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 * optimized 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),
210 #if defined(CONFIG_RCU_USER_QS) && !defined(CONFIG_RCU_USER_QS_FORCE)
211 .ignore_user_qs
= true,
215 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
216 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
217 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
219 module_param(blimit
, long, 0444);
220 module_param(qhimark
, long, 0444);
221 module_param(qlowmark
, long, 0444);
223 int rcu_cpu_stall_suppress __read_mostly
; /* 1 = suppress stall warnings. */
224 int rcu_cpu_stall_timeout __read_mostly
= CONFIG_RCU_CPU_STALL_TIMEOUT
;
226 module_param(rcu_cpu_stall_suppress
, int, 0644);
227 module_param(rcu_cpu_stall_timeout
, int, 0644);
229 static ulong jiffies_till_first_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
230 static ulong jiffies_till_next_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
232 module_param(jiffies_till_first_fqs
, ulong
, 0644);
233 module_param(jiffies_till_next_fqs
, ulong
, 0644);
235 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*));
236 static void force_quiescent_state(struct rcu_state
*rsp
);
237 static int rcu_pending(int cpu
);
240 * Return the number of RCU-sched batches processed thus far for debug & stats.
242 long rcu_batches_completed_sched(void)
244 return rcu_sched_state
.completed
;
246 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
249 * Return the number of RCU BH batches processed thus far for debug & stats.
251 long rcu_batches_completed_bh(void)
253 return rcu_bh_state
.completed
;
255 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
258 * Force a quiescent state for RCU BH.
260 void rcu_bh_force_quiescent_state(void)
262 force_quiescent_state(&rcu_bh_state
);
264 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
267 * Record the number of times rcutorture tests have been initiated and
268 * terminated. This information allows the debugfs tracing stats to be
269 * correlated to the rcutorture messages, even when the rcutorture module
270 * is being repeatedly loaded and unloaded. In other words, we cannot
271 * store this state in rcutorture itself.
273 void rcutorture_record_test_transition(void)
275 rcutorture_testseq
++;
276 rcutorture_vernum
= 0;
278 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
281 * Record the number of writer passes through the current rcutorture test.
282 * This is also used to correlate debugfs tracing stats with the rcutorture
285 void rcutorture_record_progress(unsigned long vernum
)
289 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
292 * Force a quiescent state for RCU-sched.
294 void rcu_sched_force_quiescent_state(void)
296 force_quiescent_state(&rcu_sched_state
);
298 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
301 * Does the CPU have callbacks ready to be invoked?
304 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
306 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
307 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
311 * Does the current CPU require a yet-as-unscheduled grace period?
314 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
316 struct rcu_head
**ntp
;
318 ntp
= rdp
->nxttail
[RCU_DONE_TAIL
+
319 (ACCESS_ONCE(rsp
->completed
) != rdp
->completed
)];
320 return rdp
->nxttail
[RCU_DONE_TAIL
] && ntp
&& *ntp
&&
321 !rcu_gp_in_progress(rsp
);
325 * Return the root node of the specified rcu_state structure.
327 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
329 return &rsp
->node
[0];
333 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
335 * If the new value of the ->dynticks_nesting counter now is zero,
336 * we really have entered idle, and must do the appropriate accounting.
337 * The caller must have disabled interrupts.
339 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
342 trace_rcu_dyntick("Start", oldval
, 0);
343 if (!user
&& !is_idle_task(current
)) {
344 struct task_struct
*idle
= idle_task(smp_processor_id());
346 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
347 ftrace_dump(DUMP_ORIG
);
348 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
349 current
->pid
, current
->comm
,
350 idle
->pid
, idle
->comm
); /* must be idle task! */
352 rcu_prepare_for_idle(smp_processor_id());
353 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
354 smp_mb__before_atomic_inc(); /* See above. */
355 atomic_inc(&rdtp
->dynticks
);
356 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
357 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
360 * It is illegal to enter an extended quiescent state while
361 * in an RCU read-side critical section.
363 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
364 "Illegal idle entry in RCU read-side critical section.");
365 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
366 "Illegal idle entry in RCU-bh read-side critical section.");
367 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
368 "Illegal idle entry in RCU-sched read-side critical section.");
372 * Enter an RCU extended quiescent state, which can be either the
373 * idle loop or adaptive-tickless usermode execution.
375 static void rcu_eqs_enter(bool user
)
378 struct rcu_dynticks
*rdtp
;
380 rdtp
= &__get_cpu_var(rcu_dynticks
);
381 oldval
= rdtp
->dynticks_nesting
;
382 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
383 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
384 rdtp
->dynticks_nesting
= 0;
386 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
387 rcu_eqs_enter_common(rdtp
, oldval
, user
);
391 * rcu_idle_enter - inform RCU that current CPU is entering idle
393 * Enter idle mode, in other words, -leave- the mode in which RCU
394 * read-side critical sections can occur. (Though RCU read-side
395 * critical sections can occur in irq handlers in idle, a possibility
396 * handled by irq_enter() and irq_exit().)
398 * We crowbar the ->dynticks_nesting field to zero to allow for
399 * the possibility of usermode upcalls having messed up our count
400 * of interrupt nesting level during the prior busy period.
402 void rcu_idle_enter(void)
406 local_irq_save(flags
);
407 rcu_eqs_enter(false);
408 local_irq_restore(flags
);
410 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
412 #ifdef CONFIG_RCU_USER_QS
414 * rcu_user_enter - inform RCU that we are resuming userspace.
416 * Enter RCU idle mode right before resuming userspace. No use of RCU
417 * is permitted between this call and rcu_user_exit(). This way the
418 * CPU doesn't need to maintain the tick for RCU maintenance purposes
419 * when the CPU runs in userspace.
421 void rcu_user_enter(void)
424 struct rcu_dynticks
*rdtp
;
427 * Some contexts may involve an exception occuring in an irq,
428 * leading to that nesting:
429 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
430 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
431 * helpers are enough to protect RCU uses inside the exception. So
432 * just return immediately if we detect we are in an IRQ.
437 WARN_ON_ONCE(!current
->mm
);
439 local_irq_save(flags
);
440 rdtp
= &__get_cpu_var(rcu_dynticks
);
441 if (!rdtp
->ignore_user_qs
&& !rdtp
->in_user
) {
442 rdtp
->in_user
= true;
445 local_irq_restore(flags
);
449 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
450 * after the current irq returns.
452 * This is similar to rcu_user_enter() but in the context of a non-nesting
453 * irq. After this call, RCU enters into idle mode when the interrupt
456 void rcu_user_enter_after_irq(void)
459 struct rcu_dynticks
*rdtp
;
461 local_irq_save(flags
);
462 rdtp
= &__get_cpu_var(rcu_dynticks
);
463 /* Ensure this irq is interrupting a non-idle RCU state. */
464 WARN_ON_ONCE(!(rdtp
->dynticks_nesting
& DYNTICK_TASK_MASK
));
465 rdtp
->dynticks_nesting
= 1;
466 local_irq_restore(flags
);
468 #endif /* CONFIG_RCU_USER_QS */
471 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
473 * Exit from an interrupt handler, which might possibly result in entering
474 * idle mode, in other words, leaving the mode in which read-side critical
475 * sections can occur.
477 * This code assumes that the idle loop never does anything that might
478 * result in unbalanced calls to irq_enter() and irq_exit(). If your
479 * architecture violates this assumption, RCU will give you what you
480 * deserve, good and hard. But very infrequently and irreproducibly.
482 * Use things like work queues to work around this limitation.
484 * You have been warned.
486 void rcu_irq_exit(void)
490 struct rcu_dynticks
*rdtp
;
492 local_irq_save(flags
);
493 rdtp
= &__get_cpu_var(rcu_dynticks
);
494 oldval
= rdtp
->dynticks_nesting
;
495 rdtp
->dynticks_nesting
--;
496 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
497 if (rdtp
->dynticks_nesting
)
498 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
500 rcu_eqs_enter_common(rdtp
, oldval
, true);
501 local_irq_restore(flags
);
505 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
507 * If the new value of the ->dynticks_nesting counter was previously zero,
508 * we really have exited idle, and must do the appropriate accounting.
509 * The caller must have disabled interrupts.
511 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
514 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
515 atomic_inc(&rdtp
->dynticks
);
516 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
517 smp_mb__after_atomic_inc(); /* See above. */
518 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
519 rcu_cleanup_after_idle(smp_processor_id());
520 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
521 if (!user
&& !is_idle_task(current
)) {
522 struct task_struct
*idle
= idle_task(smp_processor_id());
524 trace_rcu_dyntick("Error on exit: not idle task",
525 oldval
, rdtp
->dynticks_nesting
);
526 ftrace_dump(DUMP_ORIG
);
527 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
528 current
->pid
, current
->comm
,
529 idle
->pid
, idle
->comm
); /* must be idle task! */
534 * Exit an RCU extended quiescent state, which can be either the
535 * idle loop or adaptive-tickless usermode execution.
537 static void rcu_eqs_exit(bool user
)
539 struct rcu_dynticks
*rdtp
;
542 rdtp
= &__get_cpu_var(rcu_dynticks
);
543 oldval
= rdtp
->dynticks_nesting
;
544 WARN_ON_ONCE(oldval
< 0);
545 if (oldval
& DYNTICK_TASK_NEST_MASK
)
546 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
548 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
549 rcu_eqs_exit_common(rdtp
, oldval
, user
);
553 * rcu_idle_exit - inform RCU that current CPU is leaving idle
555 * Exit idle mode, in other words, -enter- the mode in which RCU
556 * read-side critical sections can occur.
558 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
559 * allow for the possibility of usermode upcalls messing up our count
560 * of interrupt nesting level during the busy period that is just
563 void rcu_idle_exit(void)
567 local_irq_save(flags
);
569 local_irq_restore(flags
);
571 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
573 #ifdef CONFIG_RCU_USER_QS
575 * rcu_user_exit - inform RCU that we are exiting userspace.
577 * Exit RCU idle mode while entering the kernel because it can
578 * run a RCU read side critical section anytime.
580 void rcu_user_exit(void)
583 struct rcu_dynticks
*rdtp
;
586 * Some contexts may involve an exception occuring in an irq,
587 * leading to that nesting:
588 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
589 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
590 * helpers are enough to protect RCU uses inside the exception. So
591 * just return immediately if we detect we are in an IRQ.
596 local_irq_save(flags
);
597 rdtp
= &__get_cpu_var(rcu_dynticks
);
599 rdtp
->in_user
= false;
602 local_irq_restore(flags
);
606 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
607 * idle mode after the current non-nesting irq returns.
609 * This is similar to rcu_user_exit() but in the context of an irq.
610 * This is called when the irq has interrupted a userspace RCU idle mode
611 * context. When the current non-nesting interrupt returns after this call,
612 * the CPU won't restore the RCU idle mode.
614 void rcu_user_exit_after_irq(void)
617 struct rcu_dynticks
*rdtp
;
619 local_irq_save(flags
);
620 rdtp
= &__get_cpu_var(rcu_dynticks
);
621 /* Ensure we are interrupting an RCU idle mode. */
622 WARN_ON_ONCE(rdtp
->dynticks_nesting
& DYNTICK_TASK_NEST_MASK
);
623 rdtp
->dynticks_nesting
+= DYNTICK_TASK_EXIT_IDLE
;
624 local_irq_restore(flags
);
626 #endif /* CONFIG_RCU_USER_QS */
629 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
631 * Enter an interrupt handler, which might possibly result in exiting
632 * idle mode, in other words, entering the mode in which read-side critical
633 * sections can occur.
635 * Note that the Linux kernel is fully capable of entering an interrupt
636 * handler that it never exits, for example when doing upcalls to
637 * user mode! This code assumes that the idle loop never does upcalls to
638 * user mode. If your architecture does do upcalls from the idle loop (or
639 * does anything else that results in unbalanced calls to the irq_enter()
640 * and irq_exit() functions), RCU will give you what you deserve, good
641 * and hard. But very infrequently and irreproducibly.
643 * Use things like work queues to work around this limitation.
645 * You have been warned.
647 void rcu_irq_enter(void)
650 struct rcu_dynticks
*rdtp
;
653 local_irq_save(flags
);
654 rdtp
= &__get_cpu_var(rcu_dynticks
);
655 oldval
= rdtp
->dynticks_nesting
;
656 rdtp
->dynticks_nesting
++;
657 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
659 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
661 rcu_eqs_exit_common(rdtp
, oldval
, true);
662 local_irq_restore(flags
);
666 * rcu_nmi_enter - inform RCU of entry to NMI context
668 * If the CPU was idle with dynamic ticks active, and there is no
669 * irq handler running, this updates rdtp->dynticks_nmi to let the
670 * RCU grace-period handling know that the CPU is active.
672 void rcu_nmi_enter(void)
674 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
676 if (rdtp
->dynticks_nmi_nesting
== 0 &&
677 (atomic_read(&rdtp
->dynticks
) & 0x1))
679 rdtp
->dynticks_nmi_nesting
++;
680 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
681 atomic_inc(&rdtp
->dynticks
);
682 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
683 smp_mb__after_atomic_inc(); /* See above. */
684 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
688 * rcu_nmi_exit - inform RCU of exit from NMI context
690 * If the CPU was idle with dynamic ticks active, and there is no
691 * irq handler running, this updates rdtp->dynticks_nmi to let the
692 * RCU grace-period handling know that the CPU is no longer active.
694 void rcu_nmi_exit(void)
696 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
698 if (rdtp
->dynticks_nmi_nesting
== 0 ||
699 --rdtp
->dynticks_nmi_nesting
!= 0)
701 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
702 smp_mb__before_atomic_inc(); /* See above. */
703 atomic_inc(&rdtp
->dynticks
);
704 smp_mb__after_atomic_inc(); /* Force delay to next write. */
705 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
709 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
711 * If the current CPU is in its idle loop and is neither in an interrupt
712 * or NMI handler, return true.
714 int rcu_is_cpu_idle(void)
719 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
723 EXPORT_SYMBOL(rcu_is_cpu_idle
);
725 #ifdef CONFIG_RCU_USER_QS
726 void rcu_user_hooks_switch(struct task_struct
*prev
,
727 struct task_struct
*next
)
729 struct rcu_dynticks
*rdtp
;
731 /* Interrupts are disabled in context switch */
732 rdtp
= &__get_cpu_var(rcu_dynticks
);
733 if (!rdtp
->ignore_user_qs
) {
734 clear_tsk_thread_flag(prev
, TIF_NOHZ
);
735 set_tsk_thread_flag(next
, TIF_NOHZ
);
738 #endif /* #ifdef CONFIG_RCU_USER_QS */
740 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
743 * Is the current CPU online? Disable preemption to avoid false positives
744 * that could otherwise happen due to the current CPU number being sampled,
745 * this task being preempted, its old CPU being taken offline, resuming
746 * on some other CPU, then determining that its old CPU is now offline.
747 * It is OK to use RCU on an offline processor during initial boot, hence
748 * the check for rcu_scheduler_fully_active. Note also that it is OK
749 * for a CPU coming online to use RCU for one jiffy prior to marking itself
750 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
751 * offline to continue to use RCU for one jiffy after marking itself
752 * offline in the cpu_online_mask. This leniency is necessary given the
753 * non-atomic nature of the online and offline processing, for example,
754 * the fact that a CPU enters the scheduler after completing the CPU_DYING
757 * This is also why RCU internally marks CPUs online during the
758 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
760 * Disable checking if in an NMI handler because we cannot safely report
761 * errors from NMI handlers anyway.
763 bool rcu_lockdep_current_cpu_online(void)
765 struct rcu_data
*rdp
;
766 struct rcu_node
*rnp
;
772 rdp
= &__get_cpu_var(rcu_sched_data
);
774 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
775 !rcu_scheduler_fully_active
;
779 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
781 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
784 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
786 * If the current CPU is idle or running at a first-level (not nested)
787 * interrupt from idle, return true. The caller must have at least
788 * disabled preemption.
790 int rcu_is_cpu_rrupt_from_idle(void)
792 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
796 * Snapshot the specified CPU's dynticks counter so that we can later
797 * credit them with an implicit quiescent state. Return 1 if this CPU
798 * is in dynticks idle mode, which is an extended quiescent state.
800 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
802 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
803 return (rdp
->dynticks_snap
& 0x1) == 0;
807 * Return true if the specified CPU has passed through a quiescent
808 * state by virtue of being in or having passed through an dynticks
809 * idle state since the last call to dyntick_save_progress_counter()
810 * for this same CPU, or by virtue of having been offline.
812 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
817 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
818 snap
= (unsigned int)rdp
->dynticks_snap
;
821 * If the CPU passed through or entered a dynticks idle phase with
822 * no active irq/NMI handlers, then we can safely pretend that the CPU
823 * already acknowledged the request to pass through a quiescent
824 * state. Either way, that CPU cannot possibly be in an RCU
825 * read-side critical section that started before the beginning
826 * of the current RCU grace period.
828 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
829 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
835 * Check for the CPU being offline, but only if the grace period
836 * is old enough. We don't need to worry about the CPU changing
837 * state: If we see it offline even once, it has been through a
840 * The reason for insisting that the grace period be at least
841 * one jiffy old is that CPUs that are not quite online and that
842 * have just gone offline can still execute RCU read-side critical
845 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
846 return 0; /* Grace period is not old enough. */
848 if (cpu_is_offline(rdp
->cpu
)) {
849 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
856 static int jiffies_till_stall_check(void)
858 int till_stall_check
= ACCESS_ONCE(rcu_cpu_stall_timeout
);
861 * Limit check must be consistent with the Kconfig limits
862 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
864 if (till_stall_check
< 3) {
865 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 3;
866 till_stall_check
= 3;
867 } else if (till_stall_check
> 300) {
868 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 300;
869 till_stall_check
= 300;
871 return till_stall_check
* HZ
+ RCU_STALL_DELAY_DELTA
;
874 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
876 rsp
->gp_start
= jiffies
;
877 rsp
->jiffies_stall
= jiffies
+ jiffies_till_stall_check();
881 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
882 * for architectures that do not implement trigger_all_cpu_backtrace().
883 * The NMI-triggered stack traces are more accurate because they are
884 * printed by the target CPU.
886 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
890 struct rcu_node
*rnp
;
892 rcu_for_each_leaf_node(rsp
, rnp
) {
893 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
894 if (rnp
->qsmask
!= 0) {
895 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
896 if (rnp
->qsmask
& (1UL << cpu
))
897 dump_cpu_task(rnp
->grplo
+ cpu
);
899 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
903 static void print_other_cpu_stall(struct rcu_state
*rsp
)
909 struct rcu_node
*rnp
= rcu_get_root(rsp
);
912 /* Only let one CPU complain about others per time interval. */
914 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
915 delta
= jiffies
- rsp
->jiffies_stall
;
916 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
917 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
920 rsp
->jiffies_stall
= jiffies
+ 3 * jiffies_till_stall_check() + 3;
921 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
924 * OK, time to rat on our buddy...
925 * See Documentation/RCU/stallwarn.txt for info on how to debug
926 * RCU CPU stall warnings.
928 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
930 print_cpu_stall_info_begin();
931 rcu_for_each_leaf_node(rsp
, rnp
) {
932 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
933 ndetected
+= rcu_print_task_stall(rnp
);
934 if (rnp
->qsmask
!= 0) {
935 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
936 if (rnp
->qsmask
& (1UL << cpu
)) {
937 print_cpu_stall_info(rsp
,
942 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
946 * Now rat on any tasks that got kicked up to the root rcu_node
947 * due to CPU offlining.
949 rnp
= rcu_get_root(rsp
);
950 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
951 ndetected
+= rcu_print_task_stall(rnp
);
952 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
954 print_cpu_stall_info_end();
955 for_each_possible_cpu(cpu
)
956 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
957 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
958 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
959 rsp
->gpnum
, rsp
->completed
, totqlen
);
961 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
962 else if (!trigger_all_cpu_backtrace())
963 rcu_dump_cpu_stacks(rsp
);
965 /* Complain about tasks blocking the grace period. */
967 rcu_print_detail_task_stall(rsp
);
969 force_quiescent_state(rsp
); /* Kick them all. */
972 static void print_cpu_stall(struct rcu_state
*rsp
)
976 struct rcu_node
*rnp
= rcu_get_root(rsp
);
980 * OK, time to rat on ourselves...
981 * See Documentation/RCU/stallwarn.txt for info on how to debug
982 * RCU CPU stall warnings.
984 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
985 print_cpu_stall_info_begin();
986 print_cpu_stall_info(rsp
, smp_processor_id());
987 print_cpu_stall_info_end();
988 for_each_possible_cpu(cpu
)
989 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
990 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
991 jiffies
- rsp
->gp_start
, rsp
->gpnum
, rsp
->completed
, totqlen
);
992 if (!trigger_all_cpu_backtrace())
995 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
996 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
997 rsp
->jiffies_stall
= jiffies
+
998 3 * jiffies_till_stall_check() + 3;
999 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1001 set_need_resched(); /* kick ourselves to get things going. */
1004 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1008 struct rcu_node
*rnp
;
1010 if (rcu_cpu_stall_suppress
)
1012 j
= ACCESS_ONCE(jiffies
);
1013 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
1015 if (rcu_gp_in_progress(rsp
) &&
1016 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
1018 /* We haven't checked in, so go dump stack. */
1019 print_cpu_stall(rsp
);
1021 } else if (rcu_gp_in_progress(rsp
) &&
1022 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1024 /* They had a few time units to dump stack, so complain. */
1025 print_other_cpu_stall(rsp
);
1029 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
1031 rcu_cpu_stall_suppress
= 1;
1036 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1038 * Set the stall-warning timeout way off into the future, thus preventing
1039 * any RCU CPU stall-warning messages from appearing in the current set of
1040 * RCU grace periods.
1042 * The caller must disable hard irqs.
1044 void rcu_cpu_stall_reset(void)
1046 struct rcu_state
*rsp
;
1048 for_each_rcu_flavor(rsp
)
1049 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
1052 static struct notifier_block rcu_panic_block
= {
1053 .notifier_call
= rcu_panic
,
1056 static void __init
check_cpu_stall_init(void)
1058 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
1062 * Update CPU-local rcu_data state to record the newly noticed grace period.
1063 * This is used both when we started the grace period and when we notice
1064 * that someone else started the grace period. The caller must hold the
1065 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
1066 * and must have irqs disabled.
1068 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1070 if (rdp
->gpnum
!= rnp
->gpnum
) {
1072 * If the current grace period is waiting for this CPU,
1073 * set up to detect a quiescent state, otherwise don't
1074 * go looking for one.
1076 rdp
->gpnum
= rnp
->gpnum
;
1077 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
1078 rdp
->passed_quiesce
= 0;
1079 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1080 zero_cpu_stall_ticks(rdp
);
1084 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1086 unsigned long flags
;
1087 struct rcu_node
*rnp
;
1089 local_irq_save(flags
);
1091 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
1092 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1093 local_irq_restore(flags
);
1096 __note_new_gpnum(rsp
, rnp
, rdp
);
1097 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1101 * Did someone else start a new RCU grace period start since we last
1102 * checked? Update local state appropriately if so. Must be called
1103 * on the CPU corresponding to rdp.
1106 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1108 unsigned long flags
;
1111 local_irq_save(flags
);
1112 if (rdp
->gpnum
!= rsp
->gpnum
) {
1113 note_new_gpnum(rsp
, rdp
);
1116 local_irq_restore(flags
);
1121 * Initialize the specified rcu_data structure's callback list to empty.
1123 static void init_callback_list(struct rcu_data
*rdp
)
1127 rdp
->nxtlist
= NULL
;
1128 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1129 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1130 init_nocb_callback_list(rdp
);
1134 * Advance this CPU's callbacks, but only if the current grace period
1135 * has ended. This may be called only from the CPU to whom the rdp
1136 * belongs. In addition, the corresponding leaf rcu_node structure's
1137 * ->lock must be held by the caller, with irqs disabled.
1140 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1142 /* Did another grace period end? */
1143 if (rdp
->completed
!= rnp
->completed
) {
1145 /* Advance callbacks. No harm if list empty. */
1146 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
1147 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
1148 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1150 /* Remember that we saw this grace-period completion. */
1151 rdp
->completed
= rnp
->completed
;
1152 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
1155 * If we were in an extended quiescent state, we may have
1156 * missed some grace periods that others CPUs handled on
1157 * our behalf. Catch up with this state to avoid noting
1158 * spurious new grace periods. If another grace period
1159 * has started, then rnp->gpnum will have advanced, so
1160 * we will detect this later on. Of course, any quiescent
1161 * states we found for the old GP are now invalid.
1163 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
)) {
1164 rdp
->gpnum
= rdp
->completed
;
1165 rdp
->passed_quiesce
= 0;
1169 * If RCU does not need a quiescent state from this CPU,
1170 * then make sure that this CPU doesn't go looking for one.
1172 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1173 rdp
->qs_pending
= 0;
1178 * Advance this CPU's callbacks, but only if the current grace period
1179 * has ended. This may be called only from the CPU to whom the rdp
1183 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1185 unsigned long flags
;
1186 struct rcu_node
*rnp
;
1188 local_irq_save(flags
);
1190 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1191 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1192 local_irq_restore(flags
);
1195 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1196 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1200 * Do per-CPU grace-period initialization for running CPU. The caller
1201 * must hold the lock of the leaf rcu_node structure corresponding to
1205 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1207 /* Prior grace period ended, so advance callbacks for current CPU. */
1208 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1210 /* Set state so that this CPU will detect the next quiescent state. */
1211 __note_new_gpnum(rsp
, rnp
, rdp
);
1215 * Initialize a new grace period.
1217 static int rcu_gp_init(struct rcu_state
*rsp
)
1219 struct rcu_data
*rdp
;
1220 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1222 raw_spin_lock_irq(&rnp
->lock
);
1223 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1225 if (rcu_gp_in_progress(rsp
)) {
1226 /* Grace period already in progress, don't start another. */
1227 raw_spin_unlock_irq(&rnp
->lock
);
1231 /* Advance to a new grace period and initialize state. */
1233 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1234 record_gp_stall_check_time(rsp
);
1235 raw_spin_unlock_irq(&rnp
->lock
);
1237 /* Exclude any concurrent CPU-hotplug operations. */
1238 mutex_lock(&rsp
->onoff_mutex
);
1241 * Set the quiescent-state-needed bits in all the rcu_node
1242 * structures for all currently online CPUs in breadth-first order,
1243 * starting from the root rcu_node structure, relying on the layout
1244 * of the tree within the rsp->node[] array. Note that other CPUs
1245 * will access only the leaves of the hierarchy, thus seeing that no
1246 * grace period is in progress, at least until the corresponding
1247 * leaf node has been initialized. In addition, we have excluded
1248 * CPU-hotplug operations.
1250 * The grace period cannot complete until the initialization
1251 * process finishes, because this kthread handles both.
1253 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1254 raw_spin_lock_irq(&rnp
->lock
);
1255 rdp
= this_cpu_ptr(rsp
->rda
);
1256 rcu_preempt_check_blocked_tasks(rnp
);
1257 rnp
->qsmask
= rnp
->qsmaskinit
;
1258 rnp
->gpnum
= rsp
->gpnum
;
1259 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1260 rnp
->completed
= rsp
->completed
;
1261 if (rnp
== rdp
->mynode
)
1262 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1263 rcu_preempt_boost_start_gp(rnp
);
1264 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1265 rnp
->level
, rnp
->grplo
,
1266 rnp
->grphi
, rnp
->qsmask
);
1267 raw_spin_unlock_irq(&rnp
->lock
);
1268 #ifdef CONFIG_PROVE_RCU_DELAY
1269 if ((random32() % (rcu_num_nodes
* 8)) == 0)
1270 schedule_timeout_uninterruptible(2);
1271 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1275 mutex_unlock(&rsp
->onoff_mutex
);
1280 * Do one round of quiescent-state forcing.
1282 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1284 int fqs_state
= fqs_state_in
;
1285 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1288 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1289 /* Collect dyntick-idle snapshots. */
1290 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1291 fqs_state
= RCU_FORCE_QS
;
1293 /* Handle dyntick-idle and offline CPUs. */
1294 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1296 /* Clear flag to prevent immediate re-entry. */
1297 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1298 raw_spin_lock_irq(&rnp
->lock
);
1299 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1300 raw_spin_unlock_irq(&rnp
->lock
);
1306 * Clean up after the old grace period.
1308 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1310 unsigned long gp_duration
;
1311 struct rcu_data
*rdp
;
1312 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1314 raw_spin_lock_irq(&rnp
->lock
);
1315 gp_duration
= jiffies
- rsp
->gp_start
;
1316 if (gp_duration
> rsp
->gp_max
)
1317 rsp
->gp_max
= gp_duration
;
1320 * We know the grace period is complete, but to everyone else
1321 * it appears to still be ongoing. But it is also the case
1322 * that to everyone else it looks like there is nothing that
1323 * they can do to advance the grace period. It is therefore
1324 * safe for us to drop the lock in order to mark the grace
1325 * period as completed in all of the rcu_node structures.
1327 raw_spin_unlock_irq(&rnp
->lock
);
1330 * Propagate new ->completed value to rcu_node structures so
1331 * that other CPUs don't have to wait until the start of the next
1332 * grace period to process their callbacks. This also avoids
1333 * some nasty RCU grace-period initialization races by forcing
1334 * the end of the current grace period to be completely recorded in
1335 * all of the rcu_node structures before the beginning of the next
1336 * grace period is recorded in any of the rcu_node structures.
1338 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1339 raw_spin_lock_irq(&rnp
->lock
);
1340 rnp
->completed
= rsp
->gpnum
;
1341 raw_spin_unlock_irq(&rnp
->lock
);
1344 rnp
= rcu_get_root(rsp
);
1345 raw_spin_lock_irq(&rnp
->lock
);
1347 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1348 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1349 rsp
->fqs_state
= RCU_GP_IDLE
;
1350 rdp
= this_cpu_ptr(rsp
->rda
);
1351 if (cpu_needs_another_gp(rsp
, rdp
))
1353 raw_spin_unlock_irq(&rnp
->lock
);
1357 * Body of kthread that handles grace periods.
1359 static int __noreturn
rcu_gp_kthread(void *arg
)
1364 struct rcu_state
*rsp
= arg
;
1365 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1369 /* Handle grace-period start. */
1371 wait_event_interruptible(rsp
->gp_wq
,
1374 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1378 flush_signals(current
);
1381 /* Handle quiescent-state forcing. */
1382 fqs_state
= RCU_SAVE_DYNTICK
;
1383 j
= jiffies_till_first_fqs
;
1386 jiffies_till_first_fqs
= HZ
;
1389 rsp
->jiffies_force_qs
= jiffies
+ j
;
1390 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1391 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1392 (!ACCESS_ONCE(rnp
->qsmask
) &&
1393 !rcu_preempt_blocked_readers_cgp(rnp
)),
1395 /* If grace period done, leave loop. */
1396 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1397 !rcu_preempt_blocked_readers_cgp(rnp
))
1399 /* If time for quiescent-state forcing, do it. */
1400 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1401 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1404 /* Deal with stray signal. */
1406 flush_signals(current
);
1408 j
= jiffies_till_next_fqs
;
1411 jiffies_till_next_fqs
= HZ
;
1414 jiffies_till_next_fqs
= 1;
1418 /* Handle grace-period end. */
1419 rcu_gp_cleanup(rsp
);
1424 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1425 * in preparation for detecting the next grace period. The caller must hold
1426 * the root node's ->lock, which is released before return. Hard irqs must
1429 * Note that it is legal for a dying CPU (which is marked as offline) to
1430 * invoke this function. This can happen when the dying CPU reports its
1434 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1435 __releases(rcu_get_root(rsp
)->lock
)
1437 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1438 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1440 if (!rsp
->gp_kthread
||
1441 !cpu_needs_another_gp(rsp
, rdp
)) {
1443 * Either we have not yet spawned the grace-period
1444 * task, this CPU does not need another grace period,
1445 * or a grace period is already in progress.
1446 * Either way, don't start a new grace period.
1448 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1453 * Because there is no grace period in progress right now,
1454 * any callbacks we have up to this point will be satisfied
1455 * by the next grace period. So promote all callbacks to be
1456 * handled after the end of the next grace period. If the
1457 * CPU is not yet aware of the end of the previous grace period,
1458 * we need to allow for the callback advancement that will
1459 * occur when it does become aware. Deadlock prevents us from
1460 * making it aware at this point: We cannot acquire a leaf
1461 * rcu_node ->lock while holding the root rcu_node ->lock.
1463 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1464 if (rdp
->completed
== rsp
->completed
)
1465 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1467 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1468 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
1470 /* Ensure that CPU is aware of completion of last grace period. */
1471 rcu_process_gp_end(rsp
, rdp
);
1472 local_irq_restore(flags
);
1474 /* Wake up rcu_gp_kthread() to start the grace period. */
1475 wake_up(&rsp
->gp_wq
);
1479 * Report a full set of quiescent states to the specified rcu_state
1480 * data structure. This involves cleaning up after the prior grace
1481 * period and letting rcu_start_gp() start up the next grace period
1482 * if one is needed. Note that the caller must hold rnp->lock, as
1483 * required by rcu_start_gp(), which will release it.
1485 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1486 __releases(rcu_get_root(rsp
)->lock
)
1488 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1489 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1490 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1494 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1495 * Allows quiescent states for a group of CPUs to be reported at one go
1496 * to the specified rcu_node structure, though all the CPUs in the group
1497 * must be represented by the same rcu_node structure (which need not be
1498 * a leaf rcu_node structure, though it often will be). That structure's
1499 * lock must be held upon entry, and it is released before return.
1502 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1503 struct rcu_node
*rnp
, unsigned long flags
)
1504 __releases(rnp
->lock
)
1506 struct rcu_node
*rnp_c
;
1508 /* Walk up the rcu_node hierarchy. */
1510 if (!(rnp
->qsmask
& mask
)) {
1512 /* Our bit has already been cleared, so done. */
1513 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1516 rnp
->qsmask
&= ~mask
;
1517 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1518 mask
, rnp
->qsmask
, rnp
->level
,
1519 rnp
->grplo
, rnp
->grphi
,
1521 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1523 /* Other bits still set at this level, so done. */
1524 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1527 mask
= rnp
->grpmask
;
1528 if (rnp
->parent
== NULL
) {
1530 /* No more levels. Exit loop holding root lock. */
1534 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1537 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1538 WARN_ON_ONCE(rnp_c
->qsmask
);
1542 * Get here if we are the last CPU to pass through a quiescent
1543 * state for this grace period. Invoke rcu_report_qs_rsp()
1544 * to clean up and start the next grace period if one is needed.
1546 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1550 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1551 * structure. This must be either called from the specified CPU, or
1552 * called when the specified CPU is known to be offline (and when it is
1553 * also known that no other CPU is concurrently trying to help the offline
1554 * CPU). The lastcomp argument is used to make sure we are still in the
1555 * grace period of interest. We don't want to end the current grace period
1556 * based on quiescent states detected in an earlier grace period!
1559 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1561 unsigned long flags
;
1563 struct rcu_node
*rnp
;
1566 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1567 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1568 rnp
->completed
== rnp
->gpnum
) {
1571 * The grace period in which this quiescent state was
1572 * recorded has ended, so don't report it upwards.
1573 * We will instead need a new quiescent state that lies
1574 * within the current grace period.
1576 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1577 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1580 mask
= rdp
->grpmask
;
1581 if ((rnp
->qsmask
& mask
) == 0) {
1582 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1584 rdp
->qs_pending
= 0;
1587 * This GP can't end until cpu checks in, so all of our
1588 * callbacks can be processed during the next GP.
1590 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1592 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1597 * Check to see if there is a new grace period of which this CPU
1598 * is not yet aware, and if so, set up local rcu_data state for it.
1599 * Otherwise, see if this CPU has just passed through its first
1600 * quiescent state for this grace period, and record that fact if so.
1603 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1605 /* If there is now a new grace period, record and return. */
1606 if (check_for_new_grace_period(rsp
, rdp
))
1610 * Does this CPU still need to do its part for current grace period?
1611 * If no, return and let the other CPUs do their part as well.
1613 if (!rdp
->qs_pending
)
1617 * Was there a quiescent state since the beginning of the grace
1618 * period? If no, then exit and wait for the next call.
1620 if (!rdp
->passed_quiesce
)
1624 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1627 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1630 #ifdef CONFIG_HOTPLUG_CPU
1633 * Send the specified CPU's RCU callbacks to the orphanage. The
1634 * specified CPU must be offline, and the caller must hold the
1638 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1639 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1641 /* No-CBs CPUs do not have orphanable callbacks. */
1642 if (is_nocb_cpu(rdp
->cpu
))
1646 * Orphan the callbacks. First adjust the counts. This is safe
1647 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1648 * cannot be running now. Thus no memory barrier is required.
1650 if (rdp
->nxtlist
!= NULL
) {
1651 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1652 rsp
->qlen
+= rdp
->qlen
;
1653 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1655 ACCESS_ONCE(rdp
->qlen
) = 0;
1659 * Next, move those callbacks still needing a grace period to
1660 * the orphanage, where some other CPU will pick them up.
1661 * Some of the callbacks might have gone partway through a grace
1662 * period, but that is too bad. They get to start over because we
1663 * cannot assume that grace periods are synchronized across CPUs.
1664 * We don't bother updating the ->nxttail[] array yet, instead
1665 * we just reset the whole thing later on.
1667 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1668 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1669 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1670 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1674 * Then move the ready-to-invoke callbacks to the orphanage,
1675 * where some other CPU will pick them up. These will not be
1676 * required to pass though another grace period: They are done.
1678 if (rdp
->nxtlist
!= NULL
) {
1679 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1680 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1683 /* Finally, initialize the rcu_data structure's list to empty. */
1684 init_callback_list(rdp
);
1688 * Adopt the RCU callbacks from the specified rcu_state structure's
1689 * orphanage. The caller must hold the ->orphan_lock.
1691 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1694 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1696 /* No-CBs CPUs are handled specially. */
1697 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
))
1700 /* Do the accounting first. */
1701 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1702 rdp
->qlen
+= rsp
->qlen
;
1703 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1704 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1705 rcu_idle_count_callbacks_posted();
1710 * We do not need a memory barrier here because the only way we
1711 * can get here if there is an rcu_barrier() in flight is if
1712 * we are the task doing the rcu_barrier().
1715 /* First adopt the ready-to-invoke callbacks. */
1716 if (rsp
->orphan_donelist
!= NULL
) {
1717 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1718 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1719 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1720 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1721 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1722 rsp
->orphan_donelist
= NULL
;
1723 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1726 /* And then adopt the callbacks that still need a grace period. */
1727 if (rsp
->orphan_nxtlist
!= NULL
) {
1728 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1729 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1730 rsp
->orphan_nxtlist
= NULL
;
1731 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1736 * Trace the fact that this CPU is going offline.
1738 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1740 RCU_TRACE(unsigned long mask
);
1741 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1742 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1744 RCU_TRACE(mask
= rdp
->grpmask
);
1745 trace_rcu_grace_period(rsp
->name
,
1746 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1751 * The CPU has been completely removed, and some other CPU is reporting
1752 * this fact from process context. Do the remainder of the cleanup,
1753 * including orphaning the outgoing CPU's RCU callbacks, and also
1754 * adopting them. There can only be one CPU hotplug operation at a time,
1755 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1757 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1759 unsigned long flags
;
1761 int need_report
= 0;
1762 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1763 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1765 /* Adjust any no-longer-needed kthreads. */
1766 rcu_boost_kthread_setaffinity(rnp
, -1);
1768 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1770 /* Exclude any attempts to start a new grace period. */
1771 mutex_lock(&rsp
->onoff_mutex
);
1772 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
1774 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1775 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1776 rcu_adopt_orphan_cbs(rsp
);
1778 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1779 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1781 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1782 rnp
->qsmaskinit
&= ~mask
;
1783 if (rnp
->qsmaskinit
!= 0) {
1784 if (rnp
!= rdp
->mynode
)
1785 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1788 if (rnp
== rdp
->mynode
)
1789 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1791 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1792 mask
= rnp
->grpmask
;
1794 } while (rnp
!= NULL
);
1797 * We still hold the leaf rcu_node structure lock here, and
1798 * irqs are still disabled. The reason for this subterfuge is
1799 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1800 * held leads to deadlock.
1802 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
1804 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1805 rcu_report_unblock_qs_rnp(rnp
, flags
);
1807 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1808 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1809 rcu_report_exp_rnp(rsp
, rnp
, true);
1810 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1811 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1812 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1813 init_callback_list(rdp
);
1814 /* Disallow further callbacks on this CPU. */
1815 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1816 mutex_unlock(&rsp
->onoff_mutex
);
1819 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1821 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1825 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1829 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1832 * Invoke any RCU callbacks that have made it to the end of their grace
1833 * period. Thottle as specified by rdp->blimit.
1835 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1837 unsigned long flags
;
1838 struct rcu_head
*next
, *list
, **tail
;
1839 long bl
, count
, count_lazy
;
1842 /* If no callbacks are ready, just return.*/
1843 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1844 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1845 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1846 need_resched(), is_idle_task(current
),
1847 rcu_is_callbacks_kthread());
1852 * Extract the list of ready callbacks, disabling to prevent
1853 * races with call_rcu() from interrupt handlers.
1855 local_irq_save(flags
);
1856 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1858 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1859 list
= rdp
->nxtlist
;
1860 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1861 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1862 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1863 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1864 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1865 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1866 local_irq_restore(flags
);
1868 /* Invoke callbacks. */
1869 count
= count_lazy
= 0;
1873 debug_rcu_head_unqueue(list
);
1874 if (__rcu_reclaim(rsp
->name
, list
))
1877 /* Stop only if limit reached and CPU has something to do. */
1878 if (++count
>= bl
&&
1880 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1884 local_irq_save(flags
);
1885 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1886 is_idle_task(current
),
1887 rcu_is_callbacks_kthread());
1889 /* Update count, and requeue any remaining callbacks. */
1891 *tail
= rdp
->nxtlist
;
1892 rdp
->nxtlist
= list
;
1893 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1894 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1895 rdp
->nxttail
[i
] = tail
;
1899 smp_mb(); /* List handling before counting for rcu_barrier(). */
1900 rdp
->qlen_lazy
-= count_lazy
;
1901 ACCESS_ONCE(rdp
->qlen
) -= count
;
1902 rdp
->n_cbs_invoked
+= count
;
1904 /* Reinstate batch limit if we have worked down the excess. */
1905 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1906 rdp
->blimit
= blimit
;
1908 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1909 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1910 rdp
->qlen_last_fqs_check
= 0;
1911 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1912 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1913 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1914 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
1916 local_irq_restore(flags
);
1918 /* Re-invoke RCU core processing if there are callbacks remaining. */
1919 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1924 * Check to see if this CPU is in a non-context-switch quiescent state
1925 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1926 * Also schedule RCU core processing.
1928 * This function must be called from hardirq context. It is normally
1929 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1930 * false, there is no point in invoking rcu_check_callbacks().
1932 void rcu_check_callbacks(int cpu
, int user
)
1934 trace_rcu_utilization("Start scheduler-tick");
1935 increment_cpu_stall_ticks();
1936 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1939 * Get here if this CPU took its interrupt from user
1940 * mode or from the idle loop, and if this is not a
1941 * nested interrupt. In this case, the CPU is in
1942 * a quiescent state, so note it.
1944 * No memory barrier is required here because both
1945 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1946 * variables that other CPUs neither access nor modify,
1947 * at least not while the corresponding CPU is online.
1953 } else if (!in_softirq()) {
1956 * Get here if this CPU did not take its interrupt from
1957 * softirq, in other words, if it is not interrupting
1958 * a rcu_bh read-side critical section. This is an _bh
1959 * critical section, so note it.
1964 rcu_preempt_check_callbacks(cpu
);
1965 if (rcu_pending(cpu
))
1967 trace_rcu_utilization("End scheduler-tick");
1971 * Scan the leaf rcu_node structures, processing dyntick state for any that
1972 * have not yet encountered a quiescent state, using the function specified.
1973 * Also initiate boosting for any threads blocked on the root rcu_node.
1975 * The caller must have suppressed start of new grace periods.
1977 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1981 unsigned long flags
;
1983 struct rcu_node
*rnp
;
1985 rcu_for_each_leaf_node(rsp
, rnp
) {
1988 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1989 if (!rcu_gp_in_progress(rsp
)) {
1990 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1993 if (rnp
->qsmask
== 0) {
1994 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1999 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2000 if ((rnp
->qsmask
& bit
) != 0 &&
2001 f(per_cpu_ptr(rsp
->rda
, cpu
)))
2006 /* rcu_report_qs_rnp() releases rnp->lock. */
2007 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
2010 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2012 rnp
= rcu_get_root(rsp
);
2013 if (rnp
->qsmask
== 0) {
2014 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2015 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
2020 * Force quiescent states on reluctant CPUs, and also detect which
2021 * CPUs are in dyntick-idle mode.
2023 static void force_quiescent_state(struct rcu_state
*rsp
)
2025 unsigned long flags
;
2027 struct rcu_node
*rnp
;
2028 struct rcu_node
*rnp_old
= NULL
;
2030 /* Funnel through hierarchy to reduce memory contention. */
2031 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
2032 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2033 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2034 !raw_spin_trylock(&rnp
->fqslock
);
2035 if (rnp_old
!= NULL
)
2036 raw_spin_unlock(&rnp_old
->fqslock
);
2038 rsp
->n_force_qs_lh
++;
2043 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2045 /* Reached the root of the rcu_node tree, acquire lock. */
2046 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2047 raw_spin_unlock(&rnp_old
->fqslock
);
2048 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2049 rsp
->n_force_qs_lh
++;
2050 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2051 return; /* Someone beat us to it. */
2053 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
2054 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2055 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
2059 * This does the RCU core processing work for the specified rcu_state
2060 * and rcu_data structures. This may be called only from the CPU to
2061 * whom the rdp belongs.
2064 __rcu_process_callbacks(struct rcu_state
*rsp
)
2066 unsigned long flags
;
2067 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2069 WARN_ON_ONCE(rdp
->beenonline
== 0);
2072 * Advance callbacks in response to end of earlier grace
2073 * period that some other CPU ended.
2075 rcu_process_gp_end(rsp
, rdp
);
2077 /* Update RCU state based on any recent quiescent states. */
2078 rcu_check_quiescent_state(rsp
, rdp
);
2080 /* Does this CPU require a not-yet-started grace period? */
2081 if (cpu_needs_another_gp(rsp
, rdp
)) {
2082 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
2083 rcu_start_gp(rsp
, flags
); /* releases above lock */
2086 /* If there are callbacks ready, invoke them. */
2087 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2088 invoke_rcu_callbacks(rsp
, rdp
);
2092 * Do RCU core processing for the current CPU.
2094 static void rcu_process_callbacks(struct softirq_action
*unused
)
2096 struct rcu_state
*rsp
;
2098 if (cpu_is_offline(smp_processor_id()))
2100 trace_rcu_utilization("Start RCU core");
2101 for_each_rcu_flavor(rsp
)
2102 __rcu_process_callbacks(rsp
);
2103 trace_rcu_utilization("End RCU core");
2107 * Schedule RCU callback invocation. If the specified type of RCU
2108 * does not support RCU priority boosting, just do a direct call,
2109 * otherwise wake up the per-CPU kernel kthread. Note that because we
2110 * are running on the current CPU with interrupts disabled, the
2111 * rcu_cpu_kthread_task cannot disappear out from under us.
2113 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2115 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2117 if (likely(!rsp
->boost
)) {
2118 rcu_do_batch(rsp
, rdp
);
2121 invoke_rcu_callbacks_kthread();
2124 static void invoke_rcu_core(void)
2126 raise_softirq(RCU_SOFTIRQ
);
2130 * Handle any core-RCU processing required by a call_rcu() invocation.
2132 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2133 struct rcu_head
*head
, unsigned long flags
)
2136 * If called from an extended quiescent state, invoke the RCU
2137 * core in order to force a re-evaluation of RCU's idleness.
2139 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2142 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2143 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2147 * Force the grace period if too many callbacks or too long waiting.
2148 * Enforce hysteresis, and don't invoke force_quiescent_state()
2149 * if some other CPU has recently done so. Also, don't bother
2150 * invoking force_quiescent_state() if the newly enqueued callback
2151 * is the only one waiting for a grace period to complete.
2153 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2155 /* Are we ignoring a completed grace period? */
2156 rcu_process_gp_end(rsp
, rdp
);
2157 check_for_new_grace_period(rsp
, rdp
);
2159 /* Start a new grace period if one not already started. */
2160 if (!rcu_gp_in_progress(rsp
)) {
2161 unsigned long nestflag
;
2162 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2164 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
2165 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
2167 /* Give the grace period a kick. */
2168 rdp
->blimit
= LONG_MAX
;
2169 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2170 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2171 force_quiescent_state(rsp
);
2172 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2173 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2179 * Helper function for call_rcu() and friends. The cpu argument will
2180 * normally be -1, indicating "currently running CPU". It may specify
2181 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2182 * is expected to specify a CPU.
2185 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2186 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2188 unsigned long flags
;
2189 struct rcu_data
*rdp
;
2191 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2192 debug_rcu_head_queue(head
);
2197 * Opportunistically note grace-period endings and beginnings.
2198 * Note that we might see a beginning right after we see an
2199 * end, but never vice versa, since this CPU has to pass through
2200 * a quiescent state betweentimes.
2202 local_irq_save(flags
);
2203 rdp
= this_cpu_ptr(rsp
->rda
);
2205 /* Add the callback to our list. */
2206 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2210 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2211 offline
= !__call_rcu_nocb(rdp
, head
, lazy
);
2212 WARN_ON_ONCE(offline
);
2213 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2214 local_irq_restore(flags
);
2217 ACCESS_ONCE(rdp
->qlen
)++;
2221 rcu_idle_count_callbacks_posted();
2222 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2223 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2224 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2226 if (__is_kfree_rcu_offset((unsigned long)func
))
2227 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2228 rdp
->qlen_lazy
, rdp
->qlen
);
2230 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2232 /* Go handle any RCU core processing required. */
2233 __call_rcu_core(rsp
, rdp
, head
, flags
);
2234 local_irq_restore(flags
);
2238 * Queue an RCU-sched callback for invocation after a grace period.
2240 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2242 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2244 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2247 * Queue an RCU callback for invocation after a quicker grace period.
2249 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2251 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2253 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2256 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2257 * any blocking grace-period wait automatically implies a grace period
2258 * if there is only one CPU online at any point time during execution
2259 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2260 * occasionally incorrectly indicate that there are multiple CPUs online
2261 * when there was in fact only one the whole time, as this just adds
2262 * some overhead: RCU still operates correctly.
2264 static inline int rcu_blocking_is_gp(void)
2268 might_sleep(); /* Check for RCU read-side critical section. */
2270 ret
= num_online_cpus() <= 1;
2276 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2278 * Control will return to the caller some time after a full rcu-sched
2279 * grace period has elapsed, in other words after all currently executing
2280 * rcu-sched read-side critical sections have completed. These read-side
2281 * critical sections are delimited by rcu_read_lock_sched() and
2282 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2283 * local_irq_disable(), and so on may be used in place of
2284 * rcu_read_lock_sched().
2286 * This means that all preempt_disable code sequences, including NMI and
2287 * non-threaded hardware-interrupt handlers, in progress on entry will
2288 * have completed before this primitive returns. However, this does not
2289 * guarantee that softirq handlers will have completed, since in some
2290 * kernels, these handlers can run in process context, and can block.
2292 * Note that this guarantee implies further memory-ordering guarantees.
2293 * On systems with more than one CPU, when synchronize_sched() returns,
2294 * each CPU is guaranteed to have executed a full memory barrier since the
2295 * end of its last RCU-sched read-side critical section whose beginning
2296 * preceded the call to synchronize_sched(). In addition, each CPU having
2297 * an RCU read-side critical section that extends beyond the return from
2298 * synchronize_sched() is guaranteed to have executed a full memory barrier
2299 * after the beginning of synchronize_sched() and before the beginning of
2300 * that RCU read-side critical section. Note that these guarantees include
2301 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2302 * that are executing in the kernel.
2304 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2305 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2306 * to have executed a full memory barrier during the execution of
2307 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2308 * again only if the system has more than one CPU).
2310 * This primitive provides the guarantees made by the (now removed)
2311 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2312 * guarantees that rcu_read_lock() sections will have completed.
2313 * In "classic RCU", these two guarantees happen to be one and
2314 * the same, but can differ in realtime RCU implementations.
2316 void synchronize_sched(void)
2318 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2319 !lock_is_held(&rcu_lock_map
) &&
2320 !lock_is_held(&rcu_sched_lock_map
),
2321 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2322 if (rcu_blocking_is_gp())
2325 synchronize_sched_expedited();
2327 wait_rcu_gp(call_rcu_sched
);
2329 EXPORT_SYMBOL_GPL(synchronize_sched
);
2332 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2334 * Control will return to the caller some time after a full rcu_bh grace
2335 * period has elapsed, in other words after all currently executing rcu_bh
2336 * read-side critical sections have completed. RCU read-side critical
2337 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2338 * and may be nested.
2340 * See the description of synchronize_sched() for more detailed information
2341 * on memory ordering guarantees.
2343 void synchronize_rcu_bh(void)
2345 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2346 !lock_is_held(&rcu_lock_map
) &&
2347 !lock_is_held(&rcu_sched_lock_map
),
2348 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2349 if (rcu_blocking_is_gp())
2352 synchronize_rcu_bh_expedited();
2354 wait_rcu_gp(call_rcu_bh
);
2356 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2358 static int synchronize_sched_expedited_cpu_stop(void *data
)
2361 * There must be a full memory barrier on each affected CPU
2362 * between the time that try_stop_cpus() is called and the
2363 * time that it returns.
2365 * In the current initial implementation of cpu_stop, the
2366 * above condition is already met when the control reaches
2367 * this point and the following smp_mb() is not strictly
2368 * necessary. Do smp_mb() anyway for documentation and
2369 * robustness against future implementation changes.
2371 smp_mb(); /* See above comment block. */
2376 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2378 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2379 * approach to force the grace period to end quickly. This consumes
2380 * significant time on all CPUs and is unfriendly to real-time workloads,
2381 * so is thus not recommended for any sort of common-case code. In fact,
2382 * if you are using synchronize_sched_expedited() in a loop, please
2383 * restructure your code to batch your updates, and then use a single
2384 * synchronize_sched() instead.
2386 * Note that it is illegal to call this function while holding any lock
2387 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2388 * to call this function from a CPU-hotplug notifier. Failing to observe
2389 * these restriction will result in deadlock.
2391 * This implementation can be thought of as an application of ticket
2392 * locking to RCU, with sync_sched_expedited_started and
2393 * sync_sched_expedited_done taking on the roles of the halves
2394 * of the ticket-lock word. Each task atomically increments
2395 * sync_sched_expedited_started upon entry, snapshotting the old value,
2396 * then attempts to stop all the CPUs. If this succeeds, then each
2397 * CPU will have executed a context switch, resulting in an RCU-sched
2398 * grace period. We are then done, so we use atomic_cmpxchg() to
2399 * update sync_sched_expedited_done to match our snapshot -- but
2400 * only if someone else has not already advanced past our snapshot.
2402 * On the other hand, if try_stop_cpus() fails, we check the value
2403 * of sync_sched_expedited_done. If it has advanced past our
2404 * initial snapshot, then someone else must have forced a grace period
2405 * some time after we took our snapshot. In this case, our work is
2406 * done for us, and we can simply return. Otherwise, we try again,
2407 * but keep our initial snapshot for purposes of checking for someone
2408 * doing our work for us.
2410 * If we fail too many times in a row, we fall back to synchronize_sched().
2412 void synchronize_sched_expedited(void)
2414 long firstsnap
, s
, snap
;
2416 struct rcu_state
*rsp
= &rcu_sched_state
;
2419 * If we are in danger of counter wrap, just do synchronize_sched().
2420 * By allowing sync_sched_expedited_started to advance no more than
2421 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2422 * that more than 3.5 billion CPUs would be required to force a
2423 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2424 * course be required on a 64-bit system.
2426 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
2427 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
2429 synchronize_sched();
2430 atomic_long_inc(&rsp
->expedited_wrap
);
2435 * Take a ticket. Note that atomic_inc_return() implies a
2436 * full memory barrier.
2438 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
2441 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2444 * Each pass through the following loop attempts to force a
2445 * context switch on each CPU.
2447 while (try_stop_cpus(cpu_online_mask
,
2448 synchronize_sched_expedited_cpu_stop
,
2451 atomic_long_inc(&rsp
->expedited_tryfail
);
2453 /* Check to see if someone else did our work for us. */
2454 s
= atomic_long_read(&rsp
->expedited_done
);
2455 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2456 /* ensure test happens before caller kfree */
2457 smp_mb__before_atomic_inc(); /* ^^^ */
2458 atomic_long_inc(&rsp
->expedited_workdone1
);
2462 /* No joy, try again later. Or just synchronize_sched(). */
2463 if (trycount
++ < 10) {
2464 udelay(trycount
* num_online_cpus());
2466 wait_rcu_gp(call_rcu_sched
);
2467 atomic_long_inc(&rsp
->expedited_normal
);
2471 /* Recheck to see if someone else did our work for us. */
2472 s
= atomic_long_read(&rsp
->expedited_done
);
2473 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2474 /* ensure test happens before caller kfree */
2475 smp_mb__before_atomic_inc(); /* ^^^ */
2476 atomic_long_inc(&rsp
->expedited_workdone2
);
2481 * Refetching sync_sched_expedited_started allows later
2482 * callers to piggyback on our grace period. We retry
2483 * after they started, so our grace period works for them,
2484 * and they started after our first try, so their grace
2485 * period works for us.
2488 snap
= atomic_long_read(&rsp
->expedited_start
);
2489 smp_mb(); /* ensure read is before try_stop_cpus(). */
2491 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
2494 * Everyone up to our most recent fetch is covered by our grace
2495 * period. Update the counter, but only if our work is still
2496 * relevant -- which it won't be if someone who started later
2497 * than we did already did their update.
2500 atomic_long_inc(&rsp
->expedited_done_tries
);
2501 s
= atomic_long_read(&rsp
->expedited_done
);
2502 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
2503 /* ensure test happens before caller kfree */
2504 smp_mb__before_atomic_inc(); /* ^^^ */
2505 atomic_long_inc(&rsp
->expedited_done_lost
);
2508 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
2509 atomic_long_inc(&rsp
->expedited_done_exit
);
2513 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2516 * Check to see if there is any immediate RCU-related work to be done
2517 * by the current CPU, for the specified type of RCU, returning 1 if so.
2518 * The checks are in order of increasing expense: checks that can be
2519 * carried out against CPU-local state are performed first. However,
2520 * we must check for CPU stalls first, else we might not get a chance.
2522 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2524 struct rcu_node
*rnp
= rdp
->mynode
;
2526 rdp
->n_rcu_pending
++;
2528 /* Check for CPU stalls, if enabled. */
2529 check_cpu_stall(rsp
, rdp
);
2531 /* Is the RCU core waiting for a quiescent state from this CPU? */
2532 if (rcu_scheduler_fully_active
&&
2533 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2534 rdp
->n_rp_qs_pending
++;
2535 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2536 rdp
->n_rp_report_qs
++;
2540 /* Does this CPU have callbacks ready to invoke? */
2541 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2542 rdp
->n_rp_cb_ready
++;
2546 /* Has RCU gone idle with this CPU needing another grace period? */
2547 if (cpu_needs_another_gp(rsp
, rdp
)) {
2548 rdp
->n_rp_cpu_needs_gp
++;
2552 /* Has another RCU grace period completed? */
2553 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2554 rdp
->n_rp_gp_completed
++;
2558 /* Has a new RCU grace period started? */
2559 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2560 rdp
->n_rp_gp_started
++;
2565 rdp
->n_rp_need_nothing
++;
2570 * Check to see if there is any immediate RCU-related work to be done
2571 * by the current CPU, returning 1 if so. This function is part of the
2572 * RCU implementation; it is -not- an exported member of the RCU API.
2574 static int rcu_pending(int cpu
)
2576 struct rcu_state
*rsp
;
2578 for_each_rcu_flavor(rsp
)
2579 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2585 * Check to see if any future RCU-related work will need to be done
2586 * by the current CPU, even if none need be done immediately, returning
2589 static int rcu_cpu_has_callbacks(int cpu
)
2591 struct rcu_state
*rsp
;
2593 /* RCU callbacks either ready or pending? */
2594 for_each_rcu_flavor(rsp
)
2595 if (per_cpu_ptr(rsp
->rda
, cpu
)->nxtlist
)
2601 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2602 * the compiler is expected to optimize this away.
2604 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2605 int cpu
, unsigned long done
)
2607 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2608 atomic_read(&rsp
->barrier_cpu_count
), done
);
2612 * RCU callback function for _rcu_barrier(). If we are last, wake
2613 * up the task executing _rcu_barrier().
2615 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2617 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2618 struct rcu_state
*rsp
= rdp
->rsp
;
2620 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2621 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2622 complete(&rsp
->barrier_completion
);
2624 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2629 * Called with preemption disabled, and from cross-cpu IRQ context.
2631 static void rcu_barrier_func(void *type
)
2633 struct rcu_state
*rsp
= type
;
2634 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2636 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2637 atomic_inc(&rsp
->barrier_cpu_count
);
2638 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2642 * Orchestrate the specified type of RCU barrier, waiting for all
2643 * RCU callbacks of the specified type to complete.
2645 static void _rcu_barrier(struct rcu_state
*rsp
)
2648 struct rcu_data
*rdp
;
2649 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2650 unsigned long snap_done
;
2652 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2654 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2655 mutex_lock(&rsp
->barrier_mutex
);
2658 * Ensure that all prior references, including to ->n_barrier_done,
2659 * are ordered before the _rcu_barrier() machinery.
2661 smp_mb(); /* See above block comment. */
2664 * Recheck ->n_barrier_done to see if others did our work for us.
2665 * This means checking ->n_barrier_done for an even-to-odd-to-even
2666 * transition. The "if" expression below therefore rounds the old
2667 * value up to the next even number and adds two before comparing.
2669 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2670 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2671 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2672 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2673 smp_mb(); /* caller's subsequent code after above check. */
2674 mutex_unlock(&rsp
->barrier_mutex
);
2679 * Increment ->n_barrier_done to avoid duplicate work. Use
2680 * ACCESS_ONCE() to prevent the compiler from speculating
2681 * the increment to precede the early-exit check.
2683 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2684 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2685 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2686 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2689 * Initialize the count to one rather than to zero in order to
2690 * avoid a too-soon return to zero in case of a short grace period
2691 * (or preemption of this task). Exclude CPU-hotplug operations
2692 * to ensure that no offline CPU has callbacks queued.
2694 init_completion(&rsp
->barrier_completion
);
2695 atomic_set(&rsp
->barrier_cpu_count
, 1);
2699 * Force each CPU with callbacks to register a new callback.
2700 * When that callback is invoked, we will know that all of the
2701 * corresponding CPU's preceding callbacks have been invoked.
2703 for_each_possible_cpu(cpu
) {
2704 if (!cpu_online(cpu
) && !is_nocb_cpu(cpu
))
2706 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2707 if (is_nocb_cpu(cpu
)) {
2708 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
2709 rsp
->n_barrier_done
);
2710 atomic_inc(&rsp
->barrier_cpu_count
);
2711 __call_rcu(&rdp
->barrier_head
, rcu_barrier_callback
,
2713 } else if (ACCESS_ONCE(rdp
->qlen
)) {
2714 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2715 rsp
->n_barrier_done
);
2716 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2718 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2719 rsp
->n_barrier_done
);
2725 * Now that we have an rcu_barrier_callback() callback on each
2726 * CPU, and thus each counted, remove the initial count.
2728 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2729 complete(&rsp
->barrier_completion
);
2731 /* Increment ->n_barrier_done to prevent duplicate work. */
2732 smp_mb(); /* Keep increment after above mechanism. */
2733 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2734 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2735 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2736 smp_mb(); /* Keep increment before caller's subsequent code. */
2738 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2739 wait_for_completion(&rsp
->barrier_completion
);
2741 /* Other rcu_barrier() invocations can now safely proceed. */
2742 mutex_unlock(&rsp
->barrier_mutex
);
2746 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2748 void rcu_barrier_bh(void)
2750 _rcu_barrier(&rcu_bh_state
);
2752 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2755 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2757 void rcu_barrier_sched(void)
2759 _rcu_barrier(&rcu_sched_state
);
2761 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2764 * Do boot-time initialization of a CPU's per-CPU RCU data.
2767 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2769 unsigned long flags
;
2770 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2771 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2773 /* Set up local state, ensuring consistent view of global state. */
2774 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2775 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2776 init_callback_list(rdp
);
2778 ACCESS_ONCE(rdp
->qlen
) = 0;
2779 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2780 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2781 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2782 #ifdef CONFIG_RCU_USER_QS
2783 WARN_ON_ONCE(rdp
->dynticks
->in_user
);
2787 rcu_boot_init_nocb_percpu_data(rdp
);
2788 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2792 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2793 * offline event can be happening at a given time. Note also that we
2794 * can accept some slop in the rsp->completed access due to the fact
2795 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2797 static void __cpuinit
2798 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2800 unsigned long flags
;
2802 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2803 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2805 /* Exclude new grace periods. */
2806 mutex_lock(&rsp
->onoff_mutex
);
2808 /* Set up local state, ensuring consistent view of global state. */
2809 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2810 rdp
->beenonline
= 1; /* We have now been online. */
2811 rdp
->preemptible
= preemptible
;
2812 rdp
->qlen_last_fqs_check
= 0;
2813 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2814 rdp
->blimit
= blimit
;
2815 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
2816 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2817 atomic_set(&rdp
->dynticks
->dynticks
,
2818 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2819 rcu_prepare_for_idle_init(cpu
);
2820 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2822 /* Add CPU to rcu_node bitmasks. */
2824 mask
= rdp
->grpmask
;
2826 /* Exclude any attempts to start a new GP on small systems. */
2827 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2828 rnp
->qsmaskinit
|= mask
;
2829 mask
= rnp
->grpmask
;
2830 if (rnp
== rdp
->mynode
) {
2832 * If there is a grace period in progress, we will
2833 * set up to wait for it next time we run the
2836 rdp
->gpnum
= rnp
->completed
;
2837 rdp
->completed
= rnp
->completed
;
2838 rdp
->passed_quiesce
= 0;
2839 rdp
->qs_pending
= 0;
2840 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2842 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2844 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2845 local_irq_restore(flags
);
2847 mutex_unlock(&rsp
->onoff_mutex
);
2850 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2852 struct rcu_state
*rsp
;
2854 for_each_rcu_flavor(rsp
)
2855 rcu_init_percpu_data(cpu
, rsp
,
2856 strcmp(rsp
->name
, "rcu_preempt") == 0);
2860 * Handle CPU online/offline notification events.
2862 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2863 unsigned long action
, void *hcpu
)
2865 long cpu
= (long)hcpu
;
2866 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2867 struct rcu_node
*rnp
= rdp
->mynode
;
2868 struct rcu_state
*rsp
;
2869 int ret
= NOTIFY_OK
;
2871 trace_rcu_utilization("Start CPU hotplug");
2873 case CPU_UP_PREPARE
:
2874 case CPU_UP_PREPARE_FROZEN
:
2875 rcu_prepare_cpu(cpu
);
2876 rcu_prepare_kthreads(cpu
);
2879 case CPU_DOWN_FAILED
:
2880 rcu_boost_kthread_setaffinity(rnp
, -1);
2882 case CPU_DOWN_PREPARE
:
2883 if (nocb_cpu_expendable(cpu
))
2884 rcu_boost_kthread_setaffinity(rnp
, cpu
);
2889 case CPU_DYING_FROZEN
:
2891 * The whole machine is "stopped" except this CPU, so we can
2892 * touch any data without introducing corruption. We send the
2893 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2895 for_each_rcu_flavor(rsp
)
2896 rcu_cleanup_dying_cpu(rsp
);
2897 rcu_cleanup_after_idle(cpu
);
2900 case CPU_DEAD_FROZEN
:
2901 case CPU_UP_CANCELED
:
2902 case CPU_UP_CANCELED_FROZEN
:
2903 for_each_rcu_flavor(rsp
)
2904 rcu_cleanup_dead_cpu(cpu
, rsp
);
2909 trace_rcu_utilization("End CPU hotplug");
2914 * Spawn the kthread that handles this RCU flavor's grace periods.
2916 static int __init
rcu_spawn_gp_kthread(void)
2918 unsigned long flags
;
2919 struct rcu_node
*rnp
;
2920 struct rcu_state
*rsp
;
2921 struct task_struct
*t
;
2923 for_each_rcu_flavor(rsp
) {
2924 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
2926 rnp
= rcu_get_root(rsp
);
2927 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2928 rsp
->gp_kthread
= t
;
2929 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2930 rcu_spawn_nocb_kthreads(rsp
);
2934 early_initcall(rcu_spawn_gp_kthread
);
2937 * This function is invoked towards the end of the scheduler's initialization
2938 * process. Before this is called, the idle task might contain
2939 * RCU read-side critical sections (during which time, this idle
2940 * task is booting the system). After this function is called, the
2941 * idle tasks are prohibited from containing RCU read-side critical
2942 * sections. This function also enables RCU lockdep checking.
2944 void rcu_scheduler_starting(void)
2946 WARN_ON(num_online_cpus() != 1);
2947 WARN_ON(nr_context_switches() > 0);
2948 rcu_scheduler_active
= 1;
2952 * Compute the per-level fanout, either using the exact fanout specified
2953 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2955 #ifdef CONFIG_RCU_FANOUT_EXACT
2956 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2960 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
2961 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2962 rsp
->levelspread
[0] = rcu_fanout_leaf
;
2964 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2965 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2972 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2973 ccur
= rsp
->levelcnt
[i
];
2974 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2978 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2981 * Helper function for rcu_init() that initializes one rcu_state structure.
2983 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2984 struct rcu_data __percpu
*rda
)
2986 static char *buf
[] = { "rcu_node_0",
2989 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2990 static char *fqs
[] = { "rcu_node_fqs_0",
2993 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2997 struct rcu_node
*rnp
;
2999 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3001 /* Initialize the level-tracking arrays. */
3003 for (i
= 0; i
< rcu_num_lvls
; i
++)
3004 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3005 for (i
= 1; i
< rcu_num_lvls
; i
++)
3006 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3007 rcu_init_levelspread(rsp
);
3009 /* Initialize the elements themselves, starting from the leaves. */
3011 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3012 cpustride
*= rsp
->levelspread
[i
];
3013 rnp
= rsp
->level
[i
];
3014 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3015 raw_spin_lock_init(&rnp
->lock
);
3016 lockdep_set_class_and_name(&rnp
->lock
,
3017 &rcu_node_class
[i
], buf
[i
]);
3018 raw_spin_lock_init(&rnp
->fqslock
);
3019 lockdep_set_class_and_name(&rnp
->fqslock
,
3020 &rcu_fqs_class
[i
], fqs
[i
]);
3021 rnp
->gpnum
= rsp
->gpnum
;
3022 rnp
->completed
= rsp
->completed
;
3024 rnp
->qsmaskinit
= 0;
3025 rnp
->grplo
= j
* cpustride
;
3026 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
3027 if (rnp
->grphi
>= NR_CPUS
)
3028 rnp
->grphi
= NR_CPUS
- 1;
3034 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
3035 rnp
->grpmask
= 1UL << rnp
->grpnum
;
3036 rnp
->parent
= rsp
->level
[i
- 1] +
3037 j
/ rsp
->levelspread
[i
- 1];
3040 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3045 init_waitqueue_head(&rsp
->gp_wq
);
3046 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3047 for_each_possible_cpu(i
) {
3048 while (i
> rnp
->grphi
)
3050 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3051 rcu_boot_init_percpu_data(i
, rsp
);
3053 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3057 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3058 * replace the definitions in rcutree.h because those are needed to size
3059 * the ->node array in the rcu_state structure.
3061 static void __init
rcu_init_geometry(void)
3066 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3068 /* If the compile-time values are accurate, just leave. */
3069 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3070 nr_cpu_ids
== NR_CPUS
)
3074 * Compute number of nodes that can be handled an rcu_node tree
3075 * with the given number of levels. Setting rcu_capacity[0] makes
3076 * some of the arithmetic easier.
3078 rcu_capacity
[0] = 1;
3079 rcu_capacity
[1] = rcu_fanout_leaf
;
3080 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3081 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3084 * The boot-time rcu_fanout_leaf parameter is only permitted
3085 * to increase the leaf-level fanout, not decrease it. Of course,
3086 * the leaf-level fanout cannot exceed the number of bits in
3087 * the rcu_node masks. Finally, the tree must be able to accommodate
3088 * the configured number of CPUs. Complain and fall back to the
3089 * compile-time values if these limits are exceeded.
3091 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3092 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3093 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3098 /* Calculate the number of rcu_nodes at each level of the tree. */
3099 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3100 if (n
<= rcu_capacity
[i
]) {
3101 for (j
= 0; j
<= i
; j
++)
3103 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3105 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3110 /* Calculate the total number of rcu_node structures. */
3112 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3113 rcu_num_nodes
+= num_rcu_lvl
[i
];
3117 void __init
rcu_init(void)
3121 rcu_bootup_announce();
3122 rcu_init_geometry();
3123 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3124 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3125 __rcu_init_preempt();
3127 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3130 * We don't need protection against CPU-hotplug here because
3131 * this is called early in boot, before either interrupts
3132 * or the scheduler are operational.
3134 cpu_notifier(rcu_cpu_notify
, 0);
3135 for_each_online_cpu(cpu
)
3136 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3137 check_cpu_stall_init();
3140 #include "rcutree_plugin.h"