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/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
32 #define RCU_KTHREAD_PRIO 1
34 #ifdef CONFIG_RCU_BOOST
35 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
37 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
40 #ifdef CONFIG_RCU_NOCB_CPU
41 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
42 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
43 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
44 static char __initdata nocb_buf[NR_CPUS * 5];
45 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
48 * Check the RCU kernel configuration parameters and print informative
49 * messages about anything out of the ordinary. If you like #ifdef, you
50 * will love this function.
52 static void __init rcu_bootup_announce_oddness(void)
54 #ifdef CONFIG_RCU_TRACE
55 printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n");
57 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
58 printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
61 #ifdef CONFIG_RCU_FANOUT_EXACT
62 printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n");
64 #ifdef CONFIG_RCU_FAST_NO_HZ
66 "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
68 #ifdef CONFIG_PROVE_RCU
69 printk(KERN_INFO "\tRCU lockdep checking is enabled.\n");
71 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
72 printk(KERN_INFO "\tRCU torture testing starts during boot.\n");
74 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
75 printk(KERN_INFO "\tDump stacks of tasks blocking RCU-preempt GP.\n");
77 #if defined(CONFIG_RCU_CPU_STALL_INFO)
78 printk(KERN_INFO "\tAdditional per-CPU info printed with stalls.\n");
80 #if NUM_RCU_LVL_4 != 0
81 printk(KERN_INFO "\tFour-level hierarchy is enabled.\n");
83 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
84 printk(KERN_INFO "\tExperimental boot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
85 if (nr_cpu_ids != NR_CPUS)
86 printk(KERN_INFO "\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
87 #ifdef CONFIG_RCU_NOCB_CPU
88 #ifndef CONFIG_RCU_NOCB_CPU_NONE
89 if (!have_rcu_nocb_mask) {
90 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
91 have_rcu_nocb_mask = true;
93 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
94 pr_info("\tExperimental no-CBs CPU 0\n");
95 cpumask_set_cpu(0, rcu_nocb_mask);
96 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
97 #ifdef CONFIG_RCU_NOCB_CPU_ALL
98 pr_info("\tExperimental no-CBs for all CPUs\n");
99 cpumask_setall(rcu_nocb_mask);
100 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
101 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
102 if (have_rcu_nocb_mask) {
103 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
104 pr_info("\tExperimental no-CBs CPUs: %s.\n", nocb_buf);
106 pr_info("\tExperimental polled no-CBs CPUs.\n");
108 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
111 #ifdef CONFIG_TREE_PREEMPT_RCU
113 struct rcu_state rcu_preempt_state =
114 RCU_STATE_INITIALIZER(rcu_preempt, call_rcu);
115 DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
116 static struct rcu_state *rcu_state = &rcu_preempt_state;
118 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
121 * Tell them what RCU they are running.
123 static void __init rcu_bootup_announce(void)
125 printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n");
126 rcu_bootup_announce_oddness();
130 * Return the number of RCU-preempt batches processed thus far
131 * for debug and statistics.
133 long rcu_batches_completed_preempt(void)
135 return rcu_preempt_state.completed;
137 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
140 * Return the number of RCU batches processed thus far for debug & stats.
142 long rcu_batches_completed(void)
144 return rcu_batches_completed_preempt();
146 EXPORT_SYMBOL_GPL(rcu_batches_completed);
149 * Force a quiescent state for preemptible RCU.
151 void rcu_force_quiescent_state(void)
153 force_quiescent_state(&rcu_preempt_state);
155 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
158 * Record a preemptible-RCU quiescent state for the specified CPU. Note
159 * that this just means that the task currently running on the CPU is
160 * not in a quiescent state. There might be any number of tasks blocked
161 * while in an RCU read-side critical section.
163 * Unlike the other rcu_*_qs() functions, callers to this function
164 * must disable irqs in order to protect the assignment to
165 * ->rcu_read_unlock_special.
167 static void rcu_preempt_qs(int cpu)
169 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
171 if (rdp->passed_quiesce == 0)
172 trace_rcu_grace_period("rcu_preempt", rdp->gpnum, "cpuqs");
173 rdp->passed_quiesce = 1;
174 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
178 * We have entered the scheduler, and the current task might soon be
179 * context-switched away from. If this task is in an RCU read-side
180 * critical section, we will no longer be able to rely on the CPU to
181 * record that fact, so we enqueue the task on the blkd_tasks list.
182 * The task will dequeue itself when it exits the outermost enclosing
183 * RCU read-side critical section. Therefore, the current grace period
184 * cannot be permitted to complete until the blkd_tasks list entries
185 * predating the current grace period drain, in other words, until
186 * rnp->gp_tasks becomes NULL.
188 * Caller must disable preemption.
190 static void rcu_preempt_note_context_switch(int cpu)
192 struct task_struct *t = current;
194 struct rcu_data *rdp;
195 struct rcu_node *rnp;
197 if (t->rcu_read_lock_nesting > 0 &&
198 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
200 /* Possibly blocking in an RCU read-side critical section. */
201 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
203 raw_spin_lock_irqsave(&rnp->lock, flags);
204 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
205 t->rcu_blocked_node = rnp;
208 * If this CPU has already checked in, then this task
209 * will hold up the next grace period rather than the
210 * current grace period. Queue the task accordingly.
211 * If the task is queued for the current grace period
212 * (i.e., this CPU has not yet passed through a quiescent
213 * state for the current grace period), then as long
214 * as that task remains queued, the current grace period
215 * cannot end. Note that there is some uncertainty as
216 * to exactly when the current grace period started.
217 * We take a conservative approach, which can result
218 * in unnecessarily waiting on tasks that started very
219 * slightly after the current grace period began. C'est
222 * But first, note that the current CPU must still be
225 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
226 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
227 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
228 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
229 rnp->gp_tasks = &t->rcu_node_entry;
230 #ifdef CONFIG_RCU_BOOST
231 if (rnp->boost_tasks != NULL)
232 rnp->boost_tasks = rnp->gp_tasks;
233 #endif /* #ifdef CONFIG_RCU_BOOST */
235 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
236 if (rnp->qsmask & rdp->grpmask)
237 rnp->gp_tasks = &t->rcu_node_entry;
239 trace_rcu_preempt_task(rdp->rsp->name,
241 (rnp->qsmask & rdp->grpmask)
244 raw_spin_unlock_irqrestore(&rnp->lock, flags);
245 } else if (t->rcu_read_lock_nesting < 0 &&
246 t->rcu_read_unlock_special) {
249 * Complete exit from RCU read-side critical section on
250 * behalf of preempted instance of __rcu_read_unlock().
252 rcu_read_unlock_special(t);
256 * Either we were not in an RCU read-side critical section to
257 * begin with, or we have now recorded that critical section
258 * globally. Either way, we can now note a quiescent state
259 * for this CPU. Again, if we were in an RCU read-side critical
260 * section, and if that critical section was blocking the current
261 * grace period, then the fact that the task has been enqueued
262 * means that we continue to block the current grace period.
264 local_irq_save(flags);
266 local_irq_restore(flags);
270 * Check for preempted RCU readers blocking the current grace period
271 * for the specified rcu_node structure. If the caller needs a reliable
272 * answer, it must hold the rcu_node's ->lock.
274 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
276 return rnp->gp_tasks != NULL;
280 * Record a quiescent state for all tasks that were previously queued
281 * on the specified rcu_node structure and that were blocking the current
282 * RCU grace period. The caller must hold the specified rnp->lock with
283 * irqs disabled, and this lock is released upon return, but irqs remain
286 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
287 __releases(rnp->lock)
290 struct rcu_node *rnp_p;
292 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
293 raw_spin_unlock_irqrestore(&rnp->lock, flags);
294 return; /* Still need more quiescent states! */
300 * Either there is only one rcu_node in the tree,
301 * or tasks were kicked up to root rcu_node due to
302 * CPUs going offline.
304 rcu_report_qs_rsp(&rcu_preempt_state, flags);
308 /* Report up the rest of the hierarchy. */
310 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
311 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
312 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
316 * Advance a ->blkd_tasks-list pointer to the next entry, instead
317 * returning NULL if at the end of the list.
319 static struct list_head *rcu_next_node_entry(struct task_struct *t,
320 struct rcu_node *rnp)
322 struct list_head *np;
324 np = t->rcu_node_entry.next;
325 if (np == &rnp->blkd_tasks)
331 * Handle special cases during rcu_read_unlock(), such as needing to
332 * notify RCU core processing or task having blocked during the RCU
333 * read-side critical section.
335 void rcu_read_unlock_special(struct task_struct *t)
341 struct list_head *np;
342 #ifdef CONFIG_RCU_BOOST
343 struct rt_mutex *rbmp = NULL;
344 #endif /* #ifdef CONFIG_RCU_BOOST */
345 struct rcu_node *rnp;
348 /* NMI handlers cannot block and cannot safely manipulate state. */
352 local_irq_save(flags);
355 * If RCU core is waiting for this CPU to exit critical section,
356 * let it know that we have done so.
358 special = t->rcu_read_unlock_special;
359 if (special & RCU_READ_UNLOCK_NEED_QS) {
360 rcu_preempt_qs(smp_processor_id());
363 /* Hardware IRQ handlers cannot block. */
364 if (in_irq() || in_serving_softirq()) {
365 local_irq_restore(flags);
369 /* Clean up if blocked during RCU read-side critical section. */
370 if (special & RCU_READ_UNLOCK_BLOCKED) {
371 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
374 * Remove this task from the list it blocked on. The
375 * task can migrate while we acquire the lock, but at
376 * most one time. So at most two passes through loop.
379 rnp = t->rcu_blocked_node;
380 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
381 if (rnp == t->rcu_blocked_node)
383 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
385 empty = !rcu_preempt_blocked_readers_cgp(rnp);
386 empty_exp = !rcu_preempted_readers_exp(rnp);
387 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
388 np = rcu_next_node_entry(t, rnp);
389 list_del_init(&t->rcu_node_entry);
390 t->rcu_blocked_node = NULL;
391 trace_rcu_unlock_preempted_task("rcu_preempt",
393 if (&t->rcu_node_entry == rnp->gp_tasks)
395 if (&t->rcu_node_entry == rnp->exp_tasks)
397 #ifdef CONFIG_RCU_BOOST
398 if (&t->rcu_node_entry == rnp->boost_tasks)
399 rnp->boost_tasks = np;
400 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
401 if (t->rcu_boost_mutex) {
402 rbmp = t->rcu_boost_mutex;
403 t->rcu_boost_mutex = NULL;
405 #endif /* #ifdef CONFIG_RCU_BOOST */
408 * If this was the last task on the current list, and if
409 * we aren't waiting on any CPUs, report the quiescent state.
410 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
411 * so we must take a snapshot of the expedited state.
413 empty_exp_now = !rcu_preempted_readers_exp(rnp);
414 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
415 trace_rcu_quiescent_state_report("preempt_rcu",
422 rcu_report_unblock_qs_rnp(rnp, flags);
424 raw_spin_unlock_irqrestore(&rnp->lock, flags);
427 #ifdef CONFIG_RCU_BOOST
428 /* Unboost if we were boosted. */
430 rt_mutex_unlock(rbmp);
431 #endif /* #ifdef CONFIG_RCU_BOOST */
434 * If this was the last task on the expedited lists,
435 * then we need to report up the rcu_node hierarchy.
437 if (!empty_exp && empty_exp_now)
438 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
440 local_irq_restore(flags);
444 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
447 * Dump detailed information for all tasks blocking the current RCU
448 * grace period on the specified rcu_node structure.
450 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
453 struct task_struct *t;
455 raw_spin_lock_irqsave(&rnp->lock, flags);
456 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
457 raw_spin_unlock_irqrestore(&rnp->lock, flags);
460 t = list_entry(rnp->gp_tasks,
461 struct task_struct, rcu_node_entry);
462 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
464 raw_spin_unlock_irqrestore(&rnp->lock, flags);
468 * Dump detailed information for all tasks blocking the current RCU
471 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
473 struct rcu_node *rnp = rcu_get_root(rsp);
475 rcu_print_detail_task_stall_rnp(rnp);
476 rcu_for_each_leaf_node(rsp, rnp)
477 rcu_print_detail_task_stall_rnp(rnp);
480 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
482 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
486 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
488 #ifdef CONFIG_RCU_CPU_STALL_INFO
490 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
492 printk(KERN_ERR "\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
493 rnp->level, rnp->grplo, rnp->grphi);
496 static void rcu_print_task_stall_end(void)
498 printk(KERN_CONT "\n");
501 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
503 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
507 static void rcu_print_task_stall_end(void)
511 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
514 * Scan the current list of tasks blocked within RCU read-side critical
515 * sections, printing out the tid of each.
517 static int rcu_print_task_stall(struct rcu_node *rnp)
519 struct task_struct *t;
522 if (!rcu_preempt_blocked_readers_cgp(rnp))
524 rcu_print_task_stall_begin(rnp);
525 t = list_entry(rnp->gp_tasks,
526 struct task_struct, rcu_node_entry);
527 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
528 printk(KERN_CONT " P%d", t->pid);
531 rcu_print_task_stall_end();
536 * Check that the list of blocked tasks for the newly completed grace
537 * period is in fact empty. It is a serious bug to complete a grace
538 * period that still has RCU readers blocked! This function must be
539 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
540 * must be held by the caller.
542 * Also, if there are blocked tasks on the list, they automatically
543 * block the newly created grace period, so set up ->gp_tasks accordingly.
545 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
547 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
548 if (!list_empty(&rnp->blkd_tasks))
549 rnp->gp_tasks = rnp->blkd_tasks.next;
550 WARN_ON_ONCE(rnp->qsmask);
553 #ifdef CONFIG_HOTPLUG_CPU
556 * Handle tasklist migration for case in which all CPUs covered by the
557 * specified rcu_node have gone offline. Move them up to the root
558 * rcu_node. The reason for not just moving them to the immediate
559 * parent is to remove the need for rcu_read_unlock_special() to
560 * make more than two attempts to acquire the target rcu_node's lock.
561 * Returns true if there were tasks blocking the current RCU grace
564 * Returns 1 if there was previously a task blocking the current grace
565 * period on the specified rcu_node structure.
567 * The caller must hold rnp->lock with irqs disabled.
569 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
570 struct rcu_node *rnp,
571 struct rcu_data *rdp)
573 struct list_head *lp;
574 struct list_head *lp_root;
576 struct rcu_node *rnp_root = rcu_get_root(rsp);
577 struct task_struct *t;
579 if (rnp == rnp_root) {
580 WARN_ONCE(1, "Last CPU thought to be offlined?");
581 return 0; /* Shouldn't happen: at least one CPU online. */
584 /* If we are on an internal node, complain bitterly. */
585 WARN_ON_ONCE(rnp != rdp->mynode);
588 * Move tasks up to root rcu_node. Don't try to get fancy for
589 * this corner-case operation -- just put this node's tasks
590 * at the head of the root node's list, and update the root node's
591 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
592 * if non-NULL. This might result in waiting for more tasks than
593 * absolutely necessary, but this is a good performance/complexity
596 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
597 retval |= RCU_OFL_TASKS_NORM_GP;
598 if (rcu_preempted_readers_exp(rnp))
599 retval |= RCU_OFL_TASKS_EXP_GP;
600 lp = &rnp->blkd_tasks;
601 lp_root = &rnp_root->blkd_tasks;
602 while (!list_empty(lp)) {
603 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
604 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
605 list_del(&t->rcu_node_entry);
606 t->rcu_blocked_node = rnp_root;
607 list_add(&t->rcu_node_entry, lp_root);
608 if (&t->rcu_node_entry == rnp->gp_tasks)
609 rnp_root->gp_tasks = rnp->gp_tasks;
610 if (&t->rcu_node_entry == rnp->exp_tasks)
611 rnp_root->exp_tasks = rnp->exp_tasks;
612 #ifdef CONFIG_RCU_BOOST
613 if (&t->rcu_node_entry == rnp->boost_tasks)
614 rnp_root->boost_tasks = rnp->boost_tasks;
615 #endif /* #ifdef CONFIG_RCU_BOOST */
616 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
619 rnp->gp_tasks = NULL;
620 rnp->exp_tasks = NULL;
621 #ifdef CONFIG_RCU_BOOST
622 rnp->boost_tasks = NULL;
624 * In case root is being boosted and leaf was not. Make sure
625 * that we boost the tasks blocking the current grace period
628 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
629 if (rnp_root->boost_tasks != NULL &&
630 rnp_root->boost_tasks != rnp_root->gp_tasks &&
631 rnp_root->boost_tasks != rnp_root->exp_tasks)
632 rnp_root->boost_tasks = rnp_root->gp_tasks;
633 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
634 #endif /* #ifdef CONFIG_RCU_BOOST */
639 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
642 * Check for a quiescent state from the current CPU. When a task blocks,
643 * the task is recorded in the corresponding CPU's rcu_node structure,
644 * which is checked elsewhere.
646 * Caller must disable hard irqs.
648 static void rcu_preempt_check_callbacks(int cpu)
650 struct task_struct *t = current;
652 if (t->rcu_read_lock_nesting == 0) {
656 if (t->rcu_read_lock_nesting > 0 &&
657 per_cpu(rcu_preempt_data, cpu).qs_pending)
658 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
661 #ifdef CONFIG_RCU_BOOST
663 static void rcu_preempt_do_callbacks(void)
665 rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
668 #endif /* #ifdef CONFIG_RCU_BOOST */
671 * Queue a preemptible-RCU callback for invocation after a grace period.
673 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
675 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
677 EXPORT_SYMBOL_GPL(call_rcu);
680 * Queue an RCU callback for lazy invocation after a grace period.
681 * This will likely be later named something like "call_rcu_lazy()",
682 * but this change will require some way of tagging the lazy RCU
683 * callbacks in the list of pending callbacks. Until then, this
684 * function may only be called from __kfree_rcu().
686 void kfree_call_rcu(struct rcu_head *head,
687 void (*func)(struct rcu_head *rcu))
689 __call_rcu(head, func, &rcu_preempt_state, -1, 1);
691 EXPORT_SYMBOL_GPL(kfree_call_rcu);
694 * synchronize_rcu - wait until a grace period has elapsed.
696 * Control will return to the caller some time after a full grace
697 * period has elapsed, in other words after all currently executing RCU
698 * read-side critical sections have completed. Note, however, that
699 * upon return from synchronize_rcu(), the caller might well be executing
700 * concurrently with new RCU read-side critical sections that began while
701 * synchronize_rcu() was waiting. RCU read-side critical sections are
702 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
704 * See the description of synchronize_sched() for more detailed information
705 * on memory ordering guarantees.
707 void synchronize_rcu(void)
709 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
710 !lock_is_held(&rcu_lock_map) &&
711 !lock_is_held(&rcu_sched_lock_map),
712 "Illegal synchronize_rcu() in RCU read-side critical section");
713 if (!rcu_scheduler_active)
716 synchronize_rcu_expedited();
718 wait_rcu_gp(call_rcu);
720 EXPORT_SYMBOL_GPL(synchronize_rcu);
722 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
723 static unsigned long sync_rcu_preempt_exp_count;
724 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
727 * Return non-zero if there are any tasks in RCU read-side critical
728 * sections blocking the current preemptible-RCU expedited grace period.
729 * If there is no preemptible-RCU expedited grace period currently in
730 * progress, returns zero unconditionally.
732 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
734 return rnp->exp_tasks != NULL;
738 * return non-zero if there is no RCU expedited grace period in progress
739 * for the specified rcu_node structure, in other words, if all CPUs and
740 * tasks covered by the specified rcu_node structure have done their bit
741 * for the current expedited grace period. Works only for preemptible
742 * RCU -- other RCU implementation use other means.
744 * Caller must hold sync_rcu_preempt_exp_mutex.
746 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
748 return !rcu_preempted_readers_exp(rnp) &&
749 ACCESS_ONCE(rnp->expmask) == 0;
753 * Report the exit from RCU read-side critical section for the last task
754 * that queued itself during or before the current expedited preemptible-RCU
755 * grace period. This event is reported either to the rcu_node structure on
756 * which the task was queued or to one of that rcu_node structure's ancestors,
757 * recursively up the tree. (Calm down, calm down, we do the recursion
760 * Most callers will set the "wake" flag, but the task initiating the
761 * expedited grace period need not wake itself.
763 * Caller must hold sync_rcu_preempt_exp_mutex.
765 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
771 raw_spin_lock_irqsave(&rnp->lock, flags);
773 if (!sync_rcu_preempt_exp_done(rnp)) {
774 raw_spin_unlock_irqrestore(&rnp->lock, flags);
777 if (rnp->parent == NULL) {
778 raw_spin_unlock_irqrestore(&rnp->lock, flags);
780 wake_up(&sync_rcu_preempt_exp_wq);
784 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
786 raw_spin_lock(&rnp->lock); /* irqs already disabled */
787 rnp->expmask &= ~mask;
792 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
793 * grace period for the specified rcu_node structure. If there are no such
794 * tasks, report it up the rcu_node hierarchy.
796 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
797 * CPU hotplug operations.
800 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
805 raw_spin_lock_irqsave(&rnp->lock, flags);
806 if (list_empty(&rnp->blkd_tasks)) {
807 raw_spin_unlock_irqrestore(&rnp->lock, flags);
809 rnp->exp_tasks = rnp->blkd_tasks.next;
810 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
814 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
818 * synchronize_rcu_expedited - Brute-force RCU grace period
820 * Wait for an RCU-preempt grace period, but expedite it. The basic
821 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
822 * the ->blkd_tasks lists and wait for this list to drain. This consumes
823 * significant time on all CPUs and is unfriendly to real-time workloads,
824 * so is thus not recommended for any sort of common-case code.
825 * In fact, if you are using synchronize_rcu_expedited() in a loop,
826 * please restructure your code to batch your updates, and then Use a
827 * single synchronize_rcu() instead.
829 * Note that it is illegal to call this function while holding any lock
830 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
831 * to call this function from a CPU-hotplug notifier. Failing to observe
832 * these restriction will result in deadlock.
834 void synchronize_rcu_expedited(void)
837 struct rcu_node *rnp;
838 struct rcu_state *rsp = &rcu_preempt_state;
842 smp_mb(); /* Caller's modifications seen first by other CPUs. */
843 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
844 smp_mb(); /* Above access cannot bleed into critical section. */
847 * Block CPU-hotplug operations. This means that any CPU-hotplug
848 * operation that finds an rcu_node structure with tasks in the
849 * process of being boosted will know that all tasks blocking
850 * this expedited grace period will already be in the process of
851 * being boosted. This simplifies the process of moving tasks
852 * from leaf to root rcu_node structures.
857 * Acquire lock, falling back to synchronize_rcu() if too many
858 * lock-acquisition failures. Of course, if someone does the
859 * expedited grace period for us, just leave.
861 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
862 if (ULONG_CMP_LT(snap,
863 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
865 goto mb_ret; /* Others did our work for us. */
867 if (trycount++ < 10) {
868 udelay(trycount * num_online_cpus());
871 wait_rcu_gp(call_rcu);
875 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
877 goto unlock_mb_ret; /* Others did our work for us. */
880 /* force all RCU readers onto ->blkd_tasks lists. */
881 synchronize_sched_expedited();
883 /* Initialize ->expmask for all non-leaf rcu_node structures. */
884 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
885 raw_spin_lock_irqsave(&rnp->lock, flags);
886 rnp->expmask = rnp->qsmaskinit;
887 raw_spin_unlock_irqrestore(&rnp->lock, flags);
890 /* Snapshot current state of ->blkd_tasks lists. */
891 rcu_for_each_leaf_node(rsp, rnp)
892 sync_rcu_preempt_exp_init(rsp, rnp);
893 if (NUM_RCU_NODES > 1)
894 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
898 /* Wait for snapshotted ->blkd_tasks lists to drain. */
899 rnp = rcu_get_root(rsp);
900 wait_event(sync_rcu_preempt_exp_wq,
901 sync_rcu_preempt_exp_done(rnp));
903 /* Clean up and exit. */
904 smp_mb(); /* ensure expedited GP seen before counter increment. */
905 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
907 mutex_unlock(&sync_rcu_preempt_exp_mutex);
909 smp_mb(); /* ensure subsequent action seen after grace period. */
911 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
914 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
916 * Note that this primitive does not necessarily wait for an RCU grace period
917 * to complete. For example, if there are no RCU callbacks queued anywhere
918 * in the system, then rcu_barrier() is within its rights to return
919 * immediately, without waiting for anything, much less an RCU grace period.
921 void rcu_barrier(void)
923 _rcu_barrier(&rcu_preempt_state);
925 EXPORT_SYMBOL_GPL(rcu_barrier);
928 * Initialize preemptible RCU's state structures.
930 static void __init __rcu_init_preempt(void)
932 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
935 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
937 static struct rcu_state *rcu_state = &rcu_sched_state;
940 * Tell them what RCU they are running.
942 static void __init rcu_bootup_announce(void)
944 printk(KERN_INFO "Hierarchical RCU implementation.\n");
945 rcu_bootup_announce_oddness();
949 * Return the number of RCU batches processed thus far for debug & stats.
951 long rcu_batches_completed(void)
953 return rcu_batches_completed_sched();
955 EXPORT_SYMBOL_GPL(rcu_batches_completed);
958 * Force a quiescent state for RCU, which, because there is no preemptible
959 * RCU, becomes the same as rcu-sched.
961 void rcu_force_quiescent_state(void)
963 rcu_sched_force_quiescent_state();
965 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
968 * Because preemptible RCU does not exist, we never have to check for
969 * CPUs being in quiescent states.
971 static void rcu_preempt_note_context_switch(int cpu)
976 * Because preemptible RCU does not exist, there are never any preempted
979 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
984 #ifdef CONFIG_HOTPLUG_CPU
986 /* Because preemptible RCU does not exist, no quieting of tasks. */
987 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
989 raw_spin_unlock_irqrestore(&rnp->lock, flags);
992 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
995 * Because preemptible RCU does not exist, we never have to check for
996 * tasks blocked within RCU read-side critical sections.
998 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1003 * Because preemptible RCU does not exist, we never have to check for
1004 * tasks blocked within RCU read-side critical sections.
1006 static int rcu_print_task_stall(struct rcu_node *rnp)
1012 * Because there is no preemptible RCU, there can be no readers blocked,
1013 * so there is no need to check for blocked tasks. So check only for
1014 * bogus qsmask values.
1016 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1018 WARN_ON_ONCE(rnp->qsmask);
1021 #ifdef CONFIG_HOTPLUG_CPU
1024 * Because preemptible RCU does not exist, it never needs to migrate
1025 * tasks that were blocked within RCU read-side critical sections, and
1026 * such non-existent tasks cannot possibly have been blocking the current
1029 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1030 struct rcu_node *rnp,
1031 struct rcu_data *rdp)
1036 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1039 * Because preemptible RCU does not exist, it never has any callbacks
1042 static void rcu_preempt_check_callbacks(int cpu)
1047 * Queue an RCU callback for lazy invocation after a grace period.
1048 * This will likely be later named something like "call_rcu_lazy()",
1049 * but this change will require some way of tagging the lazy RCU
1050 * callbacks in the list of pending callbacks. Until then, this
1051 * function may only be called from __kfree_rcu().
1053 * Because there is no preemptible RCU, we use RCU-sched instead.
1055 void kfree_call_rcu(struct rcu_head *head,
1056 void (*func)(struct rcu_head *rcu))
1058 __call_rcu(head, func, &rcu_sched_state, -1, 1);
1060 EXPORT_SYMBOL_GPL(kfree_call_rcu);
1063 * Wait for an rcu-preempt grace period, but make it happen quickly.
1064 * But because preemptible RCU does not exist, map to rcu-sched.
1066 void synchronize_rcu_expedited(void)
1068 synchronize_sched_expedited();
1070 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1072 #ifdef CONFIG_HOTPLUG_CPU
1075 * Because preemptible RCU does not exist, there is never any need to
1076 * report on tasks preempted in RCU read-side critical sections during
1077 * expedited RCU grace periods.
1079 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1084 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1087 * Because preemptible RCU does not exist, rcu_barrier() is just
1088 * another name for rcu_barrier_sched().
1090 void rcu_barrier(void)
1092 rcu_barrier_sched();
1094 EXPORT_SYMBOL_GPL(rcu_barrier);
1097 * Because preemptible RCU does not exist, it need not be initialized.
1099 static void __init __rcu_init_preempt(void)
1103 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1105 #ifdef CONFIG_RCU_BOOST
1107 #include "rtmutex_common.h"
1109 #ifdef CONFIG_RCU_TRACE
1111 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1113 if (list_empty(&rnp->blkd_tasks))
1114 rnp->n_balk_blkd_tasks++;
1115 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1116 rnp->n_balk_exp_gp_tasks++;
1117 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1118 rnp->n_balk_boost_tasks++;
1119 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1120 rnp->n_balk_notblocked++;
1121 else if (rnp->gp_tasks != NULL &&
1122 ULONG_CMP_LT(jiffies, rnp->boost_time))
1123 rnp->n_balk_notyet++;
1128 #else /* #ifdef CONFIG_RCU_TRACE */
1130 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1134 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1136 static void rcu_wake_cond(struct task_struct *t, int status)
1139 * If the thread is yielding, only wake it when this
1140 * is invoked from idle
1142 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1147 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1148 * or ->boost_tasks, advancing the pointer to the next task in the
1149 * ->blkd_tasks list.
1151 * Note that irqs must be enabled: boosting the task can block.
1152 * Returns 1 if there are more tasks needing to be boosted.
1154 static int rcu_boost(struct rcu_node *rnp)
1156 unsigned long flags;
1157 struct rt_mutex mtx;
1158 struct task_struct *t;
1159 struct list_head *tb;
1161 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1162 return 0; /* Nothing left to boost. */
1164 raw_spin_lock_irqsave(&rnp->lock, flags);
1167 * Recheck under the lock: all tasks in need of boosting
1168 * might exit their RCU read-side critical sections on their own.
1170 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1171 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1176 * Preferentially boost tasks blocking expedited grace periods.
1177 * This cannot starve the normal grace periods because a second
1178 * expedited grace period must boost all blocked tasks, including
1179 * those blocking the pre-existing normal grace period.
1181 if (rnp->exp_tasks != NULL) {
1182 tb = rnp->exp_tasks;
1183 rnp->n_exp_boosts++;
1185 tb = rnp->boost_tasks;
1186 rnp->n_normal_boosts++;
1188 rnp->n_tasks_boosted++;
1191 * We boost task t by manufacturing an rt_mutex that appears to
1192 * be held by task t. We leave a pointer to that rt_mutex where
1193 * task t can find it, and task t will release the mutex when it
1194 * exits its outermost RCU read-side critical section. Then
1195 * simply acquiring this artificial rt_mutex will boost task
1196 * t's priority. (Thanks to tglx for suggesting this approach!)
1198 * Note that task t must acquire rnp->lock to remove itself from
1199 * the ->blkd_tasks list, which it will do from exit() if from
1200 * nowhere else. We therefore are guaranteed that task t will
1201 * stay around at least until we drop rnp->lock. Note that
1202 * rnp->lock also resolves races between our priority boosting
1203 * and task t's exiting its outermost RCU read-side critical
1206 t = container_of(tb, struct task_struct, rcu_node_entry);
1207 rt_mutex_init_proxy_locked(&mtx, t);
1208 t->rcu_boost_mutex = &mtx;
1209 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1210 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1211 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1213 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1214 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1218 * Priority-boosting kthread. One per leaf rcu_node and one for the
1221 static int rcu_boost_kthread(void *arg)
1223 struct rcu_node *rnp = (struct rcu_node *)arg;
1227 trace_rcu_utilization("Start boost kthread@init");
1229 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1230 trace_rcu_utilization("End boost kthread@rcu_wait");
1231 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1232 trace_rcu_utilization("Start boost kthread@rcu_wait");
1233 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1234 more2boost = rcu_boost(rnp);
1240 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1241 trace_rcu_utilization("End boost kthread@rcu_yield");
1242 schedule_timeout_interruptible(2);
1243 trace_rcu_utilization("Start boost kthread@rcu_yield");
1248 trace_rcu_utilization("End boost kthread@notreached");
1253 * Check to see if it is time to start boosting RCU readers that are
1254 * blocking the current grace period, and, if so, tell the per-rcu_node
1255 * kthread to start boosting them. If there is an expedited grace
1256 * period in progress, it is always time to boost.
1258 * The caller must hold rnp->lock, which this function releases.
1259 * The ->boost_kthread_task is immortal, so we don't need to worry
1260 * about it going away.
1262 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1264 struct task_struct *t;
1266 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1267 rnp->n_balk_exp_gp_tasks++;
1268 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1271 if (rnp->exp_tasks != NULL ||
1272 (rnp->gp_tasks != NULL &&
1273 rnp->boost_tasks == NULL &&
1275 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1276 if (rnp->exp_tasks == NULL)
1277 rnp->boost_tasks = rnp->gp_tasks;
1278 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1279 t = rnp->boost_kthread_task;
1281 rcu_wake_cond(t, rnp->boost_kthread_status);
1283 rcu_initiate_boost_trace(rnp);
1284 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1289 * Wake up the per-CPU kthread to invoke RCU callbacks.
1291 static void invoke_rcu_callbacks_kthread(void)
1293 unsigned long flags;
1295 local_irq_save(flags);
1296 __this_cpu_write(rcu_cpu_has_work, 1);
1297 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1298 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1299 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1300 __this_cpu_read(rcu_cpu_kthread_status));
1302 local_irq_restore(flags);
1306 * Is the current CPU running the RCU-callbacks kthread?
1307 * Caller must have preemption disabled.
1309 static bool rcu_is_callbacks_kthread(void)
1311 return __get_cpu_var(rcu_cpu_kthread_task) == current;
1314 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1317 * Do priority-boost accounting for the start of a new grace period.
1319 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1321 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1325 * Create an RCU-boost kthread for the specified node if one does not
1326 * already exist. We only create this kthread for preemptible RCU.
1327 * Returns zero if all is well, a negated errno otherwise.
1329 static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1330 struct rcu_node *rnp)
1332 int rnp_index = rnp - &rsp->node[0];
1333 unsigned long flags;
1334 struct sched_param sp;
1335 struct task_struct *t;
1337 if (&rcu_preempt_state != rsp)
1340 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1344 if (rnp->boost_kthread_task != NULL)
1346 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1347 "rcub/%d", rnp_index);
1350 raw_spin_lock_irqsave(&rnp->lock, flags);
1351 rnp->boost_kthread_task = t;
1352 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1353 sp.sched_priority = RCU_BOOST_PRIO;
1354 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1355 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1359 static void rcu_kthread_do_work(void)
1361 rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1362 rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1363 rcu_preempt_do_callbacks();
1366 static void rcu_cpu_kthread_setup(unsigned int cpu)
1368 struct sched_param sp;
1370 sp.sched_priority = RCU_KTHREAD_PRIO;
1371 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1374 static void rcu_cpu_kthread_park(unsigned int cpu)
1376 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1379 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1381 return __get_cpu_var(rcu_cpu_has_work);
1385 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1386 * RCU softirq used in flavors and configurations of RCU that do not
1387 * support RCU priority boosting.
1389 static void rcu_cpu_kthread(unsigned int cpu)
1391 unsigned int *statusp = &__get_cpu_var(rcu_cpu_kthread_status);
1392 char work, *workp = &__get_cpu_var(rcu_cpu_has_work);
1395 for (spincnt = 0; spincnt < 10; spincnt++) {
1396 trace_rcu_utilization("Start CPU kthread@rcu_wait");
1398 *statusp = RCU_KTHREAD_RUNNING;
1399 this_cpu_inc(rcu_cpu_kthread_loops);
1400 local_irq_disable();
1405 rcu_kthread_do_work();
1408 trace_rcu_utilization("End CPU kthread@rcu_wait");
1409 *statusp = RCU_KTHREAD_WAITING;
1413 *statusp = RCU_KTHREAD_YIELDING;
1414 trace_rcu_utilization("Start CPU kthread@rcu_yield");
1415 schedule_timeout_interruptible(2);
1416 trace_rcu_utilization("End CPU kthread@rcu_yield");
1417 *statusp = RCU_KTHREAD_WAITING;
1421 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1422 * served by the rcu_node in question. The CPU hotplug lock is still
1423 * held, so the value of rnp->qsmaskinit will be stable.
1425 * We don't include outgoingcpu in the affinity set, use -1 if there is
1426 * no outgoing CPU. If there are no CPUs left in the affinity set,
1427 * this function allows the kthread to execute on any CPU.
1429 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1431 struct task_struct *t = rnp->boost_kthread_task;
1432 unsigned long mask = rnp->qsmaskinit;
1438 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1440 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1441 if ((mask & 0x1) && cpu != outgoingcpu)
1442 cpumask_set_cpu(cpu, cm);
1443 if (cpumask_weight(cm) == 0) {
1445 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1446 cpumask_clear_cpu(cpu, cm);
1447 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1449 set_cpus_allowed_ptr(t, cm);
1450 free_cpumask_var(cm);
1453 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1454 .store = &rcu_cpu_kthread_task,
1455 .thread_should_run = rcu_cpu_kthread_should_run,
1456 .thread_fn = rcu_cpu_kthread,
1457 .thread_comm = "rcuc/%u",
1458 .setup = rcu_cpu_kthread_setup,
1459 .park = rcu_cpu_kthread_park,
1463 * Spawn all kthreads -- called as soon as the scheduler is running.
1465 static int __init rcu_spawn_kthreads(void)
1467 struct rcu_node *rnp;
1470 rcu_scheduler_fully_active = 1;
1471 for_each_possible_cpu(cpu)
1472 per_cpu(rcu_cpu_has_work, cpu) = 0;
1473 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1474 rnp = rcu_get_root(rcu_state);
1475 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1476 if (NUM_RCU_NODES > 1) {
1477 rcu_for_each_leaf_node(rcu_state, rnp)
1478 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1482 early_initcall(rcu_spawn_kthreads);
1484 static void __cpuinit rcu_prepare_kthreads(int cpu)
1486 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1487 struct rcu_node *rnp = rdp->mynode;
1489 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1490 if (rcu_scheduler_fully_active)
1491 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1494 #else /* #ifdef CONFIG_RCU_BOOST */
1496 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1498 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1501 static void invoke_rcu_callbacks_kthread(void)
1506 static bool rcu_is_callbacks_kthread(void)
1511 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1515 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1519 static int __init rcu_scheduler_really_started(void)
1521 rcu_scheduler_fully_active = 1;
1524 early_initcall(rcu_scheduler_really_started);
1526 static void __cpuinit rcu_prepare_kthreads(int cpu)
1530 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1532 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1535 * Check to see if any future RCU-related work will need to be done
1536 * by the current CPU, even if none need be done immediately, returning
1537 * 1 if so. This function is part of the RCU implementation; it is -not-
1538 * an exported member of the RCU API.
1540 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1541 * any flavor of RCU.
1543 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1545 *delta_jiffies = ULONG_MAX;
1546 return rcu_cpu_has_callbacks(cpu);
1550 * Because we do not have RCU_FAST_NO_HZ, don't bother initializing for it.
1552 static void rcu_prepare_for_idle_init(int cpu)
1557 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1560 static void rcu_cleanup_after_idle(int cpu)
1565 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1568 static void rcu_prepare_for_idle(int cpu)
1573 * Don't bother keeping a running count of the number of RCU callbacks
1574 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1576 static void rcu_idle_count_callbacks_posted(void)
1580 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1583 * This code is invoked when a CPU goes idle, at which point we want
1584 * to have the CPU do everything required for RCU so that it can enter
1585 * the energy-efficient dyntick-idle mode. This is handled by a
1586 * state machine implemented by rcu_prepare_for_idle() below.
1588 * The following three proprocessor symbols control this state machine:
1590 * RCU_IDLE_FLUSHES gives the maximum number of times that we will attempt
1591 * to satisfy RCU. Beyond this point, it is better to incur a periodic
1592 * scheduling-clock interrupt than to loop through the state machine
1594 * RCU_IDLE_OPT_FLUSHES gives the number of RCU_IDLE_FLUSHES that are
1595 * optional if RCU does not need anything immediately from this
1596 * CPU, even if this CPU still has RCU callbacks queued. The first
1597 * times through the state machine are mandatory: we need to give
1598 * the state machine a chance to communicate a quiescent state
1600 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1601 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1602 * is sized to be roughly one RCU grace period. Those energy-efficiency
1603 * benchmarkers who might otherwise be tempted to set this to a large
1604 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1605 * system. And if you are -that- concerned about energy efficiency,
1606 * just power the system down and be done with it!
1607 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1608 * permitted to sleep in dyntick-idle mode with only lazy RCU
1609 * callbacks pending. Setting this too high can OOM your system.
1611 * The values below work well in practice. If future workloads require
1612 * adjustment, they can be converted into kernel config parameters, though
1613 * making the state machine smarter might be a better option.
1615 #define RCU_IDLE_FLUSHES 5 /* Number of dyntick-idle tries. */
1616 #define RCU_IDLE_OPT_FLUSHES 3 /* Optional dyntick-idle tries. */
1617 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1618 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1620 extern int tick_nohz_enabled;
1623 * Does the specified flavor of RCU have non-lazy callbacks pending on
1624 * the specified CPU? Both RCU flavor and CPU are specified by the
1625 * rcu_data structure.
1627 static bool __rcu_cpu_has_nonlazy_callbacks(struct rcu_data *rdp)
1629 return rdp->qlen != rdp->qlen_lazy;
1632 #ifdef CONFIG_TREE_PREEMPT_RCU
1635 * Are there non-lazy RCU-preempt callbacks? (There cannot be if there
1636 * is no RCU-preempt in the kernel.)
1638 static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu)
1640 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
1642 return __rcu_cpu_has_nonlazy_callbacks(rdp);
1645 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
1647 static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu)
1652 #endif /* else #ifdef CONFIG_TREE_PREEMPT_RCU */
1655 * Does any flavor of RCU have non-lazy callbacks on the specified CPU?
1657 static bool rcu_cpu_has_nonlazy_callbacks(int cpu)
1659 return __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_sched_data, cpu)) ||
1660 __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_bh_data, cpu)) ||
1661 rcu_preempt_cpu_has_nonlazy_callbacks(cpu);
1665 * Allow the CPU to enter dyntick-idle mode if either: (1) There are no
1666 * callbacks on this CPU, (2) this CPU has not yet attempted to enter
1667 * dyntick-idle mode, or (3) this CPU is in the process of attempting to
1668 * enter dyntick-idle mode. Otherwise, if we have recently tried and failed
1669 * to enter dyntick-idle mode, we refuse to try to enter it. After all,
1670 * it is better to incur scheduling-clock interrupts than to spin
1671 * continuously for the same time duration!
1673 * The delta_jiffies argument is used to store the time when RCU is
1674 * going to need the CPU again if it still has callbacks. The reason
1675 * for this is that rcu_prepare_for_idle() might need to post a timer,
1676 * but if so, it will do so after tick_nohz_stop_sched_tick() has set
1677 * the wakeup time for this CPU. This means that RCU's timer can be
1678 * delayed until the wakeup time, which defeats the purpose of posting
1681 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1683 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1685 /* Flag a new idle sojourn to the idle-entry state machine. */
1686 rdtp->idle_first_pass = 1;
1687 /* If no callbacks, RCU doesn't need the CPU. */
1688 if (!rcu_cpu_has_callbacks(cpu)) {
1689 *delta_jiffies = ULONG_MAX;
1692 if (rdtp->dyntick_holdoff == jiffies) {
1693 /* RCU recently tried and failed, so don't try again. */
1697 /* Set up for the possibility that RCU will post a timer. */
1698 if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
1699 *delta_jiffies = round_up(RCU_IDLE_GP_DELAY + jiffies,
1700 RCU_IDLE_GP_DELAY) - jiffies;
1702 *delta_jiffies = jiffies + RCU_IDLE_LAZY_GP_DELAY;
1703 *delta_jiffies = round_jiffies(*delta_jiffies) - jiffies;
1709 * Handler for smp_call_function_single(). The only point of this
1710 * handler is to wake the CPU up, so the handler does only tracing.
1712 void rcu_idle_demigrate(void *unused)
1714 trace_rcu_prep_idle("Demigrate");
1718 * Timer handler used to force CPU to start pushing its remaining RCU
1719 * callbacks in the case where it entered dyntick-idle mode with callbacks
1720 * pending. The hander doesn't really need to do anything because the
1721 * real work is done upon re-entry to idle, or by the next scheduling-clock
1722 * interrupt should idle not be re-entered.
1724 * One special case: the timer gets migrated without awakening the CPU
1725 * on which the timer was scheduled on. In this case, we must wake up
1726 * that CPU. We do so with smp_call_function_single().
1728 static void rcu_idle_gp_timer_func(unsigned long cpu_in)
1730 int cpu = (int)cpu_in;
1732 trace_rcu_prep_idle("Timer");
1733 if (cpu != smp_processor_id())
1734 smp_call_function_single(cpu, rcu_idle_demigrate, NULL, 0);
1736 WARN_ON_ONCE(1); /* Getting here can hang the system... */
1740 * Initialize the timer used to pull CPUs out of dyntick-idle mode.
1742 static void rcu_prepare_for_idle_init(int cpu)
1744 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1746 rdtp->dyntick_holdoff = jiffies - 1;
1747 setup_timer(&rdtp->idle_gp_timer, rcu_idle_gp_timer_func, cpu);
1748 rdtp->idle_gp_timer_expires = jiffies - 1;
1749 rdtp->idle_first_pass = 1;
1753 * Clean up for exit from idle. Because we are exiting from idle, there
1754 * is no longer any point to ->idle_gp_timer, so cancel it. This will
1755 * do nothing if this timer is not active, so just cancel it unconditionally.
1757 static void rcu_cleanup_after_idle(int cpu)
1759 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1761 del_timer(&rdtp->idle_gp_timer);
1762 trace_rcu_prep_idle("Cleanup after idle");
1763 rdtp->tick_nohz_enabled_snap = ACCESS_ONCE(tick_nohz_enabled);
1767 * Check to see if any RCU-related work can be done by the current CPU,
1768 * and if so, schedule a softirq to get it done. This function is part
1769 * of the RCU implementation; it is -not- an exported member of the RCU API.
1771 * The idea is for the current CPU to clear out all work required by the
1772 * RCU core for the current grace period, so that this CPU can be permitted
1773 * to enter dyntick-idle mode. In some cases, it will need to be awakened
1774 * at the end of the grace period by whatever CPU ends the grace period.
1775 * This allows CPUs to go dyntick-idle more quickly, and to reduce the
1776 * number of wakeups by a modest integer factor.
1778 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1779 * disabled, we do one pass of force_quiescent_state(), then do a
1780 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
1781 * later. The ->dyntick_drain field controls the sequencing.
1783 * The caller must have disabled interrupts.
1785 static void rcu_prepare_for_idle(int cpu)
1787 struct timer_list *tp;
1788 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1791 /* Handle nohz enablement switches conservatively. */
1792 tne = ACCESS_ONCE(tick_nohz_enabled);
1793 if (tne != rdtp->tick_nohz_enabled_snap) {
1794 if (rcu_cpu_has_callbacks(cpu))
1795 invoke_rcu_core(); /* force nohz to see update. */
1796 rdtp->tick_nohz_enabled_snap = tne;
1802 /* Adaptive-tick mode, where usermode execution is idle to RCU. */
1803 if (!is_idle_task(current)) {
1804 rdtp->dyntick_holdoff = jiffies - 1;
1805 if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
1806 trace_rcu_prep_idle("User dyntick with callbacks");
1807 rdtp->idle_gp_timer_expires =
1808 round_up(jiffies + RCU_IDLE_GP_DELAY,
1810 } else if (rcu_cpu_has_callbacks(cpu)) {
1811 rdtp->idle_gp_timer_expires =
1812 round_jiffies(jiffies + RCU_IDLE_LAZY_GP_DELAY);
1813 trace_rcu_prep_idle("User dyntick with lazy callbacks");
1817 tp = &rdtp->idle_gp_timer;
1818 mod_timer_pinned(tp, rdtp->idle_gp_timer_expires);
1823 * If this is an idle re-entry, for example, due to use of
1824 * RCU_NONIDLE() or the new idle-loop tracing API within the idle
1825 * loop, then don't take any state-machine actions, unless the
1826 * momentary exit from idle queued additional non-lazy callbacks.
1827 * Instead, repost the ->idle_gp_timer if this CPU has callbacks
1830 if (!rdtp->idle_first_pass &&
1831 (rdtp->nonlazy_posted == rdtp->nonlazy_posted_snap)) {
1832 if (rcu_cpu_has_callbacks(cpu)) {
1833 tp = &rdtp->idle_gp_timer;
1834 mod_timer_pinned(tp, rdtp->idle_gp_timer_expires);
1838 rdtp->idle_first_pass = 0;
1839 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted - 1;
1842 * If there are no callbacks on this CPU, enter dyntick-idle mode.
1843 * Also reset state to avoid prejudicing later attempts.
1845 if (!rcu_cpu_has_callbacks(cpu)) {
1846 rdtp->dyntick_holdoff = jiffies - 1;
1847 rdtp->dyntick_drain = 0;
1848 trace_rcu_prep_idle("No callbacks");
1853 * If in holdoff mode, just return. We will presumably have
1854 * refrained from disabling the scheduling-clock tick.
1856 if (rdtp->dyntick_holdoff == jiffies) {
1857 trace_rcu_prep_idle("In holdoff");
1861 /* Check and update the ->dyntick_drain sequencing. */
1862 if (rdtp->dyntick_drain <= 0) {
1863 /* First time through, initialize the counter. */
1864 rdtp->dyntick_drain = RCU_IDLE_FLUSHES;
1865 } else if (rdtp->dyntick_drain <= RCU_IDLE_OPT_FLUSHES &&
1866 !rcu_pending(cpu) &&
1867 !local_softirq_pending()) {
1868 /* Can we go dyntick-idle despite still having callbacks? */
1869 rdtp->dyntick_drain = 0;
1870 rdtp->dyntick_holdoff = jiffies;
1871 if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
1872 trace_rcu_prep_idle("Dyntick with callbacks");
1873 rdtp->idle_gp_timer_expires =
1874 round_up(jiffies + RCU_IDLE_GP_DELAY,
1877 rdtp->idle_gp_timer_expires =
1878 round_jiffies(jiffies + RCU_IDLE_LAZY_GP_DELAY);
1879 trace_rcu_prep_idle("Dyntick with lazy callbacks");
1881 tp = &rdtp->idle_gp_timer;
1882 mod_timer_pinned(tp, rdtp->idle_gp_timer_expires);
1883 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1884 return; /* Nothing more to do immediately. */
1885 } else if (--(rdtp->dyntick_drain) <= 0) {
1886 /* We have hit the limit, so time to give up. */
1887 rdtp->dyntick_holdoff = jiffies;
1888 trace_rcu_prep_idle("Begin holdoff");
1889 invoke_rcu_core(); /* Force the CPU out of dyntick-idle. */
1894 * Do one step of pushing the remaining RCU callbacks through
1895 * the RCU core state machine.
1897 #ifdef CONFIG_TREE_PREEMPT_RCU
1898 if (per_cpu(rcu_preempt_data, cpu).nxtlist) {
1899 rcu_preempt_qs(cpu);
1900 force_quiescent_state(&rcu_preempt_state);
1902 #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
1903 if (per_cpu(rcu_sched_data, cpu).nxtlist) {
1905 force_quiescent_state(&rcu_sched_state);
1907 if (per_cpu(rcu_bh_data, cpu).nxtlist) {
1909 force_quiescent_state(&rcu_bh_state);
1913 * If RCU callbacks are still pending, RCU still needs this CPU.
1914 * So try forcing the callbacks through the grace period.
1916 if (rcu_cpu_has_callbacks(cpu)) {
1917 trace_rcu_prep_idle("More callbacks");
1920 trace_rcu_prep_idle("Callbacks drained");
1925 * Keep a running count of the number of non-lazy callbacks posted
1926 * on this CPU. This running counter (which is never decremented) allows
1927 * rcu_prepare_for_idle() to detect when something out of the idle loop
1928 * posts a callback, even if an equal number of callbacks are invoked.
1929 * Of course, callbacks should only be posted from within a trace event
1930 * designed to be called from idle or from within RCU_NONIDLE().
1932 static void rcu_idle_count_callbacks_posted(void)
1934 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1938 * Data for flushing lazy RCU callbacks at OOM time.
1940 static atomic_t oom_callback_count;
1941 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1944 * RCU OOM callback -- decrement the outstanding count and deliver the
1945 * wake-up if we are the last one.
1947 static void rcu_oom_callback(struct rcu_head *rhp)
1949 if (atomic_dec_and_test(&oom_callback_count))
1950 wake_up(&oom_callback_wq);
1954 * Post an rcu_oom_notify callback on the current CPU if it has at
1955 * least one lazy callback. This will unnecessarily post callbacks
1956 * to CPUs that already have a non-lazy callback at the end of their
1957 * callback list, but this is an infrequent operation, so accept some
1958 * extra overhead to keep things simple.
1960 static void rcu_oom_notify_cpu(void *unused)
1962 struct rcu_state *rsp;
1963 struct rcu_data *rdp;
1965 for_each_rcu_flavor(rsp) {
1966 rdp = __this_cpu_ptr(rsp->rda);
1967 if (rdp->qlen_lazy != 0) {
1968 atomic_inc(&oom_callback_count);
1969 rsp->call(&rdp->oom_head, rcu_oom_callback);
1975 * If low on memory, ensure that each CPU has a non-lazy callback.
1976 * This will wake up CPUs that have only lazy callbacks, in turn
1977 * ensuring that they free up the corresponding memory in a timely manner.
1978 * Because an uncertain amount of memory will be freed in some uncertain
1979 * timeframe, we do not claim to have freed anything.
1981 static int rcu_oom_notify(struct notifier_block *self,
1982 unsigned long notused, void *nfreed)
1986 /* Wait for callbacks from earlier instance to complete. */
1987 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1990 * Prevent premature wakeup: ensure that all increments happen
1991 * before there is a chance of the counter reaching zero.
1993 atomic_set(&oom_callback_count, 1);
1996 for_each_online_cpu(cpu) {
1997 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
2002 /* Unconditionally decrement: no need to wake ourselves up. */
2003 atomic_dec(&oom_callback_count);
2008 static struct notifier_block rcu_oom_nb = {
2009 .notifier_call = rcu_oom_notify
2012 static int __init rcu_register_oom_notifier(void)
2014 register_oom_notifier(&rcu_oom_nb);
2017 early_initcall(rcu_register_oom_notifier);
2019 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
2021 #ifdef CONFIG_RCU_CPU_STALL_INFO
2023 #ifdef CONFIG_RCU_FAST_NO_HZ
2025 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
2027 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
2028 struct timer_list *tltp = &rdtp->idle_gp_timer;
2031 c = rdtp->dyntick_holdoff == jiffies ? 'H' : '.';
2032 if (timer_pending(tltp))
2033 sprintf(cp, "drain=%d %c timer=%lu",
2034 rdtp->dyntick_drain, c, tltp->expires - jiffies);
2036 sprintf(cp, "drain=%d %c timer not pending",
2037 rdtp->dyntick_drain, c);
2040 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
2042 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
2047 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
2049 /* Initiate the stall-info list. */
2050 static void print_cpu_stall_info_begin(void)
2052 printk(KERN_CONT "\n");
2056 * Print out diagnostic information for the specified stalled CPU.
2058 * If the specified CPU is aware of the current RCU grace period
2059 * (flavor specified by rsp), then print the number of scheduling
2060 * clock interrupts the CPU has taken during the time that it has
2061 * been aware. Otherwise, print the number of RCU grace periods
2062 * that this CPU is ignorant of, for example, "1" if the CPU was
2063 * aware of the previous grace period.
2065 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
2067 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
2069 char fast_no_hz[72];
2070 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2071 struct rcu_dynticks *rdtp = rdp->dynticks;
2073 unsigned long ticks_value;
2075 if (rsp->gpnum == rdp->gpnum) {
2076 ticks_title = "ticks this GP";
2077 ticks_value = rdp->ticks_this_gp;
2079 ticks_title = "GPs behind";
2080 ticks_value = rsp->gpnum - rdp->gpnum;
2082 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
2083 printk(KERN_ERR "\t%d: (%lu %s) idle=%03x/%llx/%d %s\n",
2084 cpu, ticks_value, ticks_title,
2085 atomic_read(&rdtp->dynticks) & 0xfff,
2086 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
2090 /* Terminate the stall-info list. */
2091 static void print_cpu_stall_info_end(void)
2093 printk(KERN_ERR "\t");
2096 /* Zero ->ticks_this_gp for all flavors of RCU. */
2097 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2099 rdp->ticks_this_gp = 0;
2102 /* Increment ->ticks_this_gp for all flavors of RCU. */
2103 static void increment_cpu_stall_ticks(void)
2105 struct rcu_state *rsp;
2107 for_each_rcu_flavor(rsp)
2108 __this_cpu_ptr(rsp->rda)->ticks_this_gp++;
2111 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
2113 static void print_cpu_stall_info_begin(void)
2115 printk(KERN_CONT " {");
2118 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
2120 printk(KERN_CONT " %d", cpu);
2123 static void print_cpu_stall_info_end(void)
2125 printk(KERN_CONT "} ");
2128 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2132 static void increment_cpu_stall_ticks(void)
2136 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
2138 #ifdef CONFIG_RCU_NOCB_CPU
2141 * Offload callback processing from the boot-time-specified set of CPUs
2142 * specified by rcu_nocb_mask. For each CPU in the set, there is a
2143 * kthread created that pulls the callbacks from the corresponding CPU,
2144 * waits for a grace period to elapse, and invokes the callbacks.
2145 * The no-CBs CPUs do a wake_up() on their kthread when they insert
2146 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2147 * has been specified, in which case each kthread actively polls its
2148 * CPU. (Which isn't so great for energy efficiency, but which does
2149 * reduce RCU's overhead on that CPU.)
2151 * This is intended to be used in conjunction with Frederic Weisbecker's
2152 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2153 * running CPU-bound user-mode computations.
2155 * Offloading of callback processing could also in theory be used as
2156 * an energy-efficiency measure because CPUs with no RCU callbacks
2157 * queued are more aggressive about entering dyntick-idle mode.
2161 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2162 static int __init rcu_nocb_setup(char *str)
2164 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
2165 have_rcu_nocb_mask = true;
2166 cpulist_parse(str, rcu_nocb_mask);
2169 __setup("rcu_nocbs=", rcu_nocb_setup);
2171 static int __init parse_rcu_nocb_poll(char *arg)
2176 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2179 * Do any no-CBs CPUs need another grace period?
2181 * Interrupts must be disabled. If the caller does not hold the root
2182 * rnp_node structure's ->lock, the results are advisory only.
2184 static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2186 struct rcu_node *rnp = rcu_get_root(rsp);
2188 return rnp->n_nocb_gp_requests[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
2192 * Clean up this rcu_node structure's no-CBs state at the end of
2193 * a grace period, and also return whether any no-CBs CPU associated
2194 * with this rcu_node structure needs another grace period.
2196 static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2198 int c = rnp->completed;
2201 wake_up_all(&rnp->nocb_gp_wq[c & 0x1]);
2202 rnp->n_nocb_gp_requests[c & 0x1] = 0;
2203 needmore = rnp->n_nocb_gp_requests[(c + 1) & 0x1];
2208 * Set the root rcu_node structure's ->n_nocb_gp_requests field
2209 * based on the sum of those of all rcu_node structures. This does
2210 * double-count the root rcu_node structure's requests, but this
2211 * is necessary to handle the possibility of a rcu_nocb_kthread()
2212 * having awakened during the time that the rcu_node structures
2213 * were being updated for the end of the previous grace period.
2215 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2217 rnp->n_nocb_gp_requests[(rnp->completed + 1) & 0x1] += nrq;
2220 static void rcu_init_one_nocb(struct rcu_node *rnp)
2222 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2223 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2226 /* Is the specified CPU a no-CPUs CPU? */
2227 static bool is_nocb_cpu(int cpu)
2229 if (have_rcu_nocb_mask)
2230 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2235 * Enqueue the specified string of rcu_head structures onto the specified
2236 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2237 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2238 * counts are supplied by rhcount and rhcount_lazy.
2240 * If warranted, also wake up the kthread servicing this CPUs queues.
2242 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2243 struct rcu_head *rhp,
2244 struct rcu_head **rhtp,
2245 int rhcount, int rhcount_lazy)
2248 struct rcu_head **old_rhpp;
2249 struct task_struct *t;
2251 /* Enqueue the callback on the nocb list and update counts. */
2252 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2253 ACCESS_ONCE(*old_rhpp) = rhp;
2254 atomic_long_add(rhcount, &rdp->nocb_q_count);
2255 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2257 /* If we are not being polled and there is a kthread, awaken it ... */
2258 t = ACCESS_ONCE(rdp->nocb_kthread);
2259 if (rcu_nocb_poll | !t)
2261 len = atomic_long_read(&rdp->nocb_q_count);
2262 if (old_rhpp == &rdp->nocb_head) {
2263 wake_up(&rdp->nocb_wq); /* ... only if queue was empty ... */
2264 rdp->qlen_last_fqs_check = 0;
2265 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2266 wake_up_process(t); /* ... or if many callbacks queued. */
2267 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2273 * This is a helper for __call_rcu(), which invokes this when the normal
2274 * callback queue is inoperable. If this is not a no-CBs CPU, this
2275 * function returns failure back to __call_rcu(), which can complain
2278 * Otherwise, this function queues the callback where the corresponding
2279 * "rcuo" kthread can find it.
2281 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2285 if (!is_nocb_cpu(rdp->cpu))
2287 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy);
2288 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2289 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2290 (unsigned long)rhp->func,
2291 rdp->qlen_lazy, rdp->qlen);
2293 trace_rcu_callback(rdp->rsp->name, rhp,
2294 rdp->qlen_lazy, rdp->qlen);
2299 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2302 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2303 struct rcu_data *rdp)
2305 long ql = rsp->qlen;
2306 long qll = rsp->qlen_lazy;
2308 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2309 if (!is_nocb_cpu(smp_processor_id()))
2314 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2315 if (rsp->orphan_donelist != NULL) {
2316 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2317 rsp->orphan_donetail, ql, qll);
2319 rsp->orphan_donelist = NULL;
2320 rsp->orphan_donetail = &rsp->orphan_donelist;
2322 if (rsp->orphan_nxtlist != NULL) {
2323 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2324 rsp->orphan_nxttail, ql, qll);
2326 rsp->orphan_nxtlist = NULL;
2327 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2333 * If necessary, kick off a new grace period, and either way wait
2334 * for a subsequent grace period to complete.
2336 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2340 unsigned long flags;
2341 unsigned long flags1;
2342 struct rcu_node *rnp = rdp->mynode;
2343 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
2345 raw_spin_lock_irqsave(&rnp->lock, flags);
2346 c = rnp->completed + 2;
2348 /* Count our request for a grace period. */
2349 rnp->n_nocb_gp_requests[c & 0x1]++;
2351 if (rnp->gpnum != rnp->completed) {
2354 * This rcu_node structure believes that a grace period
2355 * is in progress, so we are done. When this grace
2356 * period ends, our request will be acted upon.
2358 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2363 * Might not be a grace period, check root rcu_node
2364 * structure to see if we must start one.
2366 if (rnp != rnp_root)
2367 raw_spin_lock(&rnp_root->lock); /* irqs disabled. */
2368 if (rnp_root->gpnum != rnp_root->completed) {
2369 raw_spin_unlock(&rnp_root->lock); /* irqs disabled. */
2373 * No grace period, so we need to start one.
2374 * The good news is that we can wait for exactly
2375 * one grace period instead of part of the current
2376 * grace period and all of the next grace period.
2377 * Adjust counters accordingly and start the
2378 * needed grace period.
2380 rnp->n_nocb_gp_requests[c & 0x1]--;
2381 c = rnp_root->completed + 1;
2382 rnp->n_nocb_gp_requests[c & 0x1]++;
2383 rnp_root->n_nocb_gp_requests[c & 0x1]++;
2384 local_save_flags(flags1);
2385 rcu_start_gp(rdp->rsp, flags1); /* Rlses ->lock. */
2388 /* Clean up locking and irq state. */
2389 if (rnp != rnp_root)
2390 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2392 local_irq_restore(flags);
2396 * Wait for the grace period. Do so interruptibly to avoid messing
2397 * up the load average.
2400 wait_event_interruptible(
2401 rnp->nocb_gp_wq[c & 0x1],
2402 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2405 flush_signals(current);
2407 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2411 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2412 * callbacks queued by the corresponding no-CBs CPU.
2414 static int rcu_nocb_kthread(void *arg)
2417 struct rcu_head *list;
2418 struct rcu_head *next;
2419 struct rcu_head **tail;
2420 struct rcu_data *rdp = arg;
2422 /* Each pass through this loop invokes one batch of callbacks */
2424 /* If not polling, wait for next batch of callbacks. */
2426 wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head);
2427 list = ACCESS_ONCE(rdp->nocb_head);
2429 schedule_timeout_interruptible(1);
2430 flush_signals(current);
2435 * Extract queued callbacks, update counts, and wait
2436 * for a grace period to elapse.
2438 ACCESS_ONCE(rdp->nocb_head) = NULL;
2439 tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2440 c = atomic_long_xchg(&rdp->nocb_q_count, 0);
2441 cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2442 ACCESS_ONCE(rdp->nocb_p_count) += c;
2443 ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl;
2444 rcu_nocb_wait_gp(rdp);
2446 /* Each pass through the following loop invokes a callback. */
2447 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2451 /* Wait for enqueuing to complete, if needed. */
2452 while (next == NULL && &list->next != tail) {
2453 schedule_timeout_interruptible(1);
2456 debug_rcu_head_unqueue(list);
2458 if (__rcu_reclaim(rdp->rsp->name, list))
2464 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2465 ACCESS_ONCE(rdp->nocb_p_count) -= c;
2466 ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2467 rdp->n_nocbs_invoked += c;
2472 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2473 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2475 rdp->nocb_tail = &rdp->nocb_head;
2476 init_waitqueue_head(&rdp->nocb_wq);
2479 /* Create a kthread for each RCU flavor for each no-CBs CPU. */
2480 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2483 struct rcu_data *rdp;
2484 struct task_struct *t;
2486 if (rcu_nocb_mask == NULL)
2488 for_each_cpu(cpu, rcu_nocb_mask) {
2489 rdp = per_cpu_ptr(rsp->rda, cpu);
2490 t = kthread_run(rcu_nocb_kthread, rdp, "rcuo%d", cpu);
2492 ACCESS_ONCE(rdp->nocb_kthread) = t;
2496 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2497 static bool init_nocb_callback_list(struct rcu_data *rdp)
2499 if (rcu_nocb_mask == NULL ||
2500 !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2502 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2506 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2508 static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2513 static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2518 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2522 static void rcu_init_one_nocb(struct rcu_node *rnp)
2526 static bool is_nocb_cpu(int cpu)
2531 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2537 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2538 struct rcu_data *rdp)
2543 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2547 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2551 static bool init_nocb_callback_list(struct rcu_data *rdp)
2556 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */