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 #define RCU_KTHREAD_PRIO 1
35 #ifdef CONFIG_RCU_BOOST
36 #include "../locking/rtmutex_common.h"
37 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
39 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
42 #ifdef CONFIG_RCU_NOCB_CPU
43 static cpumask_var_t rcu_nocb_mask
; /* CPUs to have callbacks offloaded. */
44 static bool have_rcu_nocb_mask
; /* Was rcu_nocb_mask allocated? */
45 static bool __read_mostly rcu_nocb_poll
; /* Offload kthread are to poll. */
46 static char __initdata nocb_buf
[NR_CPUS
* 5];
47 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
50 * Check the RCU kernel configuration parameters and print informative
51 * messages about anything out of the ordinary. If you like #ifdef, you
52 * will love this function.
54 static void __init
rcu_bootup_announce_oddness(void)
56 #ifdef CONFIG_RCU_TRACE
57 pr_info("\tRCU debugfs-based tracing is enabled.\n");
59 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
60 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
63 #ifdef CONFIG_RCU_FANOUT_EXACT
64 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
66 #ifdef CONFIG_RCU_FAST_NO_HZ
67 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
69 #ifdef CONFIG_PROVE_RCU
70 pr_info("\tRCU lockdep checking is enabled.\n");
72 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
73 pr_info("\tRCU torture testing starts during boot.\n");
75 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
76 pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
78 #if defined(CONFIG_RCU_CPU_STALL_INFO)
79 pr_info("\tAdditional per-CPU info printed with stalls.\n");
81 #if NUM_RCU_LVL_4 != 0
82 pr_info("\tFour-level hierarchy is enabled.\n");
84 if (rcu_fanout_leaf
!= CONFIG_RCU_FANOUT_LEAF
)
85 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf
);
86 if (nr_cpu_ids
!= NR_CPUS
)
87 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS
, nr_cpu_ids
);
88 #ifdef CONFIG_RCU_NOCB_CPU
89 #ifndef CONFIG_RCU_NOCB_CPU_NONE
90 if (!have_rcu_nocb_mask
) {
91 zalloc_cpumask_var(&rcu_nocb_mask
, GFP_KERNEL
);
92 have_rcu_nocb_mask
= true;
94 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
95 pr_info("\tOffload RCU callbacks from CPU 0\n");
96 cpumask_set_cpu(0, rcu_nocb_mask
);
97 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
98 #ifdef CONFIG_RCU_NOCB_CPU_ALL
99 pr_info("\tOffload RCU callbacks from all CPUs\n");
100 cpumask_copy(rcu_nocb_mask
, cpu_possible_mask
);
101 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
102 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
103 if (have_rcu_nocb_mask
) {
104 if (!cpumask_subset(rcu_nocb_mask
, cpu_possible_mask
)) {
105 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
106 cpumask_and(rcu_nocb_mask
, cpu_possible_mask
,
109 cpulist_scnprintf(nocb_buf
, sizeof(nocb_buf
), rcu_nocb_mask
);
110 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf
);
112 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
114 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
117 #ifdef CONFIG_TREE_PREEMPT_RCU
119 RCU_STATE_INITIALIZER(rcu_preempt
, 'p', call_rcu
);
120 static struct rcu_state
*rcu_state_p
= &rcu_preempt_state
;
122 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
);
125 * Tell them what RCU they are running.
127 static void __init
rcu_bootup_announce(void)
129 pr_info("Preemptible hierarchical RCU implementation.\n");
130 rcu_bootup_announce_oddness();
134 * Return the number of RCU-preempt batches processed thus far
135 * for debug and statistics.
137 long rcu_batches_completed_preempt(void)
139 return rcu_preempt_state
.completed
;
141 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt
);
144 * Return the number of RCU batches processed thus far for debug & stats.
146 long rcu_batches_completed(void)
148 return rcu_batches_completed_preempt();
150 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
153 * Record a preemptible-RCU quiescent state for the specified CPU. Note
154 * that this just means that the task currently running on the CPU is
155 * not in a quiescent state. There might be any number of tasks blocked
156 * while in an RCU read-side critical section.
158 * Unlike the other rcu_*_qs() functions, callers to this function
159 * must disable irqs in order to protect the assignment to
160 * ->rcu_read_unlock_special.
162 static void rcu_preempt_qs(int cpu
)
164 struct rcu_data
*rdp
= &per_cpu(rcu_preempt_data
, cpu
);
166 if (rdp
->passed_quiesce
== 0)
167 trace_rcu_grace_period(TPS("rcu_preempt"), rdp
->gpnum
, TPS("cpuqs"));
168 rdp
->passed_quiesce
= 1;
169 current
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_NEED_QS
;
173 * We have entered the scheduler, and the current task might soon be
174 * context-switched away from. If this task is in an RCU read-side
175 * critical section, we will no longer be able to rely on the CPU to
176 * record that fact, so we enqueue the task on the blkd_tasks list.
177 * The task will dequeue itself when it exits the outermost enclosing
178 * RCU read-side critical section. Therefore, the current grace period
179 * cannot be permitted to complete until the blkd_tasks list entries
180 * predating the current grace period drain, in other words, until
181 * rnp->gp_tasks becomes NULL.
183 * Caller must disable preemption.
185 static void rcu_preempt_note_context_switch(int cpu
)
187 struct task_struct
*t
= current
;
189 struct rcu_data
*rdp
;
190 struct rcu_node
*rnp
;
192 if (t
->rcu_read_lock_nesting
> 0 &&
193 (t
->rcu_read_unlock_special
& RCU_READ_UNLOCK_BLOCKED
) == 0) {
195 /* Possibly blocking in an RCU read-side critical section. */
196 rdp
= per_cpu_ptr(rcu_preempt_state
.rda
, cpu
);
198 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
199 smp_mb__after_unlock_lock();
200 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_BLOCKED
;
201 t
->rcu_blocked_node
= rnp
;
204 * If this CPU has already checked in, then this task
205 * will hold up the next grace period rather than the
206 * current grace period. Queue the task accordingly.
207 * If the task is queued for the current grace period
208 * (i.e., this CPU has not yet passed through a quiescent
209 * state for the current grace period), then as long
210 * as that task remains queued, the current grace period
211 * cannot end. Note that there is some uncertainty as
212 * to exactly when the current grace period started.
213 * We take a conservative approach, which can result
214 * in unnecessarily waiting on tasks that started very
215 * slightly after the current grace period began. C'est
218 * But first, note that the current CPU must still be
221 WARN_ON_ONCE((rdp
->grpmask
& rnp
->qsmaskinit
) == 0);
222 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
223 if ((rnp
->qsmask
& rdp
->grpmask
) && rnp
->gp_tasks
!= NULL
) {
224 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
->prev
);
225 rnp
->gp_tasks
= &t
->rcu_node_entry
;
226 #ifdef CONFIG_RCU_BOOST
227 if (rnp
->boost_tasks
!= NULL
)
228 rnp
->boost_tasks
= rnp
->gp_tasks
;
229 #endif /* #ifdef CONFIG_RCU_BOOST */
231 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
232 if (rnp
->qsmask
& rdp
->grpmask
)
233 rnp
->gp_tasks
= &t
->rcu_node_entry
;
235 trace_rcu_preempt_task(rdp
->rsp
->name
,
237 (rnp
->qsmask
& rdp
->grpmask
)
240 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
241 } else if (t
->rcu_read_lock_nesting
< 0 &&
242 t
->rcu_read_unlock_special
) {
245 * Complete exit from RCU read-side critical section on
246 * behalf of preempted instance of __rcu_read_unlock().
248 rcu_read_unlock_special(t
);
252 * Either we were not in an RCU read-side critical section to
253 * begin with, or we have now recorded that critical section
254 * globally. Either way, we can now note a quiescent state
255 * for this CPU. Again, if we were in an RCU read-side critical
256 * section, and if that critical section was blocking the current
257 * grace period, then the fact that the task has been enqueued
258 * means that we continue to block the current grace period.
260 local_irq_save(flags
);
262 local_irq_restore(flags
);
266 * Check for preempted RCU readers blocking the current grace period
267 * for the specified rcu_node structure. If the caller needs a reliable
268 * answer, it must hold the rcu_node's ->lock.
270 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
272 return rnp
->gp_tasks
!= NULL
;
276 * Record a quiescent state for all tasks that were previously queued
277 * on the specified rcu_node structure and that were blocking the current
278 * RCU grace period. The caller must hold the specified rnp->lock with
279 * irqs disabled, and this lock is released upon return, but irqs remain
282 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
283 __releases(rnp
->lock
)
286 struct rcu_node
*rnp_p
;
288 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
289 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
290 return; /* Still need more quiescent states! */
296 * Either there is only one rcu_node in the tree,
297 * or tasks were kicked up to root rcu_node due to
298 * CPUs going offline.
300 rcu_report_qs_rsp(&rcu_preempt_state
, flags
);
304 /* Report up the rest of the hierarchy. */
306 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
307 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
308 smp_mb__after_unlock_lock();
309 rcu_report_qs_rnp(mask
, &rcu_preempt_state
, rnp_p
, flags
);
313 * Advance a ->blkd_tasks-list pointer to the next entry, instead
314 * returning NULL if at the end of the list.
316 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
317 struct rcu_node
*rnp
)
319 struct list_head
*np
;
321 np
= t
->rcu_node_entry
.next
;
322 if (np
== &rnp
->blkd_tasks
)
328 * Handle special cases during rcu_read_unlock(), such as needing to
329 * notify RCU core processing or task having blocked during the RCU
330 * read-side critical section.
332 void rcu_read_unlock_special(struct task_struct
*t
)
338 struct list_head
*np
;
339 #ifdef CONFIG_RCU_BOOST
340 bool drop_boost_mutex
= false;
341 #endif /* #ifdef CONFIG_RCU_BOOST */
342 struct rcu_node
*rnp
;
345 /* NMI handlers cannot block and cannot safely manipulate state. */
349 local_irq_save(flags
);
352 * If RCU core is waiting for this CPU to exit critical section,
353 * let it know that we have done so.
355 special
= t
->rcu_read_unlock_special
;
356 if (special
& RCU_READ_UNLOCK_NEED_QS
) {
357 rcu_preempt_qs(smp_processor_id());
358 if (!t
->rcu_read_unlock_special
) {
359 local_irq_restore(flags
);
364 /* Hardware IRQ handlers cannot block, complain if they get here. */
365 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
366 local_irq_restore(flags
);
370 /* Clean up if blocked during RCU read-side critical section. */
371 if (special
& RCU_READ_UNLOCK_BLOCKED
) {
372 t
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_BLOCKED
;
375 * Remove this task from the list it blocked on. The
376 * task can migrate while we acquire the lock, but at
377 * most one time. So at most two passes through loop.
380 rnp
= t
->rcu_blocked_node
;
381 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
382 smp_mb__after_unlock_lock();
383 if (rnp
== t
->rcu_blocked_node
)
385 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
387 empty
= !rcu_preempt_blocked_readers_cgp(rnp
);
388 empty_exp
= !rcu_preempted_readers_exp(rnp
);
389 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
390 np
= rcu_next_node_entry(t
, rnp
);
391 list_del_init(&t
->rcu_node_entry
);
392 t
->rcu_blocked_node
= NULL
;
393 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
395 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
397 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
399 #ifdef CONFIG_RCU_BOOST
400 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
401 rnp
->boost_tasks
= np
;
402 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
403 drop_boost_mutex
= rt_mutex_owner(&rnp
->boost_mtx
) == t
;
404 #endif /* #ifdef CONFIG_RCU_BOOST */
407 * If this was the last task on the current list, and if
408 * we aren't waiting on any CPUs, report the quiescent state.
409 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
410 * so we must take a snapshot of the expedited state.
412 empty_exp_now
= !rcu_preempted_readers_exp(rnp
);
413 if (!empty
&& !rcu_preempt_blocked_readers_cgp(rnp
)) {
414 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
421 rcu_report_unblock_qs_rnp(rnp
, flags
);
423 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
426 #ifdef CONFIG_RCU_BOOST
427 /* Unboost if we were boosted. */
428 if (drop_boost_mutex
) {
429 rt_mutex_unlock(&rnp
->boost_mtx
);
430 complete(&rnp
->boost_completion
);
432 #endif /* #ifdef CONFIG_RCU_BOOST */
435 * If this was the last task on the expedited lists,
436 * then we need to report up the rcu_node hierarchy.
438 if (!empty_exp
&& empty_exp_now
)
439 rcu_report_exp_rnp(&rcu_preempt_state
, rnp
, true);
441 local_irq_restore(flags
);
445 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
448 * Dump detailed information for all tasks blocking the current RCU
449 * grace period on the specified rcu_node structure.
451 static void rcu_print_detail_task_stall_rnp(struct rcu_node
*rnp
)
454 struct task_struct
*t
;
456 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
457 if (!rcu_preempt_blocked_readers_cgp(rnp
)) {
458 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
461 t
= list_entry(rnp
->gp_tasks
,
462 struct task_struct
, rcu_node_entry
);
463 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
465 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
469 * Dump detailed information for all tasks blocking the current RCU
472 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
474 struct rcu_node
*rnp
= rcu_get_root(rsp
);
476 rcu_print_detail_task_stall_rnp(rnp
);
477 rcu_for_each_leaf_node(rsp
, rnp
)
478 rcu_print_detail_task_stall_rnp(rnp
);
481 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
483 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
487 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
489 #ifdef CONFIG_RCU_CPU_STALL_INFO
491 static void rcu_print_task_stall_begin(struct rcu_node
*rnp
)
493 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
494 rnp
->level
, rnp
->grplo
, rnp
->grphi
);
497 static void rcu_print_task_stall_end(void)
502 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
504 static void rcu_print_task_stall_begin(struct rcu_node
*rnp
)
508 static void rcu_print_task_stall_end(void)
512 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
515 * Scan the current list of tasks blocked within RCU read-side critical
516 * sections, printing out the tid of each.
518 static int rcu_print_task_stall(struct rcu_node
*rnp
)
520 struct task_struct
*t
;
523 if (!rcu_preempt_blocked_readers_cgp(rnp
))
525 rcu_print_task_stall_begin(rnp
);
526 t
= list_entry(rnp
->gp_tasks
,
527 struct task_struct
, rcu_node_entry
);
528 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
) {
529 pr_cont(" P%d", t
->pid
);
532 rcu_print_task_stall_end();
537 * Check that the list of blocked tasks for the newly completed grace
538 * period is in fact empty. It is a serious bug to complete a grace
539 * period that still has RCU readers blocked! This function must be
540 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
541 * must be held by the caller.
543 * Also, if there are blocked tasks on the list, they automatically
544 * block the newly created grace period, so set up ->gp_tasks accordingly.
546 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
548 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
549 if (!list_empty(&rnp
->blkd_tasks
))
550 rnp
->gp_tasks
= rnp
->blkd_tasks
.next
;
551 WARN_ON_ONCE(rnp
->qsmask
);
554 #ifdef CONFIG_HOTPLUG_CPU
557 * Handle tasklist migration for case in which all CPUs covered by the
558 * specified rcu_node have gone offline. Move them up to the root
559 * rcu_node. The reason for not just moving them to the immediate
560 * parent is to remove the need for rcu_read_unlock_special() to
561 * make more than two attempts to acquire the target rcu_node's lock.
562 * Returns true if there were tasks blocking the current RCU grace
565 * Returns 1 if there was previously a task blocking the current grace
566 * period on the specified rcu_node structure.
568 * The caller must hold rnp->lock with irqs disabled.
570 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
571 struct rcu_node
*rnp
,
572 struct rcu_data
*rdp
)
574 struct list_head
*lp
;
575 struct list_head
*lp_root
;
577 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
578 struct task_struct
*t
;
580 if (rnp
== rnp_root
) {
581 WARN_ONCE(1, "Last CPU thought to be offlined?");
582 return 0; /* Shouldn't happen: at least one CPU online. */
585 /* If we are on an internal node, complain bitterly. */
586 WARN_ON_ONCE(rnp
!= rdp
->mynode
);
589 * Move tasks up to root rcu_node. Don't try to get fancy for
590 * this corner-case operation -- just put this node's tasks
591 * at the head of the root node's list, and update the root node's
592 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
593 * if non-NULL. This might result in waiting for more tasks than
594 * absolutely necessary, but this is a good performance/complexity
597 if (rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->qsmask
== 0)
598 retval
|= RCU_OFL_TASKS_NORM_GP
;
599 if (rcu_preempted_readers_exp(rnp
))
600 retval
|= RCU_OFL_TASKS_EXP_GP
;
601 lp
= &rnp
->blkd_tasks
;
602 lp_root
= &rnp_root
->blkd_tasks
;
603 while (!list_empty(lp
)) {
604 t
= list_entry(lp
->next
, typeof(*t
), rcu_node_entry
);
605 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
606 smp_mb__after_unlock_lock();
607 list_del(&t
->rcu_node_entry
);
608 t
->rcu_blocked_node
= rnp_root
;
609 list_add(&t
->rcu_node_entry
, lp_root
);
610 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
611 rnp_root
->gp_tasks
= rnp
->gp_tasks
;
612 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
613 rnp_root
->exp_tasks
= rnp
->exp_tasks
;
614 #ifdef CONFIG_RCU_BOOST
615 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
616 rnp_root
->boost_tasks
= rnp
->boost_tasks
;
617 #endif /* #ifdef CONFIG_RCU_BOOST */
618 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
621 rnp
->gp_tasks
= NULL
;
622 rnp
->exp_tasks
= NULL
;
623 #ifdef CONFIG_RCU_BOOST
624 rnp
->boost_tasks
= NULL
;
626 * In case root is being boosted and leaf was not. Make sure
627 * that we boost the tasks blocking the current grace period
630 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
631 smp_mb__after_unlock_lock();
632 if (rnp_root
->boost_tasks
!= NULL
&&
633 rnp_root
->boost_tasks
!= rnp_root
->gp_tasks
&&
634 rnp_root
->boost_tasks
!= rnp_root
->exp_tasks
)
635 rnp_root
->boost_tasks
= rnp_root
->gp_tasks
;
636 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
637 #endif /* #ifdef CONFIG_RCU_BOOST */
642 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
645 * Check for a quiescent state from the current CPU. When a task blocks,
646 * the task is recorded in the corresponding CPU's rcu_node structure,
647 * which is checked elsewhere.
649 * Caller must disable hard irqs.
651 static void rcu_preempt_check_callbacks(int cpu
)
653 struct task_struct
*t
= current
;
655 if (t
->rcu_read_lock_nesting
== 0) {
659 if (t
->rcu_read_lock_nesting
> 0 &&
660 per_cpu(rcu_preempt_data
, cpu
).qs_pending
)
661 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_NEED_QS
;
664 #ifdef CONFIG_RCU_BOOST
666 static void rcu_preempt_do_callbacks(void)
668 rcu_do_batch(&rcu_preempt_state
, this_cpu_ptr(&rcu_preempt_data
));
671 #endif /* #ifdef CONFIG_RCU_BOOST */
674 * Queue a preemptible-RCU callback for invocation after a grace period.
676 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
678 __call_rcu(head
, func
, &rcu_preempt_state
, -1, 0);
680 EXPORT_SYMBOL_GPL(call_rcu
);
683 * synchronize_rcu - wait until a grace period has elapsed.
685 * Control will return to the caller some time after a full grace
686 * period has elapsed, in other words after all currently executing RCU
687 * read-side critical sections have completed. Note, however, that
688 * upon return from synchronize_rcu(), the caller might well be executing
689 * concurrently with new RCU read-side critical sections that began while
690 * synchronize_rcu() was waiting. RCU read-side critical sections are
691 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
693 * See the description of synchronize_sched() for more detailed information
694 * on memory ordering guarantees.
696 void synchronize_rcu(void)
698 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
699 !lock_is_held(&rcu_lock_map
) &&
700 !lock_is_held(&rcu_sched_lock_map
),
701 "Illegal synchronize_rcu() in RCU read-side critical section");
702 if (!rcu_scheduler_active
)
705 synchronize_rcu_expedited();
707 wait_rcu_gp(call_rcu
);
709 EXPORT_SYMBOL_GPL(synchronize_rcu
);
711 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq
);
712 static unsigned long sync_rcu_preempt_exp_count
;
713 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex
);
716 * Return non-zero if there are any tasks in RCU read-side critical
717 * sections blocking the current preemptible-RCU expedited grace period.
718 * If there is no preemptible-RCU expedited grace period currently in
719 * progress, returns zero unconditionally.
721 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
)
723 return rnp
->exp_tasks
!= NULL
;
727 * return non-zero if there is no RCU expedited grace period in progress
728 * for the specified rcu_node structure, in other words, if all CPUs and
729 * tasks covered by the specified rcu_node structure have done their bit
730 * for the current expedited grace period. Works only for preemptible
731 * RCU -- other RCU implementation use other means.
733 * Caller must hold sync_rcu_preempt_exp_mutex.
735 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
737 return !rcu_preempted_readers_exp(rnp
) &&
738 ACCESS_ONCE(rnp
->expmask
) == 0;
742 * Report the exit from RCU read-side critical section for the last task
743 * that queued itself during or before the current expedited preemptible-RCU
744 * grace period. This event is reported either to the rcu_node structure on
745 * which the task was queued or to one of that rcu_node structure's ancestors,
746 * recursively up the tree. (Calm down, calm down, we do the recursion
749 * Most callers will set the "wake" flag, but the task initiating the
750 * expedited grace period need not wake itself.
752 * Caller must hold sync_rcu_preempt_exp_mutex.
754 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
760 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
761 smp_mb__after_unlock_lock();
763 if (!sync_rcu_preempt_exp_done(rnp
)) {
764 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
767 if (rnp
->parent
== NULL
) {
768 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
770 smp_mb(); /* EGP done before wake_up(). */
771 wake_up(&sync_rcu_preempt_exp_wq
);
776 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
778 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
779 smp_mb__after_unlock_lock();
780 rnp
->expmask
&= ~mask
;
785 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
786 * grace period for the specified rcu_node structure. If there are no such
787 * tasks, report it up the rcu_node hierarchy.
789 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
790 * CPU hotplug operations.
793 sync_rcu_preempt_exp_init(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
798 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
799 smp_mb__after_unlock_lock();
800 if (list_empty(&rnp
->blkd_tasks
)) {
801 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
803 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
804 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
808 rcu_report_exp_rnp(rsp
, rnp
, false); /* Don't wake self. */
812 * synchronize_rcu_expedited - Brute-force RCU grace period
814 * Wait for an RCU-preempt grace period, but expedite it. The basic
815 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
816 * the ->blkd_tasks lists and wait for this list to drain. This consumes
817 * significant time on all CPUs and is unfriendly to real-time workloads,
818 * so is thus not recommended for any sort of common-case code.
819 * In fact, if you are using synchronize_rcu_expedited() in a loop,
820 * please restructure your code to batch your updates, and then Use a
821 * single synchronize_rcu() instead.
823 * Note that it is illegal to call this function while holding any lock
824 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
825 * to call this function from a CPU-hotplug notifier. Failing to observe
826 * these restriction will result in deadlock.
828 void synchronize_rcu_expedited(void)
831 struct rcu_node
*rnp
;
832 struct rcu_state
*rsp
= &rcu_preempt_state
;
836 smp_mb(); /* Caller's modifications seen first by other CPUs. */
837 snap
= ACCESS_ONCE(sync_rcu_preempt_exp_count
) + 1;
838 smp_mb(); /* Above access cannot bleed into critical section. */
841 * Block CPU-hotplug operations. This means that any CPU-hotplug
842 * operation that finds an rcu_node structure with tasks in the
843 * process of being boosted will know that all tasks blocking
844 * this expedited grace period will already be in the process of
845 * being boosted. This simplifies the process of moving tasks
846 * from leaf to root rcu_node structures.
851 * Acquire lock, falling back to synchronize_rcu() if too many
852 * lock-acquisition failures. Of course, if someone does the
853 * expedited grace period for us, just leave.
855 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex
)) {
856 if (ULONG_CMP_LT(snap
,
857 ACCESS_ONCE(sync_rcu_preempt_exp_count
))) {
859 goto mb_ret
; /* Others did our work for us. */
861 if (trycount
++ < 10) {
862 udelay(trycount
* num_online_cpus());
865 wait_rcu_gp(call_rcu
);
869 if (ULONG_CMP_LT(snap
, ACCESS_ONCE(sync_rcu_preempt_exp_count
))) {
871 goto unlock_mb_ret
; /* Others did our work for us. */
874 /* force all RCU readers onto ->blkd_tasks lists. */
875 synchronize_sched_expedited();
877 /* Initialize ->expmask for all non-leaf rcu_node structures. */
878 rcu_for_each_nonleaf_node_breadth_first(rsp
, rnp
) {
879 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
880 smp_mb__after_unlock_lock();
881 rnp
->expmask
= rnp
->qsmaskinit
;
882 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
885 /* Snapshot current state of ->blkd_tasks lists. */
886 rcu_for_each_leaf_node(rsp
, rnp
)
887 sync_rcu_preempt_exp_init(rsp
, rnp
);
888 if (NUM_RCU_NODES
> 1)
889 sync_rcu_preempt_exp_init(rsp
, rcu_get_root(rsp
));
893 /* Wait for snapshotted ->blkd_tasks lists to drain. */
894 rnp
= rcu_get_root(rsp
);
895 wait_event(sync_rcu_preempt_exp_wq
,
896 sync_rcu_preempt_exp_done(rnp
));
898 /* Clean up and exit. */
899 smp_mb(); /* ensure expedited GP seen before counter increment. */
900 ACCESS_ONCE(sync_rcu_preempt_exp_count
) =
901 sync_rcu_preempt_exp_count
+ 1;
903 mutex_unlock(&sync_rcu_preempt_exp_mutex
);
905 smp_mb(); /* ensure subsequent action seen after grace period. */
907 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
910 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
912 * Note that this primitive does not necessarily wait for an RCU grace period
913 * to complete. For example, if there are no RCU callbacks queued anywhere
914 * in the system, then rcu_barrier() is within its rights to return
915 * immediately, without waiting for anything, much less an RCU grace period.
917 void rcu_barrier(void)
919 _rcu_barrier(&rcu_preempt_state
);
921 EXPORT_SYMBOL_GPL(rcu_barrier
);
924 * Initialize preemptible RCU's state structures.
926 static void __init
__rcu_init_preempt(void)
928 rcu_init_one(&rcu_preempt_state
, &rcu_preempt_data
);
932 * Check for a task exiting while in a preemptible-RCU read-side
933 * critical section, clean up if so. No need to issue warnings,
934 * as debug_check_no_locks_held() already does this if lockdep
939 struct task_struct
*t
= current
;
941 if (likely(list_empty(¤t
->rcu_node_entry
)))
943 t
->rcu_read_lock_nesting
= 1;
945 t
->rcu_read_unlock_special
= RCU_READ_UNLOCK_BLOCKED
;
949 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
951 static struct rcu_state
*rcu_state_p
= &rcu_sched_state
;
954 * Tell them what RCU they are running.
956 static void __init
rcu_bootup_announce(void)
958 pr_info("Hierarchical RCU implementation.\n");
959 rcu_bootup_announce_oddness();
963 * Return the number of RCU batches processed thus far for debug & stats.
965 long rcu_batches_completed(void)
967 return rcu_batches_completed_sched();
969 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
972 * Because preemptible RCU does not exist, we never have to check for
973 * CPUs being in quiescent states.
975 static void rcu_preempt_note_context_switch(int cpu
)
980 * Because preemptible RCU does not exist, there are never any preempted
983 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
988 #ifdef CONFIG_HOTPLUG_CPU
990 /* Because preemptible RCU does not exist, no quieting of tasks. */
991 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
992 __releases(rnp
->lock
)
994 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
997 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1000 * Because preemptible RCU does not exist, we never have to check for
1001 * tasks blocked within RCU read-side critical sections.
1003 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
1008 * Because preemptible RCU does not exist, we never have to check for
1009 * tasks blocked within RCU read-side critical sections.
1011 static int rcu_print_task_stall(struct rcu_node
*rnp
)
1017 * Because there is no preemptible RCU, there can be no readers blocked,
1018 * so there is no need to check for blocked tasks. So check only for
1019 * bogus qsmask values.
1021 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
1023 WARN_ON_ONCE(rnp
->qsmask
);
1026 #ifdef CONFIG_HOTPLUG_CPU
1029 * Because preemptible RCU does not exist, it never needs to migrate
1030 * tasks that were blocked within RCU read-side critical sections, and
1031 * such non-existent tasks cannot possibly have been blocking the current
1034 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
1035 struct rcu_node
*rnp
,
1036 struct rcu_data
*rdp
)
1041 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1044 * Because preemptible RCU does not exist, it never has any callbacks
1047 static void rcu_preempt_check_callbacks(int cpu
)
1052 * Wait for an rcu-preempt grace period, but make it happen quickly.
1053 * But because preemptible RCU does not exist, map to rcu-sched.
1055 void synchronize_rcu_expedited(void)
1057 synchronize_sched_expedited();
1059 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
1061 #ifdef CONFIG_HOTPLUG_CPU
1064 * Because preemptible RCU does not exist, there is never any need to
1065 * report on tasks preempted in RCU read-side critical sections during
1066 * expedited RCU grace periods.
1068 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1073 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1076 * Because preemptible RCU does not exist, rcu_barrier() is just
1077 * another name for rcu_barrier_sched().
1079 void rcu_barrier(void)
1081 rcu_barrier_sched();
1083 EXPORT_SYMBOL_GPL(rcu_barrier
);
1086 * Because preemptible RCU does not exist, it need not be initialized.
1088 static void __init
__rcu_init_preempt(void)
1093 * Because preemptible RCU does not exist, tasks cannot possibly exit
1094 * while in preemptible RCU read-side critical sections.
1100 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1102 #ifdef CONFIG_RCU_BOOST
1104 #include "../locking/rtmutex_common.h"
1106 #ifdef CONFIG_RCU_TRACE
1108 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1110 if (list_empty(&rnp
->blkd_tasks
))
1111 rnp
->n_balk_blkd_tasks
++;
1112 else if (rnp
->exp_tasks
== NULL
&& rnp
->gp_tasks
== NULL
)
1113 rnp
->n_balk_exp_gp_tasks
++;
1114 else if (rnp
->gp_tasks
!= NULL
&& rnp
->boost_tasks
!= NULL
)
1115 rnp
->n_balk_boost_tasks
++;
1116 else if (rnp
->gp_tasks
!= NULL
&& rnp
->qsmask
!= 0)
1117 rnp
->n_balk_notblocked
++;
1118 else if (rnp
->gp_tasks
!= NULL
&&
1119 ULONG_CMP_LT(jiffies
, rnp
->boost_time
))
1120 rnp
->n_balk_notyet
++;
1125 #else /* #ifdef CONFIG_RCU_TRACE */
1127 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1131 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1133 static void rcu_wake_cond(struct task_struct
*t
, int status
)
1136 * If the thread is yielding, only wake it when this
1137 * is invoked from idle
1139 if (status
!= RCU_KTHREAD_YIELDING
|| is_idle_task(current
))
1144 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1145 * or ->boost_tasks, advancing the pointer to the next task in the
1146 * ->blkd_tasks list.
1148 * Note that irqs must be enabled: boosting the task can block.
1149 * Returns 1 if there are more tasks needing to be boosted.
1151 static int rcu_boost(struct rcu_node
*rnp
)
1153 unsigned long flags
;
1154 struct task_struct
*t
;
1155 struct list_head
*tb
;
1157 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
)
1158 return 0; /* Nothing left to boost. */
1160 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1161 smp_mb__after_unlock_lock();
1164 * Recheck under the lock: all tasks in need of boosting
1165 * might exit their RCU read-side critical sections on their own.
1167 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
1168 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1173 * Preferentially boost tasks blocking expedited grace periods.
1174 * This cannot starve the normal grace periods because a second
1175 * expedited grace period must boost all blocked tasks, including
1176 * those blocking the pre-existing normal grace period.
1178 if (rnp
->exp_tasks
!= NULL
) {
1179 tb
= rnp
->exp_tasks
;
1180 rnp
->n_exp_boosts
++;
1182 tb
= rnp
->boost_tasks
;
1183 rnp
->n_normal_boosts
++;
1185 rnp
->n_tasks_boosted
++;
1188 * We boost task t by manufacturing an rt_mutex that appears to
1189 * be held by task t. We leave a pointer to that rt_mutex where
1190 * task t can find it, and task t will release the mutex when it
1191 * exits its outermost RCU read-side critical section. Then
1192 * simply acquiring this artificial rt_mutex will boost task
1193 * t's priority. (Thanks to tglx for suggesting this approach!)
1195 * Note that task t must acquire rnp->lock to remove itself from
1196 * the ->blkd_tasks list, which it will do from exit() if from
1197 * nowhere else. We therefore are guaranteed that task t will
1198 * stay around at least until we drop rnp->lock. Note that
1199 * rnp->lock also resolves races between our priority boosting
1200 * and task t's exiting its outermost RCU read-side critical
1203 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1204 rt_mutex_init_proxy_locked(&rnp
->boost_mtx
, t
);
1205 init_completion(&rnp
->boost_completion
);
1206 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1207 /* Lock only for side effect: boosts task t's priority. */
1208 rt_mutex_lock(&rnp
->boost_mtx
);
1209 rt_mutex_unlock(&rnp
->boost_mtx
); /* Then keep lockdep happy. */
1211 /* Wait for boostee to be done w/boost_mtx before reinitializing. */
1212 wait_for_completion(&rnp
->boost_completion
);
1214 return ACCESS_ONCE(rnp
->exp_tasks
) != NULL
||
1215 ACCESS_ONCE(rnp
->boost_tasks
) != NULL
;
1219 * Priority-boosting kthread. One per leaf rcu_node and one for the
1222 static int rcu_boost_kthread(void *arg
)
1224 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1228 trace_rcu_utilization(TPS("Start boost kthread@init"));
1230 rnp
->boost_kthread_status
= RCU_KTHREAD_WAITING
;
1231 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1232 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1233 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1234 rnp
->boost_kthread_status
= RCU_KTHREAD_RUNNING
;
1235 more2boost
= rcu_boost(rnp
);
1241 rnp
->boost_kthread_status
= RCU_KTHREAD_YIELDING
;
1242 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1243 schedule_timeout_interruptible(2);
1244 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1249 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1254 * Check to see if it is time to start boosting RCU readers that are
1255 * blocking the current grace period, and, if so, tell the per-rcu_node
1256 * kthread to start boosting them. If there is an expedited grace
1257 * period in progress, it is always time to boost.
1259 * The caller must hold rnp->lock, which this function releases.
1260 * The ->boost_kthread_task is immortal, so we don't need to worry
1261 * about it going away.
1263 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1264 __releases(rnp
->lock
)
1266 struct task_struct
*t
;
1268 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1269 rnp
->n_balk_exp_gp_tasks
++;
1270 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1273 if (rnp
->exp_tasks
!= NULL
||
1274 (rnp
->gp_tasks
!= NULL
&&
1275 rnp
->boost_tasks
== NULL
&&
1277 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))) {
1278 if (rnp
->exp_tasks
== NULL
)
1279 rnp
->boost_tasks
= rnp
->gp_tasks
;
1280 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1281 t
= rnp
->boost_kthread_task
;
1283 rcu_wake_cond(t
, rnp
->boost_kthread_status
);
1285 rcu_initiate_boost_trace(rnp
);
1286 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1291 * Wake up the per-CPU kthread to invoke RCU callbacks.
1293 static void invoke_rcu_callbacks_kthread(void)
1295 unsigned long flags
;
1297 local_irq_save(flags
);
1298 __this_cpu_write(rcu_cpu_has_work
, 1);
1299 if (__this_cpu_read(rcu_cpu_kthread_task
) != NULL
&&
1300 current
!= __this_cpu_read(rcu_cpu_kthread_task
)) {
1301 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task
),
1302 __this_cpu_read(rcu_cpu_kthread_status
));
1304 local_irq_restore(flags
);
1308 * Is the current CPU running the RCU-callbacks kthread?
1309 * Caller must have preemption disabled.
1311 static bool rcu_is_callbacks_kthread(void)
1313 return __this_cpu_read(rcu_cpu_kthread_task
) == current
;
1316 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1319 * Do priority-boost accounting for the start of a new grace period.
1321 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1323 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1327 * Create an RCU-boost kthread for the specified node if one does not
1328 * already exist. We only create this kthread for preemptible RCU.
1329 * Returns zero if all is well, a negated errno otherwise.
1331 static int rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1332 struct rcu_node
*rnp
)
1334 int rnp_index
= rnp
- &rsp
->node
[0];
1335 unsigned long flags
;
1336 struct sched_param sp
;
1337 struct task_struct
*t
;
1339 if (&rcu_preempt_state
!= rsp
)
1342 if (!rcu_scheduler_fully_active
|| rnp
->qsmaskinit
== 0)
1346 if (rnp
->boost_kthread_task
!= NULL
)
1348 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1349 "rcub/%d", rnp_index
);
1352 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1353 smp_mb__after_unlock_lock();
1354 rnp
->boost_kthread_task
= t
;
1355 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1356 sp
.sched_priority
= RCU_BOOST_PRIO
;
1357 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1358 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1362 static void rcu_kthread_do_work(void)
1364 rcu_do_batch(&rcu_sched_state
, this_cpu_ptr(&rcu_sched_data
));
1365 rcu_do_batch(&rcu_bh_state
, this_cpu_ptr(&rcu_bh_data
));
1366 rcu_preempt_do_callbacks();
1369 static void rcu_cpu_kthread_setup(unsigned int cpu
)
1371 struct sched_param sp
;
1373 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1374 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1377 static void rcu_cpu_kthread_park(unsigned int cpu
)
1379 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1382 static int rcu_cpu_kthread_should_run(unsigned int cpu
)
1384 return __this_cpu_read(rcu_cpu_has_work
);
1388 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1389 * RCU softirq used in flavors and configurations of RCU that do not
1390 * support RCU priority boosting.
1392 static void rcu_cpu_kthread(unsigned int cpu
)
1394 unsigned int *statusp
= this_cpu_ptr(&rcu_cpu_kthread_status
);
1395 char work
, *workp
= this_cpu_ptr(&rcu_cpu_has_work
);
1398 for (spincnt
= 0; spincnt
< 10; spincnt
++) {
1399 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1401 *statusp
= RCU_KTHREAD_RUNNING
;
1402 this_cpu_inc(rcu_cpu_kthread_loops
);
1403 local_irq_disable();
1408 rcu_kthread_do_work();
1411 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1412 *statusp
= RCU_KTHREAD_WAITING
;
1416 *statusp
= RCU_KTHREAD_YIELDING
;
1417 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1418 schedule_timeout_interruptible(2);
1419 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1420 *statusp
= RCU_KTHREAD_WAITING
;
1424 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1425 * served by the rcu_node in question. The CPU hotplug lock is still
1426 * held, so the value of rnp->qsmaskinit will be stable.
1428 * We don't include outgoingcpu in the affinity set, use -1 if there is
1429 * no outgoing CPU. If there are no CPUs left in the affinity set,
1430 * this function allows the kthread to execute on any CPU.
1432 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1434 struct task_struct
*t
= rnp
->boost_kthread_task
;
1435 unsigned long mask
= rnp
->qsmaskinit
;
1441 if (!zalloc_cpumask_var(&cm
, GFP_KERNEL
))
1443 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1444 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1445 cpumask_set_cpu(cpu
, cm
);
1446 if (cpumask_weight(cm
) == 0) {
1448 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1449 cpumask_clear_cpu(cpu
, cm
);
1450 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1452 set_cpus_allowed_ptr(t
, cm
);
1453 free_cpumask_var(cm
);
1456 static struct smp_hotplug_thread rcu_cpu_thread_spec
= {
1457 .store
= &rcu_cpu_kthread_task
,
1458 .thread_should_run
= rcu_cpu_kthread_should_run
,
1459 .thread_fn
= rcu_cpu_kthread
,
1460 .thread_comm
= "rcuc/%u",
1461 .setup
= rcu_cpu_kthread_setup
,
1462 .park
= rcu_cpu_kthread_park
,
1466 * Spawn all kthreads -- called as soon as the scheduler is running.
1468 static int __init
rcu_spawn_kthreads(void)
1470 struct rcu_node
*rnp
;
1473 rcu_scheduler_fully_active
= 1;
1474 for_each_possible_cpu(cpu
)
1475 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1476 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec
));
1477 rnp
= rcu_get_root(rcu_state_p
);
1478 (void)rcu_spawn_one_boost_kthread(rcu_state_p
, rnp
);
1479 if (NUM_RCU_NODES
> 1) {
1480 rcu_for_each_leaf_node(rcu_state_p
, rnp
)
1481 (void)rcu_spawn_one_boost_kthread(rcu_state_p
, rnp
);
1485 early_initcall(rcu_spawn_kthreads
);
1487 static void rcu_prepare_kthreads(int cpu
)
1489 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
1490 struct rcu_node
*rnp
= rdp
->mynode
;
1492 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1493 if (rcu_scheduler_fully_active
)
1494 (void)rcu_spawn_one_boost_kthread(rcu_state_p
, rnp
);
1497 #else /* #ifdef CONFIG_RCU_BOOST */
1499 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1500 __releases(rnp
->lock
)
1502 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1505 static void invoke_rcu_callbacks_kthread(void)
1510 static bool rcu_is_callbacks_kthread(void)
1515 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1519 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1523 static int __init
rcu_scheduler_really_started(void)
1525 rcu_scheduler_fully_active
= 1;
1528 early_initcall(rcu_scheduler_really_started
);
1530 static void rcu_prepare_kthreads(int cpu
)
1534 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1536 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1539 * Check to see if any future RCU-related work will need to be done
1540 * by the current CPU, even if none need be done immediately, returning
1541 * 1 if so. This function is part of the RCU implementation; it is -not-
1542 * an exported member of the RCU API.
1544 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1545 * any flavor of RCU.
1547 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1548 int rcu_needs_cpu(int cpu
, unsigned long *delta_jiffies
)
1550 *delta_jiffies
= ULONG_MAX
;
1551 return rcu_cpu_has_callbacks(cpu
, NULL
);
1553 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1556 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1559 static void rcu_cleanup_after_idle(int cpu
)
1564 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1567 static void rcu_prepare_for_idle(int cpu
)
1572 * Don't bother keeping a running count of the number of RCU callbacks
1573 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1575 static void rcu_idle_count_callbacks_posted(void)
1579 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1582 * This code is invoked when a CPU goes idle, at which point we want
1583 * to have the CPU do everything required for RCU so that it can enter
1584 * the energy-efficient dyntick-idle mode. This is handled by a
1585 * state machine implemented by rcu_prepare_for_idle() below.
1587 * The following three proprocessor symbols control this state machine:
1589 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1590 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1591 * is sized to be roughly one RCU grace period. Those energy-efficiency
1592 * benchmarkers who might otherwise be tempted to set this to a large
1593 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1594 * system. And if you are -that- concerned about energy efficiency,
1595 * just power the system down and be done with it!
1596 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1597 * permitted to sleep in dyntick-idle mode with only lazy RCU
1598 * callbacks pending. Setting this too high can OOM your system.
1600 * The values below work well in practice. If future workloads require
1601 * adjustment, they can be converted into kernel config parameters, though
1602 * making the state machine smarter might be a better option.
1604 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1605 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1607 static int rcu_idle_gp_delay
= RCU_IDLE_GP_DELAY
;
1608 module_param(rcu_idle_gp_delay
, int, 0644);
1609 static int rcu_idle_lazy_gp_delay
= RCU_IDLE_LAZY_GP_DELAY
;
1610 module_param(rcu_idle_lazy_gp_delay
, int, 0644);
1612 extern int tick_nohz_active
;
1615 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1616 * only if it has been awhile since the last time we did so. Afterwards,
1617 * if there are any callbacks ready for immediate invocation, return true.
1619 static bool __maybe_unused
rcu_try_advance_all_cbs(void)
1621 bool cbs_ready
= false;
1622 struct rcu_data
*rdp
;
1623 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1624 struct rcu_node
*rnp
;
1625 struct rcu_state
*rsp
;
1627 /* Exit early if we advanced recently. */
1628 if (jiffies
== rdtp
->last_advance_all
)
1630 rdtp
->last_advance_all
= jiffies
;
1632 for_each_rcu_flavor(rsp
) {
1633 rdp
= this_cpu_ptr(rsp
->rda
);
1637 * Don't bother checking unless a grace period has
1638 * completed since we last checked and there are
1639 * callbacks not yet ready to invoke.
1641 if (rdp
->completed
!= rnp
->completed
&&
1642 rdp
->nxttail
[RCU_DONE_TAIL
] != rdp
->nxttail
[RCU_NEXT_TAIL
])
1643 note_gp_changes(rsp
, rdp
);
1645 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1652 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1653 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1654 * caller to set the timeout based on whether or not there are non-lazy
1657 * The caller must have disabled interrupts.
1659 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1660 int rcu_needs_cpu(int cpu
, unsigned long *dj
)
1662 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1664 /* Snapshot to detect later posting of non-lazy callback. */
1665 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1667 /* If no callbacks, RCU doesn't need the CPU. */
1668 if (!rcu_cpu_has_callbacks(cpu
, &rdtp
->all_lazy
)) {
1673 /* Attempt to advance callbacks. */
1674 if (rcu_try_advance_all_cbs()) {
1675 /* Some ready to invoke, so initiate later invocation. */
1679 rdtp
->last_accelerate
= jiffies
;
1681 /* Request timer delay depending on laziness, and round. */
1682 if (!rdtp
->all_lazy
) {
1683 *dj
= round_up(rcu_idle_gp_delay
+ jiffies
,
1684 rcu_idle_gp_delay
) - jiffies
;
1686 *dj
= round_jiffies(rcu_idle_lazy_gp_delay
+ jiffies
) - jiffies
;
1690 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1693 * Prepare a CPU for idle from an RCU perspective. The first major task
1694 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1695 * The second major task is to check to see if a non-lazy callback has
1696 * arrived at a CPU that previously had only lazy callbacks. The third
1697 * major task is to accelerate (that is, assign grace-period numbers to)
1698 * any recently arrived callbacks.
1700 * The caller must have disabled interrupts.
1702 static void rcu_prepare_for_idle(int cpu
)
1704 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1706 struct rcu_data
*rdp
;
1707 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1708 struct rcu_node
*rnp
;
1709 struct rcu_state
*rsp
;
1712 /* Handle nohz enablement switches conservatively. */
1713 tne
= ACCESS_ONCE(tick_nohz_active
);
1714 if (tne
!= rdtp
->tick_nohz_enabled_snap
) {
1715 if (rcu_cpu_has_callbacks(cpu
, NULL
))
1716 invoke_rcu_core(); /* force nohz to see update. */
1717 rdtp
->tick_nohz_enabled_snap
= tne
;
1723 /* If this is a no-CBs CPU, no callbacks, just return. */
1724 if (rcu_is_nocb_cpu(cpu
))
1728 * If a non-lazy callback arrived at a CPU having only lazy
1729 * callbacks, invoke RCU core for the side-effect of recalculating
1730 * idle duration on re-entry to idle.
1732 if (rdtp
->all_lazy
&&
1733 rdtp
->nonlazy_posted
!= rdtp
->nonlazy_posted_snap
) {
1734 rdtp
->all_lazy
= false;
1735 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1741 * If we have not yet accelerated this jiffy, accelerate all
1742 * callbacks on this CPU.
1744 if (rdtp
->last_accelerate
== jiffies
)
1746 rdtp
->last_accelerate
= jiffies
;
1747 for_each_rcu_flavor(rsp
) {
1748 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1749 if (!*rdp
->nxttail
[RCU_DONE_TAIL
])
1752 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1753 smp_mb__after_unlock_lock();
1754 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1755 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1757 rcu_gp_kthread_wake(rsp
);
1759 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1763 * Clean up for exit from idle. Attempt to advance callbacks based on
1764 * any grace periods that elapsed while the CPU was idle, and if any
1765 * callbacks are now ready to invoke, initiate invocation.
1767 static void rcu_cleanup_after_idle(int cpu
)
1769 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1770 if (rcu_is_nocb_cpu(cpu
))
1772 if (rcu_try_advance_all_cbs())
1774 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1778 * Keep a running count of the number of non-lazy callbacks posted
1779 * on this CPU. This running counter (which is never decremented) allows
1780 * rcu_prepare_for_idle() to detect when something out of the idle loop
1781 * posts a callback, even if an equal number of callbacks are invoked.
1782 * Of course, callbacks should only be posted from within a trace event
1783 * designed to be called from idle or from within RCU_NONIDLE().
1785 static void rcu_idle_count_callbacks_posted(void)
1787 __this_cpu_add(rcu_dynticks
.nonlazy_posted
, 1);
1791 * Data for flushing lazy RCU callbacks at OOM time.
1793 static atomic_t oom_callback_count
;
1794 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq
);
1797 * RCU OOM callback -- decrement the outstanding count and deliver the
1798 * wake-up if we are the last one.
1800 static void rcu_oom_callback(struct rcu_head
*rhp
)
1802 if (atomic_dec_and_test(&oom_callback_count
))
1803 wake_up(&oom_callback_wq
);
1807 * Post an rcu_oom_notify callback on the current CPU if it has at
1808 * least one lazy callback. This will unnecessarily post callbacks
1809 * to CPUs that already have a non-lazy callback at the end of their
1810 * callback list, but this is an infrequent operation, so accept some
1811 * extra overhead to keep things simple.
1813 static void rcu_oom_notify_cpu(void *unused
)
1815 struct rcu_state
*rsp
;
1816 struct rcu_data
*rdp
;
1818 for_each_rcu_flavor(rsp
) {
1819 rdp
= raw_cpu_ptr(rsp
->rda
);
1820 if (rdp
->qlen_lazy
!= 0) {
1821 atomic_inc(&oom_callback_count
);
1822 rsp
->call(&rdp
->oom_head
, rcu_oom_callback
);
1828 * If low on memory, ensure that each CPU has a non-lazy callback.
1829 * This will wake up CPUs that have only lazy callbacks, in turn
1830 * ensuring that they free up the corresponding memory in a timely manner.
1831 * Because an uncertain amount of memory will be freed in some uncertain
1832 * timeframe, we do not claim to have freed anything.
1834 static int rcu_oom_notify(struct notifier_block
*self
,
1835 unsigned long notused
, void *nfreed
)
1839 /* Wait for callbacks from earlier instance to complete. */
1840 wait_event(oom_callback_wq
, atomic_read(&oom_callback_count
) == 0);
1841 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1844 * Prevent premature wakeup: ensure that all increments happen
1845 * before there is a chance of the counter reaching zero.
1847 atomic_set(&oom_callback_count
, 1);
1850 for_each_online_cpu(cpu
) {
1851 smp_call_function_single(cpu
, rcu_oom_notify_cpu
, NULL
, 1);
1856 /* Unconditionally decrement: no need to wake ourselves up. */
1857 atomic_dec(&oom_callback_count
);
1862 static struct notifier_block rcu_oom_nb
= {
1863 .notifier_call
= rcu_oom_notify
1866 static int __init
rcu_register_oom_notifier(void)
1868 register_oom_notifier(&rcu_oom_nb
);
1871 early_initcall(rcu_register_oom_notifier
);
1873 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1875 #ifdef CONFIG_RCU_CPU_STALL_INFO
1877 #ifdef CONFIG_RCU_FAST_NO_HZ
1879 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1881 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1882 unsigned long nlpd
= rdtp
->nonlazy_posted
- rdtp
->nonlazy_posted_snap
;
1884 sprintf(cp
, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1885 rdtp
->last_accelerate
& 0xffff, jiffies
& 0xffff,
1887 rdtp
->all_lazy
? 'L' : '.',
1888 rdtp
->tick_nohz_enabled_snap
? '.' : 'D');
1891 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1893 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1898 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1900 /* Initiate the stall-info list. */
1901 static void print_cpu_stall_info_begin(void)
1907 * Print out diagnostic information for the specified stalled CPU.
1909 * If the specified CPU is aware of the current RCU grace period
1910 * (flavor specified by rsp), then print the number of scheduling
1911 * clock interrupts the CPU has taken during the time that it has
1912 * been aware. Otherwise, print the number of RCU grace periods
1913 * that this CPU is ignorant of, for example, "1" if the CPU was
1914 * aware of the previous grace period.
1916 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1918 static void print_cpu_stall_info(struct rcu_state
*rsp
, int cpu
)
1920 char fast_no_hz
[72];
1921 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1922 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
1924 unsigned long ticks_value
;
1926 if (rsp
->gpnum
== rdp
->gpnum
) {
1927 ticks_title
= "ticks this GP";
1928 ticks_value
= rdp
->ticks_this_gp
;
1930 ticks_title
= "GPs behind";
1931 ticks_value
= rsp
->gpnum
- rdp
->gpnum
;
1933 print_cpu_stall_fast_no_hz(fast_no_hz
, cpu
);
1934 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1935 cpu
, ticks_value
, ticks_title
,
1936 atomic_read(&rdtp
->dynticks
) & 0xfff,
1937 rdtp
->dynticks_nesting
, rdtp
->dynticks_nmi_nesting
,
1938 rdp
->softirq_snap
, kstat_softirqs_cpu(RCU_SOFTIRQ
, cpu
),
1942 /* Terminate the stall-info list. */
1943 static void print_cpu_stall_info_end(void)
1948 /* Zero ->ticks_this_gp for all flavors of RCU. */
1949 static void zero_cpu_stall_ticks(struct rcu_data
*rdp
)
1951 rdp
->ticks_this_gp
= 0;
1952 rdp
->softirq_snap
= kstat_softirqs_cpu(RCU_SOFTIRQ
, smp_processor_id());
1955 /* Increment ->ticks_this_gp for all flavors of RCU. */
1956 static void increment_cpu_stall_ticks(void)
1958 struct rcu_state
*rsp
;
1960 for_each_rcu_flavor(rsp
)
1961 raw_cpu_inc(rsp
->rda
->ticks_this_gp
);
1964 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1966 static void print_cpu_stall_info_begin(void)
1971 static void print_cpu_stall_info(struct rcu_state
*rsp
, int cpu
)
1973 pr_cont(" %d", cpu
);
1976 static void print_cpu_stall_info_end(void)
1981 static void zero_cpu_stall_ticks(struct rcu_data
*rdp
)
1985 static void increment_cpu_stall_ticks(void)
1989 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1991 #ifdef CONFIG_RCU_NOCB_CPU
1994 * Offload callback processing from the boot-time-specified set of CPUs
1995 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1996 * kthread created that pulls the callbacks from the corresponding CPU,
1997 * waits for a grace period to elapse, and invokes the callbacks.
1998 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1999 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2000 * has been specified, in which case each kthread actively polls its
2001 * CPU. (Which isn't so great for energy efficiency, but which does
2002 * reduce RCU's overhead on that CPU.)
2004 * This is intended to be used in conjunction with Frederic Weisbecker's
2005 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2006 * running CPU-bound user-mode computations.
2008 * Offloading of callback processing could also in theory be used as
2009 * an energy-efficiency measure because CPUs with no RCU callbacks
2010 * queued are more aggressive about entering dyntick-idle mode.
2014 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2015 static int __init
rcu_nocb_setup(char *str
)
2017 alloc_bootmem_cpumask_var(&rcu_nocb_mask
);
2018 have_rcu_nocb_mask
= true;
2019 cpulist_parse(str
, rcu_nocb_mask
);
2022 __setup("rcu_nocbs=", rcu_nocb_setup
);
2024 static int __init
parse_rcu_nocb_poll(char *arg
)
2029 early_param("rcu_nocb_poll", parse_rcu_nocb_poll
);
2032 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2035 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
2037 wake_up_all(&rnp
->nocb_gp_wq
[rnp
->completed
& 0x1]);
2041 * Set the root rcu_node structure's ->need_future_gp field
2042 * based on the sum of those of all rcu_node structures. This does
2043 * double-count the root rcu_node structure's requests, but this
2044 * is necessary to handle the possibility of a rcu_nocb_kthread()
2045 * having awakened during the time that the rcu_node structures
2046 * were being updated for the end of the previous grace period.
2048 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
2050 rnp
->need_future_gp
[(rnp
->completed
+ 1) & 0x1] += nrq
;
2053 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2055 init_waitqueue_head(&rnp
->nocb_gp_wq
[0]);
2056 init_waitqueue_head(&rnp
->nocb_gp_wq
[1]);
2059 #ifndef CONFIG_RCU_NOCB_CPU_ALL
2060 /* Is the specified CPU a no-CBs CPU? */
2061 bool rcu_is_nocb_cpu(int cpu
)
2063 if (have_rcu_nocb_mask
)
2064 return cpumask_test_cpu(cpu
, rcu_nocb_mask
);
2067 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
2070 * Kick the leader kthread for this NOCB group.
2072 static void wake_nocb_leader(struct rcu_data
*rdp
, bool force
)
2074 struct rcu_data
*rdp_leader
= rdp
->nocb_leader
;
2076 if (!ACCESS_ONCE(rdp_leader
->nocb_kthread
))
2078 if (ACCESS_ONCE(rdp_leader
->nocb_leader_sleep
) || force
) {
2079 /* Prior xchg orders against prior callback enqueue. */
2080 ACCESS_ONCE(rdp_leader
->nocb_leader_sleep
) = false;
2081 wake_up(&rdp_leader
->nocb_wq
);
2086 * Enqueue the specified string of rcu_head structures onto the specified
2087 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2088 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2089 * counts are supplied by rhcount and rhcount_lazy.
2091 * If warranted, also wake up the kthread servicing this CPUs queues.
2093 static void __call_rcu_nocb_enqueue(struct rcu_data
*rdp
,
2094 struct rcu_head
*rhp
,
2095 struct rcu_head
**rhtp
,
2096 int rhcount
, int rhcount_lazy
,
2097 unsigned long flags
)
2100 struct rcu_head
**old_rhpp
;
2101 struct task_struct
*t
;
2103 /* Enqueue the callback on the nocb list and update counts. */
2104 old_rhpp
= xchg(&rdp
->nocb_tail
, rhtp
);
2105 ACCESS_ONCE(*old_rhpp
) = rhp
;
2106 atomic_long_add(rhcount
, &rdp
->nocb_q_count
);
2107 atomic_long_add(rhcount_lazy
, &rdp
->nocb_q_count_lazy
);
2109 /* If we are not being polled and there is a kthread, awaken it ... */
2110 t
= ACCESS_ONCE(rdp
->nocb_kthread
);
2111 if (rcu_nocb_poll
|| !t
) {
2112 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2113 TPS("WakeNotPoll"));
2116 len
= atomic_long_read(&rdp
->nocb_q_count
);
2117 if (old_rhpp
== &rdp
->nocb_head
) {
2118 if (!irqs_disabled_flags(flags
)) {
2119 /* ... if queue was empty ... */
2120 wake_nocb_leader(rdp
, false);
2121 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2124 rdp
->nocb_defer_wakeup
= true;
2125 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2126 TPS("WakeEmptyIsDeferred"));
2128 rdp
->qlen_last_fqs_check
= 0;
2129 } else if (len
> rdp
->qlen_last_fqs_check
+ qhimark
) {
2130 /* ... or if many callbacks queued. */
2131 wake_nocb_leader(rdp
, true);
2132 rdp
->qlen_last_fqs_check
= LONG_MAX
/ 2;
2133 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("WakeOvf"));
2135 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("WakeNot"));
2141 * This is a helper for __call_rcu(), which invokes this when the normal
2142 * callback queue is inoperable. If this is not a no-CBs CPU, this
2143 * function returns failure back to __call_rcu(), which can complain
2146 * Otherwise, this function queues the callback where the corresponding
2147 * "rcuo" kthread can find it.
2149 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2150 bool lazy
, unsigned long flags
)
2153 if (!rcu_is_nocb_cpu(rdp
->cpu
))
2155 __call_rcu_nocb_enqueue(rdp
, rhp
, &rhp
->next
, 1, lazy
, flags
);
2156 if (__is_kfree_rcu_offset((unsigned long)rhp
->func
))
2157 trace_rcu_kfree_callback(rdp
->rsp
->name
, rhp
,
2158 (unsigned long)rhp
->func
,
2159 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
2160 -atomic_long_read(&rdp
->nocb_q_count
));
2162 trace_rcu_callback(rdp
->rsp
->name
, rhp
,
2163 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
2164 -atomic_long_read(&rdp
->nocb_q_count
));
2169 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2172 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
2173 struct rcu_data
*rdp
,
2174 unsigned long flags
)
2176 long ql
= rsp
->qlen
;
2177 long qll
= rsp
->qlen_lazy
;
2179 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2180 if (!rcu_is_nocb_cpu(smp_processor_id()))
2185 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2186 if (rsp
->orphan_donelist
!= NULL
) {
2187 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_donelist
,
2188 rsp
->orphan_donetail
, ql
, qll
, flags
);
2190 rsp
->orphan_donelist
= NULL
;
2191 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2193 if (rsp
->orphan_nxtlist
!= NULL
) {
2194 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_nxtlist
,
2195 rsp
->orphan_nxttail
, ql
, qll
, flags
);
2197 rsp
->orphan_nxtlist
= NULL
;
2198 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2204 * If necessary, kick off a new grace period, and either way wait
2205 * for a subsequent grace period to complete.
2207 static void rcu_nocb_wait_gp(struct rcu_data
*rdp
)
2211 unsigned long flags
;
2213 struct rcu_node
*rnp
= rdp
->mynode
;
2215 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2216 smp_mb__after_unlock_lock();
2217 needwake
= rcu_start_future_gp(rnp
, rdp
, &c
);
2218 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2220 rcu_gp_kthread_wake(rdp
->rsp
);
2223 * Wait for the grace period. Do so interruptibly to avoid messing
2224 * up the load average.
2226 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("StartWait"));
2228 wait_event_interruptible(
2229 rnp
->nocb_gp_wq
[c
& 0x1],
2230 (d
= ULONG_CMP_GE(ACCESS_ONCE(rnp
->completed
), c
)));
2233 flush_signals(current
);
2234 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("ResumeWait"));
2236 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("EndWait"));
2237 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2241 * Leaders come here to wait for additional callbacks to show up.
2242 * This function does not return until callbacks appear.
2244 static void nocb_leader_wait(struct rcu_data
*my_rdp
)
2246 bool firsttime
= true;
2248 struct rcu_data
*rdp
;
2249 struct rcu_head
**tail
;
2253 /* Wait for callbacks to appear. */
2254 if (!rcu_nocb_poll
) {
2255 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
, "Sleep");
2256 wait_event_interruptible(my_rdp
->nocb_wq
,
2257 !ACCESS_ONCE(my_rdp
->nocb_leader_sleep
));
2258 /* Memory barrier handled by smp_mb() calls below and repoll. */
2259 } else if (firsttime
) {
2260 firsttime
= false; /* Don't drown trace log with "Poll"! */
2261 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
, "Poll");
2265 * Each pass through the following loop checks a follower for CBs.
2266 * We are our own first follower. Any CBs found are moved to
2267 * nocb_gp_head, where they await a grace period.
2270 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
) {
2271 rdp
->nocb_gp_head
= ACCESS_ONCE(rdp
->nocb_head
);
2272 if (!rdp
->nocb_gp_head
)
2273 continue; /* No CBs here, try next follower. */
2275 /* Move callbacks to wait-for-GP list, which is empty. */
2276 ACCESS_ONCE(rdp
->nocb_head
) = NULL
;
2277 rdp
->nocb_gp_tail
= xchg(&rdp
->nocb_tail
, &rdp
->nocb_head
);
2278 rdp
->nocb_gp_count
= atomic_long_xchg(&rdp
->nocb_q_count
, 0);
2279 rdp
->nocb_gp_count_lazy
=
2280 atomic_long_xchg(&rdp
->nocb_q_count_lazy
, 0);
2285 * If there were no callbacks, sleep a bit, rescan after a
2286 * memory barrier, and go retry.
2288 if (unlikely(!gotcbs
)) {
2290 trace_rcu_nocb_wake(my_rdp
->rsp
->name
, my_rdp
->cpu
,
2292 flush_signals(current
);
2293 schedule_timeout_interruptible(1);
2295 /* Rescan in case we were a victim of memory ordering. */
2296 my_rdp
->nocb_leader_sleep
= true;
2297 smp_mb(); /* Ensure _sleep true before scan. */
2298 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
)
2299 if (ACCESS_ONCE(rdp
->nocb_head
)) {
2300 /* Found CB, so short-circuit next wait. */
2301 my_rdp
->nocb_leader_sleep
= false;
2307 /* Wait for one grace period. */
2308 rcu_nocb_wait_gp(my_rdp
);
2311 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2312 * We set it now, but recheck for new callbacks while
2313 * traversing our follower list.
2315 my_rdp
->nocb_leader_sleep
= true;
2316 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2318 /* Each pass through the following loop wakes a follower, if needed. */
2319 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_follower
) {
2320 if (ACCESS_ONCE(rdp
->nocb_head
))
2321 my_rdp
->nocb_leader_sleep
= false;/* No need to sleep.*/
2322 if (!rdp
->nocb_gp_head
)
2323 continue; /* No CBs, so no need to wake follower. */
2325 /* Append callbacks to follower's "done" list. */
2326 tail
= xchg(&rdp
->nocb_follower_tail
, rdp
->nocb_gp_tail
);
2327 *tail
= rdp
->nocb_gp_head
;
2328 atomic_long_add(rdp
->nocb_gp_count
, &rdp
->nocb_follower_count
);
2329 atomic_long_add(rdp
->nocb_gp_count_lazy
,
2330 &rdp
->nocb_follower_count_lazy
);
2331 if (rdp
!= my_rdp
&& tail
== &rdp
->nocb_follower_head
) {
2333 * List was empty, wake up the follower.
2334 * Memory barriers supplied by atomic_long_add().
2336 wake_up(&rdp
->nocb_wq
);
2340 /* If we (the leader) don't have CBs, go wait some more. */
2341 if (!my_rdp
->nocb_follower_head
)
2346 * Followers come here to wait for additional callbacks to show up.
2347 * This function does not return until callbacks appear.
2349 static void nocb_follower_wait(struct rcu_data
*rdp
)
2351 bool firsttime
= true;
2354 if (!rcu_nocb_poll
) {
2355 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2357 wait_event_interruptible(rdp
->nocb_wq
,
2358 ACCESS_ONCE(rdp
->nocb_follower_head
));
2359 } else if (firsttime
) {
2360 /* Don't drown trace log with "Poll"! */
2362 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, "Poll");
2364 if (smp_load_acquire(&rdp
->nocb_follower_head
)) {
2365 /* ^^^ Ensure CB invocation follows _head test. */
2369 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2371 flush_signals(current
);
2372 schedule_timeout_interruptible(1);
2377 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2378 * callbacks queued by the corresponding no-CBs CPU, however, there is
2379 * an optional leader-follower relationship so that the grace-period
2380 * kthreads don't have to do quite so many wakeups.
2382 static int rcu_nocb_kthread(void *arg
)
2385 struct rcu_head
*list
;
2386 struct rcu_head
*next
;
2387 struct rcu_head
**tail
;
2388 struct rcu_data
*rdp
= arg
;
2390 /* Each pass through this loop invokes one batch of callbacks */
2392 /* Wait for callbacks. */
2393 if (rdp
->nocb_leader
== rdp
)
2394 nocb_leader_wait(rdp
);
2396 nocb_follower_wait(rdp
);
2398 /* Pull the ready-to-invoke callbacks onto local list. */
2399 list
= ACCESS_ONCE(rdp
->nocb_follower_head
);
2401 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, "WokeNonEmpty");
2402 ACCESS_ONCE(rdp
->nocb_follower_head
) = NULL
;
2403 tail
= xchg(&rdp
->nocb_follower_tail
, &rdp
->nocb_follower_head
);
2404 c
= atomic_long_xchg(&rdp
->nocb_follower_count
, 0);
2405 cl
= atomic_long_xchg(&rdp
->nocb_follower_count_lazy
, 0);
2406 rdp
->nocb_p_count
+= c
;
2407 rdp
->nocb_p_count_lazy
+= cl
;
2409 /* Each pass through the following loop invokes a callback. */
2410 trace_rcu_batch_start(rdp
->rsp
->name
, cl
, c
, -1);
2414 /* Wait for enqueuing to complete, if needed. */
2415 while (next
== NULL
&& &list
->next
!= tail
) {
2416 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2418 schedule_timeout_interruptible(1);
2419 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2423 debug_rcu_head_unqueue(list
);
2425 if (__rcu_reclaim(rdp
->rsp
->name
, list
))
2431 trace_rcu_batch_end(rdp
->rsp
->name
, c
, !!list
, 0, 0, 1);
2432 ACCESS_ONCE(rdp
->nocb_p_count
) = rdp
->nocb_p_count
- c
;
2433 ACCESS_ONCE(rdp
->nocb_p_count_lazy
) =
2434 rdp
->nocb_p_count_lazy
- cl
;
2435 rdp
->n_nocbs_invoked
+= c
;
2440 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2441 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2443 return ACCESS_ONCE(rdp
->nocb_defer_wakeup
);
2446 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2447 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2449 if (!rcu_nocb_need_deferred_wakeup(rdp
))
2451 ACCESS_ONCE(rdp
->nocb_defer_wakeup
) = false;
2452 wake_nocb_leader(rdp
, false);
2453 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("DeferredWakeEmpty"));
2456 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2457 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2459 rdp
->nocb_tail
= &rdp
->nocb_head
;
2460 init_waitqueue_head(&rdp
->nocb_wq
);
2461 rdp
->nocb_follower_tail
= &rdp
->nocb_follower_head
;
2464 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2465 static int rcu_nocb_leader_stride
= -1;
2466 module_param(rcu_nocb_leader_stride
, int, 0444);
2469 * Create a kthread for each RCU flavor for each no-CBs CPU.
2470 * Also initialize leader-follower relationships.
2472 static void __init
rcu_spawn_nocb_kthreads(struct rcu_state
*rsp
)
2475 int ls
= rcu_nocb_leader_stride
;
2476 int nl
= 0; /* Next leader. */
2477 struct rcu_data
*rdp
;
2478 struct rcu_data
*rdp_leader
= NULL
; /* Suppress misguided gcc warn. */
2479 struct rcu_data
*rdp_prev
= NULL
;
2480 struct task_struct
*t
;
2482 if (rcu_nocb_mask
== NULL
)
2484 #if defined(CONFIG_NO_HZ_FULL) && !defined(CONFIG_NO_HZ_FULL_ALL)
2485 if (tick_nohz_full_running
)
2486 cpumask_or(rcu_nocb_mask
, rcu_nocb_mask
, tick_nohz_full_mask
);
2487 #endif /* #if defined(CONFIG_NO_HZ_FULL) && !defined(CONFIG_NO_HZ_FULL_ALL) */
2489 ls
= int_sqrt(nr_cpu_ids
);
2490 rcu_nocb_leader_stride
= ls
;
2494 * Each pass through this loop sets up one rcu_data structure and
2495 * spawns one rcu_nocb_kthread().
2497 for_each_cpu(cpu
, rcu_nocb_mask
) {
2498 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2499 if (rdp
->cpu
>= nl
) {
2500 /* New leader, set up for followers & next leader. */
2501 nl
= DIV_ROUND_UP(rdp
->cpu
+ 1, ls
) * ls
;
2502 rdp
->nocb_leader
= rdp
;
2505 /* Another follower, link to previous leader. */
2506 rdp
->nocb_leader
= rdp_leader
;
2507 rdp_prev
->nocb_next_follower
= rdp
;
2511 /* Spawn the kthread for this CPU. */
2512 t
= kthread_run(rcu_nocb_kthread
, rdp
,
2513 "rcuo%c/%d", rsp
->abbr
, cpu
);
2515 ACCESS_ONCE(rdp
->nocb_kthread
) = t
;
2519 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2520 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2522 if (rcu_nocb_mask
== NULL
||
2523 !cpumask_test_cpu(rdp
->cpu
, rcu_nocb_mask
))
2525 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2529 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2531 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
2535 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
2539 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2543 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2544 bool lazy
, unsigned long flags
)
2549 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
2550 struct rcu_data
*rdp
,
2551 unsigned long flags
)
2556 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2560 static bool rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2565 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2569 static void __init
rcu_spawn_nocb_kthreads(struct rcu_state
*rsp
)
2573 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2578 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2581 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2582 * arbitrarily long period of time with the scheduling-clock tick turned
2583 * off. RCU will be paying attention to this CPU because it is in the
2584 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2585 * machine because the scheduling-clock tick has been disabled. Therefore,
2586 * if an adaptive-ticks CPU is failing to respond to the current grace
2587 * period and has not be idle from an RCU perspective, kick it.
2589 static void __maybe_unused
rcu_kick_nohz_cpu(int cpu
)
2591 #ifdef CONFIG_NO_HZ_FULL
2592 if (tick_nohz_full_cpu(cpu
))
2593 smp_send_reschedule(cpu
);
2594 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2598 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2601 * Define RCU flavor that holds sysidle state. This needs to be the
2602 * most active flavor of RCU.
2604 #ifdef CONFIG_PREEMPT_RCU
2605 static struct rcu_state
*rcu_sysidle_state
= &rcu_preempt_state
;
2606 #else /* #ifdef CONFIG_PREEMPT_RCU */
2607 static struct rcu_state
*rcu_sysidle_state
= &rcu_sched_state
;
2608 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2610 static int full_sysidle_state
; /* Current system-idle state. */
2611 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2612 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2613 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2614 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2615 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2618 * Invoked to note exit from irq or task transition to idle. Note that
2619 * usermode execution does -not- count as idle here! After all, we want
2620 * to detect full-system idle states, not RCU quiescent states and grace
2621 * periods. The caller must have disabled interrupts.
2623 static void rcu_sysidle_enter(struct rcu_dynticks
*rdtp
, int irq
)
2627 /* Adjust nesting, check for fully idle. */
2629 rdtp
->dynticks_idle_nesting
--;
2630 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2631 if (rdtp
->dynticks_idle_nesting
!= 0)
2632 return; /* Still not fully idle. */
2634 if ((rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) ==
2635 DYNTICK_TASK_NEST_VALUE
) {
2636 rdtp
->dynticks_idle_nesting
= 0;
2638 rdtp
->dynticks_idle_nesting
-= DYNTICK_TASK_NEST_VALUE
;
2639 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2640 return; /* Still not fully idle. */
2644 /* Record start of fully idle period. */
2646 ACCESS_ONCE(rdtp
->dynticks_idle_jiffies
) = j
;
2647 smp_mb__before_atomic();
2648 atomic_inc(&rdtp
->dynticks_idle
);
2649 smp_mb__after_atomic();
2650 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks_idle
) & 0x1);
2654 * Unconditionally force exit from full system-idle state. This is
2655 * invoked when a normal CPU exits idle, but must be called separately
2656 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2657 * is that the timekeeping CPU is permitted to take scheduling-clock
2658 * interrupts while the system is in system-idle state, and of course
2659 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2660 * interrupt from any other type of interrupt.
2662 void rcu_sysidle_force_exit(void)
2664 int oldstate
= ACCESS_ONCE(full_sysidle_state
);
2668 * Each pass through the following loop attempts to exit full
2669 * system-idle state. If contention proves to be a problem,
2670 * a trylock-based contention tree could be used here.
2672 while (oldstate
> RCU_SYSIDLE_SHORT
) {
2673 newoldstate
= cmpxchg(&full_sysidle_state
,
2674 oldstate
, RCU_SYSIDLE_NOT
);
2675 if (oldstate
== newoldstate
&&
2676 oldstate
== RCU_SYSIDLE_FULL_NOTED
) {
2677 rcu_kick_nohz_cpu(tick_do_timer_cpu
);
2678 return; /* We cleared it, done! */
2680 oldstate
= newoldstate
;
2682 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2686 * Invoked to note entry to irq or task transition from idle. Note that
2687 * usermode execution does -not- count as idle here! The caller must
2688 * have disabled interrupts.
2690 static void rcu_sysidle_exit(struct rcu_dynticks
*rdtp
, int irq
)
2692 /* Adjust nesting, check for already non-idle. */
2694 rdtp
->dynticks_idle_nesting
++;
2695 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2696 if (rdtp
->dynticks_idle_nesting
!= 1)
2697 return; /* Already non-idle. */
2700 * Allow for irq misnesting. Yes, it really is possible
2701 * to enter an irq handler then never leave it, and maybe
2702 * also vice versa. Handle both possibilities.
2704 if (rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) {
2705 rdtp
->dynticks_idle_nesting
+= DYNTICK_TASK_NEST_VALUE
;
2706 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2707 return; /* Already non-idle. */
2709 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2713 /* Record end of idle period. */
2714 smp_mb__before_atomic();
2715 atomic_inc(&rdtp
->dynticks_idle
);
2716 smp_mb__after_atomic();
2717 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks_idle
) & 0x1));
2720 * If we are the timekeeping CPU, we are permitted to be non-idle
2721 * during a system-idle state. This must be the case, because
2722 * the timekeeping CPU has to take scheduling-clock interrupts
2723 * during the time that the system is transitioning to full
2724 * system-idle state. This means that the timekeeping CPU must
2725 * invoke rcu_sysidle_force_exit() directly if it does anything
2726 * more than take a scheduling-clock interrupt.
2728 if (smp_processor_id() == tick_do_timer_cpu
)
2731 /* Update system-idle state: We are clearly no longer fully idle! */
2732 rcu_sysidle_force_exit();
2736 * Check to see if the current CPU is idle. Note that usermode execution
2737 * does not count as idle. The caller must have disabled interrupts.
2739 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
2740 unsigned long *maxj
)
2744 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
2747 * If some other CPU has already reported non-idle, if this is
2748 * not the flavor of RCU that tracks sysidle state, or if this
2749 * is an offline or the timekeeping CPU, nothing to do.
2751 if (!*isidle
|| rdp
->rsp
!= rcu_sysidle_state
||
2752 cpu_is_offline(rdp
->cpu
) || rdp
->cpu
== tick_do_timer_cpu
)
2754 if (rcu_gp_in_progress(rdp
->rsp
))
2755 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
);
2757 /* Pick up current idle and NMI-nesting counter and check. */
2758 cur
= atomic_read(&rdtp
->dynticks_idle
);
2760 *isidle
= false; /* We are not idle! */
2763 smp_mb(); /* Read counters before timestamps. */
2765 /* Pick up timestamps. */
2766 j
= ACCESS_ONCE(rdtp
->dynticks_idle_jiffies
);
2767 /* If this CPU entered idle more recently, update maxj timestamp. */
2768 if (ULONG_CMP_LT(*maxj
, j
))
2773 * Is this the flavor of RCU that is handling full-system idle?
2775 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
2777 return rsp
== rcu_sysidle_state
;
2781 * Return a delay in jiffies based on the number of CPUs, rcu_node
2782 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2783 * systems more time to transition to full-idle state in order to
2784 * avoid the cache thrashing that otherwise occur on the state variable.
2785 * Really small systems (less than a couple of tens of CPUs) should
2786 * instead use a single global atomically incremented counter, and later
2787 * versions of this will automatically reconfigure themselves accordingly.
2789 static unsigned long rcu_sysidle_delay(void)
2791 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2793 return DIV_ROUND_UP(nr_cpu_ids
* HZ
, rcu_fanout_leaf
* 1000);
2797 * Advance the full-system-idle state. This is invoked when all of
2798 * the non-timekeeping CPUs are idle.
2800 static void rcu_sysidle(unsigned long j
)
2802 /* Check the current state. */
2803 switch (ACCESS_ONCE(full_sysidle_state
)) {
2804 case RCU_SYSIDLE_NOT
:
2806 /* First time all are idle, so note a short idle period. */
2807 ACCESS_ONCE(full_sysidle_state
) = RCU_SYSIDLE_SHORT
;
2810 case RCU_SYSIDLE_SHORT
:
2813 * Idle for a bit, time to advance to next state?
2814 * cmpxchg failure means race with non-idle, let them win.
2816 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2817 (void)cmpxchg(&full_sysidle_state
,
2818 RCU_SYSIDLE_SHORT
, RCU_SYSIDLE_LONG
);
2821 case RCU_SYSIDLE_LONG
:
2824 * Do an additional check pass before advancing to full.
2825 * cmpxchg failure means race with non-idle, let them win.
2827 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2828 (void)cmpxchg(&full_sysidle_state
,
2829 RCU_SYSIDLE_LONG
, RCU_SYSIDLE_FULL
);
2838 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2839 * back to the beginning.
2841 static void rcu_sysidle_cancel(void)
2844 if (full_sysidle_state
> RCU_SYSIDLE_SHORT
)
2845 ACCESS_ONCE(full_sysidle_state
) = RCU_SYSIDLE_NOT
;
2849 * Update the sysidle state based on the results of a force-quiescent-state
2850 * scan of the CPUs' dyntick-idle state.
2852 static void rcu_sysidle_report(struct rcu_state
*rsp
, int isidle
,
2853 unsigned long maxj
, bool gpkt
)
2855 if (rsp
!= rcu_sysidle_state
)
2856 return; /* Wrong flavor, ignore. */
2857 if (gpkt
&& nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2858 return; /* Running state machine from timekeeping CPU. */
2860 rcu_sysidle(maxj
); /* More idle! */
2862 rcu_sysidle_cancel(); /* Idle is over. */
2866 * Wrapper for rcu_sysidle_report() when called from the grace-period
2867 * kthread's context.
2869 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
2872 rcu_sysidle_report(rsp
, isidle
, maxj
, true);
2875 /* Callback and function for forcing an RCU grace period. */
2876 struct rcu_sysidle_head
{
2881 static void rcu_sysidle_cb(struct rcu_head
*rhp
)
2883 struct rcu_sysidle_head
*rshp
;
2886 * The following memory barrier is needed to replace the
2887 * memory barriers that would normally be in the memory
2890 smp_mb(); /* grace period precedes setting inuse. */
2892 rshp
= container_of(rhp
, struct rcu_sysidle_head
, rh
);
2893 ACCESS_ONCE(rshp
->inuse
) = 0;
2897 * Check to see if the system is fully idle, other than the timekeeping CPU.
2898 * The caller must have disabled interrupts.
2900 bool rcu_sys_is_idle(void)
2902 static struct rcu_sysidle_head rsh
;
2903 int rss
= ACCESS_ONCE(full_sysidle_state
);
2905 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
))
2908 /* Handle small-system case by doing a full scan of CPUs. */
2909 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
) {
2910 int oldrss
= rss
- 1;
2913 * One pass to advance to each state up to _FULL.
2914 * Give up if any pass fails to advance the state.
2916 while (rss
< RCU_SYSIDLE_FULL
&& oldrss
< rss
) {
2919 unsigned long maxj
= jiffies
- ULONG_MAX
/ 4;
2920 struct rcu_data
*rdp
;
2922 /* Scan all the CPUs looking for nonidle CPUs. */
2923 for_each_possible_cpu(cpu
) {
2924 rdp
= per_cpu_ptr(rcu_sysidle_state
->rda
, cpu
);
2925 rcu_sysidle_check_cpu(rdp
, &isidle
, &maxj
);
2929 rcu_sysidle_report(rcu_sysidle_state
,
2930 isidle
, maxj
, false);
2932 rss
= ACCESS_ONCE(full_sysidle_state
);
2936 /* If this is the first observation of an idle period, record it. */
2937 if (rss
== RCU_SYSIDLE_FULL
) {
2938 rss
= cmpxchg(&full_sysidle_state
,
2939 RCU_SYSIDLE_FULL
, RCU_SYSIDLE_FULL_NOTED
);
2940 return rss
== RCU_SYSIDLE_FULL
;
2943 smp_mb(); /* ensure rss load happens before later caller actions. */
2945 /* If already fully idle, tell the caller (in case of races). */
2946 if (rss
== RCU_SYSIDLE_FULL_NOTED
)
2950 * If we aren't there yet, and a grace period is not in flight,
2951 * initiate a grace period. Either way, tell the caller that
2952 * we are not there yet. We use an xchg() rather than an assignment
2953 * to make up for the memory barriers that would otherwise be
2954 * provided by the memory allocator.
2956 if (nr_cpu_ids
> CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
&&
2957 !rcu_gp_in_progress(rcu_sysidle_state
) &&
2958 !rsh
.inuse
&& xchg(&rsh
.inuse
, 1) == 0)
2959 call_rcu(&rsh
.rh
, rcu_sysidle_cb
);
2964 * Initialize dynticks sysidle state for CPUs coming online.
2966 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
2968 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
;
2971 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2973 static void rcu_sysidle_enter(struct rcu_dynticks
*rdtp
, int irq
)
2977 static void rcu_sysidle_exit(struct rcu_dynticks
*rdtp
, int irq
)
2981 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
2982 unsigned long *maxj
)
2986 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
2991 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
2996 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
3000 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3003 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3004 * grace-period kthread will do force_quiescent_state() processing?
3005 * The idea is to avoid waking up RCU core processing on such a
3006 * CPU unless the grace period has extended for too long.
3008 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3009 * CONFIG_RCU_NOCB_CPU CPUs.
3011 static bool rcu_nohz_full_cpu(struct rcu_state
*rsp
)
3013 #ifdef CONFIG_NO_HZ_FULL
3014 if (tick_nohz_full_cpu(smp_processor_id()) &&
3015 (!rcu_gp_in_progress(rsp
) ||
3016 ULONG_CMP_LT(jiffies
, ACCESS_ONCE(rsp
->gp_start
) + HZ
)))
3018 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3023 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3026 static void rcu_bind_gp_kthread(void)
3028 int __maybe_unused cpu
;
3030 if (!tick_nohz_full_enabled())
3032 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3033 cpu
= tick_do_timer_cpu
;
3034 if (cpu
>= 0 && cpu
< nr_cpu_ids
&& raw_smp_processor_id() != cpu
)
3035 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
3036 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3037 if (!is_housekeeping_cpu(raw_smp_processor_id()))
3038 housekeeping_affine(current
);
3039 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */