2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <asm/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
55 /* Data structures. */
57 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
59 #define RCU_STATE_INITIALIZER(structname) { \
60 .level = { &structname.node[0] }, \
62 NUM_RCU_LVL_0, /* root of hierarchy. */ \
66 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
68 .signaled = RCU_GP_IDLE, \
71 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
72 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
74 .n_force_qs_ngp = 0, \
75 .name = #structname, \
78 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched_state
);
79 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
81 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh_state
);
82 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
84 static struct rcu_state
*rcu_state
;
86 int rcu_scheduler_active __read_mostly
;
87 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
90 * Control variables for per-CPU and per-rcu_node kthreads. These
91 * handle all flavors of RCU.
93 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
94 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
95 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
96 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
97 static DEFINE_PER_CPU(wait_queue_head_t
, rcu_cpu_wq
);
98 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
99 static char rcu_kthreads_spawnable
;
101 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
102 static void invoke_rcu_cpu_kthread(void);
104 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
107 * Track the rcutorture test sequence number and the update version
108 * number within a given test. The rcutorture_testseq is incremented
109 * on every rcutorture module load and unload, so has an odd value
110 * when a test is running. The rcutorture_vernum is set to zero
111 * when rcutorture starts and is incremented on each rcutorture update.
112 * These variables enable correlating rcutorture output with the
113 * RCU tracing information.
115 unsigned long rcutorture_testseq
;
116 unsigned long rcutorture_vernum
;
119 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
120 * permit this function to be invoked without holding the root rcu_node
121 * structure's ->lock, but of course results can be subject to change.
123 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
125 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
129 * Note a quiescent state. Because we do not need to know
130 * how many quiescent states passed, just if there was at least
131 * one since the start of the grace period, this just sets a flag.
133 void rcu_sched_qs(int cpu
)
135 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
137 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
139 rdp
->passed_quiesc
= 1;
142 void rcu_bh_qs(int cpu
)
144 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
146 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
148 rdp
->passed_quiesc
= 1;
152 * Note a context switch. This is a quiescent state for RCU-sched,
153 * and requires special handling for preemptible RCU.
155 void rcu_note_context_switch(int cpu
)
158 rcu_preempt_note_context_switch(cpu
);
160 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
163 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
164 .dynticks_nesting
= 1,
167 #endif /* #ifdef CONFIG_NO_HZ */
169 static int blimit
= 10; /* Maximum callbacks per softirq. */
170 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
171 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
173 module_param(blimit
, int, 0);
174 module_param(qhimark
, int, 0);
175 module_param(qlowmark
, int, 0);
177 int rcu_cpu_stall_suppress __read_mostly
;
178 module_param(rcu_cpu_stall_suppress
, int, 0644);
180 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
181 static int rcu_pending(int cpu
);
184 * Return the number of RCU-sched batches processed thus far for debug & stats.
186 long rcu_batches_completed_sched(void)
188 return rcu_sched_state
.completed
;
190 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
193 * Return the number of RCU BH batches processed thus far for debug & stats.
195 long rcu_batches_completed_bh(void)
197 return rcu_bh_state
.completed
;
199 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
202 * Force a quiescent state for RCU BH.
204 void rcu_bh_force_quiescent_state(void)
206 force_quiescent_state(&rcu_bh_state
, 0);
208 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
211 * Record the number of times rcutorture tests have been initiated and
212 * terminated. This information allows the debugfs tracing stats to be
213 * correlated to the rcutorture messages, even when the rcutorture module
214 * is being repeatedly loaded and unloaded. In other words, we cannot
215 * store this state in rcutorture itself.
217 void rcutorture_record_test_transition(void)
219 rcutorture_testseq
++;
220 rcutorture_vernum
= 0;
222 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
225 * Record the number of writer passes through the current rcutorture test.
226 * This is also used to correlate debugfs tracing stats with the rcutorture
229 void rcutorture_record_progress(unsigned long vernum
)
233 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
236 * Force a quiescent state for RCU-sched.
238 void rcu_sched_force_quiescent_state(void)
240 force_quiescent_state(&rcu_sched_state
, 0);
242 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
245 * Does the CPU have callbacks ready to be invoked?
248 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
250 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
254 * Does the current CPU require a yet-as-unscheduled grace period?
257 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
259 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
263 * Return the root node of the specified rcu_state structure.
265 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
267 return &rsp
->node
[0];
273 * If the specified CPU is offline, tell the caller that it is in
274 * a quiescent state. Otherwise, whack it with a reschedule IPI.
275 * Grace periods can end up waiting on an offline CPU when that
276 * CPU is in the process of coming online -- it will be added to the
277 * rcu_node bitmasks before it actually makes it online. The same thing
278 * can happen while a CPU is in the process of coming online. Because this
279 * race is quite rare, we check for it after detecting that the grace
280 * period has been delayed rather than checking each and every CPU
281 * each and every time we start a new grace period.
283 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
286 * If the CPU is offline, it is in a quiescent state. We can
287 * trust its state not to change because interrupts are disabled.
289 if (cpu_is_offline(rdp
->cpu
)) {
294 /* If preemptible RCU, no point in sending reschedule IPI. */
295 if (rdp
->preemptible
)
298 /* The CPU is online, so send it a reschedule IPI. */
299 if (rdp
->cpu
!= smp_processor_id())
300 smp_send_reschedule(rdp
->cpu
);
307 #endif /* #ifdef CONFIG_SMP */
312 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
314 * Enter nohz mode, in other words, -leave- the mode in which RCU
315 * read-side critical sections can occur. (Though RCU read-side
316 * critical sections can occur in irq handlers in nohz mode, a possibility
317 * handled by rcu_irq_enter() and rcu_irq_exit()).
319 void rcu_enter_nohz(void)
322 struct rcu_dynticks
*rdtp
;
324 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
325 local_irq_save(flags
);
326 rdtp
= &__get_cpu_var(rcu_dynticks
);
328 rdtp
->dynticks_nesting
--;
329 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
330 local_irq_restore(flags
);
334 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
336 * Exit nohz mode, in other words, -enter- the mode in which RCU
337 * read-side critical sections normally occur.
339 void rcu_exit_nohz(void)
342 struct rcu_dynticks
*rdtp
;
344 local_irq_save(flags
);
345 rdtp
= &__get_cpu_var(rcu_dynticks
);
347 rdtp
->dynticks_nesting
++;
348 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
349 local_irq_restore(flags
);
350 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
354 * rcu_nmi_enter - inform RCU of entry to NMI context
356 * If the CPU was idle with dynamic ticks active, and there is no
357 * irq handler running, this updates rdtp->dynticks_nmi to let the
358 * RCU grace-period handling know that the CPU is active.
360 void rcu_nmi_enter(void)
362 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
364 if (rdtp
->dynticks
& 0x1)
366 rdtp
->dynticks_nmi
++;
367 WARN_ON_ONCE(!(rdtp
->dynticks_nmi
& 0x1));
368 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
372 * rcu_nmi_exit - inform RCU of exit from NMI context
374 * If the CPU was idle with dynamic ticks active, and there is no
375 * irq handler running, this updates rdtp->dynticks_nmi to let the
376 * RCU grace-period handling know that the CPU is no longer active.
378 void rcu_nmi_exit(void)
380 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
382 if (rdtp
->dynticks
& 0x1)
384 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
385 rdtp
->dynticks_nmi
++;
386 WARN_ON_ONCE(rdtp
->dynticks_nmi
& 0x1);
390 * rcu_irq_enter - inform RCU of entry to hard irq context
392 * If the CPU was idle with dynamic ticks active, this updates the
393 * rdtp->dynticks to let the RCU handling know that the CPU is active.
395 void rcu_irq_enter(void)
397 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
399 if (rdtp
->dynticks_nesting
++)
402 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
403 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
407 * rcu_irq_exit - inform RCU of exit from hard irq context
409 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
410 * to put let the RCU handling be aware that the CPU is going back to idle
413 void rcu_irq_exit(void)
415 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
417 if (--rdtp
->dynticks_nesting
)
419 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
421 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
423 /* If the interrupt queued a callback, get out of dyntick mode. */
424 if (__this_cpu_read(rcu_sched_data
.nxtlist
) ||
425 __this_cpu_read(rcu_bh_data
.nxtlist
))
432 * Snapshot the specified CPU's dynticks counter so that we can later
433 * credit them with an implicit quiescent state. Return 1 if this CPU
434 * is in dynticks idle mode, which is an extended quiescent state.
436 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
442 snap
= rdp
->dynticks
->dynticks
;
443 snap_nmi
= rdp
->dynticks
->dynticks_nmi
;
444 smp_mb(); /* Order sampling of snap with end of grace period. */
445 rdp
->dynticks_snap
= snap
;
446 rdp
->dynticks_nmi_snap
= snap_nmi
;
447 ret
= ((snap
& 0x1) == 0) && ((snap_nmi
& 0x1) == 0);
454 * Return true if the specified CPU has passed through a quiescent
455 * state by virtue of being in or having passed through an dynticks
456 * idle state since the last call to dyntick_save_progress_counter()
459 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
466 curr
= rdp
->dynticks
->dynticks
;
467 snap
= rdp
->dynticks_snap
;
468 curr_nmi
= rdp
->dynticks
->dynticks_nmi
;
469 snap_nmi
= rdp
->dynticks_nmi_snap
;
470 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
473 * If the CPU passed through or entered a dynticks idle phase with
474 * no active irq/NMI handlers, then we can safely pretend that the CPU
475 * already acknowledged the request to pass through a quiescent
476 * state. Either way, that CPU cannot possibly be in an RCU
477 * read-side critical section that started before the beginning
478 * of the current RCU grace period.
480 if ((curr
!= snap
|| (curr
& 0x1) == 0) &&
481 (curr_nmi
!= snap_nmi
|| (curr_nmi
& 0x1) == 0)) {
486 /* Go check for the CPU being offline. */
487 return rcu_implicit_offline_qs(rdp
);
490 #endif /* #ifdef CONFIG_SMP */
492 #else /* #ifdef CONFIG_NO_HZ */
496 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
501 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
503 return rcu_implicit_offline_qs(rdp
);
506 #endif /* #ifdef CONFIG_SMP */
508 #endif /* #else #ifdef CONFIG_NO_HZ */
510 int rcu_cpu_stall_suppress __read_mostly
;
512 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
514 rsp
->gp_start
= jiffies
;
515 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
518 static void print_other_cpu_stall(struct rcu_state
*rsp
)
523 struct rcu_node
*rnp
= rcu_get_root(rsp
);
525 /* Only let one CPU complain about others per time interval. */
527 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
528 delta
= jiffies
- rsp
->jiffies_stall
;
529 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
530 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
533 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
536 * Now rat on any tasks that got kicked up to the root rcu_node
537 * due to CPU offlining.
539 rcu_print_task_stall(rnp
);
540 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
543 * OK, time to rat on our buddy...
544 * See Documentation/RCU/stallwarn.txt for info on how to debug
545 * RCU CPU stall warnings.
547 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
549 rcu_for_each_leaf_node(rsp
, rnp
) {
550 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
551 rcu_print_task_stall(rnp
);
552 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
553 if (rnp
->qsmask
== 0)
555 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
556 if (rnp
->qsmask
& (1UL << cpu
))
557 printk(" %d", rnp
->grplo
+ cpu
);
559 printk("} (detected by %d, t=%ld jiffies)\n",
560 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
561 trigger_all_cpu_backtrace();
563 /* If so configured, complain about tasks blocking the grace period. */
565 rcu_print_detail_task_stall(rsp
);
567 force_quiescent_state(rsp
, 0); /* Kick them all. */
570 static void print_cpu_stall(struct rcu_state
*rsp
)
573 struct rcu_node
*rnp
= rcu_get_root(rsp
);
576 * OK, time to rat on ourselves...
577 * See Documentation/RCU/stallwarn.txt for info on how to debug
578 * RCU CPU stall warnings.
580 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
581 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
582 trigger_all_cpu_backtrace();
584 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
585 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
587 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
588 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
590 set_need_resched(); /* kick ourselves to get things going. */
593 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
597 struct rcu_node
*rnp
;
599 if (rcu_cpu_stall_suppress
)
601 j
= ACCESS_ONCE(jiffies
);
602 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
604 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
606 /* We haven't checked in, so go dump stack. */
607 print_cpu_stall(rsp
);
609 } else if (rcu_gp_in_progress(rsp
) &&
610 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
612 /* They had a few time units to dump stack, so complain. */
613 print_other_cpu_stall(rsp
);
617 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
619 rcu_cpu_stall_suppress
= 1;
624 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
626 * Set the stall-warning timeout way off into the future, thus preventing
627 * any RCU CPU stall-warning messages from appearing in the current set of
630 * The caller must disable hard irqs.
632 void rcu_cpu_stall_reset(void)
634 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
635 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
636 rcu_preempt_stall_reset();
639 static struct notifier_block rcu_panic_block
= {
640 .notifier_call
= rcu_panic
,
643 static void __init
check_cpu_stall_init(void)
645 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
649 * Update CPU-local rcu_data state to record the newly noticed grace period.
650 * This is used both when we started the grace period and when we notice
651 * that someone else started the grace period. The caller must hold the
652 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
653 * and must have irqs disabled.
655 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
657 if (rdp
->gpnum
!= rnp
->gpnum
) {
659 * If the current grace period is waiting for this CPU,
660 * set up to detect a quiescent state, otherwise don't
661 * go looking for one.
663 rdp
->gpnum
= rnp
->gpnum
;
664 if (rnp
->qsmask
& rdp
->grpmask
) {
666 rdp
->passed_quiesc
= 0;
672 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
675 struct rcu_node
*rnp
;
677 local_irq_save(flags
);
679 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
680 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
681 local_irq_restore(flags
);
684 __note_new_gpnum(rsp
, rnp
, rdp
);
685 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
689 * Did someone else start a new RCU grace period start since we last
690 * checked? Update local state appropriately if so. Must be called
691 * on the CPU corresponding to rdp.
694 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
699 local_irq_save(flags
);
700 if (rdp
->gpnum
!= rsp
->gpnum
) {
701 note_new_gpnum(rsp
, rdp
);
704 local_irq_restore(flags
);
709 * Advance this CPU's callbacks, but only if the current grace period
710 * has ended. This may be called only from the CPU to whom the rdp
711 * belongs. In addition, the corresponding leaf rcu_node structure's
712 * ->lock must be held by the caller, with irqs disabled.
715 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
717 /* Did another grace period end? */
718 if (rdp
->completed
!= rnp
->completed
) {
720 /* Advance callbacks. No harm if list empty. */
721 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
722 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
723 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
725 /* Remember that we saw this grace-period completion. */
726 rdp
->completed
= rnp
->completed
;
729 * If we were in an extended quiescent state, we may have
730 * missed some grace periods that others CPUs handled on
731 * our behalf. Catch up with this state to avoid noting
732 * spurious new grace periods. If another grace period
733 * has started, then rnp->gpnum will have advanced, so
734 * we will detect this later on.
736 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
737 rdp
->gpnum
= rdp
->completed
;
740 * If RCU does not need a quiescent state from this CPU,
741 * then make sure that this CPU doesn't go looking for one.
743 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
749 * Advance this CPU's callbacks, but only if the current grace period
750 * has ended. This may be called only from the CPU to whom the rdp
754 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
757 struct rcu_node
*rnp
;
759 local_irq_save(flags
);
761 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
762 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
763 local_irq_restore(flags
);
766 __rcu_process_gp_end(rsp
, rnp
, rdp
);
767 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
771 * Do per-CPU grace-period initialization for running CPU. The caller
772 * must hold the lock of the leaf rcu_node structure corresponding to
776 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
778 /* Prior grace period ended, so advance callbacks for current CPU. */
779 __rcu_process_gp_end(rsp
, rnp
, rdp
);
782 * Because this CPU just now started the new grace period, we know
783 * that all of its callbacks will be covered by this upcoming grace
784 * period, even the ones that were registered arbitrarily recently.
785 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
787 * Other CPUs cannot be sure exactly when the grace period started.
788 * Therefore, their recently registered callbacks must pass through
789 * an additional RCU_NEXT_READY stage, so that they will be handled
790 * by the next RCU grace period.
792 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
793 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
795 /* Set state so that this CPU will detect the next quiescent state. */
796 __note_new_gpnum(rsp
, rnp
, rdp
);
800 * Start a new RCU grace period if warranted, re-initializing the hierarchy
801 * in preparation for detecting the next grace period. The caller must hold
802 * the root node's ->lock, which is released before return. Hard irqs must
806 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
807 __releases(rcu_get_root(rsp
)->lock
)
809 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
810 struct rcu_node
*rnp
= rcu_get_root(rsp
);
812 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
813 if (cpu_needs_another_gp(rsp
, rdp
))
814 rsp
->fqs_need_gp
= 1;
815 if (rnp
->completed
== rsp
->completed
) {
816 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
819 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
822 * Propagate new ->completed value to rcu_node structures
823 * so that other CPUs don't have to wait until the start
824 * of the next grace period to process their callbacks.
826 rcu_for_each_node_breadth_first(rsp
, rnp
) {
827 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
828 rnp
->completed
= rsp
->completed
;
829 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
831 local_irq_restore(flags
);
835 /* Advance to a new grace period and initialize state. */
837 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
838 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
839 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
840 record_gp_stall_check_time(rsp
);
842 /* Special-case the common single-level case. */
843 if (NUM_RCU_NODES
== 1) {
844 rcu_preempt_check_blocked_tasks(rnp
);
845 rnp
->qsmask
= rnp
->qsmaskinit
;
846 rnp
->gpnum
= rsp
->gpnum
;
847 rnp
->completed
= rsp
->completed
;
848 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
849 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
850 rcu_preempt_boost_start_gp(rnp
);
851 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
855 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
858 /* Exclude any concurrent CPU-hotplug operations. */
859 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
862 * Set the quiescent-state-needed bits in all the rcu_node
863 * structures for all currently online CPUs in breadth-first
864 * order, starting from the root rcu_node structure. This
865 * operation relies on the layout of the hierarchy within the
866 * rsp->node[] array. Note that other CPUs will access only
867 * the leaves of the hierarchy, which still indicate that no
868 * grace period is in progress, at least until the corresponding
869 * leaf node has been initialized. In addition, we have excluded
870 * CPU-hotplug operations.
872 * Note that the grace period cannot complete until we finish
873 * the initialization process, as there will be at least one
874 * qsmask bit set in the root node until that time, namely the
875 * one corresponding to this CPU, due to the fact that we have
878 rcu_for_each_node_breadth_first(rsp
, rnp
) {
879 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
880 rcu_preempt_check_blocked_tasks(rnp
);
881 rnp
->qsmask
= rnp
->qsmaskinit
;
882 rnp
->gpnum
= rsp
->gpnum
;
883 rnp
->completed
= rsp
->completed
;
884 if (rnp
== rdp
->mynode
)
885 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
886 rcu_preempt_boost_start_gp(rnp
);
887 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
890 rnp
= rcu_get_root(rsp
);
891 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
892 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
893 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
894 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
898 * Report a full set of quiescent states to the specified rcu_state
899 * data structure. This involves cleaning up after the prior grace
900 * period and letting rcu_start_gp() start up the next grace period
901 * if one is needed. Note that the caller must hold rnp->lock, as
902 * required by rcu_start_gp(), which will release it.
904 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
905 __releases(rcu_get_root(rsp
)->lock
)
907 unsigned long gp_duration
;
909 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
910 gp_duration
= jiffies
- rsp
->gp_start
;
911 if (gp_duration
> rsp
->gp_max
)
912 rsp
->gp_max
= gp_duration
;
913 rsp
->completed
= rsp
->gpnum
;
914 rsp
->signaled
= RCU_GP_IDLE
;
915 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
919 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
920 * Allows quiescent states for a group of CPUs to be reported at one go
921 * to the specified rcu_node structure, though all the CPUs in the group
922 * must be represented by the same rcu_node structure (which need not be
923 * a leaf rcu_node structure, though it often will be). That structure's
924 * lock must be held upon entry, and it is released before return.
927 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
928 struct rcu_node
*rnp
, unsigned long flags
)
929 __releases(rnp
->lock
)
931 struct rcu_node
*rnp_c
;
933 /* Walk up the rcu_node hierarchy. */
935 if (!(rnp
->qsmask
& mask
)) {
937 /* Our bit has already been cleared, so done. */
938 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
941 rnp
->qsmask
&= ~mask
;
942 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
944 /* Other bits still set at this level, so done. */
945 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
949 if (rnp
->parent
== NULL
) {
951 /* No more levels. Exit loop holding root lock. */
955 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
958 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
959 WARN_ON_ONCE(rnp_c
->qsmask
);
963 * Get here if we are the last CPU to pass through a quiescent
964 * state for this grace period. Invoke rcu_report_qs_rsp()
965 * to clean up and start the next grace period if one is needed.
967 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
971 * Record a quiescent state for the specified CPU to that CPU's rcu_data
972 * structure. This must be either called from the specified CPU, or
973 * called when the specified CPU is known to be offline (and when it is
974 * also known that no other CPU is concurrently trying to help the offline
975 * CPU). The lastcomp argument is used to make sure we are still in the
976 * grace period of interest. We don't want to end the current grace period
977 * based on quiescent states detected in an earlier grace period!
980 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
984 struct rcu_node
*rnp
;
987 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
988 if (lastcomp
!= rnp
->completed
) {
991 * Someone beat us to it for this grace period, so leave.
992 * The race with GP start is resolved by the fact that we
993 * hold the leaf rcu_node lock, so that the per-CPU bits
994 * cannot yet be initialized -- so we would simply find our
995 * CPU's bit already cleared in rcu_report_qs_rnp() if this
998 rdp
->passed_quiesc
= 0; /* try again later! */
999 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1002 mask
= rdp
->grpmask
;
1003 if ((rnp
->qsmask
& mask
) == 0) {
1004 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1006 rdp
->qs_pending
= 0;
1009 * This GP can't end until cpu checks in, so all of our
1010 * callbacks can be processed during the next GP.
1012 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1014 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1019 * Check to see if there is a new grace period of which this CPU
1020 * is not yet aware, and if so, set up local rcu_data state for it.
1021 * Otherwise, see if this CPU has just passed through its first
1022 * quiescent state for this grace period, and record that fact if so.
1025 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1027 /* If there is now a new grace period, record and return. */
1028 if (check_for_new_grace_period(rsp
, rdp
))
1032 * Does this CPU still need to do its part for current grace period?
1033 * If no, return and let the other CPUs do their part as well.
1035 if (!rdp
->qs_pending
)
1039 * Was there a quiescent state since the beginning of the grace
1040 * period? If no, then exit and wait for the next call.
1042 if (!rdp
->passed_quiesc
)
1046 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1049 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
1052 #ifdef CONFIG_HOTPLUG_CPU
1055 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1056 * Synchronization is not required because this function executes
1057 * in stop_machine() context.
1059 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1062 /* current DYING CPU is cleared in the cpu_online_mask */
1063 int receive_cpu
= cpumask_any(cpu_online_mask
);
1064 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1065 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1067 if (rdp
->nxtlist
== NULL
)
1068 return; /* irqs disabled, so comparison is stable. */
1070 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1071 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1072 receive_rdp
->qlen
+= rdp
->qlen
;
1073 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1074 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1076 rdp
->nxtlist
= NULL
;
1077 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1078 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1083 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1084 * and move all callbacks from the outgoing CPU to the current one.
1085 * There can only be one CPU hotplug operation at a time, so no other
1086 * CPU can be attempting to update rcu_cpu_kthread_task.
1088 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1090 unsigned long flags
;
1092 int need_report
= 0;
1093 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1094 struct rcu_node
*rnp
;
1095 struct task_struct
*t
;
1097 /* Stop the CPU's kthread. */
1098 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1100 per_cpu(rcu_cpu_kthread_task
, cpu
) = NULL
;
1104 /* Exclude any attempts to start a new grace period. */
1105 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1107 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1108 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1109 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1111 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1112 rnp
->qsmaskinit
&= ~mask
;
1113 if (rnp
->qsmaskinit
!= 0) {
1114 if (rnp
!= rdp
->mynode
)
1115 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1118 if (rnp
== rdp
->mynode
)
1119 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1121 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1122 mask
= rnp
->grpmask
;
1124 } while (rnp
!= NULL
);
1127 * We still hold the leaf rcu_node structure lock here, and
1128 * irqs are still disabled. The reason for this subterfuge is
1129 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1130 * held leads to deadlock.
1132 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1134 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1135 rcu_report_unblock_qs_rnp(rnp
, flags
);
1137 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1138 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1139 rcu_report_exp_rnp(rsp
, rnp
);
1140 rcu_node_kthread_setaffinity(rnp
, -1);
1144 * Remove the specified CPU from the RCU hierarchy and move any pending
1145 * callbacks that it might have to the current CPU. This code assumes
1146 * that at least one CPU in the system will remain running at all times.
1147 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1149 static void rcu_offline_cpu(int cpu
)
1151 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1152 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1153 rcu_preempt_offline_cpu(cpu
);
1156 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1158 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1162 static void rcu_offline_cpu(int cpu
)
1166 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1169 * Invoke any RCU callbacks that have made it to the end of their grace
1170 * period. Thottle as specified by rdp->blimit.
1172 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1174 unsigned long flags
;
1175 struct rcu_head
*next
, *list
, **tail
;
1178 /* If no callbacks are ready, just return.*/
1179 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1183 * Extract the list of ready callbacks, disabling to prevent
1184 * races with call_rcu() from interrupt handlers.
1186 local_irq_save(flags
);
1187 list
= rdp
->nxtlist
;
1188 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1189 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1190 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1191 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1192 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1193 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1194 local_irq_restore(flags
);
1196 /* Invoke callbacks. */
1201 debug_rcu_head_unqueue(list
);
1202 __rcu_reclaim(list
);
1204 if (++count
>= rdp
->blimit
)
1208 local_irq_save(flags
);
1210 /* Update count, and requeue any remaining callbacks. */
1212 rdp
->n_cbs_invoked
+= count
;
1214 *tail
= rdp
->nxtlist
;
1215 rdp
->nxtlist
= list
;
1216 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1217 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1218 rdp
->nxttail
[count
] = tail
;
1223 /* Reinstate batch limit if we have worked down the excess. */
1224 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1225 rdp
->blimit
= blimit
;
1227 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1228 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1229 rdp
->qlen_last_fqs_check
= 0;
1230 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1231 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1232 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1234 local_irq_restore(flags
);
1236 /* Re-raise the RCU softirq if there are callbacks remaining. */
1237 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1238 invoke_rcu_cpu_kthread();
1242 * Check to see if this CPU is in a non-context-switch quiescent state
1243 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1244 * Also schedule the RCU softirq handler.
1246 * This function must be called with hardirqs disabled. It is normally
1247 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1248 * false, there is no point in invoking rcu_check_callbacks().
1250 void rcu_check_callbacks(int cpu
, int user
)
1253 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1254 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1257 * Get here if this CPU took its interrupt from user
1258 * mode or from the idle loop, and if this is not a
1259 * nested interrupt. In this case, the CPU is in
1260 * a quiescent state, so note it.
1262 * No memory barrier is required here because both
1263 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1264 * variables that other CPUs neither access nor modify,
1265 * at least not while the corresponding CPU is online.
1271 } else if (!in_softirq()) {
1274 * Get here if this CPU did not take its interrupt from
1275 * softirq, in other words, if it is not interrupting
1276 * a rcu_bh read-side critical section. This is an _bh
1277 * critical section, so note it.
1282 rcu_preempt_check_callbacks(cpu
);
1283 if (rcu_pending(cpu
))
1284 invoke_rcu_cpu_kthread();
1290 * Scan the leaf rcu_node structures, processing dyntick state for any that
1291 * have not yet encountered a quiescent state, using the function specified.
1292 * Also initiate boosting for any threads blocked on the root rcu_node.
1294 * The caller must have suppressed start of new grace periods.
1296 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1300 unsigned long flags
;
1302 struct rcu_node
*rnp
;
1304 rcu_for_each_leaf_node(rsp
, rnp
) {
1306 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1307 if (!rcu_gp_in_progress(rsp
)) {
1308 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1311 if (rnp
->qsmask
== 0) {
1312 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1317 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1318 if ((rnp
->qsmask
& bit
) != 0 &&
1319 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1324 /* rcu_report_qs_rnp() releases rnp->lock. */
1325 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1328 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1330 rnp
= rcu_get_root(rsp
);
1331 if (rnp
->qsmask
== 0) {
1332 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1333 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1338 * Force quiescent states on reluctant CPUs, and also detect which
1339 * CPUs are in dyntick-idle mode.
1341 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1343 unsigned long flags
;
1344 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1346 if (!rcu_gp_in_progress(rsp
))
1347 return; /* No grace period in progress, nothing to force. */
1348 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1349 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1350 return; /* Someone else is already on the job. */
1352 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1353 goto unlock_fqs_ret
; /* no emergency and done recently. */
1355 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1356 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1357 if(!rcu_gp_in_progress(rsp
)) {
1358 rsp
->n_force_qs_ngp
++;
1359 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1360 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1362 rsp
->fqs_active
= 1;
1363 switch (rsp
->signaled
) {
1367 break; /* grace period idle or initializing, ignore. */
1369 case RCU_SAVE_DYNTICK
:
1370 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1371 break; /* So gcc recognizes the dead code. */
1373 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1375 /* Record dyntick-idle state. */
1376 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1377 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1378 if (rcu_gp_in_progress(rsp
))
1379 rsp
->signaled
= RCU_FORCE_QS
;
1384 /* Check dyntick-idle state, send IPI to laggarts. */
1385 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1386 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1388 /* Leave state in case more forcing is required. */
1390 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1393 rsp
->fqs_active
= 0;
1394 if (rsp
->fqs_need_gp
) {
1395 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1396 rsp
->fqs_need_gp
= 0;
1397 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1400 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1402 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1405 #else /* #ifdef CONFIG_SMP */
1407 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1412 #endif /* #else #ifdef CONFIG_SMP */
1415 * This does the RCU processing work from softirq context for the
1416 * specified rcu_state and rcu_data structures. This may be called
1417 * only from the CPU to whom the rdp belongs.
1420 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1422 unsigned long flags
;
1424 WARN_ON_ONCE(rdp
->beenonline
== 0);
1427 * If an RCU GP has gone long enough, go check for dyntick
1428 * idle CPUs and, if needed, send resched IPIs.
1430 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1431 force_quiescent_state(rsp
, 1);
1434 * Advance callbacks in response to end of earlier grace
1435 * period that some other CPU ended.
1437 rcu_process_gp_end(rsp
, rdp
);
1439 /* Update RCU state based on any recent quiescent states. */
1440 rcu_check_quiescent_state(rsp
, rdp
);
1442 /* Does this CPU require a not-yet-started grace period? */
1443 if (cpu_needs_another_gp(rsp
, rdp
)) {
1444 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1445 rcu_start_gp(rsp
, flags
); /* releases above lock */
1448 /* If there are callbacks ready, invoke them. */
1449 rcu_do_batch(rsp
, rdp
);
1453 * Do softirq processing for the current CPU.
1455 static void rcu_process_callbacks(void)
1458 * Memory references from any prior RCU read-side critical sections
1459 * executed by the interrupted code must be seen before any RCU
1460 * grace-period manipulations below.
1462 smp_mb(); /* See above block comment. */
1464 __rcu_process_callbacks(&rcu_sched_state
,
1465 &__get_cpu_var(rcu_sched_data
));
1466 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1467 rcu_preempt_process_callbacks();
1470 * Memory references from any later RCU read-side critical sections
1471 * executed by the interrupted code must be seen after any RCU
1472 * grace-period manipulations above.
1474 smp_mb(); /* See above block comment. */
1476 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1477 rcu_needs_cpu_flush();
1481 * Wake up the current CPU's kthread. This replaces raise_softirq()
1482 * in earlier versions of RCU. Note that because we are running on
1483 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1484 * cannot disappear out from under us.
1486 static void invoke_rcu_cpu_kthread(void)
1488 unsigned long flags
;
1490 local_irq_save(flags
);
1491 __this_cpu_write(rcu_cpu_has_work
, 1);
1492 if (__this_cpu_read(rcu_cpu_kthread_task
) == NULL
) {
1493 local_irq_restore(flags
);
1496 wake_up(&__get_cpu_var(rcu_cpu_wq
));
1497 local_irq_restore(flags
);
1501 * Wake up the specified per-rcu_node-structure kthread.
1502 * Because the per-rcu_node kthreads are immortal, we don't need
1503 * to do anything to keep them alive.
1505 static void invoke_rcu_node_kthread(struct rcu_node
*rnp
)
1507 struct task_struct
*t
;
1509 t
= rnp
->node_kthread_task
;
1515 * Set the specified CPU's kthread to run RT or not, as specified by
1516 * the to_rt argument. The CPU-hotplug locks are held, so the task
1517 * is not going away.
1519 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1522 struct sched_param sp
;
1523 struct task_struct
*t
;
1525 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1529 policy
= SCHED_FIFO
;
1530 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1532 policy
= SCHED_NORMAL
;
1533 sp
.sched_priority
= 0;
1535 sched_setscheduler_nocheck(t
, policy
, &sp
);
1539 * Timer handler to initiate the waking up of per-CPU kthreads that
1540 * have yielded the CPU due to excess numbers of RCU callbacks.
1541 * We wake up the per-rcu_node kthread, which in turn will wake up
1542 * the booster kthread.
1544 static void rcu_cpu_kthread_timer(unsigned long arg
)
1546 unsigned long flags
;
1547 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, arg
);
1548 struct rcu_node
*rnp
= rdp
->mynode
;
1550 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1551 rnp
->wakemask
|= rdp
->grpmask
;
1552 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1553 invoke_rcu_node_kthread(rnp
);
1557 * Drop to non-real-time priority and yield, but only after posting a
1558 * timer that will cause us to regain our real-time priority if we
1559 * remain preempted. Either way, we restore our real-time priority
1562 static void rcu_yield(void (*f
)(unsigned long), unsigned long arg
)
1564 struct sched_param sp
;
1565 struct timer_list yield_timer
;
1567 setup_timer_on_stack(&yield_timer
, f
, arg
);
1568 mod_timer(&yield_timer
, jiffies
+ 2);
1569 sp
.sched_priority
= 0;
1570 sched_setscheduler_nocheck(current
, SCHED_NORMAL
, &sp
);
1571 set_user_nice(current
, 19);
1573 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1574 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1575 del_timer(&yield_timer
);
1579 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1580 * This can happen while the corresponding CPU is either coming online
1581 * or going offline. We cannot wait until the CPU is fully online
1582 * before starting the kthread, because the various notifier functions
1583 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1584 * the corresponding CPU is online.
1586 * Return 1 if the kthread needs to stop, 0 otherwise.
1588 * Caller must disable bh. This function can momentarily enable it.
1590 static int rcu_cpu_kthread_should_stop(int cpu
)
1592 while (cpu_is_offline(cpu
) ||
1593 !cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)) ||
1594 smp_processor_id() != cpu
) {
1595 if (kthread_should_stop())
1597 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1598 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = raw_smp_processor_id();
1600 schedule_timeout_uninterruptible(1);
1601 if (!cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)))
1602 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
1605 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1610 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1611 * earlier RCU softirq.
1613 static int rcu_cpu_kthread(void *arg
)
1615 int cpu
= (int)(long)arg
;
1616 unsigned long flags
;
1618 unsigned int *statusp
= &per_cpu(rcu_cpu_kthread_status
, cpu
);
1619 wait_queue_head_t
*wqp
= &per_cpu(rcu_cpu_wq
, cpu
);
1621 char *workp
= &per_cpu(rcu_cpu_has_work
, cpu
);
1624 *statusp
= RCU_KTHREAD_WAITING
;
1625 wait_event_interruptible(*wqp
,
1626 *workp
!= 0 || kthread_should_stop());
1628 if (rcu_cpu_kthread_should_stop(cpu
)) {
1632 *statusp
= RCU_KTHREAD_RUNNING
;
1633 per_cpu(rcu_cpu_kthread_loops
, cpu
)++;
1634 local_irq_save(flags
);
1637 local_irq_restore(flags
);
1639 rcu_process_callbacks();
1646 *statusp
= RCU_KTHREAD_YIELDING
;
1647 rcu_yield(rcu_cpu_kthread_timer
, (unsigned long)cpu
);
1651 *statusp
= RCU_KTHREAD_STOPPED
;
1656 * Spawn a per-CPU kthread, setting up affinity and priority.
1657 * Because the CPU hotplug lock is held, no other CPU will be attempting
1658 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1659 * attempting to access it during boot, but the locking in kthread_bind()
1660 * will enforce sufficient ordering.
1662 static int __cpuinit
rcu_spawn_one_cpu_kthread(int cpu
)
1664 struct sched_param sp
;
1665 struct task_struct
*t
;
1667 if (!rcu_kthreads_spawnable
||
1668 per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
)
1670 t
= kthread_create(rcu_cpu_kthread
, (void *)(long)cpu
, "rcuc%d", cpu
);
1673 kthread_bind(t
, cpu
);
1674 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1675 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
);
1676 per_cpu(rcu_cpu_kthread_task
, cpu
) = t
;
1678 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1679 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1684 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1685 * kthreads when needed. We ignore requests to wake up kthreads
1686 * for offline CPUs, which is OK because force_quiescent_state()
1687 * takes care of this case.
1689 static int rcu_node_kthread(void *arg
)
1692 unsigned long flags
;
1694 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1695 struct sched_param sp
;
1696 struct task_struct
*t
;
1699 rnp
->node_kthread_status
= RCU_KTHREAD_WAITING
;
1700 wait_event_interruptible(rnp
->node_wq
, rnp
->wakemask
!= 0);
1701 rnp
->node_kthread_status
= RCU_KTHREAD_RUNNING
;
1702 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1703 mask
= rnp
->wakemask
;
1705 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1706 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1) {
1707 if ((mask
& 0x1) == 0)
1710 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1711 if (!cpu_online(cpu
) || t
== NULL
) {
1715 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1716 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1717 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1722 rnp
->node_kthread_status
= RCU_KTHREAD_STOPPED
;
1727 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1728 * served by the rcu_node in question. The CPU hotplug lock is still
1729 * held, so the value of rnp->qsmaskinit will be stable.
1731 * We don't include outgoingcpu in the affinity set, use -1 if there is
1732 * no outgoing CPU. If there are no CPUs left in the affinity set,
1733 * this function allows the kthread to execute on any CPU.
1735 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1739 unsigned long mask
= rnp
->qsmaskinit
;
1741 if (rnp
->node_kthread_task
== NULL
)
1743 if (!alloc_cpumask_var(&cm
, GFP_KERNEL
))
1746 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1747 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1748 cpumask_set_cpu(cpu
, cm
);
1749 if (cpumask_weight(cm
) == 0) {
1751 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1752 cpumask_clear_cpu(cpu
, cm
);
1753 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1755 set_cpus_allowed_ptr(rnp
->node_kthread_task
, cm
);
1756 rcu_boost_kthread_setaffinity(rnp
, cm
);
1757 free_cpumask_var(cm
);
1761 * Spawn a per-rcu_node kthread, setting priority and affinity.
1762 * Called during boot before online/offline can happen, or, if
1763 * during runtime, with the main CPU-hotplug locks held. So only
1764 * one of these can be executing at a time.
1766 static int __cpuinit
rcu_spawn_one_node_kthread(struct rcu_state
*rsp
,
1767 struct rcu_node
*rnp
)
1769 unsigned long flags
;
1770 int rnp_index
= rnp
- &rsp
->node
[0];
1771 struct sched_param sp
;
1772 struct task_struct
*t
;
1774 if (!rcu_kthreads_spawnable
||
1775 rnp
->qsmaskinit
== 0)
1777 if (rnp
->node_kthread_task
== NULL
) {
1778 t
= kthread_create(rcu_node_kthread
, (void *)rnp
,
1779 "rcun%d", rnp_index
);
1782 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1783 rnp
->node_kthread_task
= t
;
1784 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1786 sp
.sched_priority
= 99;
1787 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1789 return rcu_spawn_one_boost_kthread(rsp
, rnp
, rnp_index
);
1793 * Spawn all kthreads -- called as soon as the scheduler is running.
1795 static int __init
rcu_spawn_kthreads(void)
1798 struct rcu_node
*rnp
;
1800 rcu_kthreads_spawnable
= 1;
1801 for_each_possible_cpu(cpu
) {
1802 init_waitqueue_head(&per_cpu(rcu_cpu_wq
, cpu
));
1803 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1804 if (cpu_online(cpu
))
1805 (void)rcu_spawn_one_cpu_kthread(cpu
);
1807 rnp
= rcu_get_root(rcu_state
);
1808 init_waitqueue_head(&rnp
->node_wq
);
1809 rcu_init_boost_waitqueue(rnp
);
1810 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1811 if (NUM_RCU_NODES
> 1)
1812 rcu_for_each_leaf_node(rcu_state
, rnp
) {
1813 init_waitqueue_head(&rnp
->node_wq
);
1814 rcu_init_boost_waitqueue(rnp
);
1815 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1819 early_initcall(rcu_spawn_kthreads
);
1822 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1823 struct rcu_state
*rsp
)
1825 unsigned long flags
;
1826 struct rcu_data
*rdp
;
1828 debug_rcu_head_queue(head
);
1832 smp_mb(); /* Ensure RCU update seen before callback registry. */
1835 * Opportunistically note grace-period endings and beginnings.
1836 * Note that we might see a beginning right after we see an
1837 * end, but never vice versa, since this CPU has to pass through
1838 * a quiescent state betweentimes.
1840 local_irq_save(flags
);
1841 rdp
= this_cpu_ptr(rsp
->rda
);
1843 /* Add the callback to our list. */
1844 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1845 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1848 /* If interrupts were disabled, don't dive into RCU core. */
1849 if (irqs_disabled_flags(flags
)) {
1850 local_irq_restore(flags
);
1855 * Force the grace period if too many callbacks or too long waiting.
1856 * Enforce hysteresis, and don't invoke force_quiescent_state()
1857 * if some other CPU has recently done so. Also, don't bother
1858 * invoking force_quiescent_state() if the newly enqueued callback
1859 * is the only one waiting for a grace period to complete.
1861 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1863 /* Are we ignoring a completed grace period? */
1864 rcu_process_gp_end(rsp
, rdp
);
1865 check_for_new_grace_period(rsp
, rdp
);
1867 /* Start a new grace period if one not already started. */
1868 if (!rcu_gp_in_progress(rsp
)) {
1869 unsigned long nestflag
;
1870 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1872 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1873 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1875 /* Give the grace period a kick. */
1876 rdp
->blimit
= LONG_MAX
;
1877 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1878 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1879 force_quiescent_state(rsp
, 0);
1880 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1881 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1883 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1884 force_quiescent_state(rsp
, 1);
1885 local_irq_restore(flags
);
1889 * Queue an RCU-sched callback for invocation after a grace period.
1891 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1893 __call_rcu(head
, func
, &rcu_sched_state
);
1895 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1898 * Queue an RCU for invocation after a quicker grace period.
1900 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1902 __call_rcu(head
, func
, &rcu_bh_state
);
1904 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1907 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1909 * Control will return to the caller some time after a full rcu-sched
1910 * grace period has elapsed, in other words after all currently executing
1911 * rcu-sched read-side critical sections have completed. These read-side
1912 * critical sections are delimited by rcu_read_lock_sched() and
1913 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1914 * local_irq_disable(), and so on may be used in place of
1915 * rcu_read_lock_sched().
1917 * This means that all preempt_disable code sequences, including NMI and
1918 * hardware-interrupt handlers, in progress on entry will have completed
1919 * before this primitive returns. However, this does not guarantee that
1920 * softirq handlers will have completed, since in some kernels, these
1921 * handlers can run in process context, and can block.
1923 * This primitive provides the guarantees made by the (now removed)
1924 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1925 * guarantees that rcu_read_lock() sections will have completed.
1926 * In "classic RCU", these two guarantees happen to be one and
1927 * the same, but can differ in realtime RCU implementations.
1929 void synchronize_sched(void)
1931 struct rcu_synchronize rcu
;
1933 if (rcu_blocking_is_gp())
1936 init_rcu_head_on_stack(&rcu
.head
);
1937 init_completion(&rcu
.completion
);
1938 /* Will wake me after RCU finished. */
1939 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1941 wait_for_completion(&rcu
.completion
);
1942 destroy_rcu_head_on_stack(&rcu
.head
);
1944 EXPORT_SYMBOL_GPL(synchronize_sched
);
1947 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1949 * Control will return to the caller some time after a full rcu_bh grace
1950 * period has elapsed, in other words after all currently executing rcu_bh
1951 * read-side critical sections have completed. RCU read-side critical
1952 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1953 * and may be nested.
1955 void synchronize_rcu_bh(void)
1957 struct rcu_synchronize rcu
;
1959 if (rcu_blocking_is_gp())
1962 init_rcu_head_on_stack(&rcu
.head
);
1963 init_completion(&rcu
.completion
);
1964 /* Will wake me after RCU finished. */
1965 call_rcu_bh(&rcu
.head
, wakeme_after_rcu
);
1967 wait_for_completion(&rcu
.completion
);
1968 destroy_rcu_head_on_stack(&rcu
.head
);
1970 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1973 * Check to see if there is any immediate RCU-related work to be done
1974 * by the current CPU, for the specified type of RCU, returning 1 if so.
1975 * The checks are in order of increasing expense: checks that can be
1976 * carried out against CPU-local state are performed first. However,
1977 * we must check for CPU stalls first, else we might not get a chance.
1979 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1981 struct rcu_node
*rnp
= rdp
->mynode
;
1983 rdp
->n_rcu_pending
++;
1985 /* Check for CPU stalls, if enabled. */
1986 check_cpu_stall(rsp
, rdp
);
1988 /* Is the RCU core waiting for a quiescent state from this CPU? */
1989 if (rdp
->qs_pending
&& !rdp
->passed_quiesc
) {
1992 * If force_quiescent_state() coming soon and this CPU
1993 * needs a quiescent state, and this is either RCU-sched
1994 * or RCU-bh, force a local reschedule.
1996 rdp
->n_rp_qs_pending
++;
1997 if (!rdp
->preemptible
&&
1998 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
2001 } else if (rdp
->qs_pending
&& rdp
->passed_quiesc
) {
2002 rdp
->n_rp_report_qs
++;
2006 /* Does this CPU have callbacks ready to invoke? */
2007 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2008 rdp
->n_rp_cb_ready
++;
2012 /* Has RCU gone idle with this CPU needing another grace period? */
2013 if (cpu_needs_another_gp(rsp
, rdp
)) {
2014 rdp
->n_rp_cpu_needs_gp
++;
2018 /* Has another RCU grace period completed? */
2019 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2020 rdp
->n_rp_gp_completed
++;
2024 /* Has a new RCU grace period started? */
2025 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2026 rdp
->n_rp_gp_started
++;
2030 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2031 if (rcu_gp_in_progress(rsp
) &&
2032 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
2033 rdp
->n_rp_need_fqs
++;
2038 rdp
->n_rp_need_nothing
++;
2043 * Check to see if there is any immediate RCU-related work to be done
2044 * by the current CPU, returning 1 if so. This function is part of the
2045 * RCU implementation; it is -not- an exported member of the RCU API.
2047 static int rcu_pending(int cpu
)
2049 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
2050 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
2051 rcu_preempt_pending(cpu
);
2055 * Check to see if any future RCU-related work will need to be done
2056 * by the current CPU, even if none need be done immediately, returning
2059 static int rcu_needs_cpu_quick_check(int cpu
)
2061 /* RCU callbacks either ready or pending? */
2062 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
2063 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
2064 rcu_preempt_needs_cpu(cpu
);
2067 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
2068 static atomic_t rcu_barrier_cpu_count
;
2069 static DEFINE_MUTEX(rcu_barrier_mutex
);
2070 static struct completion rcu_barrier_completion
;
2072 static void rcu_barrier_callback(struct rcu_head
*notused
)
2074 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2075 complete(&rcu_barrier_completion
);
2079 * Called with preemption disabled, and from cross-cpu IRQ context.
2081 static void rcu_barrier_func(void *type
)
2083 int cpu
= smp_processor_id();
2084 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
2085 void (*call_rcu_func
)(struct rcu_head
*head
,
2086 void (*func
)(struct rcu_head
*head
));
2088 atomic_inc(&rcu_barrier_cpu_count
);
2089 call_rcu_func
= type
;
2090 call_rcu_func(head
, rcu_barrier_callback
);
2094 * Orchestrate the specified type of RCU barrier, waiting for all
2095 * RCU callbacks of the specified type to complete.
2097 static void _rcu_barrier(struct rcu_state
*rsp
,
2098 void (*call_rcu_func
)(struct rcu_head
*head
,
2099 void (*func
)(struct rcu_head
*head
)))
2101 BUG_ON(in_interrupt());
2102 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2103 mutex_lock(&rcu_barrier_mutex
);
2104 init_completion(&rcu_barrier_completion
);
2106 * Initialize rcu_barrier_cpu_count to 1, then invoke
2107 * rcu_barrier_func() on each CPU, so that each CPU also has
2108 * incremented rcu_barrier_cpu_count. Only then is it safe to
2109 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2110 * might complete its grace period before all of the other CPUs
2111 * did their increment, causing this function to return too
2112 * early. Note that on_each_cpu() disables irqs, which prevents
2113 * any CPUs from coming online or going offline until each online
2114 * CPU has queued its RCU-barrier callback.
2116 atomic_set(&rcu_barrier_cpu_count
, 1);
2117 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
2118 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2119 complete(&rcu_barrier_completion
);
2120 wait_for_completion(&rcu_barrier_completion
);
2121 mutex_unlock(&rcu_barrier_mutex
);
2125 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2127 void rcu_barrier_bh(void)
2129 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
2131 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2134 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2136 void rcu_barrier_sched(void)
2138 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2140 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2143 * Do boot-time initialization of a CPU's per-CPU RCU data.
2146 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2148 unsigned long flags
;
2150 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2151 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2153 /* Set up local state, ensuring consistent view of global state. */
2154 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2155 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2156 rdp
->nxtlist
= NULL
;
2157 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2158 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2161 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2162 #endif /* #ifdef CONFIG_NO_HZ */
2164 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2168 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2169 * offline event can be happening at a given time. Note also that we
2170 * can accept some slop in the rsp->completed access due to the fact
2171 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2173 static void __cpuinit
2174 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2176 unsigned long flags
;
2178 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2179 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2181 /* Set up local state, ensuring consistent view of global state. */
2182 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2183 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
2184 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
2185 rdp
->beenonline
= 1; /* We have now been online. */
2186 rdp
->preemptible
= preemptible
;
2187 rdp
->qlen_last_fqs_check
= 0;
2188 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2189 rdp
->blimit
= blimit
;
2190 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2193 * A new grace period might start here. If so, we won't be part
2194 * of it, but that is OK, as we are currently in a quiescent state.
2197 /* Exclude any attempts to start a new GP on large systems. */
2198 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2200 /* Add CPU to rcu_node bitmasks. */
2202 mask
= rdp
->grpmask
;
2204 /* Exclude any attempts to start a new GP on small systems. */
2205 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2206 rnp
->qsmaskinit
|= mask
;
2207 mask
= rnp
->grpmask
;
2208 if (rnp
== rdp
->mynode
) {
2209 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
2210 rdp
->completed
= rnp
->completed
;
2211 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
2213 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2215 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2217 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2220 static void __cpuinit
rcu_online_cpu(int cpu
)
2222 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2223 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2224 rcu_preempt_init_percpu_data(cpu
);
2227 static void __cpuinit
rcu_online_kthreads(int cpu
)
2229 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2230 struct rcu_node
*rnp
= rdp
->mynode
;
2232 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
2233 if (rcu_kthreads_spawnable
) {
2234 (void)rcu_spawn_one_cpu_kthread(cpu
);
2235 if (rnp
->node_kthread_task
== NULL
)
2236 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
2241 * Handle CPU online/offline notification events.
2243 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2244 unsigned long action
, void *hcpu
)
2246 long cpu
= (long)hcpu
;
2247 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2248 struct rcu_node
*rnp
= rdp
->mynode
;
2251 case CPU_UP_PREPARE
:
2252 case CPU_UP_PREPARE_FROZEN
:
2253 rcu_online_cpu(cpu
);
2254 rcu_online_kthreads(cpu
);
2257 case CPU_DOWN_FAILED
:
2258 rcu_node_kthread_setaffinity(rnp
, -1);
2259 rcu_cpu_kthread_setrt(cpu
, 1);
2261 case CPU_DOWN_PREPARE
:
2262 rcu_node_kthread_setaffinity(rnp
, cpu
);
2263 rcu_cpu_kthread_setrt(cpu
, 0);
2266 case CPU_DYING_FROZEN
:
2268 * The whole machine is "stopped" except this CPU, so we can
2269 * touch any data without introducing corruption. We send the
2270 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2272 rcu_send_cbs_to_online(&rcu_bh_state
);
2273 rcu_send_cbs_to_online(&rcu_sched_state
);
2274 rcu_preempt_send_cbs_to_online();
2277 case CPU_DEAD_FROZEN
:
2278 case CPU_UP_CANCELED
:
2279 case CPU_UP_CANCELED_FROZEN
:
2280 rcu_offline_cpu(cpu
);
2289 * This function is invoked towards the end of the scheduler's initialization
2290 * process. Before this is called, the idle task might contain
2291 * RCU read-side critical sections (during which time, this idle
2292 * task is booting the system). After this function is called, the
2293 * idle tasks are prohibited from containing RCU read-side critical
2294 * sections. This function also enables RCU lockdep checking.
2296 void rcu_scheduler_starting(void)
2298 WARN_ON(num_online_cpus() != 1);
2299 WARN_ON(nr_context_switches() > 0);
2300 rcu_scheduler_active
= 1;
2304 * Compute the per-level fanout, either using the exact fanout specified
2305 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2307 #ifdef CONFIG_RCU_FANOUT_EXACT
2308 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2312 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2313 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2314 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
2316 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2317 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2324 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2325 ccur
= rsp
->levelcnt
[i
];
2326 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2330 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2333 * Helper function for rcu_init() that initializes one rcu_state structure.
2335 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2336 struct rcu_data __percpu
*rda
)
2338 static char *buf
[] = { "rcu_node_level_0",
2341 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2345 struct rcu_node
*rnp
;
2347 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2349 /* Initialize the level-tracking arrays. */
2351 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2352 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2353 rcu_init_levelspread(rsp
);
2355 /* Initialize the elements themselves, starting from the leaves. */
2357 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2358 cpustride
*= rsp
->levelspread
[i
];
2359 rnp
= rsp
->level
[i
];
2360 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2361 raw_spin_lock_init(&rnp
->lock
);
2362 lockdep_set_class_and_name(&rnp
->lock
,
2363 &rcu_node_class
[i
], buf
[i
]);
2366 rnp
->qsmaskinit
= 0;
2367 rnp
->grplo
= j
* cpustride
;
2368 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2369 if (rnp
->grphi
>= NR_CPUS
)
2370 rnp
->grphi
= NR_CPUS
- 1;
2376 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2377 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2378 rnp
->parent
= rsp
->level
[i
- 1] +
2379 j
/ rsp
->levelspread
[i
- 1];
2382 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2387 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2388 for_each_possible_cpu(i
) {
2389 while (i
> rnp
->grphi
)
2391 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2392 rcu_boot_init_percpu_data(i
, rsp
);
2396 void __init
rcu_init(void)
2400 rcu_bootup_announce();
2401 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2402 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2403 __rcu_init_preempt();
2406 * We don't need protection against CPU-hotplug here because
2407 * this is called early in boot, before either interrupts
2408 * or the scheduler are operational.
2410 cpu_notifier(rcu_cpu_notify
, 0);
2411 for_each_online_cpu(cpu
)
2412 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2413 check_cpu_stall_init();
2416 #include "rcutree_plugin.h"