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>
30 #define RCU_KTHREAD_PRIO 1
32 #ifdef CONFIG_RCU_BOOST
33 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
35 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
39 * Check the RCU kernel configuration parameters and print informative
40 * messages about anything out of the ordinary. If you like #ifdef, you
41 * will love this function.
43 static void __init
rcu_bootup_announce_oddness(void)
45 #ifdef CONFIG_RCU_TRACE
46 printk(KERN_INFO
"\tRCU debugfs-based tracing is enabled.\n");
48 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
49 printk(KERN_INFO
"\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
52 #ifdef CONFIG_RCU_FANOUT_EXACT
53 printk(KERN_INFO
"\tHierarchical RCU autobalancing is disabled.\n");
55 #ifdef CONFIG_RCU_FAST_NO_HZ
57 "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
59 #ifdef CONFIG_PROVE_RCU
60 printk(KERN_INFO
"\tRCU lockdep checking is enabled.\n");
62 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
63 printk(KERN_INFO
"\tRCU torture testing starts during boot.\n");
65 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
66 printk(KERN_INFO
"\tVerbose stalled-CPUs detection is disabled.\n");
68 #if NUM_RCU_LVL_4 != 0
69 printk(KERN_INFO
"\tExperimental four-level hierarchy is enabled.\n");
73 #ifdef CONFIG_TREE_PREEMPT_RCU
75 struct rcu_state rcu_preempt_state
= RCU_STATE_INITIALIZER(rcu_preempt
);
76 DEFINE_PER_CPU(struct rcu_data
, rcu_preempt_data
);
77 static struct rcu_state
*rcu_state
= &rcu_preempt_state
;
79 static void rcu_read_unlock_special(struct task_struct
*t
);
80 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
);
83 * Tell them what RCU they are running.
85 static void __init
rcu_bootup_announce(void)
87 printk(KERN_INFO
"Preemptible hierarchical RCU implementation.\n");
88 rcu_bootup_announce_oddness();
92 * Return the number of RCU-preempt batches processed thus far
93 * for debug and statistics.
95 long rcu_batches_completed_preempt(void)
97 return rcu_preempt_state
.completed
;
99 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt
);
102 * Return the number of RCU batches processed thus far for debug & stats.
104 long rcu_batches_completed(void)
106 return rcu_batches_completed_preempt();
108 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
111 * Force a quiescent state for preemptible RCU.
113 void rcu_force_quiescent_state(void)
115 force_quiescent_state(&rcu_preempt_state
, 0);
117 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
120 * Record a preemptible-RCU quiescent state for the specified CPU. Note
121 * that this just means that the task currently running on the CPU is
122 * not in a quiescent state. There might be any number of tasks blocked
123 * while in an RCU read-side critical section.
125 * Unlike the other rcu_*_qs() functions, callers to this function
126 * must disable irqs in order to protect the assignment to
127 * ->rcu_read_unlock_special.
129 static void rcu_preempt_qs(int cpu
)
131 struct rcu_data
*rdp
= &per_cpu(rcu_preempt_data
, cpu
);
133 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
135 if (rdp
->passed_quiesce
== 0)
136 trace_rcu_grace_period("rcu_preempt", rdp
->gpnum
, "cpuqs");
137 rdp
->passed_quiesce
= 1;
138 current
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_NEED_QS
;
142 * We have entered the scheduler, and the current task might soon be
143 * context-switched away from. If this task is in an RCU read-side
144 * critical section, we will no longer be able to rely on the CPU to
145 * record that fact, so we enqueue the task on the blkd_tasks list.
146 * The task will dequeue itself when it exits the outermost enclosing
147 * RCU read-side critical section. Therefore, the current grace period
148 * cannot be permitted to complete until the blkd_tasks list entries
149 * predating the current grace period drain, in other words, until
150 * rnp->gp_tasks becomes NULL.
152 * Caller must disable preemption.
154 static void rcu_preempt_note_context_switch(int cpu
)
156 struct task_struct
*t
= current
;
158 struct rcu_data
*rdp
;
159 struct rcu_node
*rnp
;
161 if (t
->rcu_read_lock_nesting
> 0 &&
162 (t
->rcu_read_unlock_special
& RCU_READ_UNLOCK_BLOCKED
) == 0) {
164 /* Possibly blocking in an RCU read-side critical section. */
165 rdp
= per_cpu_ptr(rcu_preempt_state
.rda
, cpu
);
167 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
168 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_BLOCKED
;
169 t
->rcu_blocked_node
= rnp
;
172 * If this CPU has already checked in, then this task
173 * will hold up the next grace period rather than the
174 * current grace period. Queue the task accordingly.
175 * If the task is queued for the current grace period
176 * (i.e., this CPU has not yet passed through a quiescent
177 * state for the current grace period), then as long
178 * as that task remains queued, the current grace period
179 * cannot end. Note that there is some uncertainty as
180 * to exactly when the current grace period started.
181 * We take a conservative approach, which can result
182 * in unnecessarily waiting on tasks that started very
183 * slightly after the current grace period began. C'est
186 * But first, note that the current CPU must still be
189 WARN_ON_ONCE((rdp
->grpmask
& rnp
->qsmaskinit
) == 0);
190 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
191 if ((rnp
->qsmask
& rdp
->grpmask
) && rnp
->gp_tasks
!= NULL
) {
192 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
->prev
);
193 rnp
->gp_tasks
= &t
->rcu_node_entry
;
194 #ifdef CONFIG_RCU_BOOST
195 if (rnp
->boost_tasks
!= NULL
)
196 rnp
->boost_tasks
= rnp
->gp_tasks
;
197 #endif /* #ifdef CONFIG_RCU_BOOST */
199 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
200 if (rnp
->qsmask
& rdp
->grpmask
)
201 rnp
->gp_tasks
= &t
->rcu_node_entry
;
203 trace_rcu_preempt_task(rdp
->rsp
->name
,
205 (rnp
->qsmask
& rdp
->grpmask
)
208 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
209 } else if (t
->rcu_read_lock_nesting
< 0 &&
210 t
->rcu_read_unlock_special
) {
213 * Complete exit from RCU read-side critical section on
214 * behalf of preempted instance of __rcu_read_unlock().
216 rcu_read_unlock_special(t
);
220 * Either we were not in an RCU read-side critical section to
221 * begin with, or we have now recorded that critical section
222 * globally. Either way, we can now note a quiescent state
223 * for this CPU. Again, if we were in an RCU read-side critical
224 * section, and if that critical section was blocking the current
225 * grace period, then the fact that the task has been enqueued
226 * means that we continue to block the current grace period.
228 local_irq_save(flags
);
230 local_irq_restore(flags
);
234 * Tree-preemptible RCU implementation for rcu_read_lock().
235 * Just increment ->rcu_read_lock_nesting, shared state will be updated
238 void __rcu_read_lock(void)
240 current
->rcu_read_lock_nesting
++;
241 barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
243 EXPORT_SYMBOL_GPL(__rcu_read_lock
);
246 * Check for preempted RCU readers blocking the current grace period
247 * for the specified rcu_node structure. If the caller needs a reliable
248 * answer, it must hold the rcu_node's ->lock.
250 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
252 return rnp
->gp_tasks
!= NULL
;
256 * Record a quiescent state for all tasks that were previously queued
257 * on the specified rcu_node structure and that were blocking the current
258 * RCU grace period. The caller must hold the specified rnp->lock with
259 * irqs disabled, and this lock is released upon return, but irqs remain
262 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
263 __releases(rnp
->lock
)
266 struct rcu_node
*rnp_p
;
268 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
269 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
270 return; /* Still need more quiescent states! */
276 * Either there is only one rcu_node in the tree,
277 * or tasks were kicked up to root rcu_node due to
278 * CPUs going offline.
280 rcu_report_qs_rsp(&rcu_preempt_state
, flags
);
284 /* Report up the rest of the hierarchy. */
286 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
287 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
288 rcu_report_qs_rnp(mask
, &rcu_preempt_state
, rnp_p
, flags
);
292 * Advance a ->blkd_tasks-list pointer to the next entry, instead
293 * returning NULL if at the end of the list.
295 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
296 struct rcu_node
*rnp
)
298 struct list_head
*np
;
300 np
= t
->rcu_node_entry
.next
;
301 if (np
== &rnp
->blkd_tasks
)
307 * Handle special cases during rcu_read_unlock(), such as needing to
308 * notify RCU core processing or task having blocked during the RCU
309 * read-side critical section.
311 static noinline
void rcu_read_unlock_special(struct task_struct
*t
)
316 struct list_head
*np
;
317 #ifdef CONFIG_RCU_BOOST
318 struct rt_mutex
*rbmp
= NULL
;
319 #endif /* #ifdef CONFIG_RCU_BOOST */
320 struct rcu_node
*rnp
;
323 /* NMI handlers cannot block and cannot safely manipulate state. */
327 local_irq_save(flags
);
330 * If RCU core is waiting for this CPU to exit critical section,
331 * let it know that we have done so.
333 special
= t
->rcu_read_unlock_special
;
334 if (special
& RCU_READ_UNLOCK_NEED_QS
) {
335 rcu_preempt_qs(smp_processor_id());
338 /* Hardware IRQ handlers cannot block. */
339 if (in_irq() || in_serving_softirq()) {
340 local_irq_restore(flags
);
344 /* Clean up if blocked during RCU read-side critical section. */
345 if (special
& RCU_READ_UNLOCK_BLOCKED
) {
346 t
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_BLOCKED
;
349 * Remove this task from the list it blocked on. The
350 * task can migrate while we acquire the lock, but at
351 * most one time. So at most two passes through loop.
354 rnp
= t
->rcu_blocked_node
;
355 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
356 if (rnp
== t
->rcu_blocked_node
)
358 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
360 empty
= !rcu_preempt_blocked_readers_cgp(rnp
);
361 empty_exp
= !rcu_preempted_readers_exp(rnp
);
362 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
363 np
= rcu_next_node_entry(t
, rnp
);
364 list_del_init(&t
->rcu_node_entry
);
365 t
->rcu_blocked_node
= NULL
;
366 trace_rcu_unlock_preempted_task("rcu_preempt",
368 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
370 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
372 #ifdef CONFIG_RCU_BOOST
373 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
374 rnp
->boost_tasks
= np
;
375 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
376 if (t
->rcu_boost_mutex
) {
377 rbmp
= t
->rcu_boost_mutex
;
378 t
->rcu_boost_mutex
= NULL
;
380 #endif /* #ifdef CONFIG_RCU_BOOST */
383 * If this was the last task on the current list, and if
384 * we aren't waiting on any CPUs, report the quiescent state.
385 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
387 if (!empty
&& !rcu_preempt_blocked_readers_cgp(rnp
)) {
388 trace_rcu_quiescent_state_report("preempt_rcu",
395 rcu_report_unblock_qs_rnp(rnp
, flags
);
397 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
399 #ifdef CONFIG_RCU_BOOST
400 /* Unboost if we were boosted. */
402 rt_mutex_unlock(rbmp
);
403 #endif /* #ifdef CONFIG_RCU_BOOST */
406 * If this was the last task on the expedited lists,
407 * then we need to report up the rcu_node hierarchy.
409 if (!empty_exp
&& !rcu_preempted_readers_exp(rnp
))
410 rcu_report_exp_rnp(&rcu_preempt_state
, rnp
);
412 local_irq_restore(flags
);
417 * Tree-preemptible RCU implementation for rcu_read_unlock().
418 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
419 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
420 * invoke rcu_read_unlock_special() to clean up after a context switch
421 * in an RCU read-side critical section and other special cases.
423 void __rcu_read_unlock(void)
425 struct task_struct
*t
= current
;
427 if (t
->rcu_read_lock_nesting
!= 1)
428 --t
->rcu_read_lock_nesting
;
430 barrier(); /* critical section before exit code. */
431 t
->rcu_read_lock_nesting
= INT_MIN
;
432 barrier(); /* assign before ->rcu_read_unlock_special load */
433 if (unlikely(ACCESS_ONCE(t
->rcu_read_unlock_special
)))
434 rcu_read_unlock_special(t
);
435 barrier(); /* ->rcu_read_unlock_special load before assign */
436 t
->rcu_read_lock_nesting
= 0;
438 #ifdef CONFIG_PROVE_LOCKING
440 int rrln
= ACCESS_ONCE(t
->rcu_read_lock_nesting
);
442 WARN_ON_ONCE(rrln
< 0 && rrln
> INT_MIN
/ 2);
444 #endif /* #ifdef CONFIG_PROVE_LOCKING */
446 EXPORT_SYMBOL_GPL(__rcu_read_unlock
);
448 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
451 * Dump detailed information for all tasks blocking the current RCU
452 * grace period on the specified rcu_node structure.
454 static void rcu_print_detail_task_stall_rnp(struct rcu_node
*rnp
)
457 struct task_struct
*t
;
459 if (!rcu_preempt_blocked_readers_cgp(rnp
))
461 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
462 t
= list_entry(rnp
->gp_tasks
,
463 struct task_struct
, rcu_node_entry
);
464 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
466 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
470 * Dump detailed information for all tasks blocking the current RCU
473 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
475 struct rcu_node
*rnp
= rcu_get_root(rsp
);
477 rcu_print_detail_task_stall_rnp(rnp
);
478 rcu_for_each_leaf_node(rsp
, rnp
)
479 rcu_print_detail_task_stall_rnp(rnp
);
482 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
484 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
488 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
491 * Scan the current list of tasks blocked within RCU read-side critical
492 * sections, printing out the tid of each.
494 static int rcu_print_task_stall(struct rcu_node
*rnp
)
496 struct task_struct
*t
;
499 if (!rcu_preempt_blocked_readers_cgp(rnp
))
501 t
= list_entry(rnp
->gp_tasks
,
502 struct task_struct
, rcu_node_entry
);
503 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
) {
504 printk(" P%d", t
->pid
);
511 * Suppress preemptible RCU's CPU stall warnings by pushing the
512 * time of the next stall-warning message comfortably far into the
515 static void rcu_preempt_stall_reset(void)
517 rcu_preempt_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
521 * Check that the list of blocked tasks for the newly completed grace
522 * period is in fact empty. It is a serious bug to complete a grace
523 * period that still has RCU readers blocked! This function must be
524 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
525 * must be held by the caller.
527 * Also, if there are blocked tasks on the list, they automatically
528 * block the newly created grace period, so set up ->gp_tasks accordingly.
530 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
532 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
533 if (!list_empty(&rnp
->blkd_tasks
))
534 rnp
->gp_tasks
= rnp
->blkd_tasks
.next
;
535 WARN_ON_ONCE(rnp
->qsmask
);
538 #ifdef CONFIG_HOTPLUG_CPU
541 * Handle tasklist migration for case in which all CPUs covered by the
542 * specified rcu_node have gone offline. Move them up to the root
543 * rcu_node. The reason for not just moving them to the immediate
544 * parent is to remove the need for rcu_read_unlock_special() to
545 * make more than two attempts to acquire the target rcu_node's lock.
546 * Returns true if there were tasks blocking the current RCU grace
549 * Returns 1 if there was previously a task blocking the current grace
550 * period on the specified rcu_node structure.
552 * The caller must hold rnp->lock with irqs disabled.
554 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
555 struct rcu_node
*rnp
,
556 struct rcu_data
*rdp
)
558 struct list_head
*lp
;
559 struct list_head
*lp_root
;
561 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
562 struct task_struct
*t
;
564 if (rnp
== rnp_root
) {
565 WARN_ONCE(1, "Last CPU thought to be offlined?");
566 return 0; /* Shouldn't happen: at least one CPU online. */
569 /* If we are on an internal node, complain bitterly. */
570 WARN_ON_ONCE(rnp
!= rdp
->mynode
);
573 * Move tasks up to root rcu_node. Don't try to get fancy for
574 * this corner-case operation -- just put this node's tasks
575 * at the head of the root node's list, and update the root node's
576 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
577 * if non-NULL. This might result in waiting for more tasks than
578 * absolutely necessary, but this is a good performance/complexity
581 if (rcu_preempt_blocked_readers_cgp(rnp
))
582 retval
|= RCU_OFL_TASKS_NORM_GP
;
583 if (rcu_preempted_readers_exp(rnp
))
584 retval
|= RCU_OFL_TASKS_EXP_GP
;
585 lp
= &rnp
->blkd_tasks
;
586 lp_root
= &rnp_root
->blkd_tasks
;
587 while (!list_empty(lp
)) {
588 t
= list_entry(lp
->next
, typeof(*t
), rcu_node_entry
);
589 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
590 list_del(&t
->rcu_node_entry
);
591 t
->rcu_blocked_node
= rnp_root
;
592 list_add(&t
->rcu_node_entry
, lp_root
);
593 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
594 rnp_root
->gp_tasks
= rnp
->gp_tasks
;
595 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
596 rnp_root
->exp_tasks
= rnp
->exp_tasks
;
597 #ifdef CONFIG_RCU_BOOST
598 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
599 rnp_root
->boost_tasks
= rnp
->boost_tasks
;
600 #endif /* #ifdef CONFIG_RCU_BOOST */
601 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
604 #ifdef CONFIG_RCU_BOOST
605 /* In case root is being boosted and leaf is not. */
606 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
607 if (rnp_root
->boost_tasks
!= NULL
&&
608 rnp_root
->boost_tasks
!= rnp_root
->gp_tasks
)
609 rnp_root
->boost_tasks
= rnp_root
->gp_tasks
;
610 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
611 #endif /* #ifdef CONFIG_RCU_BOOST */
613 rnp
->gp_tasks
= NULL
;
614 rnp
->exp_tasks
= NULL
;
619 * Do CPU-offline processing for preemptible RCU.
621 static void rcu_preempt_offline_cpu(int cpu
)
623 __rcu_offline_cpu(cpu
, &rcu_preempt_state
);
626 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
629 * Check for a quiescent state from the current CPU. When a task blocks,
630 * the task is recorded in the corresponding CPU's rcu_node structure,
631 * which is checked elsewhere.
633 * Caller must disable hard irqs.
635 static void rcu_preempt_check_callbacks(int cpu
)
637 struct task_struct
*t
= current
;
639 if (t
->rcu_read_lock_nesting
== 0) {
643 if (t
->rcu_read_lock_nesting
> 0 &&
644 per_cpu(rcu_preempt_data
, cpu
).qs_pending
)
645 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_NEED_QS
;
649 * Process callbacks for preemptible RCU.
651 static void rcu_preempt_process_callbacks(void)
653 __rcu_process_callbacks(&rcu_preempt_state
,
654 &__get_cpu_var(rcu_preempt_data
));
657 #ifdef CONFIG_RCU_BOOST
659 static void rcu_preempt_do_callbacks(void)
661 rcu_do_batch(&rcu_preempt_state
, &__get_cpu_var(rcu_preempt_data
));
664 #endif /* #ifdef CONFIG_RCU_BOOST */
667 * Queue a preemptible-RCU callback for invocation after a grace period.
669 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
671 __call_rcu(head
, func
, &rcu_preempt_state
);
673 EXPORT_SYMBOL_GPL(call_rcu
);
676 * synchronize_rcu - wait until a grace period has elapsed.
678 * Control will return to the caller some time after a full grace
679 * period has elapsed, in other words after all currently executing RCU
680 * read-side critical sections have completed. Note, however, that
681 * upon return from synchronize_rcu(), the caller might well be executing
682 * concurrently with new RCU read-side critical sections that began while
683 * synchronize_rcu() was waiting. RCU read-side critical sections are
684 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
686 void synchronize_rcu(void)
688 if (!rcu_scheduler_active
)
690 wait_rcu_gp(call_rcu
);
692 EXPORT_SYMBOL_GPL(synchronize_rcu
);
694 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq
);
695 static long sync_rcu_preempt_exp_count
;
696 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex
);
699 * Return non-zero if there are any tasks in RCU read-side critical
700 * sections blocking the current preemptible-RCU expedited grace period.
701 * If there is no preemptible-RCU expedited grace period currently in
702 * progress, returns zero unconditionally.
704 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
)
706 return rnp
->exp_tasks
!= NULL
;
710 * return non-zero if there is no RCU expedited grace period in progress
711 * for the specified rcu_node structure, in other words, if all CPUs and
712 * tasks covered by the specified rcu_node structure have done their bit
713 * for the current expedited grace period. Works only for preemptible
714 * RCU -- other RCU implementation use other means.
716 * Caller must hold sync_rcu_preempt_exp_mutex.
718 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
720 return !rcu_preempted_readers_exp(rnp
) &&
721 ACCESS_ONCE(rnp
->expmask
) == 0;
725 * Report the exit from RCU read-side critical section for the last task
726 * that queued itself during or before the current expedited preemptible-RCU
727 * grace period. This event is reported either to the rcu_node structure on
728 * which the task was queued or to one of that rcu_node structure's ancestors,
729 * recursively up the tree. (Calm down, calm down, we do the recursion
732 * Caller must hold sync_rcu_preempt_exp_mutex.
734 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
739 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
741 if (!sync_rcu_preempt_exp_done(rnp
)) {
742 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
745 if (rnp
->parent
== NULL
) {
746 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
747 wake_up(&sync_rcu_preempt_exp_wq
);
751 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
753 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
754 rnp
->expmask
&= ~mask
;
759 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
760 * grace period for the specified rcu_node structure. If there are no such
761 * tasks, report it up the rcu_node hierarchy.
763 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
766 sync_rcu_preempt_exp_init(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
771 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
772 if (list_empty(&rnp
->blkd_tasks
))
773 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
775 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
776 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
780 rcu_report_exp_rnp(rsp
, rnp
);
784 * Wait for an rcu-preempt grace period, but expedite it. The basic idea
785 * is to invoke synchronize_sched_expedited() to push all the tasks to
786 * the ->blkd_tasks lists and wait for this list to drain.
788 void synchronize_rcu_expedited(void)
791 struct rcu_node
*rnp
;
792 struct rcu_state
*rsp
= &rcu_preempt_state
;
796 smp_mb(); /* Caller's modifications seen first by other CPUs. */
797 snap
= ACCESS_ONCE(sync_rcu_preempt_exp_count
) + 1;
798 smp_mb(); /* Above access cannot bleed into critical section. */
801 * Acquire lock, falling back to synchronize_rcu() if too many
802 * lock-acquisition failures. Of course, if someone does the
803 * expedited grace period for us, just leave.
805 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex
)) {
807 udelay(trycount
* num_online_cpus());
812 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count
) - snap
) > 0)
813 goto mb_ret
; /* Others did our work for us. */
815 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count
) - snap
) > 0)
816 goto unlock_mb_ret
; /* Others did our work for us. */
818 /* force all RCU readers onto ->blkd_tasks lists. */
819 synchronize_sched_expedited();
821 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
823 /* Initialize ->expmask for all non-leaf rcu_node structures. */
824 rcu_for_each_nonleaf_node_breadth_first(rsp
, rnp
) {
825 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
826 rnp
->expmask
= rnp
->qsmaskinit
;
827 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
830 /* Snapshot current state of ->blkd_tasks lists. */
831 rcu_for_each_leaf_node(rsp
, rnp
)
832 sync_rcu_preempt_exp_init(rsp
, rnp
);
833 if (NUM_RCU_NODES
> 1)
834 sync_rcu_preempt_exp_init(rsp
, rcu_get_root(rsp
));
836 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
838 /* Wait for snapshotted ->blkd_tasks lists to drain. */
839 rnp
= rcu_get_root(rsp
);
840 wait_event(sync_rcu_preempt_exp_wq
,
841 sync_rcu_preempt_exp_done(rnp
));
843 /* Clean up and exit. */
844 smp_mb(); /* ensure expedited GP seen before counter increment. */
845 ACCESS_ONCE(sync_rcu_preempt_exp_count
)++;
847 mutex_unlock(&sync_rcu_preempt_exp_mutex
);
849 smp_mb(); /* ensure subsequent action seen after grace period. */
851 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
854 * Check to see if there is any immediate preemptible-RCU-related work
857 static int rcu_preempt_pending(int cpu
)
859 return __rcu_pending(&rcu_preempt_state
,
860 &per_cpu(rcu_preempt_data
, cpu
));
864 * Does preemptible RCU need the CPU to stay out of dynticks mode?
866 static int rcu_preempt_needs_cpu(int cpu
)
868 return !!per_cpu(rcu_preempt_data
, cpu
).nxtlist
;
872 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
874 void rcu_barrier(void)
876 _rcu_barrier(&rcu_preempt_state
, call_rcu
);
878 EXPORT_SYMBOL_GPL(rcu_barrier
);
881 * Initialize preemptible RCU's per-CPU data.
883 static void __cpuinit
rcu_preempt_init_percpu_data(int cpu
)
885 rcu_init_percpu_data(cpu
, &rcu_preempt_state
, 1);
889 * Move preemptible RCU's callbacks from dying CPU to other online CPU.
891 static void rcu_preempt_send_cbs_to_online(void)
893 rcu_send_cbs_to_online(&rcu_preempt_state
);
897 * Initialize preemptible RCU's state structures.
899 static void __init
__rcu_init_preempt(void)
901 rcu_init_one(&rcu_preempt_state
, &rcu_preempt_data
);
905 * Check for a task exiting while in a preemptible-RCU read-side
906 * critical section, clean up if so. No need to issue warnings,
907 * as debug_check_no_locks_held() already does this if lockdep
912 struct task_struct
*t
= current
;
914 if (t
->rcu_read_lock_nesting
== 0)
916 t
->rcu_read_lock_nesting
= 1;
920 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
922 static struct rcu_state
*rcu_state
= &rcu_sched_state
;
925 * Tell them what RCU they are running.
927 static void __init
rcu_bootup_announce(void)
929 printk(KERN_INFO
"Hierarchical RCU implementation.\n");
930 rcu_bootup_announce_oddness();
934 * Return the number of RCU batches processed thus far for debug & stats.
936 long rcu_batches_completed(void)
938 return rcu_batches_completed_sched();
940 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
943 * Force a quiescent state for RCU, which, because there is no preemptible
944 * RCU, becomes the same as rcu-sched.
946 void rcu_force_quiescent_state(void)
948 rcu_sched_force_quiescent_state();
950 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
953 * Because preemptible RCU does not exist, we never have to check for
954 * CPUs being in quiescent states.
956 static void rcu_preempt_note_context_switch(int cpu
)
961 * Because preemptible RCU does not exist, there are never any preempted
964 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
969 #ifdef CONFIG_HOTPLUG_CPU
971 /* Because preemptible RCU does not exist, no quieting of tasks. */
972 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
974 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
977 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
980 * Because preemptible RCU does not exist, we never have to check for
981 * tasks blocked within RCU read-side critical sections.
983 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
988 * Because preemptible RCU does not exist, we never have to check for
989 * tasks blocked within RCU read-side critical sections.
991 static int rcu_print_task_stall(struct rcu_node
*rnp
)
997 * Because preemptible RCU does not exist, there is no need to suppress
998 * its CPU stall warnings.
1000 static void rcu_preempt_stall_reset(void)
1005 * Because there is no preemptible RCU, there can be no readers blocked,
1006 * so there is no need to check for blocked tasks. So check only for
1007 * bogus qsmask values.
1009 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
1011 WARN_ON_ONCE(rnp
->qsmask
);
1014 #ifdef CONFIG_HOTPLUG_CPU
1017 * Because preemptible RCU does not exist, it never needs to migrate
1018 * tasks that were blocked within RCU read-side critical sections, and
1019 * such non-existent tasks cannot possibly have been blocking the current
1022 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
1023 struct rcu_node
*rnp
,
1024 struct rcu_data
*rdp
)
1030 * Because preemptible RCU does not exist, it never needs CPU-offline
1033 static void rcu_preempt_offline_cpu(int cpu
)
1037 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1040 * Because preemptible RCU does not exist, it never has any callbacks
1043 static void rcu_preempt_check_callbacks(int cpu
)
1048 * Because preemptible RCU does not exist, it never has any callbacks
1051 static void rcu_preempt_process_callbacks(void)
1056 * Wait for an rcu-preempt grace period, but make it happen quickly.
1057 * But because preemptible RCU does not exist, map to rcu-sched.
1059 void synchronize_rcu_expedited(void)
1061 synchronize_sched_expedited();
1063 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
1065 #ifdef CONFIG_HOTPLUG_CPU
1068 * Because preemptible RCU does not exist, there is never any need to
1069 * report on tasks preempted in RCU read-side critical sections during
1070 * expedited RCU grace periods.
1072 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1077 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1080 * Because preemptible RCU does not exist, it never has any work to do.
1082 static int rcu_preempt_pending(int cpu
)
1088 * Because preemptible RCU does not exist, it never needs any CPU.
1090 static int rcu_preempt_needs_cpu(int cpu
)
1096 * Because preemptible RCU does not exist, rcu_barrier() is just
1097 * another name for rcu_barrier_sched().
1099 void rcu_barrier(void)
1101 rcu_barrier_sched();
1103 EXPORT_SYMBOL_GPL(rcu_barrier
);
1106 * Because preemptible RCU does not exist, there is no per-CPU
1107 * data to initialize.
1109 static void __cpuinit
rcu_preempt_init_percpu_data(int cpu
)
1114 * Because there is no preemptible RCU, there are no callbacks to move.
1116 static void rcu_preempt_send_cbs_to_online(void)
1121 * Because preemptible RCU does not exist, it need not be initialized.
1123 static void __init
__rcu_init_preempt(void)
1127 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1129 #ifdef CONFIG_RCU_BOOST
1131 #include "rtmutex_common.h"
1133 #ifdef CONFIG_RCU_TRACE
1135 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1137 if (list_empty(&rnp
->blkd_tasks
))
1138 rnp
->n_balk_blkd_tasks
++;
1139 else if (rnp
->exp_tasks
== NULL
&& rnp
->gp_tasks
== NULL
)
1140 rnp
->n_balk_exp_gp_tasks
++;
1141 else if (rnp
->gp_tasks
!= NULL
&& rnp
->boost_tasks
!= NULL
)
1142 rnp
->n_balk_boost_tasks
++;
1143 else if (rnp
->gp_tasks
!= NULL
&& rnp
->qsmask
!= 0)
1144 rnp
->n_balk_notblocked
++;
1145 else if (rnp
->gp_tasks
!= NULL
&&
1146 ULONG_CMP_LT(jiffies
, rnp
->boost_time
))
1147 rnp
->n_balk_notyet
++;
1152 #else /* #ifdef CONFIG_RCU_TRACE */
1154 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1158 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1160 static struct lock_class_key rcu_boost_class
;
1163 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1164 * or ->boost_tasks, advancing the pointer to the next task in the
1165 * ->blkd_tasks list.
1167 * Note that irqs must be enabled: boosting the task can block.
1168 * Returns 1 if there are more tasks needing to be boosted.
1170 static int rcu_boost(struct rcu_node
*rnp
)
1172 unsigned long flags
;
1173 struct rt_mutex mtx
;
1174 struct task_struct
*t
;
1175 struct list_head
*tb
;
1177 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
)
1178 return 0; /* Nothing left to boost. */
1180 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1183 * Recheck under the lock: all tasks in need of boosting
1184 * might exit their RCU read-side critical sections on their own.
1186 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
1187 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1192 * Preferentially boost tasks blocking expedited grace periods.
1193 * This cannot starve the normal grace periods because a second
1194 * expedited grace period must boost all blocked tasks, including
1195 * those blocking the pre-existing normal grace period.
1197 if (rnp
->exp_tasks
!= NULL
) {
1198 tb
= rnp
->exp_tasks
;
1199 rnp
->n_exp_boosts
++;
1201 tb
= rnp
->boost_tasks
;
1202 rnp
->n_normal_boosts
++;
1204 rnp
->n_tasks_boosted
++;
1207 * We boost task t by manufacturing an rt_mutex that appears to
1208 * be held by task t. We leave a pointer to that rt_mutex where
1209 * task t can find it, and task t will release the mutex when it
1210 * exits its outermost RCU read-side critical section. Then
1211 * simply acquiring this artificial rt_mutex will boost task
1212 * t's priority. (Thanks to tglx for suggesting this approach!)
1214 * Note that task t must acquire rnp->lock to remove itself from
1215 * the ->blkd_tasks list, which it will do from exit() if from
1216 * nowhere else. We therefore are guaranteed that task t will
1217 * stay around at least until we drop rnp->lock. Note that
1218 * rnp->lock also resolves races between our priority boosting
1219 * and task t's exiting its outermost RCU read-side critical
1222 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1223 rt_mutex_init_proxy_locked(&mtx
, t
);
1224 /* Avoid lockdep false positives. This rt_mutex is its own thing. */
1225 lockdep_set_class_and_name(&mtx
.wait_lock
, &rcu_boost_class
,
1227 t
->rcu_boost_mutex
= &mtx
;
1228 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1229 rt_mutex_lock(&mtx
); /* Side effect: boosts task t's priority. */
1230 rt_mutex_unlock(&mtx
); /* Keep lockdep happy. */
1231 local_irq_restore(flags
);
1233 return rnp
->exp_tasks
!= NULL
|| rnp
->boost_tasks
!= NULL
;
1237 * Timer handler to initiate waking up of boost kthreads that
1238 * have yielded the CPU due to excessive numbers of tasks to
1239 * boost. We wake up the per-rcu_node kthread, which in turn
1240 * will wake up the booster kthread.
1242 static void rcu_boost_kthread_timer(unsigned long arg
)
1244 invoke_rcu_node_kthread((struct rcu_node
*)arg
);
1248 * Priority-boosting kthread. One per leaf rcu_node and one for the
1251 static int rcu_boost_kthread(void *arg
)
1253 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1257 trace_rcu_utilization("Start boost kthread@init");
1259 rnp
->boost_kthread_status
= RCU_KTHREAD_WAITING
;
1260 trace_rcu_utilization("End boost kthread@rcu_wait");
1261 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1262 trace_rcu_utilization("Start boost kthread@rcu_wait");
1263 rnp
->boost_kthread_status
= RCU_KTHREAD_RUNNING
;
1264 more2boost
= rcu_boost(rnp
);
1270 trace_rcu_utilization("End boost kthread@rcu_yield");
1271 rcu_yield(rcu_boost_kthread_timer
, (unsigned long)rnp
);
1272 trace_rcu_utilization("Start boost kthread@rcu_yield");
1277 trace_rcu_utilization("End boost kthread@notreached");
1282 * Check to see if it is time to start boosting RCU readers that are
1283 * blocking the current grace period, and, if so, tell the per-rcu_node
1284 * kthread to start boosting them. If there is an expedited grace
1285 * period in progress, it is always time to boost.
1287 * The caller must hold rnp->lock, which this function releases,
1288 * but irqs remain disabled. The ->boost_kthread_task is immortal,
1289 * so we don't need to worry about it going away.
1291 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1293 struct task_struct
*t
;
1295 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1296 rnp
->n_balk_exp_gp_tasks
++;
1297 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1300 if (rnp
->exp_tasks
!= NULL
||
1301 (rnp
->gp_tasks
!= NULL
&&
1302 rnp
->boost_tasks
== NULL
&&
1304 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))) {
1305 if (rnp
->exp_tasks
== NULL
)
1306 rnp
->boost_tasks
= rnp
->gp_tasks
;
1307 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1308 t
= rnp
->boost_kthread_task
;
1312 rcu_initiate_boost_trace(rnp
);
1313 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1318 * Wake up the per-CPU kthread to invoke RCU callbacks.
1320 static void invoke_rcu_callbacks_kthread(void)
1322 unsigned long flags
;
1324 local_irq_save(flags
);
1325 __this_cpu_write(rcu_cpu_has_work
, 1);
1326 if (__this_cpu_read(rcu_cpu_kthread_task
) != NULL
&&
1327 current
!= __this_cpu_read(rcu_cpu_kthread_task
))
1328 wake_up_process(__this_cpu_read(rcu_cpu_kthread_task
));
1329 local_irq_restore(flags
);
1333 * Set the affinity of the boost kthread. The CPU-hotplug locks are
1334 * held, so no one should be messing with the existence of the boost
1337 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
,
1340 struct task_struct
*t
;
1342 t
= rnp
->boost_kthread_task
;
1344 set_cpus_allowed_ptr(rnp
->boost_kthread_task
, cm
);
1347 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1350 * Do priority-boost accounting for the start of a new grace period.
1352 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1354 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1358 * Create an RCU-boost kthread for the specified node if one does not
1359 * already exist. We only create this kthread for preemptible RCU.
1360 * Returns zero if all is well, a negated errno otherwise.
1362 static int __cpuinit
rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1363 struct rcu_node
*rnp
,
1366 unsigned long flags
;
1367 struct sched_param sp
;
1368 struct task_struct
*t
;
1370 if (&rcu_preempt_state
!= rsp
)
1373 if (rnp
->boost_kthread_task
!= NULL
)
1375 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1376 "rcub/%d", rnp_index
);
1379 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1380 rnp
->boost_kthread_task
= t
;
1381 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1382 sp
.sched_priority
= RCU_BOOST_PRIO
;
1383 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1384 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1388 #ifdef CONFIG_HOTPLUG_CPU
1391 * Stop the RCU's per-CPU kthread when its CPU goes offline,.
1393 static void rcu_stop_cpu_kthread(int cpu
)
1395 struct task_struct
*t
;
1397 /* Stop the CPU's kthread. */
1398 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1400 per_cpu(rcu_cpu_kthread_task
, cpu
) = NULL
;
1405 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1407 static void rcu_kthread_do_work(void)
1409 rcu_do_batch(&rcu_sched_state
, &__get_cpu_var(rcu_sched_data
));
1410 rcu_do_batch(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1411 rcu_preempt_do_callbacks();
1415 * Wake up the specified per-rcu_node-structure kthread.
1416 * Because the per-rcu_node kthreads are immortal, we don't need
1417 * to do anything to keep them alive.
1419 static void invoke_rcu_node_kthread(struct rcu_node
*rnp
)
1421 struct task_struct
*t
;
1423 t
= rnp
->node_kthread_task
;
1429 * Set the specified CPU's kthread to run RT or not, as specified by
1430 * the to_rt argument. The CPU-hotplug locks are held, so the task
1431 * is not going away.
1433 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1436 struct sched_param sp
;
1437 struct task_struct
*t
;
1439 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1443 policy
= SCHED_FIFO
;
1444 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1446 policy
= SCHED_NORMAL
;
1447 sp
.sched_priority
= 0;
1449 sched_setscheduler_nocheck(t
, policy
, &sp
);
1453 * Timer handler to initiate the waking up of per-CPU kthreads that
1454 * have yielded the CPU due to excess numbers of RCU callbacks.
1455 * We wake up the per-rcu_node kthread, which in turn will wake up
1456 * the booster kthread.
1458 static void rcu_cpu_kthread_timer(unsigned long arg
)
1460 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, arg
);
1461 struct rcu_node
*rnp
= rdp
->mynode
;
1463 atomic_or(rdp
->grpmask
, &rnp
->wakemask
);
1464 invoke_rcu_node_kthread(rnp
);
1468 * Drop to non-real-time priority and yield, but only after posting a
1469 * timer that will cause us to regain our real-time priority if we
1470 * remain preempted. Either way, we restore our real-time priority
1473 static void rcu_yield(void (*f
)(unsigned long), unsigned long arg
)
1475 struct sched_param sp
;
1476 struct timer_list yield_timer
;
1477 int prio
= current
->normal_prio
;
1479 setup_timer_on_stack(&yield_timer
, f
, arg
);
1480 mod_timer(&yield_timer
, jiffies
+ 2);
1481 sp
.sched_priority
= 0;
1482 sched_setscheduler_nocheck(current
, SCHED_NORMAL
, &sp
);
1483 set_user_nice(current
, 19);
1485 set_user_nice(current
, 0);
1486 sp
.sched_priority
= prio
;
1487 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1488 del_timer(&yield_timer
);
1492 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1493 * This can happen while the corresponding CPU is either coming online
1494 * or going offline. We cannot wait until the CPU is fully online
1495 * before starting the kthread, because the various notifier functions
1496 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1497 * the corresponding CPU is online.
1499 * Return 1 if the kthread needs to stop, 0 otherwise.
1501 * Caller must disable bh. This function can momentarily enable it.
1503 static int rcu_cpu_kthread_should_stop(int cpu
)
1505 while (cpu_is_offline(cpu
) ||
1506 !cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)) ||
1507 smp_processor_id() != cpu
) {
1508 if (kthread_should_stop())
1510 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1511 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = raw_smp_processor_id();
1513 schedule_timeout_uninterruptible(1);
1514 if (!cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)))
1515 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
1518 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1523 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1524 * RCU softirq used in flavors and configurations of RCU that do not
1525 * support RCU priority boosting.
1527 static int rcu_cpu_kthread(void *arg
)
1529 int cpu
= (int)(long)arg
;
1530 unsigned long flags
;
1532 unsigned int *statusp
= &per_cpu(rcu_cpu_kthread_status
, cpu
);
1534 char *workp
= &per_cpu(rcu_cpu_has_work
, cpu
);
1536 trace_rcu_utilization("Start CPU kthread@init");
1538 *statusp
= RCU_KTHREAD_WAITING
;
1539 trace_rcu_utilization("End CPU kthread@rcu_wait");
1540 rcu_wait(*workp
!= 0 || kthread_should_stop());
1541 trace_rcu_utilization("Start CPU kthread@rcu_wait");
1543 if (rcu_cpu_kthread_should_stop(cpu
)) {
1547 *statusp
= RCU_KTHREAD_RUNNING
;
1548 per_cpu(rcu_cpu_kthread_loops
, cpu
)++;
1549 local_irq_save(flags
);
1552 local_irq_restore(flags
);
1554 rcu_kthread_do_work();
1561 *statusp
= RCU_KTHREAD_YIELDING
;
1562 trace_rcu_utilization("End CPU kthread@rcu_yield");
1563 rcu_yield(rcu_cpu_kthread_timer
, (unsigned long)cpu
);
1564 trace_rcu_utilization("Start CPU kthread@rcu_yield");
1568 *statusp
= RCU_KTHREAD_STOPPED
;
1569 trace_rcu_utilization("End CPU kthread@term");
1574 * Spawn a per-CPU kthread, setting up affinity and priority.
1575 * Because the CPU hotplug lock is held, no other CPU will be attempting
1576 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1577 * attempting to access it during boot, but the locking in kthread_bind()
1578 * will enforce sufficient ordering.
1580 * Please note that we cannot simply refuse to wake up the per-CPU
1581 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1582 * which can result in softlockup complaints if the task ends up being
1583 * idle for more than a couple of minutes.
1585 * However, please note also that we cannot bind the per-CPU kthread to its
1586 * CPU until that CPU is fully online. We also cannot wait until the
1587 * CPU is fully online before we create its per-CPU kthread, as this would
1588 * deadlock the system when CPU notifiers tried waiting for grace
1589 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1590 * is online. If its CPU is not yet fully online, then the code in
1591 * rcu_cpu_kthread() will wait until it is fully online, and then do
1594 static int __cpuinit
rcu_spawn_one_cpu_kthread(int cpu
)
1596 struct sched_param sp
;
1597 struct task_struct
*t
;
1599 if (!rcu_scheduler_fully_active
||
1600 per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
)
1602 t
= kthread_create_on_node(rcu_cpu_kthread
,
1608 if (cpu_online(cpu
))
1609 kthread_bind(t
, cpu
);
1610 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1611 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
);
1612 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1613 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1614 per_cpu(rcu_cpu_kthread_task
, cpu
) = t
;
1615 wake_up_process(t
); /* Get to TASK_INTERRUPTIBLE quickly. */
1620 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1621 * kthreads when needed. We ignore requests to wake up kthreads
1622 * for offline CPUs, which is OK because force_quiescent_state()
1623 * takes care of this case.
1625 static int rcu_node_kthread(void *arg
)
1628 unsigned long flags
;
1630 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1631 struct sched_param sp
;
1632 struct task_struct
*t
;
1635 rnp
->node_kthread_status
= RCU_KTHREAD_WAITING
;
1636 rcu_wait(atomic_read(&rnp
->wakemask
) != 0);
1637 rnp
->node_kthread_status
= RCU_KTHREAD_RUNNING
;
1638 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1639 smp_mb(); /* Work around some architectures weak impls. */
1640 mask
= atomic_xchg(&rnp
->wakemask
, 0);
1641 smp_mb(); /* Work around some architectures weak impls. */
1642 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1643 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1) {
1644 if ((mask
& 0x1) == 0)
1647 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1648 if (!cpu_online(cpu
) || t
== NULL
) {
1652 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1653 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1654 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1659 rnp
->node_kthread_status
= RCU_KTHREAD_STOPPED
;
1664 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1665 * served by the rcu_node in question. The CPU hotplug lock is still
1666 * held, so the value of rnp->qsmaskinit will be stable.
1668 * We don't include outgoingcpu in the affinity set, use -1 if there is
1669 * no outgoing CPU. If there are no CPUs left in the affinity set,
1670 * this function allows the kthread to execute on any CPU.
1672 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1676 unsigned long mask
= rnp
->qsmaskinit
;
1678 if (rnp
->node_kthread_task
== NULL
)
1680 if (!alloc_cpumask_var(&cm
, GFP_KERNEL
))
1683 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1684 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1685 cpumask_set_cpu(cpu
, cm
);
1686 if (cpumask_weight(cm
) == 0) {
1688 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1689 cpumask_clear_cpu(cpu
, cm
);
1690 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1692 set_cpus_allowed_ptr(rnp
->node_kthread_task
, cm
);
1693 rcu_boost_kthread_setaffinity(rnp
, cm
);
1694 free_cpumask_var(cm
);
1698 * Spawn a per-rcu_node kthread, setting priority and affinity.
1699 * Called during boot before online/offline can happen, or, if
1700 * during runtime, with the main CPU-hotplug locks held. So only
1701 * one of these can be executing at a time.
1703 static int __cpuinit
rcu_spawn_one_node_kthread(struct rcu_state
*rsp
,
1704 struct rcu_node
*rnp
)
1706 unsigned long flags
;
1707 int rnp_index
= rnp
- &rsp
->node
[0];
1708 struct sched_param sp
;
1709 struct task_struct
*t
;
1711 if (!rcu_scheduler_fully_active
||
1712 rnp
->qsmaskinit
== 0)
1714 if (rnp
->node_kthread_task
== NULL
) {
1715 t
= kthread_create(rcu_node_kthread
, (void *)rnp
,
1716 "rcun/%d", rnp_index
);
1719 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1720 rnp
->node_kthread_task
= t
;
1721 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1722 sp
.sched_priority
= 99;
1723 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1724 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1726 return rcu_spawn_one_boost_kthread(rsp
, rnp
, rnp_index
);
1730 * Spawn all kthreads -- called as soon as the scheduler is running.
1732 static int __init
rcu_spawn_kthreads(void)
1735 struct rcu_node
*rnp
;
1737 rcu_scheduler_fully_active
= 1;
1738 for_each_possible_cpu(cpu
) {
1739 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1740 if (cpu_online(cpu
))
1741 (void)rcu_spawn_one_cpu_kthread(cpu
);
1743 rnp
= rcu_get_root(rcu_state
);
1744 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1745 if (NUM_RCU_NODES
> 1) {
1746 rcu_for_each_leaf_node(rcu_state
, rnp
)
1747 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1751 early_initcall(rcu_spawn_kthreads
);
1753 static void __cpuinit
rcu_prepare_kthreads(int cpu
)
1755 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
1756 struct rcu_node
*rnp
= rdp
->mynode
;
1758 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1759 if (rcu_scheduler_fully_active
) {
1760 (void)rcu_spawn_one_cpu_kthread(cpu
);
1761 if (rnp
->node_kthread_task
== NULL
)
1762 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1766 #else /* #ifdef CONFIG_RCU_BOOST */
1768 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1770 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1773 static void invoke_rcu_callbacks_kthread(void)
1778 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1782 #ifdef CONFIG_HOTPLUG_CPU
1784 static void rcu_stop_cpu_kthread(int cpu
)
1788 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1790 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1794 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1798 static int __init
rcu_scheduler_really_started(void)
1800 rcu_scheduler_fully_active
= 1;
1803 early_initcall(rcu_scheduler_really_started
);
1805 static void __cpuinit
rcu_prepare_kthreads(int cpu
)
1809 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1813 void synchronize_sched_expedited(void)
1817 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
1819 #else /* #ifndef CONFIG_SMP */
1821 static atomic_t sync_sched_expedited_started
= ATOMIC_INIT(0);
1822 static atomic_t sync_sched_expedited_done
= ATOMIC_INIT(0);
1824 static int synchronize_sched_expedited_cpu_stop(void *data
)
1827 * There must be a full memory barrier on each affected CPU
1828 * between the time that try_stop_cpus() is called and the
1829 * time that it returns.
1831 * In the current initial implementation of cpu_stop, the
1832 * above condition is already met when the control reaches
1833 * this point and the following smp_mb() is not strictly
1834 * necessary. Do smp_mb() anyway for documentation and
1835 * robustness against future implementation changes.
1837 smp_mb(); /* See above comment block. */
1842 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
1843 * approach to force grace period to end quickly. This consumes
1844 * significant time on all CPUs, and is thus not recommended for
1845 * any sort of common-case code.
1847 * Note that it is illegal to call this function while holding any
1848 * lock that is acquired by a CPU-hotplug notifier. Failing to
1849 * observe this restriction will result in deadlock.
1851 * This implementation can be thought of as an application of ticket
1852 * locking to RCU, with sync_sched_expedited_started and
1853 * sync_sched_expedited_done taking on the roles of the halves
1854 * of the ticket-lock word. Each task atomically increments
1855 * sync_sched_expedited_started upon entry, snapshotting the old value,
1856 * then attempts to stop all the CPUs. If this succeeds, then each
1857 * CPU will have executed a context switch, resulting in an RCU-sched
1858 * grace period. We are then done, so we use atomic_cmpxchg() to
1859 * update sync_sched_expedited_done to match our snapshot -- but
1860 * only if someone else has not already advanced past our snapshot.
1862 * On the other hand, if try_stop_cpus() fails, we check the value
1863 * of sync_sched_expedited_done. If it has advanced past our
1864 * initial snapshot, then someone else must have forced a grace period
1865 * some time after we took our snapshot. In this case, our work is
1866 * done for us, and we can simply return. Otherwise, we try again,
1867 * but keep our initial snapshot for purposes of checking for someone
1868 * doing our work for us.
1870 * If we fail too many times in a row, we fall back to synchronize_sched().
1872 void synchronize_sched_expedited(void)
1874 int firstsnap
, s
, snap
, trycount
= 0;
1876 /* Note that atomic_inc_return() implies full memory barrier. */
1877 smp_mb(); /* Work around some architectures weak impls. */
1878 firstsnap
= snap
= atomic_inc_return(&sync_sched_expedited_started
);
1879 smp_mb(); /* Work around some architectures weak impls. */
1883 * Each pass through the following loop attempts to force a
1884 * context switch on each CPU.
1886 while (try_stop_cpus(cpu_online_mask
,
1887 synchronize_sched_expedited_cpu_stop
,
1891 /* No joy, try again later. Or just synchronize_sched(). */
1892 if (trycount
++ < 10)
1893 udelay(trycount
* num_online_cpus());
1895 synchronize_sched();
1899 /* Check to see if someone else did our work for us. */
1900 s
= atomic_read(&sync_sched_expedited_done
);
1901 if (UINT_CMP_GE((unsigned)s
, (unsigned)firstsnap
)) {
1902 smp_mb(); /* ensure test happens before caller kfree */
1907 * Refetching sync_sched_expedited_started allows later
1908 * callers to piggyback on our grace period. We subtract
1909 * 1 to get the same token that the last incrementer got.
1910 * We retry after they started, so our grace period works
1911 * for them, and they started after our first try, so their
1912 * grace period works for us.
1915 snap
= atomic_read(&sync_sched_expedited_started
) - 1;
1916 smp_mb(); /* ensure read is before try_stop_cpus(). */
1920 * Everyone up to our most recent fetch is covered by our grace
1921 * period. Update the counter, but only if our work is still
1922 * relevant -- which it won't be if someone who started later
1923 * than we did beat us to the punch.
1926 s
= atomic_read(&sync_sched_expedited_done
);
1927 if (UINT_CMP_GE((unsigned)s
, (unsigned)snap
)) {
1928 smp_mb(); /* ensure test happens before caller kfree */
1931 } while (atomic_cmpxchg(&sync_sched_expedited_done
, s
, snap
) != s
);
1936 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
1938 #endif /* #else #ifndef CONFIG_SMP */
1940 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1943 * Check to see if any future RCU-related work will need to be done
1944 * by the current CPU, even if none need be done immediately, returning
1945 * 1 if so. This function is part of the RCU implementation; it is -not-
1946 * an exported member of the RCU API.
1948 * Because we have preemptible RCU, just check whether this CPU needs
1949 * any flavor of RCU. Do not chew up lots of CPU cycles with preemption
1950 * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
1952 int rcu_needs_cpu(int cpu
)
1954 return rcu_needs_cpu_quick_check(cpu
);
1957 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1959 #define RCU_NEEDS_CPU_FLUSHES 5
1960 static DEFINE_PER_CPU(int, rcu_dyntick_drain
);
1961 static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff
);
1964 * Check to see if any future RCU-related work will need to be done
1965 * by the current CPU, even if none need be done immediately, returning
1966 * 1 if so. This function is part of the RCU implementation; it is -not-
1967 * an exported member of the RCU API.
1969 * Because we are not supporting preemptible RCU, attempt to accelerate
1970 * any current grace periods so that RCU no longer needs this CPU, but
1971 * only if all other CPUs are already in dynticks-idle mode. This will
1972 * allow the CPU cores to be powered down immediately, as opposed to after
1973 * waiting many milliseconds for grace periods to elapse.
1975 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1976 * disabled, we do one pass of force_quiescent_state(), then do a
1977 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
1978 * later. The per-cpu rcu_dyntick_drain variable controls the sequencing.
1980 int rcu_needs_cpu(int cpu
)
1986 /* Check for being in the holdoff period. */
1987 if (per_cpu(rcu_dyntick_holdoff
, cpu
) == jiffies
)
1988 return rcu_needs_cpu_quick_check(cpu
);
1990 /* Don't bother unless we are the last non-dyntick-idle CPU. */
1991 for_each_online_cpu(thatcpu
) {
1994 smp_mb(); /* Work around some architectures weak impls. */
1995 snap
= atomic_add_return(0, &per_cpu(rcu_dynticks
,
1997 smp_mb(); /* Work around some architectures weak impls. */
1998 if ((snap
& 0x1) != 0) {
1999 per_cpu(rcu_dyntick_drain
, cpu
) = 0;
2000 per_cpu(rcu_dyntick_holdoff
, cpu
) = jiffies
- 1;
2001 return rcu_needs_cpu_quick_check(cpu
);
2005 /* Check and update the rcu_dyntick_drain sequencing. */
2006 if (per_cpu(rcu_dyntick_drain
, cpu
) <= 0) {
2007 /* First time through, initialize the counter. */
2008 per_cpu(rcu_dyntick_drain
, cpu
) = RCU_NEEDS_CPU_FLUSHES
;
2009 } else if (--per_cpu(rcu_dyntick_drain
, cpu
) <= 0) {
2010 /* We have hit the limit, so time to give up. */
2011 per_cpu(rcu_dyntick_holdoff
, cpu
) = jiffies
;
2012 return rcu_needs_cpu_quick_check(cpu
);
2015 /* Do one step pushing remaining RCU callbacks through. */
2016 if (per_cpu(rcu_sched_data
, cpu
).nxtlist
) {
2018 force_quiescent_state(&rcu_sched_state
, 0);
2019 c
= c
|| per_cpu(rcu_sched_data
, cpu
).nxtlist
;
2021 if (per_cpu(rcu_bh_data
, cpu
).nxtlist
) {
2023 force_quiescent_state(&rcu_bh_state
, 0);
2024 c
= c
|| per_cpu(rcu_bh_data
, cpu
).nxtlist
;
2027 /* If RCU callbacks are still pending, RCU still needs this CPU. */
2033 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */