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_quiesce_gpnum = rdp->gpnum;
127 if (rdp->passed_quiesce == 0)
128 trace_rcu_grace_period("rcu_preempt", rdp->gpnum, "cpuqs");
129 rdp->passed_quiesce = 1;
130 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
134 * We have entered the scheduler, and the current task might soon be
135 * context-switched away from. If this task is in an RCU read-side
136 * critical section, we will no longer be able to rely on the CPU to
137 * record that fact, so we enqueue the task on the blkd_tasks list.
138 * The task will dequeue itself when it exits the outermost enclosing
139 * RCU read-side critical section. Therefore, the current grace period
140 * cannot be permitted to complete until the blkd_tasks list entries
141 * predating the current grace period drain, in other words, until
142 * rnp->gp_tasks becomes NULL.
144 * Caller must disable preemption.
146 static void rcu_preempt_note_context_switch(int cpu)
148 struct task_struct *t = current;
150 struct rcu_data *rdp;
151 struct rcu_node *rnp;
153 if (t->rcu_read_lock_nesting > 0 &&
154 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
156 /* Possibly blocking in an RCU read-side critical section. */
157 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
159 raw_spin_lock_irqsave(&rnp->lock, flags);
160 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
161 t->rcu_blocked_node = rnp;
164 * If this CPU has already checked in, then this task
165 * will hold up the next grace period rather than the
166 * current grace period. Queue the task accordingly.
167 * If the task is queued for the current grace period
168 * (i.e., this CPU has not yet passed through a quiescent
169 * state for the current grace period), then as long
170 * as that task remains queued, the current grace period
171 * cannot end. Note that there is some uncertainty as
172 * to exactly when the current grace period started.
173 * We take a conservative approach, which can result
174 * in unnecessarily waiting on tasks that started very
175 * slightly after the current grace period began. C'est
178 * But first, note that the current CPU must still be
181 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
182 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
183 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
184 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
185 rnp->gp_tasks = &t->rcu_node_entry;
186 #ifdef CONFIG_RCU_BOOST
187 if (rnp->boost_tasks != NULL)
188 rnp->boost_tasks = rnp->gp_tasks;
189 #endif /* #ifdef CONFIG_RCU_BOOST */
191 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
192 if (rnp->qsmask & rdp->grpmask)
193 rnp->gp_tasks = &t->rcu_node_entry;
195 trace_rcu_preempt_task(rdp->rsp->name,
197 (rnp->qsmask & rdp->grpmask)
200 raw_spin_unlock_irqrestore(&rnp->lock, flags);
201 } else if (t->rcu_read_lock_nesting < 0 &&
202 t->rcu_read_unlock_special) {
205 * Complete exit from RCU read-side critical section on
206 * behalf of preempted instance of __rcu_read_unlock().
208 rcu_read_unlock_special(t);
212 * Either we were not in an RCU read-side critical section to
213 * begin with, or we have now recorded that critical section
214 * globally. Either way, we can now note a quiescent state
215 * for this CPU. Again, if we were in an RCU read-side critical
216 * section, and if that critical section was blocking the current
217 * grace period, then the fact that the task has been enqueued
218 * means that we continue to block the current grace period.
220 local_irq_save(flags);
222 local_irq_restore(flags);
226 * Tree-preemptible RCU implementation for rcu_read_lock().
227 * Just increment ->rcu_read_lock_nesting, shared state will be updated
230 void __rcu_read_lock(void)
232 current->rcu_read_lock_nesting++;
233 barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
235 EXPORT_SYMBOL_GPL(__rcu_read_lock);
238 * Check for preempted RCU readers blocking the current grace period
239 * for the specified rcu_node structure. If the caller needs a reliable
240 * answer, it must hold the rcu_node's ->lock.
242 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
244 return rnp->gp_tasks != NULL;
248 * Record a quiescent state for all tasks that were previously queued
249 * on the specified rcu_node structure and that were blocking the current
250 * RCU grace period. The caller must hold the specified rnp->lock with
251 * irqs disabled, and this lock is released upon return, but irqs remain
254 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
255 __releases(rnp->lock)
258 struct rcu_node *rnp_p;
260 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
261 raw_spin_unlock_irqrestore(&rnp->lock, flags);
262 return; /* Still need more quiescent states! */
268 * Either there is only one rcu_node in the tree,
269 * or tasks were kicked up to root rcu_node due to
270 * CPUs going offline.
272 rcu_report_qs_rsp(&rcu_preempt_state, flags);
276 /* Report up the rest of the hierarchy. */
278 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
279 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
280 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
284 * Advance a ->blkd_tasks-list pointer to the next entry, instead
285 * returning NULL if at the end of the list.
287 static struct list_head *rcu_next_node_entry(struct task_struct *t,
288 struct rcu_node *rnp)
290 struct list_head *np;
292 np = t->rcu_node_entry.next;
293 if (np == &rnp->blkd_tasks)
299 * Handle special cases during rcu_read_unlock(), such as needing to
300 * notify RCU core processing or task having blocked during the RCU
301 * read-side critical section.
303 static noinline void rcu_read_unlock_special(struct task_struct *t)
308 struct list_head *np;
309 #ifdef CONFIG_RCU_BOOST
310 struct rt_mutex *rbmp = NULL;
311 #endif /* #ifdef CONFIG_RCU_BOOST */
312 struct rcu_node *rnp;
315 /* NMI handlers cannot block and cannot safely manipulate state. */
319 local_irq_save(flags);
322 * If RCU core is waiting for this CPU to exit critical section,
323 * let it know that we have done so.
325 special = t->rcu_read_unlock_special;
326 if (special & RCU_READ_UNLOCK_NEED_QS) {
327 rcu_preempt_qs(smp_processor_id());
330 /* Hardware IRQ handlers cannot block. */
331 if (in_irq() || in_serving_softirq()) {
332 local_irq_restore(flags);
336 /* Clean up if blocked during RCU read-side critical section. */
337 if (special & RCU_READ_UNLOCK_BLOCKED) {
338 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
341 * Remove this task from the list it blocked on. The
342 * task can migrate while we acquire the lock, but at
343 * most one time. So at most two passes through loop.
346 rnp = t->rcu_blocked_node;
347 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
348 if (rnp == t->rcu_blocked_node)
350 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
352 empty = !rcu_preempt_blocked_readers_cgp(rnp);
353 empty_exp = !rcu_preempted_readers_exp(rnp);
354 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
355 np = rcu_next_node_entry(t, rnp);
356 list_del_init(&t->rcu_node_entry);
357 t->rcu_blocked_node = NULL;
358 trace_rcu_unlock_preempted_task("rcu_preempt",
360 if (&t->rcu_node_entry == rnp->gp_tasks)
362 if (&t->rcu_node_entry == rnp->exp_tasks)
364 #ifdef CONFIG_RCU_BOOST
365 if (&t->rcu_node_entry == rnp->boost_tasks)
366 rnp->boost_tasks = np;
367 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
368 if (t->rcu_boost_mutex) {
369 rbmp = t->rcu_boost_mutex;
370 t->rcu_boost_mutex = NULL;
372 #endif /* #ifdef CONFIG_RCU_BOOST */
375 * If this was the last task on the current list, and if
376 * we aren't waiting on any CPUs, report the quiescent state.
377 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
379 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
380 trace_rcu_quiescent_state_report("preempt_rcu",
387 rcu_report_unblock_qs_rnp(rnp, flags);
389 raw_spin_unlock_irqrestore(&rnp->lock, flags);
391 #ifdef CONFIG_RCU_BOOST
392 /* Unboost if we were boosted. */
394 rt_mutex_unlock(rbmp);
395 #endif /* #ifdef CONFIG_RCU_BOOST */
398 * If this was the last task on the expedited lists,
399 * then we need to report up the rcu_node hierarchy.
401 if (!empty_exp && !rcu_preempted_readers_exp(rnp))
402 rcu_report_exp_rnp(&rcu_preempt_state, rnp);
404 local_irq_restore(flags);
409 * Tree-preemptible RCU implementation for rcu_read_unlock().
410 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
411 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
412 * invoke rcu_read_unlock_special() to clean up after a context switch
413 * in an RCU read-side critical section and other special cases.
415 void __rcu_read_unlock(void)
417 struct task_struct *t = current;
419 if (t->rcu_read_lock_nesting != 1)
420 --t->rcu_read_lock_nesting;
422 barrier(); /* critical section before exit code. */
423 t->rcu_read_lock_nesting = INT_MIN;
424 barrier(); /* assign before ->rcu_read_unlock_special load */
425 if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
426 rcu_read_unlock_special(t);
427 barrier(); /* ->rcu_read_unlock_special load before assign */
428 t->rcu_read_lock_nesting = 0;
430 #ifdef CONFIG_PROVE_LOCKING
432 int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting);
434 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
436 #endif /* #ifdef CONFIG_PROVE_LOCKING */
438 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
440 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
443 * Dump detailed information for all tasks blocking the current RCU
444 * grace period on the specified rcu_node structure.
446 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
449 struct task_struct *t;
451 if (!rcu_preempt_blocked_readers_cgp(rnp))
453 raw_spin_lock_irqsave(&rnp->lock, flags);
454 t = list_entry(rnp->gp_tasks,
455 struct task_struct, rcu_node_entry);
456 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
458 raw_spin_unlock_irqrestore(&rnp->lock, flags);
462 * Dump detailed information for all tasks blocking the current RCU
465 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
467 struct rcu_node *rnp = rcu_get_root(rsp);
469 rcu_print_detail_task_stall_rnp(rnp);
470 rcu_for_each_leaf_node(rsp, rnp)
471 rcu_print_detail_task_stall_rnp(rnp);
474 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
476 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
480 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
483 * Scan the current list of tasks blocked within RCU read-side critical
484 * sections, printing out the tid of each.
486 static void rcu_print_task_stall(struct rcu_node *rnp)
488 struct task_struct *t;
490 if (!rcu_preempt_blocked_readers_cgp(rnp))
492 t = list_entry(rnp->gp_tasks,
493 struct task_struct, rcu_node_entry);
494 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
495 printk(" P%d", t->pid);
499 * Suppress preemptible RCU's CPU stall warnings by pushing the
500 * time of the next stall-warning message comfortably far into the
503 static void rcu_preempt_stall_reset(void)
505 rcu_preempt_state.jiffies_stall = jiffies + ULONG_MAX / 2;
509 * Check that the list of blocked tasks for the newly completed grace
510 * period is in fact empty. It is a serious bug to complete a grace
511 * period that still has RCU readers blocked! This function must be
512 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
513 * must be held by the caller.
515 * Also, if there are blocked tasks on the list, they automatically
516 * block the newly created grace period, so set up ->gp_tasks accordingly.
518 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
520 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
521 if (!list_empty(&rnp->blkd_tasks))
522 rnp->gp_tasks = rnp->blkd_tasks.next;
523 WARN_ON_ONCE(rnp->qsmask);
526 #ifdef CONFIG_HOTPLUG_CPU
529 * Handle tasklist migration for case in which all CPUs covered by the
530 * specified rcu_node have gone offline. Move them up to the root
531 * rcu_node. The reason for not just moving them to the immediate
532 * parent is to remove the need for rcu_read_unlock_special() to
533 * make more than two attempts to acquire the target rcu_node's lock.
534 * Returns true if there were tasks blocking the current RCU grace
537 * Returns 1 if there was previously a task blocking the current grace
538 * period on the specified rcu_node structure.
540 * The caller must hold rnp->lock with irqs disabled.
542 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
543 struct rcu_node *rnp,
544 struct rcu_data *rdp)
546 struct list_head *lp;
547 struct list_head *lp_root;
549 struct rcu_node *rnp_root = rcu_get_root(rsp);
550 struct task_struct *t;
552 if (rnp == rnp_root) {
553 WARN_ONCE(1, "Last CPU thought to be offlined?");
554 return 0; /* Shouldn't happen: at least one CPU online. */
557 /* If we are on an internal node, complain bitterly. */
558 WARN_ON_ONCE(rnp != rdp->mynode);
561 * Move tasks up to root rcu_node. Don't try to get fancy for
562 * this corner-case operation -- just put this node's tasks
563 * at the head of the root node's list, and update the root node's
564 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
565 * if non-NULL. This might result in waiting for more tasks than
566 * absolutely necessary, but this is a good performance/complexity
569 if (rcu_preempt_blocked_readers_cgp(rnp))
570 retval |= RCU_OFL_TASKS_NORM_GP;
571 if (rcu_preempted_readers_exp(rnp))
572 retval |= RCU_OFL_TASKS_EXP_GP;
573 lp = &rnp->blkd_tasks;
574 lp_root = &rnp_root->blkd_tasks;
575 while (!list_empty(lp)) {
576 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
577 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
578 list_del(&t->rcu_node_entry);
579 t->rcu_blocked_node = rnp_root;
580 list_add(&t->rcu_node_entry, lp_root);
581 if (&t->rcu_node_entry == rnp->gp_tasks)
582 rnp_root->gp_tasks = rnp->gp_tasks;
583 if (&t->rcu_node_entry == rnp->exp_tasks)
584 rnp_root->exp_tasks = rnp->exp_tasks;
585 #ifdef CONFIG_RCU_BOOST
586 if (&t->rcu_node_entry == rnp->boost_tasks)
587 rnp_root->boost_tasks = rnp->boost_tasks;
588 #endif /* #ifdef CONFIG_RCU_BOOST */
589 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
592 #ifdef CONFIG_RCU_BOOST
593 /* In case root is being boosted and leaf is not. */
594 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
595 if (rnp_root->boost_tasks != NULL &&
596 rnp_root->boost_tasks != rnp_root->gp_tasks)
597 rnp_root->boost_tasks = rnp_root->gp_tasks;
598 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
599 #endif /* #ifdef CONFIG_RCU_BOOST */
601 rnp->gp_tasks = NULL;
602 rnp->exp_tasks = NULL;
607 * Do CPU-offline processing for preemptible RCU.
609 static void rcu_preempt_offline_cpu(int cpu)
611 __rcu_offline_cpu(cpu, &rcu_preempt_state);
614 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
617 * Check for a quiescent state from the current CPU. When a task blocks,
618 * the task is recorded in the corresponding CPU's rcu_node structure,
619 * which is checked elsewhere.
621 * Caller must disable hard irqs.
623 static void rcu_preempt_check_callbacks(int cpu)
625 struct task_struct *t = current;
627 if (t->rcu_read_lock_nesting == 0) {
631 if (t->rcu_read_lock_nesting > 0 &&
632 per_cpu(rcu_preempt_data, cpu).qs_pending)
633 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
637 * Process callbacks for preemptible RCU.
639 static void rcu_preempt_process_callbacks(void)
641 __rcu_process_callbacks(&rcu_preempt_state,
642 &__get_cpu_var(rcu_preempt_data));
645 #ifdef CONFIG_RCU_BOOST
647 static void rcu_preempt_do_callbacks(void)
649 rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
652 #endif /* #ifdef CONFIG_RCU_BOOST */
655 * Queue a preemptible-RCU callback for invocation after a grace period.
657 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
659 __call_rcu(head, func, &rcu_preempt_state);
661 EXPORT_SYMBOL_GPL(call_rcu);
664 * synchronize_rcu - wait until a grace period has elapsed.
666 * Control will return to the caller some time after a full grace
667 * period has elapsed, in other words after all currently executing RCU
668 * read-side critical sections have completed. Note, however, that
669 * upon return from synchronize_rcu(), the caller might well be executing
670 * concurrently with new RCU read-side critical sections that began while
671 * synchronize_rcu() was waiting. RCU read-side critical sections are
672 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
674 void synchronize_rcu(void)
676 if (!rcu_scheduler_active)
678 wait_rcu_gp(call_rcu);
680 EXPORT_SYMBOL_GPL(synchronize_rcu);
682 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
683 static long sync_rcu_preempt_exp_count;
684 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
687 * Return non-zero if there are any tasks in RCU read-side critical
688 * sections blocking the current preemptible-RCU expedited grace period.
689 * If there is no preemptible-RCU expedited grace period currently in
690 * progress, returns zero unconditionally.
692 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
694 return rnp->exp_tasks != NULL;
698 * return non-zero if there is no RCU expedited grace period in progress
699 * for the specified rcu_node structure, in other words, if all CPUs and
700 * tasks covered by the specified rcu_node structure have done their bit
701 * for the current expedited grace period. Works only for preemptible
702 * RCU -- other RCU implementation use other means.
704 * Caller must hold sync_rcu_preempt_exp_mutex.
706 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
708 return !rcu_preempted_readers_exp(rnp) &&
709 ACCESS_ONCE(rnp->expmask) == 0;
713 * Report the exit from RCU read-side critical section for the last task
714 * that queued itself during or before the current expedited preemptible-RCU
715 * grace period. This event is reported either to the rcu_node structure on
716 * which the task was queued or to one of that rcu_node structure's ancestors,
717 * recursively up the tree. (Calm down, calm down, we do the recursion
720 * Caller must hold sync_rcu_preempt_exp_mutex.
722 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
727 raw_spin_lock_irqsave(&rnp->lock, flags);
729 if (!sync_rcu_preempt_exp_done(rnp)) {
730 raw_spin_unlock_irqrestore(&rnp->lock, flags);
733 if (rnp->parent == NULL) {
734 raw_spin_unlock_irqrestore(&rnp->lock, flags);
735 wake_up(&sync_rcu_preempt_exp_wq);
739 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
741 raw_spin_lock(&rnp->lock); /* irqs already disabled */
742 rnp->expmask &= ~mask;
747 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
748 * grace period for the specified rcu_node structure. If there are no such
749 * tasks, report it up the rcu_node hierarchy.
751 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
754 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
759 raw_spin_lock_irqsave(&rnp->lock, flags);
760 if (list_empty(&rnp->blkd_tasks))
761 raw_spin_unlock_irqrestore(&rnp->lock, flags);
763 rnp->exp_tasks = rnp->blkd_tasks.next;
764 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
768 rcu_report_exp_rnp(rsp, rnp);
772 * Wait for an rcu-preempt grace period, but expedite it. The basic idea
773 * is to invoke synchronize_sched_expedited() to push all the tasks to
774 * the ->blkd_tasks lists and wait for this list to drain.
776 void synchronize_rcu_expedited(void)
779 struct rcu_node *rnp;
780 struct rcu_state *rsp = &rcu_preempt_state;
784 smp_mb(); /* Caller's modifications seen first by other CPUs. */
785 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
786 smp_mb(); /* Above access cannot bleed into critical section. */
789 * Acquire lock, falling back to synchronize_rcu() if too many
790 * lock-acquisition failures. Of course, if someone does the
791 * expedited grace period for us, just leave.
793 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
795 udelay(trycount * num_online_cpus());
800 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
801 goto mb_ret; /* Others did our work for us. */
803 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
804 goto unlock_mb_ret; /* Others did our work for us. */
806 /* force all RCU readers onto ->blkd_tasks lists. */
807 synchronize_sched_expedited();
809 raw_spin_lock_irqsave(&rsp->onofflock, flags);
811 /* Initialize ->expmask for all non-leaf rcu_node structures. */
812 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
813 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
814 rnp->expmask = rnp->qsmaskinit;
815 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
818 /* Snapshot current state of ->blkd_tasks lists. */
819 rcu_for_each_leaf_node(rsp, rnp)
820 sync_rcu_preempt_exp_init(rsp, rnp);
821 if (NUM_RCU_NODES > 1)
822 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
824 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
826 /* Wait for snapshotted ->blkd_tasks lists to drain. */
827 rnp = rcu_get_root(rsp);
828 wait_event(sync_rcu_preempt_exp_wq,
829 sync_rcu_preempt_exp_done(rnp));
831 /* Clean up and exit. */
832 smp_mb(); /* ensure expedited GP seen before counter increment. */
833 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
835 mutex_unlock(&sync_rcu_preempt_exp_mutex);
837 smp_mb(); /* ensure subsequent action seen after grace period. */
839 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
842 * Check to see if there is any immediate preemptible-RCU-related work
845 static int rcu_preempt_pending(int cpu)
847 return __rcu_pending(&rcu_preempt_state,
848 &per_cpu(rcu_preempt_data, cpu));
852 * Does preemptible RCU need the CPU to stay out of dynticks mode?
854 static int rcu_preempt_needs_cpu(int cpu)
856 return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
860 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
862 void rcu_barrier(void)
864 _rcu_barrier(&rcu_preempt_state, call_rcu);
866 EXPORT_SYMBOL_GPL(rcu_barrier);
869 * Initialize preemptible RCU's per-CPU data.
871 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
873 rcu_init_percpu_data(cpu, &rcu_preempt_state, 1);
877 * Move preemptible RCU's callbacks from dying CPU to other online CPU.
879 static void rcu_preempt_send_cbs_to_online(void)
881 rcu_send_cbs_to_online(&rcu_preempt_state);
885 * Initialize preemptible RCU's state structures.
887 static void __init __rcu_init_preempt(void)
889 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
893 * Check for a task exiting while in a preemptible-RCU read-side
894 * critical section, clean up if so. No need to issue warnings,
895 * as debug_check_no_locks_held() already does this if lockdep
900 struct task_struct *t = current;
902 if (t->rcu_read_lock_nesting == 0)
904 t->rcu_read_lock_nesting = 1;
908 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
910 static struct rcu_state *rcu_state = &rcu_sched_state;
913 * Tell them what RCU they are running.
915 static void __init rcu_bootup_announce(void)
917 printk(KERN_INFO "Hierarchical RCU implementation.\n");
918 rcu_bootup_announce_oddness();
922 * Return the number of RCU batches processed thus far for debug & stats.
924 long rcu_batches_completed(void)
926 return rcu_batches_completed_sched();
928 EXPORT_SYMBOL_GPL(rcu_batches_completed);
931 * Force a quiescent state for RCU, which, because there is no preemptible
932 * RCU, becomes the same as rcu-sched.
934 void rcu_force_quiescent_state(void)
936 rcu_sched_force_quiescent_state();
938 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
941 * Because preemptible RCU does not exist, we never have to check for
942 * CPUs being in quiescent states.
944 static void rcu_preempt_note_context_switch(int cpu)
949 * Because preemptible RCU does not exist, there are never any preempted
952 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
957 #ifdef CONFIG_HOTPLUG_CPU
959 /* Because preemptible RCU does not exist, no quieting of tasks. */
960 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
962 raw_spin_unlock_irqrestore(&rnp->lock, flags);
965 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
968 * Because preemptible RCU does not exist, we never have to check for
969 * tasks blocked within RCU read-side critical sections.
971 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
976 * Because preemptible RCU does not exist, we never have to check for
977 * tasks blocked within RCU read-side critical sections.
979 static void rcu_print_task_stall(struct rcu_node *rnp)
984 * Because preemptible RCU does not exist, there is no need to suppress
985 * its CPU stall warnings.
987 static void rcu_preempt_stall_reset(void)
992 * Because there is no preemptible RCU, there can be no readers blocked,
993 * so there is no need to check for blocked tasks. So check only for
994 * bogus qsmask values.
996 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
998 WARN_ON_ONCE(rnp->qsmask);
1001 #ifdef CONFIG_HOTPLUG_CPU
1004 * Because preemptible RCU does not exist, it never needs to migrate
1005 * tasks that were blocked within RCU read-side critical sections, and
1006 * such non-existent tasks cannot possibly have been blocking the current
1009 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1010 struct rcu_node *rnp,
1011 struct rcu_data *rdp)
1017 * Because preemptible RCU does not exist, it never needs CPU-offline
1020 static void rcu_preempt_offline_cpu(int cpu)
1024 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1027 * Because preemptible RCU does not exist, it never has any callbacks
1030 static void rcu_preempt_check_callbacks(int cpu)
1035 * Because preemptible RCU does not exist, it never has any callbacks
1038 static void rcu_preempt_process_callbacks(void)
1043 * Wait for an rcu-preempt grace period, but make it happen quickly.
1044 * But because preemptible RCU does not exist, map to rcu-sched.
1046 void synchronize_rcu_expedited(void)
1048 synchronize_sched_expedited();
1050 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1052 #ifdef CONFIG_HOTPLUG_CPU
1055 * Because preemptible RCU does not exist, there is never any need to
1056 * report on tasks preempted in RCU read-side critical sections during
1057 * expedited RCU grace periods.
1059 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
1064 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1067 * Because preemptible RCU does not exist, it never has any work to do.
1069 static int rcu_preempt_pending(int cpu)
1075 * Because preemptible RCU does not exist, it never needs any CPU.
1077 static int rcu_preempt_needs_cpu(int cpu)
1083 * Because preemptible RCU does not exist, rcu_barrier() is just
1084 * another name for rcu_barrier_sched().
1086 void rcu_barrier(void)
1088 rcu_barrier_sched();
1090 EXPORT_SYMBOL_GPL(rcu_barrier);
1093 * Because preemptible RCU does not exist, there is no per-CPU
1094 * data to initialize.
1096 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
1101 * Because there is no preemptible RCU, there are no callbacks to move.
1103 static void rcu_preempt_send_cbs_to_online(void)
1108 * Because preemptible RCU does not exist, it need not be initialized.
1110 static void __init __rcu_init_preempt(void)
1114 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1116 #ifdef CONFIG_RCU_BOOST
1118 #include "rtmutex_common.h"
1120 #ifdef CONFIG_RCU_TRACE
1122 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1124 if (list_empty(&rnp->blkd_tasks))
1125 rnp->n_balk_blkd_tasks++;
1126 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1127 rnp->n_balk_exp_gp_tasks++;
1128 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1129 rnp->n_balk_boost_tasks++;
1130 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1131 rnp->n_balk_notblocked++;
1132 else if (rnp->gp_tasks != NULL &&
1133 ULONG_CMP_LT(jiffies, rnp->boost_time))
1134 rnp->n_balk_notyet++;
1139 #else /* #ifdef CONFIG_RCU_TRACE */
1141 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1145 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1148 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1149 * or ->boost_tasks, advancing the pointer to the next task in the
1150 * ->blkd_tasks list.
1152 * Note that irqs must be enabled: boosting the task can block.
1153 * Returns 1 if there are more tasks needing to be boosted.
1155 static int rcu_boost(struct rcu_node *rnp)
1157 unsigned long flags;
1158 struct rt_mutex mtx;
1159 struct task_struct *t;
1160 struct list_head *tb;
1162 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1163 return 0; /* Nothing left to boost. */
1165 raw_spin_lock_irqsave(&rnp->lock, flags);
1168 * Recheck under the lock: all tasks in need of boosting
1169 * might exit their RCU read-side critical sections on their own.
1171 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1172 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1177 * Preferentially boost tasks blocking expedited grace periods.
1178 * This cannot starve the normal grace periods because a second
1179 * expedited grace period must boost all blocked tasks, including
1180 * those blocking the pre-existing normal grace period.
1182 if (rnp->exp_tasks != NULL) {
1183 tb = rnp->exp_tasks;
1184 rnp->n_exp_boosts++;
1186 tb = rnp->boost_tasks;
1187 rnp->n_normal_boosts++;
1189 rnp->n_tasks_boosted++;
1192 * We boost task t by manufacturing an rt_mutex that appears to
1193 * be held by task t. We leave a pointer to that rt_mutex where
1194 * task t can find it, and task t will release the mutex when it
1195 * exits its outermost RCU read-side critical section. Then
1196 * simply acquiring this artificial rt_mutex will boost task
1197 * t's priority. (Thanks to tglx for suggesting this approach!)
1199 * Note that task t must acquire rnp->lock to remove itself from
1200 * the ->blkd_tasks list, which it will do from exit() if from
1201 * nowhere else. We therefore are guaranteed that task t will
1202 * stay around at least until we drop rnp->lock. Note that
1203 * rnp->lock also resolves races between our priority boosting
1204 * and task t's exiting its outermost RCU read-side critical
1207 t = container_of(tb, struct task_struct, rcu_node_entry);
1208 rt_mutex_init_proxy_locked(&mtx, t);
1209 t->rcu_boost_mutex = &mtx;
1210 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1211 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1212 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1214 return rnp->exp_tasks != NULL || rnp->boost_tasks != NULL;
1218 * Timer handler to initiate waking up of boost kthreads that
1219 * have yielded the CPU due to excessive numbers of tasks to
1220 * boost. We wake up the per-rcu_node kthread, which in turn
1221 * will wake up the booster kthread.
1223 static void rcu_boost_kthread_timer(unsigned long arg)
1225 invoke_rcu_node_kthread((struct rcu_node *)arg);
1229 * Priority-boosting kthread. One per leaf rcu_node and one for the
1232 static int rcu_boost_kthread(void *arg)
1234 struct rcu_node *rnp = (struct rcu_node *)arg;
1238 trace_rcu_utilization("Start boost kthread@init");
1240 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1241 trace_rcu_utilization("End boost kthread@rcu_wait");
1242 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1243 trace_rcu_utilization("Start boost kthread@rcu_wait");
1244 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1245 more2boost = rcu_boost(rnp);
1251 trace_rcu_utilization("End boost kthread@rcu_yield");
1252 rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp);
1253 trace_rcu_utilization("Start boost kthread@rcu_yield");
1258 trace_rcu_utilization("End boost kthread@notreached");
1263 * Check to see if it is time to start boosting RCU readers that are
1264 * blocking the current grace period, and, if so, tell the per-rcu_node
1265 * kthread to start boosting them. If there is an expedited grace
1266 * period in progress, it is always time to boost.
1268 * The caller must hold rnp->lock, which this function releases,
1269 * but irqs remain disabled. The ->boost_kthread_task is immortal,
1270 * so we don't need to worry about it going away.
1272 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1274 struct task_struct *t;
1276 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1277 rnp->n_balk_exp_gp_tasks++;
1278 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1281 if (rnp->exp_tasks != NULL ||
1282 (rnp->gp_tasks != NULL &&
1283 rnp->boost_tasks == NULL &&
1285 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1286 if (rnp->exp_tasks == NULL)
1287 rnp->boost_tasks = rnp->gp_tasks;
1288 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1289 t = rnp->boost_kthread_task;
1293 rcu_initiate_boost_trace(rnp);
1294 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1299 * Wake up the per-CPU kthread to invoke RCU callbacks.
1301 static void invoke_rcu_callbacks_kthread(void)
1303 unsigned long flags;
1305 local_irq_save(flags);
1306 __this_cpu_write(rcu_cpu_has_work, 1);
1307 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1308 current != __this_cpu_read(rcu_cpu_kthread_task))
1309 wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
1310 local_irq_restore(flags);
1314 * Set the affinity of the boost kthread. The CPU-hotplug locks are
1315 * held, so no one should be messing with the existence of the boost
1318 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
1321 struct task_struct *t;
1323 t = rnp->boost_kthread_task;
1325 set_cpus_allowed_ptr(rnp->boost_kthread_task, cm);
1328 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1331 * Do priority-boost accounting for the start of a new grace period.
1333 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1335 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1339 * Create an RCU-boost kthread for the specified node if one does not
1340 * already exist. We only create this kthread for preemptible RCU.
1341 * Returns zero if all is well, a negated errno otherwise.
1343 static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1344 struct rcu_node *rnp,
1347 unsigned long flags;
1348 struct sched_param sp;
1349 struct task_struct *t;
1351 if (&rcu_preempt_state != rsp)
1354 if (rnp->boost_kthread_task != NULL)
1356 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1357 "rcub%d", rnp_index);
1360 raw_spin_lock_irqsave(&rnp->lock, flags);
1361 rnp->boost_kthread_task = t;
1362 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1363 sp.sched_priority = RCU_KTHREAD_PRIO;
1364 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1365 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1369 #ifdef CONFIG_HOTPLUG_CPU
1372 * Stop the RCU's per-CPU kthread when its CPU goes offline,.
1374 static void rcu_stop_cpu_kthread(int cpu)
1376 struct task_struct *t;
1378 /* Stop the CPU's kthread. */
1379 t = per_cpu(rcu_cpu_kthread_task, cpu);
1381 per_cpu(rcu_cpu_kthread_task, cpu) = NULL;
1386 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1388 static void rcu_kthread_do_work(void)
1390 rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1391 rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1392 rcu_preempt_do_callbacks();
1396 * Wake up the specified per-rcu_node-structure kthread.
1397 * Because the per-rcu_node kthreads are immortal, we don't need
1398 * to do anything to keep them alive.
1400 static void invoke_rcu_node_kthread(struct rcu_node *rnp)
1402 struct task_struct *t;
1404 t = rnp->node_kthread_task;
1410 * Set the specified CPU's kthread to run RT or not, as specified by
1411 * the to_rt argument. The CPU-hotplug locks are held, so the task
1412 * is not going away.
1414 static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1417 struct sched_param sp;
1418 struct task_struct *t;
1420 t = per_cpu(rcu_cpu_kthread_task, cpu);
1424 policy = SCHED_FIFO;
1425 sp.sched_priority = RCU_KTHREAD_PRIO;
1427 policy = SCHED_NORMAL;
1428 sp.sched_priority = 0;
1430 sched_setscheduler_nocheck(t, policy, &sp);
1434 * Timer handler to initiate the waking up of per-CPU kthreads that
1435 * have yielded the CPU due to excess numbers of RCU callbacks.
1436 * We wake up the per-rcu_node kthread, which in turn will wake up
1437 * the booster kthread.
1439 static void rcu_cpu_kthread_timer(unsigned long arg)
1441 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
1442 struct rcu_node *rnp = rdp->mynode;
1444 atomic_or(rdp->grpmask, &rnp->wakemask);
1445 invoke_rcu_node_kthread(rnp);
1449 * Drop to non-real-time priority and yield, but only after posting a
1450 * timer that will cause us to regain our real-time priority if we
1451 * remain preempted. Either way, we restore our real-time priority
1454 static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
1456 struct sched_param sp;
1457 struct timer_list yield_timer;
1459 setup_timer_on_stack(&yield_timer, f, arg);
1460 mod_timer(&yield_timer, jiffies + 2);
1461 sp.sched_priority = 0;
1462 sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
1463 set_user_nice(current, 19);
1465 sp.sched_priority = RCU_KTHREAD_PRIO;
1466 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1467 del_timer(&yield_timer);
1471 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1472 * This can happen while the corresponding CPU is either coming online
1473 * or going offline. We cannot wait until the CPU is fully online
1474 * before starting the kthread, because the various notifier functions
1475 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1476 * the corresponding CPU is online.
1478 * Return 1 if the kthread needs to stop, 0 otherwise.
1480 * Caller must disable bh. This function can momentarily enable it.
1482 static int rcu_cpu_kthread_should_stop(int cpu)
1484 while (cpu_is_offline(cpu) ||
1485 !cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)) ||
1486 smp_processor_id() != cpu) {
1487 if (kthread_should_stop())
1489 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1490 per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id();
1492 schedule_timeout_uninterruptible(1);
1493 if (!cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)))
1494 set_cpus_allowed_ptr(current, cpumask_of(cpu));
1497 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1502 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1503 * RCU softirq used in flavors and configurations of RCU that do not
1504 * support RCU priority boosting.
1506 static int rcu_cpu_kthread(void *arg)
1508 int cpu = (int)(long)arg;
1509 unsigned long flags;
1511 unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu);
1513 char *workp = &per_cpu(rcu_cpu_has_work, cpu);
1515 trace_rcu_utilization("Start CPU kthread@init");
1517 *statusp = RCU_KTHREAD_WAITING;
1518 trace_rcu_utilization("End CPU kthread@rcu_wait");
1519 rcu_wait(*workp != 0 || kthread_should_stop());
1520 trace_rcu_utilization("Start CPU kthread@rcu_wait");
1522 if (rcu_cpu_kthread_should_stop(cpu)) {
1526 *statusp = RCU_KTHREAD_RUNNING;
1527 per_cpu(rcu_cpu_kthread_loops, cpu)++;
1528 local_irq_save(flags);
1531 local_irq_restore(flags);
1533 rcu_kthread_do_work();
1540 *statusp = RCU_KTHREAD_YIELDING;
1541 trace_rcu_utilization("End CPU kthread@rcu_yield");
1542 rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
1543 trace_rcu_utilization("Start CPU kthread@rcu_yield");
1547 *statusp = RCU_KTHREAD_STOPPED;
1548 trace_rcu_utilization("End CPU kthread@term");
1553 * Spawn a per-CPU kthread, setting up affinity and priority.
1554 * Because the CPU hotplug lock is held, no other CPU will be attempting
1555 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1556 * attempting to access it during boot, but the locking in kthread_bind()
1557 * will enforce sufficient ordering.
1559 * Please note that we cannot simply refuse to wake up the per-CPU
1560 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1561 * which can result in softlockup complaints if the task ends up being
1562 * idle for more than a couple of minutes.
1564 * However, please note also that we cannot bind the per-CPU kthread to its
1565 * CPU until that CPU is fully online. We also cannot wait until the
1566 * CPU is fully online before we create its per-CPU kthread, as this would
1567 * deadlock the system when CPU notifiers tried waiting for grace
1568 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1569 * is online. If its CPU is not yet fully online, then the code in
1570 * rcu_cpu_kthread() will wait until it is fully online, and then do
1573 static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu)
1575 struct sched_param sp;
1576 struct task_struct *t;
1578 if (!rcu_scheduler_fully_active ||
1579 per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
1581 t = kthread_create_on_node(rcu_cpu_kthread,
1587 if (cpu_online(cpu))
1588 kthread_bind(t, cpu);
1589 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1590 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL);
1591 sp.sched_priority = RCU_KTHREAD_PRIO;
1592 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1593 per_cpu(rcu_cpu_kthread_task, cpu) = t;
1594 wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */
1599 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1600 * kthreads when needed. We ignore requests to wake up kthreads
1601 * for offline CPUs, which is OK because force_quiescent_state()
1602 * takes care of this case.
1604 static int rcu_node_kthread(void *arg)
1607 unsigned long flags;
1609 struct rcu_node *rnp = (struct rcu_node *)arg;
1610 struct sched_param sp;
1611 struct task_struct *t;
1614 rnp->node_kthread_status = RCU_KTHREAD_WAITING;
1615 rcu_wait(atomic_read(&rnp->wakemask) != 0);
1616 rnp->node_kthread_status = RCU_KTHREAD_RUNNING;
1617 raw_spin_lock_irqsave(&rnp->lock, flags);
1618 mask = atomic_xchg(&rnp->wakemask, 0);
1619 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1620 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
1621 if ((mask & 0x1) == 0)
1624 t = per_cpu(rcu_cpu_kthread_task, cpu);
1625 if (!cpu_online(cpu) || t == NULL) {
1629 per_cpu(rcu_cpu_has_work, cpu) = 1;
1630 sp.sched_priority = RCU_KTHREAD_PRIO;
1631 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1636 rnp->node_kthread_status = RCU_KTHREAD_STOPPED;
1641 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1642 * served by the rcu_node in question. The CPU hotplug lock is still
1643 * held, so the value of rnp->qsmaskinit will be stable.
1645 * We don't include outgoingcpu in the affinity set, use -1 if there is
1646 * no outgoing CPU. If there are no CPUs left in the affinity set,
1647 * this function allows the kthread to execute on any CPU.
1649 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1653 unsigned long mask = rnp->qsmaskinit;
1655 if (rnp->node_kthread_task == NULL)
1657 if (!alloc_cpumask_var(&cm, GFP_KERNEL))
1660 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1661 if ((mask & 0x1) && cpu != outgoingcpu)
1662 cpumask_set_cpu(cpu, cm);
1663 if (cpumask_weight(cm) == 0) {
1665 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1666 cpumask_clear_cpu(cpu, cm);
1667 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1669 set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
1670 rcu_boost_kthread_setaffinity(rnp, cm);
1671 free_cpumask_var(cm);
1675 * Spawn a per-rcu_node kthread, setting priority and affinity.
1676 * Called during boot before online/offline can happen, or, if
1677 * during runtime, with the main CPU-hotplug locks held. So only
1678 * one of these can be executing at a time.
1680 static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
1681 struct rcu_node *rnp)
1683 unsigned long flags;
1684 int rnp_index = rnp - &rsp->node[0];
1685 struct sched_param sp;
1686 struct task_struct *t;
1688 if (!rcu_scheduler_fully_active ||
1689 rnp->qsmaskinit == 0)
1691 if (rnp->node_kthread_task == NULL) {
1692 t = kthread_create(rcu_node_kthread, (void *)rnp,
1693 "rcun%d", rnp_index);
1696 raw_spin_lock_irqsave(&rnp->lock, flags);
1697 rnp->node_kthread_task = t;
1698 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1699 sp.sched_priority = 99;
1700 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1701 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1703 return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
1707 * Spawn all kthreads -- called as soon as the scheduler is running.
1709 static int __init rcu_spawn_kthreads(void)
1712 struct rcu_node *rnp;
1714 rcu_scheduler_fully_active = 1;
1715 for_each_possible_cpu(cpu) {
1716 per_cpu(rcu_cpu_has_work, cpu) = 0;
1717 if (cpu_online(cpu))
1718 (void)rcu_spawn_one_cpu_kthread(cpu);
1720 rnp = rcu_get_root(rcu_state);
1721 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1722 if (NUM_RCU_NODES > 1) {
1723 rcu_for_each_leaf_node(rcu_state, rnp)
1724 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1728 early_initcall(rcu_spawn_kthreads);
1730 static void __cpuinit rcu_prepare_kthreads(int cpu)
1732 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1733 struct rcu_node *rnp = rdp->mynode;
1735 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1736 if (rcu_scheduler_fully_active) {
1737 (void)rcu_spawn_one_cpu_kthread(cpu);
1738 if (rnp->node_kthread_task == NULL)
1739 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1743 #else /* #ifdef CONFIG_RCU_BOOST */
1745 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1747 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1750 static void invoke_rcu_callbacks_kthread(void)
1755 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1759 #ifdef CONFIG_HOTPLUG_CPU
1761 static void rcu_stop_cpu_kthread(int cpu)
1765 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1767 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1771 static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1775 static int __init rcu_scheduler_really_started(void)
1777 rcu_scheduler_fully_active = 1;
1780 early_initcall(rcu_scheduler_really_started);
1782 static void __cpuinit rcu_prepare_kthreads(int cpu)
1786 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1790 void synchronize_sched_expedited(void)
1794 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
1796 #else /* #ifndef CONFIG_SMP */
1798 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
1799 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
1801 static int synchronize_sched_expedited_cpu_stop(void *data)
1804 * There must be a full memory barrier on each affected CPU
1805 * between the time that try_stop_cpus() is called and the
1806 * time that it returns.
1808 * In the current initial implementation of cpu_stop, the
1809 * above condition is already met when the control reaches
1810 * this point and the following smp_mb() is not strictly
1811 * necessary. Do smp_mb() anyway for documentation and
1812 * robustness against future implementation changes.
1814 smp_mb(); /* See above comment block. */
1819 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
1820 * approach to force grace period to end quickly. This consumes
1821 * significant time on all CPUs, and is thus not recommended for
1822 * any sort of common-case code.
1824 * Note that it is illegal to call this function while holding any
1825 * lock that is acquired by a CPU-hotplug notifier. Failing to
1826 * observe this restriction will result in deadlock.
1828 * This implementation can be thought of as an application of ticket
1829 * locking to RCU, with sync_sched_expedited_started and
1830 * sync_sched_expedited_done taking on the roles of the halves
1831 * of the ticket-lock word. Each task atomically increments
1832 * sync_sched_expedited_started upon entry, snapshotting the old value,
1833 * then attempts to stop all the CPUs. If this succeeds, then each
1834 * CPU will have executed a context switch, resulting in an RCU-sched
1835 * grace period. We are then done, so we use atomic_cmpxchg() to
1836 * update sync_sched_expedited_done to match our snapshot -- but
1837 * only if someone else has not already advanced past our snapshot.
1839 * On the other hand, if try_stop_cpus() fails, we check the value
1840 * of sync_sched_expedited_done. If it has advanced past our
1841 * initial snapshot, then someone else must have forced a grace period
1842 * some time after we took our snapshot. In this case, our work is
1843 * done for us, and we can simply return. Otherwise, we try again,
1844 * but keep our initial snapshot for purposes of checking for someone
1845 * doing our work for us.
1847 * If we fail too many times in a row, we fall back to synchronize_sched().
1849 void synchronize_sched_expedited(void)
1851 int firstsnap, s, snap, trycount = 0;
1853 /* Note that atomic_inc_return() implies full memory barrier. */
1854 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
1858 * Each pass through the following loop attempts to force a
1859 * context switch on each CPU.
1861 while (try_stop_cpus(cpu_online_mask,
1862 synchronize_sched_expedited_cpu_stop,
1866 /* No joy, try again later. Or just synchronize_sched(). */
1867 if (trycount++ < 10)
1868 udelay(trycount * num_online_cpus());
1870 synchronize_sched();
1874 /* Check to see if someone else did our work for us. */
1875 s = atomic_read(&sync_sched_expedited_done);
1876 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
1877 smp_mb(); /* ensure test happens before caller kfree */
1882 * Refetching sync_sched_expedited_started allows later
1883 * callers to piggyback on our grace period. We subtract
1884 * 1 to get the same token that the last incrementer got.
1885 * We retry after they started, so our grace period works
1886 * for them, and they started after our first try, so their
1887 * grace period works for us.
1890 snap = atomic_read(&sync_sched_expedited_started) - 1;
1891 smp_mb(); /* ensure read is before try_stop_cpus(). */
1895 * Everyone up to our most recent fetch is covered by our grace
1896 * period. Update the counter, but only if our work is still
1897 * relevant -- which it won't be if someone who started later
1898 * than we did beat us to the punch.
1901 s = atomic_read(&sync_sched_expedited_done);
1902 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
1903 smp_mb(); /* ensure test happens before caller kfree */
1906 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
1910 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
1912 #endif /* #else #ifndef CONFIG_SMP */
1914 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1917 * Check to see if any future RCU-related work will need to be done
1918 * by the current CPU, even if none need be done immediately, returning
1919 * 1 if so. This function is part of the RCU implementation; it is -not-
1920 * an exported member of the RCU API.
1922 * Because we have preemptible RCU, just check whether this CPU needs
1923 * any flavor of RCU. Do not chew up lots of CPU cycles with preemption
1924 * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
1926 int rcu_needs_cpu(int cpu)
1928 return rcu_needs_cpu_quick_check(cpu);
1932 * Check to see if we need to continue a callback-flush operations to
1933 * allow the last CPU to enter dyntick-idle mode. But fast dyntick-idle
1934 * entry is not configured, so we never do need to.
1936 static void rcu_needs_cpu_flush(void)
1940 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1942 #define RCU_NEEDS_CPU_FLUSHES 5
1943 static DEFINE_PER_CPU(int, rcu_dyntick_drain);
1944 static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff);
1947 * Check to see if any future RCU-related work will need to be done
1948 * by the current CPU, even if none need be done immediately, returning
1949 * 1 if so. This function is part of the RCU implementation; it is -not-
1950 * an exported member of the RCU API.
1952 * Because we are not supporting preemptible RCU, attempt to accelerate
1953 * any current grace periods so that RCU no longer needs this CPU, but
1954 * only if all other CPUs are already in dynticks-idle mode. This will
1955 * allow the CPU cores to be powered down immediately, as opposed to after
1956 * waiting many milliseconds for grace periods to elapse.
1958 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1959 * disabled, we do one pass of force_quiescent_state(), then do a
1960 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
1961 * later. The per-cpu rcu_dyntick_drain variable controls the sequencing.
1963 int rcu_needs_cpu(int cpu)
1969 /* Check for being in the holdoff period. */
1970 if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies)
1971 return rcu_needs_cpu_quick_check(cpu);
1973 /* Don't bother unless we are the last non-dyntick-idle CPU. */
1974 for_each_online_cpu(thatcpu) {
1977 snap = atomic_add_return(0, &per_cpu(rcu_dynticks,
1979 smp_mb(); /* Order sampling of snap with end of grace period. */
1980 if ((snap & 0x1) != 0) {
1981 per_cpu(rcu_dyntick_drain, cpu) = 0;
1982 per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1;
1983 return rcu_needs_cpu_quick_check(cpu);
1987 /* Check and update the rcu_dyntick_drain sequencing. */
1988 if (per_cpu(rcu_dyntick_drain, cpu) <= 0) {
1989 /* First time through, initialize the counter. */
1990 per_cpu(rcu_dyntick_drain, cpu) = RCU_NEEDS_CPU_FLUSHES;
1991 } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) {
1992 /* We have hit the limit, so time to give up. */
1993 per_cpu(rcu_dyntick_holdoff, cpu) = jiffies;
1994 return rcu_needs_cpu_quick_check(cpu);
1997 /* Do one step pushing remaining RCU callbacks through. */
1998 if (per_cpu(rcu_sched_data, cpu).nxtlist) {
2000 force_quiescent_state(&rcu_sched_state, 0);
2001 c = c || per_cpu(rcu_sched_data, cpu).nxtlist;
2003 if (per_cpu(rcu_bh_data, cpu).nxtlist) {
2005 force_quiescent_state(&rcu_bh_state, 0);
2006 c = c || per_cpu(rcu_bh_data, cpu).nxtlist;
2009 /* If RCU callbacks are still pending, RCU still needs this CPU. */
2016 * Check to see if we need to continue a callback-flush operations to
2017 * allow the last CPU to enter dyntick-idle mode.
2019 static void rcu_needs_cpu_flush(void)
2021 int cpu = smp_processor_id();
2022 unsigned long flags;
2024 if (per_cpu(rcu_dyntick_drain, cpu) <= 0)
2026 local_irq_save(flags);
2027 (void)rcu_needs_cpu(cpu);
2028 local_irq_restore(flags);
2031 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */