x86/xen: resume timer irqs early
[linux/fpc-iii.git] / kernel / rcutree_plugin.h
blob511e6b47c594b9ec6dba64ef1b91c098f4e75d82
1 /*
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>
31 #include "time/tick-internal.h"
33 #define RCU_KTHREAD_PRIO 1
35 #ifdef CONFIG_RCU_BOOST
36 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
37 #else
38 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
39 #endif
41 #ifdef CONFIG_RCU_NOCB_CPU
42 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
43 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
44 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
45 static char __initdata nocb_buf[NR_CPUS * 5];
46 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
49 * Check the RCU kernel configuration parameters and print informative
50 * messages about anything out of the ordinary. If you like #ifdef, you
51 * will love this function.
53 static void __init rcu_bootup_announce_oddness(void)
55 #ifdef CONFIG_RCU_TRACE
56 pr_info("\tRCU debugfs-based tracing is enabled.\n");
57 #endif
58 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
59 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
60 CONFIG_RCU_FANOUT);
61 #endif
62 #ifdef CONFIG_RCU_FANOUT_EXACT
63 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
64 #endif
65 #ifdef CONFIG_RCU_FAST_NO_HZ
66 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
67 #endif
68 #ifdef CONFIG_PROVE_RCU
69 pr_info("\tRCU lockdep checking is enabled.\n");
70 #endif
71 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
72 pr_info("\tRCU torture testing starts during boot.\n");
73 #endif
74 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
75 pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
76 #endif
77 #if defined(CONFIG_RCU_CPU_STALL_INFO)
78 pr_info("\tAdditional per-CPU info printed with stalls.\n");
79 #endif
80 #if NUM_RCU_LVL_4 != 0
81 pr_info("\tFour-level hierarchy is enabled.\n");
82 #endif
83 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
84 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
85 if (nr_cpu_ids != NR_CPUS)
86 pr_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 zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL);
91 have_rcu_nocb_mask = true;
93 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
94 pr_info("\tOffload RCU callbacks from 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("\tOffload RCU callbacks from 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("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
105 if (rcu_nocb_poll)
106 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
108 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
111 #ifdef CONFIG_TREE_PREEMPT_RCU
113 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
114 static struct rcu_state *rcu_state = &rcu_preempt_state;
116 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
119 * Tell them what RCU they are running.
121 static void __init rcu_bootup_announce(void)
123 pr_info("Preemptible hierarchical RCU implementation.\n");
124 rcu_bootup_announce_oddness();
128 * Return the number of RCU-preempt batches processed thus far
129 * for debug and statistics.
131 long rcu_batches_completed_preempt(void)
133 return rcu_preempt_state.completed;
135 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
138 * Return the number of RCU batches processed thus far for debug & stats.
140 long rcu_batches_completed(void)
142 return rcu_batches_completed_preempt();
144 EXPORT_SYMBOL_GPL(rcu_batches_completed);
147 * Force a quiescent state for preemptible RCU.
149 void rcu_force_quiescent_state(void)
151 force_quiescent_state(&rcu_preempt_state);
153 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
156 * Record a preemptible-RCU quiescent state for the specified CPU. Note
157 * that this just means that the task currently running on the CPU is
158 * not in a quiescent state. There might be any number of tasks blocked
159 * while in an RCU read-side critical section.
161 * Unlike the other rcu_*_qs() functions, callers to this function
162 * must disable irqs in order to protect the assignment to
163 * ->rcu_read_unlock_special.
165 static void rcu_preempt_qs(int cpu)
167 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
169 if (rdp->passed_quiesce == 0)
170 trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs"));
171 rdp->passed_quiesce = 1;
172 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
176 * We have entered the scheduler, and the current task might soon be
177 * context-switched away from. If this task is in an RCU read-side
178 * critical section, we will no longer be able to rely on the CPU to
179 * record that fact, so we enqueue the task on the blkd_tasks list.
180 * The task will dequeue itself when it exits the outermost enclosing
181 * RCU read-side critical section. Therefore, the current grace period
182 * cannot be permitted to complete until the blkd_tasks list entries
183 * predating the current grace period drain, in other words, until
184 * rnp->gp_tasks becomes NULL.
186 * Caller must disable preemption.
188 static void rcu_preempt_note_context_switch(int cpu)
190 struct task_struct *t = current;
191 unsigned long flags;
192 struct rcu_data *rdp;
193 struct rcu_node *rnp;
195 if (t->rcu_read_lock_nesting > 0 &&
196 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
198 /* Possibly blocking in an RCU read-side critical section. */
199 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
200 rnp = rdp->mynode;
201 raw_spin_lock_irqsave(&rnp->lock, flags);
202 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
203 t->rcu_blocked_node = rnp;
206 * If this CPU has already checked in, then this task
207 * will hold up the next grace period rather than the
208 * current grace period. Queue the task accordingly.
209 * If the task is queued for the current grace period
210 * (i.e., this CPU has not yet passed through a quiescent
211 * state for the current grace period), then as long
212 * as that task remains queued, the current grace period
213 * cannot end. Note that there is some uncertainty as
214 * to exactly when the current grace period started.
215 * We take a conservative approach, which can result
216 * in unnecessarily waiting on tasks that started very
217 * slightly after the current grace period began. C'est
218 * la vie!!!
220 * But first, note that the current CPU must still be
221 * on line!
223 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
224 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
225 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
226 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
227 rnp->gp_tasks = &t->rcu_node_entry;
228 #ifdef CONFIG_RCU_BOOST
229 if (rnp->boost_tasks != NULL)
230 rnp->boost_tasks = rnp->gp_tasks;
231 #endif /* #ifdef CONFIG_RCU_BOOST */
232 } else {
233 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
234 if (rnp->qsmask & rdp->grpmask)
235 rnp->gp_tasks = &t->rcu_node_entry;
237 trace_rcu_preempt_task(rdp->rsp->name,
238 t->pid,
239 (rnp->qsmask & rdp->grpmask)
240 ? rnp->gpnum
241 : rnp->gpnum + 1);
242 raw_spin_unlock_irqrestore(&rnp->lock, flags);
243 } else if (t->rcu_read_lock_nesting < 0 &&
244 t->rcu_read_unlock_special) {
247 * Complete exit from RCU read-side critical section on
248 * behalf of preempted instance of __rcu_read_unlock().
250 rcu_read_unlock_special(t);
254 * Either we were not in an RCU read-side critical section to
255 * begin with, or we have now recorded that critical section
256 * globally. Either way, we can now note a quiescent state
257 * for this CPU. Again, if we were in an RCU read-side critical
258 * section, and if that critical section was blocking the current
259 * grace period, then the fact that the task has been enqueued
260 * means that we continue to block the current grace period.
262 local_irq_save(flags);
263 rcu_preempt_qs(cpu);
264 local_irq_restore(flags);
268 * Check for preempted RCU readers blocking the current grace period
269 * for the specified rcu_node structure. If the caller needs a reliable
270 * answer, it must hold the rcu_node's ->lock.
272 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
274 return rnp->gp_tasks != NULL;
278 * Record a quiescent state for all tasks that were previously queued
279 * on the specified rcu_node structure and that were blocking the current
280 * RCU grace period. The caller must hold the specified rnp->lock with
281 * irqs disabled, and this lock is released upon return, but irqs remain
282 * disabled.
284 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
285 __releases(rnp->lock)
287 unsigned long mask;
288 struct rcu_node *rnp_p;
290 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
291 raw_spin_unlock_irqrestore(&rnp->lock, flags);
292 return; /* Still need more quiescent states! */
295 rnp_p = rnp->parent;
296 if (rnp_p == NULL) {
298 * Either there is only one rcu_node in the tree,
299 * or tasks were kicked up to root rcu_node due to
300 * CPUs going offline.
302 rcu_report_qs_rsp(&rcu_preempt_state, flags);
303 return;
306 /* Report up the rest of the hierarchy. */
307 mask = rnp->grpmask;
308 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
309 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
310 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
314 * Advance a ->blkd_tasks-list pointer to the next entry, instead
315 * returning NULL if at the end of the list.
317 static struct list_head *rcu_next_node_entry(struct task_struct *t,
318 struct rcu_node *rnp)
320 struct list_head *np;
322 np = t->rcu_node_entry.next;
323 if (np == &rnp->blkd_tasks)
324 np = NULL;
325 return np;
329 * Handle special cases during rcu_read_unlock(), such as needing to
330 * notify RCU core processing or task having blocked during the RCU
331 * read-side critical section.
333 void rcu_read_unlock_special(struct task_struct *t)
335 int empty;
336 int empty_exp;
337 int empty_exp_now;
338 unsigned long flags;
339 struct list_head *np;
340 #ifdef CONFIG_RCU_BOOST
341 struct rt_mutex *rbmp = NULL;
342 #endif /* #ifdef CONFIG_RCU_BOOST */
343 struct rcu_node *rnp;
344 int special;
346 /* NMI handlers cannot block and cannot safely manipulate state. */
347 if (in_nmi())
348 return;
350 local_irq_save(flags);
353 * If RCU core is waiting for this CPU to exit critical section,
354 * let it know that we have done so.
356 special = t->rcu_read_unlock_special;
357 if (special & RCU_READ_UNLOCK_NEED_QS) {
358 rcu_preempt_qs(smp_processor_id());
361 /* Hardware IRQ handlers cannot block. */
362 if (in_irq() || in_serving_softirq()) {
363 local_irq_restore(flags);
364 return;
367 /* Clean up if blocked during RCU read-side critical section. */
368 if (special & RCU_READ_UNLOCK_BLOCKED) {
369 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
372 * Remove this task from the list it blocked on. The
373 * task can migrate while we acquire the lock, but at
374 * most one time. So at most two passes through loop.
376 for (;;) {
377 rnp = t->rcu_blocked_node;
378 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
379 if (rnp == t->rcu_blocked_node)
380 break;
381 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
383 empty = !rcu_preempt_blocked_readers_cgp(rnp);
384 empty_exp = !rcu_preempted_readers_exp(rnp);
385 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
386 np = rcu_next_node_entry(t, rnp);
387 list_del_init(&t->rcu_node_entry);
388 t->rcu_blocked_node = NULL;
389 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
390 rnp->gpnum, t->pid);
391 if (&t->rcu_node_entry == rnp->gp_tasks)
392 rnp->gp_tasks = np;
393 if (&t->rcu_node_entry == rnp->exp_tasks)
394 rnp->exp_tasks = np;
395 #ifdef CONFIG_RCU_BOOST
396 if (&t->rcu_node_entry == rnp->boost_tasks)
397 rnp->boost_tasks = np;
398 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
399 if (t->rcu_boost_mutex) {
400 rbmp = t->rcu_boost_mutex;
401 t->rcu_boost_mutex = NULL;
403 #endif /* #ifdef CONFIG_RCU_BOOST */
406 * If this was the last task on the current list, and if
407 * we aren't waiting on any CPUs, report the quiescent state.
408 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
409 * so we must take a snapshot of the expedited state.
411 empty_exp_now = !rcu_preempted_readers_exp(rnp);
412 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
413 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
414 rnp->gpnum,
415 0, rnp->qsmask,
416 rnp->level,
417 rnp->grplo,
418 rnp->grphi,
419 !!rnp->gp_tasks);
420 rcu_report_unblock_qs_rnp(rnp, flags);
421 } else {
422 raw_spin_unlock_irqrestore(&rnp->lock, flags);
425 #ifdef CONFIG_RCU_BOOST
426 /* Unboost if we were boosted. */
427 if (rbmp)
428 rt_mutex_unlock(rbmp);
429 #endif /* #ifdef CONFIG_RCU_BOOST */
432 * If this was the last task on the expedited lists,
433 * then we need to report up the rcu_node hierarchy.
435 if (!empty_exp && empty_exp_now)
436 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
437 } else {
438 local_irq_restore(flags);
442 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
445 * Dump detailed information for all tasks blocking the current RCU
446 * grace period on the specified rcu_node structure.
448 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
450 unsigned long flags;
451 struct task_struct *t;
453 raw_spin_lock_irqsave(&rnp->lock, flags);
454 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
455 raw_spin_unlock_irqrestore(&rnp->lock, flags);
456 return;
458 t = list_entry(rnp->gp_tasks,
459 struct task_struct, rcu_node_entry);
460 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
461 sched_show_task(t);
462 raw_spin_unlock_irqrestore(&rnp->lock, flags);
466 * Dump detailed information for all tasks blocking the current RCU
467 * grace period.
469 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
471 struct rcu_node *rnp = rcu_get_root(rsp);
473 rcu_print_detail_task_stall_rnp(rnp);
474 rcu_for_each_leaf_node(rsp, rnp)
475 rcu_print_detail_task_stall_rnp(rnp);
478 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
480 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
484 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
486 #ifdef CONFIG_RCU_CPU_STALL_INFO
488 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
490 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
491 rnp->level, rnp->grplo, rnp->grphi);
494 static void rcu_print_task_stall_end(void)
496 pr_cont("\n");
499 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
501 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
505 static void rcu_print_task_stall_end(void)
509 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
512 * Scan the current list of tasks blocked within RCU read-side critical
513 * sections, printing out the tid of each.
515 static int rcu_print_task_stall(struct rcu_node *rnp)
517 struct task_struct *t;
518 int ndetected = 0;
520 if (!rcu_preempt_blocked_readers_cgp(rnp))
521 return 0;
522 rcu_print_task_stall_begin(rnp);
523 t = list_entry(rnp->gp_tasks,
524 struct task_struct, rcu_node_entry);
525 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
526 pr_cont(" P%d", t->pid);
527 ndetected++;
529 rcu_print_task_stall_end();
530 return ndetected;
534 * Check that the list of blocked tasks for the newly completed grace
535 * period is in fact empty. It is a serious bug to complete a grace
536 * period that still has RCU readers blocked! This function must be
537 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
538 * must be held by the caller.
540 * Also, if there are blocked tasks on the list, they automatically
541 * block the newly created grace period, so set up ->gp_tasks accordingly.
543 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
545 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
546 if (!list_empty(&rnp->blkd_tasks))
547 rnp->gp_tasks = rnp->blkd_tasks.next;
548 WARN_ON_ONCE(rnp->qsmask);
551 #ifdef CONFIG_HOTPLUG_CPU
554 * Handle tasklist migration for case in which all CPUs covered by the
555 * specified rcu_node have gone offline. Move them up to the root
556 * rcu_node. The reason for not just moving them to the immediate
557 * parent is to remove the need for rcu_read_unlock_special() to
558 * make more than two attempts to acquire the target rcu_node's lock.
559 * Returns true if there were tasks blocking the current RCU grace
560 * period.
562 * Returns 1 if there was previously a task blocking the current grace
563 * period on the specified rcu_node structure.
565 * The caller must hold rnp->lock with irqs disabled.
567 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
568 struct rcu_node *rnp,
569 struct rcu_data *rdp)
571 struct list_head *lp;
572 struct list_head *lp_root;
573 int retval = 0;
574 struct rcu_node *rnp_root = rcu_get_root(rsp);
575 struct task_struct *t;
577 if (rnp == rnp_root) {
578 WARN_ONCE(1, "Last CPU thought to be offlined?");
579 return 0; /* Shouldn't happen: at least one CPU online. */
582 /* If we are on an internal node, complain bitterly. */
583 WARN_ON_ONCE(rnp != rdp->mynode);
586 * Move tasks up to root rcu_node. Don't try to get fancy for
587 * this corner-case operation -- just put this node's tasks
588 * at the head of the root node's list, and update the root node's
589 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
590 * if non-NULL. This might result in waiting for more tasks than
591 * absolutely necessary, but this is a good performance/complexity
592 * tradeoff.
594 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
595 retval |= RCU_OFL_TASKS_NORM_GP;
596 if (rcu_preempted_readers_exp(rnp))
597 retval |= RCU_OFL_TASKS_EXP_GP;
598 lp = &rnp->blkd_tasks;
599 lp_root = &rnp_root->blkd_tasks;
600 while (!list_empty(lp)) {
601 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
602 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
603 list_del(&t->rcu_node_entry);
604 t->rcu_blocked_node = rnp_root;
605 list_add(&t->rcu_node_entry, lp_root);
606 if (&t->rcu_node_entry == rnp->gp_tasks)
607 rnp_root->gp_tasks = rnp->gp_tasks;
608 if (&t->rcu_node_entry == rnp->exp_tasks)
609 rnp_root->exp_tasks = rnp->exp_tasks;
610 #ifdef CONFIG_RCU_BOOST
611 if (&t->rcu_node_entry == rnp->boost_tasks)
612 rnp_root->boost_tasks = rnp->boost_tasks;
613 #endif /* #ifdef CONFIG_RCU_BOOST */
614 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
617 rnp->gp_tasks = NULL;
618 rnp->exp_tasks = NULL;
619 #ifdef CONFIG_RCU_BOOST
620 rnp->boost_tasks = NULL;
622 * In case root is being boosted and leaf was not. Make sure
623 * that we boost the tasks blocking the current grace period
624 * in this case.
626 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
627 if (rnp_root->boost_tasks != NULL &&
628 rnp_root->boost_tasks != rnp_root->gp_tasks &&
629 rnp_root->boost_tasks != rnp_root->exp_tasks)
630 rnp_root->boost_tasks = rnp_root->gp_tasks;
631 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
632 #endif /* #ifdef CONFIG_RCU_BOOST */
634 return retval;
637 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
640 * Check for a quiescent state from the current CPU. When a task blocks,
641 * the task is recorded in the corresponding CPU's rcu_node structure,
642 * which is checked elsewhere.
644 * Caller must disable hard irqs.
646 static void rcu_preempt_check_callbacks(int cpu)
648 struct task_struct *t = current;
650 if (t->rcu_read_lock_nesting == 0) {
651 rcu_preempt_qs(cpu);
652 return;
654 if (t->rcu_read_lock_nesting > 0 &&
655 per_cpu(rcu_preempt_data, cpu).qs_pending)
656 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
659 #ifdef CONFIG_RCU_BOOST
661 static void rcu_preempt_do_callbacks(void)
663 rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
666 #endif /* #ifdef CONFIG_RCU_BOOST */
669 * Queue a preemptible-RCU callback for invocation after a grace period.
671 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
673 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
675 EXPORT_SYMBOL_GPL(call_rcu);
678 * Queue an RCU callback for lazy invocation after a grace period.
679 * This will likely be later named something like "call_rcu_lazy()",
680 * but this change will require some way of tagging the lazy RCU
681 * callbacks in the list of pending callbacks. Until then, this
682 * function may only be called from __kfree_rcu().
684 void kfree_call_rcu(struct rcu_head *head,
685 void (*func)(struct rcu_head *rcu))
687 __call_rcu(head, func, &rcu_preempt_state, -1, 1);
689 EXPORT_SYMBOL_GPL(kfree_call_rcu);
692 * synchronize_rcu - wait until a grace period has elapsed.
694 * Control will return to the caller some time after a full grace
695 * period has elapsed, in other words after all currently executing RCU
696 * read-side critical sections have completed. Note, however, that
697 * upon return from synchronize_rcu(), the caller might well be executing
698 * concurrently with new RCU read-side critical sections that began while
699 * synchronize_rcu() was waiting. RCU read-side critical sections are
700 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
702 * See the description of synchronize_sched() for more detailed information
703 * on memory ordering guarantees.
705 void synchronize_rcu(void)
707 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
708 !lock_is_held(&rcu_lock_map) &&
709 !lock_is_held(&rcu_sched_lock_map),
710 "Illegal synchronize_rcu() in RCU read-side critical section");
711 if (!rcu_scheduler_active)
712 return;
713 if (rcu_expedited)
714 synchronize_rcu_expedited();
715 else
716 wait_rcu_gp(call_rcu);
718 EXPORT_SYMBOL_GPL(synchronize_rcu);
720 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
721 static unsigned long sync_rcu_preempt_exp_count;
722 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
725 * Return non-zero if there are any tasks in RCU read-side critical
726 * sections blocking the current preemptible-RCU expedited grace period.
727 * If there is no preemptible-RCU expedited grace period currently in
728 * progress, returns zero unconditionally.
730 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
732 return rnp->exp_tasks != NULL;
736 * return non-zero if there is no RCU expedited grace period in progress
737 * for the specified rcu_node structure, in other words, if all CPUs and
738 * tasks covered by the specified rcu_node structure have done their bit
739 * for the current expedited grace period. Works only for preemptible
740 * RCU -- other RCU implementation use other means.
742 * Caller must hold sync_rcu_preempt_exp_mutex.
744 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
746 return !rcu_preempted_readers_exp(rnp) &&
747 ACCESS_ONCE(rnp->expmask) == 0;
751 * Report the exit from RCU read-side critical section for the last task
752 * that queued itself during or before the current expedited preemptible-RCU
753 * grace period. This event is reported either to the rcu_node structure on
754 * which the task was queued or to one of that rcu_node structure's ancestors,
755 * recursively up the tree. (Calm down, calm down, we do the recursion
756 * iteratively!)
758 * Most callers will set the "wake" flag, but the task initiating the
759 * expedited grace period need not wake itself.
761 * Caller must hold sync_rcu_preempt_exp_mutex.
763 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
764 bool wake)
766 unsigned long flags;
767 unsigned long mask;
769 raw_spin_lock_irqsave(&rnp->lock, flags);
770 for (;;) {
771 if (!sync_rcu_preempt_exp_done(rnp)) {
772 raw_spin_unlock_irqrestore(&rnp->lock, flags);
773 break;
775 if (rnp->parent == NULL) {
776 raw_spin_unlock_irqrestore(&rnp->lock, flags);
777 if (wake)
778 wake_up(&sync_rcu_preempt_exp_wq);
779 break;
781 mask = rnp->grpmask;
782 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
783 rnp = rnp->parent;
784 raw_spin_lock(&rnp->lock); /* irqs already disabled */
785 rnp->expmask &= ~mask;
790 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
791 * grace period for the specified rcu_node structure. If there are no such
792 * tasks, report it up the rcu_node hierarchy.
794 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
795 * CPU hotplug operations.
797 static void
798 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
800 unsigned long flags;
801 int must_wait = 0;
803 raw_spin_lock_irqsave(&rnp->lock, flags);
804 if (list_empty(&rnp->blkd_tasks)) {
805 raw_spin_unlock_irqrestore(&rnp->lock, flags);
806 } else {
807 rnp->exp_tasks = rnp->blkd_tasks.next;
808 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
809 must_wait = 1;
811 if (!must_wait)
812 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
816 * synchronize_rcu_expedited - Brute-force RCU grace period
818 * Wait for an RCU-preempt grace period, but expedite it. The basic
819 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
820 * the ->blkd_tasks lists and wait for this list to drain. This consumes
821 * significant time on all CPUs and is unfriendly to real-time workloads,
822 * so is thus not recommended for any sort of common-case code.
823 * In fact, if you are using synchronize_rcu_expedited() in a loop,
824 * please restructure your code to batch your updates, and then Use a
825 * single synchronize_rcu() instead.
827 * Note that it is illegal to call this function while holding any lock
828 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
829 * to call this function from a CPU-hotplug notifier. Failing to observe
830 * these restriction will result in deadlock.
832 void synchronize_rcu_expedited(void)
834 unsigned long flags;
835 struct rcu_node *rnp;
836 struct rcu_state *rsp = &rcu_preempt_state;
837 unsigned long snap;
838 int trycount = 0;
840 smp_mb(); /* Caller's modifications seen first by other CPUs. */
841 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
842 smp_mb(); /* Above access cannot bleed into critical section. */
845 * Block CPU-hotplug operations. This means that any CPU-hotplug
846 * operation that finds an rcu_node structure with tasks in the
847 * process of being boosted will know that all tasks blocking
848 * this expedited grace period will already be in the process of
849 * being boosted. This simplifies the process of moving tasks
850 * from leaf to root rcu_node structures.
852 get_online_cpus();
855 * Acquire lock, falling back to synchronize_rcu() if too many
856 * lock-acquisition failures. Of course, if someone does the
857 * expedited grace period for us, just leave.
859 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
860 if (ULONG_CMP_LT(snap,
861 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
862 put_online_cpus();
863 goto mb_ret; /* Others did our work for us. */
865 if (trycount++ < 10) {
866 udelay(trycount * num_online_cpus());
867 } else {
868 put_online_cpus();
869 wait_rcu_gp(call_rcu);
870 return;
873 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
874 put_online_cpus();
875 goto unlock_mb_ret; /* Others did our work for us. */
878 /* force all RCU readers onto ->blkd_tasks lists. */
879 synchronize_sched_expedited();
881 /* Initialize ->expmask for all non-leaf rcu_node structures. */
882 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
883 raw_spin_lock_irqsave(&rnp->lock, flags);
884 rnp->expmask = rnp->qsmaskinit;
885 raw_spin_unlock_irqrestore(&rnp->lock, flags);
888 /* Snapshot current state of ->blkd_tasks lists. */
889 rcu_for_each_leaf_node(rsp, rnp)
890 sync_rcu_preempt_exp_init(rsp, rnp);
891 if (NUM_RCU_NODES > 1)
892 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
894 put_online_cpus();
896 /* Wait for snapshotted ->blkd_tasks lists to drain. */
897 rnp = rcu_get_root(rsp);
898 wait_event(sync_rcu_preempt_exp_wq,
899 sync_rcu_preempt_exp_done(rnp));
901 /* Clean up and exit. */
902 smp_mb(); /* ensure expedited GP seen before counter increment. */
903 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
904 unlock_mb_ret:
905 mutex_unlock(&sync_rcu_preempt_exp_mutex);
906 mb_ret:
907 smp_mb(); /* ensure subsequent action seen after grace period. */
909 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
912 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
914 * Note that this primitive does not necessarily wait for an RCU grace period
915 * to complete. For example, if there are no RCU callbacks queued anywhere
916 * in the system, then rcu_barrier() is within its rights to return
917 * immediately, without waiting for anything, much less an RCU grace period.
919 void rcu_barrier(void)
921 _rcu_barrier(&rcu_preempt_state);
923 EXPORT_SYMBOL_GPL(rcu_barrier);
926 * Initialize preemptible RCU's state structures.
928 static void __init __rcu_init_preempt(void)
930 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
934 * Check for a task exiting while in a preemptible-RCU read-side
935 * critical section, clean up if so. No need to issue warnings,
936 * as debug_check_no_locks_held() already does this if lockdep
937 * is enabled.
939 void exit_rcu(void)
941 struct task_struct *t = current;
943 if (likely(list_empty(&current->rcu_node_entry)))
944 return;
945 t->rcu_read_lock_nesting = 1;
946 barrier();
947 t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED;
948 __rcu_read_unlock();
951 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
953 static struct rcu_state *rcu_state = &rcu_sched_state;
956 * Tell them what RCU they are running.
958 static void __init rcu_bootup_announce(void)
960 pr_info("Hierarchical RCU implementation.\n");
961 rcu_bootup_announce_oddness();
965 * Return the number of RCU batches processed thus far for debug & stats.
967 long rcu_batches_completed(void)
969 return rcu_batches_completed_sched();
971 EXPORT_SYMBOL_GPL(rcu_batches_completed);
974 * Force a quiescent state for RCU, which, because there is no preemptible
975 * RCU, becomes the same as rcu-sched.
977 void rcu_force_quiescent_state(void)
979 rcu_sched_force_quiescent_state();
981 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
984 * Because preemptible RCU does not exist, we never have to check for
985 * CPUs being in quiescent states.
987 static void rcu_preempt_note_context_switch(int cpu)
992 * Because preemptible RCU does not exist, there are never any preempted
993 * RCU readers.
995 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
997 return 0;
1000 #ifdef CONFIG_HOTPLUG_CPU
1002 /* Because preemptible RCU does not exist, no quieting of tasks. */
1003 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1005 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1008 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1011 * Because preemptible RCU does not exist, we never have to check for
1012 * tasks blocked within RCU read-side critical sections.
1014 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1019 * Because preemptible RCU does not exist, we never have to check for
1020 * tasks blocked within RCU read-side critical sections.
1022 static int rcu_print_task_stall(struct rcu_node *rnp)
1024 return 0;
1028 * Because there is no preemptible RCU, there can be no readers blocked,
1029 * so there is no need to check for blocked tasks. So check only for
1030 * bogus qsmask values.
1032 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1034 WARN_ON_ONCE(rnp->qsmask);
1037 #ifdef CONFIG_HOTPLUG_CPU
1040 * Because preemptible RCU does not exist, it never needs to migrate
1041 * tasks that were blocked within RCU read-side critical sections, and
1042 * such non-existent tasks cannot possibly have been blocking the current
1043 * grace period.
1045 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1046 struct rcu_node *rnp,
1047 struct rcu_data *rdp)
1049 return 0;
1052 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1055 * Because preemptible RCU does not exist, it never has any callbacks
1056 * to check.
1058 static void rcu_preempt_check_callbacks(int cpu)
1063 * Queue an RCU callback for lazy invocation after a grace period.
1064 * This will likely be later named something like "call_rcu_lazy()",
1065 * but this change will require some way of tagging the lazy RCU
1066 * callbacks in the list of pending callbacks. Until then, this
1067 * function may only be called from __kfree_rcu().
1069 * Because there is no preemptible RCU, we use RCU-sched instead.
1071 void kfree_call_rcu(struct rcu_head *head,
1072 void (*func)(struct rcu_head *rcu))
1074 __call_rcu(head, func, &rcu_sched_state, -1, 1);
1076 EXPORT_SYMBOL_GPL(kfree_call_rcu);
1079 * Wait for an rcu-preempt grace period, but make it happen quickly.
1080 * But because preemptible RCU does not exist, map to rcu-sched.
1082 void synchronize_rcu_expedited(void)
1084 synchronize_sched_expedited();
1086 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1088 #ifdef CONFIG_HOTPLUG_CPU
1091 * Because preemptible RCU does not exist, there is never any need to
1092 * report on tasks preempted in RCU read-side critical sections during
1093 * expedited RCU grace periods.
1095 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1096 bool wake)
1100 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1103 * Because preemptible RCU does not exist, rcu_barrier() is just
1104 * another name for rcu_barrier_sched().
1106 void rcu_barrier(void)
1108 rcu_barrier_sched();
1110 EXPORT_SYMBOL_GPL(rcu_barrier);
1113 * Because preemptible RCU does not exist, it need not be initialized.
1115 static void __init __rcu_init_preempt(void)
1120 * Because preemptible RCU does not exist, tasks cannot possibly exit
1121 * while in preemptible RCU read-side critical sections.
1123 void exit_rcu(void)
1127 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1129 #ifdef CONFIG_RCU_BOOST
1131 #include "rtmutex_common.h"
1133 #ifdef CONFIG_RCU_TRACE
1135 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1137 if (list_empty(&rnp->blkd_tasks))
1138 rnp->n_balk_blkd_tasks++;
1139 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1140 rnp->n_balk_exp_gp_tasks++;
1141 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1142 rnp->n_balk_boost_tasks++;
1143 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1144 rnp->n_balk_notblocked++;
1145 else if (rnp->gp_tasks != NULL &&
1146 ULONG_CMP_LT(jiffies, rnp->boost_time))
1147 rnp->n_balk_notyet++;
1148 else
1149 rnp->n_balk_nos++;
1152 #else /* #ifdef CONFIG_RCU_TRACE */
1154 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1158 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1160 static void rcu_wake_cond(struct task_struct *t, int status)
1163 * If the thread is yielding, only wake it when this
1164 * is invoked from idle
1166 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1167 wake_up_process(t);
1171 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1172 * or ->boost_tasks, advancing the pointer to the next task in the
1173 * ->blkd_tasks list.
1175 * Note that irqs must be enabled: boosting the task can block.
1176 * Returns 1 if there are more tasks needing to be boosted.
1178 static int rcu_boost(struct rcu_node *rnp)
1180 unsigned long flags;
1181 struct rt_mutex mtx;
1182 struct task_struct *t;
1183 struct list_head *tb;
1185 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1186 return 0; /* Nothing left to boost. */
1188 raw_spin_lock_irqsave(&rnp->lock, flags);
1191 * Recheck under the lock: all tasks in need of boosting
1192 * might exit their RCU read-side critical sections on their own.
1194 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1195 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1196 return 0;
1200 * Preferentially boost tasks blocking expedited grace periods.
1201 * This cannot starve the normal grace periods because a second
1202 * expedited grace period must boost all blocked tasks, including
1203 * those blocking the pre-existing normal grace period.
1205 if (rnp->exp_tasks != NULL) {
1206 tb = rnp->exp_tasks;
1207 rnp->n_exp_boosts++;
1208 } else {
1209 tb = rnp->boost_tasks;
1210 rnp->n_normal_boosts++;
1212 rnp->n_tasks_boosted++;
1215 * We boost task t by manufacturing an rt_mutex that appears to
1216 * be held by task t. We leave a pointer to that rt_mutex where
1217 * task t can find it, and task t will release the mutex when it
1218 * exits its outermost RCU read-side critical section. Then
1219 * simply acquiring this artificial rt_mutex will boost task
1220 * t's priority. (Thanks to tglx for suggesting this approach!)
1222 * Note that task t must acquire rnp->lock to remove itself from
1223 * the ->blkd_tasks list, which it will do from exit() if from
1224 * nowhere else. We therefore are guaranteed that task t will
1225 * stay around at least until we drop rnp->lock. Note that
1226 * rnp->lock also resolves races between our priority boosting
1227 * and task t's exiting its outermost RCU read-side critical
1228 * section.
1230 t = container_of(tb, struct task_struct, rcu_node_entry);
1231 rt_mutex_init_proxy_locked(&mtx, t);
1232 t->rcu_boost_mutex = &mtx;
1233 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1234 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1235 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1237 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1238 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1242 * Priority-boosting kthread. One per leaf rcu_node and one for the
1243 * root rcu_node.
1245 static int rcu_boost_kthread(void *arg)
1247 struct rcu_node *rnp = (struct rcu_node *)arg;
1248 int spincnt = 0;
1249 int more2boost;
1251 trace_rcu_utilization(TPS("Start boost kthread@init"));
1252 for (;;) {
1253 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1254 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1255 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1256 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1257 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1258 more2boost = rcu_boost(rnp);
1259 if (more2boost)
1260 spincnt++;
1261 else
1262 spincnt = 0;
1263 if (spincnt > 10) {
1264 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1265 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1266 schedule_timeout_interruptible(2);
1267 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1268 spincnt = 0;
1271 /* NOTREACHED */
1272 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1273 return 0;
1277 * Check to see if it is time to start boosting RCU readers that are
1278 * blocking the current grace period, and, if so, tell the per-rcu_node
1279 * kthread to start boosting them. If there is an expedited grace
1280 * period in progress, it is always time to boost.
1282 * The caller must hold rnp->lock, which this function releases.
1283 * The ->boost_kthread_task is immortal, so we don't need to worry
1284 * about it going away.
1286 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1288 struct task_struct *t;
1290 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1291 rnp->n_balk_exp_gp_tasks++;
1292 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1293 return;
1295 if (rnp->exp_tasks != NULL ||
1296 (rnp->gp_tasks != NULL &&
1297 rnp->boost_tasks == NULL &&
1298 rnp->qsmask == 0 &&
1299 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1300 if (rnp->exp_tasks == NULL)
1301 rnp->boost_tasks = rnp->gp_tasks;
1302 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1303 t = rnp->boost_kthread_task;
1304 if (t)
1305 rcu_wake_cond(t, rnp->boost_kthread_status);
1306 } else {
1307 rcu_initiate_boost_trace(rnp);
1308 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1313 * Wake up the per-CPU kthread to invoke RCU callbacks.
1315 static void invoke_rcu_callbacks_kthread(void)
1317 unsigned long flags;
1319 local_irq_save(flags);
1320 __this_cpu_write(rcu_cpu_has_work, 1);
1321 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1322 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1323 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1324 __this_cpu_read(rcu_cpu_kthread_status));
1326 local_irq_restore(flags);
1330 * Is the current CPU running the RCU-callbacks kthread?
1331 * Caller must have preemption disabled.
1333 static bool rcu_is_callbacks_kthread(void)
1335 return __get_cpu_var(rcu_cpu_kthread_task) == current;
1338 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1341 * Do priority-boost accounting for the start of a new grace period.
1343 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1345 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1349 * Create an RCU-boost kthread for the specified node if one does not
1350 * already exist. We only create this kthread for preemptible RCU.
1351 * Returns zero if all is well, a negated errno otherwise.
1353 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1354 struct rcu_node *rnp)
1356 int rnp_index = rnp - &rsp->node[0];
1357 unsigned long flags;
1358 struct sched_param sp;
1359 struct task_struct *t;
1361 if (&rcu_preempt_state != rsp)
1362 return 0;
1364 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1365 return 0;
1367 rsp->boost = 1;
1368 if (rnp->boost_kthread_task != NULL)
1369 return 0;
1370 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1371 "rcub/%d", rnp_index);
1372 if (IS_ERR(t))
1373 return PTR_ERR(t);
1374 raw_spin_lock_irqsave(&rnp->lock, flags);
1375 rnp->boost_kthread_task = t;
1376 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1377 sp.sched_priority = RCU_BOOST_PRIO;
1378 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1379 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1380 return 0;
1383 static void rcu_kthread_do_work(void)
1385 rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1386 rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1387 rcu_preempt_do_callbacks();
1390 static void rcu_cpu_kthread_setup(unsigned int cpu)
1392 struct sched_param sp;
1394 sp.sched_priority = RCU_KTHREAD_PRIO;
1395 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1398 static void rcu_cpu_kthread_park(unsigned int cpu)
1400 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1403 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1405 return __get_cpu_var(rcu_cpu_has_work);
1409 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1410 * RCU softirq used in flavors and configurations of RCU that do not
1411 * support RCU priority boosting.
1413 static void rcu_cpu_kthread(unsigned int cpu)
1415 unsigned int *statusp = &__get_cpu_var(rcu_cpu_kthread_status);
1416 char work, *workp = &__get_cpu_var(rcu_cpu_has_work);
1417 int spincnt;
1419 for (spincnt = 0; spincnt < 10; spincnt++) {
1420 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1421 local_bh_disable();
1422 *statusp = RCU_KTHREAD_RUNNING;
1423 this_cpu_inc(rcu_cpu_kthread_loops);
1424 local_irq_disable();
1425 work = *workp;
1426 *workp = 0;
1427 local_irq_enable();
1428 if (work)
1429 rcu_kthread_do_work();
1430 local_bh_enable();
1431 if (*workp == 0) {
1432 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1433 *statusp = RCU_KTHREAD_WAITING;
1434 return;
1437 *statusp = RCU_KTHREAD_YIELDING;
1438 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1439 schedule_timeout_interruptible(2);
1440 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1441 *statusp = RCU_KTHREAD_WAITING;
1445 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1446 * served by the rcu_node in question. The CPU hotplug lock is still
1447 * held, so the value of rnp->qsmaskinit will be stable.
1449 * We don't include outgoingcpu in the affinity set, use -1 if there is
1450 * no outgoing CPU. If there are no CPUs left in the affinity set,
1451 * this function allows the kthread to execute on any CPU.
1453 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1455 struct task_struct *t = rnp->boost_kthread_task;
1456 unsigned long mask = rnp->qsmaskinit;
1457 cpumask_var_t cm;
1458 int cpu;
1460 if (!t)
1461 return;
1462 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1463 return;
1464 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1465 if ((mask & 0x1) && cpu != outgoingcpu)
1466 cpumask_set_cpu(cpu, cm);
1467 if (cpumask_weight(cm) == 0) {
1468 cpumask_setall(cm);
1469 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1470 cpumask_clear_cpu(cpu, cm);
1471 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1473 set_cpus_allowed_ptr(t, cm);
1474 free_cpumask_var(cm);
1477 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1478 .store = &rcu_cpu_kthread_task,
1479 .thread_should_run = rcu_cpu_kthread_should_run,
1480 .thread_fn = rcu_cpu_kthread,
1481 .thread_comm = "rcuc/%u",
1482 .setup = rcu_cpu_kthread_setup,
1483 .park = rcu_cpu_kthread_park,
1487 * Spawn all kthreads -- called as soon as the scheduler is running.
1489 static int __init rcu_spawn_kthreads(void)
1491 struct rcu_node *rnp;
1492 int cpu;
1494 rcu_scheduler_fully_active = 1;
1495 for_each_possible_cpu(cpu)
1496 per_cpu(rcu_cpu_has_work, cpu) = 0;
1497 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1498 rnp = rcu_get_root(rcu_state);
1499 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1500 if (NUM_RCU_NODES > 1) {
1501 rcu_for_each_leaf_node(rcu_state, rnp)
1502 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1504 return 0;
1506 early_initcall(rcu_spawn_kthreads);
1508 static void rcu_prepare_kthreads(int cpu)
1510 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1511 struct rcu_node *rnp = rdp->mynode;
1513 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1514 if (rcu_scheduler_fully_active)
1515 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1518 #else /* #ifdef CONFIG_RCU_BOOST */
1520 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1522 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1525 static void invoke_rcu_callbacks_kthread(void)
1527 WARN_ON_ONCE(1);
1530 static bool rcu_is_callbacks_kthread(void)
1532 return false;
1535 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1539 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1543 static int __init rcu_scheduler_really_started(void)
1545 rcu_scheduler_fully_active = 1;
1546 return 0;
1548 early_initcall(rcu_scheduler_really_started);
1550 static void rcu_prepare_kthreads(int cpu)
1554 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1556 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1559 * Check to see if any future RCU-related work will need to be done
1560 * by the current CPU, even if none need be done immediately, returning
1561 * 1 if so. This function is part of the RCU implementation; it is -not-
1562 * an exported member of the RCU API.
1564 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1565 * any flavor of RCU.
1567 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1569 *delta_jiffies = ULONG_MAX;
1570 return rcu_cpu_has_callbacks(cpu, NULL);
1574 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1575 * after it.
1577 static void rcu_cleanup_after_idle(int cpu)
1582 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1583 * is nothing.
1585 static void rcu_prepare_for_idle(int cpu)
1590 * Don't bother keeping a running count of the number of RCU callbacks
1591 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1593 static void rcu_idle_count_callbacks_posted(void)
1597 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1600 * This code is invoked when a CPU goes idle, at which point we want
1601 * to have the CPU do everything required for RCU so that it can enter
1602 * the energy-efficient dyntick-idle mode. This is handled by a
1603 * state machine implemented by rcu_prepare_for_idle() below.
1605 * The following three proprocessor symbols control this state machine:
1607 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1608 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1609 * is sized to be roughly one RCU grace period. Those energy-efficiency
1610 * benchmarkers who might otherwise be tempted to set this to a large
1611 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1612 * system. And if you are -that- concerned about energy efficiency,
1613 * just power the system down and be done with it!
1614 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1615 * permitted to sleep in dyntick-idle mode with only lazy RCU
1616 * callbacks pending. Setting this too high can OOM your system.
1618 * The values below work well in practice. If future workloads require
1619 * adjustment, they can be converted into kernel config parameters, though
1620 * making the state machine smarter might be a better option.
1622 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1623 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1625 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1626 module_param(rcu_idle_gp_delay, int, 0644);
1627 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1628 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1630 extern int tick_nohz_active;
1633 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1634 * only if it has been awhile since the last time we did so. Afterwards,
1635 * if there are any callbacks ready for immediate invocation, return true.
1637 static bool rcu_try_advance_all_cbs(void)
1639 bool cbs_ready = false;
1640 struct rcu_data *rdp;
1641 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1642 struct rcu_node *rnp;
1643 struct rcu_state *rsp;
1645 /* Exit early if we advanced recently. */
1646 if (jiffies == rdtp->last_advance_all)
1647 return 0;
1648 rdtp->last_advance_all = jiffies;
1650 for_each_rcu_flavor(rsp) {
1651 rdp = this_cpu_ptr(rsp->rda);
1652 rnp = rdp->mynode;
1655 * Don't bother checking unless a grace period has
1656 * completed since we last checked and there are
1657 * callbacks not yet ready to invoke.
1659 if (rdp->completed != rnp->completed &&
1660 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1661 note_gp_changes(rsp, rdp);
1663 if (cpu_has_callbacks_ready_to_invoke(rdp))
1664 cbs_ready = true;
1666 return cbs_ready;
1670 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1671 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1672 * caller to set the timeout based on whether or not there are non-lazy
1673 * callbacks.
1675 * The caller must have disabled interrupts.
1677 int rcu_needs_cpu(int cpu, unsigned long *dj)
1679 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1681 /* Snapshot to detect later posting of non-lazy callback. */
1682 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1684 /* If no callbacks, RCU doesn't need the CPU. */
1685 if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1686 *dj = ULONG_MAX;
1687 return 0;
1690 /* Attempt to advance callbacks. */
1691 if (rcu_try_advance_all_cbs()) {
1692 /* Some ready to invoke, so initiate later invocation. */
1693 invoke_rcu_core();
1694 return 1;
1696 rdtp->last_accelerate = jiffies;
1698 /* Request timer delay depending on laziness, and round. */
1699 if (!rdtp->all_lazy) {
1700 *dj = round_up(rcu_idle_gp_delay + jiffies,
1701 rcu_idle_gp_delay) - jiffies;
1702 } else {
1703 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1705 return 0;
1709 * Prepare a CPU for idle from an RCU perspective. The first major task
1710 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1711 * The second major task is to check to see if a non-lazy callback has
1712 * arrived at a CPU that previously had only lazy callbacks. The third
1713 * major task is to accelerate (that is, assign grace-period numbers to)
1714 * any recently arrived callbacks.
1716 * The caller must have disabled interrupts.
1718 static void rcu_prepare_for_idle(int cpu)
1720 struct rcu_data *rdp;
1721 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1722 struct rcu_node *rnp;
1723 struct rcu_state *rsp;
1724 int tne;
1726 /* Handle nohz enablement switches conservatively. */
1727 tne = ACCESS_ONCE(tick_nohz_active);
1728 if (tne != rdtp->tick_nohz_enabled_snap) {
1729 if (rcu_cpu_has_callbacks(cpu, NULL))
1730 invoke_rcu_core(); /* force nohz to see update. */
1731 rdtp->tick_nohz_enabled_snap = tne;
1732 return;
1734 if (!tne)
1735 return;
1737 /* If this is a no-CBs CPU, no callbacks, just return. */
1738 if (rcu_is_nocb_cpu(cpu))
1739 return;
1742 * If a non-lazy callback arrived at a CPU having only lazy
1743 * callbacks, invoke RCU core for the side-effect of recalculating
1744 * idle duration on re-entry to idle.
1746 if (rdtp->all_lazy &&
1747 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1748 rdtp->all_lazy = false;
1749 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1750 invoke_rcu_core();
1751 return;
1755 * If we have not yet accelerated this jiffy, accelerate all
1756 * callbacks on this CPU.
1758 if (rdtp->last_accelerate == jiffies)
1759 return;
1760 rdtp->last_accelerate = jiffies;
1761 for_each_rcu_flavor(rsp) {
1762 rdp = per_cpu_ptr(rsp->rda, cpu);
1763 if (!*rdp->nxttail[RCU_DONE_TAIL])
1764 continue;
1765 rnp = rdp->mynode;
1766 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1767 rcu_accelerate_cbs(rsp, rnp, rdp);
1768 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1773 * Clean up for exit from idle. Attempt to advance callbacks based on
1774 * any grace periods that elapsed while the CPU was idle, and if any
1775 * callbacks are now ready to invoke, initiate invocation.
1777 static void rcu_cleanup_after_idle(int cpu)
1779 struct rcu_data *rdp;
1780 struct rcu_state *rsp;
1782 if (rcu_is_nocb_cpu(cpu))
1783 return;
1784 rcu_try_advance_all_cbs();
1785 for_each_rcu_flavor(rsp) {
1786 rdp = per_cpu_ptr(rsp->rda, cpu);
1787 if (cpu_has_callbacks_ready_to_invoke(rdp))
1788 invoke_rcu_core();
1793 * Keep a running count of the number of non-lazy callbacks posted
1794 * on this CPU. This running counter (which is never decremented) allows
1795 * rcu_prepare_for_idle() to detect when something out of the idle loop
1796 * posts a callback, even if an equal number of callbacks are invoked.
1797 * Of course, callbacks should only be posted from within a trace event
1798 * designed to be called from idle or from within RCU_NONIDLE().
1800 static void rcu_idle_count_callbacks_posted(void)
1802 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1806 * Data for flushing lazy RCU callbacks at OOM time.
1808 static atomic_t oom_callback_count;
1809 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1812 * RCU OOM callback -- decrement the outstanding count and deliver the
1813 * wake-up if we are the last one.
1815 static void rcu_oom_callback(struct rcu_head *rhp)
1817 if (atomic_dec_and_test(&oom_callback_count))
1818 wake_up(&oom_callback_wq);
1822 * Post an rcu_oom_notify callback on the current CPU if it has at
1823 * least one lazy callback. This will unnecessarily post callbacks
1824 * to CPUs that already have a non-lazy callback at the end of their
1825 * callback list, but this is an infrequent operation, so accept some
1826 * extra overhead to keep things simple.
1828 static void rcu_oom_notify_cpu(void *unused)
1830 struct rcu_state *rsp;
1831 struct rcu_data *rdp;
1833 for_each_rcu_flavor(rsp) {
1834 rdp = __this_cpu_ptr(rsp->rda);
1835 if (rdp->qlen_lazy != 0) {
1836 atomic_inc(&oom_callback_count);
1837 rsp->call(&rdp->oom_head, rcu_oom_callback);
1843 * If low on memory, ensure that each CPU has a non-lazy callback.
1844 * This will wake up CPUs that have only lazy callbacks, in turn
1845 * ensuring that they free up the corresponding memory in a timely manner.
1846 * Because an uncertain amount of memory will be freed in some uncertain
1847 * timeframe, we do not claim to have freed anything.
1849 static int rcu_oom_notify(struct notifier_block *self,
1850 unsigned long notused, void *nfreed)
1852 int cpu;
1854 /* Wait for callbacks from earlier instance to complete. */
1855 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1858 * Prevent premature wakeup: ensure that all increments happen
1859 * before there is a chance of the counter reaching zero.
1861 atomic_set(&oom_callback_count, 1);
1863 get_online_cpus();
1864 for_each_online_cpu(cpu) {
1865 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1866 cond_resched();
1868 put_online_cpus();
1870 /* Unconditionally decrement: no need to wake ourselves up. */
1871 atomic_dec(&oom_callback_count);
1873 return NOTIFY_OK;
1876 static struct notifier_block rcu_oom_nb = {
1877 .notifier_call = rcu_oom_notify
1880 static int __init rcu_register_oom_notifier(void)
1882 register_oom_notifier(&rcu_oom_nb);
1883 return 0;
1885 early_initcall(rcu_register_oom_notifier);
1887 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1889 #ifdef CONFIG_RCU_CPU_STALL_INFO
1891 #ifdef CONFIG_RCU_FAST_NO_HZ
1893 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1895 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1896 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1898 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1899 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1900 ulong2long(nlpd),
1901 rdtp->all_lazy ? 'L' : '.',
1902 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1905 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1907 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1909 *cp = '\0';
1912 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1914 /* Initiate the stall-info list. */
1915 static void print_cpu_stall_info_begin(void)
1917 pr_cont("\n");
1921 * Print out diagnostic information for the specified stalled CPU.
1923 * If the specified CPU is aware of the current RCU grace period
1924 * (flavor specified by rsp), then print the number of scheduling
1925 * clock interrupts the CPU has taken during the time that it has
1926 * been aware. Otherwise, print the number of RCU grace periods
1927 * that this CPU is ignorant of, for example, "1" if the CPU was
1928 * aware of the previous grace period.
1930 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1932 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1934 char fast_no_hz[72];
1935 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1936 struct rcu_dynticks *rdtp = rdp->dynticks;
1937 char *ticks_title;
1938 unsigned long ticks_value;
1940 if (rsp->gpnum == rdp->gpnum) {
1941 ticks_title = "ticks this GP";
1942 ticks_value = rdp->ticks_this_gp;
1943 } else {
1944 ticks_title = "GPs behind";
1945 ticks_value = rsp->gpnum - rdp->gpnum;
1947 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1948 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1949 cpu, ticks_value, ticks_title,
1950 atomic_read(&rdtp->dynticks) & 0xfff,
1951 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1952 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1953 fast_no_hz);
1956 /* Terminate the stall-info list. */
1957 static void print_cpu_stall_info_end(void)
1959 pr_err("\t");
1962 /* Zero ->ticks_this_gp for all flavors of RCU. */
1963 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1965 rdp->ticks_this_gp = 0;
1966 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1969 /* Increment ->ticks_this_gp for all flavors of RCU. */
1970 static void increment_cpu_stall_ticks(void)
1972 struct rcu_state *rsp;
1974 for_each_rcu_flavor(rsp)
1975 __this_cpu_ptr(rsp->rda)->ticks_this_gp++;
1978 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1980 static void print_cpu_stall_info_begin(void)
1982 pr_cont(" {");
1985 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1987 pr_cont(" %d", cpu);
1990 static void print_cpu_stall_info_end(void)
1992 pr_cont("} ");
1995 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1999 static void increment_cpu_stall_ticks(void)
2003 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
2005 #ifdef CONFIG_RCU_NOCB_CPU
2008 * Offload callback processing from the boot-time-specified set of CPUs
2009 * specified by rcu_nocb_mask. For each CPU in the set, there is a
2010 * kthread created that pulls the callbacks from the corresponding CPU,
2011 * waits for a grace period to elapse, and invokes the callbacks.
2012 * The no-CBs CPUs do a wake_up() on their kthread when they insert
2013 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2014 * has been specified, in which case each kthread actively polls its
2015 * CPU. (Which isn't so great for energy efficiency, but which does
2016 * reduce RCU's overhead on that CPU.)
2018 * This is intended to be used in conjunction with Frederic Weisbecker's
2019 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2020 * running CPU-bound user-mode computations.
2022 * Offloading of callback processing could also in theory be used as
2023 * an energy-efficiency measure because CPUs with no RCU callbacks
2024 * queued are more aggressive about entering dyntick-idle mode.
2028 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2029 static int __init rcu_nocb_setup(char *str)
2031 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
2032 have_rcu_nocb_mask = true;
2033 cpulist_parse(str, rcu_nocb_mask);
2034 return 1;
2036 __setup("rcu_nocbs=", rcu_nocb_setup);
2038 static int __init parse_rcu_nocb_poll(char *arg)
2040 rcu_nocb_poll = 1;
2041 return 0;
2043 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2046 * Do any no-CBs CPUs need another grace period?
2048 * Interrupts must be disabled. If the caller does not hold the root
2049 * rnp_node structure's ->lock, the results are advisory only.
2051 static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2053 struct rcu_node *rnp = rcu_get_root(rsp);
2055 return rnp->need_future_gp[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
2059 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2060 * grace period.
2062 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2064 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2068 * Set the root rcu_node structure's ->need_future_gp field
2069 * based on the sum of those of all rcu_node structures. This does
2070 * double-count the root rcu_node structure's requests, but this
2071 * is necessary to handle the possibility of a rcu_nocb_kthread()
2072 * having awakened during the time that the rcu_node structures
2073 * were being updated for the end of the previous grace period.
2075 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2077 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2080 static void rcu_init_one_nocb(struct rcu_node *rnp)
2082 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2083 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2086 /* Is the specified CPU a no-CPUs CPU? */
2087 bool rcu_is_nocb_cpu(int cpu)
2089 if (have_rcu_nocb_mask)
2090 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2091 return false;
2095 * Enqueue the specified string of rcu_head structures onto the specified
2096 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2097 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2098 * counts are supplied by rhcount and rhcount_lazy.
2100 * If warranted, also wake up the kthread servicing this CPUs queues.
2102 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2103 struct rcu_head *rhp,
2104 struct rcu_head **rhtp,
2105 int rhcount, int rhcount_lazy)
2107 int len;
2108 struct rcu_head **old_rhpp;
2109 struct task_struct *t;
2111 /* Enqueue the callback on the nocb list and update counts. */
2112 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2113 ACCESS_ONCE(*old_rhpp) = rhp;
2114 atomic_long_add(rhcount, &rdp->nocb_q_count);
2115 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2117 /* If we are not being polled and there is a kthread, awaken it ... */
2118 t = ACCESS_ONCE(rdp->nocb_kthread);
2119 if (rcu_nocb_poll | !t)
2120 return;
2121 len = atomic_long_read(&rdp->nocb_q_count);
2122 if (old_rhpp == &rdp->nocb_head) {
2123 wake_up(&rdp->nocb_wq); /* ... only if queue was empty ... */
2124 rdp->qlen_last_fqs_check = 0;
2125 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2126 wake_up_process(t); /* ... or if many callbacks queued. */
2127 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2129 return;
2133 * This is a helper for __call_rcu(), which invokes this when the normal
2134 * callback queue is inoperable. If this is not a no-CBs CPU, this
2135 * function returns failure back to __call_rcu(), which can complain
2136 * appropriately.
2138 * Otherwise, this function queues the callback where the corresponding
2139 * "rcuo" kthread can find it.
2141 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2142 bool lazy)
2145 if (!rcu_is_nocb_cpu(rdp->cpu))
2146 return 0;
2147 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy);
2148 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2149 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2150 (unsigned long)rhp->func,
2151 rdp->qlen_lazy, rdp->qlen);
2152 else
2153 trace_rcu_callback(rdp->rsp->name, rhp,
2154 rdp->qlen_lazy, rdp->qlen);
2155 return 1;
2159 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2160 * not a no-CBs CPU.
2162 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2163 struct rcu_data *rdp)
2165 long ql = rsp->qlen;
2166 long qll = rsp->qlen_lazy;
2168 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2169 if (!rcu_is_nocb_cpu(smp_processor_id()))
2170 return 0;
2171 rsp->qlen = 0;
2172 rsp->qlen_lazy = 0;
2174 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2175 if (rsp->orphan_donelist != NULL) {
2176 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2177 rsp->orphan_donetail, ql, qll);
2178 ql = qll = 0;
2179 rsp->orphan_donelist = NULL;
2180 rsp->orphan_donetail = &rsp->orphan_donelist;
2182 if (rsp->orphan_nxtlist != NULL) {
2183 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2184 rsp->orphan_nxttail, ql, qll);
2185 ql = qll = 0;
2186 rsp->orphan_nxtlist = NULL;
2187 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2189 return 1;
2193 * If necessary, kick off a new grace period, and either way wait
2194 * for a subsequent grace period to complete.
2196 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2198 unsigned long c;
2199 bool d;
2200 unsigned long flags;
2201 struct rcu_node *rnp = rdp->mynode;
2203 raw_spin_lock_irqsave(&rnp->lock, flags);
2204 c = rcu_start_future_gp(rnp, rdp);
2205 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2208 * Wait for the grace period. Do so interruptibly to avoid messing
2209 * up the load average.
2211 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2212 for (;;) {
2213 wait_event_interruptible(
2214 rnp->nocb_gp_wq[c & 0x1],
2215 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2216 if (likely(d))
2217 break;
2218 flush_signals(current);
2219 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2221 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2222 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2226 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2227 * callbacks queued by the corresponding no-CBs CPU.
2229 static int rcu_nocb_kthread(void *arg)
2231 int c, cl;
2232 struct rcu_head *list;
2233 struct rcu_head *next;
2234 struct rcu_head **tail;
2235 struct rcu_data *rdp = arg;
2237 /* Each pass through this loop invokes one batch of callbacks */
2238 for (;;) {
2239 /* If not polling, wait for next batch of callbacks. */
2240 if (!rcu_nocb_poll)
2241 wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head);
2242 list = ACCESS_ONCE(rdp->nocb_head);
2243 if (!list) {
2244 schedule_timeout_interruptible(1);
2245 flush_signals(current);
2246 continue;
2250 * Extract queued callbacks, update counts, and wait
2251 * for a grace period to elapse.
2253 ACCESS_ONCE(rdp->nocb_head) = NULL;
2254 tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2255 c = atomic_long_xchg(&rdp->nocb_q_count, 0);
2256 cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2257 ACCESS_ONCE(rdp->nocb_p_count) += c;
2258 ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl;
2259 rcu_nocb_wait_gp(rdp);
2261 /* Each pass through the following loop invokes a callback. */
2262 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2263 c = cl = 0;
2264 while (list) {
2265 next = list->next;
2266 /* Wait for enqueuing to complete, if needed. */
2267 while (next == NULL && &list->next != tail) {
2268 schedule_timeout_interruptible(1);
2269 next = list->next;
2271 debug_rcu_head_unqueue(list);
2272 local_bh_disable();
2273 if (__rcu_reclaim(rdp->rsp->name, list))
2274 cl++;
2275 c++;
2276 local_bh_enable();
2277 list = next;
2279 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2280 ACCESS_ONCE(rdp->nocb_p_count) -= c;
2281 ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2282 rdp->n_nocbs_invoked += c;
2284 return 0;
2287 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2288 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2290 rdp->nocb_tail = &rdp->nocb_head;
2291 init_waitqueue_head(&rdp->nocb_wq);
2294 /* Create a kthread for each RCU flavor for each no-CBs CPU. */
2295 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2297 int cpu;
2298 struct rcu_data *rdp;
2299 struct task_struct *t;
2301 if (rcu_nocb_mask == NULL)
2302 return;
2303 for_each_cpu(cpu, rcu_nocb_mask) {
2304 rdp = per_cpu_ptr(rsp->rda, cpu);
2305 t = kthread_run(rcu_nocb_kthread, rdp,
2306 "rcuo%c/%d", rsp->abbr, cpu);
2307 BUG_ON(IS_ERR(t));
2308 ACCESS_ONCE(rdp->nocb_kthread) = t;
2312 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2313 static bool init_nocb_callback_list(struct rcu_data *rdp)
2315 if (rcu_nocb_mask == NULL ||
2316 !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2317 return false;
2318 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2319 return true;
2322 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2324 static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2326 return 0;
2329 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2333 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2337 static void rcu_init_one_nocb(struct rcu_node *rnp)
2341 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2342 bool lazy)
2344 return 0;
2347 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2348 struct rcu_data *rdp)
2350 return 0;
2353 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2357 static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2361 static bool init_nocb_callback_list(struct rcu_data *rdp)
2363 return false;
2366 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2369 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2370 * arbitrarily long period of time with the scheduling-clock tick turned
2371 * off. RCU will be paying attention to this CPU because it is in the
2372 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2373 * machine because the scheduling-clock tick has been disabled. Therefore,
2374 * if an adaptive-ticks CPU is failing to respond to the current grace
2375 * period and has not be idle from an RCU perspective, kick it.
2377 static void rcu_kick_nohz_cpu(int cpu)
2379 #ifdef CONFIG_NO_HZ_FULL
2380 if (tick_nohz_full_cpu(cpu))
2381 smp_send_reschedule(cpu);
2382 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2386 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2389 * Define RCU flavor that holds sysidle state. This needs to be the
2390 * most active flavor of RCU.
2392 #ifdef CONFIG_PREEMPT_RCU
2393 static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state;
2394 #else /* #ifdef CONFIG_PREEMPT_RCU */
2395 static struct rcu_state *rcu_sysidle_state = &rcu_sched_state;
2396 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2398 static int full_sysidle_state; /* Current system-idle state. */
2399 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2400 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2401 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2402 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2403 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2406 * Invoked to note exit from irq or task transition to idle. Note that
2407 * usermode execution does -not- count as idle here! After all, we want
2408 * to detect full-system idle states, not RCU quiescent states and grace
2409 * periods. The caller must have disabled interrupts.
2411 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2413 unsigned long j;
2415 /* Adjust nesting, check for fully idle. */
2416 if (irq) {
2417 rdtp->dynticks_idle_nesting--;
2418 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2419 if (rdtp->dynticks_idle_nesting != 0)
2420 return; /* Still not fully idle. */
2421 } else {
2422 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2423 DYNTICK_TASK_NEST_VALUE) {
2424 rdtp->dynticks_idle_nesting = 0;
2425 } else {
2426 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2427 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2428 return; /* Still not fully idle. */
2432 /* Record start of fully idle period. */
2433 j = jiffies;
2434 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2435 smp_mb__before_atomic_inc();
2436 atomic_inc(&rdtp->dynticks_idle);
2437 smp_mb__after_atomic_inc();
2438 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2442 * Unconditionally force exit from full system-idle state. This is
2443 * invoked when a normal CPU exits idle, but must be called separately
2444 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2445 * is that the timekeeping CPU is permitted to take scheduling-clock
2446 * interrupts while the system is in system-idle state, and of course
2447 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2448 * interrupt from any other type of interrupt.
2450 void rcu_sysidle_force_exit(void)
2452 int oldstate = ACCESS_ONCE(full_sysidle_state);
2453 int newoldstate;
2456 * Each pass through the following loop attempts to exit full
2457 * system-idle state. If contention proves to be a problem,
2458 * a trylock-based contention tree could be used here.
2460 while (oldstate > RCU_SYSIDLE_SHORT) {
2461 newoldstate = cmpxchg(&full_sysidle_state,
2462 oldstate, RCU_SYSIDLE_NOT);
2463 if (oldstate == newoldstate &&
2464 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2465 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2466 return; /* We cleared it, done! */
2468 oldstate = newoldstate;
2470 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2474 * Invoked to note entry to irq or task transition from idle. Note that
2475 * usermode execution does -not- count as idle here! The caller must
2476 * have disabled interrupts.
2478 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2480 /* Adjust nesting, check for already non-idle. */
2481 if (irq) {
2482 rdtp->dynticks_idle_nesting++;
2483 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2484 if (rdtp->dynticks_idle_nesting != 1)
2485 return; /* Already non-idle. */
2486 } else {
2488 * Allow for irq misnesting. Yes, it really is possible
2489 * to enter an irq handler then never leave it, and maybe
2490 * also vice versa. Handle both possibilities.
2492 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2493 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2494 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2495 return; /* Already non-idle. */
2496 } else {
2497 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2501 /* Record end of idle period. */
2502 smp_mb__before_atomic_inc();
2503 atomic_inc(&rdtp->dynticks_idle);
2504 smp_mb__after_atomic_inc();
2505 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2508 * If we are the timekeeping CPU, we are permitted to be non-idle
2509 * during a system-idle state. This must be the case, because
2510 * the timekeeping CPU has to take scheduling-clock interrupts
2511 * during the time that the system is transitioning to full
2512 * system-idle state. This means that the timekeeping CPU must
2513 * invoke rcu_sysidle_force_exit() directly if it does anything
2514 * more than take a scheduling-clock interrupt.
2516 if (smp_processor_id() == tick_do_timer_cpu)
2517 return;
2519 /* Update system-idle state: We are clearly no longer fully idle! */
2520 rcu_sysidle_force_exit();
2524 * Check to see if the current CPU is idle. Note that usermode execution
2525 * does not count as idle. The caller must have disabled interrupts.
2527 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2528 unsigned long *maxj)
2530 int cur;
2531 unsigned long j;
2532 struct rcu_dynticks *rdtp = rdp->dynticks;
2535 * If some other CPU has already reported non-idle, if this is
2536 * not the flavor of RCU that tracks sysidle state, or if this
2537 * is an offline or the timekeeping CPU, nothing to do.
2539 if (!*isidle || rdp->rsp != rcu_sysidle_state ||
2540 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2541 return;
2542 if (rcu_gp_in_progress(rdp->rsp))
2543 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2545 /* Pick up current idle and NMI-nesting counter and check. */
2546 cur = atomic_read(&rdtp->dynticks_idle);
2547 if (cur & 0x1) {
2548 *isidle = false; /* We are not idle! */
2549 return;
2551 smp_mb(); /* Read counters before timestamps. */
2553 /* Pick up timestamps. */
2554 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2555 /* If this CPU entered idle more recently, update maxj timestamp. */
2556 if (ULONG_CMP_LT(*maxj, j))
2557 *maxj = j;
2561 * Is this the flavor of RCU that is handling full-system idle?
2563 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2565 return rsp == rcu_sysidle_state;
2569 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2570 * timekeeping CPU.
2572 static void rcu_bind_gp_kthread(void)
2574 int cpu = ACCESS_ONCE(tick_do_timer_cpu);
2576 if (cpu < 0 || cpu >= nr_cpu_ids)
2577 return;
2578 if (raw_smp_processor_id() != cpu)
2579 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2583 * Return a delay in jiffies based on the number of CPUs, rcu_node
2584 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2585 * systems more time to transition to full-idle state in order to
2586 * avoid the cache thrashing that otherwise occur on the state variable.
2587 * Really small systems (less than a couple of tens of CPUs) should
2588 * instead use a single global atomically incremented counter, and later
2589 * versions of this will automatically reconfigure themselves accordingly.
2591 static unsigned long rcu_sysidle_delay(void)
2593 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2594 return 0;
2595 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2599 * Advance the full-system-idle state. This is invoked when all of
2600 * the non-timekeeping CPUs are idle.
2602 static void rcu_sysidle(unsigned long j)
2604 /* Check the current state. */
2605 switch (ACCESS_ONCE(full_sysidle_state)) {
2606 case RCU_SYSIDLE_NOT:
2608 /* First time all are idle, so note a short idle period. */
2609 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2610 break;
2612 case RCU_SYSIDLE_SHORT:
2615 * Idle for a bit, time to advance to next state?
2616 * cmpxchg failure means race with non-idle, let them win.
2618 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2619 (void)cmpxchg(&full_sysidle_state,
2620 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2621 break;
2623 case RCU_SYSIDLE_LONG:
2626 * Do an additional check pass before advancing to full.
2627 * cmpxchg failure means race with non-idle, let them win.
2629 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2630 (void)cmpxchg(&full_sysidle_state,
2631 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2632 break;
2634 default:
2635 break;
2640 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2641 * back to the beginning.
2643 static void rcu_sysidle_cancel(void)
2645 smp_mb();
2646 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2650 * Update the sysidle state based on the results of a force-quiescent-state
2651 * scan of the CPUs' dyntick-idle state.
2653 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2654 unsigned long maxj, bool gpkt)
2656 if (rsp != rcu_sysidle_state)
2657 return; /* Wrong flavor, ignore. */
2658 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2659 return; /* Running state machine from timekeeping CPU. */
2660 if (isidle)
2661 rcu_sysidle(maxj); /* More idle! */
2662 else
2663 rcu_sysidle_cancel(); /* Idle is over. */
2667 * Wrapper for rcu_sysidle_report() when called from the grace-period
2668 * kthread's context.
2670 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2671 unsigned long maxj)
2673 rcu_sysidle_report(rsp, isidle, maxj, true);
2676 /* Callback and function for forcing an RCU grace period. */
2677 struct rcu_sysidle_head {
2678 struct rcu_head rh;
2679 int inuse;
2682 static void rcu_sysidle_cb(struct rcu_head *rhp)
2684 struct rcu_sysidle_head *rshp;
2687 * The following memory barrier is needed to replace the
2688 * memory barriers that would normally be in the memory
2689 * allocator.
2691 smp_mb(); /* grace period precedes setting inuse. */
2693 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2694 ACCESS_ONCE(rshp->inuse) = 0;
2698 * Check to see if the system is fully idle, other than the timekeeping CPU.
2699 * The caller must have disabled interrupts.
2701 bool rcu_sys_is_idle(void)
2703 static struct rcu_sysidle_head rsh;
2704 int rss = ACCESS_ONCE(full_sysidle_state);
2706 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2707 return false;
2709 /* Handle small-system case by doing a full scan of CPUs. */
2710 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2711 int oldrss = rss - 1;
2714 * One pass to advance to each state up to _FULL.
2715 * Give up if any pass fails to advance the state.
2717 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2718 int cpu;
2719 bool isidle = true;
2720 unsigned long maxj = jiffies - ULONG_MAX / 4;
2721 struct rcu_data *rdp;
2723 /* Scan all the CPUs looking for nonidle CPUs. */
2724 for_each_possible_cpu(cpu) {
2725 rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu);
2726 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2727 if (!isidle)
2728 break;
2730 rcu_sysidle_report(rcu_sysidle_state,
2731 isidle, maxj, false);
2732 oldrss = rss;
2733 rss = ACCESS_ONCE(full_sysidle_state);
2737 /* If this is the first observation of an idle period, record it. */
2738 if (rss == RCU_SYSIDLE_FULL) {
2739 rss = cmpxchg(&full_sysidle_state,
2740 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2741 return rss == RCU_SYSIDLE_FULL;
2744 smp_mb(); /* ensure rss load happens before later caller actions. */
2746 /* If already fully idle, tell the caller (in case of races). */
2747 if (rss == RCU_SYSIDLE_FULL_NOTED)
2748 return true;
2751 * If we aren't there yet, and a grace period is not in flight,
2752 * initiate a grace period. Either way, tell the caller that
2753 * we are not there yet. We use an xchg() rather than an assignment
2754 * to make up for the memory barriers that would otherwise be
2755 * provided by the memory allocator.
2757 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2758 !rcu_gp_in_progress(rcu_sysidle_state) &&
2759 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2760 call_rcu(&rsh.rh, rcu_sysidle_cb);
2761 return false;
2765 * Initialize dynticks sysidle state for CPUs coming online.
2767 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2769 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2772 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2774 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2778 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2782 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2783 unsigned long *maxj)
2787 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2789 return false;
2792 static void rcu_bind_gp_kthread(void)
2796 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2797 unsigned long maxj)
2801 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2805 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */