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, write to the Free Software
18 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
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 #define RCU_KTHREAD_PRIO 1
35 #ifdef CONFIG_RCU_BOOST
36 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
38 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
41 #ifdef CONFIG_RCU_NOCB_CPU
42 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
43 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
44 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
45 static char __initdata nocb_buf[NR_CPUS * 5];
46 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
49 * Check the RCU kernel configuration parameters and print informative
50 * messages about anything out of the ordinary. If you like #ifdef, you
51 * will love this function.
53 static void __init rcu_bootup_announce_oddness(void)
55 #ifdef CONFIG_RCU_TRACE
56 pr_info("\tRCU debugfs-based tracing is enabled.\n");
58 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
59 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
62 #ifdef CONFIG_RCU_FANOUT_EXACT
63 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
65 #ifdef CONFIG_RCU_FAST_NO_HZ
66 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
68 #ifdef CONFIG_PROVE_RCU
69 pr_info("\tRCU lockdep checking is enabled.\n");
71 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
72 pr_info("\tRCU torture testing starts during boot.\n");
74 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
75 pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
77 #if defined(CONFIG_RCU_CPU_STALL_INFO)
78 pr_info("\tAdditional per-CPU info printed with stalls.\n");
80 #if NUM_RCU_LVL_4 != 0
81 pr_info("\tFour-level hierarchy is enabled.\n");
83 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
84 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
85 if (nr_cpu_ids != NR_CPUS)
86 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
87 #ifdef CONFIG_RCU_NOCB_CPU
88 #ifndef CONFIG_RCU_NOCB_CPU_NONE
89 if (!have_rcu_nocb_mask) {
90 zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL);
91 have_rcu_nocb_mask = true;
93 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
94 pr_info("\tOffload RCU callbacks from CPU 0\n");
95 cpumask_set_cpu(0, rcu_nocb_mask);
96 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
97 #ifdef CONFIG_RCU_NOCB_CPU_ALL
98 pr_info("\tOffload RCU callbacks from all CPUs\n");
99 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
100 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
101 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
102 if (have_rcu_nocb_mask) {
103 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
104 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
105 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
108 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
109 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
111 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
113 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
116 #ifdef CONFIG_TREE_PREEMPT_RCU
118 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
119 static struct rcu_state *rcu_state = &rcu_preempt_state;
121 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
124 * Tell them what RCU they are running.
126 static void __init rcu_bootup_announce(void)
128 pr_info("Preemptible hierarchical RCU implementation.\n");
129 rcu_bootup_announce_oddness();
133 * Return the number of RCU-preempt batches processed thus far
134 * for debug and statistics.
136 long rcu_batches_completed_preempt(void)
138 return rcu_preempt_state.completed;
140 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
143 * Return the number of RCU batches processed thus far for debug & stats.
145 long rcu_batches_completed(void)
147 return rcu_batches_completed_preempt();
149 EXPORT_SYMBOL_GPL(rcu_batches_completed);
152 * Force a quiescent state for preemptible RCU.
154 void rcu_force_quiescent_state(void)
156 force_quiescent_state(&rcu_preempt_state);
158 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
161 * Record a preemptible-RCU quiescent state for the specified CPU. Note
162 * that this just means that the task currently running on the CPU is
163 * not in a quiescent state. There might be any number of tasks blocked
164 * while in an RCU read-side critical section.
166 * Unlike the other rcu_*_qs() functions, callers to this function
167 * must disable irqs in order to protect the assignment to
168 * ->rcu_read_unlock_special.
170 static void rcu_preempt_qs(int cpu)
172 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
174 if (rdp->passed_quiesce == 0)
175 trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs"));
176 rdp->passed_quiesce = 1;
177 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
181 * We have entered the scheduler, and the current task might soon be
182 * context-switched away from. If this task is in an RCU read-side
183 * critical section, we will no longer be able to rely on the CPU to
184 * record that fact, so we enqueue the task on the blkd_tasks list.
185 * The task will dequeue itself when it exits the outermost enclosing
186 * RCU read-side critical section. Therefore, the current grace period
187 * cannot be permitted to complete until the blkd_tasks list entries
188 * predating the current grace period drain, in other words, until
189 * rnp->gp_tasks becomes NULL.
191 * Caller must disable preemption.
193 static void rcu_preempt_note_context_switch(int cpu)
195 struct task_struct *t = current;
197 struct rcu_data *rdp;
198 struct rcu_node *rnp;
200 if (t->rcu_read_lock_nesting > 0 &&
201 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
203 /* Possibly blocking in an RCU read-side critical section. */
204 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
206 raw_spin_lock_irqsave(&rnp->lock, flags);
207 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
208 t->rcu_blocked_node = rnp;
211 * If this CPU has already checked in, then this task
212 * will hold up the next grace period rather than the
213 * current grace period. Queue the task accordingly.
214 * If the task is queued for the current grace period
215 * (i.e., this CPU has not yet passed through a quiescent
216 * state for the current grace period), then as long
217 * as that task remains queued, the current grace period
218 * cannot end. Note that there is some uncertainty as
219 * to exactly when the current grace period started.
220 * We take a conservative approach, which can result
221 * in unnecessarily waiting on tasks that started very
222 * slightly after the current grace period began. C'est
225 * But first, note that the current CPU must still be
228 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
229 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
230 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
231 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
232 rnp->gp_tasks = &t->rcu_node_entry;
233 #ifdef CONFIG_RCU_BOOST
234 if (rnp->boost_tasks != NULL)
235 rnp->boost_tasks = rnp->gp_tasks;
236 #endif /* #ifdef CONFIG_RCU_BOOST */
238 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
239 if (rnp->qsmask & rdp->grpmask)
240 rnp->gp_tasks = &t->rcu_node_entry;
242 trace_rcu_preempt_task(rdp->rsp->name,
244 (rnp->qsmask & rdp->grpmask)
247 raw_spin_unlock_irqrestore(&rnp->lock, flags);
248 } else if (t->rcu_read_lock_nesting < 0 &&
249 t->rcu_read_unlock_special) {
252 * Complete exit from RCU read-side critical section on
253 * behalf of preempted instance of __rcu_read_unlock().
255 rcu_read_unlock_special(t);
259 * Either we were not in an RCU read-side critical section to
260 * begin with, or we have now recorded that critical section
261 * globally. Either way, we can now note a quiescent state
262 * for this CPU. Again, if we were in an RCU read-side critical
263 * section, and if that critical section was blocking the current
264 * grace period, then the fact that the task has been enqueued
265 * means that we continue to block the current grace period.
267 local_irq_save(flags);
269 local_irq_restore(flags);
273 * Check for preempted RCU readers blocking the current grace period
274 * for the specified rcu_node structure. If the caller needs a reliable
275 * answer, it must hold the rcu_node's ->lock.
277 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
279 return rnp->gp_tasks != NULL;
283 * Record a quiescent state for all tasks that were previously queued
284 * on the specified rcu_node structure and that were blocking the current
285 * RCU grace period. The caller must hold the specified rnp->lock with
286 * irqs disabled, and this lock is released upon return, but irqs remain
289 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
290 __releases(rnp->lock)
293 struct rcu_node *rnp_p;
295 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
296 raw_spin_unlock_irqrestore(&rnp->lock, flags);
297 return; /* Still need more quiescent states! */
303 * Either there is only one rcu_node in the tree,
304 * or tasks were kicked up to root rcu_node due to
305 * CPUs going offline.
307 rcu_report_qs_rsp(&rcu_preempt_state, flags);
311 /* Report up the rest of the hierarchy. */
313 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
314 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
315 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
319 * Advance a ->blkd_tasks-list pointer to the next entry, instead
320 * returning NULL if at the end of the list.
322 static struct list_head *rcu_next_node_entry(struct task_struct *t,
323 struct rcu_node *rnp)
325 struct list_head *np;
327 np = t->rcu_node_entry.next;
328 if (np == &rnp->blkd_tasks)
334 * Handle special cases during rcu_read_unlock(), such as needing to
335 * notify RCU core processing or task having blocked during the RCU
336 * read-side critical section.
338 void rcu_read_unlock_special(struct task_struct *t)
344 struct list_head *np;
345 #ifdef CONFIG_RCU_BOOST
346 struct rt_mutex *rbmp = NULL;
347 #endif /* #ifdef CONFIG_RCU_BOOST */
348 struct rcu_node *rnp;
351 /* NMI handlers cannot block and cannot safely manipulate state. */
355 local_irq_save(flags);
358 * If RCU core is waiting for this CPU to exit critical section,
359 * let it know that we have done so.
361 special = t->rcu_read_unlock_special;
362 if (special & RCU_READ_UNLOCK_NEED_QS) {
363 rcu_preempt_qs(smp_processor_id());
364 if (!t->rcu_read_unlock_special) {
365 local_irq_restore(flags);
370 /* Hardware IRQ handlers cannot block, complain if they get here. */
371 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
372 local_irq_restore(flags);
376 /* Clean up if blocked during RCU read-side critical section. */
377 if (special & RCU_READ_UNLOCK_BLOCKED) {
378 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
381 * Remove this task from the list it blocked on. The
382 * task can migrate while we acquire the lock, but at
383 * most one time. So at most two passes through loop.
386 rnp = t->rcu_blocked_node;
387 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
388 if (rnp == t->rcu_blocked_node)
390 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
392 empty = !rcu_preempt_blocked_readers_cgp(rnp);
393 empty_exp = !rcu_preempted_readers_exp(rnp);
394 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
395 np = rcu_next_node_entry(t, rnp);
396 list_del_init(&t->rcu_node_entry);
397 t->rcu_blocked_node = NULL;
398 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
400 if (&t->rcu_node_entry == rnp->gp_tasks)
402 if (&t->rcu_node_entry == rnp->exp_tasks)
404 #ifdef CONFIG_RCU_BOOST
405 if (&t->rcu_node_entry == rnp->boost_tasks)
406 rnp->boost_tasks = np;
407 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
408 if (t->rcu_boost_mutex) {
409 rbmp = t->rcu_boost_mutex;
410 t->rcu_boost_mutex = NULL;
412 #endif /* #ifdef CONFIG_RCU_BOOST */
415 * If this was the last task on the current list, and if
416 * we aren't waiting on any CPUs, report the quiescent state.
417 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
418 * so we must take a snapshot of the expedited state.
420 empty_exp_now = !rcu_preempted_readers_exp(rnp);
421 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
422 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
429 rcu_report_unblock_qs_rnp(rnp, flags);
431 raw_spin_unlock_irqrestore(&rnp->lock, flags);
434 #ifdef CONFIG_RCU_BOOST
435 /* Unboost if we were boosted. */
437 rt_mutex_unlock(rbmp);
438 #endif /* #ifdef CONFIG_RCU_BOOST */
441 * If this was the last task on the expedited lists,
442 * then we need to report up the rcu_node hierarchy.
444 if (!empty_exp && empty_exp_now)
445 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
447 local_irq_restore(flags);
451 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
454 * Dump detailed information for all tasks blocking the current RCU
455 * grace period on the specified rcu_node structure.
457 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
460 struct task_struct *t;
462 raw_spin_lock_irqsave(&rnp->lock, flags);
463 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
464 raw_spin_unlock_irqrestore(&rnp->lock, flags);
467 t = list_entry(rnp->gp_tasks,
468 struct task_struct, rcu_node_entry);
469 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
471 raw_spin_unlock_irqrestore(&rnp->lock, flags);
475 * Dump detailed information for all tasks blocking the current RCU
478 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
480 struct rcu_node *rnp = rcu_get_root(rsp);
482 rcu_print_detail_task_stall_rnp(rnp);
483 rcu_for_each_leaf_node(rsp, rnp)
484 rcu_print_detail_task_stall_rnp(rnp);
487 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
489 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
493 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
495 #ifdef CONFIG_RCU_CPU_STALL_INFO
497 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
499 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
500 rnp->level, rnp->grplo, rnp->grphi);
503 static void rcu_print_task_stall_end(void)
508 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
510 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
514 static void rcu_print_task_stall_end(void)
518 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
521 * Scan the current list of tasks blocked within RCU read-side critical
522 * sections, printing out the tid of each.
524 static int rcu_print_task_stall(struct rcu_node *rnp)
526 struct task_struct *t;
529 if (!rcu_preempt_blocked_readers_cgp(rnp))
531 rcu_print_task_stall_begin(rnp);
532 t = list_entry(rnp->gp_tasks,
533 struct task_struct, rcu_node_entry);
534 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
535 pr_cont(" P%d", t->pid);
538 rcu_print_task_stall_end();
543 * Check that the list of blocked tasks for the newly completed grace
544 * period is in fact empty. It is a serious bug to complete a grace
545 * period that still has RCU readers blocked! This function must be
546 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
547 * must be held by the caller.
549 * Also, if there are blocked tasks on the list, they automatically
550 * block the newly created grace period, so set up ->gp_tasks accordingly.
552 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
554 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
555 if (!list_empty(&rnp->blkd_tasks))
556 rnp->gp_tasks = rnp->blkd_tasks.next;
557 WARN_ON_ONCE(rnp->qsmask);
560 #ifdef CONFIG_HOTPLUG_CPU
563 * Handle tasklist migration for case in which all CPUs covered by the
564 * specified rcu_node have gone offline. Move them up to the root
565 * rcu_node. The reason for not just moving them to the immediate
566 * parent is to remove the need for rcu_read_unlock_special() to
567 * make more than two attempts to acquire the target rcu_node's lock.
568 * Returns true if there were tasks blocking the current RCU grace
571 * Returns 1 if there was previously a task blocking the current grace
572 * period on the specified rcu_node structure.
574 * The caller must hold rnp->lock with irqs disabled.
576 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
577 struct rcu_node *rnp,
578 struct rcu_data *rdp)
580 struct list_head *lp;
581 struct list_head *lp_root;
583 struct rcu_node *rnp_root = rcu_get_root(rsp);
584 struct task_struct *t;
586 if (rnp == rnp_root) {
587 WARN_ONCE(1, "Last CPU thought to be offlined?");
588 return 0; /* Shouldn't happen: at least one CPU online. */
591 /* If we are on an internal node, complain bitterly. */
592 WARN_ON_ONCE(rnp != rdp->mynode);
595 * Move tasks up to root rcu_node. Don't try to get fancy for
596 * this corner-case operation -- just put this node's tasks
597 * at the head of the root node's list, and update the root node's
598 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
599 * if non-NULL. This might result in waiting for more tasks than
600 * absolutely necessary, but this is a good performance/complexity
603 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
604 retval |= RCU_OFL_TASKS_NORM_GP;
605 if (rcu_preempted_readers_exp(rnp))
606 retval |= RCU_OFL_TASKS_EXP_GP;
607 lp = &rnp->blkd_tasks;
608 lp_root = &rnp_root->blkd_tasks;
609 while (!list_empty(lp)) {
610 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
611 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
612 list_del(&t->rcu_node_entry);
613 t->rcu_blocked_node = rnp_root;
614 list_add(&t->rcu_node_entry, lp_root);
615 if (&t->rcu_node_entry == rnp->gp_tasks)
616 rnp_root->gp_tasks = rnp->gp_tasks;
617 if (&t->rcu_node_entry == rnp->exp_tasks)
618 rnp_root->exp_tasks = rnp->exp_tasks;
619 #ifdef CONFIG_RCU_BOOST
620 if (&t->rcu_node_entry == rnp->boost_tasks)
621 rnp_root->boost_tasks = rnp->boost_tasks;
622 #endif /* #ifdef CONFIG_RCU_BOOST */
623 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
626 rnp->gp_tasks = NULL;
627 rnp->exp_tasks = NULL;
628 #ifdef CONFIG_RCU_BOOST
629 rnp->boost_tasks = NULL;
631 * In case root is being boosted and leaf was not. Make sure
632 * that we boost the tasks blocking the current grace period
635 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
636 if (rnp_root->boost_tasks != NULL &&
637 rnp_root->boost_tasks != rnp_root->gp_tasks &&
638 rnp_root->boost_tasks != rnp_root->exp_tasks)
639 rnp_root->boost_tasks = rnp_root->gp_tasks;
640 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
641 #endif /* #ifdef CONFIG_RCU_BOOST */
646 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
649 * Check for a quiescent state from the current CPU. When a task blocks,
650 * the task is recorded in the corresponding CPU's rcu_node structure,
651 * which is checked elsewhere.
653 * Caller must disable hard irqs.
655 static void rcu_preempt_check_callbacks(int cpu)
657 struct task_struct *t = current;
659 if (t->rcu_read_lock_nesting == 0) {
663 if (t->rcu_read_lock_nesting > 0 &&
664 per_cpu(rcu_preempt_data, cpu).qs_pending)
665 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
668 #ifdef CONFIG_RCU_BOOST
670 static void rcu_preempt_do_callbacks(void)
672 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
675 #endif /* #ifdef CONFIG_RCU_BOOST */
678 * Queue a preemptible-RCU callback for invocation after a grace period.
680 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
682 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
684 EXPORT_SYMBOL_GPL(call_rcu);
687 * Queue an RCU callback for lazy invocation after a grace period.
688 * This will likely be later named something like "call_rcu_lazy()",
689 * but this change will require some way of tagging the lazy RCU
690 * callbacks in the list of pending callbacks. Until then, this
691 * function may only be called from __kfree_rcu().
693 void kfree_call_rcu(struct rcu_head *head,
694 void (*func)(struct rcu_head *rcu))
696 __call_rcu(head, func, &rcu_preempt_state, -1, 1);
698 EXPORT_SYMBOL_GPL(kfree_call_rcu);
701 * synchronize_rcu - wait until a grace period has elapsed.
703 * Control will return to the caller some time after a full grace
704 * period has elapsed, in other words after all currently executing RCU
705 * read-side critical sections have completed. Note, however, that
706 * upon return from synchronize_rcu(), the caller might well be executing
707 * concurrently with new RCU read-side critical sections that began while
708 * synchronize_rcu() was waiting. RCU read-side critical sections are
709 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
711 * See the description of synchronize_sched() for more detailed information
712 * on memory ordering guarantees.
714 void synchronize_rcu(void)
716 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
717 !lock_is_held(&rcu_lock_map) &&
718 !lock_is_held(&rcu_sched_lock_map),
719 "Illegal synchronize_rcu() in RCU read-side critical section");
720 if (!rcu_scheduler_active)
723 synchronize_rcu_expedited();
725 wait_rcu_gp(call_rcu);
727 EXPORT_SYMBOL_GPL(synchronize_rcu);
729 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
730 static unsigned long sync_rcu_preempt_exp_count;
731 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
734 * Return non-zero if there are any tasks in RCU read-side critical
735 * sections blocking the current preemptible-RCU expedited grace period.
736 * If there is no preemptible-RCU expedited grace period currently in
737 * progress, returns zero unconditionally.
739 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
741 return rnp->exp_tasks != NULL;
745 * return non-zero if there is no RCU expedited grace period in progress
746 * for the specified rcu_node structure, in other words, if all CPUs and
747 * tasks covered by the specified rcu_node structure have done their bit
748 * for the current expedited grace period. Works only for preemptible
749 * RCU -- other RCU implementation use other means.
751 * Caller must hold sync_rcu_preempt_exp_mutex.
753 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
755 return !rcu_preempted_readers_exp(rnp) &&
756 ACCESS_ONCE(rnp->expmask) == 0;
760 * Report the exit from RCU read-side critical section for the last task
761 * that queued itself during or before the current expedited preemptible-RCU
762 * grace period. This event is reported either to the rcu_node structure on
763 * which the task was queued or to one of that rcu_node structure's ancestors,
764 * recursively up the tree. (Calm down, calm down, we do the recursion
767 * Most callers will set the "wake" flag, but the task initiating the
768 * expedited grace period need not wake itself.
770 * Caller must hold sync_rcu_preempt_exp_mutex.
772 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
778 raw_spin_lock_irqsave(&rnp->lock, flags);
780 if (!sync_rcu_preempt_exp_done(rnp)) {
781 raw_spin_unlock_irqrestore(&rnp->lock, flags);
784 if (rnp->parent == NULL) {
785 raw_spin_unlock_irqrestore(&rnp->lock, flags);
787 smp_mb(); /* EGP done before wake_up(). */
788 wake_up(&sync_rcu_preempt_exp_wq);
793 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
795 raw_spin_lock(&rnp->lock); /* irqs already disabled */
796 rnp->expmask &= ~mask;
801 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
802 * grace period for the specified rcu_node structure. If there are no such
803 * tasks, report it up the rcu_node hierarchy.
805 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
806 * CPU hotplug operations.
809 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
814 raw_spin_lock_irqsave(&rnp->lock, flags);
815 if (list_empty(&rnp->blkd_tasks)) {
816 raw_spin_unlock_irqrestore(&rnp->lock, flags);
818 rnp->exp_tasks = rnp->blkd_tasks.next;
819 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
823 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
827 * synchronize_rcu_expedited - Brute-force RCU grace period
829 * Wait for an RCU-preempt grace period, but expedite it. The basic
830 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
831 * the ->blkd_tasks lists and wait for this list to drain. This consumes
832 * significant time on all CPUs and is unfriendly to real-time workloads,
833 * so is thus not recommended for any sort of common-case code.
834 * In fact, if you are using synchronize_rcu_expedited() in a loop,
835 * please restructure your code to batch your updates, and then Use a
836 * single synchronize_rcu() instead.
838 * Note that it is illegal to call this function while holding any lock
839 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
840 * to call this function from a CPU-hotplug notifier. Failing to observe
841 * these restriction will result in deadlock.
843 void synchronize_rcu_expedited(void)
846 struct rcu_node *rnp;
847 struct rcu_state *rsp = &rcu_preempt_state;
851 smp_mb(); /* Caller's modifications seen first by other CPUs. */
852 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
853 smp_mb(); /* Above access cannot bleed into critical section. */
856 * Block CPU-hotplug operations. This means that any CPU-hotplug
857 * operation that finds an rcu_node structure with tasks in the
858 * process of being boosted will know that all tasks blocking
859 * this expedited grace period will already be in the process of
860 * being boosted. This simplifies the process of moving tasks
861 * from leaf to root rcu_node structures.
866 * Acquire lock, falling back to synchronize_rcu() if too many
867 * lock-acquisition failures. Of course, if someone does the
868 * expedited grace period for us, just leave.
870 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
871 if (ULONG_CMP_LT(snap,
872 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
874 goto mb_ret; /* Others did our work for us. */
876 if (trycount++ < 10) {
877 udelay(trycount * num_online_cpus());
880 wait_rcu_gp(call_rcu);
884 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
886 goto unlock_mb_ret; /* Others did our work for us. */
889 /* force all RCU readers onto ->blkd_tasks lists. */
890 synchronize_sched_expedited();
892 /* Initialize ->expmask for all non-leaf rcu_node structures. */
893 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
894 raw_spin_lock_irqsave(&rnp->lock, flags);
895 rnp->expmask = rnp->qsmaskinit;
896 raw_spin_unlock_irqrestore(&rnp->lock, flags);
899 /* Snapshot current state of ->blkd_tasks lists. */
900 rcu_for_each_leaf_node(rsp, rnp)
901 sync_rcu_preempt_exp_init(rsp, rnp);
902 if (NUM_RCU_NODES > 1)
903 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
907 /* Wait for snapshotted ->blkd_tasks lists to drain. */
908 rnp = rcu_get_root(rsp);
909 wait_event(sync_rcu_preempt_exp_wq,
910 sync_rcu_preempt_exp_done(rnp));
912 /* Clean up and exit. */
913 smp_mb(); /* ensure expedited GP seen before counter increment. */
914 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
916 mutex_unlock(&sync_rcu_preempt_exp_mutex);
918 smp_mb(); /* ensure subsequent action seen after grace period. */
920 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
923 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
925 * Note that this primitive does not necessarily wait for an RCU grace period
926 * to complete. For example, if there are no RCU callbacks queued anywhere
927 * in the system, then rcu_barrier() is within its rights to return
928 * immediately, without waiting for anything, much less an RCU grace period.
930 void rcu_barrier(void)
932 _rcu_barrier(&rcu_preempt_state);
934 EXPORT_SYMBOL_GPL(rcu_barrier);
937 * Initialize preemptible RCU's state structures.
939 static void __init __rcu_init_preempt(void)
941 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
945 * Check for a task exiting while in a preemptible-RCU read-side
946 * critical section, clean up if so. No need to issue warnings,
947 * as debug_check_no_locks_held() already does this if lockdep
952 struct task_struct *t = current;
954 if (likely(list_empty(¤t->rcu_node_entry)))
956 t->rcu_read_lock_nesting = 1;
958 t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED;
962 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
964 static struct rcu_state *rcu_state = &rcu_sched_state;
967 * Tell them what RCU they are running.
969 static void __init rcu_bootup_announce(void)
971 pr_info("Hierarchical RCU implementation.\n");
972 rcu_bootup_announce_oddness();
976 * Return the number of RCU batches processed thus far for debug & stats.
978 long rcu_batches_completed(void)
980 return rcu_batches_completed_sched();
982 EXPORT_SYMBOL_GPL(rcu_batches_completed);
985 * Force a quiescent state for RCU, which, because there is no preemptible
986 * RCU, becomes the same as rcu-sched.
988 void rcu_force_quiescent_state(void)
990 rcu_sched_force_quiescent_state();
992 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
995 * Because preemptible RCU does not exist, we never have to check for
996 * CPUs being in quiescent states.
998 static void rcu_preempt_note_context_switch(int cpu)
1003 * Because preemptible RCU does not exist, there are never any preempted
1006 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
1011 #ifdef CONFIG_HOTPLUG_CPU
1013 /* Because preemptible RCU does not exist, no quieting of tasks. */
1014 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1016 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1019 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1022 * Because preemptible RCU does not exist, we never have to check for
1023 * tasks blocked within RCU read-side critical sections.
1025 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1030 * Because preemptible RCU does not exist, we never have to check for
1031 * tasks blocked within RCU read-side critical sections.
1033 static int rcu_print_task_stall(struct rcu_node *rnp)
1039 * Because there is no preemptible RCU, there can be no readers blocked,
1040 * so there is no need to check for blocked tasks. So check only for
1041 * bogus qsmask values.
1043 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1045 WARN_ON_ONCE(rnp->qsmask);
1048 #ifdef CONFIG_HOTPLUG_CPU
1051 * Because preemptible RCU does not exist, it never needs to migrate
1052 * tasks that were blocked within RCU read-side critical sections, and
1053 * such non-existent tasks cannot possibly have been blocking the current
1056 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1057 struct rcu_node *rnp,
1058 struct rcu_data *rdp)
1063 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1066 * Because preemptible RCU does not exist, it never has any callbacks
1069 static void rcu_preempt_check_callbacks(int cpu)
1074 * Queue an RCU callback for lazy invocation after a grace period.
1075 * This will likely be later named something like "call_rcu_lazy()",
1076 * but this change will require some way of tagging the lazy RCU
1077 * callbacks in the list of pending callbacks. Until then, this
1078 * function may only be called from __kfree_rcu().
1080 * Because there is no preemptible RCU, we use RCU-sched instead.
1082 void kfree_call_rcu(struct rcu_head *head,
1083 void (*func)(struct rcu_head *rcu))
1085 __call_rcu(head, func, &rcu_sched_state, -1, 1);
1087 EXPORT_SYMBOL_GPL(kfree_call_rcu);
1090 * Wait for an rcu-preempt grace period, but make it happen quickly.
1091 * But because preemptible RCU does not exist, map to rcu-sched.
1093 void synchronize_rcu_expedited(void)
1095 synchronize_sched_expedited();
1097 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1099 #ifdef CONFIG_HOTPLUG_CPU
1102 * Because preemptible RCU does not exist, there is never any need to
1103 * report on tasks preempted in RCU read-side critical sections during
1104 * expedited RCU grace periods.
1106 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1111 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1114 * Because preemptible RCU does not exist, rcu_barrier() is just
1115 * another name for rcu_barrier_sched().
1117 void rcu_barrier(void)
1119 rcu_barrier_sched();
1121 EXPORT_SYMBOL_GPL(rcu_barrier);
1124 * Because preemptible RCU does not exist, it need not be initialized.
1126 static void __init __rcu_init_preempt(void)
1131 * Because preemptible RCU does not exist, tasks cannot possibly exit
1132 * while in preemptible RCU read-side critical sections.
1138 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1140 #ifdef CONFIG_RCU_BOOST
1142 #include "../locking/rtmutex_common.h"
1144 #ifdef CONFIG_RCU_TRACE
1146 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1148 if (list_empty(&rnp->blkd_tasks))
1149 rnp->n_balk_blkd_tasks++;
1150 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1151 rnp->n_balk_exp_gp_tasks++;
1152 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1153 rnp->n_balk_boost_tasks++;
1154 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1155 rnp->n_balk_notblocked++;
1156 else if (rnp->gp_tasks != NULL &&
1157 ULONG_CMP_LT(jiffies, rnp->boost_time))
1158 rnp->n_balk_notyet++;
1163 #else /* #ifdef CONFIG_RCU_TRACE */
1165 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1169 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1171 static void rcu_wake_cond(struct task_struct *t, int status)
1174 * If the thread is yielding, only wake it when this
1175 * is invoked from idle
1177 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1182 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1183 * or ->boost_tasks, advancing the pointer to the next task in the
1184 * ->blkd_tasks list.
1186 * Note that irqs must be enabled: boosting the task can block.
1187 * Returns 1 if there are more tasks needing to be boosted.
1189 static int rcu_boost(struct rcu_node *rnp)
1191 unsigned long flags;
1192 struct rt_mutex mtx;
1193 struct task_struct *t;
1194 struct list_head *tb;
1196 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1197 return 0; /* Nothing left to boost. */
1199 raw_spin_lock_irqsave(&rnp->lock, flags);
1202 * Recheck under the lock: all tasks in need of boosting
1203 * might exit their RCU read-side critical sections on their own.
1205 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1206 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1211 * Preferentially boost tasks blocking expedited grace periods.
1212 * This cannot starve the normal grace periods because a second
1213 * expedited grace period must boost all blocked tasks, including
1214 * those blocking the pre-existing normal grace period.
1216 if (rnp->exp_tasks != NULL) {
1217 tb = rnp->exp_tasks;
1218 rnp->n_exp_boosts++;
1220 tb = rnp->boost_tasks;
1221 rnp->n_normal_boosts++;
1223 rnp->n_tasks_boosted++;
1226 * We boost task t by manufacturing an rt_mutex that appears to
1227 * be held by task t. We leave a pointer to that rt_mutex where
1228 * task t can find it, and task t will release the mutex when it
1229 * exits its outermost RCU read-side critical section. Then
1230 * simply acquiring this artificial rt_mutex will boost task
1231 * t's priority. (Thanks to tglx for suggesting this approach!)
1233 * Note that task t must acquire rnp->lock to remove itself from
1234 * the ->blkd_tasks list, which it will do from exit() if from
1235 * nowhere else. We therefore are guaranteed that task t will
1236 * stay around at least until we drop rnp->lock. Note that
1237 * rnp->lock also resolves races between our priority boosting
1238 * and task t's exiting its outermost RCU read-side critical
1241 t = container_of(tb, struct task_struct, rcu_node_entry);
1242 rt_mutex_init_proxy_locked(&mtx, t);
1243 t->rcu_boost_mutex = &mtx;
1244 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1245 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1246 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1248 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1249 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1253 * Priority-boosting kthread. One per leaf rcu_node and one for the
1256 static int rcu_boost_kthread(void *arg)
1258 struct rcu_node *rnp = (struct rcu_node *)arg;
1262 trace_rcu_utilization(TPS("Start boost kthread@init"));
1264 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1265 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1266 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1267 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1268 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1269 more2boost = rcu_boost(rnp);
1275 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1276 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1277 schedule_timeout_interruptible(2);
1278 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1283 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1288 * Check to see if it is time to start boosting RCU readers that are
1289 * blocking the current grace period, and, if so, tell the per-rcu_node
1290 * kthread to start boosting them. If there is an expedited grace
1291 * period in progress, it is always time to boost.
1293 * The caller must hold rnp->lock, which this function releases.
1294 * The ->boost_kthread_task is immortal, so we don't need to worry
1295 * about it going away.
1297 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1299 struct task_struct *t;
1301 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1302 rnp->n_balk_exp_gp_tasks++;
1303 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1306 if (rnp->exp_tasks != NULL ||
1307 (rnp->gp_tasks != NULL &&
1308 rnp->boost_tasks == NULL &&
1310 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1311 if (rnp->exp_tasks == NULL)
1312 rnp->boost_tasks = rnp->gp_tasks;
1313 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1314 t = rnp->boost_kthread_task;
1316 rcu_wake_cond(t, rnp->boost_kthread_status);
1318 rcu_initiate_boost_trace(rnp);
1319 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1324 * Wake up the per-CPU kthread to invoke RCU callbacks.
1326 static void invoke_rcu_callbacks_kthread(void)
1328 unsigned long flags;
1330 local_irq_save(flags);
1331 __this_cpu_write(rcu_cpu_has_work, 1);
1332 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1333 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1334 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1335 __this_cpu_read(rcu_cpu_kthread_status));
1337 local_irq_restore(flags);
1341 * Is the current CPU running the RCU-callbacks kthread?
1342 * Caller must have preemption disabled.
1344 static bool rcu_is_callbacks_kthread(void)
1346 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1349 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1352 * Do priority-boost accounting for the start of a new grace period.
1354 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1356 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1360 * Create an RCU-boost kthread for the specified node if one does not
1361 * already exist. We only create this kthread for preemptible RCU.
1362 * Returns zero if all is well, a negated errno otherwise.
1364 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1365 struct rcu_node *rnp)
1367 int rnp_index = rnp - &rsp->node[0];
1368 unsigned long flags;
1369 struct sched_param sp;
1370 struct task_struct *t;
1372 if (&rcu_preempt_state != rsp)
1375 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1379 if (rnp->boost_kthread_task != NULL)
1381 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1382 "rcub/%d", rnp_index);
1385 raw_spin_lock_irqsave(&rnp->lock, flags);
1386 rnp->boost_kthread_task = t;
1387 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1388 sp.sched_priority = RCU_BOOST_PRIO;
1389 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1390 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1394 static void rcu_kthread_do_work(void)
1396 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1397 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1398 rcu_preempt_do_callbacks();
1401 static void rcu_cpu_kthread_setup(unsigned int cpu)
1403 struct sched_param sp;
1405 sp.sched_priority = RCU_KTHREAD_PRIO;
1406 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1409 static void rcu_cpu_kthread_park(unsigned int cpu)
1411 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1414 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1416 return __this_cpu_read(rcu_cpu_has_work);
1420 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1421 * RCU softirq used in flavors and configurations of RCU that do not
1422 * support RCU priority boosting.
1424 static void rcu_cpu_kthread(unsigned int cpu)
1426 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1427 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1430 for (spincnt = 0; spincnt < 10; spincnt++) {
1431 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1433 *statusp = RCU_KTHREAD_RUNNING;
1434 this_cpu_inc(rcu_cpu_kthread_loops);
1435 local_irq_disable();
1440 rcu_kthread_do_work();
1443 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1444 *statusp = RCU_KTHREAD_WAITING;
1448 *statusp = RCU_KTHREAD_YIELDING;
1449 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1450 schedule_timeout_interruptible(2);
1451 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1452 *statusp = RCU_KTHREAD_WAITING;
1456 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1457 * served by the rcu_node in question. The CPU hotplug lock is still
1458 * held, so the value of rnp->qsmaskinit will be stable.
1460 * We don't include outgoingcpu in the affinity set, use -1 if there is
1461 * no outgoing CPU. If there are no CPUs left in the affinity set,
1462 * this function allows the kthread to execute on any CPU.
1464 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1466 struct task_struct *t = rnp->boost_kthread_task;
1467 unsigned long mask = rnp->qsmaskinit;
1473 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1475 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1476 if ((mask & 0x1) && cpu != outgoingcpu)
1477 cpumask_set_cpu(cpu, cm);
1478 if (cpumask_weight(cm) == 0) {
1480 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1481 cpumask_clear_cpu(cpu, cm);
1482 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1484 set_cpus_allowed_ptr(t, cm);
1485 free_cpumask_var(cm);
1488 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1489 .store = &rcu_cpu_kthread_task,
1490 .thread_should_run = rcu_cpu_kthread_should_run,
1491 .thread_fn = rcu_cpu_kthread,
1492 .thread_comm = "rcuc/%u",
1493 .setup = rcu_cpu_kthread_setup,
1494 .park = rcu_cpu_kthread_park,
1498 * Spawn all kthreads -- called as soon as the scheduler is running.
1500 static int __init rcu_spawn_kthreads(void)
1502 struct rcu_node *rnp;
1505 rcu_scheduler_fully_active = 1;
1506 for_each_possible_cpu(cpu)
1507 per_cpu(rcu_cpu_has_work, cpu) = 0;
1508 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1509 rnp = rcu_get_root(rcu_state);
1510 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1511 if (NUM_RCU_NODES > 1) {
1512 rcu_for_each_leaf_node(rcu_state, rnp)
1513 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1517 early_initcall(rcu_spawn_kthreads);
1519 static void rcu_prepare_kthreads(int cpu)
1521 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1522 struct rcu_node *rnp = rdp->mynode;
1524 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1525 if (rcu_scheduler_fully_active)
1526 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1529 #else /* #ifdef CONFIG_RCU_BOOST */
1531 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1533 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1536 static void invoke_rcu_callbacks_kthread(void)
1541 static bool rcu_is_callbacks_kthread(void)
1546 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1550 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1554 static int __init rcu_scheduler_really_started(void)
1556 rcu_scheduler_fully_active = 1;
1559 early_initcall(rcu_scheduler_really_started);
1561 static void rcu_prepare_kthreads(int cpu)
1565 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1567 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1570 * Check to see if any future RCU-related work will need to be done
1571 * by the current CPU, even if none need be done immediately, returning
1572 * 1 if so. This function is part of the RCU implementation; it is -not-
1573 * an exported member of the RCU API.
1575 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1576 * any flavor of RCU.
1578 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1580 *delta_jiffies = ULONG_MAX;
1581 return rcu_cpu_has_callbacks(cpu, NULL);
1585 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1588 static void rcu_cleanup_after_idle(int cpu)
1593 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1596 static void rcu_prepare_for_idle(int cpu)
1601 * Don't bother keeping a running count of the number of RCU callbacks
1602 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1604 static void rcu_idle_count_callbacks_posted(void)
1608 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1611 * This code is invoked when a CPU goes idle, at which point we want
1612 * to have the CPU do everything required for RCU so that it can enter
1613 * the energy-efficient dyntick-idle mode. This is handled by a
1614 * state machine implemented by rcu_prepare_for_idle() below.
1616 * The following three proprocessor symbols control this state machine:
1618 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1619 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1620 * is sized to be roughly one RCU grace period. Those energy-efficiency
1621 * benchmarkers who might otherwise be tempted to set this to a large
1622 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1623 * system. And if you are -that- concerned about energy efficiency,
1624 * just power the system down and be done with it!
1625 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1626 * permitted to sleep in dyntick-idle mode with only lazy RCU
1627 * callbacks pending. Setting this too high can OOM your system.
1629 * The values below work well in practice. If future workloads require
1630 * adjustment, they can be converted into kernel config parameters, though
1631 * making the state machine smarter might be a better option.
1633 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1634 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1636 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1637 module_param(rcu_idle_gp_delay, int, 0644);
1638 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1639 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1641 extern int tick_nohz_enabled;
1644 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1645 * only if it has been awhile since the last time we did so. Afterwards,
1646 * if there are any callbacks ready for immediate invocation, return true.
1648 static bool rcu_try_advance_all_cbs(void)
1650 bool cbs_ready = false;
1651 struct rcu_data *rdp;
1652 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1653 struct rcu_node *rnp;
1654 struct rcu_state *rsp;
1656 /* Exit early if we advanced recently. */
1657 if (jiffies == rdtp->last_advance_all)
1659 rdtp->last_advance_all = jiffies;
1661 for_each_rcu_flavor(rsp) {
1662 rdp = this_cpu_ptr(rsp->rda);
1666 * Don't bother checking unless a grace period has
1667 * completed since we last checked and there are
1668 * callbacks not yet ready to invoke.
1670 if (rdp->completed != rnp->completed &&
1671 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1672 note_gp_changes(rsp, rdp);
1674 if (cpu_has_callbacks_ready_to_invoke(rdp))
1681 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1682 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1683 * caller to set the timeout based on whether or not there are non-lazy
1686 * The caller must have disabled interrupts.
1688 int rcu_needs_cpu(int cpu, unsigned long *dj)
1690 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1692 /* Snapshot to detect later posting of non-lazy callback. */
1693 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1695 /* If no callbacks, RCU doesn't need the CPU. */
1696 if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1701 /* Attempt to advance callbacks. */
1702 if (rcu_try_advance_all_cbs()) {
1703 /* Some ready to invoke, so initiate later invocation. */
1707 rdtp->last_accelerate = jiffies;
1709 /* Request timer delay depending on laziness, and round. */
1710 if (!rdtp->all_lazy) {
1711 *dj = round_up(rcu_idle_gp_delay + jiffies,
1712 rcu_idle_gp_delay) - jiffies;
1714 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1720 * Prepare a CPU for idle from an RCU perspective. The first major task
1721 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1722 * The second major task is to check to see if a non-lazy callback has
1723 * arrived at a CPU that previously had only lazy callbacks. The third
1724 * major task is to accelerate (that is, assign grace-period numbers to)
1725 * any recently arrived callbacks.
1727 * The caller must have disabled interrupts.
1729 static void rcu_prepare_for_idle(int cpu)
1731 struct rcu_data *rdp;
1732 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1733 struct rcu_node *rnp;
1734 struct rcu_state *rsp;
1737 /* Handle nohz enablement switches conservatively. */
1738 tne = ACCESS_ONCE(tick_nohz_enabled);
1739 if (tne != rdtp->tick_nohz_enabled_snap) {
1740 if (rcu_cpu_has_callbacks(cpu, NULL))
1741 invoke_rcu_core(); /* force nohz to see update. */
1742 rdtp->tick_nohz_enabled_snap = tne;
1748 /* If this is a no-CBs CPU, no callbacks, just return. */
1749 if (rcu_is_nocb_cpu(cpu))
1753 * If a non-lazy callback arrived at a CPU having only lazy
1754 * callbacks, invoke RCU core for the side-effect of recalculating
1755 * idle duration on re-entry to idle.
1757 if (rdtp->all_lazy &&
1758 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1759 rdtp->all_lazy = false;
1760 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1766 * If we have not yet accelerated this jiffy, accelerate all
1767 * callbacks on this CPU.
1769 if (rdtp->last_accelerate == jiffies)
1771 rdtp->last_accelerate = jiffies;
1772 for_each_rcu_flavor(rsp) {
1773 rdp = per_cpu_ptr(rsp->rda, cpu);
1774 if (!*rdp->nxttail[RCU_DONE_TAIL])
1777 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1778 rcu_accelerate_cbs(rsp, rnp, rdp);
1779 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1784 * Clean up for exit from idle. Attempt to advance callbacks based on
1785 * any grace periods that elapsed while the CPU was idle, and if any
1786 * callbacks are now ready to invoke, initiate invocation.
1788 static void rcu_cleanup_after_idle(int cpu)
1791 if (rcu_is_nocb_cpu(cpu))
1793 if (rcu_try_advance_all_cbs())
1798 * Keep a running count of the number of non-lazy callbacks posted
1799 * on this CPU. This running counter (which is never decremented) allows
1800 * rcu_prepare_for_idle() to detect when something out of the idle loop
1801 * posts a callback, even if an equal number of callbacks are invoked.
1802 * Of course, callbacks should only be posted from within a trace event
1803 * designed to be called from idle or from within RCU_NONIDLE().
1805 static void rcu_idle_count_callbacks_posted(void)
1807 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1811 * Data for flushing lazy RCU callbacks at OOM time.
1813 static atomic_t oom_callback_count;
1814 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1817 * RCU OOM callback -- decrement the outstanding count and deliver the
1818 * wake-up if we are the last one.
1820 static void rcu_oom_callback(struct rcu_head *rhp)
1822 if (atomic_dec_and_test(&oom_callback_count))
1823 wake_up(&oom_callback_wq);
1827 * Post an rcu_oom_notify callback on the current CPU if it has at
1828 * least one lazy callback. This will unnecessarily post callbacks
1829 * to CPUs that already have a non-lazy callback at the end of their
1830 * callback list, but this is an infrequent operation, so accept some
1831 * extra overhead to keep things simple.
1833 static void rcu_oom_notify_cpu(void *unused)
1835 struct rcu_state *rsp;
1836 struct rcu_data *rdp;
1838 for_each_rcu_flavor(rsp) {
1839 rdp = __this_cpu_ptr(rsp->rda);
1840 if (rdp->qlen_lazy != 0) {
1841 atomic_inc(&oom_callback_count);
1842 rsp->call(&rdp->oom_head, rcu_oom_callback);
1848 * If low on memory, ensure that each CPU has a non-lazy callback.
1849 * This will wake up CPUs that have only lazy callbacks, in turn
1850 * ensuring that they free up the corresponding memory in a timely manner.
1851 * Because an uncertain amount of memory will be freed in some uncertain
1852 * timeframe, we do not claim to have freed anything.
1854 static int rcu_oom_notify(struct notifier_block *self,
1855 unsigned long notused, void *nfreed)
1859 /* Wait for callbacks from earlier instance to complete. */
1860 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1861 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1864 * Prevent premature wakeup: ensure that all increments happen
1865 * before there is a chance of the counter reaching zero.
1867 atomic_set(&oom_callback_count, 1);
1870 for_each_online_cpu(cpu) {
1871 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1876 /* Unconditionally decrement: no need to wake ourselves up. */
1877 atomic_dec(&oom_callback_count);
1882 static struct notifier_block rcu_oom_nb = {
1883 .notifier_call = rcu_oom_notify
1886 static int __init rcu_register_oom_notifier(void)
1888 register_oom_notifier(&rcu_oom_nb);
1891 early_initcall(rcu_register_oom_notifier);
1893 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1895 #ifdef CONFIG_RCU_CPU_STALL_INFO
1897 #ifdef CONFIG_RCU_FAST_NO_HZ
1899 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1901 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1902 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1904 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1905 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1907 rdtp->all_lazy ? 'L' : '.',
1908 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1911 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1913 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1918 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1920 /* Initiate the stall-info list. */
1921 static void print_cpu_stall_info_begin(void)
1927 * Print out diagnostic information for the specified stalled CPU.
1929 * If the specified CPU is aware of the current RCU grace period
1930 * (flavor specified by rsp), then print the number of scheduling
1931 * clock interrupts the CPU has taken during the time that it has
1932 * been aware. Otherwise, print the number of RCU grace periods
1933 * that this CPU is ignorant of, for example, "1" if the CPU was
1934 * aware of the previous grace period.
1936 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1938 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1940 char fast_no_hz[72];
1941 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1942 struct rcu_dynticks *rdtp = rdp->dynticks;
1944 unsigned long ticks_value;
1946 if (rsp->gpnum == rdp->gpnum) {
1947 ticks_title = "ticks this GP";
1948 ticks_value = rdp->ticks_this_gp;
1950 ticks_title = "GPs behind";
1951 ticks_value = rsp->gpnum - rdp->gpnum;
1953 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1954 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1955 cpu, ticks_value, ticks_title,
1956 atomic_read(&rdtp->dynticks) & 0xfff,
1957 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1958 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1962 /* Terminate the stall-info list. */
1963 static void print_cpu_stall_info_end(void)
1968 /* Zero ->ticks_this_gp for all flavors of RCU. */
1969 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1971 rdp->ticks_this_gp = 0;
1972 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1975 /* Increment ->ticks_this_gp for all flavors of RCU. */
1976 static void increment_cpu_stall_ticks(void)
1978 struct rcu_state *rsp;
1980 for_each_rcu_flavor(rsp)
1981 __this_cpu_ptr(rsp->rda)->ticks_this_gp++;
1984 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1986 static void print_cpu_stall_info_begin(void)
1991 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1993 pr_cont(" %d", cpu);
1996 static void print_cpu_stall_info_end(void)
2001 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2005 static void increment_cpu_stall_ticks(void)
2009 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
2011 #ifdef CONFIG_RCU_NOCB_CPU
2014 * Offload callback processing from the boot-time-specified set of CPUs
2015 * specified by rcu_nocb_mask. For each CPU in the set, there is a
2016 * kthread created that pulls the callbacks from the corresponding CPU,
2017 * waits for a grace period to elapse, and invokes the callbacks.
2018 * The no-CBs CPUs do a wake_up() on their kthread when they insert
2019 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2020 * has been specified, in which case each kthread actively polls its
2021 * CPU. (Which isn't so great for energy efficiency, but which does
2022 * reduce RCU's overhead on that CPU.)
2024 * This is intended to be used in conjunction with Frederic Weisbecker's
2025 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2026 * running CPU-bound user-mode computations.
2028 * Offloading of callback processing could also in theory be used as
2029 * an energy-efficiency measure because CPUs with no RCU callbacks
2030 * queued are more aggressive about entering dyntick-idle mode.
2034 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2035 static int __init rcu_nocb_setup(char *str)
2037 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
2038 have_rcu_nocb_mask = true;
2039 cpulist_parse(str, rcu_nocb_mask);
2042 __setup("rcu_nocbs=", rcu_nocb_setup);
2044 static int __init parse_rcu_nocb_poll(char *arg)
2049 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2052 * Do any no-CBs CPUs need another grace period?
2054 * Interrupts must be disabled. If the caller does not hold the root
2055 * rnp_node structure's ->lock, the results are advisory only.
2057 static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2059 struct rcu_node *rnp = rcu_get_root(rsp);
2061 return rnp->need_future_gp[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
2065 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2068 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2070 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2074 * Set the root rcu_node structure's ->need_future_gp field
2075 * based on the sum of those of all rcu_node structures. This does
2076 * double-count the root rcu_node structure's requests, but this
2077 * is necessary to handle the possibility of a rcu_nocb_kthread()
2078 * having awakened during the time that the rcu_node structures
2079 * were being updated for the end of the previous grace period.
2081 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2083 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2086 static void rcu_init_one_nocb(struct rcu_node *rnp)
2088 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2089 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2092 /* Is the specified CPU a no-CPUs CPU? */
2093 bool rcu_is_nocb_cpu(int cpu)
2095 if (have_rcu_nocb_mask)
2096 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2101 * Enqueue the specified string of rcu_head structures onto the specified
2102 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2103 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2104 * counts are supplied by rhcount and rhcount_lazy.
2106 * If warranted, also wake up the kthread servicing this CPUs queues.
2108 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2109 struct rcu_head *rhp,
2110 struct rcu_head **rhtp,
2111 int rhcount, int rhcount_lazy,
2112 unsigned long flags)
2115 struct rcu_head **old_rhpp;
2116 struct task_struct *t;
2118 /* Enqueue the callback on the nocb list and update counts. */
2119 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2120 ACCESS_ONCE(*old_rhpp) = rhp;
2121 atomic_long_add(rhcount, &rdp->nocb_q_count);
2122 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2124 /* If we are not being polled and there is a kthread, awaken it ... */
2125 t = ACCESS_ONCE(rdp->nocb_kthread);
2126 if (rcu_nocb_poll || !t) {
2127 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2128 TPS("WakeNotPoll"));
2131 len = atomic_long_read(&rdp->nocb_q_count);
2132 if (old_rhpp == &rdp->nocb_head) {
2133 if (!irqs_disabled_flags(flags)) {
2134 wake_up(&rdp->nocb_wq); /* ... if queue was empty ... */
2135 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2138 rdp->nocb_defer_wakeup = true;
2139 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2140 TPS("WakeEmptyIsDeferred"));
2142 rdp->qlen_last_fqs_check = 0;
2143 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2144 wake_up_process(t); /* ... or if many callbacks queued. */
2145 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2146 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf"));
2148 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2154 * This is a helper for __call_rcu(), which invokes this when the normal
2155 * callback queue is inoperable. If this is not a no-CBs CPU, this
2156 * function returns failure back to __call_rcu(), which can complain
2159 * Otherwise, this function queues the callback where the corresponding
2160 * "rcuo" kthread can find it.
2162 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2163 bool lazy, unsigned long flags)
2166 if (!rcu_is_nocb_cpu(rdp->cpu))
2168 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2169 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2170 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2171 (unsigned long)rhp->func,
2172 -atomic_long_read(&rdp->nocb_q_count_lazy),
2173 -atomic_long_read(&rdp->nocb_q_count));
2175 trace_rcu_callback(rdp->rsp->name, rhp,
2176 -atomic_long_read(&rdp->nocb_q_count_lazy),
2177 -atomic_long_read(&rdp->nocb_q_count));
2182 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2185 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2186 struct rcu_data *rdp,
2187 unsigned long flags)
2189 long ql = rsp->qlen;
2190 long qll = rsp->qlen_lazy;
2192 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2193 if (!rcu_is_nocb_cpu(smp_processor_id()))
2198 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2199 if (rsp->orphan_donelist != NULL) {
2200 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2201 rsp->orphan_donetail, ql, qll, flags);
2203 rsp->orphan_donelist = NULL;
2204 rsp->orphan_donetail = &rsp->orphan_donelist;
2206 if (rsp->orphan_nxtlist != NULL) {
2207 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2208 rsp->orphan_nxttail, ql, qll, flags);
2210 rsp->orphan_nxtlist = NULL;
2211 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2217 * If necessary, kick off a new grace period, and either way wait
2218 * for a subsequent grace period to complete.
2220 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2224 unsigned long flags;
2225 struct rcu_node *rnp = rdp->mynode;
2227 raw_spin_lock_irqsave(&rnp->lock, flags);
2228 c = rcu_start_future_gp(rnp, rdp);
2229 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2232 * Wait for the grace period. Do so interruptibly to avoid messing
2233 * up the load average.
2235 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2237 wait_event_interruptible(
2238 rnp->nocb_gp_wq[c & 0x1],
2239 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2242 flush_signals(current);
2243 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2245 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2246 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2250 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2251 * callbacks queued by the corresponding no-CBs CPU.
2253 static int rcu_nocb_kthread(void *arg)
2257 struct rcu_head *list;
2258 struct rcu_head *next;
2259 struct rcu_head **tail;
2260 struct rcu_data *rdp = arg;
2262 /* Each pass through this loop invokes one batch of callbacks */
2264 /* If not polling, wait for next batch of callbacks. */
2265 if (!rcu_nocb_poll) {
2266 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2268 wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head);
2269 /* Memory barrier provide by xchg() below. */
2270 } else if (firsttime) {
2272 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2275 list = ACCESS_ONCE(rdp->nocb_head);
2278 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2280 schedule_timeout_interruptible(1);
2281 flush_signals(current);
2285 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2286 TPS("WokeNonEmpty"));
2289 * Extract queued callbacks, update counts, and wait
2290 * for a grace period to elapse.
2292 ACCESS_ONCE(rdp->nocb_head) = NULL;
2293 tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2294 c = atomic_long_xchg(&rdp->nocb_q_count, 0);
2295 cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2296 ACCESS_ONCE(rdp->nocb_p_count) += c;
2297 ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl;
2298 rcu_nocb_wait_gp(rdp);
2300 /* Each pass through the following loop invokes a callback. */
2301 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2305 /* Wait for enqueuing to complete, if needed. */
2306 while (next == NULL && &list->next != tail) {
2307 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2309 schedule_timeout_interruptible(1);
2310 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2314 debug_rcu_head_unqueue(list);
2316 if (__rcu_reclaim(rdp->rsp->name, list))
2322 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2323 ACCESS_ONCE(rdp->nocb_p_count) -= c;
2324 ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2325 rdp->n_nocbs_invoked += c;
2330 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2331 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2333 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2336 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2337 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2339 if (!rcu_nocb_need_deferred_wakeup(rdp))
2341 ACCESS_ONCE(rdp->nocb_defer_wakeup) = false;
2342 wake_up(&rdp->nocb_wq);
2343 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWakeEmpty"));
2346 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2347 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2349 rdp->nocb_tail = &rdp->nocb_head;
2350 init_waitqueue_head(&rdp->nocb_wq);
2353 /* Create a kthread for each RCU flavor for each no-CBs CPU. */
2354 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2357 struct rcu_data *rdp;
2358 struct task_struct *t;
2360 if (rcu_nocb_mask == NULL)
2362 for_each_cpu(cpu, rcu_nocb_mask) {
2363 rdp = per_cpu_ptr(rsp->rda, cpu);
2364 t = kthread_run(rcu_nocb_kthread, rdp,
2365 "rcuo%c/%d", rsp->abbr, cpu);
2367 ACCESS_ONCE(rdp->nocb_kthread) = t;
2371 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2372 static bool init_nocb_callback_list(struct rcu_data *rdp)
2374 if (rcu_nocb_mask == NULL ||
2375 !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2377 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2381 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2383 static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2388 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2392 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2396 static void rcu_init_one_nocb(struct rcu_node *rnp)
2400 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2401 bool lazy, unsigned long flags)
2406 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2407 struct rcu_data *rdp,
2408 unsigned long flags)
2413 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2417 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2422 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2426 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2430 static bool init_nocb_callback_list(struct rcu_data *rdp)
2435 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2438 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2439 * arbitrarily long period of time with the scheduling-clock tick turned
2440 * off. RCU will be paying attention to this CPU because it is in the
2441 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2442 * machine because the scheduling-clock tick has been disabled. Therefore,
2443 * if an adaptive-ticks CPU is failing to respond to the current grace
2444 * period and has not be idle from an RCU perspective, kick it.
2446 static void rcu_kick_nohz_cpu(int cpu)
2448 #ifdef CONFIG_NO_HZ_FULL
2449 if (tick_nohz_full_cpu(cpu))
2450 smp_send_reschedule(cpu);
2451 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2455 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2458 * Define RCU flavor that holds sysidle state. This needs to be the
2459 * most active flavor of RCU.
2461 #ifdef CONFIG_PREEMPT_RCU
2462 static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state;
2463 #else /* #ifdef CONFIG_PREEMPT_RCU */
2464 static struct rcu_state *rcu_sysidle_state = &rcu_sched_state;
2465 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2467 static int full_sysidle_state; /* Current system-idle state. */
2468 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2469 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2470 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2471 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2472 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2475 * Invoked to note exit from irq or task transition to idle. Note that
2476 * usermode execution does -not- count as idle here! After all, we want
2477 * to detect full-system idle states, not RCU quiescent states and grace
2478 * periods. The caller must have disabled interrupts.
2480 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2484 /* Adjust nesting, check for fully idle. */
2486 rdtp->dynticks_idle_nesting--;
2487 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2488 if (rdtp->dynticks_idle_nesting != 0)
2489 return; /* Still not fully idle. */
2491 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2492 DYNTICK_TASK_NEST_VALUE) {
2493 rdtp->dynticks_idle_nesting = 0;
2495 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2496 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2497 return; /* Still not fully idle. */
2501 /* Record start of fully idle period. */
2503 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2504 smp_mb__before_atomic_inc();
2505 atomic_inc(&rdtp->dynticks_idle);
2506 smp_mb__after_atomic_inc();
2507 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2511 * Unconditionally force exit from full system-idle state. This is
2512 * invoked when a normal CPU exits idle, but must be called separately
2513 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2514 * is that the timekeeping CPU is permitted to take scheduling-clock
2515 * interrupts while the system is in system-idle state, and of course
2516 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2517 * interrupt from any other type of interrupt.
2519 void rcu_sysidle_force_exit(void)
2521 int oldstate = ACCESS_ONCE(full_sysidle_state);
2525 * Each pass through the following loop attempts to exit full
2526 * system-idle state. If contention proves to be a problem,
2527 * a trylock-based contention tree could be used here.
2529 while (oldstate > RCU_SYSIDLE_SHORT) {
2530 newoldstate = cmpxchg(&full_sysidle_state,
2531 oldstate, RCU_SYSIDLE_NOT);
2532 if (oldstate == newoldstate &&
2533 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2534 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2535 return; /* We cleared it, done! */
2537 oldstate = newoldstate;
2539 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2543 * Invoked to note entry to irq or task transition from idle. Note that
2544 * usermode execution does -not- count as idle here! The caller must
2545 * have disabled interrupts.
2547 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2549 /* Adjust nesting, check for already non-idle. */
2551 rdtp->dynticks_idle_nesting++;
2552 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2553 if (rdtp->dynticks_idle_nesting != 1)
2554 return; /* Already non-idle. */
2557 * Allow for irq misnesting. Yes, it really is possible
2558 * to enter an irq handler then never leave it, and maybe
2559 * also vice versa. Handle both possibilities.
2561 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2562 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2563 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2564 return; /* Already non-idle. */
2566 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2570 /* Record end of idle period. */
2571 smp_mb__before_atomic_inc();
2572 atomic_inc(&rdtp->dynticks_idle);
2573 smp_mb__after_atomic_inc();
2574 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2577 * If we are the timekeeping CPU, we are permitted to be non-idle
2578 * during a system-idle state. This must be the case, because
2579 * the timekeeping CPU has to take scheduling-clock interrupts
2580 * during the time that the system is transitioning to full
2581 * system-idle state. This means that the timekeeping CPU must
2582 * invoke rcu_sysidle_force_exit() directly if it does anything
2583 * more than take a scheduling-clock interrupt.
2585 if (smp_processor_id() == tick_do_timer_cpu)
2588 /* Update system-idle state: We are clearly no longer fully idle! */
2589 rcu_sysidle_force_exit();
2593 * Check to see if the current CPU is idle. Note that usermode execution
2594 * does not count as idle. The caller must have disabled interrupts.
2596 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2597 unsigned long *maxj)
2601 struct rcu_dynticks *rdtp = rdp->dynticks;
2604 * If some other CPU has already reported non-idle, if this is
2605 * not the flavor of RCU that tracks sysidle state, or if this
2606 * is an offline or the timekeeping CPU, nothing to do.
2608 if (!*isidle || rdp->rsp != rcu_sysidle_state ||
2609 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2611 if (rcu_gp_in_progress(rdp->rsp))
2612 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2614 /* Pick up current idle and NMI-nesting counter and check. */
2615 cur = atomic_read(&rdtp->dynticks_idle);
2617 *isidle = false; /* We are not idle! */
2620 smp_mb(); /* Read counters before timestamps. */
2622 /* Pick up timestamps. */
2623 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2624 /* If this CPU entered idle more recently, update maxj timestamp. */
2625 if (ULONG_CMP_LT(*maxj, j))
2630 * Is this the flavor of RCU that is handling full-system idle?
2632 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2634 return rsp == rcu_sysidle_state;
2638 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2641 static void rcu_bind_gp_kthread(void)
2643 int cpu = ACCESS_ONCE(tick_do_timer_cpu);
2645 if (cpu < 0 || cpu >= nr_cpu_ids)
2647 if (raw_smp_processor_id() != cpu)
2648 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2652 * Return a delay in jiffies based on the number of CPUs, rcu_node
2653 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2654 * systems more time to transition to full-idle state in order to
2655 * avoid the cache thrashing that otherwise occur on the state variable.
2656 * Really small systems (less than a couple of tens of CPUs) should
2657 * instead use a single global atomically incremented counter, and later
2658 * versions of this will automatically reconfigure themselves accordingly.
2660 static unsigned long rcu_sysidle_delay(void)
2662 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2664 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2668 * Advance the full-system-idle state. This is invoked when all of
2669 * the non-timekeeping CPUs are idle.
2671 static void rcu_sysidle(unsigned long j)
2673 /* Check the current state. */
2674 switch (ACCESS_ONCE(full_sysidle_state)) {
2675 case RCU_SYSIDLE_NOT:
2677 /* First time all are idle, so note a short idle period. */
2678 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2681 case RCU_SYSIDLE_SHORT:
2684 * Idle for a bit, time to advance to next state?
2685 * cmpxchg failure means race with non-idle, let them win.
2687 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2688 (void)cmpxchg(&full_sysidle_state,
2689 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2692 case RCU_SYSIDLE_LONG:
2695 * Do an additional check pass before advancing to full.
2696 * cmpxchg failure means race with non-idle, let them win.
2698 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2699 (void)cmpxchg(&full_sysidle_state,
2700 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2709 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2710 * back to the beginning.
2712 static void rcu_sysidle_cancel(void)
2715 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2719 * Update the sysidle state based on the results of a force-quiescent-state
2720 * scan of the CPUs' dyntick-idle state.
2722 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2723 unsigned long maxj, bool gpkt)
2725 if (rsp != rcu_sysidle_state)
2726 return; /* Wrong flavor, ignore. */
2727 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2728 return; /* Running state machine from timekeeping CPU. */
2730 rcu_sysidle(maxj); /* More idle! */
2732 rcu_sysidle_cancel(); /* Idle is over. */
2736 * Wrapper for rcu_sysidle_report() when called from the grace-period
2737 * kthread's context.
2739 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2742 rcu_sysidle_report(rsp, isidle, maxj, true);
2745 /* Callback and function for forcing an RCU grace period. */
2746 struct rcu_sysidle_head {
2751 static void rcu_sysidle_cb(struct rcu_head *rhp)
2753 struct rcu_sysidle_head *rshp;
2756 * The following memory barrier is needed to replace the
2757 * memory barriers that would normally be in the memory
2760 smp_mb(); /* grace period precedes setting inuse. */
2762 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2763 ACCESS_ONCE(rshp->inuse) = 0;
2767 * Check to see if the system is fully idle, other than the timekeeping CPU.
2768 * The caller must have disabled interrupts.
2770 bool rcu_sys_is_idle(void)
2772 static struct rcu_sysidle_head rsh;
2773 int rss = ACCESS_ONCE(full_sysidle_state);
2775 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2778 /* Handle small-system case by doing a full scan of CPUs. */
2779 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2780 int oldrss = rss - 1;
2783 * One pass to advance to each state up to _FULL.
2784 * Give up if any pass fails to advance the state.
2786 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2789 unsigned long maxj = jiffies - ULONG_MAX / 4;
2790 struct rcu_data *rdp;
2792 /* Scan all the CPUs looking for nonidle CPUs. */
2793 for_each_possible_cpu(cpu) {
2794 rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu);
2795 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2799 rcu_sysidle_report(rcu_sysidle_state,
2800 isidle, maxj, false);
2802 rss = ACCESS_ONCE(full_sysidle_state);
2806 /* If this is the first observation of an idle period, record it. */
2807 if (rss == RCU_SYSIDLE_FULL) {
2808 rss = cmpxchg(&full_sysidle_state,
2809 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2810 return rss == RCU_SYSIDLE_FULL;
2813 smp_mb(); /* ensure rss load happens before later caller actions. */
2815 /* If already fully idle, tell the caller (in case of races). */
2816 if (rss == RCU_SYSIDLE_FULL_NOTED)
2820 * If we aren't there yet, and a grace period is not in flight,
2821 * initiate a grace period. Either way, tell the caller that
2822 * we are not there yet. We use an xchg() rather than an assignment
2823 * to make up for the memory barriers that would otherwise be
2824 * provided by the memory allocator.
2826 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2827 !rcu_gp_in_progress(rcu_sysidle_state) &&
2828 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2829 call_rcu(&rsh.rh, rcu_sysidle_cb);
2834 * Initialize dynticks sysidle state for CPUs coming online.
2836 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2838 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2841 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2843 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2847 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2851 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2852 unsigned long *maxj)
2856 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2861 static void rcu_bind_gp_kthread(void)
2865 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2870 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2874 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */