 unsigned long maxj, bool gpkt)
2850 if (rsp != rcu_state_p)
2851 return; /* Wrong flavor, ignore. */
2852 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2853 return; /* Running state machine from timekeeping CPU. */
2855 rcu_sysidle(maxj); /* More idle! */
2857 rcu_sysidle_cancel(); /* Idle is over. */
2861 * Wrapper for rcu_sysidle_report() when called from the grace-period
2862 * kthread's context.
2864 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2867 /* If there are no nohz_full= CPUs, no need to track this. */
2868 if (!tick_nohz_full_enabled())
2871 rcu_sysidle_report(rsp, isidle, maxj, true);
2874 /* Callback and function for forcing an RCU grace period. */
2875 struct rcu_sysidle_head {
2880 static void rcu_sysidle_cb(struct rcu_head *rhp)
2882 struct rcu_sysidle_head *rshp;
2885 * The following memory barrier is needed to replace the
2886 * memory barriers that would normally be in the memory
2889 smp_mb(); /* grace period precedes setting inuse. */
2891 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2892 WRITE_ONCE(rshp->inuse, 0);
2896 * Check to see if the system is fully idle, other than the timekeeping CPU.
2897 * The caller must have disabled interrupts. This is not intended to be
2898 * called unless tick_nohz_full_enabled().
2900 bool rcu_sys_is_idle(void)
2902 static struct rcu_sysidle_head rsh;
2903 int rss = READ_ONCE(full_sysidle_state);
2905 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2908 /* Handle small-system case by doing a full scan of CPUs. */
2909 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2910 int oldrss = rss - 1;
2913 * One pass to advance to each state up to _FULL.
2914 * Give up if any pass fails to advance the state.
2916 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2919 unsigned long maxj = jiffies - ULONG_MAX / 4;
2920 struct rcu_data *rdp;
2922 /* Scan all the CPUs looking for nonidle CPUs. */
2923 for_each_possible_cpu(cpu) {
2924 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2925 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2929 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2931 rss = READ_ONCE(full_sysidle_state);
2935 /* If this is the first observation of an idle period, record it. */
2936 if (rss == RCU_SYSIDLE_FULL) {
2937 rss = cmpxchg(&full_sysidle_state,
2938 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2939 return rss == RCU_SYSIDLE_FULL;
2942 smp_mb(); /* ensure rss load happens before later caller actions. */
2944 /* If already fully idle, tell the caller (in case of races). */
2945 if (rss == RCU_SYSIDLE_FULL_NOTED)
2949 * If we aren't there yet, and a grace period is not in flight,
2950 * initiate a grace period. Either way, tell the caller that
2951 * we are not there yet. We use an xchg() rather than an assignment
2952 * to make up for the memory barriers that would otherwise be
2953 * provided by the memory allocator.
2955 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2956 !rcu_gp_in_progress(rcu_state_p) &&
2957 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2958 call_rcu(&rsh.rh, rcu_sysidle_cb);
2963 * Initialize dynticks sysidle state for CPUs coming online.
2965 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2967 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2970 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2972 static void rcu_sysidle_enter(int irq)
2976 static void rcu_sysidle_exit(int irq)
2980 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2981 unsigned long *maxj)
2985 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2990 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2995 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2999 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3002 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3003 * grace-period kthread will do force_quiescent_state() processing?
3004 * The idea is to avoid waking up RCU core processing on such a
3005 * CPU unless the grace period has extended for too long.
3007 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3008 * CONFIG_RCU_NOCB_CPU CPUs.
3010 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3012 #ifdef CONFIG_NO_HZ_FULL
3013 if (tick_nohz_full_cpu(smp_processor_id()) &&
3014 (!rcu_gp_in_progress(rsp) ||
3015 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
3017 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3022 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3025 static void rcu_bind_gp_kthread(void)
3027 int __maybe_unused cpu;
3029 if (!tick_nohz_full_enabled())
3031 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3032 cpu = tick_do_timer_cpu;
3033 if (cpu >= 0 && cpu < nr_cpu_ids)
3034 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3035 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3036 housekeeping_affine(current);
3037 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3040 /* Record the current task on dyntick-idle entry. */
3041 static void rcu_dynticks_task_enter(void)
3043 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3044 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
3045 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3048 /* Record no current task on dyntick-idle exit. */
3049 static void rcu_dynticks_task_exit(void)
3051 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3052 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
3053 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */