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 static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
19 * In order to read the offloaded state of an rdp is a safe
20 * and stable way and prevent from its value to be changed
21 * under us, we must either hold the barrier mutex, the cpu
22 * hotplug lock (read or write) or the nocb lock. Local
23 * non-preemptible reads are also safe. NOCB kthreads and
24 * timers have their own means of synchronization against the
25 * offloaded state updaters.
28 !(lockdep_is_held(&rcu_state.barrier_mutex) ||
29 (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
30 rcu_lockdep_is_held_nocb(rdp) ||
31 (rdp == this_cpu_ptr(&rcu_data) &&
32 !(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible())) ||
33 rcu_current_is_nocb_kthread(rdp)),
34 "Unsafe read of RCU_NOCB offloaded state"
37 return rcu_segcblist_is_offloaded(&rdp->cblist);
41 * Check the RCU kernel configuration parameters and print informative
42 * messages about anything out of the ordinary.
44 static void __init rcu_bootup_announce_oddness(void)
46 if (IS_ENABLED(CONFIG_RCU_TRACE))
47 pr_info("\tRCU event tracing is enabled.\n");
48 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
49 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
50 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
53 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
54 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
55 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
56 if (IS_ENABLED(CONFIG_PROVE_RCU))
57 pr_info("\tRCU lockdep checking is enabled.\n");
58 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
59 pr_info("\tRCU strict (and thus non-scalable) grace periods enabled.\n");
60 if (RCU_NUM_LVLS >= 4)
61 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
62 if (RCU_FANOUT_LEAF != 16)
63 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
65 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
66 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
68 if (nr_cpu_ids != NR_CPUS)
69 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
70 #ifdef CONFIG_RCU_BOOST
71 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
72 kthread_prio, CONFIG_RCU_BOOST_DELAY);
74 if (blimit != DEFAULT_RCU_BLIMIT)
75 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
76 if (qhimark != DEFAULT_RCU_QHIMARK)
77 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
78 if (qlowmark != DEFAULT_RCU_QLOMARK)
79 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
80 if (qovld != DEFAULT_RCU_QOVLD)
81 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
82 if (jiffies_till_first_fqs != ULONG_MAX)
83 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
84 if (jiffies_till_next_fqs != ULONG_MAX)
85 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
86 if (jiffies_till_sched_qs != ULONG_MAX)
87 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
88 if (rcu_kick_kthreads)
89 pr_info("\tKick kthreads if too-long grace period.\n");
90 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
91 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
93 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
95 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
97 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
99 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
100 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
101 pr_info("\tRCU debug extended QS entry/exit.\n");
102 rcupdate_announce_bootup_oddness();
105 #ifdef CONFIG_PREEMPT_RCU
107 static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
108 static void rcu_read_unlock_special(struct task_struct *t);
111 * Tell them what RCU they are running.
113 static void __init rcu_bootup_announce(void)
115 pr_info("Preemptible hierarchical RCU implementation.\n");
116 rcu_bootup_announce_oddness();
119 /* Flags for rcu_preempt_ctxt_queue() decision table. */
120 #define RCU_GP_TASKS 0x8
121 #define RCU_EXP_TASKS 0x4
122 #define RCU_GP_BLKD 0x2
123 #define RCU_EXP_BLKD 0x1
126 * Queues a task preempted within an RCU-preempt read-side critical
127 * section into the appropriate location within the ->blkd_tasks list,
128 * depending on the states of any ongoing normal and expedited grace
129 * periods. The ->gp_tasks pointer indicates which element the normal
130 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
131 * indicates which element the expedited grace period is waiting on (again,
132 * NULL if none). If a grace period is waiting on a given element in the
133 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
134 * adding a task to the tail of the list blocks any grace period that is
135 * already waiting on one of the elements. In contrast, adding a task
136 * to the head of the list won't block any grace period that is already
137 * waiting on one of the elements.
139 * This queuing is imprecise, and can sometimes make an ongoing grace
140 * period wait for a task that is not strictly speaking blocking it.
141 * Given the choice, we needlessly block a normal grace period rather than
142 * blocking an expedited grace period.
144 * Note that an endless sequence of expedited grace periods still cannot
145 * indefinitely postpone a normal grace period. Eventually, all of the
146 * fixed number of preempted tasks blocking the normal grace period that are
147 * not also blocking the expedited grace period will resume and complete
148 * their RCU read-side critical sections. At that point, the ->gp_tasks
149 * pointer will equal the ->exp_tasks pointer, at which point the end of
150 * the corresponding expedited grace period will also be the end of the
151 * normal grace period.
153 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
154 __releases(rnp->lock) /* But leaves rrupts disabled. */
156 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
157 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
158 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
159 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
160 struct task_struct *t = current;
162 raw_lockdep_assert_held_rcu_node(rnp);
163 WARN_ON_ONCE(rdp->mynode != rnp);
164 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
165 /* RCU better not be waiting on newly onlined CPUs! */
166 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
170 * Decide where to queue the newly blocked task. In theory,
171 * this could be an if-statement. In practice, when I tried
172 * that, it was quite messy.
174 switch (blkd_state) {
177 case RCU_EXP_TASKS + RCU_GP_BLKD:
179 case RCU_GP_TASKS + RCU_EXP_TASKS:
182 * Blocking neither GP, or first task blocking the normal
183 * GP but not blocking the already-waiting expedited GP.
184 * Queue at the head of the list to avoid unnecessarily
185 * blocking the already-waiting GPs.
187 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
192 case RCU_GP_BLKD + RCU_EXP_BLKD:
193 case RCU_GP_TASKS + RCU_EXP_BLKD:
194 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
195 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
198 * First task arriving that blocks either GP, or first task
199 * arriving that blocks the expedited GP (with the normal
200 * GP already waiting), or a task arriving that blocks
201 * both GPs with both GPs already waiting. Queue at the
202 * tail of the list to avoid any GP waiting on any of the
203 * already queued tasks that are not blocking it.
205 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
208 case RCU_EXP_TASKS + RCU_EXP_BLKD:
209 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
210 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
213 * Second or subsequent task blocking the expedited GP.
214 * The task either does not block the normal GP, or is the
215 * first task blocking the normal GP. Queue just after
216 * the first task blocking the expedited GP.
218 list_add(&t->rcu_node_entry, rnp->exp_tasks);
221 case RCU_GP_TASKS + RCU_GP_BLKD:
222 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
225 * Second or subsequent task blocking the normal GP.
226 * The task does not block the expedited GP. Queue just
227 * after the first task blocking the normal GP.
229 list_add(&t->rcu_node_entry, rnp->gp_tasks);
234 /* Yet another exercise in excessive paranoia. */
240 * We have now queued the task. If it was the first one to
241 * block either grace period, update the ->gp_tasks and/or
242 * ->exp_tasks pointers, respectively, to reference the newly
245 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
246 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
247 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
249 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
250 WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
251 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
252 !(rnp->qsmask & rdp->grpmask));
253 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
254 !(rnp->expmask & rdp->grpmask));
255 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
258 * Report the quiescent state for the expedited GP. This expedited
259 * GP should not be able to end until we report, so there should be
260 * no need to check for a subsequent expedited GP. (Though we are
261 * still in a quiescent state in any case.)
263 if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
264 rcu_report_exp_rdp(rdp);
266 WARN_ON_ONCE(rdp->exp_deferred_qs);
270 * Record a preemptible-RCU quiescent state for the specified CPU.
271 * Note that this does not necessarily mean that the task currently running
272 * on the CPU is in a quiescent state: Instead, it means that the current
273 * grace period need not wait on any RCU read-side critical section that
274 * starts later on this CPU. It also means that if the current task is
275 * in an RCU read-side critical section, it has already added itself to
276 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
277 * current task, there might be any number of other tasks blocked while
278 * in an RCU read-side critical section.
280 * Callers to this function must disable preemption.
282 static void rcu_qs(void)
284 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
285 if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
286 trace_rcu_grace_period(TPS("rcu_preempt"),
287 __this_cpu_read(rcu_data.gp_seq),
289 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
290 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
291 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
296 * We have entered the scheduler, and the current task might soon be
297 * context-switched away from. If this task is in an RCU read-side
298 * critical section, we will no longer be able to rely on the CPU to
299 * record that fact, so we enqueue the task on the blkd_tasks list.
300 * The task will dequeue itself when it exits the outermost enclosing
301 * RCU read-side critical section. Therefore, the current grace period
302 * cannot be permitted to complete until the blkd_tasks list entries
303 * predating the current grace period drain, in other words, until
304 * rnp->gp_tasks becomes NULL.
306 * Caller must disable interrupts.
308 void rcu_note_context_switch(bool preempt)
310 struct task_struct *t = current;
311 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
312 struct rcu_node *rnp;
314 trace_rcu_utilization(TPS("Start context switch"));
315 lockdep_assert_irqs_disabled();
316 WARN_ONCE(!preempt && rcu_preempt_depth() > 0, "Voluntary context switch within RCU read-side critical section!");
317 if (rcu_preempt_depth() > 0 &&
318 !t->rcu_read_unlock_special.b.blocked) {
320 /* Possibly blocking in an RCU read-side critical section. */
322 raw_spin_lock_rcu_node(rnp);
323 t->rcu_read_unlock_special.b.blocked = true;
324 t->rcu_blocked_node = rnp;
327 * Verify the CPU's sanity, trace the preemption, and
328 * then queue the task as required based on the states
329 * of any ongoing and expedited grace periods.
331 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
332 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
333 trace_rcu_preempt_task(rcu_state.name,
335 (rnp->qsmask & rdp->grpmask)
337 : rcu_seq_snap(&rnp->gp_seq));
338 rcu_preempt_ctxt_queue(rnp, rdp);
340 rcu_preempt_deferred_qs(t);
344 * Either we were not in an RCU read-side critical section to
345 * begin with, or we have now recorded that critical section
346 * globally. Either way, we can now note a quiescent state
347 * for this CPU. Again, if we were in an RCU read-side critical
348 * section, and if that critical section was blocking the current
349 * grace period, then the fact that the task has been enqueued
350 * means that we continue to block the current grace period.
353 if (rdp->exp_deferred_qs)
354 rcu_report_exp_rdp(rdp);
355 rcu_tasks_qs(current, preempt);
356 trace_rcu_utilization(TPS("End context switch"));
358 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
361 * Check for preempted RCU readers blocking the current grace period
362 * for the specified rcu_node structure. If the caller needs a reliable
363 * answer, it must hold the rcu_node's ->lock.
365 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
367 return READ_ONCE(rnp->gp_tasks) != NULL;
370 /* limit value for ->rcu_read_lock_nesting. */
371 #define RCU_NEST_PMAX (INT_MAX / 2)
373 static void rcu_preempt_read_enter(void)
375 WRITE_ONCE(current->rcu_read_lock_nesting, READ_ONCE(current->rcu_read_lock_nesting) + 1);
378 static int rcu_preempt_read_exit(void)
380 int ret = READ_ONCE(current->rcu_read_lock_nesting) - 1;
382 WRITE_ONCE(current->rcu_read_lock_nesting, ret);
386 static void rcu_preempt_depth_set(int val)
388 WRITE_ONCE(current->rcu_read_lock_nesting, val);
392 * Preemptible RCU implementation for rcu_read_lock().
393 * Just increment ->rcu_read_lock_nesting, shared state will be updated
396 void __rcu_read_lock(void)
398 rcu_preempt_read_enter();
399 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
400 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
401 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
402 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
403 barrier(); /* critical section after entry code. */
405 EXPORT_SYMBOL_GPL(__rcu_read_lock);
408 * Preemptible RCU implementation for rcu_read_unlock().
409 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
410 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
411 * invoke rcu_read_unlock_special() to clean up after a context switch
412 * in an RCU read-side critical section and other special cases.
414 void __rcu_read_unlock(void)
416 struct task_struct *t = current;
418 barrier(); // critical section before exit code.
419 if (rcu_preempt_read_exit() == 0) {
420 barrier(); // critical-section exit before .s check.
421 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
422 rcu_read_unlock_special(t);
424 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
425 int rrln = rcu_preempt_depth();
427 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
430 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
433 * Advance a ->blkd_tasks-list pointer to the next entry, instead
434 * returning NULL if at the end of the list.
436 static struct list_head *rcu_next_node_entry(struct task_struct *t,
437 struct rcu_node *rnp)
439 struct list_head *np;
441 np = t->rcu_node_entry.next;
442 if (np == &rnp->blkd_tasks)
448 * Return true if the specified rcu_node structure has tasks that were
449 * preempted within an RCU read-side critical section.
451 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
453 return !list_empty(&rnp->blkd_tasks);
457 * Report deferred quiescent states. The deferral time can
458 * be quite short, for example, in the case of the call from
459 * rcu_read_unlock_special().
462 rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
467 struct list_head *np;
468 bool drop_boost_mutex = false;
469 struct rcu_data *rdp;
470 struct rcu_node *rnp;
471 union rcu_special special;
474 * If RCU core is waiting for this CPU to exit its critical section,
475 * report the fact that it has exited. Because irqs are disabled,
476 * t->rcu_read_unlock_special cannot change.
478 special = t->rcu_read_unlock_special;
479 rdp = this_cpu_ptr(&rcu_data);
480 if (!special.s && !rdp->exp_deferred_qs) {
481 local_irq_restore(flags);
484 t->rcu_read_unlock_special.s = 0;
485 if (special.b.need_qs) {
486 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
487 rcu_report_qs_rdp(rdp);
488 udelay(rcu_unlock_delay);
495 * Respond to a request by an expedited grace period for a
496 * quiescent state from this CPU. Note that requests from
497 * tasks are handled when removing the task from the
498 * blocked-tasks list below.
500 if (rdp->exp_deferred_qs)
501 rcu_report_exp_rdp(rdp);
503 /* Clean up if blocked during RCU read-side critical section. */
504 if (special.b.blocked) {
507 * Remove this task from the list it blocked on. The task
508 * now remains queued on the rcu_node corresponding to the
509 * CPU it first blocked on, so there is no longer any need
510 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
512 rnp = t->rcu_blocked_node;
513 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
514 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
515 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
516 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
517 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
518 (!empty_norm || rnp->qsmask));
519 empty_exp = sync_rcu_exp_done(rnp);
520 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
521 np = rcu_next_node_entry(t, rnp);
522 list_del_init(&t->rcu_node_entry);
523 t->rcu_blocked_node = NULL;
524 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
525 rnp->gp_seq, t->pid);
526 if (&t->rcu_node_entry == rnp->gp_tasks)
527 WRITE_ONCE(rnp->gp_tasks, np);
528 if (&t->rcu_node_entry == rnp->exp_tasks)
529 WRITE_ONCE(rnp->exp_tasks, np);
530 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
531 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
532 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx.rtmutex) == t;
533 if (&t->rcu_node_entry == rnp->boost_tasks)
534 WRITE_ONCE(rnp->boost_tasks, np);
538 * If this was the last task on the current list, and if
539 * we aren't waiting on any CPUs, report the quiescent state.
540 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
541 * so we must take a snapshot of the expedited state.
543 empty_exp_now = sync_rcu_exp_done(rnp);
544 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
545 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
552 rcu_report_unblock_qs_rnp(rnp, flags);
554 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
558 * If this was the last task on the expedited lists,
559 * then we need to report up the rcu_node hierarchy.
561 if (!empty_exp && empty_exp_now)
562 rcu_report_exp_rnp(rnp, true);
564 /* Unboost if we were boosted. */
565 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
566 rt_mutex_futex_unlock(&rnp->boost_mtx.rtmutex);
568 local_irq_restore(flags);
573 * Is a deferred quiescent-state pending, and are we also not in
574 * an RCU read-side critical section? It is the caller's responsibility
575 * to ensure it is otherwise safe to report any deferred quiescent
576 * states. The reason for this is that it is safe to report a
577 * quiescent state during context switch even though preemption
578 * is disabled. This function cannot be expected to understand these
579 * nuances, so the caller must handle them.
581 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
583 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
584 READ_ONCE(t->rcu_read_unlock_special.s)) &&
585 rcu_preempt_depth() == 0;
589 * Report a deferred quiescent state if needed and safe to do so.
590 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
591 * not being in an RCU read-side critical section. The caller must
592 * evaluate safety in terms of interrupt, softirq, and preemption
595 static void rcu_preempt_deferred_qs(struct task_struct *t)
599 if (!rcu_preempt_need_deferred_qs(t))
601 local_irq_save(flags);
602 rcu_preempt_deferred_qs_irqrestore(t, flags);
606 * Minimal handler to give the scheduler a chance to re-evaluate.
608 static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
610 struct rcu_data *rdp;
612 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
613 rdp->defer_qs_iw_pending = false;
617 * Handle special cases during rcu_read_unlock(), such as needing to
618 * notify RCU core processing or task having blocked during the RCU
619 * read-side critical section.
621 static void rcu_read_unlock_special(struct task_struct *t)
624 bool irqs_were_disabled;
625 bool preempt_bh_were_disabled =
626 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
628 /* NMI handlers cannot block and cannot safely manipulate state. */
632 local_irq_save(flags);
633 irqs_were_disabled = irqs_disabled_flags(flags);
634 if (preempt_bh_were_disabled || irqs_were_disabled) {
635 bool expboost; // Expedited GP in flight or possible boosting.
636 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
637 struct rcu_node *rnp = rdp->mynode;
639 expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
640 (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
641 (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
642 ((rdp->grpmask & READ_ONCE(rnp->qsmask)) || t->rcu_blocked_node)) ||
643 (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
644 t->rcu_blocked_node);
645 // Need to defer quiescent state until everything is enabled.
646 if (use_softirq && (in_irq() || (expboost && !irqs_were_disabled))) {
647 // Using softirq, safe to awaken, and either the
648 // wakeup is free or there is either an expedited
649 // GP in flight or a potential need to deboost.
650 raise_softirq_irqoff(RCU_SOFTIRQ);
652 // Enabling BH or preempt does reschedule, so...
653 // Also if no expediting and no possible deboosting,
654 // slow is OK. Plus nohz_full CPUs eventually get
656 set_tsk_need_resched(current);
657 set_preempt_need_resched();
658 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
659 expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
660 // Get scheduler to re-evaluate and call hooks.
661 // If !IRQ_WORK, FQS scan will eventually IPI.
662 init_irq_work(&rdp->defer_qs_iw, rcu_preempt_deferred_qs_handler);
663 rdp->defer_qs_iw_pending = true;
664 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
667 local_irq_restore(flags);
670 rcu_preempt_deferred_qs_irqrestore(t, flags);
674 * Check that the list of blocked tasks for the newly completed grace
675 * period is in fact empty. It is a serious bug to complete a grace
676 * period that still has RCU readers blocked! This function must be
677 * invoked -before- updating this rnp's ->gp_seq.
679 * Also, if there are blocked tasks on the list, they automatically
680 * block the newly created grace period, so set up ->gp_tasks accordingly.
682 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
684 struct task_struct *t;
686 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
687 raw_lockdep_assert_held_rcu_node(rnp);
688 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
689 dump_blkd_tasks(rnp, 10);
690 if (rcu_preempt_has_tasks(rnp) &&
691 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
692 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
693 t = container_of(rnp->gp_tasks, struct task_struct,
695 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
696 rnp->gp_seq, t->pid);
698 WARN_ON_ONCE(rnp->qsmask);
702 * Check for a quiescent state from the current CPU, including voluntary
703 * context switches for Tasks RCU. When a task blocks, the task is
704 * recorded in the corresponding CPU's rcu_node structure, which is checked
705 * elsewhere, hence this function need only check for quiescent states
706 * related to the current CPU, not to those related to tasks.
708 static void rcu_flavor_sched_clock_irq(int user)
710 struct task_struct *t = current;
712 lockdep_assert_irqs_disabled();
713 if (user || rcu_is_cpu_rrupt_from_idle()) {
714 rcu_note_voluntary_context_switch(current);
716 if (rcu_preempt_depth() > 0 ||
717 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
718 /* No QS, force context switch if deferred. */
719 if (rcu_preempt_need_deferred_qs(t)) {
720 set_tsk_need_resched(t);
721 set_preempt_need_resched();
723 } else if (rcu_preempt_need_deferred_qs(t)) {
724 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
726 } else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
727 rcu_qs(); /* Report immediate QS. */
731 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
732 if (rcu_preempt_depth() > 0 &&
733 __this_cpu_read(rcu_data.core_needs_qs) &&
734 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
735 !t->rcu_read_unlock_special.b.need_qs &&
736 time_after(jiffies, rcu_state.gp_start + HZ))
737 t->rcu_read_unlock_special.b.need_qs = true;
741 * Check for a task exiting while in a preemptible-RCU read-side
742 * critical section, clean up if so. No need to issue warnings, as
743 * debug_check_no_locks_held() already does this if lockdep is enabled.
744 * Besides, if this function does anything other than just immediately
745 * return, there was a bug of some sort. Spewing warnings from this
746 * function is like as not to simply obscure important prior warnings.
750 struct task_struct *t = current;
752 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
753 rcu_preempt_depth_set(1);
755 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
756 } else if (unlikely(rcu_preempt_depth())) {
757 rcu_preempt_depth_set(1);
762 rcu_preempt_deferred_qs(current);
766 * Dump the blocked-tasks state, but limit the list dump to the
767 * specified number of elements.
770 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
774 struct list_head *lhp;
776 struct rcu_data *rdp;
777 struct rcu_node *rnp1;
779 raw_lockdep_assert_held_rcu_node(rnp);
780 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
781 __func__, rnp->grplo, rnp->grphi, rnp->level,
782 (long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
783 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
784 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
785 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
786 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
787 __func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
788 READ_ONCE(rnp->exp_tasks));
789 pr_info("%s: ->blkd_tasks", __func__);
791 list_for_each(lhp, &rnp->blkd_tasks) {
797 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
798 rdp = per_cpu_ptr(&rcu_data, cpu);
799 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
800 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
802 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
803 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
807 #else /* #ifdef CONFIG_PREEMPT_RCU */
810 * If strict grace periods are enabled, and if the calling
811 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
812 * report that quiescent state and, if requested, spin for a bit.
814 void rcu_read_unlock_strict(void)
816 struct rcu_data *rdp;
818 if (!IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ||
819 irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
821 rdp = this_cpu_ptr(&rcu_data);
822 rcu_report_qs_rdp(rdp);
823 udelay(rcu_unlock_delay);
825 EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
828 * Tell them what RCU they are running.
830 static void __init rcu_bootup_announce(void)
832 pr_info("Hierarchical RCU implementation.\n");
833 rcu_bootup_announce_oddness();
837 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
838 * how many quiescent states passed, just if there was at least one since
839 * the start of the grace period, this just sets a flag. The caller must
840 * have disabled preemption.
842 static void rcu_qs(void)
844 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
845 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
847 trace_rcu_grace_period(TPS("rcu_sched"),
848 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
849 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
850 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
852 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
853 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
857 * Register an urgently needed quiescent state. If there is an
858 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
859 * dyntick-idle quiescent state visible to other CPUs, which will in
860 * some cases serve for expedited as well as normal grace periods.
861 * Either way, register a lightweight quiescent state.
863 void rcu_all_qs(void)
867 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
870 /* Load rcu_urgent_qs before other flags. */
871 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
875 this_cpu_write(rcu_data.rcu_urgent_qs, false);
876 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
877 local_irq_save(flags);
878 rcu_momentary_dyntick_idle();
879 local_irq_restore(flags);
884 EXPORT_SYMBOL_GPL(rcu_all_qs);
887 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
889 void rcu_note_context_switch(bool preempt)
891 trace_rcu_utilization(TPS("Start context switch"));
893 /* Load rcu_urgent_qs before other flags. */
894 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
896 this_cpu_write(rcu_data.rcu_urgent_qs, false);
897 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
898 rcu_momentary_dyntick_idle();
899 rcu_tasks_qs(current, preempt);
901 trace_rcu_utilization(TPS("End context switch"));
903 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
906 * Because preemptible RCU does not exist, there are never any preempted
909 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
915 * Because there is no preemptible RCU, there can be no readers blocked.
917 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
923 * Because there is no preemptible RCU, there can be no deferred quiescent
926 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
930 static void rcu_preempt_deferred_qs(struct task_struct *t) { }
933 * Because there is no preemptible RCU, there can be no readers blocked,
934 * so there is no need to check for blocked tasks. So check only for
935 * bogus qsmask values.
937 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
939 WARN_ON_ONCE(rnp->qsmask);
943 * Check to see if this CPU is in a non-context-switch quiescent state,
944 * namely user mode and idle loop.
946 static void rcu_flavor_sched_clock_irq(int user)
948 if (user || rcu_is_cpu_rrupt_from_idle()) {
951 * Get here if this CPU took its interrupt from user
952 * mode or from the idle loop, and if this is not a
953 * nested interrupt. In this case, the CPU is in
954 * a quiescent state, so note it.
956 * No memory barrier is required here because rcu_qs()
957 * references only CPU-local variables that other CPUs
958 * neither access nor modify, at least not while the
959 * corresponding CPU is online.
967 * Because preemptible RCU does not exist, tasks cannot possibly exit
968 * while in preemptible RCU read-side critical sections.
975 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
978 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
980 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
983 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
986 * If boosting, set rcuc kthreads to realtime priority.
988 static void rcu_cpu_kthread_setup(unsigned int cpu)
990 #ifdef CONFIG_RCU_BOOST
991 struct sched_param sp;
993 sp.sched_priority = kthread_prio;
994 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
995 #endif /* #ifdef CONFIG_RCU_BOOST */
998 #ifdef CONFIG_RCU_BOOST
1001 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1002 * or ->boost_tasks, advancing the pointer to the next task in the
1003 * ->blkd_tasks list.
1005 * Note that irqs must be enabled: boosting the task can block.
1006 * Returns 1 if there are more tasks needing to be boosted.
1008 static int rcu_boost(struct rcu_node *rnp)
1010 unsigned long flags;
1011 struct task_struct *t;
1012 struct list_head *tb;
1014 if (READ_ONCE(rnp->exp_tasks) == NULL &&
1015 READ_ONCE(rnp->boost_tasks) == NULL)
1016 return 0; /* Nothing left to boost. */
1018 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1021 * Recheck under the lock: all tasks in need of boosting
1022 * might exit their RCU read-side critical sections on their own.
1024 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1025 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1030 * Preferentially boost tasks blocking expedited grace periods.
1031 * This cannot starve the normal grace periods because a second
1032 * expedited grace period must boost all blocked tasks, including
1033 * those blocking the pre-existing normal grace period.
1035 if (rnp->exp_tasks != NULL)
1036 tb = rnp->exp_tasks;
1038 tb = rnp->boost_tasks;
1041 * We boost task t by manufacturing an rt_mutex that appears to
1042 * be held by task t. We leave a pointer to that rt_mutex where
1043 * task t can find it, and task t will release the mutex when it
1044 * exits its outermost RCU read-side critical section. Then
1045 * simply acquiring this artificial rt_mutex will boost task
1046 * t's priority. (Thanks to tglx for suggesting this approach!)
1048 * Note that task t must acquire rnp->lock to remove itself from
1049 * the ->blkd_tasks list, which it will do from exit() if from
1050 * nowhere else. We therefore are guaranteed that task t will
1051 * stay around at least until we drop rnp->lock. Note that
1052 * rnp->lock also resolves races between our priority boosting
1053 * and task t's exiting its outermost RCU read-side critical
1056 t = container_of(tb, struct task_struct, rcu_node_entry);
1057 rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
1058 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1059 /* Lock only for side effect: boosts task t's priority. */
1060 rt_mutex_lock(&rnp->boost_mtx);
1061 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1064 return READ_ONCE(rnp->exp_tasks) != NULL ||
1065 READ_ONCE(rnp->boost_tasks) != NULL;
1069 * Priority-boosting kthread, one per leaf rcu_node.
1071 static int rcu_boost_kthread(void *arg)
1073 struct rcu_node *rnp = (struct rcu_node *)arg;
1077 trace_rcu_utilization(TPS("Start boost kthread@init"));
1079 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1080 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1081 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1082 READ_ONCE(rnp->exp_tasks));
1083 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1084 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1085 more2boost = rcu_boost(rnp);
1091 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1092 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1093 schedule_timeout_idle(2);
1094 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1099 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1104 * Check to see if it is time to start boosting RCU readers that are
1105 * blocking the current grace period, and, if so, tell the per-rcu_node
1106 * kthread to start boosting them. If there is an expedited grace
1107 * period in progress, it is always time to boost.
1109 * The caller must hold rnp->lock, which this function releases.
1110 * The ->boost_kthread_task is immortal, so we don't need to worry
1111 * about it going away.
1113 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1114 __releases(rnp->lock)
1116 raw_lockdep_assert_held_rcu_node(rnp);
1117 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1118 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1121 if (rnp->exp_tasks != NULL ||
1122 (rnp->gp_tasks != NULL &&
1123 rnp->boost_tasks == NULL &&
1125 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld))) {
1126 if (rnp->exp_tasks == NULL)
1127 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1128 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1129 rcu_wake_cond(rnp->boost_kthread_task,
1130 READ_ONCE(rnp->boost_kthread_status));
1132 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1137 * Is the current CPU running the RCU-callbacks kthread?
1138 * Caller must have preemption disabled.
1140 static bool rcu_is_callbacks_kthread(void)
1142 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1145 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1148 * Do priority-boost accounting for the start of a new grace period.
1150 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1152 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1156 * Create an RCU-boost kthread for the specified node if one does not
1157 * already exist. We only create this kthread for preemptible RCU.
1158 * Returns zero if all is well, a negated errno otherwise.
1160 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1162 unsigned long flags;
1163 int rnp_index = rnp - rcu_get_root();
1164 struct sched_param sp;
1165 struct task_struct *t;
1167 if (rnp->boost_kthread_task || !rcu_scheduler_fully_active)
1170 rcu_state.boost = 1;
1172 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1173 "rcub/%d", rnp_index);
1174 if (WARN_ON_ONCE(IS_ERR(t)))
1177 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1178 rnp->boost_kthread_task = t;
1179 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1180 sp.sched_priority = kthread_prio;
1181 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1182 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1186 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1187 * served by the rcu_node in question. The CPU hotplug lock is still
1188 * held, so the value of rnp->qsmaskinit will be stable.
1190 * We don't include outgoingcpu in the affinity set, use -1 if there is
1191 * no outgoing CPU. If there are no CPUs left in the affinity set,
1192 * this function allows the kthread to execute on any CPU.
1194 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1196 struct task_struct *t = rnp->boost_kthread_task;
1197 unsigned long mask = rcu_rnp_online_cpus(rnp);
1203 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1205 for_each_leaf_node_possible_cpu(rnp, cpu)
1206 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1208 cpumask_set_cpu(cpu, cm);
1209 if (cpumask_weight(cm) == 0)
1211 set_cpus_allowed_ptr(t, cm);
1212 free_cpumask_var(cm);
1216 * Spawn boost kthreads -- called as soon as the scheduler is running.
1218 static void __init rcu_spawn_boost_kthreads(void)
1220 struct rcu_node *rnp;
1222 rcu_for_each_leaf_node(rnp)
1223 if (rcu_rnp_online_cpus(rnp))
1224 rcu_spawn_one_boost_kthread(rnp);
1227 #else /* #ifdef CONFIG_RCU_BOOST */
1229 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1230 __releases(rnp->lock)
1232 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1235 static bool rcu_is_callbacks_kthread(void)
1240 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1244 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1248 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1252 static void __init rcu_spawn_boost_kthreads(void)
1256 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1258 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1261 * Check to see if any future non-offloaded RCU-related work will need
1262 * to be done by the current CPU, even if none need be done immediately,
1263 * returning 1 if so. This function is part of the RCU implementation;
1264 * it is -not- an exported member of the RCU API.
1266 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1267 * CPU has RCU callbacks queued.
1269 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1271 *nextevt = KTIME_MAX;
1272 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1273 !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
1277 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1280 static void rcu_cleanup_after_idle(void)
1285 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1288 static void rcu_prepare_for_idle(void)
1292 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1295 * This code is invoked when a CPU goes idle, at which point we want
1296 * to have the CPU do everything required for RCU so that it can enter
1297 * the energy-efficient dyntick-idle mode.
1299 * The following preprocessor symbol controls this:
1301 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1302 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1303 * is sized to be roughly one RCU grace period. Those energy-efficiency
1304 * benchmarkers who might otherwise be tempted to set this to a large
1305 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1306 * system. And if you are -that- concerned about energy efficiency,
1307 * just power the system down and be done with it!
1309 * The value below works well in practice. If future workloads require
1310 * adjustment, they can be converted into kernel config parameters, though
1311 * making the state machine smarter might be a better option.
1313 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1315 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1316 module_param(rcu_idle_gp_delay, int, 0644);
1319 * Try to advance callbacks on the current CPU, but only if it has been
1320 * awhile since the last time we did so. Afterwards, if there are any
1321 * callbacks ready for immediate invocation, return true.
1323 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1325 bool cbs_ready = false;
1326 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1327 struct rcu_node *rnp;
1329 /* Exit early if we advanced recently. */
1330 if (jiffies == rdp->last_advance_all)
1332 rdp->last_advance_all = jiffies;
1337 * Don't bother checking unless a grace period has
1338 * completed since we last checked and there are
1339 * callbacks not yet ready to invoke.
1341 if ((rcu_seq_completed_gp(rdp->gp_seq,
1342 rcu_seq_current(&rnp->gp_seq)) ||
1343 unlikely(READ_ONCE(rdp->gpwrap))) &&
1344 rcu_segcblist_pend_cbs(&rdp->cblist))
1345 note_gp_changes(rdp);
1347 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1353 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1354 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1355 * caller about what to set the timeout.
1357 * The caller must have disabled interrupts.
1359 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1361 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1364 lockdep_assert_irqs_disabled();
1366 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1367 if (rcu_segcblist_empty(&rdp->cblist) ||
1368 rcu_rdp_is_offloaded(rdp)) {
1369 *nextevt = KTIME_MAX;
1373 /* Attempt to advance callbacks. */
1374 if (rcu_try_advance_all_cbs()) {
1375 /* Some ready to invoke, so initiate later invocation. */
1379 rdp->last_accelerate = jiffies;
1381 /* Request timer and round. */
1382 dj = round_up(rcu_idle_gp_delay + jiffies, rcu_idle_gp_delay) - jiffies;
1384 *nextevt = basemono + dj * TICK_NSEC;
1389 * Prepare a CPU for idle from an RCU perspective. The first major task is to
1390 * sense whether nohz mode has been enabled or disabled via sysfs. The second
1391 * major task is to accelerate (that is, assign grace-period numbers to) any
1392 * recently arrived callbacks.
1394 * The caller must have disabled interrupts.
1396 static void rcu_prepare_for_idle(void)
1399 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1400 struct rcu_node *rnp;
1403 lockdep_assert_irqs_disabled();
1404 if (rcu_rdp_is_offloaded(rdp))
1407 /* Handle nohz enablement switches conservatively. */
1408 tne = READ_ONCE(tick_nohz_active);
1409 if (tne != rdp->tick_nohz_enabled_snap) {
1410 if (!rcu_segcblist_empty(&rdp->cblist))
1411 invoke_rcu_core(); /* force nohz to see update. */
1412 rdp->tick_nohz_enabled_snap = tne;
1419 * If we have not yet accelerated this jiffy, accelerate all
1420 * callbacks on this CPU.
1422 if (rdp->last_accelerate == jiffies)
1424 rdp->last_accelerate = jiffies;
1425 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1427 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1428 needwake = rcu_accelerate_cbs(rnp, rdp);
1429 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1431 rcu_gp_kthread_wake();
1436 * Clean up for exit from idle. Attempt to advance callbacks based on
1437 * any grace periods that elapsed while the CPU was idle, and if any
1438 * callbacks are now ready to invoke, initiate invocation.
1440 static void rcu_cleanup_after_idle(void)
1442 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1444 lockdep_assert_irqs_disabled();
1445 if (rcu_rdp_is_offloaded(rdp))
1447 if (rcu_try_advance_all_cbs())
1451 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1454 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
1455 * grace-period kthread will do force_quiescent_state() processing?
1456 * The idea is to avoid waking up RCU core processing on such a
1457 * CPU unless the grace period has extended for too long.
1459 * This code relies on the fact that all NO_HZ_FULL CPUs are also
1460 * CONFIG_RCU_NOCB_CPU CPUs.
1462 static bool rcu_nohz_full_cpu(void)
1464 #ifdef CONFIG_NO_HZ_FULL
1465 if (tick_nohz_full_cpu(smp_processor_id()) &&
1466 (!rcu_gp_in_progress() ||
1467 time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
1469 #endif /* #ifdef CONFIG_NO_HZ_FULL */
1474 * Bind the RCU grace-period kthreads to the housekeeping CPU.
1476 static void rcu_bind_gp_kthread(void)
1478 if (!tick_nohz_full_enabled())
1480 housekeeping_affine(current, HK_FLAG_RCU);
1483 /* Record the current task on dyntick-idle entry. */
1484 static __always_inline void rcu_dynticks_task_enter(void)
1486 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
1487 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
1488 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
1491 /* Record no current task on dyntick-idle exit. */
1492 static __always_inline void rcu_dynticks_task_exit(void)
1494 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
1495 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
1496 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
1499 /* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
1500 static __always_inline void rcu_dynticks_task_trace_enter(void)
1502 #ifdef CONFIG_TASKS_TRACE_RCU
1503 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1504 current->trc_reader_special.b.need_mb = true;
1505 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
1508 /* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
1509 static __always_inline void rcu_dynticks_task_trace_exit(void)
1511 #ifdef CONFIG_TASKS_TRACE_RCU
1512 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1513 current->trc_reader_special.b.need_mb = false;
1514 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */