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>
53 /* Data structures. */
55 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
57 #define RCU_STATE_INITIALIZER(structname) { \
58 .level = { &structname.node[0] }, \
60 NUM_RCU_LVL_0, /* root of hierarchy. */ \
64 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
66 .signaled = RCU_GP_IDLE, \
69 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
70 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
72 .n_force_qs_ngp = 0, \
73 .name = #structname, \
76 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched_state
);
77 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
79 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh_state
);
80 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
82 int rcu_scheduler_active __read_mostly
;
83 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
86 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
87 * permit this function to be invoked without holding the root rcu_node
88 * structure's ->lock, but of course results can be subject to change.
90 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
92 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
96 * Note a quiescent state. Because we do not need to know
97 * how many quiescent states passed, just if there was at least
98 * one since the start of the grace period, this just sets a flag.
100 void rcu_sched_qs(int cpu
)
102 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
104 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
106 rdp
->passed_quiesc
= 1;
109 void rcu_bh_qs(int cpu
)
111 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
113 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
115 rdp
->passed_quiesc
= 1;
119 * Note a context switch. This is a quiescent state for RCU-sched,
120 * and requires special handling for preemptible RCU.
122 void rcu_note_context_switch(int cpu
)
125 rcu_preempt_note_context_switch(cpu
);
129 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
130 .dynticks_nesting
= 1,
133 #endif /* #ifdef CONFIG_NO_HZ */
135 static int blimit
= 10; /* Maximum callbacks per softirq. */
136 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
137 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
139 module_param(blimit
, int, 0);
140 module_param(qhimark
, int, 0);
141 module_param(qlowmark
, int, 0);
143 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
144 int rcu_cpu_stall_suppress __read_mostly
= RCU_CPU_STALL_SUPPRESS_INIT
;
145 module_param(rcu_cpu_stall_suppress
, int, 0644);
146 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
148 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
149 static int rcu_pending(int cpu
);
152 * Return the number of RCU-sched batches processed thus far for debug & stats.
154 long rcu_batches_completed_sched(void)
156 return rcu_sched_state
.completed
;
158 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
161 * Return the number of RCU BH batches processed thus far for debug & stats.
163 long rcu_batches_completed_bh(void)
165 return rcu_bh_state
.completed
;
167 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
170 * Force a quiescent state for RCU BH.
172 void rcu_bh_force_quiescent_state(void)
174 force_quiescent_state(&rcu_bh_state
, 0);
176 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
179 * Force a quiescent state for RCU-sched.
181 void rcu_sched_force_quiescent_state(void)
183 force_quiescent_state(&rcu_sched_state
, 0);
185 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
188 * Does the CPU have callbacks ready to be invoked?
191 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
193 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
197 * Does the current CPU require a yet-as-unscheduled grace period?
200 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
202 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
206 * Return the root node of the specified rcu_state structure.
208 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
210 return &rsp
->node
[0];
216 * If the specified CPU is offline, tell the caller that it is in
217 * a quiescent state. Otherwise, whack it with a reschedule IPI.
218 * Grace periods can end up waiting on an offline CPU when that
219 * CPU is in the process of coming online -- it will be added to the
220 * rcu_node bitmasks before it actually makes it online. The same thing
221 * can happen while a CPU is in the process of coming online. Because this
222 * race is quite rare, we check for it after detecting that the grace
223 * period has been delayed rather than checking each and every CPU
224 * each and every time we start a new grace period.
226 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
229 * If the CPU is offline, it is in a quiescent state. We can
230 * trust its state not to change because interrupts are disabled.
232 if (cpu_is_offline(rdp
->cpu
)) {
237 /* If preemptable RCU, no point in sending reschedule IPI. */
238 if (rdp
->preemptable
)
241 /* The CPU is online, so send it a reschedule IPI. */
242 if (rdp
->cpu
!= smp_processor_id())
243 smp_send_reschedule(rdp
->cpu
);
250 #endif /* #ifdef CONFIG_SMP */
255 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
257 * Enter nohz mode, in other words, -leave- the mode in which RCU
258 * read-side critical sections can occur. (Though RCU read-side
259 * critical sections can occur in irq handlers in nohz mode, a possibility
260 * handled by rcu_irq_enter() and rcu_irq_exit()).
262 void rcu_enter_nohz(void)
265 struct rcu_dynticks
*rdtp
;
267 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
268 local_irq_save(flags
);
269 rdtp
= &__get_cpu_var(rcu_dynticks
);
271 rdtp
->dynticks_nesting
--;
272 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
273 local_irq_restore(flags
);
277 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
279 * Exit nohz mode, in other words, -enter- the mode in which RCU
280 * read-side critical sections normally occur.
282 void rcu_exit_nohz(void)
285 struct rcu_dynticks
*rdtp
;
287 local_irq_save(flags
);
288 rdtp
= &__get_cpu_var(rcu_dynticks
);
290 rdtp
->dynticks_nesting
++;
291 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
292 local_irq_restore(flags
);
293 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
297 * rcu_nmi_enter - inform RCU of entry to NMI context
299 * If the CPU was idle with dynamic ticks active, and there is no
300 * irq handler running, this updates rdtp->dynticks_nmi to let the
301 * RCU grace-period handling know that the CPU is active.
303 void rcu_nmi_enter(void)
305 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
307 if (rdtp
->dynticks
& 0x1)
309 rdtp
->dynticks_nmi
++;
310 WARN_ON_ONCE(!(rdtp
->dynticks_nmi
& 0x1));
311 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
315 * rcu_nmi_exit - inform RCU of exit from NMI context
317 * If the CPU was idle with dynamic ticks active, and there is no
318 * irq handler running, this updates rdtp->dynticks_nmi to let the
319 * RCU grace-period handling know that the CPU is no longer active.
321 void rcu_nmi_exit(void)
323 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
325 if (rdtp
->dynticks
& 0x1)
327 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
328 rdtp
->dynticks_nmi
++;
329 WARN_ON_ONCE(rdtp
->dynticks_nmi
& 0x1);
333 * rcu_irq_enter - inform RCU of entry to hard irq context
335 * If the CPU was idle with dynamic ticks active, this updates the
336 * rdtp->dynticks to let the RCU handling know that the CPU is active.
338 void rcu_irq_enter(void)
340 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
342 if (rdtp
->dynticks_nesting
++)
345 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
346 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
350 * rcu_irq_exit - inform RCU of exit from hard irq context
352 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
353 * to put let the RCU handling be aware that the CPU is going back to idle
356 void rcu_irq_exit(void)
358 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
360 if (--rdtp
->dynticks_nesting
)
362 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
364 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
366 /* If the interrupt queued a callback, get out of dyntick mode. */
367 if (__this_cpu_read(rcu_sched_data
.nxtlist
) ||
368 __this_cpu_read(rcu_bh_data
.nxtlist
))
375 * Snapshot the specified CPU's dynticks counter so that we can later
376 * credit them with an implicit quiescent state. Return 1 if this CPU
377 * is in dynticks idle mode, which is an extended quiescent state.
379 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
385 snap
= rdp
->dynticks
->dynticks
;
386 snap_nmi
= rdp
->dynticks
->dynticks_nmi
;
387 smp_mb(); /* Order sampling of snap with end of grace period. */
388 rdp
->dynticks_snap
= snap
;
389 rdp
->dynticks_nmi_snap
= snap_nmi
;
390 ret
= ((snap
& 0x1) == 0) && ((snap_nmi
& 0x1) == 0);
397 * Return true if the specified CPU has passed through a quiescent
398 * state by virtue of being in or having passed through an dynticks
399 * idle state since the last call to dyntick_save_progress_counter()
402 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
409 curr
= rdp
->dynticks
->dynticks
;
410 snap
= rdp
->dynticks_snap
;
411 curr_nmi
= rdp
->dynticks
->dynticks_nmi
;
412 snap_nmi
= rdp
->dynticks_nmi_snap
;
413 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
416 * If the CPU passed through or entered a dynticks idle phase with
417 * no active irq/NMI handlers, then we can safely pretend that the CPU
418 * already acknowledged the request to pass through a quiescent
419 * state. Either way, that CPU cannot possibly be in an RCU
420 * read-side critical section that started before the beginning
421 * of the current RCU grace period.
423 if ((curr
!= snap
|| (curr
& 0x1) == 0) &&
424 (curr_nmi
!= snap_nmi
|| (curr_nmi
& 0x1) == 0)) {
429 /* Go check for the CPU being offline. */
430 return rcu_implicit_offline_qs(rdp
);
433 #endif /* #ifdef CONFIG_SMP */
435 #else /* #ifdef CONFIG_NO_HZ */
439 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
444 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
446 return rcu_implicit_offline_qs(rdp
);
449 #endif /* #ifdef CONFIG_SMP */
451 #endif /* #else #ifdef CONFIG_NO_HZ */
453 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
455 int rcu_cpu_stall_suppress __read_mostly
;
457 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
459 rsp
->gp_start
= jiffies
;
460 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
463 static void print_other_cpu_stall(struct rcu_state
*rsp
)
468 struct rcu_node
*rnp
= rcu_get_root(rsp
);
470 /* Only let one CPU complain about others per time interval. */
472 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
473 delta
= jiffies
- rsp
->jiffies_stall
;
474 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
475 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
478 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
481 * Now rat on any tasks that got kicked up to the root rcu_node
482 * due to CPU offlining.
484 rcu_print_task_stall(rnp
);
485 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
488 * OK, time to rat on our buddy...
489 * See Documentation/RCU/stallwarn.txt for info on how to debug
490 * RCU CPU stall warnings.
492 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
494 rcu_for_each_leaf_node(rsp
, rnp
) {
495 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
496 rcu_print_task_stall(rnp
);
497 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
498 if (rnp
->qsmask
== 0)
500 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
501 if (rnp
->qsmask
& (1UL << cpu
))
502 printk(" %d", rnp
->grplo
+ cpu
);
504 printk("} (detected by %d, t=%ld jiffies)\n",
505 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
506 trigger_all_cpu_backtrace();
508 /* If so configured, complain about tasks blocking the grace period. */
510 rcu_print_detail_task_stall(rsp
);
512 force_quiescent_state(rsp
, 0); /* Kick them all. */
515 static void print_cpu_stall(struct rcu_state
*rsp
)
518 struct rcu_node
*rnp
= rcu_get_root(rsp
);
521 * OK, time to rat on ourselves...
522 * See Documentation/RCU/stallwarn.txt for info on how to debug
523 * RCU CPU stall warnings.
525 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
526 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
527 trigger_all_cpu_backtrace();
529 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
530 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
532 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
533 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
535 set_need_resched(); /* kick ourselves to get things going. */
538 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
541 struct rcu_node
*rnp
;
543 if (rcu_cpu_stall_suppress
)
545 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
547 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && delta
>= 0) {
549 /* We haven't checked in, so go dump stack. */
550 print_cpu_stall(rsp
);
552 } else if (rcu_gp_in_progress(rsp
) && delta
>= RCU_STALL_RAT_DELAY
) {
554 /* They had two time units to dump stack, so complain. */
555 print_other_cpu_stall(rsp
);
559 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
561 rcu_cpu_stall_suppress
= 1;
566 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
568 * Set the stall-warning timeout way off into the future, thus preventing
569 * any RCU CPU stall-warning messages from appearing in the current set of
572 * The caller must disable hard irqs.
574 void rcu_cpu_stall_reset(void)
576 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
577 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
578 rcu_preempt_stall_reset();
581 static struct notifier_block rcu_panic_block
= {
582 .notifier_call
= rcu_panic
,
585 static void __init
check_cpu_stall_init(void)
587 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
590 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
592 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
596 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
600 void rcu_cpu_stall_reset(void)
604 static void __init
check_cpu_stall_init(void)
608 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
611 * Update CPU-local rcu_data state to record the newly noticed grace period.
612 * This is used both when we started the grace period and when we notice
613 * that someone else started the grace period. The caller must hold the
614 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
615 * and must have irqs disabled.
617 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
619 if (rdp
->gpnum
!= rnp
->gpnum
) {
621 * If the current grace period is waiting for this CPU,
622 * set up to detect a quiescent state, otherwise don't
623 * go looking for one.
625 rdp
->gpnum
= rnp
->gpnum
;
626 if (rnp
->qsmask
& rdp
->grpmask
) {
628 rdp
->passed_quiesc
= 0;
634 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
637 struct rcu_node
*rnp
;
639 local_irq_save(flags
);
641 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
642 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
643 local_irq_restore(flags
);
646 __note_new_gpnum(rsp
, rnp
, rdp
);
647 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
651 * Did someone else start a new RCU grace period start since we last
652 * checked? Update local state appropriately if so. Must be called
653 * on the CPU corresponding to rdp.
656 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
661 local_irq_save(flags
);
662 if (rdp
->gpnum
!= rsp
->gpnum
) {
663 note_new_gpnum(rsp
, rdp
);
666 local_irq_restore(flags
);
671 * Advance this CPU's callbacks, but only if the current grace period
672 * has ended. This may be called only from the CPU to whom the rdp
673 * belongs. In addition, the corresponding leaf rcu_node structure's
674 * ->lock must be held by the caller, with irqs disabled.
677 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
679 /* Did another grace period end? */
680 if (rdp
->completed
!= rnp
->completed
) {
682 /* Advance callbacks. No harm if list empty. */
683 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
684 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
685 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
687 /* Remember that we saw this grace-period completion. */
688 rdp
->completed
= rnp
->completed
;
691 * If we were in an extended quiescent state, we may have
692 * missed some grace periods that others CPUs handled on
693 * our behalf. Catch up with this state to avoid noting
694 * spurious new grace periods. If another grace period
695 * has started, then rnp->gpnum will have advanced, so
696 * we will detect this later on.
698 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
699 rdp
->gpnum
= rdp
->completed
;
702 * If RCU does not need a quiescent state from this CPU,
703 * then make sure that this CPU doesn't go looking for one.
705 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
711 * Advance this CPU's callbacks, but only if the current grace period
712 * has ended. This may be called only from the CPU to whom the rdp
716 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
719 struct rcu_node
*rnp
;
721 local_irq_save(flags
);
723 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
724 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
725 local_irq_restore(flags
);
728 __rcu_process_gp_end(rsp
, rnp
, rdp
);
729 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
733 * Do per-CPU grace-period initialization for running CPU. The caller
734 * must hold the lock of the leaf rcu_node structure corresponding to
738 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
740 /* Prior grace period ended, so advance callbacks for current CPU. */
741 __rcu_process_gp_end(rsp
, rnp
, rdp
);
744 * Because this CPU just now started the new grace period, we know
745 * that all of its callbacks will be covered by this upcoming grace
746 * period, even the ones that were registered arbitrarily recently.
747 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
749 * Other CPUs cannot be sure exactly when the grace period started.
750 * Therefore, their recently registered callbacks must pass through
751 * an additional RCU_NEXT_READY stage, so that they will be handled
752 * by the next RCU grace period.
754 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
755 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
757 /* Set state so that this CPU will detect the next quiescent state. */
758 __note_new_gpnum(rsp
, rnp
, rdp
);
762 * Start a new RCU grace period if warranted, re-initializing the hierarchy
763 * in preparation for detecting the next grace period. The caller must hold
764 * the root node's ->lock, which is released before return. Hard irqs must
768 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
769 __releases(rcu_get_root(rsp
)->lock
)
771 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
772 struct rcu_node
*rnp
= rcu_get_root(rsp
);
774 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
775 if (cpu_needs_another_gp(rsp
, rdp
))
776 rsp
->fqs_need_gp
= 1;
777 if (rnp
->completed
== rsp
->completed
) {
778 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
781 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
784 * Propagate new ->completed value to rcu_node structures
785 * so that other CPUs don't have to wait until the start
786 * of the next grace period to process their callbacks.
788 rcu_for_each_node_breadth_first(rsp
, rnp
) {
789 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
790 rnp
->completed
= rsp
->completed
;
791 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
793 local_irq_restore(flags
);
797 /* Advance to a new grace period and initialize state. */
799 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
800 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
801 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
802 record_gp_stall_check_time(rsp
);
804 /* Special-case the common single-level case. */
805 if (NUM_RCU_NODES
== 1) {
806 rcu_preempt_check_blocked_tasks(rnp
);
807 rnp
->qsmask
= rnp
->qsmaskinit
;
808 rnp
->gpnum
= rsp
->gpnum
;
809 rnp
->completed
= rsp
->completed
;
810 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
811 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
812 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
816 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
819 /* Exclude any concurrent CPU-hotplug operations. */
820 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
823 * Set the quiescent-state-needed bits in all the rcu_node
824 * structures for all currently online CPUs in breadth-first
825 * order, starting from the root rcu_node structure. This
826 * operation relies on the layout of the hierarchy within the
827 * rsp->node[] array. Note that other CPUs will access only
828 * the leaves of the hierarchy, which still indicate that no
829 * grace period is in progress, at least until the corresponding
830 * leaf node has been initialized. In addition, we have excluded
831 * CPU-hotplug operations.
833 * Note that the grace period cannot complete until we finish
834 * the initialization process, as there will be at least one
835 * qsmask bit set in the root node until that time, namely the
836 * one corresponding to this CPU, due to the fact that we have
839 rcu_for_each_node_breadth_first(rsp
, rnp
) {
840 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
841 rcu_preempt_check_blocked_tasks(rnp
);
842 rnp
->qsmask
= rnp
->qsmaskinit
;
843 rnp
->gpnum
= rsp
->gpnum
;
844 rnp
->completed
= rsp
->completed
;
845 if (rnp
== rdp
->mynode
)
846 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
847 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
850 rnp
= rcu_get_root(rsp
);
851 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
852 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
853 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
854 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
858 * Report a full set of quiescent states to the specified rcu_state
859 * data structure. This involves cleaning up after the prior grace
860 * period and letting rcu_start_gp() start up the next grace period
861 * if one is needed. Note that the caller must hold rnp->lock, as
862 * required by rcu_start_gp(), which will release it.
864 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
865 __releases(rcu_get_root(rsp
)->lock
)
867 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
868 rsp
->completed
= rsp
->gpnum
;
869 rsp
->signaled
= RCU_GP_IDLE
;
870 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
874 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
875 * Allows quiescent states for a group of CPUs to be reported at one go
876 * to the specified rcu_node structure, though all the CPUs in the group
877 * must be represented by the same rcu_node structure (which need not be
878 * a leaf rcu_node structure, though it often will be). That structure's
879 * lock must be held upon entry, and it is released before return.
882 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
883 struct rcu_node
*rnp
, unsigned long flags
)
884 __releases(rnp
->lock
)
886 struct rcu_node
*rnp_c
;
888 /* Walk up the rcu_node hierarchy. */
890 if (!(rnp
->qsmask
& mask
)) {
892 /* Our bit has already been cleared, so done. */
893 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
896 rnp
->qsmask
&= ~mask
;
897 if (rnp
->qsmask
!= 0 || rcu_preempted_readers(rnp
)) {
899 /* Other bits still set at this level, so done. */
900 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
904 if (rnp
->parent
== NULL
) {
906 /* No more levels. Exit loop holding root lock. */
910 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
913 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
914 WARN_ON_ONCE(rnp_c
->qsmask
);
918 * Get here if we are the last CPU to pass through a quiescent
919 * state for this grace period. Invoke rcu_report_qs_rsp()
920 * to clean up and start the next grace period if one is needed.
922 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
926 * Record a quiescent state for the specified CPU to that CPU's rcu_data
927 * structure. This must be either called from the specified CPU, or
928 * called when the specified CPU is known to be offline (and when it is
929 * also known that no other CPU is concurrently trying to help the offline
930 * CPU). The lastcomp argument is used to make sure we are still in the
931 * grace period of interest. We don't want to end the current grace period
932 * based on quiescent states detected in an earlier grace period!
935 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
939 struct rcu_node
*rnp
;
942 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
943 if (lastcomp
!= rnp
->completed
) {
946 * Someone beat us to it for this grace period, so leave.
947 * The race with GP start is resolved by the fact that we
948 * hold the leaf rcu_node lock, so that the per-CPU bits
949 * cannot yet be initialized -- so we would simply find our
950 * CPU's bit already cleared in rcu_report_qs_rnp() if this
953 rdp
->passed_quiesc
= 0; /* try again later! */
954 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
958 if ((rnp
->qsmask
& mask
) == 0) {
959 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
964 * This GP can't end until cpu checks in, so all of our
965 * callbacks can be processed during the next GP.
967 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
969 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
974 * Check to see if there is a new grace period of which this CPU
975 * is not yet aware, and if so, set up local rcu_data state for it.
976 * Otherwise, see if this CPU has just passed through its first
977 * quiescent state for this grace period, and record that fact if so.
980 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
982 /* If there is now a new grace period, record and return. */
983 if (check_for_new_grace_period(rsp
, rdp
))
987 * Does this CPU still need to do its part for current grace period?
988 * If no, return and let the other CPUs do their part as well.
990 if (!rdp
->qs_pending
)
994 * Was there a quiescent state since the beginning of the grace
995 * period? If no, then exit and wait for the next call.
997 if (!rdp
->passed_quiesc
)
1001 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1004 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
1007 #ifdef CONFIG_HOTPLUG_CPU
1010 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1011 * Synchronization is not required because this function executes
1012 * in stop_machine() context.
1014 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1017 /* current DYING CPU is cleared in the cpu_online_mask */
1018 int receive_cpu
= cpumask_any(cpu_online_mask
);
1019 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1020 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1022 if (rdp
->nxtlist
== NULL
)
1023 return; /* irqs disabled, so comparison is stable. */
1025 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1026 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1027 receive_rdp
->qlen
+= rdp
->qlen
;
1028 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1029 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1031 rdp
->nxtlist
= NULL
;
1032 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1033 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1038 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1039 * and move all callbacks from the outgoing CPU to the current one.
1041 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1043 unsigned long flags
;
1045 int need_report
= 0;
1046 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1047 struct rcu_node
*rnp
;
1049 /* Exclude any attempts to start a new grace period. */
1050 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1052 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1053 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1054 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1056 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1057 rnp
->qsmaskinit
&= ~mask
;
1058 if (rnp
->qsmaskinit
!= 0) {
1059 if (rnp
!= rdp
->mynode
)
1060 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1063 if (rnp
== rdp
->mynode
)
1064 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1066 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1067 mask
= rnp
->grpmask
;
1069 } while (rnp
!= NULL
);
1072 * We still hold the leaf rcu_node structure lock here, and
1073 * irqs are still disabled. The reason for this subterfuge is
1074 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1075 * held leads to deadlock.
1077 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1079 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1080 rcu_report_unblock_qs_rnp(rnp
, flags
);
1082 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1083 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1084 rcu_report_exp_rnp(rsp
, rnp
);
1088 * Remove the specified CPU from the RCU hierarchy and move any pending
1089 * callbacks that it might have to the current CPU. This code assumes
1090 * that at least one CPU in the system will remain running at all times.
1091 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1093 static void rcu_offline_cpu(int cpu
)
1095 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1096 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1097 rcu_preempt_offline_cpu(cpu
);
1100 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1102 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1106 static void rcu_offline_cpu(int cpu
)
1110 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1113 * Invoke any RCU callbacks that have made it to the end of their grace
1114 * period. Thottle as specified by rdp->blimit.
1116 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1118 unsigned long flags
;
1119 struct rcu_head
*next
, *list
, **tail
;
1122 /* If no callbacks are ready, just return.*/
1123 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1127 * Extract the list of ready callbacks, disabling to prevent
1128 * races with call_rcu() from interrupt handlers.
1130 local_irq_save(flags
);
1131 list
= rdp
->nxtlist
;
1132 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1133 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1134 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1135 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1136 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1137 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1138 local_irq_restore(flags
);
1140 /* Invoke callbacks. */
1145 debug_rcu_head_unqueue(list
);
1148 if (++count
>= rdp
->blimit
)
1152 local_irq_save(flags
);
1154 /* Update count, and requeue any remaining callbacks. */
1156 rdp
->n_cbs_invoked
+= count
;
1158 *tail
= rdp
->nxtlist
;
1159 rdp
->nxtlist
= list
;
1160 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1161 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1162 rdp
->nxttail
[count
] = tail
;
1167 /* Reinstate batch limit if we have worked down the excess. */
1168 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1169 rdp
->blimit
= blimit
;
1171 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1172 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1173 rdp
->qlen_last_fqs_check
= 0;
1174 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1175 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1176 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1178 local_irq_restore(flags
);
1180 /* Re-raise the RCU softirq if there are callbacks remaining. */
1181 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1182 raise_softirq(RCU_SOFTIRQ
);
1186 * Check to see if this CPU is in a non-context-switch quiescent state
1187 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1188 * Also schedule the RCU softirq handler.
1190 * This function must be called with hardirqs disabled. It is normally
1191 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1192 * false, there is no point in invoking rcu_check_callbacks().
1194 void rcu_check_callbacks(int cpu
, int user
)
1197 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1198 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1201 * Get here if this CPU took its interrupt from user
1202 * mode or from the idle loop, and if this is not a
1203 * nested interrupt. In this case, the CPU is in
1204 * a quiescent state, so note it.
1206 * No memory barrier is required here because both
1207 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1208 * variables that other CPUs neither access nor modify,
1209 * at least not while the corresponding CPU is online.
1215 } else if (!in_softirq()) {
1218 * Get here if this CPU did not take its interrupt from
1219 * softirq, in other words, if it is not interrupting
1220 * a rcu_bh read-side critical section. This is an _bh
1221 * critical section, so note it.
1226 rcu_preempt_check_callbacks(cpu
);
1227 if (rcu_pending(cpu
))
1228 raise_softirq(RCU_SOFTIRQ
);
1234 * Scan the leaf rcu_node structures, processing dyntick state for any that
1235 * have not yet encountered a quiescent state, using the function specified.
1236 * The caller must have suppressed start of new grace periods.
1238 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1242 unsigned long flags
;
1244 struct rcu_node
*rnp
;
1246 rcu_for_each_leaf_node(rsp
, rnp
) {
1248 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1249 if (!rcu_gp_in_progress(rsp
)) {
1250 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1253 if (rnp
->qsmask
== 0) {
1254 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1259 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1260 if ((rnp
->qsmask
& bit
) != 0 &&
1261 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1266 /* rcu_report_qs_rnp() releases rnp->lock. */
1267 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1270 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1275 * Force quiescent states on reluctant CPUs, and also detect which
1276 * CPUs are in dyntick-idle mode.
1278 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1280 unsigned long flags
;
1281 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1283 if (!rcu_gp_in_progress(rsp
))
1284 return; /* No grace period in progress, nothing to force. */
1285 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1286 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1287 return; /* Someone else is already on the job. */
1289 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1290 goto unlock_fqs_ret
; /* no emergency and done recently. */
1292 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1293 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1294 if(!rcu_gp_in_progress(rsp
)) {
1295 rsp
->n_force_qs_ngp
++;
1296 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1297 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1299 rsp
->fqs_active
= 1;
1300 switch (rsp
->signaled
) {
1304 break; /* grace period idle or initializing, ignore. */
1306 case RCU_SAVE_DYNTICK
:
1307 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1308 break; /* So gcc recognizes the dead code. */
1310 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1312 /* Record dyntick-idle state. */
1313 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1314 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1315 if (rcu_gp_in_progress(rsp
))
1316 rsp
->signaled
= RCU_FORCE_QS
;
1321 /* Check dyntick-idle state, send IPI to laggarts. */
1322 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1323 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1325 /* Leave state in case more forcing is required. */
1327 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1330 rsp
->fqs_active
= 0;
1331 if (rsp
->fqs_need_gp
) {
1332 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1333 rsp
->fqs_need_gp
= 0;
1334 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1337 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1339 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1342 #else /* #ifdef CONFIG_SMP */
1344 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1349 #endif /* #else #ifdef CONFIG_SMP */
1352 * This does the RCU processing work from softirq context for the
1353 * specified rcu_state and rcu_data structures. This may be called
1354 * only from the CPU to whom the rdp belongs.
1357 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1359 unsigned long flags
;
1361 WARN_ON_ONCE(rdp
->beenonline
== 0);
1364 * If an RCU GP has gone long enough, go check for dyntick
1365 * idle CPUs and, if needed, send resched IPIs.
1367 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1368 force_quiescent_state(rsp
, 1);
1371 * Advance callbacks in response to end of earlier grace
1372 * period that some other CPU ended.
1374 rcu_process_gp_end(rsp
, rdp
);
1376 /* Update RCU state based on any recent quiescent states. */
1377 rcu_check_quiescent_state(rsp
, rdp
);
1379 /* Does this CPU require a not-yet-started grace period? */
1380 if (cpu_needs_another_gp(rsp
, rdp
)) {
1381 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1382 rcu_start_gp(rsp
, flags
); /* releases above lock */
1385 /* If there are callbacks ready, invoke them. */
1386 rcu_do_batch(rsp
, rdp
);
1390 * Do softirq processing for the current CPU.
1392 static void rcu_process_callbacks(struct softirq_action
*unused
)
1395 * Memory references from any prior RCU read-side critical sections
1396 * executed by the interrupted code must be seen before any RCU
1397 * grace-period manipulations below.
1399 smp_mb(); /* See above block comment. */
1401 __rcu_process_callbacks(&rcu_sched_state
,
1402 &__get_cpu_var(rcu_sched_data
));
1403 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1404 rcu_preempt_process_callbacks();
1407 * Memory references from any later RCU read-side critical sections
1408 * executed by the interrupted code must be seen after any RCU
1409 * grace-period manipulations above.
1411 smp_mb(); /* See above block comment. */
1413 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1414 rcu_needs_cpu_flush();
1418 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1419 struct rcu_state
*rsp
)
1421 unsigned long flags
;
1422 struct rcu_data
*rdp
;
1424 debug_rcu_head_queue(head
);
1428 smp_mb(); /* Ensure RCU update seen before callback registry. */
1431 * Opportunistically note grace-period endings and beginnings.
1432 * Note that we might see a beginning right after we see an
1433 * end, but never vice versa, since this CPU has to pass through
1434 * a quiescent state betweentimes.
1436 local_irq_save(flags
);
1437 rdp
= this_cpu_ptr(rsp
->rda
);
1439 /* Add the callback to our list. */
1440 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1441 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1444 * Force the grace period if too many callbacks or too long waiting.
1445 * Enforce hysteresis, and don't invoke force_quiescent_state()
1446 * if some other CPU has recently done so. Also, don't bother
1447 * invoking force_quiescent_state() if the newly enqueued callback
1448 * is the only one waiting for a grace period to complete.
1450 if (unlikely(++rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1452 /* Are we ignoring a completed grace period? */
1453 rcu_process_gp_end(rsp
, rdp
);
1454 check_for_new_grace_period(rsp
, rdp
);
1456 /* Start a new grace period if one not already started. */
1457 if (!rcu_gp_in_progress(rsp
)) {
1458 unsigned long nestflag
;
1459 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1461 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1462 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1464 /* Give the grace period a kick. */
1465 rdp
->blimit
= LONG_MAX
;
1466 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1467 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1468 force_quiescent_state(rsp
, 0);
1469 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1470 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1472 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1473 force_quiescent_state(rsp
, 1);
1474 local_irq_restore(flags
);
1478 * Queue an RCU-sched callback for invocation after a grace period.
1480 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1482 __call_rcu(head
, func
, &rcu_sched_state
);
1484 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1487 * Queue an RCU for invocation after a quicker grace period.
1489 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1491 __call_rcu(head
, func
, &rcu_bh_state
);
1493 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1496 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1498 * Control will return to the caller some time after a full rcu-sched
1499 * grace period has elapsed, in other words after all currently executing
1500 * rcu-sched read-side critical sections have completed. These read-side
1501 * critical sections are delimited by rcu_read_lock_sched() and
1502 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1503 * local_irq_disable(), and so on may be used in place of
1504 * rcu_read_lock_sched().
1506 * This means that all preempt_disable code sequences, including NMI and
1507 * hardware-interrupt handlers, in progress on entry will have completed
1508 * before this primitive returns. However, this does not guarantee that
1509 * softirq handlers will have completed, since in some kernels, these
1510 * handlers can run in process context, and can block.
1512 * This primitive provides the guarantees made by the (now removed)
1513 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1514 * guarantees that rcu_read_lock() sections will have completed.
1515 * In "classic RCU", these two guarantees happen to be one and
1516 * the same, but can differ in realtime RCU implementations.
1518 void synchronize_sched(void)
1520 struct rcu_synchronize rcu
;
1522 if (rcu_blocking_is_gp())
1525 init_rcu_head_on_stack(&rcu
.head
);
1526 init_completion(&rcu
.completion
);
1527 /* Will wake me after RCU finished. */
1528 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1530 wait_for_completion(&rcu
.completion
);
1531 destroy_rcu_head_on_stack(&rcu
.head
);
1533 EXPORT_SYMBOL_GPL(synchronize_sched
);
1536 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1538 * Control will return to the caller some time after a full rcu_bh grace
1539 * period has elapsed, in other words after all currently executing rcu_bh
1540 * read-side critical sections have completed. RCU read-side critical
1541 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1542 * and may be nested.
1544 void synchronize_rcu_bh(void)
1546 struct rcu_synchronize rcu
;
1548 if (rcu_blocking_is_gp())
1551 init_rcu_head_on_stack(&rcu
.head
);
1552 init_completion(&rcu
.completion
);
1553 /* Will wake me after RCU finished. */
1554 call_rcu_bh(&rcu
.head
, wakeme_after_rcu
);
1556 wait_for_completion(&rcu
.completion
);
1557 destroy_rcu_head_on_stack(&rcu
.head
);
1559 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1562 * Check to see if there is any immediate RCU-related work to be done
1563 * by the current CPU, for the specified type of RCU, returning 1 if so.
1564 * The checks are in order of increasing expense: checks that can be
1565 * carried out against CPU-local state are performed first. However,
1566 * we must check for CPU stalls first, else we might not get a chance.
1568 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1570 struct rcu_node
*rnp
= rdp
->mynode
;
1572 rdp
->n_rcu_pending
++;
1574 /* Check for CPU stalls, if enabled. */
1575 check_cpu_stall(rsp
, rdp
);
1577 /* Is the RCU core waiting for a quiescent state from this CPU? */
1578 if (rdp
->qs_pending
&& !rdp
->passed_quiesc
) {
1581 * If force_quiescent_state() coming soon and this CPU
1582 * needs a quiescent state, and this is either RCU-sched
1583 * or RCU-bh, force a local reschedule.
1585 rdp
->n_rp_qs_pending
++;
1586 if (!rdp
->preemptable
&&
1587 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
1590 } else if (rdp
->qs_pending
&& rdp
->passed_quiesc
) {
1591 rdp
->n_rp_report_qs
++;
1595 /* Does this CPU have callbacks ready to invoke? */
1596 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1597 rdp
->n_rp_cb_ready
++;
1601 /* Has RCU gone idle with this CPU needing another grace period? */
1602 if (cpu_needs_another_gp(rsp
, rdp
)) {
1603 rdp
->n_rp_cpu_needs_gp
++;
1607 /* Has another RCU grace period completed? */
1608 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
1609 rdp
->n_rp_gp_completed
++;
1613 /* Has a new RCU grace period started? */
1614 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
1615 rdp
->n_rp_gp_started
++;
1619 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1620 if (rcu_gp_in_progress(rsp
) &&
1621 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
1622 rdp
->n_rp_need_fqs
++;
1627 rdp
->n_rp_need_nothing
++;
1632 * Check to see if there is any immediate RCU-related work to be done
1633 * by the current CPU, returning 1 if so. This function is part of the
1634 * RCU implementation; it is -not- an exported member of the RCU API.
1636 static int rcu_pending(int cpu
)
1638 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
1639 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
1640 rcu_preempt_pending(cpu
);
1644 * Check to see if any future RCU-related work will need to be done
1645 * by the current CPU, even if none need be done immediately, returning
1648 static int rcu_needs_cpu_quick_check(int cpu
)
1650 /* RCU callbacks either ready or pending? */
1651 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
1652 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
1653 rcu_preempt_needs_cpu(cpu
);
1656 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
1657 static atomic_t rcu_barrier_cpu_count
;
1658 static DEFINE_MUTEX(rcu_barrier_mutex
);
1659 static struct completion rcu_barrier_completion
;
1661 static void rcu_barrier_callback(struct rcu_head
*notused
)
1663 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1664 complete(&rcu_barrier_completion
);
1668 * Called with preemption disabled, and from cross-cpu IRQ context.
1670 static void rcu_barrier_func(void *type
)
1672 int cpu
= smp_processor_id();
1673 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
1674 void (*call_rcu_func
)(struct rcu_head
*head
,
1675 void (*func
)(struct rcu_head
*head
));
1677 atomic_inc(&rcu_barrier_cpu_count
);
1678 call_rcu_func
= type
;
1679 call_rcu_func(head
, rcu_barrier_callback
);
1683 * Orchestrate the specified type of RCU barrier, waiting for all
1684 * RCU callbacks of the specified type to complete.
1686 static void _rcu_barrier(struct rcu_state
*rsp
,
1687 void (*call_rcu_func
)(struct rcu_head
*head
,
1688 void (*func
)(struct rcu_head
*head
)))
1690 BUG_ON(in_interrupt());
1691 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1692 mutex_lock(&rcu_barrier_mutex
);
1693 init_completion(&rcu_barrier_completion
);
1695 * Initialize rcu_barrier_cpu_count to 1, then invoke
1696 * rcu_barrier_func() on each CPU, so that each CPU also has
1697 * incremented rcu_barrier_cpu_count. Only then is it safe to
1698 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1699 * might complete its grace period before all of the other CPUs
1700 * did their increment, causing this function to return too
1701 * early. Note that on_each_cpu() disables irqs, which prevents
1702 * any CPUs from coming online or going offline until each online
1703 * CPU has queued its RCU-barrier callback.
1705 atomic_set(&rcu_barrier_cpu_count
, 1);
1706 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
1707 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1708 complete(&rcu_barrier_completion
);
1709 wait_for_completion(&rcu_barrier_completion
);
1710 mutex_unlock(&rcu_barrier_mutex
);
1714 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1716 void rcu_barrier_bh(void)
1718 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
1720 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
1723 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1725 void rcu_barrier_sched(void)
1727 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
1729 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
1732 * Do boot-time initialization of a CPU's per-CPU RCU data.
1735 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
1737 unsigned long flags
;
1739 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1740 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1742 /* Set up local state, ensuring consistent view of global state. */
1743 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1744 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
1745 rdp
->nxtlist
= NULL
;
1746 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1747 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1750 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
1751 #endif /* #ifdef CONFIG_NO_HZ */
1753 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1757 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1758 * offline event can be happening at a given time. Note also that we
1759 * can accept some slop in the rsp->completed access due to the fact
1760 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1762 static void __cpuinit
1763 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptable
)
1765 unsigned long flags
;
1767 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1768 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1770 /* Set up local state, ensuring consistent view of global state. */
1771 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1772 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
1773 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
1774 rdp
->beenonline
= 1; /* We have now been online. */
1775 rdp
->preemptable
= preemptable
;
1776 rdp
->qlen_last_fqs_check
= 0;
1777 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1778 rdp
->blimit
= blimit
;
1779 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1782 * A new grace period might start here. If so, we won't be part
1783 * of it, but that is OK, as we are currently in a quiescent state.
1786 /* Exclude any attempts to start a new GP on large systems. */
1787 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
1789 /* Add CPU to rcu_node bitmasks. */
1791 mask
= rdp
->grpmask
;
1793 /* Exclude any attempts to start a new GP on small systems. */
1794 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1795 rnp
->qsmaskinit
|= mask
;
1796 mask
= rnp
->grpmask
;
1797 if (rnp
== rdp
->mynode
) {
1798 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
1799 rdp
->completed
= rnp
->completed
;
1800 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
1802 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
1804 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
1806 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
1809 static void __cpuinit
rcu_online_cpu(int cpu
)
1811 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
1812 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
1813 rcu_preempt_init_percpu_data(cpu
);
1817 * Handle CPU online/offline notification events.
1819 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
1820 unsigned long action
, void *hcpu
)
1822 long cpu
= (long)hcpu
;
1825 case CPU_UP_PREPARE
:
1826 case CPU_UP_PREPARE_FROZEN
:
1827 rcu_online_cpu(cpu
);
1830 case CPU_DYING_FROZEN
:
1832 * The whole machine is "stopped" except this CPU, so we can
1833 * touch any data without introducing corruption. We send the
1834 * dying CPU's callbacks to an arbitrarily chosen online CPU.
1836 rcu_send_cbs_to_online(&rcu_bh_state
);
1837 rcu_send_cbs_to_online(&rcu_sched_state
);
1838 rcu_preempt_send_cbs_to_online();
1841 case CPU_DEAD_FROZEN
:
1842 case CPU_UP_CANCELED
:
1843 case CPU_UP_CANCELED_FROZEN
:
1844 rcu_offline_cpu(cpu
);
1853 * This function is invoked towards the end of the scheduler's initialization
1854 * process. Before this is called, the idle task might contain
1855 * RCU read-side critical sections (during which time, this idle
1856 * task is booting the system). After this function is called, the
1857 * idle tasks are prohibited from containing RCU read-side critical
1858 * sections. This function also enables RCU lockdep checking.
1860 void rcu_scheduler_starting(void)
1862 WARN_ON(num_online_cpus() != 1);
1863 WARN_ON(nr_context_switches() > 0);
1864 rcu_scheduler_active
= 1;
1868 * Compute the per-level fanout, either using the exact fanout specified
1869 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1871 #ifdef CONFIG_RCU_FANOUT_EXACT
1872 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
1876 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
1877 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
1878 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
1880 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1881 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
1888 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
1889 ccur
= rsp
->levelcnt
[i
];
1890 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
1894 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1897 * Helper function for rcu_init() that initializes one rcu_state structure.
1899 static void __init
rcu_init_one(struct rcu_state
*rsp
,
1900 struct rcu_data __percpu
*rda
)
1902 static char *buf
[] = { "rcu_node_level_0",
1905 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1909 struct rcu_node
*rnp
;
1911 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
1913 /* Initialize the level-tracking arrays. */
1915 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
1916 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
1917 rcu_init_levelspread(rsp
);
1919 /* Initialize the elements themselves, starting from the leaves. */
1921 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
1922 cpustride
*= rsp
->levelspread
[i
];
1923 rnp
= rsp
->level
[i
];
1924 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
1925 raw_spin_lock_init(&rnp
->lock
);
1926 lockdep_set_class_and_name(&rnp
->lock
,
1927 &rcu_node_class
[i
], buf
[i
]);
1930 rnp
->qsmaskinit
= 0;
1931 rnp
->grplo
= j
* cpustride
;
1932 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
1933 if (rnp
->grphi
>= NR_CPUS
)
1934 rnp
->grphi
= NR_CPUS
- 1;
1940 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
1941 rnp
->grpmask
= 1UL << rnp
->grpnum
;
1942 rnp
->parent
= rsp
->level
[i
- 1] +
1943 j
/ rsp
->levelspread
[i
- 1];
1946 INIT_LIST_HEAD(&rnp
->blocked_tasks
[0]);
1947 INIT_LIST_HEAD(&rnp
->blocked_tasks
[1]);
1948 INIT_LIST_HEAD(&rnp
->blocked_tasks
[2]);
1949 INIT_LIST_HEAD(&rnp
->blocked_tasks
[3]);
1954 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
1955 for_each_possible_cpu(i
) {
1956 while (i
> rnp
->grphi
)
1958 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
1959 rcu_boot_init_percpu_data(i
, rsp
);
1963 void __init
rcu_init(void)
1967 rcu_bootup_announce();
1968 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
1969 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
1970 __rcu_init_preempt();
1971 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
1974 * We don't need protection against CPU-hotplug here because
1975 * this is called early in boot, before either interrupts
1976 * or the scheduler are operational.
1978 cpu_notifier(rcu_cpu_notify
, 0);
1979 for_each_online_cpu(cpu
)
1980 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1981 check_cpu_stall_init();
1984 #include "rcutree_plugin.h"