2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
31 #include "../time/tick-internal.h"
33 #ifdef CONFIG_RCU_BOOST
35 #include "../locking/rtmutex_common.h"
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
41 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
42 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
43 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
44 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
46 #else /* #ifdef CONFIG_RCU_BOOST */
49 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
50 * all uses are in dead code. Provide a definition to keep the compiler
51 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
52 * This probably needs to be excluded from -rt builds.
54 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
56 #endif /* #else #ifdef CONFIG_RCU_BOOST */
58 #ifdef CONFIG_RCU_NOCB_CPU
59 static cpumask_var_t rcu_nocb_mask
; /* CPUs to have callbacks offloaded. */
60 static bool have_rcu_nocb_mask
; /* Was rcu_nocb_mask allocated? */
61 static bool __read_mostly rcu_nocb_poll
; /* Offload kthread are to poll. */
62 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
65 * Check the RCU kernel configuration parameters and print informative
66 * messages about anything out of the ordinary. If you like #ifdef, you
67 * will love this function.
69 static void __init
rcu_bootup_announce_oddness(void)
71 if (IS_ENABLED(CONFIG_RCU_TRACE
))
72 pr_info("\tRCU debugfs-based tracing is enabled.\n");
73 if ((IS_ENABLED(CONFIG_64BIT
) && RCU_FANOUT
!= 64) ||
74 (!IS_ENABLED(CONFIG_64BIT
) && RCU_FANOUT
!= 32))
75 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
78 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
79 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ
))
80 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
81 if (IS_ENABLED(CONFIG_PROVE_RCU
))
82 pr_info("\tRCU lockdep checking is enabled.\n");
83 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE
))
84 pr_info("\tRCU torture testing starts during boot.\n");
85 if (RCU_NUM_LVLS
>= 4)
86 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
87 if (RCU_FANOUT_LEAF
!= 16)
88 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
90 if (rcu_fanout_leaf
!= RCU_FANOUT_LEAF
)
91 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf
);
92 if (nr_cpu_ids
!= NR_CPUS
)
93 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS
, nr_cpu_ids
);
94 if (IS_ENABLED(CONFIG_RCU_BOOST
))
95 pr_info("\tRCU kthread priority: %d.\n", kthread_prio
);
98 #ifdef CONFIG_PREEMPT_RCU
100 RCU_STATE_INITIALIZER(rcu_preempt
, 'p', call_rcu
);
101 static struct rcu_state
*const rcu_state_p
= &rcu_preempt_state
;
102 static struct rcu_data __percpu
*const rcu_data_p
= &rcu_preempt_data
;
104 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
);
105 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
109 * Tell them what RCU they are running.
111 static void __init
rcu_bootup_announce(void)
113 pr_info("Preemptible hierarchical RCU implementation.\n");
114 rcu_bootup_announce_oddness();
118 * Record a preemptible-RCU quiescent state for the specified CPU. Note
119 * that this just means that the task currently running on the CPU is
120 * not in a quiescent state. There might be any number of tasks blocked
121 * while in an RCU read-side critical section.
123 * As with the other rcu_*_qs() functions, callers to this function
124 * must disable preemption.
126 static void rcu_preempt_qs(void)
128 if (!__this_cpu_read(rcu_data_p
->passed_quiesce
)) {
129 trace_rcu_grace_period(TPS("rcu_preempt"),
130 __this_cpu_read(rcu_data_p
->gpnum
),
132 __this_cpu_write(rcu_data_p
->passed_quiesce
, 1);
133 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
134 current
->rcu_read_unlock_special
.b
.need_qs
= false;
139 * We have entered the scheduler, and the current task might soon be
140 * context-switched away from. If this task is in an RCU read-side
141 * critical section, we will no longer be able to rely on the CPU to
142 * record that fact, so we enqueue the task on the blkd_tasks list.
143 * The task will dequeue itself when it exits the outermost enclosing
144 * RCU read-side critical section. Therefore, the current grace period
145 * cannot be permitted to complete until the blkd_tasks list entries
146 * predating the current grace period drain, in other words, until
147 * rnp->gp_tasks becomes NULL.
149 * Caller must disable preemption.
151 static void rcu_preempt_note_context_switch(void)
153 struct task_struct
*t
= current
;
155 struct rcu_data
*rdp
;
156 struct rcu_node
*rnp
;
158 if (t
->rcu_read_lock_nesting
> 0 &&
159 !t
->rcu_read_unlock_special
.b
.blocked
) {
161 /* Possibly blocking in an RCU read-side critical section. */
162 rdp
= this_cpu_ptr(rcu_state_p
->rda
);
164 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
165 smp_mb__after_unlock_lock();
166 t
->rcu_read_unlock_special
.b
.blocked
= true;
167 t
->rcu_blocked_node
= rnp
;
170 * If this CPU has already checked in, then this task
171 * will hold up the next grace period rather than the
172 * current grace period. Queue the task accordingly.
173 * If the task is queued for the current grace period
174 * (i.e., this CPU has not yet passed through a quiescent
175 * state for the current grace period), then as long
176 * as that task remains queued, the current grace period
177 * cannot end. Note that there is some uncertainty as
178 * to exactly when the current grace period started.
179 * We take a conservative approach, which can result
180 * in unnecessarily waiting on tasks that started very
181 * slightly after the current grace period began. C'est
184 * But first, note that the current CPU must still be
187 WARN_ON_ONCE((rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) == 0);
188 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
189 if ((rnp
->qsmask
& rdp
->grpmask
) && rnp
->gp_tasks
!= NULL
) {
190 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
->prev
);
191 rnp
->gp_tasks
= &t
->rcu_node_entry
;
192 if (IS_ENABLED(CONFIG_RCU_BOOST
) &&
193 rnp
->boost_tasks
!= NULL
)
194 rnp
->boost_tasks
= rnp
->gp_tasks
;
196 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
197 if (rnp
->qsmask
& rdp
->grpmask
)
198 rnp
->gp_tasks
= &t
->rcu_node_entry
;
200 trace_rcu_preempt_task(rdp
->rsp
->name
,
202 (rnp
->qsmask
& rdp
->grpmask
)
205 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
206 } else if (t
->rcu_read_lock_nesting
< 0 &&
207 t
->rcu_read_unlock_special
.s
) {
210 * Complete exit from RCU read-side critical section on
211 * behalf of preempted instance of __rcu_read_unlock().
213 rcu_read_unlock_special(t
);
217 * Either we were not in an RCU read-side critical section to
218 * begin with, or we have now recorded that critical section
219 * globally. Either way, we can now note a quiescent state
220 * for this CPU. Again, if we were in an RCU read-side critical
221 * section, and if that critical section was blocking the current
222 * grace period, then the fact that the task has been enqueued
223 * means that we continue to block the current grace period.
229 * Check for preempted RCU readers blocking the current grace period
230 * for the specified rcu_node structure. If the caller needs a reliable
231 * answer, it must hold the rcu_node's ->lock.
233 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
235 return rnp
->gp_tasks
!= NULL
;
239 * Advance a ->blkd_tasks-list pointer to the next entry, instead
240 * returning NULL if at the end of the list.
242 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
243 struct rcu_node
*rnp
)
245 struct list_head
*np
;
247 np
= t
->rcu_node_entry
.next
;
248 if (np
== &rnp
->blkd_tasks
)
254 * Return true if the specified rcu_node structure has tasks that were
255 * preempted within an RCU read-side critical section.
257 static bool rcu_preempt_has_tasks(struct rcu_node
*rnp
)
259 return !list_empty(&rnp
->blkd_tasks
);
263 * Handle special cases during rcu_read_unlock(), such as needing to
264 * notify RCU core processing or task having blocked during the RCU
265 * read-side critical section.
267 void rcu_read_unlock_special(struct task_struct
*t
)
273 struct list_head
*np
;
274 bool drop_boost_mutex
= false;
275 struct rcu_node
*rnp
;
276 union rcu_special special
;
278 /* NMI handlers cannot block and cannot safely manipulate state. */
282 local_irq_save(flags
);
285 * If RCU core is waiting for this CPU to exit critical section,
286 * let it know that we have done so. Because irqs are disabled,
287 * t->rcu_read_unlock_special cannot change.
289 special
= t
->rcu_read_unlock_special
;
290 if (special
.b
.need_qs
) {
292 t
->rcu_read_unlock_special
.b
.need_qs
= false;
293 if (!t
->rcu_read_unlock_special
.s
) {
294 local_irq_restore(flags
);
299 /* Hardware IRQ handlers cannot block, complain if they get here. */
300 if (in_irq() || in_serving_softirq()) {
301 lockdep_rcu_suspicious(__FILE__
, __LINE__
,
302 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
303 pr_alert("->rcu_read_unlock_special: %#x (b: %d, nq: %d)\n",
304 t
->rcu_read_unlock_special
.s
,
305 t
->rcu_read_unlock_special
.b
.blocked
,
306 t
->rcu_read_unlock_special
.b
.need_qs
);
307 local_irq_restore(flags
);
311 /* Clean up if blocked during RCU read-side critical section. */
312 if (special
.b
.blocked
) {
313 t
->rcu_read_unlock_special
.b
.blocked
= false;
316 * Remove this task from the list it blocked on. The task
317 * now remains queued on the rcu_node corresponding to
318 * the CPU it first blocked on, so the first attempt to
319 * acquire the task's rcu_node's ->lock will succeed.
320 * Keep the loop and add a WARN_ON() out of sheer paranoia.
323 rnp
= t
->rcu_blocked_node
;
324 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
325 smp_mb__after_unlock_lock();
326 if (rnp
== t
->rcu_blocked_node
)
329 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
331 empty_norm
= !rcu_preempt_blocked_readers_cgp(rnp
);
332 empty_exp
= !rcu_preempted_readers_exp(rnp
);
333 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
334 np
= rcu_next_node_entry(t
, rnp
);
335 list_del_init(&t
->rcu_node_entry
);
336 t
->rcu_blocked_node
= NULL
;
337 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
339 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
341 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
343 if (IS_ENABLED(CONFIG_RCU_BOOST
)) {
344 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
345 rnp
->boost_tasks
= np
;
346 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
347 drop_boost_mutex
= rt_mutex_owner(&rnp
->boost_mtx
) == t
;
351 * If this was the last task on the current list, and if
352 * we aren't waiting on any CPUs, report the quiescent state.
353 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
354 * so we must take a snapshot of the expedited state.
356 empty_exp_now
= !rcu_preempted_readers_exp(rnp
);
357 if (!empty_norm
&& !rcu_preempt_blocked_readers_cgp(rnp
)) {
358 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
365 rcu_report_unblock_qs_rnp(rcu_state_p
, rnp
, flags
);
367 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
370 /* Unboost if we were boosted. */
371 if (IS_ENABLED(CONFIG_RCU_BOOST
) && drop_boost_mutex
)
372 rt_mutex_unlock(&rnp
->boost_mtx
);
375 * If this was the last task on the expedited lists,
376 * then we need to report up the rcu_node hierarchy.
378 if (!empty_exp
&& empty_exp_now
)
379 rcu_report_exp_rnp(rcu_state_p
, rnp
, true);
381 local_irq_restore(flags
);
386 * Dump detailed information for all tasks blocking the current RCU
387 * grace period on the specified rcu_node structure.
389 static void rcu_print_detail_task_stall_rnp(struct rcu_node
*rnp
)
392 struct task_struct
*t
;
394 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
395 if (!rcu_preempt_blocked_readers_cgp(rnp
)) {
396 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
399 t
= list_entry(rnp
->gp_tasks
->prev
,
400 struct task_struct
, rcu_node_entry
);
401 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
403 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
407 * Dump detailed information for all tasks blocking the current RCU
410 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
412 struct rcu_node
*rnp
= rcu_get_root(rsp
);
414 rcu_print_detail_task_stall_rnp(rnp
);
415 rcu_for_each_leaf_node(rsp
, rnp
)
416 rcu_print_detail_task_stall_rnp(rnp
);
419 static void rcu_print_task_stall_begin(struct rcu_node
*rnp
)
421 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
422 rnp
->level
, rnp
->grplo
, rnp
->grphi
);
425 static void rcu_print_task_stall_end(void)
431 * Scan the current list of tasks blocked within RCU read-side critical
432 * sections, printing out the tid of each.
434 static int rcu_print_task_stall(struct rcu_node
*rnp
)
436 struct task_struct
*t
;
439 if (!rcu_preempt_blocked_readers_cgp(rnp
))
441 rcu_print_task_stall_begin(rnp
);
442 t
= list_entry(rnp
->gp_tasks
->prev
,
443 struct task_struct
, rcu_node_entry
);
444 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
) {
445 pr_cont(" P%d", t
->pid
);
448 rcu_print_task_stall_end();
453 * Check that the list of blocked tasks for the newly completed grace
454 * period is in fact empty. It is a serious bug to complete a grace
455 * period that still has RCU readers blocked! This function must be
456 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
457 * must be held by the caller.
459 * Also, if there are blocked tasks on the list, they automatically
460 * block the newly created grace period, so set up ->gp_tasks accordingly.
462 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
464 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
465 if (rcu_preempt_has_tasks(rnp
))
466 rnp
->gp_tasks
= rnp
->blkd_tasks
.next
;
467 WARN_ON_ONCE(rnp
->qsmask
);
471 * Check for a quiescent state from the current CPU. When a task blocks,
472 * the task is recorded in the corresponding CPU's rcu_node structure,
473 * which is checked elsewhere.
475 * Caller must disable hard irqs.
477 static void rcu_preempt_check_callbacks(void)
479 struct task_struct
*t
= current
;
481 if (t
->rcu_read_lock_nesting
== 0) {
485 if (t
->rcu_read_lock_nesting
> 0 &&
486 __this_cpu_read(rcu_data_p
->qs_pending
) &&
487 !__this_cpu_read(rcu_data_p
->passed_quiesce
))
488 t
->rcu_read_unlock_special
.b
.need_qs
= true;
491 #ifdef CONFIG_RCU_BOOST
493 static void rcu_preempt_do_callbacks(void)
495 rcu_do_batch(rcu_state_p
, this_cpu_ptr(rcu_data_p
));
498 #endif /* #ifdef CONFIG_RCU_BOOST */
501 * Queue a preemptible-RCU callback for invocation after a grace period.
503 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
505 __call_rcu(head
, func
, rcu_state_p
, -1, 0);
507 EXPORT_SYMBOL_GPL(call_rcu
);
510 * synchronize_rcu - wait until a grace period has elapsed.
512 * Control will return to the caller some time after a full grace
513 * period has elapsed, in other words after all currently executing RCU
514 * read-side critical sections have completed. Note, however, that
515 * upon return from synchronize_rcu(), the caller might well be executing
516 * concurrently with new RCU read-side critical sections that began while
517 * synchronize_rcu() was waiting. RCU read-side critical sections are
518 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
520 * See the description of synchronize_sched() for more detailed information
521 * on memory ordering guarantees.
523 void synchronize_rcu(void)
525 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
526 lock_is_held(&rcu_lock_map
) ||
527 lock_is_held(&rcu_sched_lock_map
),
528 "Illegal synchronize_rcu() in RCU read-side critical section");
529 if (!rcu_scheduler_active
)
531 if (rcu_gp_is_expedited())
532 synchronize_rcu_expedited();
534 wait_rcu_gp(call_rcu
);
536 EXPORT_SYMBOL_GPL(synchronize_rcu
);
538 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq
);
541 * Return non-zero if there are any tasks in RCU read-side critical
542 * sections blocking the current preemptible-RCU expedited grace period.
543 * If there is no preemptible-RCU expedited grace period currently in
544 * progress, returns zero unconditionally.
546 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
)
548 return rnp
->exp_tasks
!= NULL
;
552 * return non-zero if there is no RCU expedited grace period in progress
553 * for the specified rcu_node structure, in other words, if all CPUs and
554 * tasks covered by the specified rcu_node structure have done their bit
555 * for the current expedited grace period. Works only for preemptible
556 * RCU -- other RCU implementation use other means.
558 * Caller must hold the root rcu_node's exp_funnel_mutex.
560 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
562 return !rcu_preempted_readers_exp(rnp
) &&
563 READ_ONCE(rnp
->expmask
) == 0;
567 * Report the exit from RCU read-side critical section for the last task
568 * that queued itself during or before the current expedited preemptible-RCU
569 * grace period. This event is reported either to the rcu_node structure on
570 * which the task was queued or to one of that rcu_node structure's ancestors,
571 * recursively up the tree. (Calm down, calm down, we do the recursion
574 * Caller must hold the root rcu_node's exp_funnel_mutex.
576 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
582 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
583 smp_mb__after_unlock_lock();
585 if (!sync_rcu_preempt_exp_done(rnp
)) {
586 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
589 if (rnp
->parent
== NULL
) {
590 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
592 smp_mb(); /* EGP done before wake_up(). */
593 wake_up(&sync_rcu_preempt_exp_wq
);
598 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
600 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
601 smp_mb__after_unlock_lock();
602 rnp
->expmask
&= ~mask
;
607 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
608 * grace period for the specified rcu_node structure, phase 1. If there
609 * are such tasks, set the ->expmask bits up the rcu_node tree and also
610 * set the ->expmask bits on the leaf rcu_node structures to tell phase 2
611 * that work is needed here.
613 * Caller must hold the root rcu_node's exp_funnel_mutex.
616 sync_rcu_preempt_exp_init1(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
620 struct rcu_node
*rnp_up
;
622 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
623 smp_mb__after_unlock_lock();
624 WARN_ON_ONCE(rnp
->expmask
);
625 WARN_ON_ONCE(rnp
->exp_tasks
);
626 if (!rcu_preempt_has_tasks(rnp
)) {
627 /* No blocked tasks, nothing to do. */
628 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
631 /* Call for Phase 2 and propagate ->expmask bits up the tree. */
634 while (rnp_up
->parent
) {
635 mask
= rnp_up
->grpmask
;
636 rnp_up
= rnp_up
->parent
;
637 if (rnp_up
->expmask
& mask
)
639 raw_spin_lock(&rnp_up
->lock
); /* irqs already off */
640 smp_mb__after_unlock_lock();
641 rnp_up
->expmask
|= mask
;
642 raw_spin_unlock(&rnp_up
->lock
); /* irqs still off */
644 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
648 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
649 * grace period for the specified rcu_node structure, phase 2. If the
650 * leaf rcu_node structure has its ->expmask field set, check for tasks.
651 * If there are some, clear ->expmask and set ->exp_tasks accordingly,
652 * then initiate RCU priority boosting. Otherwise, clear ->expmask and
653 * invoke rcu_report_exp_rnp() to clear out the upper-level ->expmask bits,
654 * enabling rcu_read_unlock_special() to do the bit-clearing.
656 * Caller must hold the root rcu_node's exp_funnel_mutex.
659 sync_rcu_preempt_exp_init2(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
663 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
664 smp_mb__after_unlock_lock();
666 /* Phase 1 didn't do anything, so Phase 2 doesn't either. */
667 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
671 /* Phase 1 is over. */
675 * If there are still blocked tasks, set up ->exp_tasks so that
676 * rcu_read_unlock_special() will wake us and then boost them.
678 if (rcu_preempt_has_tasks(rnp
)) {
679 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
680 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
684 /* No longer any blocked tasks, so undo bit setting. */
685 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
686 rcu_report_exp_rnp(rsp
, rnp
, false);
690 * synchronize_rcu_expedited - Brute-force RCU grace period
692 * Wait for an RCU-preempt grace period, but expedite it. The basic
693 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
694 * the ->blkd_tasks lists and wait for this list to drain. This consumes
695 * significant time on all CPUs and is unfriendly to real-time workloads,
696 * so is thus not recommended for any sort of common-case code.
697 * In fact, if you are using synchronize_rcu_expedited() in a loop,
698 * please restructure your code to batch your updates, and then Use a
699 * single synchronize_rcu() instead.
701 void synchronize_rcu_expedited(void)
703 struct rcu_node
*rnp
;
704 struct rcu_node
*rnp_unlock
;
705 struct rcu_state
*rsp
= rcu_state_p
;
708 s
= rcu_exp_gp_seq_snap(rsp
);
710 rnp_unlock
= exp_funnel_lock(rsp
, s
);
711 if (rnp_unlock
== NULL
)
712 return; /* Someone else did our work for us. */
714 rcu_exp_gp_seq_start(rsp
);
716 /* force all RCU readers onto ->blkd_tasks lists. */
717 synchronize_sched_expedited();
720 * Snapshot current state of ->blkd_tasks lists into ->expmask.
721 * Phase 1 sets bits and phase 2 permits rcu_read_unlock_special()
722 * to start clearing them. Doing this in one phase leads to
723 * strange races between setting and clearing bits, so just say "no"!
725 rcu_for_each_leaf_node(rsp
, rnp
)
726 sync_rcu_preempt_exp_init1(rsp
, rnp
);
727 rcu_for_each_leaf_node(rsp
, rnp
)
728 sync_rcu_preempt_exp_init2(rsp
, rnp
);
730 /* Wait for snapshotted ->blkd_tasks lists to drain. */
731 rnp
= rcu_get_root(rsp
);
732 wait_event(sync_rcu_preempt_exp_wq
,
733 sync_rcu_preempt_exp_done(rnp
));
735 /* Clean up and exit. */
736 rcu_exp_gp_seq_end(rsp
);
737 mutex_unlock(&rnp_unlock
->exp_funnel_mutex
);
739 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
742 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
744 * Note that this primitive does not necessarily wait for an RCU grace period
745 * to complete. For example, if there are no RCU callbacks queued anywhere
746 * in the system, then rcu_barrier() is within its rights to return
747 * immediately, without waiting for anything, much less an RCU grace period.
749 void rcu_barrier(void)
751 _rcu_barrier(rcu_state_p
);
753 EXPORT_SYMBOL_GPL(rcu_barrier
);
756 * Initialize preemptible RCU's state structures.
758 static void __init
__rcu_init_preempt(void)
760 rcu_init_one(rcu_state_p
, rcu_data_p
);
764 * Check for a task exiting while in a preemptible-RCU read-side
765 * critical section, clean up if so. No need to issue warnings,
766 * as debug_check_no_locks_held() already does this if lockdep
771 struct task_struct
*t
= current
;
773 if (likely(list_empty(¤t
->rcu_node_entry
)))
775 t
->rcu_read_lock_nesting
= 1;
777 t
->rcu_read_unlock_special
.b
.blocked
= true;
781 #else /* #ifdef CONFIG_PREEMPT_RCU */
783 static struct rcu_state
*const rcu_state_p
= &rcu_sched_state
;
784 static struct rcu_data __percpu
*const rcu_data_p
= &rcu_sched_data
;
787 * Tell them what RCU they are running.
789 static void __init
rcu_bootup_announce(void)
791 pr_info("Hierarchical RCU implementation.\n");
792 rcu_bootup_announce_oddness();
796 * Because preemptible RCU does not exist, we never have to check for
797 * CPUs being in quiescent states.
799 static void rcu_preempt_note_context_switch(void)
804 * Because preemptible RCU does not exist, there are never any preempted
807 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
813 * Because there is no preemptible RCU, there can be no readers blocked.
815 static bool rcu_preempt_has_tasks(struct rcu_node
*rnp
)
821 * Because preemptible RCU does not exist, we never have to check for
822 * tasks blocked within RCU read-side critical sections.
824 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
829 * Because preemptible RCU does not exist, we never have to check for
830 * tasks blocked within RCU read-side critical sections.
832 static int rcu_print_task_stall(struct rcu_node
*rnp
)
838 * Because there is no preemptible RCU, there can be no readers blocked,
839 * so there is no need to check for blocked tasks. So check only for
840 * bogus qsmask values.
842 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
844 WARN_ON_ONCE(rnp
->qsmask
);
848 * Because preemptible RCU does not exist, it never has any callbacks
851 static void rcu_preempt_check_callbacks(void)
856 * Wait for an rcu-preempt grace period, but make it happen quickly.
857 * But because preemptible RCU does not exist, map to rcu-sched.
859 void synchronize_rcu_expedited(void)
861 synchronize_sched_expedited();
863 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
866 * Because preemptible RCU does not exist, rcu_barrier() is just
867 * another name for rcu_barrier_sched().
869 void rcu_barrier(void)
873 EXPORT_SYMBOL_GPL(rcu_barrier
);
876 * Because preemptible RCU does not exist, it need not be initialized.
878 static void __init
__rcu_init_preempt(void)
883 * Because preemptible RCU does not exist, tasks cannot possibly exit
884 * while in preemptible RCU read-side critical sections.
890 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
892 #ifdef CONFIG_RCU_BOOST
894 #include "../locking/rtmutex_common.h"
896 #ifdef CONFIG_RCU_TRACE
898 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
900 if (!rcu_preempt_has_tasks(rnp
))
901 rnp
->n_balk_blkd_tasks
++;
902 else if (rnp
->exp_tasks
== NULL
&& rnp
->gp_tasks
== NULL
)
903 rnp
->n_balk_exp_gp_tasks
++;
904 else if (rnp
->gp_tasks
!= NULL
&& rnp
->boost_tasks
!= NULL
)
905 rnp
->n_balk_boost_tasks
++;
906 else if (rnp
->gp_tasks
!= NULL
&& rnp
->qsmask
!= 0)
907 rnp
->n_balk_notblocked
++;
908 else if (rnp
->gp_tasks
!= NULL
&&
909 ULONG_CMP_LT(jiffies
, rnp
->boost_time
))
910 rnp
->n_balk_notyet
++;
915 #else /* #ifdef CONFIG_RCU_TRACE */
917 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
921 #endif /* #else #ifdef CONFIG_RCU_TRACE */
923 static void rcu_wake_cond(struct task_struct
*t
, int status
)
926 * If the thread is yielding, only wake it when this
927 * is invoked from idle
929 if (status
!= RCU_KTHREAD_YIELDING
|| is_idle_task(current
))
934 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
935 * or ->boost_tasks, advancing the pointer to the next task in the
938 * Note that irqs must be enabled: boosting the task can block.
939 * Returns 1 if there are more tasks needing to be boosted.
941 static int rcu_boost(struct rcu_node
*rnp
)
944 struct task_struct
*t
;
945 struct list_head
*tb
;
947 if (READ_ONCE(rnp
->exp_tasks
) == NULL
&&
948 READ_ONCE(rnp
->boost_tasks
) == NULL
)
949 return 0; /* Nothing left to boost. */
951 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
952 smp_mb__after_unlock_lock();
955 * Recheck under the lock: all tasks in need of boosting
956 * might exit their RCU read-side critical sections on their own.
958 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
959 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
964 * Preferentially boost tasks blocking expedited grace periods.
965 * This cannot starve the normal grace periods because a second
966 * expedited grace period must boost all blocked tasks, including
967 * those blocking the pre-existing normal grace period.
969 if (rnp
->exp_tasks
!= NULL
) {
973 tb
= rnp
->boost_tasks
;
974 rnp
->n_normal_boosts
++;
976 rnp
->n_tasks_boosted
++;
979 * We boost task t by manufacturing an rt_mutex that appears to
980 * be held by task t. We leave a pointer to that rt_mutex where
981 * task t can find it, and task t will release the mutex when it
982 * exits its outermost RCU read-side critical section. Then
983 * simply acquiring this artificial rt_mutex will boost task
984 * t's priority. (Thanks to tglx for suggesting this approach!)
986 * Note that task t must acquire rnp->lock to remove itself from
987 * the ->blkd_tasks list, which it will do from exit() if from
988 * nowhere else. We therefore are guaranteed that task t will
989 * stay around at least until we drop rnp->lock. Note that
990 * rnp->lock also resolves races between our priority boosting
991 * and task t's exiting its outermost RCU read-side critical
994 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
995 rt_mutex_init_proxy_locked(&rnp
->boost_mtx
, t
);
996 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
997 /* Lock only for side effect: boosts task t's priority. */
998 rt_mutex_lock(&rnp
->boost_mtx
);
999 rt_mutex_unlock(&rnp
->boost_mtx
); /* Then keep lockdep happy. */
1001 return READ_ONCE(rnp
->exp_tasks
) != NULL
||
1002 READ_ONCE(rnp
->boost_tasks
) != NULL
;
1006 * Priority-boosting kthread, one per leaf rcu_node.
1008 static int rcu_boost_kthread(void *arg
)
1010 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1014 trace_rcu_utilization(TPS("Start boost kthread@init"));
1016 rnp
->boost_kthread_status
= RCU_KTHREAD_WAITING
;
1017 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1018 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1019 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1020 rnp
->boost_kthread_status
= RCU_KTHREAD_RUNNING
;
1021 more2boost
= rcu_boost(rnp
);
1027 rnp
->boost_kthread_status
= RCU_KTHREAD_YIELDING
;
1028 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1029 schedule_timeout_interruptible(2);
1030 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1035 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1040 * Check to see if it is time to start boosting RCU readers that are
1041 * blocking the current grace period, and, if so, tell the per-rcu_node
1042 * kthread to start boosting them. If there is an expedited grace
1043 * period in progress, it is always time to boost.
1045 * The caller must hold rnp->lock, which this function releases.
1046 * The ->boost_kthread_task is immortal, so we don't need to worry
1047 * about it going away.
1049 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1050 __releases(rnp
->lock
)
1052 struct task_struct
*t
;
1054 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1055 rnp
->n_balk_exp_gp_tasks
++;
1056 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1059 if (rnp
->exp_tasks
!= NULL
||
1060 (rnp
->gp_tasks
!= NULL
&&
1061 rnp
->boost_tasks
== NULL
&&
1063 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))) {
1064 if (rnp
->exp_tasks
== NULL
)
1065 rnp
->boost_tasks
= rnp
->gp_tasks
;
1066 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1067 t
= rnp
->boost_kthread_task
;
1069 rcu_wake_cond(t
, rnp
->boost_kthread_status
);
1071 rcu_initiate_boost_trace(rnp
);
1072 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1077 * Wake up the per-CPU kthread to invoke RCU callbacks.
1079 static void invoke_rcu_callbacks_kthread(void)
1081 unsigned long flags
;
1083 local_irq_save(flags
);
1084 __this_cpu_write(rcu_cpu_has_work
, 1);
1085 if (__this_cpu_read(rcu_cpu_kthread_task
) != NULL
&&
1086 current
!= __this_cpu_read(rcu_cpu_kthread_task
)) {
1087 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task
),
1088 __this_cpu_read(rcu_cpu_kthread_status
));
1090 local_irq_restore(flags
);
1094 * Is the current CPU running the RCU-callbacks kthread?
1095 * Caller must have preemption disabled.
1097 static bool rcu_is_callbacks_kthread(void)
1099 return __this_cpu_read(rcu_cpu_kthread_task
) == current
;
1102 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1105 * Do priority-boost accounting for the start of a new grace period.
1107 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1109 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1113 * Create an RCU-boost kthread for the specified node if one does not
1114 * already exist. We only create this kthread for preemptible RCU.
1115 * Returns zero if all is well, a negated errno otherwise.
1117 static int rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1118 struct rcu_node
*rnp
)
1120 int rnp_index
= rnp
- &rsp
->node
[0];
1121 unsigned long flags
;
1122 struct sched_param sp
;
1123 struct task_struct
*t
;
1125 if (rcu_state_p
!= rsp
)
1128 if (!rcu_scheduler_fully_active
|| rcu_rnp_online_cpus(rnp
) == 0)
1132 if (rnp
->boost_kthread_task
!= NULL
)
1134 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1135 "rcub/%d", rnp_index
);
1138 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1139 smp_mb__after_unlock_lock();
1140 rnp
->boost_kthread_task
= t
;
1141 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1142 sp
.sched_priority
= kthread_prio
;
1143 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1144 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1148 static void rcu_kthread_do_work(void)
1150 rcu_do_batch(&rcu_sched_state
, this_cpu_ptr(&rcu_sched_data
));
1151 rcu_do_batch(&rcu_bh_state
, this_cpu_ptr(&rcu_bh_data
));
1152 rcu_preempt_do_callbacks();
1155 static void rcu_cpu_kthread_setup(unsigned int cpu
)
1157 struct sched_param sp
;
1159 sp
.sched_priority
= kthread_prio
;
1160 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1163 static void rcu_cpu_kthread_park(unsigned int cpu
)
1165 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1168 static int rcu_cpu_kthread_should_run(unsigned int cpu
)
1170 return __this_cpu_read(rcu_cpu_has_work
);
1174 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1175 * RCU softirq used in flavors and configurations of RCU that do not
1176 * support RCU priority boosting.
1178 static void rcu_cpu_kthread(unsigned int cpu
)
1180 unsigned int *statusp
= this_cpu_ptr(&rcu_cpu_kthread_status
);
1181 char work
, *workp
= this_cpu_ptr(&rcu_cpu_has_work
);
1184 for (spincnt
= 0; spincnt
< 10; spincnt
++) {
1185 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1187 *statusp
= RCU_KTHREAD_RUNNING
;
1188 this_cpu_inc(rcu_cpu_kthread_loops
);
1189 local_irq_disable();
1194 rcu_kthread_do_work();
1197 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1198 *statusp
= RCU_KTHREAD_WAITING
;
1202 *statusp
= RCU_KTHREAD_YIELDING
;
1203 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1204 schedule_timeout_interruptible(2);
1205 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1206 *statusp
= RCU_KTHREAD_WAITING
;
1210 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1211 * served by the rcu_node in question. The CPU hotplug lock is still
1212 * held, so the value of rnp->qsmaskinit will be stable.
1214 * We don't include outgoingcpu in the affinity set, use -1 if there is
1215 * no outgoing CPU. If there are no CPUs left in the affinity set,
1216 * this function allows the kthread to execute on any CPU.
1218 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1220 struct task_struct
*t
= rnp
->boost_kthread_task
;
1221 unsigned long mask
= rcu_rnp_online_cpus(rnp
);
1227 if (!zalloc_cpumask_var(&cm
, GFP_KERNEL
))
1229 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1230 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1231 cpumask_set_cpu(cpu
, cm
);
1232 if (cpumask_weight(cm
) == 0)
1234 set_cpus_allowed_ptr(t
, cm
);
1235 free_cpumask_var(cm
);
1238 static struct smp_hotplug_thread rcu_cpu_thread_spec
= {
1239 .store
= &rcu_cpu_kthread_task
,
1240 .thread_should_run
= rcu_cpu_kthread_should_run
,
1241 .thread_fn
= rcu_cpu_kthread
,
1242 .thread_comm
= "rcuc/%u",
1243 .setup
= rcu_cpu_kthread_setup
,
1244 .park
= rcu_cpu_kthread_park
,
1248 * Spawn boost kthreads -- called as soon as the scheduler is running.
1250 static void __init
rcu_spawn_boost_kthreads(void)
1252 struct rcu_node
*rnp
;
1255 for_each_possible_cpu(cpu
)
1256 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1257 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec
));
1258 rcu_for_each_leaf_node(rcu_state_p
, rnp
)
1259 (void)rcu_spawn_one_boost_kthread(rcu_state_p
, rnp
);
1262 static void rcu_prepare_kthreads(int cpu
)
1264 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
1265 struct rcu_node
*rnp
= rdp
->mynode
;
1267 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1268 if (rcu_scheduler_fully_active
)
1269 (void)rcu_spawn_one_boost_kthread(rcu_state_p
, rnp
);
1272 #else /* #ifdef CONFIG_RCU_BOOST */
1274 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1275 __releases(rnp
->lock
)
1277 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1280 static void invoke_rcu_callbacks_kthread(void)
1285 static bool rcu_is_callbacks_kthread(void)
1290 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1294 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1298 static void __init
rcu_spawn_boost_kthreads(void)
1302 static void rcu_prepare_kthreads(int cpu
)
1306 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1308 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1311 * Check to see if any future RCU-related work will need to be done
1312 * by the current CPU, even if none need be done immediately, returning
1313 * 1 if so. This function is part of the RCU implementation; it is -not-
1314 * an exported member of the RCU API.
1316 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1317 * any flavor of RCU.
1319 int rcu_needs_cpu(u64 basemono
, u64
*nextevt
)
1321 *nextevt
= KTIME_MAX
;
1322 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL
)
1323 ? 0 : rcu_cpu_has_callbacks(NULL
);
1327 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1330 static void rcu_cleanup_after_idle(void)
1335 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1338 static void rcu_prepare_for_idle(void)
1343 * Don't bother keeping a running count of the number of RCU callbacks
1344 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1346 static void rcu_idle_count_callbacks_posted(void)
1350 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1353 * This code is invoked when a CPU goes idle, at which point we want
1354 * to have the CPU do everything required for RCU so that it can enter
1355 * the energy-efficient dyntick-idle mode. This is handled by a
1356 * state machine implemented by rcu_prepare_for_idle() below.
1358 * The following three proprocessor symbols control this state machine:
1360 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1361 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1362 * is sized to be roughly one RCU grace period. Those energy-efficiency
1363 * benchmarkers who might otherwise be tempted to set this to a large
1364 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1365 * system. And if you are -that- concerned about energy efficiency,
1366 * just power the system down and be done with it!
1367 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1368 * permitted to sleep in dyntick-idle mode with only lazy RCU
1369 * callbacks pending. Setting this too high can OOM your system.
1371 * The values below work well in practice. If future workloads require
1372 * adjustment, they can be converted into kernel config parameters, though
1373 * making the state machine smarter might be a better option.
1375 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1376 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1378 static int rcu_idle_gp_delay
= RCU_IDLE_GP_DELAY
;
1379 module_param(rcu_idle_gp_delay
, int, 0644);
1380 static int rcu_idle_lazy_gp_delay
= RCU_IDLE_LAZY_GP_DELAY
;
1381 module_param(rcu_idle_lazy_gp_delay
, int, 0644);
1384 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1385 * only if it has been awhile since the last time we did so. Afterwards,
1386 * if there are any callbacks ready for immediate invocation, return true.
1388 static bool __maybe_unused
rcu_try_advance_all_cbs(void)
1390 bool cbs_ready
= false;
1391 struct rcu_data
*rdp
;
1392 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1393 struct rcu_node
*rnp
;
1394 struct rcu_state
*rsp
;
1396 /* Exit early if we advanced recently. */
1397 if (jiffies
== rdtp
->last_advance_all
)
1399 rdtp
->last_advance_all
= jiffies
;
1401 for_each_rcu_flavor(rsp
) {
1402 rdp
= this_cpu_ptr(rsp
->rda
);
1406 * Don't bother checking unless a grace period has
1407 * completed since we last checked and there are
1408 * callbacks not yet ready to invoke.
1410 if ((rdp
->completed
!= rnp
->completed
||
1411 unlikely(READ_ONCE(rdp
->gpwrap
))) &&
1412 rdp
->nxttail
[RCU_DONE_TAIL
] != rdp
->nxttail
[RCU_NEXT_TAIL
])
1413 note_gp_changes(rsp
, rdp
);
1415 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1422 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1423 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1424 * caller to set the timeout based on whether or not there are non-lazy
1427 * The caller must have disabled interrupts.
1429 int rcu_needs_cpu(u64 basemono
, u64
*nextevt
)
1431 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1434 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL
)) {
1435 *nextevt
= KTIME_MAX
;
1439 /* Snapshot to detect later posting of non-lazy callback. */
1440 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1442 /* If no callbacks, RCU doesn't need the CPU. */
1443 if (!rcu_cpu_has_callbacks(&rdtp
->all_lazy
)) {
1444 *nextevt
= KTIME_MAX
;
1448 /* Attempt to advance callbacks. */
1449 if (rcu_try_advance_all_cbs()) {
1450 /* Some ready to invoke, so initiate later invocation. */
1454 rdtp
->last_accelerate
= jiffies
;
1456 /* Request timer delay depending on laziness, and round. */
1457 if (!rdtp
->all_lazy
) {
1458 dj
= round_up(rcu_idle_gp_delay
+ jiffies
,
1459 rcu_idle_gp_delay
) - jiffies
;
1461 dj
= round_jiffies(rcu_idle_lazy_gp_delay
+ jiffies
) - jiffies
;
1463 *nextevt
= basemono
+ dj
* TICK_NSEC
;
1468 * Prepare a CPU for idle from an RCU perspective. The first major task
1469 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1470 * The second major task is to check to see if a non-lazy callback has
1471 * arrived at a CPU that previously had only lazy callbacks. The third
1472 * major task is to accelerate (that is, assign grace-period numbers to)
1473 * any recently arrived callbacks.
1475 * The caller must have disabled interrupts.
1477 static void rcu_prepare_for_idle(void)
1480 struct rcu_data
*rdp
;
1481 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1482 struct rcu_node
*rnp
;
1483 struct rcu_state
*rsp
;
1486 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL
))
1489 /* Handle nohz enablement switches conservatively. */
1490 tne
= READ_ONCE(tick_nohz_active
);
1491 if (tne
!= rdtp
->tick_nohz_enabled_snap
) {
1492 if (rcu_cpu_has_callbacks(NULL
))
1493 invoke_rcu_core(); /* force nohz to see update. */
1494 rdtp
->tick_nohz_enabled_snap
= tne
;
1500 /* If this is a no-CBs CPU, no callbacks, just return. */
1501 if (rcu_is_nocb_cpu(smp_processor_id()))
1505 * If a non-lazy callback arrived at a CPU having only lazy
1506 * callbacks, invoke RCU core for the side-effect of recalculating
1507 * idle duration on re-entry to idle.
1509 if (rdtp
->all_lazy
&&
1510 rdtp
->nonlazy_posted
!= rdtp
->nonlazy_posted_snap
) {
1511 rdtp
->all_lazy
= false;
1512 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1518 * If we have not yet accelerated this jiffy, accelerate all
1519 * callbacks on this CPU.
1521 if (rdtp
->last_accelerate
== jiffies
)
1523 rdtp
->last_accelerate
= jiffies
;
1524 for_each_rcu_flavor(rsp
) {
1525 rdp
= this_cpu_ptr(rsp
->rda
);
1526 if (!*rdp
->nxttail
[RCU_DONE_TAIL
])
1529 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1530 smp_mb__after_unlock_lock();
1531 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1532 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1534 rcu_gp_kthread_wake(rsp
);
1539 * Clean up for exit from idle. Attempt to advance callbacks based on
1540 * any grace periods that elapsed while the CPU was idle, and if any
1541 * callbacks are now ready to invoke, initiate invocation.
1543 static void rcu_cleanup_after_idle(void)
1545 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL
) ||
1546 rcu_is_nocb_cpu(smp_processor_id()))
1548 if (rcu_try_advance_all_cbs())
1553 * Keep a running count of the number of non-lazy callbacks posted
1554 * on this CPU. This running counter (which is never decremented) allows
1555 * rcu_prepare_for_idle() to detect when something out of the idle loop
1556 * posts a callback, even if an equal number of callbacks are invoked.
1557 * Of course, callbacks should only be posted from within a trace event
1558 * designed to be called from idle or from within RCU_NONIDLE().
1560 static void rcu_idle_count_callbacks_posted(void)
1562 __this_cpu_add(rcu_dynticks
.nonlazy_posted
, 1);
1566 * Data for flushing lazy RCU callbacks at OOM time.
1568 static atomic_t oom_callback_count
;
1569 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq
);
1572 * RCU OOM callback -- decrement the outstanding count and deliver the
1573 * wake-up if we are the last one.
1575 static void rcu_oom_callback(struct rcu_head
*rhp
)
1577 if (atomic_dec_and_test(&oom_callback_count
))
1578 wake_up(&oom_callback_wq
);
1582 * Post an rcu_oom_notify callback on the current CPU if it has at
1583 * least one lazy callback. This will unnecessarily post callbacks
1584 * to CPUs that already have a non-lazy callback at the end of their
1585 * callback list, but this is an infrequent operation, so accept some
1586 * extra overhead to keep things simple.
1588 static void rcu_oom_notify_cpu(void *unused
)
1590 struct rcu_state
*rsp
;
1591 struct rcu_data
*rdp
;
1593 for_each_rcu_flavor(rsp
) {
1594 rdp
= raw_cpu_ptr(rsp
->rda
);
1595 if (rdp
->qlen_lazy
!= 0) {
1596 atomic_inc(&oom_callback_count
);
1597 rsp
->call(&rdp
->oom_head
, rcu_oom_callback
);
1603 * If low on memory, ensure that each CPU has a non-lazy callback.
1604 * This will wake up CPUs that have only lazy callbacks, in turn
1605 * ensuring that they free up the corresponding memory in a timely manner.
1606 * Because an uncertain amount of memory will be freed in some uncertain
1607 * timeframe, we do not claim to have freed anything.
1609 static int rcu_oom_notify(struct notifier_block
*self
,
1610 unsigned long notused
, void *nfreed
)
1614 /* Wait for callbacks from earlier instance to complete. */
1615 wait_event(oom_callback_wq
, atomic_read(&oom_callback_count
) == 0);
1616 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1619 * Prevent premature wakeup: ensure that all increments happen
1620 * before there is a chance of the counter reaching zero.
1622 atomic_set(&oom_callback_count
, 1);
1624 for_each_online_cpu(cpu
) {
1625 smp_call_function_single(cpu
, rcu_oom_notify_cpu
, NULL
, 1);
1626 cond_resched_rcu_qs();
1629 /* Unconditionally decrement: no need to wake ourselves up. */
1630 atomic_dec(&oom_callback_count
);
1635 static struct notifier_block rcu_oom_nb
= {
1636 .notifier_call
= rcu_oom_notify
1639 static int __init
rcu_register_oom_notifier(void)
1641 register_oom_notifier(&rcu_oom_nb
);
1644 early_initcall(rcu_register_oom_notifier
);
1646 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1648 #ifdef CONFIG_RCU_FAST_NO_HZ
1650 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1652 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1653 unsigned long nlpd
= rdtp
->nonlazy_posted
- rdtp
->nonlazy_posted_snap
;
1655 sprintf(cp
, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1656 rdtp
->last_accelerate
& 0xffff, jiffies
& 0xffff,
1658 rdtp
->all_lazy
? 'L' : '.',
1659 rdtp
->tick_nohz_enabled_snap
? '.' : 'D');
1662 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1664 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1669 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1671 /* Initiate the stall-info list. */
1672 static void print_cpu_stall_info_begin(void)
1678 * Print out diagnostic information for the specified stalled CPU.
1680 * If the specified CPU is aware of the current RCU grace period
1681 * (flavor specified by rsp), then print the number of scheduling
1682 * clock interrupts the CPU has taken during the time that it has
1683 * been aware. Otherwise, print the number of RCU grace periods
1684 * that this CPU is ignorant of, for example, "1" if the CPU was
1685 * aware of the previous grace period.
1687 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1689 static void print_cpu_stall_info(struct rcu_state
*rsp
, int cpu
)
1691 char fast_no_hz
[72];
1692 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1693 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
1695 unsigned long ticks_value
;
1697 if (rsp
->gpnum
== rdp
->gpnum
) {
1698 ticks_title
= "ticks this GP";
1699 ticks_value
= rdp
->ticks_this_gp
;
1701 ticks_title
= "GPs behind";
1702 ticks_value
= rsp
->gpnum
- rdp
->gpnum
;
1704 print_cpu_stall_fast_no_hz(fast_no_hz
, cpu
);
1705 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1706 cpu
, ticks_value
, ticks_title
,
1707 atomic_read(&rdtp
->dynticks
) & 0xfff,
1708 rdtp
->dynticks_nesting
, rdtp
->dynticks_nmi_nesting
,
1709 rdp
->softirq_snap
, kstat_softirqs_cpu(RCU_SOFTIRQ
, cpu
),
1710 READ_ONCE(rsp
->n_force_qs
) - rsp
->n_force_qs_gpstart
,
1714 /* Terminate the stall-info list. */
1715 static void print_cpu_stall_info_end(void)
1720 /* Zero ->ticks_this_gp for all flavors of RCU. */
1721 static void zero_cpu_stall_ticks(struct rcu_data
*rdp
)
1723 rdp
->ticks_this_gp
= 0;
1724 rdp
->softirq_snap
= kstat_softirqs_cpu(RCU_SOFTIRQ
, smp_processor_id());
1727 /* Increment ->ticks_this_gp for all flavors of RCU. */
1728 static void increment_cpu_stall_ticks(void)
1730 struct rcu_state
*rsp
;
1732 for_each_rcu_flavor(rsp
)
1733 raw_cpu_inc(rsp
->rda
->ticks_this_gp
);
1736 #ifdef CONFIG_RCU_NOCB_CPU
1739 * Offload callback processing from the boot-time-specified set of CPUs
1740 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1741 * kthread created that pulls the callbacks from the corresponding CPU,
1742 * waits for a grace period to elapse, and invokes the callbacks.
1743 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1744 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1745 * has been specified, in which case each kthread actively polls its
1746 * CPU. (Which isn't so great for energy efficiency, but which does
1747 * reduce RCU's overhead on that CPU.)
1749 * This is intended to be used in conjunction with Frederic Weisbecker's
1750 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1751 * running CPU-bound user-mode computations.
1753 * Offloading of callback processing could also in theory be used as
1754 * an energy-efficiency measure because CPUs with no RCU callbacks
1755 * queued are more aggressive about entering dyntick-idle mode.
1759 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1760 static int __init
rcu_nocb_setup(char *str
)
1762 alloc_bootmem_cpumask_var(&rcu_nocb_mask
);
1763 have_rcu_nocb_mask
= true;
1764 cpulist_parse(str
, rcu_nocb_mask
);
1767 __setup("rcu_nocbs=", rcu_nocb_setup
);
1769 static int __init
parse_rcu_nocb_poll(char *arg
)
1774 early_param("rcu_nocb_poll", parse_rcu_nocb_poll
);
1777 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1780 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1782 wake_up_all(&rnp
->nocb_gp_wq
[rnp
->completed
& 0x1]);
1786 * Set the root rcu_node structure's ->need_future_gp field
1787 * based on the sum of those of all rcu_node structures. This does
1788 * double-count the root rcu_node structure's requests, but this
1789 * is necessary to handle the possibility of a rcu_nocb_kthread()
1790 * having awakened during the time that the rcu_node structures
1791 * were being updated for the end of the previous grace period.
1793 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
1795 rnp
->need_future_gp
[(rnp
->completed
+ 1) & 0x1] += nrq
;
1798 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
1800 init_waitqueue_head(&rnp
->nocb_gp_wq
[0]);
1801 init_waitqueue_head(&rnp
->nocb_gp_wq
[1]);
1804 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1805 /* Is the specified CPU a no-CBs CPU? */
1806 bool rcu_is_nocb_cpu(int cpu
)
1808 if (have_rcu_nocb_mask
)
1809 return cpumask_test_cpu(cpu
, rcu_nocb_mask
);
1812 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1815 * Kick the leader kthread for this NOCB group.
1817 static void wake_nocb_leader(struct rcu_data
*rdp
, bool force
)
1819 struct rcu_data
*rdp_leader
= rdp
->nocb_leader
;
1821 if (!READ_ONCE(rdp_leader
->nocb_kthread
))
1823 if (READ_ONCE(rdp_leader
->nocb_leader_sleep
) || force
) {
1824 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1825 WRITE_ONCE(rdp_leader
->nocb_leader_sleep
, false);
1826 wake_up(&rdp_leader
->nocb_wq
);
1831 * Does the specified CPU need an RCU callback for the specified flavor
1834 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state
*rsp
, int cpu
)
1836 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1838 #ifdef CONFIG_PROVE_RCU
1839 struct rcu_head
*rhp
;
1840 #endif /* #ifdef CONFIG_PROVE_RCU */
1843 * Check count of all no-CBs callbacks awaiting invocation.
1844 * There needs to be a barrier before this function is called,
1845 * but associated with a prior determination that no more
1846 * callbacks would be posted. In the worst case, the first
1847 * barrier in _rcu_barrier() suffices (but the caller cannot
1848 * necessarily rely on this, not a substitute for the caller
1849 * getting the concurrency design right!). There must also be
1850 * a barrier between the following load an posting of a callback
1851 * (if a callback is in fact needed). This is associated with an
1852 * atomic_inc() in the caller.
1854 ret
= atomic_long_read(&rdp
->nocb_q_count
);
1856 #ifdef CONFIG_PROVE_RCU
1857 rhp
= READ_ONCE(rdp
->nocb_head
);
1859 rhp
= READ_ONCE(rdp
->nocb_gp_head
);
1861 rhp
= READ_ONCE(rdp
->nocb_follower_head
);
1863 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1864 if (!READ_ONCE(rdp
->nocb_kthread
) && rhp
&&
1865 rcu_scheduler_fully_active
) {
1866 /* RCU callback enqueued before CPU first came online??? */
1867 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1871 #endif /* #ifdef CONFIG_PROVE_RCU */
1877 * Enqueue the specified string of rcu_head structures onto the specified
1878 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1879 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1880 * counts are supplied by rhcount and rhcount_lazy.
1882 * If warranted, also wake up the kthread servicing this CPUs queues.
1884 static void __call_rcu_nocb_enqueue(struct rcu_data
*rdp
,
1885 struct rcu_head
*rhp
,
1886 struct rcu_head
**rhtp
,
1887 int rhcount
, int rhcount_lazy
,
1888 unsigned long flags
)
1891 struct rcu_head
**old_rhpp
;
1892 struct task_struct
*t
;
1894 /* Enqueue the callback on the nocb list and update counts. */
1895 atomic_long_add(rhcount
, &rdp
->nocb_q_count
);
1896 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1897 old_rhpp
= xchg(&rdp
->nocb_tail
, rhtp
);
1898 WRITE_ONCE(*old_rhpp
, rhp
);
1899 atomic_long_add(rhcount_lazy
, &rdp
->nocb_q_count_lazy
);
1900 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1902 /* If we are not being polled and there is a kthread, awaken it ... */
1903 t
= READ_ONCE(rdp
->nocb_kthread
);
1904 if (rcu_nocb_poll
|| !t
) {
1905 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
1906 TPS("WakeNotPoll"));
1909 len
= atomic_long_read(&rdp
->nocb_q_count
);
1910 if (old_rhpp
== &rdp
->nocb_head
) {
1911 if (!irqs_disabled_flags(flags
)) {
1912 /* ... if queue was empty ... */
1913 wake_nocb_leader(rdp
, false);
1914 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
1917 rdp
->nocb_defer_wakeup
= RCU_NOGP_WAKE
;
1918 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
1919 TPS("WakeEmptyIsDeferred"));
1921 rdp
->qlen_last_fqs_check
= 0;
1922 } else if (len
> rdp
->qlen_last_fqs_check
+ qhimark
) {
1923 /* ... or if many callbacks queued. */
1924 if (!irqs_disabled_flags(flags
)) {
1925 wake_nocb_leader(rdp
, true);
1926 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
1929 rdp
->nocb_defer_wakeup
= RCU_NOGP_WAKE_FORCE
;
1930 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
1931 TPS("WakeOvfIsDeferred"));
1933 rdp
->qlen_last_fqs_check
= LONG_MAX
/ 2;
1935 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("WakeNot"));
1941 * This is a helper for __call_rcu(), which invokes this when the normal
1942 * callback queue is inoperable. If this is not a no-CBs CPU, this
1943 * function returns failure back to __call_rcu(), which can complain
1946 * Otherwise, this function queues the callback where the corresponding
1947 * "rcuo" kthread can find it.
1949 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
1950 bool lazy
, unsigned long flags
)
1953 if (!rcu_is_nocb_cpu(rdp
->cpu
))
1955 __call_rcu_nocb_enqueue(rdp
, rhp
, &rhp
->next
, 1, lazy
, flags
);
1956 if (__is_kfree_rcu_offset((unsigned long)rhp
->func
))
1957 trace_rcu_kfree_callback(rdp
->rsp
->name
, rhp
,
1958 (unsigned long)rhp
->func
,
1959 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
1960 -atomic_long_read(&rdp
->nocb_q_count
));
1962 trace_rcu_callback(rdp
->rsp
->name
, rhp
,
1963 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
1964 -atomic_long_read(&rdp
->nocb_q_count
));
1967 * If called from an extended quiescent state with interrupts
1968 * disabled, invoke the RCU core in order to allow the idle-entry
1969 * deferred-wakeup check to function.
1971 if (irqs_disabled_flags(flags
) &&
1972 !rcu_is_watching() &&
1973 cpu_online(smp_processor_id()))
1980 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
1983 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
1984 struct rcu_data
*rdp
,
1985 unsigned long flags
)
1987 long ql
= rsp
->qlen
;
1988 long qll
= rsp
->qlen_lazy
;
1990 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
1991 if (!rcu_is_nocb_cpu(smp_processor_id()))
1996 /* First, enqueue the donelist, if any. This preserves CB ordering. */
1997 if (rsp
->orphan_donelist
!= NULL
) {
1998 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_donelist
,
1999 rsp
->orphan_donetail
, ql
, qll
, flags
);
2001 rsp
->orphan_donelist
= NULL
;
2002 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2004 if (rsp
->orphan_nxtlist
!= NULL
) {
2005 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_nxtlist
,
2006 rsp
->orphan_nxttail
, ql
, qll
, flags
);
2008 rsp
->orphan_nxtlist
= NULL
;
2009 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2015 * If necessary, kick off a new grace period, and either way wait
2016 * for a subsequent grace period to complete.
2018 static void rcu_nocb_wait_gp(struct rcu_data
*rdp
)
2022 unsigned long flags
;
2024 struct rcu_node
*rnp
= rdp
->mynode
;
2026 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2027 smp_mb__after_unlock_lock();
2028 needwake
= rcu_start_future_gp(rnp
, rdp
, &c
);
2029 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2031 rcu_gp_kthread_wake(rdp
->rsp
);
2034 * Wait for the grace period. Do so interruptibly to avoid messing
2035 * up the load average.
2037 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("StartWait"));
2039 wait_event_interruptible(
2040 rnp
->nocb_gp_wq
[c
& 0x1],
2041 (d
= ULONG_CMP_GE(READ_ONCE(rnp
->completed
), c
)));
2044 WARN_ON(signal_pending(current
));
2045 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("ResumeWait"));
2047 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("EndWait"));
2048 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2052 * Leaders come here to wait for additional callbacks to show up.
2053 * This function does not return until callbacks appear.
2055 static void nocb_leader_wait(struct rcu_data
*my_rdp
)
2057 bool firsttime
= true;
2059 struct rcu_data
*rdp
;
2060 struct rcu_head
**tail
;
2064 /* Wait for callbacks to appear. */
2065 if (!rcu_nocb_poll
) {
2066 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
, "Sleep");
2067 wait_event_interruptible(my_rdp
->nocb_wq
,
2068 !READ_ONCE(my_rdp
->nocb_leader_sleep
));
2069 /* Memory barrier handled by smp_mb() calls below and repoll. */
2070 } else if (firsttime
) {
2071 firsttime
= false; /* Don't drown trace log with "Poll"! */
2072 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
, "Poll");
2076 * Each pass through the following loop checks a follower for CBs.
2077 * We are our own first follower. Any CBs found are moved to
2078 * nocb_gp_head, where they await a grace period.
2081 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
) {
2082 rdp
->nocb_gp_head
= READ_ONCE(rdp
->nocb_head
);
2083 if (!rdp
->nocb_gp_head
)
2084 continue; /* No CBs here, try next follower. */
2086 /* Move callbacks to wait-for-GP list, which is empty. */
2087 WRITE_ONCE(rdp
->nocb_head
, NULL
);
2088 rdp
->nocb_gp_tail
= xchg(&rdp
->nocb_tail
, &rdp
->nocb_head
);
2093 * If there were no callbacks, sleep a bit, rescan after a
2094 * memory barrier, and go retry.
2096 if (unlikely(!gotcbs
)) {
2098 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
,
2100 WARN_ON(signal_pending(current
));
2101 schedule_timeout_interruptible(1);
2103 /* Rescan in case we were a victim of memory ordering. */
2104 my_rdp
->nocb_leader_sleep
= true;
2105 smp_mb(); /* Ensure _sleep true before scan. */
2106 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
)
2107 if (READ_ONCE(rdp
->nocb_head
)) {
2108 /* Found CB, so short-circuit next wait. */
2109 my_rdp
->nocb_leader_sleep
= false;
2115 /* Wait for one grace period. */
2116 rcu_nocb_wait_gp(my_rdp
);
2119 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2120 * We set it now, but recheck for new callbacks while
2121 * traversing our follower list.
2123 my_rdp
->nocb_leader_sleep
= true;
2124 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2126 /* Each pass through the following loop wakes a follower, if needed. */
2127 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
) {
2128 if (READ_ONCE(rdp
->nocb_head
))
2129 my_rdp
->nocb_leader_sleep
= false;/* No need to sleep.*/
2130 if (!rdp
->nocb_gp_head
)
2131 continue; /* No CBs, so no need to wake follower. */
2133 /* Append callbacks to follower's "done" list. */
2134 tail
= xchg(&rdp
->nocb_follower_tail
, rdp
->nocb_gp_tail
);
2135 *tail
= rdp
->nocb_gp_head
;
2136 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2137 if (rdp
!= my_rdp
&& tail
== &rdp
->nocb_follower_head
) {
2139 * List was empty, wake up the follower.
2140 * Memory barriers supplied by atomic_long_add().
2142 wake_up(&rdp
->nocb_wq
);
2146 /* If we (the leader) don't have CBs, go wait some more. */
2147 if (!my_rdp
->nocb_follower_head
)
2152 * Followers come here to wait for additional callbacks to show up.
2153 * This function does not return until callbacks appear.
2155 static void nocb_follower_wait(struct rcu_data
*rdp
)
2157 bool firsttime
= true;
2160 if (!rcu_nocb_poll
) {
2161 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2163 wait_event_interruptible(rdp
->nocb_wq
,
2164 READ_ONCE(rdp
->nocb_follower_head
));
2165 } else if (firsttime
) {
2166 /* Don't drown trace log with "Poll"! */
2168 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, "Poll");
2170 if (smp_load_acquire(&rdp
->nocb_follower_head
)) {
2171 /* ^^^ Ensure CB invocation follows _head test. */
2175 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2177 WARN_ON(signal_pending(current
));
2178 schedule_timeout_interruptible(1);
2183 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2184 * callbacks queued by the corresponding no-CBs CPU, however, there is
2185 * an optional leader-follower relationship so that the grace-period
2186 * kthreads don't have to do quite so many wakeups.
2188 static int rcu_nocb_kthread(void *arg
)
2191 struct rcu_head
*list
;
2192 struct rcu_head
*next
;
2193 struct rcu_head
**tail
;
2194 struct rcu_data
*rdp
= arg
;
2196 /* Each pass through this loop invokes one batch of callbacks */
2198 /* Wait for callbacks. */
2199 if (rdp
->nocb_leader
== rdp
)
2200 nocb_leader_wait(rdp
);
2202 nocb_follower_wait(rdp
);
2204 /* Pull the ready-to-invoke callbacks onto local list. */
2205 list
= READ_ONCE(rdp
->nocb_follower_head
);
2207 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, "WokeNonEmpty");
2208 WRITE_ONCE(rdp
->nocb_follower_head
, NULL
);
2209 tail
= xchg(&rdp
->nocb_follower_tail
, &rdp
->nocb_follower_head
);
2211 /* Each pass through the following loop invokes a callback. */
2212 trace_rcu_batch_start(rdp
->rsp
->name
,
2213 atomic_long_read(&rdp
->nocb_q_count_lazy
),
2214 atomic_long_read(&rdp
->nocb_q_count
), -1);
2218 /* Wait for enqueuing to complete, if needed. */
2219 while (next
== NULL
&& &list
->next
!= tail
) {
2220 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2222 schedule_timeout_interruptible(1);
2223 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2227 debug_rcu_head_unqueue(list
);
2229 if (__rcu_reclaim(rdp
->rsp
->name
, list
))
2235 trace_rcu_batch_end(rdp
->rsp
->name
, c
, !!list
, 0, 0, 1);
2236 smp_mb__before_atomic(); /* _add after CB invocation. */
2237 atomic_long_add(-c
, &rdp
->nocb_q_count
);
2238 atomic_long_add(-cl
, &rdp
->nocb_q_count_lazy
);
2239 rdp
->n_nocbs_invoked
+= c
;
2244 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2245 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2247 return READ_ONCE(rdp
->nocb_defer_wakeup
);
2250 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2251 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2255 if (!rcu_nocb_need_deferred_wakeup(rdp
))
2257 ndw
= READ_ONCE(rdp
->nocb_defer_wakeup
);
2258 WRITE_ONCE(rdp
->nocb_defer_wakeup
, RCU_NOGP_WAKE_NOT
);
2259 wake_nocb_leader(rdp
, ndw
== RCU_NOGP_WAKE_FORCE
);
2260 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("DeferredWake"));
2263 void __init
rcu_init_nohz(void)
2266 bool need_rcu_nocb_mask
= true;
2267 struct rcu_state
*rsp
;
2269 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2270 need_rcu_nocb_mask
= false;
2271 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2273 #if defined(CONFIG_NO_HZ_FULL)
2274 if (tick_nohz_full_running
&& cpumask_weight(tick_nohz_full_mask
))
2275 need_rcu_nocb_mask
= true;
2276 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2278 if (!have_rcu_nocb_mask
&& need_rcu_nocb_mask
) {
2279 if (!zalloc_cpumask_var(&rcu_nocb_mask
, GFP_KERNEL
)) {
2280 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2283 have_rcu_nocb_mask
= true;
2285 if (!have_rcu_nocb_mask
)
2288 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2289 pr_info("\tOffload RCU callbacks from CPU 0\n");
2290 cpumask_set_cpu(0, rcu_nocb_mask
);
2291 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2292 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2293 pr_info("\tOffload RCU callbacks from all CPUs\n");
2294 cpumask_copy(rcu_nocb_mask
, cpu_possible_mask
);
2295 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2296 #if defined(CONFIG_NO_HZ_FULL)
2297 if (tick_nohz_full_running
)
2298 cpumask_or(rcu_nocb_mask
, rcu_nocb_mask
, tick_nohz_full_mask
);
2299 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2301 if (!cpumask_subset(rcu_nocb_mask
, cpu_possible_mask
)) {
2302 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2303 cpumask_and(rcu_nocb_mask
, cpu_possible_mask
,
2306 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2307 cpumask_pr_args(rcu_nocb_mask
));
2309 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2311 for_each_rcu_flavor(rsp
) {
2312 for_each_cpu(cpu
, rcu_nocb_mask
)
2313 init_nocb_callback_list(per_cpu_ptr(rsp
->rda
, cpu
));
2314 rcu_organize_nocb_kthreads(rsp
);
2318 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2319 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2321 rdp
->nocb_tail
= &rdp
->nocb_head
;
2322 init_waitqueue_head(&rdp
->nocb_wq
);
2323 rdp
->nocb_follower_tail
= &rdp
->nocb_follower_head
;
2327 * If the specified CPU is a no-CBs CPU that does not already have its
2328 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2329 * brought online out of order, this can require re-organizing the
2330 * leader-follower relationships.
2332 static void rcu_spawn_one_nocb_kthread(struct rcu_state
*rsp
, int cpu
)
2334 struct rcu_data
*rdp
;
2335 struct rcu_data
*rdp_last
;
2336 struct rcu_data
*rdp_old_leader
;
2337 struct rcu_data
*rdp_spawn
= per_cpu_ptr(rsp
->rda
, cpu
);
2338 struct task_struct
*t
;
2341 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2342 * then nothing to do.
2344 if (!rcu_is_nocb_cpu(cpu
) || rdp_spawn
->nocb_kthread
)
2347 /* If we didn't spawn the leader first, reorganize! */
2348 rdp_old_leader
= rdp_spawn
->nocb_leader
;
2349 if (rdp_old_leader
!= rdp_spawn
&& !rdp_old_leader
->nocb_kthread
) {
2351 rdp
= rdp_old_leader
;
2353 rdp
->nocb_leader
= rdp_spawn
;
2354 if (rdp_last
&& rdp
!= rdp_spawn
)
2355 rdp_last
->nocb_next_follower
= rdp
;
2356 if (rdp
== rdp_spawn
) {
2357 rdp
= rdp
->nocb_next_follower
;
2360 rdp
= rdp
->nocb_next_follower
;
2361 rdp_last
->nocb_next_follower
= NULL
;
2364 rdp_spawn
->nocb_next_follower
= rdp_old_leader
;
2367 /* Spawn the kthread for this CPU and RCU flavor. */
2368 t
= kthread_run(rcu_nocb_kthread
, rdp_spawn
,
2369 "rcuo%c/%d", rsp
->abbr
, cpu
);
2371 WRITE_ONCE(rdp_spawn
->nocb_kthread
, t
);
2375 * If the specified CPU is a no-CBs CPU that does not already have its
2376 * rcuo kthreads, spawn them.
2378 static void rcu_spawn_all_nocb_kthreads(int cpu
)
2380 struct rcu_state
*rsp
;
2382 if (rcu_scheduler_fully_active
)
2383 for_each_rcu_flavor(rsp
)
2384 rcu_spawn_one_nocb_kthread(rsp
, cpu
);
2388 * Once the scheduler is running, spawn rcuo kthreads for all online
2389 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2390 * non-boot CPUs come online -- if this changes, we will need to add
2391 * some mutual exclusion.
2393 static void __init
rcu_spawn_nocb_kthreads(void)
2397 for_each_online_cpu(cpu
)
2398 rcu_spawn_all_nocb_kthreads(cpu
);
2401 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2402 static int rcu_nocb_leader_stride
= -1;
2403 module_param(rcu_nocb_leader_stride
, int, 0444);
2406 * Initialize leader-follower relationships for all no-CBs CPU.
2408 static void __init
rcu_organize_nocb_kthreads(struct rcu_state
*rsp
)
2411 int ls
= rcu_nocb_leader_stride
;
2412 int nl
= 0; /* Next leader. */
2413 struct rcu_data
*rdp
;
2414 struct rcu_data
*rdp_leader
= NULL
; /* Suppress misguided gcc warn. */
2415 struct rcu_data
*rdp_prev
= NULL
;
2417 if (!have_rcu_nocb_mask
)
2420 ls
= int_sqrt(nr_cpu_ids
);
2421 rcu_nocb_leader_stride
= ls
;
2425 * Each pass through this loop sets up one rcu_data structure and
2426 * spawns one rcu_nocb_kthread().
2428 for_each_cpu(cpu
, rcu_nocb_mask
) {
2429 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2430 if (rdp
->cpu
>= nl
) {
2431 /* New leader, set up for followers & next leader. */
2432 nl
= DIV_ROUND_UP(rdp
->cpu
+ 1, ls
) * ls
;
2433 rdp
->nocb_leader
= rdp
;
2436 /* Another follower, link to previous leader. */
2437 rdp
->nocb_leader
= rdp_leader
;
2438 rdp_prev
->nocb_next_follower
= rdp
;
2444 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2445 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2447 if (!rcu_is_nocb_cpu(rdp
->cpu
))
2450 /* If there are early-boot callbacks, move them to nocb lists. */
2452 rdp
->nocb_head
= rdp
->nxtlist
;
2453 rdp
->nocb_tail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2454 atomic_long_set(&rdp
->nocb_q_count
, rdp
->qlen
);
2455 atomic_long_set(&rdp
->nocb_q_count_lazy
, rdp
->qlen_lazy
);
2456 rdp
->nxtlist
= NULL
;
2460 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2464 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2466 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state
*rsp
, int cpu
)
2468 WARN_ON_ONCE(1); /* Should be dead code. */
2472 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
2476 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
2480 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2484 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2485 bool lazy
, unsigned long flags
)
2490 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
2491 struct rcu_data
*rdp
,
2492 unsigned long flags
)
2497 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2501 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2506 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2510 static void rcu_spawn_all_nocb_kthreads(int cpu
)
2514 static void __init
rcu_spawn_nocb_kthreads(void)
2518 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2523 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2526 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2527 * arbitrarily long period of time with the scheduling-clock tick turned
2528 * off. RCU will be paying attention to this CPU because it is in the
2529 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2530 * machine because the scheduling-clock tick has been disabled. Therefore,
2531 * if an adaptive-ticks CPU is failing to respond to the current grace
2532 * period and has not be idle from an RCU perspective, kick it.
2534 static void __maybe_unused
rcu_kick_nohz_cpu(int cpu
)
2536 #ifdef CONFIG_NO_HZ_FULL
2537 if (tick_nohz_full_cpu(cpu
))
2538 smp_send_reschedule(cpu
);
2539 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2543 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2545 static int full_sysidle_state
; /* Current system-idle state. */
2546 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2547 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2548 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2549 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2550 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2553 * Invoked to note exit from irq or task transition to idle. Note that
2554 * usermode execution does -not- count as idle here! After all, we want
2555 * to detect full-system idle states, not RCU quiescent states and grace
2556 * periods. The caller must have disabled interrupts.
2558 static void rcu_sysidle_enter(int irq
)
2561 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
2563 /* If there are no nohz_full= CPUs, no need to track this. */
2564 if (!tick_nohz_full_enabled())
2567 /* Adjust nesting, check for fully idle. */
2569 rdtp
->dynticks_idle_nesting
--;
2570 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2571 if (rdtp
->dynticks_idle_nesting
!= 0)
2572 return; /* Still not fully idle. */
2574 if ((rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) ==
2575 DYNTICK_TASK_NEST_VALUE
) {
2576 rdtp
->dynticks_idle_nesting
= 0;
2578 rdtp
->dynticks_idle_nesting
-= DYNTICK_TASK_NEST_VALUE
;
2579 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2580 return; /* Still not fully idle. */
2584 /* Record start of fully idle period. */
2586 WRITE_ONCE(rdtp
->dynticks_idle_jiffies
, j
);
2587 smp_mb__before_atomic();
2588 atomic_inc(&rdtp
->dynticks_idle
);
2589 smp_mb__after_atomic();
2590 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks_idle
) & 0x1);
2594 * Unconditionally force exit from full system-idle state. This is
2595 * invoked when a normal CPU exits idle, but must be called separately
2596 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2597 * is that the timekeeping CPU is permitted to take scheduling-clock
2598 * interrupts while the system is in system-idle state, and of course
2599 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2600 * interrupt from any other type of interrupt.
2602 void rcu_sysidle_force_exit(void)
2604 int oldstate
= READ_ONCE(full_sysidle_state
);
2608 * Each pass through the following loop attempts to exit full
2609 * system-idle state. If contention proves to be a problem,
2610 * a trylock-based contention tree could be used here.
2612 while (oldstate
> RCU_SYSIDLE_SHORT
) {
2613 newoldstate
= cmpxchg(&full_sysidle_state
,
2614 oldstate
, RCU_SYSIDLE_NOT
);
2615 if (oldstate
== newoldstate
&&
2616 oldstate
== RCU_SYSIDLE_FULL_NOTED
) {
2617 rcu_kick_nohz_cpu(tick_do_timer_cpu
);
2618 return; /* We cleared it, done! */
2620 oldstate
= newoldstate
;
2622 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2626 * Invoked to note entry to irq or task transition from idle. Note that
2627 * usermode execution does -not- count as idle here! The caller must
2628 * have disabled interrupts.
2630 static void rcu_sysidle_exit(int irq
)
2632 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
2634 /* If there are no nohz_full= CPUs, no need to track this. */
2635 if (!tick_nohz_full_enabled())
2638 /* Adjust nesting, check for already non-idle. */
2640 rdtp
->dynticks_idle_nesting
++;
2641 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2642 if (rdtp
->dynticks_idle_nesting
!= 1)
2643 return; /* Already non-idle. */
2646 * Allow for irq misnesting. Yes, it really is possible
2647 * to enter an irq handler then never leave it, and maybe
2648 * also vice versa. Handle both possibilities.
2650 if (rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) {
2651 rdtp
->dynticks_idle_nesting
+= DYNTICK_TASK_NEST_VALUE
;
2652 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2653 return; /* Already non-idle. */
2655 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2659 /* Record end of idle period. */
2660 smp_mb__before_atomic();
2661 atomic_inc(&rdtp
->dynticks_idle
);
2662 smp_mb__after_atomic();
2663 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks_idle
) & 0x1));
2666 * If we are the timekeeping CPU, we are permitted to be non-idle
2667 * during a system-idle state. This must be the case, because
2668 * the timekeeping CPU has to take scheduling-clock interrupts
2669 * during the time that the system is transitioning to full
2670 * system-idle state. This means that the timekeeping CPU must
2671 * invoke rcu_sysidle_force_exit() directly if it does anything
2672 * more than take a scheduling-clock interrupt.
2674 if (smp_processor_id() == tick_do_timer_cpu
)
2677 /* Update system-idle state: We are clearly no longer fully idle! */
2678 rcu_sysidle_force_exit();
2682 * Check to see if the current CPU is idle. Note that usermode execution
2683 * does not count as idle. The caller must have disabled interrupts,
2684 * and must be running on tick_do_timer_cpu.
2686 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
2687 unsigned long *maxj
)
2691 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
2693 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2694 if (!tick_nohz_full_enabled())
2698 * If some other CPU has already reported non-idle, if this is
2699 * not the flavor of RCU that tracks sysidle state, or if this
2700 * is an offline or the timekeeping CPU, nothing to do.
2702 if (!*isidle
|| rdp
->rsp
!= rcu_state_p
||
2703 cpu_is_offline(rdp
->cpu
) || rdp
->cpu
== tick_do_timer_cpu
)
2705 /* Verify affinity of current kthread. */
2706 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
);
2708 /* Pick up current idle and NMI-nesting counter and check. */
2709 cur
= atomic_read(&rdtp
->dynticks_idle
);
2711 *isidle
= false; /* We are not idle! */
2714 smp_mb(); /* Read counters before timestamps. */
2716 /* Pick up timestamps. */
2717 j
= READ_ONCE(rdtp
->dynticks_idle_jiffies
);
2718 /* If this CPU entered idle more recently, update maxj timestamp. */
2719 if (ULONG_CMP_LT(*maxj
, j
))
2724 * Is this the flavor of RCU that is handling full-system idle?
2726 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
2728 return rsp
== rcu_state_p
;
2732 * Return a delay in jiffies based on the number of CPUs, rcu_node
2733 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2734 * systems more time to transition to full-idle state in order to
2735 * avoid the cache thrashing that otherwise occur on the state variable.
2736 * Really small systems (less than a couple of tens of CPUs) should
2737 * instead use a single global atomically incremented counter, and later
2738 * versions of this will automatically reconfigure themselves accordingly.
2740 static unsigned long rcu_sysidle_delay(void)
2742 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2744 return DIV_ROUND_UP(nr_cpu_ids
* HZ
, rcu_fanout_leaf
* 1000);
2748 * Advance the full-system-idle state. This is invoked when all of
2749 * the non-timekeeping CPUs are idle.
2751 static void rcu_sysidle(unsigned long j
)
2753 /* Check the current state. */
2754 switch (READ_ONCE(full_sysidle_state
)) {
2755 case RCU_SYSIDLE_NOT
:
2757 /* First time all are idle, so note a short idle period. */
2758 WRITE_ONCE(full_sysidle_state
, RCU_SYSIDLE_SHORT
);
2761 case RCU_SYSIDLE_SHORT
:
2764 * Idle for a bit, time to advance to next state?
2765 * cmpxchg failure means race with non-idle, let them win.
2767 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2768 (void)cmpxchg(&full_sysidle_state
,
2769 RCU_SYSIDLE_SHORT
, RCU_SYSIDLE_LONG
);
2772 case RCU_SYSIDLE_LONG
:
2775 * Do an additional check pass before advancing to full.
2776 * cmpxchg failure means race with non-idle, let them win.
2778 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2779 (void)cmpxchg(&full_sysidle_state
,
2780 RCU_SYSIDLE_LONG
, RCU_SYSIDLE_FULL
);
2789 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2790 * back to the beginning.
2792 static void rcu_sysidle_cancel(void)
2795 if (full_sysidle_state
> RCU_SYSIDLE_SHORT
)
2796 WRITE_ONCE(full_sysidle_state
, RCU_SYSIDLE_NOT
);
2800 * Update the sysidle state based on the results of a force-quiescent-state
2801 * scan of the CPUs' dyntick-idle state.
2803 static void rcu_sysidle_report(struct rcu_state
*rsp
, int isidle
,
2804 unsigned long maxj
, bool gpkt
)
2806 if (rsp
!= rcu_state_p
)
2807 return; /* Wrong flavor, ignore. */
2808 if (gpkt
&& nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2809 return; /* Running state machine from timekeeping CPU. */
2811 rcu_sysidle(maxj
); /* More idle! */
2813 rcu_sysidle_cancel(); /* Idle is over. */
2817 * Wrapper for rcu_sysidle_report() when called from the grace-period
2818 * kthread's context.
2820 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
2823 /* If there are no nohz_full= CPUs, no need to track this. */
2824 if (!tick_nohz_full_enabled())
2827 rcu_sysidle_report(rsp
, isidle
, maxj
, true);
2830 /* Callback and function for forcing an RCU grace period. */
2831 struct rcu_sysidle_head
{
2836 static void rcu_sysidle_cb(struct rcu_head
*rhp
)
2838 struct rcu_sysidle_head
*rshp
;
2841 * The following memory barrier is needed to replace the
2842 * memory barriers that would normally be in the memory
2845 smp_mb(); /* grace period precedes setting inuse. */
2847 rshp
= container_of(rhp
, struct rcu_sysidle_head
, rh
);
2848 WRITE_ONCE(rshp
->inuse
, 0);
2852 * Check to see if the system is fully idle, other than the timekeeping CPU.
2853 * The caller must have disabled interrupts. This is not intended to be
2854 * called unless tick_nohz_full_enabled().
2856 bool rcu_sys_is_idle(void)
2858 static struct rcu_sysidle_head rsh
;
2859 int rss
= READ_ONCE(full_sysidle_state
);
2861 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
))
2864 /* Handle small-system case by doing a full scan of CPUs. */
2865 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
) {
2866 int oldrss
= rss
- 1;
2869 * One pass to advance to each state up to _FULL.
2870 * Give up if any pass fails to advance the state.
2872 while (rss
< RCU_SYSIDLE_FULL
&& oldrss
< rss
) {
2875 unsigned long maxj
= jiffies
- ULONG_MAX
/ 4;
2876 struct rcu_data
*rdp
;
2878 /* Scan all the CPUs looking for nonidle CPUs. */
2879 for_each_possible_cpu(cpu
) {
2880 rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
2881 rcu_sysidle_check_cpu(rdp
, &isidle
, &maxj
);
2885 rcu_sysidle_report(rcu_state_p
, isidle
, maxj
, false);
2887 rss
= READ_ONCE(full_sysidle_state
);
2891 /* If this is the first observation of an idle period, record it. */
2892 if (rss
== RCU_SYSIDLE_FULL
) {
2893 rss
= cmpxchg(&full_sysidle_state
,
2894 RCU_SYSIDLE_FULL
, RCU_SYSIDLE_FULL_NOTED
);
2895 return rss
== RCU_SYSIDLE_FULL
;
2898 smp_mb(); /* ensure rss load happens before later caller actions. */
2900 /* If already fully idle, tell the caller (in case of races). */
2901 if (rss
== RCU_SYSIDLE_FULL_NOTED
)
2905 * If we aren't there yet, and a grace period is not in flight,
2906 * initiate a grace period. Either way, tell the caller that
2907 * we are not there yet. We use an xchg() rather than an assignment
2908 * to make up for the memory barriers that would otherwise be
2909 * provided by the memory allocator.
2911 if (nr_cpu_ids
> CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
&&
2912 !rcu_gp_in_progress(rcu_state_p
) &&
2913 !rsh
.inuse
&& xchg(&rsh
.inuse
, 1) == 0)
2914 call_rcu(&rsh
.rh
, rcu_sysidle_cb
);
2919 * Initialize dynticks sysidle state for CPUs coming online.
2921 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
2923 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
;
2926 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2928 static void rcu_sysidle_enter(int irq
)
2932 static void rcu_sysidle_exit(int irq
)
2936 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
2937 unsigned long *maxj
)
2941 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
2946 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
2951 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
2955 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2958 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2959 * grace-period kthread will do force_quiescent_state() processing?
2960 * The idea is to avoid waking up RCU core processing on such a
2961 * CPU unless the grace period has extended for too long.
2963 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2964 * CONFIG_RCU_NOCB_CPU CPUs.
2966 static bool rcu_nohz_full_cpu(struct rcu_state
*rsp
)
2968 #ifdef CONFIG_NO_HZ_FULL
2969 if (tick_nohz_full_cpu(smp_processor_id()) &&
2970 (!rcu_gp_in_progress(rsp
) ||
2971 ULONG_CMP_LT(jiffies
, READ_ONCE(rsp
->gp_start
) + HZ
)))
2973 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2978 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2981 static void rcu_bind_gp_kthread(void)
2983 int __maybe_unused cpu
;
2985 if (!tick_nohz_full_enabled())
2987 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2988 cpu
= tick_do_timer_cpu
;
2989 if (cpu
>= 0 && cpu
< nr_cpu_ids
)
2990 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
2991 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2992 housekeeping_affine(current
);
2993 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2996 /* Record the current task on dyntick-idle entry. */
2997 static void rcu_dynticks_task_enter(void)
2999 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3000 WRITE_ONCE(current
->rcu_tasks_idle_cpu
, smp_processor_id());
3001 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3004 /* Record no current task on dyntick-idle exit. */
3005 static void rcu_dynticks_task_exit(void)
3007 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3008 WRITE_ONCE(current
->rcu_tasks_idle_cpu
, -1);
3009 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */