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>
29 #define RCU_KTHREAD_PRIO 1
31 #ifdef CONFIG_RCU_BOOST
32 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
34 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
38 * Check the RCU kernel configuration parameters and print informative
39 * messages about anything out of the ordinary. If you like #ifdef, you
40 * will love this function.
42 static void __init rcu_bootup_announce_oddness(void)
44 #ifdef CONFIG_RCU_TRACE
45 printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n");
47 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
48 printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
51 #ifdef CONFIG_RCU_FANOUT_EXACT
52 printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n");
54 #ifdef CONFIG_RCU_FAST_NO_HZ
56 "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
58 #ifdef CONFIG_PROVE_RCU
59 printk(KERN_INFO "\tRCU lockdep checking is enabled.\n");
61 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
62 printk(KERN_INFO "\tRCU torture testing starts during boot.\n");
64 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
65 printk(KERN_INFO "\tDump stacks of tasks blocking RCU-preempt GP.\n");
67 #if defined(CONFIG_RCU_CPU_STALL_INFO)
68 printk(KERN_INFO "\tAdditional per-CPU info printed with stalls.\n");
70 #if NUM_RCU_LVL_4 != 0
71 printk(KERN_INFO "\tFour-level hierarchy is enabled.\n");
73 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
74 printk(KERN_INFO "\tExperimental boot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
75 if (nr_cpu_ids != NR_CPUS)
76 printk(KERN_INFO "\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
79 #ifdef CONFIG_TREE_PREEMPT_RCU
81 struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt);
82 DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
83 static struct rcu_state *rcu_state = &rcu_preempt_state;
85 static void rcu_read_unlock_special(struct task_struct *t);
86 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
89 * Tell them what RCU they are running.
91 static void __init rcu_bootup_announce(void)
93 printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n");
94 rcu_bootup_announce_oddness();
98 * Return the number of RCU-preempt batches processed thus far
99 * for debug and statistics.
101 long rcu_batches_completed_preempt(void)
103 return rcu_preempt_state.completed;
105 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
108 * Return the number of RCU batches processed thus far for debug & stats.
110 long rcu_batches_completed(void)
112 return rcu_batches_completed_preempt();
114 EXPORT_SYMBOL_GPL(rcu_batches_completed);
117 * Force a quiescent state for preemptible RCU.
119 void rcu_force_quiescent_state(void)
121 force_quiescent_state(&rcu_preempt_state, 0);
123 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
126 * Record a preemptible-RCU quiescent state for the specified CPU. Note
127 * that this just means that the task currently running on the CPU is
128 * not in a quiescent state. There might be any number of tasks blocked
129 * while in an RCU read-side critical section.
131 * Unlike the other rcu_*_qs() functions, callers to this function
132 * must disable irqs in order to protect the assignment to
133 * ->rcu_read_unlock_special.
135 static void rcu_preempt_qs(int cpu)
137 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
139 rdp->passed_quiesce_gpnum = rdp->gpnum;
141 if (rdp->passed_quiesce == 0)
142 trace_rcu_grace_period("rcu_preempt", rdp->gpnum, "cpuqs");
143 rdp->passed_quiesce = 1;
144 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
148 * We have entered the scheduler, and the current task might soon be
149 * context-switched away from. If this task is in an RCU read-side
150 * critical section, we will no longer be able to rely on the CPU to
151 * record that fact, so we enqueue the task on the blkd_tasks list.
152 * The task will dequeue itself when it exits the outermost enclosing
153 * RCU read-side critical section. Therefore, the current grace period
154 * cannot be permitted to complete until the blkd_tasks list entries
155 * predating the current grace period drain, in other words, until
156 * rnp->gp_tasks becomes NULL.
158 * Caller must disable preemption.
160 static void rcu_preempt_note_context_switch(int cpu)
162 struct task_struct *t = current;
164 struct rcu_data *rdp;
165 struct rcu_node *rnp;
167 if (t->rcu_read_lock_nesting > 0 &&
168 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
170 /* Possibly blocking in an RCU read-side critical section. */
171 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
173 raw_spin_lock_irqsave(&rnp->lock, flags);
174 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
175 t->rcu_blocked_node = rnp;
178 * If this CPU has already checked in, then this task
179 * will hold up the next grace period rather than the
180 * current grace period. Queue the task accordingly.
181 * If the task is queued for the current grace period
182 * (i.e., this CPU has not yet passed through a quiescent
183 * state for the current grace period), then as long
184 * as that task remains queued, the current grace period
185 * cannot end. Note that there is some uncertainty as
186 * to exactly when the current grace period started.
187 * We take a conservative approach, which can result
188 * in unnecessarily waiting on tasks that started very
189 * slightly after the current grace period began. C'est
192 * But first, note that the current CPU must still be
195 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
196 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
197 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
198 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
199 rnp->gp_tasks = &t->rcu_node_entry;
200 #ifdef CONFIG_RCU_BOOST
201 if (rnp->boost_tasks != NULL)
202 rnp->boost_tasks = rnp->gp_tasks;
203 #endif /* #ifdef CONFIG_RCU_BOOST */
205 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
206 if (rnp->qsmask & rdp->grpmask)
207 rnp->gp_tasks = &t->rcu_node_entry;
209 trace_rcu_preempt_task(rdp->rsp->name,
211 (rnp->qsmask & rdp->grpmask)
214 raw_spin_unlock_irqrestore(&rnp->lock, flags);
215 } else if (t->rcu_read_lock_nesting < 0 &&
216 t->rcu_read_unlock_special) {
219 * Complete exit from RCU read-side critical section on
220 * behalf of preempted instance of __rcu_read_unlock().
222 rcu_read_unlock_special(t);
226 * Either we were not in an RCU read-side critical section to
227 * begin with, or we have now recorded that critical section
228 * globally. Either way, we can now note a quiescent state
229 * for this CPU. Again, if we were in an RCU read-side critical
230 * section, and if that critical section was blocking the current
231 * grace period, then the fact that the task has been enqueued
232 * means that we continue to block the current grace period.
234 local_irq_save(flags);
236 local_irq_restore(flags);
240 * Tree-preemptible RCU implementation for rcu_read_lock().
241 * Just increment ->rcu_read_lock_nesting, shared state will be updated
244 void __rcu_read_lock(void)
246 current->rcu_read_lock_nesting++;
247 barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
249 EXPORT_SYMBOL_GPL(__rcu_read_lock);
252 * Check for preempted RCU readers blocking the current grace period
253 * for the specified rcu_node structure. If the caller needs a reliable
254 * answer, it must hold the rcu_node's ->lock.
256 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
258 return rnp->gp_tasks != NULL;
262 * Record a quiescent state for all tasks that were previously queued
263 * on the specified rcu_node structure and that were blocking the current
264 * RCU grace period. The caller must hold the specified rnp->lock with
265 * irqs disabled, and this lock is released upon return, but irqs remain
268 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
269 __releases(rnp->lock)
272 struct rcu_node *rnp_p;
274 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
275 raw_spin_unlock_irqrestore(&rnp->lock, flags);
276 return; /* Still need more quiescent states! */
282 * Either there is only one rcu_node in the tree,
283 * or tasks were kicked up to root rcu_node due to
284 * CPUs going offline.
286 rcu_report_qs_rsp(&rcu_preempt_state, flags);
290 /* Report up the rest of the hierarchy. */
292 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
293 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
294 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
298 * Advance a ->blkd_tasks-list pointer to the next entry, instead
299 * returning NULL if at the end of the list.
301 static struct list_head *rcu_next_node_entry(struct task_struct *t,
302 struct rcu_node *rnp)
304 struct list_head *np;
306 np = t->rcu_node_entry.next;
307 if (np == &rnp->blkd_tasks)
313 * Handle special cases during rcu_read_unlock(), such as needing to
314 * notify RCU core processing or task having blocked during the RCU
315 * read-side critical section.
317 static noinline void rcu_read_unlock_special(struct task_struct *t)
323 struct list_head *np;
324 #ifdef CONFIG_RCU_BOOST
325 struct rt_mutex *rbmp = NULL;
326 #endif /* #ifdef CONFIG_RCU_BOOST */
327 struct rcu_node *rnp;
330 /* NMI handlers cannot block and cannot safely manipulate state. */
334 local_irq_save(flags);
337 * If RCU core is waiting for this CPU to exit critical section,
338 * let it know that we have done so.
340 special = t->rcu_read_unlock_special;
341 if (special & RCU_READ_UNLOCK_NEED_QS) {
342 rcu_preempt_qs(smp_processor_id());
345 /* Hardware IRQ handlers cannot block. */
346 if (in_irq() || in_serving_softirq()) {
347 local_irq_restore(flags);
351 /* Clean up if blocked during RCU read-side critical section. */
352 if (special & RCU_READ_UNLOCK_BLOCKED) {
353 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
356 * Remove this task from the list it blocked on. The
357 * task can migrate while we acquire the lock, but at
358 * most one time. So at most two passes through loop.
361 rnp = t->rcu_blocked_node;
362 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
363 if (rnp == t->rcu_blocked_node)
365 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
367 empty = !rcu_preempt_blocked_readers_cgp(rnp);
368 empty_exp = !rcu_preempted_readers_exp(rnp);
369 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
370 np = rcu_next_node_entry(t, rnp);
371 list_del_init(&t->rcu_node_entry);
372 t->rcu_blocked_node = NULL;
373 trace_rcu_unlock_preempted_task("rcu_preempt",
375 if (&t->rcu_node_entry == rnp->gp_tasks)
377 if (&t->rcu_node_entry == rnp->exp_tasks)
379 #ifdef CONFIG_RCU_BOOST
380 if (&t->rcu_node_entry == rnp->boost_tasks)
381 rnp->boost_tasks = np;
382 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
383 if (t->rcu_boost_mutex) {
384 rbmp = t->rcu_boost_mutex;
385 t->rcu_boost_mutex = NULL;
387 #endif /* #ifdef CONFIG_RCU_BOOST */
390 * If this was the last task on the current list, and if
391 * we aren't waiting on any CPUs, report the quiescent state.
392 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
393 * so we must take a snapshot of the expedited state.
395 empty_exp_now = !rcu_preempted_readers_exp(rnp);
396 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
397 trace_rcu_quiescent_state_report("preempt_rcu",
404 rcu_report_unblock_qs_rnp(rnp, flags);
406 raw_spin_unlock_irqrestore(&rnp->lock, flags);
408 #ifdef CONFIG_RCU_BOOST
409 /* Unboost if we were boosted. */
411 rt_mutex_unlock(rbmp);
412 #endif /* #ifdef CONFIG_RCU_BOOST */
415 * If this was the last task on the expedited lists,
416 * then we need to report up the rcu_node hierarchy.
418 if (!empty_exp && empty_exp_now)
419 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
421 local_irq_restore(flags);
426 * Tree-preemptible RCU implementation for rcu_read_unlock().
427 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
428 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
429 * invoke rcu_read_unlock_special() to clean up after a context switch
430 * in an RCU read-side critical section and other special cases.
432 void __rcu_read_unlock(void)
434 struct task_struct *t = current;
436 if (t->rcu_read_lock_nesting != 1)
437 --t->rcu_read_lock_nesting;
439 barrier(); /* critical section before exit code. */
440 t->rcu_read_lock_nesting = INT_MIN;
441 barrier(); /* assign before ->rcu_read_unlock_special load */
442 if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
443 rcu_read_unlock_special(t);
444 barrier(); /* ->rcu_read_unlock_special load before assign */
445 t->rcu_read_lock_nesting = 0;
447 #ifdef CONFIG_PROVE_LOCKING
449 int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting);
451 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
453 #endif /* #ifdef CONFIG_PROVE_LOCKING */
455 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
457 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
460 * Dump detailed information for all tasks blocking the current RCU
461 * grace period on the specified rcu_node structure.
463 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
466 struct task_struct *t;
468 if (!rcu_preempt_blocked_readers_cgp(rnp))
470 raw_spin_lock_irqsave(&rnp->lock, flags);
471 t = list_entry(rnp->gp_tasks,
472 struct task_struct, rcu_node_entry);
473 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
475 raw_spin_unlock_irqrestore(&rnp->lock, flags);
479 * Dump detailed information for all tasks blocking the current RCU
482 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
484 struct rcu_node *rnp = rcu_get_root(rsp);
486 rcu_print_detail_task_stall_rnp(rnp);
487 rcu_for_each_leaf_node(rsp, rnp)
488 rcu_print_detail_task_stall_rnp(rnp);
491 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
493 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
497 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
499 #ifdef CONFIG_RCU_CPU_STALL_INFO
501 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
503 printk(KERN_ERR "\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
504 rnp->level, rnp->grplo, rnp->grphi);
507 static void rcu_print_task_stall_end(void)
509 printk(KERN_CONT "\n");
512 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
514 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
518 static void rcu_print_task_stall_end(void)
522 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
525 * Scan the current list of tasks blocked within RCU read-side critical
526 * sections, printing out the tid of each.
528 static int rcu_print_task_stall(struct rcu_node *rnp)
530 struct task_struct *t;
533 if (!rcu_preempt_blocked_readers_cgp(rnp))
535 rcu_print_task_stall_begin(rnp);
536 t = list_entry(rnp->gp_tasks,
537 struct task_struct, rcu_node_entry);
538 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
539 printk(KERN_CONT " P%d", t->pid);
542 rcu_print_task_stall_end();
547 * Suppress preemptible RCU's CPU stall warnings by pushing the
548 * time of the next stall-warning message comfortably far into the
551 static void rcu_preempt_stall_reset(void)
553 rcu_preempt_state.jiffies_stall = jiffies + ULONG_MAX / 2;
557 * Check that the list of blocked tasks for the newly completed grace
558 * period is in fact empty. It is a serious bug to complete a grace
559 * period that still has RCU readers blocked! This function must be
560 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
561 * must be held by the caller.
563 * Also, if there are blocked tasks on the list, they automatically
564 * block the newly created grace period, so set up ->gp_tasks accordingly.
566 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
568 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
569 if (!list_empty(&rnp->blkd_tasks))
570 rnp->gp_tasks = rnp->blkd_tasks.next;
571 WARN_ON_ONCE(rnp->qsmask);
574 #ifdef CONFIG_HOTPLUG_CPU
577 * Handle tasklist migration for case in which all CPUs covered by the
578 * specified rcu_node have gone offline. Move them up to the root
579 * rcu_node. The reason for not just moving them to the immediate
580 * parent is to remove the need for rcu_read_unlock_special() to
581 * make more than two attempts to acquire the target rcu_node's lock.
582 * Returns true if there were tasks blocking the current RCU grace
585 * Returns 1 if there was previously a task blocking the current grace
586 * period on the specified rcu_node structure.
588 * The caller must hold rnp->lock with irqs disabled.
590 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
591 struct rcu_node *rnp,
592 struct rcu_data *rdp)
594 struct list_head *lp;
595 struct list_head *lp_root;
597 struct rcu_node *rnp_root = rcu_get_root(rsp);
598 struct task_struct *t;
600 if (rnp == rnp_root) {
601 WARN_ONCE(1, "Last CPU thought to be offlined?");
602 return 0; /* Shouldn't happen: at least one CPU online. */
605 /* If we are on an internal node, complain bitterly. */
606 WARN_ON_ONCE(rnp != rdp->mynode);
609 * Move tasks up to root rcu_node. Don't try to get fancy for
610 * this corner-case operation -- just put this node's tasks
611 * at the head of the root node's list, and update the root node's
612 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
613 * if non-NULL. This might result in waiting for more tasks than
614 * absolutely necessary, but this is a good performance/complexity
617 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
618 retval |= RCU_OFL_TASKS_NORM_GP;
619 if (rcu_preempted_readers_exp(rnp))
620 retval |= RCU_OFL_TASKS_EXP_GP;
621 lp = &rnp->blkd_tasks;
622 lp_root = &rnp_root->blkd_tasks;
623 while (!list_empty(lp)) {
624 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
625 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
626 list_del(&t->rcu_node_entry);
627 t->rcu_blocked_node = rnp_root;
628 list_add(&t->rcu_node_entry, lp_root);
629 if (&t->rcu_node_entry == rnp->gp_tasks)
630 rnp_root->gp_tasks = rnp->gp_tasks;
631 if (&t->rcu_node_entry == rnp->exp_tasks)
632 rnp_root->exp_tasks = rnp->exp_tasks;
633 #ifdef CONFIG_RCU_BOOST
634 if (&t->rcu_node_entry == rnp->boost_tasks)
635 rnp_root->boost_tasks = rnp->boost_tasks;
636 #endif /* #ifdef CONFIG_RCU_BOOST */
637 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
640 #ifdef CONFIG_RCU_BOOST
641 /* In case root is being boosted and leaf is not. */
642 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
643 if (rnp_root->boost_tasks != NULL &&
644 rnp_root->boost_tasks != rnp_root->gp_tasks)
645 rnp_root->boost_tasks = rnp_root->gp_tasks;
646 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
647 #endif /* #ifdef CONFIG_RCU_BOOST */
649 rnp->gp_tasks = NULL;
650 rnp->exp_tasks = NULL;
654 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
657 * Do CPU-offline processing for preemptible RCU.
659 static void rcu_preempt_cleanup_dead_cpu(int cpu)
661 rcu_cleanup_dead_cpu(cpu, &rcu_preempt_state);
665 * Check for a quiescent state from the current CPU. When a task blocks,
666 * the task is recorded in the corresponding CPU's rcu_node structure,
667 * which is checked elsewhere.
669 * Caller must disable hard irqs.
671 static void rcu_preempt_check_callbacks(int cpu)
673 struct task_struct *t = current;
675 if (t->rcu_read_lock_nesting == 0) {
679 if (t->rcu_read_lock_nesting > 0 &&
680 per_cpu(rcu_preempt_data, cpu).qs_pending)
681 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
685 * Process callbacks for preemptible RCU.
687 static void rcu_preempt_process_callbacks(void)
689 __rcu_process_callbacks(&rcu_preempt_state,
690 &__get_cpu_var(rcu_preempt_data));
693 #ifdef CONFIG_RCU_BOOST
695 static void rcu_preempt_do_callbacks(void)
697 rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
700 #endif /* #ifdef CONFIG_RCU_BOOST */
703 * Queue a preemptible-RCU callback for invocation after a grace period.
705 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
707 __call_rcu(head, func, &rcu_preempt_state, 0);
709 EXPORT_SYMBOL_GPL(call_rcu);
712 * Queue an RCU callback for lazy invocation after a grace period.
713 * This will likely be later named something like "call_rcu_lazy()",
714 * but this change will require some way of tagging the lazy RCU
715 * callbacks in the list of pending callbacks. Until then, this
716 * function may only be called from __kfree_rcu().
718 void kfree_call_rcu(struct rcu_head *head,
719 void (*func)(struct rcu_head *rcu))
721 __call_rcu(head, func, &rcu_preempt_state, 1);
723 EXPORT_SYMBOL_GPL(kfree_call_rcu);
726 * synchronize_rcu - wait until a grace period has elapsed.
728 * Control will return to the caller some time after a full grace
729 * period has elapsed, in other words after all currently executing RCU
730 * read-side critical sections have completed. Note, however, that
731 * upon return from synchronize_rcu(), the caller might well be executing
732 * concurrently with new RCU read-side critical sections that began while
733 * synchronize_rcu() was waiting. RCU read-side critical sections are
734 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
736 void synchronize_rcu(void)
738 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
739 !lock_is_held(&rcu_lock_map) &&
740 !lock_is_held(&rcu_sched_lock_map),
741 "Illegal synchronize_rcu() in RCU read-side critical section");
742 if (!rcu_scheduler_active)
744 wait_rcu_gp(call_rcu);
746 EXPORT_SYMBOL_GPL(synchronize_rcu);
748 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
749 static long sync_rcu_preempt_exp_count;
750 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
753 * Return non-zero if there are any tasks in RCU read-side critical
754 * sections blocking the current preemptible-RCU expedited grace period.
755 * If there is no preemptible-RCU expedited grace period currently in
756 * progress, returns zero unconditionally.
758 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
760 return rnp->exp_tasks != NULL;
764 * return non-zero if there is no RCU expedited grace period in progress
765 * for the specified rcu_node structure, in other words, if all CPUs and
766 * tasks covered by the specified rcu_node structure have done their bit
767 * for the current expedited grace period. Works only for preemptible
768 * RCU -- other RCU implementation use other means.
770 * Caller must hold sync_rcu_preempt_exp_mutex.
772 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
774 return !rcu_preempted_readers_exp(rnp) &&
775 ACCESS_ONCE(rnp->expmask) == 0;
779 * Report the exit from RCU read-side critical section for the last task
780 * that queued itself during or before the current expedited preemptible-RCU
781 * grace period. This event is reported either to the rcu_node structure on
782 * which the task was queued or to one of that rcu_node structure's ancestors,
783 * recursively up the tree. (Calm down, calm down, we do the recursion
786 * Most callers will set the "wake" flag, but the task initiating the
787 * expedited grace period need not wake itself.
789 * Caller must hold sync_rcu_preempt_exp_mutex.
791 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
797 raw_spin_lock_irqsave(&rnp->lock, flags);
799 if (!sync_rcu_preempt_exp_done(rnp)) {
800 raw_spin_unlock_irqrestore(&rnp->lock, flags);
803 if (rnp->parent == NULL) {
804 raw_spin_unlock_irqrestore(&rnp->lock, flags);
806 wake_up(&sync_rcu_preempt_exp_wq);
810 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
812 raw_spin_lock(&rnp->lock); /* irqs already disabled */
813 rnp->expmask &= ~mask;
818 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
819 * grace period for the specified rcu_node structure. If there are no such
820 * tasks, report it up the rcu_node hierarchy.
822 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
825 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
830 raw_spin_lock_irqsave(&rnp->lock, flags);
831 if (list_empty(&rnp->blkd_tasks))
832 raw_spin_unlock_irqrestore(&rnp->lock, flags);
834 rnp->exp_tasks = rnp->blkd_tasks.next;
835 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
839 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
843 * synchronize_rcu_expedited - Brute-force RCU grace period
845 * Wait for an RCU-preempt grace period, but expedite it. The basic
846 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
847 * the ->blkd_tasks lists and wait for this list to drain. This consumes
848 * significant time on all CPUs and is unfriendly to real-time workloads,
849 * so is thus not recommended for any sort of common-case code.
850 * In fact, if you are using synchronize_rcu_expedited() in a loop,
851 * please restructure your code to batch your updates, and then Use a
852 * single synchronize_rcu() instead.
854 * Note that it is illegal to call this function while holding any lock
855 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
856 * to call this function from a CPU-hotplug notifier. Failing to observe
857 * these restriction will result in deadlock.
859 void synchronize_rcu_expedited(void)
862 struct rcu_node *rnp;
863 struct rcu_state *rsp = &rcu_preempt_state;
867 smp_mb(); /* Caller's modifications seen first by other CPUs. */
868 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
869 smp_mb(); /* Above access cannot bleed into critical section. */
872 * Acquire lock, falling back to synchronize_rcu() if too many
873 * lock-acquisition failures. Of course, if someone does the
874 * expedited grace period for us, just leave.
876 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
878 udelay(trycount * num_online_cpus());
883 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
884 goto mb_ret; /* Others did our work for us. */
886 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
887 goto unlock_mb_ret; /* Others did our work for us. */
889 /* force all RCU readers onto ->blkd_tasks lists. */
890 synchronize_sched_expedited();
892 raw_spin_lock_irqsave(&rsp->onofflock, flags);
894 /* Initialize ->expmask for all non-leaf rcu_node structures. */
895 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
896 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
897 rnp->expmask = rnp->qsmaskinit;
898 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
901 /* Snapshot current state of ->blkd_tasks lists. */
902 rcu_for_each_leaf_node(rsp, rnp)
903 sync_rcu_preempt_exp_init(rsp, rnp);
904 if (NUM_RCU_NODES > 1)
905 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
907 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
909 /* Wait for snapshotted ->blkd_tasks lists to drain. */
910 rnp = rcu_get_root(rsp);
911 wait_event(sync_rcu_preempt_exp_wq,
912 sync_rcu_preempt_exp_done(rnp));
914 /* Clean up and exit. */
915 smp_mb(); /* ensure expedited GP seen before counter increment. */
916 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
918 mutex_unlock(&sync_rcu_preempt_exp_mutex);
920 smp_mb(); /* ensure subsequent action seen after grace period. */
922 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
925 * Check to see if there is any immediate preemptible-RCU-related work
928 static int rcu_preempt_pending(int cpu)
930 return __rcu_pending(&rcu_preempt_state,
931 &per_cpu(rcu_preempt_data, cpu));
935 * Does preemptible RCU have callbacks on this CPU?
937 static int rcu_preempt_cpu_has_callbacks(int cpu)
939 return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
943 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
945 void rcu_barrier(void)
947 _rcu_barrier(&rcu_preempt_state, call_rcu);
949 EXPORT_SYMBOL_GPL(rcu_barrier);
952 * Initialize preemptible RCU's per-CPU data.
954 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
956 rcu_init_percpu_data(cpu, &rcu_preempt_state, 1);
960 * Move preemptible RCU's callbacks from dying CPU to other online CPU
961 * and record a quiescent state.
963 static void rcu_preempt_cleanup_dying_cpu(void)
965 rcu_cleanup_dying_cpu(&rcu_preempt_state);
969 * Initialize preemptible RCU's state structures.
971 static void __init __rcu_init_preempt(void)
973 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
976 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
978 static struct rcu_state *rcu_state = &rcu_sched_state;
981 * Tell them what RCU they are running.
983 static void __init rcu_bootup_announce(void)
985 printk(KERN_INFO "Hierarchical RCU implementation.\n");
986 rcu_bootup_announce_oddness();
990 * Return the number of RCU batches processed thus far for debug & stats.
992 long rcu_batches_completed(void)
994 return rcu_batches_completed_sched();
996 EXPORT_SYMBOL_GPL(rcu_batches_completed);
999 * Force a quiescent state for RCU, which, because there is no preemptible
1000 * RCU, becomes the same as rcu-sched.
1002 void rcu_force_quiescent_state(void)
1004 rcu_sched_force_quiescent_state();
1006 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
1009 * Because preemptible RCU does not exist, we never have to check for
1010 * CPUs being in quiescent states.
1012 static void rcu_preempt_note_context_switch(int cpu)
1017 * Because preemptible RCU does not exist, there are never any preempted
1020 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
1025 #ifdef CONFIG_HOTPLUG_CPU
1027 /* Because preemptible RCU does not exist, no quieting of tasks. */
1028 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1030 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1033 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1036 * Because preemptible RCU does not exist, we never have to check for
1037 * tasks blocked within RCU read-side critical sections.
1039 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1044 * Because preemptible RCU does not exist, we never have to check for
1045 * tasks blocked within RCU read-side critical sections.
1047 static int rcu_print_task_stall(struct rcu_node *rnp)
1053 * Because preemptible RCU does not exist, there is no need to suppress
1054 * its CPU stall warnings.
1056 static void rcu_preempt_stall_reset(void)
1061 * Because there is no preemptible RCU, there can be no readers blocked,
1062 * so there is no need to check for blocked tasks. So check only for
1063 * bogus qsmask values.
1065 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1067 WARN_ON_ONCE(rnp->qsmask);
1070 #ifdef CONFIG_HOTPLUG_CPU
1073 * Because preemptible RCU does not exist, it never needs to migrate
1074 * tasks that were blocked within RCU read-side critical sections, and
1075 * such non-existent tasks cannot possibly have been blocking the current
1078 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1079 struct rcu_node *rnp,
1080 struct rcu_data *rdp)
1085 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1088 * Because preemptible RCU does not exist, it never needs CPU-offline
1091 static void rcu_preempt_cleanup_dead_cpu(int cpu)
1096 * Because preemptible RCU does not exist, it never has any callbacks
1099 static void rcu_preempt_check_callbacks(int cpu)
1104 * Because preemptible RCU does not exist, it never has any callbacks
1107 static void rcu_preempt_process_callbacks(void)
1112 * Queue an RCU callback for lazy invocation after a grace period.
1113 * This will likely be later named something like "call_rcu_lazy()",
1114 * but this change will require some way of tagging the lazy RCU
1115 * callbacks in the list of pending callbacks. Until then, this
1116 * function may only be called from __kfree_rcu().
1118 * Because there is no preemptible RCU, we use RCU-sched instead.
1120 void kfree_call_rcu(struct rcu_head *head,
1121 void (*func)(struct rcu_head *rcu))
1123 __call_rcu(head, func, &rcu_sched_state, 1);
1125 EXPORT_SYMBOL_GPL(kfree_call_rcu);
1128 * Wait for an rcu-preempt grace period, but make it happen quickly.
1129 * But because preemptible RCU does not exist, map to rcu-sched.
1131 void synchronize_rcu_expedited(void)
1133 synchronize_sched_expedited();
1135 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1137 #ifdef CONFIG_HOTPLUG_CPU
1140 * Because preemptible RCU does not exist, there is never any need to
1141 * report on tasks preempted in RCU read-side critical sections during
1142 * expedited RCU grace periods.
1144 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1149 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1152 * Because preemptible RCU does not exist, it never has any work to do.
1154 static int rcu_preempt_pending(int cpu)
1160 * Because preemptible RCU does not exist, it never has callbacks
1162 static int rcu_preempt_cpu_has_callbacks(int cpu)
1168 * Because preemptible RCU does not exist, rcu_barrier() is just
1169 * another name for rcu_barrier_sched().
1171 void rcu_barrier(void)
1173 rcu_barrier_sched();
1175 EXPORT_SYMBOL_GPL(rcu_barrier);
1178 * Because preemptible RCU does not exist, there is no per-CPU
1179 * data to initialize.
1181 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
1186 * Because there is no preemptible RCU, there is no cleanup to do.
1188 static void rcu_preempt_cleanup_dying_cpu(void)
1193 * Because preemptible RCU does not exist, it need not be initialized.
1195 static void __init __rcu_init_preempt(void)
1199 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1201 #ifdef CONFIG_RCU_BOOST
1203 #include "rtmutex_common.h"
1205 #ifdef CONFIG_RCU_TRACE
1207 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1209 if (list_empty(&rnp->blkd_tasks))
1210 rnp->n_balk_blkd_tasks++;
1211 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1212 rnp->n_balk_exp_gp_tasks++;
1213 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1214 rnp->n_balk_boost_tasks++;
1215 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1216 rnp->n_balk_notblocked++;
1217 else if (rnp->gp_tasks != NULL &&
1218 ULONG_CMP_LT(jiffies, rnp->boost_time))
1219 rnp->n_balk_notyet++;
1224 #else /* #ifdef CONFIG_RCU_TRACE */
1226 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1230 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1233 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1234 * or ->boost_tasks, advancing the pointer to the next task in the
1235 * ->blkd_tasks list.
1237 * Note that irqs must be enabled: boosting the task can block.
1238 * Returns 1 if there are more tasks needing to be boosted.
1240 static int rcu_boost(struct rcu_node *rnp)
1242 unsigned long flags;
1243 struct rt_mutex mtx;
1244 struct task_struct *t;
1245 struct list_head *tb;
1247 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1248 return 0; /* Nothing left to boost. */
1250 raw_spin_lock_irqsave(&rnp->lock, flags);
1253 * Recheck under the lock: all tasks in need of boosting
1254 * might exit their RCU read-side critical sections on their own.
1256 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1257 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1262 * Preferentially boost tasks blocking expedited grace periods.
1263 * This cannot starve the normal grace periods because a second
1264 * expedited grace period must boost all blocked tasks, including
1265 * those blocking the pre-existing normal grace period.
1267 if (rnp->exp_tasks != NULL) {
1268 tb = rnp->exp_tasks;
1269 rnp->n_exp_boosts++;
1271 tb = rnp->boost_tasks;
1272 rnp->n_normal_boosts++;
1274 rnp->n_tasks_boosted++;
1277 * We boost task t by manufacturing an rt_mutex that appears to
1278 * be held by task t. We leave a pointer to that rt_mutex where
1279 * task t can find it, and task t will release the mutex when it
1280 * exits its outermost RCU read-side critical section. Then
1281 * simply acquiring this artificial rt_mutex will boost task
1282 * t's priority. (Thanks to tglx for suggesting this approach!)
1284 * Note that task t must acquire rnp->lock to remove itself from
1285 * the ->blkd_tasks list, which it will do from exit() if from
1286 * nowhere else. We therefore are guaranteed that task t will
1287 * stay around at least until we drop rnp->lock. Note that
1288 * rnp->lock also resolves races between our priority boosting
1289 * and task t's exiting its outermost RCU read-side critical
1292 t = container_of(tb, struct task_struct, rcu_node_entry);
1293 rt_mutex_init_proxy_locked(&mtx, t);
1294 t->rcu_boost_mutex = &mtx;
1295 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1296 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1297 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1299 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1300 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1304 * Timer handler to initiate waking up of boost kthreads that
1305 * have yielded the CPU due to excessive numbers of tasks to
1306 * boost. We wake up the per-rcu_node kthread, which in turn
1307 * will wake up the booster kthread.
1309 static void rcu_boost_kthread_timer(unsigned long arg)
1311 invoke_rcu_node_kthread((struct rcu_node *)arg);
1315 * Priority-boosting kthread. One per leaf rcu_node and one for the
1318 static int rcu_boost_kthread(void *arg)
1320 struct rcu_node *rnp = (struct rcu_node *)arg;
1324 trace_rcu_utilization("Start boost kthread@init");
1326 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1327 trace_rcu_utilization("End boost kthread@rcu_wait");
1328 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1329 trace_rcu_utilization("Start boost kthread@rcu_wait");
1330 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1331 more2boost = rcu_boost(rnp);
1337 trace_rcu_utilization("End boost kthread@rcu_yield");
1338 rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp);
1339 trace_rcu_utilization("Start boost kthread@rcu_yield");
1344 trace_rcu_utilization("End boost kthread@notreached");
1349 * Check to see if it is time to start boosting RCU readers that are
1350 * blocking the current grace period, and, if so, tell the per-rcu_node
1351 * kthread to start boosting them. If there is an expedited grace
1352 * period in progress, it is always time to boost.
1354 * The caller must hold rnp->lock, which this function releases,
1355 * but irqs remain disabled. The ->boost_kthread_task is immortal,
1356 * so we don't need to worry about it going away.
1358 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1360 struct task_struct *t;
1362 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1363 rnp->n_balk_exp_gp_tasks++;
1364 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1367 if (rnp->exp_tasks != NULL ||
1368 (rnp->gp_tasks != NULL &&
1369 rnp->boost_tasks == NULL &&
1371 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1372 if (rnp->exp_tasks == NULL)
1373 rnp->boost_tasks = rnp->gp_tasks;
1374 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1375 t = rnp->boost_kthread_task;
1379 rcu_initiate_boost_trace(rnp);
1380 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1385 * Wake up the per-CPU kthread to invoke RCU callbacks.
1387 static void invoke_rcu_callbacks_kthread(void)
1389 unsigned long flags;
1391 local_irq_save(flags);
1392 __this_cpu_write(rcu_cpu_has_work, 1);
1393 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1394 current != __this_cpu_read(rcu_cpu_kthread_task))
1395 wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
1396 local_irq_restore(flags);
1400 * Is the current CPU running the RCU-callbacks kthread?
1401 * Caller must have preemption disabled.
1403 static bool rcu_is_callbacks_kthread(void)
1405 return __get_cpu_var(rcu_cpu_kthread_task) == current;
1409 * Set the affinity of the boost kthread. The CPU-hotplug locks are
1410 * held, so no one should be messing with the existence of the boost
1413 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
1416 struct task_struct *t;
1418 t = rnp->boost_kthread_task;
1420 set_cpus_allowed_ptr(rnp->boost_kthread_task, cm);
1423 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1426 * Do priority-boost accounting for the start of a new grace period.
1428 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1430 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1434 * Create an RCU-boost kthread for the specified node if one does not
1435 * already exist. We only create this kthread for preemptible RCU.
1436 * Returns zero if all is well, a negated errno otherwise.
1438 static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1439 struct rcu_node *rnp,
1442 unsigned long flags;
1443 struct sched_param sp;
1444 struct task_struct *t;
1446 if (&rcu_preempt_state != rsp)
1449 if (rnp->boost_kthread_task != NULL)
1451 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1452 "rcub/%d", rnp_index);
1455 raw_spin_lock_irqsave(&rnp->lock, flags);
1456 rnp->boost_kthread_task = t;
1457 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1458 sp.sched_priority = RCU_BOOST_PRIO;
1459 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1460 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1464 #ifdef CONFIG_HOTPLUG_CPU
1467 * Stop the RCU's per-CPU kthread when its CPU goes offline,.
1469 static void rcu_stop_cpu_kthread(int cpu)
1471 struct task_struct *t;
1473 /* Stop the CPU's kthread. */
1474 t = per_cpu(rcu_cpu_kthread_task, cpu);
1476 per_cpu(rcu_cpu_kthread_task, cpu) = NULL;
1481 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1483 static void rcu_kthread_do_work(void)
1485 rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1486 rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1487 rcu_preempt_do_callbacks();
1491 * Wake up the specified per-rcu_node-structure kthread.
1492 * Because the per-rcu_node kthreads are immortal, we don't need
1493 * to do anything to keep them alive.
1495 static void invoke_rcu_node_kthread(struct rcu_node *rnp)
1497 struct task_struct *t;
1499 t = rnp->node_kthread_task;
1505 * Set the specified CPU's kthread to run RT or not, as specified by
1506 * the to_rt argument. The CPU-hotplug locks are held, so the task
1507 * is not going away.
1509 static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1512 struct sched_param sp;
1513 struct task_struct *t;
1515 t = per_cpu(rcu_cpu_kthread_task, cpu);
1519 policy = SCHED_FIFO;
1520 sp.sched_priority = RCU_KTHREAD_PRIO;
1522 policy = SCHED_NORMAL;
1523 sp.sched_priority = 0;
1525 sched_setscheduler_nocheck(t, policy, &sp);
1529 * Timer handler to initiate the waking up of per-CPU kthreads that
1530 * have yielded the CPU due to excess numbers of RCU callbacks.
1531 * We wake up the per-rcu_node kthread, which in turn will wake up
1532 * the booster kthread.
1534 static void rcu_cpu_kthread_timer(unsigned long arg)
1536 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
1537 struct rcu_node *rnp = rdp->mynode;
1539 atomic_or(rdp->grpmask, &rnp->wakemask);
1540 invoke_rcu_node_kthread(rnp);
1544 * Drop to non-real-time priority and yield, but only after posting a
1545 * timer that will cause us to regain our real-time priority if we
1546 * remain preempted. Either way, we restore our real-time priority
1549 static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
1551 struct sched_param sp;
1552 struct timer_list yield_timer;
1553 int prio = current->rt_priority;
1555 setup_timer_on_stack(&yield_timer, f, arg);
1556 mod_timer(&yield_timer, jiffies + 2);
1557 sp.sched_priority = 0;
1558 sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
1559 set_user_nice(current, 19);
1561 set_user_nice(current, 0);
1562 sp.sched_priority = prio;
1563 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1564 del_timer(&yield_timer);
1568 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1569 * This can happen while the corresponding CPU is either coming online
1570 * or going offline. We cannot wait until the CPU is fully online
1571 * before starting the kthread, because the various notifier functions
1572 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1573 * the corresponding CPU is online.
1575 * Return 1 if the kthread needs to stop, 0 otherwise.
1577 * Caller must disable bh. This function can momentarily enable it.
1579 static int rcu_cpu_kthread_should_stop(int cpu)
1581 while (cpu_is_offline(cpu) ||
1582 !cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)) ||
1583 smp_processor_id() != cpu) {
1584 if (kthread_should_stop())
1586 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1587 per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id();
1589 schedule_timeout_uninterruptible(1);
1590 if (!cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)))
1591 set_cpus_allowed_ptr(current, cpumask_of(cpu));
1594 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1599 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1600 * RCU softirq used in flavors and configurations of RCU that do not
1601 * support RCU priority boosting.
1603 static int rcu_cpu_kthread(void *arg)
1605 int cpu = (int)(long)arg;
1606 unsigned long flags;
1608 unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu);
1610 char *workp = &per_cpu(rcu_cpu_has_work, cpu);
1612 trace_rcu_utilization("Start CPU kthread@init");
1614 *statusp = RCU_KTHREAD_WAITING;
1615 trace_rcu_utilization("End CPU kthread@rcu_wait");
1616 rcu_wait(*workp != 0 || kthread_should_stop());
1617 trace_rcu_utilization("Start CPU kthread@rcu_wait");
1619 if (rcu_cpu_kthread_should_stop(cpu)) {
1623 *statusp = RCU_KTHREAD_RUNNING;
1624 per_cpu(rcu_cpu_kthread_loops, cpu)++;
1625 local_irq_save(flags);
1628 local_irq_restore(flags);
1630 rcu_kthread_do_work();
1637 *statusp = RCU_KTHREAD_YIELDING;
1638 trace_rcu_utilization("End CPU kthread@rcu_yield");
1639 rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
1640 trace_rcu_utilization("Start CPU kthread@rcu_yield");
1644 *statusp = RCU_KTHREAD_STOPPED;
1645 trace_rcu_utilization("End CPU kthread@term");
1650 * Spawn a per-CPU kthread, setting up affinity and priority.
1651 * Because the CPU hotplug lock is held, no other CPU will be attempting
1652 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1653 * attempting to access it during boot, but the locking in kthread_bind()
1654 * will enforce sufficient ordering.
1656 * Please note that we cannot simply refuse to wake up the per-CPU
1657 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1658 * which can result in softlockup complaints if the task ends up being
1659 * idle for more than a couple of minutes.
1661 * However, please note also that we cannot bind the per-CPU kthread to its
1662 * CPU until that CPU is fully online. We also cannot wait until the
1663 * CPU is fully online before we create its per-CPU kthread, as this would
1664 * deadlock the system when CPU notifiers tried waiting for grace
1665 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1666 * is online. If its CPU is not yet fully online, then the code in
1667 * rcu_cpu_kthread() will wait until it is fully online, and then do
1670 static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu)
1672 struct sched_param sp;
1673 struct task_struct *t;
1675 if (!rcu_scheduler_fully_active ||
1676 per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
1678 t = kthread_create_on_node(rcu_cpu_kthread,
1684 if (cpu_online(cpu))
1685 kthread_bind(t, cpu);
1686 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1687 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL);
1688 sp.sched_priority = RCU_KTHREAD_PRIO;
1689 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1690 per_cpu(rcu_cpu_kthread_task, cpu) = t;
1691 wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */
1696 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1697 * kthreads when needed. We ignore requests to wake up kthreads
1698 * for offline CPUs, which is OK because force_quiescent_state()
1699 * takes care of this case.
1701 static int rcu_node_kthread(void *arg)
1704 unsigned long flags;
1706 struct rcu_node *rnp = (struct rcu_node *)arg;
1707 struct sched_param sp;
1708 struct task_struct *t;
1711 rnp->node_kthread_status = RCU_KTHREAD_WAITING;
1712 rcu_wait(atomic_read(&rnp->wakemask) != 0);
1713 rnp->node_kthread_status = RCU_KTHREAD_RUNNING;
1714 raw_spin_lock_irqsave(&rnp->lock, flags);
1715 mask = atomic_xchg(&rnp->wakemask, 0);
1716 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1717 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
1718 if ((mask & 0x1) == 0)
1721 t = per_cpu(rcu_cpu_kthread_task, cpu);
1722 if (!cpu_online(cpu) || t == NULL) {
1726 per_cpu(rcu_cpu_has_work, cpu) = 1;
1727 sp.sched_priority = RCU_KTHREAD_PRIO;
1728 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1733 rnp->node_kthread_status = RCU_KTHREAD_STOPPED;
1738 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1739 * served by the rcu_node in question. The CPU hotplug lock is still
1740 * held, so the value of rnp->qsmaskinit will be stable.
1742 * We don't include outgoingcpu in the affinity set, use -1 if there is
1743 * no outgoing CPU. If there are no CPUs left in the affinity set,
1744 * this function allows the kthread to execute on any CPU.
1746 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1750 unsigned long mask = rnp->qsmaskinit;
1752 if (rnp->node_kthread_task == NULL)
1754 if (!alloc_cpumask_var(&cm, GFP_KERNEL))
1757 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1758 if ((mask & 0x1) && cpu != outgoingcpu)
1759 cpumask_set_cpu(cpu, cm);
1760 if (cpumask_weight(cm) == 0) {
1762 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1763 cpumask_clear_cpu(cpu, cm);
1764 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1766 set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
1767 rcu_boost_kthread_setaffinity(rnp, cm);
1768 free_cpumask_var(cm);
1772 * Spawn a per-rcu_node kthread, setting priority and affinity.
1773 * Called during boot before online/offline can happen, or, if
1774 * during runtime, with the main CPU-hotplug locks held. So only
1775 * one of these can be executing at a time.
1777 static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
1778 struct rcu_node *rnp)
1780 unsigned long flags;
1781 int rnp_index = rnp - &rsp->node[0];
1782 struct sched_param sp;
1783 struct task_struct *t;
1785 if (!rcu_scheduler_fully_active ||
1786 rnp->qsmaskinit == 0)
1788 if (rnp->node_kthread_task == NULL) {
1789 t = kthread_create(rcu_node_kthread, (void *)rnp,
1790 "rcun/%d", rnp_index);
1793 raw_spin_lock_irqsave(&rnp->lock, flags);
1794 rnp->node_kthread_task = t;
1795 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1796 sp.sched_priority = 99;
1797 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1798 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1800 return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
1804 * Spawn all kthreads -- called as soon as the scheduler is running.
1806 static int __init rcu_spawn_kthreads(void)
1809 struct rcu_node *rnp;
1811 rcu_scheduler_fully_active = 1;
1812 for_each_possible_cpu(cpu) {
1813 per_cpu(rcu_cpu_has_work, cpu) = 0;
1814 if (cpu_online(cpu))
1815 (void)rcu_spawn_one_cpu_kthread(cpu);
1817 rnp = rcu_get_root(rcu_state);
1818 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1819 if (NUM_RCU_NODES > 1) {
1820 rcu_for_each_leaf_node(rcu_state, rnp)
1821 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1825 early_initcall(rcu_spawn_kthreads);
1827 static void __cpuinit rcu_prepare_kthreads(int cpu)
1829 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1830 struct rcu_node *rnp = rdp->mynode;
1832 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1833 if (rcu_scheduler_fully_active) {
1834 (void)rcu_spawn_one_cpu_kthread(cpu);
1835 if (rnp->node_kthread_task == NULL)
1836 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1840 #else /* #ifdef CONFIG_RCU_BOOST */
1842 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1844 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1847 static void invoke_rcu_callbacks_kthread(void)
1852 static bool rcu_is_callbacks_kthread(void)
1857 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1861 #ifdef CONFIG_HOTPLUG_CPU
1863 static void rcu_stop_cpu_kthread(int cpu)
1867 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1869 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1873 static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1877 static int __init rcu_scheduler_really_started(void)
1879 rcu_scheduler_fully_active = 1;
1882 early_initcall(rcu_scheduler_really_started);
1884 static void __cpuinit rcu_prepare_kthreads(int cpu)
1888 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1890 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1893 * Check to see if any future RCU-related work will need to be done
1894 * by the current CPU, even if none need be done immediately, returning
1895 * 1 if so. This function is part of the RCU implementation; it is -not-
1896 * an exported member of the RCU API.
1898 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1899 * any flavor of RCU.
1901 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1903 *delta_jiffies = ULONG_MAX;
1904 return rcu_cpu_has_callbacks(cpu);
1908 * Because we do not have RCU_FAST_NO_HZ, don't bother initializing for it.
1910 static void rcu_prepare_for_idle_init(int cpu)
1915 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1918 static void rcu_cleanup_after_idle(int cpu)
1923 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1926 static void rcu_prepare_for_idle(int cpu)
1931 * Don't bother keeping a running count of the number of RCU callbacks
1932 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1934 static void rcu_idle_count_callbacks_posted(void)
1938 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1941 * This code is invoked when a CPU goes idle, at which point we want
1942 * to have the CPU do everything required for RCU so that it can enter
1943 * the energy-efficient dyntick-idle mode. This is handled by a
1944 * state machine implemented by rcu_prepare_for_idle() below.
1946 * The following three proprocessor symbols control this state machine:
1948 * RCU_IDLE_FLUSHES gives the maximum number of times that we will attempt
1949 * to satisfy RCU. Beyond this point, it is better to incur a periodic
1950 * scheduling-clock interrupt than to loop through the state machine
1952 * RCU_IDLE_OPT_FLUSHES gives the number of RCU_IDLE_FLUSHES that are
1953 * optional if RCU does not need anything immediately from this
1954 * CPU, even if this CPU still has RCU callbacks queued. The first
1955 * times through the state machine are mandatory: we need to give
1956 * the state machine a chance to communicate a quiescent state
1958 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1959 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1960 * is sized to be roughly one RCU grace period. Those energy-efficiency
1961 * benchmarkers who might otherwise be tempted to set this to a large
1962 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1963 * system. And if you are -that- concerned about energy efficiency,
1964 * just power the system down and be done with it!
1965 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1966 * permitted to sleep in dyntick-idle mode with only lazy RCU
1967 * callbacks pending. Setting this too high can OOM your system.
1969 * The values below work well in practice. If future workloads require
1970 * adjustment, they can be converted into kernel config parameters, though
1971 * making the state machine smarter might be a better option.
1973 #define RCU_IDLE_FLUSHES 5 /* Number of dyntick-idle tries. */
1974 #define RCU_IDLE_OPT_FLUSHES 3 /* Optional dyntick-idle tries. */
1975 #define RCU_IDLE_GP_DELAY 6 /* Roughly one grace period. */
1976 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1979 * Does the specified flavor of RCU have non-lazy callbacks pending on
1980 * the specified CPU? Both RCU flavor and CPU are specified by the
1981 * rcu_data structure.
1983 static bool __rcu_cpu_has_nonlazy_callbacks(struct rcu_data *rdp)
1985 return rdp->qlen != rdp->qlen_lazy;
1988 #ifdef CONFIG_TREE_PREEMPT_RCU
1991 * Are there non-lazy RCU-preempt callbacks? (There cannot be if there
1992 * is no RCU-preempt in the kernel.)
1994 static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu)
1996 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
1998 return __rcu_cpu_has_nonlazy_callbacks(rdp);
2001 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
2003 static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu)
2008 #endif /* else #ifdef CONFIG_TREE_PREEMPT_RCU */
2011 * Does any flavor of RCU have non-lazy callbacks on the specified CPU?
2013 static bool rcu_cpu_has_nonlazy_callbacks(int cpu)
2015 return __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_sched_data, cpu)) ||
2016 __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_bh_data, cpu)) ||
2017 rcu_preempt_cpu_has_nonlazy_callbacks(cpu);
2021 * Allow the CPU to enter dyntick-idle mode if either: (1) There are no
2022 * callbacks on this CPU, (2) this CPU has not yet attempted to enter
2023 * dyntick-idle mode, or (3) this CPU is in the process of attempting to
2024 * enter dyntick-idle mode. Otherwise, if we have recently tried and failed
2025 * to enter dyntick-idle mode, we refuse to try to enter it. After all,
2026 * it is better to incur scheduling-clock interrupts than to spin
2027 * continuously for the same time duration!
2029 * The delta_jiffies argument is used to store the time when RCU is
2030 * going to need the CPU again if it still has callbacks. The reason
2031 * for this is that rcu_prepare_for_idle() might need to post a timer,
2032 * but if so, it will do so after tick_nohz_stop_sched_tick() has set
2033 * the wakeup time for this CPU. This means that RCU's timer can be
2034 * delayed until the wakeup time, which defeats the purpose of posting
2037 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
2039 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
2041 /* Flag a new idle sojourn to the idle-entry state machine. */
2042 rdtp->idle_first_pass = 1;
2043 /* If no callbacks, RCU doesn't need the CPU. */
2044 if (!rcu_cpu_has_callbacks(cpu)) {
2045 *delta_jiffies = ULONG_MAX;
2048 if (rdtp->dyntick_holdoff == jiffies) {
2049 /* RCU recently tried and failed, so don't try again. */
2053 /* Set up for the possibility that RCU will post a timer. */
2054 if (rcu_cpu_has_nonlazy_callbacks(cpu))
2055 *delta_jiffies = RCU_IDLE_GP_DELAY;
2057 *delta_jiffies = RCU_IDLE_LAZY_GP_DELAY;
2062 * Handler for smp_call_function_single(). The only point of this
2063 * handler is to wake the CPU up, so the handler does only tracing.
2065 void rcu_idle_demigrate(void *unused)
2067 trace_rcu_prep_idle("Demigrate");
2071 * Timer handler used to force CPU to start pushing its remaining RCU
2072 * callbacks in the case where it entered dyntick-idle mode with callbacks
2073 * pending. The hander doesn't really need to do anything because the
2074 * real work is done upon re-entry to idle, or by the next scheduling-clock
2075 * interrupt should idle not be re-entered.
2077 * One special case: the timer gets migrated without awakening the CPU
2078 * on which the timer was scheduled on. In this case, we must wake up
2079 * that CPU. We do so with smp_call_function_single().
2081 static void rcu_idle_gp_timer_func(unsigned long cpu_in)
2083 int cpu = (int)cpu_in;
2085 trace_rcu_prep_idle("Timer");
2086 if (cpu != smp_processor_id())
2087 smp_call_function_single(cpu, rcu_idle_demigrate, NULL, 0);
2089 WARN_ON_ONCE(1); /* Getting here can hang the system... */
2093 * Initialize the timer used to pull CPUs out of dyntick-idle mode.
2095 static void rcu_prepare_for_idle_init(int cpu)
2097 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
2099 rdtp->dyntick_holdoff = jiffies - 1;
2100 setup_timer(&rdtp->idle_gp_timer, rcu_idle_gp_timer_func, cpu);
2101 rdtp->idle_gp_timer_expires = jiffies - 1;
2102 rdtp->idle_first_pass = 1;
2106 * Clean up for exit from idle. Because we are exiting from idle, there
2107 * is no longer any point to ->idle_gp_timer, so cancel it. This will
2108 * do nothing if this timer is not active, so just cancel it unconditionally.
2110 static void rcu_cleanup_after_idle(int cpu)
2112 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
2114 del_timer(&rdtp->idle_gp_timer);
2115 trace_rcu_prep_idle("Cleanup after idle");
2119 * Check to see if any RCU-related work can be done by the current CPU,
2120 * and if so, schedule a softirq to get it done. This function is part
2121 * of the RCU implementation; it is -not- an exported member of the RCU API.
2123 * The idea is for the current CPU to clear out all work required by the
2124 * RCU core for the current grace period, so that this CPU can be permitted
2125 * to enter dyntick-idle mode. In some cases, it will need to be awakened
2126 * at the end of the grace period by whatever CPU ends the grace period.
2127 * This allows CPUs to go dyntick-idle more quickly, and to reduce the
2128 * number of wakeups by a modest integer factor.
2130 * Because it is not legal to invoke rcu_process_callbacks() with irqs
2131 * disabled, we do one pass of force_quiescent_state(), then do a
2132 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
2133 * later. The ->dyntick_drain field controls the sequencing.
2135 * The caller must have disabled interrupts.
2137 static void rcu_prepare_for_idle(int cpu)
2139 struct timer_list *tp;
2140 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
2143 * If this is an idle re-entry, for example, due to use of
2144 * RCU_NONIDLE() or the new idle-loop tracing API within the idle
2145 * loop, then don't take any state-machine actions, unless the
2146 * momentary exit from idle queued additional non-lazy callbacks.
2147 * Instead, repost the ->idle_gp_timer if this CPU has callbacks
2150 if (!rdtp->idle_first_pass &&
2151 (rdtp->nonlazy_posted == rdtp->nonlazy_posted_snap)) {
2152 if (rcu_cpu_has_callbacks(cpu)) {
2153 tp = &rdtp->idle_gp_timer;
2154 mod_timer_pinned(tp, rdtp->idle_gp_timer_expires);
2158 rdtp->idle_first_pass = 0;
2159 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted - 1;
2162 * If there are no callbacks on this CPU, enter dyntick-idle mode.
2163 * Also reset state to avoid prejudicing later attempts.
2165 if (!rcu_cpu_has_callbacks(cpu)) {
2166 rdtp->dyntick_holdoff = jiffies - 1;
2167 rdtp->dyntick_drain = 0;
2168 trace_rcu_prep_idle("No callbacks");
2173 * If in holdoff mode, just return. We will presumably have
2174 * refrained from disabling the scheduling-clock tick.
2176 if (rdtp->dyntick_holdoff == jiffies) {
2177 trace_rcu_prep_idle("In holdoff");
2181 /* Check and update the ->dyntick_drain sequencing. */
2182 if (rdtp->dyntick_drain <= 0) {
2183 /* First time through, initialize the counter. */
2184 rdtp->dyntick_drain = RCU_IDLE_FLUSHES;
2185 } else if (rdtp->dyntick_drain <= RCU_IDLE_OPT_FLUSHES &&
2186 !rcu_pending(cpu) &&
2187 !local_softirq_pending()) {
2188 /* Can we go dyntick-idle despite still having callbacks? */
2189 rdtp->dyntick_drain = 0;
2190 rdtp->dyntick_holdoff = jiffies;
2191 if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
2192 trace_rcu_prep_idle("Dyntick with callbacks");
2193 rdtp->idle_gp_timer_expires =
2194 jiffies + RCU_IDLE_GP_DELAY;
2196 rdtp->idle_gp_timer_expires =
2197 jiffies + RCU_IDLE_LAZY_GP_DELAY;
2198 trace_rcu_prep_idle("Dyntick with lazy callbacks");
2200 tp = &rdtp->idle_gp_timer;
2201 mod_timer_pinned(tp, rdtp->idle_gp_timer_expires);
2202 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
2203 return; /* Nothing more to do immediately. */
2204 } else if (--(rdtp->dyntick_drain) <= 0) {
2205 /* We have hit the limit, so time to give up. */
2206 rdtp->dyntick_holdoff = jiffies;
2207 trace_rcu_prep_idle("Begin holdoff");
2208 invoke_rcu_core(); /* Force the CPU out of dyntick-idle. */
2213 * Do one step of pushing the remaining RCU callbacks through
2214 * the RCU core state machine.
2216 #ifdef CONFIG_TREE_PREEMPT_RCU
2217 if (per_cpu(rcu_preempt_data, cpu).nxtlist) {
2218 rcu_preempt_qs(cpu);
2219 force_quiescent_state(&rcu_preempt_state, 0);
2221 #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
2222 if (per_cpu(rcu_sched_data, cpu).nxtlist) {
2224 force_quiescent_state(&rcu_sched_state, 0);
2226 if (per_cpu(rcu_bh_data, cpu).nxtlist) {
2228 force_quiescent_state(&rcu_bh_state, 0);
2232 * If RCU callbacks are still pending, RCU still needs this CPU.
2233 * So try forcing the callbacks through the grace period.
2235 if (rcu_cpu_has_callbacks(cpu)) {
2236 trace_rcu_prep_idle("More callbacks");
2239 trace_rcu_prep_idle("Callbacks drained");
2243 * Keep a running count of the number of non-lazy callbacks posted
2244 * on this CPU. This running counter (which is never decremented) allows
2245 * rcu_prepare_for_idle() to detect when something out of the idle loop
2246 * posts a callback, even if an equal number of callbacks are invoked.
2247 * Of course, callbacks should only be posted from within a trace event
2248 * designed to be called from idle or from within RCU_NONIDLE().
2250 static void rcu_idle_count_callbacks_posted(void)
2252 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
2255 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
2257 #ifdef CONFIG_RCU_CPU_STALL_INFO
2259 #ifdef CONFIG_RCU_FAST_NO_HZ
2261 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
2263 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
2264 struct timer_list *tltp = &rdtp->idle_gp_timer;
2266 sprintf(cp, "drain=%d %c timer=%lu",
2267 rdtp->dyntick_drain,
2268 rdtp->dyntick_holdoff == jiffies ? 'H' : '.',
2269 timer_pending(tltp) ? tltp->expires - jiffies : -1);
2272 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
2274 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
2278 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
2280 /* Initiate the stall-info list. */
2281 static void print_cpu_stall_info_begin(void)
2283 printk(KERN_CONT "\n");
2287 * Print out diagnostic information for the specified stalled CPU.
2289 * If the specified CPU is aware of the current RCU grace period
2290 * (flavor specified by rsp), then print the number of scheduling
2291 * clock interrupts the CPU has taken during the time that it has
2292 * been aware. Otherwise, print the number of RCU grace periods
2293 * that this CPU is ignorant of, for example, "1" if the CPU was
2294 * aware of the previous grace period.
2296 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
2298 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
2300 char fast_no_hz[72];
2301 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2302 struct rcu_dynticks *rdtp = rdp->dynticks;
2304 unsigned long ticks_value;
2306 if (rsp->gpnum == rdp->gpnum) {
2307 ticks_title = "ticks this GP";
2308 ticks_value = rdp->ticks_this_gp;
2310 ticks_title = "GPs behind";
2311 ticks_value = rsp->gpnum - rdp->gpnum;
2313 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
2314 printk(KERN_ERR "\t%d: (%lu %s) idle=%03x/%llx/%d %s\n",
2315 cpu, ticks_value, ticks_title,
2316 atomic_read(&rdtp->dynticks) & 0xfff,
2317 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
2321 /* Terminate the stall-info list. */
2322 static void print_cpu_stall_info_end(void)
2324 printk(KERN_ERR "\t");
2327 /* Zero ->ticks_this_gp for all flavors of RCU. */
2328 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2330 rdp->ticks_this_gp = 0;
2333 /* Increment ->ticks_this_gp for all flavors of RCU. */
2334 static void increment_cpu_stall_ticks(void)
2336 __get_cpu_var(rcu_sched_data).ticks_this_gp++;
2337 __get_cpu_var(rcu_bh_data).ticks_this_gp++;
2338 #ifdef CONFIG_TREE_PREEMPT_RCU
2339 __get_cpu_var(rcu_preempt_data).ticks_this_gp++;
2340 #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
2343 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
2345 static void print_cpu_stall_info_begin(void)
2347 printk(KERN_CONT " {");
2350 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
2352 printk(KERN_CONT " %d", cpu);
2355 static void print_cpu_stall_info_end(void)
2357 printk(KERN_CONT "} ");
2360 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2364 static void increment_cpu_stall_ticks(void)
2368 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */