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 static DEFINE_PER_CPU(wait_queue_head_t
, rcu_cpu_wq
);
97 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
98 static char rcu_kthreads_spawnable
;
100 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
101 static void invoke_rcu_cpu_kthread(void);
103 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
106 * Track the rcutorture test sequence number and the update version
107 * number within a given test. The rcutorture_testseq is incremented
108 * on every rcutorture module load and unload, so has an odd value
109 * when a test is running. The rcutorture_vernum is set to zero
110 * when rcutorture starts and is incremented on each rcutorture update.
111 * These variables enable correlating rcutorture output with the
112 * RCU tracing information.
114 unsigned long rcutorture_testseq
;
115 unsigned long rcutorture_vernum
;
118 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
119 * permit this function to be invoked without holding the root rcu_node
120 * structure's ->lock, but of course results can be subject to change.
122 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
124 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
128 * Note a quiescent state. Because we do not need to know
129 * how many quiescent states passed, just if there was at least
130 * one since the start of the grace period, this just sets a flag.
132 void rcu_sched_qs(int cpu
)
134 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
136 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
138 rdp
->passed_quiesc
= 1;
141 void rcu_bh_qs(int cpu
)
143 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
145 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
147 rdp
->passed_quiesc
= 1;
151 * Note a context switch. This is a quiescent state for RCU-sched,
152 * and requires special handling for preemptible RCU.
154 void rcu_note_context_switch(int cpu
)
157 rcu_preempt_note_context_switch(cpu
);
161 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
162 .dynticks_nesting
= 1,
163 .dynticks
= ATOMIC_INIT(1),
165 #endif /* #ifdef CONFIG_NO_HZ */
167 static int blimit
= 10; /* Maximum callbacks per softirq. */
168 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
169 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
171 module_param(blimit
, int, 0);
172 module_param(qhimark
, int, 0);
173 module_param(qlowmark
, int, 0);
175 int rcu_cpu_stall_suppress __read_mostly
;
176 module_param(rcu_cpu_stall_suppress
, int, 0644);
178 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
179 static int rcu_pending(int cpu
);
182 * Return the number of RCU-sched batches processed thus far for debug & stats.
184 long rcu_batches_completed_sched(void)
186 return rcu_sched_state
.completed
;
188 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
191 * Return the number of RCU BH batches processed thus far for debug & stats.
193 long rcu_batches_completed_bh(void)
195 return rcu_bh_state
.completed
;
197 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
200 * Force a quiescent state for RCU BH.
202 void rcu_bh_force_quiescent_state(void)
204 force_quiescent_state(&rcu_bh_state
, 0);
206 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
209 * Record the number of times rcutorture tests have been initiated and
210 * terminated. This information allows the debugfs tracing stats to be
211 * correlated to the rcutorture messages, even when the rcutorture module
212 * is being repeatedly loaded and unloaded. In other words, we cannot
213 * store this state in rcutorture itself.
215 void rcutorture_record_test_transition(void)
217 rcutorture_testseq
++;
218 rcutorture_vernum
= 0;
220 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
223 * Record the number of writer passes through the current rcutorture test.
224 * This is also used to correlate debugfs tracing stats with the rcutorture
227 void rcutorture_record_progress(unsigned long vernum
)
231 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
234 * Force a quiescent state for RCU-sched.
236 void rcu_sched_force_quiescent_state(void)
238 force_quiescent_state(&rcu_sched_state
, 0);
240 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
243 * Does the CPU have callbacks ready to be invoked?
246 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
248 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
252 * Does the current CPU require a yet-as-unscheduled grace period?
255 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
257 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
261 * Return the root node of the specified rcu_state structure.
263 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
265 return &rsp
->node
[0];
271 * If the specified CPU is offline, tell the caller that it is in
272 * a quiescent state. Otherwise, whack it with a reschedule IPI.
273 * Grace periods can end up waiting on an offline CPU when that
274 * CPU is in the process of coming online -- it will be added to the
275 * rcu_node bitmasks before it actually makes it online. The same thing
276 * can happen while a CPU is in the process of coming online. Because this
277 * race is quite rare, we check for it after detecting that the grace
278 * period has been delayed rather than checking each and every CPU
279 * each and every time we start a new grace period.
281 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
284 * If the CPU is offline, it is in a quiescent state. We can
285 * trust its state not to change because interrupts are disabled.
287 if (cpu_is_offline(rdp
->cpu
)) {
292 /* If preemptable RCU, no point in sending reschedule IPI. */
293 if (rdp
->preemptable
)
296 /* The CPU is online, so send it a reschedule IPI. */
297 if (rdp
->cpu
!= smp_processor_id())
298 smp_send_reschedule(rdp
->cpu
);
305 #endif /* #ifdef CONFIG_SMP */
310 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
312 * Enter nohz mode, in other words, -leave- the mode in which RCU
313 * read-side critical sections can occur. (Though RCU read-side
314 * critical sections can occur in irq handlers in nohz mode, a possibility
315 * handled by rcu_irq_enter() and rcu_irq_exit()).
317 void rcu_enter_nohz(void)
320 struct rcu_dynticks
*rdtp
;
322 local_irq_save(flags
);
323 rdtp
= &__get_cpu_var(rcu_dynticks
);
324 if (--rdtp
->dynticks_nesting
) {
325 local_irq_restore(flags
);
328 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
329 smp_mb__before_atomic_inc(); /* See above. */
330 atomic_inc(&rdtp
->dynticks
);
331 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
332 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
333 local_irq_restore(flags
);
335 /* If the interrupt queued a callback, get out of dyntick mode. */
337 (__get_cpu_var(rcu_sched_data
).nxtlist
||
338 __get_cpu_var(rcu_bh_data
).nxtlist
||
339 rcu_preempt_needs_cpu(smp_processor_id())))
344 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
346 * Exit nohz mode, in other words, -enter- the mode in which RCU
347 * read-side critical sections normally occur.
349 void rcu_exit_nohz(void)
352 struct rcu_dynticks
*rdtp
;
354 local_irq_save(flags
);
355 rdtp
= &__get_cpu_var(rcu_dynticks
);
356 if (rdtp
->dynticks_nesting
++) {
357 local_irq_restore(flags
);
360 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
361 atomic_inc(&rdtp
->dynticks
);
362 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
363 smp_mb__after_atomic_inc(); /* See above. */
364 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
365 local_irq_restore(flags
);
369 * rcu_nmi_enter - inform RCU of entry to NMI context
371 * If the CPU was idle with dynamic ticks active, and there is no
372 * irq handler running, this updates rdtp->dynticks_nmi to let the
373 * RCU grace-period handling know that the CPU is active.
375 void rcu_nmi_enter(void)
377 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
379 if (rdtp
->dynticks_nmi_nesting
== 0 &&
380 (atomic_read(&rdtp
->dynticks
) & 0x1))
382 rdtp
->dynticks_nmi_nesting
++;
383 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
384 atomic_inc(&rdtp
->dynticks
);
385 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
386 smp_mb__after_atomic_inc(); /* See above. */
387 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
391 * rcu_nmi_exit - inform RCU of exit from NMI context
393 * If the CPU was idle with dynamic ticks active, and there is no
394 * irq handler running, this updates rdtp->dynticks_nmi to let the
395 * RCU grace-period handling know that the CPU is no longer active.
397 void rcu_nmi_exit(void)
399 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
401 if (rdtp
->dynticks_nmi_nesting
== 0 ||
402 --rdtp
->dynticks_nmi_nesting
!= 0)
404 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
405 smp_mb__before_atomic_inc(); /* See above. */
406 atomic_inc(&rdtp
->dynticks
);
407 smp_mb__after_atomic_inc(); /* Force delay to next write. */
408 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
412 * rcu_irq_enter - inform RCU of entry to hard irq context
414 * If the CPU was idle with dynamic ticks active, this updates the
415 * rdtp->dynticks to let the RCU handling know that the CPU is active.
417 void rcu_irq_enter(void)
423 * rcu_irq_exit - inform RCU of exit from hard irq context
425 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
426 * to put let the RCU handling be aware that the CPU is going back to idle
429 void rcu_irq_exit(void)
437 * Snapshot the specified CPU's dynticks counter so that we can later
438 * credit them with an implicit quiescent state. Return 1 if this CPU
439 * is in dynticks idle mode, which is an extended quiescent state.
441 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
443 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
448 * Return true if the specified CPU has passed through a quiescent
449 * state by virtue of being in or having passed through an dynticks
450 * idle state since the last call to dyntick_save_progress_counter()
453 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
458 curr
= (unsigned long)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
459 snap
= (unsigned long)rdp
->dynticks_snap
;
462 * If the CPU passed through or entered a dynticks idle phase with
463 * no active irq/NMI handlers, then we can safely pretend that the CPU
464 * already acknowledged the request to pass through a quiescent
465 * state. Either way, that CPU cannot possibly be in an RCU
466 * read-side critical section that started before the beginning
467 * of the current RCU grace period.
469 if ((curr
& 0x1) == 0 || ULONG_CMP_GE(curr
, snap
+ 2)) {
474 /* Go check for the CPU being offline. */
475 return rcu_implicit_offline_qs(rdp
);
478 #endif /* #ifdef CONFIG_SMP */
480 #else /* #ifdef CONFIG_NO_HZ */
484 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
489 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
491 return rcu_implicit_offline_qs(rdp
);
494 #endif /* #ifdef CONFIG_SMP */
496 #endif /* #else #ifdef CONFIG_NO_HZ */
498 int rcu_cpu_stall_suppress __read_mostly
;
500 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
502 rsp
->gp_start
= jiffies
;
503 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
506 static void print_other_cpu_stall(struct rcu_state
*rsp
)
511 struct rcu_node
*rnp
= rcu_get_root(rsp
);
513 /* Only let one CPU complain about others per time interval. */
515 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
516 delta
= jiffies
- rsp
->jiffies_stall
;
517 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
518 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
521 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
524 * Now rat on any tasks that got kicked up to the root rcu_node
525 * due to CPU offlining.
527 rcu_print_task_stall(rnp
);
528 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
531 * OK, time to rat on our buddy...
532 * See Documentation/RCU/stallwarn.txt for info on how to debug
533 * RCU CPU stall warnings.
535 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
537 rcu_for_each_leaf_node(rsp
, rnp
) {
538 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
539 rcu_print_task_stall(rnp
);
540 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
541 if (rnp
->qsmask
== 0)
543 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
544 if (rnp
->qsmask
& (1UL << cpu
))
545 printk(" %d", rnp
->grplo
+ cpu
);
547 printk("} (detected by %d, t=%ld jiffies)\n",
548 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
549 trigger_all_cpu_backtrace();
551 /* If so configured, complain about tasks blocking the grace period. */
553 rcu_print_detail_task_stall(rsp
);
555 force_quiescent_state(rsp
, 0); /* Kick them all. */
558 static void print_cpu_stall(struct rcu_state
*rsp
)
561 struct rcu_node
*rnp
= rcu_get_root(rsp
);
564 * OK, time to rat on ourselves...
565 * See Documentation/RCU/stallwarn.txt for info on how to debug
566 * RCU CPU stall warnings.
568 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
569 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
570 trigger_all_cpu_backtrace();
572 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
573 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
575 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
576 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
578 set_need_resched(); /* kick ourselves to get things going. */
581 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
584 struct rcu_node
*rnp
;
586 if (rcu_cpu_stall_suppress
)
588 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
590 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && delta
>= 0) {
592 /* We haven't checked in, so go dump stack. */
593 print_cpu_stall(rsp
);
595 } else if (rcu_gp_in_progress(rsp
) && delta
>= RCU_STALL_RAT_DELAY
) {
597 /* They had two time units to dump stack, so complain. */
598 print_other_cpu_stall(rsp
);
602 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
604 rcu_cpu_stall_suppress
= 1;
609 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
611 * Set the stall-warning timeout way off into the future, thus preventing
612 * any RCU CPU stall-warning messages from appearing in the current set of
615 * The caller must disable hard irqs.
617 void rcu_cpu_stall_reset(void)
619 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
620 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
621 rcu_preempt_stall_reset();
624 static struct notifier_block rcu_panic_block
= {
625 .notifier_call
= rcu_panic
,
628 static void __init
check_cpu_stall_init(void)
630 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
634 * Update CPU-local rcu_data state to record the newly noticed grace period.
635 * This is used both when we started the grace period and when we notice
636 * that someone else started the grace period. The caller must hold the
637 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
638 * and must have irqs disabled.
640 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
642 if (rdp
->gpnum
!= rnp
->gpnum
) {
644 * If the current grace period is waiting for this CPU,
645 * set up to detect a quiescent state, otherwise don't
646 * go looking for one.
648 rdp
->gpnum
= rnp
->gpnum
;
649 if (rnp
->qsmask
& rdp
->grpmask
) {
651 rdp
->passed_quiesc
= 0;
657 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
660 struct rcu_node
*rnp
;
662 local_irq_save(flags
);
664 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
665 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
666 local_irq_restore(flags
);
669 __note_new_gpnum(rsp
, rnp
, rdp
);
670 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
674 * Did someone else start a new RCU grace period start since we last
675 * checked? Update local state appropriately if so. Must be called
676 * on the CPU corresponding to rdp.
679 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
684 local_irq_save(flags
);
685 if (rdp
->gpnum
!= rsp
->gpnum
) {
686 note_new_gpnum(rsp
, rdp
);
689 local_irq_restore(flags
);
694 * Advance this CPU's callbacks, but only if the current grace period
695 * has ended. This may be called only from the CPU to whom the rdp
696 * belongs. In addition, the corresponding leaf rcu_node structure's
697 * ->lock must be held by the caller, with irqs disabled.
700 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
702 /* Did another grace period end? */
703 if (rdp
->completed
!= rnp
->completed
) {
705 /* Advance callbacks. No harm if list empty. */
706 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
707 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
708 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
710 /* Remember that we saw this grace-period completion. */
711 rdp
->completed
= rnp
->completed
;
714 * If we were in an extended quiescent state, we may have
715 * missed some grace periods that others CPUs handled on
716 * our behalf. Catch up with this state to avoid noting
717 * spurious new grace periods. If another grace period
718 * has started, then rnp->gpnum will have advanced, so
719 * we will detect this later on.
721 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
722 rdp
->gpnum
= rdp
->completed
;
725 * If RCU does not need a quiescent state from this CPU,
726 * then make sure that this CPU doesn't go looking for one.
728 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
734 * Advance this CPU's callbacks, but only if the current grace period
735 * has ended. This may be called only from the CPU to whom the rdp
739 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
742 struct rcu_node
*rnp
;
744 local_irq_save(flags
);
746 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
747 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
748 local_irq_restore(flags
);
751 __rcu_process_gp_end(rsp
, rnp
, rdp
);
752 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
756 * Do per-CPU grace-period initialization for running CPU. The caller
757 * must hold the lock of the leaf rcu_node structure corresponding to
761 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
763 /* Prior grace period ended, so advance callbacks for current CPU. */
764 __rcu_process_gp_end(rsp
, rnp
, rdp
);
767 * Because this CPU just now started the new grace period, we know
768 * that all of its callbacks will be covered by this upcoming grace
769 * period, even the ones that were registered arbitrarily recently.
770 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
772 * Other CPUs cannot be sure exactly when the grace period started.
773 * Therefore, their recently registered callbacks must pass through
774 * an additional RCU_NEXT_READY stage, so that they will be handled
775 * by the next RCU grace period.
777 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
778 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
780 /* Set state so that this CPU will detect the next quiescent state. */
781 __note_new_gpnum(rsp
, rnp
, rdp
);
785 * Start a new RCU grace period if warranted, re-initializing the hierarchy
786 * in preparation for detecting the next grace period. The caller must hold
787 * the root node's ->lock, which is released before return. Hard irqs must
791 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
792 __releases(rcu_get_root(rsp
)->lock
)
794 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
795 struct rcu_node
*rnp
= rcu_get_root(rsp
);
797 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
798 if (cpu_needs_another_gp(rsp
, rdp
))
799 rsp
->fqs_need_gp
= 1;
800 if (rnp
->completed
== rsp
->completed
) {
801 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
804 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
807 * Propagate new ->completed value to rcu_node structures
808 * so that other CPUs don't have to wait until the start
809 * of the next grace period to process their callbacks.
811 rcu_for_each_node_breadth_first(rsp
, rnp
) {
812 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
813 rnp
->completed
= rsp
->completed
;
814 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
816 local_irq_restore(flags
);
820 /* Advance to a new grace period and initialize state. */
822 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
823 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
824 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
825 record_gp_stall_check_time(rsp
);
827 /* Special-case the common single-level case. */
828 if (NUM_RCU_NODES
== 1) {
829 rcu_preempt_check_blocked_tasks(rnp
);
830 rnp
->qsmask
= rnp
->qsmaskinit
;
831 rnp
->gpnum
= rsp
->gpnum
;
832 rnp
->completed
= rsp
->completed
;
833 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
834 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
835 rcu_preempt_boost_start_gp(rnp
);
836 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
840 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
843 /* Exclude any concurrent CPU-hotplug operations. */
844 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
847 * Set the quiescent-state-needed bits in all the rcu_node
848 * structures for all currently online CPUs in breadth-first
849 * order, starting from the root rcu_node structure. This
850 * operation relies on the layout of the hierarchy within the
851 * rsp->node[] array. Note that other CPUs will access only
852 * the leaves of the hierarchy, which still indicate that no
853 * grace period is in progress, at least until the corresponding
854 * leaf node has been initialized. In addition, we have excluded
855 * CPU-hotplug operations.
857 * Note that the grace period cannot complete until we finish
858 * the initialization process, as there will be at least one
859 * qsmask bit set in the root node until that time, namely the
860 * one corresponding to this CPU, due to the fact that we have
863 rcu_for_each_node_breadth_first(rsp
, rnp
) {
864 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
865 rcu_preempt_check_blocked_tasks(rnp
);
866 rnp
->qsmask
= rnp
->qsmaskinit
;
867 rnp
->gpnum
= rsp
->gpnum
;
868 rnp
->completed
= rsp
->completed
;
869 if (rnp
== rdp
->mynode
)
870 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
871 rcu_preempt_boost_start_gp(rnp
);
872 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
875 rnp
= rcu_get_root(rsp
);
876 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
877 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
878 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
879 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
883 * Report a full set of quiescent states to the specified rcu_state
884 * data structure. This involves cleaning up after the prior grace
885 * period and letting rcu_start_gp() start up the next grace period
886 * if one is needed. Note that the caller must hold rnp->lock, as
887 * required by rcu_start_gp(), which will release it.
889 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
890 __releases(rcu_get_root(rsp
)->lock
)
892 unsigned long gp_duration
;
894 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
897 * Ensure that all grace-period and pre-grace-period activity
898 * is seen before the assignment to rsp->completed.
900 smp_mb(); /* See above block comment. */
901 gp_duration
= jiffies
- rsp
->gp_start
;
902 if (gp_duration
> rsp
->gp_max
)
903 rsp
->gp_max
= gp_duration
;
904 rsp
->completed
= rsp
->gpnum
;
905 rsp
->signaled
= RCU_GP_IDLE
;
906 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
910 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
911 * Allows quiescent states for a group of CPUs to be reported at one go
912 * to the specified rcu_node structure, though all the CPUs in the group
913 * must be represented by the same rcu_node structure (which need not be
914 * a leaf rcu_node structure, though it often will be). That structure's
915 * lock must be held upon entry, and it is released before return.
918 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
919 struct rcu_node
*rnp
, unsigned long flags
)
920 __releases(rnp
->lock
)
922 struct rcu_node
*rnp_c
;
924 /* Walk up the rcu_node hierarchy. */
926 if (!(rnp
->qsmask
& mask
)) {
928 /* Our bit has already been cleared, so done. */
929 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
932 rnp
->qsmask
&= ~mask
;
933 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
935 /* Other bits still set at this level, so done. */
936 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
940 if (rnp
->parent
== NULL
) {
942 /* No more levels. Exit loop holding root lock. */
946 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
949 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
950 WARN_ON_ONCE(rnp_c
->qsmask
);
954 * Get here if we are the last CPU to pass through a quiescent
955 * state for this grace period. Invoke rcu_report_qs_rsp()
956 * to clean up and start the next grace period if one is needed.
958 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
962 * Record a quiescent state for the specified CPU to that CPU's rcu_data
963 * structure. This must be either called from the specified CPU, or
964 * called when the specified CPU is known to be offline (and when it is
965 * also known that no other CPU is concurrently trying to help the offline
966 * CPU). The lastcomp argument is used to make sure we are still in the
967 * grace period of interest. We don't want to end the current grace period
968 * based on quiescent states detected in an earlier grace period!
971 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
975 struct rcu_node
*rnp
;
978 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
979 if (lastcomp
!= rnp
->completed
) {
982 * Someone beat us to it for this grace period, so leave.
983 * The race with GP start is resolved by the fact that we
984 * hold the leaf rcu_node lock, so that the per-CPU bits
985 * cannot yet be initialized -- so we would simply find our
986 * CPU's bit already cleared in rcu_report_qs_rnp() if this
989 rdp
->passed_quiesc
= 0; /* try again later! */
990 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
994 if ((rnp
->qsmask
& mask
) == 0) {
995 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1000 * This GP can't end until cpu checks in, so all of our
1001 * callbacks can be processed during the next GP.
1003 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1005 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1010 * Check to see if there is a new grace period of which this CPU
1011 * is not yet aware, and if so, set up local rcu_data state for it.
1012 * Otherwise, see if this CPU has just passed through its first
1013 * quiescent state for this grace period, and record that fact if so.
1016 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1018 /* If there is now a new grace period, record and return. */
1019 if (check_for_new_grace_period(rsp
, rdp
))
1023 * Does this CPU still need to do its part for current grace period?
1024 * If no, return and let the other CPUs do their part as well.
1026 if (!rdp
->qs_pending
)
1030 * Was there a quiescent state since the beginning of the grace
1031 * period? If no, then exit and wait for the next call.
1033 if (!rdp
->passed_quiesc
)
1037 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1040 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
1043 #ifdef CONFIG_HOTPLUG_CPU
1046 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1047 * Synchronization is not required because this function executes
1048 * in stop_machine() context.
1050 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1053 /* current DYING CPU is cleared in the cpu_online_mask */
1054 int receive_cpu
= cpumask_any(cpu_online_mask
);
1055 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1056 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1058 if (rdp
->nxtlist
== NULL
)
1059 return; /* irqs disabled, so comparison is stable. */
1061 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1062 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1063 receive_rdp
->qlen
+= rdp
->qlen
;
1064 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1065 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1067 rdp
->nxtlist
= NULL
;
1068 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1069 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1074 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1075 * and move all callbacks from the outgoing CPU to the current one.
1076 * There can only be one CPU hotplug operation at a time, so no other
1077 * CPU can be attempting to update rcu_cpu_kthread_task.
1079 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1081 unsigned long flags
;
1083 int need_report
= 0;
1084 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1085 struct rcu_node
*rnp
;
1086 struct task_struct
*t
;
1088 /* Stop the CPU's kthread. */
1089 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1091 per_cpu(rcu_cpu_kthread_task
, cpu
) = NULL
;
1095 /* Exclude any attempts to start a new grace period. */
1096 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1098 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1099 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1100 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1102 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1103 rnp
->qsmaskinit
&= ~mask
;
1104 if (rnp
->qsmaskinit
!= 0) {
1105 if (rnp
!= rdp
->mynode
)
1106 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1109 if (rnp
== rdp
->mynode
)
1110 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1112 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1113 mask
= rnp
->grpmask
;
1115 } while (rnp
!= NULL
);
1118 * We still hold the leaf rcu_node structure lock here, and
1119 * irqs are still disabled. The reason for this subterfuge is
1120 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1121 * held leads to deadlock.
1123 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1125 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1126 rcu_report_unblock_qs_rnp(rnp
, flags
);
1128 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1129 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1130 rcu_report_exp_rnp(rsp
, rnp
);
1133 * If there are no more online CPUs for this rcu_node structure,
1134 * kill the rcu_node structure's kthread. Otherwise, adjust its
1137 t
= rnp
->node_kthread_task
;
1139 rnp
->qsmaskinit
== 0) {
1140 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1141 rnp
->node_kthread_task
= NULL
;
1142 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1144 rcu_stop_boost_kthread(rnp
);
1146 rcu_node_kthread_setaffinity(rnp
, -1);
1150 * Remove the specified CPU from the RCU hierarchy and move any pending
1151 * callbacks that it might have to the current CPU. This code assumes
1152 * that at least one CPU in the system will remain running at all times.
1153 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1155 static void rcu_offline_cpu(int cpu
)
1157 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1158 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1159 rcu_preempt_offline_cpu(cpu
);
1162 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1164 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1168 static void rcu_offline_cpu(int cpu
)
1172 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1175 * Invoke any RCU callbacks that have made it to the end of their grace
1176 * period. Thottle as specified by rdp->blimit.
1178 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1180 unsigned long flags
;
1181 struct rcu_head
*next
, *list
, **tail
;
1184 /* If no callbacks are ready, just return.*/
1185 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1189 * Extract the list of ready callbacks, disabling to prevent
1190 * races with call_rcu() from interrupt handlers.
1192 local_irq_save(flags
);
1193 list
= rdp
->nxtlist
;
1194 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1195 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1196 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1197 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1198 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1199 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1200 local_irq_restore(flags
);
1202 /* Invoke callbacks. */
1207 debug_rcu_head_unqueue(list
);
1210 if (++count
>= rdp
->blimit
)
1214 local_irq_save(flags
);
1216 /* Update count, and requeue any remaining callbacks. */
1218 rdp
->n_cbs_invoked
+= count
;
1220 *tail
= rdp
->nxtlist
;
1221 rdp
->nxtlist
= list
;
1222 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1223 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1224 rdp
->nxttail
[count
] = tail
;
1229 /* Reinstate batch limit if we have worked down the excess. */
1230 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1231 rdp
->blimit
= blimit
;
1233 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1234 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1235 rdp
->qlen_last_fqs_check
= 0;
1236 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1237 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1238 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1240 local_irq_restore(flags
);
1242 /* Re-raise the RCU softirq if there are callbacks remaining. */
1243 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1244 invoke_rcu_cpu_kthread();
1248 * Check to see if this CPU is in a non-context-switch quiescent state
1249 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1250 * Also schedule the RCU softirq handler.
1252 * This function must be called with hardirqs disabled. It is normally
1253 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1254 * false, there is no point in invoking rcu_check_callbacks().
1256 void rcu_check_callbacks(int cpu
, int user
)
1259 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1260 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1263 * Get here if this CPU took its interrupt from user
1264 * mode or from the idle loop, and if this is not a
1265 * nested interrupt. In this case, the CPU is in
1266 * a quiescent state, so note it.
1268 * No memory barrier is required here because both
1269 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1270 * variables that other CPUs neither access nor modify,
1271 * at least not while the corresponding CPU is online.
1277 } else if (!in_softirq()) {
1280 * Get here if this CPU did not take its interrupt from
1281 * softirq, in other words, if it is not interrupting
1282 * a rcu_bh read-side critical section. This is an _bh
1283 * critical section, so note it.
1288 rcu_preempt_check_callbacks(cpu
);
1289 if (rcu_pending(cpu
))
1290 invoke_rcu_cpu_kthread();
1296 * Scan the leaf rcu_node structures, processing dyntick state for any that
1297 * have not yet encountered a quiescent state, using the function specified.
1298 * Also initiate boosting for any threads blocked on the root rcu_node.
1300 * The caller must have suppressed start of new grace periods.
1302 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1306 unsigned long flags
;
1308 struct rcu_node
*rnp
;
1310 rcu_for_each_leaf_node(rsp
, rnp
) {
1312 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1313 if (!rcu_gp_in_progress(rsp
)) {
1314 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1317 if (rnp
->qsmask
== 0) {
1318 rcu_initiate_boost(rnp
);
1319 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1324 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1325 if ((rnp
->qsmask
& bit
) != 0 &&
1326 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1331 /* rcu_report_qs_rnp() releases rnp->lock. */
1332 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1335 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1337 rnp
= rcu_get_root(rsp
);
1338 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1339 if (rnp
->qsmask
== 0)
1340 rcu_initiate_boost(rnp
);
1341 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1345 * Force quiescent states on reluctant CPUs, and also detect which
1346 * CPUs are in dyntick-idle mode.
1348 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1350 unsigned long flags
;
1351 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1353 if (!rcu_gp_in_progress(rsp
))
1354 return; /* No grace period in progress, nothing to force. */
1355 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1356 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1357 return; /* Someone else is already on the job. */
1359 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1360 goto unlock_fqs_ret
; /* no emergency and done recently. */
1362 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1363 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1364 if(!rcu_gp_in_progress(rsp
)) {
1365 rsp
->n_force_qs_ngp
++;
1366 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1367 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1369 rsp
->fqs_active
= 1;
1370 switch (rsp
->signaled
) {
1374 break; /* grace period idle or initializing, ignore. */
1376 case RCU_SAVE_DYNTICK
:
1377 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1378 break; /* So gcc recognizes the dead code. */
1380 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1382 /* Record dyntick-idle state. */
1383 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1384 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1385 if (rcu_gp_in_progress(rsp
))
1386 rsp
->signaled
= RCU_FORCE_QS
;
1391 /* Check dyntick-idle state, send IPI to laggarts. */
1392 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1393 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1395 /* Leave state in case more forcing is required. */
1397 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1400 rsp
->fqs_active
= 0;
1401 if (rsp
->fqs_need_gp
) {
1402 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1403 rsp
->fqs_need_gp
= 0;
1404 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1407 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1409 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1412 #else /* #ifdef CONFIG_SMP */
1414 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1419 #endif /* #else #ifdef CONFIG_SMP */
1422 * This does the RCU processing work from softirq context for the
1423 * specified rcu_state and rcu_data structures. This may be called
1424 * only from the CPU to whom the rdp belongs.
1427 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1429 unsigned long flags
;
1431 WARN_ON_ONCE(rdp
->beenonline
== 0);
1434 * If an RCU GP has gone long enough, go check for dyntick
1435 * idle CPUs and, if needed, send resched IPIs.
1437 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1438 force_quiescent_state(rsp
, 1);
1441 * Advance callbacks in response to end of earlier grace
1442 * period that some other CPU ended.
1444 rcu_process_gp_end(rsp
, rdp
);
1446 /* Update RCU state based on any recent quiescent states. */
1447 rcu_check_quiescent_state(rsp
, rdp
);
1449 /* Does this CPU require a not-yet-started grace period? */
1450 if (cpu_needs_another_gp(rsp
, rdp
)) {
1451 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1452 rcu_start_gp(rsp
, flags
); /* releases above lock */
1455 /* If there are callbacks ready, invoke them. */
1456 rcu_do_batch(rsp
, rdp
);
1460 * Do softirq processing for the current CPU.
1462 static void rcu_process_callbacks(void)
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();
1469 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1470 rcu_needs_cpu_flush();
1474 * Wake up the current CPU's kthread. This replaces raise_softirq()
1475 * in earlier versions of RCU. Note that because we are running on
1476 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1477 * cannot disappear out from under us.
1479 static void invoke_rcu_cpu_kthread(void)
1481 unsigned long flags
;
1482 wait_queue_head_t
*q
;
1485 local_irq_save(flags
);
1486 cpu
= smp_processor_id();
1487 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1488 if (per_cpu(rcu_cpu_kthread_task
, cpu
) == NULL
) {
1489 local_irq_restore(flags
);
1492 q
= &per_cpu(rcu_cpu_wq
, cpu
);
1494 local_irq_restore(flags
);
1498 * Wake up the specified per-rcu_node-structure kthread.
1499 * The caller must hold ->lock.
1501 static void invoke_rcu_node_kthread(struct rcu_node
*rnp
)
1503 struct task_struct
*t
;
1505 t
= rnp
->node_kthread_task
;
1511 * Set the specified CPU's kthread to run RT or not, as specified by
1512 * the to_rt argument. The CPU-hotplug locks are held, so the task
1513 * is not going away.
1515 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1518 struct sched_param sp
;
1519 struct task_struct
*t
;
1521 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1525 policy
= SCHED_FIFO
;
1526 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1528 policy
= SCHED_NORMAL
;
1529 sp
.sched_priority
= 0;
1531 sched_setscheduler_nocheck(t
, policy
, &sp
);
1535 * Timer handler to initiate the waking up of per-CPU kthreads that
1536 * have yielded the CPU due to excess numbers of RCU callbacks.
1537 * We wake up the per-rcu_node kthread, which in turn will wake up
1538 * the booster kthread.
1540 static void rcu_cpu_kthread_timer(unsigned long arg
)
1542 unsigned long flags
;
1543 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, arg
);
1544 struct rcu_node
*rnp
= rdp
->mynode
;
1546 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1547 rnp
->wakemask
|= rdp
->grpmask
;
1548 invoke_rcu_node_kthread(rnp
);
1549 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1553 * Drop to non-real-time priority and yield, but only after posting a
1554 * timer that will cause us to regain our real-time priority if we
1555 * remain preempted. Either way, we restore our real-time priority
1558 static void rcu_yield(void (*f
)(unsigned long), unsigned long arg
)
1560 struct sched_param sp
;
1561 struct timer_list yield_timer
;
1563 setup_timer_on_stack(&yield_timer
, f
, arg
);
1564 mod_timer(&yield_timer
, jiffies
+ 2);
1565 sp
.sched_priority
= 0;
1566 sched_setscheduler_nocheck(current
, SCHED_NORMAL
, &sp
);
1568 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1569 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1570 del_timer(&yield_timer
);
1574 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1575 * This can happen while the corresponding CPU is either coming online
1576 * or going offline. We cannot wait until the CPU is fully online
1577 * before starting the kthread, because the various notifier functions
1578 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1579 * the corresponding CPU is online.
1581 * Return 1 if the kthread needs to stop, 0 otherwise.
1583 * Caller must disable bh. This function can momentarily enable it.
1585 static int rcu_cpu_kthread_should_stop(int cpu
)
1587 while (cpu_is_offline(cpu
) ||
1588 !cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)) ||
1589 smp_processor_id() != cpu
) {
1590 if (kthread_should_stop())
1592 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1593 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = raw_smp_processor_id();
1595 schedule_timeout_uninterruptible(1);
1596 if (!cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)))
1597 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
1600 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1605 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1606 * earlier RCU softirq.
1608 static int rcu_cpu_kthread(void *arg
)
1610 int cpu
= (int)(long)arg
;
1611 unsigned long flags
;
1613 unsigned int *statusp
= &per_cpu(rcu_cpu_kthread_status
, cpu
);
1614 wait_queue_head_t
*wqp
= &per_cpu(rcu_cpu_wq
, cpu
);
1616 char *workp
= &per_cpu(rcu_cpu_has_work
, cpu
);
1619 *statusp
= RCU_KTHREAD_WAITING
;
1620 wait_event_interruptible(*wqp
,
1621 *workp
!= 0 || kthread_should_stop());
1623 if (rcu_cpu_kthread_should_stop(cpu
)) {
1627 *statusp
= RCU_KTHREAD_RUNNING
;
1628 local_irq_save(flags
);
1631 local_irq_restore(flags
);
1633 rcu_process_callbacks();
1640 *statusp
= RCU_KTHREAD_YIELDING
;
1641 rcu_yield(rcu_cpu_kthread_timer
, (unsigned long)cpu
);
1645 *statusp
= RCU_KTHREAD_STOPPED
;
1650 * Spawn a per-CPU kthread, setting up affinity and priority.
1651 * Because the CPU hotplug lock is held, no other CPU will be attempting
1652 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1653 * attempting to access it during boot, but the locking in kthread_bind()
1654 * will enforce sufficient ordering.
1656 static int __cpuinit
rcu_spawn_one_cpu_kthread(int cpu
)
1658 struct sched_param sp
;
1659 struct task_struct
*t
;
1661 if (!rcu_kthreads_spawnable
||
1662 per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
)
1664 t
= kthread_create(rcu_cpu_kthread
, (void *)(long)cpu
, "rcuc%d", cpu
);
1667 kthread_bind(t
, cpu
);
1668 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1669 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
);
1670 per_cpu(rcu_cpu_kthread_task
, cpu
) = t
;
1672 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1673 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1678 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1679 * kthreads when needed. We ignore requests to wake up kthreads
1680 * for offline CPUs, which is OK because force_quiescent_state()
1681 * takes care of this case.
1683 static int rcu_node_kthread(void *arg
)
1686 unsigned long flags
;
1688 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1689 struct sched_param sp
;
1690 struct task_struct
*t
;
1693 rnp
->node_kthread_status
= RCU_KTHREAD_WAITING
;
1694 wait_event_interruptible(rnp
->node_wq
, rnp
->wakemask
!= 0 ||
1695 kthread_should_stop());
1696 if (kthread_should_stop())
1698 rnp
->node_kthread_status
= RCU_KTHREAD_RUNNING
;
1699 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1700 mask
= rnp
->wakemask
;
1702 rcu_initiate_boost(rnp
);
1703 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1704 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1) {
1705 if ((mask
& 0x1) == 0)
1708 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1709 if (!cpu_online(cpu
) || t
== NULL
) {
1713 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1714 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1715 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1719 rnp
->node_kthread_status
= RCU_KTHREAD_STOPPED
;
1724 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1725 * served by the rcu_node in question. The CPU hotplug lock is still
1726 * held, so the value of rnp->qsmaskinit will be stable.
1728 * We don't include outgoingcpu in the affinity set, use -1 if there is
1729 * no outgoing CPU. If there are no CPUs left in the affinity set,
1730 * this function allows the kthread to execute on any CPU.
1732 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1736 unsigned long mask
= rnp
->qsmaskinit
;
1738 if (rnp
->node_kthread_task
== NULL
|| mask
== 0)
1740 if (!alloc_cpumask_var(&cm
, GFP_KERNEL
))
1743 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1744 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1745 cpumask_set_cpu(cpu
, cm
);
1746 if (cpumask_weight(cm
) == 0) {
1748 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1749 cpumask_clear_cpu(cpu
, cm
);
1750 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1752 set_cpus_allowed_ptr(rnp
->node_kthread_task
, cm
);
1753 rcu_boost_kthread_setaffinity(rnp
, cm
);
1754 free_cpumask_var(cm
);
1758 * Spawn a per-rcu_node kthread, setting priority and affinity.
1759 * Called during boot before online/offline can happen, or, if
1760 * during runtime, with the main CPU-hotplug locks held. So only
1761 * one of these can be executing at a time.
1763 static int __cpuinit
rcu_spawn_one_node_kthread(struct rcu_state
*rsp
,
1764 struct rcu_node
*rnp
)
1766 unsigned long flags
;
1767 int rnp_index
= rnp
- &rsp
->node
[0];
1768 struct sched_param sp
;
1769 struct task_struct
*t
;
1771 if (!rcu_kthreads_spawnable
||
1772 rnp
->qsmaskinit
== 0)
1774 if (rnp
->node_kthread_task
== NULL
) {
1775 t
= kthread_create(rcu_node_kthread
, (void *)rnp
,
1776 "rcun%d", rnp_index
);
1779 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1780 rnp
->node_kthread_task
= t
;
1781 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1783 sp
.sched_priority
= 99;
1784 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1786 return rcu_spawn_one_boost_kthread(rsp
, rnp
, rnp_index
);
1790 * Spawn all kthreads -- called as soon as the scheduler is running.
1792 static int __init
rcu_spawn_kthreads(void)
1795 struct rcu_node
*rnp
;
1797 rcu_kthreads_spawnable
= 1;
1798 for_each_possible_cpu(cpu
) {
1799 init_waitqueue_head(&per_cpu(rcu_cpu_wq
, cpu
));
1800 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1801 if (cpu_online(cpu
))
1802 (void)rcu_spawn_one_cpu_kthread(cpu
);
1804 rnp
= rcu_get_root(rcu_state
);
1805 init_waitqueue_head(&rnp
->node_wq
);
1806 rcu_init_boost_waitqueue(rnp
);
1807 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1808 if (NUM_RCU_NODES
> 1)
1809 rcu_for_each_leaf_node(rcu_state
, rnp
) {
1810 init_waitqueue_head(&rnp
->node_wq
);
1811 rcu_init_boost_waitqueue(rnp
);
1812 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1816 early_initcall(rcu_spawn_kthreads
);
1819 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1820 struct rcu_state
*rsp
)
1822 unsigned long flags
;
1823 struct rcu_data
*rdp
;
1825 debug_rcu_head_queue(head
);
1829 smp_mb(); /* Ensure RCU update seen before callback registry. */
1832 * Opportunistically note grace-period endings and beginnings.
1833 * Note that we might see a beginning right after we see an
1834 * end, but never vice versa, since this CPU has to pass through
1835 * a quiescent state betweentimes.
1837 local_irq_save(flags
);
1838 rdp
= this_cpu_ptr(rsp
->rda
);
1840 /* Add the callback to our list. */
1841 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1842 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1845 * Force the grace period if too many callbacks or too long waiting.
1846 * Enforce hysteresis, and don't invoke force_quiescent_state()
1847 * if some other CPU has recently done so. Also, don't bother
1848 * invoking force_quiescent_state() if the newly enqueued callback
1849 * is the only one waiting for a grace period to complete.
1851 if (unlikely(++rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1853 /* Are we ignoring a completed grace period? */
1854 rcu_process_gp_end(rsp
, rdp
);
1855 check_for_new_grace_period(rsp
, rdp
);
1857 /* Start a new grace period if one not already started. */
1858 if (!rcu_gp_in_progress(rsp
)) {
1859 unsigned long nestflag
;
1860 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1862 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1863 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1865 /* Give the grace period a kick. */
1866 rdp
->blimit
= LONG_MAX
;
1867 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1868 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1869 force_quiescent_state(rsp
, 0);
1870 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1871 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1873 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1874 force_quiescent_state(rsp
, 1);
1875 local_irq_restore(flags
);
1879 * Queue an RCU-sched callback for invocation after a grace period.
1881 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1883 __call_rcu(head
, func
, &rcu_sched_state
);
1885 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1888 * Queue an RCU for invocation after a quicker grace period.
1890 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1892 __call_rcu(head
, func
, &rcu_bh_state
);
1894 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1897 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1899 * Control will return to the caller some time after a full rcu-sched
1900 * grace period has elapsed, in other words after all currently executing
1901 * rcu-sched read-side critical sections have completed. These read-side
1902 * critical sections are delimited by rcu_read_lock_sched() and
1903 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1904 * local_irq_disable(), and so on may be used in place of
1905 * rcu_read_lock_sched().
1907 * This means that all preempt_disable code sequences, including NMI and
1908 * hardware-interrupt handlers, in progress on entry will have completed
1909 * before this primitive returns. However, this does not guarantee that
1910 * softirq handlers will have completed, since in some kernels, these
1911 * handlers can run in process context, and can block.
1913 * This primitive provides the guarantees made by the (now removed)
1914 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1915 * guarantees that rcu_read_lock() sections will have completed.
1916 * In "classic RCU", these two guarantees happen to be one and
1917 * the same, but can differ in realtime RCU implementations.
1919 void synchronize_sched(void)
1921 struct rcu_synchronize rcu
;
1923 if (rcu_blocking_is_gp())
1926 init_rcu_head_on_stack(&rcu
.head
);
1927 init_completion(&rcu
.completion
);
1928 /* Will wake me after RCU finished. */
1929 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1931 wait_for_completion(&rcu
.completion
);
1932 destroy_rcu_head_on_stack(&rcu
.head
);
1934 EXPORT_SYMBOL_GPL(synchronize_sched
);
1937 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1939 * Control will return to the caller some time after a full rcu_bh grace
1940 * period has elapsed, in other words after all currently executing rcu_bh
1941 * read-side critical sections have completed. RCU read-side critical
1942 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1943 * and may be nested.
1945 void synchronize_rcu_bh(void)
1947 struct rcu_synchronize rcu
;
1949 if (rcu_blocking_is_gp())
1952 init_rcu_head_on_stack(&rcu
.head
);
1953 init_completion(&rcu
.completion
);
1954 /* Will wake me after RCU finished. */
1955 call_rcu_bh(&rcu
.head
, wakeme_after_rcu
);
1957 wait_for_completion(&rcu
.completion
);
1958 destroy_rcu_head_on_stack(&rcu
.head
);
1960 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1963 * Check to see if there is any immediate RCU-related work to be done
1964 * by the current CPU, for the specified type of RCU, returning 1 if so.
1965 * The checks are in order of increasing expense: checks that can be
1966 * carried out against CPU-local state are performed first. However,
1967 * we must check for CPU stalls first, else we might not get a chance.
1969 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1971 struct rcu_node
*rnp
= rdp
->mynode
;
1973 rdp
->n_rcu_pending
++;
1975 /* Check for CPU stalls, if enabled. */
1976 check_cpu_stall(rsp
, rdp
);
1978 /* Is the RCU core waiting for a quiescent state from this CPU? */
1979 if (rdp
->qs_pending
&& !rdp
->passed_quiesc
) {
1982 * If force_quiescent_state() coming soon and this CPU
1983 * needs a quiescent state, and this is either RCU-sched
1984 * or RCU-bh, force a local reschedule.
1986 rdp
->n_rp_qs_pending
++;
1987 if (!rdp
->preemptable
&&
1988 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
1991 } else if (rdp
->qs_pending
&& rdp
->passed_quiesc
) {
1992 rdp
->n_rp_report_qs
++;
1996 /* Does this CPU have callbacks ready to invoke? */
1997 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1998 rdp
->n_rp_cb_ready
++;
2002 /* Has RCU gone idle with this CPU needing another grace period? */
2003 if (cpu_needs_another_gp(rsp
, rdp
)) {
2004 rdp
->n_rp_cpu_needs_gp
++;
2008 /* Has another RCU grace period completed? */
2009 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2010 rdp
->n_rp_gp_completed
++;
2014 /* Has a new RCU grace period started? */
2015 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2016 rdp
->n_rp_gp_started
++;
2020 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2021 if (rcu_gp_in_progress(rsp
) &&
2022 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
2023 rdp
->n_rp_need_fqs
++;
2028 rdp
->n_rp_need_nothing
++;
2033 * Check to see if there is any immediate RCU-related work to be done
2034 * by the current CPU, returning 1 if so. This function is part of the
2035 * RCU implementation; it is -not- an exported member of the RCU API.
2037 static int rcu_pending(int cpu
)
2039 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
2040 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
2041 rcu_preempt_pending(cpu
);
2045 * Check to see if any future RCU-related work will need to be done
2046 * by the current CPU, even if none need be done immediately, returning
2049 static int rcu_needs_cpu_quick_check(int cpu
)
2051 /* RCU callbacks either ready or pending? */
2052 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
2053 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
2054 rcu_preempt_needs_cpu(cpu
);
2057 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
2058 static atomic_t rcu_barrier_cpu_count
;
2059 static DEFINE_MUTEX(rcu_barrier_mutex
);
2060 static struct completion rcu_barrier_completion
;
2062 static void rcu_barrier_callback(struct rcu_head
*notused
)
2064 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2065 complete(&rcu_barrier_completion
);
2069 * Called with preemption disabled, and from cross-cpu IRQ context.
2071 static void rcu_barrier_func(void *type
)
2073 int cpu
= smp_processor_id();
2074 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
2075 void (*call_rcu_func
)(struct rcu_head
*head
,
2076 void (*func
)(struct rcu_head
*head
));
2078 atomic_inc(&rcu_barrier_cpu_count
);
2079 call_rcu_func
= type
;
2080 call_rcu_func(head
, rcu_barrier_callback
);
2084 * Orchestrate the specified type of RCU barrier, waiting for all
2085 * RCU callbacks of the specified type to complete.
2087 static void _rcu_barrier(struct rcu_state
*rsp
,
2088 void (*call_rcu_func
)(struct rcu_head
*head
,
2089 void (*func
)(struct rcu_head
*head
)))
2091 BUG_ON(in_interrupt());
2092 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2093 mutex_lock(&rcu_barrier_mutex
);
2094 init_completion(&rcu_barrier_completion
);
2096 * Initialize rcu_barrier_cpu_count to 1, then invoke
2097 * rcu_barrier_func() on each CPU, so that each CPU also has
2098 * incremented rcu_barrier_cpu_count. Only then is it safe to
2099 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2100 * might complete its grace period before all of the other CPUs
2101 * did their increment, causing this function to return too
2102 * early. Note that on_each_cpu() disables irqs, which prevents
2103 * any CPUs from coming online or going offline until each online
2104 * CPU has queued its RCU-barrier callback.
2106 atomic_set(&rcu_barrier_cpu_count
, 1);
2107 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
2108 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2109 complete(&rcu_barrier_completion
);
2110 wait_for_completion(&rcu_barrier_completion
);
2111 mutex_unlock(&rcu_barrier_mutex
);
2115 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2117 void rcu_barrier_bh(void)
2119 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
2121 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2124 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2126 void rcu_barrier_sched(void)
2128 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2130 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2133 * Do boot-time initialization of a CPU's per-CPU RCU data.
2136 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2138 unsigned long flags
;
2140 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2141 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2143 /* Set up local state, ensuring consistent view of global state. */
2144 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2145 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2146 rdp
->nxtlist
= NULL
;
2147 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2148 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2151 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2152 #endif /* #ifdef CONFIG_NO_HZ */
2154 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2158 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2159 * offline event can be happening at a given time. Note also that we
2160 * can accept some slop in the rsp->completed access due to the fact
2161 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2163 static void __cpuinit
2164 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptable
)
2166 unsigned long flags
;
2168 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2169 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2171 /* Set up local state, ensuring consistent view of global state. */
2172 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2173 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
2174 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
2175 rdp
->beenonline
= 1; /* We have now been online. */
2176 rdp
->preemptable
= preemptable
;
2177 rdp
->qlen_last_fqs_check
= 0;
2178 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2179 rdp
->blimit
= blimit
;
2180 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2183 * A new grace period might start here. If so, we won't be part
2184 * of it, but that is OK, as we are currently in a quiescent state.
2187 /* Exclude any attempts to start a new GP on large systems. */
2188 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2190 /* Add CPU to rcu_node bitmasks. */
2192 mask
= rdp
->grpmask
;
2194 /* Exclude any attempts to start a new GP on small systems. */
2195 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2196 rnp
->qsmaskinit
|= mask
;
2197 mask
= rnp
->grpmask
;
2198 if (rnp
== rdp
->mynode
) {
2199 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
2200 rdp
->completed
= rnp
->completed
;
2201 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
2203 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2205 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2207 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2210 static void __cpuinit
rcu_online_cpu(int cpu
)
2212 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2213 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2214 rcu_preempt_init_percpu_data(cpu
);
2217 static void __cpuinit
rcu_online_kthreads(int cpu
)
2219 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2220 struct rcu_node
*rnp
= rdp
->mynode
;
2222 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
2223 if (rcu_kthreads_spawnable
) {
2224 (void)rcu_spawn_one_cpu_kthread(cpu
);
2225 if (rnp
->node_kthread_task
== NULL
)
2226 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
2231 * Handle CPU online/offline notification events.
2233 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2234 unsigned long action
, void *hcpu
)
2236 long cpu
= (long)hcpu
;
2237 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2238 struct rcu_node
*rnp
= rdp
->mynode
;
2241 case CPU_UP_PREPARE
:
2242 case CPU_UP_PREPARE_FROZEN
:
2243 rcu_online_cpu(cpu
);
2244 rcu_online_kthreads(cpu
);
2247 case CPU_DOWN_FAILED
:
2248 rcu_node_kthread_setaffinity(rnp
, -1);
2249 rcu_cpu_kthread_setrt(cpu
, 1);
2251 case CPU_DOWN_PREPARE
:
2252 rcu_node_kthread_setaffinity(rnp
, cpu
);
2253 rcu_cpu_kthread_setrt(cpu
, 0);
2256 case CPU_DYING_FROZEN
:
2258 * The whole machine is "stopped" except this CPU, so we can
2259 * touch any data without introducing corruption. We send the
2260 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2262 rcu_send_cbs_to_online(&rcu_bh_state
);
2263 rcu_send_cbs_to_online(&rcu_sched_state
);
2264 rcu_preempt_send_cbs_to_online();
2267 case CPU_DEAD_FROZEN
:
2268 case CPU_UP_CANCELED
:
2269 case CPU_UP_CANCELED_FROZEN
:
2270 rcu_offline_cpu(cpu
);
2279 * This function is invoked towards the end of the scheduler's initialization
2280 * process. Before this is called, the idle task might contain
2281 * RCU read-side critical sections (during which time, this idle
2282 * task is booting the system). After this function is called, the
2283 * idle tasks are prohibited from containing RCU read-side critical
2284 * sections. This function also enables RCU lockdep checking.
2286 void rcu_scheduler_starting(void)
2288 WARN_ON(num_online_cpus() != 1);
2289 WARN_ON(nr_context_switches() > 0);
2290 rcu_scheduler_active
= 1;
2294 * Compute the per-level fanout, either using the exact fanout specified
2295 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2297 #ifdef CONFIG_RCU_FANOUT_EXACT
2298 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2302 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2303 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2304 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
2306 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2307 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2314 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2315 ccur
= rsp
->levelcnt
[i
];
2316 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2320 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2323 * Helper function for rcu_init() that initializes one rcu_state structure.
2325 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2326 struct rcu_data __percpu
*rda
)
2328 static char *buf
[] = { "rcu_node_level_0",
2331 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2335 struct rcu_node
*rnp
;
2337 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2339 /* Initialize the level-tracking arrays. */
2341 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2342 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2343 rcu_init_levelspread(rsp
);
2345 /* Initialize the elements themselves, starting from the leaves. */
2347 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2348 cpustride
*= rsp
->levelspread
[i
];
2349 rnp
= rsp
->level
[i
];
2350 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2351 raw_spin_lock_init(&rnp
->lock
);
2352 lockdep_set_class_and_name(&rnp
->lock
,
2353 &rcu_node_class
[i
], buf
[i
]);
2356 rnp
->qsmaskinit
= 0;
2357 rnp
->grplo
= j
* cpustride
;
2358 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2359 if (rnp
->grphi
>= NR_CPUS
)
2360 rnp
->grphi
= NR_CPUS
- 1;
2366 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2367 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2368 rnp
->parent
= rsp
->level
[i
- 1] +
2369 j
/ rsp
->levelspread
[i
- 1];
2372 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2377 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2378 for_each_possible_cpu(i
) {
2379 while (i
> rnp
->grphi
)
2381 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2382 rcu_boot_init_percpu_data(i
, rsp
);
2386 void __init
rcu_init(void)
2390 rcu_bootup_announce();
2391 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2392 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2393 __rcu_init_preempt();
2396 * We don't need protection against CPU-hotplug here because
2397 * this is called early in boot, before either interrupts
2398 * or the scheduler are operational.
2400 cpu_notifier(rcu_cpu_notify
, 0);
2401 for_each_online_cpu(cpu
)
2402 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2403 check_cpu_stall_init();
2406 #include "rcutree_plugin.h"