2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
31 #include "../time/tick-internal.h"
33 #ifdef CONFIG_RCU_BOOST
35 #include "../locking/rtmutex_common.h"
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
41 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
42 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
43 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
44 DEFINE_PER_CPU(char, rcu_cpu_has_work);
46 #else /* #ifdef CONFIG_RCU_BOOST */
49 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
50 * all uses are in dead code. Provide a definition to keep the compiler
51 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
52 * This probably needs to be excluded from -rt builds.
54 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
56 #endif /* #else #ifdef CONFIG_RCU_BOOST */
58 #ifdef CONFIG_RCU_NOCB_CPU
59 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
60 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
61 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
62 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
65 * Check the RCU kernel configuration parameters and print informative
66 * messages about anything out of the ordinary. If you like #ifdef, you
67 * will love this function.
69 static void __init rcu_bootup_announce_oddness(void)
71 if (IS_ENABLED(CONFIG_RCU_TRACE))
72 pr_info("\tRCU debugfs-based tracing is enabled.\n");
73 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
74 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
75 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
78 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
79 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
80 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
81 if (IS_ENABLED(CONFIG_PROVE_RCU))
82 pr_info("\tRCU lockdep checking is enabled.\n");
83 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE))
84 pr_info("\tRCU torture testing starts during boot.\n");
85 if (RCU_NUM_LVLS >= 4)
86 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
87 if (RCU_FANOUT_LEAF != 16)
88 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
90 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
91 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
92 if (nr_cpu_ids != NR_CPUS)
93 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
94 if (IS_ENABLED(CONFIG_RCU_BOOST))
95 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
98 #ifdef CONFIG_PREEMPT_RCU
100 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
101 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
102 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
104 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
108 * Tell them what RCU they are running.
110 static void __init rcu_bootup_announce(void)
112 pr_info("Preemptible hierarchical RCU implementation.\n");
113 rcu_bootup_announce_oddness();
116 /* Flags for rcu_preempt_ctxt_queue() decision table. */
117 #define RCU_GP_TASKS 0x8
118 #define RCU_EXP_TASKS 0x4
119 #define RCU_GP_BLKD 0x2
120 #define RCU_EXP_BLKD 0x1
123 * Queues a task preempted within an RCU-preempt read-side critical
124 * section into the appropriate location within the ->blkd_tasks list,
125 * depending on the states of any ongoing normal and expedited grace
126 * periods. The ->gp_tasks pointer indicates which element the normal
127 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
128 * indicates which element the expedited grace period is waiting on (again,
129 * NULL if none). If a grace period is waiting on a given element in the
130 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
131 * adding a task to the tail of the list blocks any grace period that is
132 * already waiting on one of the elements. In contrast, adding a task
133 * to the head of the list won't block any grace period that is already
134 * waiting on one of the elements.
136 * This queuing is imprecise, and can sometimes make an ongoing grace
137 * period wait for a task that is not strictly speaking blocking it.
138 * Given the choice, we needlessly block a normal grace period rather than
139 * blocking an expedited grace period.
141 * Note that an endless sequence of expedited grace periods still cannot
142 * indefinitely postpone a normal grace period. Eventually, all of the
143 * fixed number of preempted tasks blocking the normal grace period that are
144 * not also blocking the expedited grace period will resume and complete
145 * their RCU read-side critical sections. At that point, the ->gp_tasks
146 * pointer will equal the ->exp_tasks pointer, at which point the end of
147 * the corresponding expedited grace period will also be the end of the
148 * normal grace period.
150 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp,
151 unsigned long flags) __releases(rnp->lock)
153 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
154 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
155 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
156 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
157 struct task_struct *t = current;
160 * Decide where to queue the newly blocked task. In theory,
161 * this could be an if-statement. In practice, when I tried
162 * that, it was quite messy.
164 switch (blkd_state) {
167 case RCU_EXP_TASKS + RCU_GP_BLKD:
169 case RCU_GP_TASKS + RCU_EXP_TASKS:
172 * Blocking neither GP, or first task blocking the normal
173 * GP but not blocking the already-waiting expedited GP.
174 * Queue at the head of the list to avoid unnecessarily
175 * blocking the already-waiting GPs.
177 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
182 case RCU_GP_BLKD + RCU_EXP_BLKD:
183 case RCU_GP_TASKS + RCU_EXP_BLKD:
184 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
185 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
188 * First task arriving that blocks either GP, or first task
189 * arriving that blocks the expedited GP (with the normal
190 * GP already waiting), or a task arriving that blocks
191 * both GPs with both GPs already waiting. Queue at the
192 * tail of the list to avoid any GP waiting on any of the
193 * already queued tasks that are not blocking it.
195 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
198 case RCU_EXP_TASKS + RCU_EXP_BLKD:
199 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
200 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
203 * Second or subsequent task blocking the expedited GP.
204 * The task either does not block the normal GP, or is the
205 * first task blocking the normal GP. Queue just after
206 * the first task blocking the expedited GP.
208 list_add(&t->rcu_node_entry, rnp->exp_tasks);
211 case RCU_GP_TASKS + RCU_GP_BLKD:
212 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
215 * Second or subsequent task blocking the normal GP.
216 * The task does not block the expedited GP. Queue just
217 * after the first task blocking the normal GP.
219 list_add(&t->rcu_node_entry, rnp->gp_tasks);
224 /* Yet another exercise in excessive paranoia. */
230 * We have now queued the task. If it was the first one to
231 * block either grace period, update the ->gp_tasks and/or
232 * ->exp_tasks pointers, respectively, to reference the newly
235 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD))
236 rnp->gp_tasks = &t->rcu_node_entry;
237 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
238 rnp->exp_tasks = &t->rcu_node_entry;
239 raw_spin_unlock(&rnp->lock);
242 * Report the quiescent state for the expedited GP. This expedited
243 * GP should not be able to end until we report, so there should be
244 * no need to check for a subsequent expedited GP. (Though we are
245 * still in a quiescent state in any case.)
247 if (blkd_state & RCU_EXP_BLKD &&
248 t->rcu_read_unlock_special.b.exp_need_qs) {
249 t->rcu_read_unlock_special.b.exp_need_qs = false;
250 rcu_report_exp_rdp(rdp->rsp, rdp, true);
252 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
254 local_irq_restore(flags);
258 * Record a preemptible-RCU quiescent state for the specified CPU. Note
259 * that this just means that the task currently running on the CPU is
260 * not in a quiescent state. There might be any number of tasks blocked
261 * while in an RCU read-side critical section.
263 * As with the other rcu_*_qs() functions, callers to this function
264 * must disable preemption.
266 static void rcu_preempt_qs(void)
268 if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
269 trace_rcu_grace_period(TPS("rcu_preempt"),
270 __this_cpu_read(rcu_data_p->gpnum),
272 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
273 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
274 current->rcu_read_unlock_special.b.need_qs = false;
279 * We have entered the scheduler, and the current task might soon be
280 * context-switched away from. If this task is in an RCU read-side
281 * critical section, we will no longer be able to rely on the CPU to
282 * record that fact, so we enqueue the task on the blkd_tasks list.
283 * The task will dequeue itself when it exits the outermost enclosing
284 * RCU read-side critical section. Therefore, the current grace period
285 * cannot be permitted to complete until the blkd_tasks list entries
286 * predating the current grace period drain, in other words, until
287 * rnp->gp_tasks becomes NULL.
289 * Caller must disable preemption.
291 static void rcu_preempt_note_context_switch(void)
293 struct task_struct *t = current;
295 struct rcu_data *rdp;
296 struct rcu_node *rnp;
298 if (t->rcu_read_lock_nesting > 0 &&
299 !t->rcu_read_unlock_special.b.blocked) {
301 /* Possibly blocking in an RCU read-side critical section. */
302 rdp = this_cpu_ptr(rcu_state_p->rda);
304 raw_spin_lock_irqsave(&rnp->lock, flags);
305 smp_mb__after_unlock_lock();
306 t->rcu_read_unlock_special.b.blocked = true;
307 t->rcu_blocked_node = rnp;
310 * Verify the CPU's sanity, trace the preemption, and
311 * then queue the task as required based on the states
312 * of any ongoing and expedited grace periods.
314 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
315 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
316 trace_rcu_preempt_task(rdp->rsp->name,
318 (rnp->qsmask & rdp->grpmask)
321 rcu_preempt_ctxt_queue(rnp, rdp, flags);
322 } else if (t->rcu_read_lock_nesting < 0 &&
323 t->rcu_read_unlock_special.s) {
326 * Complete exit from RCU read-side critical section on
327 * behalf of preempted instance of __rcu_read_unlock().
329 rcu_read_unlock_special(t);
333 * Either we were not in an RCU read-side critical section to
334 * begin with, or we have now recorded that critical section
335 * globally. Either way, we can now note a quiescent state
336 * for this CPU. Again, if we were in an RCU read-side critical
337 * section, and if that critical section was blocking the current
338 * grace period, then the fact that the task has been enqueued
339 * means that we continue to block the current grace period.
345 * Check for preempted RCU readers blocking the current grace period
346 * for the specified rcu_node structure. If the caller needs a reliable
347 * answer, it must hold the rcu_node's ->lock.
349 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
351 return rnp->gp_tasks != NULL;
355 * Advance a ->blkd_tasks-list pointer to the next entry, instead
356 * returning NULL if at the end of the list.
358 static struct list_head *rcu_next_node_entry(struct task_struct *t,
359 struct rcu_node *rnp)
361 struct list_head *np;
363 np = t->rcu_node_entry.next;
364 if (np == &rnp->blkd_tasks)
370 * Return true if the specified rcu_node structure has tasks that were
371 * preempted within an RCU read-side critical section.
373 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
375 return !list_empty(&rnp->blkd_tasks);
379 * Handle special cases during rcu_read_unlock(), such as needing to
380 * notify RCU core processing or task having blocked during the RCU
381 * read-side critical section.
383 void rcu_read_unlock_special(struct task_struct *t)
389 struct list_head *np;
390 bool drop_boost_mutex = false;
391 struct rcu_data *rdp;
392 struct rcu_node *rnp;
393 union rcu_special special;
395 /* NMI handlers cannot block and cannot safely manipulate state. */
399 local_irq_save(flags);
402 * If RCU core is waiting for this CPU to exit its critical section,
403 * report the fact that it has exited. Because irqs are disabled,
404 * t->rcu_read_unlock_special cannot change.
406 special = t->rcu_read_unlock_special;
407 if (special.b.need_qs) {
409 t->rcu_read_unlock_special.b.need_qs = false;
410 if (!t->rcu_read_unlock_special.s) {
411 local_irq_restore(flags);
417 * Respond to a request for an expedited grace period, but only if
418 * we were not preempted, meaning that we were running on the same
419 * CPU throughout. If we were preempted, the exp_need_qs flag
420 * would have been cleared at the time of the first preemption,
421 * and the quiescent state would be reported when we were dequeued.
423 if (special.b.exp_need_qs) {
424 WARN_ON_ONCE(special.b.blocked);
425 t->rcu_read_unlock_special.b.exp_need_qs = false;
426 rdp = this_cpu_ptr(rcu_state_p->rda);
427 rcu_report_exp_rdp(rcu_state_p, rdp, true);
428 if (!t->rcu_read_unlock_special.s) {
429 local_irq_restore(flags);
434 /* Hardware IRQ handlers cannot block, complain if they get here. */
435 if (in_irq() || in_serving_softirq()) {
436 lockdep_rcu_suspicious(__FILE__, __LINE__,
437 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
438 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
439 t->rcu_read_unlock_special.s,
440 t->rcu_read_unlock_special.b.blocked,
441 t->rcu_read_unlock_special.b.exp_need_qs,
442 t->rcu_read_unlock_special.b.need_qs);
443 local_irq_restore(flags);
447 /* Clean up if blocked during RCU read-side critical section. */
448 if (special.b.blocked) {
449 t->rcu_read_unlock_special.b.blocked = false;
452 * Remove this task from the list it blocked on. The task
453 * now remains queued on the rcu_node corresponding to
454 * the CPU it first blocked on, so the first attempt to
455 * acquire the task's rcu_node's ->lock will succeed.
456 * Keep the loop and add a WARN_ON() out of sheer paranoia.
459 rnp = t->rcu_blocked_node;
460 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
461 smp_mb__after_unlock_lock();
462 if (rnp == t->rcu_blocked_node)
465 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
467 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
468 empty_exp = sync_rcu_preempt_exp_done(rnp);
469 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
470 np = rcu_next_node_entry(t, rnp);
471 list_del_init(&t->rcu_node_entry);
472 t->rcu_blocked_node = NULL;
473 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
475 if (&t->rcu_node_entry == rnp->gp_tasks)
477 if (&t->rcu_node_entry == rnp->exp_tasks)
479 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
480 if (&t->rcu_node_entry == rnp->boost_tasks)
481 rnp->boost_tasks = np;
482 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
483 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
487 * If this was the last task on the current list, and if
488 * we aren't waiting on any CPUs, report the quiescent state.
489 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
490 * so we must take a snapshot of the expedited state.
492 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
493 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
494 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
501 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
503 raw_spin_unlock_irqrestore(&rnp->lock, flags);
506 /* Unboost if we were boosted. */
507 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
508 rt_mutex_unlock(&rnp->boost_mtx);
511 * If this was the last task on the expedited lists,
512 * then we need to report up the rcu_node hierarchy.
514 if (!empty_exp && empty_exp_now)
515 rcu_report_exp_rnp(rcu_state_p, rnp, true);
517 local_irq_restore(flags);
522 * Dump detailed information for all tasks blocking the current RCU
523 * grace period on the specified rcu_node structure.
525 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
528 struct task_struct *t;
530 raw_spin_lock_irqsave(&rnp->lock, flags);
531 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
532 raw_spin_unlock_irqrestore(&rnp->lock, flags);
535 t = list_entry(rnp->gp_tasks->prev,
536 struct task_struct, rcu_node_entry);
537 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
539 raw_spin_unlock_irqrestore(&rnp->lock, flags);
543 * Dump detailed information for all tasks blocking the current RCU
546 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
548 struct rcu_node *rnp = rcu_get_root(rsp);
550 rcu_print_detail_task_stall_rnp(rnp);
551 rcu_for_each_leaf_node(rsp, rnp)
552 rcu_print_detail_task_stall_rnp(rnp);
555 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
557 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
558 rnp->level, rnp->grplo, rnp->grphi);
561 static void rcu_print_task_stall_end(void)
567 * Scan the current list of tasks blocked within RCU read-side critical
568 * sections, printing out the tid of each.
570 static int rcu_print_task_stall(struct rcu_node *rnp)
572 struct task_struct *t;
575 if (!rcu_preempt_blocked_readers_cgp(rnp))
577 rcu_print_task_stall_begin(rnp);
578 t = list_entry(rnp->gp_tasks->prev,
579 struct task_struct, rcu_node_entry);
580 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
581 pr_cont(" P%d", t->pid);
584 rcu_print_task_stall_end();
589 * Scan the current list of tasks blocked within RCU read-side critical
590 * sections, printing out the tid of each that is blocking the current
591 * expedited grace period.
593 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
595 struct task_struct *t;
600 t = list_entry(rnp->exp_tasks->prev,
601 struct task_struct, rcu_node_entry);
602 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
603 pr_cont(" P%d", t->pid);
610 * Check that the list of blocked tasks for the newly completed grace
611 * period is in fact empty. It is a serious bug to complete a grace
612 * period that still has RCU readers blocked! This function must be
613 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
614 * must be held by the caller.
616 * Also, if there are blocked tasks on the list, they automatically
617 * block the newly created grace period, so set up ->gp_tasks accordingly.
619 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
621 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
622 if (rcu_preempt_has_tasks(rnp))
623 rnp->gp_tasks = rnp->blkd_tasks.next;
624 WARN_ON_ONCE(rnp->qsmask);
628 * Check for a quiescent state from the current CPU. When a task blocks,
629 * the task is recorded in the corresponding CPU's rcu_node structure,
630 * which is checked elsewhere.
632 * Caller must disable hard irqs.
634 static void rcu_preempt_check_callbacks(void)
636 struct task_struct *t = current;
638 if (t->rcu_read_lock_nesting == 0) {
642 if (t->rcu_read_lock_nesting > 0 &&
643 __this_cpu_read(rcu_data_p->core_needs_qs) &&
644 __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm))
645 t->rcu_read_unlock_special.b.need_qs = true;
648 #ifdef CONFIG_RCU_BOOST
650 static void rcu_preempt_do_callbacks(void)
652 rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p));
655 #endif /* #ifdef CONFIG_RCU_BOOST */
658 * Queue a preemptible-RCU callback for invocation after a grace period.
660 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
662 __call_rcu(head, func, rcu_state_p, -1, 0);
664 EXPORT_SYMBOL_GPL(call_rcu);
667 * synchronize_rcu - wait until a grace period has elapsed.
669 * Control will return to the caller some time after a full grace
670 * period has elapsed, in other words after all currently executing RCU
671 * read-side critical sections have completed. Note, however, that
672 * upon return from synchronize_rcu(), the caller might well be executing
673 * concurrently with new RCU read-side critical sections that began while
674 * synchronize_rcu() was waiting. RCU read-side critical sections are
675 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
677 * See the description of synchronize_sched() for more detailed information
678 * on memory ordering guarantees.
680 void synchronize_rcu(void)
682 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
683 lock_is_held(&rcu_lock_map) ||
684 lock_is_held(&rcu_sched_lock_map),
685 "Illegal synchronize_rcu() in RCU read-side critical section");
686 if (!rcu_scheduler_active)
688 if (rcu_gp_is_expedited())
689 synchronize_rcu_expedited();
691 wait_rcu_gp(call_rcu);
693 EXPORT_SYMBOL_GPL(synchronize_rcu);
696 * Remote handler for smp_call_function_single(). If there is an
697 * RCU read-side critical section in effect, request that the
698 * next rcu_read_unlock() record the quiescent state up the
699 * ->expmask fields in the rcu_node tree. Otherwise, immediately
700 * report the quiescent state.
702 static void sync_rcu_exp_handler(void *info)
704 struct rcu_data *rdp;
705 struct rcu_state *rsp = info;
706 struct task_struct *t = current;
709 * Within an RCU read-side critical section, request that the next
710 * rcu_read_unlock() report. Unless this RCU read-side critical
711 * section has already blocked, in which case it is already set
712 * up for the expedited grace period to wait on it.
714 if (t->rcu_read_lock_nesting > 0 &&
715 !t->rcu_read_unlock_special.b.blocked) {
716 t->rcu_read_unlock_special.b.exp_need_qs = true;
721 * We are either exiting an RCU read-side critical section (negative
722 * values of t->rcu_read_lock_nesting) or are not in one at all
723 * (zero value of t->rcu_read_lock_nesting). Or we are in an RCU
724 * read-side critical section that blocked before this expedited
725 * grace period started. Either way, we can immediately report
726 * the quiescent state.
728 rdp = this_cpu_ptr(rsp->rda);
729 rcu_report_exp_rdp(rsp, rdp, true);
733 * synchronize_rcu_expedited - Brute-force RCU grace period
735 * Wait for an RCU-preempt grace period, but expedite it. The basic
736 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
737 * the ->blkd_tasks lists and wait for this list to drain. This consumes
738 * significant time on all CPUs and is unfriendly to real-time workloads,
739 * so is thus not recommended for any sort of common-case code.
740 * In fact, if you are using synchronize_rcu_expedited() in a loop,
741 * please restructure your code to batch your updates, and then Use a
742 * single synchronize_rcu() instead.
744 void synchronize_rcu_expedited(void)
746 struct rcu_node *rnp;
747 struct rcu_node *rnp_unlock;
748 struct rcu_state *rsp = rcu_state_p;
751 s = rcu_exp_gp_seq_snap(rsp);
753 rnp_unlock = exp_funnel_lock(rsp, s);
754 if (rnp_unlock == NULL)
755 return; /* Someone else did our work for us. */
757 rcu_exp_gp_seq_start(rsp);
759 /* Initialize the rcu_node tree in preparation for the wait. */
760 sync_rcu_exp_select_cpus(rsp, sync_rcu_exp_handler);
762 /* Wait for snapshotted ->blkd_tasks lists to drain. */
763 rnp = rcu_get_root(rsp);
764 wait_event(rsp->expedited_wq,
765 sync_rcu_preempt_exp_done(rnp));
767 /* Clean up and exit. */
768 rcu_exp_gp_seq_end(rsp);
769 mutex_unlock(&rnp_unlock->exp_funnel_mutex);
771 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
774 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
776 * Note that this primitive does not necessarily wait for an RCU grace period
777 * to complete. For example, if there are no RCU callbacks queued anywhere
778 * in the system, then rcu_barrier() is within its rights to return
779 * immediately, without waiting for anything, much less an RCU grace period.
781 void rcu_barrier(void)
783 _rcu_barrier(rcu_state_p);
785 EXPORT_SYMBOL_GPL(rcu_barrier);
788 * Initialize preemptible RCU's state structures.
790 static void __init __rcu_init_preempt(void)
792 rcu_init_one(rcu_state_p, rcu_data_p);
796 * Check for a task exiting while in a preemptible-RCU read-side
797 * critical section, clean up if so. No need to issue warnings,
798 * as debug_check_no_locks_held() already does this if lockdep
803 struct task_struct *t = current;
805 if (likely(list_empty(¤t->rcu_node_entry)))
807 t->rcu_read_lock_nesting = 1;
809 t->rcu_read_unlock_special.b.blocked = true;
813 #else /* #ifdef CONFIG_PREEMPT_RCU */
815 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
816 static struct rcu_data __percpu *const rcu_data_p = &rcu_sched_data;
819 * Tell them what RCU they are running.
821 static void __init rcu_bootup_announce(void)
823 pr_info("Hierarchical RCU implementation.\n");
824 rcu_bootup_announce_oddness();
828 * Because preemptible RCU does not exist, we never have to check for
829 * CPUs being in quiescent states.
831 static void rcu_preempt_note_context_switch(void)
836 * Because preemptible RCU does not exist, there are never any preempted
839 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
845 * Because there is no preemptible RCU, there can be no readers blocked.
847 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
853 * Because preemptible RCU does not exist, we never have to check for
854 * tasks blocked within RCU read-side critical sections.
856 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
861 * Because preemptible RCU does not exist, we never have to check for
862 * tasks blocked within RCU read-side critical sections.
864 static int rcu_print_task_stall(struct rcu_node *rnp)
870 * Because preemptible RCU does not exist, we never have to check for
871 * tasks blocked within RCU read-side critical sections that are
872 * blocking the current expedited grace period.
874 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
880 * Because there is no preemptible RCU, there can be no readers blocked,
881 * so there is no need to check for blocked tasks. So check only for
882 * bogus qsmask values.
884 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
886 WARN_ON_ONCE(rnp->qsmask);
890 * Because preemptible RCU does not exist, it never has any callbacks
893 static void rcu_preempt_check_callbacks(void)
898 * Wait for an rcu-preempt grace period, but make it happen quickly.
899 * But because preemptible RCU does not exist, map to rcu-sched.
901 void synchronize_rcu_expedited(void)
903 synchronize_sched_expedited();
905 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
908 * Because preemptible RCU does not exist, rcu_barrier() is just
909 * another name for rcu_barrier_sched().
911 void rcu_barrier(void)
915 EXPORT_SYMBOL_GPL(rcu_barrier);
918 * Because preemptible RCU does not exist, it need not be initialized.
920 static void __init __rcu_init_preempt(void)
925 * Because preemptible RCU does not exist, tasks cannot possibly exit
926 * while in preemptible RCU read-side critical sections.
932 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
934 #ifdef CONFIG_RCU_BOOST
936 #include "../locking/rtmutex_common.h"
938 #ifdef CONFIG_RCU_TRACE
940 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
942 if (!rcu_preempt_has_tasks(rnp))
943 rnp->n_balk_blkd_tasks++;
944 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
945 rnp->n_balk_exp_gp_tasks++;
946 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
947 rnp->n_balk_boost_tasks++;
948 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
949 rnp->n_balk_notblocked++;
950 else if (rnp->gp_tasks != NULL &&
951 ULONG_CMP_LT(jiffies, rnp->boost_time))
952 rnp->n_balk_notyet++;
957 #else /* #ifdef CONFIG_RCU_TRACE */
959 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
963 #endif /* #else #ifdef CONFIG_RCU_TRACE */
965 static void rcu_wake_cond(struct task_struct *t, int status)
968 * If the thread is yielding, only wake it when this
969 * is invoked from idle
971 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
976 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
977 * or ->boost_tasks, advancing the pointer to the next task in the
980 * Note that irqs must be enabled: boosting the task can block.
981 * Returns 1 if there are more tasks needing to be boosted.
983 static int rcu_boost(struct rcu_node *rnp)
986 struct task_struct *t;
987 struct list_head *tb;
989 if (READ_ONCE(rnp->exp_tasks) == NULL &&
990 READ_ONCE(rnp->boost_tasks) == NULL)
991 return 0; /* Nothing left to boost. */
993 raw_spin_lock_irqsave(&rnp->lock, flags);
994 smp_mb__after_unlock_lock();
997 * Recheck under the lock: all tasks in need of boosting
998 * might exit their RCU read-side critical sections on their own.
1000 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1001 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1006 * Preferentially boost tasks blocking expedited grace periods.
1007 * This cannot starve the normal grace periods because a second
1008 * expedited grace period must boost all blocked tasks, including
1009 * those blocking the pre-existing normal grace period.
1011 if (rnp->exp_tasks != NULL) {
1012 tb = rnp->exp_tasks;
1013 rnp->n_exp_boosts++;
1015 tb = rnp->boost_tasks;
1016 rnp->n_normal_boosts++;
1018 rnp->n_tasks_boosted++;
1021 * We boost task t by manufacturing an rt_mutex that appears to
1022 * be held by task t. We leave a pointer to that rt_mutex where
1023 * task t can find it, and task t will release the mutex when it
1024 * exits its outermost RCU read-side critical section. Then
1025 * simply acquiring this artificial rt_mutex will boost task
1026 * t's priority. (Thanks to tglx for suggesting this approach!)
1028 * Note that task t must acquire rnp->lock to remove itself from
1029 * the ->blkd_tasks list, which it will do from exit() if from
1030 * nowhere else. We therefore are guaranteed that task t will
1031 * stay around at least until we drop rnp->lock. Note that
1032 * rnp->lock also resolves races between our priority boosting
1033 * and task t's exiting its outermost RCU read-side critical
1036 t = container_of(tb, struct task_struct, rcu_node_entry);
1037 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1038 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1039 /* Lock only for side effect: boosts task t's priority. */
1040 rt_mutex_lock(&rnp->boost_mtx);
1041 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1043 return READ_ONCE(rnp->exp_tasks) != NULL ||
1044 READ_ONCE(rnp->boost_tasks) != NULL;
1048 * Priority-boosting kthread, one per leaf rcu_node.
1050 static int rcu_boost_kthread(void *arg)
1052 struct rcu_node *rnp = (struct rcu_node *)arg;
1056 trace_rcu_utilization(TPS("Start boost kthread@init"));
1058 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1059 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1060 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1061 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1062 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1063 more2boost = rcu_boost(rnp);
1069 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1070 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1071 schedule_timeout_interruptible(2);
1072 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1077 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1082 * Check to see if it is time to start boosting RCU readers that are
1083 * blocking the current grace period, and, if so, tell the per-rcu_node
1084 * kthread to start boosting them. If there is an expedited grace
1085 * period in progress, it is always time to boost.
1087 * The caller must hold rnp->lock, which this function releases.
1088 * The ->boost_kthread_task is immortal, so we don't need to worry
1089 * about it going away.
1091 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1092 __releases(rnp->lock)
1094 struct task_struct *t;
1096 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1097 rnp->n_balk_exp_gp_tasks++;
1098 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1101 if (rnp->exp_tasks != NULL ||
1102 (rnp->gp_tasks != NULL &&
1103 rnp->boost_tasks == NULL &&
1105 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1106 if (rnp->exp_tasks == NULL)
1107 rnp->boost_tasks = rnp->gp_tasks;
1108 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1109 t = rnp->boost_kthread_task;
1111 rcu_wake_cond(t, rnp->boost_kthread_status);
1113 rcu_initiate_boost_trace(rnp);
1114 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1119 * Wake up the per-CPU kthread to invoke RCU callbacks.
1121 static void invoke_rcu_callbacks_kthread(void)
1123 unsigned long flags;
1125 local_irq_save(flags);
1126 __this_cpu_write(rcu_cpu_has_work, 1);
1127 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1128 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1129 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1130 __this_cpu_read(rcu_cpu_kthread_status));
1132 local_irq_restore(flags);
1136 * Is the current CPU running the RCU-callbacks kthread?
1137 * Caller must have preemption disabled.
1139 static bool rcu_is_callbacks_kthread(void)
1141 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1144 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1147 * Do priority-boost accounting for the start of a new grace period.
1149 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1151 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1155 * Create an RCU-boost kthread for the specified node if one does not
1156 * already exist. We only create this kthread for preemptible RCU.
1157 * Returns zero if all is well, a negated errno otherwise.
1159 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1160 struct rcu_node *rnp)
1162 int rnp_index = rnp - &rsp->node[0];
1163 unsigned long flags;
1164 struct sched_param sp;
1165 struct task_struct *t;
1167 if (rcu_state_p != rsp)
1170 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1174 if (rnp->boost_kthread_task != NULL)
1176 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1177 "rcub/%d", rnp_index);
1180 raw_spin_lock_irqsave(&rnp->lock, flags);
1181 smp_mb__after_unlock_lock();
1182 rnp->boost_kthread_task = t;
1183 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1184 sp.sched_priority = kthread_prio;
1185 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1186 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1190 static void rcu_kthread_do_work(void)
1192 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1193 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1194 rcu_preempt_do_callbacks();
1197 static void rcu_cpu_kthread_setup(unsigned int cpu)
1199 struct sched_param sp;
1201 sp.sched_priority = kthread_prio;
1202 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1205 static void rcu_cpu_kthread_park(unsigned int cpu)
1207 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1210 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1212 return __this_cpu_read(rcu_cpu_has_work);
1216 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1217 * RCU softirq used in flavors and configurations of RCU that do not
1218 * support RCU priority boosting.
1220 static void rcu_cpu_kthread(unsigned int cpu)
1222 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1223 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1226 for (spincnt = 0; spincnt < 10; spincnt++) {
1227 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1229 *statusp = RCU_KTHREAD_RUNNING;
1230 this_cpu_inc(rcu_cpu_kthread_loops);
1231 local_irq_disable();
1236 rcu_kthread_do_work();
1239 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1240 *statusp = RCU_KTHREAD_WAITING;
1244 *statusp = RCU_KTHREAD_YIELDING;
1245 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1246 schedule_timeout_interruptible(2);
1247 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1248 *statusp = RCU_KTHREAD_WAITING;
1252 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1253 * served by the rcu_node in question. The CPU hotplug lock is still
1254 * held, so the value of rnp->qsmaskinit will be stable.
1256 * We don't include outgoingcpu in the affinity set, use -1 if there is
1257 * no outgoing CPU. If there are no CPUs left in the affinity set,
1258 * this function allows the kthread to execute on any CPU.
1260 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1262 struct task_struct *t = rnp->boost_kthread_task;
1263 unsigned long mask = rcu_rnp_online_cpus(rnp);
1269 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1271 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1272 if ((mask & 0x1) && cpu != outgoingcpu)
1273 cpumask_set_cpu(cpu, cm);
1274 if (cpumask_weight(cm) == 0)
1276 set_cpus_allowed_ptr(t, cm);
1277 free_cpumask_var(cm);
1280 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1281 .store = &rcu_cpu_kthread_task,
1282 .thread_should_run = rcu_cpu_kthread_should_run,
1283 .thread_fn = rcu_cpu_kthread,
1284 .thread_comm = "rcuc/%u",
1285 .setup = rcu_cpu_kthread_setup,
1286 .park = rcu_cpu_kthread_park,
1290 * Spawn boost kthreads -- called as soon as the scheduler is running.
1292 static void __init rcu_spawn_boost_kthreads(void)
1294 struct rcu_node *rnp;
1297 for_each_possible_cpu(cpu)
1298 per_cpu(rcu_cpu_has_work, cpu) = 0;
1299 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1300 rcu_for_each_leaf_node(rcu_state_p, rnp)
1301 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1304 static void rcu_prepare_kthreads(int cpu)
1306 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1307 struct rcu_node *rnp = rdp->mynode;
1309 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1310 if (rcu_scheduler_fully_active)
1311 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1314 #else /* #ifdef CONFIG_RCU_BOOST */
1316 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1317 __releases(rnp->lock)
1319 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1322 static void invoke_rcu_callbacks_kthread(void)
1327 static bool rcu_is_callbacks_kthread(void)
1332 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1336 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1340 static void __init rcu_spawn_boost_kthreads(void)
1344 static void rcu_prepare_kthreads(int cpu)
1348 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1350 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1353 * Check to see if any future RCU-related work will need to be done
1354 * by the current CPU, even if none need be done immediately, returning
1355 * 1 if so. This function is part of the RCU implementation; it is -not-
1356 * an exported member of the RCU API.
1358 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1359 * any flavor of RCU.
1361 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1363 *nextevt = KTIME_MAX;
1364 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1365 ? 0 : rcu_cpu_has_callbacks(NULL);
1369 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1372 static void rcu_cleanup_after_idle(void)
1377 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1380 static void rcu_prepare_for_idle(void)
1385 * Don't bother keeping a running count of the number of RCU callbacks
1386 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1388 static void rcu_idle_count_callbacks_posted(void)
1392 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1395 * This code is invoked when a CPU goes idle, at which point we want
1396 * to have the CPU do everything required for RCU so that it can enter
1397 * the energy-efficient dyntick-idle mode. This is handled by a
1398 * state machine implemented by rcu_prepare_for_idle() below.
1400 * The following three proprocessor symbols control this state machine:
1402 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1403 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1404 * is sized to be roughly one RCU grace period. Those energy-efficiency
1405 * benchmarkers who might otherwise be tempted to set this to a large
1406 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1407 * system. And if you are -that- concerned about energy efficiency,
1408 * just power the system down and be done with it!
1409 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1410 * permitted to sleep in dyntick-idle mode with only lazy RCU
1411 * callbacks pending. Setting this too high can OOM your system.
1413 * The values below work well in practice. If future workloads require
1414 * adjustment, they can be converted into kernel config parameters, though
1415 * making the state machine smarter might be a better option.
1417 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1418 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1420 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1421 module_param(rcu_idle_gp_delay, int, 0644);
1422 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1423 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1426 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1427 * only if it has been awhile since the last time we did so. Afterwards,
1428 * if there are any callbacks ready for immediate invocation, return true.
1430 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1432 bool cbs_ready = false;
1433 struct rcu_data *rdp;
1434 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1435 struct rcu_node *rnp;
1436 struct rcu_state *rsp;
1438 /* Exit early if we advanced recently. */
1439 if (jiffies == rdtp->last_advance_all)
1441 rdtp->last_advance_all = jiffies;
1443 for_each_rcu_flavor(rsp) {
1444 rdp = this_cpu_ptr(rsp->rda);
1448 * Don't bother checking unless a grace period has
1449 * completed since we last checked and there are
1450 * callbacks not yet ready to invoke.
1452 if ((rdp->completed != rnp->completed ||
1453 unlikely(READ_ONCE(rdp->gpwrap))) &&
1454 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1455 note_gp_changes(rsp, rdp);
1457 if (cpu_has_callbacks_ready_to_invoke(rdp))
1464 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1465 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1466 * caller to set the timeout based on whether or not there are non-lazy
1469 * The caller must have disabled interrupts.
1471 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1473 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1476 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1477 *nextevt = KTIME_MAX;
1481 /* Snapshot to detect later posting of non-lazy callback. */
1482 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1484 /* If no callbacks, RCU doesn't need the CPU. */
1485 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1486 *nextevt = KTIME_MAX;
1490 /* Attempt to advance callbacks. */
1491 if (rcu_try_advance_all_cbs()) {
1492 /* Some ready to invoke, so initiate later invocation. */
1496 rdtp->last_accelerate = jiffies;
1498 /* Request timer delay depending on laziness, and round. */
1499 if (!rdtp->all_lazy) {
1500 dj = round_up(rcu_idle_gp_delay + jiffies,
1501 rcu_idle_gp_delay) - jiffies;
1503 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1505 *nextevt = basemono + dj * TICK_NSEC;
1510 * Prepare a CPU for idle from an RCU perspective. The first major task
1511 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1512 * The second major task is to check to see if a non-lazy callback has
1513 * arrived at a CPU that previously had only lazy callbacks. The third
1514 * major task is to accelerate (that is, assign grace-period numbers to)
1515 * any recently arrived callbacks.
1517 * The caller must have disabled interrupts.
1519 static void rcu_prepare_for_idle(void)
1522 struct rcu_data *rdp;
1523 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1524 struct rcu_node *rnp;
1525 struct rcu_state *rsp;
1528 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL))
1531 /* Handle nohz enablement switches conservatively. */
1532 tne = READ_ONCE(tick_nohz_active);
1533 if (tne != rdtp->tick_nohz_enabled_snap) {
1534 if (rcu_cpu_has_callbacks(NULL))
1535 invoke_rcu_core(); /* force nohz to see update. */
1536 rdtp->tick_nohz_enabled_snap = tne;
1542 /* If this is a no-CBs CPU, no callbacks, just return. */
1543 if (rcu_is_nocb_cpu(smp_processor_id()))
1547 * If a non-lazy callback arrived at a CPU having only lazy
1548 * callbacks, invoke RCU core for the side-effect of recalculating
1549 * idle duration on re-entry to idle.
1551 if (rdtp->all_lazy &&
1552 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1553 rdtp->all_lazy = false;
1554 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1560 * If we have not yet accelerated this jiffy, accelerate all
1561 * callbacks on this CPU.
1563 if (rdtp->last_accelerate == jiffies)
1565 rdtp->last_accelerate = jiffies;
1566 for_each_rcu_flavor(rsp) {
1567 rdp = this_cpu_ptr(rsp->rda);
1568 if (!*rdp->nxttail[RCU_DONE_TAIL])
1571 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1572 smp_mb__after_unlock_lock();
1573 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1574 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1576 rcu_gp_kthread_wake(rsp);
1581 * Clean up for exit from idle. Attempt to advance callbacks based on
1582 * any grace periods that elapsed while the CPU was idle, and if any
1583 * callbacks are now ready to invoke, initiate invocation.
1585 static void rcu_cleanup_after_idle(void)
1587 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1588 rcu_is_nocb_cpu(smp_processor_id()))
1590 if (rcu_try_advance_all_cbs())
1595 * Keep a running count of the number of non-lazy callbacks posted
1596 * on this CPU. This running counter (which is never decremented) allows
1597 * rcu_prepare_for_idle() to detect when something out of the idle loop
1598 * posts a callback, even if an equal number of callbacks are invoked.
1599 * Of course, callbacks should only be posted from within a trace event
1600 * designed to be called from idle or from within RCU_NONIDLE().
1602 static void rcu_idle_count_callbacks_posted(void)
1604 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1608 * Data for flushing lazy RCU callbacks at OOM time.
1610 static atomic_t oom_callback_count;
1611 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1614 * RCU OOM callback -- decrement the outstanding count and deliver the
1615 * wake-up if we are the last one.
1617 static void rcu_oom_callback(struct rcu_head *rhp)
1619 if (atomic_dec_and_test(&oom_callback_count))
1620 wake_up(&oom_callback_wq);
1624 * Post an rcu_oom_notify callback on the current CPU if it has at
1625 * least one lazy callback. This will unnecessarily post callbacks
1626 * to CPUs that already have a non-lazy callback at the end of their
1627 * callback list, but this is an infrequent operation, so accept some
1628 * extra overhead to keep things simple.
1630 static void rcu_oom_notify_cpu(void *unused)
1632 struct rcu_state *rsp;
1633 struct rcu_data *rdp;
1635 for_each_rcu_flavor(rsp) {
1636 rdp = raw_cpu_ptr(rsp->rda);
1637 if (rdp->qlen_lazy != 0) {
1638 atomic_inc(&oom_callback_count);
1639 rsp->call(&rdp->oom_head, rcu_oom_callback);
1645 * If low on memory, ensure that each CPU has a non-lazy callback.
1646 * This will wake up CPUs that have only lazy callbacks, in turn
1647 * ensuring that they free up the corresponding memory in a timely manner.
1648 * Because an uncertain amount of memory will be freed in some uncertain
1649 * timeframe, we do not claim to have freed anything.
1651 static int rcu_oom_notify(struct notifier_block *self,
1652 unsigned long notused, void *nfreed)
1656 /* Wait for callbacks from earlier instance to complete. */
1657 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1658 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1661 * Prevent premature wakeup: ensure that all increments happen
1662 * before there is a chance of the counter reaching zero.
1664 atomic_set(&oom_callback_count, 1);
1666 for_each_online_cpu(cpu) {
1667 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1668 cond_resched_rcu_qs();
1671 /* Unconditionally decrement: no need to wake ourselves up. */
1672 atomic_dec(&oom_callback_count);
1677 static struct notifier_block rcu_oom_nb = {
1678 .notifier_call = rcu_oom_notify
1681 static int __init rcu_register_oom_notifier(void)
1683 register_oom_notifier(&rcu_oom_nb);
1686 early_initcall(rcu_register_oom_notifier);
1688 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1690 #ifdef CONFIG_RCU_FAST_NO_HZ
1692 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1694 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1695 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1697 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1698 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1700 rdtp->all_lazy ? 'L' : '.',
1701 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1704 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1706 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1711 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1713 /* Initiate the stall-info list. */
1714 static void print_cpu_stall_info_begin(void)
1720 * Print out diagnostic information for the specified stalled CPU.
1722 * If the specified CPU is aware of the current RCU grace period
1723 * (flavor specified by rsp), then print the number of scheduling
1724 * clock interrupts the CPU has taken during the time that it has
1725 * been aware. Otherwise, print the number of RCU grace periods
1726 * that this CPU is ignorant of, for example, "1" if the CPU was
1727 * aware of the previous grace period.
1729 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1731 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1733 char fast_no_hz[72];
1734 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1735 struct rcu_dynticks *rdtp = rdp->dynticks;
1737 unsigned long ticks_value;
1739 if (rsp->gpnum == rdp->gpnum) {
1740 ticks_title = "ticks this GP";
1741 ticks_value = rdp->ticks_this_gp;
1743 ticks_title = "GPs behind";
1744 ticks_value = rsp->gpnum - rdp->gpnum;
1746 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1747 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1748 cpu, ticks_value, ticks_title,
1749 atomic_read(&rdtp->dynticks) & 0xfff,
1750 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1751 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1752 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1756 /* Terminate the stall-info list. */
1757 static void print_cpu_stall_info_end(void)
1762 /* Zero ->ticks_this_gp for all flavors of RCU. */
1763 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1765 rdp->ticks_this_gp = 0;
1766 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1769 /* Increment ->ticks_this_gp for all flavors of RCU. */
1770 static void increment_cpu_stall_ticks(void)
1772 struct rcu_state *rsp;
1774 for_each_rcu_flavor(rsp)
1775 raw_cpu_inc(rsp->rda->ticks_this_gp);
1778 #ifdef CONFIG_RCU_NOCB_CPU
1781 * Offload callback processing from the boot-time-specified set of CPUs
1782 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1783 * kthread created that pulls the callbacks from the corresponding CPU,
1784 * waits for a grace period to elapse, and invokes the callbacks.
1785 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1786 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1787 * has been specified, in which case each kthread actively polls its
1788 * CPU. (Which isn't so great for energy efficiency, but which does
1789 * reduce RCU's overhead on that CPU.)
1791 * This is intended to be used in conjunction with Frederic Weisbecker's
1792 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1793 * running CPU-bound user-mode computations.
1795 * Offloading of callback processing could also in theory be used as
1796 * an energy-efficiency measure because CPUs with no RCU callbacks
1797 * queued are more aggressive about entering dyntick-idle mode.
1801 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1802 static int __init rcu_nocb_setup(char *str)
1804 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1805 have_rcu_nocb_mask = true;
1806 cpulist_parse(str, rcu_nocb_mask);
1809 __setup("rcu_nocbs=", rcu_nocb_setup);
1811 static int __init parse_rcu_nocb_poll(char *arg)
1816 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1819 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1822 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1824 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
1828 * Set the root rcu_node structure's ->need_future_gp field
1829 * based on the sum of those of all rcu_node structures. This does
1830 * double-count the root rcu_node structure's requests, but this
1831 * is necessary to handle the possibility of a rcu_nocb_kthread()
1832 * having awakened during the time that the rcu_node structures
1833 * were being updated for the end of the previous grace period.
1835 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1837 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1840 static void rcu_init_one_nocb(struct rcu_node *rnp)
1842 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
1843 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
1846 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1847 /* Is the specified CPU a no-CBs CPU? */
1848 bool rcu_is_nocb_cpu(int cpu)
1850 if (have_rcu_nocb_mask)
1851 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1854 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1857 * Kick the leader kthread for this NOCB group.
1859 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1861 struct rcu_data *rdp_leader = rdp->nocb_leader;
1863 if (!READ_ONCE(rdp_leader->nocb_kthread))
1865 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1866 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1867 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1868 wake_up(&rdp_leader->nocb_wq);
1873 * Does the specified CPU need an RCU callback for the specified flavor
1876 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1878 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1880 #ifdef CONFIG_PROVE_RCU
1881 struct rcu_head *rhp;
1882 #endif /* #ifdef CONFIG_PROVE_RCU */
1885 * Check count of all no-CBs callbacks awaiting invocation.
1886 * There needs to be a barrier before this function is called,
1887 * but associated with a prior determination that no more
1888 * callbacks would be posted. In the worst case, the first
1889 * barrier in _rcu_barrier() suffices (but the caller cannot
1890 * necessarily rely on this, not a substitute for the caller
1891 * getting the concurrency design right!). There must also be
1892 * a barrier between the following load an posting of a callback
1893 * (if a callback is in fact needed). This is associated with an
1894 * atomic_inc() in the caller.
1896 ret = atomic_long_read(&rdp->nocb_q_count);
1898 #ifdef CONFIG_PROVE_RCU
1899 rhp = READ_ONCE(rdp->nocb_head);
1901 rhp = READ_ONCE(rdp->nocb_gp_head);
1903 rhp = READ_ONCE(rdp->nocb_follower_head);
1905 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1906 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1907 rcu_scheduler_fully_active) {
1908 /* RCU callback enqueued before CPU first came online??? */
1909 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1913 #endif /* #ifdef CONFIG_PROVE_RCU */
1919 * Enqueue the specified string of rcu_head structures onto the specified
1920 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1921 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1922 * counts are supplied by rhcount and rhcount_lazy.
1924 * If warranted, also wake up the kthread servicing this CPUs queues.
1926 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1927 struct rcu_head *rhp,
1928 struct rcu_head **rhtp,
1929 int rhcount, int rhcount_lazy,
1930 unsigned long flags)
1933 struct rcu_head **old_rhpp;
1934 struct task_struct *t;
1936 /* Enqueue the callback on the nocb list and update counts. */
1937 atomic_long_add(rhcount, &rdp->nocb_q_count);
1938 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1939 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1940 WRITE_ONCE(*old_rhpp, rhp);
1941 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1942 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1944 /* If we are not being polled and there is a kthread, awaken it ... */
1945 t = READ_ONCE(rdp->nocb_kthread);
1946 if (rcu_nocb_poll || !t) {
1947 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1948 TPS("WakeNotPoll"));
1951 len = atomic_long_read(&rdp->nocb_q_count);
1952 if (old_rhpp == &rdp->nocb_head) {
1953 if (!irqs_disabled_flags(flags)) {
1954 /* ... if queue was empty ... */
1955 wake_nocb_leader(rdp, false);
1956 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1959 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1960 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1961 TPS("WakeEmptyIsDeferred"));
1963 rdp->qlen_last_fqs_check = 0;
1964 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1965 /* ... or if many callbacks queued. */
1966 if (!irqs_disabled_flags(flags)) {
1967 wake_nocb_leader(rdp, true);
1968 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1971 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1972 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1973 TPS("WakeOvfIsDeferred"));
1975 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1977 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1983 * This is a helper for __call_rcu(), which invokes this when the normal
1984 * callback queue is inoperable. If this is not a no-CBs CPU, this
1985 * function returns failure back to __call_rcu(), which can complain
1988 * Otherwise, this function queues the callback where the corresponding
1989 * "rcuo" kthread can find it.
1991 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1992 bool lazy, unsigned long flags)
1995 if (!rcu_is_nocb_cpu(rdp->cpu))
1997 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1998 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1999 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2000 (unsigned long)rhp->func,
2001 -atomic_long_read(&rdp->nocb_q_count_lazy),
2002 -atomic_long_read(&rdp->nocb_q_count));
2004 trace_rcu_callback(rdp->rsp->name, rhp,
2005 -atomic_long_read(&rdp->nocb_q_count_lazy),
2006 -atomic_long_read(&rdp->nocb_q_count));
2009 * If called from an extended quiescent state with interrupts
2010 * disabled, invoke the RCU core in order to allow the idle-entry
2011 * deferred-wakeup check to function.
2013 if (irqs_disabled_flags(flags) &&
2014 !rcu_is_watching() &&
2015 cpu_online(smp_processor_id()))
2022 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2025 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2026 struct rcu_data *rdp,
2027 unsigned long flags)
2029 long ql = rsp->qlen;
2030 long qll = rsp->qlen_lazy;
2032 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2033 if (!rcu_is_nocb_cpu(smp_processor_id()))
2038 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2039 if (rsp->orphan_donelist != NULL) {
2040 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2041 rsp->orphan_donetail, ql, qll, flags);
2043 rsp->orphan_donelist = NULL;
2044 rsp->orphan_donetail = &rsp->orphan_donelist;
2046 if (rsp->orphan_nxtlist != NULL) {
2047 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2048 rsp->orphan_nxttail, ql, qll, flags);
2050 rsp->orphan_nxtlist = NULL;
2051 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2057 * If necessary, kick off a new grace period, and either way wait
2058 * for a subsequent grace period to complete.
2060 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2064 unsigned long flags;
2066 struct rcu_node *rnp = rdp->mynode;
2068 raw_spin_lock_irqsave(&rnp->lock, flags);
2069 smp_mb__after_unlock_lock();
2070 needwake = rcu_start_future_gp(rnp, rdp, &c);
2071 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2073 rcu_gp_kthread_wake(rdp->rsp);
2076 * Wait for the grace period. Do so interruptibly to avoid messing
2077 * up the load average.
2079 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2081 wait_event_interruptible(
2082 rnp->nocb_gp_wq[c & 0x1],
2083 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
2086 WARN_ON(signal_pending(current));
2087 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2089 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2090 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2094 * Leaders come here to wait for additional callbacks to show up.
2095 * This function does not return until callbacks appear.
2097 static void nocb_leader_wait(struct rcu_data *my_rdp)
2099 bool firsttime = true;
2101 struct rcu_data *rdp;
2102 struct rcu_head **tail;
2106 /* Wait for callbacks to appear. */
2107 if (!rcu_nocb_poll) {
2108 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2109 wait_event_interruptible(my_rdp->nocb_wq,
2110 !READ_ONCE(my_rdp->nocb_leader_sleep));
2111 /* Memory barrier handled by smp_mb() calls below and repoll. */
2112 } else if (firsttime) {
2113 firsttime = false; /* Don't drown trace log with "Poll"! */
2114 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2118 * Each pass through the following loop checks a follower for CBs.
2119 * We are our own first follower. Any CBs found are moved to
2120 * nocb_gp_head, where they await a grace period.
2123 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2124 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2125 if (!rdp->nocb_gp_head)
2126 continue; /* No CBs here, try next follower. */
2128 /* Move callbacks to wait-for-GP list, which is empty. */
2129 WRITE_ONCE(rdp->nocb_head, NULL);
2130 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2135 * If there were no callbacks, sleep a bit, rescan after a
2136 * memory barrier, and go retry.
2138 if (unlikely(!gotcbs)) {
2140 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2142 WARN_ON(signal_pending(current));
2143 schedule_timeout_interruptible(1);
2145 /* Rescan in case we were a victim of memory ordering. */
2146 my_rdp->nocb_leader_sleep = true;
2147 smp_mb(); /* Ensure _sleep true before scan. */
2148 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2149 if (READ_ONCE(rdp->nocb_head)) {
2150 /* Found CB, so short-circuit next wait. */
2151 my_rdp->nocb_leader_sleep = false;
2157 /* Wait for one grace period. */
2158 rcu_nocb_wait_gp(my_rdp);
2161 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2162 * We set it now, but recheck for new callbacks while
2163 * traversing our follower list.
2165 my_rdp->nocb_leader_sleep = true;
2166 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2168 /* Each pass through the following loop wakes a follower, if needed. */
2169 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2170 if (READ_ONCE(rdp->nocb_head))
2171 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2172 if (!rdp->nocb_gp_head)
2173 continue; /* No CBs, so no need to wake follower. */
2175 /* Append callbacks to follower's "done" list. */
2176 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2177 *tail = rdp->nocb_gp_head;
2178 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2179 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2181 * List was empty, wake up the follower.
2182 * Memory barriers supplied by atomic_long_add().
2184 wake_up(&rdp->nocb_wq);
2188 /* If we (the leader) don't have CBs, go wait some more. */
2189 if (!my_rdp->nocb_follower_head)
2194 * Followers come here to wait for additional callbacks to show up.
2195 * This function does not return until callbacks appear.
2197 static void nocb_follower_wait(struct rcu_data *rdp)
2199 bool firsttime = true;
2202 if (!rcu_nocb_poll) {
2203 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2205 wait_event_interruptible(rdp->nocb_wq,
2206 READ_ONCE(rdp->nocb_follower_head));
2207 } else if (firsttime) {
2208 /* Don't drown trace log with "Poll"! */
2210 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2212 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2213 /* ^^^ Ensure CB invocation follows _head test. */
2217 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2219 WARN_ON(signal_pending(current));
2220 schedule_timeout_interruptible(1);
2225 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2226 * callbacks queued by the corresponding no-CBs CPU, however, there is
2227 * an optional leader-follower relationship so that the grace-period
2228 * kthreads don't have to do quite so many wakeups.
2230 static int rcu_nocb_kthread(void *arg)
2233 struct rcu_head *list;
2234 struct rcu_head *next;
2235 struct rcu_head **tail;
2236 struct rcu_data *rdp = arg;
2238 /* Each pass through this loop invokes one batch of callbacks */
2240 /* Wait for callbacks. */
2241 if (rdp->nocb_leader == rdp)
2242 nocb_leader_wait(rdp);
2244 nocb_follower_wait(rdp);
2246 /* Pull the ready-to-invoke callbacks onto local list. */
2247 list = READ_ONCE(rdp->nocb_follower_head);
2249 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2250 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2251 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2253 /* Each pass through the following loop invokes a callback. */
2254 trace_rcu_batch_start(rdp->rsp->name,
2255 atomic_long_read(&rdp->nocb_q_count_lazy),
2256 atomic_long_read(&rdp->nocb_q_count), -1);
2260 /* Wait for enqueuing to complete, if needed. */
2261 while (next == NULL && &list->next != tail) {
2262 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2264 schedule_timeout_interruptible(1);
2265 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2269 debug_rcu_head_unqueue(list);
2271 if (__rcu_reclaim(rdp->rsp->name, list))
2277 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2278 smp_mb__before_atomic(); /* _add after CB invocation. */
2279 atomic_long_add(-c, &rdp->nocb_q_count);
2280 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2281 rdp->n_nocbs_invoked += c;
2286 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2287 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2289 return READ_ONCE(rdp->nocb_defer_wakeup);
2292 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2293 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2297 if (!rcu_nocb_need_deferred_wakeup(rdp))
2299 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2300 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2301 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2302 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2305 void __init rcu_init_nohz(void)
2308 bool need_rcu_nocb_mask = true;
2309 struct rcu_state *rsp;
2311 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2312 need_rcu_nocb_mask = false;
2313 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2315 #if defined(CONFIG_NO_HZ_FULL)
2316 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2317 need_rcu_nocb_mask = true;
2318 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2320 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2321 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2322 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2325 have_rcu_nocb_mask = true;
2327 if (!have_rcu_nocb_mask)
2330 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2331 pr_info("\tOffload RCU callbacks from CPU 0\n");
2332 cpumask_set_cpu(0, rcu_nocb_mask);
2333 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2334 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2335 pr_info("\tOffload RCU callbacks from all CPUs\n");
2336 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2337 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2338 #if defined(CONFIG_NO_HZ_FULL)
2339 if (tick_nohz_full_running)
2340 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2341 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2343 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2344 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2345 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2348 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2349 cpumask_pr_args(rcu_nocb_mask));
2351 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2353 for_each_rcu_flavor(rsp) {
2354 for_each_cpu(cpu, rcu_nocb_mask)
2355 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2356 rcu_organize_nocb_kthreads(rsp);
2360 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2361 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2363 rdp->nocb_tail = &rdp->nocb_head;
2364 init_waitqueue_head(&rdp->nocb_wq);
2365 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2369 * If the specified CPU is a no-CBs CPU that does not already have its
2370 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2371 * brought online out of order, this can require re-organizing the
2372 * leader-follower relationships.
2374 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2376 struct rcu_data *rdp;
2377 struct rcu_data *rdp_last;
2378 struct rcu_data *rdp_old_leader;
2379 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2380 struct task_struct *t;
2383 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2384 * then nothing to do.
2386 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2389 /* If we didn't spawn the leader first, reorganize! */
2390 rdp_old_leader = rdp_spawn->nocb_leader;
2391 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2393 rdp = rdp_old_leader;
2395 rdp->nocb_leader = rdp_spawn;
2396 if (rdp_last && rdp != rdp_spawn)
2397 rdp_last->nocb_next_follower = rdp;
2398 if (rdp == rdp_spawn) {
2399 rdp = rdp->nocb_next_follower;
2402 rdp = rdp->nocb_next_follower;
2403 rdp_last->nocb_next_follower = NULL;
2406 rdp_spawn->nocb_next_follower = rdp_old_leader;
2409 /* Spawn the kthread for this CPU and RCU flavor. */
2410 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2411 "rcuo%c/%d", rsp->abbr, cpu);
2413 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2417 * If the specified CPU is a no-CBs CPU that does not already have its
2418 * rcuo kthreads, spawn them.
2420 static void rcu_spawn_all_nocb_kthreads(int cpu)
2422 struct rcu_state *rsp;
2424 if (rcu_scheduler_fully_active)
2425 for_each_rcu_flavor(rsp)
2426 rcu_spawn_one_nocb_kthread(rsp, cpu);
2430 * Once the scheduler is running, spawn rcuo kthreads for all online
2431 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2432 * non-boot CPUs come online -- if this changes, we will need to add
2433 * some mutual exclusion.
2435 static void __init rcu_spawn_nocb_kthreads(void)
2439 for_each_online_cpu(cpu)
2440 rcu_spawn_all_nocb_kthreads(cpu);
2443 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2444 static int rcu_nocb_leader_stride = -1;
2445 module_param(rcu_nocb_leader_stride, int, 0444);
2448 * Initialize leader-follower relationships for all no-CBs CPU.
2450 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2453 int ls = rcu_nocb_leader_stride;
2454 int nl = 0; /* Next leader. */
2455 struct rcu_data *rdp;
2456 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2457 struct rcu_data *rdp_prev = NULL;
2459 if (!have_rcu_nocb_mask)
2462 ls = int_sqrt(nr_cpu_ids);
2463 rcu_nocb_leader_stride = ls;
2467 * Each pass through this loop sets up one rcu_data structure and
2468 * spawns one rcu_nocb_kthread().
2470 for_each_cpu(cpu, rcu_nocb_mask) {
2471 rdp = per_cpu_ptr(rsp->rda, cpu);
2472 if (rdp->cpu >= nl) {
2473 /* New leader, set up for followers & next leader. */
2474 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2475 rdp->nocb_leader = rdp;
2478 /* Another follower, link to previous leader. */
2479 rdp->nocb_leader = rdp_leader;
2480 rdp_prev->nocb_next_follower = rdp;
2486 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2487 static bool init_nocb_callback_list(struct rcu_data *rdp)
2489 if (!rcu_is_nocb_cpu(rdp->cpu))
2492 /* If there are early-boot callbacks, move them to nocb lists. */
2494 rdp->nocb_head = rdp->nxtlist;
2495 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2496 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2497 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2498 rdp->nxtlist = NULL;
2502 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2506 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2508 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2510 WARN_ON_ONCE(1); /* Should be dead code. */
2514 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2518 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2522 static void rcu_init_one_nocb(struct rcu_node *rnp)
2526 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2527 bool lazy, unsigned long flags)
2532 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2533 struct rcu_data *rdp,
2534 unsigned long flags)
2539 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2543 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2548 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2552 static void rcu_spawn_all_nocb_kthreads(int cpu)
2556 static void __init rcu_spawn_nocb_kthreads(void)
2560 static bool init_nocb_callback_list(struct rcu_data *rdp)
2565 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2568 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2569 * arbitrarily long period of time with the scheduling-clock tick turned
2570 * off. RCU will be paying attention to this CPU because it is in the
2571 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2572 * machine because the scheduling-clock tick has been disabled. Therefore,
2573 * if an adaptive-ticks CPU is failing to respond to the current grace
2574 * period and has not be idle from an RCU perspective, kick it.
2576 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2578 #ifdef CONFIG_NO_HZ_FULL
2579 if (tick_nohz_full_cpu(cpu))
2580 smp_send_reschedule(cpu);
2581 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2585 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2587 static int full_sysidle_state; /* Current system-idle state. */
2588 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2589 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2590 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2591 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2592 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2595 * Invoked to note exit from irq or task transition to idle. Note that
2596 * usermode execution does -not- count as idle here! After all, we want
2597 * to detect full-system idle states, not RCU quiescent states and grace
2598 * periods. The caller must have disabled interrupts.
2600 static void rcu_sysidle_enter(int irq)
2603 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2605 /* If there are no nohz_full= CPUs, no need to track this. */
2606 if (!tick_nohz_full_enabled())
2609 /* Adjust nesting, check for fully idle. */
2611 rdtp->dynticks_idle_nesting--;
2612 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2613 if (rdtp->dynticks_idle_nesting != 0)
2614 return; /* Still not fully idle. */
2616 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2617 DYNTICK_TASK_NEST_VALUE) {
2618 rdtp->dynticks_idle_nesting = 0;
2620 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2621 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2622 return; /* Still not fully idle. */
2626 /* Record start of fully idle period. */
2628 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2629 smp_mb__before_atomic();
2630 atomic_inc(&rdtp->dynticks_idle);
2631 smp_mb__after_atomic();
2632 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2636 * Unconditionally force exit from full system-idle state. This is
2637 * invoked when a normal CPU exits idle, but must be called separately
2638 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2639 * is that the timekeeping CPU is permitted to take scheduling-clock
2640 * interrupts while the system is in system-idle state, and of course
2641 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2642 * interrupt from any other type of interrupt.
2644 void rcu_sysidle_force_exit(void)
2646 int oldstate = READ_ONCE(full_sysidle_state);
2650 * Each pass through the following loop attempts to exit full
2651 * system-idle state. If contention proves to be a problem,
2652 * a trylock-based contention tree could be used here.
2654 while (oldstate > RCU_SYSIDLE_SHORT) {
2655 newoldstate = cmpxchg(&full_sysidle_state,
2656 oldstate, RCU_SYSIDLE_NOT);
2657 if (oldstate == newoldstate &&
2658 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2659 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2660 return; /* We cleared it, done! */
2662 oldstate = newoldstate;
2664 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2668 * Invoked to note entry to irq or task transition from idle. Note that
2669 * usermode execution does -not- count as idle here! The caller must
2670 * have disabled interrupts.
2672 static void rcu_sysidle_exit(int irq)
2674 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2676 /* If there are no nohz_full= CPUs, no need to track this. */
2677 if (!tick_nohz_full_enabled())
2680 /* Adjust nesting, check for already non-idle. */
2682 rdtp->dynticks_idle_nesting++;
2683 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2684 if (rdtp->dynticks_idle_nesting != 1)
2685 return; /* Already non-idle. */
2688 * Allow for irq misnesting. Yes, it really is possible
2689 * to enter an irq handler then never leave it, and maybe
2690 * also vice versa. Handle both possibilities.
2692 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2693 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2694 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2695 return; /* Already non-idle. */
2697 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2701 /* Record end of idle period. */
2702 smp_mb__before_atomic();
2703 atomic_inc(&rdtp->dynticks_idle);
2704 smp_mb__after_atomic();
2705 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2708 * If we are the timekeeping CPU, we are permitted to be non-idle
2709 * during a system-idle state. This must be the case, because
2710 * the timekeeping CPU has to take scheduling-clock interrupts
2711 * during the time that the system is transitioning to full
2712 * system-idle state. This means that the timekeeping CPU must
2713 * invoke rcu_sysidle_force_exit() directly if it does anything
2714 * more than take a scheduling-clock interrupt.
2716 if (smp_processor_id() == tick_do_timer_cpu)
2719 /* Update system-idle state: We are clearly no longer fully idle! */
2720 rcu_sysidle_force_exit();
2724 * Check to see if the current CPU is idle. Note that usermode execution
2725 * does not count as idle. The caller must have disabled interrupts,
2726 * and must be running on tick_do_timer_cpu.
2728 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2729 unsigned long *maxj)
2733 struct rcu_dynticks *rdtp = rdp->dynticks;
2735 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2736 if (!tick_nohz_full_enabled())
2740 * If some other CPU has already reported non-idle, if this is
2741 * not the flavor of RCU that tracks sysidle state, or if this
2742 * is an offline or the timekeeping CPU, nothing to do.
2744 if (!*isidle || rdp->rsp != rcu_state_p ||
2745 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2747 /* Verify affinity of current kthread. */
2748 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2750 /* Pick up current idle and NMI-nesting counter and check. */
2751 cur = atomic_read(&rdtp->dynticks_idle);
2753 *isidle = false; /* We are not idle! */
2756 smp_mb(); /* Read counters before timestamps. */
2758 /* Pick up timestamps. */
2759 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2760 /* If this CPU entered idle more recently, update maxj timestamp. */
2761 if (ULONG_CMP_LT(*maxj, j))
2766 * Is this the flavor of RCU that is handling full-system idle?
2768 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2770 return rsp == rcu_state_p;
2774 * Return a delay in jiffies based on the number of CPUs, rcu_node
2775 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2776 * systems more time to transition to full-idle state in order to
2777 * avoid the cache thrashing that otherwise occur on the state variable.
2778 * Really small systems (less than a couple of tens of CPUs) should
2779 * instead use a single global atomically incremented counter, and later
2780 * versions of this will automatically reconfigure themselves accordingly.
2782 static unsigned long rcu_sysidle_delay(void)
2784 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2786 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2790 * Advance the full-system-idle state. This is invoked when all of
2791 * the non-timekeeping CPUs are idle.
2793 static void rcu_sysidle(unsigned long j)
2795 /* Check the current state. */
2796 switch (READ_ONCE(full_sysidle_state)) {
2797 case RCU_SYSIDLE_NOT:
2799 /* First time all are idle, so note a short idle period. */
2800 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2803 case RCU_SYSIDLE_SHORT:
2806 * Idle for a bit, time to advance to next state?
2807 * cmpxchg failure means race with non-idle, let them win.
2809 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2810 (void)cmpxchg(&full_sysidle_state,
2811 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2814 case RCU_SYSIDLE_LONG:
2817 * Do an additional check pass before advancing to full.
2818 * cmpxchg failure means race with non-idle, let them win.
2820 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2821 (void)cmpxchg(&full_sysidle_state,
2822 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2831 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2832 * back to the beginning.
2834 static void rcu_sysidle_cancel(void)
2837 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2838 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2842 * Update the sysidle state based on the results of a force-quiescent-state
2843 * scan of the CPUs' dyntick-idle state.
2845 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2846 unsigned long maxj, bool gpkt)
2848 if (rsp != rcu_state_p)
2849 return; /* Wrong flavor, ignore. */
2850 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2851 return; /* Running state machine from timekeeping CPU. */
2853 rcu_sysidle(maxj); /* More idle! */
2855 rcu_sysidle_cancel(); /* Idle is over. */
2859 * Wrapper for rcu_sysidle_report() when called from the grace-period
2860 * kthread's context.
2862 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2865 /* If there are no nohz_full= CPUs, no need to track this. */
2866 if (!tick_nohz_full_enabled())
2869 rcu_sysidle_report(rsp, isidle, maxj, true);
2872 /* Callback and function for forcing an RCU grace period. */
2873 struct rcu_sysidle_head {
2878 static void rcu_sysidle_cb(struct rcu_head *rhp)
2880 struct rcu_sysidle_head *rshp;
2883 * The following memory barrier is needed to replace the
2884 * memory barriers that would normally be in the memory
2887 smp_mb(); /* grace period precedes setting inuse. */
2889 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2890 WRITE_ONCE(rshp->inuse, 0);
2894 * Check to see if the system is fully idle, other than the timekeeping CPU.
2895 * The caller must have disabled interrupts. This is not intended to be
2896 * called unless tick_nohz_full_enabled().
2898 bool rcu_sys_is_idle(void)
2900 static struct rcu_sysidle_head rsh;
2901 int rss = READ_ONCE(full_sysidle_state);
2903 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2906 /* Handle small-system case by doing a full scan of CPUs. */
2907 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2908 int oldrss = rss - 1;
2911 * One pass to advance to each state up to _FULL.
2912 * Give up if any pass fails to advance the state.
2914 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2917 unsigned long maxj = jiffies - ULONG_MAX / 4;
2918 struct rcu_data *rdp;
2920 /* Scan all the CPUs looking for nonidle CPUs. */
2921 for_each_possible_cpu(cpu) {
2922 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2923 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2927 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2929 rss = READ_ONCE(full_sysidle_state);
2933 /* If this is the first observation of an idle period, record it. */
2934 if (rss == RCU_SYSIDLE_FULL) {
2935 rss = cmpxchg(&full_sysidle_state,
2936 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2937 return rss == RCU_SYSIDLE_FULL;
2940 smp_mb(); /* ensure rss load happens before later caller actions. */
2942 /* If already fully idle, tell the caller (in case of races). */
2943 if (rss == RCU_SYSIDLE_FULL_NOTED)
2947 * If we aren't there yet, and a grace period is not in flight,
2948 * initiate a grace period. Either way, tell the caller that
2949 * we are not there yet. We use an xchg() rather than an assignment
2950 * to make up for the memory barriers that would otherwise be
2951 * provided by the memory allocator.
2953 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2954 !rcu_gp_in_progress(rcu_state_p) &&
2955 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2956 call_rcu(&rsh.rh, rcu_sysidle_cb);
2961 * Initialize dynticks sysidle state for CPUs coming online.
2963 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2965 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2968 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2970 static void rcu_sysidle_enter(int irq)
2974 static void rcu_sysidle_exit(int irq)
2978 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2979 unsigned long *maxj)
2983 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2988 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2993 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2997 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3000 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3001 * grace-period kthread will do force_quiescent_state() processing?
3002 * The idea is to avoid waking up RCU core processing on such a
3003 * CPU unless the grace period has extended for too long.
3005 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3006 * CONFIG_RCU_NOCB_CPU CPUs.
3008 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3010 #ifdef CONFIG_NO_HZ_FULL
3011 if (tick_nohz_full_cpu(smp_processor_id()) &&
3012 (!rcu_gp_in_progress(rsp) ||
3013 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
3015 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3020 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3023 static void rcu_bind_gp_kthread(void)
3025 int __maybe_unused cpu;
3027 if (!tick_nohz_full_enabled())
3029 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3030 cpu = tick_do_timer_cpu;
3031 if (cpu >= 0 && cpu < nr_cpu_ids)
3032 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3033 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3034 housekeeping_affine(current);
3035 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3038 /* Record the current task on dyntick-idle entry. */
3039 static void rcu_dynticks_task_enter(void)
3041 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3042 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
3043 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3046 /* Record no current task on dyntick-idle exit. */
3047 static void rcu_dynticks_task_exit(void)
3049 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3050 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
3051 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */