2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <asm/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
55 /* Data structures. */
57 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
59 #define RCU_STATE_INITIALIZER(structname) { \
60 .level = { &structname.node[0] }, \
62 NUM_RCU_LVL_0, /* root of hierarchy. */ \
66 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
68 .signaled = RCU_GP_IDLE, \
71 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
72 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
74 .n_force_qs_ngp = 0, \
75 .name = #structname, \
78 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched_state
);
79 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
81 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh_state
);
82 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
84 static struct rcu_state
*rcu_state
;
86 int rcu_scheduler_active __read_mostly
;
87 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
90 * Control variables for per-CPU and per-rcu_node kthreads. These
91 * handle all flavors of RCU.
93 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
94 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
95 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
96 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
97 static DEFINE_PER_CPU(wait_queue_head_t
, rcu_cpu_wq
);
98 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
99 static char rcu_kthreads_spawnable
;
101 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
102 static void invoke_rcu_cpu_kthread(void);
104 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
107 * Track the rcutorture test sequence number and the update version
108 * number within a given test. The rcutorture_testseq is incremented
109 * on every rcutorture module load and unload, so has an odd value
110 * when a test is running. The rcutorture_vernum is set to zero
111 * when rcutorture starts and is incremented on each rcutorture update.
112 * These variables enable correlating rcutorture output with the
113 * RCU tracing information.
115 unsigned long rcutorture_testseq
;
116 unsigned long rcutorture_vernum
;
119 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
120 * permit this function to be invoked without holding the root rcu_node
121 * structure's ->lock, but of course results can be subject to change.
123 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
125 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
129 * Note a quiescent state. Because we do not need to know
130 * how many quiescent states passed, just if there was at least
131 * one since the start of the grace period, this just sets a flag.
133 void rcu_sched_qs(int cpu
)
135 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
137 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
139 rdp
->passed_quiesc
= 1;
142 void rcu_bh_qs(int cpu
)
144 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
146 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
148 rdp
->passed_quiesc
= 1;
152 * Note a context switch. This is a quiescent state for RCU-sched,
153 * and requires special handling for preemptible RCU.
155 void rcu_note_context_switch(int cpu
)
158 rcu_preempt_note_context_switch(cpu
);
160 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
163 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
164 .dynticks_nesting
= 1,
167 #endif /* #ifdef CONFIG_NO_HZ */
169 static int blimit
= 10; /* Maximum callbacks per softirq. */
170 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
171 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
173 module_param(blimit
, int, 0);
174 module_param(qhimark
, int, 0);
175 module_param(qlowmark
, int, 0);
177 int rcu_cpu_stall_suppress __read_mostly
;
178 module_param(rcu_cpu_stall_suppress
, int, 0644);
180 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
181 static int rcu_pending(int cpu
);
184 * Return the number of RCU-sched batches processed thus far for debug & stats.
186 long rcu_batches_completed_sched(void)
188 return rcu_sched_state
.completed
;
190 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
193 * Return the number of RCU BH batches processed thus far for debug & stats.
195 long rcu_batches_completed_bh(void)
197 return rcu_bh_state
.completed
;
199 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
202 * Force a quiescent state for RCU BH.
204 void rcu_bh_force_quiescent_state(void)
206 force_quiescent_state(&rcu_bh_state
, 0);
208 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
211 * Record the number of times rcutorture tests have been initiated and
212 * terminated. This information allows the debugfs tracing stats to be
213 * correlated to the rcutorture messages, even when the rcutorture module
214 * is being repeatedly loaded and unloaded. In other words, we cannot
215 * store this state in rcutorture itself.
217 void rcutorture_record_test_transition(void)
219 rcutorture_testseq
++;
220 rcutorture_vernum
= 0;
222 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
225 * Record the number of writer passes through the current rcutorture test.
226 * This is also used to correlate debugfs tracing stats with the rcutorture
229 void rcutorture_record_progress(unsigned long vernum
)
233 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
236 * Force a quiescent state for RCU-sched.
238 void rcu_sched_force_quiescent_state(void)
240 force_quiescent_state(&rcu_sched_state
, 0);
242 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
245 * Does the CPU have callbacks ready to be invoked?
248 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
250 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
254 * Does the current CPU require a yet-as-unscheduled grace period?
257 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
259 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
263 * Return the root node of the specified rcu_state structure.
265 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
267 return &rsp
->node
[0];
273 * If the specified CPU is offline, tell the caller that it is in
274 * a quiescent state. Otherwise, whack it with a reschedule IPI.
275 * Grace periods can end up waiting on an offline CPU when that
276 * CPU is in the process of coming online -- it will be added to the
277 * rcu_node bitmasks before it actually makes it online. The same thing
278 * can happen while a CPU is in the process of coming online. Because this
279 * race is quite rare, we check for it after detecting that the grace
280 * period has been delayed rather than checking each and every CPU
281 * each and every time we start a new grace period.
283 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
286 * If the CPU is offline, it is in a quiescent state. We can
287 * trust its state not to change because interrupts are disabled.
289 if (cpu_is_offline(rdp
->cpu
)) {
294 /* If preemptible RCU, no point in sending reschedule IPI. */
295 if (rdp
->preemptible
)
298 /* The CPU is online, so send it a reschedule IPI. */
299 if (rdp
->cpu
!= smp_processor_id())
300 smp_send_reschedule(rdp
->cpu
);
307 #endif /* #ifdef CONFIG_SMP */
312 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
314 * Enter nohz mode, in other words, -leave- the mode in which RCU
315 * read-side critical sections can occur. (Though RCU read-side
316 * critical sections can occur in irq handlers in nohz mode, a possibility
317 * handled by rcu_irq_enter() and rcu_irq_exit()).
319 void rcu_enter_nohz(void)
322 struct rcu_dynticks
*rdtp
;
324 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
325 local_irq_save(flags
);
326 rdtp
= &__get_cpu_var(rcu_dynticks
);
327 if (--rdtp
->dynticks_nesting
== 0)
329 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
330 local_irq_restore(flags
);
334 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
336 * Exit nohz mode, in other words, -enter- the mode in which RCU
337 * read-side critical sections normally occur.
339 void rcu_exit_nohz(void)
342 struct rcu_dynticks
*rdtp
;
344 local_irq_save(flags
);
345 rdtp
= &__get_cpu_var(rcu_dynticks
);
347 rdtp
->dynticks_nesting
++;
348 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
349 local_irq_restore(flags
);
350 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
354 * rcu_nmi_enter - inform RCU of entry to NMI context
356 * If the CPU was idle with dynamic ticks active, and there is no
357 * irq handler running, this updates rdtp->dynticks_nmi to let the
358 * RCU grace-period handling know that the CPU is active.
360 void rcu_nmi_enter(void)
362 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
364 if (rdtp
->dynticks
& 0x1)
366 rdtp
->dynticks_nmi
++;
367 WARN_ON_ONCE(!(rdtp
->dynticks_nmi
& 0x1));
368 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
372 * rcu_nmi_exit - inform RCU of exit from NMI context
374 * If the CPU was idle with dynamic ticks active, and there is no
375 * irq handler running, this updates rdtp->dynticks_nmi to let the
376 * RCU grace-period handling know that the CPU is no longer active.
378 void rcu_nmi_exit(void)
380 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
382 if (rdtp
->dynticks
& 0x1)
384 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
385 rdtp
->dynticks_nmi
++;
386 WARN_ON_ONCE(rdtp
->dynticks_nmi
& 0x1);
390 * rcu_irq_enter - inform RCU of entry to hard irq context
392 * If the CPU was idle with dynamic ticks active, this updates the
393 * rdtp->dynticks to let the RCU handling know that the CPU is active.
395 void rcu_irq_enter(void)
397 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
399 if (rdtp
->dynticks_nesting
++)
402 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
403 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
407 * rcu_irq_exit - inform RCU of exit from hard irq context
409 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
410 * to put let the RCU handling be aware that the CPU is going back to idle
413 void rcu_irq_exit(void)
415 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
417 if (--rdtp
->dynticks_nesting
)
419 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
421 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
423 /* If the interrupt queued a callback, get out of dyntick mode. */
425 (__this_cpu_read(rcu_sched_data
.nxtlist
) ||
426 __this_cpu_read(rcu_bh_data
.nxtlist
)))
433 * Snapshot the specified CPU's dynticks counter so that we can later
434 * credit them with an implicit quiescent state. Return 1 if this CPU
435 * is in dynticks idle mode, which is an extended quiescent state.
437 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
443 snap
= rdp
->dynticks
->dynticks
;
444 snap_nmi
= rdp
->dynticks
->dynticks_nmi
;
445 smp_mb(); /* Order sampling of snap with end of grace period. */
446 rdp
->dynticks_snap
= snap
;
447 rdp
->dynticks_nmi_snap
= snap_nmi
;
448 ret
= ((snap
& 0x1) == 0) && ((snap_nmi
& 0x1) == 0);
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
)
467 curr
= rdp
->dynticks
->dynticks
;
468 snap
= rdp
->dynticks_snap
;
469 curr_nmi
= rdp
->dynticks
->dynticks_nmi
;
470 snap_nmi
= rdp
->dynticks_nmi_snap
;
471 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
474 * If the CPU passed through or entered a dynticks idle phase with
475 * no active irq/NMI handlers, then we can safely pretend that the CPU
476 * already acknowledged the request to pass through a quiescent
477 * state. Either way, that CPU cannot possibly be in an RCU
478 * read-side critical section that started before the beginning
479 * of the current RCU grace period.
481 if ((curr
!= snap
|| (curr
& 0x1) == 0) &&
482 (curr_nmi
!= snap_nmi
|| (curr_nmi
& 0x1) == 0)) {
487 /* Go check for the CPU being offline. */
488 return rcu_implicit_offline_qs(rdp
);
491 #endif /* #ifdef CONFIG_SMP */
493 #else /* #ifdef CONFIG_NO_HZ */
497 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
502 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
504 return rcu_implicit_offline_qs(rdp
);
507 #endif /* #ifdef CONFIG_SMP */
509 #endif /* #else #ifdef CONFIG_NO_HZ */
511 int rcu_cpu_stall_suppress __read_mostly
;
513 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
515 rsp
->gp_start
= jiffies
;
516 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
519 static void print_other_cpu_stall(struct rcu_state
*rsp
)
524 struct rcu_node
*rnp
= rcu_get_root(rsp
);
526 /* Only let one CPU complain about others per time interval. */
528 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
529 delta
= jiffies
- rsp
->jiffies_stall
;
530 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
531 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
534 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
537 * Now rat on any tasks that got kicked up to the root rcu_node
538 * due to CPU offlining.
540 rcu_print_task_stall(rnp
);
541 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
544 * OK, time to rat on our buddy...
545 * See Documentation/RCU/stallwarn.txt for info on how to debug
546 * RCU CPU stall warnings.
548 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
550 rcu_for_each_leaf_node(rsp
, rnp
) {
551 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
552 rcu_print_task_stall(rnp
);
553 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
554 if (rnp
->qsmask
== 0)
556 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
557 if (rnp
->qsmask
& (1UL << cpu
))
558 printk(" %d", rnp
->grplo
+ cpu
);
560 printk("} (detected by %d, t=%ld jiffies)\n",
561 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
562 trigger_all_cpu_backtrace();
564 /* If so configured, complain about tasks blocking the grace period. */
566 rcu_print_detail_task_stall(rsp
);
568 force_quiescent_state(rsp
, 0); /* Kick them all. */
571 static void print_cpu_stall(struct rcu_state
*rsp
)
574 struct rcu_node
*rnp
= rcu_get_root(rsp
);
577 * OK, time to rat on ourselves...
578 * See Documentation/RCU/stallwarn.txt for info on how to debug
579 * RCU CPU stall warnings.
581 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
582 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
583 trigger_all_cpu_backtrace();
585 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
586 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
588 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
589 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
591 set_need_resched(); /* kick ourselves to get things going. */
594 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
598 struct rcu_node
*rnp
;
600 if (rcu_cpu_stall_suppress
)
602 j
= ACCESS_ONCE(jiffies
);
603 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
605 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
607 /* We haven't checked in, so go dump stack. */
608 print_cpu_stall(rsp
);
610 } else if (rcu_gp_in_progress(rsp
) &&
611 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
613 /* They had a few time units to dump stack, so complain. */
614 print_other_cpu_stall(rsp
);
618 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
620 rcu_cpu_stall_suppress
= 1;
625 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
627 * Set the stall-warning timeout way off into the future, thus preventing
628 * any RCU CPU stall-warning messages from appearing in the current set of
631 * The caller must disable hard irqs.
633 void rcu_cpu_stall_reset(void)
635 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
636 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
637 rcu_preempt_stall_reset();
640 static struct notifier_block rcu_panic_block
= {
641 .notifier_call
= rcu_panic
,
644 static void __init
check_cpu_stall_init(void)
646 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
650 * Update CPU-local rcu_data state to record the newly noticed grace period.
651 * This is used both when we started the grace period and when we notice
652 * that someone else started the grace period. The caller must hold the
653 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
654 * and must have irqs disabled.
656 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
658 if (rdp
->gpnum
!= rnp
->gpnum
) {
660 * If the current grace period is waiting for this CPU,
661 * set up to detect a quiescent state, otherwise don't
662 * go looking for one.
664 rdp
->gpnum
= rnp
->gpnum
;
665 if (rnp
->qsmask
& rdp
->grpmask
) {
667 rdp
->passed_quiesc
= 0;
673 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
676 struct rcu_node
*rnp
;
678 local_irq_save(flags
);
680 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
681 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
682 local_irq_restore(flags
);
685 __note_new_gpnum(rsp
, rnp
, rdp
);
686 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
690 * Did someone else start a new RCU grace period start since we last
691 * checked? Update local state appropriately if so. Must be called
692 * on the CPU corresponding to rdp.
695 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
700 local_irq_save(flags
);
701 if (rdp
->gpnum
!= rsp
->gpnum
) {
702 note_new_gpnum(rsp
, rdp
);
705 local_irq_restore(flags
);
710 * Advance this CPU's callbacks, but only if the current grace period
711 * has ended. This may be called only from the CPU to whom the rdp
712 * belongs. In addition, the corresponding leaf rcu_node structure's
713 * ->lock must be held by the caller, with irqs disabled.
716 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
718 /* Did another grace period end? */
719 if (rdp
->completed
!= rnp
->completed
) {
721 /* Advance callbacks. No harm if list empty. */
722 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
723 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
724 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
726 /* Remember that we saw this grace-period completion. */
727 rdp
->completed
= rnp
->completed
;
730 * If we were in an extended quiescent state, we may have
731 * missed some grace periods that others CPUs handled on
732 * our behalf. Catch up with this state to avoid noting
733 * spurious new grace periods. If another grace period
734 * has started, then rnp->gpnum will have advanced, so
735 * we will detect this later on.
737 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
738 rdp
->gpnum
= rdp
->completed
;
741 * If RCU does not need a quiescent state from this CPU,
742 * then make sure that this CPU doesn't go looking for one.
744 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
750 * Advance this CPU's callbacks, but only if the current grace period
751 * has ended. This may be called only from the CPU to whom the rdp
755 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
758 struct rcu_node
*rnp
;
760 local_irq_save(flags
);
762 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
763 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
764 local_irq_restore(flags
);
767 __rcu_process_gp_end(rsp
, rnp
, rdp
);
768 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
772 * Do per-CPU grace-period initialization for running CPU. The caller
773 * must hold the lock of the leaf rcu_node structure corresponding to
777 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
779 /* Prior grace period ended, so advance callbacks for current CPU. */
780 __rcu_process_gp_end(rsp
, rnp
, rdp
);
783 * Because this CPU just now started the new grace period, we know
784 * that all of its callbacks will be covered by this upcoming grace
785 * period, even the ones that were registered arbitrarily recently.
786 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
788 * Other CPUs cannot be sure exactly when the grace period started.
789 * Therefore, their recently registered callbacks must pass through
790 * an additional RCU_NEXT_READY stage, so that they will be handled
791 * by the next RCU grace period.
793 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
794 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
796 /* Set state so that this CPU will detect the next quiescent state. */
797 __note_new_gpnum(rsp
, rnp
, rdp
);
801 * Start a new RCU grace period if warranted, re-initializing the hierarchy
802 * in preparation for detecting the next grace period. The caller must hold
803 * the root node's ->lock, which is released before return. Hard irqs must
807 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
808 __releases(rcu_get_root(rsp
)->lock
)
810 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
811 struct rcu_node
*rnp
= rcu_get_root(rsp
);
813 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
814 if (cpu_needs_another_gp(rsp
, rdp
))
815 rsp
->fqs_need_gp
= 1;
816 if (rnp
->completed
== rsp
->completed
) {
817 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
820 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
823 * Propagate new ->completed value to rcu_node structures
824 * so that other CPUs don't have to wait until the start
825 * of the next grace period to process their callbacks.
827 rcu_for_each_node_breadth_first(rsp
, rnp
) {
828 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
829 rnp
->completed
= rsp
->completed
;
830 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
832 local_irq_restore(flags
);
836 /* Advance to a new grace period and initialize state. */
838 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
839 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
840 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
841 record_gp_stall_check_time(rsp
);
843 /* Special-case the common single-level case. */
844 if (NUM_RCU_NODES
== 1) {
845 rcu_preempt_check_blocked_tasks(rnp
);
846 rnp
->qsmask
= rnp
->qsmaskinit
;
847 rnp
->gpnum
= rsp
->gpnum
;
848 rnp
->completed
= rsp
->completed
;
849 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
850 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
851 rcu_preempt_boost_start_gp(rnp
);
852 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
856 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
859 /* Exclude any concurrent CPU-hotplug operations. */
860 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
863 * Set the quiescent-state-needed bits in all the rcu_node
864 * structures for all currently online CPUs in breadth-first
865 * order, starting from the root rcu_node structure. This
866 * operation relies on the layout of the hierarchy within the
867 * rsp->node[] array. Note that other CPUs will access only
868 * the leaves of the hierarchy, which still indicate that no
869 * grace period is in progress, at least until the corresponding
870 * leaf node has been initialized. In addition, we have excluded
871 * CPU-hotplug operations.
873 * Note that the grace period cannot complete until we finish
874 * the initialization process, as there will be at least one
875 * qsmask bit set in the root node until that time, namely the
876 * one corresponding to this CPU, due to the fact that we have
879 rcu_for_each_node_breadth_first(rsp
, rnp
) {
880 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
881 rcu_preempt_check_blocked_tasks(rnp
);
882 rnp
->qsmask
= rnp
->qsmaskinit
;
883 rnp
->gpnum
= rsp
->gpnum
;
884 rnp
->completed
= rsp
->completed
;
885 if (rnp
== rdp
->mynode
)
886 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
887 rcu_preempt_boost_start_gp(rnp
);
888 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
891 rnp
= rcu_get_root(rsp
);
892 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
893 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
894 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
895 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
899 * Report a full set of quiescent states to the specified rcu_state
900 * data structure. This involves cleaning up after the prior grace
901 * period and letting rcu_start_gp() start up the next grace period
902 * if one is needed. Note that the caller must hold rnp->lock, as
903 * required by rcu_start_gp(), which will release it.
905 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
906 __releases(rcu_get_root(rsp
)->lock
)
908 unsigned long gp_duration
;
910 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
913 * Ensure that all grace-period and pre-grace-period activity
914 * is seen before the assignment to rsp->completed.
916 smp_mb(); /* See above block comment. */
917 gp_duration
= jiffies
- rsp
->gp_start
;
918 if (gp_duration
> rsp
->gp_max
)
919 rsp
->gp_max
= gp_duration
;
920 rsp
->completed
= rsp
->gpnum
;
921 rsp
->signaled
= RCU_GP_IDLE
;
922 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
926 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
927 * Allows quiescent states for a group of CPUs to be reported at one go
928 * to the specified rcu_node structure, though all the CPUs in the group
929 * must be represented by the same rcu_node structure (which need not be
930 * a leaf rcu_node structure, though it often will be). That structure's
931 * lock must be held upon entry, and it is released before return.
934 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
935 struct rcu_node
*rnp
, unsigned long flags
)
936 __releases(rnp
->lock
)
938 struct rcu_node
*rnp_c
;
940 /* Walk up the rcu_node hierarchy. */
942 if (!(rnp
->qsmask
& mask
)) {
944 /* Our bit has already been cleared, so done. */
945 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
948 rnp
->qsmask
&= ~mask
;
949 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
951 /* Other bits still set at this level, so done. */
952 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
956 if (rnp
->parent
== NULL
) {
958 /* No more levels. Exit loop holding root lock. */
962 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
965 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
966 WARN_ON_ONCE(rnp_c
->qsmask
);
970 * Get here if we are the last CPU to pass through a quiescent
971 * state for this grace period. Invoke rcu_report_qs_rsp()
972 * to clean up and start the next grace period if one is needed.
974 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
978 * Record a quiescent state for the specified CPU to that CPU's rcu_data
979 * structure. This must be either called from the specified CPU, or
980 * called when the specified CPU is known to be offline (and when it is
981 * also known that no other CPU is concurrently trying to help the offline
982 * CPU). The lastcomp argument is used to make sure we are still in the
983 * grace period of interest. We don't want to end the current grace period
984 * based on quiescent states detected in an earlier grace period!
987 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
991 struct rcu_node
*rnp
;
994 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
995 if (lastcomp
!= rnp
->completed
) {
998 * Someone beat us to it for this grace period, so leave.
999 * The race with GP start is resolved by the fact that we
1000 * hold the leaf rcu_node lock, so that the per-CPU bits
1001 * cannot yet be initialized -- so we would simply find our
1002 * CPU's bit already cleared in rcu_report_qs_rnp() if this
1005 rdp
->passed_quiesc
= 0; /* try again later! */
1006 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1009 mask
= rdp
->grpmask
;
1010 if ((rnp
->qsmask
& mask
) == 0) {
1011 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1013 rdp
->qs_pending
= 0;
1016 * This GP can't end until cpu checks in, so all of our
1017 * callbacks can be processed during the next GP.
1019 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1021 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1026 * Check to see if there is a new grace period of which this CPU
1027 * is not yet aware, and if so, set up local rcu_data state for it.
1028 * Otherwise, see if this CPU has just passed through its first
1029 * quiescent state for this grace period, and record that fact if so.
1032 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1034 /* If there is now a new grace period, record and return. */
1035 if (check_for_new_grace_period(rsp
, rdp
))
1039 * Does this CPU still need to do its part for current grace period?
1040 * If no, return and let the other CPUs do their part as well.
1042 if (!rdp
->qs_pending
)
1046 * Was there a quiescent state since the beginning of the grace
1047 * period? If no, then exit and wait for the next call.
1049 if (!rdp
->passed_quiesc
)
1053 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1056 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
1059 #ifdef CONFIG_HOTPLUG_CPU
1062 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1063 * Synchronization is not required because this function executes
1064 * in stop_machine() context.
1066 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1069 /* current DYING CPU is cleared in the cpu_online_mask */
1070 int receive_cpu
= cpumask_any(cpu_online_mask
);
1071 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1072 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1074 if (rdp
->nxtlist
== NULL
)
1075 return; /* irqs disabled, so comparison is stable. */
1077 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1078 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1079 receive_rdp
->qlen
+= rdp
->qlen
;
1080 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1081 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1083 rdp
->nxtlist
= NULL
;
1084 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1085 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1090 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1091 * and move all callbacks from the outgoing CPU to the current one.
1092 * There can only be one CPU hotplug operation at a time, so no other
1093 * CPU can be attempting to update rcu_cpu_kthread_task.
1095 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1097 unsigned long flags
;
1099 int need_report
= 0;
1100 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1101 struct rcu_node
*rnp
;
1102 struct task_struct
*t
;
1104 /* Stop the CPU's kthread. */
1105 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1107 per_cpu(rcu_cpu_kthread_task
, cpu
) = NULL
;
1111 /* Exclude any attempts to start a new grace period. */
1112 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1114 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1115 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1116 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1118 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1119 rnp
->qsmaskinit
&= ~mask
;
1120 if (rnp
->qsmaskinit
!= 0) {
1121 if (rnp
!= rdp
->mynode
)
1122 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1125 if (rnp
== rdp
->mynode
)
1126 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1128 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1129 mask
= rnp
->grpmask
;
1131 } while (rnp
!= NULL
);
1134 * We still hold the leaf rcu_node structure lock here, and
1135 * irqs are still disabled. The reason for this subterfuge is
1136 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1137 * held leads to deadlock.
1139 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1141 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1142 rcu_report_unblock_qs_rnp(rnp
, flags
);
1144 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1145 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1146 rcu_report_exp_rnp(rsp
, rnp
);
1147 rcu_node_kthread_setaffinity(rnp
, -1);
1151 * Remove the specified CPU from the RCU hierarchy and move any pending
1152 * callbacks that it might have to the current CPU. This code assumes
1153 * that at least one CPU in the system will remain running at all times.
1154 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1156 static void rcu_offline_cpu(int cpu
)
1158 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1159 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1160 rcu_preempt_offline_cpu(cpu
);
1163 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1165 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1169 static void rcu_offline_cpu(int cpu
)
1173 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1176 * Invoke any RCU callbacks that have made it to the end of their grace
1177 * period. Thottle as specified by rdp->blimit.
1179 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1181 unsigned long flags
;
1182 struct rcu_head
*next
, *list
, **tail
;
1185 /* If no callbacks are ready, just return.*/
1186 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1190 * Extract the list of ready callbacks, disabling to prevent
1191 * races with call_rcu() from interrupt handlers.
1193 local_irq_save(flags
);
1194 list
= rdp
->nxtlist
;
1195 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1196 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1197 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1198 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1199 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1200 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1201 local_irq_restore(flags
);
1203 /* Invoke callbacks. */
1208 debug_rcu_head_unqueue(list
);
1209 __rcu_reclaim(list
);
1211 if (++count
>= rdp
->blimit
)
1215 local_irq_save(flags
);
1217 /* Update count, and requeue any remaining callbacks. */
1219 rdp
->n_cbs_invoked
+= count
;
1221 *tail
= rdp
->nxtlist
;
1222 rdp
->nxtlist
= list
;
1223 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1224 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1225 rdp
->nxttail
[count
] = tail
;
1230 /* Reinstate batch limit if we have worked down the excess. */
1231 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1232 rdp
->blimit
= blimit
;
1234 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1235 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1236 rdp
->qlen_last_fqs_check
= 0;
1237 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1238 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1239 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1241 local_irq_restore(flags
);
1243 /* Re-raise the RCU softirq if there are callbacks remaining. */
1244 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1245 invoke_rcu_cpu_kthread();
1249 * Check to see if this CPU is in a non-context-switch quiescent state
1250 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1251 * Also schedule the RCU softirq handler.
1253 * This function must be called with hardirqs disabled. It is normally
1254 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1255 * false, there is no point in invoking rcu_check_callbacks().
1257 void rcu_check_callbacks(int cpu
, int user
)
1260 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1261 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1264 * Get here if this CPU took its interrupt from user
1265 * mode or from the idle loop, and if this is not a
1266 * nested interrupt. In this case, the CPU is in
1267 * a quiescent state, so note it.
1269 * No memory barrier is required here because both
1270 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1271 * variables that other CPUs neither access nor modify,
1272 * at least not while the corresponding CPU is online.
1278 } else if (!in_softirq()) {
1281 * Get here if this CPU did not take its interrupt from
1282 * softirq, in other words, if it is not interrupting
1283 * a rcu_bh read-side critical section. This is an _bh
1284 * critical section, so note it.
1289 rcu_preempt_check_callbacks(cpu
);
1290 if (rcu_pending(cpu
))
1291 invoke_rcu_cpu_kthread();
1297 * Scan the leaf rcu_node structures, processing dyntick state for any that
1298 * have not yet encountered a quiescent state, using the function specified.
1299 * Also initiate boosting for any threads blocked on the root rcu_node.
1301 * The caller must have suppressed start of new grace periods.
1303 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1307 unsigned long flags
;
1309 struct rcu_node
*rnp
;
1311 rcu_for_each_leaf_node(rsp
, rnp
) {
1313 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1314 if (!rcu_gp_in_progress(rsp
)) {
1315 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1318 if (rnp
->qsmask
== 0) {
1319 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1324 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1325 if ((rnp
->qsmask
& bit
) != 0 &&
1326 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1331 /* rcu_report_qs_rnp() releases rnp->lock. */
1332 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1335 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1337 rnp
= rcu_get_root(rsp
);
1338 if (rnp
->qsmask
== 0) {
1339 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1340 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1345 * Force quiescent states on reluctant CPUs, and also detect which
1346 * CPUs are in dyntick-idle mode.
1348 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1350 unsigned long flags
;
1351 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1353 if (!rcu_gp_in_progress(rsp
))
1354 return; /* No grace period in progress, nothing to force. */
1355 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1356 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1357 return; /* Someone else is already on the job. */
1359 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1360 goto unlock_fqs_ret
; /* no emergency and done recently. */
1362 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1363 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1364 if(!rcu_gp_in_progress(rsp
)) {
1365 rsp
->n_force_qs_ngp
++;
1366 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1367 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1369 rsp
->fqs_active
= 1;
1370 switch (rsp
->signaled
) {
1374 break; /* grace period idle or initializing, ignore. */
1376 case RCU_SAVE_DYNTICK
:
1377 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1378 break; /* So gcc recognizes the dead code. */
1380 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1382 /* Record dyntick-idle state. */
1383 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1384 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1385 if (rcu_gp_in_progress(rsp
))
1386 rsp
->signaled
= RCU_FORCE_QS
;
1391 /* Check dyntick-idle state, send IPI to laggarts. */
1392 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1393 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1395 /* Leave state in case more forcing is required. */
1397 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1400 rsp
->fqs_active
= 0;
1401 if (rsp
->fqs_need_gp
) {
1402 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1403 rsp
->fqs_need_gp
= 0;
1404 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1407 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1409 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1412 #else /* #ifdef CONFIG_SMP */
1414 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1419 #endif /* #else #ifdef CONFIG_SMP */
1422 * This does the RCU processing work from softirq context for the
1423 * specified rcu_state and rcu_data structures. This may be called
1424 * only from the CPU to whom the rdp belongs.
1427 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1429 unsigned long flags
;
1431 WARN_ON_ONCE(rdp
->beenonline
== 0);
1434 * If an RCU GP has gone long enough, go check for dyntick
1435 * idle CPUs and, if needed, send resched IPIs.
1437 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1438 force_quiescent_state(rsp
, 1);
1441 * Advance callbacks in response to end of earlier grace
1442 * period that some other CPU ended.
1444 rcu_process_gp_end(rsp
, rdp
);
1446 /* Update RCU state based on any recent quiescent states. */
1447 rcu_check_quiescent_state(rsp
, rdp
);
1449 /* Does this CPU require a not-yet-started grace period? */
1450 if (cpu_needs_another_gp(rsp
, rdp
)) {
1451 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1452 rcu_start_gp(rsp
, flags
); /* releases above lock */
1455 /* If there are callbacks ready, invoke them. */
1456 rcu_do_batch(rsp
, rdp
);
1460 * Do softirq processing for the current CPU.
1462 static void rcu_process_callbacks(void)
1464 __rcu_process_callbacks(&rcu_sched_state
,
1465 &__get_cpu_var(rcu_sched_data
));
1466 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1467 rcu_preempt_process_callbacks();
1469 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1470 rcu_needs_cpu_flush();
1474 * Wake up the current CPU's kthread. This replaces raise_softirq()
1475 * in earlier versions of RCU. Note that because we are running on
1476 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1477 * cannot disappear out from under us.
1479 static void invoke_rcu_cpu_kthread(void)
1481 unsigned long flags
;
1483 local_irq_save(flags
);
1484 __this_cpu_write(rcu_cpu_has_work
, 1);
1485 if (__this_cpu_read(rcu_cpu_kthread_task
) == NULL
) {
1486 local_irq_restore(flags
);
1489 wake_up(&__get_cpu_var(rcu_cpu_wq
));
1490 local_irq_restore(flags
);
1494 * Wake up the specified per-rcu_node-structure kthread.
1495 * Because the per-rcu_node kthreads are immortal, we don't need
1496 * to do anything to keep them alive.
1498 static void invoke_rcu_node_kthread(struct rcu_node
*rnp
)
1500 struct task_struct
*t
;
1502 t
= rnp
->node_kthread_task
;
1508 * Set the specified CPU's kthread to run RT or not, as specified by
1509 * the to_rt argument. The CPU-hotplug locks are held, so the task
1510 * is not going away.
1512 static void rcu_cpu_kthread_setrt(int cpu
, int to_rt
)
1515 struct sched_param sp
;
1516 struct task_struct
*t
;
1518 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1522 policy
= SCHED_FIFO
;
1523 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1525 policy
= SCHED_NORMAL
;
1526 sp
.sched_priority
= 0;
1528 sched_setscheduler_nocheck(t
, policy
, &sp
);
1532 * Timer handler to initiate the waking up of per-CPU kthreads that
1533 * have yielded the CPU due to excess numbers of RCU callbacks.
1534 * We wake up the per-rcu_node kthread, which in turn will wake up
1535 * the booster kthread.
1537 static void rcu_cpu_kthread_timer(unsigned long arg
)
1539 unsigned long flags
;
1540 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, arg
);
1541 struct rcu_node
*rnp
= rdp
->mynode
;
1543 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1544 rnp
->wakemask
|= rdp
->grpmask
;
1545 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1546 invoke_rcu_node_kthread(rnp
);
1550 * Drop to non-real-time priority and yield, but only after posting a
1551 * timer that will cause us to regain our real-time priority if we
1552 * remain preempted. Either way, we restore our real-time priority
1555 static void rcu_yield(void (*f
)(unsigned long), unsigned long arg
)
1557 struct sched_param sp
;
1558 struct timer_list yield_timer
;
1560 setup_timer_on_stack(&yield_timer
, f
, arg
);
1561 mod_timer(&yield_timer
, jiffies
+ 2);
1562 sp
.sched_priority
= 0;
1563 sched_setscheduler_nocheck(current
, SCHED_NORMAL
, &sp
);
1564 set_user_nice(current
, 19);
1566 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1567 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
1568 del_timer(&yield_timer
);
1572 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1573 * This can happen while the corresponding CPU is either coming online
1574 * or going offline. We cannot wait until the CPU is fully online
1575 * before starting the kthread, because the various notifier functions
1576 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1577 * the corresponding CPU is online.
1579 * Return 1 if the kthread needs to stop, 0 otherwise.
1581 * Caller must disable bh. This function can momentarily enable it.
1583 static int rcu_cpu_kthread_should_stop(int cpu
)
1585 while (cpu_is_offline(cpu
) ||
1586 !cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)) ||
1587 smp_processor_id() != cpu
) {
1588 if (kthread_should_stop())
1590 per_cpu(rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
1591 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = raw_smp_processor_id();
1593 schedule_timeout_uninterruptible(1);
1594 if (!cpumask_equal(¤t
->cpus_allowed
, cpumask_of(cpu
)))
1595 set_cpus_allowed_ptr(current
, cpumask_of(cpu
));
1598 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1603 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1604 * earlier RCU softirq.
1606 static int rcu_cpu_kthread(void *arg
)
1608 int cpu
= (int)(long)arg
;
1609 unsigned long flags
;
1611 unsigned int *statusp
= &per_cpu(rcu_cpu_kthread_status
, cpu
);
1612 wait_queue_head_t
*wqp
= &per_cpu(rcu_cpu_wq
, cpu
);
1614 char *workp
= &per_cpu(rcu_cpu_has_work
, cpu
);
1617 *statusp
= RCU_KTHREAD_WAITING
;
1618 wait_event_interruptible(*wqp
,
1619 *workp
!= 0 || kthread_should_stop());
1621 if (rcu_cpu_kthread_should_stop(cpu
)) {
1625 *statusp
= RCU_KTHREAD_RUNNING
;
1626 per_cpu(rcu_cpu_kthread_loops
, cpu
)++;
1627 local_irq_save(flags
);
1630 local_irq_restore(flags
);
1632 rcu_process_callbacks();
1639 *statusp
= RCU_KTHREAD_YIELDING
;
1640 rcu_yield(rcu_cpu_kthread_timer
, (unsigned long)cpu
);
1644 *statusp
= RCU_KTHREAD_STOPPED
;
1649 * Spawn a per-CPU kthread, setting up affinity and priority.
1650 * Because the CPU hotplug lock is held, no other CPU will be attempting
1651 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1652 * attempting to access it during boot, but the locking in kthread_bind()
1653 * will enforce sufficient ordering.
1655 static int __cpuinit
rcu_spawn_one_cpu_kthread(int cpu
)
1657 struct sched_param sp
;
1658 struct task_struct
*t
;
1660 if (!rcu_kthreads_spawnable
||
1661 per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
)
1663 t
= kthread_create(rcu_cpu_kthread
, (void *)(long)cpu
, "rcuc%d", cpu
);
1666 kthread_bind(t
, cpu
);
1667 per_cpu(rcu_cpu_kthread_cpu
, cpu
) = cpu
;
1668 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task
, cpu
) != NULL
);
1669 per_cpu(rcu_cpu_kthread_task
, cpu
) = t
;
1671 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1672 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1677 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1678 * kthreads when needed. We ignore requests to wake up kthreads
1679 * for offline CPUs, which is OK because force_quiescent_state()
1680 * takes care of this case.
1682 static int rcu_node_kthread(void *arg
)
1685 unsigned long flags
;
1687 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1688 struct sched_param sp
;
1689 struct task_struct
*t
;
1692 rnp
->node_kthread_status
= RCU_KTHREAD_WAITING
;
1693 wait_event_interruptible(rnp
->node_wq
, rnp
->wakemask
!= 0);
1694 rnp
->node_kthread_status
= RCU_KTHREAD_RUNNING
;
1695 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1696 mask
= rnp
->wakemask
;
1698 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1699 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1) {
1700 if ((mask
& 0x1) == 0)
1703 t
= per_cpu(rcu_cpu_kthread_task
, cpu
);
1704 if (!cpu_online(cpu
) || t
== NULL
) {
1708 per_cpu(rcu_cpu_has_work
, cpu
) = 1;
1709 sp
.sched_priority
= RCU_KTHREAD_PRIO
;
1710 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1715 rnp
->node_kthread_status
= RCU_KTHREAD_STOPPED
;
1720 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1721 * served by the rcu_node in question. The CPU hotplug lock is still
1722 * held, so the value of rnp->qsmaskinit will be stable.
1724 * We don't include outgoingcpu in the affinity set, use -1 if there is
1725 * no outgoing CPU. If there are no CPUs left in the affinity set,
1726 * this function allows the kthread to execute on any CPU.
1728 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1732 unsigned long mask
= rnp
->qsmaskinit
;
1734 if (rnp
->node_kthread_task
== NULL
)
1736 if (!alloc_cpumask_var(&cm
, GFP_KERNEL
))
1739 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
>>= 1)
1740 if ((mask
& 0x1) && cpu
!= outgoingcpu
)
1741 cpumask_set_cpu(cpu
, cm
);
1742 if (cpumask_weight(cm
) == 0) {
1744 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++)
1745 cpumask_clear_cpu(cpu
, cm
);
1746 WARN_ON_ONCE(cpumask_weight(cm
) == 0);
1748 set_cpus_allowed_ptr(rnp
->node_kthread_task
, cm
);
1749 rcu_boost_kthread_setaffinity(rnp
, cm
);
1750 free_cpumask_var(cm
);
1754 * Spawn a per-rcu_node kthread, setting priority and affinity.
1755 * Called during boot before online/offline can happen, or, if
1756 * during runtime, with the main CPU-hotplug locks held. So only
1757 * one of these can be executing at a time.
1759 static int __cpuinit
rcu_spawn_one_node_kthread(struct rcu_state
*rsp
,
1760 struct rcu_node
*rnp
)
1762 unsigned long flags
;
1763 int rnp_index
= rnp
- &rsp
->node
[0];
1764 struct sched_param sp
;
1765 struct task_struct
*t
;
1767 if (!rcu_kthreads_spawnable
||
1768 rnp
->qsmaskinit
== 0)
1770 if (rnp
->node_kthread_task
== NULL
) {
1771 t
= kthread_create(rcu_node_kthread
, (void *)rnp
,
1772 "rcun%d", rnp_index
);
1775 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1776 rnp
->node_kthread_task
= t
;
1777 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1779 sp
.sched_priority
= 99;
1780 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1782 return rcu_spawn_one_boost_kthread(rsp
, rnp
, rnp_index
);
1786 * Spawn all kthreads -- called as soon as the scheduler is running.
1788 static int __init
rcu_spawn_kthreads(void)
1791 struct rcu_node
*rnp
;
1793 rcu_kthreads_spawnable
= 1;
1794 for_each_possible_cpu(cpu
) {
1795 init_waitqueue_head(&per_cpu(rcu_cpu_wq
, cpu
));
1796 per_cpu(rcu_cpu_has_work
, cpu
) = 0;
1797 if (cpu_online(cpu
))
1798 (void)rcu_spawn_one_cpu_kthread(cpu
);
1800 rnp
= rcu_get_root(rcu_state
);
1801 init_waitqueue_head(&rnp
->node_wq
);
1802 rcu_init_boost_waitqueue(rnp
);
1803 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1804 if (NUM_RCU_NODES
> 1)
1805 rcu_for_each_leaf_node(rcu_state
, rnp
) {
1806 init_waitqueue_head(&rnp
->node_wq
);
1807 rcu_init_boost_waitqueue(rnp
);
1808 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
1812 early_initcall(rcu_spawn_kthreads
);
1815 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1816 struct rcu_state
*rsp
)
1818 unsigned long flags
;
1819 struct rcu_data
*rdp
;
1821 debug_rcu_head_queue(head
);
1825 smp_mb(); /* Ensure RCU update seen before callback registry. */
1828 * Opportunistically note grace-period endings and beginnings.
1829 * Note that we might see a beginning right after we see an
1830 * end, but never vice versa, since this CPU has to pass through
1831 * a quiescent state betweentimes.
1833 local_irq_save(flags
);
1834 rdp
= this_cpu_ptr(rsp
->rda
);
1836 /* Add the callback to our list. */
1837 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1838 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1841 /* If interrupts were disabled, don't dive into RCU core. */
1842 if (irqs_disabled_flags(flags
)) {
1843 local_irq_restore(flags
);
1848 * Force the grace period if too many callbacks or too long waiting.
1849 * Enforce hysteresis, and don't invoke force_quiescent_state()
1850 * if some other CPU has recently done so. Also, don't bother
1851 * invoking force_quiescent_state() if the newly enqueued callback
1852 * is the only one waiting for a grace period to complete.
1854 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1856 /* Are we ignoring a completed grace period? */
1857 rcu_process_gp_end(rsp
, rdp
);
1858 check_for_new_grace_period(rsp
, rdp
);
1860 /* Start a new grace period if one not already started. */
1861 if (!rcu_gp_in_progress(rsp
)) {
1862 unsigned long nestflag
;
1863 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1865 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1866 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1868 /* Give the grace period a kick. */
1869 rdp
->blimit
= LONG_MAX
;
1870 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1871 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1872 force_quiescent_state(rsp
, 0);
1873 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1874 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1876 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1877 force_quiescent_state(rsp
, 1);
1878 local_irq_restore(flags
);
1882 * Queue an RCU-sched callback for invocation after a grace period.
1884 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1886 __call_rcu(head
, func
, &rcu_sched_state
);
1888 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1891 * Queue an RCU for invocation after a quicker grace period.
1893 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1895 __call_rcu(head
, func
, &rcu_bh_state
);
1897 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1900 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1902 * Control will return to the caller some time after a full rcu-sched
1903 * grace period has elapsed, in other words after all currently executing
1904 * rcu-sched read-side critical sections have completed. These read-side
1905 * critical sections are delimited by rcu_read_lock_sched() and
1906 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1907 * local_irq_disable(), and so on may be used in place of
1908 * rcu_read_lock_sched().
1910 * This means that all preempt_disable code sequences, including NMI and
1911 * hardware-interrupt handlers, in progress on entry will have completed
1912 * before this primitive returns. However, this does not guarantee that
1913 * softirq handlers will have completed, since in some kernels, these
1914 * handlers can run in process context, and can block.
1916 * This primitive provides the guarantees made by the (now removed)
1917 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1918 * guarantees that rcu_read_lock() sections will have completed.
1919 * In "classic RCU", these two guarantees happen to be one and
1920 * the same, but can differ in realtime RCU implementations.
1922 void synchronize_sched(void)
1924 struct rcu_synchronize rcu
;
1926 if (rcu_blocking_is_gp())
1929 init_rcu_head_on_stack(&rcu
.head
);
1930 init_completion(&rcu
.completion
);
1931 /* Will wake me after RCU finished. */
1932 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1934 wait_for_completion(&rcu
.completion
);
1935 destroy_rcu_head_on_stack(&rcu
.head
);
1937 EXPORT_SYMBOL_GPL(synchronize_sched
);
1940 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1942 * Control will return to the caller some time after a full rcu_bh grace
1943 * period has elapsed, in other words after all currently executing rcu_bh
1944 * read-side critical sections have completed. RCU read-side critical
1945 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1946 * and may be nested.
1948 void synchronize_rcu_bh(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_bh(&rcu
.head
, wakeme_after_rcu
);
1960 wait_for_completion(&rcu
.completion
);
1961 destroy_rcu_head_on_stack(&rcu
.head
);
1963 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1966 * Check to see if there is any immediate RCU-related work to be done
1967 * by the current CPU, for the specified type of RCU, returning 1 if so.
1968 * The checks are in order of increasing expense: checks that can be
1969 * carried out against CPU-local state are performed first. However,
1970 * we must check for CPU stalls first, else we might not get a chance.
1972 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1974 struct rcu_node
*rnp
= rdp
->mynode
;
1976 rdp
->n_rcu_pending
++;
1978 /* Check for CPU stalls, if enabled. */
1979 check_cpu_stall(rsp
, rdp
);
1981 /* Is the RCU core waiting for a quiescent state from this CPU? */
1982 if (rdp
->qs_pending
&& !rdp
->passed_quiesc
) {
1985 * If force_quiescent_state() coming soon and this CPU
1986 * needs a quiescent state, and this is either RCU-sched
1987 * or RCU-bh, force a local reschedule.
1989 rdp
->n_rp_qs_pending
++;
1990 if (!rdp
->preemptible
&&
1991 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
1994 } else if (rdp
->qs_pending
&& rdp
->passed_quiesc
) {
1995 rdp
->n_rp_report_qs
++;
1999 /* Does this CPU have callbacks ready to invoke? */
2000 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2001 rdp
->n_rp_cb_ready
++;
2005 /* Has RCU gone idle with this CPU needing another grace period? */
2006 if (cpu_needs_another_gp(rsp
, rdp
)) {
2007 rdp
->n_rp_cpu_needs_gp
++;
2011 /* Has another RCU grace period completed? */
2012 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2013 rdp
->n_rp_gp_completed
++;
2017 /* Has a new RCU grace period started? */
2018 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2019 rdp
->n_rp_gp_started
++;
2023 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2024 if (rcu_gp_in_progress(rsp
) &&
2025 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
2026 rdp
->n_rp_need_fqs
++;
2031 rdp
->n_rp_need_nothing
++;
2036 * Check to see if there is any immediate RCU-related work to be done
2037 * by the current CPU, returning 1 if so. This function is part of the
2038 * RCU implementation; it is -not- an exported member of the RCU API.
2040 static int rcu_pending(int cpu
)
2042 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
2043 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
2044 rcu_preempt_pending(cpu
);
2048 * Check to see if any future RCU-related work will need to be done
2049 * by the current CPU, even if none need be done immediately, returning
2052 static int rcu_needs_cpu_quick_check(int cpu
)
2054 /* RCU callbacks either ready or pending? */
2055 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
2056 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
2057 rcu_preempt_needs_cpu(cpu
);
2060 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
2061 static atomic_t rcu_barrier_cpu_count
;
2062 static DEFINE_MUTEX(rcu_barrier_mutex
);
2063 static struct completion rcu_barrier_completion
;
2065 static void rcu_barrier_callback(struct rcu_head
*notused
)
2067 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2068 complete(&rcu_barrier_completion
);
2072 * Called with preemption disabled, and from cross-cpu IRQ context.
2074 static void rcu_barrier_func(void *type
)
2076 int cpu
= smp_processor_id();
2077 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
2078 void (*call_rcu_func
)(struct rcu_head
*head
,
2079 void (*func
)(struct rcu_head
*head
));
2081 atomic_inc(&rcu_barrier_cpu_count
);
2082 call_rcu_func
= type
;
2083 call_rcu_func(head
, rcu_barrier_callback
);
2087 * Orchestrate the specified type of RCU barrier, waiting for all
2088 * RCU callbacks of the specified type to complete.
2090 static void _rcu_barrier(struct rcu_state
*rsp
,
2091 void (*call_rcu_func
)(struct rcu_head
*head
,
2092 void (*func
)(struct rcu_head
*head
)))
2094 BUG_ON(in_interrupt());
2095 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2096 mutex_lock(&rcu_barrier_mutex
);
2097 init_completion(&rcu_barrier_completion
);
2099 * Initialize rcu_barrier_cpu_count to 1, then invoke
2100 * rcu_barrier_func() on each CPU, so that each CPU also has
2101 * incremented rcu_barrier_cpu_count. Only then is it safe to
2102 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2103 * might complete its grace period before all of the other CPUs
2104 * did their increment, causing this function to return too
2105 * early. Note that on_each_cpu() disables irqs, which prevents
2106 * any CPUs from coming online or going offline until each online
2107 * CPU has queued its RCU-barrier callback.
2109 atomic_set(&rcu_barrier_cpu_count
, 1);
2110 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
2111 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2112 complete(&rcu_barrier_completion
);
2113 wait_for_completion(&rcu_barrier_completion
);
2114 mutex_unlock(&rcu_barrier_mutex
);
2118 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2120 void rcu_barrier_bh(void)
2122 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
2124 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2127 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2129 void rcu_barrier_sched(void)
2131 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2133 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2136 * Do boot-time initialization of a CPU's per-CPU RCU data.
2139 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2141 unsigned long flags
;
2143 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2144 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2146 /* Set up local state, ensuring consistent view of global state. */
2147 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2148 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2149 rdp
->nxtlist
= NULL
;
2150 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2151 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2154 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2155 #endif /* #ifdef CONFIG_NO_HZ */
2157 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2161 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2162 * offline event can be happening at a given time. Note also that we
2163 * can accept some slop in the rsp->completed access due to the fact
2164 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2166 static void __cpuinit
2167 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2169 unsigned long flags
;
2171 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2172 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2174 /* Set up local state, ensuring consistent view of global state. */
2175 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2176 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
2177 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
2178 rdp
->beenonline
= 1; /* We have now been online. */
2179 rdp
->preemptible
= preemptible
;
2180 rdp
->qlen_last_fqs_check
= 0;
2181 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2182 rdp
->blimit
= blimit
;
2183 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2186 * A new grace period might start here. If so, we won't be part
2187 * of it, but that is OK, as we are currently in a quiescent state.
2190 /* Exclude any attempts to start a new GP on large systems. */
2191 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2193 /* Add CPU to rcu_node bitmasks. */
2195 mask
= rdp
->grpmask
;
2197 /* Exclude any attempts to start a new GP on small systems. */
2198 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2199 rnp
->qsmaskinit
|= mask
;
2200 mask
= rnp
->grpmask
;
2201 if (rnp
== rdp
->mynode
) {
2202 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
2203 rdp
->completed
= rnp
->completed
;
2204 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
2206 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2208 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2210 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2213 static void __cpuinit
rcu_online_cpu(int cpu
)
2215 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2216 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2217 rcu_preempt_init_percpu_data(cpu
);
2220 static void __cpuinit
rcu_online_kthreads(int cpu
)
2222 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2223 struct rcu_node
*rnp
= rdp
->mynode
;
2225 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
2226 if (rcu_kthreads_spawnable
) {
2227 (void)rcu_spawn_one_cpu_kthread(cpu
);
2228 if (rnp
->node_kthread_task
== NULL
)
2229 (void)rcu_spawn_one_node_kthread(rcu_state
, rnp
);
2234 * Handle CPU online/offline notification events.
2236 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2237 unsigned long action
, void *hcpu
)
2239 long cpu
= (long)hcpu
;
2240 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2241 struct rcu_node
*rnp
= rdp
->mynode
;
2244 case CPU_UP_PREPARE
:
2245 case CPU_UP_PREPARE_FROZEN
:
2246 rcu_online_cpu(cpu
);
2247 rcu_online_kthreads(cpu
);
2250 case CPU_DOWN_FAILED
:
2251 rcu_node_kthread_setaffinity(rnp
, -1);
2252 rcu_cpu_kthread_setrt(cpu
, 1);
2254 case CPU_DOWN_PREPARE
:
2255 rcu_node_kthread_setaffinity(rnp
, cpu
);
2256 rcu_cpu_kthread_setrt(cpu
, 0);
2259 case CPU_DYING_FROZEN
:
2261 * The whole machine is "stopped" except this CPU, so we can
2262 * touch any data without introducing corruption. We send the
2263 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2265 rcu_send_cbs_to_online(&rcu_bh_state
);
2266 rcu_send_cbs_to_online(&rcu_sched_state
);
2267 rcu_preempt_send_cbs_to_online();
2270 case CPU_DEAD_FROZEN
:
2271 case CPU_UP_CANCELED
:
2272 case CPU_UP_CANCELED_FROZEN
:
2273 rcu_offline_cpu(cpu
);
2282 * This function is invoked towards the end of the scheduler's initialization
2283 * process. Before this is called, the idle task might contain
2284 * RCU read-side critical sections (during which time, this idle
2285 * task is booting the system). After this function is called, the
2286 * idle tasks are prohibited from containing RCU read-side critical
2287 * sections. This function also enables RCU lockdep checking.
2289 void rcu_scheduler_starting(void)
2291 WARN_ON(num_online_cpus() != 1);
2292 WARN_ON(nr_context_switches() > 0);
2293 rcu_scheduler_active
= 1;
2297 * Compute the per-level fanout, either using the exact fanout specified
2298 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2300 #ifdef CONFIG_RCU_FANOUT_EXACT
2301 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2305 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2306 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2307 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
2309 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2310 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2317 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2318 ccur
= rsp
->levelcnt
[i
];
2319 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2323 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2326 * Helper function for rcu_init() that initializes one rcu_state structure.
2328 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2329 struct rcu_data __percpu
*rda
)
2331 static char *buf
[] = { "rcu_node_level_0",
2334 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2338 struct rcu_node
*rnp
;
2340 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2342 /* Initialize the level-tracking arrays. */
2344 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2345 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2346 rcu_init_levelspread(rsp
);
2348 /* Initialize the elements themselves, starting from the leaves. */
2350 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2351 cpustride
*= rsp
->levelspread
[i
];
2352 rnp
= rsp
->level
[i
];
2353 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2354 raw_spin_lock_init(&rnp
->lock
);
2355 lockdep_set_class_and_name(&rnp
->lock
,
2356 &rcu_node_class
[i
], buf
[i
]);
2359 rnp
->qsmaskinit
= 0;
2360 rnp
->grplo
= j
* cpustride
;
2361 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2362 if (rnp
->grphi
>= NR_CPUS
)
2363 rnp
->grphi
= NR_CPUS
- 1;
2369 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2370 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2371 rnp
->parent
= rsp
->level
[i
- 1] +
2372 j
/ rsp
->levelspread
[i
- 1];
2375 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2380 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2381 for_each_possible_cpu(i
) {
2382 while (i
> rnp
->grphi
)
2384 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2385 rcu_boot_init_percpu_data(i
, rsp
);
2389 void __init
rcu_init(void)
2393 rcu_bootup_announce();
2394 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2395 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2396 __rcu_init_preempt();
2399 * We don't need protection against CPU-hotplug here because
2400 * this is called early in boot, before either interrupts
2401 * or the scheduler are operational.
2403 cpu_notifier(rcu_cpu_notify
, 0);
2404 for_each_online_cpu(cpu
)
2405 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2406 check_cpu_stall_init();
2409 #include "rcutree_plugin.h"