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 #endif /* #ifdef CONFIG_RCU_BOOST */
48 #ifdef CONFIG_RCU_NOCB_CPU
49 static cpumask_var_t rcu_nocb_mask
; /* CPUs to have callbacks offloaded. */
50 static bool have_rcu_nocb_mask
; /* Was rcu_nocb_mask allocated? */
51 static bool __read_mostly rcu_nocb_poll
; /* Offload kthread are to poll. */
52 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
55 * Check the RCU kernel configuration parameters and print informative
56 * messages about anything out of the ordinary. If you like #ifdef, you
57 * will love this function.
59 static void __init
rcu_bootup_announce_oddness(void)
61 #ifdef CONFIG_RCU_TRACE
62 pr_info("\tRCU debugfs-based tracing is enabled.\n");
64 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
65 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
68 #ifdef CONFIG_RCU_FANOUT_EXACT
69 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
71 #ifdef CONFIG_RCU_FAST_NO_HZ
72 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
74 #ifdef CONFIG_PROVE_RCU
75 pr_info("\tRCU lockdep checking is enabled.\n");
77 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
78 pr_info("\tRCU torture testing starts during boot.\n");
80 #if defined(CONFIG_RCU_CPU_STALL_INFO)
81 pr_info("\tAdditional per-CPU info printed with stalls.\n");
83 #if NUM_RCU_LVL_4 != 0
84 pr_info("\tFour-level hierarchy is enabled.\n");
86 if (rcu_fanout_leaf
!= CONFIG_RCU_FANOUT_LEAF
)
87 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf
);
88 if (nr_cpu_ids
!= NR_CPUS
)
89 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS
, nr_cpu_ids
);
90 #ifdef CONFIG_RCU_BOOST
91 pr_info("\tRCU kthread priority: %d.\n", kthread_prio
);
95 #ifdef CONFIG_PREEMPT_RCU
97 RCU_STATE_INITIALIZER(rcu_preempt
, 'p', call_rcu
);
98 static struct rcu_state
*rcu_state_p
= &rcu_preempt_state
;
100 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
);
101 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
105 * Tell them what RCU they are running.
107 static void __init
rcu_bootup_announce(void)
109 pr_info("Preemptible hierarchical RCU implementation.\n");
110 rcu_bootup_announce_oddness();
114 * Record a preemptible-RCU quiescent state for the specified CPU. Note
115 * that this just means that the task currently running on the CPU is
116 * not in a quiescent state. There might be any number of tasks blocked
117 * while in an RCU read-side critical section.
119 * As with the other rcu_*_qs() functions, callers to this function
120 * must disable preemption.
122 static void rcu_preempt_qs(void)
124 if (!__this_cpu_read(rcu_preempt_data
.passed_quiesce
)) {
125 trace_rcu_grace_period(TPS("rcu_preempt"),
126 __this_cpu_read(rcu_preempt_data
.gpnum
),
128 __this_cpu_write(rcu_preempt_data
.passed_quiesce
, 1);
129 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
130 current
->rcu_read_unlock_special
.b
.need_qs
= false;
135 * We have entered the scheduler, and the current task might soon be
136 * context-switched away from. If this task is in an RCU read-side
137 * critical section, we will no longer be able to rely on the CPU to
138 * record that fact, so we enqueue the task on the blkd_tasks list.
139 * The task will dequeue itself when it exits the outermost enclosing
140 * RCU read-side critical section. Therefore, the current grace period
141 * cannot be permitted to complete until the blkd_tasks list entries
142 * predating the current grace period drain, in other words, until
143 * rnp->gp_tasks becomes NULL.
145 * Caller must disable preemption.
147 static void rcu_preempt_note_context_switch(void)
149 struct task_struct
*t
= current
;
151 struct rcu_data
*rdp
;
152 struct rcu_node
*rnp
;
154 if (t
->rcu_read_lock_nesting
> 0 &&
155 !t
->rcu_read_unlock_special
.b
.blocked
) {
157 /* Possibly blocking in an RCU read-side critical section. */
158 rdp
= this_cpu_ptr(rcu_preempt_state
.rda
);
160 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
161 smp_mb__after_unlock_lock();
162 t
->rcu_read_unlock_special
.b
.blocked
= true;
163 t
->rcu_blocked_node
= rnp
;
166 * If this CPU has already checked in, then this task
167 * will hold up the next grace period rather than the
168 * current grace period. Queue the task accordingly.
169 * If the task is queued for the current grace period
170 * (i.e., this CPU has not yet passed through a quiescent
171 * state for the current grace period), then as long
172 * as that task remains queued, the current grace period
173 * cannot end. Note that there is some uncertainty as
174 * to exactly when the current grace period started.
175 * We take a conservative approach, which can result
176 * in unnecessarily waiting on tasks that started very
177 * slightly after the current grace period began. C'est
180 * But first, note that the current CPU must still be
183 WARN_ON_ONCE((rdp
->grpmask
& rnp
->qsmaskinit
) == 0);
184 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
185 if ((rnp
->qsmask
& rdp
->grpmask
) && rnp
->gp_tasks
!= NULL
) {
186 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
->prev
);
187 rnp
->gp_tasks
= &t
->rcu_node_entry
;
188 #ifdef CONFIG_RCU_BOOST
189 if (rnp
->boost_tasks
!= NULL
)
190 rnp
->boost_tasks
= rnp
->gp_tasks
;
191 #endif /* #ifdef CONFIG_RCU_BOOST */
193 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
194 if (rnp
->qsmask
& rdp
->grpmask
)
195 rnp
->gp_tasks
= &t
->rcu_node_entry
;
197 trace_rcu_preempt_task(rdp
->rsp
->name
,
199 (rnp
->qsmask
& rdp
->grpmask
)
202 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
203 } else if (t
->rcu_read_lock_nesting
< 0 &&
204 t
->rcu_read_unlock_special
.s
) {
207 * Complete exit from RCU read-side critical section on
208 * behalf of preempted instance of __rcu_read_unlock().
210 rcu_read_unlock_special(t
);
214 * Either we were not in an RCU read-side critical section to
215 * begin with, or we have now recorded that critical section
216 * globally. Either way, we can now note a quiescent state
217 * for this CPU. Again, if we were in an RCU read-side critical
218 * section, and if that critical section was blocking the current
219 * grace period, then the fact that the task has been enqueued
220 * means that we continue to block the current grace period.
226 * Check for preempted RCU readers blocking the current grace period
227 * for the specified rcu_node structure. If the caller needs a reliable
228 * answer, it must hold the rcu_node's ->lock.
230 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
232 return rnp
->gp_tasks
!= NULL
;
236 * Record a quiescent state for all tasks that were previously queued
237 * on the specified rcu_node structure and that were blocking the current
238 * RCU grace period. The caller must hold the specified rnp->lock with
239 * irqs disabled, and this lock is released upon return, but irqs remain
242 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
243 __releases(rnp
->lock
)
246 struct rcu_node
*rnp_p
;
248 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
249 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
250 return; /* Still need more quiescent states! */
256 * Either there is only one rcu_node in the tree,
257 * or tasks were kicked up to root rcu_node due to
258 * CPUs going offline.
260 rcu_report_qs_rsp(&rcu_preempt_state
, flags
);
264 /* Report up the rest of the hierarchy. */
266 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
267 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
268 smp_mb__after_unlock_lock();
269 rcu_report_qs_rnp(mask
, &rcu_preempt_state
, rnp_p
, flags
);
273 * Advance a ->blkd_tasks-list pointer to the next entry, instead
274 * returning NULL if at the end of the list.
276 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
277 struct rcu_node
*rnp
)
279 struct list_head
*np
;
281 np
= t
->rcu_node_entry
.next
;
282 if (np
== &rnp
->blkd_tasks
)
288 * Return true if the specified rcu_node structure has tasks that were
289 * preempted within an RCU read-side critical section.
291 static bool rcu_preempt_has_tasks(struct rcu_node
*rnp
)
293 return !list_empty(&rnp
->blkd_tasks
);
297 * Handle special cases during rcu_read_unlock(), such as needing to
298 * notify RCU core processing or task having blocked during the RCU
299 * read-side critical section.
301 void rcu_read_unlock_special(struct task_struct
*t
)
308 struct list_head
*np
;
309 #ifdef CONFIG_RCU_BOOST
310 bool drop_boost_mutex
= false;
311 #endif /* #ifdef CONFIG_RCU_BOOST */
312 struct rcu_node
*rnp
;
313 union rcu_special special
;
315 /* NMI handlers cannot block and cannot safely manipulate state. */
319 local_irq_save(flags
);
322 * If RCU core is waiting for this CPU to exit critical section,
323 * let it know that we have done so. Because irqs are disabled,
324 * t->rcu_read_unlock_special cannot change.
326 special
= t
->rcu_read_unlock_special
;
327 if (special
.b
.need_qs
) {
329 if (!t
->rcu_read_unlock_special
.s
) {
330 local_irq_restore(flags
);
335 /* Hardware IRQ handlers cannot block, complain if they get here. */
336 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
337 local_irq_restore(flags
);
341 /* Clean up if blocked during RCU read-side critical section. */
342 if (special
.b
.blocked
) {
343 t
->rcu_read_unlock_special
.b
.blocked
= false;
346 * Remove this task from the list it blocked on. The
347 * task can migrate while we acquire the lock, but at
348 * most one time. So at most two passes through loop.
351 rnp
= t
->rcu_blocked_node
;
352 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
353 smp_mb__after_unlock_lock();
354 if (rnp
== t
->rcu_blocked_node
)
356 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
358 empty
= !rcu_preempt_has_tasks(rnp
);
359 empty_norm
= !rcu_preempt_blocked_readers_cgp(rnp
);
360 empty_exp
= !rcu_preempted_readers_exp(rnp
);
361 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
362 np
= rcu_next_node_entry(t
, rnp
);
363 list_del_init(&t
->rcu_node_entry
);
364 t
->rcu_blocked_node
= NULL
;
365 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
367 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
369 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
371 #ifdef CONFIG_RCU_BOOST
372 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
373 rnp
->boost_tasks
= np
;
374 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
375 drop_boost_mutex
= rt_mutex_owner(&rnp
->boost_mtx
) == t
;
376 #endif /* #ifdef CONFIG_RCU_BOOST */
379 * If this was the last task on the list, go see if we
380 * need to propagate ->qsmaskinit bit clearing up the
383 if (!empty
&& !rcu_preempt_has_tasks(rnp
))
384 rcu_cleanup_dead_rnp(rnp
);
387 * If this was the last task on the current list, and if
388 * we aren't waiting on any CPUs, report the quiescent state.
389 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
390 * so we must take a snapshot of the expedited state.
392 empty_exp_now
= !rcu_preempted_readers_exp(rnp
);
393 if (!empty_norm
&& !rcu_preempt_blocked_readers_cgp(rnp
)) {
394 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
401 rcu_report_unblock_qs_rnp(rnp
, flags
);
403 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
406 #ifdef CONFIG_RCU_BOOST
407 /* Unboost if we were boosted. */
408 if (drop_boost_mutex
)
409 rt_mutex_unlock(&rnp
->boost_mtx
);
410 #endif /* #ifdef CONFIG_RCU_BOOST */
413 * If this was the last task on the expedited lists,
414 * then we need to report up the rcu_node hierarchy.
416 if (!empty_exp
&& empty_exp_now
)
417 rcu_report_exp_rnp(&rcu_preempt_state
, rnp
, true);
419 local_irq_restore(flags
);
424 * Dump detailed information for all tasks blocking the current RCU
425 * grace period on the specified rcu_node structure.
427 static void rcu_print_detail_task_stall_rnp(struct rcu_node
*rnp
)
430 struct task_struct
*t
;
432 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
433 if (!rcu_preempt_blocked_readers_cgp(rnp
)) {
434 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
437 t
= list_entry(rnp
->gp_tasks
,
438 struct task_struct
, rcu_node_entry
);
439 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
441 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
445 * Dump detailed information for all tasks blocking the current RCU
448 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
450 struct rcu_node
*rnp
= rcu_get_root(rsp
);
452 rcu_print_detail_task_stall_rnp(rnp
);
453 rcu_for_each_leaf_node(rsp
, rnp
)
454 rcu_print_detail_task_stall_rnp(rnp
);
457 #ifdef CONFIG_RCU_CPU_STALL_INFO
459 static void rcu_print_task_stall_begin(struct rcu_node
*rnp
)
461 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
462 rnp
->level
, rnp
->grplo
, rnp
->grphi
);
465 static void rcu_print_task_stall_end(void)
470 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
472 static void rcu_print_task_stall_begin(struct rcu_node
*rnp
)
476 static void rcu_print_task_stall_end(void)
480 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
483 * Scan the current list of tasks blocked within RCU read-side critical
484 * sections, printing out the tid of each.
486 static int rcu_print_task_stall(struct rcu_node
*rnp
)
488 struct task_struct
*t
;
491 if (!rcu_preempt_blocked_readers_cgp(rnp
))
493 rcu_print_task_stall_begin(rnp
);
494 t
= list_entry(rnp
->gp_tasks
,
495 struct task_struct
, rcu_node_entry
);
496 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
) {
497 pr_cont(" P%d", t
->pid
);
500 rcu_print_task_stall_end();
505 * Check that the list of blocked tasks for the newly completed grace
506 * period is in fact empty. It is a serious bug to complete a grace
507 * period that still has RCU readers blocked! This function must be
508 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
509 * must be held by the caller.
511 * Also, if there are blocked tasks on the list, they automatically
512 * block the newly created grace period, so set up ->gp_tasks accordingly.
514 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
516 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
517 if (rcu_preempt_has_tasks(rnp
))
518 rnp
->gp_tasks
= rnp
->blkd_tasks
.next
;
519 WARN_ON_ONCE(rnp
->qsmask
);
522 #ifdef CONFIG_HOTPLUG_CPU
524 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
527 * Check for a quiescent state from the current CPU. When a task blocks,
528 * the task is recorded in the corresponding CPU's rcu_node structure,
529 * which is checked elsewhere.
531 * Caller must disable hard irqs.
533 static void rcu_preempt_check_callbacks(void)
535 struct task_struct
*t
= current
;
537 if (t
->rcu_read_lock_nesting
== 0) {
541 if (t
->rcu_read_lock_nesting
> 0 &&
542 __this_cpu_read(rcu_preempt_data
.qs_pending
) &&
543 !__this_cpu_read(rcu_preempt_data
.passed_quiesce
))
544 t
->rcu_read_unlock_special
.b
.need_qs
= true;
547 #ifdef CONFIG_RCU_BOOST
549 static void rcu_preempt_do_callbacks(void)
551 rcu_do_batch(&rcu_preempt_state
, this_cpu_ptr(&rcu_preempt_data
));
554 #endif /* #ifdef CONFIG_RCU_BOOST */
557 * Queue a preemptible-RCU callback for invocation after a grace period.
559 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
561 __call_rcu(head
, func
, &rcu_preempt_state
, -1, 0);
563 EXPORT_SYMBOL_GPL(call_rcu
);
566 * synchronize_rcu - wait until a grace period has elapsed.
568 * Control will return to the caller some time after a full grace
569 * period has elapsed, in other words after all currently executing RCU
570 * read-side critical sections have completed. Note, however, that
571 * upon return from synchronize_rcu(), the caller might well be executing
572 * concurrently with new RCU read-side critical sections that began while
573 * synchronize_rcu() was waiting. RCU read-side critical sections are
574 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
576 * See the description of synchronize_sched() for more detailed information
577 * on memory ordering guarantees.
579 void synchronize_rcu(void)
581 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
582 !lock_is_held(&rcu_lock_map
) &&
583 !lock_is_held(&rcu_sched_lock_map
),
584 "Illegal synchronize_rcu() in RCU read-side critical section");
585 if (!rcu_scheduler_active
)
588 synchronize_rcu_expedited();
590 wait_rcu_gp(call_rcu
);
592 EXPORT_SYMBOL_GPL(synchronize_rcu
);
594 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq
);
595 static unsigned long sync_rcu_preempt_exp_count
;
596 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex
);
599 * Return non-zero if there are any tasks in RCU read-side critical
600 * sections blocking the current preemptible-RCU expedited grace period.
601 * If there is no preemptible-RCU expedited grace period currently in
602 * progress, returns zero unconditionally.
604 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
)
606 return rnp
->exp_tasks
!= NULL
;
610 * return non-zero if there is no RCU expedited grace period in progress
611 * for the specified rcu_node structure, in other words, if all CPUs and
612 * tasks covered by the specified rcu_node structure have done their bit
613 * for the current expedited grace period. Works only for preemptible
614 * RCU -- other RCU implementation use other means.
616 * Caller must hold sync_rcu_preempt_exp_mutex.
618 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
620 return !rcu_preempted_readers_exp(rnp
) &&
621 ACCESS_ONCE(rnp
->expmask
) == 0;
625 * Report the exit from RCU read-side critical section for the last task
626 * that queued itself during or before the current expedited preemptible-RCU
627 * grace period. This event is reported either to the rcu_node structure on
628 * which the task was queued or to one of that rcu_node structure's ancestors,
629 * recursively up the tree. (Calm down, calm down, we do the recursion
632 * Most callers will set the "wake" flag, but the task initiating the
633 * expedited grace period need not wake itself.
635 * Caller must hold sync_rcu_preempt_exp_mutex.
637 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
643 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
644 smp_mb__after_unlock_lock();
646 if (!sync_rcu_preempt_exp_done(rnp
)) {
647 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
650 if (rnp
->parent
== NULL
) {
651 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
653 smp_mb(); /* EGP done before wake_up(). */
654 wake_up(&sync_rcu_preempt_exp_wq
);
659 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
661 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
662 smp_mb__after_unlock_lock();
663 rnp
->expmask
&= ~mask
;
668 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
669 * grace period for the specified rcu_node structure. If there are no such
670 * tasks, report it up the rcu_node hierarchy.
672 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
673 * CPU hotplug operations.
676 sync_rcu_preempt_exp_init(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
681 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
682 smp_mb__after_unlock_lock();
683 if (!rcu_preempt_has_tasks(rnp
)) {
684 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
686 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
687 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
691 rcu_report_exp_rnp(rsp
, rnp
, false); /* Don't wake self. */
695 * synchronize_rcu_expedited - Brute-force RCU grace period
697 * Wait for an RCU-preempt grace period, but expedite it. The basic
698 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
699 * the ->blkd_tasks lists and wait for this list to drain. This consumes
700 * significant time on all CPUs and is unfriendly to real-time workloads,
701 * so is thus not recommended for any sort of common-case code.
702 * In fact, if you are using synchronize_rcu_expedited() in a loop,
703 * please restructure your code to batch your updates, and then Use a
704 * single synchronize_rcu() instead.
706 void synchronize_rcu_expedited(void)
709 struct rcu_node
*rnp
;
710 struct rcu_state
*rsp
= &rcu_preempt_state
;
714 smp_mb(); /* Caller's modifications seen first by other CPUs. */
715 snap
= ACCESS_ONCE(sync_rcu_preempt_exp_count
) + 1;
716 smp_mb(); /* Above access cannot bleed into critical section. */
719 * Block CPU-hotplug operations. This means that any CPU-hotplug
720 * operation that finds an rcu_node structure with tasks in the
721 * process of being boosted will know that all tasks blocking
722 * this expedited grace period will already be in the process of
723 * being boosted. This simplifies the process of moving tasks
724 * from leaf to root rcu_node structures.
726 if (!try_get_online_cpus()) {
727 /* CPU-hotplug operation in flight, fall back to normal GP. */
728 wait_rcu_gp(call_rcu
);
733 * Acquire lock, falling back to synchronize_rcu() if too many
734 * lock-acquisition failures. Of course, if someone does the
735 * expedited grace period for us, just leave.
737 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex
)) {
738 if (ULONG_CMP_LT(snap
,
739 ACCESS_ONCE(sync_rcu_preempt_exp_count
))) {
741 goto mb_ret
; /* Others did our work for us. */
743 if (trycount
++ < 10) {
744 udelay(trycount
* num_online_cpus());
747 wait_rcu_gp(call_rcu
);
751 if (ULONG_CMP_LT(snap
, ACCESS_ONCE(sync_rcu_preempt_exp_count
))) {
753 goto unlock_mb_ret
; /* Others did our work for us. */
756 /* force all RCU readers onto ->blkd_tasks lists. */
757 synchronize_sched_expedited();
759 /* Initialize ->expmask for all non-leaf rcu_node structures. */
760 rcu_for_each_nonleaf_node_breadth_first(rsp
, rnp
) {
761 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
762 smp_mb__after_unlock_lock();
763 rnp
->expmask
= rnp
->qsmaskinit
;
764 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
767 /* Snapshot current state of ->blkd_tasks lists. */
768 rcu_for_each_leaf_node(rsp
, rnp
)
769 sync_rcu_preempt_exp_init(rsp
, rnp
);
770 if (NUM_RCU_NODES
> 1)
771 sync_rcu_preempt_exp_init(rsp
, rcu_get_root(rsp
));
775 /* Wait for snapshotted ->blkd_tasks lists to drain. */
776 rnp
= rcu_get_root(rsp
);
777 wait_event(sync_rcu_preempt_exp_wq
,
778 sync_rcu_preempt_exp_done(rnp
));
780 /* Clean up and exit. */
781 smp_mb(); /* ensure expedited GP seen before counter increment. */
782 ACCESS_ONCE(sync_rcu_preempt_exp_count
) =
783 sync_rcu_preempt_exp_count
+ 1;
785 mutex_unlock(&sync_rcu_preempt_exp_mutex
);
787 smp_mb(); /* ensure subsequent action seen after grace period. */
789 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
792 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
794 * Note that this primitive does not necessarily wait for an RCU grace period
795 * to complete. For example, if there are no RCU callbacks queued anywhere
796 * in the system, then rcu_barrier() is within its rights to return
797 * immediately, without waiting for anything, much less an RCU grace period.
799 void rcu_barrier(void)
801 _rcu_barrier(&rcu_preempt_state
);
803 EXPORT_SYMBOL_GPL(rcu_barrier
);
806 * Initialize preemptible RCU's state structures.
808 static void __init
__rcu_init_preempt(void)
810 rcu_init_one(&rcu_preempt_state
, &rcu_preempt_data
);
814 * Check for a task exiting while in a preemptible-RCU read-side
815 * critical section, clean up if so. No need to issue warnings,
816 * as debug_check_no_locks_held() already does this if lockdep
821 struct task_struct
*t
= current
;
823 if (likely(list_empty(¤t
->rcu_node_entry
)))
825 t
->rcu_read_lock_nesting
= 1;
827 t
->rcu_read_unlock_special
.b
.blocked
= true;
831 #else /* #ifdef CONFIG_PREEMPT_RCU */
833 static struct rcu_state
*rcu_state_p
= &rcu_sched_state
;
836 * Tell them what RCU they are running.
838 static void __init
rcu_bootup_announce(void)
840 pr_info("Hierarchical RCU implementation.\n");
841 rcu_bootup_announce_oddness();
845 * Because preemptible RCU does not exist, we never have to check for
846 * CPUs being in quiescent states.
848 static void rcu_preempt_note_context_switch(void)
853 * Because preemptible RCU does not exist, there are never any preempted
856 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
861 #ifdef CONFIG_HOTPLUG_CPU
864 * Because there is no preemptible RCU, there can be no readers blocked.
866 static bool rcu_preempt_has_tasks(struct rcu_node
*rnp
)
871 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
874 * Because preemptible RCU does not exist, we never have to check for
875 * tasks blocked within RCU read-side critical sections.
877 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
882 * Because preemptible RCU does not exist, we never have to check for
883 * tasks blocked within RCU read-side critical sections.
885 static int rcu_print_task_stall(struct rcu_node
*rnp
)
891 * Because there is no preemptible RCU, there can be no readers blocked,
892 * so there is no need to check for blocked tasks. So check only for
893 * bogus qsmask values.
895 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
897 WARN_ON_ONCE(rnp
->qsmask
);
901 * Because preemptible RCU does not exist, it never has any callbacks
904 static void rcu_preempt_check_callbacks(void)
909 * Wait for an rcu-preempt grace period, but make it happen quickly.
910 * But because preemptible RCU does not exist, map to rcu-sched.
912 void synchronize_rcu_expedited(void)
914 synchronize_sched_expedited();
916 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
919 * Because preemptible RCU does not exist, rcu_barrier() is just
920 * another name for rcu_barrier_sched().
922 void rcu_barrier(void)
926 EXPORT_SYMBOL_GPL(rcu_barrier
);
929 * Because preemptible RCU does not exist, it need not be initialized.
931 static void __init
__rcu_init_preempt(void)
936 * Because preemptible RCU does not exist, tasks cannot possibly exit
937 * while in preemptible RCU read-side critical sections.
943 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
945 #ifdef CONFIG_RCU_BOOST
947 #include "../locking/rtmutex_common.h"
949 #ifdef CONFIG_RCU_TRACE
951 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
953 if (!rcu_preempt_has_tasks(rnp
))
954 rnp
->n_balk_blkd_tasks
++;
955 else if (rnp
->exp_tasks
== NULL
&& rnp
->gp_tasks
== NULL
)
956 rnp
->n_balk_exp_gp_tasks
++;
957 else if (rnp
->gp_tasks
!= NULL
&& rnp
->boost_tasks
!= NULL
)
958 rnp
->n_balk_boost_tasks
++;
959 else if (rnp
->gp_tasks
!= NULL
&& rnp
->qsmask
!= 0)
960 rnp
->n_balk_notblocked
++;
961 else if (rnp
->gp_tasks
!= NULL
&&
962 ULONG_CMP_LT(jiffies
, rnp
->boost_time
))
963 rnp
->n_balk_notyet
++;
968 #else /* #ifdef CONFIG_RCU_TRACE */
970 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
974 #endif /* #else #ifdef CONFIG_RCU_TRACE */
976 static void rcu_wake_cond(struct task_struct
*t
, int status
)
979 * If the thread is yielding, only wake it when this
980 * is invoked from idle
982 if (status
!= RCU_KTHREAD_YIELDING
|| is_idle_task(current
))
987 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
988 * or ->boost_tasks, advancing the pointer to the next task in the
991 * Note that irqs must be enabled: boosting the task can block.
992 * Returns 1 if there are more tasks needing to be boosted.
994 static int rcu_boost(struct rcu_node
*rnp
)
997 struct task_struct
*t
;
998 struct list_head
*tb
;
1000 if (ACCESS_ONCE(rnp
->exp_tasks
) == NULL
&&
1001 ACCESS_ONCE(rnp
->boost_tasks
) == NULL
)
1002 return 0; /* Nothing left to boost. */
1004 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1005 smp_mb__after_unlock_lock();
1008 * Recheck under the lock: all tasks in need of boosting
1009 * might exit their RCU read-side critical sections on their own.
1011 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
1012 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1017 * Preferentially boost tasks blocking expedited grace periods.
1018 * This cannot starve the normal grace periods because a second
1019 * expedited grace period must boost all blocked tasks, including
1020 * those blocking the pre-existing normal grace period.
1022 if (rnp
->exp_tasks
!= NULL
) {
1023 tb
= rnp
->exp_tasks
;
1024 rnp
->n_exp_boosts
++;
1026 tb
= rnp
->boost_tasks
;
1027 rnp
->n_normal_boosts
++;
1029 rnp
->n_tasks_boosted
++;
1032 * We boost task t by manufacturing an rt_mutex that appears to
1033 * be held by task t. We leave a pointer to that rt_mutex where
1034 * task t can find it, and task t will release the mutex when it
1035 * exits its outermost RCU read-side critical section. Then
1036 * simply acquiring this artificial rt_mutex will boost task
1037 * t's priority. (Thanks to tglx for suggesting this approach!)
1039 * Note that task t must acquire rnp->lock to remove itself from
1040 * the ->blkd_tasks list, which it will do from exit() if from
1041 * nowhere else. We therefore are guaranteed that task t will
1042 * stay around at least until we drop rnp->lock. Note that
1043 * rnp->lock also resolves races between our priority boosting
1044 * and task t's exiting its outermost RCU read-side critical
1047 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1048 rt_mutex_init_proxy_locked(&rnp
->boost_mtx
, t
);
1049 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1050 /* Lock only for side effect: boosts task t's priority. */
1051 rt_mutex_lock(&rnp
->boost_mtx
);
1052 rt_mutex_unlock(&rnp
->boost_mtx
); /* Then keep lockdep happy. */
1054 return ACCESS_ONCE(rnp
->exp_tasks
) != NULL
||
1055 ACCESS_ONCE(rnp
->boost_tasks
) != NULL
;
1059 * Priority-boosting kthread. One per leaf rcu_node and one for the
1062 static int rcu_boost_kthread(void *arg
)
1064 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1068 trace_rcu_utilization(TPS("Start boost kthread@init"));
1070 rnp
->boost_kthread_status
= RCU_KTHREAD_WAITING
;
1071 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1072 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1073 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1074 rnp
->boost_kthread_status
= RCU_KTHREAD_RUNNING
;
1075 more2boost
= rcu_boost(rnp
);
1081 rnp
->boost_kthread_status
= RCU_KTHREAD_YIELDING
;
1082 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1083 schedule_timeout_interruptible(2);
1084 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1089 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1094 * Check to see if it is time to start boosting RCU readers that are
1095 * blocking the current grace period, and, if so, tell the per-rcu_node
1096 * kthread to start boosting them. If there is an expedited grace
1097 * period in progress, it is always time to boost.
1099 * The caller must hold rnp->lock, which this function releases.
1100 * The ->boost_kthread_task is immortal, so we don't need to worry
1101 * about it going away.
1103 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1104 __releases(rnp
->lock
)
1106 struct task_struct
*t
;
1108 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1109 rnp
->n_balk_exp_gp_tasks
++;
1110 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1113 if (rnp
->exp_tasks
!= NULL
||
1114 (rnp
->gp_tasks
!= NULL
&&
1115 rnp
->boost_tasks
== NULL
&&
1117 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))) {
1118 if (rnp
->exp_tasks
== NULL
)
1119 rnp
->boost_tasks
= rnp
->gp_tasks
;
1120 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1121 t
= rnp
->boost_kthread_task
;
1123 rcu_wake_cond(t
, rnp
->boost_kthread_status
);
1125 rcu_initiate_boost_trace(rnp
);
1126 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1131 * Wake up the per-CPU kthread to invoke RCU callbacks.
1133 static void invoke_rcu_callbacks_kthread(void)
1135 unsigned long flags
;
1137 local_irq_save(flags
);
1138 __this_cpu_write(rcu_cpu_has_work
, 1);
1139 if (__this_cpu_read(rcu_cpu_kthread_task
) != NULL
&&
1140 current
!= __this_cpu_read(rcu_cpu_kthread_task
)) {
1141 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task
),
1142 __this_cpu_read(rcu_cpu_kthread_status
));
1144 local_irq_restore(flags
);
1148 * Is the current CPU running the RCU-callbacks kthread?
1149 * Caller must have preemption disabled.
1151 static bool rcu_is_callbacks_kthread(void)
1153 return __this_cpu_read(rcu_cpu_kthread_task
) == current
;
1156 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1159 * Do priority-boost accounting for the start of a new grace period.
1161 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1163 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1167 * Create an RCU-boost kthread for the specified node if one does not
1168 * already exist. We only create this kthread for preemptible RCU.
1169 * Returns zero if all is well, a negated errno otherwise.
1171 static int rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1172 struct rcu_node
*rnp
)
1174 int rnp_index
= rnp
- &rsp
->node
[0];
1175 unsigned long flags
;
1176 struct sched_param sp
;
1177 struct task_struct
*t
;
1179 if (&rcu_preempt_state
!= rsp
)
1182 if (!rcu_scheduler_fully_active
|| rnp
->qsmaskinit
== 0)
1186 if (rnp
->boost_kthread_task
!= NULL
)
1188 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1189 "rcub/%d", rnp_index
);
1192 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1193 smp_mb__after_unlock_lock();
1194 rnp
->boost_kthread_task
= t
;
1195 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1196 sp
.sched_priority
= kthread_prio
;
1197 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1198 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1202 static void rcu_kthread_do_work(void)
1204 rcu_do_batch(&rcu_sched_state
, this_cpu_ptr(&rcu_sched_data
));
1205 rcu_do_batch(&rcu_bh_state
, this_cpu_ptr(&rcu_bh_data
));
1206 rcu_preempt_do_callbacks();
1209 static void rcu_cpu_kthread_setup(unsigned int cpu
)
1211 struct sched_param sp
;
1213 sp
.sched_priority
= kthread_prio
;
1214 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1217 static void rcu_cpu_kthread_park(unsigned int cpu
)
1219 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1222 static int rcu_cpu_kthread_should_run(unsigned int cpu
)
1224 return __this_cpu_read(rcu_cpu_has_work
);
1228 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1229 * RCU softirq used in flavors and configurations of RCU that do not
1230 * support RCU priority boosting.
1232 static void rcu_cpu_kthread(unsigned int cpu
)
1234 unsigned int *statusp
= this_cpu_ptr(&rcu_cpu_kthread_status
);
1235 char work
, *workp
= this_cpu_ptr(&rcu_cpu_has_work
);
1238 for (spincnt
= 0; spincnt
< 10; spincnt
++) {
1239 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1241 *statusp
= RCU_KTHREAD_RUNNING
;
1242 this_cpu_inc(rcu_cpu_kthread_loops
);
1243 local_irq_disable();
1248 rcu_kthread_do_work();
1251 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1252 *statusp
= RCU_KTHREAD_WAITING
;
1256 *statusp
= RCU_KTHREAD_YIELDING
;
1257 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1258 schedule_timeout_interruptible(2);
1259 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1260 *statusp
= RCU_KTHREAD_WAITING
;
1264 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1265 * served by the rcu_node in question. The CPU hotplug lock is still
1266 * held, so the value of rnp->qsmaskinit will be stable.
1268 * We don't include outgoingcpu in the affinity set, use -1 if there is
1269 * no outgoing CPU. If there are no CPUs left in the affinity set,
1270 * this function allows the kthread to execute on any CPU.
1272 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1274 struct task_struct
*t
= rnp
->boost_kthread_task
;
1275 unsigned long mask
= rnp
->qsmaskinit
;
1281 if (!zalloc_cpumask_var(&cm
, GFP_KERNEL
))
1283 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1284 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1285 cpumask_set_cpu(cpu
, cm
);
1286 if (cpumask_weight(cm
) == 0)
1288 set_cpus_allowed_ptr(t
, cm
);
1289 free_cpumask_var(cm
);
1292 static struct smp_hotplug_thread rcu_cpu_thread_spec
= {
1293 .store
= &rcu_cpu_kthread_task
,
1294 .thread_should_run
= rcu_cpu_kthread_should_run
,
1295 .thread_fn
= rcu_cpu_kthread
,
1296 .thread_comm
= "rcuc/%u",
1297 .setup
= rcu_cpu_kthread_setup
,
1298 .park
= rcu_cpu_kthread_park
,
1302 * Spawn boost kthreads -- called as soon as the scheduler is running.
1304 static void __init
rcu_spawn_boost_kthreads(void)
1306 struct rcu_node
*rnp
;
1309 for_each_possible_cpu(cpu
)
1310 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1311 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec
));
1312 rcu_for_each_leaf_node(rcu_state_p
, rnp
)
1313 (void)rcu_spawn_one_boost_kthread(rcu_state_p
, rnp
);
1316 static void rcu_prepare_kthreads(int cpu
)
1318 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
1319 struct rcu_node
*rnp
= rdp
->mynode
;
1321 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1322 if (rcu_scheduler_fully_active
)
1323 (void)rcu_spawn_one_boost_kthread(rcu_state_p
, rnp
);
1326 #else /* #ifdef CONFIG_RCU_BOOST */
1328 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1329 __releases(rnp
->lock
)
1331 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1334 static void invoke_rcu_callbacks_kthread(void)
1339 static bool rcu_is_callbacks_kthread(void)
1344 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1348 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1352 static void __init
rcu_spawn_boost_kthreads(void)
1356 static void rcu_prepare_kthreads(int cpu
)
1360 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1362 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1365 * Check to see if any future RCU-related work will need to be done
1366 * by the current CPU, even if none need be done immediately, returning
1367 * 1 if so. This function is part of the RCU implementation; it is -not-
1368 * an exported member of the RCU API.
1370 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1371 * any flavor of RCU.
1373 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1374 int rcu_needs_cpu(unsigned long *delta_jiffies
)
1376 *delta_jiffies
= ULONG_MAX
;
1377 return rcu_cpu_has_callbacks(NULL
);
1379 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1382 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1385 static void rcu_cleanup_after_idle(void)
1390 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1393 static void rcu_prepare_for_idle(void)
1398 * Don't bother keeping a running count of the number of RCU callbacks
1399 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1401 static void rcu_idle_count_callbacks_posted(void)
1405 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1408 * This code is invoked when a CPU goes idle, at which point we want
1409 * to have the CPU do everything required for RCU so that it can enter
1410 * the energy-efficient dyntick-idle mode. This is handled by a
1411 * state machine implemented by rcu_prepare_for_idle() below.
1413 * The following three proprocessor symbols control this state machine:
1415 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1416 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1417 * is sized to be roughly one RCU grace period. Those energy-efficiency
1418 * benchmarkers who might otherwise be tempted to set this to a large
1419 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1420 * system. And if you are -that- concerned about energy efficiency,
1421 * just power the system down and be done with it!
1422 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1423 * permitted to sleep in dyntick-idle mode with only lazy RCU
1424 * callbacks pending. Setting this too high can OOM your system.
1426 * The values below work well in practice. If future workloads require
1427 * adjustment, they can be converted into kernel config parameters, though
1428 * making the state machine smarter might be a better option.
1430 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1431 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1433 static int rcu_idle_gp_delay
= RCU_IDLE_GP_DELAY
;
1434 module_param(rcu_idle_gp_delay
, int, 0644);
1435 static int rcu_idle_lazy_gp_delay
= RCU_IDLE_LAZY_GP_DELAY
;
1436 module_param(rcu_idle_lazy_gp_delay
, int, 0644);
1438 extern int tick_nohz_active
;
1441 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1442 * only if it has been awhile since the last time we did so. Afterwards,
1443 * if there are any callbacks ready for immediate invocation, return true.
1445 static bool __maybe_unused
rcu_try_advance_all_cbs(void)
1447 bool cbs_ready
= false;
1448 struct rcu_data
*rdp
;
1449 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1450 struct rcu_node
*rnp
;
1451 struct rcu_state
*rsp
;
1453 /* Exit early if we advanced recently. */
1454 if (jiffies
== rdtp
->last_advance_all
)
1456 rdtp
->last_advance_all
= jiffies
;
1458 for_each_rcu_flavor(rsp
) {
1459 rdp
= this_cpu_ptr(rsp
->rda
);
1463 * Don't bother checking unless a grace period has
1464 * completed since we last checked and there are
1465 * callbacks not yet ready to invoke.
1467 if ((rdp
->completed
!= rnp
->completed
||
1468 unlikely(ACCESS_ONCE(rdp
->gpwrap
))) &&
1469 rdp
->nxttail
[RCU_DONE_TAIL
] != rdp
->nxttail
[RCU_NEXT_TAIL
])
1470 note_gp_changes(rsp
, rdp
);
1472 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1479 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1480 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1481 * caller to set the timeout based on whether or not there are non-lazy
1484 * The caller must have disabled interrupts.
1486 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1487 int rcu_needs_cpu(unsigned long *dj
)
1489 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1491 /* Snapshot to detect later posting of non-lazy callback. */
1492 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1494 /* If no callbacks, RCU doesn't need the CPU. */
1495 if (!rcu_cpu_has_callbacks(&rdtp
->all_lazy
)) {
1500 /* Attempt to advance callbacks. */
1501 if (rcu_try_advance_all_cbs()) {
1502 /* Some ready to invoke, so initiate later invocation. */
1506 rdtp
->last_accelerate
= jiffies
;
1508 /* Request timer delay depending on laziness, and round. */
1509 if (!rdtp
->all_lazy
) {
1510 *dj
= round_up(rcu_idle_gp_delay
+ jiffies
,
1511 rcu_idle_gp_delay
) - jiffies
;
1513 *dj
= round_jiffies(rcu_idle_lazy_gp_delay
+ jiffies
) - jiffies
;
1517 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1520 * Prepare a CPU for idle from an RCU perspective. The first major task
1521 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1522 * The second major task is to check to see if a non-lazy callback has
1523 * arrived at a CPU that previously had only lazy callbacks. The third
1524 * major task is to accelerate (that is, assign grace-period numbers to)
1525 * any recently arrived callbacks.
1527 * The caller must have disabled interrupts.
1529 static void rcu_prepare_for_idle(void)
1531 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1533 struct rcu_data
*rdp
;
1534 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1535 struct rcu_node
*rnp
;
1536 struct rcu_state
*rsp
;
1539 /* Handle nohz enablement switches conservatively. */
1540 tne
= ACCESS_ONCE(tick_nohz_active
);
1541 if (tne
!= rdtp
->tick_nohz_enabled_snap
) {
1542 if (rcu_cpu_has_callbacks(NULL
))
1543 invoke_rcu_core(); /* force nohz to see update. */
1544 rdtp
->tick_nohz_enabled_snap
= tne
;
1550 /* If this is a no-CBs CPU, no callbacks, just return. */
1551 if (rcu_is_nocb_cpu(smp_processor_id()))
1555 * If a non-lazy callback arrived at a CPU having only lazy
1556 * callbacks, invoke RCU core for the side-effect of recalculating
1557 * idle duration on re-entry to idle.
1559 if (rdtp
->all_lazy
&&
1560 rdtp
->nonlazy_posted
!= rdtp
->nonlazy_posted_snap
) {
1561 rdtp
->all_lazy
= false;
1562 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1568 * If we have not yet accelerated this jiffy, accelerate all
1569 * callbacks on this CPU.
1571 if (rdtp
->last_accelerate
== jiffies
)
1573 rdtp
->last_accelerate
= jiffies
;
1574 for_each_rcu_flavor(rsp
) {
1575 rdp
= this_cpu_ptr(rsp
->rda
);
1576 if (!*rdp
->nxttail
[RCU_DONE_TAIL
])
1579 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1580 smp_mb__after_unlock_lock();
1581 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1582 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1584 rcu_gp_kthread_wake(rsp
);
1586 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1590 * Clean up for exit from idle. Attempt to advance callbacks based on
1591 * any grace periods that elapsed while the CPU was idle, and if any
1592 * callbacks are now ready to invoke, initiate invocation.
1594 static void rcu_cleanup_after_idle(void)
1596 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1597 if (rcu_is_nocb_cpu(smp_processor_id()))
1599 if (rcu_try_advance_all_cbs())
1601 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1605 * Keep a running count of the number of non-lazy callbacks posted
1606 * on this CPU. This running counter (which is never decremented) allows
1607 * rcu_prepare_for_idle() to detect when something out of the idle loop
1608 * posts a callback, even if an equal number of callbacks are invoked.
1609 * Of course, callbacks should only be posted from within a trace event
1610 * designed to be called from idle or from within RCU_NONIDLE().
1612 static void rcu_idle_count_callbacks_posted(void)
1614 __this_cpu_add(rcu_dynticks
.nonlazy_posted
, 1);
1618 * Data for flushing lazy RCU callbacks at OOM time.
1620 static atomic_t oom_callback_count
;
1621 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq
);
1624 * RCU OOM callback -- decrement the outstanding count and deliver the
1625 * wake-up if we are the last one.
1627 static void rcu_oom_callback(struct rcu_head
*rhp
)
1629 if (atomic_dec_and_test(&oom_callback_count
))
1630 wake_up(&oom_callback_wq
);
1634 * Post an rcu_oom_notify callback on the current CPU if it has at
1635 * least one lazy callback. This will unnecessarily post callbacks
1636 * to CPUs that already have a non-lazy callback at the end of their
1637 * callback list, but this is an infrequent operation, so accept some
1638 * extra overhead to keep things simple.
1640 static void rcu_oom_notify_cpu(void *unused
)
1642 struct rcu_state
*rsp
;
1643 struct rcu_data
*rdp
;
1645 for_each_rcu_flavor(rsp
) {
1646 rdp
= raw_cpu_ptr(rsp
->rda
);
1647 if (rdp
->qlen_lazy
!= 0) {
1648 atomic_inc(&oom_callback_count
);
1649 rsp
->call(&rdp
->oom_head
, rcu_oom_callback
);
1655 * If low on memory, ensure that each CPU has a non-lazy callback.
1656 * This will wake up CPUs that have only lazy callbacks, in turn
1657 * ensuring that they free up the corresponding memory in a timely manner.
1658 * Because an uncertain amount of memory will be freed in some uncertain
1659 * timeframe, we do not claim to have freed anything.
1661 static int rcu_oom_notify(struct notifier_block
*self
,
1662 unsigned long notused
, void *nfreed
)
1666 /* Wait for callbacks from earlier instance to complete. */
1667 wait_event(oom_callback_wq
, atomic_read(&oom_callback_count
) == 0);
1668 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1671 * Prevent premature wakeup: ensure that all increments happen
1672 * before there is a chance of the counter reaching zero.
1674 atomic_set(&oom_callback_count
, 1);
1677 for_each_online_cpu(cpu
) {
1678 smp_call_function_single(cpu
, rcu_oom_notify_cpu
, NULL
, 1);
1679 cond_resched_rcu_qs();
1683 /* Unconditionally decrement: no need to wake ourselves up. */
1684 atomic_dec(&oom_callback_count
);
1689 static struct notifier_block rcu_oom_nb
= {
1690 .notifier_call
= rcu_oom_notify
1693 static int __init
rcu_register_oom_notifier(void)
1695 register_oom_notifier(&rcu_oom_nb
);
1698 early_initcall(rcu_register_oom_notifier
);
1700 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1702 #ifdef CONFIG_RCU_CPU_STALL_INFO
1704 #ifdef CONFIG_RCU_FAST_NO_HZ
1706 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1708 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1709 unsigned long nlpd
= rdtp
->nonlazy_posted
- rdtp
->nonlazy_posted_snap
;
1711 sprintf(cp
, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1712 rdtp
->last_accelerate
& 0xffff, jiffies
& 0xffff,
1714 rdtp
->all_lazy
? 'L' : '.',
1715 rdtp
->tick_nohz_enabled_snap
? '.' : 'D');
1718 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1720 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1725 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1727 /* Initiate the stall-info list. */
1728 static void print_cpu_stall_info_begin(void)
1734 * Print out diagnostic information for the specified stalled CPU.
1736 * If the specified CPU is aware of the current RCU grace period
1737 * (flavor specified by rsp), then print the number of scheduling
1738 * clock interrupts the CPU has taken during the time that it has
1739 * been aware. Otherwise, print the number of RCU grace periods
1740 * that this CPU is ignorant of, for example, "1" if the CPU was
1741 * aware of the previous grace period.
1743 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1745 static void print_cpu_stall_info(struct rcu_state
*rsp
, int cpu
)
1747 char fast_no_hz
[72];
1748 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1749 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
1751 unsigned long ticks_value
;
1753 if (rsp
->gpnum
== rdp
->gpnum
) {
1754 ticks_title
= "ticks this GP";
1755 ticks_value
= rdp
->ticks_this_gp
;
1757 ticks_title
= "GPs behind";
1758 ticks_value
= rsp
->gpnum
- rdp
->gpnum
;
1760 print_cpu_stall_fast_no_hz(fast_no_hz
, cpu
);
1761 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1762 cpu
, ticks_value
, ticks_title
,
1763 atomic_read(&rdtp
->dynticks
) & 0xfff,
1764 rdtp
->dynticks_nesting
, rdtp
->dynticks_nmi_nesting
,
1765 rdp
->softirq_snap
, kstat_softirqs_cpu(RCU_SOFTIRQ
, cpu
),
1766 ACCESS_ONCE(rsp
->n_force_qs
) - rsp
->n_force_qs_gpstart
,
1770 /* Terminate the stall-info list. */
1771 static void print_cpu_stall_info_end(void)
1776 /* Zero ->ticks_this_gp for all flavors of RCU. */
1777 static void zero_cpu_stall_ticks(struct rcu_data
*rdp
)
1779 rdp
->ticks_this_gp
= 0;
1780 rdp
->softirq_snap
= kstat_softirqs_cpu(RCU_SOFTIRQ
, smp_processor_id());
1783 /* Increment ->ticks_this_gp for all flavors of RCU. */
1784 static void increment_cpu_stall_ticks(void)
1786 struct rcu_state
*rsp
;
1788 for_each_rcu_flavor(rsp
)
1789 raw_cpu_inc(rsp
->rda
->ticks_this_gp
);
1792 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1794 static void print_cpu_stall_info_begin(void)
1799 static void print_cpu_stall_info(struct rcu_state
*rsp
, int cpu
)
1801 pr_cont(" %d", cpu
);
1804 static void print_cpu_stall_info_end(void)
1809 static void zero_cpu_stall_ticks(struct rcu_data
*rdp
)
1813 static void increment_cpu_stall_ticks(void)
1817 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1819 #ifdef CONFIG_RCU_NOCB_CPU
1822 * Offload callback processing from the boot-time-specified set of CPUs
1823 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1824 * kthread created that pulls the callbacks from the corresponding CPU,
1825 * waits for a grace period to elapse, and invokes the callbacks.
1826 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1827 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1828 * has been specified, in which case each kthread actively polls its
1829 * CPU. (Which isn't so great for energy efficiency, but which does
1830 * reduce RCU's overhead on that CPU.)
1832 * This is intended to be used in conjunction with Frederic Weisbecker's
1833 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1834 * running CPU-bound user-mode computations.
1836 * Offloading of callback processing could also in theory be used as
1837 * an energy-efficiency measure because CPUs with no RCU callbacks
1838 * queued are more aggressive about entering dyntick-idle mode.
1842 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1843 static int __init
rcu_nocb_setup(char *str
)
1845 alloc_bootmem_cpumask_var(&rcu_nocb_mask
);
1846 have_rcu_nocb_mask
= true;
1847 cpulist_parse(str
, rcu_nocb_mask
);
1850 __setup("rcu_nocbs=", rcu_nocb_setup
);
1852 static int __init
parse_rcu_nocb_poll(char *arg
)
1857 early_param("rcu_nocb_poll", parse_rcu_nocb_poll
);
1860 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1863 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1865 wake_up_all(&rnp
->nocb_gp_wq
[rnp
->completed
& 0x1]);
1869 * Set the root rcu_node structure's ->need_future_gp field
1870 * based on the sum of those of all rcu_node structures. This does
1871 * double-count the root rcu_node structure's requests, but this
1872 * is necessary to handle the possibility of a rcu_nocb_kthread()
1873 * having awakened during the time that the rcu_node structures
1874 * were being updated for the end of the previous grace period.
1876 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
1878 rnp
->need_future_gp
[(rnp
->completed
+ 1) & 0x1] += nrq
;
1881 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
1883 init_waitqueue_head(&rnp
->nocb_gp_wq
[0]);
1884 init_waitqueue_head(&rnp
->nocb_gp_wq
[1]);
1887 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1888 /* Is the specified CPU a no-CBs CPU? */
1889 bool rcu_is_nocb_cpu(int cpu
)
1891 if (have_rcu_nocb_mask
)
1892 return cpumask_test_cpu(cpu
, rcu_nocb_mask
);
1895 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1898 * Kick the leader kthread for this NOCB group.
1900 static void wake_nocb_leader(struct rcu_data
*rdp
, bool force
)
1902 struct rcu_data
*rdp_leader
= rdp
->nocb_leader
;
1904 if (!ACCESS_ONCE(rdp_leader
->nocb_kthread
))
1906 if (ACCESS_ONCE(rdp_leader
->nocb_leader_sleep
) || force
) {
1907 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1908 ACCESS_ONCE(rdp_leader
->nocb_leader_sleep
) = false;
1909 wake_up(&rdp_leader
->nocb_wq
);
1914 * Does the specified CPU need an RCU callback for the specified flavor
1917 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state
*rsp
, int cpu
)
1919 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1921 #ifdef CONFIG_PROVE_RCU
1922 struct rcu_head
*rhp
;
1923 #endif /* #ifdef CONFIG_PROVE_RCU */
1926 * Check count of all no-CBs callbacks awaiting invocation.
1927 * There needs to be a barrier before this function is called,
1928 * but associated with a prior determination that no more
1929 * callbacks would be posted. In the worst case, the first
1930 * barrier in _rcu_barrier() suffices (but the caller cannot
1931 * necessarily rely on this, not a substitute for the caller
1932 * getting the concurrency design right!). There must also be
1933 * a barrier between the following load an posting of a callback
1934 * (if a callback is in fact needed). This is associated with an
1935 * atomic_inc() in the caller.
1937 ret
= atomic_long_read(&rdp
->nocb_q_count
);
1939 #ifdef CONFIG_PROVE_RCU
1940 rhp
= ACCESS_ONCE(rdp
->nocb_head
);
1942 rhp
= ACCESS_ONCE(rdp
->nocb_gp_head
);
1944 rhp
= ACCESS_ONCE(rdp
->nocb_follower_head
);
1946 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1947 if (!ACCESS_ONCE(rdp
->nocb_kthread
) && rhp
) {
1948 /* RCU callback enqueued before CPU first came online??? */
1949 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1953 #endif /* #ifdef CONFIG_PROVE_RCU */
1959 * Enqueue the specified string of rcu_head structures onto the specified
1960 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1961 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1962 * counts are supplied by rhcount and rhcount_lazy.
1964 * If warranted, also wake up the kthread servicing this CPUs queues.
1966 static void __call_rcu_nocb_enqueue(struct rcu_data
*rdp
,
1967 struct rcu_head
*rhp
,
1968 struct rcu_head
**rhtp
,
1969 int rhcount
, int rhcount_lazy
,
1970 unsigned long flags
)
1973 struct rcu_head
**old_rhpp
;
1974 struct task_struct
*t
;
1976 /* Enqueue the callback on the nocb list and update counts. */
1977 atomic_long_add(rhcount
, &rdp
->nocb_q_count
);
1978 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1979 old_rhpp
= xchg(&rdp
->nocb_tail
, rhtp
);
1980 ACCESS_ONCE(*old_rhpp
) = rhp
;
1981 atomic_long_add(rhcount_lazy
, &rdp
->nocb_q_count_lazy
);
1982 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1984 /* If we are not being polled and there is a kthread, awaken it ... */
1985 t
= ACCESS_ONCE(rdp
->nocb_kthread
);
1986 if (rcu_nocb_poll
|| !t
) {
1987 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
1988 TPS("WakeNotPoll"));
1991 len
= atomic_long_read(&rdp
->nocb_q_count
);
1992 if (old_rhpp
== &rdp
->nocb_head
) {
1993 if (!irqs_disabled_flags(flags
)) {
1994 /* ... if queue was empty ... */
1995 wake_nocb_leader(rdp
, false);
1996 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
1999 rdp
->nocb_defer_wakeup
= RCU_NOGP_WAKE
;
2000 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2001 TPS("WakeEmptyIsDeferred"));
2003 rdp
->qlen_last_fqs_check
= 0;
2004 } else if (len
> rdp
->qlen_last_fqs_check
+ qhimark
) {
2005 /* ... or if many callbacks queued. */
2006 if (!irqs_disabled_flags(flags
)) {
2007 wake_nocb_leader(rdp
, true);
2008 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2011 rdp
->nocb_defer_wakeup
= RCU_NOGP_WAKE_FORCE
;
2012 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2013 TPS("WakeOvfIsDeferred"));
2015 rdp
->qlen_last_fqs_check
= LONG_MAX
/ 2;
2017 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("WakeNot"));
2023 * This is a helper for __call_rcu(), which invokes this when the normal
2024 * callback queue is inoperable. If this is not a no-CBs CPU, this
2025 * function returns failure back to __call_rcu(), which can complain
2028 * Otherwise, this function queues the callback where the corresponding
2029 * "rcuo" kthread can find it.
2031 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2032 bool lazy
, unsigned long flags
)
2035 if (!rcu_is_nocb_cpu(rdp
->cpu
))
2037 __call_rcu_nocb_enqueue(rdp
, rhp
, &rhp
->next
, 1, lazy
, flags
);
2038 if (__is_kfree_rcu_offset((unsigned long)rhp
->func
))
2039 trace_rcu_kfree_callback(rdp
->rsp
->name
, rhp
,
2040 (unsigned long)rhp
->func
,
2041 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
2042 -atomic_long_read(&rdp
->nocb_q_count
));
2044 trace_rcu_callback(rdp
->rsp
->name
, rhp
,
2045 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
2046 -atomic_long_read(&rdp
->nocb_q_count
));
2049 * If called from an extended quiescent state with interrupts
2050 * disabled, invoke the RCU core in order to allow the idle-entry
2051 * deferred-wakeup check to function.
2053 if (irqs_disabled_flags(flags
) &&
2054 !rcu_is_watching() &&
2055 cpu_online(smp_processor_id()))
2062 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2065 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
2066 struct rcu_data
*rdp
,
2067 unsigned long flags
)
2069 long ql
= rsp
->qlen
;
2070 long qll
= rsp
->qlen_lazy
;
2072 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2073 if (!rcu_is_nocb_cpu(smp_processor_id()))
2078 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2079 if (rsp
->orphan_donelist
!= NULL
) {
2080 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_donelist
,
2081 rsp
->orphan_donetail
, ql
, qll
, flags
);
2083 rsp
->orphan_donelist
= NULL
;
2084 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2086 if (rsp
->orphan_nxtlist
!= NULL
) {
2087 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_nxtlist
,
2088 rsp
->orphan_nxttail
, ql
, qll
, flags
);
2090 rsp
->orphan_nxtlist
= NULL
;
2091 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2097 * If necessary, kick off a new grace period, and either way wait
2098 * for a subsequent grace period to complete.
2100 static void rcu_nocb_wait_gp(struct rcu_data
*rdp
)
2104 unsigned long flags
;
2106 struct rcu_node
*rnp
= rdp
->mynode
;
2108 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2109 smp_mb__after_unlock_lock();
2110 needwake
= rcu_start_future_gp(rnp
, rdp
, &c
);
2111 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2113 rcu_gp_kthread_wake(rdp
->rsp
);
2116 * Wait for the grace period. Do so interruptibly to avoid messing
2117 * up the load average.
2119 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("StartWait"));
2121 wait_event_interruptible(
2122 rnp
->nocb_gp_wq
[c
& 0x1],
2123 (d
= ULONG_CMP_GE(ACCESS_ONCE(rnp
->completed
), c
)));
2126 WARN_ON(signal_pending(current
));
2127 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("ResumeWait"));
2129 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("EndWait"));
2130 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2134 * Leaders come here to wait for additional callbacks to show up.
2135 * This function does not return until callbacks appear.
2137 static void nocb_leader_wait(struct rcu_data
*my_rdp
)
2139 bool firsttime
= true;
2141 struct rcu_data
*rdp
;
2142 struct rcu_head
**tail
;
2146 /* Wait for callbacks to appear. */
2147 if (!rcu_nocb_poll
) {
2148 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
, "Sleep");
2149 wait_event_interruptible(my_rdp
->nocb_wq
,
2150 !ACCESS_ONCE(my_rdp
->nocb_leader_sleep
));
2151 /* Memory barrier handled by smp_mb() calls below and repoll. */
2152 } else if (firsttime
) {
2153 firsttime
= false; /* Don't drown trace log with "Poll"! */
2154 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
, "Poll");
2158 * Each pass through the following loop checks a follower for CBs.
2159 * We are our own first follower. Any CBs found are moved to
2160 * nocb_gp_head, where they await a grace period.
2163 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
) {
2164 rdp
->nocb_gp_head
= ACCESS_ONCE(rdp
->nocb_head
);
2165 if (!rdp
->nocb_gp_head
)
2166 continue; /* No CBs here, try next follower. */
2168 /* Move callbacks to wait-for-GP list, which is empty. */
2169 ACCESS_ONCE(rdp
->nocb_head
) = NULL
;
2170 rdp
->nocb_gp_tail
= xchg(&rdp
->nocb_tail
, &rdp
->nocb_head
);
2175 * If there were no callbacks, sleep a bit, rescan after a
2176 * memory barrier, and go retry.
2178 if (unlikely(!gotcbs
)) {
2180 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
,
2182 WARN_ON(signal_pending(current
));
2183 schedule_timeout_interruptible(1);
2185 /* Rescan in case we were a victim of memory ordering. */
2186 my_rdp
->nocb_leader_sleep
= true;
2187 smp_mb(); /* Ensure _sleep true before scan. */
2188 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
)
2189 if (ACCESS_ONCE(rdp
->nocb_head
)) {
2190 /* Found CB, so short-circuit next wait. */
2191 my_rdp
->nocb_leader_sleep
= false;
2197 /* Wait for one grace period. */
2198 rcu_nocb_wait_gp(my_rdp
);
2201 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2202 * We set it now, but recheck for new callbacks while
2203 * traversing our follower list.
2205 my_rdp
->nocb_leader_sleep
= true;
2206 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2208 /* Each pass through the following loop wakes a follower, if needed. */
2209 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
) {
2210 if (ACCESS_ONCE(rdp
->nocb_head
))
2211 my_rdp
->nocb_leader_sleep
= false;/* No need to sleep.*/
2212 if (!rdp
->nocb_gp_head
)
2213 continue; /* No CBs, so no need to wake follower. */
2215 /* Append callbacks to follower's "done" list. */
2216 tail
= xchg(&rdp
->nocb_follower_tail
, rdp
->nocb_gp_tail
);
2217 *tail
= rdp
->nocb_gp_head
;
2218 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2219 if (rdp
!= my_rdp
&& tail
== &rdp
->nocb_follower_head
) {
2221 * List was empty, wake up the follower.
2222 * Memory barriers supplied by atomic_long_add().
2224 wake_up(&rdp
->nocb_wq
);
2228 /* If we (the leader) don't have CBs, go wait some more. */
2229 if (!my_rdp
->nocb_follower_head
)
2234 * Followers come here to wait for additional callbacks to show up.
2235 * This function does not return until callbacks appear.
2237 static void nocb_follower_wait(struct rcu_data
*rdp
)
2239 bool firsttime
= true;
2242 if (!rcu_nocb_poll
) {
2243 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2245 wait_event_interruptible(rdp
->nocb_wq
,
2246 ACCESS_ONCE(rdp
->nocb_follower_head
));
2247 } else if (firsttime
) {
2248 /* Don't drown trace log with "Poll"! */
2250 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, "Poll");
2252 if (smp_load_acquire(&rdp
->nocb_follower_head
)) {
2253 /* ^^^ Ensure CB invocation follows _head test. */
2257 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2259 WARN_ON(signal_pending(current
));
2260 schedule_timeout_interruptible(1);
2265 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2266 * callbacks queued by the corresponding no-CBs CPU, however, there is
2267 * an optional leader-follower relationship so that the grace-period
2268 * kthreads don't have to do quite so many wakeups.
2270 static int rcu_nocb_kthread(void *arg
)
2273 struct rcu_head
*list
;
2274 struct rcu_head
*next
;
2275 struct rcu_head
**tail
;
2276 struct rcu_data
*rdp
= arg
;
2278 /* Each pass through this loop invokes one batch of callbacks */
2280 /* Wait for callbacks. */
2281 if (rdp
->nocb_leader
== rdp
)
2282 nocb_leader_wait(rdp
);
2284 nocb_follower_wait(rdp
);
2286 /* Pull the ready-to-invoke callbacks onto local list. */
2287 list
= ACCESS_ONCE(rdp
->nocb_follower_head
);
2289 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, "WokeNonEmpty");
2290 ACCESS_ONCE(rdp
->nocb_follower_head
) = NULL
;
2291 tail
= xchg(&rdp
->nocb_follower_tail
, &rdp
->nocb_follower_head
);
2293 /* Each pass through the following loop invokes a callback. */
2294 trace_rcu_batch_start(rdp
->rsp
->name
,
2295 atomic_long_read(&rdp
->nocb_q_count_lazy
),
2296 atomic_long_read(&rdp
->nocb_q_count
), -1);
2300 /* Wait for enqueuing to complete, if needed. */
2301 while (next
== NULL
&& &list
->next
!= tail
) {
2302 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2304 schedule_timeout_interruptible(1);
2305 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2309 debug_rcu_head_unqueue(list
);
2311 if (__rcu_reclaim(rdp
->rsp
->name
, list
))
2317 trace_rcu_batch_end(rdp
->rsp
->name
, c
, !!list
, 0, 0, 1);
2318 smp_mb__before_atomic(); /* _add after CB invocation. */
2319 atomic_long_add(-c
, &rdp
->nocb_q_count
);
2320 atomic_long_add(-cl
, &rdp
->nocb_q_count_lazy
);
2321 rdp
->n_nocbs_invoked
+= c
;
2326 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2327 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2329 return ACCESS_ONCE(rdp
->nocb_defer_wakeup
);
2332 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2333 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2337 if (!rcu_nocb_need_deferred_wakeup(rdp
))
2339 ndw
= ACCESS_ONCE(rdp
->nocb_defer_wakeup
);
2340 ACCESS_ONCE(rdp
->nocb_defer_wakeup
) = RCU_NOGP_WAKE_NOT
;
2341 wake_nocb_leader(rdp
, ndw
== RCU_NOGP_WAKE_FORCE
);
2342 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("DeferredWake"));
2345 void __init
rcu_init_nohz(void)
2348 bool need_rcu_nocb_mask
= true;
2349 struct rcu_state
*rsp
;
2351 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2352 need_rcu_nocb_mask
= false;
2353 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2355 #if defined(CONFIG_NO_HZ_FULL)
2356 if (tick_nohz_full_running
&& cpumask_weight(tick_nohz_full_mask
))
2357 need_rcu_nocb_mask
= true;
2358 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2360 if (!have_rcu_nocb_mask
&& need_rcu_nocb_mask
) {
2361 if (!zalloc_cpumask_var(&rcu_nocb_mask
, GFP_KERNEL
)) {
2362 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2365 have_rcu_nocb_mask
= true;
2367 if (!have_rcu_nocb_mask
)
2370 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2371 pr_info("\tOffload RCU callbacks from CPU 0\n");
2372 cpumask_set_cpu(0, rcu_nocb_mask
);
2373 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2374 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2375 pr_info("\tOffload RCU callbacks from all CPUs\n");
2376 cpumask_copy(rcu_nocb_mask
, cpu_possible_mask
);
2377 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2378 #if defined(CONFIG_NO_HZ_FULL)
2379 if (tick_nohz_full_running
)
2380 cpumask_or(rcu_nocb_mask
, rcu_nocb_mask
, tick_nohz_full_mask
);
2381 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2383 if (!cpumask_subset(rcu_nocb_mask
, cpu_possible_mask
)) {
2384 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2385 cpumask_and(rcu_nocb_mask
, cpu_possible_mask
,
2388 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2389 cpumask_pr_args(rcu_nocb_mask
));
2391 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2393 for_each_rcu_flavor(rsp
) {
2394 for_each_cpu(cpu
, rcu_nocb_mask
) {
2395 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2398 * If there are early callbacks, they will need
2399 * to be moved to the nocb lists.
2401 WARN_ON_ONCE(rdp
->nxttail
[RCU_NEXT_TAIL
] !=
2403 rdp
->nxttail
[RCU_NEXT_TAIL
] != NULL
);
2404 init_nocb_callback_list(rdp
);
2406 rcu_organize_nocb_kthreads(rsp
);
2410 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2411 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2413 rdp
->nocb_tail
= &rdp
->nocb_head
;
2414 init_waitqueue_head(&rdp
->nocb_wq
);
2415 rdp
->nocb_follower_tail
= &rdp
->nocb_follower_head
;
2419 * If the specified CPU is a no-CBs CPU that does not already have its
2420 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2421 * brought online out of order, this can require re-organizing the
2422 * leader-follower relationships.
2424 static void rcu_spawn_one_nocb_kthread(struct rcu_state
*rsp
, int cpu
)
2426 struct rcu_data
*rdp
;
2427 struct rcu_data
*rdp_last
;
2428 struct rcu_data
*rdp_old_leader
;
2429 struct rcu_data
*rdp_spawn
= per_cpu_ptr(rsp
->rda
, cpu
);
2430 struct task_struct
*t
;
2433 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2434 * then nothing to do.
2436 if (!rcu_is_nocb_cpu(cpu
) || rdp_spawn
->nocb_kthread
)
2439 /* If we didn't spawn the leader first, reorganize! */
2440 rdp_old_leader
= rdp_spawn
->nocb_leader
;
2441 if (rdp_old_leader
!= rdp_spawn
&& !rdp_old_leader
->nocb_kthread
) {
2443 rdp
= rdp_old_leader
;
2445 rdp
->nocb_leader
= rdp_spawn
;
2446 if (rdp_last
&& rdp
!= rdp_spawn
)
2447 rdp_last
->nocb_next_follower
= rdp
;
2448 if (rdp
== rdp_spawn
) {
2449 rdp
= rdp
->nocb_next_follower
;
2452 rdp
= rdp
->nocb_next_follower
;
2453 rdp_last
->nocb_next_follower
= NULL
;
2456 rdp_spawn
->nocb_next_follower
= rdp_old_leader
;
2459 /* Spawn the kthread for this CPU and RCU flavor. */
2460 t
= kthread_run(rcu_nocb_kthread
, rdp_spawn
,
2461 "rcuo%c/%d", rsp
->abbr
, cpu
);
2463 ACCESS_ONCE(rdp_spawn
->nocb_kthread
) = t
;
2467 * If the specified CPU is a no-CBs CPU that does not already have its
2468 * rcuo kthreads, spawn them.
2470 static void rcu_spawn_all_nocb_kthreads(int cpu
)
2472 struct rcu_state
*rsp
;
2474 if (rcu_scheduler_fully_active
)
2475 for_each_rcu_flavor(rsp
)
2476 rcu_spawn_one_nocb_kthread(rsp
, cpu
);
2480 * Once the scheduler is running, spawn rcuo kthreads for all online
2481 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2482 * non-boot CPUs come online -- if this changes, we will need to add
2483 * some mutual exclusion.
2485 static void __init
rcu_spawn_nocb_kthreads(void)
2489 for_each_online_cpu(cpu
)
2490 rcu_spawn_all_nocb_kthreads(cpu
);
2493 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2494 static int rcu_nocb_leader_stride
= -1;
2495 module_param(rcu_nocb_leader_stride
, int, 0444);
2498 * Initialize leader-follower relationships for all no-CBs CPU.
2500 static void __init
rcu_organize_nocb_kthreads(struct rcu_state
*rsp
)
2503 int ls
= rcu_nocb_leader_stride
;
2504 int nl
= 0; /* Next leader. */
2505 struct rcu_data
*rdp
;
2506 struct rcu_data
*rdp_leader
= NULL
; /* Suppress misguided gcc warn. */
2507 struct rcu_data
*rdp_prev
= NULL
;
2509 if (!have_rcu_nocb_mask
)
2512 ls
= int_sqrt(nr_cpu_ids
);
2513 rcu_nocb_leader_stride
= ls
;
2517 * Each pass through this loop sets up one rcu_data structure and
2518 * spawns one rcu_nocb_kthread().
2520 for_each_cpu(cpu
, rcu_nocb_mask
) {
2521 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2522 if (rdp
->cpu
>= nl
) {
2523 /* New leader, set up for followers & next leader. */
2524 nl
= DIV_ROUND_UP(rdp
->cpu
+ 1, ls
) * ls
;
2525 rdp
->nocb_leader
= rdp
;
2528 /* Another follower, link to previous leader. */
2529 rdp
->nocb_leader
= rdp_leader
;
2530 rdp_prev
->nocb_next_follower
= rdp
;
2536 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2537 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2539 if (!rcu_is_nocb_cpu(rdp
->cpu
))
2542 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2546 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2548 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state
*rsp
, int cpu
)
2550 WARN_ON_ONCE(1); /* Should be dead code. */
2554 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
2558 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
2562 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2566 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2567 bool lazy
, unsigned long flags
)
2572 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
2573 struct rcu_data
*rdp
,
2574 unsigned long flags
)
2579 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2583 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2588 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2592 static void rcu_spawn_all_nocb_kthreads(int cpu
)
2596 static void __init
rcu_spawn_nocb_kthreads(void)
2600 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2605 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2608 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2609 * arbitrarily long period of time with the scheduling-clock tick turned
2610 * off. RCU will be paying attention to this CPU because it is in the
2611 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2612 * machine because the scheduling-clock tick has been disabled. Therefore,
2613 * if an adaptive-ticks CPU is failing to respond to the current grace
2614 * period and has not be idle from an RCU perspective, kick it.
2616 static void __maybe_unused
rcu_kick_nohz_cpu(int cpu
)
2618 #ifdef CONFIG_NO_HZ_FULL
2619 if (tick_nohz_full_cpu(cpu
))
2620 smp_send_reschedule(cpu
);
2621 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2625 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2627 static int full_sysidle_state
; /* Current system-idle state. */
2628 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2629 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2630 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2631 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2632 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2635 * Invoked to note exit from irq or task transition to idle. Note that
2636 * usermode execution does -not- count as idle here! After all, we want
2637 * to detect full-system idle states, not RCU quiescent states and grace
2638 * periods. The caller must have disabled interrupts.
2640 static void rcu_sysidle_enter(int irq
)
2643 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
2645 /* If there are no nohz_full= CPUs, no need to track this. */
2646 if (!tick_nohz_full_enabled())
2649 /* Adjust nesting, check for fully idle. */
2651 rdtp
->dynticks_idle_nesting
--;
2652 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2653 if (rdtp
->dynticks_idle_nesting
!= 0)
2654 return; /* Still not fully idle. */
2656 if ((rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) ==
2657 DYNTICK_TASK_NEST_VALUE
) {
2658 rdtp
->dynticks_idle_nesting
= 0;
2660 rdtp
->dynticks_idle_nesting
-= DYNTICK_TASK_NEST_VALUE
;
2661 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2662 return; /* Still not fully idle. */
2666 /* Record start of fully idle period. */
2668 ACCESS_ONCE(rdtp
->dynticks_idle_jiffies
) = j
;
2669 smp_mb__before_atomic();
2670 atomic_inc(&rdtp
->dynticks_idle
);
2671 smp_mb__after_atomic();
2672 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks_idle
) & 0x1);
2676 * Unconditionally force exit from full system-idle state. This is
2677 * invoked when a normal CPU exits idle, but must be called separately
2678 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2679 * is that the timekeeping CPU is permitted to take scheduling-clock
2680 * interrupts while the system is in system-idle state, and of course
2681 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2682 * interrupt from any other type of interrupt.
2684 void rcu_sysidle_force_exit(void)
2686 int oldstate
= ACCESS_ONCE(full_sysidle_state
);
2690 * Each pass through the following loop attempts to exit full
2691 * system-idle state. If contention proves to be a problem,
2692 * a trylock-based contention tree could be used here.
2694 while (oldstate
> RCU_SYSIDLE_SHORT
) {
2695 newoldstate
= cmpxchg(&full_sysidle_state
,
2696 oldstate
, RCU_SYSIDLE_NOT
);
2697 if (oldstate
== newoldstate
&&
2698 oldstate
== RCU_SYSIDLE_FULL_NOTED
) {
2699 rcu_kick_nohz_cpu(tick_do_timer_cpu
);
2700 return; /* We cleared it, done! */
2702 oldstate
= newoldstate
;
2704 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2708 * Invoked to note entry to irq or task transition from idle. Note that
2709 * usermode execution does -not- count as idle here! The caller must
2710 * have disabled interrupts.
2712 static void rcu_sysidle_exit(int irq
)
2714 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
2716 /* If there are no nohz_full= CPUs, no need to track this. */
2717 if (!tick_nohz_full_enabled())
2720 /* Adjust nesting, check for already non-idle. */
2722 rdtp
->dynticks_idle_nesting
++;
2723 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2724 if (rdtp
->dynticks_idle_nesting
!= 1)
2725 return; /* Already non-idle. */
2728 * Allow for irq misnesting. Yes, it really is possible
2729 * to enter an irq handler then never leave it, and maybe
2730 * also vice versa. Handle both possibilities.
2732 if (rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) {
2733 rdtp
->dynticks_idle_nesting
+= DYNTICK_TASK_NEST_VALUE
;
2734 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2735 return; /* Already non-idle. */
2737 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2741 /* Record end of idle period. */
2742 smp_mb__before_atomic();
2743 atomic_inc(&rdtp
->dynticks_idle
);
2744 smp_mb__after_atomic();
2745 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks_idle
) & 0x1));
2748 * If we are the timekeeping CPU, we are permitted to be non-idle
2749 * during a system-idle state. This must be the case, because
2750 * the timekeeping CPU has to take scheduling-clock interrupts
2751 * during the time that the system is transitioning to full
2752 * system-idle state. This means that the timekeeping CPU must
2753 * invoke rcu_sysidle_force_exit() directly if it does anything
2754 * more than take a scheduling-clock interrupt.
2756 if (smp_processor_id() == tick_do_timer_cpu
)
2759 /* Update system-idle state: We are clearly no longer fully idle! */
2760 rcu_sysidle_force_exit();
2764 * Check to see if the current CPU is idle. Note that usermode execution
2765 * does not count as idle. The caller must have disabled interrupts.
2767 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
2768 unsigned long *maxj
)
2772 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
2774 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2775 if (!tick_nohz_full_enabled())
2779 * If some other CPU has already reported non-idle, if this is
2780 * not the flavor of RCU that tracks sysidle state, or if this
2781 * is an offline or the timekeeping CPU, nothing to do.
2783 if (!*isidle
|| rdp
->rsp
!= rcu_state_p
||
2784 cpu_is_offline(rdp
->cpu
) || rdp
->cpu
== tick_do_timer_cpu
)
2786 if (rcu_gp_in_progress(rdp
->rsp
))
2787 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
);
2789 /* Pick up current idle and NMI-nesting counter and check. */
2790 cur
= atomic_read(&rdtp
->dynticks_idle
);
2792 *isidle
= false; /* We are not idle! */
2795 smp_mb(); /* Read counters before timestamps. */
2797 /* Pick up timestamps. */
2798 j
= ACCESS_ONCE(rdtp
->dynticks_idle_jiffies
);
2799 /* If this CPU entered idle more recently, update maxj timestamp. */
2800 if (ULONG_CMP_LT(*maxj
, j
))
2805 * Is this the flavor of RCU that is handling full-system idle?
2807 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
2809 return rsp
== rcu_state_p
;
2813 * Return a delay in jiffies based on the number of CPUs, rcu_node
2814 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2815 * systems more time to transition to full-idle state in order to
2816 * avoid the cache thrashing that otherwise occur on the state variable.
2817 * Really small systems (less than a couple of tens of CPUs) should
2818 * instead use a single global atomically incremented counter, and later
2819 * versions of this will automatically reconfigure themselves accordingly.
2821 static unsigned long rcu_sysidle_delay(void)
2823 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2825 return DIV_ROUND_UP(nr_cpu_ids
* HZ
, rcu_fanout_leaf
* 1000);
2829 * Advance the full-system-idle state. This is invoked when all of
2830 * the non-timekeeping CPUs are idle.
2832 static void rcu_sysidle(unsigned long j
)
2834 /* Check the current state. */
2835 switch (ACCESS_ONCE(full_sysidle_state
)) {
2836 case RCU_SYSIDLE_NOT
:
2838 /* First time all are idle, so note a short idle period. */
2839 ACCESS_ONCE(full_sysidle_state
) = RCU_SYSIDLE_SHORT
;
2842 case RCU_SYSIDLE_SHORT
:
2845 * Idle for a bit, time to advance to next state?
2846 * cmpxchg failure means race with non-idle, let them win.
2848 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2849 (void)cmpxchg(&full_sysidle_state
,
2850 RCU_SYSIDLE_SHORT
, RCU_SYSIDLE_LONG
);
2853 case RCU_SYSIDLE_LONG
:
2856 * Do an additional check pass before advancing to full.
2857 * cmpxchg failure means race with non-idle, let them win.
2859 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2860 (void)cmpxchg(&full_sysidle_state
,
2861 RCU_SYSIDLE_LONG
, RCU_SYSIDLE_FULL
);
2870 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2871 * back to the beginning.
2873 static void rcu_sysidle_cancel(void)
2876 if (full_sysidle_state
> RCU_SYSIDLE_SHORT
)
2877 ACCESS_ONCE(full_sysidle_state
) = RCU_SYSIDLE_NOT
;
2881 * Update the sysidle state based on the results of a force-quiescent-state
2882 * scan of the CPUs' dyntick-idle state.
2884 static void rcu_sysidle_report(struct rcu_state
*rsp
, int isidle
,
2885 unsigned long maxj
, bool gpkt
)
2887 if (rsp
!= rcu_state_p
)
2888 return; /* Wrong flavor, ignore. */
2889 if (gpkt
&& nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2890 return; /* Running state machine from timekeeping CPU. */
2892 rcu_sysidle(maxj
); /* More idle! */
2894 rcu_sysidle_cancel(); /* Idle is over. */
2898 * Wrapper for rcu_sysidle_report() when called from the grace-period
2899 * kthread's context.
2901 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
2904 /* If there are no nohz_full= CPUs, no need to track this. */
2905 if (!tick_nohz_full_enabled())
2908 rcu_sysidle_report(rsp
, isidle
, maxj
, true);
2911 /* Callback and function for forcing an RCU grace period. */
2912 struct rcu_sysidle_head
{
2917 static void rcu_sysidle_cb(struct rcu_head
*rhp
)
2919 struct rcu_sysidle_head
*rshp
;
2922 * The following memory barrier is needed to replace the
2923 * memory barriers that would normally be in the memory
2926 smp_mb(); /* grace period precedes setting inuse. */
2928 rshp
= container_of(rhp
, struct rcu_sysidle_head
, rh
);
2929 ACCESS_ONCE(rshp
->inuse
) = 0;
2933 * Check to see if the system is fully idle, other than the timekeeping CPU.
2934 * The caller must have disabled interrupts. This is not intended to be
2935 * called unless tick_nohz_full_enabled().
2937 bool rcu_sys_is_idle(void)
2939 static struct rcu_sysidle_head rsh
;
2940 int rss
= ACCESS_ONCE(full_sysidle_state
);
2942 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
))
2945 /* Handle small-system case by doing a full scan of CPUs. */
2946 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
) {
2947 int oldrss
= rss
- 1;
2950 * One pass to advance to each state up to _FULL.
2951 * Give up if any pass fails to advance the state.
2953 while (rss
< RCU_SYSIDLE_FULL
&& oldrss
< rss
) {
2956 unsigned long maxj
= jiffies
- ULONG_MAX
/ 4;
2957 struct rcu_data
*rdp
;
2959 /* Scan all the CPUs looking for nonidle CPUs. */
2960 for_each_possible_cpu(cpu
) {
2961 rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
2962 rcu_sysidle_check_cpu(rdp
, &isidle
, &maxj
);
2966 rcu_sysidle_report(rcu_state_p
, isidle
, maxj
, false);
2968 rss
= ACCESS_ONCE(full_sysidle_state
);
2972 /* If this is the first observation of an idle period, record it. */
2973 if (rss
== RCU_SYSIDLE_FULL
) {
2974 rss
= cmpxchg(&full_sysidle_state
,
2975 RCU_SYSIDLE_FULL
, RCU_SYSIDLE_FULL_NOTED
);
2976 return rss
== RCU_SYSIDLE_FULL
;
2979 smp_mb(); /* ensure rss load happens before later caller actions. */
2981 /* If already fully idle, tell the caller (in case of races). */
2982 if (rss
== RCU_SYSIDLE_FULL_NOTED
)
2986 * If we aren't there yet, and a grace period is not in flight,
2987 * initiate a grace period. Either way, tell the caller that
2988 * we are not there yet. We use an xchg() rather than an assignment
2989 * to make up for the memory barriers that would otherwise be
2990 * provided by the memory allocator.
2992 if (nr_cpu_ids
> CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
&&
2993 !rcu_gp_in_progress(rcu_state_p
) &&
2994 !rsh
.inuse
&& xchg(&rsh
.inuse
, 1) == 0)
2995 call_rcu(&rsh
.rh
, rcu_sysidle_cb
);
3000 * Initialize dynticks sysidle state for CPUs coming online.
3002 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
3004 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
;
3007 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3009 static void rcu_sysidle_enter(int irq
)
3013 static void rcu_sysidle_exit(int irq
)
3017 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
3018 unsigned long *maxj
)
3022 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
3027 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
3032 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
3036 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3039 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3040 * grace-period kthread will do force_quiescent_state() processing?
3041 * The idea is to avoid waking up RCU core processing on such a
3042 * CPU unless the grace period has extended for too long.
3044 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3045 * CONFIG_RCU_NOCB_CPU CPUs.
3047 static bool rcu_nohz_full_cpu(struct rcu_state
*rsp
)
3049 #ifdef CONFIG_NO_HZ_FULL
3050 if (tick_nohz_full_cpu(smp_processor_id()) &&
3051 (!rcu_gp_in_progress(rsp
) ||
3052 ULONG_CMP_LT(jiffies
, ACCESS_ONCE(rsp
->gp_start
) + HZ
)))
3054 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3059 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3062 static void rcu_bind_gp_kthread(void)
3064 int __maybe_unused cpu
;
3066 if (!tick_nohz_full_enabled())
3068 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3069 cpu
= tick_do_timer_cpu
;
3070 if (cpu
>= 0 && cpu
< nr_cpu_ids
&& raw_smp_processor_id() != cpu
)
3071 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
3072 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3073 if (!is_housekeeping_cpu(raw_smp_processor_id()))
3074 housekeeping_affine(current
);
3075 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3078 /* Record the current task on dyntick-idle entry. */
3079 static void rcu_dynticks_task_enter(void)
3081 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3082 ACCESS_ONCE(current
->rcu_tasks_idle_cpu
) = smp_processor_id();
3083 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3086 /* Record no current task on dyntick-idle exit. */
3087 static void rcu_dynticks_task_exit(void)
3089 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3090 ACCESS_ONCE(current
->rcu_tasks_idle_cpu
) = -1;
3091 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */