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>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
58 #include <trace/events/rcu.h>
62 /* Data structures. */
64 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
65 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
67 #define RCU_STATE_INITIALIZER(sname, cr) { \
68 .level = { &sname##_state.node[0] }, \
70 .fqs_state = RCU_GP_IDLE, \
71 .gpnum = 0UL - 300UL, \
72 .completed = 0UL - 300UL, \
73 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
77 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
81 struct rcu_state rcu_sched_state
=
82 RCU_STATE_INITIALIZER(rcu_sched
, call_rcu_sched
);
83 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
85 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
, call_rcu_bh
);
86 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
88 static struct rcu_state
*rcu_state
;
89 LIST_HEAD(rcu_struct_flavors
);
91 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
92 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
93 module_param(rcu_fanout_leaf
, int, 0444);
94 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
95 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
102 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
105 * The rcu_scheduler_active variable transitions from zero to one just
106 * before the first task is spawned. So when this variable is zero, RCU
107 * can assume that there is but one task, allowing RCU to (for example)
108 * optimize synchronize_sched() to a simple barrier(). When this variable
109 * is one, RCU must actually do all the hard work required to detect real
110 * grace periods. This variable is also used to suppress boot-time false
111 * positives from lockdep-RCU error checking.
113 int rcu_scheduler_active __read_mostly
;
114 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
117 * The rcu_scheduler_fully_active variable transitions from zero to one
118 * during the early_initcall() processing, which is after the scheduler
119 * is capable of creating new tasks. So RCU processing (for example,
120 * creating tasks for RCU priority boosting) must be delayed until after
121 * rcu_scheduler_fully_active transitions from zero to one. We also
122 * currently delay invocation of any RCU callbacks until after this point.
124 * It might later prove better for people registering RCU callbacks during
125 * early boot to take responsibility for these callbacks, but one step at
128 static int rcu_scheduler_fully_active __read_mostly
;
130 #ifdef CONFIG_RCU_BOOST
133 * Control variables for per-CPU and per-rcu_node kthreads. These
134 * handle all flavors of RCU.
136 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
139 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
141 #endif /* #ifdef CONFIG_RCU_BOOST */
143 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
144 static void invoke_rcu_core(void);
145 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
148 * Track the rcutorture test sequence number and the update version
149 * number within a given test. The rcutorture_testseq is incremented
150 * on every rcutorture module load and unload, so has an odd value
151 * when a test is running. The rcutorture_vernum is set to zero
152 * when rcutorture starts and is incremented on each rcutorture update.
153 * These variables enable correlating rcutorture output with the
154 * RCU tracing information.
156 unsigned long rcutorture_testseq
;
157 unsigned long rcutorture_vernum
;
160 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
161 * permit this function to be invoked without holding the root rcu_node
162 * structure's ->lock, but of course results can be subject to change.
164 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
166 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
170 * Note a quiescent state. Because we do not need to know
171 * how many quiescent states passed, just if there was at least
172 * one since the start of the grace period, this just sets a flag.
173 * The caller must have disabled preemption.
175 void rcu_sched_qs(int cpu
)
177 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
179 if (rdp
->passed_quiesce
== 0)
180 trace_rcu_grace_period("rcu_sched", rdp
->gpnum
, "cpuqs");
181 rdp
->passed_quiesce
= 1;
184 void rcu_bh_qs(int cpu
)
186 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
188 if (rdp
->passed_quiesce
== 0)
189 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
190 rdp
->passed_quiesce
= 1;
194 * Note a context switch. This is a quiescent state for RCU-sched,
195 * and requires special handling for preemptible RCU.
196 * The caller must have disabled preemption.
198 void rcu_note_context_switch(int cpu
)
200 trace_rcu_utilization("Start context switch");
202 rcu_preempt_note_context_switch(cpu
);
203 trace_rcu_utilization("End context switch");
205 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
207 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
208 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
209 .dynticks
= ATOMIC_INIT(1),
212 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
213 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
214 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
216 module_param(blimit
, long, 0444);
217 module_param(qhimark
, long, 0444);
218 module_param(qlowmark
, long, 0444);
220 static ulong jiffies_till_first_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
221 static ulong jiffies_till_next_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
223 module_param(jiffies_till_first_fqs
, ulong
, 0644);
224 module_param(jiffies_till_next_fqs
, ulong
, 0644);
226 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*));
227 static void force_quiescent_state(struct rcu_state
*rsp
);
228 static int rcu_pending(int cpu
);
231 * Return the number of RCU-sched batches processed thus far for debug & stats.
233 long rcu_batches_completed_sched(void)
235 return rcu_sched_state
.completed
;
237 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
240 * Return the number of RCU BH batches processed thus far for debug & stats.
242 long rcu_batches_completed_bh(void)
244 return rcu_bh_state
.completed
;
246 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
249 * Force a quiescent state for RCU BH.
251 void rcu_bh_force_quiescent_state(void)
253 force_quiescent_state(&rcu_bh_state
);
255 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
258 * Record the number of times rcutorture tests have been initiated and
259 * terminated. This information allows the debugfs tracing stats to be
260 * correlated to the rcutorture messages, even when the rcutorture module
261 * is being repeatedly loaded and unloaded. In other words, we cannot
262 * store this state in rcutorture itself.
264 void rcutorture_record_test_transition(void)
266 rcutorture_testseq
++;
267 rcutorture_vernum
= 0;
269 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
272 * Record the number of writer passes through the current rcutorture test.
273 * This is also used to correlate debugfs tracing stats with the rcutorture
276 void rcutorture_record_progress(unsigned long vernum
)
280 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
283 * Force a quiescent state for RCU-sched.
285 void rcu_sched_force_quiescent_state(void)
287 force_quiescent_state(&rcu_sched_state
);
289 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
292 * Does the CPU have callbacks ready to be invoked?
295 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
297 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
298 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
302 * Does the current CPU require a not-yet-started grace period?
303 * The caller must have disabled interrupts to prevent races with
304 * normal callback registry.
307 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
311 if (rcu_gp_in_progress(rsp
))
312 return 0; /* No, a grace period is already in progress. */
313 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
314 return 0; /* No, this is a no-CBs (or offline) CPU. */
315 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
316 return 1; /* Yes, this CPU has newly registered callbacks. */
317 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
318 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
319 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
320 rdp
->nxtcompleted
[i
]))
321 return 1; /* Yes, CBs for future grace period. */
322 return 0; /* No grace period needed. */
326 * Return the root node of the specified rcu_state structure.
328 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
330 return &rsp
->node
[0];
334 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
336 * If the new value of the ->dynticks_nesting counter now is zero,
337 * we really have entered idle, and must do the appropriate accounting.
338 * The caller must have disabled interrupts.
340 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
343 trace_rcu_dyntick("Start", oldval
, rdtp
->dynticks_nesting
);
344 if (!user
&& !is_idle_task(current
)) {
345 struct task_struct
*idle
= idle_task(smp_processor_id());
347 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
348 ftrace_dump(DUMP_ORIG
);
349 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
350 current
->pid
, current
->comm
,
351 idle
->pid
, idle
->comm
); /* must be idle task! */
353 rcu_prepare_for_idle(smp_processor_id());
354 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
355 smp_mb__before_atomic_inc(); /* See above. */
356 atomic_inc(&rdtp
->dynticks
);
357 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
358 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
361 * It is illegal to enter an extended quiescent state while
362 * in an RCU read-side critical section.
364 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
365 "Illegal idle entry in RCU read-side critical section.");
366 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
367 "Illegal idle entry in RCU-bh read-side critical section.");
368 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
369 "Illegal idle entry in RCU-sched read-side critical section.");
373 * Enter an RCU extended quiescent state, which can be either the
374 * idle loop or adaptive-tickless usermode execution.
376 static void rcu_eqs_enter(bool user
)
379 struct rcu_dynticks
*rdtp
;
381 rdtp
= &__get_cpu_var(rcu_dynticks
);
382 oldval
= rdtp
->dynticks_nesting
;
383 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
384 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
385 rdtp
->dynticks_nesting
= 0;
387 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
388 rcu_eqs_enter_common(rdtp
, oldval
, user
);
392 * rcu_idle_enter - inform RCU that current CPU is entering idle
394 * Enter idle mode, in other words, -leave- the mode in which RCU
395 * read-side critical sections can occur. (Though RCU read-side
396 * critical sections can occur in irq handlers in idle, a possibility
397 * handled by irq_enter() and irq_exit().)
399 * We crowbar the ->dynticks_nesting field to zero to allow for
400 * the possibility of usermode upcalls having messed up our count
401 * of interrupt nesting level during the prior busy period.
403 void rcu_idle_enter(void)
407 local_irq_save(flags
);
408 rcu_eqs_enter(false);
409 local_irq_restore(flags
);
411 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
413 #ifdef CONFIG_RCU_USER_QS
415 * rcu_user_enter - inform RCU that we are resuming userspace.
417 * Enter RCU idle mode right before resuming userspace. No use of RCU
418 * is permitted between this call and rcu_user_exit(). This way the
419 * CPU doesn't need to maintain the tick for RCU maintenance purposes
420 * when the CPU runs in userspace.
422 void rcu_user_enter(void)
428 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
429 * after the current irq returns.
431 * This is similar to rcu_user_enter() but in the context of a non-nesting
432 * irq. After this call, RCU enters into idle mode when the interrupt
435 void rcu_user_enter_after_irq(void)
438 struct rcu_dynticks
*rdtp
;
440 local_irq_save(flags
);
441 rdtp
= &__get_cpu_var(rcu_dynticks
);
442 /* Ensure this irq is interrupting a non-idle RCU state. */
443 WARN_ON_ONCE(!(rdtp
->dynticks_nesting
& DYNTICK_TASK_MASK
));
444 rdtp
->dynticks_nesting
= 1;
445 local_irq_restore(flags
);
447 #endif /* CONFIG_RCU_USER_QS */
450 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
452 * Exit from an interrupt handler, which might possibly result in entering
453 * idle mode, in other words, leaving the mode in which read-side critical
454 * sections can occur.
456 * This code assumes that the idle loop never does anything that might
457 * result in unbalanced calls to irq_enter() and irq_exit(). If your
458 * architecture violates this assumption, RCU will give you what you
459 * deserve, good and hard. But very infrequently and irreproducibly.
461 * Use things like work queues to work around this limitation.
463 * You have been warned.
465 void rcu_irq_exit(void)
469 struct rcu_dynticks
*rdtp
;
471 local_irq_save(flags
);
472 rdtp
= &__get_cpu_var(rcu_dynticks
);
473 oldval
= rdtp
->dynticks_nesting
;
474 rdtp
->dynticks_nesting
--;
475 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
476 if (rdtp
->dynticks_nesting
)
477 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
479 rcu_eqs_enter_common(rdtp
, oldval
, true);
480 local_irq_restore(flags
);
484 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
486 * If the new value of the ->dynticks_nesting counter was previously zero,
487 * we really have exited idle, and must do the appropriate accounting.
488 * The caller must have disabled interrupts.
490 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
493 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
494 atomic_inc(&rdtp
->dynticks
);
495 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
496 smp_mb__after_atomic_inc(); /* See above. */
497 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
498 rcu_cleanup_after_idle(smp_processor_id());
499 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
500 if (!user
&& !is_idle_task(current
)) {
501 struct task_struct
*idle
= idle_task(smp_processor_id());
503 trace_rcu_dyntick("Error on exit: not idle task",
504 oldval
, rdtp
->dynticks_nesting
);
505 ftrace_dump(DUMP_ORIG
);
506 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
507 current
->pid
, current
->comm
,
508 idle
->pid
, idle
->comm
); /* must be idle task! */
513 * Exit an RCU extended quiescent state, which can be either the
514 * idle loop or adaptive-tickless usermode execution.
516 static void rcu_eqs_exit(bool user
)
518 struct rcu_dynticks
*rdtp
;
521 rdtp
= &__get_cpu_var(rcu_dynticks
);
522 oldval
= rdtp
->dynticks_nesting
;
523 WARN_ON_ONCE(oldval
< 0);
524 if (oldval
& DYNTICK_TASK_NEST_MASK
)
525 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
527 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
528 rcu_eqs_exit_common(rdtp
, oldval
, user
);
532 * rcu_idle_exit - inform RCU that current CPU is leaving idle
534 * Exit idle mode, in other words, -enter- the mode in which RCU
535 * read-side critical sections can occur.
537 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
538 * allow for the possibility of usermode upcalls messing up our count
539 * of interrupt nesting level during the busy period that is just
542 void rcu_idle_exit(void)
546 local_irq_save(flags
);
548 local_irq_restore(flags
);
550 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
552 #ifdef CONFIG_RCU_USER_QS
554 * rcu_user_exit - inform RCU that we are exiting userspace.
556 * Exit RCU idle mode while entering the kernel because it can
557 * run a RCU read side critical section anytime.
559 void rcu_user_exit(void)
565 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
566 * idle mode after the current non-nesting irq returns.
568 * This is similar to rcu_user_exit() but in the context of an irq.
569 * This is called when the irq has interrupted a userspace RCU idle mode
570 * context. When the current non-nesting interrupt returns after this call,
571 * the CPU won't restore the RCU idle mode.
573 void rcu_user_exit_after_irq(void)
576 struct rcu_dynticks
*rdtp
;
578 local_irq_save(flags
);
579 rdtp
= &__get_cpu_var(rcu_dynticks
);
580 /* Ensure we are interrupting an RCU idle mode. */
581 WARN_ON_ONCE(rdtp
->dynticks_nesting
& DYNTICK_TASK_NEST_MASK
);
582 rdtp
->dynticks_nesting
+= DYNTICK_TASK_EXIT_IDLE
;
583 local_irq_restore(flags
);
585 #endif /* CONFIG_RCU_USER_QS */
588 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
590 * Enter an interrupt handler, which might possibly result in exiting
591 * idle mode, in other words, entering the mode in which read-side critical
592 * sections can occur.
594 * Note that the Linux kernel is fully capable of entering an interrupt
595 * handler that it never exits, for example when doing upcalls to
596 * user mode! This code assumes that the idle loop never does upcalls to
597 * user mode. If your architecture does do upcalls from the idle loop (or
598 * does anything else that results in unbalanced calls to the irq_enter()
599 * and irq_exit() functions), RCU will give you what you deserve, good
600 * and hard. But very infrequently and irreproducibly.
602 * Use things like work queues to work around this limitation.
604 * You have been warned.
606 void rcu_irq_enter(void)
609 struct rcu_dynticks
*rdtp
;
612 local_irq_save(flags
);
613 rdtp
= &__get_cpu_var(rcu_dynticks
);
614 oldval
= rdtp
->dynticks_nesting
;
615 rdtp
->dynticks_nesting
++;
616 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
618 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
620 rcu_eqs_exit_common(rdtp
, oldval
, true);
621 local_irq_restore(flags
);
625 * rcu_nmi_enter - inform RCU of entry to NMI context
627 * If the CPU was idle with dynamic ticks active, and there is no
628 * irq handler running, this updates rdtp->dynticks_nmi to let the
629 * RCU grace-period handling know that the CPU is active.
631 void rcu_nmi_enter(void)
633 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
635 if (rdtp
->dynticks_nmi_nesting
== 0 &&
636 (atomic_read(&rdtp
->dynticks
) & 0x1))
638 rdtp
->dynticks_nmi_nesting
++;
639 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
640 atomic_inc(&rdtp
->dynticks
);
641 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
642 smp_mb__after_atomic_inc(); /* See above. */
643 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
647 * rcu_nmi_exit - inform RCU of exit from NMI context
649 * If the CPU was idle with dynamic ticks active, and there is no
650 * irq handler running, this updates rdtp->dynticks_nmi to let the
651 * RCU grace-period handling know that the CPU is no longer active.
653 void rcu_nmi_exit(void)
655 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
657 if (rdtp
->dynticks_nmi_nesting
== 0 ||
658 --rdtp
->dynticks_nmi_nesting
!= 0)
660 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
661 smp_mb__before_atomic_inc(); /* See above. */
662 atomic_inc(&rdtp
->dynticks
);
663 smp_mb__after_atomic_inc(); /* Force delay to next write. */
664 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
668 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
670 * If the current CPU is in its idle loop and is neither in an interrupt
671 * or NMI handler, return true.
673 int rcu_is_cpu_idle(void)
678 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
682 EXPORT_SYMBOL(rcu_is_cpu_idle
);
684 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
687 * Is the current CPU online? Disable preemption to avoid false positives
688 * that could otherwise happen due to the current CPU number being sampled,
689 * this task being preempted, its old CPU being taken offline, resuming
690 * on some other CPU, then determining that its old CPU is now offline.
691 * It is OK to use RCU on an offline processor during initial boot, hence
692 * the check for rcu_scheduler_fully_active. Note also that it is OK
693 * for a CPU coming online to use RCU for one jiffy prior to marking itself
694 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
695 * offline to continue to use RCU for one jiffy after marking itself
696 * offline in the cpu_online_mask. This leniency is necessary given the
697 * non-atomic nature of the online and offline processing, for example,
698 * the fact that a CPU enters the scheduler after completing the CPU_DYING
701 * This is also why RCU internally marks CPUs online during the
702 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
704 * Disable checking if in an NMI handler because we cannot safely report
705 * errors from NMI handlers anyway.
707 bool rcu_lockdep_current_cpu_online(void)
709 struct rcu_data
*rdp
;
710 struct rcu_node
*rnp
;
716 rdp
= &__get_cpu_var(rcu_sched_data
);
718 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
719 !rcu_scheduler_fully_active
;
723 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
725 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
728 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
730 * If the current CPU is idle or running at a first-level (not nested)
731 * interrupt from idle, return true. The caller must have at least
732 * disabled preemption.
734 static int rcu_is_cpu_rrupt_from_idle(void)
736 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
740 * Snapshot the specified CPU's dynticks counter so that we can later
741 * credit them with an implicit quiescent state. Return 1 if this CPU
742 * is in dynticks idle mode, which is an extended quiescent state.
744 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
746 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
747 return (rdp
->dynticks_snap
& 0x1) == 0;
751 * Return true if the specified CPU has passed through a quiescent
752 * state by virtue of being in or having passed through an dynticks
753 * idle state since the last call to dyntick_save_progress_counter()
754 * for this same CPU, or by virtue of having been offline.
756 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
761 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
762 snap
= (unsigned int)rdp
->dynticks_snap
;
765 * If the CPU passed through or entered a dynticks idle phase with
766 * no active irq/NMI handlers, then we can safely pretend that the CPU
767 * already acknowledged the request to pass through a quiescent
768 * state. Either way, that CPU cannot possibly be in an RCU
769 * read-side critical section that started before the beginning
770 * of the current RCU grace period.
772 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
773 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
779 * Check for the CPU being offline, but only if the grace period
780 * is old enough. We don't need to worry about the CPU changing
781 * state: If we see it offline even once, it has been through a
784 * The reason for insisting that the grace period be at least
785 * one jiffy old is that CPUs that are not quite online and that
786 * have just gone offline can still execute RCU read-side critical
789 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
790 return 0; /* Grace period is not old enough. */
792 if (cpu_is_offline(rdp
->cpu
)) {
793 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
799 * There is a possibility that a CPU in adaptive-ticks state
800 * might run in the kernel with the scheduling-clock tick disabled
801 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
802 * force the CPU to restart the scheduling-clock tick in this
803 * CPU is in this state.
805 rcu_kick_nohz_cpu(rdp
->cpu
);
810 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
812 rsp
->gp_start
= jiffies
;
813 rsp
->jiffies_stall
= jiffies
+ rcu_jiffies_till_stall_check();
817 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
818 * for architectures that do not implement trigger_all_cpu_backtrace().
819 * The NMI-triggered stack traces are more accurate because they are
820 * printed by the target CPU.
822 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
826 struct rcu_node
*rnp
;
828 rcu_for_each_leaf_node(rsp
, rnp
) {
829 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
830 if (rnp
->qsmask
!= 0) {
831 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
832 if (rnp
->qsmask
& (1UL << cpu
))
833 dump_cpu_task(rnp
->grplo
+ cpu
);
835 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
839 static void print_other_cpu_stall(struct rcu_state
*rsp
)
845 struct rcu_node
*rnp
= rcu_get_root(rsp
);
848 /* Only let one CPU complain about others per time interval. */
850 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
851 delta
= jiffies
- rsp
->jiffies_stall
;
852 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
853 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
856 rsp
->jiffies_stall
= jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
857 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
860 * OK, time to rat on our buddy...
861 * See Documentation/RCU/stallwarn.txt for info on how to debug
862 * RCU CPU stall warnings.
864 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
866 print_cpu_stall_info_begin();
867 rcu_for_each_leaf_node(rsp
, rnp
) {
868 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
869 ndetected
+= rcu_print_task_stall(rnp
);
870 if (rnp
->qsmask
!= 0) {
871 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
872 if (rnp
->qsmask
& (1UL << cpu
)) {
873 print_cpu_stall_info(rsp
,
878 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
882 * Now rat on any tasks that got kicked up to the root rcu_node
883 * due to CPU offlining.
885 rnp
= rcu_get_root(rsp
);
886 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
887 ndetected
+= rcu_print_task_stall(rnp
);
888 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
890 print_cpu_stall_info_end();
891 for_each_possible_cpu(cpu
)
892 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
893 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
894 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
895 rsp
->gpnum
, rsp
->completed
, totqlen
);
897 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
898 else if (!trigger_all_cpu_backtrace())
899 rcu_dump_cpu_stacks(rsp
);
901 /* Complain about tasks blocking the grace period. */
903 rcu_print_detail_task_stall(rsp
);
905 force_quiescent_state(rsp
); /* Kick them all. */
908 static void print_cpu_stall(struct rcu_state
*rsp
)
912 struct rcu_node
*rnp
= rcu_get_root(rsp
);
916 * OK, time to rat on ourselves...
917 * See Documentation/RCU/stallwarn.txt for info on how to debug
918 * RCU CPU stall warnings.
920 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
921 print_cpu_stall_info_begin();
922 print_cpu_stall_info(rsp
, smp_processor_id());
923 print_cpu_stall_info_end();
924 for_each_possible_cpu(cpu
)
925 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
926 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
927 jiffies
- rsp
->gp_start
, rsp
->gpnum
, rsp
->completed
, totqlen
);
928 if (!trigger_all_cpu_backtrace())
931 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
932 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
933 rsp
->jiffies_stall
= jiffies
+
934 3 * rcu_jiffies_till_stall_check() + 3;
935 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
937 set_need_resched(); /* kick ourselves to get things going. */
940 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
944 struct rcu_node
*rnp
;
946 if (rcu_cpu_stall_suppress
)
948 j
= ACCESS_ONCE(jiffies
);
949 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
951 if (rcu_gp_in_progress(rsp
) &&
952 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
954 /* We haven't checked in, so go dump stack. */
955 print_cpu_stall(rsp
);
957 } else if (rcu_gp_in_progress(rsp
) &&
958 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
960 /* They had a few time units to dump stack, so complain. */
961 print_other_cpu_stall(rsp
);
966 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
968 * Set the stall-warning timeout way off into the future, thus preventing
969 * any RCU CPU stall-warning messages from appearing in the current set of
972 * The caller must disable hard irqs.
974 void rcu_cpu_stall_reset(void)
976 struct rcu_state
*rsp
;
978 for_each_rcu_flavor(rsp
)
979 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
983 * Update CPU-local rcu_data state to record the newly noticed grace period.
984 * This is used both when we started the grace period and when we notice
985 * that someone else started the grace period. The caller must hold the
986 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
987 * and must have irqs disabled.
989 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
991 if (rdp
->gpnum
!= rnp
->gpnum
) {
993 * If the current grace period is waiting for this CPU,
994 * set up to detect a quiescent state, otherwise don't
995 * go looking for one.
997 rdp
->gpnum
= rnp
->gpnum
;
998 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
999 rdp
->passed_quiesce
= 0;
1000 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1001 zero_cpu_stall_ticks(rdp
);
1005 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1007 unsigned long flags
;
1008 struct rcu_node
*rnp
;
1010 local_irq_save(flags
);
1012 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
1013 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1014 local_irq_restore(flags
);
1017 __note_new_gpnum(rsp
, rnp
, rdp
);
1018 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1022 * Did someone else start a new RCU grace period start since we last
1023 * checked? Update local state appropriately if so. Must be called
1024 * on the CPU corresponding to rdp.
1027 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1029 unsigned long flags
;
1032 local_irq_save(flags
);
1033 if (rdp
->gpnum
!= rsp
->gpnum
) {
1034 note_new_gpnum(rsp
, rdp
);
1037 local_irq_restore(flags
);
1042 * Initialize the specified rcu_data structure's callback list to empty.
1044 static void init_callback_list(struct rcu_data
*rdp
)
1048 rdp
->nxtlist
= NULL
;
1049 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1050 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1051 init_nocb_callback_list(rdp
);
1055 * Determine the value that ->completed will have at the end of the
1056 * next subsequent grace period. This is used to tag callbacks so that
1057 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1058 * been dyntick-idle for an extended period with callbacks under the
1059 * influence of RCU_FAST_NO_HZ.
1061 * The caller must hold rnp->lock with interrupts disabled.
1063 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1064 struct rcu_node
*rnp
)
1067 * If RCU is idle, we just wait for the next grace period.
1068 * But we can only be sure that RCU is idle if we are looking
1069 * at the root rcu_node structure -- otherwise, a new grace
1070 * period might have started, but just not yet gotten around
1071 * to initializing the current non-root rcu_node structure.
1073 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1074 return rnp
->completed
+ 1;
1077 * Otherwise, wait for a possible partial grace period and
1078 * then the subsequent full grace period.
1080 return rnp
->completed
+ 2;
1084 * If there is room, assign a ->completed number to any callbacks on
1085 * this CPU that have not already been assigned. Also accelerate any
1086 * callbacks that were previously assigned a ->completed number that has
1087 * since proven to be too conservative, which can happen if callbacks get
1088 * assigned a ->completed number while RCU is idle, but with reference to
1089 * a non-root rcu_node structure. This function is idempotent, so it does
1090 * not hurt to call it repeatedly.
1092 * The caller must hold rnp->lock with interrupts disabled.
1094 static void rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1095 struct rcu_data
*rdp
)
1100 /* If the CPU has no callbacks, nothing to do. */
1101 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1105 * Starting from the sublist containing the callbacks most
1106 * recently assigned a ->completed number and working down, find the
1107 * first sublist that is not assignable to an upcoming grace period.
1108 * Such a sublist has something in it (first two tests) and has
1109 * a ->completed number assigned that will complete sooner than
1110 * the ->completed number for newly arrived callbacks (last test).
1112 * The key point is that any later sublist can be assigned the
1113 * same ->completed number as the newly arrived callbacks, which
1114 * means that the callbacks in any of these later sublist can be
1115 * grouped into a single sublist, whether or not they have already
1116 * been assigned a ->completed number.
1118 c
= rcu_cbs_completed(rsp
, rnp
);
1119 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1120 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1121 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1125 * If there are no sublist for unassigned callbacks, leave.
1126 * At the same time, advance "i" one sublist, so that "i" will
1127 * index into the sublist where all the remaining callbacks should
1130 if (++i
>= RCU_NEXT_TAIL
)
1134 * Assign all subsequent callbacks' ->completed number to the next
1135 * full grace period and group them all in the sublist initially
1138 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1139 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1140 rdp
->nxtcompleted
[i
] = c
;
1143 /* Trace depending on how much we were able to accelerate. */
1144 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1145 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "AccWaitCB");
1147 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "AccReadyCB");
1151 * Move any callbacks whose grace period has completed to the
1152 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1153 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1154 * sublist. This function is idempotent, so it does not hurt to
1155 * invoke it repeatedly. As long as it is not invoked -too- often...
1157 * The caller must hold rnp->lock with interrupts disabled.
1159 static void rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1160 struct rcu_data
*rdp
)
1164 /* If the CPU has no callbacks, nothing to do. */
1165 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1169 * Find all callbacks whose ->completed numbers indicate that they
1170 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1172 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1173 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1175 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1177 /* Clean up any sublist tail pointers that were misordered above. */
1178 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1179 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1181 /* Copy down callbacks to fill in empty sublists. */
1182 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1183 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1185 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1186 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1189 /* Classify any remaining callbacks. */
1190 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1194 * Advance this CPU's callbacks, but only if the current grace period
1195 * has ended. This may be called only from the CPU to whom the rdp
1196 * belongs. In addition, the corresponding leaf rcu_node structure's
1197 * ->lock must be held by the caller, with irqs disabled.
1200 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1202 /* Did another grace period end? */
1203 if (rdp
->completed
== rnp
->completed
) {
1205 /* No, so just accelerate recent callbacks. */
1206 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1210 /* Advance callbacks. */
1211 rcu_advance_cbs(rsp
, rnp
, rdp
);
1213 /* Remember that we saw this grace-period completion. */
1214 rdp
->completed
= rnp
->completed
;
1215 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
1218 * If we were in an extended quiescent state, we may have
1219 * missed some grace periods that others CPUs handled on
1220 * our behalf. Catch up with this state to avoid noting
1221 * spurious new grace periods. If another grace period
1222 * has started, then rnp->gpnum will have advanced, so
1223 * we will detect this later on. Of course, any quiescent
1224 * states we found for the old GP are now invalid.
1226 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
)) {
1227 rdp
->gpnum
= rdp
->completed
;
1228 rdp
->passed_quiesce
= 0;
1232 * If RCU does not need a quiescent state from this CPU,
1233 * then make sure that this CPU doesn't go looking for one.
1235 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1236 rdp
->qs_pending
= 0;
1241 * Advance this CPU's callbacks, but only if the current grace period
1242 * has ended. This may be called only from the CPU to whom the rdp
1246 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1248 unsigned long flags
;
1249 struct rcu_node
*rnp
;
1251 local_irq_save(flags
);
1253 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1254 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1255 local_irq_restore(flags
);
1258 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1259 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1263 * Do per-CPU grace-period initialization for running CPU. The caller
1264 * must hold the lock of the leaf rcu_node structure corresponding to
1268 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1270 /* Prior grace period ended, so advance callbacks for current CPU. */
1271 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1273 /* Set state so that this CPU will detect the next quiescent state. */
1274 __note_new_gpnum(rsp
, rnp
, rdp
);
1278 * Initialize a new grace period.
1280 static int rcu_gp_init(struct rcu_state
*rsp
)
1282 struct rcu_data
*rdp
;
1283 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1285 raw_spin_lock_irq(&rnp
->lock
);
1286 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1288 if (rcu_gp_in_progress(rsp
)) {
1289 /* Grace period already in progress, don't start another. */
1290 raw_spin_unlock_irq(&rnp
->lock
);
1294 /* Advance to a new grace period and initialize state. */
1296 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1297 record_gp_stall_check_time(rsp
);
1298 raw_spin_unlock_irq(&rnp
->lock
);
1300 /* Exclude any concurrent CPU-hotplug operations. */
1301 mutex_lock(&rsp
->onoff_mutex
);
1304 * Set the quiescent-state-needed bits in all the rcu_node
1305 * structures for all currently online CPUs in breadth-first order,
1306 * starting from the root rcu_node structure, relying on the layout
1307 * of the tree within the rsp->node[] array. Note that other CPUs
1308 * will access only the leaves of the hierarchy, thus seeing that no
1309 * grace period is in progress, at least until the corresponding
1310 * leaf node has been initialized. In addition, we have excluded
1311 * CPU-hotplug operations.
1313 * The grace period cannot complete until the initialization
1314 * process finishes, because this kthread handles both.
1316 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1317 raw_spin_lock_irq(&rnp
->lock
);
1318 rdp
= this_cpu_ptr(rsp
->rda
);
1319 rcu_preempt_check_blocked_tasks(rnp
);
1320 rnp
->qsmask
= rnp
->qsmaskinit
;
1321 rnp
->gpnum
= rsp
->gpnum
;
1322 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1323 rnp
->completed
= rsp
->completed
;
1324 if (rnp
== rdp
->mynode
)
1325 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1326 rcu_preempt_boost_start_gp(rnp
);
1327 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1328 rnp
->level
, rnp
->grplo
,
1329 rnp
->grphi
, rnp
->qsmask
);
1330 raw_spin_unlock_irq(&rnp
->lock
);
1331 #ifdef CONFIG_PROVE_RCU_DELAY
1332 if ((random32() % (rcu_num_nodes
* 8)) == 0)
1333 schedule_timeout_uninterruptible(2);
1334 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1338 mutex_unlock(&rsp
->onoff_mutex
);
1343 * Do one round of quiescent-state forcing.
1345 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1347 int fqs_state
= fqs_state_in
;
1348 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1351 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1352 /* Collect dyntick-idle snapshots. */
1353 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1354 fqs_state
= RCU_FORCE_QS
;
1356 /* Handle dyntick-idle and offline CPUs. */
1357 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1359 /* Clear flag to prevent immediate re-entry. */
1360 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1361 raw_spin_lock_irq(&rnp
->lock
);
1362 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1363 raw_spin_unlock_irq(&rnp
->lock
);
1369 * Clean up after the old grace period.
1371 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1373 unsigned long gp_duration
;
1374 struct rcu_data
*rdp
;
1375 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1377 raw_spin_lock_irq(&rnp
->lock
);
1378 gp_duration
= jiffies
- rsp
->gp_start
;
1379 if (gp_duration
> rsp
->gp_max
)
1380 rsp
->gp_max
= gp_duration
;
1383 * We know the grace period is complete, but to everyone else
1384 * it appears to still be ongoing. But it is also the case
1385 * that to everyone else it looks like there is nothing that
1386 * they can do to advance the grace period. It is therefore
1387 * safe for us to drop the lock in order to mark the grace
1388 * period as completed in all of the rcu_node structures.
1390 raw_spin_unlock_irq(&rnp
->lock
);
1393 * Propagate new ->completed value to rcu_node structures so
1394 * that other CPUs don't have to wait until the start of the next
1395 * grace period to process their callbacks. This also avoids
1396 * some nasty RCU grace-period initialization races by forcing
1397 * the end of the current grace period to be completely recorded in
1398 * all of the rcu_node structures before the beginning of the next
1399 * grace period is recorded in any of the rcu_node structures.
1401 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1402 raw_spin_lock_irq(&rnp
->lock
);
1403 rnp
->completed
= rsp
->gpnum
;
1404 raw_spin_unlock_irq(&rnp
->lock
);
1407 rnp
= rcu_get_root(rsp
);
1408 raw_spin_lock_irq(&rnp
->lock
);
1410 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1411 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1412 rsp
->fqs_state
= RCU_GP_IDLE
;
1413 rdp
= this_cpu_ptr(rsp
->rda
);
1414 if (cpu_needs_another_gp(rsp
, rdp
))
1416 raw_spin_unlock_irq(&rnp
->lock
);
1420 * Body of kthread that handles grace periods.
1422 static int __noreturn
rcu_gp_kthread(void *arg
)
1427 struct rcu_state
*rsp
= arg
;
1428 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1432 /* Handle grace-period start. */
1434 wait_event_interruptible(rsp
->gp_wq
,
1437 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1441 flush_signals(current
);
1444 /* Handle quiescent-state forcing. */
1445 fqs_state
= RCU_SAVE_DYNTICK
;
1446 j
= jiffies_till_first_fqs
;
1449 jiffies_till_first_fqs
= HZ
;
1452 rsp
->jiffies_force_qs
= jiffies
+ j
;
1453 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1454 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1455 (!ACCESS_ONCE(rnp
->qsmask
) &&
1456 !rcu_preempt_blocked_readers_cgp(rnp
)),
1458 /* If grace period done, leave loop. */
1459 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1460 !rcu_preempt_blocked_readers_cgp(rnp
))
1462 /* If time for quiescent-state forcing, do it. */
1463 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1464 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1467 /* Deal with stray signal. */
1469 flush_signals(current
);
1471 j
= jiffies_till_next_fqs
;
1474 jiffies_till_next_fqs
= HZ
;
1477 jiffies_till_next_fqs
= 1;
1481 /* Handle grace-period end. */
1482 rcu_gp_cleanup(rsp
);
1487 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1488 * in preparation for detecting the next grace period. The caller must hold
1489 * the root node's ->lock, which is released before return. Hard irqs must
1492 * Note that it is legal for a dying CPU (which is marked as offline) to
1493 * invoke this function. This can happen when the dying CPU reports its
1497 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1498 __releases(rcu_get_root(rsp
)->lock
)
1500 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1501 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1503 if (!rsp
->gp_kthread
||
1504 !cpu_needs_another_gp(rsp
, rdp
)) {
1506 * Either we have not yet spawned the grace-period
1507 * task, this CPU does not need another grace period,
1508 * or a grace period is already in progress.
1509 * Either way, don't start a new grace period.
1511 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1516 * Because there is no grace period in progress right now,
1517 * any callbacks we have up to this point will be satisfied
1518 * by the next grace period. So this is a good place to
1519 * assign a grace period number to recently posted callbacks.
1521 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1523 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1524 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
1526 /* Ensure that CPU is aware of completion of last grace period. */
1527 rcu_process_gp_end(rsp
, rdp
);
1528 local_irq_restore(flags
);
1530 /* Wake up rcu_gp_kthread() to start the grace period. */
1531 wake_up(&rsp
->gp_wq
);
1535 * Report a full set of quiescent states to the specified rcu_state
1536 * data structure. This involves cleaning up after the prior grace
1537 * period and letting rcu_start_gp() start up the next grace period
1538 * if one is needed. Note that the caller must hold rnp->lock, as
1539 * required by rcu_start_gp(), which will release it.
1541 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1542 __releases(rcu_get_root(rsp
)->lock
)
1544 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1545 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1546 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1550 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1551 * Allows quiescent states for a group of CPUs to be reported at one go
1552 * to the specified rcu_node structure, though all the CPUs in the group
1553 * must be represented by the same rcu_node structure (which need not be
1554 * a leaf rcu_node structure, though it often will be). That structure's
1555 * lock must be held upon entry, and it is released before return.
1558 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1559 struct rcu_node
*rnp
, unsigned long flags
)
1560 __releases(rnp
->lock
)
1562 struct rcu_node
*rnp_c
;
1564 /* Walk up the rcu_node hierarchy. */
1566 if (!(rnp
->qsmask
& mask
)) {
1568 /* Our bit has already been cleared, so done. */
1569 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1572 rnp
->qsmask
&= ~mask
;
1573 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1574 mask
, rnp
->qsmask
, rnp
->level
,
1575 rnp
->grplo
, rnp
->grphi
,
1577 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1579 /* Other bits still set at this level, so done. */
1580 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1583 mask
= rnp
->grpmask
;
1584 if (rnp
->parent
== NULL
) {
1586 /* No more levels. Exit loop holding root lock. */
1590 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1593 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1594 WARN_ON_ONCE(rnp_c
->qsmask
);
1598 * Get here if we are the last CPU to pass through a quiescent
1599 * state for this grace period. Invoke rcu_report_qs_rsp()
1600 * to clean up and start the next grace period if one is needed.
1602 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1606 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1607 * structure. This must be either called from the specified CPU, or
1608 * called when the specified CPU is known to be offline (and when it is
1609 * also known that no other CPU is concurrently trying to help the offline
1610 * CPU). The lastcomp argument is used to make sure we are still in the
1611 * grace period of interest. We don't want to end the current grace period
1612 * based on quiescent states detected in an earlier grace period!
1615 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1617 unsigned long flags
;
1619 struct rcu_node
*rnp
;
1622 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1623 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1624 rnp
->completed
== rnp
->gpnum
) {
1627 * The grace period in which this quiescent state was
1628 * recorded has ended, so don't report it upwards.
1629 * We will instead need a new quiescent state that lies
1630 * within the current grace period.
1632 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1633 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1636 mask
= rdp
->grpmask
;
1637 if ((rnp
->qsmask
& mask
) == 0) {
1638 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1640 rdp
->qs_pending
= 0;
1643 * This GP can't end until cpu checks in, so all of our
1644 * callbacks can be processed during the next GP.
1646 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1648 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1653 * Check to see if there is a new grace period of which this CPU
1654 * is not yet aware, and if so, set up local rcu_data state for it.
1655 * Otherwise, see if this CPU has just passed through its first
1656 * quiescent state for this grace period, and record that fact if so.
1659 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1661 /* If there is now a new grace period, record and return. */
1662 if (check_for_new_grace_period(rsp
, rdp
))
1666 * Does this CPU still need to do its part for current grace period?
1667 * If no, return and let the other CPUs do their part as well.
1669 if (!rdp
->qs_pending
)
1673 * Was there a quiescent state since the beginning of the grace
1674 * period? If no, then exit and wait for the next call.
1676 if (!rdp
->passed_quiesce
)
1680 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1683 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1686 #ifdef CONFIG_HOTPLUG_CPU
1689 * Send the specified CPU's RCU callbacks to the orphanage. The
1690 * specified CPU must be offline, and the caller must hold the
1694 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1695 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1697 /* No-CBs CPUs do not have orphanable callbacks. */
1698 if (is_nocb_cpu(rdp
->cpu
))
1702 * Orphan the callbacks. First adjust the counts. This is safe
1703 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1704 * cannot be running now. Thus no memory barrier is required.
1706 if (rdp
->nxtlist
!= NULL
) {
1707 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1708 rsp
->qlen
+= rdp
->qlen
;
1709 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1711 ACCESS_ONCE(rdp
->qlen
) = 0;
1715 * Next, move those callbacks still needing a grace period to
1716 * the orphanage, where some other CPU will pick them up.
1717 * Some of the callbacks might have gone partway through a grace
1718 * period, but that is too bad. They get to start over because we
1719 * cannot assume that grace periods are synchronized across CPUs.
1720 * We don't bother updating the ->nxttail[] array yet, instead
1721 * we just reset the whole thing later on.
1723 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1724 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1725 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1726 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1730 * Then move the ready-to-invoke callbacks to the orphanage,
1731 * where some other CPU will pick them up. These will not be
1732 * required to pass though another grace period: They are done.
1734 if (rdp
->nxtlist
!= NULL
) {
1735 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1736 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1739 /* Finally, initialize the rcu_data structure's list to empty. */
1740 init_callback_list(rdp
);
1744 * Adopt the RCU callbacks from the specified rcu_state structure's
1745 * orphanage. The caller must hold the ->orphan_lock.
1747 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1750 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1752 /* No-CBs CPUs are handled specially. */
1753 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
))
1756 /* Do the accounting first. */
1757 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1758 rdp
->qlen
+= rsp
->qlen
;
1759 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1760 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1761 rcu_idle_count_callbacks_posted();
1766 * We do not need a memory barrier here because the only way we
1767 * can get here if there is an rcu_barrier() in flight is if
1768 * we are the task doing the rcu_barrier().
1771 /* First adopt the ready-to-invoke callbacks. */
1772 if (rsp
->orphan_donelist
!= NULL
) {
1773 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1774 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1775 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1776 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1777 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1778 rsp
->orphan_donelist
= NULL
;
1779 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1782 /* And then adopt the callbacks that still need a grace period. */
1783 if (rsp
->orphan_nxtlist
!= NULL
) {
1784 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1785 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1786 rsp
->orphan_nxtlist
= NULL
;
1787 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1792 * Trace the fact that this CPU is going offline.
1794 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1796 RCU_TRACE(unsigned long mask
);
1797 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1798 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1800 RCU_TRACE(mask
= rdp
->grpmask
);
1801 trace_rcu_grace_period(rsp
->name
,
1802 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1807 * The CPU has been completely removed, and some other CPU is reporting
1808 * this fact from process context. Do the remainder of the cleanup,
1809 * including orphaning the outgoing CPU's RCU callbacks, and also
1810 * adopting them. There can only be one CPU hotplug operation at a time,
1811 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1813 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1815 unsigned long flags
;
1817 int need_report
= 0;
1818 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1819 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1821 /* Adjust any no-longer-needed kthreads. */
1822 rcu_boost_kthread_setaffinity(rnp
, -1);
1824 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1826 /* Exclude any attempts to start a new grace period. */
1827 mutex_lock(&rsp
->onoff_mutex
);
1828 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
1830 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1831 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1832 rcu_adopt_orphan_cbs(rsp
);
1834 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1835 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1837 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1838 rnp
->qsmaskinit
&= ~mask
;
1839 if (rnp
->qsmaskinit
!= 0) {
1840 if (rnp
!= rdp
->mynode
)
1841 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1844 if (rnp
== rdp
->mynode
)
1845 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1847 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1848 mask
= rnp
->grpmask
;
1850 } while (rnp
!= NULL
);
1853 * We still hold the leaf rcu_node structure lock here, and
1854 * irqs are still disabled. The reason for this subterfuge is
1855 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1856 * held leads to deadlock.
1858 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
1860 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1861 rcu_report_unblock_qs_rnp(rnp
, flags
);
1863 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1864 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1865 rcu_report_exp_rnp(rsp
, rnp
, true);
1866 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1867 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1868 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1869 init_callback_list(rdp
);
1870 /* Disallow further callbacks on this CPU. */
1871 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1872 mutex_unlock(&rsp
->onoff_mutex
);
1875 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1877 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1881 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1885 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1888 * Invoke any RCU callbacks that have made it to the end of their grace
1889 * period. Thottle as specified by rdp->blimit.
1891 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1893 unsigned long flags
;
1894 struct rcu_head
*next
, *list
, **tail
;
1895 long bl
, count
, count_lazy
;
1898 /* If no callbacks are ready, just return. */
1899 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1900 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1901 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1902 need_resched(), is_idle_task(current
),
1903 rcu_is_callbacks_kthread());
1908 * Extract the list of ready callbacks, disabling to prevent
1909 * races with call_rcu() from interrupt handlers.
1911 local_irq_save(flags
);
1912 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1914 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1915 list
= rdp
->nxtlist
;
1916 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1917 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1918 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1919 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1920 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1921 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1922 local_irq_restore(flags
);
1924 /* Invoke callbacks. */
1925 count
= count_lazy
= 0;
1929 debug_rcu_head_unqueue(list
);
1930 if (__rcu_reclaim(rsp
->name
, list
))
1933 /* Stop only if limit reached and CPU has something to do. */
1934 if (++count
>= bl
&&
1936 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1940 local_irq_save(flags
);
1941 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1942 is_idle_task(current
),
1943 rcu_is_callbacks_kthread());
1945 /* Update count, and requeue any remaining callbacks. */
1947 *tail
= rdp
->nxtlist
;
1948 rdp
->nxtlist
= list
;
1949 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1950 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1951 rdp
->nxttail
[i
] = tail
;
1955 smp_mb(); /* List handling before counting for rcu_barrier(). */
1956 rdp
->qlen_lazy
-= count_lazy
;
1957 ACCESS_ONCE(rdp
->qlen
) -= count
;
1958 rdp
->n_cbs_invoked
+= count
;
1960 /* Reinstate batch limit if we have worked down the excess. */
1961 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1962 rdp
->blimit
= blimit
;
1964 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1965 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1966 rdp
->qlen_last_fqs_check
= 0;
1967 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1968 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1969 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1970 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
1972 local_irq_restore(flags
);
1974 /* Re-invoke RCU core processing if there are callbacks remaining. */
1975 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1980 * Check to see if this CPU is in a non-context-switch quiescent state
1981 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1982 * Also schedule RCU core processing.
1984 * This function must be called from hardirq context. It is normally
1985 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1986 * false, there is no point in invoking rcu_check_callbacks().
1988 void rcu_check_callbacks(int cpu
, int user
)
1990 trace_rcu_utilization("Start scheduler-tick");
1991 increment_cpu_stall_ticks();
1992 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1995 * Get here if this CPU took its interrupt from user
1996 * mode or from the idle loop, and if this is not a
1997 * nested interrupt. In this case, the CPU is in
1998 * a quiescent state, so note it.
2000 * No memory barrier is required here because both
2001 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2002 * variables that other CPUs neither access nor modify,
2003 * at least not while the corresponding CPU is online.
2009 } else if (!in_softirq()) {
2012 * Get here if this CPU did not take its interrupt from
2013 * softirq, in other words, if it is not interrupting
2014 * a rcu_bh read-side critical section. This is an _bh
2015 * critical section, so note it.
2020 rcu_preempt_check_callbacks(cpu
);
2021 if (rcu_pending(cpu
))
2023 trace_rcu_utilization("End scheduler-tick");
2027 * Scan the leaf rcu_node structures, processing dyntick state for any that
2028 * have not yet encountered a quiescent state, using the function specified.
2029 * Also initiate boosting for any threads blocked on the root rcu_node.
2031 * The caller must have suppressed start of new grace periods.
2033 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
2037 unsigned long flags
;
2039 struct rcu_node
*rnp
;
2041 rcu_for_each_leaf_node(rsp
, rnp
) {
2044 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2045 if (!rcu_gp_in_progress(rsp
)) {
2046 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2049 if (rnp
->qsmask
== 0) {
2050 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
2055 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2056 if ((rnp
->qsmask
& bit
) != 0 &&
2057 f(per_cpu_ptr(rsp
->rda
, cpu
)))
2062 /* rcu_report_qs_rnp() releases rnp->lock. */
2063 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
2066 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2068 rnp
= rcu_get_root(rsp
);
2069 if (rnp
->qsmask
== 0) {
2070 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2071 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
2076 * Force quiescent states on reluctant CPUs, and also detect which
2077 * CPUs are in dyntick-idle mode.
2079 static void force_quiescent_state(struct rcu_state
*rsp
)
2081 unsigned long flags
;
2083 struct rcu_node
*rnp
;
2084 struct rcu_node
*rnp_old
= NULL
;
2086 /* Funnel through hierarchy to reduce memory contention. */
2087 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
2088 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2089 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2090 !raw_spin_trylock(&rnp
->fqslock
);
2091 if (rnp_old
!= NULL
)
2092 raw_spin_unlock(&rnp_old
->fqslock
);
2094 rsp
->n_force_qs_lh
++;
2099 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2101 /* Reached the root of the rcu_node tree, acquire lock. */
2102 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2103 raw_spin_unlock(&rnp_old
->fqslock
);
2104 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2105 rsp
->n_force_qs_lh
++;
2106 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2107 return; /* Someone beat us to it. */
2109 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
2110 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2111 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
2115 * This does the RCU core processing work for the specified rcu_state
2116 * and rcu_data structures. This may be called only from the CPU to
2117 * whom the rdp belongs.
2120 __rcu_process_callbacks(struct rcu_state
*rsp
)
2122 unsigned long flags
;
2123 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2125 WARN_ON_ONCE(rdp
->beenonline
== 0);
2127 /* Handle the end of a grace period that some other CPU ended. */
2128 rcu_process_gp_end(rsp
, rdp
);
2130 /* Update RCU state based on any recent quiescent states. */
2131 rcu_check_quiescent_state(rsp
, rdp
);
2133 /* Does this CPU require a not-yet-started grace period? */
2134 local_irq_save(flags
);
2135 if (cpu_needs_another_gp(rsp
, rdp
)) {
2136 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2137 rcu_start_gp(rsp
, flags
); /* releases above lock */
2139 local_irq_restore(flags
);
2142 /* If there are callbacks ready, invoke them. */
2143 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2144 invoke_rcu_callbacks(rsp
, rdp
);
2148 * Do RCU core processing for the current CPU.
2150 static void rcu_process_callbacks(struct softirq_action
*unused
)
2152 struct rcu_state
*rsp
;
2154 if (cpu_is_offline(smp_processor_id()))
2156 trace_rcu_utilization("Start RCU core");
2157 for_each_rcu_flavor(rsp
)
2158 __rcu_process_callbacks(rsp
);
2159 trace_rcu_utilization("End RCU core");
2163 * Schedule RCU callback invocation. If the specified type of RCU
2164 * does not support RCU priority boosting, just do a direct call,
2165 * otherwise wake up the per-CPU kernel kthread. Note that because we
2166 * are running on the current CPU with interrupts disabled, the
2167 * rcu_cpu_kthread_task cannot disappear out from under us.
2169 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2171 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2173 if (likely(!rsp
->boost
)) {
2174 rcu_do_batch(rsp
, rdp
);
2177 invoke_rcu_callbacks_kthread();
2180 static void invoke_rcu_core(void)
2182 raise_softirq(RCU_SOFTIRQ
);
2186 * Handle any core-RCU processing required by a call_rcu() invocation.
2188 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2189 struct rcu_head
*head
, unsigned long flags
)
2192 * If called from an extended quiescent state, invoke the RCU
2193 * core in order to force a re-evaluation of RCU's idleness.
2195 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2198 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2199 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2203 * Force the grace period if too many callbacks or too long waiting.
2204 * Enforce hysteresis, and don't invoke force_quiescent_state()
2205 * if some other CPU has recently done so. Also, don't bother
2206 * invoking force_quiescent_state() if the newly enqueued callback
2207 * is the only one waiting for a grace period to complete.
2209 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2211 /* Are we ignoring a completed grace period? */
2212 rcu_process_gp_end(rsp
, rdp
);
2213 check_for_new_grace_period(rsp
, rdp
);
2215 /* Start a new grace period if one not already started. */
2216 if (!rcu_gp_in_progress(rsp
)) {
2217 unsigned long nestflag
;
2218 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2220 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
2221 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
2223 /* Give the grace period a kick. */
2224 rdp
->blimit
= LONG_MAX
;
2225 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2226 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2227 force_quiescent_state(rsp
);
2228 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2229 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2235 * Helper function for call_rcu() and friends. The cpu argument will
2236 * normally be -1, indicating "currently running CPU". It may specify
2237 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2238 * is expected to specify a CPU.
2241 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2242 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2244 unsigned long flags
;
2245 struct rcu_data
*rdp
;
2247 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2248 debug_rcu_head_queue(head
);
2253 * Opportunistically note grace-period endings and beginnings.
2254 * Note that we might see a beginning right after we see an
2255 * end, but never vice versa, since this CPU has to pass through
2256 * a quiescent state betweentimes.
2258 local_irq_save(flags
);
2259 rdp
= this_cpu_ptr(rsp
->rda
);
2261 /* Add the callback to our list. */
2262 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2266 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2267 offline
= !__call_rcu_nocb(rdp
, head
, lazy
);
2268 WARN_ON_ONCE(offline
);
2269 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2270 local_irq_restore(flags
);
2273 ACCESS_ONCE(rdp
->qlen
)++;
2277 rcu_idle_count_callbacks_posted();
2278 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2279 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2280 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2282 if (__is_kfree_rcu_offset((unsigned long)func
))
2283 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2284 rdp
->qlen_lazy
, rdp
->qlen
);
2286 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2288 /* Go handle any RCU core processing required. */
2289 __call_rcu_core(rsp
, rdp
, head
, flags
);
2290 local_irq_restore(flags
);
2294 * Queue an RCU-sched callback for invocation after a grace period.
2296 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2298 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2300 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2303 * Queue an RCU callback for invocation after a quicker grace period.
2305 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2307 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2309 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2312 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2313 * any blocking grace-period wait automatically implies a grace period
2314 * if there is only one CPU online at any point time during execution
2315 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2316 * occasionally incorrectly indicate that there are multiple CPUs online
2317 * when there was in fact only one the whole time, as this just adds
2318 * some overhead: RCU still operates correctly.
2320 static inline int rcu_blocking_is_gp(void)
2324 might_sleep(); /* Check for RCU read-side critical section. */
2326 ret
= num_online_cpus() <= 1;
2332 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2334 * Control will return to the caller some time after a full rcu-sched
2335 * grace period has elapsed, in other words after all currently executing
2336 * rcu-sched read-side critical sections have completed. These read-side
2337 * critical sections are delimited by rcu_read_lock_sched() and
2338 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2339 * local_irq_disable(), and so on may be used in place of
2340 * rcu_read_lock_sched().
2342 * This means that all preempt_disable code sequences, including NMI and
2343 * non-threaded hardware-interrupt handlers, in progress on entry will
2344 * have completed before this primitive returns. However, this does not
2345 * guarantee that softirq handlers will have completed, since in some
2346 * kernels, these handlers can run in process context, and can block.
2348 * Note that this guarantee implies further memory-ordering guarantees.
2349 * On systems with more than one CPU, when synchronize_sched() returns,
2350 * each CPU is guaranteed to have executed a full memory barrier since the
2351 * end of its last RCU-sched read-side critical section whose beginning
2352 * preceded the call to synchronize_sched(). In addition, each CPU having
2353 * an RCU read-side critical section that extends beyond the return from
2354 * synchronize_sched() is guaranteed to have executed a full memory barrier
2355 * after the beginning of synchronize_sched() and before the beginning of
2356 * that RCU read-side critical section. Note that these guarantees include
2357 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2358 * that are executing in the kernel.
2360 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2361 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2362 * to have executed a full memory barrier during the execution of
2363 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2364 * again only if the system has more than one CPU).
2366 * This primitive provides the guarantees made by the (now removed)
2367 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2368 * guarantees that rcu_read_lock() sections will have completed.
2369 * In "classic RCU", these two guarantees happen to be one and
2370 * the same, but can differ in realtime RCU implementations.
2372 void synchronize_sched(void)
2374 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2375 !lock_is_held(&rcu_lock_map
) &&
2376 !lock_is_held(&rcu_sched_lock_map
),
2377 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2378 if (rcu_blocking_is_gp())
2381 synchronize_sched_expedited();
2383 wait_rcu_gp(call_rcu_sched
);
2385 EXPORT_SYMBOL_GPL(synchronize_sched
);
2388 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2390 * Control will return to the caller some time after a full rcu_bh grace
2391 * period has elapsed, in other words after all currently executing rcu_bh
2392 * read-side critical sections have completed. RCU read-side critical
2393 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2394 * and may be nested.
2396 * See the description of synchronize_sched() for more detailed information
2397 * on memory ordering guarantees.
2399 void synchronize_rcu_bh(void)
2401 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2402 !lock_is_held(&rcu_lock_map
) &&
2403 !lock_is_held(&rcu_sched_lock_map
),
2404 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2405 if (rcu_blocking_is_gp())
2408 synchronize_rcu_bh_expedited();
2410 wait_rcu_gp(call_rcu_bh
);
2412 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2414 static int synchronize_sched_expedited_cpu_stop(void *data
)
2417 * There must be a full memory barrier on each affected CPU
2418 * between the time that try_stop_cpus() is called and the
2419 * time that it returns.
2421 * In the current initial implementation of cpu_stop, the
2422 * above condition is already met when the control reaches
2423 * this point and the following smp_mb() is not strictly
2424 * necessary. Do smp_mb() anyway for documentation and
2425 * robustness against future implementation changes.
2427 smp_mb(); /* See above comment block. */
2432 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2434 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2435 * approach to force the grace period to end quickly. This consumes
2436 * significant time on all CPUs and is unfriendly to real-time workloads,
2437 * so is thus not recommended for any sort of common-case code. In fact,
2438 * if you are using synchronize_sched_expedited() in a loop, please
2439 * restructure your code to batch your updates, and then use a single
2440 * synchronize_sched() instead.
2442 * Note that it is illegal to call this function while holding any lock
2443 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2444 * to call this function from a CPU-hotplug notifier. Failing to observe
2445 * these restriction will result in deadlock.
2447 * This implementation can be thought of as an application of ticket
2448 * locking to RCU, with sync_sched_expedited_started and
2449 * sync_sched_expedited_done taking on the roles of the halves
2450 * of the ticket-lock word. Each task atomically increments
2451 * sync_sched_expedited_started upon entry, snapshotting the old value,
2452 * then attempts to stop all the CPUs. If this succeeds, then each
2453 * CPU will have executed a context switch, resulting in an RCU-sched
2454 * grace period. We are then done, so we use atomic_cmpxchg() to
2455 * update sync_sched_expedited_done to match our snapshot -- but
2456 * only if someone else has not already advanced past our snapshot.
2458 * On the other hand, if try_stop_cpus() fails, we check the value
2459 * of sync_sched_expedited_done. If it has advanced past our
2460 * initial snapshot, then someone else must have forced a grace period
2461 * some time after we took our snapshot. In this case, our work is
2462 * done for us, and we can simply return. Otherwise, we try again,
2463 * but keep our initial snapshot for purposes of checking for someone
2464 * doing our work for us.
2466 * If we fail too many times in a row, we fall back to synchronize_sched().
2468 void synchronize_sched_expedited(void)
2470 long firstsnap
, s
, snap
;
2472 struct rcu_state
*rsp
= &rcu_sched_state
;
2475 * If we are in danger of counter wrap, just do synchronize_sched().
2476 * By allowing sync_sched_expedited_started to advance no more than
2477 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2478 * that more than 3.5 billion CPUs would be required to force a
2479 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2480 * course be required on a 64-bit system.
2482 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
2483 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
2485 synchronize_sched();
2486 atomic_long_inc(&rsp
->expedited_wrap
);
2491 * Take a ticket. Note that atomic_inc_return() implies a
2492 * full memory barrier.
2494 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
2497 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2500 * Each pass through the following loop attempts to force a
2501 * context switch on each CPU.
2503 while (try_stop_cpus(cpu_online_mask
,
2504 synchronize_sched_expedited_cpu_stop
,
2507 atomic_long_inc(&rsp
->expedited_tryfail
);
2509 /* Check to see if someone else did our work for us. */
2510 s
= atomic_long_read(&rsp
->expedited_done
);
2511 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2512 /* ensure test happens before caller kfree */
2513 smp_mb__before_atomic_inc(); /* ^^^ */
2514 atomic_long_inc(&rsp
->expedited_workdone1
);
2518 /* No joy, try again later. Or just synchronize_sched(). */
2519 if (trycount
++ < 10) {
2520 udelay(trycount
* num_online_cpus());
2522 wait_rcu_gp(call_rcu_sched
);
2523 atomic_long_inc(&rsp
->expedited_normal
);
2527 /* Recheck to see if someone else did our work for us. */
2528 s
= atomic_long_read(&rsp
->expedited_done
);
2529 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2530 /* ensure test happens before caller kfree */
2531 smp_mb__before_atomic_inc(); /* ^^^ */
2532 atomic_long_inc(&rsp
->expedited_workdone2
);
2537 * Refetching sync_sched_expedited_started allows later
2538 * callers to piggyback on our grace period. We retry
2539 * after they started, so our grace period works for them,
2540 * and they started after our first try, so their grace
2541 * period works for us.
2544 snap
= atomic_long_read(&rsp
->expedited_start
);
2545 smp_mb(); /* ensure read is before try_stop_cpus(). */
2547 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
2550 * Everyone up to our most recent fetch is covered by our grace
2551 * period. Update the counter, but only if our work is still
2552 * relevant -- which it won't be if someone who started later
2553 * than we did already did their update.
2556 atomic_long_inc(&rsp
->expedited_done_tries
);
2557 s
= atomic_long_read(&rsp
->expedited_done
);
2558 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
2559 /* ensure test happens before caller kfree */
2560 smp_mb__before_atomic_inc(); /* ^^^ */
2561 atomic_long_inc(&rsp
->expedited_done_lost
);
2564 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
2565 atomic_long_inc(&rsp
->expedited_done_exit
);
2569 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2572 * Check to see if there is any immediate RCU-related work to be done
2573 * by the current CPU, for the specified type of RCU, returning 1 if so.
2574 * The checks are in order of increasing expense: checks that can be
2575 * carried out against CPU-local state are performed first. However,
2576 * we must check for CPU stalls first, else we might not get a chance.
2578 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2580 struct rcu_node
*rnp
= rdp
->mynode
;
2582 rdp
->n_rcu_pending
++;
2584 /* Check for CPU stalls, if enabled. */
2585 check_cpu_stall(rsp
, rdp
);
2587 /* Is the RCU core waiting for a quiescent state from this CPU? */
2588 if (rcu_scheduler_fully_active
&&
2589 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2590 rdp
->n_rp_qs_pending
++;
2591 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2592 rdp
->n_rp_report_qs
++;
2596 /* Does this CPU have callbacks ready to invoke? */
2597 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2598 rdp
->n_rp_cb_ready
++;
2602 /* Has RCU gone idle with this CPU needing another grace period? */
2603 if (cpu_needs_another_gp(rsp
, rdp
)) {
2604 rdp
->n_rp_cpu_needs_gp
++;
2608 /* Has another RCU grace period completed? */
2609 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2610 rdp
->n_rp_gp_completed
++;
2614 /* Has a new RCU grace period started? */
2615 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2616 rdp
->n_rp_gp_started
++;
2621 rdp
->n_rp_need_nothing
++;
2626 * Check to see if there is any immediate RCU-related work to be done
2627 * by the current CPU, returning 1 if so. This function is part of the
2628 * RCU implementation; it is -not- an exported member of the RCU API.
2630 static int rcu_pending(int cpu
)
2632 struct rcu_state
*rsp
;
2634 for_each_rcu_flavor(rsp
)
2635 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2641 * Check to see if any future RCU-related work will need to be done
2642 * by the current CPU, even if none need be done immediately, returning
2645 static int rcu_cpu_has_callbacks(int cpu
)
2647 struct rcu_state
*rsp
;
2649 /* RCU callbacks either ready or pending? */
2650 for_each_rcu_flavor(rsp
)
2651 if (per_cpu_ptr(rsp
->rda
, cpu
)->nxtlist
)
2657 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2658 * the compiler is expected to optimize this away.
2660 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2661 int cpu
, unsigned long done
)
2663 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2664 atomic_read(&rsp
->barrier_cpu_count
), done
);
2668 * RCU callback function for _rcu_barrier(). If we are last, wake
2669 * up the task executing _rcu_barrier().
2671 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2673 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2674 struct rcu_state
*rsp
= rdp
->rsp
;
2676 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2677 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2678 complete(&rsp
->barrier_completion
);
2680 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2685 * Called with preemption disabled, and from cross-cpu IRQ context.
2687 static void rcu_barrier_func(void *type
)
2689 struct rcu_state
*rsp
= type
;
2690 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2692 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2693 atomic_inc(&rsp
->barrier_cpu_count
);
2694 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2698 * Orchestrate the specified type of RCU barrier, waiting for all
2699 * RCU callbacks of the specified type to complete.
2701 static void _rcu_barrier(struct rcu_state
*rsp
)
2704 struct rcu_data
*rdp
;
2705 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2706 unsigned long snap_done
;
2708 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2710 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2711 mutex_lock(&rsp
->barrier_mutex
);
2714 * Ensure that all prior references, including to ->n_barrier_done,
2715 * are ordered before the _rcu_barrier() machinery.
2717 smp_mb(); /* See above block comment. */
2720 * Recheck ->n_barrier_done to see if others did our work for us.
2721 * This means checking ->n_barrier_done for an even-to-odd-to-even
2722 * transition. The "if" expression below therefore rounds the old
2723 * value up to the next even number and adds two before comparing.
2725 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2726 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2727 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2728 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2729 smp_mb(); /* caller's subsequent code after above check. */
2730 mutex_unlock(&rsp
->barrier_mutex
);
2735 * Increment ->n_barrier_done to avoid duplicate work. Use
2736 * ACCESS_ONCE() to prevent the compiler from speculating
2737 * the increment to precede the early-exit check.
2739 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2740 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2741 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2742 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2745 * Initialize the count to one rather than to zero in order to
2746 * avoid a too-soon return to zero in case of a short grace period
2747 * (or preemption of this task). Exclude CPU-hotplug operations
2748 * to ensure that no offline CPU has callbacks queued.
2750 init_completion(&rsp
->barrier_completion
);
2751 atomic_set(&rsp
->barrier_cpu_count
, 1);
2755 * Force each CPU with callbacks to register a new callback.
2756 * When that callback is invoked, we will know that all of the
2757 * corresponding CPU's preceding callbacks have been invoked.
2759 for_each_possible_cpu(cpu
) {
2760 if (!cpu_online(cpu
) && !is_nocb_cpu(cpu
))
2762 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2763 if (is_nocb_cpu(cpu
)) {
2764 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
2765 rsp
->n_barrier_done
);
2766 atomic_inc(&rsp
->barrier_cpu_count
);
2767 __call_rcu(&rdp
->barrier_head
, rcu_barrier_callback
,
2769 } else if (ACCESS_ONCE(rdp
->qlen
)) {
2770 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2771 rsp
->n_barrier_done
);
2772 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2774 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2775 rsp
->n_barrier_done
);
2781 * Now that we have an rcu_barrier_callback() callback on each
2782 * CPU, and thus each counted, remove the initial count.
2784 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2785 complete(&rsp
->barrier_completion
);
2787 /* Increment ->n_barrier_done to prevent duplicate work. */
2788 smp_mb(); /* Keep increment after above mechanism. */
2789 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2790 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2791 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2792 smp_mb(); /* Keep increment before caller's subsequent code. */
2794 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2795 wait_for_completion(&rsp
->barrier_completion
);
2797 /* Other rcu_barrier() invocations can now safely proceed. */
2798 mutex_unlock(&rsp
->barrier_mutex
);
2802 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2804 void rcu_barrier_bh(void)
2806 _rcu_barrier(&rcu_bh_state
);
2808 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2811 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2813 void rcu_barrier_sched(void)
2815 _rcu_barrier(&rcu_sched_state
);
2817 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2820 * Do boot-time initialization of a CPU's per-CPU RCU data.
2823 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2825 unsigned long flags
;
2826 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2827 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2829 /* Set up local state, ensuring consistent view of global state. */
2830 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2831 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2832 init_callback_list(rdp
);
2834 ACCESS_ONCE(rdp
->qlen
) = 0;
2835 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2836 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2837 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2840 rcu_boot_init_nocb_percpu_data(rdp
);
2841 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2845 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2846 * offline event can be happening at a given time. Note also that we
2847 * can accept some slop in the rsp->completed access due to the fact
2848 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2850 static void __cpuinit
2851 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2853 unsigned long flags
;
2855 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2856 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2858 /* Exclude new grace periods. */
2859 mutex_lock(&rsp
->onoff_mutex
);
2861 /* Set up local state, ensuring consistent view of global state. */
2862 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2863 rdp
->beenonline
= 1; /* We have now been online. */
2864 rdp
->preemptible
= preemptible
;
2865 rdp
->qlen_last_fqs_check
= 0;
2866 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2867 rdp
->blimit
= blimit
;
2868 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
2869 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2870 atomic_set(&rdp
->dynticks
->dynticks
,
2871 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2872 rcu_prepare_for_idle_init(cpu
);
2873 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2875 /* Add CPU to rcu_node bitmasks. */
2877 mask
= rdp
->grpmask
;
2879 /* Exclude any attempts to start a new GP on small systems. */
2880 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2881 rnp
->qsmaskinit
|= mask
;
2882 mask
= rnp
->grpmask
;
2883 if (rnp
== rdp
->mynode
) {
2885 * If there is a grace period in progress, we will
2886 * set up to wait for it next time we run the
2889 rdp
->gpnum
= rnp
->completed
;
2890 rdp
->completed
= rnp
->completed
;
2891 rdp
->passed_quiesce
= 0;
2892 rdp
->qs_pending
= 0;
2893 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2895 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2897 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2898 local_irq_restore(flags
);
2900 mutex_unlock(&rsp
->onoff_mutex
);
2903 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2905 struct rcu_state
*rsp
;
2907 for_each_rcu_flavor(rsp
)
2908 rcu_init_percpu_data(cpu
, rsp
,
2909 strcmp(rsp
->name
, "rcu_preempt") == 0);
2913 * Handle CPU online/offline notification events.
2915 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2916 unsigned long action
, void *hcpu
)
2918 long cpu
= (long)hcpu
;
2919 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2920 struct rcu_node
*rnp
= rdp
->mynode
;
2921 struct rcu_state
*rsp
;
2922 int ret
= NOTIFY_OK
;
2924 trace_rcu_utilization("Start CPU hotplug");
2926 case CPU_UP_PREPARE
:
2927 case CPU_UP_PREPARE_FROZEN
:
2928 rcu_prepare_cpu(cpu
);
2929 rcu_prepare_kthreads(cpu
);
2932 case CPU_DOWN_FAILED
:
2933 rcu_boost_kthread_setaffinity(rnp
, -1);
2935 case CPU_DOWN_PREPARE
:
2936 if (nocb_cpu_expendable(cpu
))
2937 rcu_boost_kthread_setaffinity(rnp
, cpu
);
2942 case CPU_DYING_FROZEN
:
2944 * The whole machine is "stopped" except this CPU, so we can
2945 * touch any data without introducing corruption. We send the
2946 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2948 for_each_rcu_flavor(rsp
)
2949 rcu_cleanup_dying_cpu(rsp
);
2950 rcu_cleanup_after_idle(cpu
);
2953 case CPU_DEAD_FROZEN
:
2954 case CPU_UP_CANCELED
:
2955 case CPU_UP_CANCELED_FROZEN
:
2956 for_each_rcu_flavor(rsp
)
2957 rcu_cleanup_dead_cpu(cpu
, rsp
);
2962 trace_rcu_utilization("End CPU hotplug");
2967 * Spawn the kthread that handles this RCU flavor's grace periods.
2969 static int __init
rcu_spawn_gp_kthread(void)
2971 unsigned long flags
;
2972 struct rcu_node
*rnp
;
2973 struct rcu_state
*rsp
;
2974 struct task_struct
*t
;
2976 for_each_rcu_flavor(rsp
) {
2977 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
2979 rnp
= rcu_get_root(rsp
);
2980 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2981 rsp
->gp_kthread
= t
;
2982 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2983 rcu_spawn_nocb_kthreads(rsp
);
2987 early_initcall(rcu_spawn_gp_kthread
);
2990 * This function is invoked towards the end of the scheduler's initialization
2991 * process. Before this is called, the idle task might contain
2992 * RCU read-side critical sections (during which time, this idle
2993 * task is booting the system). After this function is called, the
2994 * idle tasks are prohibited from containing RCU read-side critical
2995 * sections. This function also enables RCU lockdep checking.
2997 void rcu_scheduler_starting(void)
2999 WARN_ON(num_online_cpus() != 1);
3000 WARN_ON(nr_context_switches() > 0);
3001 rcu_scheduler_active
= 1;
3005 * Compute the per-level fanout, either using the exact fanout specified
3006 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3008 #ifdef CONFIG_RCU_FANOUT_EXACT
3009 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3013 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
3014 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3015 rsp
->levelspread
[0] = rcu_fanout_leaf
;
3017 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3018 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3025 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3026 ccur
= rsp
->levelcnt
[i
];
3027 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3031 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3034 * Helper function for rcu_init() that initializes one rcu_state structure.
3036 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3037 struct rcu_data __percpu
*rda
)
3039 static char *buf
[] = { "rcu_node_0",
3042 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3043 static char *fqs
[] = { "rcu_node_fqs_0",
3046 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3050 struct rcu_node
*rnp
;
3052 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3054 /* Silence gcc 4.8 warning about array index out of range. */
3055 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3056 panic("rcu_init_one: rcu_num_lvls overflow");
3058 /* Initialize the level-tracking arrays. */
3060 for (i
= 0; i
< rcu_num_lvls
; i
++)
3061 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3062 for (i
= 1; i
< rcu_num_lvls
; i
++)
3063 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3064 rcu_init_levelspread(rsp
);
3066 /* Initialize the elements themselves, starting from the leaves. */
3068 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3069 cpustride
*= rsp
->levelspread
[i
];
3070 rnp
= rsp
->level
[i
];
3071 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3072 raw_spin_lock_init(&rnp
->lock
);
3073 lockdep_set_class_and_name(&rnp
->lock
,
3074 &rcu_node_class
[i
], buf
[i
]);
3075 raw_spin_lock_init(&rnp
->fqslock
);
3076 lockdep_set_class_and_name(&rnp
->fqslock
,
3077 &rcu_fqs_class
[i
], fqs
[i
]);
3078 rnp
->gpnum
= rsp
->gpnum
;
3079 rnp
->completed
= rsp
->completed
;
3081 rnp
->qsmaskinit
= 0;
3082 rnp
->grplo
= j
* cpustride
;
3083 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
3084 if (rnp
->grphi
>= NR_CPUS
)
3085 rnp
->grphi
= NR_CPUS
- 1;
3091 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
3092 rnp
->grpmask
= 1UL << rnp
->grpnum
;
3093 rnp
->parent
= rsp
->level
[i
- 1] +
3094 j
/ rsp
->levelspread
[i
- 1];
3097 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3102 init_waitqueue_head(&rsp
->gp_wq
);
3103 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3104 for_each_possible_cpu(i
) {
3105 while (i
> rnp
->grphi
)
3107 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3108 rcu_boot_init_percpu_data(i
, rsp
);
3110 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3114 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3115 * replace the definitions in rcutree.h because those are needed to size
3116 * the ->node array in the rcu_state structure.
3118 static void __init
rcu_init_geometry(void)
3123 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3125 /* If the compile-time values are accurate, just leave. */
3126 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3127 nr_cpu_ids
== NR_CPUS
)
3131 * Compute number of nodes that can be handled an rcu_node tree
3132 * with the given number of levels. Setting rcu_capacity[0] makes
3133 * some of the arithmetic easier.
3135 rcu_capacity
[0] = 1;
3136 rcu_capacity
[1] = rcu_fanout_leaf
;
3137 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3138 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3141 * The boot-time rcu_fanout_leaf parameter is only permitted
3142 * to increase the leaf-level fanout, not decrease it. Of course,
3143 * the leaf-level fanout cannot exceed the number of bits in
3144 * the rcu_node masks. Finally, the tree must be able to accommodate
3145 * the configured number of CPUs. Complain and fall back to the
3146 * compile-time values if these limits are exceeded.
3148 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3149 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3150 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3155 /* Calculate the number of rcu_nodes at each level of the tree. */
3156 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3157 if (n
<= rcu_capacity
[i
]) {
3158 for (j
= 0; j
<= i
; j
++)
3160 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3162 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3167 /* Calculate the total number of rcu_node structures. */
3169 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3170 rcu_num_nodes
+= num_rcu_lvl
[i
];
3174 void __init
rcu_init(void)
3178 rcu_bootup_announce();
3179 rcu_init_geometry();
3180 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3181 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3182 __rcu_init_preempt();
3184 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3187 * We don't need protection against CPU-hotplug here because
3188 * this is called early in boot, before either interrupts
3189 * or the scheduler are operational.
3191 cpu_notifier(rcu_cpu_notify
, 0);
3192 for_each_online_cpu(cpu
)
3193 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3196 #include "rcutree_plugin.h"