2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptable semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/stop_machine.h>
31 * Check the RCU kernel configuration parameters and print informative
32 * messages about anything out of the ordinary. If you like #ifdef, you
33 * will love this function.
35 static void __init
rcu_bootup_announce_oddness(void)
37 #ifdef CONFIG_RCU_TRACE
38 printk(KERN_INFO
"\tRCU debugfs-based tracing is enabled.\n");
40 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
41 printk(KERN_INFO
"\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
44 #ifdef CONFIG_RCU_FANOUT_EXACT
45 printk(KERN_INFO
"\tHierarchical RCU autobalancing is disabled.\n");
47 #ifdef CONFIG_RCU_FAST_NO_HZ
49 "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
51 #ifdef CONFIG_PROVE_RCU
52 printk(KERN_INFO
"\tRCU lockdep checking is enabled.\n");
54 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
55 printk(KERN_INFO
"\tRCU torture testing starts during boot.\n");
57 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
58 printk(KERN_INFO
"\tVerbose stalled-CPUs detection is disabled.\n");
60 #if NUM_RCU_LVL_4 != 0
61 printk(KERN_INFO
"\tExperimental four-level hierarchy is enabled.\n");
65 #ifdef CONFIG_TREE_PREEMPT_RCU
67 struct rcu_state rcu_preempt_state
= RCU_STATE_INITIALIZER(rcu_preempt_state
);
68 DEFINE_PER_CPU(struct rcu_data
, rcu_preempt_data
);
69 static struct rcu_state
*rcu_state
= &rcu_preempt_state
;
71 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
);
74 * Tell them what RCU they are running.
76 static void __init
rcu_bootup_announce(void)
78 printk(KERN_INFO
"Preemptable hierarchical RCU implementation.\n");
79 rcu_bootup_announce_oddness();
83 * Return the number of RCU-preempt batches processed thus far
84 * for debug and statistics.
86 long rcu_batches_completed_preempt(void)
88 return rcu_preempt_state
.completed
;
90 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt
);
93 * Return the number of RCU batches processed thus far for debug & stats.
95 long rcu_batches_completed(void)
97 return rcu_batches_completed_preempt();
99 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
102 * Force a quiescent state for preemptible RCU.
104 void rcu_force_quiescent_state(void)
106 force_quiescent_state(&rcu_preempt_state
, 0);
108 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
111 * Record a preemptable-RCU quiescent state for the specified CPU. Note
112 * that this just means that the task currently running on the CPU is
113 * not in a quiescent state. There might be any number of tasks blocked
114 * while in an RCU read-side critical section.
116 * Unlike the other rcu_*_qs() functions, callers to this function
117 * must disable irqs in order to protect the assignment to
118 * ->rcu_read_unlock_special.
120 static void rcu_preempt_qs(int cpu
)
122 struct rcu_data
*rdp
= &per_cpu(rcu_preempt_data
, cpu
);
124 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
126 rdp
->passed_quiesc
= 1;
127 current
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_NEED_QS
;
131 * We have entered the scheduler, and the current task might soon be
132 * context-switched away from. If this task is in an RCU read-side
133 * critical section, we will no longer be able to rely on the CPU to
134 * record that fact, so we enqueue the task on the blkd_tasks list.
135 * The task will dequeue itself when it exits the outermost enclosing
136 * RCU read-side critical section. Therefore, the current grace period
137 * cannot be permitted to complete until the blkd_tasks list entries
138 * predating the current grace period drain, in other words, until
139 * rnp->gp_tasks becomes NULL.
141 * Caller must disable preemption.
143 static void rcu_preempt_note_context_switch(int cpu
)
145 struct task_struct
*t
= current
;
147 struct rcu_data
*rdp
;
148 struct rcu_node
*rnp
;
150 if (t
->rcu_read_lock_nesting
&&
151 (t
->rcu_read_unlock_special
& RCU_READ_UNLOCK_BLOCKED
) == 0) {
153 /* Possibly blocking in an RCU read-side critical section. */
154 rdp
= per_cpu_ptr(rcu_preempt_state
.rda
, cpu
);
156 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
157 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_BLOCKED
;
158 t
->rcu_blocked_node
= rnp
;
161 * If this CPU has already checked in, then this task
162 * will hold up the next grace period rather than the
163 * current grace period. Queue the task accordingly.
164 * If the task is queued for the current grace period
165 * (i.e., this CPU has not yet passed through a quiescent
166 * state for the current grace period), then as long
167 * as that task remains queued, the current grace period
168 * cannot end. Note that there is some uncertainty as
169 * to exactly when the current grace period started.
170 * We take a conservative approach, which can result
171 * in unnecessarily waiting on tasks that started very
172 * slightly after the current grace period began. C'est
175 * But first, note that the current CPU must still be
178 WARN_ON_ONCE((rdp
->grpmask
& rnp
->qsmaskinit
) == 0);
179 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
180 if ((rnp
->qsmask
& rdp
->grpmask
) && rnp
->gp_tasks
!= NULL
) {
181 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
->prev
);
182 rnp
->gp_tasks
= &t
->rcu_node_entry
;
183 #ifdef CONFIG_RCU_BOOST
184 if (rnp
->boost_tasks
!= NULL
)
185 rnp
->boost_tasks
= rnp
->gp_tasks
;
186 #endif /* #ifdef CONFIG_RCU_BOOST */
188 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
189 if (rnp
->qsmask
& rdp
->grpmask
)
190 rnp
->gp_tasks
= &t
->rcu_node_entry
;
192 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
196 * Either we were not in an RCU read-side critical section to
197 * begin with, or we have now recorded that critical section
198 * globally. Either way, we can now note a quiescent state
199 * for this CPU. Again, if we were in an RCU read-side critical
200 * section, and if that critical section was blocking the current
201 * grace period, then the fact that the task has been enqueued
202 * means that we continue to block the current grace period.
204 local_irq_save(flags
);
206 local_irq_restore(flags
);
210 * Tree-preemptable RCU implementation for rcu_read_lock().
211 * Just increment ->rcu_read_lock_nesting, shared state will be updated
214 void __rcu_read_lock(void)
216 current
->rcu_read_lock_nesting
++;
217 barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
219 EXPORT_SYMBOL_GPL(__rcu_read_lock
);
222 * Check for preempted RCU readers blocking the current grace period
223 * for the specified rcu_node structure. If the caller needs a reliable
224 * answer, it must hold the rcu_node's ->lock.
226 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
228 return rnp
->gp_tasks
!= NULL
;
232 * Record a quiescent state for all tasks that were previously queued
233 * on the specified rcu_node structure and that were blocking the current
234 * RCU grace period. The caller must hold the specified rnp->lock with
235 * irqs disabled, and this lock is released upon return, but irqs remain
238 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
239 __releases(rnp
->lock
)
242 struct rcu_node
*rnp_p
;
244 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
245 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
246 return; /* Still need more quiescent states! */
252 * Either there is only one rcu_node in the tree,
253 * or tasks were kicked up to root rcu_node due to
254 * CPUs going offline.
256 rcu_report_qs_rsp(&rcu_preempt_state
, flags
);
260 /* Report up the rest of the hierarchy. */
262 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
263 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
264 rcu_report_qs_rnp(mask
, &rcu_preempt_state
, rnp_p
, flags
);
268 * Advance a ->blkd_tasks-list pointer to the next entry, instead
269 * returning NULL if at the end of the list.
271 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
272 struct rcu_node
*rnp
)
274 struct list_head
*np
;
276 np
= t
->rcu_node_entry
.next
;
277 if (np
== &rnp
->blkd_tasks
)
283 * Handle special cases during rcu_read_unlock(), such as needing to
284 * notify RCU core processing or task having blocked during the RCU
285 * read-side critical section.
287 static void rcu_read_unlock_special(struct task_struct
*t
)
292 struct list_head
*np
;
293 struct rcu_node
*rnp
;
296 /* NMI handlers cannot block and cannot safely manipulate state. */
300 local_irq_save(flags
);
303 * If RCU core is waiting for this CPU to exit critical section,
304 * let it know that we have done so.
306 special
= t
->rcu_read_unlock_special
;
307 if (special
& RCU_READ_UNLOCK_NEED_QS
) {
308 rcu_preempt_qs(smp_processor_id());
311 /* Hardware IRQ handlers cannot block. */
313 local_irq_restore(flags
);
317 /* Clean up if blocked during RCU read-side critical section. */
318 if (special
& RCU_READ_UNLOCK_BLOCKED
) {
319 t
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_BLOCKED
;
322 * Remove this task from the list it blocked on. The
323 * task can migrate while we acquire the lock, but at
324 * most one time. So at most two passes through loop.
327 rnp
= t
->rcu_blocked_node
;
328 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
329 if (rnp
== t
->rcu_blocked_node
)
331 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
333 empty
= !rcu_preempt_blocked_readers_cgp(rnp
);
334 empty_exp
= !rcu_preempted_readers_exp(rnp
);
335 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
336 np
= rcu_next_node_entry(t
, rnp
);
337 list_del_init(&t
->rcu_node_entry
);
338 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
340 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
342 #ifdef CONFIG_RCU_BOOST
343 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
344 rnp
->boost_tasks
= np
;
345 #endif /* #ifdef CONFIG_RCU_BOOST */
346 t
->rcu_blocked_node
= NULL
;
349 * If this was the last task on the current list, and if
350 * we aren't waiting on any CPUs, report the quiescent state.
351 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
354 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
356 rcu_report_unblock_qs_rnp(rnp
, flags
);
358 #ifdef CONFIG_RCU_BOOST
359 /* Unboost if we were boosted. */
360 if (special
& RCU_READ_UNLOCK_BOOSTED
) {
361 t
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_BOOSTED
;
362 rt_mutex_unlock(t
->rcu_boost_mutex
);
363 t
->rcu_boost_mutex
= NULL
;
365 #endif /* #ifdef CONFIG_RCU_BOOST */
368 * If this was the last task on the expedited lists,
369 * then we need to report up the rcu_node hierarchy.
371 if (!empty_exp
&& !rcu_preempted_readers_exp(rnp
))
372 rcu_report_exp_rnp(&rcu_preempt_state
, rnp
);
374 local_irq_restore(flags
);
379 * Tree-preemptable RCU implementation for rcu_read_unlock().
380 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
381 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
382 * invoke rcu_read_unlock_special() to clean up after a context switch
383 * in an RCU read-side critical section and other special cases.
385 void __rcu_read_unlock(void)
387 struct task_struct
*t
= current
;
389 barrier(); /* needed if we ever invoke rcu_read_unlock in rcutree.c */
390 --t
->rcu_read_lock_nesting
;
391 barrier(); /* decrement before load of ->rcu_read_unlock_special */
392 if (t
->rcu_read_lock_nesting
== 0 &&
393 unlikely(ACCESS_ONCE(t
->rcu_read_unlock_special
)))
394 rcu_read_unlock_special(t
);
395 #ifdef CONFIG_PROVE_LOCKING
396 WARN_ON_ONCE(ACCESS_ONCE(t
->rcu_read_lock_nesting
) < 0);
397 #endif /* #ifdef CONFIG_PROVE_LOCKING */
399 EXPORT_SYMBOL_GPL(__rcu_read_unlock
);
401 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
404 * Dump detailed information for all tasks blocking the current RCU
405 * grace period on the specified rcu_node structure.
407 static void rcu_print_detail_task_stall_rnp(struct rcu_node
*rnp
)
410 struct task_struct
*t
;
412 if (!rcu_preempt_blocked_readers_cgp(rnp
))
414 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
415 t
= list_entry(rnp
->gp_tasks
,
416 struct task_struct
, rcu_node_entry
);
417 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
419 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
423 * Dump detailed information for all tasks blocking the current RCU
426 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
428 struct rcu_node
*rnp
= rcu_get_root(rsp
);
430 rcu_print_detail_task_stall_rnp(rnp
);
431 rcu_for_each_leaf_node(rsp
, rnp
)
432 rcu_print_detail_task_stall_rnp(rnp
);
435 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
437 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
441 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
444 * Scan the current list of tasks blocked within RCU read-side critical
445 * sections, printing out the tid of each.
447 static void rcu_print_task_stall(struct rcu_node
*rnp
)
449 struct task_struct
*t
;
451 if (!rcu_preempt_blocked_readers_cgp(rnp
))
453 t
= list_entry(rnp
->gp_tasks
,
454 struct task_struct
, rcu_node_entry
);
455 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
456 printk(" P%d", t
->pid
);
460 * Suppress preemptible RCU's CPU stall warnings by pushing the
461 * time of the next stall-warning message comfortably far into the
464 static void rcu_preempt_stall_reset(void)
466 rcu_preempt_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
470 * Check that the list of blocked tasks for the newly completed grace
471 * period is in fact empty. It is a serious bug to complete a grace
472 * period that still has RCU readers blocked! This function must be
473 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
474 * must be held by the caller.
476 * Also, if there are blocked tasks on the list, they automatically
477 * block the newly created grace period, so set up ->gp_tasks accordingly.
479 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
481 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
482 if (!list_empty(&rnp
->blkd_tasks
))
483 rnp
->gp_tasks
= rnp
->blkd_tasks
.next
;
484 WARN_ON_ONCE(rnp
->qsmask
);
487 #ifdef CONFIG_HOTPLUG_CPU
490 * Handle tasklist migration for case in which all CPUs covered by the
491 * specified rcu_node have gone offline. Move them up to the root
492 * rcu_node. The reason for not just moving them to the immediate
493 * parent is to remove the need for rcu_read_unlock_special() to
494 * make more than two attempts to acquire the target rcu_node's lock.
495 * Returns true if there were tasks blocking the current RCU grace
498 * Returns 1 if there was previously a task blocking the current grace
499 * period on the specified rcu_node structure.
501 * The caller must hold rnp->lock with irqs disabled.
503 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
504 struct rcu_node
*rnp
,
505 struct rcu_data
*rdp
)
507 struct list_head
*lp
;
508 struct list_head
*lp_root
;
510 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
511 struct task_struct
*t
;
513 if (rnp
== rnp_root
) {
514 WARN_ONCE(1, "Last CPU thought to be offlined?");
515 return 0; /* Shouldn't happen: at least one CPU online. */
518 /* If we are on an internal node, complain bitterly. */
519 WARN_ON_ONCE(rnp
!= rdp
->mynode
);
522 * Move tasks up to root rcu_node. Don't try to get fancy for
523 * this corner-case operation -- just put this node's tasks
524 * at the head of the root node's list, and update the root node's
525 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
526 * if non-NULL. This might result in waiting for more tasks than
527 * absolutely necessary, but this is a good performance/complexity
530 if (rcu_preempt_blocked_readers_cgp(rnp
))
531 retval
|= RCU_OFL_TASKS_NORM_GP
;
532 if (rcu_preempted_readers_exp(rnp
))
533 retval
|= RCU_OFL_TASKS_EXP_GP
;
534 lp
= &rnp
->blkd_tasks
;
535 lp_root
= &rnp_root
->blkd_tasks
;
536 while (!list_empty(lp
)) {
537 t
= list_entry(lp
->next
, typeof(*t
), rcu_node_entry
);
538 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
539 list_del(&t
->rcu_node_entry
);
540 t
->rcu_blocked_node
= rnp_root
;
541 list_add(&t
->rcu_node_entry
, lp_root
);
542 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
543 rnp_root
->gp_tasks
= rnp
->gp_tasks
;
544 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
545 rnp_root
->exp_tasks
= rnp
->exp_tasks
;
546 #ifdef CONFIG_RCU_BOOST
547 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
548 rnp_root
->boost_tasks
= rnp
->boost_tasks
;
549 #endif /* #ifdef CONFIG_RCU_BOOST */
550 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
553 #ifdef CONFIG_RCU_BOOST
554 /* In case root is being boosted and leaf is not. */
555 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
556 if (rnp_root
->boost_tasks
!= NULL
&&
557 rnp_root
->boost_tasks
!= rnp_root
->gp_tasks
)
558 rnp_root
->boost_tasks
= rnp_root
->gp_tasks
;
559 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
560 #endif /* #ifdef CONFIG_RCU_BOOST */
562 rnp
->gp_tasks
= NULL
;
563 rnp
->exp_tasks
= NULL
;
568 * Do CPU-offline processing for preemptable RCU.
570 static void rcu_preempt_offline_cpu(int cpu
)
572 __rcu_offline_cpu(cpu
, &rcu_preempt_state
);
575 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
578 * Check for a quiescent state from the current CPU. When a task blocks,
579 * the task is recorded in the corresponding CPU's rcu_node structure,
580 * which is checked elsewhere.
582 * Caller must disable hard irqs.
584 static void rcu_preempt_check_callbacks(int cpu
)
586 struct task_struct
*t
= current
;
588 if (t
->rcu_read_lock_nesting
== 0) {
592 if (per_cpu(rcu_preempt_data
, cpu
).qs_pending
)
593 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_NEED_QS
;
597 * Process callbacks for preemptable RCU.
599 static void rcu_preempt_process_callbacks(void)
601 __rcu_process_callbacks(&rcu_preempt_state
,
602 &__get_cpu_var(rcu_preempt_data
));
606 * Queue a preemptable-RCU callback for invocation after a grace period.
608 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
610 __call_rcu(head
, func
, &rcu_preempt_state
);
612 EXPORT_SYMBOL_GPL(call_rcu
);
615 * synchronize_rcu - wait until a grace period has elapsed.
617 * Control will return to the caller some time after a full grace
618 * period has elapsed, in other words after all currently executing RCU
619 * read-side critical sections have completed. Note, however, that
620 * upon return from synchronize_rcu(), the caller might well be executing
621 * concurrently with new RCU read-side critical sections that began while
622 * synchronize_rcu() was waiting. RCU read-side critical sections are
623 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
625 void synchronize_rcu(void)
627 struct rcu_synchronize rcu
;
629 if (!rcu_scheduler_active
)
632 init_rcu_head_on_stack(&rcu
.head
);
633 init_completion(&rcu
.completion
);
634 /* Will wake me after RCU finished. */
635 call_rcu(&rcu
.head
, wakeme_after_rcu
);
637 wait_for_completion(&rcu
.completion
);
638 destroy_rcu_head_on_stack(&rcu
.head
);
640 EXPORT_SYMBOL_GPL(synchronize_rcu
);
642 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq
);
643 static long sync_rcu_preempt_exp_count
;
644 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex
);
647 * Return non-zero if there are any tasks in RCU read-side critical
648 * sections blocking the current preemptible-RCU expedited grace period.
649 * If there is no preemptible-RCU expedited grace period currently in
650 * progress, returns zero unconditionally.
652 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
)
654 return rnp
->exp_tasks
!= NULL
;
658 * return non-zero if there is no RCU expedited grace period in progress
659 * for the specified rcu_node structure, in other words, if all CPUs and
660 * tasks covered by the specified rcu_node structure have done their bit
661 * for the current expedited grace period. Works only for preemptible
662 * RCU -- other RCU implementation use other means.
664 * Caller must hold sync_rcu_preempt_exp_mutex.
666 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
668 return !rcu_preempted_readers_exp(rnp
) &&
669 ACCESS_ONCE(rnp
->expmask
) == 0;
673 * Report the exit from RCU read-side critical section for the last task
674 * that queued itself during or before the current expedited preemptible-RCU
675 * grace period. This event is reported either to the rcu_node structure on
676 * which the task was queued or to one of that rcu_node structure's ancestors,
677 * recursively up the tree. (Calm down, calm down, we do the recursion
680 * Caller must hold sync_rcu_preempt_exp_mutex.
682 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
687 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
689 if (!sync_rcu_preempt_exp_done(rnp
))
691 if (rnp
->parent
== NULL
) {
692 wake_up(&sync_rcu_preempt_exp_wq
);
696 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
698 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
699 rnp
->expmask
&= ~mask
;
701 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
705 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
706 * grace period for the specified rcu_node structure. If there are no such
707 * tasks, report it up the rcu_node hierarchy.
709 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
712 sync_rcu_preempt_exp_init(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
716 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
717 if (!list_empty(&rnp
->blkd_tasks
)) {
718 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
719 rcu_initiate_boost(rnp
);
722 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
724 rcu_report_exp_rnp(rsp
, rnp
);
728 * Wait for an rcu-preempt grace period, but expedite it. The basic idea
729 * is to invoke synchronize_sched_expedited() to push all the tasks to
730 * the ->blkd_tasks lists and wait for this list to drain.
732 void synchronize_rcu_expedited(void)
735 struct rcu_node
*rnp
;
736 struct rcu_state
*rsp
= &rcu_preempt_state
;
740 smp_mb(); /* Caller's modifications seen first by other CPUs. */
741 snap
= ACCESS_ONCE(sync_rcu_preempt_exp_count
) + 1;
742 smp_mb(); /* Above access cannot bleed into critical section. */
745 * Acquire lock, falling back to synchronize_rcu() if too many
746 * lock-acquisition failures. Of course, if someone does the
747 * expedited grace period for us, just leave.
749 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex
)) {
751 udelay(trycount
* num_online_cpus());
756 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count
) - snap
) > 0)
757 goto mb_ret
; /* Others did our work for us. */
759 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count
) - snap
) > 0)
760 goto unlock_mb_ret
; /* Others did our work for us. */
762 /* force all RCU readers onto ->blkd_tasks lists. */
763 synchronize_sched_expedited();
765 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
767 /* Initialize ->expmask for all non-leaf rcu_node structures. */
768 rcu_for_each_nonleaf_node_breadth_first(rsp
, rnp
) {
769 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
770 rnp
->expmask
= rnp
->qsmaskinit
;
771 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
774 /* Snapshot current state of ->blkd_tasks lists. */
775 rcu_for_each_leaf_node(rsp
, rnp
)
776 sync_rcu_preempt_exp_init(rsp
, rnp
);
777 if (NUM_RCU_NODES
> 1)
778 sync_rcu_preempt_exp_init(rsp
, rcu_get_root(rsp
));
780 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
782 /* Wait for snapshotted ->blkd_tasks lists to drain. */
783 rnp
= rcu_get_root(rsp
);
784 wait_event(sync_rcu_preempt_exp_wq
,
785 sync_rcu_preempt_exp_done(rnp
));
787 /* Clean up and exit. */
788 smp_mb(); /* ensure expedited GP seen before counter increment. */
789 ACCESS_ONCE(sync_rcu_preempt_exp_count
)++;
791 mutex_unlock(&sync_rcu_preempt_exp_mutex
);
793 smp_mb(); /* ensure subsequent action seen after grace period. */
795 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
798 * Check to see if there is any immediate preemptable-RCU-related work
801 static int rcu_preempt_pending(int cpu
)
803 return __rcu_pending(&rcu_preempt_state
,
804 &per_cpu(rcu_preempt_data
, cpu
));
808 * Does preemptable RCU need the CPU to stay out of dynticks mode?
810 static int rcu_preempt_needs_cpu(int cpu
)
812 return !!per_cpu(rcu_preempt_data
, cpu
).nxtlist
;
816 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
818 void rcu_barrier(void)
820 _rcu_barrier(&rcu_preempt_state
, call_rcu
);
822 EXPORT_SYMBOL_GPL(rcu_barrier
);
825 * Initialize preemptable RCU's per-CPU data.
827 static void __cpuinit
rcu_preempt_init_percpu_data(int cpu
)
829 rcu_init_percpu_data(cpu
, &rcu_preempt_state
, 1);
833 * Move preemptable RCU's callbacks from dying CPU to other online CPU.
835 static void rcu_preempt_send_cbs_to_online(void)
837 rcu_send_cbs_to_online(&rcu_preempt_state
);
841 * Initialize preemptable RCU's state structures.
843 static void __init
__rcu_init_preempt(void)
845 rcu_init_one(&rcu_preempt_state
, &rcu_preempt_data
);
849 * Check for a task exiting while in a preemptable-RCU read-side
850 * critical section, clean up if so. No need to issue warnings,
851 * as debug_check_no_locks_held() already does this if lockdep
856 struct task_struct
*t
= current
;
858 if (t
->rcu_read_lock_nesting
== 0)
860 t
->rcu_read_lock_nesting
= 1;
864 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
866 static struct rcu_state
*rcu_state
= &rcu_sched_state
;
869 * Tell them what RCU they are running.
871 static void __init
rcu_bootup_announce(void)
873 printk(KERN_INFO
"Hierarchical RCU implementation.\n");
874 rcu_bootup_announce_oddness();
878 * Return the number of RCU batches processed thus far for debug & stats.
880 long rcu_batches_completed(void)
882 return rcu_batches_completed_sched();
884 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
887 * Force a quiescent state for RCU, which, because there is no preemptible
888 * RCU, becomes the same as rcu-sched.
890 void rcu_force_quiescent_state(void)
892 rcu_sched_force_quiescent_state();
894 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
897 * Because preemptable RCU does not exist, we never have to check for
898 * CPUs being in quiescent states.
900 static void rcu_preempt_note_context_switch(int cpu
)
905 * Because preemptable RCU does not exist, there are never any preempted
908 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
913 #ifdef CONFIG_HOTPLUG_CPU
915 /* Because preemptible RCU does not exist, no quieting of tasks. */
916 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
918 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
921 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
924 * Because preemptable RCU does not exist, we never have to check for
925 * tasks blocked within RCU read-side critical sections.
927 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
932 * Because preemptable RCU does not exist, we never have to check for
933 * tasks blocked within RCU read-side critical sections.
935 static void rcu_print_task_stall(struct rcu_node
*rnp
)
940 * Because preemptible RCU does not exist, there is no need to suppress
941 * its CPU stall warnings.
943 static void rcu_preempt_stall_reset(void)
948 * Because there is no preemptable RCU, there can be no readers blocked,
949 * so there is no need to check for blocked tasks. So check only for
950 * bogus qsmask values.
952 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
954 WARN_ON_ONCE(rnp
->qsmask
);
957 #ifdef CONFIG_HOTPLUG_CPU
960 * Because preemptable RCU does not exist, it never needs to migrate
961 * tasks that were blocked within RCU read-side critical sections, and
962 * such non-existent tasks cannot possibly have been blocking the current
965 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
966 struct rcu_node
*rnp
,
967 struct rcu_data
*rdp
)
973 * Because preemptable RCU does not exist, it never needs CPU-offline
976 static void rcu_preempt_offline_cpu(int cpu
)
980 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
983 * Because preemptable RCU does not exist, it never has any callbacks
986 static void rcu_preempt_check_callbacks(int cpu
)
991 * Because preemptable RCU does not exist, it never has any callbacks
994 static void rcu_preempt_process_callbacks(void)
999 * Wait for an rcu-preempt grace period, but make it happen quickly.
1000 * But because preemptable RCU does not exist, map to rcu-sched.
1002 void synchronize_rcu_expedited(void)
1004 synchronize_sched_expedited();
1006 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
1008 #ifdef CONFIG_HOTPLUG_CPU
1011 * Because preemptable RCU does not exist, there is never any need to
1012 * report on tasks preempted in RCU read-side critical sections during
1013 * expedited RCU grace periods.
1015 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1020 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1023 * Because preemptable RCU does not exist, it never has any work to do.
1025 static int rcu_preempt_pending(int cpu
)
1031 * Because preemptable RCU does not exist, it never needs any CPU.
1033 static int rcu_preempt_needs_cpu(int cpu
)
1039 * Because preemptable RCU does not exist, rcu_barrier() is just
1040 * another name for rcu_barrier_sched().
1042 void rcu_barrier(void)
1044 rcu_barrier_sched();
1046 EXPORT_SYMBOL_GPL(rcu_barrier
);
1049 * Because preemptable RCU does not exist, there is no per-CPU
1050 * data to initialize.
1052 static void __cpuinit
rcu_preempt_init_percpu_data(int cpu
)
1057 * Because there is no preemptable RCU, there are no callbacks to move.
1059 static void rcu_preempt_send_cbs_to_online(void)
1064 * Because preemptable RCU does not exist, it need not be initialized.
1066 static void __init
__rcu_init_preempt(void)
1070 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1072 #ifdef CONFIG_RCU_BOOST
1074 #include "rtmutex_common.h"
1076 #ifdef CONFIG_RCU_TRACE
1078 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1080 if (list_empty(&rnp
->blkd_tasks
))
1081 rnp
->n_balk_blkd_tasks
++;
1082 else if (rnp
->exp_tasks
== NULL
&& rnp
->gp_tasks
== NULL
)
1083 rnp
->n_balk_exp_gp_tasks
++;
1084 else if (rnp
->gp_tasks
!= NULL
&& rnp
->boost_tasks
!= NULL
)
1085 rnp
->n_balk_boost_tasks
++;
1086 else if (rnp
->gp_tasks
!= NULL
&& rnp
->qsmask
!= 0)
1087 rnp
->n_balk_notblocked
++;
1088 else if (rnp
->gp_tasks
!= NULL
&&
1089 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))
1090 rnp
->n_balk_notyet
++;
1095 #else /* #ifdef CONFIG_RCU_TRACE */
1097 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1101 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1104 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1105 * or ->boost_tasks, advancing the pointer to the next task in the
1106 * ->blkd_tasks list.
1108 * Note that irqs must be enabled: boosting the task can block.
1109 * Returns 1 if there are more tasks needing to be boosted.
1111 static int rcu_boost(struct rcu_node
*rnp
)
1113 unsigned long flags
;
1114 struct rt_mutex mtx
;
1115 struct task_struct
*t
;
1116 struct list_head
*tb
;
1118 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
)
1119 return 0; /* Nothing left to boost. */
1121 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1124 * Recheck under the lock: all tasks in need of boosting
1125 * might exit their RCU read-side critical sections on their own.
1127 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
1128 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1133 * Preferentially boost tasks blocking expedited grace periods.
1134 * This cannot starve the normal grace periods because a second
1135 * expedited grace period must boost all blocked tasks, including
1136 * those blocking the pre-existing normal grace period.
1138 if (rnp
->exp_tasks
!= NULL
) {
1139 tb
= rnp
->exp_tasks
;
1140 rnp
->n_exp_boosts
++;
1142 tb
= rnp
->boost_tasks
;
1143 rnp
->n_normal_boosts
++;
1145 rnp
->n_tasks_boosted
++;
1148 * We boost task t by manufacturing an rt_mutex that appears to
1149 * be held by task t. We leave a pointer to that rt_mutex where
1150 * task t can find it, and task t will release the mutex when it
1151 * exits its outermost RCU read-side critical section. Then
1152 * simply acquiring this artificial rt_mutex will boost task
1153 * t's priority. (Thanks to tglx for suggesting this approach!)
1155 * Note that task t must acquire rnp->lock to remove itself from
1156 * the ->blkd_tasks list, which it will do from exit() if from
1157 * nowhere else. We therefore are guaranteed that task t will
1158 * stay around at least until we drop rnp->lock. Note that
1159 * rnp->lock also resolves races between our priority boosting
1160 * and task t's exiting its outermost RCU read-side critical
1163 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1164 rt_mutex_init_proxy_locked(&mtx
, t
);
1165 t
->rcu_boost_mutex
= &mtx
;
1166 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_BOOSTED
;
1167 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1168 rt_mutex_lock(&mtx
); /* Side effect: boosts task t's priority. */
1169 rt_mutex_unlock(&mtx
); /* Keep lockdep happy. */
1171 return rnp
->exp_tasks
!= NULL
|| rnp
->boost_tasks
!= NULL
;
1175 * Timer handler to initiate waking up of boost kthreads that
1176 * have yielded the CPU due to excessive numbers of tasks to
1177 * boost. We wake up the per-rcu_node kthread, which in turn
1178 * will wake up the booster kthread.
1180 static void rcu_boost_kthread_timer(unsigned long arg
)
1182 unsigned long flags
;
1183 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1185 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1186 invoke_rcu_node_kthread(rnp
);
1187 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1191 * Priority-boosting kthread. One per leaf rcu_node and one for the
1194 static int rcu_boost_kthread(void *arg
)
1196 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1201 rnp
->boost_kthread_status
= RCU_KTHREAD_WAITING
;
1202 wait_event_interruptible(rnp
->boost_wq
, rnp
->boost_tasks
||
1204 kthread_should_stop());
1205 if (kthread_should_stop())
1207 rnp
->boost_kthread_status
= RCU_KTHREAD_RUNNING
;
1208 more2boost
= rcu_boost(rnp
);
1214 rcu_yield(rcu_boost_kthread_timer
, (unsigned long)rnp
);
1218 rnp
->boost_kthread_status
= RCU_KTHREAD_STOPPED
;
1223 * Check to see if it is time to start boosting RCU readers that are
1224 * blocking the current grace period, and, if so, tell the per-rcu_node
1225 * kthread to start boosting them. If there is an expedited grace
1226 * period in progress, it is always time to boost.
1228 * The caller must hold rnp->lock.
1230 static void rcu_initiate_boost(struct rcu_node
*rnp
)
1232 struct task_struct
*t
;
1234 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1235 rnp
->n_balk_exp_gp_tasks
++;
1238 if (rnp
->exp_tasks
!= NULL
||
1239 (rnp
->gp_tasks
!= NULL
&&
1240 rnp
->boost_tasks
== NULL
&&
1242 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))) {
1243 if (rnp
->exp_tasks
== NULL
)
1244 rnp
->boost_tasks
= rnp
->gp_tasks
;
1245 t
= rnp
->boost_kthread_task
;
1249 rcu_initiate_boost_trace(rnp
);
1253 * Set the affinity of the boost kthread. The CPU-hotplug locks are
1254 * held, so no one should be messing with the existence of the boost
1257 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
,
1260 struct task_struct
*t
;
1262 t
= rnp
->boost_kthread_task
;
1264 set_cpus_allowed_ptr(rnp
->boost_kthread_task
, cm
);
1267 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1270 * Do priority-boost accounting for the start of a new grace period.
1272 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1274 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1278 * Initialize the RCU-boost waitqueue.
1280 static void __init
rcu_init_boost_waitqueue(struct rcu_node
*rnp
)
1282 init_waitqueue_head(&rnp
->boost_wq
);
1286 * Create an RCU-boost kthread for the specified node if one does not
1287 * already exist. We only create this kthread for preemptible RCU.
1288 * Returns zero if all is well, a negated errno otherwise.
1290 static int __cpuinit
rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1291 struct rcu_node
*rnp
,
1294 unsigned long flags
;
1295 struct sched_param sp
;
1296 struct task_struct
*t
;
1298 if (&rcu_preempt_state
!= rsp
)
1300 if (rnp
->boost_kthread_task
!= NULL
)
1302 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1303 "rcub%d", rnp_index
);
1306 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1307 rnp
->boost_kthread_task
= t
;
1308 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1310 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1311 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1315 #ifdef CONFIG_HOTPLUG_CPU
1317 static void rcu_stop_boost_kthread(struct rcu_node
*rnp
)
1319 unsigned long flags
;
1320 struct task_struct
*t
;
1322 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1323 t
= rnp
->boost_kthread_task
;
1324 rnp
->boost_kthread_task
= NULL
;
1325 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1330 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1332 #else /* #ifdef CONFIG_RCU_BOOST */
1334 static void rcu_initiate_boost(struct rcu_node
*rnp
)
1338 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
,
1343 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1347 static void __init
rcu_init_boost_waitqueue(struct rcu_node
*rnp
)
1351 static int __cpuinit
rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1352 struct rcu_node
*rnp
,
1358 #ifdef CONFIG_HOTPLUG_CPU
1360 static void rcu_stop_boost_kthread(struct rcu_node
*rnp
)
1364 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1366 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1370 void synchronize_sched_expedited(void)
1374 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
1376 #else /* #ifndef CONFIG_SMP */
1378 static atomic_t sync_sched_expedited_started
= ATOMIC_INIT(0);
1379 static atomic_t sync_sched_expedited_done
= ATOMIC_INIT(0);
1381 static int synchronize_sched_expedited_cpu_stop(void *data
)
1384 * There must be a full memory barrier on each affected CPU
1385 * between the time that try_stop_cpus() is called and the
1386 * time that it returns.
1388 * In the current initial implementation of cpu_stop, the
1389 * above condition is already met when the control reaches
1390 * this point and the following smp_mb() is not strictly
1391 * necessary. Do smp_mb() anyway for documentation and
1392 * robustness against future implementation changes.
1394 smp_mb(); /* See above comment block. */
1399 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
1400 * approach to force grace period to end quickly. This consumes
1401 * significant time on all CPUs, and is thus not recommended for
1402 * any sort of common-case code.
1404 * Note that it is illegal to call this function while holding any
1405 * lock that is acquired by a CPU-hotplug notifier. Failing to
1406 * observe this restriction will result in deadlock.
1408 * This implementation can be thought of as an application of ticket
1409 * locking to RCU, with sync_sched_expedited_started and
1410 * sync_sched_expedited_done taking on the roles of the halves
1411 * of the ticket-lock word. Each task atomically increments
1412 * sync_sched_expedited_started upon entry, snapshotting the old value,
1413 * then attempts to stop all the CPUs. If this succeeds, then each
1414 * CPU will have executed a context switch, resulting in an RCU-sched
1415 * grace period. We are then done, so we use atomic_cmpxchg() to
1416 * update sync_sched_expedited_done to match our snapshot -- but
1417 * only if someone else has not already advanced past our snapshot.
1419 * On the other hand, if try_stop_cpus() fails, we check the value
1420 * of sync_sched_expedited_done. If it has advanced past our
1421 * initial snapshot, then someone else must have forced a grace period
1422 * some time after we took our snapshot. In this case, our work is
1423 * done for us, and we can simply return. Otherwise, we try again,
1424 * but keep our initial snapshot for purposes of checking for someone
1425 * doing our work for us.
1427 * If we fail too many times in a row, we fall back to synchronize_sched().
1429 void synchronize_sched_expedited(void)
1431 int firstsnap
, s
, snap
, trycount
= 0;
1433 /* Note that atomic_inc_return() implies full memory barrier. */
1434 firstsnap
= snap
= atomic_inc_return(&sync_sched_expedited_started
);
1438 * Each pass through the following loop attempts to force a
1439 * context switch on each CPU.
1441 while (try_stop_cpus(cpu_online_mask
,
1442 synchronize_sched_expedited_cpu_stop
,
1446 /* No joy, try again later. Or just synchronize_sched(). */
1447 if (trycount
++ < 10)
1448 udelay(trycount
* num_online_cpus());
1450 synchronize_sched();
1454 /* Check to see if someone else did our work for us. */
1455 s
= atomic_read(&sync_sched_expedited_done
);
1456 if (UINT_CMP_GE((unsigned)s
, (unsigned)firstsnap
)) {
1457 smp_mb(); /* ensure test happens before caller kfree */
1462 * Refetching sync_sched_expedited_started allows later
1463 * callers to piggyback on our grace period. We subtract
1464 * 1 to get the same token that the last incrementer got.
1465 * We retry after they started, so our grace period works
1466 * for them, and they started after our first try, so their
1467 * grace period works for us.
1470 snap
= atomic_read(&sync_sched_expedited_started
) - 1;
1471 smp_mb(); /* ensure read is before try_stop_cpus(). */
1475 * Everyone up to our most recent fetch is covered by our grace
1476 * period. Update the counter, but only if our work is still
1477 * relevant -- which it won't be if someone who started later
1478 * than we did beat us to the punch.
1481 s
= atomic_read(&sync_sched_expedited_done
);
1482 if (UINT_CMP_GE((unsigned)s
, (unsigned)snap
)) {
1483 smp_mb(); /* ensure test happens before caller kfree */
1486 } while (atomic_cmpxchg(&sync_sched_expedited_done
, s
, snap
) != s
);
1490 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
1492 #endif /* #else #ifndef CONFIG_SMP */
1494 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1497 * Check to see if any future RCU-related work will need to be done
1498 * by the current CPU, even if none need be done immediately, returning
1499 * 1 if so. This function is part of the RCU implementation; it is -not-
1500 * an exported member of the RCU API.
1502 * Because we have preemptible RCU, just check whether this CPU needs
1503 * any flavor of RCU. Do not chew up lots of CPU cycles with preemption
1504 * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
1506 int rcu_needs_cpu(int cpu
)
1508 return rcu_needs_cpu_quick_check(cpu
);
1512 * Check to see if we need to continue a callback-flush operations to
1513 * allow the last CPU to enter dyntick-idle mode. But fast dyntick-idle
1514 * entry is not configured, so we never do need to.
1516 static void rcu_needs_cpu_flush(void)
1520 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1522 #define RCU_NEEDS_CPU_FLUSHES 5
1523 static DEFINE_PER_CPU(int, rcu_dyntick_drain
);
1524 static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff
);
1527 * Check to see if any future RCU-related work will need to be done
1528 * by the current CPU, even if none need be done immediately, returning
1529 * 1 if so. This function is part of the RCU implementation; it is -not-
1530 * an exported member of the RCU API.
1532 * Because we are not supporting preemptible RCU, attempt to accelerate
1533 * any current grace periods so that RCU no longer needs this CPU, but
1534 * only if all other CPUs are already in dynticks-idle mode. This will
1535 * allow the CPU cores to be powered down immediately, as opposed to after
1536 * waiting many milliseconds for grace periods to elapse.
1538 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1539 * disabled, we do one pass of force_quiescent_state(), then do a
1540 * invoke_rcu_cpu_kthread() to cause rcu_process_callbacks() to be invoked
1541 * later. The per-cpu rcu_dyntick_drain variable controls the sequencing.
1543 int rcu_needs_cpu(int cpu
)
1549 /* Check for being in the holdoff period. */
1550 if (per_cpu(rcu_dyntick_holdoff
, cpu
) == jiffies
)
1551 return rcu_needs_cpu_quick_check(cpu
);
1553 /* Don't bother unless we are the last non-dyntick-idle CPU. */
1554 for_each_online_cpu(thatcpu
) {
1557 snap
= atomic_add_return(0, &per_cpu(rcu_dynticks
,
1559 smp_mb(); /* Order sampling of snap with end of grace period. */
1560 if ((snap
& 0x1) != 0) {
1561 per_cpu(rcu_dyntick_drain
, cpu
) = 0;
1562 per_cpu(rcu_dyntick_holdoff
, cpu
) = jiffies
- 1;
1563 return rcu_needs_cpu_quick_check(cpu
);
1567 /* Check and update the rcu_dyntick_drain sequencing. */
1568 if (per_cpu(rcu_dyntick_drain
, cpu
) <= 0) {
1569 /* First time through, initialize the counter. */
1570 per_cpu(rcu_dyntick_drain
, cpu
) = RCU_NEEDS_CPU_FLUSHES
;
1571 } else if (--per_cpu(rcu_dyntick_drain
, cpu
) <= 0) {
1572 /* We have hit the limit, so time to give up. */
1573 per_cpu(rcu_dyntick_holdoff
, cpu
) = jiffies
;
1574 return rcu_needs_cpu_quick_check(cpu
);
1577 /* Do one step pushing remaining RCU callbacks through. */
1578 if (per_cpu(rcu_sched_data
, cpu
).nxtlist
) {
1580 force_quiescent_state(&rcu_sched_state
, 0);
1581 c
= c
|| per_cpu(rcu_sched_data
, cpu
).nxtlist
;
1583 if (per_cpu(rcu_bh_data
, cpu
).nxtlist
) {
1585 force_quiescent_state(&rcu_bh_state
, 0);
1586 c
= c
|| per_cpu(rcu_bh_data
, cpu
).nxtlist
;
1589 /* If RCU callbacks are still pending, RCU still needs this CPU. */
1591 invoke_rcu_cpu_kthread();
1596 * Check to see if we need to continue a callback-flush operations to
1597 * allow the last CPU to enter dyntick-idle mode.
1599 static void rcu_needs_cpu_flush(void)
1601 int cpu
= smp_processor_id();
1602 unsigned long flags
;
1604 if (per_cpu(rcu_dyntick_drain
, cpu
) <= 0)
1606 local_irq_save(flags
);
1607 (void)rcu_needs_cpu(cpu
);
1608 local_irq_restore(flags
);
1611 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */