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 <linux/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>
52 #include <linux/prefetch.h>
56 /* Data structures. */
58 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
60 #define RCU_STATE_INITIALIZER(structname) { \
61 .level = { &structname.node[0] }, \
63 NUM_RCU_LVL_0, /* root of hierarchy. */ \
67 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
69 .signaled = RCU_GP_IDLE, \
72 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
73 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
75 .n_force_qs_ngp = 0, \
76 .name = #structname, \
79 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched_state
);
80 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
82 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh_state
);
83 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
85 static struct rcu_state
*rcu_state
;
87 int rcu_scheduler_active __read_mostly
;
88 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
90 #ifdef CONFIG_RCU_BOOST
93 * Control variables for per-CPU and per-rcu_node kthreads. These
94 * handle all flavors of RCU.
96 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
97 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
98 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
99 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
100 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
101 static char rcu_kthreads_spawnable
;
103 #endif /* #ifdef CONFIG_RCU_BOOST */
105 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
106 static void invoke_rcu_core(void);
107 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
109 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
112 * Track the rcutorture test sequence number and the update version
113 * number within a given test. The rcutorture_testseq is incremented
114 * on every rcutorture module load and unload, so has an odd value
115 * when a test is running. The rcutorture_vernum is set to zero
116 * when rcutorture starts and is incremented on each rcutorture update.
117 * These variables enable correlating rcutorture output with the
118 * RCU tracing information.
120 unsigned long rcutorture_testseq
;
121 unsigned long rcutorture_vernum
;
124 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
125 * permit this function to be invoked without holding the root rcu_node
126 * structure's ->lock, but of course results can be subject to change.
128 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
130 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
134 * Note a quiescent state. Because we do not need to know
135 * how many quiescent states passed, just if there was at least
136 * one since the start of the grace period, this just sets a flag.
138 void rcu_sched_qs(int cpu
)
140 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
142 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
144 rdp
->passed_quiesc
= 1;
147 void rcu_bh_qs(int cpu
)
149 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
151 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
153 rdp
->passed_quiesc
= 1;
157 * Note a context switch. This is a quiescent state for RCU-sched,
158 * and requires special handling for preemptible RCU.
160 void rcu_note_context_switch(int cpu
)
163 rcu_preempt_note_context_switch(cpu
);
165 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
168 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
169 .dynticks_nesting
= 1,
170 .dynticks
= ATOMIC_INIT(1),
172 #endif /* #ifdef CONFIG_NO_HZ */
174 static int blimit
= 10; /* Maximum callbacks per softirq. */
175 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
176 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
178 module_param(blimit
, int, 0);
179 module_param(qhimark
, int, 0);
180 module_param(qlowmark
, int, 0);
182 int rcu_cpu_stall_suppress __read_mostly
;
183 module_param(rcu_cpu_stall_suppress
, int, 0644);
185 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
186 static int rcu_pending(int cpu
);
189 * Return the number of RCU-sched batches processed thus far for debug & stats.
191 long rcu_batches_completed_sched(void)
193 return rcu_sched_state
.completed
;
195 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
198 * Return the number of RCU BH batches processed thus far for debug & stats.
200 long rcu_batches_completed_bh(void)
202 return rcu_bh_state
.completed
;
204 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
207 * Force a quiescent state for RCU BH.
209 void rcu_bh_force_quiescent_state(void)
211 force_quiescent_state(&rcu_bh_state
, 0);
213 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
216 * Record the number of times rcutorture tests have been initiated and
217 * terminated. This information allows the debugfs tracing stats to be
218 * correlated to the rcutorture messages, even when the rcutorture module
219 * is being repeatedly loaded and unloaded. In other words, we cannot
220 * store this state in rcutorture itself.
222 void rcutorture_record_test_transition(void)
224 rcutorture_testseq
++;
225 rcutorture_vernum
= 0;
227 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
230 * Record the number of writer passes through the current rcutorture test.
231 * This is also used to correlate debugfs tracing stats with the rcutorture
234 void rcutorture_record_progress(unsigned long vernum
)
238 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
241 * Force a quiescent state for RCU-sched.
243 void rcu_sched_force_quiescent_state(void)
245 force_quiescent_state(&rcu_sched_state
, 0);
247 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
250 * Does the CPU have callbacks ready to be invoked?
253 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
255 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
259 * Does the current CPU require a yet-as-unscheduled grace period?
262 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
264 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
268 * Return the root node of the specified rcu_state structure.
270 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
272 return &rsp
->node
[0];
278 * If the specified CPU is offline, tell the caller that it is in
279 * a quiescent state. Otherwise, whack it with a reschedule IPI.
280 * Grace periods can end up waiting on an offline CPU when that
281 * CPU is in the process of coming online -- it will be added to the
282 * rcu_node bitmasks before it actually makes it online. The same thing
283 * can happen while a CPU is in the process of coming online. Because this
284 * race is quite rare, we check for it after detecting that the grace
285 * period has been delayed rather than checking each and every CPU
286 * each and every time we start a new grace period.
288 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
291 * If the CPU is offline, it is in a quiescent state. We can
292 * trust its state not to change because interrupts are disabled.
294 if (cpu_is_offline(rdp
->cpu
)) {
299 /* If preemptible RCU, no point in sending reschedule IPI. */
300 if (rdp
->preemptible
)
303 /* The CPU is online, so send it a reschedule IPI. */
304 if (rdp
->cpu
!= smp_processor_id())
305 smp_send_reschedule(rdp
->cpu
);
312 #endif /* #ifdef CONFIG_SMP */
317 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
319 * Enter nohz mode, in other words, -leave- the mode in which RCU
320 * read-side critical sections can occur. (Though RCU read-side
321 * critical sections can occur in irq handlers in nohz mode, a possibility
322 * handled by rcu_irq_enter() and rcu_irq_exit()).
324 void rcu_enter_nohz(void)
327 struct rcu_dynticks
*rdtp
;
329 local_irq_save(flags
);
330 rdtp
= &__get_cpu_var(rcu_dynticks
);
331 if (--rdtp
->dynticks_nesting
) {
332 local_irq_restore(flags
);
335 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
336 smp_mb__before_atomic_inc(); /* See above. */
337 atomic_inc(&rdtp
->dynticks
);
338 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
339 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
340 local_irq_restore(flags
);
342 /* If the interrupt queued a callback, get out of dyntick mode. */
344 (__get_cpu_var(rcu_sched_data
).nxtlist
||
345 __get_cpu_var(rcu_bh_data
).nxtlist
||
346 rcu_preempt_needs_cpu(smp_processor_id())))
351 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
353 * Exit nohz mode, in other words, -enter- the mode in which RCU
354 * read-side critical sections normally occur.
356 void rcu_exit_nohz(void)
359 struct rcu_dynticks
*rdtp
;
361 local_irq_save(flags
);
362 rdtp
= &__get_cpu_var(rcu_dynticks
);
363 if (rdtp
->dynticks_nesting
++) {
364 local_irq_restore(flags
);
367 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
368 atomic_inc(&rdtp
->dynticks
);
369 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
370 smp_mb__after_atomic_inc(); /* See above. */
371 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
372 local_irq_restore(flags
);
376 * rcu_nmi_enter - inform RCU of entry to NMI context
378 * If the CPU was idle with dynamic ticks active, and there is no
379 * irq handler running, this updates rdtp->dynticks_nmi to let the
380 * RCU grace-period handling know that the CPU is active.
382 void rcu_nmi_enter(void)
384 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
386 if (rdtp
->dynticks_nmi_nesting
== 0 &&
387 (atomic_read(&rdtp
->dynticks
) & 0x1))
389 rdtp
->dynticks_nmi_nesting
++;
390 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
391 atomic_inc(&rdtp
->dynticks
);
392 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
393 smp_mb__after_atomic_inc(); /* See above. */
394 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
398 * rcu_nmi_exit - inform RCU of exit from NMI context
400 * If the CPU was idle with dynamic ticks active, and there is no
401 * irq handler running, this updates rdtp->dynticks_nmi to let the
402 * RCU grace-period handling know that the CPU is no longer active.
404 void rcu_nmi_exit(void)
406 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
408 if (rdtp
->dynticks_nmi_nesting
== 0 ||
409 --rdtp
->dynticks_nmi_nesting
!= 0)
411 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
412 smp_mb__before_atomic_inc(); /* See above. */
413 atomic_inc(&rdtp
->dynticks
);
414 smp_mb__after_atomic_inc(); /* Force delay to next write. */
415 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
419 * rcu_irq_enter - inform RCU of entry to hard irq context
421 * If the CPU was idle with dynamic ticks active, this updates the
422 * rdtp->dynticks to let the RCU handling know that the CPU is active.
424 void rcu_irq_enter(void)
430 * rcu_irq_exit - inform RCU of exit from hard irq context
432 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
433 * to put let the RCU handling be aware that the CPU is going back to idle
436 void rcu_irq_exit(void)
444 * Snapshot the specified CPU's dynticks counter so that we can later
445 * credit them with an implicit quiescent state. Return 1 if this CPU
446 * is in dynticks idle mode, which is an extended quiescent state.
448 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
450 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
455 * Return true if the specified CPU has passed through a quiescent
456 * state by virtue of being in or having passed through an dynticks
457 * idle state since the last call to dyntick_save_progress_counter()
460 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
465 curr
= (unsigned long)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
466 snap
= (unsigned long)rdp
->dynticks_snap
;
469 * If the CPU passed through or entered a dynticks idle phase with
470 * no active irq/NMI handlers, then we can safely pretend that the CPU
471 * already acknowledged the request to pass through a quiescent
472 * state. Either way, that CPU cannot possibly be in an RCU
473 * read-side critical section that started before the beginning
474 * of the current RCU grace period.
476 if ((curr
& 0x1) == 0 || ULONG_CMP_GE(curr
, snap
+ 2)) {
481 /* Go check for the CPU being offline. */
482 return rcu_implicit_offline_qs(rdp
);
485 #endif /* #ifdef CONFIG_SMP */
487 #else /* #ifdef CONFIG_NO_HZ */
491 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
496 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
498 return rcu_implicit_offline_qs(rdp
);
501 #endif /* #ifdef CONFIG_SMP */
503 #endif /* #else #ifdef CONFIG_NO_HZ */
505 int rcu_cpu_stall_suppress __read_mostly
;
507 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
509 rsp
->gp_start
= jiffies
;
510 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
513 static void print_other_cpu_stall(struct rcu_state
*rsp
)
518 struct rcu_node
*rnp
= rcu_get_root(rsp
);
520 /* Only let one CPU complain about others per time interval. */
522 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
523 delta
= jiffies
- rsp
->jiffies_stall
;
524 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
525 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
528 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
531 * Now rat on any tasks that got kicked up to the root rcu_node
532 * due to CPU offlining.
534 rcu_print_task_stall(rnp
);
535 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
538 * OK, time to rat on our buddy...
539 * See Documentation/RCU/stallwarn.txt for info on how to debug
540 * RCU CPU stall warnings.
542 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
544 rcu_for_each_leaf_node(rsp
, rnp
) {
545 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
546 rcu_print_task_stall(rnp
);
547 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
548 if (rnp
->qsmask
== 0)
550 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
551 if (rnp
->qsmask
& (1UL << cpu
))
552 printk(" %d", rnp
->grplo
+ cpu
);
554 printk("} (detected by %d, t=%ld jiffies)\n",
555 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
556 trigger_all_cpu_backtrace();
558 /* If so configured, complain about tasks blocking the grace period. */
560 rcu_print_detail_task_stall(rsp
);
562 force_quiescent_state(rsp
, 0); /* Kick them all. */
565 static void print_cpu_stall(struct rcu_state
*rsp
)
568 struct rcu_node
*rnp
= rcu_get_root(rsp
);
571 * OK, time to rat on ourselves...
572 * See Documentation/RCU/stallwarn.txt for info on how to debug
573 * RCU CPU stall warnings.
575 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
576 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
577 trigger_all_cpu_backtrace();
579 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
580 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
582 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
583 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
585 set_need_resched(); /* kick ourselves to get things going. */
588 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
592 struct rcu_node
*rnp
;
594 if (rcu_cpu_stall_suppress
)
596 j
= ACCESS_ONCE(jiffies
);
597 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
599 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
601 /* We haven't checked in, so go dump stack. */
602 print_cpu_stall(rsp
);
604 } else if (rcu_gp_in_progress(rsp
) &&
605 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
607 /* They had a few time units to dump stack, so complain. */
608 print_other_cpu_stall(rsp
);
612 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
614 rcu_cpu_stall_suppress
= 1;
619 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
621 * Set the stall-warning timeout way off into the future, thus preventing
622 * any RCU CPU stall-warning messages from appearing in the current set of
625 * The caller must disable hard irqs.
627 void rcu_cpu_stall_reset(void)
629 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
630 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
631 rcu_preempt_stall_reset();
634 static struct notifier_block rcu_panic_block
= {
635 .notifier_call
= rcu_panic
,
638 static void __init
check_cpu_stall_init(void)
640 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
644 * Update CPU-local rcu_data state to record the newly noticed grace period.
645 * This is used both when we started the grace period and when we notice
646 * that someone else started the grace period. The caller must hold the
647 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
648 * and must have irqs disabled.
650 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
652 if (rdp
->gpnum
!= rnp
->gpnum
) {
654 * If the current grace period is waiting for this CPU,
655 * set up to detect a quiescent state, otherwise don't
656 * go looking for one.
658 rdp
->gpnum
= rnp
->gpnum
;
659 if (rnp
->qsmask
& rdp
->grpmask
) {
661 rdp
->passed_quiesc
= 0;
667 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
670 struct rcu_node
*rnp
;
672 local_irq_save(flags
);
674 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
675 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
676 local_irq_restore(flags
);
679 __note_new_gpnum(rsp
, rnp
, rdp
);
680 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
684 * Did someone else start a new RCU grace period start since we last
685 * checked? Update local state appropriately if so. Must be called
686 * on the CPU corresponding to rdp.
689 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
694 local_irq_save(flags
);
695 if (rdp
->gpnum
!= rsp
->gpnum
) {
696 note_new_gpnum(rsp
, rdp
);
699 local_irq_restore(flags
);
704 * Advance this CPU's callbacks, but only if the current grace period
705 * has ended. This may be called only from the CPU to whom the rdp
706 * belongs. In addition, the corresponding leaf rcu_node structure's
707 * ->lock must be held by the caller, with irqs disabled.
710 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
712 /* Did another grace period end? */
713 if (rdp
->completed
!= rnp
->completed
) {
715 /* Advance callbacks. No harm if list empty. */
716 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
717 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
718 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
720 /* Remember that we saw this grace-period completion. */
721 rdp
->completed
= rnp
->completed
;
724 * If we were in an extended quiescent state, we may have
725 * missed some grace periods that others CPUs handled on
726 * our behalf. Catch up with this state to avoid noting
727 * spurious new grace periods. If another grace period
728 * has started, then rnp->gpnum will have advanced, so
729 * we will detect this later on.
731 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
732 rdp
->gpnum
= rdp
->completed
;
735 * If RCU does not need a quiescent state from this CPU,
736 * then make sure that this CPU doesn't go looking for one.
738 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
744 * Advance this CPU's callbacks, but only if the current grace period
745 * has ended. This may be called only from the CPU to whom the rdp
749 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
752 struct rcu_node
*rnp
;
754 local_irq_save(flags
);
756 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
757 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
758 local_irq_restore(flags
);
761 __rcu_process_gp_end(rsp
, rnp
, rdp
);
762 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
766 * Do per-CPU grace-period initialization for running CPU. The caller
767 * must hold the lock of the leaf rcu_node structure corresponding to
771 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
773 /* Prior grace period ended, so advance callbacks for current CPU. */
774 __rcu_process_gp_end(rsp
, rnp
, rdp
);
777 * Because this CPU just now started the new grace period, we know
778 * that all of its callbacks will be covered by this upcoming grace
779 * period, even the ones that were registered arbitrarily recently.
780 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
782 * Other CPUs cannot be sure exactly when the grace period started.
783 * Therefore, their recently registered callbacks must pass through
784 * an additional RCU_NEXT_READY stage, so that they will be handled
785 * by the next RCU grace period.
787 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
788 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
790 /* Set state so that this CPU will detect the next quiescent state. */
791 __note_new_gpnum(rsp
, rnp
, rdp
);
795 * Start a new RCU grace period if warranted, re-initializing the hierarchy
796 * in preparation for detecting the next grace period. The caller must hold
797 * the root node's ->lock, which is released before return. Hard irqs must
801 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
802 __releases(rcu_get_root(rsp
)->lock
)
804 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
805 struct rcu_node
*rnp
= rcu_get_root(rsp
);
807 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
808 if (cpu_needs_another_gp(rsp
, rdp
))
809 rsp
->fqs_need_gp
= 1;
810 if (rnp
->completed
== rsp
->completed
) {
811 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
814 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
817 * Propagate new ->completed value to rcu_node structures
818 * so that other CPUs don't have to wait until the start
819 * of the next grace period to process their callbacks.
821 rcu_for_each_node_breadth_first(rsp
, rnp
) {
822 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
823 rnp
->completed
= rsp
->completed
;
824 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
826 local_irq_restore(flags
);
830 /* Advance to a new grace period and initialize state. */
832 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
833 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
834 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
835 record_gp_stall_check_time(rsp
);
837 /* Special-case the common single-level case. */
838 if (NUM_RCU_NODES
== 1) {
839 rcu_preempt_check_blocked_tasks(rnp
);
840 rnp
->qsmask
= rnp
->qsmaskinit
;
841 rnp
->gpnum
= rsp
->gpnum
;
842 rnp
->completed
= rsp
->completed
;
843 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
844 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
845 rcu_preempt_boost_start_gp(rnp
);
846 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
850 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
853 /* Exclude any concurrent CPU-hotplug operations. */
854 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
857 * Set the quiescent-state-needed bits in all the rcu_node
858 * structures for all currently online CPUs in breadth-first
859 * order, starting from the root rcu_node structure. This
860 * operation relies on the layout of the hierarchy within the
861 * rsp->node[] array. Note that other CPUs will access only
862 * the leaves of the hierarchy, which still indicate that no
863 * grace period is in progress, at least until the corresponding
864 * leaf node has been initialized. In addition, we have excluded
865 * CPU-hotplug operations.
867 * Note that the grace period cannot complete until we finish
868 * the initialization process, as there will be at least one
869 * qsmask bit set in the root node until that time, namely the
870 * one corresponding to this CPU, due to the fact that we have
873 rcu_for_each_node_breadth_first(rsp
, rnp
) {
874 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
875 rcu_preempt_check_blocked_tasks(rnp
);
876 rnp
->qsmask
= rnp
->qsmaskinit
;
877 rnp
->gpnum
= rsp
->gpnum
;
878 rnp
->completed
= rsp
->completed
;
879 if (rnp
== rdp
->mynode
)
880 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
881 rcu_preempt_boost_start_gp(rnp
);
882 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
885 rnp
= rcu_get_root(rsp
);
886 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
887 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
888 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
889 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
893 * Report a full set of quiescent states to the specified rcu_state
894 * data structure. This involves cleaning up after the prior grace
895 * period and letting rcu_start_gp() start up the next grace period
896 * if one is needed. Note that the caller must hold rnp->lock, as
897 * required by rcu_start_gp(), which will release it.
899 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
900 __releases(rcu_get_root(rsp
)->lock
)
902 unsigned long gp_duration
;
904 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
907 * Ensure that all grace-period and pre-grace-period activity
908 * is seen before the assignment to rsp->completed.
910 smp_mb(); /* See above block comment. */
911 gp_duration
= jiffies
- rsp
->gp_start
;
912 if (gp_duration
> rsp
->gp_max
)
913 rsp
->gp_max
= gp_duration
;
914 rsp
->completed
= rsp
->gpnum
;
915 rsp
->signaled
= RCU_GP_IDLE
;
916 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
920 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
921 * Allows quiescent states for a group of CPUs to be reported at one go
922 * to the specified rcu_node structure, though all the CPUs in the group
923 * must be represented by the same rcu_node structure (which need not be
924 * a leaf rcu_node structure, though it often will be). That structure's
925 * lock must be held upon entry, and it is released before return.
928 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
929 struct rcu_node
*rnp
, unsigned long flags
)
930 __releases(rnp
->lock
)
932 struct rcu_node
*rnp_c
;
934 /* Walk up the rcu_node hierarchy. */
936 if (!(rnp
->qsmask
& mask
)) {
938 /* Our bit has already been cleared, so done. */
939 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
942 rnp
->qsmask
&= ~mask
;
943 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
945 /* Other bits still set at this level, so done. */
946 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
950 if (rnp
->parent
== NULL
) {
952 /* No more levels. Exit loop holding root lock. */
956 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
959 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
960 WARN_ON_ONCE(rnp_c
->qsmask
);
964 * Get here if we are the last CPU to pass through a quiescent
965 * state for this grace period. Invoke rcu_report_qs_rsp()
966 * to clean up and start the next grace period if one is needed.
968 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
972 * Record a quiescent state for the specified CPU to that CPU's rcu_data
973 * structure. This must be either called from the specified CPU, or
974 * called when the specified CPU is known to be offline (and when it is
975 * also known that no other CPU is concurrently trying to help the offline
976 * CPU). The lastcomp argument is used to make sure we are still in the
977 * grace period of interest. We don't want to end the current grace period
978 * based on quiescent states detected in an earlier grace period!
981 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
985 struct rcu_node
*rnp
;
988 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
989 if (lastcomp
!= rnp
->completed
) {
992 * Someone beat us to it for this grace period, so leave.
993 * The race with GP start is resolved by the fact that we
994 * hold the leaf rcu_node lock, so that the per-CPU bits
995 * cannot yet be initialized -- so we would simply find our
996 * CPU's bit already cleared in rcu_report_qs_rnp() if this
999 rdp
->passed_quiesc
= 0; /* try again later! */
1000 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1003 mask
= rdp
->grpmask
;
1004 if ((rnp
->qsmask
& mask
) == 0) {
1005 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1007 rdp
->qs_pending
= 0;
1010 * This GP can't end until cpu checks in, so all of our
1011 * callbacks can be processed during the next GP.
1013 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1015 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1020 * Check to see if there is a new grace period of which this CPU
1021 * is not yet aware, and if so, set up local rcu_data state for it.
1022 * Otherwise, see if this CPU has just passed through its first
1023 * quiescent state for this grace period, and record that fact if so.
1026 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1028 /* If there is now a new grace period, record and return. */
1029 if (check_for_new_grace_period(rsp
, rdp
))
1033 * Does this CPU still need to do its part for current grace period?
1034 * If no, return and let the other CPUs do their part as well.
1036 if (!rdp
->qs_pending
)
1040 * Was there a quiescent state since the beginning of the grace
1041 * period? If no, then exit and wait for the next call.
1043 if (!rdp
->passed_quiesc
)
1047 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1050 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
1053 #ifdef CONFIG_HOTPLUG_CPU
1056 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1057 * Synchronization is not required because this function executes
1058 * in stop_machine() context.
1060 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1063 /* current DYING CPU is cleared in the cpu_online_mask */
1064 int receive_cpu
= cpumask_any(cpu_online_mask
);
1065 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1066 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1068 if (rdp
->nxtlist
== NULL
)
1069 return; /* irqs disabled, so comparison is stable. */
1071 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1072 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1073 receive_rdp
->qlen
+= rdp
->qlen
;
1074 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1075 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1077 rdp
->nxtlist
= NULL
;
1078 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1079 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1084 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1085 * and move all callbacks from the outgoing CPU to the current one.
1086 * There can only be one CPU hotplug operation at a time, so no other
1087 * CPU can be attempting to update rcu_cpu_kthread_task.
1089 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1091 unsigned long flags
;
1093 int need_report
= 0;
1094 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1095 struct rcu_node
*rnp
;
1096 #ifdef CONFIG_RCU_BOOST
1097 struct task_struct
*t
;
1099 /* Stop the CPU's kthread. */
1100 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1102 per_cpu(rcu_cpu_kthread_task
, cpu
) = NULL
;
1105 #endif /* #ifdef CONFIG_RCU_BOOST */
1107 /* Exclude any attempts to start a new grace period. */
1108 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1110 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1111 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1112 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1114 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1115 rnp
->qsmaskinit
&= ~mask
;
1116 if (rnp
->qsmaskinit
!= 0) {
1117 if (rnp
!= rdp
->mynode
)
1118 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1121 if (rnp
== rdp
->mynode
)
1122 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1124 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1125 mask
= rnp
->grpmask
;
1127 } while (rnp
!= NULL
);
1130 * We still hold the leaf rcu_node structure lock here, and
1131 * irqs are still disabled. The reason for this subterfuge is
1132 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1133 * held leads to deadlock.
1135 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1137 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1138 rcu_report_unblock_qs_rnp(rnp
, flags
);
1140 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1141 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1142 rcu_report_exp_rnp(rsp
, rnp
);
1143 rcu_node_kthread_setaffinity(rnp
, -1);
1147 * Remove the specified CPU from the RCU hierarchy and move any pending
1148 * callbacks that it might have to the current CPU. This code assumes
1149 * that at least one CPU in the system will remain running at all times.
1150 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1152 static void rcu_offline_cpu(int cpu
)
1154 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1155 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1156 rcu_preempt_offline_cpu(cpu
);
1159 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1161 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1165 static void rcu_offline_cpu(int cpu
)
1169 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1172 * Invoke any RCU callbacks that have made it to the end of their grace
1173 * period. Thottle as specified by rdp->blimit.
1175 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1177 unsigned long flags
;
1178 struct rcu_head
*next
, *list
, **tail
;
1181 /* If no callbacks are ready, just return.*/
1182 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1186 * Extract the list of ready callbacks, disabling to prevent
1187 * races with call_rcu() from interrupt handlers.
1189 local_irq_save(flags
);
1190 list
= rdp
->nxtlist
;
1191 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1192 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1193 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1194 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1195 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1196 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1197 local_irq_restore(flags
);
1199 /* Invoke callbacks. */
1204 debug_rcu_head_unqueue(list
);
1205 __rcu_reclaim(list
);
1207 if (++count
>= rdp
->blimit
)
1211 local_irq_save(flags
);
1213 /* Update count, and requeue any remaining callbacks. */
1215 rdp
->n_cbs_invoked
+= count
;
1217 *tail
= rdp
->nxtlist
;
1218 rdp
->nxtlist
= list
;
1219 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1220 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1221 rdp
->nxttail
[count
] = tail
;
1226 /* Reinstate batch limit if we have worked down the excess. */
1227 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1228 rdp
->blimit
= blimit
;
1230 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1231 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1232 rdp
->qlen_last_fqs_check
= 0;
1233 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1234 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1235 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1237 local_irq_restore(flags
);
1239 /* Re-raise the RCU softirq if there are callbacks remaining. */
1240 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1245 * Check to see if this CPU is in a non-context-switch quiescent state
1246 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1247 * Also schedule the RCU softirq handler.
1249 * This function must be called with hardirqs disabled. It is normally
1250 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1251 * false, there is no point in invoking rcu_check_callbacks().
1253 void rcu_check_callbacks(int cpu
, int user
)
1256 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1257 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1260 * Get here if this CPU took its interrupt from user
1261 * mode or from the idle loop, and if this is not a
1262 * nested interrupt. In this case, the CPU is in
1263 * a quiescent state, so note it.
1265 * No memory barrier is required here because both
1266 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1267 * variables that other CPUs neither access nor modify,
1268 * at least not while the corresponding CPU is online.
1274 } else if (!in_softirq()) {
1277 * Get here if this CPU did not take its interrupt from
1278 * softirq, in other words, if it is not interrupting
1279 * a rcu_bh read-side critical section. This is an _bh
1280 * critical section, so note it.
1285 rcu_preempt_check_callbacks(cpu
);
1286 if (rcu_pending(cpu
))
1293 * Scan the leaf rcu_node structures, processing dyntick state for any that
1294 * have not yet encountered a quiescent state, using the function specified.
1295 * Also initiate boosting for any threads blocked on the root rcu_node.
1297 * The caller must have suppressed start of new grace periods.
1299 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1303 unsigned long flags
;
1305 struct rcu_node
*rnp
;
1307 rcu_for_each_leaf_node(rsp
, rnp
) {
1309 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1310 if (!rcu_gp_in_progress(rsp
)) {
1311 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1314 if (rnp
->qsmask
== 0) {
1315 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1320 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1321 if ((rnp
->qsmask
& bit
) != 0 &&
1322 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1327 /* rcu_report_qs_rnp() releases rnp->lock. */
1328 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1331 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1333 rnp
= rcu_get_root(rsp
);
1334 if (rnp
->qsmask
== 0) {
1335 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1336 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1341 * Force quiescent states on reluctant CPUs, and also detect which
1342 * CPUs are in dyntick-idle mode.
1344 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1346 unsigned long flags
;
1347 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1349 if (!rcu_gp_in_progress(rsp
))
1350 return; /* No grace period in progress, nothing to force. */
1351 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1352 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1353 return; /* Someone else is already on the job. */
1355 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1356 goto unlock_fqs_ret
; /* no emergency and done recently. */
1358 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1359 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1360 if(!rcu_gp_in_progress(rsp
)) {
1361 rsp
->n_force_qs_ngp
++;
1362 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1363 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1365 rsp
->fqs_active
= 1;
1366 switch (rsp
->signaled
) {
1370 break; /* grace period idle or initializing, ignore. */
1372 case RCU_SAVE_DYNTICK
:
1373 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1374 break; /* So gcc recognizes the dead code. */
1376 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1378 /* Record dyntick-idle state. */
1379 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1380 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1381 if (rcu_gp_in_progress(rsp
))
1382 rsp
->signaled
= RCU_FORCE_QS
;
1387 /* Check dyntick-idle state, send IPI to laggarts. */
1388 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1389 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1391 /* Leave state in case more forcing is required. */
1393 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1396 rsp
->fqs_active
= 0;
1397 if (rsp
->fqs_need_gp
) {
1398 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1399 rsp
->fqs_need_gp
= 0;
1400 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1403 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1405 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1408 #else /* #ifdef CONFIG_SMP */
1410 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1415 #endif /* #else #ifdef CONFIG_SMP */
1418 * This does the RCU processing work from softirq context for the
1419 * specified rcu_state and rcu_data structures. This may be called
1420 * only from the CPU to whom the rdp belongs.
1423 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1425 unsigned long flags
;
1427 WARN_ON_ONCE(rdp
->beenonline
== 0);
1430 * If an RCU GP has gone long enough, go check for dyntick
1431 * idle CPUs and, if needed, send resched IPIs.
1433 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1434 force_quiescent_state(rsp
, 1);
1437 * Advance callbacks in response to end of earlier grace
1438 * period that some other CPU ended.
1440 rcu_process_gp_end(rsp
, rdp
);
1442 /* Update RCU state based on any recent quiescent states. */
1443 rcu_check_quiescent_state(rsp
, rdp
);
1445 /* Does this CPU require a not-yet-started grace period? */
1446 if (cpu_needs_another_gp(rsp
, rdp
)) {
1447 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1448 rcu_start_gp(rsp
, flags
); /* releases above lock */
1451 /* If there are callbacks ready, invoke them. */
1452 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1453 invoke_rcu_callbacks(rsp
, rdp
);
1456 #ifdef CONFIG_RCU_BOOST
1458 static void rcu_kthread_do_work(void)
1460 rcu_do_batch(&rcu_sched_state
, &__get_cpu_var(rcu_sched_data
));
1461 rcu_do_batch(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1462 rcu_preempt_do_callbacks();
1465 #endif /* #ifdef CONFIG_RCU_BOOST */
1468 * Do softirq processing for the current CPU.
1470 static void rcu_process_callbacks(struct softirq_action
*unused
)
1472 __rcu_process_callbacks(&rcu_sched_state
,
1473 &__get_cpu_var(rcu_sched_data
));
1474 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1475 rcu_preempt_process_callbacks();
1477 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1478 rcu_needs_cpu_flush();
1482 * Wake up the current CPU's kthread. This replaces raise_softirq()
1483 * in earlier versions of RCU. Note that because we are running on
1484 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1485 * cannot disappear out from under us.
1487 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1489 if (likely(!rsp
->boost
)) {
1490 rcu_do_batch(rsp
, rdp
);
1493 invoke_rcu_callbacks_kthread();
1496 static void invoke_rcu_core(void)
1498 raise_softirq(RCU_SOFTIRQ
);
1501 #ifdef CONFIG_RCU_BOOST
1504 * Wake up the specified per-rcu_node-structure kthread.
1505 * Because the per-rcu_node kthreads are immortal, we don't need
1506 * to do anything to keep them alive.
1508 static void invoke_rcu_node_kthread(struct rcu_node
*rnp
)
1510 struct task_struct
*t
;
1512 t
= rnp
->node_kthread_task
;
1518 * Set the specified CPU's kthread to run RT or not, as specified by
1519 * the to_rt argument. The CPU-hotplug locks are held, so the task
1520 * is not going away.
1522 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1525 struct sched_param sp
;
1526 struct task_struct
*t
;
1528 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1532 policy
= SCHED_FIFO
;
1533 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1535 policy
= SCHED_NORMAL
;
1536 sp
.sched_priority
= 0;
1538 sched_setscheduler_nocheck(t
, policy
, &sp
);
1542 * Timer handler to initiate the waking up of per-CPU kthreads that
1543 * have yielded the CPU due to excess numbers of RCU callbacks.
1544 * We wake up the per-rcu_node kthread, which in turn will wake up
1545 * the booster kthread.
1547 static void rcu_cpu_kthread_timer(unsigned long arg
)
1549 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, arg
);
1550 struct rcu_node
*rnp
= rdp
->mynode
;
1552 atomic_or(rdp
->grpmask
, &rnp
->wakemask
);
1553 invoke_rcu_node_kthread(rnp
);
1557 * Drop to non-real-time priority and yield, but only after posting a
1558 * timer that will cause us to regain our real-time priority if we
1559 * remain preempted. Either way, we restore our real-time priority
1562 static void rcu_yield(void (*f
)(unsigned long), unsigned long arg
)
1564 struct sched_param sp
;
1565 struct timer_list yield_timer
;
1567 setup_timer_on_stack(&yield_timer
, f
, arg
);
1568 mod_timer(&yield_timer
, jiffies
+ 2);
1569 sp
.sched_priority
= 0;
1570 sched_setscheduler_nocheck(current
, SCHED_NORMAL
, &sp
);
1571 set_user_nice(current
, 19);
1573 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1574 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1575 del_timer(&yield_timer
);
1579 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1580 * This can happen while the corresponding CPU is either coming online
1581 * or going offline. We cannot wait until the CPU is fully online
1582 * before starting the kthread, because the various notifier functions
1583 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1584 * the corresponding CPU is online.
1586 * Return 1 if the kthread needs to stop, 0 otherwise.
1588 * Caller must disable bh. This function can momentarily enable it.
1590 static int rcu_cpu_kthread_should_stop(int cpu
)
1592 while (cpu_is_offline(cpu
) ||
1593 !cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)) ||
1594 smp_processor_id() != cpu
) {
1595 if (kthread_should_stop())
1597 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1598 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = raw_smp_processor_id();
1600 schedule_timeout_uninterruptible(1);
1601 if (!cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)))
1602 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
1605 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1610 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1611 * earlier RCU softirq.
1613 static int rcu_cpu_kthread(void *arg
)
1615 int cpu
= (int)(long)arg
;
1616 unsigned long flags
;
1618 unsigned int *statusp
= &per_cpu(rcu_cpu_kthread_status
, cpu
);
1620 char *workp
= &per_cpu(rcu_cpu_has_work
, cpu
);
1623 *statusp
= RCU_KTHREAD_WAITING
;
1624 rcu_wait(*workp
!= 0 || kthread_should_stop());
1626 if (rcu_cpu_kthread_should_stop(cpu
)) {
1630 *statusp
= RCU_KTHREAD_RUNNING
;
1631 per_cpu(rcu_cpu_kthread_loops
, cpu
)++;
1632 local_irq_save(flags
);
1635 local_irq_restore(flags
);
1637 rcu_kthread_do_work();
1644 *statusp
= RCU_KTHREAD_YIELDING
;
1645 rcu_yield(rcu_cpu_kthread_timer
, (unsigned long)cpu
);
1649 *statusp
= RCU_KTHREAD_STOPPED
;
1654 * Spawn a per-CPU kthread, setting up affinity and priority.
1655 * Because the CPU hotplug lock is held, no other CPU will be attempting
1656 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1657 * attempting to access it during boot, but the locking in kthread_bind()
1658 * will enforce sufficient ordering.
1660 * Please note that we cannot simply refuse to wake up the per-CPU
1661 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1662 * which can result in softlockup complaints if the task ends up being
1663 * idle for more than a couple of minutes.
1665 * However, please note also that we cannot bind the per-CPU kthread to its
1666 * CPU until that CPU is fully online. We also cannot wait until the
1667 * CPU is fully online before we create its per-CPU kthread, as this would
1668 * deadlock the system when CPU notifiers tried waiting for grace
1669 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1670 * is online. If its CPU is not yet fully online, then the code in
1671 * rcu_cpu_kthread() will wait until it is fully online, and then do
1674 static int __cpuinit
rcu_spawn_one_cpu_kthread(int cpu
)
1676 struct sched_param sp
;
1677 struct task_struct
*t
;
1679 if (!rcu_kthreads_spawnable
||
1680 per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
)
1682 t
= kthread_create(rcu_cpu_kthread
, (void *)(long)cpu
, "rcuc%d", cpu
);
1685 if (cpu_online(cpu
))
1686 kthread_bind(t
, cpu
);
1687 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1688 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
);
1689 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1690 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1691 per_cpu(rcu_cpu_kthread_task
, cpu
) = t
;
1692 wake_up_process(t
); /* Get to TASK_INTERRUPTIBLE quickly. */
1697 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1698 * kthreads when needed. We ignore requests to wake up kthreads
1699 * for offline CPUs, which is OK because force_quiescent_state()
1700 * takes care of this case.
1702 static int rcu_node_kthread(void *arg
)
1705 unsigned long flags
;
1707 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1708 struct sched_param sp
;
1709 struct task_struct
*t
;
1712 rnp
->node_kthread_status
= RCU_KTHREAD_WAITING
;
1713 rcu_wait(atomic_read(&rnp
->wakemask
) != 0);
1714 rnp
->node_kthread_status
= RCU_KTHREAD_RUNNING
;
1715 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1716 mask
= atomic_xchg(&rnp
->wakemask
, 0);
1717 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1718 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1) {
1719 if ((mask
& 0x1) == 0)
1722 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1723 if (!cpu_online(cpu
) || t
== NULL
) {
1727 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1728 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1729 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1734 rnp
->node_kthread_status
= RCU_KTHREAD_STOPPED
;
1739 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1740 * served by the rcu_node in question. The CPU hotplug lock is still
1741 * held, so the value of rnp->qsmaskinit will be stable.
1743 * We don't include outgoingcpu in the affinity set, use -1 if there is
1744 * no outgoing CPU. If there are no CPUs left in the affinity set,
1745 * this function allows the kthread to execute on any CPU.
1747 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1751 unsigned long mask
= rnp
->qsmaskinit
;
1753 if (rnp
->node_kthread_task
== NULL
)
1755 if (!alloc_cpumask_var(&cm
, GFP_KERNEL
))
1758 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1759 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1760 cpumask_set_cpu(cpu
, cm
);
1761 if (cpumask_weight(cm
) == 0) {
1763 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1764 cpumask_clear_cpu(cpu
, cm
);
1765 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1767 set_cpus_allowed_ptr(rnp
->node_kthread_task
, cm
);
1768 rcu_boost_kthread_setaffinity(rnp
, cm
);
1769 free_cpumask_var(cm
);
1773 * Spawn a per-rcu_node kthread, setting priority and affinity.
1774 * Called during boot before online/offline can happen, or, if
1775 * during runtime, with the main CPU-hotplug locks held. So only
1776 * one of these can be executing at a time.
1778 static int __cpuinit
rcu_spawn_one_node_kthread(struct rcu_state
*rsp
,
1779 struct rcu_node
*rnp
)
1781 unsigned long flags
;
1782 int rnp_index
= rnp
- &rsp
->node
[0];
1783 struct sched_param sp
;
1784 struct task_struct
*t
;
1786 if (!rcu_kthreads_spawnable
||
1787 rnp
->qsmaskinit
== 0)
1789 if (rnp
->node_kthread_task
== NULL
) {
1790 t
= kthread_create(rcu_node_kthread
, (void *)rnp
,
1791 "rcun%d", rnp_index
);
1794 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1795 rnp
->node_kthread_task
= t
;
1796 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1797 sp
.sched_priority
= 99;
1798 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1799 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1801 return rcu_spawn_one_boost_kthread(rsp
, rnp
, rnp_index
);
1805 * Spawn all kthreads -- called as soon as the scheduler is running.
1807 static int __init
rcu_spawn_kthreads(void)
1810 struct rcu_node
*rnp
;
1812 rcu_kthreads_spawnable
= 1;
1813 for_each_possible_cpu(cpu
) {
1814 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1815 if (cpu_online(cpu
))
1816 (void)rcu_spawn_one_cpu_kthread(cpu
);
1818 rnp
= rcu_get_root(rcu_state
);
1819 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1820 if (NUM_RCU_NODES
> 1) {
1821 rcu_for_each_leaf_node(rcu_state
, rnp
)
1822 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1826 early_initcall(rcu_spawn_kthreads
);
1828 #else /* #ifdef CONFIG_RCU_BOOST */
1830 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1834 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1838 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1841 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1842 struct rcu_state
*rsp
)
1844 unsigned long flags
;
1845 struct rcu_data
*rdp
;
1847 debug_rcu_head_queue(head
);
1851 smp_mb(); /* Ensure RCU update seen before callback registry. */
1854 * Opportunistically note grace-period endings and beginnings.
1855 * Note that we might see a beginning right after we see an
1856 * end, but never vice versa, since this CPU has to pass through
1857 * a quiescent state betweentimes.
1859 local_irq_save(flags
);
1860 rdp
= this_cpu_ptr(rsp
->rda
);
1862 /* Add the callback to our list. */
1863 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1864 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1867 /* If interrupts were disabled, don't dive into RCU core. */
1868 if (irqs_disabled_flags(flags
)) {
1869 local_irq_restore(flags
);
1874 * Force the grace period if too many callbacks or too long waiting.
1875 * Enforce hysteresis, and don't invoke force_quiescent_state()
1876 * if some other CPU has recently done so. Also, don't bother
1877 * invoking force_quiescent_state() if the newly enqueued callback
1878 * is the only one waiting for a grace period to complete.
1880 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1882 /* Are we ignoring a completed grace period? */
1883 rcu_process_gp_end(rsp
, rdp
);
1884 check_for_new_grace_period(rsp
, rdp
);
1886 /* Start a new grace period if one not already started. */
1887 if (!rcu_gp_in_progress(rsp
)) {
1888 unsigned long nestflag
;
1889 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1891 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1892 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1894 /* Give the grace period a kick. */
1895 rdp
->blimit
= LONG_MAX
;
1896 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1897 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1898 force_quiescent_state(rsp
, 0);
1899 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1900 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1902 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1903 force_quiescent_state(rsp
, 1);
1904 local_irq_restore(flags
);
1908 * Queue an RCU-sched callback for invocation after a grace period.
1910 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1912 __call_rcu(head
, func
, &rcu_sched_state
);
1914 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1917 * Queue an RCU for invocation after a quicker grace period.
1919 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1921 __call_rcu(head
, func
, &rcu_bh_state
);
1923 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1926 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1928 * Control will return to the caller some time after a full rcu-sched
1929 * grace period has elapsed, in other words after all currently executing
1930 * rcu-sched read-side critical sections have completed. These read-side
1931 * critical sections are delimited by rcu_read_lock_sched() and
1932 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1933 * local_irq_disable(), and so on may be used in place of
1934 * rcu_read_lock_sched().
1936 * This means that all preempt_disable code sequences, including NMI and
1937 * hardware-interrupt handlers, in progress on entry will have completed
1938 * before this primitive returns. However, this does not guarantee that
1939 * softirq handlers will have completed, since in some kernels, these
1940 * handlers can run in process context, and can block.
1942 * This primitive provides the guarantees made by the (now removed)
1943 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1944 * guarantees that rcu_read_lock() sections will have completed.
1945 * In "classic RCU", these two guarantees happen to be one and
1946 * the same, but can differ in realtime RCU implementations.
1948 void synchronize_sched(void)
1950 struct rcu_synchronize rcu
;
1952 if (rcu_blocking_is_gp())
1955 init_rcu_head_on_stack(&rcu
.head
);
1956 init_completion(&rcu
.completion
);
1957 /* Will wake me after RCU finished. */
1958 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1960 wait_for_completion(&rcu
.completion
);
1961 destroy_rcu_head_on_stack(&rcu
.head
);
1963 EXPORT_SYMBOL_GPL(synchronize_sched
);
1966 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1968 * Control will return to the caller some time after a full rcu_bh grace
1969 * period has elapsed, in other words after all currently executing rcu_bh
1970 * read-side critical sections have completed. RCU read-side critical
1971 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1972 * and may be nested.
1974 void synchronize_rcu_bh(void)
1976 struct rcu_synchronize rcu
;
1978 if (rcu_blocking_is_gp())
1981 init_rcu_head_on_stack(&rcu
.head
);
1982 init_completion(&rcu
.completion
);
1983 /* Will wake me after RCU finished. */
1984 call_rcu_bh(&rcu
.head
, wakeme_after_rcu
);
1986 wait_for_completion(&rcu
.completion
);
1987 destroy_rcu_head_on_stack(&rcu
.head
);
1989 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1992 * Check to see if there is any immediate RCU-related work to be done
1993 * by the current CPU, for the specified type of RCU, returning 1 if so.
1994 * The checks are in order of increasing expense: checks that can be
1995 * carried out against CPU-local state are performed first. However,
1996 * we must check for CPU stalls first, else we might not get a chance.
1998 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2000 struct rcu_node
*rnp
= rdp
->mynode
;
2002 rdp
->n_rcu_pending
++;
2004 /* Check for CPU stalls, if enabled. */
2005 check_cpu_stall(rsp
, rdp
);
2007 /* Is the RCU core waiting for a quiescent state from this CPU? */
2008 if (rdp
->qs_pending
&& !rdp
->passed_quiesc
) {
2011 * If force_quiescent_state() coming soon and this CPU
2012 * needs a quiescent state, and this is either RCU-sched
2013 * or RCU-bh, force a local reschedule.
2015 rdp
->n_rp_qs_pending
++;
2016 if (!rdp
->preemptible
&&
2017 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
2020 } else if (rdp
->qs_pending
&& rdp
->passed_quiesc
) {
2021 rdp
->n_rp_report_qs
++;
2025 /* Does this CPU have callbacks ready to invoke? */
2026 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2027 rdp
->n_rp_cb_ready
++;
2031 /* Has RCU gone idle with this CPU needing another grace period? */
2032 if (cpu_needs_another_gp(rsp
, rdp
)) {
2033 rdp
->n_rp_cpu_needs_gp
++;
2037 /* Has another RCU grace period completed? */
2038 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2039 rdp
->n_rp_gp_completed
++;
2043 /* Has a new RCU grace period started? */
2044 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2045 rdp
->n_rp_gp_started
++;
2049 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2050 if (rcu_gp_in_progress(rsp
) &&
2051 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
2052 rdp
->n_rp_need_fqs
++;
2057 rdp
->n_rp_need_nothing
++;
2062 * Check to see if there is any immediate RCU-related work to be done
2063 * by the current CPU, returning 1 if so. This function is part of the
2064 * RCU implementation; it is -not- an exported member of the RCU API.
2066 static int rcu_pending(int cpu
)
2068 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
2069 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
2070 rcu_preempt_pending(cpu
);
2074 * Check to see if any future RCU-related work will need to be done
2075 * by the current CPU, even if none need be done immediately, returning
2078 static int rcu_needs_cpu_quick_check(int cpu
)
2080 /* RCU callbacks either ready or pending? */
2081 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
2082 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
2083 rcu_preempt_needs_cpu(cpu
);
2086 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
2087 static atomic_t rcu_barrier_cpu_count
;
2088 static DEFINE_MUTEX(rcu_barrier_mutex
);
2089 static struct completion rcu_barrier_completion
;
2091 static void rcu_barrier_callback(struct rcu_head
*notused
)
2093 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2094 complete(&rcu_barrier_completion
);
2098 * Called with preemption disabled, and from cross-cpu IRQ context.
2100 static void rcu_barrier_func(void *type
)
2102 int cpu
= smp_processor_id();
2103 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
2104 void (*call_rcu_func
)(struct rcu_head
*head
,
2105 void (*func
)(struct rcu_head
*head
));
2107 atomic_inc(&rcu_barrier_cpu_count
);
2108 call_rcu_func
= type
;
2109 call_rcu_func(head
, rcu_barrier_callback
);
2113 * Orchestrate the specified type of RCU barrier, waiting for all
2114 * RCU callbacks of the specified type to complete.
2116 static void _rcu_barrier(struct rcu_state
*rsp
,
2117 void (*call_rcu_func
)(struct rcu_head
*head
,
2118 void (*func
)(struct rcu_head
*head
)))
2120 BUG_ON(in_interrupt());
2121 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2122 mutex_lock(&rcu_barrier_mutex
);
2123 init_completion(&rcu_barrier_completion
);
2125 * Initialize rcu_barrier_cpu_count to 1, then invoke
2126 * rcu_barrier_func() on each CPU, so that each CPU also has
2127 * incremented rcu_barrier_cpu_count. Only then is it safe to
2128 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2129 * might complete its grace period before all of the other CPUs
2130 * did their increment, causing this function to return too
2131 * early. Note that on_each_cpu() disables irqs, which prevents
2132 * any CPUs from coming online or going offline until each online
2133 * CPU has queued its RCU-barrier callback.
2135 atomic_set(&rcu_barrier_cpu_count
, 1);
2136 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
2137 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2138 complete(&rcu_barrier_completion
);
2139 wait_for_completion(&rcu_barrier_completion
);
2140 mutex_unlock(&rcu_barrier_mutex
);
2144 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2146 void rcu_barrier_bh(void)
2148 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
2150 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2153 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2155 void rcu_barrier_sched(void)
2157 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2159 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2162 * Do boot-time initialization of a CPU's per-CPU RCU data.
2165 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2167 unsigned long flags
;
2169 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2170 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2172 /* Set up local state, ensuring consistent view of global state. */
2173 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2174 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2175 rdp
->nxtlist
= NULL
;
2176 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2177 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2180 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2181 #endif /* #ifdef CONFIG_NO_HZ */
2183 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2187 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2188 * offline event can be happening at a given time. Note also that we
2189 * can accept some slop in the rsp->completed access due to the fact
2190 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2192 static void __cpuinit
2193 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2195 unsigned long flags
;
2197 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2198 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2200 /* Set up local state, ensuring consistent view of global state. */
2201 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2202 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
2203 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
2204 rdp
->beenonline
= 1; /* We have now been online. */
2205 rdp
->preemptible
= preemptible
;
2206 rdp
->qlen_last_fqs_check
= 0;
2207 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2208 rdp
->blimit
= blimit
;
2209 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2212 * A new grace period might start here. If so, we won't be part
2213 * of it, but that is OK, as we are currently in a quiescent state.
2216 /* Exclude any attempts to start a new GP on large systems. */
2217 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2219 /* Add CPU to rcu_node bitmasks. */
2221 mask
= rdp
->grpmask
;
2223 /* Exclude any attempts to start a new GP on small systems. */
2224 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2225 rnp
->qsmaskinit
|= mask
;
2226 mask
= rnp
->grpmask
;
2227 if (rnp
== rdp
->mynode
) {
2228 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
2229 rdp
->completed
= rnp
->completed
;
2230 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
2232 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2234 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2236 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2239 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2241 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2242 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2243 rcu_preempt_init_percpu_data(cpu
);
2246 #ifdef CONFIG_RCU_BOOST
2248 static void __cpuinit
rcu_prepare_kthreads(int cpu
)
2250 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2251 struct rcu_node
*rnp
= rdp
->mynode
;
2253 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
2254 if (rcu_kthreads_spawnable
) {
2255 (void)rcu_spawn_one_cpu_kthread(cpu
);
2256 if (rnp
->node_kthread_task
== NULL
)
2257 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
2261 #else /* #ifdef CONFIG_RCU_BOOST */
2263 static void __cpuinit
rcu_prepare_kthreads(int cpu
)
2267 #endif /* #else #ifdef CONFIG_RCU_BOOST */
2270 * Handle CPU online/offline notification events.
2272 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2273 unsigned long action
, void *hcpu
)
2275 long cpu
= (long)hcpu
;
2276 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2277 struct rcu_node
*rnp
= rdp
->mynode
;
2280 case CPU_UP_PREPARE
:
2281 case CPU_UP_PREPARE_FROZEN
:
2282 rcu_prepare_cpu(cpu
);
2283 rcu_prepare_kthreads(cpu
);
2286 case CPU_DOWN_FAILED
:
2287 rcu_node_kthread_setaffinity(rnp
, -1);
2288 rcu_cpu_kthread_setrt(cpu
, 1);
2290 case CPU_DOWN_PREPARE
:
2291 rcu_node_kthread_setaffinity(rnp
, cpu
);
2292 rcu_cpu_kthread_setrt(cpu
, 0);
2295 case CPU_DYING_FROZEN
:
2297 * The whole machine is "stopped" except this CPU, so we can
2298 * touch any data without introducing corruption. We send the
2299 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2301 rcu_send_cbs_to_online(&rcu_bh_state
);
2302 rcu_send_cbs_to_online(&rcu_sched_state
);
2303 rcu_preempt_send_cbs_to_online();
2306 case CPU_DEAD_FROZEN
:
2307 case CPU_UP_CANCELED
:
2308 case CPU_UP_CANCELED_FROZEN
:
2309 rcu_offline_cpu(cpu
);
2318 * This function is invoked towards the end of the scheduler's initialization
2319 * process. Before this is called, the idle task might contain
2320 * RCU read-side critical sections (during which time, this idle
2321 * task is booting the system). After this function is called, the
2322 * idle tasks are prohibited from containing RCU read-side critical
2323 * sections. This function also enables RCU lockdep checking.
2325 void rcu_scheduler_starting(void)
2327 WARN_ON(num_online_cpus() != 1);
2328 WARN_ON(nr_context_switches() > 0);
2329 rcu_scheduler_active
= 1;
2333 * Compute the per-level fanout, either using the exact fanout specified
2334 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2336 #ifdef CONFIG_RCU_FANOUT_EXACT
2337 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2341 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2342 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2343 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
2345 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2346 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2353 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2354 ccur
= rsp
->levelcnt
[i
];
2355 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2359 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2362 * Helper function for rcu_init() that initializes one rcu_state structure.
2364 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2365 struct rcu_data __percpu
*rda
)
2367 static char *buf
[] = { "rcu_node_level_0",
2370 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2374 struct rcu_node
*rnp
;
2376 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2378 /* Initialize the level-tracking arrays. */
2380 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2381 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2382 rcu_init_levelspread(rsp
);
2384 /* Initialize the elements themselves, starting from the leaves. */
2386 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2387 cpustride
*= rsp
->levelspread
[i
];
2388 rnp
= rsp
->level
[i
];
2389 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2390 raw_spin_lock_init(&rnp
->lock
);
2391 lockdep_set_class_and_name(&rnp
->lock
,
2392 &rcu_node_class
[i
], buf
[i
]);
2395 rnp
->qsmaskinit
= 0;
2396 rnp
->grplo
= j
* cpustride
;
2397 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2398 if (rnp
->grphi
>= NR_CPUS
)
2399 rnp
->grphi
= NR_CPUS
- 1;
2405 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2406 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2407 rnp
->parent
= rsp
->level
[i
- 1] +
2408 j
/ rsp
->levelspread
[i
- 1];
2411 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2416 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2417 for_each_possible_cpu(i
) {
2418 while (i
> rnp
->grphi
)
2420 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2421 rcu_boot_init_percpu_data(i
, rsp
);
2425 void __init
rcu_init(void)
2429 rcu_bootup_announce();
2430 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2431 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2432 __rcu_init_preempt();
2433 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2436 * We don't need protection against CPU-hotplug here because
2437 * this is called early in boot, before either interrupts
2438 * or the scheduler are operational.
2440 cpu_notifier(rcu_cpu_notify
, 0);
2441 for_each_online_cpu(cpu
)
2442 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2443 check_cpu_stall_init();
2446 #include "rcutree_plugin.h"