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
38 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
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");
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",
62 #ifdef CONFIG_RCU_FANOUT_EXACT
63 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
65 #ifdef CONFIG_RCU_FAST_NO_HZ
66 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
68 #ifdef CONFIG_PROVE_RCU
69 pr_info("\tRCU lockdep checking is enabled.\n");
71 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
72 pr_info("\tRCU torture testing starts during boot.\n");
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");
77 #if defined(CONFIG_RCU_CPU_STALL_INFO)
78 pr_info("\tAdditional per-CPU info printed with stalls.\n");
80 #if NUM_RCU_LVL_4 != 0
81 pr_info("\tFour-level hierarchy is enabled.\n");
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_copy(rcu_nocb_mask
, cpu_possible_mask
);
100 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
101 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
102 if (have_rcu_nocb_mask
) {
103 if (!cpumask_subset(rcu_nocb_mask
, cpu_possible_mask
)) {
104 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
105 cpumask_and(rcu_nocb_mask
, cpu_possible_mask
,
108 cpulist_scnprintf(nocb_buf
, sizeof(nocb_buf
), rcu_nocb_mask
);
109 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf
);
111 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
113 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
116 #ifdef CONFIG_TREE_PREEMPT_RCU
118 RCU_STATE_INITIALIZER(rcu_preempt
, 'p', call_rcu
);
119 static struct rcu_state
*rcu_state
= &rcu_preempt_state
;
121 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
);
124 * Tell them what RCU they are running.
126 static void __init
rcu_bootup_announce(void)
128 pr_info("Preemptible hierarchical RCU implementation.\n");
129 rcu_bootup_announce_oddness();
133 * Return the number of RCU-preempt batches processed thus far
134 * for debug and statistics.
136 long rcu_batches_completed_preempt(void)
138 return rcu_preempt_state
.completed
;
140 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt
);
143 * Return the number of RCU batches processed thus far for debug & stats.
145 long rcu_batches_completed(void)
147 return rcu_batches_completed_preempt();
149 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
152 * Force a quiescent state for preemptible RCU.
154 void rcu_force_quiescent_state(void)
156 force_quiescent_state(&rcu_preempt_state
);
158 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
161 * Record a preemptible-RCU quiescent state for the specified CPU. Note
162 * that this just means that the task currently running on the CPU is
163 * not in a quiescent state. There might be any number of tasks blocked
164 * while in an RCU read-side critical section.
166 * Unlike the other rcu_*_qs() functions, callers to this function
167 * must disable irqs in order to protect the assignment to
168 * ->rcu_read_unlock_special.
170 static void rcu_preempt_qs(int cpu
)
172 struct rcu_data
*rdp
= &per_cpu(rcu_preempt_data
, cpu
);
174 if (rdp
->passed_quiesce
== 0)
175 trace_rcu_grace_period(TPS("rcu_preempt"), rdp
->gpnum
, TPS("cpuqs"));
176 rdp
->passed_quiesce
= 1;
177 current
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_NEED_QS
;
181 * We have entered the scheduler, and the current task might soon be
182 * context-switched away from. If this task is in an RCU read-side
183 * critical section, we will no longer be able to rely on the CPU to
184 * record that fact, so we enqueue the task on the blkd_tasks list.
185 * The task will dequeue itself when it exits the outermost enclosing
186 * RCU read-side critical section. Therefore, the current grace period
187 * cannot be permitted to complete until the blkd_tasks list entries
188 * predating the current grace period drain, in other words, until
189 * rnp->gp_tasks becomes NULL.
191 * Caller must disable preemption.
193 static void rcu_preempt_note_context_switch(int cpu
)
195 struct task_struct
*t
= current
;
197 struct rcu_data
*rdp
;
198 struct rcu_node
*rnp
;
200 if (t
->rcu_read_lock_nesting
> 0 &&
201 (t
->rcu_read_unlock_special
& RCU_READ_UNLOCK_BLOCKED
) == 0) {
203 /* Possibly blocking in an RCU read-side critical section. */
204 rdp
= per_cpu_ptr(rcu_preempt_state
.rda
, cpu
);
206 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
207 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_BLOCKED
;
208 t
->rcu_blocked_node
= rnp
;
211 * If this CPU has already checked in, then this task
212 * will hold up the next grace period rather than the
213 * current grace period. Queue the task accordingly.
214 * If the task is queued for the current grace period
215 * (i.e., this CPU has not yet passed through a quiescent
216 * state for the current grace period), then as long
217 * as that task remains queued, the current grace period
218 * cannot end. Note that there is some uncertainty as
219 * to exactly when the current grace period started.
220 * We take a conservative approach, which can result
221 * in unnecessarily waiting on tasks that started very
222 * slightly after the current grace period began. C'est
225 * But first, note that the current CPU must still be
228 WARN_ON_ONCE((rdp
->grpmask
& rnp
->qsmaskinit
) == 0);
229 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
230 if ((rnp
->qsmask
& rdp
->grpmask
) && rnp
->gp_tasks
!= NULL
) {
231 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
->prev
);
232 rnp
->gp_tasks
= &t
->rcu_node_entry
;
233 #ifdef CONFIG_RCU_BOOST
234 if (rnp
->boost_tasks
!= NULL
)
235 rnp
->boost_tasks
= rnp
->gp_tasks
;
236 #endif /* #ifdef CONFIG_RCU_BOOST */
238 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
239 if (rnp
->qsmask
& rdp
->grpmask
)
240 rnp
->gp_tasks
= &t
->rcu_node_entry
;
242 trace_rcu_preempt_task(rdp
->rsp
->name
,
244 (rnp
->qsmask
& rdp
->grpmask
)
247 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
248 } else if (t
->rcu_read_lock_nesting
< 0 &&
249 t
->rcu_read_unlock_special
) {
252 * Complete exit from RCU read-side critical section on
253 * behalf of preempted instance of __rcu_read_unlock().
255 rcu_read_unlock_special(t
);
259 * Either we were not in an RCU read-side critical section to
260 * begin with, or we have now recorded that critical section
261 * globally. Either way, we can now note a quiescent state
262 * for this CPU. Again, if we were in an RCU read-side critical
263 * section, and if that critical section was blocking the current
264 * grace period, then the fact that the task has been enqueued
265 * means that we continue to block the current grace period.
267 local_irq_save(flags
);
269 local_irq_restore(flags
);
273 * Check for preempted RCU readers blocking the current grace period
274 * for the specified rcu_node structure. If the caller needs a reliable
275 * answer, it must hold the rcu_node's ->lock.
277 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
279 return rnp
->gp_tasks
!= NULL
;
283 * Record a quiescent state for all tasks that were previously queued
284 * on the specified rcu_node structure and that were blocking the current
285 * RCU grace period. The caller must hold the specified rnp->lock with
286 * irqs disabled, and this lock is released upon return, but irqs remain
289 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
290 __releases(rnp
->lock
)
293 struct rcu_node
*rnp_p
;
295 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
296 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
297 return; /* Still need more quiescent states! */
303 * Either there is only one rcu_node in the tree,
304 * or tasks were kicked up to root rcu_node due to
305 * CPUs going offline.
307 rcu_report_qs_rsp(&rcu_preempt_state
, flags
);
311 /* Report up the rest of the hierarchy. */
313 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
314 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
315 rcu_report_qs_rnp(mask
, &rcu_preempt_state
, rnp_p
, flags
);
319 * Advance a ->blkd_tasks-list pointer to the next entry, instead
320 * returning NULL if at the end of the list.
322 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
323 struct rcu_node
*rnp
)
325 struct list_head
*np
;
327 np
= t
->rcu_node_entry
.next
;
328 if (np
== &rnp
->blkd_tasks
)
334 * Handle special cases during rcu_read_unlock(), such as needing to
335 * notify RCU core processing or task having blocked during the RCU
336 * read-side critical section.
338 void rcu_read_unlock_special(struct task_struct
*t
)
344 struct list_head
*np
;
345 #ifdef CONFIG_RCU_BOOST
346 struct rt_mutex
*rbmp
= NULL
;
347 #endif /* #ifdef CONFIG_RCU_BOOST */
348 struct rcu_node
*rnp
;
351 /* NMI handlers cannot block and cannot safely manipulate state. */
355 local_irq_save(flags
);
358 * If RCU core is waiting for this CPU to exit critical section,
359 * let it know that we have done so.
361 special
= t
->rcu_read_unlock_special
;
362 if (special
& RCU_READ_UNLOCK_NEED_QS
) {
363 rcu_preempt_qs(smp_processor_id());
366 /* Hardware IRQ handlers cannot block. */
367 if (in_irq() || in_serving_softirq()) {
368 local_irq_restore(flags
);
372 /* Clean up if blocked during RCU read-side critical section. */
373 if (special
& RCU_READ_UNLOCK_BLOCKED
) {
374 t
->rcu_read_unlock_special
&= ~RCU_READ_UNLOCK_BLOCKED
;
377 * Remove this task from the list it blocked on. The
378 * task can migrate while we acquire the lock, but at
379 * most one time. So at most two passes through loop.
382 rnp
= t
->rcu_blocked_node
;
383 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
384 if (rnp
== t
->rcu_blocked_node
)
386 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
388 empty
= !rcu_preempt_blocked_readers_cgp(rnp
);
389 empty_exp
= !rcu_preempted_readers_exp(rnp
);
390 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
391 np
= rcu_next_node_entry(t
, rnp
);
392 list_del_init(&t
->rcu_node_entry
);
393 t
->rcu_blocked_node
= NULL
;
394 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
396 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
398 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
400 #ifdef CONFIG_RCU_BOOST
401 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
402 rnp
->boost_tasks
= np
;
403 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
404 if (t
->rcu_boost_mutex
) {
405 rbmp
= t
->rcu_boost_mutex
;
406 t
->rcu_boost_mutex
= NULL
;
408 #endif /* #ifdef CONFIG_RCU_BOOST */
411 * If this was the last task on the current list, and if
412 * we aren't waiting on any CPUs, report the quiescent state.
413 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
414 * so we must take a snapshot of the expedited state.
416 empty_exp_now
= !rcu_preempted_readers_exp(rnp
);
417 if (!empty
&& !rcu_preempt_blocked_readers_cgp(rnp
)) {
418 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
425 rcu_report_unblock_qs_rnp(rnp
, flags
);
427 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
430 #ifdef CONFIG_RCU_BOOST
431 /* Unboost if we were boosted. */
433 rt_mutex_unlock(rbmp
);
434 #endif /* #ifdef CONFIG_RCU_BOOST */
437 * If this was the last task on the expedited lists,
438 * then we need to report up the rcu_node hierarchy.
440 if (!empty_exp
&& empty_exp_now
)
441 rcu_report_exp_rnp(&rcu_preempt_state
, rnp
, true);
443 local_irq_restore(flags
);
447 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
450 * Dump detailed information for all tasks blocking the current RCU
451 * grace period on the specified rcu_node structure.
453 static void rcu_print_detail_task_stall_rnp(struct rcu_node
*rnp
)
456 struct task_struct
*t
;
458 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
459 if (!rcu_preempt_blocked_readers_cgp(rnp
)) {
460 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
463 t
= list_entry(rnp
->gp_tasks
,
464 struct task_struct
, rcu_node_entry
);
465 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
)
467 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
471 * Dump detailed information for all tasks blocking the current RCU
474 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
476 struct rcu_node
*rnp
= rcu_get_root(rsp
);
478 rcu_print_detail_task_stall_rnp(rnp
);
479 rcu_for_each_leaf_node(rsp
, rnp
)
480 rcu_print_detail_task_stall_rnp(rnp
);
483 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
485 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
489 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
491 #ifdef CONFIG_RCU_CPU_STALL_INFO
493 static void rcu_print_task_stall_begin(struct rcu_node
*rnp
)
495 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
496 rnp
->level
, rnp
->grplo
, rnp
->grphi
);
499 static void rcu_print_task_stall_end(void)
504 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
506 static void rcu_print_task_stall_begin(struct rcu_node
*rnp
)
510 static void rcu_print_task_stall_end(void)
514 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
517 * Scan the current list of tasks blocked within RCU read-side critical
518 * sections, printing out the tid of each.
520 static int rcu_print_task_stall(struct rcu_node
*rnp
)
522 struct task_struct
*t
;
525 if (!rcu_preempt_blocked_readers_cgp(rnp
))
527 rcu_print_task_stall_begin(rnp
);
528 t
= list_entry(rnp
->gp_tasks
,
529 struct task_struct
, rcu_node_entry
);
530 list_for_each_entry_continue(t
, &rnp
->blkd_tasks
, rcu_node_entry
) {
531 pr_cont(" P%d", t
->pid
);
534 rcu_print_task_stall_end();
539 * Check that the list of blocked tasks for the newly completed grace
540 * period is in fact empty. It is a serious bug to complete a grace
541 * period that still has RCU readers blocked! This function must be
542 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
543 * must be held by the caller.
545 * Also, if there are blocked tasks on the list, they automatically
546 * block the newly created grace period, so set up ->gp_tasks accordingly.
548 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
550 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
551 if (!list_empty(&rnp
->blkd_tasks
))
552 rnp
->gp_tasks
= rnp
->blkd_tasks
.next
;
553 WARN_ON_ONCE(rnp
->qsmask
);
556 #ifdef CONFIG_HOTPLUG_CPU
559 * Handle tasklist migration for case in which all CPUs covered by the
560 * specified rcu_node have gone offline. Move them up to the root
561 * rcu_node. The reason for not just moving them to the immediate
562 * parent is to remove the need for rcu_read_unlock_special() to
563 * make more than two attempts to acquire the target rcu_node's lock.
564 * Returns true if there were tasks blocking the current RCU grace
567 * Returns 1 if there was previously a task blocking the current grace
568 * period on the specified rcu_node structure.
570 * The caller must hold rnp->lock with irqs disabled.
572 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
573 struct rcu_node
*rnp
,
574 struct rcu_data
*rdp
)
576 struct list_head
*lp
;
577 struct list_head
*lp_root
;
579 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
580 struct task_struct
*t
;
582 if (rnp
== rnp_root
) {
583 WARN_ONCE(1, "Last CPU thought to be offlined?");
584 return 0; /* Shouldn't happen: at least one CPU online. */
587 /* If we are on an internal node, complain bitterly. */
588 WARN_ON_ONCE(rnp
!= rdp
->mynode
);
591 * Move tasks up to root rcu_node. Don't try to get fancy for
592 * this corner-case operation -- just put this node's tasks
593 * at the head of the root node's list, and update the root node's
594 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
595 * if non-NULL. This might result in waiting for more tasks than
596 * absolutely necessary, but this is a good performance/complexity
599 if (rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->qsmask
== 0)
600 retval
|= RCU_OFL_TASKS_NORM_GP
;
601 if (rcu_preempted_readers_exp(rnp
))
602 retval
|= RCU_OFL_TASKS_EXP_GP
;
603 lp
= &rnp
->blkd_tasks
;
604 lp_root
= &rnp_root
->blkd_tasks
;
605 while (!list_empty(lp
)) {
606 t
= list_entry(lp
->next
, typeof(*t
), rcu_node_entry
);
607 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
608 list_del(&t
->rcu_node_entry
);
609 t
->rcu_blocked_node
= rnp_root
;
610 list_add(&t
->rcu_node_entry
, lp_root
);
611 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
612 rnp_root
->gp_tasks
= rnp
->gp_tasks
;
613 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
614 rnp_root
->exp_tasks
= rnp
->exp_tasks
;
615 #ifdef CONFIG_RCU_BOOST
616 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
617 rnp_root
->boost_tasks
= rnp
->boost_tasks
;
618 #endif /* #ifdef CONFIG_RCU_BOOST */
619 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
622 rnp
->gp_tasks
= NULL
;
623 rnp
->exp_tasks
= NULL
;
624 #ifdef CONFIG_RCU_BOOST
625 rnp
->boost_tasks
= NULL
;
627 * In case root is being boosted and leaf was not. Make sure
628 * that we boost the tasks blocking the current grace period
631 raw_spin_lock(&rnp_root
->lock
); /* irqs already disabled */
632 if (rnp_root
->boost_tasks
!= NULL
&&
633 rnp_root
->boost_tasks
!= rnp_root
->gp_tasks
&&
634 rnp_root
->boost_tasks
!= rnp_root
->exp_tasks
)
635 rnp_root
->boost_tasks
= rnp_root
->gp_tasks
;
636 raw_spin_unlock(&rnp_root
->lock
); /* irqs still disabled */
637 #endif /* #ifdef CONFIG_RCU_BOOST */
642 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
645 * Check for a quiescent state from the current CPU. When a task blocks,
646 * the task is recorded in the corresponding CPU's rcu_node structure,
647 * which is checked elsewhere.
649 * Caller must disable hard irqs.
651 static void rcu_preempt_check_callbacks(int cpu
)
653 struct task_struct
*t
= current
;
655 if (t
->rcu_read_lock_nesting
== 0) {
659 if (t
->rcu_read_lock_nesting
> 0 &&
660 per_cpu(rcu_preempt_data
, cpu
).qs_pending
)
661 t
->rcu_read_unlock_special
|= RCU_READ_UNLOCK_NEED_QS
;
664 #ifdef CONFIG_RCU_BOOST
666 static void rcu_preempt_do_callbacks(void)
668 rcu_do_batch(&rcu_preempt_state
, this_cpu_ptr(&rcu_preempt_data
));
671 #endif /* #ifdef CONFIG_RCU_BOOST */
674 * Queue a preemptible-RCU callback for invocation after a grace period.
676 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
678 __call_rcu(head
, func
, &rcu_preempt_state
, -1, 0);
680 EXPORT_SYMBOL_GPL(call_rcu
);
683 * Queue an RCU callback for lazy invocation after a grace period.
684 * This will likely be later named something like "call_rcu_lazy()",
685 * but this change will require some way of tagging the lazy RCU
686 * callbacks in the list of pending callbacks. Until then, this
687 * function may only be called from __kfree_rcu().
689 void kfree_call_rcu(struct rcu_head
*head
,
690 void (*func
)(struct rcu_head
*rcu
))
692 __call_rcu(head
, func
, &rcu_preempt_state
, -1, 1);
694 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
697 * synchronize_rcu - wait until a grace period has elapsed.
699 * Control will return to the caller some time after a full grace
700 * period has elapsed, in other words after all currently executing RCU
701 * read-side critical sections have completed. Note, however, that
702 * upon return from synchronize_rcu(), the caller might well be executing
703 * concurrently with new RCU read-side critical sections that began while
704 * synchronize_rcu() was waiting. RCU read-side critical sections are
705 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
707 * See the description of synchronize_sched() for more detailed information
708 * on memory ordering guarantees.
710 void synchronize_rcu(void)
712 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
713 !lock_is_held(&rcu_lock_map
) &&
714 !lock_is_held(&rcu_sched_lock_map
),
715 "Illegal synchronize_rcu() in RCU read-side critical section");
716 if (!rcu_scheduler_active
)
719 synchronize_rcu_expedited();
721 wait_rcu_gp(call_rcu
);
723 EXPORT_SYMBOL_GPL(synchronize_rcu
);
725 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq
);
726 static unsigned long sync_rcu_preempt_exp_count
;
727 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex
);
730 * Return non-zero if there are any tasks in RCU read-side critical
731 * sections blocking the current preemptible-RCU expedited grace period.
732 * If there is no preemptible-RCU expedited grace period currently in
733 * progress, returns zero unconditionally.
735 static int rcu_preempted_readers_exp(struct rcu_node
*rnp
)
737 return rnp
->exp_tasks
!= NULL
;
741 * return non-zero if there is no RCU expedited grace period in progress
742 * for the specified rcu_node structure, in other words, if all CPUs and
743 * tasks covered by the specified rcu_node structure have done their bit
744 * for the current expedited grace period. Works only for preemptible
745 * RCU -- other RCU implementation use other means.
747 * Caller must hold sync_rcu_preempt_exp_mutex.
749 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
751 return !rcu_preempted_readers_exp(rnp
) &&
752 ACCESS_ONCE(rnp
->expmask
) == 0;
756 * Report the exit from RCU read-side critical section for the last task
757 * that queued itself during or before the current expedited preemptible-RCU
758 * grace period. This event is reported either to the rcu_node structure on
759 * which the task was queued or to one of that rcu_node structure's ancestors,
760 * recursively up the tree. (Calm down, calm down, we do the recursion
763 * Most callers will set the "wake" flag, but the task initiating the
764 * expedited grace period need not wake itself.
766 * Caller must hold sync_rcu_preempt_exp_mutex.
768 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
774 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
776 if (!sync_rcu_preempt_exp_done(rnp
)) {
777 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
780 if (rnp
->parent
== NULL
) {
781 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
783 wake_up(&sync_rcu_preempt_exp_wq
);
787 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
789 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
790 rnp
->expmask
&= ~mask
;
795 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
796 * grace period for the specified rcu_node structure. If there are no such
797 * tasks, report it up the rcu_node hierarchy.
799 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
800 * CPU hotplug operations.
803 sync_rcu_preempt_exp_init(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
808 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
809 if (list_empty(&rnp
->blkd_tasks
)) {
810 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
812 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
813 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
817 rcu_report_exp_rnp(rsp
, rnp
, false); /* Don't wake self. */
821 * synchronize_rcu_expedited - Brute-force RCU grace period
823 * Wait for an RCU-preempt grace period, but expedite it. The basic
824 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
825 * the ->blkd_tasks lists and wait for this list to drain. This consumes
826 * significant time on all CPUs and is unfriendly to real-time workloads,
827 * so is thus not recommended for any sort of common-case code.
828 * In fact, if you are using synchronize_rcu_expedited() in a loop,
829 * please restructure your code to batch your updates, and then Use a
830 * single synchronize_rcu() instead.
832 * Note that it is illegal to call this function while holding any lock
833 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
834 * to call this function from a CPU-hotplug notifier. Failing to observe
835 * these restriction will result in deadlock.
837 void synchronize_rcu_expedited(void)
840 struct rcu_node
*rnp
;
841 struct rcu_state
*rsp
= &rcu_preempt_state
;
845 smp_mb(); /* Caller's modifications seen first by other CPUs. */
846 snap
= ACCESS_ONCE(sync_rcu_preempt_exp_count
) + 1;
847 smp_mb(); /* Above access cannot bleed into critical section. */
850 * Block CPU-hotplug operations. This means that any CPU-hotplug
851 * operation that finds an rcu_node structure with tasks in the
852 * process of being boosted will know that all tasks blocking
853 * this expedited grace period will already be in the process of
854 * being boosted. This simplifies the process of moving tasks
855 * from leaf to root rcu_node structures.
860 * Acquire lock, falling back to synchronize_rcu() if too many
861 * lock-acquisition failures. Of course, if someone does the
862 * expedited grace period for us, just leave.
864 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex
)) {
865 if (ULONG_CMP_LT(snap
,
866 ACCESS_ONCE(sync_rcu_preempt_exp_count
))) {
868 goto mb_ret
; /* Others did our work for us. */
870 if (trycount
++ < 10) {
871 udelay(trycount
* num_online_cpus());
874 wait_rcu_gp(call_rcu
);
878 if (ULONG_CMP_LT(snap
, ACCESS_ONCE(sync_rcu_preempt_exp_count
))) {
880 goto unlock_mb_ret
; /* Others did our work for us. */
883 /* force all RCU readers onto ->blkd_tasks lists. */
884 synchronize_sched_expedited();
886 /* Initialize ->expmask for all non-leaf rcu_node structures. */
887 rcu_for_each_nonleaf_node_breadth_first(rsp
, rnp
) {
888 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
889 rnp
->expmask
= rnp
->qsmaskinit
;
890 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
893 /* Snapshot current state of ->blkd_tasks lists. */
894 rcu_for_each_leaf_node(rsp
, rnp
)
895 sync_rcu_preempt_exp_init(rsp
, rnp
);
896 if (NUM_RCU_NODES
> 1)
897 sync_rcu_preempt_exp_init(rsp
, rcu_get_root(rsp
));
901 /* Wait for snapshotted ->blkd_tasks lists to drain. */
902 rnp
= rcu_get_root(rsp
);
903 wait_event(sync_rcu_preempt_exp_wq
,
904 sync_rcu_preempt_exp_done(rnp
));
906 /* Clean up and exit. */
907 smp_mb(); /* ensure expedited GP seen before counter increment. */
908 ACCESS_ONCE(sync_rcu_preempt_exp_count
)++;
910 mutex_unlock(&sync_rcu_preempt_exp_mutex
);
912 smp_mb(); /* ensure subsequent action seen after grace period. */
914 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
917 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
919 * Note that this primitive does not necessarily wait for an RCU grace period
920 * to complete. For example, if there are no RCU callbacks queued anywhere
921 * in the system, then rcu_barrier() is within its rights to return
922 * immediately, without waiting for anything, much less an RCU grace period.
924 void rcu_barrier(void)
926 _rcu_barrier(&rcu_preempt_state
);
928 EXPORT_SYMBOL_GPL(rcu_barrier
);
931 * Initialize preemptible RCU's state structures.
933 static void __init
__rcu_init_preempt(void)
935 rcu_init_one(&rcu_preempt_state
, &rcu_preempt_data
);
939 * Check for a task exiting while in a preemptible-RCU read-side
940 * critical section, clean up if so. No need to issue warnings,
941 * as debug_check_no_locks_held() already does this if lockdep
946 struct task_struct
*t
= current
;
948 if (likely(list_empty(¤t
->rcu_node_entry
)))
950 t
->rcu_read_lock_nesting
= 1;
952 t
->rcu_read_unlock_special
= RCU_READ_UNLOCK_BLOCKED
;
956 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
958 static struct rcu_state
*rcu_state
= &rcu_sched_state
;
961 * Tell them what RCU they are running.
963 static void __init
rcu_bootup_announce(void)
965 pr_info("Hierarchical RCU implementation.\n");
966 rcu_bootup_announce_oddness();
970 * Return the number of RCU batches processed thus far for debug & stats.
972 long rcu_batches_completed(void)
974 return rcu_batches_completed_sched();
976 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
979 * Force a quiescent state for RCU, which, because there is no preemptible
980 * RCU, becomes the same as rcu-sched.
982 void rcu_force_quiescent_state(void)
984 rcu_sched_force_quiescent_state();
986 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
989 * Because preemptible RCU does not exist, we never have to check for
990 * CPUs being in quiescent states.
992 static void rcu_preempt_note_context_switch(int cpu
)
997 * Because preemptible RCU does not exist, there are never any preempted
1000 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
1005 #ifdef CONFIG_HOTPLUG_CPU
1007 /* Because preemptible RCU does not exist, no quieting of tasks. */
1008 static void rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
1010 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1013 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1016 * Because preemptible RCU does not exist, we never have to check for
1017 * tasks blocked within RCU read-side critical sections.
1019 static void rcu_print_detail_task_stall(struct rcu_state
*rsp
)
1024 * Because preemptible RCU does not exist, we never have to check for
1025 * tasks blocked within RCU read-side critical sections.
1027 static int rcu_print_task_stall(struct rcu_node
*rnp
)
1033 * Because there is no preemptible RCU, there can be no readers blocked,
1034 * so there is no need to check for blocked tasks. So check only for
1035 * bogus qsmask values.
1037 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
1039 WARN_ON_ONCE(rnp
->qsmask
);
1042 #ifdef CONFIG_HOTPLUG_CPU
1045 * Because preemptible RCU does not exist, it never needs to migrate
1046 * tasks that were blocked within RCU read-side critical sections, and
1047 * such non-existent tasks cannot possibly have been blocking the current
1050 static int rcu_preempt_offline_tasks(struct rcu_state
*rsp
,
1051 struct rcu_node
*rnp
,
1052 struct rcu_data
*rdp
)
1057 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1060 * Because preemptible RCU does not exist, it never has any callbacks
1063 static void rcu_preempt_check_callbacks(int cpu
)
1068 * Queue an RCU callback for lazy invocation after a grace period.
1069 * This will likely be later named something like "call_rcu_lazy()",
1070 * but this change will require some way of tagging the lazy RCU
1071 * callbacks in the list of pending callbacks. Until then, this
1072 * function may only be called from __kfree_rcu().
1074 * Because there is no preemptible RCU, we use RCU-sched instead.
1076 void kfree_call_rcu(struct rcu_head
*head
,
1077 void (*func
)(struct rcu_head
*rcu
))
1079 __call_rcu(head
, func
, &rcu_sched_state
, -1, 1);
1081 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
1084 * Wait for an rcu-preempt grace period, but make it happen quickly.
1085 * But because preemptible RCU does not exist, map to rcu-sched.
1087 void synchronize_rcu_expedited(void)
1089 synchronize_sched_expedited();
1091 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited
);
1093 #ifdef CONFIG_HOTPLUG_CPU
1096 * Because preemptible RCU does not exist, there is never any need to
1097 * report on tasks preempted in RCU read-side critical sections during
1098 * expedited RCU grace periods.
1100 static void rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1105 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1108 * Because preemptible RCU does not exist, rcu_barrier() is just
1109 * another name for rcu_barrier_sched().
1111 void rcu_barrier(void)
1113 rcu_barrier_sched();
1115 EXPORT_SYMBOL_GPL(rcu_barrier
);
1118 * Because preemptible RCU does not exist, it need not be initialized.
1120 static void __init
__rcu_init_preempt(void)
1125 * Because preemptible RCU does not exist, tasks cannot possibly exit
1126 * while in preemptible RCU read-side critical sections.
1132 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1134 #ifdef CONFIG_RCU_BOOST
1136 #include "../locking/rtmutex_common.h"
1138 #ifdef CONFIG_RCU_TRACE
1140 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1142 if (list_empty(&rnp
->blkd_tasks
))
1143 rnp
->n_balk_blkd_tasks
++;
1144 else if (rnp
->exp_tasks
== NULL
&& rnp
->gp_tasks
== NULL
)
1145 rnp
->n_balk_exp_gp_tasks
++;
1146 else if (rnp
->gp_tasks
!= NULL
&& rnp
->boost_tasks
!= NULL
)
1147 rnp
->n_balk_boost_tasks
++;
1148 else if (rnp
->gp_tasks
!= NULL
&& rnp
->qsmask
!= 0)
1149 rnp
->n_balk_notblocked
++;
1150 else if (rnp
->gp_tasks
!= NULL
&&
1151 ULONG_CMP_LT(jiffies
, rnp
->boost_time
))
1152 rnp
->n_balk_notyet
++;
1157 #else /* #ifdef CONFIG_RCU_TRACE */
1159 static void rcu_initiate_boost_trace(struct rcu_node
*rnp
)
1163 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1165 static void rcu_wake_cond(struct task_struct
*t
, int status
)
1168 * If the thread is yielding, only wake it when this
1169 * is invoked from idle
1171 if (status
!= RCU_KTHREAD_YIELDING
|| is_idle_task(current
))
1176 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1177 * or ->boost_tasks, advancing the pointer to the next task in the
1178 * ->blkd_tasks list.
1180 * Note that irqs must be enabled: boosting the task can block.
1181 * Returns 1 if there are more tasks needing to be boosted.
1183 static int rcu_boost(struct rcu_node
*rnp
)
1185 unsigned long flags
;
1186 struct rt_mutex mtx
;
1187 struct task_struct
*t
;
1188 struct list_head
*tb
;
1190 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
)
1191 return 0; /* Nothing left to boost. */
1193 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1196 * Recheck under the lock: all tasks in need of boosting
1197 * might exit their RCU read-side critical sections on their own.
1199 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
1200 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1205 * Preferentially boost tasks blocking expedited grace periods.
1206 * This cannot starve the normal grace periods because a second
1207 * expedited grace period must boost all blocked tasks, including
1208 * those blocking the pre-existing normal grace period.
1210 if (rnp
->exp_tasks
!= NULL
) {
1211 tb
= rnp
->exp_tasks
;
1212 rnp
->n_exp_boosts
++;
1214 tb
= rnp
->boost_tasks
;
1215 rnp
->n_normal_boosts
++;
1217 rnp
->n_tasks_boosted
++;
1220 * We boost task t by manufacturing an rt_mutex that appears to
1221 * be held by task t. We leave a pointer to that rt_mutex where
1222 * task t can find it, and task t will release the mutex when it
1223 * exits its outermost RCU read-side critical section. Then
1224 * simply acquiring this artificial rt_mutex will boost task
1225 * t's priority. (Thanks to tglx for suggesting this approach!)
1227 * Note that task t must acquire rnp->lock to remove itself from
1228 * the ->blkd_tasks list, which it will do from exit() if from
1229 * nowhere else. We therefore are guaranteed that task t will
1230 * stay around at least until we drop rnp->lock. Note that
1231 * rnp->lock also resolves races between our priority boosting
1232 * and task t's exiting its outermost RCU read-side critical
1235 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1236 rt_mutex_init_proxy_locked(&mtx
, t
);
1237 t
->rcu_boost_mutex
= &mtx
;
1238 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1239 rt_mutex_lock(&mtx
); /* Side effect: boosts task t's priority. */
1240 rt_mutex_unlock(&mtx
); /* Keep lockdep happy. */
1242 return ACCESS_ONCE(rnp
->exp_tasks
) != NULL
||
1243 ACCESS_ONCE(rnp
->boost_tasks
) != NULL
;
1247 * Priority-boosting kthread. One per leaf rcu_node and one for the
1250 static int rcu_boost_kthread(void *arg
)
1252 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1256 trace_rcu_utilization(TPS("Start boost kthread@init"));
1258 rnp
->boost_kthread_status
= RCU_KTHREAD_WAITING
;
1259 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1260 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1261 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1262 rnp
->boost_kthread_status
= RCU_KTHREAD_RUNNING
;
1263 more2boost
= rcu_boost(rnp
);
1269 rnp
->boost_kthread_status
= RCU_KTHREAD_YIELDING
;
1270 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1271 schedule_timeout_interruptible(2);
1272 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1277 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1282 * Check to see if it is time to start boosting RCU readers that are
1283 * blocking the current grace period, and, if so, tell the per-rcu_node
1284 * kthread to start boosting them. If there is an expedited grace
1285 * period in progress, it is always time to boost.
1287 * The caller must hold rnp->lock, which this function releases.
1288 * The ->boost_kthread_task is immortal, so we don't need to worry
1289 * about it going away.
1291 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1293 struct task_struct
*t
;
1295 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1296 rnp
->n_balk_exp_gp_tasks
++;
1297 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1300 if (rnp
->exp_tasks
!= NULL
||
1301 (rnp
->gp_tasks
!= NULL
&&
1302 rnp
->boost_tasks
== NULL
&&
1304 ULONG_CMP_GE(jiffies
, rnp
->boost_time
))) {
1305 if (rnp
->exp_tasks
== NULL
)
1306 rnp
->boost_tasks
= rnp
->gp_tasks
;
1307 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1308 t
= rnp
->boost_kthread_task
;
1310 rcu_wake_cond(t
, rnp
->boost_kthread_status
);
1312 rcu_initiate_boost_trace(rnp
);
1313 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1318 * Wake up the per-CPU kthread to invoke RCU callbacks.
1320 static void invoke_rcu_callbacks_kthread(void)
1322 unsigned long flags
;
1324 local_irq_save(flags
);
1325 __this_cpu_write(rcu_cpu_has_work
, 1);
1326 if (__this_cpu_read(rcu_cpu_kthread_task
) != NULL
&&
1327 current
!= __this_cpu_read(rcu_cpu_kthread_task
)) {
1328 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task
),
1329 __this_cpu_read(rcu_cpu_kthread_status
));
1331 local_irq_restore(flags
);
1335 * Is the current CPU running the RCU-callbacks kthread?
1336 * Caller must have preemption disabled.
1338 static bool rcu_is_callbacks_kthread(void)
1340 return __this_cpu_read(rcu_cpu_kthread_task
) == current
;
1343 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1346 * Do priority-boost accounting for the start of a new grace period.
1348 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1350 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1354 * Create an RCU-boost kthread for the specified node if one does not
1355 * already exist. We only create this kthread for preemptible RCU.
1356 * Returns zero if all is well, a negated errno otherwise.
1358 static int rcu_spawn_one_boost_kthread(struct rcu_state
*rsp
,
1359 struct rcu_node
*rnp
)
1361 int rnp_index
= rnp
- &rsp
->node
[0];
1362 unsigned long flags
;
1363 struct sched_param sp
;
1364 struct task_struct
*t
;
1366 if (&rcu_preempt_state
!= rsp
)
1369 if (!rcu_scheduler_fully_active
|| rnp
->qsmaskinit
== 0)
1373 if (rnp
->boost_kthread_task
!= NULL
)
1375 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1376 "rcub/%d", rnp_index
);
1379 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1380 rnp
->boost_kthread_task
= t
;
1381 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1382 sp
.sched_priority
= RCU_BOOST_PRIO
;
1383 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1384 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1388 static void rcu_kthread_do_work(void)
1390 rcu_do_batch(&rcu_sched_state
, this_cpu_ptr(&rcu_sched_data
));
1391 rcu_do_batch(&rcu_bh_state
, this_cpu_ptr(&rcu_bh_data
));
1392 rcu_preempt_do_callbacks();
1395 static void rcu_cpu_kthread_setup(unsigned int cpu
)
1397 struct sched_param sp
;
1399 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1400 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1403 static void rcu_cpu_kthread_park(unsigned int cpu
)
1405 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1408 static int rcu_cpu_kthread_should_run(unsigned int cpu
)
1410 return __this_cpu_read(rcu_cpu_has_work
);
1414 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1415 * RCU softirq used in flavors and configurations of RCU that do not
1416 * support RCU priority boosting.
1418 static void rcu_cpu_kthread(unsigned int cpu
)
1420 unsigned int *statusp
= this_cpu_ptr(&rcu_cpu_kthread_status
);
1421 char work
, *workp
= this_cpu_ptr(&rcu_cpu_has_work
);
1424 for (spincnt
= 0; spincnt
< 10; spincnt
++) {
1425 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1427 *statusp
= RCU_KTHREAD_RUNNING
;
1428 this_cpu_inc(rcu_cpu_kthread_loops
);
1429 local_irq_disable();
1434 rcu_kthread_do_work();
1437 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1438 *statusp
= RCU_KTHREAD_WAITING
;
1442 *statusp
= RCU_KTHREAD_YIELDING
;
1443 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1444 schedule_timeout_interruptible(2);
1445 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1446 *statusp
= RCU_KTHREAD_WAITING
;
1450 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1451 * served by the rcu_node in question. The CPU hotplug lock is still
1452 * held, so the value of rnp->qsmaskinit will be stable.
1454 * We don't include outgoingcpu in the affinity set, use -1 if there is
1455 * no outgoing CPU. If there are no CPUs left in the affinity set,
1456 * this function allows the kthread to execute on any CPU.
1458 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1460 struct task_struct
*t
= rnp
->boost_kthread_task
;
1461 unsigned long mask
= rnp
->qsmaskinit
;
1467 if (!zalloc_cpumask_var(&cm
, GFP_KERNEL
))
1469 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1470 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1471 cpumask_set_cpu(cpu
, cm
);
1472 if (cpumask_weight(cm
) == 0) {
1474 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1475 cpumask_clear_cpu(cpu
, cm
);
1476 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1478 set_cpus_allowed_ptr(t
, cm
);
1479 free_cpumask_var(cm
);
1482 static struct smp_hotplug_thread rcu_cpu_thread_spec
= {
1483 .store
= &rcu_cpu_kthread_task
,
1484 .thread_should_run
= rcu_cpu_kthread_should_run
,
1485 .thread_fn
= rcu_cpu_kthread
,
1486 .thread_comm
= "rcuc/%u",
1487 .setup
= rcu_cpu_kthread_setup
,
1488 .park
= rcu_cpu_kthread_park
,
1492 * Spawn all kthreads -- called as soon as the scheduler is running.
1494 static int __init
rcu_spawn_kthreads(void)
1496 struct rcu_node
*rnp
;
1499 rcu_scheduler_fully_active
= 1;
1500 for_each_possible_cpu(cpu
)
1501 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1502 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec
));
1503 rnp
= rcu_get_root(rcu_state
);
1504 (void)rcu_spawn_one_boost_kthread(rcu_state
, rnp
);
1505 if (NUM_RCU_NODES
> 1) {
1506 rcu_for_each_leaf_node(rcu_state
, rnp
)
1507 (void)rcu_spawn_one_boost_kthread(rcu_state
, rnp
);
1511 early_initcall(rcu_spawn_kthreads
);
1513 static void rcu_prepare_kthreads(int cpu
)
1515 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
1516 struct rcu_node
*rnp
= rdp
->mynode
;
1518 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1519 if (rcu_scheduler_fully_active
)
1520 (void)rcu_spawn_one_boost_kthread(rcu_state
, rnp
);
1523 #else /* #ifdef CONFIG_RCU_BOOST */
1525 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1527 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1530 static void invoke_rcu_callbacks_kthread(void)
1535 static bool rcu_is_callbacks_kthread(void)
1540 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1544 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1548 static int __init
rcu_scheduler_really_started(void)
1550 rcu_scheduler_fully_active
= 1;
1553 early_initcall(rcu_scheduler_really_started
);
1555 static void rcu_prepare_kthreads(int cpu
)
1559 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1561 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1564 * Check to see if any future RCU-related work will need to be done
1565 * by the current CPU, even if none need be done immediately, returning
1566 * 1 if so. This function is part of the RCU implementation; it is -not-
1567 * an exported member of the RCU API.
1569 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1570 * any flavor of RCU.
1572 int rcu_needs_cpu(int cpu
, unsigned long *delta_jiffies
)
1574 *delta_jiffies
= ULONG_MAX
;
1575 return rcu_cpu_has_callbacks(cpu
, NULL
);
1579 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1582 static void rcu_cleanup_after_idle(int cpu
)
1587 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1590 static void rcu_prepare_for_idle(int cpu
)
1595 * Don't bother keeping a running count of the number of RCU callbacks
1596 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1598 static void rcu_idle_count_callbacks_posted(void)
1602 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1605 * This code is invoked when a CPU goes idle, at which point we want
1606 * to have the CPU do everything required for RCU so that it can enter
1607 * the energy-efficient dyntick-idle mode. This is handled by a
1608 * state machine implemented by rcu_prepare_for_idle() below.
1610 * The following three proprocessor symbols control this state machine:
1612 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1613 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1614 * is sized to be roughly one RCU grace period. Those energy-efficiency
1615 * benchmarkers who might otherwise be tempted to set this to a large
1616 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1617 * system. And if you are -that- concerned about energy efficiency,
1618 * just power the system down and be done with it!
1619 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1620 * permitted to sleep in dyntick-idle mode with only lazy RCU
1621 * callbacks pending. Setting this too high can OOM your system.
1623 * The values below work well in practice. If future workloads require
1624 * adjustment, they can be converted into kernel config parameters, though
1625 * making the state machine smarter might be a better option.
1627 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1628 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1630 static int rcu_idle_gp_delay
= RCU_IDLE_GP_DELAY
;
1631 module_param(rcu_idle_gp_delay
, int, 0644);
1632 static int rcu_idle_lazy_gp_delay
= RCU_IDLE_LAZY_GP_DELAY
;
1633 module_param(rcu_idle_lazy_gp_delay
, int, 0644);
1635 extern int tick_nohz_enabled
;
1638 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1639 * only if it has been awhile since the last time we did so. Afterwards,
1640 * if there are any callbacks ready for immediate invocation, return true.
1642 static bool rcu_try_advance_all_cbs(void)
1644 bool cbs_ready
= false;
1645 struct rcu_data
*rdp
;
1646 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1647 struct rcu_node
*rnp
;
1648 struct rcu_state
*rsp
;
1650 /* Exit early if we advanced recently. */
1651 if (jiffies
== rdtp
->last_advance_all
)
1653 rdtp
->last_advance_all
= jiffies
;
1655 for_each_rcu_flavor(rsp
) {
1656 rdp
= this_cpu_ptr(rsp
->rda
);
1660 * Don't bother checking unless a grace period has
1661 * completed since we last checked and there are
1662 * callbacks not yet ready to invoke.
1664 if (rdp
->completed
!= rnp
->completed
&&
1665 rdp
->nxttail
[RCU_DONE_TAIL
] != rdp
->nxttail
[RCU_NEXT_TAIL
])
1666 note_gp_changes(rsp
, rdp
);
1668 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1675 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1676 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1677 * caller to set the timeout based on whether or not there are non-lazy
1680 * The caller must have disabled interrupts.
1682 int rcu_needs_cpu(int cpu
, unsigned long *dj
)
1684 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1686 /* Snapshot to detect later posting of non-lazy callback. */
1687 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1689 /* If no callbacks, RCU doesn't need the CPU. */
1690 if (!rcu_cpu_has_callbacks(cpu
, &rdtp
->all_lazy
)) {
1695 /* Attempt to advance callbacks. */
1696 if (rcu_try_advance_all_cbs()) {
1697 /* Some ready to invoke, so initiate later invocation. */
1701 rdtp
->last_accelerate
= jiffies
;
1703 /* Request timer delay depending on laziness, and round. */
1704 if (!rdtp
->all_lazy
) {
1705 *dj
= round_up(rcu_idle_gp_delay
+ jiffies
,
1706 rcu_idle_gp_delay
) - jiffies
;
1708 *dj
= round_jiffies(rcu_idle_lazy_gp_delay
+ jiffies
) - jiffies
;
1714 * Prepare a CPU for idle from an RCU perspective. The first major task
1715 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1716 * The second major task is to check to see if a non-lazy callback has
1717 * arrived at a CPU that previously had only lazy callbacks. The third
1718 * major task is to accelerate (that is, assign grace-period numbers to)
1719 * any recently arrived callbacks.
1721 * The caller must have disabled interrupts.
1723 static void rcu_prepare_for_idle(int cpu
)
1725 struct rcu_data
*rdp
;
1726 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1727 struct rcu_node
*rnp
;
1728 struct rcu_state
*rsp
;
1731 /* Handle nohz enablement switches conservatively. */
1732 tne
= ACCESS_ONCE(tick_nohz_enabled
);
1733 if (tne
!= rdtp
->tick_nohz_enabled_snap
) {
1734 if (rcu_cpu_has_callbacks(cpu
, NULL
))
1735 invoke_rcu_core(); /* force nohz to see update. */
1736 rdtp
->tick_nohz_enabled_snap
= tne
;
1742 /* If this is a no-CBs CPU, no callbacks, just return. */
1743 if (rcu_is_nocb_cpu(cpu
))
1747 * If a non-lazy callback arrived at a CPU having only lazy
1748 * callbacks, invoke RCU core for the side-effect of recalculating
1749 * idle duration on re-entry to idle.
1751 if (rdtp
->all_lazy
&&
1752 rdtp
->nonlazy_posted
!= rdtp
->nonlazy_posted_snap
) {
1753 rdtp
->all_lazy
= false;
1754 rdtp
->nonlazy_posted_snap
= rdtp
->nonlazy_posted
;
1760 * If we have not yet accelerated this jiffy, accelerate all
1761 * callbacks on this CPU.
1763 if (rdtp
->last_accelerate
== jiffies
)
1765 rdtp
->last_accelerate
= jiffies
;
1766 for_each_rcu_flavor(rsp
) {
1767 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1768 if (!*rdp
->nxttail
[RCU_DONE_TAIL
])
1771 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1772 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1773 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1778 * Clean up for exit from idle. Attempt to advance callbacks based on
1779 * any grace periods that elapsed while the CPU was idle, and if any
1780 * callbacks are now ready to invoke, initiate invocation.
1782 static void rcu_cleanup_after_idle(int cpu
)
1785 if (rcu_is_nocb_cpu(cpu
))
1787 if (rcu_try_advance_all_cbs())
1792 * Keep a running count of the number of non-lazy callbacks posted
1793 * on this CPU. This running counter (which is never decremented) allows
1794 * rcu_prepare_for_idle() to detect when something out of the idle loop
1795 * posts a callback, even if an equal number of callbacks are invoked.
1796 * Of course, callbacks should only be posted from within a trace event
1797 * designed to be called from idle or from within RCU_NONIDLE().
1799 static void rcu_idle_count_callbacks_posted(void)
1801 __this_cpu_add(rcu_dynticks
.nonlazy_posted
, 1);
1805 * Data for flushing lazy RCU callbacks at OOM time.
1807 static atomic_t oom_callback_count
;
1808 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq
);
1811 * RCU OOM callback -- decrement the outstanding count and deliver the
1812 * wake-up if we are the last one.
1814 static void rcu_oom_callback(struct rcu_head
*rhp
)
1816 if (atomic_dec_and_test(&oom_callback_count
))
1817 wake_up(&oom_callback_wq
);
1821 * Post an rcu_oom_notify callback on the current CPU if it has at
1822 * least one lazy callback. This will unnecessarily post callbacks
1823 * to CPUs that already have a non-lazy callback at the end of their
1824 * callback list, but this is an infrequent operation, so accept some
1825 * extra overhead to keep things simple.
1827 static void rcu_oom_notify_cpu(void *unused
)
1829 struct rcu_state
*rsp
;
1830 struct rcu_data
*rdp
;
1832 for_each_rcu_flavor(rsp
) {
1833 rdp
= __this_cpu_ptr(rsp
->rda
);
1834 if (rdp
->qlen_lazy
!= 0) {
1835 atomic_inc(&oom_callback_count
);
1836 rsp
->call(&rdp
->oom_head
, rcu_oom_callback
);
1842 * If low on memory, ensure that each CPU has a non-lazy callback.
1843 * This will wake up CPUs that have only lazy callbacks, in turn
1844 * ensuring that they free up the corresponding memory in a timely manner.
1845 * Because an uncertain amount of memory will be freed in some uncertain
1846 * timeframe, we do not claim to have freed anything.
1848 static int rcu_oom_notify(struct notifier_block
*self
,
1849 unsigned long notused
, void *nfreed
)
1853 /* Wait for callbacks from earlier instance to complete. */
1854 wait_event(oom_callback_wq
, atomic_read(&oom_callback_count
) == 0);
1857 * Prevent premature wakeup: ensure that all increments happen
1858 * before there is a chance of the counter reaching zero.
1860 atomic_set(&oom_callback_count
, 1);
1863 for_each_online_cpu(cpu
) {
1864 smp_call_function_single(cpu
, rcu_oom_notify_cpu
, NULL
, 1);
1869 /* Unconditionally decrement: no need to wake ourselves up. */
1870 atomic_dec(&oom_callback_count
);
1875 static struct notifier_block rcu_oom_nb
= {
1876 .notifier_call
= rcu_oom_notify
1879 static int __init
rcu_register_oom_notifier(void)
1881 register_oom_notifier(&rcu_oom_nb
);
1884 early_initcall(rcu_register_oom_notifier
);
1886 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1888 #ifdef CONFIG_RCU_CPU_STALL_INFO
1890 #ifdef CONFIG_RCU_FAST_NO_HZ
1892 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1894 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
1895 unsigned long nlpd
= rdtp
->nonlazy_posted
- rdtp
->nonlazy_posted_snap
;
1897 sprintf(cp
, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1898 rdtp
->last_accelerate
& 0xffff, jiffies
& 0xffff,
1900 rdtp
->all_lazy
? 'L' : '.',
1901 rdtp
->tick_nohz_enabled_snap
? '.' : 'D');
1904 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1906 static void print_cpu_stall_fast_no_hz(char *cp
, int cpu
)
1911 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1913 /* Initiate the stall-info list. */
1914 static void print_cpu_stall_info_begin(void)
1920 * Print out diagnostic information for the specified stalled CPU.
1922 * If the specified CPU is aware of the current RCU grace period
1923 * (flavor specified by rsp), then print the number of scheduling
1924 * clock interrupts the CPU has taken during the time that it has
1925 * been aware. Otherwise, print the number of RCU grace periods
1926 * that this CPU is ignorant of, for example, "1" if the CPU was
1927 * aware of the previous grace period.
1929 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1931 static void print_cpu_stall_info(struct rcu_state
*rsp
, int cpu
)
1933 char fast_no_hz
[72];
1934 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1935 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
1937 unsigned long ticks_value
;
1939 if (rsp
->gpnum
== rdp
->gpnum
) {
1940 ticks_title
= "ticks this GP";
1941 ticks_value
= rdp
->ticks_this_gp
;
1943 ticks_title
= "GPs behind";
1944 ticks_value
= rsp
->gpnum
- rdp
->gpnum
;
1946 print_cpu_stall_fast_no_hz(fast_no_hz
, cpu
);
1947 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1948 cpu
, ticks_value
, ticks_title
,
1949 atomic_read(&rdtp
->dynticks
) & 0xfff,
1950 rdtp
->dynticks_nesting
, rdtp
->dynticks_nmi_nesting
,
1951 rdp
->softirq_snap
, kstat_softirqs_cpu(RCU_SOFTIRQ
, cpu
),
1955 /* Terminate the stall-info list. */
1956 static void print_cpu_stall_info_end(void)
1961 /* Zero ->ticks_this_gp for all flavors of RCU. */
1962 static void zero_cpu_stall_ticks(struct rcu_data
*rdp
)
1964 rdp
->ticks_this_gp
= 0;
1965 rdp
->softirq_snap
= kstat_softirqs_cpu(RCU_SOFTIRQ
, smp_processor_id());
1968 /* Increment ->ticks_this_gp for all flavors of RCU. */
1969 static void increment_cpu_stall_ticks(void)
1971 struct rcu_state
*rsp
;
1973 for_each_rcu_flavor(rsp
)
1974 __this_cpu_ptr(rsp
->rda
)->ticks_this_gp
++;
1977 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1979 static void print_cpu_stall_info_begin(void)
1984 static void print_cpu_stall_info(struct rcu_state
*rsp
, int cpu
)
1986 pr_cont(" %d", cpu
);
1989 static void print_cpu_stall_info_end(void)
1994 static void zero_cpu_stall_ticks(struct rcu_data
*rdp
)
1998 static void increment_cpu_stall_ticks(void)
2002 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
2004 #ifdef CONFIG_RCU_NOCB_CPU
2007 * Offload callback processing from the boot-time-specified set of CPUs
2008 * specified by rcu_nocb_mask. For each CPU in the set, there is a
2009 * kthread created that pulls the callbacks from the corresponding CPU,
2010 * waits for a grace period to elapse, and invokes the callbacks.
2011 * The no-CBs CPUs do a wake_up() on their kthread when they insert
2012 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2013 * has been specified, in which case each kthread actively polls its
2014 * CPU. (Which isn't so great for energy efficiency, but which does
2015 * reduce RCU's overhead on that CPU.)
2017 * This is intended to be used in conjunction with Frederic Weisbecker's
2018 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2019 * running CPU-bound user-mode computations.
2021 * Offloading of callback processing could also in theory be used as
2022 * an energy-efficiency measure because CPUs with no RCU callbacks
2023 * queued are more aggressive about entering dyntick-idle mode.
2027 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2028 static int __init
rcu_nocb_setup(char *str
)
2030 alloc_bootmem_cpumask_var(&rcu_nocb_mask
);
2031 have_rcu_nocb_mask
= true;
2032 cpulist_parse(str
, rcu_nocb_mask
);
2035 __setup("rcu_nocbs=", rcu_nocb_setup
);
2037 static int __init
parse_rcu_nocb_poll(char *arg
)
2042 early_param("rcu_nocb_poll", parse_rcu_nocb_poll
);
2045 * Do any no-CBs CPUs need another grace period?
2047 * Interrupts must be disabled. If the caller does not hold the root
2048 * rnp_node structure's ->lock, the results are advisory only.
2050 static int rcu_nocb_needs_gp(struct rcu_state
*rsp
)
2052 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2054 return rnp
->need_future_gp
[(ACCESS_ONCE(rnp
->completed
) + 1) & 0x1];
2058 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2061 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
2063 wake_up_all(&rnp
->nocb_gp_wq
[rnp
->completed
& 0x1]);
2067 * Set the root rcu_node structure's ->need_future_gp field
2068 * based on the sum of those of all rcu_node structures. This does
2069 * double-count the root rcu_node structure's requests, but this
2070 * is necessary to handle the possibility of a rcu_nocb_kthread()
2071 * having awakened during the time that the rcu_node structures
2072 * were being updated for the end of the previous grace period.
2074 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
2076 rnp
->need_future_gp
[(rnp
->completed
+ 1) & 0x1] += nrq
;
2079 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2081 init_waitqueue_head(&rnp
->nocb_gp_wq
[0]);
2082 init_waitqueue_head(&rnp
->nocb_gp_wq
[1]);
2085 /* Is the specified CPU a no-CPUs CPU? */
2086 bool rcu_is_nocb_cpu(int cpu
)
2088 if (have_rcu_nocb_mask
)
2089 return cpumask_test_cpu(cpu
, rcu_nocb_mask
);
2094 * Enqueue the specified string of rcu_head structures onto the specified
2095 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2096 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2097 * counts are supplied by rhcount and rhcount_lazy.
2099 * If warranted, also wake up the kthread servicing this CPUs queues.
2101 static void __call_rcu_nocb_enqueue(struct rcu_data
*rdp
,
2102 struct rcu_head
*rhp
,
2103 struct rcu_head
**rhtp
,
2104 int rhcount
, int rhcount_lazy
)
2107 struct rcu_head
**old_rhpp
;
2108 struct task_struct
*t
;
2110 /* Enqueue the callback on the nocb list and update counts. */
2111 old_rhpp
= xchg(&rdp
->nocb_tail
, rhtp
);
2112 ACCESS_ONCE(*old_rhpp
) = rhp
;
2113 atomic_long_add(rhcount
, &rdp
->nocb_q_count
);
2114 atomic_long_add(rhcount_lazy
, &rdp
->nocb_q_count_lazy
);
2116 /* If we are not being polled and there is a kthread, awaken it ... */
2117 t
= ACCESS_ONCE(rdp
->nocb_kthread
);
2118 if (rcu_nocb_poll
|| !t
) {
2119 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2120 TPS("WakeNotPoll"));
2123 len
= atomic_long_read(&rdp
->nocb_q_count
);
2124 if (old_rhpp
== &rdp
->nocb_head
) {
2125 wake_up(&rdp
->nocb_wq
); /* ... only if queue was empty ... */
2126 rdp
->qlen_last_fqs_check
= 0;
2127 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("WakeEmpty"));
2128 } else if (len
> rdp
->qlen_last_fqs_check
+ qhimark
) {
2129 wake_up_process(t
); /* ... or if many callbacks queued. */
2130 rdp
->qlen_last_fqs_check
= LONG_MAX
/ 2;
2131 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("WakeOvf"));
2133 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
, TPS("WakeNot"));
2139 * This is a helper for __call_rcu(), which invokes this when the normal
2140 * callback queue is inoperable. If this is not a no-CBs CPU, this
2141 * function returns failure back to __call_rcu(), which can complain
2144 * Otherwise, this function queues the callback where the corresponding
2145 * "rcuo" kthread can find it.
2147 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2151 if (!rcu_is_nocb_cpu(rdp
->cpu
))
2153 __call_rcu_nocb_enqueue(rdp
, rhp
, &rhp
->next
, 1, lazy
);
2154 if (__is_kfree_rcu_offset((unsigned long)rhp
->func
))
2155 trace_rcu_kfree_callback(rdp
->rsp
->name
, rhp
,
2156 (unsigned long)rhp
->func
,
2157 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
2158 -atomic_long_read(&rdp
->nocb_q_count
));
2160 trace_rcu_callback(rdp
->rsp
->name
, rhp
,
2161 -atomic_long_read(&rdp
->nocb_q_count_lazy
),
2162 -atomic_long_read(&rdp
->nocb_q_count
));
2167 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2170 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
2171 struct rcu_data
*rdp
)
2173 long ql
= rsp
->qlen
;
2174 long qll
= rsp
->qlen_lazy
;
2176 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2177 if (!rcu_is_nocb_cpu(smp_processor_id()))
2182 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2183 if (rsp
->orphan_donelist
!= NULL
) {
2184 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_donelist
,
2185 rsp
->orphan_donetail
, ql
, qll
);
2187 rsp
->orphan_donelist
= NULL
;
2188 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2190 if (rsp
->orphan_nxtlist
!= NULL
) {
2191 __call_rcu_nocb_enqueue(rdp
, rsp
->orphan_nxtlist
,
2192 rsp
->orphan_nxttail
, ql
, qll
);
2194 rsp
->orphan_nxtlist
= NULL
;
2195 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2201 * If necessary, kick off a new grace period, and either way wait
2202 * for a subsequent grace period to complete.
2204 static void rcu_nocb_wait_gp(struct rcu_data
*rdp
)
2208 unsigned long flags
;
2209 struct rcu_node
*rnp
= rdp
->mynode
;
2211 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2212 c
= rcu_start_future_gp(rnp
, rdp
);
2213 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2216 * Wait for the grace period. Do so interruptibly to avoid messing
2217 * up the load average.
2219 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("StartWait"));
2221 wait_event_interruptible(
2222 rnp
->nocb_gp_wq
[c
& 0x1],
2223 (d
= ULONG_CMP_GE(ACCESS_ONCE(rnp
->completed
), c
)));
2226 flush_signals(current
);
2227 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("ResumeWait"));
2229 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("EndWait"));
2230 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2234 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2235 * callbacks queued by the corresponding no-CBs CPU.
2237 static int rcu_nocb_kthread(void *arg
)
2241 struct rcu_head
*list
;
2242 struct rcu_head
*next
;
2243 struct rcu_head
**tail
;
2244 struct rcu_data
*rdp
= arg
;
2246 /* Each pass through this loop invokes one batch of callbacks */
2248 /* If not polling, wait for next batch of callbacks. */
2249 if (!rcu_nocb_poll
) {
2250 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2252 wait_event_interruptible(rdp
->nocb_wq
, rdp
->nocb_head
);
2253 } else if (firsttime
) {
2255 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2258 list
= ACCESS_ONCE(rdp
->nocb_head
);
2261 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2263 schedule_timeout_interruptible(1);
2264 flush_signals(current
);
2268 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2269 TPS("WokeNonEmpty"));
2272 * Extract queued callbacks, update counts, and wait
2273 * for a grace period to elapse.
2275 ACCESS_ONCE(rdp
->nocb_head
) = NULL
;
2276 tail
= xchg(&rdp
->nocb_tail
, &rdp
->nocb_head
);
2277 c
= atomic_long_xchg(&rdp
->nocb_q_count
, 0);
2278 cl
= atomic_long_xchg(&rdp
->nocb_q_count_lazy
, 0);
2279 ACCESS_ONCE(rdp
->nocb_p_count
) += c
;
2280 ACCESS_ONCE(rdp
->nocb_p_count_lazy
) += cl
;
2281 rcu_nocb_wait_gp(rdp
);
2283 /* Each pass through the following loop invokes a callback. */
2284 trace_rcu_batch_start(rdp
->rsp
->name
, cl
, c
, -1);
2288 /* Wait for enqueuing to complete, if needed. */
2289 while (next
== NULL
&& &list
->next
!= tail
) {
2290 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2292 schedule_timeout_interruptible(1);
2293 trace_rcu_nocb_wake(rdp
->rsp
->name
, rdp
->cpu
,
2297 debug_rcu_head_unqueue(list
);
2299 if (__rcu_reclaim(rdp
->rsp
->name
, list
))
2305 trace_rcu_batch_end(rdp
->rsp
->name
, c
, !!list
, 0, 0, 1);
2306 ACCESS_ONCE(rdp
->nocb_p_count
) -= c
;
2307 ACCESS_ONCE(rdp
->nocb_p_count_lazy
) -= cl
;
2308 rdp
->n_nocbs_invoked
+= c
;
2313 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2314 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2316 rdp
->nocb_tail
= &rdp
->nocb_head
;
2317 init_waitqueue_head(&rdp
->nocb_wq
);
2320 /* Create a kthread for each RCU flavor for each no-CBs CPU. */
2321 static void __init
rcu_spawn_nocb_kthreads(struct rcu_state
*rsp
)
2324 struct rcu_data
*rdp
;
2325 struct task_struct
*t
;
2327 if (rcu_nocb_mask
== NULL
)
2329 for_each_cpu(cpu
, rcu_nocb_mask
) {
2330 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2331 t
= kthread_run(rcu_nocb_kthread
, rdp
,
2332 "rcuo%c/%d", rsp
->abbr
, cpu
);
2334 ACCESS_ONCE(rdp
->nocb_kthread
) = t
;
2338 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2339 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2341 if (rcu_nocb_mask
== NULL
||
2342 !cpumask_test_cpu(rdp
->cpu
, rcu_nocb_mask
))
2344 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2348 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2350 static int rcu_nocb_needs_gp(struct rcu_state
*rsp
)
2355 static void rcu_nocb_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
2359 static void rcu_nocb_gp_set(struct rcu_node
*rnp
, int nrq
)
2363 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2367 static bool __call_rcu_nocb(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2373 static bool __maybe_unused
rcu_nocb_adopt_orphan_cbs(struct rcu_state
*rsp
,
2374 struct rcu_data
*rdp
)
2379 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2383 static void __init
rcu_spawn_nocb_kthreads(struct rcu_state
*rsp
)
2387 static bool init_nocb_callback_list(struct rcu_data
*rdp
)
2392 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2395 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2396 * arbitrarily long period of time with the scheduling-clock tick turned
2397 * off. RCU will be paying attention to this CPU because it is in the
2398 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2399 * machine because the scheduling-clock tick has been disabled. Therefore,
2400 * if an adaptive-ticks CPU is failing to respond to the current grace
2401 * period and has not be idle from an RCU perspective, kick it.
2403 static void rcu_kick_nohz_cpu(int cpu
)
2405 #ifdef CONFIG_NO_HZ_FULL
2406 if (tick_nohz_full_cpu(cpu
))
2407 smp_send_reschedule(cpu
);
2408 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2412 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2415 * Define RCU flavor that holds sysidle state. This needs to be the
2416 * most active flavor of RCU.
2418 #ifdef CONFIG_PREEMPT_RCU
2419 static struct rcu_state
*rcu_sysidle_state
= &rcu_preempt_state
;
2420 #else /* #ifdef CONFIG_PREEMPT_RCU */
2421 static struct rcu_state
*rcu_sysidle_state
= &rcu_sched_state
;
2422 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2424 static int full_sysidle_state
; /* Current system-idle state. */
2425 #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2426 #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2427 #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2428 #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2429 #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2432 * Invoked to note exit from irq or task transition to idle. Note that
2433 * usermode execution does -not- count as idle here! After all, we want
2434 * to detect full-system idle states, not RCU quiescent states and grace
2435 * periods. The caller must have disabled interrupts.
2437 static void rcu_sysidle_enter(struct rcu_dynticks
*rdtp
, int irq
)
2441 /* Adjust nesting, check for fully idle. */
2443 rdtp
->dynticks_idle_nesting
--;
2444 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2445 if (rdtp
->dynticks_idle_nesting
!= 0)
2446 return; /* Still not fully idle. */
2448 if ((rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) ==
2449 DYNTICK_TASK_NEST_VALUE
) {
2450 rdtp
->dynticks_idle_nesting
= 0;
2452 rdtp
->dynticks_idle_nesting
-= DYNTICK_TASK_NEST_VALUE
;
2453 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
< 0);
2454 return; /* Still not fully idle. */
2458 /* Record start of fully idle period. */
2460 ACCESS_ONCE(rdtp
->dynticks_idle_jiffies
) = j
;
2461 smp_mb__before_atomic_inc();
2462 atomic_inc(&rdtp
->dynticks_idle
);
2463 smp_mb__after_atomic_inc();
2464 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks_idle
) & 0x1);
2468 * Unconditionally force exit from full system-idle state. This is
2469 * invoked when a normal CPU exits idle, but must be called separately
2470 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2471 * is that the timekeeping CPU is permitted to take scheduling-clock
2472 * interrupts while the system is in system-idle state, and of course
2473 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2474 * interrupt from any other type of interrupt.
2476 void rcu_sysidle_force_exit(void)
2478 int oldstate
= ACCESS_ONCE(full_sysidle_state
);
2482 * Each pass through the following loop attempts to exit full
2483 * system-idle state. If contention proves to be a problem,
2484 * a trylock-based contention tree could be used here.
2486 while (oldstate
> RCU_SYSIDLE_SHORT
) {
2487 newoldstate
= cmpxchg(&full_sysidle_state
,
2488 oldstate
, RCU_SYSIDLE_NOT
);
2489 if (oldstate
== newoldstate
&&
2490 oldstate
== RCU_SYSIDLE_FULL_NOTED
) {
2491 rcu_kick_nohz_cpu(tick_do_timer_cpu
);
2492 return; /* We cleared it, done! */
2494 oldstate
= newoldstate
;
2496 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2500 * Invoked to note entry to irq or task transition from idle. Note that
2501 * usermode execution does -not- count as idle here! The caller must
2502 * have disabled interrupts.
2504 static void rcu_sysidle_exit(struct rcu_dynticks
*rdtp
, int irq
)
2506 /* Adjust nesting, check for already non-idle. */
2508 rdtp
->dynticks_idle_nesting
++;
2509 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2510 if (rdtp
->dynticks_idle_nesting
!= 1)
2511 return; /* Already non-idle. */
2514 * Allow for irq misnesting. Yes, it really is possible
2515 * to enter an irq handler then never leave it, and maybe
2516 * also vice versa. Handle both possibilities.
2518 if (rdtp
->dynticks_idle_nesting
& DYNTICK_TASK_NEST_MASK
) {
2519 rdtp
->dynticks_idle_nesting
+= DYNTICK_TASK_NEST_VALUE
;
2520 WARN_ON_ONCE(rdtp
->dynticks_idle_nesting
<= 0);
2521 return; /* Already non-idle. */
2523 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2527 /* Record end of idle period. */
2528 smp_mb__before_atomic_inc();
2529 atomic_inc(&rdtp
->dynticks_idle
);
2530 smp_mb__after_atomic_inc();
2531 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks_idle
) & 0x1));
2534 * If we are the timekeeping CPU, we are permitted to be non-idle
2535 * during a system-idle state. This must be the case, because
2536 * the timekeeping CPU has to take scheduling-clock interrupts
2537 * during the time that the system is transitioning to full
2538 * system-idle state. This means that the timekeeping CPU must
2539 * invoke rcu_sysidle_force_exit() directly if it does anything
2540 * more than take a scheduling-clock interrupt.
2542 if (smp_processor_id() == tick_do_timer_cpu
)
2545 /* Update system-idle state: We are clearly no longer fully idle! */
2546 rcu_sysidle_force_exit();
2550 * Check to see if the current CPU is idle. Note that usermode execution
2551 * does not count as idle. The caller must have disabled interrupts.
2553 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
2554 unsigned long *maxj
)
2558 struct rcu_dynticks
*rdtp
= rdp
->dynticks
;
2561 * If some other CPU has already reported non-idle, if this is
2562 * not the flavor of RCU that tracks sysidle state, or if this
2563 * is an offline or the timekeeping CPU, nothing to do.
2565 if (!*isidle
|| rdp
->rsp
!= rcu_sysidle_state
||
2566 cpu_is_offline(rdp
->cpu
) || rdp
->cpu
== tick_do_timer_cpu
)
2568 if (rcu_gp_in_progress(rdp
->rsp
))
2569 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
);
2571 /* Pick up current idle and NMI-nesting counter and check. */
2572 cur
= atomic_read(&rdtp
->dynticks_idle
);
2574 *isidle
= false; /* We are not idle! */
2577 smp_mb(); /* Read counters before timestamps. */
2579 /* Pick up timestamps. */
2580 j
= ACCESS_ONCE(rdtp
->dynticks_idle_jiffies
);
2581 /* If this CPU entered idle more recently, update maxj timestamp. */
2582 if (ULONG_CMP_LT(*maxj
, j
))
2587 * Is this the flavor of RCU that is handling full-system idle?
2589 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
2591 return rsp
== rcu_sysidle_state
;
2595 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2598 static void rcu_bind_gp_kthread(void)
2600 int cpu
= ACCESS_ONCE(tick_do_timer_cpu
);
2602 if (cpu
< 0 || cpu
>= nr_cpu_ids
)
2604 if (raw_smp_processor_id() != cpu
)
2605 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
2609 * Return a delay in jiffies based on the number of CPUs, rcu_node
2610 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2611 * systems more time to transition to full-idle state in order to
2612 * avoid the cache thrashing that otherwise occur on the state variable.
2613 * Really small systems (less than a couple of tens of CPUs) should
2614 * instead use a single global atomically incremented counter, and later
2615 * versions of this will automatically reconfigure themselves accordingly.
2617 static unsigned long rcu_sysidle_delay(void)
2619 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2621 return DIV_ROUND_UP(nr_cpu_ids
* HZ
, rcu_fanout_leaf
* 1000);
2625 * Advance the full-system-idle state. This is invoked when all of
2626 * the non-timekeeping CPUs are idle.
2628 static void rcu_sysidle(unsigned long j
)
2630 /* Check the current state. */
2631 switch (ACCESS_ONCE(full_sysidle_state
)) {
2632 case RCU_SYSIDLE_NOT
:
2634 /* First time all are idle, so note a short idle period. */
2635 ACCESS_ONCE(full_sysidle_state
) = RCU_SYSIDLE_SHORT
;
2638 case RCU_SYSIDLE_SHORT
:
2641 * Idle for a bit, time to advance to next state?
2642 * cmpxchg failure means race with non-idle, let them win.
2644 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2645 (void)cmpxchg(&full_sysidle_state
,
2646 RCU_SYSIDLE_SHORT
, RCU_SYSIDLE_LONG
);
2649 case RCU_SYSIDLE_LONG
:
2652 * Do an additional check pass before advancing to full.
2653 * cmpxchg failure means race with non-idle, let them win.
2655 if (ULONG_CMP_GE(jiffies
, j
+ rcu_sysidle_delay()))
2656 (void)cmpxchg(&full_sysidle_state
,
2657 RCU_SYSIDLE_LONG
, RCU_SYSIDLE_FULL
);
2666 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2667 * back to the beginning.
2669 static void rcu_sysidle_cancel(void)
2672 ACCESS_ONCE(full_sysidle_state
) = RCU_SYSIDLE_NOT
;
2676 * Update the sysidle state based on the results of a force-quiescent-state
2677 * scan of the CPUs' dyntick-idle state.
2679 static void rcu_sysidle_report(struct rcu_state
*rsp
, int isidle
,
2680 unsigned long maxj
, bool gpkt
)
2682 if (rsp
!= rcu_sysidle_state
)
2683 return; /* Wrong flavor, ignore. */
2684 if (gpkt
&& nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
)
2685 return; /* Running state machine from timekeeping CPU. */
2687 rcu_sysidle(maxj
); /* More idle! */
2689 rcu_sysidle_cancel(); /* Idle is over. */
2693 * Wrapper for rcu_sysidle_report() when called from the grace-period
2694 * kthread's context.
2696 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
2699 rcu_sysidle_report(rsp
, isidle
, maxj
, true);
2702 /* Callback and function for forcing an RCU grace period. */
2703 struct rcu_sysidle_head
{
2708 static void rcu_sysidle_cb(struct rcu_head
*rhp
)
2710 struct rcu_sysidle_head
*rshp
;
2713 * The following memory barrier is needed to replace the
2714 * memory barriers that would normally be in the memory
2717 smp_mb(); /* grace period precedes setting inuse. */
2719 rshp
= container_of(rhp
, struct rcu_sysidle_head
, rh
);
2720 ACCESS_ONCE(rshp
->inuse
) = 0;
2724 * Check to see if the system is fully idle, other than the timekeeping CPU.
2725 * The caller must have disabled interrupts.
2727 bool rcu_sys_is_idle(void)
2729 static struct rcu_sysidle_head rsh
;
2730 int rss
= ACCESS_ONCE(full_sysidle_state
);
2732 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu
))
2735 /* Handle small-system case by doing a full scan of CPUs. */
2736 if (nr_cpu_ids
<= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
) {
2737 int oldrss
= rss
- 1;
2740 * One pass to advance to each state up to _FULL.
2741 * Give up if any pass fails to advance the state.
2743 while (rss
< RCU_SYSIDLE_FULL
&& oldrss
< rss
) {
2746 unsigned long maxj
= jiffies
- ULONG_MAX
/ 4;
2747 struct rcu_data
*rdp
;
2749 /* Scan all the CPUs looking for nonidle CPUs. */
2750 for_each_possible_cpu(cpu
) {
2751 rdp
= per_cpu_ptr(rcu_sysidle_state
->rda
, cpu
);
2752 rcu_sysidle_check_cpu(rdp
, &isidle
, &maxj
);
2756 rcu_sysidle_report(rcu_sysidle_state
,
2757 isidle
, maxj
, false);
2759 rss
= ACCESS_ONCE(full_sysidle_state
);
2763 /* If this is the first observation of an idle period, record it. */
2764 if (rss
== RCU_SYSIDLE_FULL
) {
2765 rss
= cmpxchg(&full_sysidle_state
,
2766 RCU_SYSIDLE_FULL
, RCU_SYSIDLE_FULL_NOTED
);
2767 return rss
== RCU_SYSIDLE_FULL
;
2770 smp_mb(); /* ensure rss load happens before later caller actions. */
2772 /* If already fully idle, tell the caller (in case of races). */
2773 if (rss
== RCU_SYSIDLE_FULL_NOTED
)
2777 * If we aren't there yet, and a grace period is not in flight,
2778 * initiate a grace period. Either way, tell the caller that
2779 * we are not there yet. We use an xchg() rather than an assignment
2780 * to make up for the memory barriers that would otherwise be
2781 * provided by the memory allocator.
2783 if (nr_cpu_ids
> CONFIG_NO_HZ_FULL_SYSIDLE_SMALL
&&
2784 !rcu_gp_in_progress(rcu_sysidle_state
) &&
2785 !rsh
.inuse
&& xchg(&rsh
.inuse
, 1) == 0)
2786 call_rcu(&rsh
.rh
, rcu_sysidle_cb
);
2791 * Initialize dynticks sysidle state for CPUs coming online.
2793 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
2795 rdtp
->dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
;
2798 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2800 static void rcu_sysidle_enter(struct rcu_dynticks
*rdtp
, int irq
)
2804 static void rcu_sysidle_exit(struct rcu_dynticks
*rdtp
, int irq
)
2808 static void rcu_sysidle_check_cpu(struct rcu_data
*rdp
, bool *isidle
,
2809 unsigned long *maxj
)
2813 static bool is_sysidle_rcu_state(struct rcu_state
*rsp
)
2818 static void rcu_bind_gp_kthread(void)
2822 static void rcu_sysidle_report_gp(struct rcu_state
*rsp
, int isidle
,
2827 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks
*rdtp
)
2831 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */