2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/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
);
68 DEFINE_PER_CPU(struct rcu_data
, rcu_preempt_data
);
69 static struct rcu_state
*rcu_state
= &rcu_preempt_state
;
71 static void rcu_read_unlock_special(struct task_struct
*t
);
72 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
);
75 * Tell them what RCU they are running.
77 static void __init
rcu_bootup_announce(void)
79 printk(KERN_INFO
"Preemptible hierarchical RCU implementation.\n");
80 rcu_bootup_announce_oddness();
84 * Return the number of RCU-preempt batches processed thus far
85 * for debug and statistics.
87 long rcu_batches_completed_preempt(void)
89 return rcu_preempt_state
.completed
;
91 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt
);
94 * Return the number of RCU batches processed thus far for debug & stats.
96 long rcu_batches_completed(void)
98 return rcu_batches_completed_preempt();
100 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
103 * Force a quiescent state for preemptible RCU.
105 void rcu_force_quiescent_state(void)
107 force_quiescent_state(&rcu_preempt_state
, 0);
109 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
112 * Record a preemptible-RCU quiescent state for the specified CPU. Note
113 * that this just means that the task currently running on the CPU is
114 * not in a quiescent state. There might be any number of tasks blocked
115 * while in an RCU read-side critical section.
117 * Unlike the other rcu_*_qs() functions, callers to this function
118 * must disable irqs in order to protect the assignment to
119 * ->rcu_read_unlock_special.
121 static void rcu_preempt_qs(int cpu
)
123 struct rcu_data
*rdp
= &per_cpu(rcu_preempt_data
, cpu
);
125 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
127 rdp
->passed_quiesc
= 1;
128 current
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_NEED_QS
;
132 * We have entered the scheduler, and the current task might soon be
133 * context-switched away from. If this task is in an RCU read-side
134 * critical section, we will no longer be able to rely on the CPU to
135 * record that fact, so we enqueue the task on the blkd_tasks list.
136 * The task will dequeue itself when it exits the outermost enclosing
137 * RCU read-side critical section. Therefore, the current grace period
138 * cannot be permitted to complete until the blkd_tasks list entries
139 * predating the current grace period drain, in other words, until
140 * rnp->gp_tasks becomes NULL.
142 * Caller must disable preemption.
144 static void rcu_preempt_note_context_switch(int cpu
)
146 struct task_struct
*t
= current
;
148 struct rcu_data
*rdp
;
149 struct rcu_node
*rnp
;
151 if (t
->rcu_read_lock_nesting
> 0 &&
152 (t
->rcu_read_unlock_special
& RCU_READ_UNLOCK_BLOCKED
) == 0) {
154 /* Possibly blocking in an RCU read-side critical section. */
155 rdp
= per_cpu_ptr(rcu_preempt_state
.rda
, cpu
);
157 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
158 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_BLOCKED
;
159 t
->rcu_blocked_node
= rnp
;
162 * If this CPU has already checked in, then this task
163 * will hold up the next grace period rather than the
164 * current grace period. Queue the task accordingly.
165 * If the task is queued for the current grace period
166 * (i.e., this CPU has not yet passed through a quiescent
167 * state for the current grace period), then as long
168 * as that task remains queued, the current grace period
169 * cannot end. Note that there is some uncertainty as
170 * to exactly when the current grace period started.
171 * We take a conservative approach, which can result
172 * in unnecessarily waiting on tasks that started very
173 * slightly after the current grace period began. C'est
176 * But first, note that the current CPU must still be
179 WARN_ON_ONCE((rdp
->grpmask
& rnp
->qsmaskinit
) == 0);
180 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
181 if ((rnp
->qsmask
& rdp
->grpmask
) && rnp
->gp_tasks
!= NULL
) {
182 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
->prev
);
183 rnp
->gp_tasks
= &t
->rcu_node_entry
;
184 #ifdef CONFIG_RCU_BOOST
185 if (rnp
->boost_tasks
!= NULL
)
186 rnp
->boost_tasks
= rnp
->gp_tasks
;
187 #endif /* #ifdef CONFIG_RCU_BOOST */
189 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
190 if (rnp
->qsmask
& rdp
->grpmask
)
191 rnp
->gp_tasks
= &t
->rcu_node_entry
;
193 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
194 } else if (t
->rcu_read_lock_nesting
< 0 &&
195 t
->rcu_read_unlock_special
) {
198 * Complete exit from RCU read-side critical section on
199 * behalf of preempted instance of __rcu_read_unlock().
201 rcu_read_unlock_special(t
);
205 * Either we were not in an RCU read-side critical section to
206 * begin with, or we have now recorded that critical section
207 * globally. Either way, we can now note a quiescent state
208 * for this CPU. Again, if we were in an RCU read-side critical
209 * section, and if that critical section was blocking the current
210 * grace period, then the fact that the task has been enqueued
211 * means that we continue to block the current grace period.
213 local_irq_save(flags
);
215 local_irq_restore(flags
);
219 * Tree-preemptible RCU implementation for rcu_read_lock().
220 * Just increment ->rcu_read_lock_nesting, shared state will be updated
223 void __rcu_read_lock(void)
225 current
->rcu_read_lock_nesting
++;
226 barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
228 EXPORT_SYMBOL_GPL(__rcu_read_lock
);
231 * Check for preempted RCU readers blocking the current grace period
232 * for the specified rcu_node structure. If the caller needs a reliable
233 * answer, it must hold the rcu_node's ->lock.
235 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
237 return rnp
->gp_tasks
!= NULL
;
241 * Record a quiescent state for all tasks that were previously queued
242 * on the specified rcu_node structure and that were blocking the current
243 * RCU grace period. The caller must hold the specified rnp->lock with
244 * irqs disabled, and this lock is released upon return, but irqs remain
247 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
248 __releases(rnp
->lock
)
251 struct rcu_node
*rnp_p
;
253 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
254 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
255 return; /* Still need more quiescent states! */
261 * Either there is only one rcu_node in the tree,
262 * or tasks were kicked up to root rcu_node due to
263 * CPUs going offline.
265 rcu_report_qs_rsp(&rcu_preempt_state
, flags
);
269 /* Report up the rest of the hierarchy. */
271 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
272 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
273 rcu_report_qs_rnp(mask
, &rcu_preempt_state
, rnp_p
, flags
);
277 * Advance a ->blkd_tasks-list pointer to the next entry, instead
278 * returning NULL if at the end of the list.
280 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
281 struct rcu_node
*rnp
)
283 struct list_head
*np
;
285 np
= t
->rcu_node_entry
.next
;
286 if (np
== &rnp
->blkd_tasks
)
292 * Handle special cases during rcu_read_unlock(), such as needing to
293 * notify RCU core processing or task having blocked during the RCU
294 * read-side critical section.
296 static noinline
void rcu_read_unlock_special(struct task_struct
*t
)
301 struct list_head
*np
;
302 struct rcu_node
*rnp
;
305 /* NMI handlers cannot block and cannot safely manipulate state. */
309 local_irq_save(flags
);
312 * If RCU core is waiting for this CPU to exit critical section,
313 * let it know that we have done so.
315 special
= t
->rcu_read_unlock_special
;
316 if (special
& RCU_READ_UNLOCK_NEED_QS
) {
317 rcu_preempt_qs(smp_processor_id());
320 /* Hardware IRQ handlers cannot block. */
321 if (in_irq() || in_serving_softirq()) {
322 local_irq_restore(flags
);
326 /* Clean up if blocked during RCU read-side critical section. */
327 if (special
& RCU_READ_UNLOCK_BLOCKED
) {
328 t
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_BLOCKED
;
331 * Remove this task from the list it blocked on. The
332 * task can migrate while we acquire the lock, but at
333 * most one time. So at most two passes through loop.
336 rnp
= t
->rcu_blocked_node
;
337 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
338 if (rnp
== t
->rcu_blocked_node
)
340 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
342 empty
= !rcu_preempt_blocked_readers_cgp(rnp
);
343 empty_exp
= !rcu_preempted_readers_exp(rnp
);
344 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
345 np
= rcu_next_node_entry(t
, rnp
);
346 list_del_init(&t
->rcu_node_entry
);
347 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
349 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
351 #ifdef CONFIG_RCU_BOOST
352 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
353 rnp
->boost_tasks
= np
;
354 /* Snapshot and clear ->rcu_boosted with rcu_node lock held. */
355 if (t
->rcu_boosted
) {
356 special
|= RCU_READ_UNLOCK_BOOSTED
;
359 #endif /* #ifdef CONFIG_RCU_BOOST */
360 t
->rcu_blocked_node
= NULL
;
363 * If this was the last task on the current list, and if
364 * we aren't waiting on any CPUs, report the quiescent state.
365 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
368 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
370 rcu_report_unblock_qs_rnp(rnp
, flags
);
372 #ifdef CONFIG_RCU_BOOST
373 /* Unboost if we were boosted. */
374 if (special
& RCU_READ_UNLOCK_BOOSTED
) {
375 rt_mutex_unlock(t
->rcu_boost_mutex
);
376 t
->rcu_boost_mutex
= NULL
;
378 #endif /* #ifdef CONFIG_RCU_BOOST */
381 * If this was the last task on the expedited lists,
382 * then we need to report up the rcu_node hierarchy.
384 if (!empty_exp
&& !rcu_preempted_readers_exp(rnp
))
385 rcu_report_exp_rnp(&rcu_preempt_state
, rnp
);
387 local_irq_restore(flags
);
392 * Tree-preemptible RCU implementation for rcu_read_unlock().
393 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
394 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
395 * invoke rcu_read_unlock_special() to clean up after a context switch
396 * in an RCU read-side critical section and other special cases.
398 void __rcu_read_unlock(void)
400 struct task_struct
*t
= current
;
402 barrier(); /* needed if we ever invoke rcu_read_unlock in rcutree.c */
403 if (t
->rcu_read_lock_nesting
!= 1)
404 --t
->rcu_read_lock_nesting
;
406 t
->rcu_read_lock_nesting
= INT_MIN
;
407 barrier(); /* assign before ->rcu_read_unlock_special load */
408 if (unlikely(ACCESS_ONCE(t
->rcu_read_unlock_special
)))
409 rcu_read_unlock_special(t
);
410 barrier(); /* ->rcu_read_unlock_special load before assign */
411 t
->rcu_read_lock_nesting
= 0;
413 #ifdef CONFIG_PROVE_LOCKING
415 int rrln
= ACCESS_ONCE(t
->rcu_read_lock_nesting
);
417 WARN_ON_ONCE(rrln
< 0 && rrln
> INT_MIN
/ 2);
419 #endif /* #ifdef CONFIG_PROVE_LOCKING */
421 EXPORT_SYMBOL_GPL(__rcu_read_unlock
);
423 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
426 * Dump detailed information for all tasks blocking the current RCU
427 * grace period on the specified rcu_node structure.
429 static void rcu_print_detail_task_stall_rnp(struct rcu_node
*rnp
)
432 struct task_struct
*t
;
434 if (!rcu_preempt_blocked_readers_cgp(rnp
))
436 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
437 t
= list_entry(rnp
->gp_tasks
,
438 struct task_struct
, rcu_node_entry
);
439 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
441 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
445 * Dump detailed information for all tasks blocking the current RCU
448 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
450 struct rcu_node
*rnp
= rcu_get_root(rsp
);
452 rcu_print_detail_task_stall_rnp(rnp
);
453 rcu_for_each_leaf_node(rsp
, rnp
)
454 rcu_print_detail_task_stall_rnp(rnp
);
457 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
459 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
463 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
466 * Scan the current list of tasks blocked within RCU read-side critical
467 * sections, printing out the tid of each.
469 static void rcu_print_task_stall(struct rcu_node
*rnp
)
471 struct task_struct
*t
;
473 if (!rcu_preempt_blocked_readers_cgp(rnp
))
475 t
= list_entry(rnp
->gp_tasks
,
476 struct task_struct
, rcu_node_entry
);
477 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
478 printk(" P%d", t
->pid
);
482 * Suppress preemptible RCU's CPU stall warnings by pushing the
483 * time of the next stall-warning message comfortably far into the
486 static void rcu_preempt_stall_reset(void)
488 rcu_preempt_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
492 * Check that the list of blocked tasks for the newly completed grace
493 * period is in fact empty. It is a serious bug to complete a grace
494 * period that still has RCU readers blocked! This function must be
495 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
496 * must be held by the caller.
498 * Also, if there are blocked tasks on the list, they automatically
499 * block the newly created grace period, so set up ->gp_tasks accordingly.
501 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
503 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
504 if (!list_empty(&rnp
->blkd_tasks
))
505 rnp
->gp_tasks
= rnp
->blkd_tasks
.next
;
506 WARN_ON_ONCE(rnp
->qsmask
);
509 #ifdef CONFIG_HOTPLUG_CPU
512 * Handle tasklist migration for case in which all CPUs covered by the
513 * specified rcu_node have gone offline. Move them up to the root
514 * rcu_node. The reason for not just moving them to the immediate
515 * parent is to remove the need for rcu_read_unlock_special() to
516 * make more than two attempts to acquire the target rcu_node's lock.
517 * Returns true if there were tasks blocking the current RCU grace
520 * Returns 1 if there was previously a task blocking the current grace
521 * period on the specified rcu_node structure.
523 * The caller must hold rnp->lock with irqs disabled.
525 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
526 struct rcu_node
*rnp
,
527 struct rcu_data
*rdp
)
529 struct list_head
*lp
;
530 struct list_head
*lp_root
;
532 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
533 struct task_struct
*t
;
535 if (rnp
== rnp_root
) {
536 WARN_ONCE(1, "Last CPU thought to be offlined?");
537 return 0; /* Shouldn't happen: at least one CPU online. */
540 /* If we are on an internal node, complain bitterly. */
541 WARN_ON_ONCE(rnp
!= rdp
->mynode
);
544 * Move tasks up to root rcu_node. Don't try to get fancy for
545 * this corner-case operation -- just put this node's tasks
546 * at the head of the root node's list, and update the root node's
547 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
548 * if non-NULL. This might result in waiting for more tasks than
549 * absolutely necessary, but this is a good performance/complexity
552 if (rcu_preempt_blocked_readers_cgp(rnp
))
553 retval
|= RCU_OFL_TASKS_NORM_GP
;
554 if (rcu_preempted_readers_exp(rnp
))
555 retval
|= RCU_OFL_TASKS_EXP_GP
;
556 lp
= &rnp
->blkd_tasks
;
557 lp_root
= &rnp_root
->blkd_tasks
;
558 while (!list_empty(lp
)) {
559 t
= list_entry(lp
->next
, typeof(*t
), rcu_node_entry
);
560 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
561 list_del(&t
->rcu_node_entry
);
562 t
->rcu_blocked_node
= rnp_root
;
563 list_add(&t
->rcu_node_entry
, lp_root
);
564 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
565 rnp_root
->gp_tasks
= rnp
->gp_tasks
;
566 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
567 rnp_root
->exp_tasks
= rnp
->exp_tasks
;
568 #ifdef CONFIG_RCU_BOOST
569 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
570 rnp_root
->boost_tasks
= rnp
->boost_tasks
;
571 #endif /* #ifdef CONFIG_RCU_BOOST */
572 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
575 #ifdef CONFIG_RCU_BOOST
576 /* In case root is being boosted and leaf is not. */
577 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
578 if (rnp_root
->boost_tasks
!= NULL
&&
579 rnp_root
->boost_tasks
!= rnp_root
->gp_tasks
)
580 rnp_root
->boost_tasks
= rnp_root
->gp_tasks
;
581 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
582 #endif /* #ifdef CONFIG_RCU_BOOST */
584 rnp
->gp_tasks
= NULL
;
585 rnp
->exp_tasks
= NULL
;
590 * Do CPU-offline processing for preemptible RCU.
592 static void rcu_preempt_offline_cpu(int cpu
)
594 __rcu_offline_cpu(cpu
, &rcu_preempt_state
);
597 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
600 * Check for a quiescent state from the current CPU. When a task blocks,
601 * the task is recorded in the corresponding CPU's rcu_node structure,
602 * which is checked elsewhere.
604 * Caller must disable hard irqs.
606 static void rcu_preempt_check_callbacks(int cpu
)
608 struct task_struct
*t
= current
;
610 if (t
->rcu_read_lock_nesting
== 0) {
614 if (t
->rcu_read_lock_nesting
> 0 &&
615 per_cpu(rcu_preempt_data
, cpu
).qs_pending
)
616 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_NEED_QS
;
620 * Process callbacks for preemptible RCU.
622 static void rcu_preempt_process_callbacks(void)
624 __rcu_process_callbacks(&rcu_preempt_state
,
625 &__get_cpu_var(rcu_preempt_data
));
628 #ifdef CONFIG_RCU_BOOST
630 static void rcu_preempt_do_callbacks(void)
632 rcu_do_batch(&rcu_preempt_state
, &__get_cpu_var(rcu_preempt_data
));
635 #endif /* #ifdef CONFIG_RCU_BOOST */
638 * Queue a preemptible-RCU callback for invocation after a grace period.
640 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
642 __call_rcu(head
, func
, &rcu_preempt_state
);
644 EXPORT_SYMBOL_GPL(call_rcu
);
647 * synchronize_rcu - wait until a grace period has elapsed.
649 * Control will return to the caller some time after a full grace
650 * period has elapsed, in other words after all currently executing RCU
651 * read-side critical sections have completed. Note, however, that
652 * upon return from synchronize_rcu(), the caller might well be executing
653 * concurrently with new RCU read-side critical sections that began while
654 * synchronize_rcu() was waiting. RCU read-side critical sections are
655 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
657 void synchronize_rcu(void)
659 if (!rcu_scheduler_active
)
661 wait_rcu_gp(call_rcu
);
663 EXPORT_SYMBOL_GPL(synchronize_rcu
);
665 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq
);
666 static long sync_rcu_preempt_exp_count
;
667 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex
);
670 * Return non-zero if there are any tasks in RCU read-side critical
671 * sections blocking the current preemptible-RCU expedited grace period.
672 * If there is no preemptible-RCU expedited grace period currently in
673 * progress, returns zero unconditionally.
675 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
)
677 return rnp
->exp_tasks
!= NULL
;
681 * return non-zero if there is no RCU expedited grace period in progress
682 * for the specified rcu_node structure, in other words, if all CPUs and
683 * tasks covered by the specified rcu_node structure have done their bit
684 * for the current expedited grace period. Works only for preemptible
685 * RCU -- other RCU implementation use other means.
687 * Caller must hold sync_rcu_preempt_exp_mutex.
689 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
691 return !rcu_preempted_readers_exp(rnp
) &&
692 ACCESS_ONCE(rnp
->expmask
) == 0;
696 * Report the exit from RCU read-side critical section for the last task
697 * that queued itself during or before the current expedited preemptible-RCU
698 * grace period. This event is reported either to the rcu_node structure on
699 * which the task was queued or to one of that rcu_node structure's ancestors,
700 * recursively up the tree. (Calm down, calm down, we do the recursion
703 * Caller must hold sync_rcu_preempt_exp_mutex.
705 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
710 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
712 if (!sync_rcu_preempt_exp_done(rnp
)) {
713 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
716 if (rnp
->parent
== NULL
) {
717 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
718 wake_up(&sync_rcu_preempt_exp_wq
);
722 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
724 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
725 rnp
->expmask
&= ~mask
;
730 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
731 * grace period for the specified rcu_node structure. If there are no such
732 * tasks, report it up the rcu_node hierarchy.
734 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
737 sync_rcu_preempt_exp_init(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
742 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
743 if (list_empty(&rnp
->blkd_tasks
))
744 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
746 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
747 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
751 rcu_report_exp_rnp(rsp
, rnp
);
755 * Wait for an rcu-preempt grace period, but expedite it. The basic idea
756 * is to invoke synchronize_sched_expedited() to push all the tasks to
757 * the ->blkd_tasks lists and wait for this list to drain.
759 void synchronize_rcu_expedited(void)
762 struct rcu_node
*rnp
;
763 struct rcu_state
*rsp
= &rcu_preempt_state
;
767 smp_mb(); /* Caller's modifications seen first by other CPUs. */
768 snap
= ACCESS_ONCE(sync_rcu_preempt_exp_count
) + 1;
769 smp_mb(); /* Above access cannot bleed into critical section. */
772 * Acquire lock, falling back to synchronize_rcu() if too many
773 * lock-acquisition failures. Of course, if someone does the
774 * expedited grace period for us, just leave.
776 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex
)) {
778 udelay(trycount
* num_online_cpus());
783 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count
) - snap
) > 0)
784 goto mb_ret
; /* Others did our work for us. */
786 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count
) - snap
) > 0)
787 goto unlock_mb_ret
; /* Others did our work for us. */
789 /* force all RCU readers onto ->blkd_tasks lists. */
790 synchronize_sched_expedited();
792 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
794 /* Initialize ->expmask for all non-leaf rcu_node structures. */
795 rcu_for_each_nonleaf_node_breadth_first(rsp
, rnp
) {
796 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
797 rnp
->expmask
= rnp
->qsmaskinit
;
798 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
801 /* Snapshot current state of ->blkd_tasks lists. */
802 rcu_for_each_leaf_node(rsp
, rnp
)
803 sync_rcu_preempt_exp_init(rsp
, rnp
);
804 if (NUM_RCU_NODES
> 1)
805 sync_rcu_preempt_exp_init(rsp
, rcu_get_root(rsp
));
807 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
809 /* Wait for snapshotted ->blkd_tasks lists to drain. */
810 rnp
= rcu_get_root(rsp
);
811 wait_event(sync_rcu_preempt_exp_wq
,
812 sync_rcu_preempt_exp_done(rnp
));
814 /* Clean up and exit. */
815 smp_mb(); /* ensure expedited GP seen before counter increment. */
816 ACCESS_ONCE(sync_rcu_preempt_exp_count
)++;
818 mutex_unlock(&sync_rcu_preempt_exp_mutex
);
820 smp_mb(); /* ensure subsequent action seen after grace period. */
822 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
825 * Check to see if there is any immediate preemptible-RCU-related work
828 static int rcu_preempt_pending(int cpu
)
830 return __rcu_pending(&rcu_preempt_state
,
831 &per_cpu(rcu_preempt_data
, cpu
));
835 * Does preemptible RCU need the CPU to stay out of dynticks mode?
837 static int rcu_preempt_needs_cpu(int cpu
)
839 return !!per_cpu(rcu_preempt_data
, cpu
).nxtlist
;
843 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
845 void rcu_barrier(void)
847 _rcu_barrier(&rcu_preempt_state
, call_rcu
);
849 EXPORT_SYMBOL_GPL(rcu_barrier
);
852 * Initialize preemptible RCU's per-CPU data.
854 static void __cpuinit
rcu_preempt_init_percpu_data(int cpu
)
856 rcu_init_percpu_data(cpu
, &rcu_preempt_state
, 1);
860 * Move preemptible RCU's callbacks from dying CPU to other online CPU.
862 static void rcu_preempt_send_cbs_to_online(void)
864 rcu_send_cbs_to_online(&rcu_preempt_state
);
868 * Initialize preemptible RCU's state structures.
870 static void __init
__rcu_init_preempt(void)
872 rcu_init_one(&rcu_preempt_state
, &rcu_preempt_data
);
876 * Check for a task exiting while in a preemptible-RCU read-side
877 * critical section, clean up if so. No need to issue warnings,
878 * as debug_check_no_locks_held() already does this if lockdep
883 struct task_struct
*t
= current
;
885 if (t
->rcu_read_lock_nesting
== 0)
887 t
->rcu_read_lock_nesting
= 1;
891 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
893 static struct rcu_state
*rcu_state
= &rcu_sched_state
;
896 * Tell them what RCU they are running.
898 static void __init
rcu_bootup_announce(void)
900 printk(KERN_INFO
"Hierarchical RCU implementation.\n");
901 rcu_bootup_announce_oddness();
905 * Return the number of RCU batches processed thus far for debug & stats.
907 long rcu_batches_completed(void)
909 return rcu_batches_completed_sched();
911 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
914 * Force a quiescent state for RCU, which, because there is no preemptible
915 * RCU, becomes the same as rcu-sched.
917 void rcu_force_quiescent_state(void)
919 rcu_sched_force_quiescent_state();
921 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
924 * Because preemptible RCU does not exist, we never have to check for
925 * CPUs being in quiescent states.
927 static void rcu_preempt_note_context_switch(int cpu
)
932 * Because preemptible RCU does not exist, there are never any preempted
935 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
940 #ifdef CONFIG_HOTPLUG_CPU
942 /* Because preemptible RCU does not exist, no quieting of tasks. */
943 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
945 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
948 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
951 * Because preemptible RCU does not exist, we never have to check for
952 * tasks blocked within RCU read-side critical sections.
954 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
959 * Because preemptible RCU does not exist, we never have to check for
960 * tasks blocked within RCU read-side critical sections.
962 static void rcu_print_task_stall(struct rcu_node
*rnp
)
967 * Because preemptible RCU does not exist, there is no need to suppress
968 * its CPU stall warnings.
970 static void rcu_preempt_stall_reset(void)
975 * Because there is no preemptible RCU, there can be no readers blocked,
976 * so there is no need to check for blocked tasks. So check only for
977 * bogus qsmask values.
979 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
981 WARN_ON_ONCE(rnp
->qsmask
);
984 #ifdef CONFIG_HOTPLUG_CPU
987 * Because preemptible RCU does not exist, it never needs to migrate
988 * tasks that were blocked within RCU read-side critical sections, and
989 * such non-existent tasks cannot possibly have been blocking the current
992 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
993 struct rcu_node
*rnp
,
994 struct rcu_data
*rdp
)
1000 * Because preemptible RCU does not exist, it never needs CPU-offline
1003 static void rcu_preempt_offline_cpu(int cpu
)
1007 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1010 * Because preemptible RCU does not exist, it never has any callbacks
1013 static void rcu_preempt_check_callbacks(int cpu
)
1018 * Because preemptible RCU does not exist, it never has any callbacks
1021 static void rcu_preempt_process_callbacks(void)
1026 * Wait for an rcu-preempt grace period, but make it happen quickly.
1027 * But because preemptible RCU does not exist, map to rcu-sched.
1029 void synchronize_rcu_expedited(void)
1031 synchronize_sched_expedited();
1033 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
1035 #ifdef CONFIG_HOTPLUG_CPU
1038 * Because preemptible RCU does not exist, there is never any need to
1039 * report on tasks preempted in RCU read-side critical sections during
1040 * expedited RCU grace periods.
1042 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1047 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1050 * Because preemptible RCU does not exist, it never has any work to do.
1052 static int rcu_preempt_pending(int cpu
)
1058 * Because preemptible RCU does not exist, it never needs any CPU.
1060 static int rcu_preempt_needs_cpu(int cpu
)
1066 * Because preemptible RCU does not exist, rcu_barrier() is just
1067 * another name for rcu_barrier_sched().
1069 void rcu_barrier(void)
1071 rcu_barrier_sched();
1073 EXPORT_SYMBOL_GPL(rcu_barrier
);
1076 * Because preemptible RCU does not exist, there is no per-CPU
1077 * data to initialize.
1079 static void __cpuinit
rcu_preempt_init_percpu_data(int cpu
)
1084 * Because there is no preemptible RCU, there are no callbacks to move.
1086 static void rcu_preempt_send_cbs_to_online(void)
1091 * Because preemptible RCU does not exist, it need not be initialized.
1093 static void __init
__rcu_init_preempt(void)
1097 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1099 #ifdef CONFIG_RCU_BOOST
1101 #include "rtmutex_common.h"
1103 #ifdef CONFIG_RCU_TRACE
1105 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1107 if (list_empty(&rnp
->blkd_tasks
))
1108 rnp
->n_balk_blkd_tasks
++;
1109 else if (rnp
->exp_tasks
== NULL
&& rnp
->gp_tasks
== NULL
)
1110 rnp
->n_balk_exp_gp_tasks
++;
1111 else if (rnp
->gp_tasks
!= NULL
&& rnp
->boost_tasks
!= NULL
)
1112 rnp
->n_balk_boost_tasks
++;
1113 else if (rnp
->gp_tasks
!= NULL
&& rnp
->qsmask
!= 0)
1114 rnp
->n_balk_notblocked
++;
1115 else if (rnp
->gp_tasks
!= NULL
&&
1116 ULONG_CMP_LT(jiffies
, rnp
->boost_time
))
1117 rnp
->n_balk_notyet
++;
1122 #else /* #ifdef CONFIG_RCU_TRACE */
1124 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1128 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1131 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1132 * or ->boost_tasks, advancing the pointer to the next task in the
1133 * ->blkd_tasks list.
1135 * Note that irqs must be enabled: boosting the task can block.
1136 * Returns 1 if there are more tasks needing to be boosted.
1138 static int rcu_boost(struct rcu_node
*rnp
)
1140 unsigned long flags
;
1141 struct rt_mutex mtx
;
1142 struct task_struct
*t
;
1143 struct list_head
*tb
;
1145 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
)
1146 return 0; /* Nothing left to boost. */
1148 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1151 * Recheck under the lock: all tasks in need of boosting
1152 * might exit their RCU read-side critical sections on their own.
1154 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
1155 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1160 * Preferentially boost tasks blocking expedited grace periods.
1161 * This cannot starve the normal grace periods because a second
1162 * expedited grace period must boost all blocked tasks, including
1163 * those blocking the pre-existing normal grace period.
1165 if (rnp
->exp_tasks
!= NULL
) {
1166 tb
= rnp
->exp_tasks
;
1167 rnp
->n_exp_boosts
++;
1169 tb
= rnp
->boost_tasks
;
1170 rnp
->n_normal_boosts
++;
1172 rnp
->n_tasks_boosted
++;
1175 * We boost task t by manufacturing an rt_mutex that appears to
1176 * be held by task t. We leave a pointer to that rt_mutex where
1177 * task t can find it, and task t will release the mutex when it
1178 * exits its outermost RCU read-side critical section. Then
1179 * simply acquiring this artificial rt_mutex will boost task
1180 * t's priority. (Thanks to tglx for suggesting this approach!)
1182 * Note that task t must acquire rnp->lock to remove itself from
1183 * the ->blkd_tasks list, which it will do from exit() if from
1184 * nowhere else. We therefore are guaranteed that task t will
1185 * stay around at least until we drop rnp->lock. Note that
1186 * rnp->lock also resolves races between our priority boosting
1187 * and task t's exiting its outermost RCU read-side critical
1190 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1191 rt_mutex_init_proxy_locked(&mtx
, t
);
1192 t
->rcu_boost_mutex
= &mtx
;
1194 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1195 rt_mutex_lock(&mtx
); /* Side effect: boosts task t's priority. */
1196 rt_mutex_unlock(&mtx
); /* Keep lockdep happy. */
1198 return rnp
->exp_tasks
!= NULL
|| rnp
->boost_tasks
!= NULL
;
1202 * Timer handler to initiate waking up of boost kthreads that
1203 * have yielded the CPU due to excessive numbers of tasks to
1204 * boost. We wake up the per-rcu_node kthread, which in turn
1205 * will wake up the booster kthread.
1207 static void rcu_boost_kthread_timer(unsigned long arg
)
1209 invoke_rcu_node_kthread((struct rcu_node
*)arg
);
1213 * Priority-boosting kthread. One per leaf rcu_node and one for the
1216 static int rcu_boost_kthread(void *arg
)
1218 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1222 trace_rcu_utilization("Start boost kthread@init");
1224 rnp
->boost_kthread_status
= RCU_KTHREAD_WAITING
;
1225 trace_rcu_utilization("End boost kthread@rcu_wait");
1226 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1227 trace_rcu_utilization("Start boost kthread@rcu_wait");
1228 rnp
->boost_kthread_status
= RCU_KTHREAD_RUNNING
;
1229 more2boost
= rcu_boost(rnp
);
1235 trace_rcu_utilization("End boost kthread@rcu_yield");
1236 rcu_yield(rcu_boost_kthread_timer
, (unsigned long)rnp
);
1237 trace_rcu_utilization("Start boost kthread@rcu_yield");
1242 trace_rcu_utilization("End boost kthread@notreached");
1247 * Check to see if it is time to start boosting RCU readers that are
1248 * blocking the current grace period, and, if so, tell the per-rcu_node
1249 * kthread to start boosting them. If there is an expedited grace
1250 * period in progress, it is always time to boost.
1252 * The caller must hold rnp->lock, which this function releases,
1253 * but irqs remain disabled. The ->boost_kthread_task is immortal,
1254 * so we don't need to worry about it going away.
1256 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1258 struct task_struct
*t
;
1260 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1261 rnp
->n_balk_exp_gp_tasks
++;
1262 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1265 if (rnp
->exp_tasks
!= NULL
||
1266 (rnp
->gp_tasks
!= NULL
&&
1267 rnp
->boost_tasks
== NULL
&&
1269 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))) {
1270 if (rnp
->exp_tasks
== NULL
)
1271 rnp
->boost_tasks
= rnp
->gp_tasks
;
1272 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1273 t
= rnp
->boost_kthread_task
;
1277 rcu_initiate_boost_trace(rnp
);
1278 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1283 * Wake up the per-CPU kthread to invoke RCU callbacks.
1285 static void invoke_rcu_callbacks_kthread(void)
1287 unsigned long flags
;
1289 local_irq_save(flags
);
1290 __this_cpu_write(rcu_cpu_has_work
, 1);
1291 if (__this_cpu_read(rcu_cpu_kthread_task
) != NULL
&&
1292 current
!= __this_cpu_read(rcu_cpu_kthread_task
))
1293 wake_up_process(__this_cpu_read(rcu_cpu_kthread_task
));
1294 local_irq_restore(flags
);
1298 * Set the affinity of the boost kthread. The CPU-hotplug locks are
1299 * held, so no one should be messing with the existence of the boost
1302 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
,
1305 struct task_struct
*t
;
1307 t
= rnp
->boost_kthread_task
;
1309 set_cpus_allowed_ptr(rnp
->boost_kthread_task
, cm
);
1312 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1315 * Do priority-boost accounting for the start of a new grace period.
1317 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1319 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1323 * Create an RCU-boost kthread for the specified node if one does not
1324 * already exist. We only create this kthread for preemptible RCU.
1325 * Returns zero if all is well, a negated errno otherwise.
1327 static int __cpuinit
rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1328 struct rcu_node
*rnp
,
1331 unsigned long flags
;
1332 struct sched_param sp
;
1333 struct task_struct
*t
;
1335 if (&rcu_preempt_state
!= rsp
)
1338 if (rnp
->boost_kthread_task
!= NULL
)
1340 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1341 "rcub%d", rnp_index
);
1344 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1345 rnp
->boost_kthread_task
= t
;
1346 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1347 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1348 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1349 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1353 #ifdef CONFIG_HOTPLUG_CPU
1356 * Stop the RCU's per-CPU kthread when its CPU goes offline,.
1358 static void rcu_stop_cpu_kthread(int cpu
)
1360 struct task_struct
*t
;
1362 /* Stop the CPU's kthread. */
1363 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1365 per_cpu(rcu_cpu_kthread_task
, cpu
) = NULL
;
1370 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1372 static void rcu_kthread_do_work(void)
1374 rcu_do_batch(&rcu_sched_state
, &__get_cpu_var(rcu_sched_data
));
1375 rcu_do_batch(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1376 rcu_preempt_do_callbacks();
1380 * Wake up the specified per-rcu_node-structure kthread.
1381 * Because the per-rcu_node kthreads are immortal, we don't need
1382 * to do anything to keep them alive.
1384 static void invoke_rcu_node_kthread(struct rcu_node
*rnp
)
1386 struct task_struct
*t
;
1388 t
= rnp
->node_kthread_task
;
1394 * Set the specified CPU's kthread to run RT or not, as specified by
1395 * the to_rt argument. The CPU-hotplug locks are held, so the task
1396 * is not going away.
1398 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1401 struct sched_param sp
;
1402 struct task_struct
*t
;
1404 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1408 policy
= SCHED_FIFO
;
1409 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1411 policy
= SCHED_NORMAL
;
1412 sp
.sched_priority
= 0;
1414 sched_setscheduler_nocheck(t
, policy
, &sp
);
1418 * Timer handler to initiate the waking up of per-CPU kthreads that
1419 * have yielded the CPU due to excess numbers of RCU callbacks.
1420 * We wake up the per-rcu_node kthread, which in turn will wake up
1421 * the booster kthread.
1423 static void rcu_cpu_kthread_timer(unsigned long arg
)
1425 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, arg
);
1426 struct rcu_node
*rnp
= rdp
->mynode
;
1428 atomic_or(rdp
->grpmask
, &rnp
->wakemask
);
1429 invoke_rcu_node_kthread(rnp
);
1433 * Drop to non-real-time priority and yield, but only after posting a
1434 * timer that will cause us to regain our real-time priority if we
1435 * remain preempted. Either way, we restore our real-time priority
1438 static void rcu_yield(void (*f
)(unsigned long), unsigned long arg
)
1440 struct sched_param sp
;
1441 struct timer_list yield_timer
;
1443 setup_timer_on_stack(&yield_timer
, f
, arg
);
1444 mod_timer(&yield_timer
, jiffies
+ 2);
1445 sp
.sched_priority
= 0;
1446 sched_setscheduler_nocheck(current
, SCHED_NORMAL
, &sp
);
1447 set_user_nice(current
, 19);
1449 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1450 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1451 del_timer(&yield_timer
);
1455 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1456 * This can happen while the corresponding CPU is either coming online
1457 * or going offline. We cannot wait until the CPU is fully online
1458 * before starting the kthread, because the various notifier functions
1459 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1460 * the corresponding CPU is online.
1462 * Return 1 if the kthread needs to stop, 0 otherwise.
1464 * Caller must disable bh. This function can momentarily enable it.
1466 static int rcu_cpu_kthread_should_stop(int cpu
)
1468 while (cpu_is_offline(cpu
) ||
1469 !cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)) ||
1470 smp_processor_id() != cpu
) {
1471 if (kthread_should_stop())
1473 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1474 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = raw_smp_processor_id();
1476 schedule_timeout_uninterruptible(1);
1477 if (!cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)))
1478 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
1481 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1486 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1487 * RCU softirq used in flavors and configurations of RCU that do not
1488 * support RCU priority boosting.
1490 static int rcu_cpu_kthread(void *arg
)
1492 int cpu
= (int)(long)arg
;
1493 unsigned long flags
;
1495 unsigned int *statusp
= &per_cpu(rcu_cpu_kthread_status
, cpu
);
1497 char *workp
= &per_cpu(rcu_cpu_has_work
, cpu
);
1499 trace_rcu_utilization("Start CPU kthread@init");
1501 *statusp
= RCU_KTHREAD_WAITING
;
1502 trace_rcu_utilization("End CPU kthread@rcu_wait");
1503 rcu_wait(*workp
!= 0 || kthread_should_stop());
1504 trace_rcu_utilization("Start CPU kthread@rcu_wait");
1506 if (rcu_cpu_kthread_should_stop(cpu
)) {
1510 *statusp
= RCU_KTHREAD_RUNNING
;
1511 per_cpu(rcu_cpu_kthread_loops
, cpu
)++;
1512 local_irq_save(flags
);
1515 local_irq_restore(flags
);
1517 rcu_kthread_do_work();
1524 *statusp
= RCU_KTHREAD_YIELDING
;
1525 trace_rcu_utilization("End CPU kthread@rcu_yield");
1526 rcu_yield(rcu_cpu_kthread_timer
, (unsigned long)cpu
);
1527 trace_rcu_utilization("Start CPU kthread@rcu_yield");
1531 *statusp
= RCU_KTHREAD_STOPPED
;
1532 trace_rcu_utilization("End CPU kthread@term");
1537 * Spawn a per-CPU kthread, setting up affinity and priority.
1538 * Because the CPU hotplug lock is held, no other CPU will be attempting
1539 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1540 * attempting to access it during boot, but the locking in kthread_bind()
1541 * will enforce sufficient ordering.
1543 * Please note that we cannot simply refuse to wake up the per-CPU
1544 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1545 * which can result in softlockup complaints if the task ends up being
1546 * idle for more than a couple of minutes.
1548 * However, please note also that we cannot bind the per-CPU kthread to its
1549 * CPU until that CPU is fully online. We also cannot wait until the
1550 * CPU is fully online before we create its per-CPU kthread, as this would
1551 * deadlock the system when CPU notifiers tried waiting for grace
1552 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1553 * is online. If its CPU is not yet fully online, then the code in
1554 * rcu_cpu_kthread() will wait until it is fully online, and then do
1557 static int __cpuinit
rcu_spawn_one_cpu_kthread(int cpu
)
1559 struct sched_param sp
;
1560 struct task_struct
*t
;
1562 if (!rcu_scheduler_fully_active
||
1563 per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
)
1565 t
= kthread_create_on_node(rcu_cpu_kthread
,
1571 if (cpu_online(cpu
))
1572 kthread_bind(t
, cpu
);
1573 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1574 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
);
1575 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1576 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1577 per_cpu(rcu_cpu_kthread_task
, cpu
) = t
;
1578 wake_up_process(t
); /* Get to TASK_INTERRUPTIBLE quickly. */
1583 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1584 * kthreads when needed. We ignore requests to wake up kthreads
1585 * for offline CPUs, which is OK because force_quiescent_state()
1586 * takes care of this case.
1588 static int rcu_node_kthread(void *arg
)
1591 unsigned long flags
;
1593 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1594 struct sched_param sp
;
1595 struct task_struct
*t
;
1598 rnp
->node_kthread_status
= RCU_KTHREAD_WAITING
;
1599 rcu_wait(atomic_read(&rnp
->wakemask
) != 0);
1600 rnp
->node_kthread_status
= RCU_KTHREAD_RUNNING
;
1601 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1602 mask
= atomic_xchg(&rnp
->wakemask
, 0);
1603 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1604 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1) {
1605 if ((mask
& 0x1) == 0)
1608 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1609 if (!cpu_online(cpu
) || t
== NULL
) {
1613 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1614 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1615 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1620 rnp
->node_kthread_status
= RCU_KTHREAD_STOPPED
;
1625 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1626 * served by the rcu_node in question. The CPU hotplug lock is still
1627 * held, so the value of rnp->qsmaskinit will be stable.
1629 * We don't include outgoingcpu in the affinity set, use -1 if there is
1630 * no outgoing CPU. If there are no CPUs left in the affinity set,
1631 * this function allows the kthread to execute on any CPU.
1633 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1637 unsigned long mask
= rnp
->qsmaskinit
;
1639 if (rnp
->node_kthread_task
== NULL
)
1641 if (!alloc_cpumask_var(&cm
, GFP_KERNEL
))
1644 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1645 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1646 cpumask_set_cpu(cpu
, cm
);
1647 if (cpumask_weight(cm
) == 0) {
1649 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1650 cpumask_clear_cpu(cpu
, cm
);
1651 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1653 set_cpus_allowed_ptr(rnp
->node_kthread_task
, cm
);
1654 rcu_boost_kthread_setaffinity(rnp
, cm
);
1655 free_cpumask_var(cm
);
1659 * Spawn a per-rcu_node kthread, setting priority and affinity.
1660 * Called during boot before online/offline can happen, or, if
1661 * during runtime, with the main CPU-hotplug locks held. So only
1662 * one of these can be executing at a time.
1664 static int __cpuinit
rcu_spawn_one_node_kthread(struct rcu_state
*rsp
,
1665 struct rcu_node
*rnp
)
1667 unsigned long flags
;
1668 int rnp_index
= rnp
- &rsp
->node
[0];
1669 struct sched_param sp
;
1670 struct task_struct
*t
;
1672 if (!rcu_scheduler_fully_active
||
1673 rnp
->qsmaskinit
== 0)
1675 if (rnp
->node_kthread_task
== NULL
) {
1676 t
= kthread_create(rcu_node_kthread
, (void *)rnp
,
1677 "rcun%d", rnp_index
);
1680 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1681 rnp
->node_kthread_task
= t
;
1682 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1683 sp
.sched_priority
= 99;
1684 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1685 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1687 return rcu_spawn_one_boost_kthread(rsp
, rnp
, rnp_index
);
1691 * Spawn all kthreads -- called as soon as the scheduler is running.
1693 static int __init
rcu_spawn_kthreads(void)
1696 struct rcu_node
*rnp
;
1698 rcu_scheduler_fully_active
= 1;
1699 for_each_possible_cpu(cpu
) {
1700 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1701 if (cpu_online(cpu
))
1702 (void)rcu_spawn_one_cpu_kthread(cpu
);
1704 rnp
= rcu_get_root(rcu_state
);
1705 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1706 if (NUM_RCU_NODES
> 1) {
1707 rcu_for_each_leaf_node(rcu_state
, rnp
)
1708 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1712 early_initcall(rcu_spawn_kthreads
);
1714 static void __cpuinit
rcu_prepare_kthreads(int cpu
)
1716 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
1717 struct rcu_node
*rnp
= rdp
->mynode
;
1719 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1720 if (rcu_scheduler_fully_active
) {
1721 (void)rcu_spawn_one_cpu_kthread(cpu
);
1722 if (rnp
->node_kthread_task
== NULL
)
1723 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1727 #else /* #ifdef CONFIG_RCU_BOOST */
1729 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1731 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1734 static void invoke_rcu_callbacks_kthread(void)
1739 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1743 #ifdef CONFIG_HOTPLUG_CPU
1745 static void rcu_stop_cpu_kthread(int cpu
)
1749 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1751 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1755 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1759 static int __init
rcu_scheduler_really_started(void)
1761 rcu_scheduler_fully_active
= 1;
1764 early_initcall(rcu_scheduler_really_started
);
1766 static void __cpuinit
rcu_prepare_kthreads(int cpu
)
1770 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1774 void synchronize_sched_expedited(void)
1778 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
1780 #else /* #ifndef CONFIG_SMP */
1782 static atomic_t sync_sched_expedited_started
= ATOMIC_INIT(0);
1783 static atomic_t sync_sched_expedited_done
= ATOMIC_INIT(0);
1785 static int synchronize_sched_expedited_cpu_stop(void *data
)
1788 * There must be a full memory barrier on each affected CPU
1789 * between the time that try_stop_cpus() is called and the
1790 * time that it returns.
1792 * In the current initial implementation of cpu_stop, the
1793 * above condition is already met when the control reaches
1794 * this point and the following smp_mb() is not strictly
1795 * necessary. Do smp_mb() anyway for documentation and
1796 * robustness against future implementation changes.
1798 smp_mb(); /* See above comment block. */
1803 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
1804 * approach to force grace period to end quickly. This consumes
1805 * significant time on all CPUs, and is thus not recommended for
1806 * any sort of common-case code.
1808 * Note that it is illegal to call this function while holding any
1809 * lock that is acquired by a CPU-hotplug notifier. Failing to
1810 * observe this restriction will result in deadlock.
1812 * This implementation can be thought of as an application of ticket
1813 * locking to RCU, with sync_sched_expedited_started and
1814 * sync_sched_expedited_done taking on the roles of the halves
1815 * of the ticket-lock word. Each task atomically increments
1816 * sync_sched_expedited_started upon entry, snapshotting the old value,
1817 * then attempts to stop all the CPUs. If this succeeds, then each
1818 * CPU will have executed a context switch, resulting in an RCU-sched
1819 * grace period. We are then done, so we use atomic_cmpxchg() to
1820 * update sync_sched_expedited_done to match our snapshot -- but
1821 * only if someone else has not already advanced past our snapshot.
1823 * On the other hand, if try_stop_cpus() fails, we check the value
1824 * of sync_sched_expedited_done. If it has advanced past our
1825 * initial snapshot, then someone else must have forced a grace period
1826 * some time after we took our snapshot. In this case, our work is
1827 * done for us, and we can simply return. Otherwise, we try again,
1828 * but keep our initial snapshot for purposes of checking for someone
1829 * doing our work for us.
1831 * If we fail too many times in a row, we fall back to synchronize_sched().
1833 void synchronize_sched_expedited(void)
1835 int firstsnap
, s
, snap
, trycount
= 0;
1837 /* Note that atomic_inc_return() implies full memory barrier. */
1838 firstsnap
= snap
= atomic_inc_return(&sync_sched_expedited_started
);
1842 * Each pass through the following loop attempts to force a
1843 * context switch on each CPU.
1845 while (try_stop_cpus(cpu_online_mask
,
1846 synchronize_sched_expedited_cpu_stop
,
1850 /* No joy, try again later. Or just synchronize_sched(). */
1851 if (trycount
++ < 10)
1852 udelay(trycount
* num_online_cpus());
1854 synchronize_sched();
1858 /* Check to see if someone else did our work for us. */
1859 s
= atomic_read(&sync_sched_expedited_done
);
1860 if (UINT_CMP_GE((unsigned)s
, (unsigned)firstsnap
)) {
1861 smp_mb(); /* ensure test happens before caller kfree */
1866 * Refetching sync_sched_expedited_started allows later
1867 * callers to piggyback on our grace period. We subtract
1868 * 1 to get the same token that the last incrementer got.
1869 * We retry after they started, so our grace period works
1870 * for them, and they started after our first try, so their
1871 * grace period works for us.
1874 snap
= atomic_read(&sync_sched_expedited_started
) - 1;
1875 smp_mb(); /* ensure read is before try_stop_cpus(). */
1879 * Everyone up to our most recent fetch is covered by our grace
1880 * period. Update the counter, but only if our work is still
1881 * relevant -- which it won't be if someone who started later
1882 * than we did beat us to the punch.
1885 s
= atomic_read(&sync_sched_expedited_done
);
1886 if (UINT_CMP_GE((unsigned)s
, (unsigned)snap
)) {
1887 smp_mb(); /* ensure test happens before caller kfree */
1890 } while (atomic_cmpxchg(&sync_sched_expedited_done
, s
, snap
) != s
);
1894 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
1896 #endif /* #else #ifndef CONFIG_SMP */
1898 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1901 * Check to see if any future RCU-related work will need to be done
1902 * by the current CPU, even if none need be done immediately, returning
1903 * 1 if so. This function is part of the RCU implementation; it is -not-
1904 * an exported member of the RCU API.
1906 * Because we have preemptible RCU, just check whether this CPU needs
1907 * any flavor of RCU. Do not chew up lots of CPU cycles with preemption
1908 * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
1910 int rcu_needs_cpu(int cpu
)
1912 return rcu_needs_cpu_quick_check(cpu
);
1916 * Check to see if we need to continue a callback-flush operations to
1917 * allow the last CPU to enter dyntick-idle mode. But fast dyntick-idle
1918 * entry is not configured, so we never do need to.
1920 static void rcu_needs_cpu_flush(void)
1924 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1926 #define RCU_NEEDS_CPU_FLUSHES 5
1927 static DEFINE_PER_CPU(int, rcu_dyntick_drain
);
1928 static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff
);
1931 * Check to see if any future RCU-related work will need to be done
1932 * by the current CPU, even if none need be done immediately, returning
1933 * 1 if so. This function is part of the RCU implementation; it is -not-
1934 * an exported member of the RCU API.
1936 * Because we are not supporting preemptible RCU, attempt to accelerate
1937 * any current grace periods so that RCU no longer needs this CPU, but
1938 * only if all other CPUs are already in dynticks-idle mode. This will
1939 * allow the CPU cores to be powered down immediately, as opposed to after
1940 * waiting many milliseconds for grace periods to elapse.
1942 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1943 * disabled, we do one pass of force_quiescent_state(), then do a
1944 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
1945 * later. The per-cpu rcu_dyntick_drain variable controls the sequencing.
1947 int rcu_needs_cpu(int cpu
)
1953 /* Check for being in the holdoff period. */
1954 if (per_cpu(rcu_dyntick_holdoff
, cpu
) == jiffies
)
1955 return rcu_needs_cpu_quick_check(cpu
);
1957 /* Don't bother unless we are the last non-dyntick-idle CPU. */
1958 for_each_online_cpu(thatcpu
) {
1961 snap
= atomic_add_return(0, &per_cpu(rcu_dynticks
,
1963 smp_mb(); /* Order sampling of snap with end of grace period. */
1964 if ((snap
& 0x1) != 0) {
1965 per_cpu(rcu_dyntick_drain
, cpu
) = 0;
1966 per_cpu(rcu_dyntick_holdoff
, cpu
) = jiffies
- 1;
1967 return rcu_needs_cpu_quick_check(cpu
);
1971 /* Check and update the rcu_dyntick_drain sequencing. */
1972 if (per_cpu(rcu_dyntick_drain
, cpu
) <= 0) {
1973 /* First time through, initialize the counter. */
1974 per_cpu(rcu_dyntick_drain
, cpu
) = RCU_NEEDS_CPU_FLUSHES
;
1975 } else if (--per_cpu(rcu_dyntick_drain
, cpu
) <= 0) {
1976 /* We have hit the limit, so time to give up. */
1977 per_cpu(rcu_dyntick_holdoff
, cpu
) = jiffies
;
1978 return rcu_needs_cpu_quick_check(cpu
);
1981 /* Do one step pushing remaining RCU callbacks through. */
1982 if (per_cpu(rcu_sched_data
, cpu
).nxtlist
) {
1984 force_quiescent_state(&rcu_sched_state
, 0);
1985 c
= c
|| per_cpu(rcu_sched_data
, cpu
).nxtlist
;
1987 if (per_cpu(rcu_bh_data
, cpu
).nxtlist
) {
1989 force_quiescent_state(&rcu_bh_state
, 0);
1990 c
= c
|| per_cpu(rcu_bh_data
, cpu
).nxtlist
;
1993 /* If RCU callbacks are still pending, RCU still needs this CPU. */
2000 * Check to see if we need to continue a callback-flush operations to
2001 * allow the last CPU to enter dyntick-idle mode.
2003 static void rcu_needs_cpu_flush(void)
2005 int cpu
= smp_processor_id();
2006 unsigned long flags
;
2008 if (per_cpu(rcu_dyntick_drain
, cpu
) <= 0)
2010 local_irq_save(flags
);
2011 (void)rcu_needs_cpu(cpu
);
2012 local_irq_restore(flags
);
2015 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */