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 (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
40 pr_info("\tRCU strict (and thus non-scalable) grace periods enabled.\n");
41 if (RCU_NUM_LVLS >= 4)
42 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
43 if (RCU_FANOUT_LEAF != 16)
44 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
46 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
47 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
49 if (nr_cpu_ids != NR_CPUS)
50 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
51 #ifdef CONFIG_RCU_BOOST
52 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
53 kthread_prio, CONFIG_RCU_BOOST_DELAY);
55 if (blimit != DEFAULT_RCU_BLIMIT)
56 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
57 if (qhimark != DEFAULT_RCU_QHIMARK)
58 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
59 if (qlowmark != DEFAULT_RCU_QLOMARK)
60 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
61 if (qovld != DEFAULT_RCU_QOVLD)
62 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
63 if (jiffies_till_first_fqs != ULONG_MAX)
64 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
65 if (jiffies_till_next_fqs != ULONG_MAX)
66 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
67 if (jiffies_till_sched_qs != ULONG_MAX)
68 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
69 if (rcu_kick_kthreads)
70 pr_info("\tKick kthreads if too-long grace period.\n");
71 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
72 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
74 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
76 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
78 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
80 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
81 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
82 pr_info("\tRCU debug extended QS entry/exit.\n");
83 rcupdate_announce_bootup_oddness();
86 #ifdef CONFIG_PREEMPT_RCU
88 static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
89 static void rcu_read_unlock_special(struct task_struct *t);
92 * Tell them what RCU they are running.
94 static void __init rcu_bootup_announce(void)
96 pr_info("Preemptible hierarchical RCU implementation.\n");
97 rcu_bootup_announce_oddness();
100 /* Flags for rcu_preempt_ctxt_queue() decision table. */
101 #define RCU_GP_TASKS 0x8
102 #define RCU_EXP_TASKS 0x4
103 #define RCU_GP_BLKD 0x2
104 #define RCU_EXP_BLKD 0x1
107 * Queues a task preempted within an RCU-preempt read-side critical
108 * section into the appropriate location within the ->blkd_tasks list,
109 * depending on the states of any ongoing normal and expedited grace
110 * periods. The ->gp_tasks pointer indicates which element the normal
111 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
112 * indicates which element the expedited grace period is waiting on (again,
113 * NULL if none). If a grace period is waiting on a given element in the
114 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
115 * adding a task to the tail of the list blocks any grace period that is
116 * already waiting on one of the elements. In contrast, adding a task
117 * to the head of the list won't block any grace period that is already
118 * waiting on one of the elements.
120 * This queuing is imprecise, and can sometimes make an ongoing grace
121 * period wait for a task that is not strictly speaking blocking it.
122 * Given the choice, we needlessly block a normal grace period rather than
123 * blocking an expedited grace period.
125 * Note that an endless sequence of expedited grace periods still cannot
126 * indefinitely postpone a normal grace period. Eventually, all of the
127 * fixed number of preempted tasks blocking the normal grace period that are
128 * not also blocking the expedited grace period will resume and complete
129 * their RCU read-side critical sections. At that point, the ->gp_tasks
130 * pointer will equal the ->exp_tasks pointer, at which point the end of
131 * the corresponding expedited grace period will also be the end of the
132 * normal grace period.
134 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
135 __releases(rnp->lock) /* But leaves rrupts disabled. */
137 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
138 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
139 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
140 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
141 struct task_struct *t = current;
143 raw_lockdep_assert_held_rcu_node(rnp);
144 WARN_ON_ONCE(rdp->mynode != rnp);
145 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
146 /* RCU better not be waiting on newly onlined CPUs! */
147 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
151 * Decide where to queue the newly blocked task. In theory,
152 * this could be an if-statement. In practice, when I tried
153 * that, it was quite messy.
155 switch (blkd_state) {
158 case RCU_EXP_TASKS + RCU_GP_BLKD:
160 case RCU_GP_TASKS + RCU_EXP_TASKS:
163 * Blocking neither GP, or first task blocking the normal
164 * GP but not blocking the already-waiting expedited GP.
165 * Queue at the head of the list to avoid unnecessarily
166 * blocking the already-waiting GPs.
168 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
173 case RCU_GP_BLKD + RCU_EXP_BLKD:
174 case RCU_GP_TASKS + RCU_EXP_BLKD:
175 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
176 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
179 * First task arriving that blocks either GP, or first task
180 * arriving that blocks the expedited GP (with the normal
181 * GP already waiting), or a task arriving that blocks
182 * both GPs with both GPs already waiting. Queue at the
183 * tail of the list to avoid any GP waiting on any of the
184 * already queued tasks that are not blocking it.
186 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
189 case RCU_EXP_TASKS + RCU_EXP_BLKD:
190 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
191 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
194 * Second or subsequent task blocking the expedited GP.
195 * The task either does not block the normal GP, or is the
196 * first task blocking the normal GP. Queue just after
197 * the first task blocking the expedited GP.
199 list_add(&t->rcu_node_entry, rnp->exp_tasks);
202 case RCU_GP_TASKS + RCU_GP_BLKD:
203 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
206 * Second or subsequent task blocking the normal GP.
207 * The task does not block the expedited GP. Queue just
208 * after the first task blocking the normal GP.
210 list_add(&t->rcu_node_entry, rnp->gp_tasks);
215 /* Yet another exercise in excessive paranoia. */
221 * We have now queued the task. If it was the first one to
222 * block either grace period, update the ->gp_tasks and/or
223 * ->exp_tasks pointers, respectively, to reference the newly
226 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
227 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
228 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
230 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
231 WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
232 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
233 !(rnp->qsmask & rdp->grpmask));
234 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
235 !(rnp->expmask & rdp->grpmask));
236 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
239 * Report the quiescent state for the expedited GP. This expedited
240 * GP should not be able to end until we report, so there should be
241 * no need to check for a subsequent expedited GP. (Though we are
242 * still in a quiescent state in any case.)
244 if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
245 rcu_report_exp_rdp(rdp);
247 WARN_ON_ONCE(rdp->exp_deferred_qs);
251 * Record a preemptible-RCU quiescent state for the specified CPU.
252 * Note that this does not necessarily mean that the task currently running
253 * on the CPU is in a quiescent state: Instead, it means that the current
254 * grace period need not wait on any RCU read-side critical section that
255 * starts later on this CPU. It also means that if the current task is
256 * in an RCU read-side critical section, it has already added itself to
257 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
258 * current task, there might be any number of other tasks blocked while
259 * in an RCU read-side critical section.
261 * Callers to this function must disable preemption.
263 static void rcu_qs(void)
265 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
266 if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
267 trace_rcu_grace_period(TPS("rcu_preempt"),
268 __this_cpu_read(rcu_data.gp_seq),
270 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
271 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
272 WRITE_ONCE(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 void rcu_note_context_switch(bool preempt)
291 struct task_struct *t = current;
292 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
293 struct rcu_node *rnp;
295 trace_rcu_utilization(TPS("Start context switch"));
296 lockdep_assert_irqs_disabled();
297 WARN_ON_ONCE(!preempt && rcu_preempt_depth() > 0);
298 if (rcu_preempt_depth() > 0 &&
299 !t->rcu_read_unlock_special.b.blocked) {
301 /* Possibly blocking in an RCU read-side critical section. */
303 raw_spin_lock_rcu_node(rnp);
304 t->rcu_read_unlock_special.b.blocked = true;
305 t->rcu_blocked_node = rnp;
308 * Verify the CPU's sanity, trace the preemption, and
309 * then queue the task as required based on the states
310 * of any ongoing and expedited grace periods.
312 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
313 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
314 trace_rcu_preempt_task(rcu_state.name,
316 (rnp->qsmask & rdp->grpmask)
318 : rcu_seq_snap(&rnp->gp_seq));
319 rcu_preempt_ctxt_queue(rnp, rdp);
321 rcu_preempt_deferred_qs(t);
325 * Either we were not in an RCU read-side critical section to
326 * begin with, or we have now recorded that critical section
327 * globally. Either way, we can now note a quiescent state
328 * for this CPU. Again, if we were in an RCU read-side critical
329 * section, and if that critical section was blocking the current
330 * grace period, then the fact that the task has been enqueued
331 * means that we continue to block the current grace period.
334 if (rdp->exp_deferred_qs)
335 rcu_report_exp_rdp(rdp);
336 rcu_tasks_qs(current, preempt);
337 trace_rcu_utilization(TPS("End context switch"));
339 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
342 * Check for preempted RCU readers blocking the current grace period
343 * for the specified rcu_node structure. If the caller needs a reliable
344 * answer, it must hold the rcu_node's ->lock.
346 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
348 return READ_ONCE(rnp->gp_tasks) != NULL;
351 /* limit value for ->rcu_read_lock_nesting. */
352 #define RCU_NEST_PMAX (INT_MAX / 2)
354 static void rcu_preempt_read_enter(void)
356 current->rcu_read_lock_nesting++;
359 static int rcu_preempt_read_exit(void)
361 return --current->rcu_read_lock_nesting;
364 static void rcu_preempt_depth_set(int val)
366 current->rcu_read_lock_nesting = val;
370 * Preemptible RCU implementation for rcu_read_lock().
371 * Just increment ->rcu_read_lock_nesting, shared state will be updated
374 void __rcu_read_lock(void)
376 rcu_preempt_read_enter();
377 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
378 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
379 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
380 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
381 barrier(); /* critical section after entry code. */
383 EXPORT_SYMBOL_GPL(__rcu_read_lock);
386 * Preemptible RCU implementation for rcu_read_unlock().
387 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
388 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
389 * invoke rcu_read_unlock_special() to clean up after a context switch
390 * in an RCU read-side critical section and other special cases.
392 void __rcu_read_unlock(void)
394 struct task_struct *t = current;
396 if (rcu_preempt_read_exit() == 0) {
397 barrier(); /* critical section before exit code. */
398 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
399 rcu_read_unlock_special(t);
401 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
402 int rrln = rcu_preempt_depth();
404 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
407 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
410 * Advance a ->blkd_tasks-list pointer to the next entry, instead
411 * returning NULL if at the end of the list.
413 static struct list_head *rcu_next_node_entry(struct task_struct *t,
414 struct rcu_node *rnp)
416 struct list_head *np;
418 np = t->rcu_node_entry.next;
419 if (np == &rnp->blkd_tasks)
425 * Return true if the specified rcu_node structure has tasks that were
426 * preempted within an RCU read-side critical section.
428 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
430 return !list_empty(&rnp->blkd_tasks);
434 * Report deferred quiescent states. The deferral time can
435 * be quite short, for example, in the case of the call from
436 * rcu_read_unlock_special().
439 rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
444 struct list_head *np;
445 bool drop_boost_mutex = false;
446 struct rcu_data *rdp;
447 struct rcu_node *rnp;
448 union rcu_special special;
451 * If RCU core is waiting for this CPU to exit its critical section,
452 * report the fact that it has exited. Because irqs are disabled,
453 * t->rcu_read_unlock_special cannot change.
455 special = t->rcu_read_unlock_special;
456 rdp = this_cpu_ptr(&rcu_data);
457 if (!special.s && !rdp->exp_deferred_qs) {
458 local_irq_restore(flags);
461 t->rcu_read_unlock_special.s = 0;
462 if (special.b.need_qs) {
463 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
464 rcu_report_qs_rdp(rdp);
465 udelay(rcu_unlock_delay);
472 * Respond to a request by an expedited grace period for a
473 * quiescent state from this CPU. Note that requests from
474 * tasks are handled when removing the task from the
475 * blocked-tasks list below.
477 if (rdp->exp_deferred_qs)
478 rcu_report_exp_rdp(rdp);
480 /* Clean up if blocked during RCU read-side critical section. */
481 if (special.b.blocked) {
484 * Remove this task from the list it blocked on. The task
485 * now remains queued on the rcu_node corresponding to the
486 * CPU it first blocked on, so there is no longer any need
487 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
489 rnp = t->rcu_blocked_node;
490 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
491 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
492 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
493 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
494 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
495 (!empty_norm || rnp->qsmask));
496 empty_exp = sync_rcu_exp_done(rnp);
497 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
498 np = rcu_next_node_entry(t, rnp);
499 list_del_init(&t->rcu_node_entry);
500 t->rcu_blocked_node = NULL;
501 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
502 rnp->gp_seq, t->pid);
503 if (&t->rcu_node_entry == rnp->gp_tasks)
504 WRITE_ONCE(rnp->gp_tasks, np);
505 if (&t->rcu_node_entry == rnp->exp_tasks)
506 WRITE_ONCE(rnp->exp_tasks, np);
507 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
508 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
509 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
510 if (&t->rcu_node_entry == rnp->boost_tasks)
511 WRITE_ONCE(rnp->boost_tasks, np);
515 * If this was the last task on the current list, and if
516 * we aren't waiting on any CPUs, report the quiescent state.
517 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
518 * so we must take a snapshot of the expedited state.
520 empty_exp_now = sync_rcu_exp_done(rnp);
521 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
522 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
529 rcu_report_unblock_qs_rnp(rnp, flags);
531 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
535 * If this was the last task on the expedited lists,
536 * then we need to report up the rcu_node hierarchy.
538 if (!empty_exp && empty_exp_now)
539 rcu_report_exp_rnp(rnp, true);
541 /* Unboost if we were boosted. */
542 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
543 rt_mutex_futex_unlock(&rnp->boost_mtx);
546 local_irq_restore(flags);
551 * Is a deferred quiescent-state pending, and are we also not in
552 * an RCU read-side critical section? It is the caller's responsibility
553 * to ensure it is otherwise safe to report any deferred quiescent
554 * states. The reason for this is that it is safe to report a
555 * quiescent state during context switch even though preemption
556 * is disabled. This function cannot be expected to understand these
557 * nuances, so the caller must handle them.
559 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
561 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
562 READ_ONCE(t->rcu_read_unlock_special.s)) &&
563 rcu_preempt_depth() == 0;
567 * Report a deferred quiescent state if needed and safe to do so.
568 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
569 * not being in an RCU read-side critical section. The caller must
570 * evaluate safety in terms of interrupt, softirq, and preemption
573 static void rcu_preempt_deferred_qs(struct task_struct *t)
577 if (!rcu_preempt_need_deferred_qs(t))
579 local_irq_save(flags);
580 rcu_preempt_deferred_qs_irqrestore(t, flags);
584 * Minimal handler to give the scheduler a chance to re-evaluate.
586 static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
588 struct rcu_data *rdp;
590 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
591 rdp->defer_qs_iw_pending = false;
595 * Handle special cases during rcu_read_unlock(), such as needing to
596 * notify RCU core processing or task having blocked during the RCU
597 * read-side critical section.
599 static void rcu_read_unlock_special(struct task_struct *t)
602 bool preempt_bh_were_disabled =
603 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
604 bool irqs_were_disabled;
606 /* NMI handlers cannot block and cannot safely manipulate state. */
610 local_irq_save(flags);
611 irqs_were_disabled = irqs_disabled_flags(flags);
612 if (preempt_bh_were_disabled || irqs_were_disabled) {
614 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
615 struct rcu_node *rnp = rdp->mynode;
617 exp = (t->rcu_blocked_node &&
618 READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
619 (rdp->grpmask & READ_ONCE(rnp->expmask));
620 // Need to defer quiescent state until everything is enabled.
621 if (use_softirq && (in_irq() || (exp && !irqs_were_disabled))) {
622 // Using softirq, safe to awaken, and either the
623 // wakeup is free or there is an expedited GP.
624 raise_softirq_irqoff(RCU_SOFTIRQ);
626 // Enabling BH or preempt does reschedule, so...
627 // Also if no expediting, slow is OK.
628 // Plus nohz_full CPUs eventually get tick enabled.
629 set_tsk_need_resched(current);
630 set_preempt_need_resched();
631 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
632 !rdp->defer_qs_iw_pending && exp && cpu_online(rdp->cpu)) {
633 // Get scheduler to re-evaluate and call hooks.
634 // If !IRQ_WORK, FQS scan will eventually IPI.
635 init_irq_work(&rdp->defer_qs_iw,
636 rcu_preempt_deferred_qs_handler);
637 rdp->defer_qs_iw_pending = true;
638 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
641 local_irq_restore(flags);
644 rcu_preempt_deferred_qs_irqrestore(t, flags);
648 * Check that the list of blocked tasks for the newly completed grace
649 * period is in fact empty. It is a serious bug to complete a grace
650 * period that still has RCU readers blocked! This function must be
651 * invoked -before- updating this rnp's ->gp_seq.
653 * Also, if there are blocked tasks on the list, they automatically
654 * block the newly created grace period, so set up ->gp_tasks accordingly.
656 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
658 struct task_struct *t;
660 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
661 raw_lockdep_assert_held_rcu_node(rnp);
662 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
663 dump_blkd_tasks(rnp, 10);
664 if (rcu_preempt_has_tasks(rnp) &&
665 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
666 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
667 t = container_of(rnp->gp_tasks, struct task_struct,
669 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
670 rnp->gp_seq, t->pid);
672 WARN_ON_ONCE(rnp->qsmask);
676 * Check for a quiescent state from the current CPU, including voluntary
677 * context switches for Tasks RCU. When a task blocks, the task is
678 * recorded in the corresponding CPU's rcu_node structure, which is checked
679 * elsewhere, hence this function need only check for quiescent states
680 * related to the current CPU, not to those related to tasks.
682 static void rcu_flavor_sched_clock_irq(int user)
684 struct task_struct *t = current;
686 lockdep_assert_irqs_disabled();
687 if (user || rcu_is_cpu_rrupt_from_idle()) {
688 rcu_note_voluntary_context_switch(current);
690 if (rcu_preempt_depth() > 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 (!WARN_ON_ONCE(rcu_preempt_depth())) {
701 rcu_qs(); /* Report immediate QS. */
705 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
706 if (rcu_preempt_depth() > 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 rcu_preempt_depth_set(1);
729 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
730 } else if (unlikely(rcu_preempt_depth())) {
731 rcu_preempt_depth_set(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)READ_ONCE(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), data_race(rnp->boost_tasks),
762 READ_ONCE(rnp->exp_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 * If strict grace periods are enabled, and if the calling
785 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
786 * report that quiescent state and, if requested, spin for a bit.
788 void rcu_read_unlock_strict(void)
790 struct rcu_data *rdp;
792 if (!IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ||
793 irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
795 rdp = this_cpu_ptr(&rcu_data);
796 rcu_report_qs_rdp(rdp);
797 udelay(rcu_unlock_delay);
799 EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
802 * Tell them what RCU they are running.
804 static void __init rcu_bootup_announce(void)
806 pr_info("Hierarchical RCU implementation.\n");
807 rcu_bootup_announce_oddness();
811 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
812 * how many quiescent states passed, just if there was at least one since
813 * the start of the grace period, this just sets a flag. The caller must
814 * have disabled preemption.
816 static void rcu_qs(void)
818 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
819 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
821 trace_rcu_grace_period(TPS("rcu_sched"),
822 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
823 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
824 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
826 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
827 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
831 * Register an urgently needed quiescent state. If there is an
832 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
833 * dyntick-idle quiescent state visible to other CPUs, which will in
834 * some cases serve for expedited as well as normal grace periods.
835 * Either way, register a lightweight quiescent state.
837 void rcu_all_qs(void)
841 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
844 /* Load rcu_urgent_qs before other flags. */
845 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
849 this_cpu_write(rcu_data.rcu_urgent_qs, false);
850 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
851 local_irq_save(flags);
852 rcu_momentary_dyntick_idle();
853 local_irq_restore(flags);
858 EXPORT_SYMBOL_GPL(rcu_all_qs);
861 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
863 void rcu_note_context_switch(bool preempt)
865 trace_rcu_utilization(TPS("Start context switch"));
867 /* Load rcu_urgent_qs before other flags. */
868 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
870 this_cpu_write(rcu_data.rcu_urgent_qs, false);
871 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
872 rcu_momentary_dyntick_idle();
873 rcu_tasks_qs(current, preempt);
875 trace_rcu_utilization(TPS("End context switch"));
877 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
880 * Because preemptible RCU does not exist, there are never any preempted
883 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
889 * Because there is no preemptible RCU, there can be no readers blocked.
891 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
897 * Because there is no preemptible RCU, there can be no deferred quiescent
900 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
904 static void rcu_preempt_deferred_qs(struct task_struct *t) { }
907 * Because there is no preemptible RCU, there can be no readers blocked,
908 * so there is no need to check for blocked tasks. So check only for
909 * bogus qsmask values.
911 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
913 WARN_ON_ONCE(rnp->qsmask);
917 * Check to see if this CPU is in a non-context-switch quiescent state,
918 * namely user mode and idle loop.
920 static void rcu_flavor_sched_clock_irq(int user)
922 if (user || rcu_is_cpu_rrupt_from_idle()) {
925 * Get here if this CPU took its interrupt from user
926 * mode or from the idle loop, and if this is not a
927 * nested interrupt. In this case, the CPU is in
928 * a quiescent state, so note it.
930 * No memory barrier is required here because rcu_qs()
931 * references only CPU-local variables that other CPUs
932 * neither access nor modify, at least not while the
933 * corresponding CPU is online.
941 * Because preemptible RCU does not exist, tasks cannot possibly exit
942 * while in preemptible RCU read-side critical sections.
949 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
952 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
954 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
957 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
960 * If boosting, set rcuc kthreads to realtime priority.
962 static void rcu_cpu_kthread_setup(unsigned int cpu)
964 #ifdef CONFIG_RCU_BOOST
965 struct sched_param sp;
967 sp.sched_priority = kthread_prio;
968 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
969 #endif /* #ifdef CONFIG_RCU_BOOST */
972 #ifdef CONFIG_RCU_BOOST
975 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
976 * or ->boost_tasks, advancing the pointer to the next task in the
979 * Note that irqs must be enabled: boosting the task can block.
980 * Returns 1 if there are more tasks needing to be boosted.
982 static int rcu_boost(struct rcu_node *rnp)
985 struct task_struct *t;
986 struct list_head *tb;
988 if (READ_ONCE(rnp->exp_tasks) == NULL &&
989 READ_ONCE(rnp->boost_tasks) == NULL)
990 return 0; /* Nothing left to boost. */
992 raw_spin_lock_irqsave_rcu_node(rnp, flags);
995 * Recheck under the lock: all tasks in need of boosting
996 * might exit their RCU read-side critical sections on their own.
998 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
999 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1004 * Preferentially boost tasks blocking expedited grace periods.
1005 * This cannot starve the normal grace periods because a second
1006 * expedited grace period must boost all blocked tasks, including
1007 * those blocking the pre-existing normal grace period.
1009 if (rnp->exp_tasks != NULL)
1010 tb = rnp->exp_tasks;
1012 tb = rnp->boost_tasks;
1015 * We boost task t by manufacturing an rt_mutex that appears to
1016 * be held by task t. We leave a pointer to that rt_mutex where
1017 * task t can find it, and task t will release the mutex when it
1018 * exits its outermost RCU read-side critical section. Then
1019 * simply acquiring this artificial rt_mutex will boost task
1020 * t's priority. (Thanks to tglx for suggesting this approach!)
1022 * Note that task t must acquire rnp->lock to remove itself from
1023 * the ->blkd_tasks list, which it will do from exit() if from
1024 * nowhere else. We therefore are guaranteed that task t will
1025 * stay around at least until we drop rnp->lock. Note that
1026 * rnp->lock also resolves races between our priority boosting
1027 * and task t's exiting its outermost RCU read-side critical
1030 t = container_of(tb, struct task_struct, rcu_node_entry);
1031 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1032 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1033 /* Lock only for side effect: boosts task t's priority. */
1034 rt_mutex_lock(&rnp->boost_mtx);
1035 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1037 return READ_ONCE(rnp->exp_tasks) != NULL ||
1038 READ_ONCE(rnp->boost_tasks) != NULL;
1042 * Priority-boosting kthread, one per leaf rcu_node.
1044 static int rcu_boost_kthread(void *arg)
1046 struct rcu_node *rnp = (struct rcu_node *)arg;
1050 trace_rcu_utilization(TPS("Start boost kthread@init"));
1052 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1053 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1054 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1055 READ_ONCE(rnp->exp_tasks));
1056 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1057 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1058 more2boost = rcu_boost(rnp);
1064 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1065 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1066 schedule_timeout_idle(2);
1067 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1072 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1077 * Check to see if it is time to start boosting RCU readers that are
1078 * blocking the current grace period, and, if so, tell the per-rcu_node
1079 * kthread to start boosting them. If there is an expedited grace
1080 * period in progress, it is always time to boost.
1082 * The caller must hold rnp->lock, which this function releases.
1083 * The ->boost_kthread_task is immortal, so we don't need to worry
1084 * about it going away.
1086 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1087 __releases(rnp->lock)
1089 raw_lockdep_assert_held_rcu_node(rnp);
1090 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1091 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1094 if (rnp->exp_tasks != NULL ||
1095 (rnp->gp_tasks != NULL &&
1096 rnp->boost_tasks == NULL &&
1098 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld))) {
1099 if (rnp->exp_tasks == NULL)
1100 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1101 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1102 rcu_wake_cond(rnp->boost_kthread_task,
1103 READ_ONCE(rnp->boost_kthread_status));
1105 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1110 * Is the current CPU running the RCU-callbacks kthread?
1111 * Caller must have preemption disabled.
1113 static bool rcu_is_callbacks_kthread(void)
1115 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1118 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1121 * Do priority-boost accounting for the start of a new grace period.
1123 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1125 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1129 * Create an RCU-boost kthread for the specified node if one does not
1130 * already exist. We only create this kthread for preemptible RCU.
1131 * Returns zero if all is well, a negated errno otherwise.
1133 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1135 int rnp_index = rnp - rcu_get_root();
1136 unsigned long flags;
1137 struct sched_param sp;
1138 struct task_struct *t;
1140 if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1143 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1146 rcu_state.boost = 1;
1148 if (rnp->boost_kthread_task != NULL)
1151 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1152 "rcub/%d", rnp_index);
1153 if (WARN_ON_ONCE(IS_ERR(t)))
1156 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1157 rnp->boost_kthread_task = t;
1158 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1159 sp.sched_priority = kthread_prio;
1160 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1161 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1165 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1166 * served by the rcu_node in question. The CPU hotplug lock is still
1167 * held, so the value of rnp->qsmaskinit will be stable.
1169 * We don't include outgoingcpu in the affinity set, use -1 if there is
1170 * no outgoing CPU. If there are no CPUs left in the affinity set,
1171 * this function allows the kthread to execute on any CPU.
1173 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1175 struct task_struct *t = rnp->boost_kthread_task;
1176 unsigned long mask = rcu_rnp_online_cpus(rnp);
1182 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1184 for_each_leaf_node_possible_cpu(rnp, cpu)
1185 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1187 cpumask_set_cpu(cpu, cm);
1188 if (cpumask_weight(cm) == 0)
1190 set_cpus_allowed_ptr(t, cm);
1191 free_cpumask_var(cm);
1195 * Spawn boost kthreads -- called as soon as the scheduler is running.
1197 static void __init rcu_spawn_boost_kthreads(void)
1199 struct rcu_node *rnp;
1201 rcu_for_each_leaf_node(rnp)
1202 rcu_spawn_one_boost_kthread(rnp);
1205 static void rcu_prepare_kthreads(int cpu)
1207 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1208 struct rcu_node *rnp = rdp->mynode;
1210 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1211 if (rcu_scheduler_fully_active)
1212 rcu_spawn_one_boost_kthread(rnp);
1215 #else /* #ifdef CONFIG_RCU_BOOST */
1217 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1218 __releases(rnp->lock)
1220 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1223 static bool rcu_is_callbacks_kthread(void)
1228 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1232 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1236 static void __init rcu_spawn_boost_kthreads(void)
1240 static void rcu_prepare_kthreads(int cpu)
1244 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1246 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1249 * Check to see if any future non-offloaded RCU-related work will need
1250 * to be done by the current CPU, even if none need be done immediately,
1251 * returning 1 if so. This function is part of the RCU implementation;
1252 * it is -not- an exported member of the RCU API.
1254 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1255 * CPU has RCU callbacks queued.
1257 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1259 *nextevt = KTIME_MAX;
1260 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1261 !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist);
1265 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1268 static void rcu_cleanup_after_idle(void)
1273 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1276 static void rcu_prepare_for_idle(void)
1280 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1283 * This code is invoked when a CPU goes idle, at which point we want
1284 * to have the CPU do everything required for RCU so that it can enter
1285 * the energy-efficient dyntick-idle mode.
1287 * The following preprocessor symbol controls this:
1289 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1290 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1291 * is sized to be roughly one RCU grace period. Those energy-efficiency
1292 * benchmarkers who might otherwise be tempted to set this to a large
1293 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1294 * system. And if you are -that- concerned about energy efficiency,
1295 * just power the system down and be done with it!
1297 * The value below works well in practice. If future workloads require
1298 * adjustment, they can be converted into kernel config parameters, though
1299 * making the state machine smarter might be a better option.
1301 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1303 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1304 module_param(rcu_idle_gp_delay, int, 0644);
1307 * Try to advance callbacks on the current CPU, but only if it has been
1308 * awhile since the last time we did so. Afterwards, if there are any
1309 * callbacks ready for immediate invocation, return true.
1311 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1313 bool cbs_ready = false;
1314 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1315 struct rcu_node *rnp;
1317 /* Exit early if we advanced recently. */
1318 if (jiffies == rdp->last_advance_all)
1320 rdp->last_advance_all = jiffies;
1325 * Don't bother checking unless a grace period has
1326 * completed since we last checked and there are
1327 * callbacks not yet ready to invoke.
1329 if ((rcu_seq_completed_gp(rdp->gp_seq,
1330 rcu_seq_current(&rnp->gp_seq)) ||
1331 unlikely(READ_ONCE(rdp->gpwrap))) &&
1332 rcu_segcblist_pend_cbs(&rdp->cblist))
1333 note_gp_changes(rdp);
1335 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1341 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1342 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1343 * caller about what to set the timeout.
1345 * The caller must have disabled interrupts.
1347 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1349 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1352 lockdep_assert_irqs_disabled();
1354 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1355 if (rcu_segcblist_empty(&rdp->cblist) ||
1356 rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist)) {
1357 *nextevt = KTIME_MAX;
1361 /* Attempt to advance callbacks. */
1362 if (rcu_try_advance_all_cbs()) {
1363 /* Some ready to invoke, so initiate later invocation. */
1367 rdp->last_accelerate = jiffies;
1369 /* Request timer and round. */
1370 dj = round_up(rcu_idle_gp_delay + jiffies, rcu_idle_gp_delay) - jiffies;
1372 *nextevt = basemono + dj * TICK_NSEC;
1377 * Prepare a CPU for idle from an RCU perspective. The first major task is to
1378 * sense whether nohz mode has been enabled or disabled via sysfs. The second
1379 * major task is to accelerate (that is, assign grace-period numbers to) any
1380 * recently arrived callbacks.
1382 * The caller must have disabled interrupts.
1384 static void rcu_prepare_for_idle(void)
1387 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1388 struct rcu_node *rnp;
1391 lockdep_assert_irqs_disabled();
1392 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1395 /* Handle nohz enablement switches conservatively. */
1396 tne = READ_ONCE(tick_nohz_active);
1397 if (tne != rdp->tick_nohz_enabled_snap) {
1398 if (!rcu_segcblist_empty(&rdp->cblist))
1399 invoke_rcu_core(); /* force nohz to see update. */
1400 rdp->tick_nohz_enabled_snap = tne;
1407 * If we have not yet accelerated this jiffy, accelerate all
1408 * callbacks on this CPU.
1410 if (rdp->last_accelerate == jiffies)
1412 rdp->last_accelerate = jiffies;
1413 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1415 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1416 needwake = rcu_accelerate_cbs(rnp, rdp);
1417 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1419 rcu_gp_kthread_wake();
1424 * Clean up for exit from idle. Attempt to advance callbacks based on
1425 * any grace periods that elapsed while the CPU was idle, and if any
1426 * callbacks are now ready to invoke, initiate invocation.
1428 static void rcu_cleanup_after_idle(void)
1430 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1432 lockdep_assert_irqs_disabled();
1433 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1435 if (rcu_try_advance_all_cbs())
1439 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1441 #ifdef CONFIG_RCU_NOCB_CPU
1444 * Offload callback processing from the boot-time-specified set of CPUs
1445 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1446 * created that pull the callbacks from the corresponding CPU, wait for
1447 * a grace period to elapse, and invoke the callbacks. These kthreads
1448 * are organized into GP kthreads, which manage incoming callbacks, wait for
1449 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1450 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1451 * do a wake_up() on their GP kthread when they insert a callback into any
1452 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1453 * in which case each kthread actively polls its CPU. (Which isn't so great
1454 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1456 * This is intended to be used in conjunction with Frederic Weisbecker's
1457 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1458 * running CPU-bound user-mode computations.
1460 * Offloading of callbacks can also be used as an energy-efficiency
1461 * measure because CPUs with no RCU callbacks queued are more aggressive
1462 * about entering dyntick-idle mode.
1467 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1468 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1469 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1470 * given, a warning is emitted and all CPUs are offloaded.
1472 static int __init rcu_nocb_setup(char *str)
1474 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1475 if (!strcasecmp(str, "all"))
1476 cpumask_setall(rcu_nocb_mask);
1478 if (cpulist_parse(str, rcu_nocb_mask)) {
1479 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1480 cpumask_setall(rcu_nocb_mask);
1484 __setup("rcu_nocbs=", rcu_nocb_setup);
1486 static int __init parse_rcu_nocb_poll(char *arg)
1488 rcu_nocb_poll = true;
1491 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1494 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1495 * After all, the main point of bypassing is to avoid lock contention
1496 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1498 int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1499 module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1502 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1503 * lock isn't immediately available, increment ->nocb_lock_contended to
1504 * flag the contention.
1506 static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1507 __acquires(&rdp->nocb_bypass_lock)
1509 lockdep_assert_irqs_disabled();
1510 if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1512 atomic_inc(&rdp->nocb_lock_contended);
1513 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1514 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1515 raw_spin_lock(&rdp->nocb_bypass_lock);
1516 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1517 atomic_dec(&rdp->nocb_lock_contended);
1521 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1522 * not contended. Please note that this is extremely special-purpose,
1523 * relying on the fact that at most two kthreads and one CPU contend for
1524 * this lock, and also that the two kthreads are guaranteed to have frequent
1525 * grace-period-duration time intervals between successive acquisitions
1526 * of the lock. This allows us to use an extremely simple throttling
1527 * mechanism, and further to apply it only to the CPU doing floods of
1528 * call_rcu() invocations. Don't try this at home!
1530 static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1532 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1533 while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1538 * Conditionally acquire the specified rcu_data structure's
1539 * ->nocb_bypass_lock.
1541 static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1543 lockdep_assert_irqs_disabled();
1544 return raw_spin_trylock(&rdp->nocb_bypass_lock);
1548 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1550 static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1551 __releases(&rdp->nocb_bypass_lock)
1553 lockdep_assert_irqs_disabled();
1554 raw_spin_unlock(&rdp->nocb_bypass_lock);
1558 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1559 * if it corresponds to a no-CBs CPU.
1561 static void rcu_nocb_lock(struct rcu_data *rdp)
1563 lockdep_assert_irqs_disabled();
1564 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1566 raw_spin_lock(&rdp->nocb_lock);
1570 * Release the specified rcu_data structure's ->nocb_lock, but only
1571 * if it corresponds to a no-CBs CPU.
1573 static void rcu_nocb_unlock(struct rcu_data *rdp)
1575 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1576 lockdep_assert_irqs_disabled();
1577 raw_spin_unlock(&rdp->nocb_lock);
1582 * Release the specified rcu_data structure's ->nocb_lock and restore
1583 * interrupts, but only if it corresponds to a no-CBs CPU.
1585 static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1586 unsigned long flags)
1588 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1589 lockdep_assert_irqs_disabled();
1590 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1592 local_irq_restore(flags);
1596 /* Lockdep check that ->cblist may be safely accessed. */
1597 static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1599 lockdep_assert_irqs_disabled();
1600 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1601 lockdep_assert_held(&rdp->nocb_lock);
1605 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1608 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1613 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1615 return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1618 static void rcu_init_one_nocb(struct rcu_node *rnp)
1620 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1621 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1624 /* Is the specified CPU a no-CBs CPU? */
1625 bool rcu_is_nocb_cpu(int cpu)
1627 if (cpumask_available(rcu_nocb_mask))
1628 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1633 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1634 * and this function releases it.
1636 static void wake_nocb_gp(struct rcu_data *rdp, bool force,
1637 unsigned long flags)
1638 __releases(rdp->nocb_lock)
1640 bool needwake = false;
1641 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1643 lockdep_assert_held(&rdp->nocb_lock);
1644 if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1645 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1646 TPS("AlreadyAwake"));
1647 rcu_nocb_unlock_irqrestore(rdp, flags);
1651 if (READ_ONCE(rdp->nocb_defer_wakeup) > RCU_NOCB_WAKE_NOT) {
1652 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
1653 del_timer(&rdp->nocb_timer);
1655 rcu_nocb_unlock_irqrestore(rdp, flags);
1656 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1657 if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1658 WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1660 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1662 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1664 wake_up_process(rdp_gp->nocb_gp_kthread);
1668 * Arrange to wake the GP kthread for this NOCB group at some future
1669 * time when it is safe to do so.
1671 static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1674 if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1675 mod_timer(&rdp->nocb_timer, jiffies + 1);
1676 if (rdp->nocb_defer_wakeup < waketype)
1677 WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1678 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1682 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1683 * However, if there is a callback to be enqueued and if ->nocb_bypass
1684 * proves to be initially empty, just return false because the no-CB GP
1685 * kthread may need to be awakened in this case.
1687 * Note that this function always returns true if rhp is NULL.
1689 static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1692 struct rcu_cblist rcl;
1694 WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp->cblist));
1695 rcu_lockdep_assert_cblist_protected(rdp);
1696 lockdep_assert_held(&rdp->nocb_bypass_lock);
1697 if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1698 raw_spin_unlock(&rdp->nocb_bypass_lock);
1701 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1703 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1704 rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1705 rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1706 WRITE_ONCE(rdp->nocb_bypass_first, j);
1707 rcu_nocb_bypass_unlock(rdp);
1712 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1713 * However, if there is a callback to be enqueued and if ->nocb_bypass
1714 * proves to be initially empty, just return false because the no-CB GP
1715 * kthread may need to be awakened in this case.
1717 * Note that this function always returns true if rhp is NULL.
1719 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1722 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1724 rcu_lockdep_assert_cblist_protected(rdp);
1725 rcu_nocb_bypass_lock(rdp);
1726 return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1730 * If the ->nocb_bypass_lock is immediately available, flush the
1731 * ->nocb_bypass queue into ->cblist.
1733 static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1735 rcu_lockdep_assert_cblist_protected(rdp);
1736 if (!rcu_segcblist_is_offloaded(&rdp->cblist) ||
1737 !rcu_nocb_bypass_trylock(rdp))
1739 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1743 * See whether it is appropriate to use the ->nocb_bypass list in order
1744 * to control contention on ->nocb_lock. A limited number of direct
1745 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1746 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1747 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1748 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1749 * used if ->cblist is empty, because otherwise callbacks can be stranded
1750 * on ->nocb_bypass because we cannot count on the current CPU ever again
1751 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1752 * non-empty, the corresponding no-CBs grace-period kthread must not be
1753 * in an indefinite sleep state.
1755 * Finally, it is not permitted to use the bypass during early boot,
1756 * as doing so would confuse the auto-initialization code. Besides
1757 * which, there is no point in worrying about lock contention while
1758 * there is only one CPU in operation.
1760 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1761 bool *was_alldone, unsigned long flags)
1764 unsigned long cur_gp_seq;
1765 unsigned long j = jiffies;
1766 long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1768 if (!rcu_segcblist_is_offloaded(&rdp->cblist)) {
1769 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1770 return false; /* Not offloaded, no bypassing. */
1772 lockdep_assert_irqs_disabled();
1774 // Don't use ->nocb_bypass during early boot.
1775 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1777 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1778 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1782 // If we have advanced to a new jiffy, reset counts to allow
1783 // moving back from ->nocb_bypass to ->cblist.
1784 if (j == rdp->nocb_nobypass_last) {
1785 c = rdp->nocb_nobypass_count + 1;
1787 WRITE_ONCE(rdp->nocb_nobypass_last, j);
1788 c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1789 if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1790 nocb_nobypass_lim_per_jiffy))
1792 else if (c > nocb_nobypass_lim_per_jiffy)
1793 c = nocb_nobypass_lim_per_jiffy;
1795 WRITE_ONCE(rdp->nocb_nobypass_count, c);
1797 // If there hasn't yet been all that many ->cblist enqueues
1798 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1799 // ->nocb_bypass first.
1800 if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1802 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1804 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1806 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1807 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1808 return false; // Caller must enqueue the callback.
1811 // If ->nocb_bypass has been used too long or is too full,
1812 // flush ->nocb_bypass to ->cblist.
1813 if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1816 if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1817 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1819 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1821 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1822 return false; // Caller must enqueue the callback.
1824 if (j != rdp->nocb_gp_adv_time &&
1825 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1826 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1827 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1828 rdp->nocb_gp_adv_time = j;
1830 rcu_nocb_unlock_irqrestore(rdp, flags);
1831 return true; // Callback already enqueued.
1834 // We need to use the bypass.
1835 rcu_nocb_wait_contended(rdp);
1836 rcu_nocb_bypass_lock(rdp);
1837 ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1838 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1839 rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1841 WRITE_ONCE(rdp->nocb_bypass_first, j);
1842 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1844 rcu_nocb_bypass_unlock(rdp);
1845 smp_mb(); /* Order enqueue before wake. */
1847 local_irq_restore(flags);
1849 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1850 rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1851 if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1852 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1853 TPS("FirstBQwake"));
1854 __call_rcu_nocb_wake(rdp, true, flags);
1856 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1857 TPS("FirstBQnoWake"));
1858 rcu_nocb_unlock_irqrestore(rdp, flags);
1861 return true; // Callback already enqueued.
1865 * Awaken the no-CBs grace-period kthead if needed, either due to it
1866 * legitimately being asleep or due to overload conditions.
1868 * If warranted, also wake up the kthread servicing this CPUs queues.
1870 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1871 unsigned long flags)
1872 __releases(rdp->nocb_lock)
1874 unsigned long cur_gp_seq;
1877 struct task_struct *t;
1879 // If we are being polled or there is no kthread, just leave.
1880 t = READ_ONCE(rdp->nocb_gp_kthread);
1881 if (rcu_nocb_poll || !t) {
1882 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1883 TPS("WakeNotPoll"));
1884 rcu_nocb_unlock_irqrestore(rdp, flags);
1887 // Need to actually to a wakeup.
1888 len = rcu_segcblist_n_cbs(&rdp->cblist);
1890 rdp->qlen_last_fqs_check = len;
1891 if (!irqs_disabled_flags(flags)) {
1892 /* ... if queue was empty ... */
1893 wake_nocb_gp(rdp, false, flags);
1894 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1897 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1898 TPS("WakeEmptyIsDeferred"));
1899 rcu_nocb_unlock_irqrestore(rdp, flags);
1901 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1902 /* ... or if many callbacks queued. */
1903 rdp->qlen_last_fqs_check = len;
1905 if (j != rdp->nocb_gp_adv_time &&
1906 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1907 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1908 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1909 rdp->nocb_gp_adv_time = j;
1911 smp_mb(); /* Enqueue before timer_pending(). */
1912 if ((rdp->nocb_cb_sleep ||
1913 !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1914 !timer_pending(&rdp->nocb_bypass_timer))
1915 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1916 TPS("WakeOvfIsDeferred"));
1917 rcu_nocb_unlock_irqrestore(rdp, flags);
1919 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1920 rcu_nocb_unlock_irqrestore(rdp, flags);
1925 /* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1926 static void do_nocb_bypass_wakeup_timer(struct timer_list *t)
1928 unsigned long flags;
1929 struct rcu_data *rdp = from_timer(rdp, t, nocb_bypass_timer);
1931 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1932 rcu_nocb_lock_irqsave(rdp, flags);
1933 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1934 __call_rcu_nocb_wake(rdp, true, flags);
1938 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1939 * or for grace periods to end.
1941 static void nocb_gp_wait(struct rcu_data *my_rdp)
1943 bool bypass = false;
1945 int __maybe_unused cpu = my_rdp->cpu;
1946 unsigned long cur_gp_seq;
1947 unsigned long flags;
1948 bool gotcbs = false;
1949 unsigned long j = jiffies;
1950 bool needwait_gp = false; // This prevents actual uninitialized use.
1953 struct rcu_data *rdp;
1954 struct rcu_node *rnp;
1955 unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
1956 bool wasempty = false;
1959 * Each pass through the following loop checks for CBs and for the
1960 * nearest grace period (if any) to wait for next. The CB kthreads
1961 * and the global grace-period kthread are awakened if needed.
1963 WARN_ON_ONCE(my_rdp->nocb_gp_rdp != my_rdp);
1964 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1965 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
1966 rcu_nocb_lock_irqsave(rdp, flags);
1967 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1969 (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
1970 bypass_ncbs > 2 * qhimark)) {
1971 // Bypass full or old, so flush it.
1972 (void)rcu_nocb_try_flush_bypass(rdp, j);
1973 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1974 } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
1975 rcu_nocb_unlock_irqrestore(rdp, flags);
1976 continue; /* No callbacks here, try next. */
1979 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1984 if (bypass) { // Avoid race with first bypass CB.
1985 WRITE_ONCE(my_rdp->nocb_defer_wakeup,
1987 del_timer(&my_rdp->nocb_timer);
1989 // Advance callbacks if helpful and low contention.
1990 needwake_gp = false;
1991 if (!rcu_segcblist_restempty(&rdp->cblist,
1992 RCU_NEXT_READY_TAIL) ||
1993 (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1994 rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
1995 raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
1996 needwake_gp = rcu_advance_cbs(rnp, rdp);
1997 wasempty = rcu_segcblist_restempty(&rdp->cblist,
1998 RCU_NEXT_READY_TAIL);
1999 raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
2001 // Need to wait on some grace period?
2002 WARN_ON_ONCE(wasempty &&
2003 !rcu_segcblist_restempty(&rdp->cblist,
2004 RCU_NEXT_READY_TAIL));
2005 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
2007 ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
2008 wait_gp_seq = cur_gp_seq;
2010 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2013 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2014 needwake = rdp->nocb_cb_sleep;
2015 WRITE_ONCE(rdp->nocb_cb_sleep, false);
2016 smp_mb(); /* CB invocation -after- GP end. */
2020 rcu_nocb_unlock_irqrestore(rdp, flags);
2022 swake_up_one(&rdp->nocb_cb_wq);
2026 rcu_gp_kthread_wake();
2029 my_rdp->nocb_gp_bypass = bypass;
2030 my_rdp->nocb_gp_gp = needwait_gp;
2031 my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
2032 if (bypass && !rcu_nocb_poll) {
2033 // At least one child with non-empty ->nocb_bypass, so set
2034 // timer in order to avoid stranding its callbacks.
2035 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2036 mod_timer(&my_rdp->nocb_bypass_timer, j + 2);
2037 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2039 if (rcu_nocb_poll) {
2040 /* Polling, so trace if first poll in the series. */
2042 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2043 schedule_timeout_idle(1);
2044 } else if (!needwait_gp) {
2045 /* Wait for callbacks to appear. */
2046 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2047 swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2048 !READ_ONCE(my_rdp->nocb_gp_sleep));
2049 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2051 rnp = my_rdp->mynode;
2052 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2053 swait_event_interruptible_exclusive(
2054 rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2055 rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2056 !READ_ONCE(my_rdp->nocb_gp_sleep));
2057 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2059 if (!rcu_nocb_poll) {
2060 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2062 del_timer(&my_rdp->nocb_bypass_timer);
2063 WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2064 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2066 my_rdp->nocb_gp_seq = -1;
2067 WARN_ON(signal_pending(current));
2071 * No-CBs grace-period-wait kthread. There is one of these per group
2072 * of CPUs, but only once at least one CPU in that group has come online
2073 * at least once since boot. This kthread checks for newly posted
2074 * callbacks from any of the CPUs it is responsible for, waits for a
2075 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2076 * that then have callback-invocation work to do.
2078 static int rcu_nocb_gp_kthread(void *arg)
2080 struct rcu_data *rdp = arg;
2083 WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2085 cond_resched_tasks_rcu_qs();
2091 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2092 * then, if there are no more, wait for more to appear.
2094 static void nocb_cb_wait(struct rcu_data *rdp)
2096 unsigned long cur_gp_seq;
2097 unsigned long flags;
2098 bool needwake_gp = false;
2099 struct rcu_node *rnp = rdp->mynode;
2101 local_irq_save(flags);
2102 rcu_momentary_dyntick_idle();
2103 local_irq_restore(flags);
2107 lockdep_assert_irqs_enabled();
2108 rcu_nocb_lock_irqsave(rdp, flags);
2109 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2110 rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2111 raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2112 needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2113 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2115 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2116 rcu_nocb_unlock_irqrestore(rdp, flags);
2118 rcu_gp_kthread_wake();
2122 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2123 WRITE_ONCE(rdp->nocb_cb_sleep, true);
2124 rcu_nocb_unlock_irqrestore(rdp, flags);
2126 rcu_gp_kthread_wake();
2127 swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2128 !READ_ONCE(rdp->nocb_cb_sleep));
2129 if (!smp_load_acquire(&rdp->nocb_cb_sleep)) { /* VVV */
2130 /* ^^^ Ensure CB invocation follows _sleep test. */
2133 WARN_ON(signal_pending(current));
2134 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2138 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2139 * nocb_cb_wait() to do the dirty work.
2141 static int rcu_nocb_cb_kthread(void *arg)
2143 struct rcu_data *rdp = arg;
2145 // Each pass through this loop does one callback batch, and,
2146 // if there are no more ready callbacks, waits for them.
2149 cond_resched_tasks_rcu_qs();
2154 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2155 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2157 return READ_ONCE(rdp->nocb_defer_wakeup);
2160 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2161 static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2163 unsigned long flags;
2166 rcu_nocb_lock_irqsave(rdp, flags);
2167 if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2168 rcu_nocb_unlock_irqrestore(rdp, flags);
2171 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2172 wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2173 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2176 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2177 static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2179 struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2181 do_nocb_deferred_wakeup_common(rdp);
2185 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2186 * This means we do an inexact common-case check. Note that if
2187 * we miss, ->nocb_timer will eventually clean things up.
2189 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2191 if (rcu_nocb_need_deferred_wakeup(rdp))
2192 do_nocb_deferred_wakeup_common(rdp);
2195 void rcu_nocb_flush_deferred_wakeup(void)
2197 do_nocb_deferred_wakeup(this_cpu_ptr(&rcu_data));
2200 void __init rcu_init_nohz(void)
2203 bool need_rcu_nocb_mask = false;
2204 struct rcu_data *rdp;
2206 #if defined(CONFIG_NO_HZ_FULL)
2207 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2208 need_rcu_nocb_mask = true;
2209 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2211 if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2212 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2213 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2217 if (!cpumask_available(rcu_nocb_mask))
2220 #if defined(CONFIG_NO_HZ_FULL)
2221 if (tick_nohz_full_running)
2222 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2223 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2225 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2226 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2227 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2230 if (cpumask_empty(rcu_nocb_mask))
2231 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2233 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2234 cpumask_pr_args(rcu_nocb_mask));
2236 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2238 for_each_cpu(cpu, rcu_nocb_mask) {
2239 rdp = per_cpu_ptr(&rcu_data, cpu);
2240 if (rcu_segcblist_empty(&rdp->cblist))
2241 rcu_segcblist_init(&rdp->cblist);
2242 rcu_segcblist_offload(&rdp->cblist);
2244 rcu_organize_nocb_kthreads();
2247 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2248 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2250 init_swait_queue_head(&rdp->nocb_cb_wq);
2251 init_swait_queue_head(&rdp->nocb_gp_wq);
2252 raw_spin_lock_init(&rdp->nocb_lock);
2253 raw_spin_lock_init(&rdp->nocb_bypass_lock);
2254 raw_spin_lock_init(&rdp->nocb_gp_lock);
2255 timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2256 timer_setup(&rdp->nocb_bypass_timer, do_nocb_bypass_wakeup_timer, 0);
2257 rcu_cblist_init(&rdp->nocb_bypass);
2261 * If the specified CPU is a no-CBs CPU that does not already have its
2262 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2263 * for this CPU's group has not yet been created, spawn it as well.
2265 static void rcu_spawn_one_nocb_kthread(int cpu)
2267 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2268 struct rcu_data *rdp_gp;
2269 struct task_struct *t;
2272 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2273 * then nothing to do.
2275 if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2278 /* If we didn't spawn the GP kthread first, reorganize! */
2279 rdp_gp = rdp->nocb_gp_rdp;
2280 if (!rdp_gp->nocb_gp_kthread) {
2281 t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2282 "rcuog/%d", rdp_gp->cpu);
2283 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2285 WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2288 /* Spawn the kthread for this CPU. */
2289 t = kthread_run(rcu_nocb_cb_kthread, rdp,
2290 "rcuo%c/%d", rcu_state.abbr, cpu);
2291 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2293 WRITE_ONCE(rdp->nocb_cb_kthread, t);
2294 WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2298 * If the specified CPU is a no-CBs CPU that does not already have its
2299 * rcuo kthread, spawn it.
2301 static void rcu_spawn_cpu_nocb_kthread(int cpu)
2303 if (rcu_scheduler_fully_active)
2304 rcu_spawn_one_nocb_kthread(cpu);
2308 * Once the scheduler is running, spawn rcuo kthreads for all online
2309 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2310 * non-boot CPUs come online -- if this changes, we will need to add
2311 * some mutual exclusion.
2313 static void __init rcu_spawn_nocb_kthreads(void)
2317 for_each_online_cpu(cpu)
2318 rcu_spawn_cpu_nocb_kthread(cpu);
2321 /* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2322 static int rcu_nocb_gp_stride = -1;
2323 module_param(rcu_nocb_gp_stride, int, 0444);
2326 * Initialize GP-CB relationships for all no-CBs CPU.
2328 static void __init rcu_organize_nocb_kthreads(void)
2331 bool firsttime = true;
2332 bool gotnocbs = false;
2333 bool gotnocbscbs = true;
2334 int ls = rcu_nocb_gp_stride;
2335 int nl = 0; /* Next GP kthread. */
2336 struct rcu_data *rdp;
2337 struct rcu_data *rdp_gp = NULL; /* Suppress misguided gcc warn. */
2338 struct rcu_data *rdp_prev = NULL;
2340 if (!cpumask_available(rcu_nocb_mask))
2343 ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2344 rcu_nocb_gp_stride = ls;
2348 * Each pass through this loop sets up one rcu_data structure.
2349 * Should the corresponding CPU come online in the future, then
2350 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2352 for_each_cpu(cpu, rcu_nocb_mask) {
2353 rdp = per_cpu_ptr(&rcu_data, cpu);
2354 if (rdp->cpu >= nl) {
2355 /* New GP kthread, set up for CBs & next GP. */
2357 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2358 rdp->nocb_gp_rdp = rdp;
2362 pr_cont("%s\n", gotnocbscbs
2363 ? "" : " (self only)");
2364 gotnocbscbs = false;
2366 pr_alert("%s: No-CB GP kthread CPU %d:",
2370 /* Another CB kthread, link to previous GP kthread. */
2372 rdp->nocb_gp_rdp = rdp_gp;
2373 rdp_prev->nocb_next_cb_rdp = rdp;
2375 pr_cont(" %d", cpu);
2379 if (gotnocbs && dump_tree)
2380 pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
2384 * Bind the current task to the offloaded CPUs. If there are no offloaded
2385 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2387 void rcu_bind_current_to_nocb(void)
2389 if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2390 WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2392 EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2395 * Dump out nocb grace-period kthread state for the specified rcu_data
2398 static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2400 struct rcu_node *rnp = rdp->mynode;
2402 pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2404 "kK"[!!rdp->nocb_gp_kthread],
2405 "lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2406 "dD"[!!rdp->nocb_defer_wakeup],
2407 "tT"[timer_pending(&rdp->nocb_timer)],
2408 "bB"[timer_pending(&rdp->nocb_bypass_timer)],
2409 "sS"[!!rdp->nocb_gp_sleep],
2410 ".W"[swait_active(&rdp->nocb_gp_wq)],
2411 ".W"[swait_active(&rnp->nocb_gp_wq[0])],
2412 ".W"[swait_active(&rnp->nocb_gp_wq[1])],
2413 ".B"[!!rdp->nocb_gp_bypass],
2414 ".G"[!!rdp->nocb_gp_gp],
2415 (long)rdp->nocb_gp_seq,
2416 rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops));
2419 /* Dump out nocb kthread state for the specified rcu_data structure. */
2420 static void show_rcu_nocb_state(struct rcu_data *rdp)
2422 struct rcu_segcblist *rsclp = &rdp->cblist;
2427 if (rdp->nocb_gp_rdp == rdp)
2428 show_rcu_nocb_gp_state(rdp);
2430 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",
2431 rdp->cpu, rdp->nocb_gp_rdp->cpu,
2432 "kK"[!!rdp->nocb_cb_kthread],
2433 "bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2434 "cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2435 "lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2436 "sS"[!!rdp->nocb_cb_sleep],
2437 ".W"[swait_active(&rdp->nocb_cb_wq)],
2438 jiffies - rdp->nocb_bypass_first,
2439 jiffies - rdp->nocb_nobypass_last,
2440 rdp->nocb_nobypass_count,
2441 ".D"[rcu_segcblist_ready_cbs(rsclp)],
2442 ".W"[!rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL)],
2443 ".R"[!rcu_segcblist_restempty(rsclp, RCU_WAIT_TAIL)],
2444 ".N"[!rcu_segcblist_restempty(rsclp, RCU_NEXT_READY_TAIL)],
2445 ".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2446 rcu_segcblist_n_cbs(&rdp->cblist));
2448 /* It is OK for GP kthreads to have GP state. */
2449 if (rdp->nocb_gp_rdp == rdp)
2452 waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2453 wastimer = timer_pending(&rdp->nocb_bypass_timer);
2454 wassleep = swait_active(&rdp->nocb_gp_wq);
2455 if (!rdp->nocb_gp_sleep && !waslocked && !wastimer && !wassleep)
2456 return; /* Nothing untowards. */
2458 pr_info(" nocb GP activity on CB-only CPU!!! %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 time_before(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 __always_inline 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 __always_inline 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) */
2595 /* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
2596 static __always_inline void rcu_dynticks_task_trace_enter(void)
2598 #ifdef CONFIG_TASKS_TRACE_RCU
2599 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2600 current->trc_reader_special.b.need_mb = true;
2601 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
2604 /* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
2605 static __always_inline void rcu_dynticks_task_trace_exit(void)
2607 #ifdef CONFIG_TASKS_TRACE_RCU
2608 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2609 current->trc_reader_special.b.need_mb = false;
2610 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */