1 /* SPDX-License-Identifier: GPL-2.0+ */
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions that provide either classic
5 * or preemptible semantics.
7 * Copyright Red Hat, 2009
8 * Copyright IBM Corporation, 2009
10 * Author: Ingo Molnar <mingo@elte.hu>
11 * Paul E. McKenney <paulmck@linux.ibm.com>
14 #include "../locking/rtmutex_common.h"
16 #ifdef CONFIG_RCU_NOCB_CPU
17 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
18 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
19 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
22 * Check the RCU kernel configuration parameters and print informative
23 * messages about anything out of the ordinary.
25 static void __init rcu_bootup_announce_oddness(void)
27 if (IS_ENABLED(CONFIG_RCU_TRACE))
28 pr_info("\tRCU event tracing is enabled.\n");
29 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
30 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
31 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
34 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
35 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
36 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
37 if (IS_ENABLED(CONFIG_PROVE_RCU))
38 pr_info("\tRCU lockdep checking is enabled.\n");
39 if (RCU_NUM_LVLS >= 4)
40 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
41 if (RCU_FANOUT_LEAF != 16)
42 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
44 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
45 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
47 if (nr_cpu_ids != NR_CPUS)
48 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
49 #ifdef CONFIG_RCU_BOOST
50 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
51 kthread_prio, CONFIG_RCU_BOOST_DELAY);
53 if (blimit != DEFAULT_RCU_BLIMIT)
54 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
55 if (qhimark != DEFAULT_RCU_QHIMARK)
56 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
57 if (qlowmark != DEFAULT_RCU_QLOMARK)
58 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
59 if (jiffies_till_first_fqs != ULONG_MAX)
60 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
61 if (jiffies_till_next_fqs != ULONG_MAX)
62 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
63 if (jiffies_till_sched_qs != ULONG_MAX)
64 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
65 if (rcu_kick_kthreads)
66 pr_info("\tKick kthreads if too-long grace period.\n");
67 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
68 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
70 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
72 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
74 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
76 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
77 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
78 pr_info("\tRCU debug extended QS entry/exit.\n");
79 rcupdate_announce_bootup_oddness();
82 #ifdef CONFIG_PREEMPT_RCU
84 static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
85 static void rcu_read_unlock_special(struct task_struct *t);
88 * Tell them what RCU they are running.
90 static void __init rcu_bootup_announce(void)
92 pr_info("Preemptible hierarchical RCU implementation.\n");
93 rcu_bootup_announce_oddness();
96 /* Flags for rcu_preempt_ctxt_queue() decision table. */
97 #define RCU_GP_TASKS 0x8
98 #define RCU_EXP_TASKS 0x4
99 #define RCU_GP_BLKD 0x2
100 #define RCU_EXP_BLKD 0x1
103 * Queues a task preempted within an RCU-preempt read-side critical
104 * section into the appropriate location within the ->blkd_tasks list,
105 * depending on the states of any ongoing normal and expedited grace
106 * periods. The ->gp_tasks pointer indicates which element the normal
107 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
108 * indicates which element the expedited grace period is waiting on (again,
109 * NULL if none). If a grace period is waiting on a given element in the
110 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
111 * adding a task to the tail of the list blocks any grace period that is
112 * already waiting on one of the elements. In contrast, adding a task
113 * to the head of the list won't block any grace period that is already
114 * waiting on one of the elements.
116 * This queuing is imprecise, and can sometimes make an ongoing grace
117 * period wait for a task that is not strictly speaking blocking it.
118 * Given the choice, we needlessly block a normal grace period rather than
119 * blocking an expedited grace period.
121 * Note that an endless sequence of expedited grace periods still cannot
122 * indefinitely postpone a normal grace period. Eventually, all of the
123 * fixed number of preempted tasks blocking the normal grace period that are
124 * not also blocking the expedited grace period will resume and complete
125 * their RCU read-side critical sections. At that point, the ->gp_tasks
126 * pointer will equal the ->exp_tasks pointer, at which point the end of
127 * the corresponding expedited grace period will also be the end of the
128 * normal grace period.
130 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
131 __releases(rnp->lock) /* But leaves rrupts disabled. */
133 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
134 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
135 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
136 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
137 struct task_struct *t = current;
139 raw_lockdep_assert_held_rcu_node(rnp);
140 WARN_ON_ONCE(rdp->mynode != rnp);
141 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
142 /* RCU better not be waiting on newly onlined CPUs! */
143 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
147 * Decide where to queue the newly blocked task. In theory,
148 * this could be an if-statement. In practice, when I tried
149 * that, it was quite messy.
151 switch (blkd_state) {
154 case RCU_EXP_TASKS + RCU_GP_BLKD:
156 case RCU_GP_TASKS + RCU_EXP_TASKS:
159 * Blocking neither GP, or first task blocking the normal
160 * GP but not blocking the already-waiting expedited GP.
161 * Queue at the head of the list to avoid unnecessarily
162 * blocking the already-waiting GPs.
164 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
169 case RCU_GP_BLKD + RCU_EXP_BLKD:
170 case RCU_GP_TASKS + RCU_EXP_BLKD:
171 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
172 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
175 * First task arriving that blocks either GP, or first task
176 * arriving that blocks the expedited GP (with the normal
177 * GP already waiting), or a task arriving that blocks
178 * both GPs with both GPs already waiting. Queue at the
179 * tail of the list to avoid any GP waiting on any of the
180 * already queued tasks that are not blocking it.
182 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
185 case RCU_EXP_TASKS + RCU_EXP_BLKD:
186 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
187 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
190 * Second or subsequent task blocking the expedited GP.
191 * The task either does not block the normal GP, or is the
192 * first task blocking the normal GP. Queue just after
193 * the first task blocking the expedited GP.
195 list_add(&t->rcu_node_entry, rnp->exp_tasks);
198 case RCU_GP_TASKS + RCU_GP_BLKD:
199 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
202 * Second or subsequent task blocking the normal GP.
203 * The task does not block the expedited GP. Queue just
204 * after the first task blocking the normal GP.
206 list_add(&t->rcu_node_entry, rnp->gp_tasks);
211 /* Yet another exercise in excessive paranoia. */
217 * We have now queued the task. If it was the first one to
218 * block either grace period, update the ->gp_tasks and/or
219 * ->exp_tasks pointers, respectively, to reference the newly
222 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
223 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
224 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
226 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
227 rnp->exp_tasks = &t->rcu_node_entry;
228 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
229 !(rnp->qsmask & rdp->grpmask));
230 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
231 !(rnp->expmask & rdp->grpmask));
232 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
235 * Report the quiescent state for the expedited GP. This expedited
236 * GP should not be able to end until we report, so there should be
237 * no need to check for a subsequent expedited GP. (Though we are
238 * still in a quiescent state in any case.)
240 if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
241 rcu_report_exp_rdp(rdp);
243 WARN_ON_ONCE(rdp->exp_deferred_qs);
247 * Record a preemptible-RCU quiescent state for the specified CPU.
248 * Note that this does not necessarily mean that the task currently running
249 * on the CPU is in a quiescent state: Instead, it means that the current
250 * grace period need not wait on any RCU read-side critical section that
251 * starts later on this CPU. It also means that if the current task is
252 * in an RCU read-side critical section, it has already added itself to
253 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
254 * current task, there might be any number of other tasks blocked while
255 * in an RCU read-side critical section.
257 * Callers to this function must disable preemption.
259 static void rcu_qs(void)
261 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
262 if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
263 trace_rcu_grace_period(TPS("rcu_preempt"),
264 __this_cpu_read(rcu_data.gp_seq),
266 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
267 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
268 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
273 * We have entered the scheduler, and the current task might soon be
274 * context-switched away from. If this task is in an RCU read-side
275 * critical section, we will no longer be able to rely on the CPU to
276 * record that fact, so we enqueue the task on the blkd_tasks list.
277 * The task will dequeue itself when it exits the outermost enclosing
278 * RCU read-side critical section. Therefore, the current grace period
279 * cannot be permitted to complete until the blkd_tasks list entries
280 * predating the current grace period drain, in other words, until
281 * rnp->gp_tasks becomes NULL.
283 * Caller must disable interrupts.
285 void rcu_note_context_switch(bool preempt)
287 struct task_struct *t = current;
288 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
289 struct rcu_node *rnp;
291 trace_rcu_utilization(TPS("Start context switch"));
292 lockdep_assert_irqs_disabled();
293 WARN_ON_ONCE(!preempt && t->rcu_read_lock_nesting > 0);
294 if (t->rcu_read_lock_nesting > 0 &&
295 !t->rcu_read_unlock_special.b.blocked) {
297 /* Possibly blocking in an RCU read-side critical section. */
299 raw_spin_lock_rcu_node(rnp);
300 t->rcu_read_unlock_special.b.blocked = true;
301 t->rcu_blocked_node = rnp;
304 * Verify the CPU's sanity, trace the preemption, and
305 * then queue the task as required based on the states
306 * of any ongoing and expedited grace periods.
308 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
309 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
310 trace_rcu_preempt_task(rcu_state.name,
312 (rnp->qsmask & rdp->grpmask)
314 : rcu_seq_snap(&rnp->gp_seq));
315 rcu_preempt_ctxt_queue(rnp, rdp);
317 rcu_preempt_deferred_qs(t);
321 * Either we were not in an RCU read-side critical section to
322 * begin with, or we have now recorded that critical section
323 * globally. Either way, we can now note a quiescent state
324 * for this CPU. Again, if we were in an RCU read-side critical
325 * section, and if that critical section was blocking the current
326 * grace period, then the fact that the task has been enqueued
327 * means that we continue to block the current grace period.
330 if (rdp->exp_deferred_qs)
331 rcu_report_exp_rdp(rdp);
332 trace_rcu_utilization(TPS("End context switch"));
334 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
337 * Check for preempted RCU readers blocking the current grace period
338 * for the specified rcu_node structure. If the caller needs a reliable
339 * answer, it must hold the rcu_node's ->lock.
341 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
343 return READ_ONCE(rnp->gp_tasks) != NULL;
346 /* Bias and limit values for ->rcu_read_lock_nesting. */
347 #define RCU_NEST_BIAS INT_MAX
348 #define RCU_NEST_NMAX (-INT_MAX / 2)
349 #define RCU_NEST_PMAX (INT_MAX / 2)
352 * Preemptible RCU implementation for rcu_read_lock().
353 * Just increment ->rcu_read_lock_nesting, shared state will be updated
356 void __rcu_read_lock(void)
358 current->rcu_read_lock_nesting++;
359 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
360 WARN_ON_ONCE(current->rcu_read_lock_nesting > RCU_NEST_PMAX);
361 barrier(); /* critical section after entry code. */
363 EXPORT_SYMBOL_GPL(__rcu_read_lock);
366 * Preemptible RCU implementation for rcu_read_unlock().
367 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
368 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
369 * invoke rcu_read_unlock_special() to clean up after a context switch
370 * in an RCU read-side critical section and other special cases.
372 void __rcu_read_unlock(void)
374 struct task_struct *t = current;
376 if (t->rcu_read_lock_nesting != 1) {
377 --t->rcu_read_lock_nesting;
379 barrier(); /* critical section before exit code. */
380 t->rcu_read_lock_nesting = -RCU_NEST_BIAS;
381 barrier(); /* assign before ->rcu_read_unlock_special load */
382 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
383 rcu_read_unlock_special(t);
384 barrier(); /* ->rcu_read_unlock_special load before assign */
385 t->rcu_read_lock_nesting = 0;
387 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
388 int rrln = t->rcu_read_lock_nesting;
390 WARN_ON_ONCE(rrln < 0 && rrln > RCU_NEST_NMAX);
393 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
396 * Advance a ->blkd_tasks-list pointer to the next entry, instead
397 * returning NULL if at the end of the list.
399 static struct list_head *rcu_next_node_entry(struct task_struct *t,
400 struct rcu_node *rnp)
402 struct list_head *np;
404 np = t->rcu_node_entry.next;
405 if (np == &rnp->blkd_tasks)
411 * Return true if the specified rcu_node structure has tasks that were
412 * preempted within an RCU read-side critical section.
414 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
416 return !list_empty(&rnp->blkd_tasks);
420 * Report deferred quiescent states. The deferral time can
421 * be quite short, for example, in the case of the call from
422 * rcu_read_unlock_special().
425 rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
430 struct list_head *np;
431 bool drop_boost_mutex = false;
432 struct rcu_data *rdp;
433 struct rcu_node *rnp;
434 union rcu_special special;
437 * If RCU core is waiting for this CPU to exit its critical section,
438 * report the fact that it has exited. Because irqs are disabled,
439 * t->rcu_read_unlock_special cannot change.
441 special = t->rcu_read_unlock_special;
442 rdp = this_cpu_ptr(&rcu_data);
443 if (!special.s && !rdp->exp_deferred_qs) {
444 local_irq_restore(flags);
447 t->rcu_read_unlock_special.b.deferred_qs = false;
448 if (special.b.need_qs) {
450 t->rcu_read_unlock_special.b.need_qs = false;
451 if (!t->rcu_read_unlock_special.s && !rdp->exp_deferred_qs) {
452 local_irq_restore(flags);
458 * Respond to a request by an expedited grace period for a
459 * quiescent state from this CPU. Note that requests from
460 * tasks are handled when removing the task from the
461 * blocked-tasks list below.
463 if (rdp->exp_deferred_qs) {
464 rcu_report_exp_rdp(rdp);
465 if (!t->rcu_read_unlock_special.s) {
466 local_irq_restore(flags);
471 /* Clean up if blocked during RCU read-side critical section. */
472 if (special.b.blocked) {
473 t->rcu_read_unlock_special.b.blocked = false;
476 * Remove this task from the list it blocked on. The task
477 * now remains queued on the rcu_node corresponding to the
478 * CPU it first blocked on, so there is no longer any need
479 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
481 rnp = t->rcu_blocked_node;
482 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
483 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
484 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
485 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
486 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
487 (!empty_norm || rnp->qsmask));
488 empty_exp = sync_rcu_preempt_exp_done(rnp);
489 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
490 np = rcu_next_node_entry(t, rnp);
491 list_del_init(&t->rcu_node_entry);
492 t->rcu_blocked_node = NULL;
493 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
494 rnp->gp_seq, t->pid);
495 if (&t->rcu_node_entry == rnp->gp_tasks)
496 WRITE_ONCE(rnp->gp_tasks, np);
497 if (&t->rcu_node_entry == rnp->exp_tasks)
499 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
500 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
501 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
502 if (&t->rcu_node_entry == rnp->boost_tasks)
503 rnp->boost_tasks = np;
507 * If this was the last task on the current list, and if
508 * we aren't waiting on any CPUs, report the quiescent state.
509 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
510 * so we must take a snapshot of the expedited state.
512 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
513 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
514 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
521 rcu_report_unblock_qs_rnp(rnp, flags);
523 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
527 * If this was the last task on the expedited lists,
528 * then we need to report up the rcu_node hierarchy.
530 if (!empty_exp && empty_exp_now)
531 rcu_report_exp_rnp(rnp, true);
533 /* Unboost if we were boosted. */
534 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
535 rt_mutex_futex_unlock(&rnp->boost_mtx);
538 local_irq_restore(flags);
543 * Is a deferred quiescent-state pending, and are we also not in
544 * an RCU read-side critical section? It is the caller's responsibility
545 * to ensure it is otherwise safe to report any deferred quiescent
546 * states. The reason for this is that it is safe to report a
547 * quiescent state during context switch even though preemption
548 * is disabled. This function cannot be expected to understand these
549 * nuances, so the caller must handle them.
551 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
553 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
554 READ_ONCE(t->rcu_read_unlock_special.s)) &&
555 t->rcu_read_lock_nesting <= 0;
559 * Report a deferred quiescent state if needed and safe to do so.
560 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
561 * not being in an RCU read-side critical section. The caller must
562 * evaluate safety in terms of interrupt, softirq, and preemption
565 static void rcu_preempt_deferred_qs(struct task_struct *t)
568 bool couldrecurse = t->rcu_read_lock_nesting >= 0;
570 if (!rcu_preempt_need_deferred_qs(t))
573 t->rcu_read_lock_nesting -= RCU_NEST_BIAS;
574 local_irq_save(flags);
575 rcu_preempt_deferred_qs_irqrestore(t, flags);
577 t->rcu_read_lock_nesting += RCU_NEST_BIAS;
581 * Minimal handler to give the scheduler a chance to re-evaluate.
583 static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
585 struct rcu_data *rdp;
587 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
588 rdp->defer_qs_iw_pending = false;
592 * Handle special cases during rcu_read_unlock(), such as needing to
593 * notify RCU core processing or task having blocked during the RCU
594 * read-side critical section.
596 static void rcu_read_unlock_special(struct task_struct *t)
599 bool preempt_bh_were_disabled =
600 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
601 bool irqs_were_disabled;
603 /* NMI handlers cannot block and cannot safely manipulate state. */
607 local_irq_save(flags);
608 irqs_were_disabled = irqs_disabled_flags(flags);
609 if (preempt_bh_were_disabled || irqs_were_disabled) {
611 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
612 struct rcu_node *rnp = rdp->mynode;
614 t->rcu_read_unlock_special.b.exp_hint = false;
615 exp = (t->rcu_blocked_node && t->rcu_blocked_node->exp_tasks) ||
616 (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
617 tick_nohz_full_cpu(rdp->cpu);
618 // Need to defer quiescent state until everything is enabled.
619 if (irqs_were_disabled && use_softirq &&
621 (exp && !t->rcu_read_unlock_special.b.deferred_qs))) {
622 // Using softirq, safe to awaken, and we get
623 // no help from enabling irqs, unlike bh/preempt.
624 raise_softirq_irqoff(RCU_SOFTIRQ);
626 // Enabling BH or preempt does reschedule, so...
627 // Also if no expediting or NO_HZ_FULL, slow is OK.
628 set_tsk_need_resched(current);
629 set_preempt_need_resched();
630 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
631 !rdp->defer_qs_iw_pending && exp) {
632 // Get scheduler to re-evaluate and call hooks.
633 // If !IRQ_WORK, FQS scan will eventually IPI.
634 init_irq_work(&rdp->defer_qs_iw,
635 rcu_preempt_deferred_qs_handler);
636 rdp->defer_qs_iw_pending = true;
637 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
640 t->rcu_read_unlock_special.b.deferred_qs = true;
641 local_irq_restore(flags);
644 WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
645 rcu_preempt_deferred_qs_irqrestore(t, flags);
649 * Check that the list of blocked tasks for the newly completed grace
650 * period is in fact empty. It is a serious bug to complete a grace
651 * period that still has RCU readers blocked! This function must be
652 * invoked -before- updating this rnp's ->gp_seq, and the rnp's ->lock
653 * must be held by the caller.
655 * Also, if there are blocked tasks on the list, they automatically
656 * block the newly created grace period, so set up ->gp_tasks accordingly.
658 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
660 struct task_struct *t;
662 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
663 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
664 dump_blkd_tasks(rnp, 10);
665 if (rcu_preempt_has_tasks(rnp) &&
666 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
667 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
668 t = container_of(rnp->gp_tasks, struct task_struct,
670 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
671 rnp->gp_seq, t->pid);
673 WARN_ON_ONCE(rnp->qsmask);
677 * Check for a quiescent state from the current CPU, including voluntary
678 * context switches for Tasks RCU. When a task blocks, the task is
679 * recorded in the corresponding CPU's rcu_node structure, which is checked
680 * elsewhere, hence this function need only check for quiescent states
681 * related to the current CPU, not to those related to tasks.
683 static void rcu_flavor_sched_clock_irq(int user)
685 struct task_struct *t = current;
687 if (user || rcu_is_cpu_rrupt_from_idle()) {
688 rcu_note_voluntary_context_switch(current);
690 if (t->rcu_read_lock_nesting > 0 ||
691 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
692 /* No QS, force context switch if deferred. */
693 if (rcu_preempt_need_deferred_qs(t)) {
694 set_tsk_need_resched(t);
695 set_preempt_need_resched();
697 } else if (rcu_preempt_need_deferred_qs(t)) {
698 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
700 } else if (!t->rcu_read_lock_nesting) {
701 rcu_qs(); /* Report immediate QS. */
705 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
706 if (t->rcu_read_lock_nesting > 0 &&
707 __this_cpu_read(rcu_data.core_needs_qs) &&
708 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
709 !t->rcu_read_unlock_special.b.need_qs &&
710 time_after(jiffies, rcu_state.gp_start + HZ))
711 t->rcu_read_unlock_special.b.need_qs = true;
715 * Check for a task exiting while in a preemptible-RCU read-side
716 * critical section, clean up if so. No need to issue warnings, as
717 * debug_check_no_locks_held() already does this if lockdep is enabled.
718 * Besides, if this function does anything other than just immediately
719 * return, there was a bug of some sort. Spewing warnings from this
720 * function is like as not to simply obscure important prior warnings.
724 struct task_struct *t = current;
726 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
727 t->rcu_read_lock_nesting = 1;
729 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
730 } else if (unlikely(t->rcu_read_lock_nesting)) {
731 t->rcu_read_lock_nesting = 1;
736 rcu_preempt_deferred_qs(current);
740 * Dump the blocked-tasks state, but limit the list dump to the
741 * specified number of elements.
744 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
748 struct list_head *lhp;
750 struct rcu_data *rdp;
751 struct rcu_node *rnp1;
753 raw_lockdep_assert_held_rcu_node(rnp);
754 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
755 __func__, rnp->grplo, rnp->grphi, rnp->level,
756 (long)rnp->gp_seq, (long)rnp->completedqs);
757 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
758 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
759 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
760 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
761 __func__, READ_ONCE(rnp->gp_tasks), rnp->boost_tasks,
763 pr_info("%s: ->blkd_tasks", __func__);
765 list_for_each(lhp, &rnp->blkd_tasks) {
771 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
772 rdp = per_cpu_ptr(&rcu_data, cpu);
773 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
774 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
776 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
777 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
781 #else /* #ifdef CONFIG_PREEMPT_RCU */
784 * Tell them what RCU they are running.
786 static void __init rcu_bootup_announce(void)
788 pr_info("Hierarchical RCU implementation.\n");
789 rcu_bootup_announce_oddness();
793 * Note a quiescent state for PREEMPT=n. Because we do not need to know
794 * how many quiescent states passed, just if there was at least one since
795 * the start of the grace period, this just sets a flag. The caller must
796 * have disabled preemption.
798 static void rcu_qs(void)
800 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
801 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
803 trace_rcu_grace_period(TPS("rcu_sched"),
804 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
805 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
806 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
808 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
809 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
813 * Register an urgently needed quiescent state. If there is an
814 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
815 * dyntick-idle quiescent state visible to other CPUs, which will in
816 * some cases serve for expedited as well as normal grace periods.
817 * Either way, register a lightweight quiescent state.
819 void rcu_all_qs(void)
823 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
826 /* Load rcu_urgent_qs before other flags. */
827 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
831 this_cpu_write(rcu_data.rcu_urgent_qs, false);
832 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
833 local_irq_save(flags);
834 rcu_momentary_dyntick_idle();
835 local_irq_restore(flags);
840 EXPORT_SYMBOL_GPL(rcu_all_qs);
843 * Note a PREEMPT=n context switch. The caller must have disabled interrupts.
845 void rcu_note_context_switch(bool preempt)
847 trace_rcu_utilization(TPS("Start context switch"));
849 /* Load rcu_urgent_qs before other flags. */
850 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
852 this_cpu_write(rcu_data.rcu_urgent_qs, false);
853 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
854 rcu_momentary_dyntick_idle();
856 rcu_tasks_qs(current);
858 trace_rcu_utilization(TPS("End context switch"));
860 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
863 * Because preemptible RCU does not exist, there are never any preempted
866 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
872 * Because there is no preemptible RCU, there can be no readers blocked.
874 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
880 * Because there is no preemptible RCU, there can be no deferred quiescent
883 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
887 static void rcu_preempt_deferred_qs(struct task_struct *t) { }
890 * Because there is no preemptible RCU, there can be no readers blocked,
891 * so there is no need to check for blocked tasks. So check only for
892 * bogus qsmask values.
894 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
896 WARN_ON_ONCE(rnp->qsmask);
900 * Check to see if this CPU is in a non-context-switch quiescent state,
901 * namely user mode and idle loop.
903 static void rcu_flavor_sched_clock_irq(int user)
905 if (user || rcu_is_cpu_rrupt_from_idle()) {
908 * Get here if this CPU took its interrupt from user
909 * mode or from the idle loop, and if this is not a
910 * nested interrupt. In this case, the CPU is in
911 * a quiescent state, so note it.
913 * No memory barrier is required here because rcu_qs()
914 * references only CPU-local variables that other CPUs
915 * neither access nor modify, at least not while the
916 * corresponding CPU is online.
924 * Because preemptible RCU does not exist, tasks cannot possibly exit
925 * while in preemptible RCU read-side critical sections.
932 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
935 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
937 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
940 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
943 * If boosting, set rcuc kthreads to realtime priority.
945 static void rcu_cpu_kthread_setup(unsigned int cpu)
947 #ifdef CONFIG_RCU_BOOST
948 struct sched_param sp;
950 sp.sched_priority = kthread_prio;
951 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
952 #endif /* #ifdef CONFIG_RCU_BOOST */
955 #ifdef CONFIG_RCU_BOOST
958 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
959 * or ->boost_tasks, advancing the pointer to the next task in the
962 * Note that irqs must be enabled: boosting the task can block.
963 * Returns 1 if there are more tasks needing to be boosted.
965 static int rcu_boost(struct rcu_node *rnp)
968 struct task_struct *t;
969 struct list_head *tb;
971 if (READ_ONCE(rnp->exp_tasks) == NULL &&
972 READ_ONCE(rnp->boost_tasks) == NULL)
973 return 0; /* Nothing left to boost. */
975 raw_spin_lock_irqsave_rcu_node(rnp, flags);
978 * Recheck under the lock: all tasks in need of boosting
979 * might exit their RCU read-side critical sections on their own.
981 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
982 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
987 * Preferentially boost tasks blocking expedited grace periods.
988 * This cannot starve the normal grace periods because a second
989 * expedited grace period must boost all blocked tasks, including
990 * those blocking the pre-existing normal grace period.
992 if (rnp->exp_tasks != NULL)
995 tb = rnp->boost_tasks;
998 * We boost task t by manufacturing an rt_mutex that appears to
999 * be held by task t. We leave a pointer to that rt_mutex where
1000 * task t can find it, and task t will release the mutex when it
1001 * exits its outermost RCU read-side critical section. Then
1002 * simply acquiring this artificial rt_mutex will boost task
1003 * t's priority. (Thanks to tglx for suggesting this approach!)
1005 * Note that task t must acquire rnp->lock to remove itself from
1006 * the ->blkd_tasks list, which it will do from exit() if from
1007 * nowhere else. We therefore are guaranteed that task t will
1008 * stay around at least until we drop rnp->lock. Note that
1009 * rnp->lock also resolves races between our priority boosting
1010 * and task t's exiting its outermost RCU read-side critical
1013 t = container_of(tb, struct task_struct, rcu_node_entry);
1014 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1015 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1016 /* Lock only for side effect: boosts task t's priority. */
1017 rt_mutex_lock(&rnp->boost_mtx);
1018 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1020 return READ_ONCE(rnp->exp_tasks) != NULL ||
1021 READ_ONCE(rnp->boost_tasks) != NULL;
1025 * Priority-boosting kthread, one per leaf rcu_node.
1027 static int rcu_boost_kthread(void *arg)
1029 struct rcu_node *rnp = (struct rcu_node *)arg;
1033 trace_rcu_utilization(TPS("Start boost kthread@init"));
1035 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1036 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1037 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1038 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1039 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1040 more2boost = rcu_boost(rnp);
1046 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1047 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1048 schedule_timeout_interruptible(2);
1049 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1054 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1059 * Check to see if it is time to start boosting RCU readers that are
1060 * blocking the current grace period, and, if so, tell the per-rcu_node
1061 * kthread to start boosting them. If there is an expedited grace
1062 * period in progress, it is always time to boost.
1064 * The caller must hold rnp->lock, which this function releases.
1065 * The ->boost_kthread_task is immortal, so we don't need to worry
1066 * about it going away.
1068 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1069 __releases(rnp->lock)
1071 raw_lockdep_assert_held_rcu_node(rnp);
1072 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1073 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1076 if (rnp->exp_tasks != NULL ||
1077 (rnp->gp_tasks != NULL &&
1078 rnp->boost_tasks == NULL &&
1080 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1081 if (rnp->exp_tasks == NULL)
1082 rnp->boost_tasks = rnp->gp_tasks;
1083 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1084 rcu_wake_cond(rnp->boost_kthread_task,
1085 rnp->boost_kthread_status);
1087 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1092 * Is the current CPU running the RCU-callbacks kthread?
1093 * Caller must have preemption disabled.
1095 static bool rcu_is_callbacks_kthread(void)
1097 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1100 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1103 * Do priority-boost accounting for the start of a new grace period.
1105 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1107 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1111 * Create an RCU-boost kthread for the specified node if one does not
1112 * already exist. We only create this kthread for preemptible RCU.
1113 * Returns zero if all is well, a negated errno otherwise.
1115 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1117 int rnp_index = rnp - rcu_get_root();
1118 unsigned long flags;
1119 struct sched_param sp;
1120 struct task_struct *t;
1122 if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1125 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1128 rcu_state.boost = 1;
1130 if (rnp->boost_kthread_task != NULL)
1133 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1134 "rcub/%d", rnp_index);
1135 if (WARN_ON_ONCE(IS_ERR(t)))
1138 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1139 rnp->boost_kthread_task = t;
1140 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1141 sp.sched_priority = kthread_prio;
1142 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1143 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1147 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1148 * served by the rcu_node in question. The CPU hotplug lock is still
1149 * held, so the value of rnp->qsmaskinit will be stable.
1151 * We don't include outgoingcpu in the affinity set, use -1 if there is
1152 * no outgoing CPU. If there are no CPUs left in the affinity set,
1153 * this function allows the kthread to execute on any CPU.
1155 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1157 struct task_struct *t = rnp->boost_kthread_task;
1158 unsigned long mask = rcu_rnp_online_cpus(rnp);
1164 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1166 for_each_leaf_node_possible_cpu(rnp, cpu)
1167 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1169 cpumask_set_cpu(cpu, cm);
1170 if (cpumask_weight(cm) == 0)
1172 set_cpus_allowed_ptr(t, cm);
1173 free_cpumask_var(cm);
1177 * Spawn boost kthreads -- called as soon as the scheduler is running.
1179 static void __init rcu_spawn_boost_kthreads(void)
1181 struct rcu_node *rnp;
1183 rcu_for_each_leaf_node(rnp)
1184 rcu_spawn_one_boost_kthread(rnp);
1187 static void rcu_prepare_kthreads(int cpu)
1189 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1190 struct rcu_node *rnp = rdp->mynode;
1192 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1193 if (rcu_scheduler_fully_active)
1194 rcu_spawn_one_boost_kthread(rnp);
1197 #else /* #ifdef CONFIG_RCU_BOOST */
1199 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1200 __releases(rnp->lock)
1202 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1205 static bool rcu_is_callbacks_kthread(void)
1210 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1214 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1218 static void __init rcu_spawn_boost_kthreads(void)
1222 static void rcu_prepare_kthreads(int cpu)
1226 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1228 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1231 * Check to see if any future non-offloaded RCU-related work will need
1232 * to be done by the current CPU, even if none need be done immediately,
1233 * returning 1 if so. This function is part of the RCU implementation;
1234 * it is -not- an exported member of the RCU API.
1236 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1237 * CPU has RCU callbacks queued.
1239 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1241 *nextevt = KTIME_MAX;
1242 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1243 !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist);
1247 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1250 static void rcu_cleanup_after_idle(void)
1255 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1258 static void rcu_prepare_for_idle(void)
1262 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1265 * This code is invoked when a CPU goes idle, at which point we want
1266 * to have the CPU do everything required for RCU so that it can enter
1267 * the energy-efficient dyntick-idle mode. This is handled by a
1268 * state machine implemented by rcu_prepare_for_idle() below.
1270 * The following three proprocessor symbols control this state machine:
1272 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1273 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1274 * is sized to be roughly one RCU grace period. Those energy-efficiency
1275 * benchmarkers who might otherwise be tempted to set this to a large
1276 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1277 * system. And if you are -that- concerned about energy efficiency,
1278 * just power the system down and be done with it!
1279 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1280 * permitted to sleep in dyntick-idle mode with only lazy RCU
1281 * callbacks pending. Setting this too high can OOM your system.
1283 * The values below work well in practice. If future workloads require
1284 * adjustment, they can be converted into kernel config parameters, though
1285 * making the state machine smarter might be a better option.
1287 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1288 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1290 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1291 module_param(rcu_idle_gp_delay, int, 0644);
1292 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1293 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1296 * Try to advance callbacks on the current CPU, but only if it has been
1297 * awhile since the last time we did so. Afterwards, if there are any
1298 * callbacks ready for immediate invocation, return true.
1300 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1302 bool cbs_ready = false;
1303 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1304 struct rcu_node *rnp;
1306 /* Exit early if we advanced recently. */
1307 if (jiffies == rdp->last_advance_all)
1309 rdp->last_advance_all = jiffies;
1314 * Don't bother checking unless a grace period has
1315 * completed since we last checked and there are
1316 * callbacks not yet ready to invoke.
1318 if ((rcu_seq_completed_gp(rdp->gp_seq,
1319 rcu_seq_current(&rnp->gp_seq)) ||
1320 unlikely(READ_ONCE(rdp->gpwrap))) &&
1321 rcu_segcblist_pend_cbs(&rdp->cblist))
1322 note_gp_changes(rdp);
1324 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1330 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1331 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1332 * caller to set the timeout based on whether or not there are non-lazy
1335 * The caller must have disabled interrupts.
1337 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1339 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1342 lockdep_assert_irqs_disabled();
1344 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1345 if (rcu_segcblist_empty(&rdp->cblist) ||
1346 rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist)) {
1347 *nextevt = KTIME_MAX;
1351 /* Attempt to advance callbacks. */
1352 if (rcu_try_advance_all_cbs()) {
1353 /* Some ready to invoke, so initiate later invocation. */
1357 rdp->last_accelerate = jiffies;
1359 /* Request timer delay depending on laziness, and round. */
1360 rdp->all_lazy = !rcu_segcblist_n_nonlazy_cbs(&rdp->cblist);
1361 if (rdp->all_lazy) {
1362 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1364 dj = round_up(rcu_idle_gp_delay + jiffies,
1365 rcu_idle_gp_delay) - jiffies;
1367 *nextevt = basemono + dj * TICK_NSEC;
1372 * Prepare a CPU for idle from an RCU perspective. The first major task
1373 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1374 * The second major task is to check to see if a non-lazy callback has
1375 * arrived at a CPU that previously had only lazy callbacks. The third
1376 * major task is to accelerate (that is, assign grace-period numbers to)
1377 * any recently arrived callbacks.
1379 * The caller must have disabled interrupts.
1381 static void rcu_prepare_for_idle(void)
1384 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1385 struct rcu_node *rnp;
1388 lockdep_assert_irqs_disabled();
1389 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1392 /* Handle nohz enablement switches conservatively. */
1393 tne = READ_ONCE(tick_nohz_active);
1394 if (tne != rdp->tick_nohz_enabled_snap) {
1395 if (!rcu_segcblist_empty(&rdp->cblist))
1396 invoke_rcu_core(); /* force nohz to see update. */
1397 rdp->tick_nohz_enabled_snap = tne;
1404 * If a non-lazy callback arrived at a CPU having only lazy
1405 * callbacks, invoke RCU core for the side-effect of recalculating
1406 * idle duration on re-entry to idle.
1408 if (rdp->all_lazy && rcu_segcblist_n_nonlazy_cbs(&rdp->cblist)) {
1409 rdp->all_lazy = false;
1415 * If we have not yet accelerated this jiffy, accelerate all
1416 * callbacks on this CPU.
1418 if (rdp->last_accelerate == jiffies)
1420 rdp->last_accelerate = jiffies;
1421 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1423 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1424 needwake = rcu_accelerate_cbs(rnp, rdp);
1425 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1427 rcu_gp_kthread_wake();
1432 * Clean up for exit from idle. Attempt to advance callbacks based on
1433 * any grace periods that elapsed while the CPU was idle, and if any
1434 * callbacks are now ready to invoke, initiate invocation.
1436 static void rcu_cleanup_after_idle(void)
1438 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1440 lockdep_assert_irqs_disabled();
1441 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1443 if (rcu_try_advance_all_cbs())
1447 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1449 #ifdef CONFIG_RCU_NOCB_CPU
1452 * Offload callback processing from the boot-time-specified set of CPUs
1453 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1454 * created that pull the callbacks from the corresponding CPU, wait for
1455 * a grace period to elapse, and invoke the callbacks. These kthreads
1456 * are organized into GP kthreads, which manage incoming callbacks, wait for
1457 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1458 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1459 * do a wake_up() on their GP kthread when they insert a callback into any
1460 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1461 * in which case each kthread actively polls its CPU. (Which isn't so great
1462 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1464 * This is intended to be used in conjunction with Frederic Weisbecker's
1465 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1466 * running CPU-bound user-mode computations.
1468 * Offloading of callbacks can also be used as an energy-efficiency
1469 * measure because CPUs with no RCU callbacks queued are more aggressive
1470 * about entering dyntick-idle mode.
1475 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1476 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1477 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1478 * given, a warning is emitted and all CPUs are offloaded.
1480 static int __init rcu_nocb_setup(char *str)
1482 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1483 if (!strcasecmp(str, "all"))
1484 cpumask_setall(rcu_nocb_mask);
1486 if (cpulist_parse(str, rcu_nocb_mask)) {
1487 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1488 cpumask_setall(rcu_nocb_mask);
1492 __setup("rcu_nocbs=", rcu_nocb_setup);
1494 static int __init parse_rcu_nocb_poll(char *arg)
1496 rcu_nocb_poll = true;
1499 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1502 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1503 * After all, the main point of bypassing is to avoid lock contention
1504 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1506 int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1507 module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1510 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1511 * lock isn't immediately available, increment ->nocb_lock_contended to
1512 * flag the contention.
1514 static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1516 lockdep_assert_irqs_disabled();
1517 if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1519 atomic_inc(&rdp->nocb_lock_contended);
1520 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1521 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1522 raw_spin_lock(&rdp->nocb_bypass_lock);
1523 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1524 atomic_dec(&rdp->nocb_lock_contended);
1528 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1529 * not contended. Please note that this is extremely special-purpose,
1530 * relying on the fact that at most two kthreads and one CPU contend for
1531 * this lock, and also that the two kthreads are guaranteed to have frequent
1532 * grace-period-duration time intervals between successive acquisitions
1533 * of the lock. This allows us to use an extremely simple throttling
1534 * mechanism, and further to apply it only to the CPU doing floods of
1535 * call_rcu() invocations. Don't try this at home!
1537 static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1539 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1540 while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1545 * Conditionally acquire the specified rcu_data structure's
1546 * ->nocb_bypass_lock.
1548 static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1550 lockdep_assert_irqs_disabled();
1551 return raw_spin_trylock(&rdp->nocb_bypass_lock);
1555 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1557 static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1559 lockdep_assert_irqs_disabled();
1560 raw_spin_unlock(&rdp->nocb_bypass_lock);
1564 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1565 * if it corresponds to a no-CBs CPU.
1567 static void rcu_nocb_lock(struct rcu_data *rdp)
1569 lockdep_assert_irqs_disabled();
1570 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1572 raw_spin_lock(&rdp->nocb_lock);
1576 * Release the specified rcu_data structure's ->nocb_lock, but only
1577 * if it corresponds to a no-CBs CPU.
1579 static void rcu_nocb_unlock(struct rcu_data *rdp)
1581 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1582 lockdep_assert_irqs_disabled();
1583 raw_spin_unlock(&rdp->nocb_lock);
1588 * Release the specified rcu_data structure's ->nocb_lock and restore
1589 * interrupts, but only if it corresponds to a no-CBs CPU.
1591 static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1592 unsigned long flags)
1594 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1595 lockdep_assert_irqs_disabled();
1596 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1598 local_irq_restore(flags);
1602 /* Lockdep check that ->cblist may be safely accessed. */
1603 static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1605 lockdep_assert_irqs_disabled();
1606 if (rcu_segcblist_is_offloaded(&rdp->cblist) &&
1607 cpu_online(rdp->cpu))
1608 lockdep_assert_held(&rdp->nocb_lock);
1612 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1615 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1620 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1622 return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1625 static void rcu_init_one_nocb(struct rcu_node *rnp)
1627 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1628 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1631 /* Is the specified CPU a no-CBs CPU? */
1632 bool rcu_is_nocb_cpu(int cpu)
1634 if (cpumask_available(rcu_nocb_mask))
1635 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1640 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1641 * and this function releases it.
1643 static void wake_nocb_gp(struct rcu_data *rdp, bool force,
1644 unsigned long flags)
1645 __releases(rdp->nocb_lock)
1647 bool needwake = false;
1648 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1650 lockdep_assert_held(&rdp->nocb_lock);
1651 if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1652 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1653 TPS("AlreadyAwake"));
1654 rcu_nocb_unlock_irqrestore(rdp, flags);
1657 del_timer(&rdp->nocb_timer);
1658 rcu_nocb_unlock_irqrestore(rdp, flags);
1659 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1660 if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1661 WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1663 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1665 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1667 wake_up_process(rdp_gp->nocb_gp_kthread);
1671 * Arrange to wake the GP kthread for this NOCB group at some future
1672 * time when it is safe to do so.
1674 static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1677 if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1678 mod_timer(&rdp->nocb_timer, jiffies + 1);
1679 if (rdp->nocb_defer_wakeup < waketype)
1680 WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1681 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1685 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1686 * However, if there is a callback to be enqueued and if ->nocb_bypass
1687 * proves to be initially empty, just return false because the no-CB GP
1688 * kthread may need to be awakened in this case.
1690 * Note that this function always returns true if rhp is NULL.
1692 static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1695 struct rcu_cblist rcl;
1697 WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp->cblist));
1698 rcu_lockdep_assert_cblist_protected(rdp);
1699 lockdep_assert_held(&rdp->nocb_bypass_lock);
1700 if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1701 raw_spin_unlock(&rdp->nocb_bypass_lock);
1704 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1706 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1707 rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1708 rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1709 WRITE_ONCE(rdp->nocb_bypass_first, j);
1710 rcu_nocb_bypass_unlock(rdp);
1715 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1716 * However, if there is a callback to be enqueued and if ->nocb_bypass
1717 * proves to be initially empty, just return false because the no-CB GP
1718 * kthread may need to be awakened in this case.
1720 * Note that this function always returns true if rhp is NULL.
1722 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1725 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1727 rcu_lockdep_assert_cblist_protected(rdp);
1728 rcu_nocb_bypass_lock(rdp);
1729 return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1733 * If the ->nocb_bypass_lock is immediately available, flush the
1734 * ->nocb_bypass queue into ->cblist.
1736 static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1738 rcu_lockdep_assert_cblist_protected(rdp);
1739 if (!rcu_segcblist_is_offloaded(&rdp->cblist) ||
1740 !rcu_nocb_bypass_trylock(rdp))
1742 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1746 * See whether it is appropriate to use the ->nocb_bypass list in order
1747 * to control contention on ->nocb_lock. A limited number of direct
1748 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1749 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1750 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1751 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1752 * used if ->cblist is empty, because otherwise callbacks can be stranded
1753 * on ->nocb_bypass because we cannot count on the current CPU ever again
1754 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1755 * non-empty, the corresponding no-CBs grace-period kthread must not be
1756 * in an indefinite sleep state.
1758 * Finally, it is not permitted to use the bypass during early boot,
1759 * as doing so would confuse the auto-initialization code. Besides
1760 * which, there is no point in worrying about lock contention while
1761 * there is only one CPU in operation.
1763 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1764 bool *was_alldone, unsigned long flags)
1767 unsigned long cur_gp_seq;
1768 unsigned long j = jiffies;
1769 long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1771 if (!rcu_segcblist_is_offloaded(&rdp->cblist)) {
1772 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1773 return false; /* Not offloaded, no bypassing. */
1775 lockdep_assert_irqs_disabled();
1777 // Don't use ->nocb_bypass during early boot.
1778 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1780 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1781 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1785 // If we have advanced to a new jiffy, reset counts to allow
1786 // moving back from ->nocb_bypass to ->cblist.
1787 if (j == rdp->nocb_nobypass_last) {
1788 c = rdp->nocb_nobypass_count + 1;
1790 WRITE_ONCE(rdp->nocb_nobypass_last, j);
1791 c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1792 if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1793 nocb_nobypass_lim_per_jiffy))
1795 else if (c > nocb_nobypass_lim_per_jiffy)
1796 c = nocb_nobypass_lim_per_jiffy;
1798 WRITE_ONCE(rdp->nocb_nobypass_count, c);
1800 // If there hasn't yet been all that many ->cblist enqueues
1801 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1802 // ->nocb_bypass first.
1803 if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1805 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1807 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1809 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1810 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1811 return false; // Caller must enqueue the callback.
1814 // If ->nocb_bypass has been used too long or is too full,
1815 // flush ->nocb_bypass to ->cblist.
1816 if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1819 if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1820 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1822 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1824 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1825 return false; // Caller must enqueue the callback.
1827 if (j != rdp->nocb_gp_adv_time &&
1828 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1829 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1830 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1831 rdp->nocb_gp_adv_time = j;
1833 rcu_nocb_unlock_irqrestore(rdp, flags);
1834 return true; // Callback already enqueued.
1837 // We need to use the bypass.
1838 rcu_nocb_wait_contended(rdp);
1839 rcu_nocb_bypass_lock(rdp);
1840 ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1841 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1842 rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1844 WRITE_ONCE(rdp->nocb_bypass_first, j);
1845 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1847 rcu_nocb_bypass_unlock(rdp);
1848 smp_mb(); /* Order enqueue before wake. */
1850 local_irq_restore(flags);
1852 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1853 rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1854 if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1855 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1856 TPS("FirstBQwake"));
1857 __call_rcu_nocb_wake(rdp, true, flags);
1859 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1860 TPS("FirstBQnoWake"));
1861 rcu_nocb_unlock_irqrestore(rdp, flags);
1864 return true; // Callback already enqueued.
1868 * Awaken the no-CBs grace-period kthead if needed, either due to it
1869 * legitimately being asleep or due to overload conditions.
1871 * If warranted, also wake up the kthread servicing this CPUs queues.
1873 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1874 unsigned long flags)
1875 __releases(rdp->nocb_lock)
1877 unsigned long cur_gp_seq;
1880 struct task_struct *t;
1882 // If we are being polled or there is no kthread, just leave.
1883 t = READ_ONCE(rdp->nocb_gp_kthread);
1884 if (rcu_nocb_poll || !t) {
1885 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1886 TPS("WakeNotPoll"));
1887 rcu_nocb_unlock_irqrestore(rdp, flags);
1890 // Need to actually to a wakeup.
1891 len = rcu_segcblist_n_cbs(&rdp->cblist);
1893 rdp->qlen_last_fqs_check = len;
1894 if (!irqs_disabled_flags(flags)) {
1895 /* ... if queue was empty ... */
1896 wake_nocb_gp(rdp, false, flags);
1897 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1900 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1901 TPS("WakeEmptyIsDeferred"));
1902 rcu_nocb_unlock_irqrestore(rdp, flags);
1904 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1905 /* ... or if many callbacks queued. */
1906 rdp->qlen_last_fqs_check = len;
1908 if (j != rdp->nocb_gp_adv_time &&
1909 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1910 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1911 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1912 rdp->nocb_gp_adv_time = j;
1914 smp_mb(); /* Enqueue before timer_pending(). */
1915 if ((rdp->nocb_cb_sleep ||
1916 !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1917 !timer_pending(&rdp->nocb_bypass_timer))
1918 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1919 TPS("WakeOvfIsDeferred"));
1920 rcu_nocb_unlock_irqrestore(rdp, flags);
1922 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1923 rcu_nocb_unlock_irqrestore(rdp, flags);
1928 /* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1929 static void do_nocb_bypass_wakeup_timer(struct timer_list *t)
1931 unsigned long flags;
1932 struct rcu_data *rdp = from_timer(rdp, t, nocb_bypass_timer);
1934 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1935 rcu_nocb_lock_irqsave(rdp, flags);
1936 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1937 __call_rcu_nocb_wake(rdp, true, flags);
1941 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1942 * or for grace periods to end.
1944 static void nocb_gp_wait(struct rcu_data *my_rdp)
1946 bool bypass = false;
1948 int __maybe_unused cpu = my_rdp->cpu;
1949 unsigned long cur_gp_seq;
1950 unsigned long flags;
1951 bool gotcbs = false;
1952 unsigned long j = jiffies;
1953 bool needwait_gp = false; // This prevents actual uninitialized use.
1956 struct rcu_data *rdp;
1957 struct rcu_node *rnp;
1958 unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
1961 * Each pass through the following loop checks for CBs and for the
1962 * nearest grace period (if any) to wait for next. The CB kthreads
1963 * and the global grace-period kthread are awakened if needed.
1965 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1966 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
1967 rcu_nocb_lock_irqsave(rdp, flags);
1968 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1970 (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
1971 bypass_ncbs > 2 * qhimark)) {
1972 // Bypass full or old, so flush it.
1973 (void)rcu_nocb_try_flush_bypass(rdp, j);
1974 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1975 } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
1976 rcu_nocb_unlock_irqrestore(rdp, flags);
1977 continue; /* No callbacks here, try next. */
1980 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1985 if (bypass) { // Avoid race with first bypass CB.
1986 WRITE_ONCE(my_rdp->nocb_defer_wakeup,
1988 del_timer(&my_rdp->nocb_timer);
1990 // Advance callbacks if helpful and low contention.
1991 needwake_gp = false;
1992 if (!rcu_segcblist_restempty(&rdp->cblist,
1993 RCU_NEXT_READY_TAIL) ||
1994 (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1995 rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
1996 raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
1997 needwake_gp = rcu_advance_cbs(rnp, rdp);
1998 raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
2000 // Need to wait on some grace period?
2001 WARN_ON_ONCE(!rcu_segcblist_restempty(&rdp->cblist,
2002 RCU_NEXT_READY_TAIL));
2003 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
2005 ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
2006 wait_gp_seq = cur_gp_seq;
2008 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2011 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2012 needwake = rdp->nocb_cb_sleep;
2013 WRITE_ONCE(rdp->nocb_cb_sleep, false);
2014 smp_mb(); /* CB invocation -after- GP end. */
2018 rcu_nocb_unlock_irqrestore(rdp, flags);
2020 swake_up_one(&rdp->nocb_cb_wq);
2024 rcu_gp_kthread_wake();
2027 my_rdp->nocb_gp_bypass = bypass;
2028 my_rdp->nocb_gp_gp = needwait_gp;
2029 my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
2030 if (bypass && !rcu_nocb_poll) {
2031 // At least one child with non-empty ->nocb_bypass, so set
2032 // timer in order to avoid stranding its callbacks.
2033 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2034 mod_timer(&my_rdp->nocb_bypass_timer, j + 2);
2035 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2037 if (rcu_nocb_poll) {
2038 /* Polling, so trace if first poll in the series. */
2040 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2041 schedule_timeout_interruptible(1);
2042 } else if (!needwait_gp) {
2043 /* Wait for callbacks to appear. */
2044 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2045 swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2046 !READ_ONCE(my_rdp->nocb_gp_sleep));
2047 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2049 rnp = my_rdp->mynode;
2050 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2051 swait_event_interruptible_exclusive(
2052 rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2053 rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2054 !READ_ONCE(my_rdp->nocb_gp_sleep));
2055 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2057 if (!rcu_nocb_poll) {
2058 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2060 del_timer(&my_rdp->nocb_bypass_timer);
2061 WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2062 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2064 my_rdp->nocb_gp_seq = -1;
2065 WARN_ON(signal_pending(current));
2069 * No-CBs grace-period-wait kthread. There is one of these per group
2070 * of CPUs, but only once at least one CPU in that group has come online
2071 * at least once since boot. This kthread checks for newly posted
2072 * callbacks from any of the CPUs it is responsible for, waits for a
2073 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2074 * that then have callback-invocation work to do.
2076 static int rcu_nocb_gp_kthread(void *arg)
2078 struct rcu_data *rdp = arg;
2081 WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2083 cond_resched_tasks_rcu_qs();
2089 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2090 * then, if there are no more, wait for more to appear.
2092 static void nocb_cb_wait(struct rcu_data *rdp)
2094 unsigned long cur_gp_seq;
2095 unsigned long flags;
2096 bool needwake_gp = false;
2097 struct rcu_node *rnp = rdp->mynode;
2099 local_irq_save(flags);
2100 rcu_momentary_dyntick_idle();
2101 local_irq_restore(flags);
2105 lockdep_assert_irqs_enabled();
2106 rcu_nocb_lock_irqsave(rdp, flags);
2107 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2108 rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2109 raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2110 needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2111 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2113 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2114 rcu_nocb_unlock_irqrestore(rdp, flags);
2116 rcu_gp_kthread_wake();
2120 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2121 WRITE_ONCE(rdp->nocb_cb_sleep, true);
2122 rcu_nocb_unlock_irqrestore(rdp, flags);
2124 rcu_gp_kthread_wake();
2125 swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2126 !READ_ONCE(rdp->nocb_cb_sleep));
2127 if (!smp_load_acquire(&rdp->nocb_cb_sleep)) { /* VVV */
2128 /* ^^^ Ensure CB invocation follows _sleep test. */
2131 WARN_ON(signal_pending(current));
2132 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2136 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2137 * nocb_cb_wait() to do the dirty work.
2139 static int rcu_nocb_cb_kthread(void *arg)
2141 struct rcu_data *rdp = arg;
2143 // Each pass through this loop does one callback batch, and,
2144 // if there are no more ready callbacks, waits for them.
2147 cond_resched_tasks_rcu_qs();
2152 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2153 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2155 return READ_ONCE(rdp->nocb_defer_wakeup);
2158 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2159 static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2161 unsigned long flags;
2164 rcu_nocb_lock_irqsave(rdp, flags);
2165 if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2166 rcu_nocb_unlock_irqrestore(rdp, flags);
2169 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2170 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2171 wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2172 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2175 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2176 static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2178 struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2180 do_nocb_deferred_wakeup_common(rdp);
2184 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2185 * This means we do an inexact common-case check. Note that if
2186 * we miss, ->nocb_timer will eventually clean things up.
2188 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2190 if (rcu_nocb_need_deferred_wakeup(rdp))
2191 do_nocb_deferred_wakeup_common(rdp);
2194 void rcu_nocb_flush_deferred_wakeup(void)
2196 do_nocb_deferred_wakeup(this_cpu_ptr(&rcu_data));
2199 void __init rcu_init_nohz(void)
2202 bool need_rcu_nocb_mask = false;
2203 struct rcu_data *rdp;
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 (!cpumask_available(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");
2216 if (!cpumask_available(rcu_nocb_mask))
2219 #if defined(CONFIG_NO_HZ_FULL)
2220 if (tick_nohz_full_running)
2221 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2222 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2224 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2225 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2226 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2229 if (cpumask_empty(rcu_nocb_mask))
2230 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2232 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2233 cpumask_pr_args(rcu_nocb_mask));
2235 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2237 for_each_cpu(cpu, rcu_nocb_mask) {
2238 rdp = per_cpu_ptr(&rcu_data, cpu);
2239 if (rcu_segcblist_empty(&rdp->cblist))
2240 rcu_segcblist_init(&rdp->cblist);
2241 rcu_segcblist_offload(&rdp->cblist);
2243 rcu_organize_nocb_kthreads();
2246 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2247 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2249 init_swait_queue_head(&rdp->nocb_cb_wq);
2250 init_swait_queue_head(&rdp->nocb_gp_wq);
2251 raw_spin_lock_init(&rdp->nocb_lock);
2252 raw_spin_lock_init(&rdp->nocb_bypass_lock);
2253 raw_spin_lock_init(&rdp->nocb_gp_lock);
2254 timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2255 timer_setup(&rdp->nocb_bypass_timer, do_nocb_bypass_wakeup_timer, 0);
2256 rcu_cblist_init(&rdp->nocb_bypass);
2260 * If the specified CPU is a no-CBs CPU that does not already have its
2261 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2262 * for this CPU's group has not yet been created, spawn it as well.
2264 static void rcu_spawn_one_nocb_kthread(int cpu)
2266 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2267 struct rcu_data *rdp_gp;
2268 struct task_struct *t;
2271 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2272 * then nothing to do.
2274 if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2277 /* If we didn't spawn the GP kthread first, reorganize! */
2278 rdp_gp = rdp->nocb_gp_rdp;
2279 if (!rdp_gp->nocb_gp_kthread) {
2280 t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2281 "rcuog/%d", rdp_gp->cpu);
2282 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2284 WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2287 /* Spawn the kthread for this CPU. */
2288 t = kthread_run(rcu_nocb_cb_kthread, rdp,
2289 "rcuo%c/%d", rcu_state.abbr, cpu);
2290 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2292 WRITE_ONCE(rdp->nocb_cb_kthread, t);
2293 WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2297 * If the specified CPU is a no-CBs CPU that does not already have its
2298 * rcuo kthread, spawn it.
2300 static void rcu_spawn_cpu_nocb_kthread(int cpu)
2302 if (rcu_scheduler_fully_active)
2303 rcu_spawn_one_nocb_kthread(cpu);
2307 * Once the scheduler is running, spawn rcuo kthreads for all online
2308 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2309 * non-boot CPUs come online -- if this changes, we will need to add
2310 * some mutual exclusion.
2312 static void __init rcu_spawn_nocb_kthreads(void)
2316 for_each_online_cpu(cpu)
2317 rcu_spawn_cpu_nocb_kthread(cpu);
2320 /* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2321 static int rcu_nocb_gp_stride = -1;
2322 module_param(rcu_nocb_gp_stride, int, 0444);
2325 * Initialize GP-CB relationships for all no-CBs CPU.
2327 static void __init rcu_organize_nocb_kthreads(void)
2330 bool firsttime = true;
2331 bool gotnocbs = false;
2332 bool gotnocbscbs = true;
2333 int ls = rcu_nocb_gp_stride;
2334 int nl = 0; /* Next GP kthread. */
2335 struct rcu_data *rdp;
2336 struct rcu_data *rdp_gp = NULL; /* Suppress misguided gcc warn. */
2337 struct rcu_data *rdp_prev = NULL;
2339 if (!cpumask_available(rcu_nocb_mask))
2342 ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2343 rcu_nocb_gp_stride = ls;
2347 * Each pass through this loop sets up one rcu_data structure.
2348 * Should the corresponding CPU come online in the future, then
2349 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2351 for_each_cpu(cpu, rcu_nocb_mask) {
2352 rdp = per_cpu_ptr(&rcu_data, cpu);
2353 if (rdp->cpu >= nl) {
2354 /* New GP kthread, set up for CBs & next GP. */
2356 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2357 rdp->nocb_gp_rdp = rdp;
2361 pr_cont("%s\n", gotnocbscbs
2362 ? "" : " (self only)");
2363 gotnocbscbs = false;
2365 pr_alert("%s: No-CB GP kthread CPU %d:",
2369 /* Another CB kthread, link to previous GP kthread. */
2371 rdp->nocb_gp_rdp = rdp_gp;
2372 rdp_prev->nocb_next_cb_rdp = rdp;
2374 pr_cont(" %d", cpu);
2378 if (gotnocbs && dump_tree)
2379 pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
2383 * Bind the current task to the offloaded CPUs. If there are no offloaded
2384 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2386 void rcu_bind_current_to_nocb(void)
2388 if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2389 WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2391 EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2394 * Dump out nocb grace-period kthread state for the specified rcu_data
2397 static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2399 struct rcu_node *rnp = rdp->mynode;
2401 pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2403 "kK"[!!rdp->nocb_gp_kthread],
2404 "lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2405 "dD"[!!rdp->nocb_defer_wakeup],
2406 "tT"[timer_pending(&rdp->nocb_timer)],
2407 "bB"[timer_pending(&rdp->nocb_bypass_timer)],
2408 "sS"[!!rdp->nocb_gp_sleep],
2409 ".W"[swait_active(&rdp->nocb_gp_wq)],
2410 ".W"[swait_active(&rnp->nocb_gp_wq[0])],
2411 ".W"[swait_active(&rnp->nocb_gp_wq[1])],
2412 ".B"[!!rdp->nocb_gp_bypass],
2413 ".G"[!!rdp->nocb_gp_gp],
2414 (long)rdp->nocb_gp_seq,
2415 rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops));
2418 /* Dump out nocb kthread state for the specified rcu_data structure. */
2419 static void show_rcu_nocb_state(struct rcu_data *rdp)
2421 struct rcu_segcblist *rsclp = &rdp->cblist;
2426 if (rdp->nocb_gp_rdp == rdp)
2427 show_rcu_nocb_gp_state(rdp);
2429 pr_info(" CB %d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%c%c%c q%ld\n",
2430 rdp->cpu, rdp->nocb_gp_rdp->cpu,
2431 "kK"[!!rdp->nocb_cb_kthread],
2432 "bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2433 "cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2434 "lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2435 "sS"[!!rdp->nocb_cb_sleep],
2436 ".W"[swait_active(&rdp->nocb_cb_wq)],
2437 jiffies - rdp->nocb_bypass_first,
2438 jiffies - rdp->nocb_nobypass_last,
2439 rdp->nocb_nobypass_count,
2440 ".D"[rcu_segcblist_ready_cbs(rsclp)],
2441 ".W"[!rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL)],
2442 ".R"[!rcu_segcblist_restempty(rsclp, RCU_WAIT_TAIL)],
2443 ".N"[!rcu_segcblist_restempty(rsclp, RCU_NEXT_READY_TAIL)],
2444 ".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2445 rcu_segcblist_n_cbs(&rdp->cblist));
2447 /* It is OK for GP kthreads to have GP state. */
2448 if (rdp->nocb_gp_rdp == rdp)
2451 waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2452 wastimer = timer_pending(&rdp->nocb_timer);
2453 wassleep = swait_active(&rdp->nocb_gp_wq);
2454 if (!rdp->nocb_defer_wakeup && !rdp->nocb_gp_sleep &&
2455 !waslocked && !wastimer && !wassleep)
2456 return; /* Nothing untowards. */
2458 pr_info(" !!! %c%c%c%c %c\n",
2460 "dD"[!!rdp->nocb_defer_wakeup],
2462 "sS"[!!rdp->nocb_gp_sleep],
2466 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2468 /* No ->nocb_lock to acquire. */
2469 static void rcu_nocb_lock(struct rcu_data *rdp)
2473 /* No ->nocb_lock to release. */
2474 static void rcu_nocb_unlock(struct rcu_data *rdp)
2478 /* No ->nocb_lock to release. */
2479 static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2480 unsigned long flags)
2482 local_irq_restore(flags);
2485 /* Lockdep check that ->cblist may be safely accessed. */
2486 static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2488 lockdep_assert_irqs_disabled();
2491 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2495 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2500 static void rcu_init_one_nocb(struct rcu_node *rnp)
2504 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2510 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2511 bool *was_alldone, unsigned long flags)
2516 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2517 unsigned long flags)
2519 WARN_ON_ONCE(1); /* Should be dead code! */
2522 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2526 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2531 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2535 static void rcu_spawn_cpu_nocb_kthread(int cpu)
2539 static void __init rcu_spawn_nocb_kthreads(void)
2543 static void show_rcu_nocb_state(struct rcu_data *rdp)
2547 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2550 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2551 * grace-period kthread will do force_quiescent_state() processing?
2552 * The idea is to avoid waking up RCU core processing on such a
2553 * CPU unless the grace period has extended for too long.
2555 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2556 * CONFIG_RCU_NOCB_CPU CPUs.
2558 static bool rcu_nohz_full_cpu(void)
2560 #ifdef CONFIG_NO_HZ_FULL
2561 if (tick_nohz_full_cpu(smp_processor_id()) &&
2562 (!rcu_gp_in_progress() ||
2563 ULONG_CMP_LT(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2565 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2570 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2572 static void rcu_bind_gp_kthread(void)
2574 if (!tick_nohz_full_enabled())
2576 housekeeping_affine(current, HK_FLAG_RCU);
2579 /* Record the current task on dyntick-idle entry. */
2580 static void rcu_dynticks_task_enter(void)
2582 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2583 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2584 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2587 /* Record no current task on dyntick-idle exit. */
2588 static void rcu_dynticks_task_exit(void)
2590 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2591 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2592 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */