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/export.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>
55 #include <trace/events/rcu.h>
59 /* Data structures. */
61 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
63 #define RCU_STATE_INITIALIZER(structname) { \
64 .level = { &structname##_state.node[0] }, \
66 NUM_RCU_LVL_0, /* root of hierarchy. */ \
70 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
72 .fqs_state = RCU_GP_IDLE, \
75 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
76 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
78 .n_force_qs_ngp = 0, \
79 .name = #structname, \
82 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched
);
83 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
85 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
);
86 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
88 static struct rcu_state
*rcu_state
;
91 * The rcu_scheduler_active variable transitions from zero to one just
92 * before the first task is spawned. So when this variable is zero, RCU
93 * can assume that there is but one task, allowing RCU to (for example)
94 * optimized synchronize_sched() to a simple barrier(). When this variable
95 * is one, RCU must actually do all the hard work required to detect real
96 * grace periods. This variable is also used to suppress boot-time false
97 * positives from lockdep-RCU error checking.
99 int rcu_scheduler_active __read_mostly
;
100 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
103 * The rcu_scheduler_fully_active variable transitions from zero to one
104 * during the early_initcall() processing, which is after the scheduler
105 * is capable of creating new tasks. So RCU processing (for example,
106 * creating tasks for RCU priority boosting) must be delayed until after
107 * rcu_scheduler_fully_active transitions from zero to one. We also
108 * currently delay invocation of any RCU callbacks until after this point.
110 * It might later prove better for people registering RCU callbacks during
111 * early boot to take responsibility for these callbacks, but one step at
114 static int rcu_scheduler_fully_active __read_mostly
;
116 #ifdef CONFIG_RCU_BOOST
119 * Control variables for per-CPU and per-rcu_node kthreads. These
120 * handle all flavors of RCU.
122 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
123 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
124 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
125 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
126 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
128 #endif /* #ifdef CONFIG_RCU_BOOST */
130 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
131 static void invoke_rcu_core(void);
132 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
135 * Track the rcutorture test sequence number and the update version
136 * number within a given test. The rcutorture_testseq is incremented
137 * on every rcutorture module load and unload, so has an odd value
138 * when a test is running. The rcutorture_vernum is set to zero
139 * when rcutorture starts and is incremented on each rcutorture update.
140 * These variables enable correlating rcutorture output with the
141 * RCU tracing information.
143 unsigned long rcutorture_testseq
;
144 unsigned long rcutorture_vernum
;
147 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
148 * permit this function to be invoked without holding the root rcu_node
149 * structure's ->lock, but of course results can be subject to change.
151 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
153 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
157 * Note a quiescent state. Because we do not need to know
158 * how many quiescent states passed, just if there was at least
159 * one since the start of the grace period, this just sets a flag.
160 * The caller must have disabled preemption.
162 void rcu_sched_qs(int cpu
)
164 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
166 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
168 if (rdp
->passed_quiesce
== 0)
169 trace_rcu_grace_period("rcu_sched", rdp
->gpnum
, "cpuqs");
170 rdp
->passed_quiesce
= 1;
173 void rcu_bh_qs(int cpu
)
175 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
177 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
179 if (rdp
->passed_quiesce
== 0)
180 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
181 rdp
->passed_quiesce
= 1;
185 * Note a context switch. This is a quiescent state for RCU-sched,
186 * and requires special handling for preemptible RCU.
187 * The caller must have disabled preemption.
189 void rcu_note_context_switch(int cpu
)
191 trace_rcu_utilization("Start context switch");
193 rcu_preempt_note_context_switch(cpu
);
194 trace_rcu_utilization("End context switch");
196 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
198 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
199 .dynticks_nesting
= DYNTICK_TASK_NESTING
,
200 .dynticks
= ATOMIC_INIT(1),
203 static int blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
204 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
205 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
207 module_param(blimit
, int, 0);
208 module_param(qhimark
, int, 0);
209 module_param(qlowmark
, int, 0);
211 int rcu_cpu_stall_suppress __read_mostly
;
212 module_param(rcu_cpu_stall_suppress
, int, 0644);
214 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
215 static int rcu_pending(int cpu
);
218 * Return the number of RCU-sched batches processed thus far for debug & stats.
220 long rcu_batches_completed_sched(void)
222 return rcu_sched_state
.completed
;
224 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
227 * Return the number of RCU BH batches processed thus far for debug & stats.
229 long rcu_batches_completed_bh(void)
231 return rcu_bh_state
.completed
;
233 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
236 * Force a quiescent state for RCU BH.
238 void rcu_bh_force_quiescent_state(void)
240 force_quiescent_state(&rcu_bh_state
, 0);
242 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
245 * Record the number of times rcutorture tests have been initiated and
246 * terminated. This information allows the debugfs tracing stats to be
247 * correlated to the rcutorture messages, even when the rcutorture module
248 * is being repeatedly loaded and unloaded. In other words, we cannot
249 * store this state in rcutorture itself.
251 void rcutorture_record_test_transition(void)
253 rcutorture_testseq
++;
254 rcutorture_vernum
= 0;
256 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
259 * Record the number of writer passes through the current rcutorture test.
260 * This is also used to correlate debugfs tracing stats with the rcutorture
263 void rcutorture_record_progress(unsigned long vernum
)
267 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
270 * Force a quiescent state for RCU-sched.
272 void rcu_sched_force_quiescent_state(void)
274 force_quiescent_state(&rcu_sched_state
, 0);
276 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
279 * Does the CPU have callbacks ready to be invoked?
282 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
284 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
288 * Does the current CPU require a yet-as-unscheduled grace period?
291 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
293 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
297 * Return the root node of the specified rcu_state structure.
299 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
301 return &rsp
->node
[0];
307 * If the specified CPU is offline, tell the caller that it is in
308 * a quiescent state. Otherwise, whack it with a reschedule IPI.
309 * Grace periods can end up waiting on an offline CPU when that
310 * CPU is in the process of coming online -- it will be added to the
311 * rcu_node bitmasks before it actually makes it online. The same thing
312 * can happen while a CPU is in the process of coming online. Because this
313 * race is quite rare, we check for it after detecting that the grace
314 * period has been delayed rather than checking each and every CPU
315 * each and every time we start a new grace period.
317 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
320 * If the CPU is offline, it is in a quiescent state. We can
321 * trust its state not to change because interrupts are disabled.
323 if (cpu_is_offline(rdp
->cpu
)) {
324 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
330 * The CPU is online, so send it a reschedule IPI. This forces
331 * it through the scheduler, and (inefficiently) also handles cases
332 * where idle loops fail to inform RCU about the CPU being idle.
334 if (rdp
->cpu
!= smp_processor_id())
335 smp_send_reschedule(rdp
->cpu
);
342 #endif /* #ifdef CONFIG_SMP */
345 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
347 * If the new value of the ->dynticks_nesting counter now is zero,
348 * we really have entered idle, and must do the appropriate accounting.
349 * The caller must have disabled interrupts.
351 static void rcu_idle_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
)
353 if (rdtp
->dynticks_nesting
) {
354 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
357 trace_rcu_dyntick("Start", oldval
, rdtp
->dynticks_nesting
);
358 if (!idle_cpu(smp_processor_id())) {
359 struct task_struct
*idle
= idle_task(smp_processor_id());
361 trace_rcu_dyntick("Error on entry: not idle task",
362 oldval
, rdtp
->dynticks_nesting
);
363 ftrace_dump(DUMP_ALL
);
364 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
365 current
->pid
, current
->comm
,
366 idle
->pid
, idle
->comm
); /* must be idle task! */
368 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
369 smp_mb__before_atomic_inc(); /* See above. */
370 atomic_inc(&rdtp
->dynticks
);
371 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
372 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
376 * rcu_idle_enter - inform RCU that current CPU is entering idle
378 * Enter idle mode, in other words, -leave- the mode in which RCU
379 * read-side critical sections can occur. (Though RCU read-side
380 * critical sections can occur in irq handlers in idle, a possibility
381 * handled by irq_enter() and irq_exit().)
383 * We crowbar the ->dynticks_nesting field to zero to allow for
384 * the possibility of usermode upcalls having messed up our count
385 * of interrupt nesting level during the prior busy period.
387 void rcu_idle_enter(void)
391 struct rcu_dynticks
*rdtp
;
393 local_irq_save(flags
);
394 rdtp
= &__get_cpu_var(rcu_dynticks
);
395 oldval
= rdtp
->dynticks_nesting
;
396 rdtp
->dynticks_nesting
= 0;
397 rcu_idle_enter_common(rdtp
, oldval
);
398 local_irq_restore(flags
);
402 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
404 * Exit from an interrupt handler, which might possibly result in entering
405 * idle mode, in other words, leaving the mode in which read-side critical
406 * sections can occur.
408 * This code assumes that the idle loop never does anything that might
409 * result in unbalanced calls to irq_enter() and irq_exit(). If your
410 * architecture violates this assumption, RCU will give you what you
411 * deserve, good and hard. But very infrequently and irreproducibly.
413 * Use things like work queues to work around this limitation.
415 * You have been warned.
417 void rcu_irq_exit(void)
421 struct rcu_dynticks
*rdtp
;
423 local_irq_save(flags
);
424 rdtp
= &__get_cpu_var(rcu_dynticks
);
425 oldval
= rdtp
->dynticks_nesting
;
426 rdtp
->dynticks_nesting
--;
427 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
428 rcu_idle_enter_common(rdtp
, oldval
);
429 local_irq_restore(flags
);
433 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
435 * If the new value of the ->dynticks_nesting counter was previously zero,
436 * we really have exited idle, and must do the appropriate accounting.
437 * The caller must have disabled interrupts.
439 static void rcu_idle_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
)
442 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
445 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
446 atomic_inc(&rdtp
->dynticks
);
447 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
448 smp_mb__after_atomic_inc(); /* See above. */
449 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
450 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
451 if (!idle_cpu(smp_processor_id())) {
452 struct task_struct
*idle
= idle_task(smp_processor_id());
454 trace_rcu_dyntick("Error on exit: not idle task",
455 oldval
, rdtp
->dynticks_nesting
);
456 ftrace_dump(DUMP_ALL
);
457 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
458 current
->pid
, current
->comm
,
459 idle
->pid
, idle
->comm
); /* must be idle task! */
464 * rcu_idle_exit - inform RCU that current CPU is leaving idle
466 * Exit idle mode, in other words, -enter- the mode in which RCU
467 * read-side critical sections can occur.
469 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NESTING to
470 * allow for the possibility of usermode upcalls messing up our count
471 * of interrupt nesting level during the busy period that is just
474 void rcu_idle_exit(void)
477 struct rcu_dynticks
*rdtp
;
480 local_irq_save(flags
);
481 rdtp
= &__get_cpu_var(rcu_dynticks
);
482 oldval
= rdtp
->dynticks_nesting
;
483 WARN_ON_ONCE(oldval
!= 0);
484 rdtp
->dynticks_nesting
= DYNTICK_TASK_NESTING
;
485 rcu_idle_exit_common(rdtp
, oldval
);
486 local_irq_restore(flags
);
490 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
492 * Enter an interrupt handler, which might possibly result in exiting
493 * idle mode, in other words, entering the mode in which read-side critical
494 * sections can occur.
496 * Note that the Linux kernel is fully capable of entering an interrupt
497 * handler that it never exits, for example when doing upcalls to
498 * user mode! This code assumes that the idle loop never does upcalls to
499 * user mode. If your architecture does do upcalls from the idle loop (or
500 * does anything else that results in unbalanced calls to the irq_enter()
501 * and irq_exit() functions), RCU will give you what you deserve, good
502 * and hard. But very infrequently and irreproducibly.
504 * Use things like work queues to work around this limitation.
506 * You have been warned.
508 void rcu_irq_enter(void)
511 struct rcu_dynticks
*rdtp
;
514 local_irq_save(flags
);
515 rdtp
= &__get_cpu_var(rcu_dynticks
);
516 oldval
= rdtp
->dynticks_nesting
;
517 rdtp
->dynticks_nesting
++;
518 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
519 rcu_idle_exit_common(rdtp
, oldval
);
520 local_irq_restore(flags
);
524 * rcu_nmi_enter - inform RCU of entry to NMI context
526 * If the CPU was idle with dynamic ticks active, and there is no
527 * irq handler running, this updates rdtp->dynticks_nmi to let the
528 * RCU grace-period handling know that the CPU is active.
530 void rcu_nmi_enter(void)
532 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
534 if (rdtp
->dynticks_nmi_nesting
== 0 &&
535 (atomic_read(&rdtp
->dynticks
) & 0x1))
537 rdtp
->dynticks_nmi_nesting
++;
538 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
539 atomic_inc(&rdtp
->dynticks
);
540 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
541 smp_mb__after_atomic_inc(); /* See above. */
542 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
546 * rcu_nmi_exit - inform RCU of exit from NMI context
548 * If the CPU was idle with dynamic ticks active, and there is no
549 * irq handler running, this updates rdtp->dynticks_nmi to let the
550 * RCU grace-period handling know that the CPU is no longer active.
552 void rcu_nmi_exit(void)
554 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
556 if (rdtp
->dynticks_nmi_nesting
== 0 ||
557 --rdtp
->dynticks_nmi_nesting
!= 0)
559 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
560 smp_mb__before_atomic_inc(); /* See above. */
561 atomic_inc(&rdtp
->dynticks
);
562 smp_mb__after_atomic_inc(); /* Force delay to next write. */
563 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
566 #ifdef CONFIG_PROVE_RCU
569 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
571 * If the current CPU is in its idle loop and is neither in an interrupt
572 * or NMI handler, return true.
574 int rcu_is_cpu_idle(void)
579 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
583 EXPORT_SYMBOL(rcu_is_cpu_idle
);
585 #endif /* #ifdef CONFIG_PROVE_RCU */
588 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
590 * If the current CPU is idle or running at a first-level (not nested)
591 * interrupt from idle, return true. The caller must have at least
592 * disabled preemption.
594 int rcu_is_cpu_rrupt_from_idle(void)
596 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
602 * Snapshot the specified CPU's dynticks counter so that we can later
603 * credit them with an implicit quiescent state. Return 1 if this CPU
604 * is in dynticks idle mode, which is an extended quiescent state.
606 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
608 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
613 * Return true if the specified CPU has passed through a quiescent
614 * state by virtue of being in or having passed through an dynticks
615 * idle state since the last call to dyntick_save_progress_counter()
618 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
623 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
624 snap
= (unsigned int)rdp
->dynticks_snap
;
627 * If the CPU passed through or entered a dynticks idle phase with
628 * no active irq/NMI handlers, then we can safely pretend that the CPU
629 * already acknowledged the request to pass through a quiescent
630 * state. Either way, that CPU cannot possibly be in an RCU
631 * read-side critical section that started before the beginning
632 * of the current RCU grace period.
634 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
635 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
640 /* Go check for the CPU being offline. */
641 return rcu_implicit_offline_qs(rdp
);
644 #endif /* #ifdef CONFIG_SMP */
646 int rcu_cpu_stall_suppress __read_mostly
;
648 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
650 rsp
->gp_start
= jiffies
;
651 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
654 static void print_other_cpu_stall(struct rcu_state
*rsp
)
660 struct rcu_node
*rnp
= rcu_get_root(rsp
);
662 /* Only let one CPU complain about others per time interval. */
664 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
665 delta
= jiffies
- rsp
->jiffies_stall
;
666 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
667 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
670 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
673 * Now rat on any tasks that got kicked up to the root rcu_node
674 * due to CPU offlining.
676 ndetected
= rcu_print_task_stall(rnp
);
677 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
680 * OK, time to rat on our buddy...
681 * See Documentation/RCU/stallwarn.txt for info on how to debug
682 * RCU CPU stall warnings.
684 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
686 rcu_for_each_leaf_node(rsp
, rnp
) {
687 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
688 ndetected
+= rcu_print_task_stall(rnp
);
689 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
690 if (rnp
->qsmask
== 0)
692 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
693 if (rnp
->qsmask
& (1UL << cpu
)) {
694 printk(" %d", rnp
->grplo
+ cpu
);
698 printk("} (detected by %d, t=%ld jiffies)\n",
699 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
701 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
702 else if (!trigger_all_cpu_backtrace())
705 /* If so configured, complain about tasks blocking the grace period. */
707 rcu_print_detail_task_stall(rsp
);
709 force_quiescent_state(rsp
, 0); /* Kick them all. */
712 static void print_cpu_stall(struct rcu_state
*rsp
)
715 struct rcu_node
*rnp
= rcu_get_root(rsp
);
718 * OK, time to rat on ourselves...
719 * See Documentation/RCU/stallwarn.txt for info on how to debug
720 * RCU CPU stall warnings.
722 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
723 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
724 if (!trigger_all_cpu_backtrace())
727 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
728 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
730 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
731 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
733 set_need_resched(); /* kick ourselves to get things going. */
736 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
740 struct rcu_node
*rnp
;
742 if (rcu_cpu_stall_suppress
)
744 j
= ACCESS_ONCE(jiffies
);
745 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
747 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
749 /* We haven't checked in, so go dump stack. */
750 print_cpu_stall(rsp
);
752 } else if (rcu_gp_in_progress(rsp
) &&
753 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
755 /* They had a few time units to dump stack, so complain. */
756 print_other_cpu_stall(rsp
);
760 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
762 rcu_cpu_stall_suppress
= 1;
767 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
769 * Set the stall-warning timeout way off into the future, thus preventing
770 * any RCU CPU stall-warning messages from appearing in the current set of
773 * The caller must disable hard irqs.
775 void rcu_cpu_stall_reset(void)
777 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
778 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
779 rcu_preempt_stall_reset();
782 static struct notifier_block rcu_panic_block
= {
783 .notifier_call
= rcu_panic
,
786 static void __init
check_cpu_stall_init(void)
788 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
792 * Update CPU-local rcu_data state to record the newly noticed grace period.
793 * This is used both when we started the grace period and when we notice
794 * that someone else started the grace period. The caller must hold the
795 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
796 * and must have irqs disabled.
798 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
800 if (rdp
->gpnum
!= rnp
->gpnum
) {
802 * If the current grace period is waiting for this CPU,
803 * set up to detect a quiescent state, otherwise don't
804 * go looking for one.
806 rdp
->gpnum
= rnp
->gpnum
;
807 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
808 if (rnp
->qsmask
& rdp
->grpmask
) {
810 rdp
->passed_quiesce
= 0;
816 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
819 struct rcu_node
*rnp
;
821 local_irq_save(flags
);
823 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
824 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
825 local_irq_restore(flags
);
828 __note_new_gpnum(rsp
, rnp
, rdp
);
829 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
833 * Did someone else start a new RCU grace period start since we last
834 * checked? Update local state appropriately if so. Must be called
835 * on the CPU corresponding to rdp.
838 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
843 local_irq_save(flags
);
844 if (rdp
->gpnum
!= rsp
->gpnum
) {
845 note_new_gpnum(rsp
, rdp
);
848 local_irq_restore(flags
);
853 * Advance this CPU's callbacks, but only if the current grace period
854 * has ended. This may be called only from the CPU to whom the rdp
855 * belongs. In addition, the corresponding leaf rcu_node structure's
856 * ->lock must be held by the caller, with irqs disabled.
859 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
861 /* Did another grace period end? */
862 if (rdp
->completed
!= rnp
->completed
) {
864 /* Advance callbacks. No harm if list empty. */
865 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
866 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
867 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
869 /* Remember that we saw this grace-period completion. */
870 rdp
->completed
= rnp
->completed
;
871 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
874 * If we were in an extended quiescent state, we may have
875 * missed some grace periods that others CPUs handled on
876 * our behalf. Catch up with this state to avoid noting
877 * spurious new grace periods. If another grace period
878 * has started, then rnp->gpnum will have advanced, so
879 * we will detect this later on.
881 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
882 rdp
->gpnum
= rdp
->completed
;
885 * If RCU does not need a quiescent state from this CPU,
886 * then make sure that this CPU doesn't go looking for one.
888 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
894 * Advance this CPU's callbacks, but only if the current grace period
895 * has ended. This may be called only from the CPU to whom the rdp
899 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
902 struct rcu_node
*rnp
;
904 local_irq_save(flags
);
906 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
907 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
908 local_irq_restore(flags
);
911 __rcu_process_gp_end(rsp
, rnp
, rdp
);
912 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
916 * Do per-CPU grace-period initialization for running CPU. The caller
917 * must hold the lock of the leaf rcu_node structure corresponding to
921 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
923 /* Prior grace period ended, so advance callbacks for current CPU. */
924 __rcu_process_gp_end(rsp
, rnp
, rdp
);
927 * Because this CPU just now started the new grace period, we know
928 * that all of its callbacks will be covered by this upcoming grace
929 * period, even the ones that were registered arbitrarily recently.
930 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
932 * Other CPUs cannot be sure exactly when the grace period started.
933 * Therefore, their recently registered callbacks must pass through
934 * an additional RCU_NEXT_READY stage, so that they will be handled
935 * by the next RCU grace period.
937 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
938 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
940 /* Set state so that this CPU will detect the next quiescent state. */
941 __note_new_gpnum(rsp
, rnp
, rdp
);
945 * Start a new RCU grace period if warranted, re-initializing the hierarchy
946 * in preparation for detecting the next grace period. The caller must hold
947 * the root node's ->lock, which is released before return. Hard irqs must
951 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
952 __releases(rcu_get_root(rsp
)->lock
)
954 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
955 struct rcu_node
*rnp
= rcu_get_root(rsp
);
957 if (!rcu_scheduler_fully_active
||
958 !cpu_needs_another_gp(rsp
, rdp
)) {
960 * Either the scheduler hasn't yet spawned the first
961 * non-idle task or this CPU does not need another
962 * grace period. Either way, don't start a new grace
965 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
969 if (rsp
->fqs_active
) {
971 * This CPU needs a grace period, but force_quiescent_state()
972 * is running. Tell it to start one on this CPU's behalf.
974 rsp
->fqs_need_gp
= 1;
975 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
979 /* Advance to a new grace period and initialize state. */
981 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
982 WARN_ON_ONCE(rsp
->fqs_state
== RCU_GP_INIT
);
983 rsp
->fqs_state
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
984 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
985 record_gp_stall_check_time(rsp
);
987 /* Special-case the common single-level case. */
988 if (NUM_RCU_NODES
== 1) {
989 rcu_preempt_check_blocked_tasks(rnp
);
990 rnp
->qsmask
= rnp
->qsmaskinit
;
991 rnp
->gpnum
= rsp
->gpnum
;
992 rnp
->completed
= rsp
->completed
;
993 rsp
->fqs_state
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK */
994 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
995 rcu_preempt_boost_start_gp(rnp
);
996 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
997 rnp
->level
, rnp
->grplo
,
998 rnp
->grphi
, rnp
->qsmask
);
999 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1003 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
1006 /* Exclude any concurrent CPU-hotplug operations. */
1007 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
1010 * Set the quiescent-state-needed bits in all the rcu_node
1011 * structures for all currently online CPUs in breadth-first
1012 * order, starting from the root rcu_node structure. This
1013 * operation relies on the layout of the hierarchy within the
1014 * rsp->node[] array. Note that other CPUs will access only
1015 * the leaves of the hierarchy, which still indicate that no
1016 * grace period is in progress, at least until the corresponding
1017 * leaf node has been initialized. In addition, we have excluded
1018 * CPU-hotplug operations.
1020 * Note that the grace period cannot complete until we finish
1021 * the initialization process, as there will be at least one
1022 * qsmask bit set in the root node until that time, namely the
1023 * one corresponding to this CPU, due to the fact that we have
1026 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1027 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1028 rcu_preempt_check_blocked_tasks(rnp
);
1029 rnp
->qsmask
= rnp
->qsmaskinit
;
1030 rnp
->gpnum
= rsp
->gpnum
;
1031 rnp
->completed
= rsp
->completed
;
1032 if (rnp
== rdp
->mynode
)
1033 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1034 rcu_preempt_boost_start_gp(rnp
);
1035 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1036 rnp
->level
, rnp
->grplo
,
1037 rnp
->grphi
, rnp
->qsmask
);
1038 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1041 rnp
= rcu_get_root(rsp
);
1042 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1043 rsp
->fqs_state
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
1044 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1045 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
1049 * Report a full set of quiescent states to the specified rcu_state
1050 * data structure. This involves cleaning up after the prior grace
1051 * period and letting rcu_start_gp() start up the next grace period
1052 * if one is needed. Note that the caller must hold rnp->lock, as
1053 * required by rcu_start_gp(), which will release it.
1055 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1056 __releases(rcu_get_root(rsp
)->lock
)
1058 unsigned long gp_duration
;
1059 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1060 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1062 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1065 * Ensure that all grace-period and pre-grace-period activity
1066 * is seen before the assignment to rsp->completed.
1068 smp_mb(); /* See above block comment. */
1069 gp_duration
= jiffies
- rsp
->gp_start
;
1070 if (gp_duration
> rsp
->gp_max
)
1071 rsp
->gp_max
= gp_duration
;
1074 * We know the grace period is complete, but to everyone else
1075 * it appears to still be ongoing. But it is also the case
1076 * that to everyone else it looks like there is nothing that
1077 * they can do to advance the grace period. It is therefore
1078 * safe for us to drop the lock in order to mark the grace
1079 * period as completed in all of the rcu_node structures.
1081 * But if this CPU needs another grace period, it will take
1082 * care of this while initializing the next grace period.
1083 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1084 * because the callbacks have not yet been advanced: Those
1085 * callbacks are waiting on the grace period that just now
1088 if (*rdp
->nxttail
[RCU_WAIT_TAIL
] == NULL
) {
1089 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1092 * Propagate new ->completed value to rcu_node structures
1093 * so that other CPUs don't have to wait until the start
1094 * of the next grace period to process their callbacks.
1096 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1097 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1098 rnp
->completed
= rsp
->gpnum
;
1099 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1101 rnp
= rcu_get_root(rsp
);
1102 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1105 rsp
->completed
= rsp
->gpnum
; /* Declare the grace period complete. */
1106 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1107 rsp
->fqs_state
= RCU_GP_IDLE
;
1108 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
1112 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1113 * Allows quiescent states for a group of CPUs to be reported at one go
1114 * to the specified rcu_node structure, though all the CPUs in the group
1115 * must be represented by the same rcu_node structure (which need not be
1116 * a leaf rcu_node structure, though it often will be). That structure's
1117 * lock must be held upon entry, and it is released before return.
1120 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1121 struct rcu_node
*rnp
, unsigned long flags
)
1122 __releases(rnp
->lock
)
1124 struct rcu_node
*rnp_c
;
1126 /* Walk up the rcu_node hierarchy. */
1128 if (!(rnp
->qsmask
& mask
)) {
1130 /* Our bit has already been cleared, so done. */
1131 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1134 rnp
->qsmask
&= ~mask
;
1135 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1136 mask
, rnp
->qsmask
, rnp
->level
,
1137 rnp
->grplo
, rnp
->grphi
,
1139 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1141 /* Other bits still set at this level, so done. */
1142 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1145 mask
= rnp
->grpmask
;
1146 if (rnp
->parent
== NULL
) {
1148 /* No more levels. Exit loop holding root lock. */
1152 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1155 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1156 WARN_ON_ONCE(rnp_c
->qsmask
);
1160 * Get here if we are the last CPU to pass through a quiescent
1161 * state for this grace period. Invoke rcu_report_qs_rsp()
1162 * to clean up and start the next grace period if one is needed.
1164 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1168 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1169 * structure. This must be either called from the specified CPU, or
1170 * called when the specified CPU is known to be offline (and when it is
1171 * also known that no other CPU is concurrently trying to help the offline
1172 * CPU). The lastcomp argument is used to make sure we are still in the
1173 * grace period of interest. We don't want to end the current grace period
1174 * based on quiescent states detected in an earlier grace period!
1177 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastgp
)
1179 unsigned long flags
;
1181 struct rcu_node
*rnp
;
1184 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1185 if (lastgp
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
) {
1188 * The grace period in which this quiescent state was
1189 * recorded has ended, so don't report it upwards.
1190 * We will instead need a new quiescent state that lies
1191 * within the current grace period.
1193 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1194 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1197 mask
= rdp
->grpmask
;
1198 if ((rnp
->qsmask
& mask
) == 0) {
1199 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1201 rdp
->qs_pending
= 0;
1204 * This GP can't end until cpu checks in, so all of our
1205 * callbacks can be processed during the next GP.
1207 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1209 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1214 * Check to see if there is a new grace period of which this CPU
1215 * is not yet aware, and if so, set up local rcu_data state for it.
1216 * Otherwise, see if this CPU has just passed through its first
1217 * quiescent state for this grace period, and record that fact if so.
1220 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1222 /* If there is now a new grace period, record and return. */
1223 if (check_for_new_grace_period(rsp
, rdp
))
1227 * Does this CPU still need to do its part for current grace period?
1228 * If no, return and let the other CPUs do their part as well.
1230 if (!rdp
->qs_pending
)
1234 * Was there a quiescent state since the beginning of the grace
1235 * period? If no, then exit and wait for the next call.
1237 if (!rdp
->passed_quiesce
)
1241 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1244 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesce_gpnum
);
1247 #ifdef CONFIG_HOTPLUG_CPU
1250 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1251 * Synchronization is not required because this function executes
1252 * in stop_machine() context.
1254 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1257 /* current DYING CPU is cleared in the cpu_online_mask */
1258 int receive_cpu
= cpumask_any(cpu_online_mask
);
1259 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1260 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1262 if (rdp
->nxtlist
== NULL
)
1263 return; /* irqs disabled, so comparison is stable. */
1265 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1266 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1267 receive_rdp
->qlen
+= rdp
->qlen
;
1268 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1269 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1271 rdp
->nxtlist
= NULL
;
1272 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1273 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1278 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1279 * and move all callbacks from the outgoing CPU to the current one.
1280 * There can only be one CPU hotplug operation at a time, so no other
1281 * CPU can be attempting to update rcu_cpu_kthread_task.
1283 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1285 unsigned long flags
;
1287 int need_report
= 0;
1288 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1289 struct rcu_node
*rnp
;
1291 rcu_stop_cpu_kthread(cpu
);
1293 /* Exclude any attempts to start a new grace period. */
1294 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1296 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1297 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1298 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1300 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1301 rnp
->qsmaskinit
&= ~mask
;
1302 if (rnp
->qsmaskinit
!= 0) {
1303 if (rnp
!= rdp
->mynode
)
1304 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1306 trace_rcu_grace_period(rsp
->name
,
1308 !!(rnp
->qsmask
& mask
),
1312 if (rnp
== rdp
->mynode
) {
1313 trace_rcu_grace_period(rsp
->name
,
1315 !!(rnp
->qsmask
& mask
),
1317 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1319 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1320 mask
= rnp
->grpmask
;
1322 } while (rnp
!= NULL
);
1325 * We still hold the leaf rcu_node structure lock here, and
1326 * irqs are still disabled. The reason for this subterfuge is
1327 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1328 * held leads to deadlock.
1330 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1332 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1333 rcu_report_unblock_qs_rnp(rnp
, flags
);
1335 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1336 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1337 rcu_report_exp_rnp(rsp
, rnp
, true);
1338 rcu_node_kthread_setaffinity(rnp
, -1);
1342 * Remove the specified CPU from the RCU hierarchy and move any pending
1343 * callbacks that it might have to the current CPU. This code assumes
1344 * that at least one CPU in the system will remain running at all times.
1345 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1347 static void rcu_offline_cpu(int cpu
)
1349 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1350 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1351 rcu_preempt_offline_cpu(cpu
);
1354 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1356 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1360 static void rcu_offline_cpu(int cpu
)
1364 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1367 * Invoke any RCU callbacks that have made it to the end of their grace
1368 * period. Thottle as specified by rdp->blimit.
1370 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1372 unsigned long flags
;
1373 struct rcu_head
*next
, *list
, **tail
;
1376 /* If no callbacks are ready, just return.*/
1377 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1378 trace_rcu_batch_start(rsp
->name
, 0, 0);
1379 trace_rcu_batch_end(rsp
->name
, 0);
1384 * Extract the list of ready callbacks, disabling to prevent
1385 * races with call_rcu() from interrupt handlers.
1387 local_irq_save(flags
);
1389 trace_rcu_batch_start(rsp
->name
, rdp
->qlen
, bl
);
1390 list
= rdp
->nxtlist
;
1391 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1392 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1393 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1394 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1395 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1396 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1397 local_irq_restore(flags
);
1399 /* Invoke callbacks. */
1404 debug_rcu_head_unqueue(list
);
1405 __rcu_reclaim(rsp
->name
, list
);
1411 local_irq_save(flags
);
1412 trace_rcu_batch_end(rsp
->name
, count
);
1414 /* Update count, and requeue any remaining callbacks. */
1416 rdp
->n_cbs_invoked
+= count
;
1418 *tail
= rdp
->nxtlist
;
1419 rdp
->nxtlist
= list
;
1420 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1421 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1422 rdp
->nxttail
[count
] = tail
;
1427 /* Reinstate batch limit if we have worked down the excess. */
1428 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1429 rdp
->blimit
= blimit
;
1431 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1432 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1433 rdp
->qlen_last_fqs_check
= 0;
1434 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1435 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1436 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1438 local_irq_restore(flags
);
1440 /* Re-invoke RCU core processing if there are callbacks remaining. */
1441 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1446 * Check to see if this CPU is in a non-context-switch quiescent state
1447 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1448 * Also schedule RCU core processing.
1450 * This function must be called from hardirq context. It is normally
1451 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1452 * false, there is no point in invoking rcu_check_callbacks().
1454 void rcu_check_callbacks(int cpu
, int user
)
1456 trace_rcu_utilization("Start scheduler-tick");
1457 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1460 * Get here if this CPU took its interrupt from user
1461 * mode or from the idle loop, and if this is not a
1462 * nested interrupt. In this case, the CPU is in
1463 * a quiescent state, so note it.
1465 * No memory barrier is required here because both
1466 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1467 * variables that other CPUs neither access nor modify,
1468 * at least not while the corresponding CPU is online.
1474 } else if (!in_softirq()) {
1477 * Get here if this CPU did not take its interrupt from
1478 * softirq, in other words, if it is not interrupting
1479 * a rcu_bh read-side critical section. This is an _bh
1480 * critical section, so note it.
1485 rcu_preempt_check_callbacks(cpu
);
1486 if (rcu_pending(cpu
))
1488 trace_rcu_utilization("End scheduler-tick");
1494 * Scan the leaf rcu_node structures, processing dyntick state for any that
1495 * have not yet encountered a quiescent state, using the function specified.
1496 * Also initiate boosting for any threads blocked on the root rcu_node.
1498 * The caller must have suppressed start of new grace periods.
1500 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1504 unsigned long flags
;
1506 struct rcu_node
*rnp
;
1508 rcu_for_each_leaf_node(rsp
, rnp
) {
1510 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1511 if (!rcu_gp_in_progress(rsp
)) {
1512 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1515 if (rnp
->qsmask
== 0) {
1516 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1521 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1522 if ((rnp
->qsmask
& bit
) != 0 &&
1523 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1528 /* rcu_report_qs_rnp() releases rnp->lock. */
1529 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1532 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1534 rnp
= rcu_get_root(rsp
);
1535 if (rnp
->qsmask
== 0) {
1536 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1537 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1542 * Force quiescent states on reluctant CPUs, and also detect which
1543 * CPUs are in dyntick-idle mode.
1545 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1547 unsigned long flags
;
1548 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1550 trace_rcu_utilization("Start fqs");
1551 if (!rcu_gp_in_progress(rsp
)) {
1552 trace_rcu_utilization("End fqs");
1553 return; /* No grace period in progress, nothing to force. */
1555 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1556 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1557 trace_rcu_utilization("End fqs");
1558 return; /* Someone else is already on the job. */
1560 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1561 goto unlock_fqs_ret
; /* no emergency and done recently. */
1563 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1564 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1565 if(!rcu_gp_in_progress(rsp
)) {
1566 rsp
->n_force_qs_ngp
++;
1567 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1568 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1570 rsp
->fqs_active
= 1;
1571 switch (rsp
->fqs_state
) {
1575 break; /* grace period idle or initializing, ignore. */
1577 case RCU_SAVE_DYNTICK
:
1578 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1579 break; /* So gcc recognizes the dead code. */
1581 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1583 /* Record dyntick-idle state. */
1584 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1585 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1586 if (rcu_gp_in_progress(rsp
))
1587 rsp
->fqs_state
= RCU_FORCE_QS
;
1592 /* Check dyntick-idle state, send IPI to laggarts. */
1593 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1594 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1596 /* Leave state in case more forcing is required. */
1598 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1601 rsp
->fqs_active
= 0;
1602 if (rsp
->fqs_need_gp
) {
1603 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1604 rsp
->fqs_need_gp
= 0;
1605 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1606 trace_rcu_utilization("End fqs");
1609 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1611 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1612 trace_rcu_utilization("End fqs");
1615 #else /* #ifdef CONFIG_SMP */
1617 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1622 #endif /* #else #ifdef CONFIG_SMP */
1625 * This does the RCU core processing work for the specified rcu_state
1626 * and rcu_data structures. This may be called only from the CPU to
1627 * whom the rdp belongs.
1630 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1632 unsigned long flags
;
1634 WARN_ON_ONCE(rdp
->beenonline
== 0);
1637 * If an RCU GP has gone long enough, go check for dyntick
1638 * idle CPUs and, if needed, send resched IPIs.
1640 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1641 force_quiescent_state(rsp
, 1);
1644 * Advance callbacks in response to end of earlier grace
1645 * period that some other CPU ended.
1647 rcu_process_gp_end(rsp
, rdp
);
1649 /* Update RCU state based on any recent quiescent states. */
1650 rcu_check_quiescent_state(rsp
, rdp
);
1652 /* Does this CPU require a not-yet-started grace period? */
1653 if (cpu_needs_another_gp(rsp
, rdp
)) {
1654 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1655 rcu_start_gp(rsp
, flags
); /* releases above lock */
1658 /* If there are callbacks ready, invoke them. */
1659 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1660 invoke_rcu_callbacks(rsp
, rdp
);
1664 * Do RCU core processing for the current CPU.
1666 static void rcu_process_callbacks(struct softirq_action
*unused
)
1668 trace_rcu_utilization("Start RCU core");
1669 __rcu_process_callbacks(&rcu_sched_state
,
1670 &__get_cpu_var(rcu_sched_data
));
1671 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1672 rcu_preempt_process_callbacks();
1673 trace_rcu_utilization("End RCU core");
1677 * Schedule RCU callback invocation. If the specified type of RCU
1678 * does not support RCU priority boosting, just do a direct call,
1679 * otherwise wake up the per-CPU kernel kthread. Note that because we
1680 * are running on the current CPU with interrupts disabled, the
1681 * rcu_cpu_kthread_task cannot disappear out from under us.
1683 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1685 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
1687 if (likely(!rsp
->boost
)) {
1688 rcu_do_batch(rsp
, rdp
);
1691 invoke_rcu_callbacks_kthread();
1694 static void invoke_rcu_core(void)
1696 raise_softirq(RCU_SOFTIRQ
);
1700 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1701 struct rcu_state
*rsp
)
1703 unsigned long flags
;
1704 struct rcu_data
*rdp
;
1706 debug_rcu_head_queue(head
);
1710 smp_mb(); /* Ensure RCU update seen before callback registry. */
1713 * Opportunistically note grace-period endings and beginnings.
1714 * Note that we might see a beginning right after we see an
1715 * end, but never vice versa, since this CPU has to pass through
1716 * a quiescent state betweentimes.
1718 local_irq_save(flags
);
1719 rdp
= this_cpu_ptr(rsp
->rda
);
1721 /* Add the callback to our list. */
1722 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1723 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1726 if (__is_kfree_rcu_offset((unsigned long)func
))
1727 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
1730 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen
);
1732 /* If interrupts were disabled, don't dive into RCU core. */
1733 if (irqs_disabled_flags(flags
)) {
1734 local_irq_restore(flags
);
1739 * Force the grace period if too many callbacks or too long waiting.
1740 * Enforce hysteresis, and don't invoke force_quiescent_state()
1741 * if some other CPU has recently done so. Also, don't bother
1742 * invoking force_quiescent_state() if the newly enqueued callback
1743 * is the only one waiting for a grace period to complete.
1745 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1747 /* Are we ignoring a completed grace period? */
1748 rcu_process_gp_end(rsp
, rdp
);
1749 check_for_new_grace_period(rsp
, rdp
);
1751 /* Start a new grace period if one not already started. */
1752 if (!rcu_gp_in_progress(rsp
)) {
1753 unsigned long nestflag
;
1754 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1756 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1757 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1759 /* Give the grace period a kick. */
1760 rdp
->blimit
= LONG_MAX
;
1761 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1762 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1763 force_quiescent_state(rsp
, 0);
1764 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1765 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1767 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1768 force_quiescent_state(rsp
, 1);
1769 local_irq_restore(flags
);
1773 * Queue an RCU-sched callback for invocation after a grace period.
1775 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1777 __call_rcu(head
, func
, &rcu_sched_state
);
1779 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1782 * Queue an RCU for invocation after a quicker grace period.
1784 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1786 __call_rcu(head
, func
, &rcu_bh_state
);
1788 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1791 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1793 * Control will return to the caller some time after a full rcu-sched
1794 * grace period has elapsed, in other words after all currently executing
1795 * rcu-sched read-side critical sections have completed. These read-side
1796 * critical sections are delimited by rcu_read_lock_sched() and
1797 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1798 * local_irq_disable(), and so on may be used in place of
1799 * rcu_read_lock_sched().
1801 * This means that all preempt_disable code sequences, including NMI and
1802 * hardware-interrupt handlers, in progress on entry will have completed
1803 * before this primitive returns. However, this does not guarantee that
1804 * softirq handlers will have completed, since in some kernels, these
1805 * handlers can run in process context, and can block.
1807 * This primitive provides the guarantees made by the (now removed)
1808 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1809 * guarantees that rcu_read_lock() sections will have completed.
1810 * In "classic RCU", these two guarantees happen to be one and
1811 * the same, but can differ in realtime RCU implementations.
1813 void synchronize_sched(void)
1815 if (rcu_blocking_is_gp())
1817 wait_rcu_gp(call_rcu_sched
);
1819 EXPORT_SYMBOL_GPL(synchronize_sched
);
1822 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1824 * Control will return to the caller some time after a full rcu_bh grace
1825 * period has elapsed, in other words after all currently executing rcu_bh
1826 * read-side critical sections have completed. RCU read-side critical
1827 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1828 * and may be nested.
1830 void synchronize_rcu_bh(void)
1832 if (rcu_blocking_is_gp())
1834 wait_rcu_gp(call_rcu_bh
);
1836 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1839 * Check to see if there is any immediate RCU-related work to be done
1840 * by the current CPU, for the specified type of RCU, returning 1 if so.
1841 * The checks are in order of increasing expense: checks that can be
1842 * carried out against CPU-local state are performed first. However,
1843 * we must check for CPU stalls first, else we might not get a chance.
1845 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1847 struct rcu_node
*rnp
= rdp
->mynode
;
1849 rdp
->n_rcu_pending
++;
1851 /* Check for CPU stalls, if enabled. */
1852 check_cpu_stall(rsp
, rdp
);
1854 /* Is the RCU core waiting for a quiescent state from this CPU? */
1855 if (rcu_scheduler_fully_active
&&
1856 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
1859 * If force_quiescent_state() coming soon and this CPU
1860 * needs a quiescent state, and this is either RCU-sched
1861 * or RCU-bh, force a local reschedule.
1863 rdp
->n_rp_qs_pending
++;
1864 if (!rdp
->preemptible
&&
1865 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
1868 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
1869 rdp
->n_rp_report_qs
++;
1873 /* Does this CPU have callbacks ready to invoke? */
1874 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1875 rdp
->n_rp_cb_ready
++;
1879 /* Has RCU gone idle with this CPU needing another grace period? */
1880 if (cpu_needs_another_gp(rsp
, rdp
)) {
1881 rdp
->n_rp_cpu_needs_gp
++;
1885 /* Has another RCU grace period completed? */
1886 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
1887 rdp
->n_rp_gp_completed
++;
1891 /* Has a new RCU grace period started? */
1892 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
1893 rdp
->n_rp_gp_started
++;
1897 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1898 if (rcu_gp_in_progress(rsp
) &&
1899 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
1900 rdp
->n_rp_need_fqs
++;
1905 rdp
->n_rp_need_nothing
++;
1910 * Check to see if there is any immediate RCU-related work to be done
1911 * by the current CPU, returning 1 if so. This function is part of the
1912 * RCU implementation; it is -not- an exported member of the RCU API.
1914 static int rcu_pending(int cpu
)
1916 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
1917 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
1918 rcu_preempt_pending(cpu
);
1922 * Check to see if any future RCU-related work will need to be done
1923 * by the current CPU, even if none need be done immediately, returning
1926 static int rcu_needs_cpu_quick_check(int cpu
)
1928 /* RCU callbacks either ready or pending? */
1929 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
1930 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
1931 rcu_preempt_needs_cpu(cpu
);
1934 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
1935 static atomic_t rcu_barrier_cpu_count
;
1936 static DEFINE_MUTEX(rcu_barrier_mutex
);
1937 static struct completion rcu_barrier_completion
;
1939 static void rcu_barrier_callback(struct rcu_head
*notused
)
1941 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1942 complete(&rcu_barrier_completion
);
1946 * Called with preemption disabled, and from cross-cpu IRQ context.
1948 static void rcu_barrier_func(void *type
)
1950 int cpu
= smp_processor_id();
1951 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
1952 void (*call_rcu_func
)(struct rcu_head
*head
,
1953 void (*func
)(struct rcu_head
*head
));
1955 atomic_inc(&rcu_barrier_cpu_count
);
1956 call_rcu_func
= type
;
1957 call_rcu_func(head
, rcu_barrier_callback
);
1961 * Orchestrate the specified type of RCU barrier, waiting for all
1962 * RCU callbacks of the specified type to complete.
1964 static void _rcu_barrier(struct rcu_state
*rsp
,
1965 void (*call_rcu_func
)(struct rcu_head
*head
,
1966 void (*func
)(struct rcu_head
*head
)))
1968 BUG_ON(in_interrupt());
1969 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1970 mutex_lock(&rcu_barrier_mutex
);
1971 init_completion(&rcu_barrier_completion
);
1973 * Initialize rcu_barrier_cpu_count to 1, then invoke
1974 * rcu_barrier_func() on each CPU, so that each CPU also has
1975 * incremented rcu_barrier_cpu_count. Only then is it safe to
1976 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1977 * might complete its grace period before all of the other CPUs
1978 * did their increment, causing this function to return too
1979 * early. Note that on_each_cpu() disables irqs, which prevents
1980 * any CPUs from coming online or going offline until each online
1981 * CPU has queued its RCU-barrier callback.
1983 atomic_set(&rcu_barrier_cpu_count
, 1);
1984 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
1985 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1986 complete(&rcu_barrier_completion
);
1987 wait_for_completion(&rcu_barrier_completion
);
1988 mutex_unlock(&rcu_barrier_mutex
);
1992 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1994 void rcu_barrier_bh(void)
1996 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
1998 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2001 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2003 void rcu_barrier_sched(void)
2005 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2007 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2010 * Do boot-time initialization of a CPU's per-CPU RCU data.
2013 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2015 unsigned long flags
;
2017 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2018 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2020 /* Set up local state, ensuring consistent view of global state. */
2021 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2022 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2023 rdp
->nxtlist
= NULL
;
2024 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2025 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2027 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2028 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_NESTING
);
2029 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2032 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2036 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2037 * offline event can be happening at a given time. Note also that we
2038 * can accept some slop in the rsp->completed access due to the fact
2039 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2041 static void __cpuinit
2042 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2044 unsigned long flags
;
2046 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2047 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2049 /* Set up local state, ensuring consistent view of global state. */
2050 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2051 rdp
->beenonline
= 1; /* We have now been online. */
2052 rdp
->preemptible
= preemptible
;
2053 rdp
->qlen_last_fqs_check
= 0;
2054 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2055 rdp
->blimit
= blimit
;
2056 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_NESTING
);
2057 WARN_ON_ONCE((atomic_read(&rdp
->dynticks
->dynticks
) & 0x1) != 1);
2058 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2061 * A new grace period might start here. If so, we won't be part
2062 * of it, but that is OK, as we are currently in a quiescent state.
2065 /* Exclude any attempts to start a new GP on large systems. */
2066 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2068 /* Add CPU to rcu_node bitmasks. */
2070 mask
= rdp
->grpmask
;
2072 /* Exclude any attempts to start a new GP on small systems. */
2073 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2074 rnp
->qsmaskinit
|= mask
;
2075 mask
= rnp
->grpmask
;
2076 if (rnp
== rdp
->mynode
) {
2078 * If there is a grace period in progress, we will
2079 * set up to wait for it next time we run the
2082 rdp
->gpnum
= rnp
->completed
;
2083 rdp
->completed
= rnp
->completed
;
2084 rdp
->passed_quiesce
= 0;
2085 rdp
->qs_pending
= 0;
2086 rdp
->passed_quiesce_gpnum
= rnp
->gpnum
- 1;
2087 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2089 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2091 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2093 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2096 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2098 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2099 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2100 rcu_preempt_init_percpu_data(cpu
);
2104 * Handle CPU online/offline notification events.
2106 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2107 unsigned long action
, void *hcpu
)
2109 long cpu
= (long)hcpu
;
2110 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2111 struct rcu_node
*rnp
= rdp
->mynode
;
2113 trace_rcu_utilization("Start CPU hotplug");
2115 case CPU_UP_PREPARE
:
2116 case CPU_UP_PREPARE_FROZEN
:
2117 rcu_prepare_cpu(cpu
);
2118 rcu_prepare_kthreads(cpu
);
2121 case CPU_DOWN_FAILED
:
2122 rcu_node_kthread_setaffinity(rnp
, -1);
2123 rcu_cpu_kthread_setrt(cpu
, 1);
2125 case CPU_DOWN_PREPARE
:
2126 rcu_node_kthread_setaffinity(rnp
, cpu
);
2127 rcu_cpu_kthread_setrt(cpu
, 0);
2130 case CPU_DYING_FROZEN
:
2132 * The whole machine is "stopped" except this CPU, so we can
2133 * touch any data without introducing corruption. We send the
2134 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2136 rcu_send_cbs_to_online(&rcu_bh_state
);
2137 rcu_send_cbs_to_online(&rcu_sched_state
);
2138 rcu_preempt_send_cbs_to_online();
2141 case CPU_DEAD_FROZEN
:
2142 case CPU_UP_CANCELED
:
2143 case CPU_UP_CANCELED_FROZEN
:
2144 rcu_offline_cpu(cpu
);
2149 trace_rcu_utilization("End CPU hotplug");
2154 * This function is invoked towards the end of the scheduler's initialization
2155 * process. Before this is called, the idle task might contain
2156 * RCU read-side critical sections (during which time, this idle
2157 * task is booting the system). After this function is called, the
2158 * idle tasks are prohibited from containing RCU read-side critical
2159 * sections. This function also enables RCU lockdep checking.
2161 void rcu_scheduler_starting(void)
2163 WARN_ON(num_online_cpus() != 1);
2164 WARN_ON(nr_context_switches() > 0);
2165 rcu_scheduler_active
= 1;
2169 * Compute the per-level fanout, either using the exact fanout specified
2170 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2172 #ifdef CONFIG_RCU_FANOUT_EXACT
2173 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2177 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2178 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2179 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
2181 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2182 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2189 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2190 ccur
= rsp
->levelcnt
[i
];
2191 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2195 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2198 * Helper function for rcu_init() that initializes one rcu_state structure.
2200 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2201 struct rcu_data __percpu
*rda
)
2203 static char *buf
[] = { "rcu_node_level_0",
2206 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2210 struct rcu_node
*rnp
;
2212 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2214 /* Initialize the level-tracking arrays. */
2216 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2217 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2218 rcu_init_levelspread(rsp
);
2220 /* Initialize the elements themselves, starting from the leaves. */
2222 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2223 cpustride
*= rsp
->levelspread
[i
];
2224 rnp
= rsp
->level
[i
];
2225 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2226 raw_spin_lock_init(&rnp
->lock
);
2227 lockdep_set_class_and_name(&rnp
->lock
,
2228 &rcu_node_class
[i
], buf
[i
]);
2231 rnp
->qsmaskinit
= 0;
2232 rnp
->grplo
= j
* cpustride
;
2233 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2234 if (rnp
->grphi
>= NR_CPUS
)
2235 rnp
->grphi
= NR_CPUS
- 1;
2241 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2242 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2243 rnp
->parent
= rsp
->level
[i
- 1] +
2244 j
/ rsp
->levelspread
[i
- 1];
2247 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2252 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2253 for_each_possible_cpu(i
) {
2254 while (i
> rnp
->grphi
)
2256 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2257 rcu_boot_init_percpu_data(i
, rsp
);
2261 void __init
rcu_init(void)
2265 rcu_bootup_announce();
2266 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2267 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2268 __rcu_init_preempt();
2269 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2272 * We don't need protection against CPU-hotplug here because
2273 * this is called early in boot, before either interrupts
2274 * or the scheduler are operational.
2276 cpu_notifier(rcu_cpu_notify
, 0);
2277 for_each_online_cpu(cpu
)
2278 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2279 check_cpu_stall_init();
2282 #include "rcutree_plugin.h"