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 static DEFINE_PER_CPU(wait_queue_head_t
, rcu_cpu_wq
);
96 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
97 static char rcu_kthreads_spawnable
;
99 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
100 static void invoke_rcu_cpu_kthread(void);
102 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
105 * Track the rcutorture test sequence number and the update version
106 * number within a given test. The rcutorture_testseq is incremented
107 * on every rcutorture module load and unload, so has an odd value
108 * when a test is running. The rcutorture_vernum is set to zero
109 * when rcutorture starts and is incremented on each rcutorture update.
110 * These variables enable correlating rcutorture output with the
111 * RCU tracing information.
113 unsigned long rcutorture_testseq
;
114 unsigned long rcutorture_vernum
;
117 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
118 * permit this function to be invoked without holding the root rcu_node
119 * structure's ->lock, but of course results can be subject to change.
121 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
123 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
127 * Note a quiescent state. Because we do not need to know
128 * how many quiescent states passed, just if there was at least
129 * one since the start of the grace period, this just sets a flag.
131 void rcu_sched_qs(int cpu
)
133 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
135 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
137 rdp
->passed_quiesc
= 1;
140 void rcu_bh_qs(int cpu
)
142 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
144 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
146 rdp
->passed_quiesc
= 1;
150 * Note a context switch. This is a quiescent state for RCU-sched,
151 * and requires special handling for preemptible RCU.
153 void rcu_note_context_switch(int cpu
)
156 rcu_preempt_note_context_switch(cpu
);
160 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
161 .dynticks_nesting
= 1,
162 .dynticks
= ATOMIC_INIT(1),
164 #endif /* #ifdef CONFIG_NO_HZ */
166 static int blimit
= 10; /* Maximum callbacks per softirq. */
167 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
168 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
170 module_param(blimit
, int, 0);
171 module_param(qhimark
, int, 0);
172 module_param(qlowmark
, int, 0);
174 int rcu_cpu_stall_suppress __read_mostly
;
175 module_param(rcu_cpu_stall_suppress
, int, 0644);
177 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
178 static int rcu_pending(int cpu
);
181 * Return the number of RCU-sched batches processed thus far for debug & stats.
183 long rcu_batches_completed_sched(void)
185 return rcu_sched_state
.completed
;
187 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
190 * Return the number of RCU BH batches processed thus far for debug & stats.
192 long rcu_batches_completed_bh(void)
194 return rcu_bh_state
.completed
;
196 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
199 * Force a quiescent state for RCU BH.
201 void rcu_bh_force_quiescent_state(void)
203 force_quiescent_state(&rcu_bh_state
, 0);
205 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
208 * Record the number of times rcutorture tests have been initiated and
209 * terminated. This information allows the debugfs tracing stats to be
210 * correlated to the rcutorture messages, even when the rcutorture module
211 * is being repeatedly loaded and unloaded. In other words, we cannot
212 * store this state in rcutorture itself.
214 void rcutorture_record_test_transition(void)
216 rcutorture_testseq
++;
217 rcutorture_vernum
= 0;
219 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
222 * Record the number of writer passes through the current rcutorture test.
223 * This is also used to correlate debugfs tracing stats with the rcutorture
226 void rcutorture_record_progress(unsigned long vernum
)
230 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
233 * Force a quiescent state for RCU-sched.
235 void rcu_sched_force_quiescent_state(void)
237 force_quiescent_state(&rcu_sched_state
, 0);
239 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
242 * Does the CPU have callbacks ready to be invoked?
245 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
247 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
251 * Does the current CPU require a yet-as-unscheduled grace period?
254 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
256 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
260 * Return the root node of the specified rcu_state structure.
262 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
264 return &rsp
->node
[0];
270 * If the specified CPU is offline, tell the caller that it is in
271 * a quiescent state. Otherwise, whack it with a reschedule IPI.
272 * Grace periods can end up waiting on an offline CPU when that
273 * CPU is in the process of coming online -- it will be added to the
274 * rcu_node bitmasks before it actually makes it online. The same thing
275 * can happen while a CPU is in the process of coming online. Because this
276 * race is quite rare, we check for it after detecting that the grace
277 * period has been delayed rather than checking each and every CPU
278 * each and every time we start a new grace period.
280 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
283 * If the CPU is offline, it is in a quiescent state. We can
284 * trust its state not to change because interrupts are disabled.
286 if (cpu_is_offline(rdp
->cpu
)) {
291 /* If preemptable RCU, no point in sending reschedule IPI. */
292 if (rdp
->preemptable
)
295 /* The CPU is online, so send it a reschedule IPI. */
296 if (rdp
->cpu
!= smp_processor_id())
297 smp_send_reschedule(rdp
->cpu
);
304 #endif /* #ifdef CONFIG_SMP */
309 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
311 * Enter nohz mode, in other words, -leave- the mode in which RCU
312 * read-side critical sections can occur. (Though RCU read-side
313 * critical sections can occur in irq handlers in nohz mode, a possibility
314 * handled by rcu_irq_enter() and rcu_irq_exit()).
316 void rcu_enter_nohz(void)
319 struct rcu_dynticks
*rdtp
;
321 local_irq_save(flags
);
322 rdtp
= &__get_cpu_var(rcu_dynticks
);
323 if (--rdtp
->dynticks_nesting
) {
324 local_irq_restore(flags
);
327 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
328 smp_mb__before_atomic_inc(); /* See above. */
329 atomic_inc(&rdtp
->dynticks
);
330 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
331 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
332 local_irq_restore(flags
);
334 /* If the interrupt queued a callback, get out of dyntick mode. */
336 (__get_cpu_var(rcu_sched_data
).nxtlist
||
337 __get_cpu_var(rcu_bh_data
).nxtlist
||
338 rcu_preempt_needs_cpu(smp_processor_id())))
343 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
345 * Exit nohz mode, in other words, -enter- the mode in which RCU
346 * read-side critical sections normally occur.
348 void rcu_exit_nohz(void)
351 struct rcu_dynticks
*rdtp
;
353 local_irq_save(flags
);
354 rdtp
= &__get_cpu_var(rcu_dynticks
);
355 if (rdtp
->dynticks_nesting
++) {
356 local_irq_restore(flags
);
359 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
360 atomic_inc(&rdtp
->dynticks
);
361 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
362 smp_mb__after_atomic_inc(); /* See above. */
363 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
364 local_irq_restore(flags
);
368 * rcu_nmi_enter - inform RCU of entry to NMI context
370 * If the CPU was idle with dynamic ticks active, and there is no
371 * irq handler running, this updates rdtp->dynticks_nmi to let the
372 * RCU grace-period handling know that the CPU is active.
374 void rcu_nmi_enter(void)
376 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
378 if (rdtp
->dynticks_nmi_nesting
== 0 &&
379 (atomic_read(&rdtp
->dynticks
) & 0x1))
381 rdtp
->dynticks_nmi_nesting
++;
382 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
383 atomic_inc(&rdtp
->dynticks
);
384 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
385 smp_mb__after_atomic_inc(); /* See above. */
386 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
390 * rcu_nmi_exit - inform RCU of exit from NMI context
392 * If the CPU was idle with dynamic ticks active, and there is no
393 * irq handler running, this updates rdtp->dynticks_nmi to let the
394 * RCU grace-period handling know that the CPU is no longer active.
396 void rcu_nmi_exit(void)
398 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
400 if (rdtp
->dynticks_nmi_nesting
== 0 ||
401 --rdtp
->dynticks_nmi_nesting
!= 0)
403 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
404 smp_mb__before_atomic_inc(); /* See above. */
405 atomic_inc(&rdtp
->dynticks
);
406 smp_mb__after_atomic_inc(); /* Force delay to next write. */
407 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
411 * rcu_irq_enter - inform RCU of entry to hard irq context
413 * If the CPU was idle with dynamic ticks active, this updates the
414 * rdtp->dynticks to let the RCU handling know that the CPU is active.
416 void rcu_irq_enter(void)
422 * rcu_irq_exit - inform RCU of exit from hard irq context
424 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
425 * to put let the RCU handling be aware that the CPU is going back to idle
428 void rcu_irq_exit(void)
436 * Snapshot the specified CPU's dynticks counter so that we can later
437 * credit them with an implicit quiescent state. Return 1 if this CPU
438 * is in dynticks idle mode, which is an extended quiescent state.
440 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
442 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
447 * Return true if the specified CPU has passed through a quiescent
448 * state by virtue of being in or having passed through an dynticks
449 * idle state since the last call to dyntick_save_progress_counter()
452 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
457 curr
= (unsigned long)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
458 snap
= (unsigned long)rdp
->dynticks_snap
;
461 * If the CPU passed through or entered a dynticks idle phase with
462 * no active irq/NMI handlers, then we can safely pretend that the CPU
463 * already acknowledged the request to pass through a quiescent
464 * state. Either way, that CPU cannot possibly be in an RCU
465 * read-side critical section that started before the beginning
466 * of the current RCU grace period.
468 if ((curr
& 0x1) == 0 || ULONG_CMP_GE(curr
, snap
+ 2)) {
473 /* Go check for the CPU being offline. */
474 return rcu_implicit_offline_qs(rdp
);
477 #endif /* #ifdef CONFIG_SMP */
479 #else /* #ifdef CONFIG_NO_HZ */
483 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
488 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
490 return rcu_implicit_offline_qs(rdp
);
493 #endif /* #ifdef CONFIG_SMP */
495 #endif /* #else #ifdef CONFIG_NO_HZ */
497 int rcu_cpu_stall_suppress __read_mostly
;
499 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
501 rsp
->gp_start
= jiffies
;
502 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
505 static void print_other_cpu_stall(struct rcu_state
*rsp
)
510 struct rcu_node
*rnp
= rcu_get_root(rsp
);
512 /* Only let one CPU complain about others per time interval. */
514 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
515 delta
= jiffies
- rsp
->jiffies_stall
;
516 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
517 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
520 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
523 * Now rat on any tasks that got kicked up to the root rcu_node
524 * due to CPU offlining.
526 rcu_print_task_stall(rnp
);
527 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
530 * OK, time to rat on our buddy...
531 * See Documentation/RCU/stallwarn.txt for info on how to debug
532 * RCU CPU stall warnings.
534 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
536 rcu_for_each_leaf_node(rsp
, rnp
) {
537 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
538 rcu_print_task_stall(rnp
);
539 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
540 if (rnp
->qsmask
== 0)
542 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
543 if (rnp
->qsmask
& (1UL << cpu
))
544 printk(" %d", rnp
->grplo
+ cpu
);
546 printk("} (detected by %d, t=%ld jiffies)\n",
547 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
548 trigger_all_cpu_backtrace();
550 /* If so configured, complain about tasks blocking the grace period. */
552 rcu_print_detail_task_stall(rsp
);
554 force_quiescent_state(rsp
, 0); /* Kick them all. */
557 static void print_cpu_stall(struct rcu_state
*rsp
)
560 struct rcu_node
*rnp
= rcu_get_root(rsp
);
563 * OK, time to rat on ourselves...
564 * See Documentation/RCU/stallwarn.txt for info on how to debug
565 * RCU CPU stall warnings.
567 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
568 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
569 trigger_all_cpu_backtrace();
571 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
572 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
574 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
575 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
577 set_need_resched(); /* kick ourselves to get things going. */
580 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
583 struct rcu_node
*rnp
;
585 if (rcu_cpu_stall_suppress
)
587 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
589 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && delta
>= 0) {
591 /* We haven't checked in, so go dump stack. */
592 print_cpu_stall(rsp
);
594 } else if (rcu_gp_in_progress(rsp
) && delta
>= RCU_STALL_RAT_DELAY
) {
596 /* They had two time units to dump stack, so complain. */
597 print_other_cpu_stall(rsp
);
601 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
603 rcu_cpu_stall_suppress
= 1;
608 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
610 * Set the stall-warning timeout way off into the future, thus preventing
611 * any RCU CPU stall-warning messages from appearing in the current set of
614 * The caller must disable hard irqs.
616 void rcu_cpu_stall_reset(void)
618 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
619 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
620 rcu_preempt_stall_reset();
623 static struct notifier_block rcu_panic_block
= {
624 .notifier_call
= rcu_panic
,
627 static void __init
check_cpu_stall_init(void)
629 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
633 * Update CPU-local rcu_data state to record the newly noticed grace period.
634 * This is used both when we started the grace period and when we notice
635 * that someone else started the grace period. The caller must hold the
636 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
637 * and must have irqs disabled.
639 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
641 if (rdp
->gpnum
!= rnp
->gpnum
) {
643 * If the current grace period is waiting for this CPU,
644 * set up to detect a quiescent state, otherwise don't
645 * go looking for one.
647 rdp
->gpnum
= rnp
->gpnum
;
648 if (rnp
->qsmask
& rdp
->grpmask
) {
650 rdp
->passed_quiesc
= 0;
656 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
659 struct rcu_node
*rnp
;
661 local_irq_save(flags
);
663 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
664 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
665 local_irq_restore(flags
);
668 __note_new_gpnum(rsp
, rnp
, rdp
);
669 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
673 * Did someone else start a new RCU grace period start since we last
674 * checked? Update local state appropriately if so. Must be called
675 * on the CPU corresponding to rdp.
678 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
683 local_irq_save(flags
);
684 if (rdp
->gpnum
!= rsp
->gpnum
) {
685 note_new_gpnum(rsp
, rdp
);
688 local_irq_restore(flags
);
693 * Advance this CPU's callbacks, but only if the current grace period
694 * has ended. This may be called only from the CPU to whom the rdp
695 * belongs. In addition, the corresponding leaf rcu_node structure's
696 * ->lock must be held by the caller, with irqs disabled.
699 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
701 /* Did another grace period end? */
702 if (rdp
->completed
!= rnp
->completed
) {
704 /* Advance callbacks. No harm if list empty. */
705 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
706 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
707 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
709 /* Remember that we saw this grace-period completion. */
710 rdp
->completed
= rnp
->completed
;
713 * If we were in an extended quiescent state, we may have
714 * missed some grace periods that others CPUs handled on
715 * our behalf. Catch up with this state to avoid noting
716 * spurious new grace periods. If another grace period
717 * has started, then rnp->gpnum will have advanced, so
718 * we will detect this later on.
720 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
721 rdp
->gpnum
= rdp
->completed
;
724 * If RCU does not need a quiescent state from this CPU,
725 * then make sure that this CPU doesn't go looking for one.
727 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
733 * Advance this CPU's callbacks, but only if the current grace period
734 * has ended. This may be called only from the CPU to whom the rdp
738 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
741 struct rcu_node
*rnp
;
743 local_irq_save(flags
);
745 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
746 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
747 local_irq_restore(flags
);
750 __rcu_process_gp_end(rsp
, rnp
, rdp
);
751 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
755 * Do per-CPU grace-period initialization for running CPU. The caller
756 * must hold the lock of the leaf rcu_node structure corresponding to
760 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
762 /* Prior grace period ended, so advance callbacks for current CPU. */
763 __rcu_process_gp_end(rsp
, rnp
, rdp
);
766 * Because this CPU just now started the new grace period, we know
767 * that all of its callbacks will be covered by this upcoming grace
768 * period, even the ones that were registered arbitrarily recently.
769 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
771 * Other CPUs cannot be sure exactly when the grace period started.
772 * Therefore, their recently registered callbacks must pass through
773 * an additional RCU_NEXT_READY stage, so that they will be handled
774 * by the next RCU grace period.
776 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
777 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
779 /* Set state so that this CPU will detect the next quiescent state. */
780 __note_new_gpnum(rsp
, rnp
, rdp
);
784 * Start a new RCU grace period if warranted, re-initializing the hierarchy
785 * in preparation for detecting the next grace period. The caller must hold
786 * the root node's ->lock, which is released before return. Hard irqs must
790 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
791 __releases(rcu_get_root(rsp
)->lock
)
793 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
794 struct rcu_node
*rnp
= rcu_get_root(rsp
);
796 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
797 if (cpu_needs_another_gp(rsp
, rdp
))
798 rsp
->fqs_need_gp
= 1;
799 if (rnp
->completed
== rsp
->completed
) {
800 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
803 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
806 * Propagate new ->completed value to rcu_node structures
807 * so that other CPUs don't have to wait until the start
808 * of the next grace period to process their callbacks.
810 rcu_for_each_node_breadth_first(rsp
, rnp
) {
811 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
812 rnp
->completed
= rsp
->completed
;
813 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
815 local_irq_restore(flags
);
819 /* Advance to a new grace period and initialize state. */
821 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
822 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
823 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
824 record_gp_stall_check_time(rsp
);
826 /* Special-case the common single-level case. */
827 if (NUM_RCU_NODES
== 1) {
828 rcu_preempt_check_blocked_tasks(rnp
);
829 rnp
->qsmask
= rnp
->qsmaskinit
;
830 rnp
->gpnum
= rsp
->gpnum
;
831 rnp
->completed
= rsp
->completed
;
832 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
833 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
834 rcu_preempt_boost_start_gp(rnp
);
835 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
839 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
842 /* Exclude any concurrent CPU-hotplug operations. */
843 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
846 * Set the quiescent-state-needed bits in all the rcu_node
847 * structures for all currently online CPUs in breadth-first
848 * order, starting from the root rcu_node structure. This
849 * operation relies on the layout of the hierarchy within the
850 * rsp->node[] array. Note that other CPUs will access only
851 * the leaves of the hierarchy, which still indicate that no
852 * grace period is in progress, at least until the corresponding
853 * leaf node has been initialized. In addition, we have excluded
854 * CPU-hotplug operations.
856 * Note that the grace period cannot complete until we finish
857 * the initialization process, as there will be at least one
858 * qsmask bit set in the root node until that time, namely the
859 * one corresponding to this CPU, due to the fact that we have
862 rcu_for_each_node_breadth_first(rsp
, rnp
) {
863 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
864 rcu_preempt_check_blocked_tasks(rnp
);
865 rnp
->qsmask
= rnp
->qsmaskinit
;
866 rnp
->gpnum
= rsp
->gpnum
;
867 rnp
->completed
= rsp
->completed
;
868 if (rnp
== rdp
->mynode
)
869 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
870 rcu_preempt_boost_start_gp(rnp
);
871 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
874 rnp
= rcu_get_root(rsp
);
875 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
876 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
877 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
878 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
882 * Report a full set of quiescent states to the specified rcu_state
883 * data structure. This involves cleaning up after the prior grace
884 * period and letting rcu_start_gp() start up the next grace period
885 * if one is needed. Note that the caller must hold rnp->lock, as
886 * required by rcu_start_gp(), which will release it.
888 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
889 __releases(rcu_get_root(rsp
)->lock
)
891 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
894 * Ensure that all grace-period and pre-grace-period activity
895 * is seen before the assignment to rsp->completed.
897 smp_mb(); /* See above block comment. */
898 rsp
->completed
= rsp
->gpnum
;
899 rsp
->signaled
= RCU_GP_IDLE
;
900 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
904 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
905 * Allows quiescent states for a group of CPUs to be reported at one go
906 * to the specified rcu_node structure, though all the CPUs in the group
907 * must be represented by the same rcu_node structure (which need not be
908 * a leaf rcu_node structure, though it often will be). That structure's
909 * lock must be held upon entry, and it is released before return.
912 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
913 struct rcu_node
*rnp
, unsigned long flags
)
914 __releases(rnp
->lock
)
916 struct rcu_node
*rnp_c
;
918 /* Walk up the rcu_node hierarchy. */
920 if (!(rnp
->qsmask
& mask
)) {
922 /* Our bit has already been cleared, so done. */
923 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
926 rnp
->qsmask
&= ~mask
;
927 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
929 /* Other bits still set at this level, so done. */
930 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
934 if (rnp
->parent
== NULL
) {
936 /* No more levels. Exit loop holding root lock. */
940 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
943 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
944 WARN_ON_ONCE(rnp_c
->qsmask
);
948 * Get here if we are the last CPU to pass through a quiescent
949 * state for this grace period. Invoke rcu_report_qs_rsp()
950 * to clean up and start the next grace period if one is needed.
952 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
956 * Record a quiescent state for the specified CPU to that CPU's rcu_data
957 * structure. This must be either called from the specified CPU, or
958 * called when the specified CPU is known to be offline (and when it is
959 * also known that no other CPU is concurrently trying to help the offline
960 * CPU). The lastcomp argument is used to make sure we are still in the
961 * grace period of interest. We don't want to end the current grace period
962 * based on quiescent states detected in an earlier grace period!
965 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
969 struct rcu_node
*rnp
;
972 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
973 if (lastcomp
!= rnp
->completed
) {
976 * Someone beat us to it for this grace period, so leave.
977 * The race with GP start is resolved by the fact that we
978 * hold the leaf rcu_node lock, so that the per-CPU bits
979 * cannot yet be initialized -- so we would simply find our
980 * CPU's bit already cleared in rcu_report_qs_rnp() if this
983 rdp
->passed_quiesc
= 0; /* try again later! */
984 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
988 if ((rnp
->qsmask
& mask
) == 0) {
989 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
994 * This GP can't end until cpu checks in, so all of our
995 * callbacks can be processed during the next GP.
997 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
999 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1004 * Check to see if there is a new grace period of which this CPU
1005 * is not yet aware, and if so, set up local rcu_data state for it.
1006 * Otherwise, see if this CPU has just passed through its first
1007 * quiescent state for this grace period, and record that fact if so.
1010 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1012 /* If there is now a new grace period, record and return. */
1013 if (check_for_new_grace_period(rsp
, rdp
))
1017 * Does this CPU still need to do its part for current grace period?
1018 * If no, return and let the other CPUs do their part as well.
1020 if (!rdp
->qs_pending
)
1024 * Was there a quiescent state since the beginning of the grace
1025 * period? If no, then exit and wait for the next call.
1027 if (!rdp
->passed_quiesc
)
1031 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1034 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
1037 #ifdef CONFIG_HOTPLUG_CPU
1040 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1041 * Synchronization is not required because this function executes
1042 * in stop_machine() context.
1044 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1047 /* current DYING CPU is cleared in the cpu_online_mask */
1048 int receive_cpu
= cpumask_any(cpu_online_mask
);
1049 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1050 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1052 if (rdp
->nxtlist
== NULL
)
1053 return; /* irqs disabled, so comparison is stable. */
1055 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1056 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1057 receive_rdp
->qlen
+= rdp
->qlen
;
1058 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1059 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1061 rdp
->nxtlist
= NULL
;
1062 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1063 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1068 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1069 * and move all callbacks from the outgoing CPU to the current one.
1070 * There can only be one CPU hotplug operation at a time, so no other
1071 * CPU can be attempting to update rcu_cpu_kthread_task.
1073 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1075 unsigned long flags
;
1077 int need_report
= 0;
1078 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1079 struct rcu_node
*rnp
;
1080 struct task_struct
*t
;
1082 /* Stop the CPU's kthread. */
1083 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1085 per_cpu(rcu_cpu_kthread_task
, cpu
) = NULL
;
1089 /* Exclude any attempts to start a new grace period. */
1090 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1092 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1093 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1094 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1096 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1097 rnp
->qsmaskinit
&= ~mask
;
1098 if (rnp
->qsmaskinit
!= 0) {
1099 if (rnp
!= rdp
->mynode
)
1100 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1103 if (rnp
== rdp
->mynode
)
1104 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1106 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1107 mask
= rnp
->grpmask
;
1109 } while (rnp
!= NULL
);
1112 * We still hold the leaf rcu_node structure lock here, and
1113 * irqs are still disabled. The reason for this subterfuge is
1114 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1115 * held leads to deadlock.
1117 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1119 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1120 rcu_report_unblock_qs_rnp(rnp
, flags
);
1122 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1123 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1124 rcu_report_exp_rnp(rsp
, rnp
);
1127 * If there are no more online CPUs for this rcu_node structure,
1128 * kill the rcu_node structure's kthread. Otherwise, adjust its
1131 t
= rnp
->node_kthread_task
;
1133 rnp
->qsmaskinit
== 0) {
1134 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1135 rnp
->node_kthread_task
= NULL
;
1136 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1138 rcu_stop_boost_kthread(rnp
);
1140 rcu_node_kthread_setaffinity(rnp
, -1);
1144 * Remove the specified CPU from the RCU hierarchy and move any pending
1145 * callbacks that it might have to the current CPU. This code assumes
1146 * that at least one CPU in the system will remain running at all times.
1147 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1149 static void rcu_offline_cpu(int cpu
)
1151 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1152 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1153 rcu_preempt_offline_cpu(cpu
);
1156 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1158 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1162 static void rcu_offline_cpu(int cpu
)
1166 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1169 * Invoke any RCU callbacks that have made it to the end of their grace
1170 * period. Thottle as specified by rdp->blimit.
1172 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1174 unsigned long flags
;
1175 struct rcu_head
*next
, *list
, **tail
;
1178 /* If no callbacks are ready, just return.*/
1179 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1183 * Extract the list of ready callbacks, disabling to prevent
1184 * races with call_rcu() from interrupt handlers.
1186 local_irq_save(flags
);
1187 list
= rdp
->nxtlist
;
1188 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1189 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1190 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1191 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1192 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1193 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1194 local_irq_restore(flags
);
1196 /* Invoke callbacks. */
1201 debug_rcu_head_unqueue(list
);
1204 if (++count
>= rdp
->blimit
)
1208 local_irq_save(flags
);
1210 /* Update count, and requeue any remaining callbacks. */
1212 rdp
->n_cbs_invoked
+= count
;
1214 *tail
= rdp
->nxtlist
;
1215 rdp
->nxtlist
= list
;
1216 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1217 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1218 rdp
->nxttail
[count
] = tail
;
1223 /* Reinstate batch limit if we have worked down the excess. */
1224 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1225 rdp
->blimit
= blimit
;
1227 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1228 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1229 rdp
->qlen_last_fqs_check
= 0;
1230 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1231 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1232 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1234 local_irq_restore(flags
);
1236 /* Re-raise the RCU softirq if there are callbacks remaining. */
1237 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1238 invoke_rcu_cpu_kthread();
1242 * Check to see if this CPU is in a non-context-switch quiescent state
1243 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1244 * Also schedule the RCU softirq handler.
1246 * This function must be called with hardirqs disabled. It is normally
1247 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1248 * false, there is no point in invoking rcu_check_callbacks().
1250 void rcu_check_callbacks(int cpu
, int user
)
1253 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1254 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1257 * Get here if this CPU took its interrupt from user
1258 * mode or from the idle loop, and if this is not a
1259 * nested interrupt. In this case, the CPU is in
1260 * a quiescent state, so note it.
1262 * No memory barrier is required here because both
1263 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1264 * variables that other CPUs neither access nor modify,
1265 * at least not while the corresponding CPU is online.
1271 } else if (!in_softirq()) {
1274 * Get here if this CPU did not take its interrupt from
1275 * softirq, in other words, if it is not interrupting
1276 * a rcu_bh read-side critical section. This is an _bh
1277 * critical section, so note it.
1282 rcu_preempt_check_callbacks(cpu
);
1283 if (rcu_pending(cpu
))
1284 invoke_rcu_cpu_kthread();
1290 * Scan the leaf rcu_node structures, processing dyntick state for any that
1291 * have not yet encountered a quiescent state, using the function specified.
1292 * Also initiate boosting for any threads blocked on the root rcu_node.
1294 * The caller must have suppressed start of new grace periods.
1296 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1300 unsigned long flags
;
1302 struct rcu_node
*rnp
;
1304 rcu_for_each_leaf_node(rsp
, rnp
) {
1306 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1307 if (!rcu_gp_in_progress(rsp
)) {
1308 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1311 if (rnp
->qsmask
== 0) {
1312 rcu_initiate_boost(rnp
);
1313 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1318 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1319 if ((rnp
->qsmask
& bit
) != 0 &&
1320 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1325 /* rcu_report_qs_rnp() releases rnp->lock. */
1326 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1329 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1331 rnp
= rcu_get_root(rsp
);
1332 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1333 if (rnp
->qsmask
== 0)
1334 rcu_initiate_boost(rnp
);
1335 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1339 * Force quiescent states on reluctant CPUs, and also detect which
1340 * CPUs are in dyntick-idle mode.
1342 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1344 unsigned long flags
;
1345 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1347 if (!rcu_gp_in_progress(rsp
))
1348 return; /* No grace period in progress, nothing to force. */
1349 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1350 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1351 return; /* Someone else is already on the job. */
1353 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1354 goto unlock_fqs_ret
; /* no emergency and done recently. */
1356 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1357 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1358 if(!rcu_gp_in_progress(rsp
)) {
1359 rsp
->n_force_qs_ngp
++;
1360 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1361 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1363 rsp
->fqs_active
= 1;
1364 switch (rsp
->signaled
) {
1368 break; /* grace period idle or initializing, ignore. */
1370 case RCU_SAVE_DYNTICK
:
1371 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1372 break; /* So gcc recognizes the dead code. */
1374 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1376 /* Record dyntick-idle state. */
1377 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1378 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1379 if (rcu_gp_in_progress(rsp
))
1380 rsp
->signaled
= RCU_FORCE_QS
;
1385 /* Check dyntick-idle state, send IPI to laggarts. */
1386 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1387 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1389 /* Leave state in case more forcing is required. */
1391 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1394 rsp
->fqs_active
= 0;
1395 if (rsp
->fqs_need_gp
) {
1396 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1397 rsp
->fqs_need_gp
= 0;
1398 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1401 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1403 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1406 #else /* #ifdef CONFIG_SMP */
1408 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1413 #endif /* #else #ifdef CONFIG_SMP */
1416 * This does the RCU processing work from softirq context for the
1417 * specified rcu_state and rcu_data structures. This may be called
1418 * only from the CPU to whom the rdp belongs.
1421 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1423 unsigned long flags
;
1425 WARN_ON_ONCE(rdp
->beenonline
== 0);
1428 * If an RCU GP has gone long enough, go check for dyntick
1429 * idle CPUs and, if needed, send resched IPIs.
1431 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1432 force_quiescent_state(rsp
, 1);
1435 * Advance callbacks in response to end of earlier grace
1436 * period that some other CPU ended.
1438 rcu_process_gp_end(rsp
, rdp
);
1440 /* Update RCU state based on any recent quiescent states. */
1441 rcu_check_quiescent_state(rsp
, rdp
);
1443 /* Does this CPU require a not-yet-started grace period? */
1444 if (cpu_needs_another_gp(rsp
, rdp
)) {
1445 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1446 rcu_start_gp(rsp
, flags
); /* releases above lock */
1449 /* If there are callbacks ready, invoke them. */
1450 rcu_do_batch(rsp
, rdp
);
1454 * Do softirq processing for the current CPU.
1456 static void rcu_process_callbacks(void)
1458 __rcu_process_callbacks(&rcu_sched_state
,
1459 &__get_cpu_var(rcu_sched_data
));
1460 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1461 rcu_preempt_process_callbacks();
1463 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1464 rcu_needs_cpu_flush();
1468 * Wake up the current CPU's kthread. This replaces raise_softirq()
1469 * in earlier versions of RCU. Note that because we are running on
1470 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1471 * cannot disappear out from under us.
1473 static void invoke_rcu_cpu_kthread(void)
1475 unsigned long flags
;
1476 wait_queue_head_t
*q
;
1479 local_irq_save(flags
);
1480 cpu
= smp_processor_id();
1481 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1482 if (per_cpu(rcu_cpu_kthread_task
, cpu
) == NULL
) {
1483 local_irq_restore(flags
);
1486 q
= &per_cpu(rcu_cpu_wq
, cpu
);
1488 local_irq_restore(flags
);
1492 * Wake up the specified per-rcu_node-structure kthread.
1493 * The caller must hold ->lock.
1495 static void invoke_rcu_node_kthread(struct rcu_node
*rnp
)
1497 struct task_struct
*t
;
1499 t
= rnp
->node_kthread_task
;
1505 * Set the specified CPU's kthread to run RT or not, as specified by
1506 * the to_rt argument. The CPU-hotplug locks are held, so the task
1507 * is not going away.
1509 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1512 struct sched_param sp
;
1513 struct task_struct
*t
;
1515 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1519 policy
= SCHED_FIFO
;
1520 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1522 policy
= SCHED_NORMAL
;
1523 sp
.sched_priority
= 0;
1525 sched_setscheduler_nocheck(t
, policy
, &sp
);
1529 * Timer handler to initiate the waking up of per-CPU kthreads that
1530 * have yielded the CPU due to excess numbers of RCU callbacks.
1531 * We wake up the per-rcu_node kthread, which in turn will wake up
1532 * the booster kthread.
1534 static void rcu_cpu_kthread_timer(unsigned long arg
)
1536 unsigned long flags
;
1537 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, arg
);
1538 struct rcu_node
*rnp
= rdp
->mynode
;
1540 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1541 rnp
->wakemask
|= rdp
->grpmask
;
1542 invoke_rcu_node_kthread(rnp
);
1543 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1547 * Drop to non-real-time priority and yield, but only after posting a
1548 * timer that will cause us to regain our real-time priority if we
1549 * remain preempted. Either way, we restore our real-time priority
1552 static void rcu_yield(void (*f
)(unsigned long), unsigned long arg
)
1554 struct sched_param sp
;
1555 struct timer_list yield_timer
;
1557 setup_timer_on_stack(&yield_timer
, f
, arg
);
1558 mod_timer(&yield_timer
, jiffies
+ 2);
1559 sp
.sched_priority
= 0;
1560 sched_setscheduler_nocheck(current
, SCHED_NORMAL
, &sp
);
1562 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1563 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1564 del_timer(&yield_timer
);
1568 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1569 * This can happen while the corresponding CPU is either coming online
1570 * or going offline. We cannot wait until the CPU is fully online
1571 * before starting the kthread, because the various notifier functions
1572 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1573 * the corresponding CPU is online.
1575 * Return 1 if the kthread needs to stop, 0 otherwise.
1577 * Caller must disable bh. This function can momentarily enable it.
1579 static int rcu_cpu_kthread_should_stop(int cpu
)
1581 while (cpu_is_offline(cpu
) ||
1582 !cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)) ||
1583 smp_processor_id() != cpu
) {
1584 if (kthread_should_stop())
1587 schedule_timeout_uninterruptible(1);
1588 if (!cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)))
1589 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
1596 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1597 * earlier RCU softirq.
1599 static int rcu_cpu_kthread(void *arg
)
1601 int cpu
= (int)(long)arg
;
1602 unsigned long flags
;
1604 unsigned int *statusp
= &per_cpu(rcu_cpu_kthread_status
, cpu
);
1605 wait_queue_head_t
*wqp
= &per_cpu(rcu_cpu_wq
, cpu
);
1607 char *workp
= &per_cpu(rcu_cpu_has_work
, cpu
);
1610 *statusp
= RCU_KTHREAD_WAITING
;
1611 wait_event_interruptible(*wqp
,
1612 *workp
!= 0 || kthread_should_stop());
1614 if (rcu_cpu_kthread_should_stop(cpu
)) {
1618 *statusp
= RCU_KTHREAD_RUNNING
;
1619 local_irq_save(flags
);
1622 local_irq_restore(flags
);
1624 rcu_process_callbacks();
1631 *statusp
= RCU_KTHREAD_YIELDING
;
1632 rcu_yield(rcu_cpu_kthread_timer
, (unsigned long)cpu
);
1636 *statusp
= RCU_KTHREAD_STOPPED
;
1641 * Spawn a per-CPU kthread, setting up affinity and priority.
1642 * Because the CPU hotplug lock is held, no other CPU will be attempting
1643 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1644 * attempting to access it during boot, but the locking in kthread_bind()
1645 * will enforce sufficient ordering.
1647 static int __cpuinit
rcu_spawn_one_cpu_kthread(int cpu
)
1649 struct sched_param sp
;
1650 struct task_struct
*t
;
1652 if (!rcu_kthreads_spawnable
||
1653 per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
)
1655 t
= kthread_create(rcu_cpu_kthread
, (void *)(long)cpu
, "rcuc%d", cpu
);
1658 kthread_bind(t
, cpu
);
1659 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
);
1660 per_cpu(rcu_cpu_kthread_task
, cpu
) = t
;
1662 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1663 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1668 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1669 * kthreads when needed. We ignore requests to wake up kthreads
1670 * for offline CPUs, which is OK because force_quiescent_state()
1671 * takes care of this case.
1673 static int rcu_node_kthread(void *arg
)
1676 unsigned long flags
;
1678 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1679 struct sched_param sp
;
1680 struct task_struct
*t
;
1683 rnp
->node_kthread_status
= RCU_KTHREAD_WAITING
;
1684 wait_event_interruptible(rnp
->node_wq
, rnp
->wakemask
!= 0 ||
1685 kthread_should_stop());
1686 if (kthread_should_stop())
1688 rnp
->node_kthread_status
= RCU_KTHREAD_RUNNING
;
1689 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1690 mask
= rnp
->wakemask
;
1692 rcu_initiate_boost(rnp
);
1693 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1694 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1) {
1695 if ((mask
& 0x1) == 0)
1698 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1699 if (!cpu_online(cpu
) || t
== NULL
) {
1703 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1704 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1705 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1709 rnp
->node_kthread_status
= RCU_KTHREAD_STOPPED
;
1714 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1715 * served by the rcu_node in question. The CPU hotplug lock is still
1716 * held, so the value of rnp->qsmaskinit will be stable.
1718 * We don't include outgoingcpu in the affinity set, use -1 if there is
1719 * no outgoing CPU. If there are no CPUs left in the affinity set,
1720 * this function allows the kthread to execute on any CPU.
1722 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1726 unsigned long mask
= rnp
->qsmaskinit
;
1728 if (rnp
->node_kthread_task
== NULL
|| mask
== 0)
1730 if (!alloc_cpumask_var(&cm
, GFP_KERNEL
))
1733 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1734 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1735 cpumask_set_cpu(cpu
, cm
);
1736 if (cpumask_weight(cm
) == 0) {
1738 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1739 cpumask_clear_cpu(cpu
, cm
);
1740 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1742 set_cpus_allowed_ptr(rnp
->node_kthread_task
, cm
);
1743 rcu_boost_kthread_setaffinity(rnp
, cm
);
1744 free_cpumask_var(cm
);
1748 * Spawn a per-rcu_node kthread, setting priority and affinity.
1749 * Called during boot before online/offline can happen, or, if
1750 * during runtime, with the main CPU-hotplug locks held. So only
1751 * one of these can be executing at a time.
1753 static int __cpuinit
rcu_spawn_one_node_kthread(struct rcu_state
*rsp
,
1754 struct rcu_node
*rnp
)
1756 unsigned long flags
;
1757 int rnp_index
= rnp
- &rsp
->node
[0];
1758 struct sched_param sp
;
1759 struct task_struct
*t
;
1761 if (!rcu_kthreads_spawnable
||
1762 rnp
->qsmaskinit
== 0)
1764 if (rnp
->node_kthread_task
== NULL
) {
1765 t
= kthread_create(rcu_node_kthread
, (void *)rnp
,
1766 "rcun%d", rnp_index
);
1769 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1770 rnp
->node_kthread_task
= t
;
1771 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1773 sp
.sched_priority
= 99;
1774 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1776 return rcu_spawn_one_boost_kthread(rsp
, rnp
, rnp_index
);
1780 * Spawn all kthreads -- called as soon as the scheduler is running.
1782 static int __init
rcu_spawn_kthreads(void)
1785 struct rcu_node
*rnp
;
1787 rcu_kthreads_spawnable
= 1;
1788 for_each_possible_cpu(cpu
) {
1789 init_waitqueue_head(&per_cpu(rcu_cpu_wq
, cpu
));
1790 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1791 if (cpu_online(cpu
))
1792 (void)rcu_spawn_one_cpu_kthread(cpu
);
1794 rnp
= rcu_get_root(rcu_state
);
1795 init_waitqueue_head(&rnp
->node_wq
);
1796 rcu_init_boost_waitqueue(rnp
);
1797 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1798 if (NUM_RCU_NODES
> 1)
1799 rcu_for_each_leaf_node(rcu_state
, rnp
) {
1800 init_waitqueue_head(&rnp
->node_wq
);
1801 rcu_init_boost_waitqueue(rnp
);
1802 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1806 early_initcall(rcu_spawn_kthreads
);
1809 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1810 struct rcu_state
*rsp
)
1812 unsigned long flags
;
1813 struct rcu_data
*rdp
;
1815 debug_rcu_head_queue(head
);
1819 smp_mb(); /* Ensure RCU update seen before callback registry. */
1822 * Opportunistically note grace-period endings and beginnings.
1823 * Note that we might see a beginning right after we see an
1824 * end, but never vice versa, since this CPU has to pass through
1825 * a quiescent state betweentimes.
1827 local_irq_save(flags
);
1828 rdp
= this_cpu_ptr(rsp
->rda
);
1830 /* Add the callback to our list. */
1831 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1832 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1835 * Force the grace period if too many callbacks or too long waiting.
1836 * Enforce hysteresis, and don't invoke force_quiescent_state()
1837 * if some other CPU has recently done so. Also, don't bother
1838 * invoking force_quiescent_state() if the newly enqueued callback
1839 * is the only one waiting for a grace period to complete.
1841 if (unlikely(++rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1843 /* Are we ignoring a completed grace period? */
1844 rcu_process_gp_end(rsp
, rdp
);
1845 check_for_new_grace_period(rsp
, rdp
);
1847 /* Start a new grace period if one not already started. */
1848 if (!rcu_gp_in_progress(rsp
)) {
1849 unsigned long nestflag
;
1850 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1852 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1853 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1855 /* Give the grace period a kick. */
1856 rdp
->blimit
= LONG_MAX
;
1857 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1858 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1859 force_quiescent_state(rsp
, 0);
1860 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1861 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1863 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1864 force_quiescent_state(rsp
, 1);
1865 local_irq_restore(flags
);
1869 * Queue an RCU-sched callback for invocation after a grace period.
1871 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1873 __call_rcu(head
, func
, &rcu_sched_state
);
1875 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1878 * Queue an RCU for invocation after a quicker grace period.
1880 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1882 __call_rcu(head
, func
, &rcu_bh_state
);
1884 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1887 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1889 * Control will return to the caller some time after a full rcu-sched
1890 * grace period has elapsed, in other words after all currently executing
1891 * rcu-sched read-side critical sections have completed. These read-side
1892 * critical sections are delimited by rcu_read_lock_sched() and
1893 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1894 * local_irq_disable(), and so on may be used in place of
1895 * rcu_read_lock_sched().
1897 * This means that all preempt_disable code sequences, including NMI and
1898 * hardware-interrupt handlers, in progress on entry will have completed
1899 * before this primitive returns. However, this does not guarantee that
1900 * softirq handlers will have completed, since in some kernels, these
1901 * handlers can run in process context, and can block.
1903 * This primitive provides the guarantees made by the (now removed)
1904 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1905 * guarantees that rcu_read_lock() sections will have completed.
1906 * In "classic RCU", these two guarantees happen to be one and
1907 * the same, but can differ in realtime RCU implementations.
1909 void synchronize_sched(void)
1911 struct rcu_synchronize rcu
;
1913 if (rcu_blocking_is_gp())
1916 init_rcu_head_on_stack(&rcu
.head
);
1917 init_completion(&rcu
.completion
);
1918 /* Will wake me after RCU finished. */
1919 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1921 wait_for_completion(&rcu
.completion
);
1922 destroy_rcu_head_on_stack(&rcu
.head
);
1924 EXPORT_SYMBOL_GPL(synchronize_sched
);
1927 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1929 * Control will return to the caller some time after a full rcu_bh grace
1930 * period has elapsed, in other words after all currently executing rcu_bh
1931 * read-side critical sections have completed. RCU read-side critical
1932 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1933 * and may be nested.
1935 void synchronize_rcu_bh(void)
1937 struct rcu_synchronize rcu
;
1939 if (rcu_blocking_is_gp())
1942 init_rcu_head_on_stack(&rcu
.head
);
1943 init_completion(&rcu
.completion
);
1944 /* Will wake me after RCU finished. */
1945 call_rcu_bh(&rcu
.head
, wakeme_after_rcu
);
1947 wait_for_completion(&rcu
.completion
);
1948 destroy_rcu_head_on_stack(&rcu
.head
);
1950 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1953 * Check to see if there is any immediate RCU-related work to be done
1954 * by the current CPU, for the specified type of RCU, returning 1 if so.
1955 * The checks are in order of increasing expense: checks that can be
1956 * carried out against CPU-local state are performed first. However,
1957 * we must check for CPU stalls first, else we might not get a chance.
1959 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1961 struct rcu_node
*rnp
= rdp
->mynode
;
1963 rdp
->n_rcu_pending
++;
1965 /* Check for CPU stalls, if enabled. */
1966 check_cpu_stall(rsp
, rdp
);
1968 /* Is the RCU core waiting for a quiescent state from this CPU? */
1969 if (rdp
->qs_pending
&& !rdp
->passed_quiesc
) {
1972 * If force_quiescent_state() coming soon and this CPU
1973 * needs a quiescent state, and this is either RCU-sched
1974 * or RCU-bh, force a local reschedule.
1976 rdp
->n_rp_qs_pending
++;
1977 if (!rdp
->preemptable
&&
1978 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
1981 } else if (rdp
->qs_pending
&& rdp
->passed_quiesc
) {
1982 rdp
->n_rp_report_qs
++;
1986 /* Does this CPU have callbacks ready to invoke? */
1987 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1988 rdp
->n_rp_cb_ready
++;
1992 /* Has RCU gone idle with this CPU needing another grace period? */
1993 if (cpu_needs_another_gp(rsp
, rdp
)) {
1994 rdp
->n_rp_cpu_needs_gp
++;
1998 /* Has another RCU grace period completed? */
1999 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2000 rdp
->n_rp_gp_completed
++;
2004 /* Has a new RCU grace period started? */
2005 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2006 rdp
->n_rp_gp_started
++;
2010 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2011 if (rcu_gp_in_progress(rsp
) &&
2012 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
2013 rdp
->n_rp_need_fqs
++;
2018 rdp
->n_rp_need_nothing
++;
2023 * Check to see if there is any immediate RCU-related work to be done
2024 * by the current CPU, returning 1 if so. This function is part of the
2025 * RCU implementation; it is -not- an exported member of the RCU API.
2027 static int rcu_pending(int cpu
)
2029 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
2030 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
2031 rcu_preempt_pending(cpu
);
2035 * Check to see if any future RCU-related work will need to be done
2036 * by the current CPU, even if none need be done immediately, returning
2039 static int rcu_needs_cpu_quick_check(int cpu
)
2041 /* RCU callbacks either ready or pending? */
2042 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
2043 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
2044 rcu_preempt_needs_cpu(cpu
);
2047 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
2048 static atomic_t rcu_barrier_cpu_count
;
2049 static DEFINE_MUTEX(rcu_barrier_mutex
);
2050 static struct completion rcu_barrier_completion
;
2052 static void rcu_barrier_callback(struct rcu_head
*notused
)
2054 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2055 complete(&rcu_barrier_completion
);
2059 * Called with preemption disabled, and from cross-cpu IRQ context.
2061 static void rcu_barrier_func(void *type
)
2063 int cpu
= smp_processor_id();
2064 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
2065 void (*call_rcu_func
)(struct rcu_head
*head
,
2066 void (*func
)(struct rcu_head
*head
));
2068 atomic_inc(&rcu_barrier_cpu_count
);
2069 call_rcu_func
= type
;
2070 call_rcu_func(head
, rcu_barrier_callback
);
2074 * Orchestrate the specified type of RCU barrier, waiting for all
2075 * RCU callbacks of the specified type to complete.
2077 static void _rcu_barrier(struct rcu_state
*rsp
,
2078 void (*call_rcu_func
)(struct rcu_head
*head
,
2079 void (*func
)(struct rcu_head
*head
)))
2081 BUG_ON(in_interrupt());
2082 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2083 mutex_lock(&rcu_barrier_mutex
);
2084 init_completion(&rcu_barrier_completion
);
2086 * Initialize rcu_barrier_cpu_count to 1, then invoke
2087 * rcu_barrier_func() on each CPU, so that each CPU also has
2088 * incremented rcu_barrier_cpu_count. Only then is it safe to
2089 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2090 * might complete its grace period before all of the other CPUs
2091 * did their increment, causing this function to return too
2092 * early. Note that on_each_cpu() disables irqs, which prevents
2093 * any CPUs from coming online or going offline until each online
2094 * CPU has queued its RCU-barrier callback.
2096 atomic_set(&rcu_barrier_cpu_count
, 1);
2097 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
2098 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2099 complete(&rcu_barrier_completion
);
2100 wait_for_completion(&rcu_barrier_completion
);
2101 mutex_unlock(&rcu_barrier_mutex
);
2105 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2107 void rcu_barrier_bh(void)
2109 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
2111 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2114 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2116 void rcu_barrier_sched(void)
2118 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2120 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2123 * Do boot-time initialization of a CPU's per-CPU RCU data.
2126 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2128 unsigned long flags
;
2130 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2131 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2133 /* Set up local state, ensuring consistent view of global state. */
2134 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2135 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2136 rdp
->nxtlist
= NULL
;
2137 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2138 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2141 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2142 #endif /* #ifdef CONFIG_NO_HZ */
2144 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2148 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2149 * offline event can be happening at a given time. Note also that we
2150 * can accept some slop in the rsp->completed access due to the fact
2151 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2153 static void __cpuinit
2154 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptable
)
2156 unsigned long flags
;
2158 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2159 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2161 /* Set up local state, ensuring consistent view of global state. */
2162 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2163 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
2164 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
2165 rdp
->beenonline
= 1; /* We have now been online. */
2166 rdp
->preemptable
= preemptable
;
2167 rdp
->qlen_last_fqs_check
= 0;
2168 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2169 rdp
->blimit
= blimit
;
2170 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2173 * A new grace period might start here. If so, we won't be part
2174 * of it, but that is OK, as we are currently in a quiescent state.
2177 /* Exclude any attempts to start a new GP on large systems. */
2178 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2180 /* Add CPU to rcu_node bitmasks. */
2182 mask
= rdp
->grpmask
;
2184 /* Exclude any attempts to start a new GP on small systems. */
2185 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2186 rnp
->qsmaskinit
|= mask
;
2187 mask
= rnp
->grpmask
;
2188 if (rnp
== rdp
->mynode
) {
2189 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
2190 rdp
->completed
= rnp
->completed
;
2191 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
2193 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2195 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2197 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2200 static void __cpuinit
rcu_online_cpu(int cpu
)
2202 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2203 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2204 rcu_preempt_init_percpu_data(cpu
);
2207 static void __cpuinit
rcu_online_kthreads(int cpu
)
2209 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2210 struct rcu_node
*rnp
= rdp
->mynode
;
2212 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
2213 if (rcu_kthreads_spawnable
) {
2214 (void)rcu_spawn_one_cpu_kthread(cpu
);
2215 if (rnp
->node_kthread_task
== NULL
)
2216 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
2221 * Handle CPU online/offline notification events.
2223 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2224 unsigned long action
, void *hcpu
)
2226 long cpu
= (long)hcpu
;
2227 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2228 struct rcu_node
*rnp
= rdp
->mynode
;
2231 case CPU_UP_PREPARE
:
2232 case CPU_UP_PREPARE_FROZEN
:
2233 rcu_online_cpu(cpu
);
2234 rcu_online_kthreads(cpu
);
2237 case CPU_DOWN_FAILED
:
2238 rcu_node_kthread_setaffinity(rnp
, -1);
2239 rcu_cpu_kthread_setrt(cpu
, 1);
2241 case CPU_DOWN_PREPARE
:
2242 rcu_node_kthread_setaffinity(rnp
, cpu
);
2243 rcu_cpu_kthread_setrt(cpu
, 0);
2246 case CPU_DYING_FROZEN
:
2248 * The whole machine is "stopped" except this CPU, so we can
2249 * touch any data without introducing corruption. We send the
2250 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2252 rcu_send_cbs_to_online(&rcu_bh_state
);
2253 rcu_send_cbs_to_online(&rcu_sched_state
);
2254 rcu_preempt_send_cbs_to_online();
2257 case CPU_DEAD_FROZEN
:
2258 case CPU_UP_CANCELED
:
2259 case CPU_UP_CANCELED_FROZEN
:
2260 rcu_offline_cpu(cpu
);
2269 * This function is invoked towards the end of the scheduler's initialization
2270 * process. Before this is called, the idle task might contain
2271 * RCU read-side critical sections (during which time, this idle
2272 * task is booting the system). After this function is called, the
2273 * idle tasks are prohibited from containing RCU read-side critical
2274 * sections. This function also enables RCU lockdep checking.
2276 void rcu_scheduler_starting(void)
2278 WARN_ON(num_online_cpus() != 1);
2279 WARN_ON(nr_context_switches() > 0);
2280 rcu_scheduler_active
= 1;
2284 * Compute the per-level fanout, either using the exact fanout specified
2285 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2287 #ifdef CONFIG_RCU_FANOUT_EXACT
2288 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2292 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2293 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2294 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
2296 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2297 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2304 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2305 ccur
= rsp
->levelcnt
[i
];
2306 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2310 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2313 * Helper function for rcu_init() that initializes one rcu_state structure.
2315 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2316 struct rcu_data __percpu
*rda
)
2318 static char *buf
[] = { "rcu_node_level_0",
2321 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2325 struct rcu_node
*rnp
;
2327 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2329 /* Initialize the level-tracking arrays. */
2331 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2332 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2333 rcu_init_levelspread(rsp
);
2335 /* Initialize the elements themselves, starting from the leaves. */
2337 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2338 cpustride
*= rsp
->levelspread
[i
];
2339 rnp
= rsp
->level
[i
];
2340 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2341 raw_spin_lock_init(&rnp
->lock
);
2342 lockdep_set_class_and_name(&rnp
->lock
,
2343 &rcu_node_class
[i
], buf
[i
]);
2346 rnp
->qsmaskinit
= 0;
2347 rnp
->grplo
= j
* cpustride
;
2348 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2349 if (rnp
->grphi
>= NR_CPUS
)
2350 rnp
->grphi
= NR_CPUS
- 1;
2356 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2357 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2358 rnp
->parent
= rsp
->level
[i
- 1] +
2359 j
/ rsp
->levelspread
[i
- 1];
2362 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2367 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2368 for_each_possible_cpu(i
) {
2369 while (i
> rnp
->grphi
)
2371 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2372 rcu_boot_init_percpu_data(i
, rsp
);
2376 void __init
rcu_init(void)
2380 rcu_bootup_announce();
2381 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2382 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2383 __rcu_init_preempt();
2386 * We don't need protection against CPU-hotplug here because
2387 * this is called early in boot, before either interrupts
2388 * or the scheduler are operational.
2390 cpu_notifier(rcu_cpu_notify
, 0);
2391 for_each_online_cpu(cpu
)
2392 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2393 check_cpu_stall_init();
2396 #include "rcutree_plugin.h"