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 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 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)
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);
526 /* Unboost if we were boosted. */
527 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
528 rt_mutex_futex_unlock(&rnp->boost_mtx);
531 * If this was the last task on the expedited lists,
532 * then we need to report up the rcu_node hierarchy.
534 if (!empty_exp && empty_exp_now)
535 rcu_report_exp_rnp(rnp, true);
537 local_irq_restore(flags);
542 * Is a deferred quiescent-state pending, and are we also not in
543 * an RCU read-side critical section? It is the caller's responsibility
544 * to ensure it is otherwise safe to report any deferred quiescent
545 * states. The reason for this is that it is safe to report a
546 * quiescent state during context switch even though preemption
547 * is disabled. This function cannot be expected to understand these
548 * nuances, so the caller must handle them.
550 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
552 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
553 READ_ONCE(t->rcu_read_unlock_special.s)) &&
554 t->rcu_read_lock_nesting <= 0;
558 * Report a deferred quiescent state if needed and safe to do so.
559 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
560 * not being in an RCU read-side critical section. The caller must
561 * evaluate safety in terms of interrupt, softirq, and preemption
564 static void rcu_preempt_deferred_qs(struct task_struct *t)
567 bool couldrecurse = t->rcu_read_lock_nesting >= 0;
569 if (!rcu_preempt_need_deferred_qs(t))
572 t->rcu_read_lock_nesting -= RCU_NEST_BIAS;
573 local_irq_save(flags);
574 rcu_preempt_deferred_qs_irqrestore(t, flags);
576 t->rcu_read_lock_nesting += RCU_NEST_BIAS;
580 * Minimal handler to give the scheduler a chance to re-evaluate.
582 static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
584 struct rcu_data *rdp;
586 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
587 rdp->defer_qs_iw_pending = false;
591 * Handle special cases during rcu_read_unlock(), such as needing to
592 * notify RCU core processing or task having blocked during the RCU
593 * read-side critical section.
595 static void rcu_read_unlock_special(struct task_struct *t)
598 bool preempt_bh_were_disabled =
599 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
600 bool irqs_were_disabled;
602 /* NMI handlers cannot block and cannot safely manipulate state. */
606 local_irq_save(flags);
607 irqs_were_disabled = irqs_disabled_flags(flags);
608 if (preempt_bh_were_disabled || irqs_were_disabled) {
610 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
611 struct rcu_node *rnp = rdp->mynode;
613 t->rcu_read_unlock_special.b.exp_hint = false;
614 exp = (t->rcu_blocked_node && t->rcu_blocked_node->exp_tasks) ||
615 (rdp->grpmask & rnp->expmask) ||
616 tick_nohz_full_cpu(rdp->cpu);
617 // Need to defer quiescent state until everything is enabled.
618 if (irqs_were_disabled && use_softirq &&
620 (exp && !t->rcu_read_unlock_special.b.deferred_qs))) {
621 // Using softirq, safe to awaken, and we get
622 // no help from enabling irqs, unlike bh/preempt.
623 raise_softirq_irqoff(RCU_SOFTIRQ);
625 // Enabling BH or preempt does reschedule, so...
626 // Also if no expediting or NO_HZ_FULL, slow is OK.
627 set_tsk_need_resched(current);
628 set_preempt_need_resched();
629 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
630 !rdp->defer_qs_iw_pending && exp) {
631 // Get scheduler to re-evaluate and call hooks.
632 // If !IRQ_WORK, FQS scan will eventually IPI.
633 init_irq_work(&rdp->defer_qs_iw,
634 rcu_preempt_deferred_qs_handler);
635 rdp->defer_qs_iw_pending = true;
636 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
639 t->rcu_read_unlock_special.b.deferred_qs = true;
640 local_irq_restore(flags);
643 WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
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, and the rnp's ->lock
652 * must be held by the caller.
654 * Also, if there are blocked tasks on the list, they automatically
655 * block the newly created grace period, so set up ->gp_tasks accordingly.
657 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
659 struct task_struct *t;
661 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
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 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 if (user || rcu_is_cpu_rrupt_from_idle()) {
687 rcu_note_voluntary_context_switch(current);
689 if (t->rcu_read_lock_nesting > 0 ||
690 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
691 /* No QS, force context switch if deferred. */
692 if (rcu_preempt_need_deferred_qs(t)) {
693 set_tsk_need_resched(t);
694 set_preempt_need_resched();
696 } else if (rcu_preempt_need_deferred_qs(t)) {
697 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
699 } else if (!t->rcu_read_lock_nesting) {
700 rcu_qs(); /* Report immediate QS. */
704 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
705 if (t->rcu_read_lock_nesting > 0 &&
706 __this_cpu_read(rcu_data.core_needs_qs) &&
707 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
708 !t->rcu_read_unlock_special.b.need_qs &&
709 time_after(jiffies, rcu_state.gp_start + HZ))
710 t->rcu_read_unlock_special.b.need_qs = true;
714 * Check for a task exiting while in a preemptible-RCU read-side
715 * critical section, clean up if so. No need to issue warnings, as
716 * debug_check_no_locks_held() already does this if lockdep is enabled.
717 * Besides, if this function does anything other than just immediately
718 * return, there was a bug of some sort. Spewing warnings from this
719 * function is like as not to simply obscure important prior warnings.
723 struct task_struct *t = current;
725 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
726 t->rcu_read_lock_nesting = 1;
728 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
729 } else if (unlikely(t->rcu_read_lock_nesting)) {
730 t->rcu_read_lock_nesting = 1;
735 rcu_preempt_deferred_qs(current);
739 * Dump the blocked-tasks state, but limit the list dump to the
740 * specified number of elements.
743 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
747 struct list_head *lhp;
749 struct rcu_data *rdp;
750 struct rcu_node *rnp1;
752 raw_lockdep_assert_held_rcu_node(rnp);
753 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
754 __func__, rnp->grplo, rnp->grphi, rnp->level,
755 (long)rnp->gp_seq, (long)rnp->completedqs);
756 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
757 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
758 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
759 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
760 __func__, rnp->gp_tasks, rnp->boost_tasks, rnp->exp_tasks);
761 pr_info("%s: ->blkd_tasks", __func__);
763 list_for_each(lhp, &rnp->blkd_tasks) {
769 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
770 rdp = per_cpu_ptr(&rcu_data, cpu);
771 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
772 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
774 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
775 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
779 #else /* #ifdef CONFIG_PREEMPT_RCU */
782 * Tell them what RCU they are running.
784 static void __init rcu_bootup_announce(void)
786 pr_info("Hierarchical RCU implementation.\n");
787 rcu_bootup_announce_oddness();
791 * Note a quiescent state for PREEMPT=n. Because we do not need to know
792 * how many quiescent states passed, just if there was at least one since
793 * the start of the grace period, this just sets a flag. The caller must
794 * have disabled preemption.
796 static void rcu_qs(void)
798 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
799 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
801 trace_rcu_grace_period(TPS("rcu_sched"),
802 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
803 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
804 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
806 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
807 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
811 * Register an urgently needed quiescent state. If there is an
812 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
813 * dyntick-idle quiescent state visible to other CPUs, which will in
814 * some cases serve for expedited as well as normal grace periods.
815 * Either way, register a lightweight quiescent state.
817 void rcu_all_qs(void)
821 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
824 /* Load rcu_urgent_qs before other flags. */
825 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
829 this_cpu_write(rcu_data.rcu_urgent_qs, false);
830 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
831 local_irq_save(flags);
832 rcu_momentary_dyntick_idle();
833 local_irq_restore(flags);
838 EXPORT_SYMBOL_GPL(rcu_all_qs);
841 * Note a PREEMPT=n context switch. The caller must have disabled interrupts.
843 void rcu_note_context_switch(bool preempt)
845 trace_rcu_utilization(TPS("Start context switch"));
847 /* Load rcu_urgent_qs before other flags. */
848 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
850 this_cpu_write(rcu_data.rcu_urgent_qs, false);
851 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
852 rcu_momentary_dyntick_idle();
854 rcu_tasks_qs(current);
856 trace_rcu_utilization(TPS("End context switch"));
858 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
861 * Because preemptible RCU does not exist, there are never any preempted
864 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
870 * Because there is no preemptible RCU, there can be no readers blocked.
872 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
878 * Because there is no preemptible RCU, there can be no deferred quiescent
881 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
885 static void rcu_preempt_deferred_qs(struct task_struct *t) { }
888 * Because there is no preemptible RCU, there can be no readers blocked,
889 * so there is no need to check for blocked tasks. So check only for
890 * bogus qsmask values.
892 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
894 WARN_ON_ONCE(rnp->qsmask);
898 * Check to see if this CPU is in a non-context-switch quiescent state,
899 * namely user mode and idle loop.
901 static void rcu_flavor_sched_clock_irq(int user)
903 if (user || rcu_is_cpu_rrupt_from_idle()) {
906 * Get here if this CPU took its interrupt from user
907 * mode or from the idle loop, and if this is not a
908 * nested interrupt. In this case, the CPU is in
909 * a quiescent state, so note it.
911 * No memory barrier is required here because rcu_qs()
912 * references only CPU-local variables that other CPUs
913 * neither access nor modify, at least not while the
914 * corresponding CPU is online.
922 * Because preemptible RCU does not exist, tasks cannot possibly exit
923 * while in preemptible RCU read-side critical sections.
930 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
933 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
935 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
938 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
941 * If boosting, set rcuc kthreads to realtime priority.
943 static void rcu_cpu_kthread_setup(unsigned int cpu)
945 #ifdef CONFIG_RCU_BOOST
946 struct sched_param sp;
948 sp.sched_priority = kthread_prio;
949 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
950 #endif /* #ifdef CONFIG_RCU_BOOST */
953 #ifdef CONFIG_RCU_BOOST
956 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
957 * or ->boost_tasks, advancing the pointer to the next task in the
960 * Note that irqs must be enabled: boosting the task can block.
961 * Returns 1 if there are more tasks needing to be boosted.
963 static int rcu_boost(struct rcu_node *rnp)
966 struct task_struct *t;
967 struct list_head *tb;
969 if (READ_ONCE(rnp->exp_tasks) == NULL &&
970 READ_ONCE(rnp->boost_tasks) == NULL)
971 return 0; /* Nothing left to boost. */
973 raw_spin_lock_irqsave_rcu_node(rnp, flags);
976 * Recheck under the lock: all tasks in need of boosting
977 * might exit their RCU read-side critical sections on their own.
979 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
980 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
985 * Preferentially boost tasks blocking expedited grace periods.
986 * This cannot starve the normal grace periods because a second
987 * expedited grace period must boost all blocked tasks, including
988 * those blocking the pre-existing normal grace period.
990 if (rnp->exp_tasks != NULL)
993 tb = rnp->boost_tasks;
996 * We boost task t by manufacturing an rt_mutex that appears to
997 * be held by task t. We leave a pointer to that rt_mutex where
998 * task t can find it, and task t will release the mutex when it
999 * exits its outermost RCU read-side critical section. Then
1000 * simply acquiring this artificial rt_mutex will boost task
1001 * t's priority. (Thanks to tglx for suggesting this approach!)
1003 * Note that task t must acquire rnp->lock to remove itself from
1004 * the ->blkd_tasks list, which it will do from exit() if from
1005 * nowhere else. We therefore are guaranteed that task t will
1006 * stay around at least until we drop rnp->lock. Note that
1007 * rnp->lock also resolves races between our priority boosting
1008 * and task t's exiting its outermost RCU read-side critical
1011 t = container_of(tb, struct task_struct, rcu_node_entry);
1012 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1013 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1014 /* Lock only for side effect: boosts task t's priority. */
1015 rt_mutex_lock(&rnp->boost_mtx);
1016 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1018 return READ_ONCE(rnp->exp_tasks) != NULL ||
1019 READ_ONCE(rnp->boost_tasks) != NULL;
1023 * Priority-boosting kthread, one per leaf rcu_node.
1025 static int rcu_boost_kthread(void *arg)
1027 struct rcu_node *rnp = (struct rcu_node *)arg;
1031 trace_rcu_utilization(TPS("Start boost kthread@init"));
1033 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1034 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1035 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1036 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1037 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1038 more2boost = rcu_boost(rnp);
1044 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1045 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1046 schedule_timeout_interruptible(2);
1047 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1052 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1057 * Check to see if it is time to start boosting RCU readers that are
1058 * blocking the current grace period, and, if so, tell the per-rcu_node
1059 * kthread to start boosting them. If there is an expedited grace
1060 * period in progress, it is always time to boost.
1062 * The caller must hold rnp->lock, which this function releases.
1063 * The ->boost_kthread_task is immortal, so we don't need to worry
1064 * about it going away.
1066 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1067 __releases(rnp->lock)
1069 raw_lockdep_assert_held_rcu_node(rnp);
1070 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1071 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1074 if (rnp->exp_tasks != NULL ||
1075 (rnp->gp_tasks != NULL &&
1076 rnp->boost_tasks == NULL &&
1078 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1079 if (rnp->exp_tasks == NULL)
1080 rnp->boost_tasks = rnp->gp_tasks;
1081 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1082 rcu_wake_cond(rnp->boost_kthread_task,
1083 rnp->boost_kthread_status);
1085 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1090 * Is the current CPU running the RCU-callbacks kthread?
1091 * Caller must have preemption disabled.
1093 static bool rcu_is_callbacks_kthread(void)
1095 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1098 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1101 * Do priority-boost accounting for the start of a new grace period.
1103 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1105 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1109 * Create an RCU-boost kthread for the specified node if one does not
1110 * already exist. We only create this kthread for preemptible RCU.
1111 * Returns zero if all is well, a negated errno otherwise.
1113 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1115 int rnp_index = rnp - rcu_get_root();
1116 unsigned long flags;
1117 struct sched_param sp;
1118 struct task_struct *t;
1120 if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1123 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1126 rcu_state.boost = 1;
1128 if (rnp->boost_kthread_task != NULL)
1131 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1132 "rcub/%d", rnp_index);
1133 if (WARN_ON_ONCE(IS_ERR(t)))
1136 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1137 rnp->boost_kthread_task = t;
1138 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1139 sp.sched_priority = kthread_prio;
1140 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1141 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1145 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1146 * served by the rcu_node in question. The CPU hotplug lock is still
1147 * held, so the value of rnp->qsmaskinit will be stable.
1149 * We don't include outgoingcpu in the affinity set, use -1 if there is
1150 * no outgoing CPU. If there are no CPUs left in the affinity set,
1151 * this function allows the kthread to execute on any CPU.
1153 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1155 struct task_struct *t = rnp->boost_kthread_task;
1156 unsigned long mask = rcu_rnp_online_cpus(rnp);
1162 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1164 for_each_leaf_node_possible_cpu(rnp, cpu)
1165 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1167 cpumask_set_cpu(cpu, cm);
1168 if (cpumask_weight(cm) == 0)
1170 set_cpus_allowed_ptr(t, cm);
1171 free_cpumask_var(cm);
1175 * Spawn boost kthreads -- called as soon as the scheduler is running.
1177 static void __init rcu_spawn_boost_kthreads(void)
1179 struct rcu_node *rnp;
1181 rcu_for_each_leaf_node(rnp)
1182 rcu_spawn_one_boost_kthread(rnp);
1185 static void rcu_prepare_kthreads(int cpu)
1187 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1188 struct rcu_node *rnp = rdp->mynode;
1190 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1191 if (rcu_scheduler_fully_active)
1192 rcu_spawn_one_boost_kthread(rnp);
1195 #else /* #ifdef CONFIG_RCU_BOOST */
1197 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1198 __releases(rnp->lock)
1200 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1203 static bool rcu_is_callbacks_kthread(void)
1208 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1212 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1216 static void __init rcu_spawn_boost_kthreads(void)
1220 static void rcu_prepare_kthreads(int cpu)
1224 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1226 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1229 * Check to see if any future RCU-related work will need to be done
1230 * by the current CPU, even if none need be done immediately, returning
1231 * 1 if so. This function is part of the RCU implementation; it is -not-
1232 * an exported member of the RCU API.
1234 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1235 * CPU has RCU callbacks queued.
1237 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1239 *nextevt = KTIME_MAX;
1240 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist);
1244 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1247 static void rcu_cleanup_after_idle(void)
1252 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1255 static void rcu_prepare_for_idle(void)
1259 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1262 * This code is invoked when a CPU goes idle, at which point we want
1263 * to have the CPU do everything required for RCU so that it can enter
1264 * the energy-efficient dyntick-idle mode. This is handled by a
1265 * state machine implemented by rcu_prepare_for_idle() below.
1267 * The following three proprocessor symbols control this state machine:
1269 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1270 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1271 * is sized to be roughly one RCU grace period. Those energy-efficiency
1272 * benchmarkers who might otherwise be tempted to set this to a large
1273 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1274 * system. And if you are -that- concerned about energy efficiency,
1275 * just power the system down and be done with it!
1276 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1277 * permitted to sleep in dyntick-idle mode with only lazy RCU
1278 * callbacks pending. Setting this too high can OOM your system.
1280 * The values below work well in practice. If future workloads require
1281 * adjustment, they can be converted into kernel config parameters, though
1282 * making the state machine smarter might be a better option.
1284 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1285 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1287 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1288 module_param(rcu_idle_gp_delay, int, 0644);
1289 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1290 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1293 * Try to advance callbacks on the current CPU, but only if it has been
1294 * awhile since the last time we did so. Afterwards, if there are any
1295 * callbacks ready for immediate invocation, return true.
1297 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1299 bool cbs_ready = false;
1300 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1301 struct rcu_node *rnp;
1303 /* Exit early if we advanced recently. */
1304 if (jiffies == rdp->last_advance_all)
1306 rdp->last_advance_all = jiffies;
1311 * Don't bother checking unless a grace period has
1312 * completed since we last checked and there are
1313 * callbacks not yet ready to invoke.
1315 if ((rcu_seq_completed_gp(rdp->gp_seq,
1316 rcu_seq_current(&rnp->gp_seq)) ||
1317 unlikely(READ_ONCE(rdp->gpwrap))) &&
1318 rcu_segcblist_pend_cbs(&rdp->cblist))
1319 note_gp_changes(rdp);
1321 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1327 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1328 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1329 * caller to set the timeout based on whether or not there are non-lazy
1332 * The caller must have disabled interrupts.
1334 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1336 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1339 lockdep_assert_irqs_disabled();
1341 /* If no callbacks, RCU doesn't need the CPU. */
1342 if (rcu_segcblist_empty(&rdp->cblist)) {
1343 *nextevt = KTIME_MAX;
1347 /* Attempt to advance callbacks. */
1348 if (rcu_try_advance_all_cbs()) {
1349 /* Some ready to invoke, so initiate later invocation. */
1353 rdp->last_accelerate = jiffies;
1355 /* Request timer delay depending on laziness, and round. */
1356 rdp->all_lazy = !rcu_segcblist_n_nonlazy_cbs(&rdp->cblist);
1357 if (rdp->all_lazy) {
1358 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1360 dj = round_up(rcu_idle_gp_delay + jiffies,
1361 rcu_idle_gp_delay) - jiffies;
1363 *nextevt = basemono + dj * TICK_NSEC;
1368 * Prepare a CPU for idle from an RCU perspective. The first major task
1369 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1370 * The second major task is to check to see if a non-lazy callback has
1371 * arrived at a CPU that previously had only lazy callbacks. The third
1372 * major task is to accelerate (that is, assign grace-period numbers to)
1373 * any recently arrived callbacks.
1375 * The caller must have disabled interrupts.
1377 static void rcu_prepare_for_idle(void)
1380 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1381 struct rcu_node *rnp;
1384 lockdep_assert_irqs_disabled();
1385 if (rcu_is_nocb_cpu(smp_processor_id()))
1388 /* Handle nohz enablement switches conservatively. */
1389 tne = READ_ONCE(tick_nohz_active);
1390 if (tne != rdp->tick_nohz_enabled_snap) {
1391 if (!rcu_segcblist_empty(&rdp->cblist))
1392 invoke_rcu_core(); /* force nohz to see update. */
1393 rdp->tick_nohz_enabled_snap = tne;
1400 * If a non-lazy callback arrived at a CPU having only lazy
1401 * callbacks, invoke RCU core for the side-effect of recalculating
1402 * idle duration on re-entry to idle.
1404 if (rdp->all_lazy && rcu_segcblist_n_nonlazy_cbs(&rdp->cblist)) {
1405 rdp->all_lazy = false;
1411 * If we have not yet accelerated this jiffy, accelerate all
1412 * callbacks on this CPU.
1414 if (rdp->last_accelerate == jiffies)
1416 rdp->last_accelerate = jiffies;
1417 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1419 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1420 needwake = rcu_accelerate_cbs(rnp, rdp);
1421 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1423 rcu_gp_kthread_wake();
1428 * Clean up for exit from idle. Attempt to advance callbacks based on
1429 * any grace periods that elapsed while the CPU was idle, and if any
1430 * callbacks are now ready to invoke, initiate invocation.
1432 static void rcu_cleanup_after_idle(void)
1434 lockdep_assert_irqs_disabled();
1435 if (rcu_is_nocb_cpu(smp_processor_id()))
1437 if (rcu_try_advance_all_cbs())
1441 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1443 #ifdef CONFIG_RCU_NOCB_CPU
1446 * Offload callback processing from the boot-time-specified set of CPUs
1447 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1448 * created that pull the callbacks from the corresponding CPU, wait for
1449 * a grace period to elapse, and invoke the callbacks. These kthreads
1450 * are organized into GP kthreads, which manage incoming callbacks, wait for
1451 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1452 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1453 * do a wake_up() on their GP kthread when they insert a callback into any
1454 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1455 * in which case each kthread actively polls its CPU. (Which isn't so great
1456 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1458 * This is intended to be used in conjunction with Frederic Weisbecker's
1459 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1460 * running CPU-bound user-mode computations.
1462 * Offloading of callbacks can also be used as an energy-efficiency
1463 * measure because CPUs with no RCU callbacks queued are more aggressive
1464 * about entering dyntick-idle mode.
1469 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1470 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1471 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1472 * given, a warning is emitted and all CPUs are offloaded.
1474 static int __init rcu_nocb_setup(char *str)
1476 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1477 if (!strcasecmp(str, "all"))
1478 cpumask_setall(rcu_nocb_mask);
1480 if (cpulist_parse(str, rcu_nocb_mask)) {
1481 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1482 cpumask_setall(rcu_nocb_mask);
1486 __setup("rcu_nocbs=", rcu_nocb_setup);
1488 static int __init parse_rcu_nocb_poll(char *arg)
1490 rcu_nocb_poll = true;
1493 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1496 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1499 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1504 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1506 return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1509 static void rcu_init_one_nocb(struct rcu_node *rnp)
1511 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1512 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1515 /* Is the specified CPU a no-CBs CPU? */
1516 bool rcu_is_nocb_cpu(int cpu)
1518 if (cpumask_available(rcu_nocb_mask))
1519 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1524 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1525 * and this function releases it.
1527 static void __wake_nocb_gp(struct rcu_data *rdp, bool force,
1528 unsigned long flags)
1529 __releases(rdp->nocb_lock)
1531 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1533 lockdep_assert_held(&rdp->nocb_lock);
1534 if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1535 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1538 if (rdp_gp->nocb_gp_sleep || force) {
1539 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1540 WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1541 del_timer(&rdp->nocb_timer);
1542 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1543 smp_mb(); /* ->nocb_gp_sleep before swake_up_one(). */
1544 swake_up_one(&rdp_gp->nocb_gp_wq);
1546 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1551 * Kick the GP kthread for this NOCB group, but caller has not
1554 static void wake_nocb_gp(struct rcu_data *rdp, bool force)
1556 unsigned long flags;
1558 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1559 __wake_nocb_gp(rdp, force, flags);
1563 * Arrange to wake the GP kthread for this NOCB group at some future
1564 * time when it is safe to do so.
1566 static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1569 unsigned long flags;
1571 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1572 if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1573 mod_timer(&rdp->nocb_timer, jiffies + 1);
1574 WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1575 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1576 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1579 /* Does rcu_barrier need to queue an RCU callback on the specified CPU? */
1580 static bool rcu_nocb_cpu_needs_barrier(int cpu)
1582 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1584 #ifdef CONFIG_PROVE_RCU
1585 struct rcu_head *rhp;
1586 #endif /* #ifdef CONFIG_PROVE_RCU */
1589 * Check count of all no-CBs callbacks awaiting invocation.
1590 * There needs to be a barrier before this function is called,
1591 * but associated with a prior determination that no more
1592 * callbacks would be posted. In the worst case, the first
1593 * barrier in rcu_barrier() suffices (but the caller cannot
1594 * necessarily rely on this, not a substitute for the caller
1595 * getting the concurrency design right!). There must also be a
1596 * barrier between the following load and posting of a callback
1597 * (if a callback is in fact needed). This is associated with an
1598 * atomic_inc() in the caller.
1600 ret = rcu_get_n_cbs_nocb_cpu(rdp);
1602 #ifdef CONFIG_PROVE_RCU
1603 rhp = READ_ONCE(rdp->nocb_head);
1605 rhp = READ_ONCE(rdp->nocb_gp_head);
1607 rhp = READ_ONCE(rdp->nocb_cb_head);
1609 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1610 if (!READ_ONCE(rdp->nocb_cb_kthread) && rhp &&
1611 rcu_scheduler_fully_active) {
1612 /* RCU callback enqueued before CPU first came online??? */
1613 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1617 #endif /* #ifdef CONFIG_PROVE_RCU */
1623 * Enqueue the specified string of rcu_head structures onto the specified
1624 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1625 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1626 * counts are supplied by rhcount and rhcount_lazy.
1628 * If warranted, also wake up the kthread servicing this CPUs queues.
1630 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1631 struct rcu_head *rhp,
1632 struct rcu_head **rhtp,
1633 int rhcount, int rhcount_lazy,
1634 unsigned long flags)
1637 struct rcu_head **old_rhpp;
1638 struct task_struct *t;
1640 /* Enqueue the callback on the nocb list and update counts. */
1641 atomic_long_add(rhcount, &rdp->nocb_q_count);
1642 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1643 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1644 WRITE_ONCE(*old_rhpp, rhp);
1645 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1646 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1648 /* If we are not being polled and there is a kthread, awaken it ... */
1649 t = READ_ONCE(rdp->nocb_gp_kthread);
1650 if (rcu_nocb_poll || !t) {
1651 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1652 TPS("WakeNotPoll"));
1655 len = rcu_get_n_cbs_nocb_cpu(rdp);
1656 if (old_rhpp == &rdp->nocb_head) {
1657 if (!irqs_disabled_flags(flags)) {
1658 /* ... if queue was empty ... */
1659 wake_nocb_gp(rdp, false);
1660 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1663 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1664 TPS("WakeEmptyIsDeferred"));
1666 rdp->qlen_last_fqs_check = 0;
1667 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1668 /* ... or if many callbacks queued. */
1669 if (!irqs_disabled_flags(flags)) {
1670 wake_nocb_gp(rdp, true);
1671 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1674 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1675 TPS("WakeOvfIsDeferred"));
1677 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1679 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1685 * This is a helper for __call_rcu(), which invokes this when the normal
1686 * callback queue is inoperable. If this is not a no-CBs CPU, this
1687 * function returns failure back to __call_rcu(), which can complain
1690 * Otherwise, this function queues the callback where the corresponding
1691 * "rcuo" kthread can find it.
1693 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1694 bool lazy, unsigned long flags)
1697 if (!rcu_is_nocb_cpu(rdp->cpu))
1699 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1700 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1701 trace_rcu_kfree_callback(rcu_state.name, rhp,
1702 (unsigned long)rhp->func,
1703 -atomic_long_read(&rdp->nocb_q_count_lazy),
1704 -rcu_get_n_cbs_nocb_cpu(rdp));
1706 trace_rcu_callback(rcu_state.name, rhp,
1707 -atomic_long_read(&rdp->nocb_q_count_lazy),
1708 -rcu_get_n_cbs_nocb_cpu(rdp));
1711 * If called from an extended quiescent state with interrupts
1712 * disabled, invoke the RCU core in order to allow the idle-entry
1713 * deferred-wakeup check to function.
1715 if (irqs_disabled_flags(flags) &&
1716 !rcu_is_watching() &&
1717 cpu_online(smp_processor_id()))
1724 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
1727 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
1728 struct rcu_data *rdp,
1729 unsigned long flags)
1731 lockdep_assert_irqs_disabled();
1732 if (!rcu_is_nocb_cpu(smp_processor_id()))
1733 return false; /* Not NOCBs CPU, caller must migrate CBs. */
1734 __call_rcu_nocb_enqueue(my_rdp, rcu_segcblist_head(&rdp->cblist),
1735 rcu_segcblist_tail(&rdp->cblist),
1736 rcu_segcblist_n_cbs(&rdp->cblist),
1737 rcu_segcblist_n_lazy_cbs(&rdp->cblist), flags);
1738 rcu_segcblist_init(&rdp->cblist);
1739 rcu_segcblist_disable(&rdp->cblist);
1744 * If necessary, kick off a new grace period, and either way wait
1745 * for a subsequent grace period to complete.
1747 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
1751 unsigned long flags;
1753 struct rcu_node *rnp = rdp->mynode;
1755 local_irq_save(flags);
1756 c = rcu_seq_snap(&rcu_state.gp_seq);
1757 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1758 local_irq_restore(flags);
1760 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1761 needwake = rcu_start_this_gp(rnp, rdp, c);
1762 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1764 rcu_gp_kthread_wake();
1768 * Wait for the grace period. Do so interruptibly to avoid messing
1769 * up the load average.
1771 trace_rcu_this_gp(rnp, rdp, c, TPS("StartWait"));
1773 swait_event_interruptible_exclusive(
1774 rnp->nocb_gp_wq[rcu_seq_ctr(c) & 0x1],
1775 (d = rcu_seq_done(&rnp->gp_seq, c)));
1778 WARN_ON(signal_pending(current));
1779 trace_rcu_this_gp(rnp, rdp, c, TPS("ResumeWait"));
1781 trace_rcu_this_gp(rnp, rdp, c, TPS("EndWait"));
1782 smp_mb(); /* Ensure that CB invocation happens after GP end. */
1786 * No-CBs GP kthreads come here to wait for additional callbacks to show up.
1787 * This function does not return until callbacks appear.
1789 static void nocb_gp_wait(struct rcu_data *my_rdp)
1791 bool firsttime = true;
1792 unsigned long flags;
1794 struct rcu_data *rdp;
1795 struct rcu_head **tail;
1797 /* Wait for callbacks to appear. */
1798 if (!rcu_nocb_poll) {
1799 trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu, TPS("Sleep"));
1800 swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
1801 !READ_ONCE(my_rdp->nocb_gp_sleep));
1802 raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
1803 my_rdp->nocb_gp_sleep = true;
1804 WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
1805 del_timer(&my_rdp->nocb_timer);
1806 raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
1807 } else if (firsttime) {
1808 firsttime = false; /* Don't drown trace log with "Poll"! */
1809 trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu, TPS("Poll"));
1813 * Each pass through the following loop checks for CBs.
1814 * We are our own first CB kthread. Any CBs found are moved to
1815 * nocb_gp_head, where they await a grace period.
1818 smp_mb(); /* wakeup and _sleep before ->nocb_head reads. */
1819 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1820 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
1821 if (!rdp->nocb_gp_head)
1822 continue; /* No CBs here, try next. */
1824 /* Move callbacks to wait-for-GP list, which is empty. */
1825 WRITE_ONCE(rdp->nocb_head, NULL);
1826 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
1830 /* No callbacks? Sleep a bit if polling, and go retry. */
1831 if (unlikely(!gotcbs)) {
1832 WARN_ON(signal_pending(current));
1833 if (rcu_nocb_poll) {
1834 schedule_timeout_interruptible(1);
1836 trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu,
1842 /* Wait for one grace period. */
1843 rcu_nocb_wait_gp(my_rdp);
1845 /* Each pass through this loop wakes a CB kthread, if needed. */
1846 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1847 if (!rcu_nocb_poll &&
1848 READ_ONCE(rdp->nocb_head) &&
1849 READ_ONCE(my_rdp->nocb_gp_sleep)) {
1850 raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
1851 my_rdp->nocb_gp_sleep = false;/* No need to sleep.*/
1852 raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
1854 if (!rdp->nocb_gp_head)
1855 continue; /* No CBs, so no need to wake kthread. */
1857 /* Append callbacks to CB kthread's "done" list. */
1858 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1859 tail = rdp->nocb_cb_tail;
1860 rdp->nocb_cb_tail = rdp->nocb_gp_tail;
1861 *tail = rdp->nocb_gp_head;
1862 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1863 if (tail == &rdp->nocb_cb_head) {
1864 /* List was empty, so wake up the kthread. */
1865 swake_up_one(&rdp->nocb_cb_wq);
1871 * No-CBs grace-period-wait kthread. There is one of these per group
1872 * of CPUs, but only once at least one CPU in that group has come online
1873 * at least once since boot. This kthread checks for newly posted
1874 * callbacks from any of the CPUs it is responsible for, waits for a
1875 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
1876 * that then have callback-invocation work to do.
1878 static int rcu_nocb_gp_kthread(void *arg)
1880 struct rcu_data *rdp = arg;
1888 * No-CBs CB kthreads come here to wait for additional callbacks to show up.
1889 * This function returns true ("keep waiting") until callbacks appear and
1890 * then false ("stop waiting") when callbacks finally do appear.
1892 static bool nocb_cb_wait(struct rcu_data *rdp)
1894 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
1895 swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
1896 READ_ONCE(rdp->nocb_cb_head));
1897 if (smp_load_acquire(&rdp->nocb_cb_head)) { /* VVV */
1898 /* ^^^ Ensure CB invocation follows _head test. */
1901 WARN_ON(signal_pending(current));
1902 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
1907 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
1908 * callbacks queued by the corresponding no-CBs CPU, however, there is an
1909 * optional GP-CB relationship so that the grace-period kthreads don't
1910 * have to do quite so many wakeups (as in they only need to wake the
1911 * no-CBs GP kthreads, not the CB kthreads).
1913 static int rcu_nocb_cb_kthread(void *arg)
1916 unsigned long flags;
1917 struct rcu_head *list;
1918 struct rcu_head *next;
1919 struct rcu_head **tail;
1920 struct rcu_data *rdp = arg;
1922 /* Each pass through this loop invokes one batch of callbacks */
1924 /* Wait for callbacks. */
1925 while (nocb_cb_wait(rdp))
1928 /* Pull the ready-to-invoke callbacks onto local list. */
1929 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1930 list = rdp->nocb_cb_head;
1931 rdp->nocb_cb_head = NULL;
1932 tail = rdp->nocb_cb_tail;
1933 rdp->nocb_cb_tail = &rdp->nocb_cb_head;
1934 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1935 if (WARN_ON_ONCE(!list))
1937 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeNonEmpty"));
1939 /* Each pass through the following loop invokes a callback. */
1940 trace_rcu_batch_start(rcu_state.name,
1941 atomic_long_read(&rdp->nocb_q_count_lazy),
1942 rcu_get_n_cbs_nocb_cpu(rdp), -1);
1946 /* Wait for enqueuing to complete, if needed. */
1947 while (next == NULL && &list->next != tail) {
1948 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1950 schedule_timeout_interruptible(1);
1951 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1955 debug_rcu_head_unqueue(list);
1957 if (__rcu_reclaim(rcu_state.name, list))
1961 cond_resched_tasks_rcu_qs();
1964 trace_rcu_batch_end(rcu_state.name, c, !!list, 0, 0, 1);
1965 smp_mb__before_atomic(); /* _add after CB invocation. */
1966 atomic_long_add(-c, &rdp->nocb_q_count);
1967 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
1972 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
1973 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
1975 return READ_ONCE(rdp->nocb_defer_wakeup);
1978 /* Do a deferred wakeup of rcu_nocb_kthread(). */
1979 static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
1981 unsigned long flags;
1984 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1985 if (!rcu_nocb_need_deferred_wakeup(rdp)) {
1986 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1989 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
1990 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
1991 __wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
1992 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
1995 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
1996 static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
1998 struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2000 do_nocb_deferred_wakeup_common(rdp);
2004 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2005 * This means we do an inexact common-case check. Note that if
2006 * we miss, ->nocb_timer will eventually clean things up.
2008 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2010 if (rcu_nocb_need_deferred_wakeup(rdp))
2011 do_nocb_deferred_wakeup_common(rdp);
2014 void __init rcu_init_nohz(void)
2017 bool need_rcu_nocb_mask = false;
2019 #if defined(CONFIG_NO_HZ_FULL)
2020 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2021 need_rcu_nocb_mask = true;
2022 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2024 if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2025 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2026 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2030 if (!cpumask_available(rcu_nocb_mask))
2033 #if defined(CONFIG_NO_HZ_FULL)
2034 if (tick_nohz_full_running)
2035 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2036 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2038 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2039 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2040 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2043 if (cpumask_empty(rcu_nocb_mask))
2044 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2046 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2047 cpumask_pr_args(rcu_nocb_mask));
2049 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2051 for_each_cpu(cpu, rcu_nocb_mask)
2052 init_nocb_callback_list(per_cpu_ptr(&rcu_data, cpu));
2053 rcu_organize_nocb_kthreads();
2056 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2057 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2059 rdp->nocb_tail = &rdp->nocb_head;
2060 init_swait_queue_head(&rdp->nocb_cb_wq);
2061 init_swait_queue_head(&rdp->nocb_gp_wq);
2062 rdp->nocb_cb_tail = &rdp->nocb_cb_head;
2063 raw_spin_lock_init(&rdp->nocb_lock);
2064 timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2068 * If the specified CPU is a no-CBs CPU that does not already have its
2069 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2070 * for this CPU's group has not yet been created, spawn it as well.
2072 static void rcu_spawn_one_nocb_kthread(int cpu)
2074 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2075 struct rcu_data *rdp_gp;
2076 struct task_struct *t;
2079 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2080 * then nothing to do.
2082 if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2085 /* If we didn't spawn the GP kthread first, reorganize! */
2086 rdp_gp = rdp->nocb_gp_rdp;
2087 if (!rdp_gp->nocb_gp_kthread) {
2088 t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2089 "rcuog/%d", rdp_gp->cpu);
2090 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2092 WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2095 /* Spawn the kthread for this CPU. */
2096 t = kthread_run(rcu_nocb_cb_kthread, rdp,
2097 "rcuo%c/%d", rcu_state.abbr, cpu);
2098 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2100 WRITE_ONCE(rdp->nocb_cb_kthread, t);
2101 WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2105 * If the specified CPU is a no-CBs CPU that does not already have its
2106 * rcuo kthread, spawn it.
2108 static void rcu_spawn_cpu_nocb_kthread(int cpu)
2110 if (rcu_scheduler_fully_active)
2111 rcu_spawn_one_nocb_kthread(cpu);
2115 * Once the scheduler is running, spawn rcuo kthreads for all online
2116 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2117 * non-boot CPUs come online -- if this changes, we will need to add
2118 * some mutual exclusion.
2120 static void __init rcu_spawn_nocb_kthreads(void)
2124 for_each_online_cpu(cpu)
2125 rcu_spawn_cpu_nocb_kthread(cpu);
2128 /* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2129 static int rcu_nocb_gp_stride = -1;
2130 module_param(rcu_nocb_gp_stride, int, 0444);
2133 * Initialize GP-CB relationships for all no-CBs CPU.
2135 static void __init rcu_organize_nocb_kthreads(void)
2138 bool firsttime = true;
2139 int ls = rcu_nocb_gp_stride;
2140 int nl = 0; /* Next GP kthread. */
2141 struct rcu_data *rdp;
2142 struct rcu_data *rdp_gp = NULL; /* Suppress misguided gcc warn. */
2143 struct rcu_data *rdp_prev = NULL;
2145 if (!cpumask_available(rcu_nocb_mask))
2148 ls = int_sqrt(nr_cpu_ids);
2149 rcu_nocb_gp_stride = ls;
2153 * Each pass through this loop sets up one rcu_data structure.
2154 * Should the corresponding CPU come online in the future, then
2155 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2157 for_each_cpu(cpu, rcu_nocb_mask) {
2158 rdp = per_cpu_ptr(&rcu_data, cpu);
2159 if (rdp->cpu >= nl) {
2160 /* New GP kthread, set up for CBs & next GP. */
2161 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2162 rdp->nocb_gp_rdp = rdp;
2164 if (!firsttime && dump_tree)
2167 pr_alert("%s: No-CB GP kthread CPU %d:", __func__, cpu);
2169 /* Another CB kthread, link to previous GP kthread. */
2170 rdp->nocb_gp_rdp = rdp_gp;
2171 rdp_prev->nocb_next_cb_rdp = rdp;
2172 pr_alert(" %d", cpu);
2178 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2179 static bool init_nocb_callback_list(struct rcu_data *rdp)
2181 if (!rcu_is_nocb_cpu(rdp->cpu))
2184 /* If there are early-boot callbacks, move them to nocb lists. */
2185 if (!rcu_segcblist_empty(&rdp->cblist)) {
2186 rdp->nocb_head = rcu_segcblist_head(&rdp->cblist);
2187 rdp->nocb_tail = rcu_segcblist_tail(&rdp->cblist);
2188 atomic_long_set(&rdp->nocb_q_count,
2189 rcu_segcblist_n_cbs(&rdp->cblist));
2190 atomic_long_set(&rdp->nocb_q_count_lazy,
2191 rcu_segcblist_n_lazy_cbs(&rdp->cblist));
2193 rcu_segcblist_init(&rdp->cblist);
2194 rcu_segcblist_disable(&rdp->cblist);
2199 * Bind the current task to the offloaded CPUs. If there are no offloaded
2200 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2202 void rcu_bind_current_to_nocb(void)
2204 if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2205 WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2207 EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2210 * Return the number of RCU callbacks still queued from the specified
2211 * CPU, which must be a nocbs CPU.
2213 static unsigned long rcu_get_n_cbs_nocb_cpu(struct rcu_data *rdp)
2215 return atomic_long_read(&rdp->nocb_q_count);
2218 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2220 static bool rcu_nocb_cpu_needs_barrier(int cpu)
2222 WARN_ON_ONCE(1); /* Should be dead code. */
2226 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2230 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2235 static void rcu_init_one_nocb(struct rcu_node *rnp)
2239 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2240 bool lazy, unsigned long flags)
2245 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
2246 struct rcu_data *rdp,
2247 unsigned long flags)
2252 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2256 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2261 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2265 static void rcu_spawn_cpu_nocb_kthread(int cpu)
2269 static void __init rcu_spawn_nocb_kthreads(void)
2273 static bool init_nocb_callback_list(struct rcu_data *rdp)
2278 static unsigned long rcu_get_n_cbs_nocb_cpu(struct rcu_data *rdp)
2283 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2286 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2287 * grace-period kthread will do force_quiescent_state() processing?
2288 * The idea is to avoid waking up RCU core processing on such a
2289 * CPU unless the grace period has extended for too long.
2291 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2292 * CONFIG_RCU_NOCB_CPU CPUs.
2294 static bool rcu_nohz_full_cpu(void)
2296 #ifdef CONFIG_NO_HZ_FULL
2297 if (tick_nohz_full_cpu(smp_processor_id()) &&
2298 (!rcu_gp_in_progress() ||
2299 ULONG_CMP_LT(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2301 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2306 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2308 static void rcu_bind_gp_kthread(void)
2310 if (!tick_nohz_full_enabled())
2312 housekeeping_affine(current, HK_FLAG_RCU);
2315 /* Record the current task on dyntick-idle entry. */
2316 static void rcu_dynticks_task_enter(void)
2318 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2319 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2320 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2323 /* Record no current task on dyntick-idle exit. */
2324 static void rcu_dynticks_task_exit(void)
2326 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2327 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2328 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */