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
];
310 * Does the current CPU require a yet-as-unscheduled grace period?
313 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
315 return *rdp
->nxttail
[RCU_DONE_TAIL
+
316 (ACCESS_ONCE(rsp
->completed
) != rdp
->completed
)] &&
317 !rcu_gp_in_progress(rsp
);
321 * Return the root node of the specified rcu_state structure.
323 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
325 return &rsp
->node
[0];
329 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
331 * If the new value of the ->dynticks_nesting counter now is zero,
332 * we really have entered idle, and must do the appropriate accounting.
333 * The caller must have disabled interrupts.
335 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
338 trace_rcu_dyntick("Start", oldval
, 0);
339 if (!user
&& !is_idle_task(current
)) {
340 struct task_struct
*idle
= idle_task(smp_processor_id());
342 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
343 ftrace_dump(DUMP_ORIG
);
344 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
345 current
->pid
, current
->comm
,
346 idle
->pid
, idle
->comm
); /* must be idle task! */
348 rcu_prepare_for_idle(smp_processor_id());
349 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
350 smp_mb__before_atomic_inc(); /* See above. */
351 atomic_inc(&rdtp
->dynticks
);
352 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
353 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
356 * It is illegal to enter an extended quiescent state while
357 * in an RCU read-side critical section.
359 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
360 "Illegal idle entry in RCU read-side critical section.");
361 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
362 "Illegal idle entry in RCU-bh read-side critical section.");
363 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
364 "Illegal idle entry in RCU-sched read-side critical section.");
368 * Enter an RCU extended quiescent state, which can be either the
369 * idle loop or adaptive-tickless usermode execution.
371 static void rcu_eqs_enter(bool user
)
374 struct rcu_dynticks
*rdtp
;
376 rdtp
= &__get_cpu_var(rcu_dynticks
);
377 oldval
= rdtp
->dynticks_nesting
;
378 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
379 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
380 rdtp
->dynticks_nesting
= 0;
382 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
383 rcu_eqs_enter_common(rdtp
, oldval
, user
);
387 * rcu_idle_enter - inform RCU that current CPU is entering idle
389 * Enter idle mode, in other words, -leave- the mode in which RCU
390 * read-side critical sections can occur. (Though RCU read-side
391 * critical sections can occur in irq handlers in idle, a possibility
392 * handled by irq_enter() and irq_exit().)
394 * We crowbar the ->dynticks_nesting field to zero to allow for
395 * the possibility of usermode upcalls having messed up our count
396 * of interrupt nesting level during the prior busy period.
398 void rcu_idle_enter(void)
402 local_irq_save(flags
);
403 rcu_eqs_enter(false);
404 local_irq_restore(flags
);
406 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
408 #ifdef CONFIG_RCU_USER_QS
410 * rcu_user_enter - inform RCU that we are resuming userspace.
412 * Enter RCU idle mode right before resuming userspace. No use of RCU
413 * is permitted between this call and rcu_user_exit(). This way the
414 * CPU doesn't need to maintain the tick for RCU maintenance purposes
415 * when the CPU runs in userspace.
417 void rcu_user_enter(void)
420 struct rcu_dynticks
*rdtp
;
423 * Some contexts may involve an exception occuring in an irq,
424 * leading to that nesting:
425 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
426 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
427 * helpers are enough to protect RCU uses inside the exception. So
428 * just return immediately if we detect we are in an IRQ.
433 WARN_ON_ONCE(!current
->mm
);
435 local_irq_save(flags
);
436 rdtp
= &__get_cpu_var(rcu_dynticks
);
437 if (!rdtp
->ignore_user_qs
&& !rdtp
->in_user
) {
438 rdtp
->in_user
= true;
441 local_irq_restore(flags
);
445 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
446 * after the current irq returns.
448 * This is similar to rcu_user_enter() but in the context of a non-nesting
449 * irq. After this call, RCU enters into idle mode when the interrupt
452 void rcu_user_enter_after_irq(void)
455 struct rcu_dynticks
*rdtp
;
457 local_irq_save(flags
);
458 rdtp
= &__get_cpu_var(rcu_dynticks
);
459 /* Ensure this irq is interrupting a non-idle RCU state. */
460 WARN_ON_ONCE(!(rdtp
->dynticks_nesting
& DYNTICK_TASK_MASK
));
461 rdtp
->dynticks_nesting
= 1;
462 local_irq_restore(flags
);
464 #endif /* CONFIG_RCU_USER_QS */
467 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
469 * Exit from an interrupt handler, which might possibly result in entering
470 * idle mode, in other words, leaving the mode in which read-side critical
471 * sections can occur.
473 * This code assumes that the idle loop never does anything that might
474 * result in unbalanced calls to irq_enter() and irq_exit(). If your
475 * architecture violates this assumption, RCU will give you what you
476 * deserve, good and hard. But very infrequently and irreproducibly.
478 * Use things like work queues to work around this limitation.
480 * You have been warned.
482 void rcu_irq_exit(void)
486 struct rcu_dynticks
*rdtp
;
488 local_irq_save(flags
);
489 rdtp
= &__get_cpu_var(rcu_dynticks
);
490 oldval
= rdtp
->dynticks_nesting
;
491 rdtp
->dynticks_nesting
--;
492 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
493 if (rdtp
->dynticks_nesting
)
494 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
496 rcu_eqs_enter_common(rdtp
, oldval
, true);
497 local_irq_restore(flags
);
501 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
503 * If the new value of the ->dynticks_nesting counter was previously zero,
504 * we really have exited idle, and must do the appropriate accounting.
505 * The caller must have disabled interrupts.
507 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
510 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
511 atomic_inc(&rdtp
->dynticks
);
512 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
513 smp_mb__after_atomic_inc(); /* See above. */
514 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
515 rcu_cleanup_after_idle(smp_processor_id());
516 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
517 if (!user
&& !is_idle_task(current
)) {
518 struct task_struct
*idle
= idle_task(smp_processor_id());
520 trace_rcu_dyntick("Error on exit: not idle task",
521 oldval
, rdtp
->dynticks_nesting
);
522 ftrace_dump(DUMP_ORIG
);
523 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
524 current
->pid
, current
->comm
,
525 idle
->pid
, idle
->comm
); /* must be idle task! */
530 * Exit an RCU extended quiescent state, which can be either the
531 * idle loop or adaptive-tickless usermode execution.
533 static void rcu_eqs_exit(bool user
)
535 struct rcu_dynticks
*rdtp
;
538 rdtp
= &__get_cpu_var(rcu_dynticks
);
539 oldval
= rdtp
->dynticks_nesting
;
540 WARN_ON_ONCE(oldval
< 0);
541 if (oldval
& DYNTICK_TASK_NEST_MASK
)
542 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
544 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
545 rcu_eqs_exit_common(rdtp
, oldval
, user
);
549 * rcu_idle_exit - inform RCU that current CPU is leaving idle
551 * Exit idle mode, in other words, -enter- the mode in which RCU
552 * read-side critical sections can occur.
554 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
555 * allow for the possibility of usermode upcalls messing up our count
556 * of interrupt nesting level during the busy period that is just
559 void rcu_idle_exit(void)
563 local_irq_save(flags
);
565 local_irq_restore(flags
);
567 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
569 #ifdef CONFIG_RCU_USER_QS
571 * rcu_user_exit - inform RCU that we are exiting userspace.
573 * Exit RCU idle mode while entering the kernel because it can
574 * run a RCU read side critical section anytime.
576 void rcu_user_exit(void)
579 struct rcu_dynticks
*rdtp
;
582 * Some contexts may involve an exception occuring in an irq,
583 * leading to that nesting:
584 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
585 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
586 * helpers are enough to protect RCU uses inside the exception. So
587 * just return immediately if we detect we are in an IRQ.
592 local_irq_save(flags
);
593 rdtp
= &__get_cpu_var(rcu_dynticks
);
595 rdtp
->in_user
= false;
598 local_irq_restore(flags
);
602 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
603 * idle mode after the current non-nesting irq returns.
605 * This is similar to rcu_user_exit() but in the context of an irq.
606 * This is called when the irq has interrupted a userspace RCU idle mode
607 * context. When the current non-nesting interrupt returns after this call,
608 * the CPU won't restore the RCU idle mode.
610 void rcu_user_exit_after_irq(void)
613 struct rcu_dynticks
*rdtp
;
615 local_irq_save(flags
);
616 rdtp
= &__get_cpu_var(rcu_dynticks
);
617 /* Ensure we are interrupting an RCU idle mode. */
618 WARN_ON_ONCE(rdtp
->dynticks_nesting
& DYNTICK_TASK_NEST_MASK
);
619 rdtp
->dynticks_nesting
+= DYNTICK_TASK_EXIT_IDLE
;
620 local_irq_restore(flags
);
622 #endif /* CONFIG_RCU_USER_QS */
625 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
627 * Enter an interrupt handler, which might possibly result in exiting
628 * idle mode, in other words, entering the mode in which read-side critical
629 * sections can occur.
631 * Note that the Linux kernel is fully capable of entering an interrupt
632 * handler that it never exits, for example when doing upcalls to
633 * user mode! This code assumes that the idle loop never does upcalls to
634 * user mode. If your architecture does do upcalls from the idle loop (or
635 * does anything else that results in unbalanced calls to the irq_enter()
636 * and irq_exit() functions), RCU will give you what you deserve, good
637 * and hard. But very infrequently and irreproducibly.
639 * Use things like work queues to work around this limitation.
641 * You have been warned.
643 void rcu_irq_enter(void)
646 struct rcu_dynticks
*rdtp
;
649 local_irq_save(flags
);
650 rdtp
= &__get_cpu_var(rcu_dynticks
);
651 oldval
= rdtp
->dynticks_nesting
;
652 rdtp
->dynticks_nesting
++;
653 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
655 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
657 rcu_eqs_exit_common(rdtp
, oldval
, true);
658 local_irq_restore(flags
);
662 * rcu_nmi_enter - inform RCU of entry to NMI context
664 * If the CPU was idle with dynamic ticks active, and there is no
665 * irq handler running, this updates rdtp->dynticks_nmi to let the
666 * RCU grace-period handling know that the CPU is active.
668 void rcu_nmi_enter(void)
670 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
672 if (rdtp
->dynticks_nmi_nesting
== 0 &&
673 (atomic_read(&rdtp
->dynticks
) & 0x1))
675 rdtp
->dynticks_nmi_nesting
++;
676 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
677 atomic_inc(&rdtp
->dynticks
);
678 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
679 smp_mb__after_atomic_inc(); /* See above. */
680 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
684 * rcu_nmi_exit - inform RCU of exit from NMI context
686 * If the CPU was idle with dynamic ticks active, and there is no
687 * irq handler running, this updates rdtp->dynticks_nmi to let the
688 * RCU grace-period handling know that the CPU is no longer active.
690 void rcu_nmi_exit(void)
692 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
694 if (rdtp
->dynticks_nmi_nesting
== 0 ||
695 --rdtp
->dynticks_nmi_nesting
!= 0)
697 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
698 smp_mb__before_atomic_inc(); /* See above. */
699 atomic_inc(&rdtp
->dynticks
);
700 smp_mb__after_atomic_inc(); /* Force delay to next write. */
701 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
705 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
707 * If the current CPU is in its idle loop and is neither in an interrupt
708 * or NMI handler, return true.
710 int rcu_is_cpu_idle(void)
715 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
719 EXPORT_SYMBOL(rcu_is_cpu_idle
);
721 #ifdef CONFIG_RCU_USER_QS
722 void rcu_user_hooks_switch(struct task_struct
*prev
,
723 struct task_struct
*next
)
725 struct rcu_dynticks
*rdtp
;
727 /* Interrupts are disabled in context switch */
728 rdtp
= &__get_cpu_var(rcu_dynticks
);
729 if (!rdtp
->ignore_user_qs
) {
730 clear_tsk_thread_flag(prev
, TIF_NOHZ
);
731 set_tsk_thread_flag(next
, TIF_NOHZ
);
734 #endif /* #ifdef CONFIG_RCU_USER_QS */
736 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
739 * Is the current CPU online? Disable preemption to avoid false positives
740 * that could otherwise happen due to the current CPU number being sampled,
741 * this task being preempted, its old CPU being taken offline, resuming
742 * on some other CPU, then determining that its old CPU is now offline.
743 * It is OK to use RCU on an offline processor during initial boot, hence
744 * the check for rcu_scheduler_fully_active. Note also that it is OK
745 * for a CPU coming online to use RCU for one jiffy prior to marking itself
746 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
747 * offline to continue to use RCU for one jiffy after marking itself
748 * offline in the cpu_online_mask. This leniency is necessary given the
749 * non-atomic nature of the online and offline processing, for example,
750 * the fact that a CPU enters the scheduler after completing the CPU_DYING
753 * This is also why RCU internally marks CPUs online during the
754 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
756 * Disable checking if in an NMI handler because we cannot safely report
757 * errors from NMI handlers anyway.
759 bool rcu_lockdep_current_cpu_online(void)
761 struct rcu_data
*rdp
;
762 struct rcu_node
*rnp
;
768 rdp
= &__get_cpu_var(rcu_sched_data
);
770 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
771 !rcu_scheduler_fully_active
;
775 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
777 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
780 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
782 * If the current CPU is idle or running at a first-level (not nested)
783 * interrupt from idle, return true. The caller must have at least
784 * disabled preemption.
786 int rcu_is_cpu_rrupt_from_idle(void)
788 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
792 * Snapshot the specified CPU's dynticks counter so that we can later
793 * credit them with an implicit quiescent state. Return 1 if this CPU
794 * is in dynticks idle mode, which is an extended quiescent state.
796 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
798 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
799 return (rdp
->dynticks_snap
& 0x1) == 0;
803 * Return true if the specified CPU has passed through a quiescent
804 * state by virtue of being in or having passed through an dynticks
805 * idle state since the last call to dyntick_save_progress_counter()
806 * for this same CPU, or by virtue of having been offline.
808 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
813 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
814 snap
= (unsigned int)rdp
->dynticks_snap
;
817 * If the CPU passed through or entered a dynticks idle phase with
818 * no active irq/NMI handlers, then we can safely pretend that the CPU
819 * already acknowledged the request to pass through a quiescent
820 * state. Either way, that CPU cannot possibly be in an RCU
821 * read-side critical section that started before the beginning
822 * of the current RCU grace period.
824 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
825 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
831 * Check for the CPU being offline, but only if the grace period
832 * is old enough. We don't need to worry about the CPU changing
833 * state: If we see it offline even once, it has been through a
836 * The reason for insisting that the grace period be at least
837 * one jiffy old is that CPUs that are not quite online and that
838 * have just gone offline can still execute RCU read-side critical
841 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
842 return 0; /* Grace period is not old enough. */
844 if (cpu_is_offline(rdp
->cpu
)) {
845 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
852 static int jiffies_till_stall_check(void)
854 int till_stall_check
= ACCESS_ONCE(rcu_cpu_stall_timeout
);
857 * Limit check must be consistent with the Kconfig limits
858 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
860 if (till_stall_check
< 3) {
861 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 3;
862 till_stall_check
= 3;
863 } else if (till_stall_check
> 300) {
864 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 300;
865 till_stall_check
= 300;
867 return till_stall_check
* HZ
+ RCU_STALL_DELAY_DELTA
;
870 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
872 rsp
->gp_start
= jiffies
;
873 rsp
->jiffies_stall
= jiffies
+ jiffies_till_stall_check();
877 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
878 * for architectures that do not implement trigger_all_cpu_backtrace().
879 * The NMI-triggered stack traces are more accurate because they are
880 * printed by the target CPU.
882 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
886 struct rcu_node
*rnp
;
888 rcu_for_each_leaf_node(rsp
, rnp
) {
889 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
890 if (rnp
->qsmask
!= 0) {
891 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
892 if (rnp
->qsmask
& (1UL << cpu
))
893 dump_cpu_task(rnp
->grplo
+ cpu
);
895 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
899 static void print_other_cpu_stall(struct rcu_state
*rsp
)
905 struct rcu_node
*rnp
= rcu_get_root(rsp
);
908 /* Only let one CPU complain about others per time interval. */
910 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
911 delta
= jiffies
- rsp
->jiffies_stall
;
912 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
913 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
916 rsp
->jiffies_stall
= jiffies
+ 3 * jiffies_till_stall_check() + 3;
917 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
920 * OK, time to rat on our buddy...
921 * See Documentation/RCU/stallwarn.txt for info on how to debug
922 * RCU CPU stall warnings.
924 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
926 print_cpu_stall_info_begin();
927 rcu_for_each_leaf_node(rsp
, rnp
) {
928 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
929 ndetected
+= rcu_print_task_stall(rnp
);
930 if (rnp
->qsmask
!= 0) {
931 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
932 if (rnp
->qsmask
& (1UL << cpu
)) {
933 print_cpu_stall_info(rsp
,
938 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
942 * Now rat on any tasks that got kicked up to the root rcu_node
943 * due to CPU offlining.
945 rnp
= rcu_get_root(rsp
);
946 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
947 ndetected
+= rcu_print_task_stall(rnp
);
948 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
950 print_cpu_stall_info_end();
951 for_each_possible_cpu(cpu
)
952 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
953 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
954 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
955 rsp
->gpnum
, rsp
->completed
, totqlen
);
957 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
958 else if (!trigger_all_cpu_backtrace())
959 rcu_dump_cpu_stacks(rsp
);
961 /* Complain about tasks blocking the grace period. */
963 rcu_print_detail_task_stall(rsp
);
965 force_quiescent_state(rsp
); /* Kick them all. */
968 static void print_cpu_stall(struct rcu_state
*rsp
)
972 struct rcu_node
*rnp
= rcu_get_root(rsp
);
976 * OK, time to rat on ourselves...
977 * See Documentation/RCU/stallwarn.txt for info on how to debug
978 * RCU CPU stall warnings.
980 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
981 print_cpu_stall_info_begin();
982 print_cpu_stall_info(rsp
, smp_processor_id());
983 print_cpu_stall_info_end();
984 for_each_possible_cpu(cpu
)
985 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
986 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
987 jiffies
- rsp
->gp_start
, rsp
->gpnum
, rsp
->completed
, totqlen
);
988 if (!trigger_all_cpu_backtrace())
991 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
992 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
993 rsp
->jiffies_stall
= jiffies
+
994 3 * jiffies_till_stall_check() + 3;
995 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
997 set_need_resched(); /* kick ourselves to get things going. */
1000 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1004 struct rcu_node
*rnp
;
1006 if (rcu_cpu_stall_suppress
)
1008 j
= ACCESS_ONCE(jiffies
);
1009 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
1011 if (rcu_gp_in_progress(rsp
) &&
1012 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
1014 /* We haven't checked in, so go dump stack. */
1015 print_cpu_stall(rsp
);
1017 } else if (rcu_gp_in_progress(rsp
) &&
1018 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1020 /* They had a few time units to dump stack, so complain. */
1021 print_other_cpu_stall(rsp
);
1025 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
1027 rcu_cpu_stall_suppress
= 1;
1032 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1034 * Set the stall-warning timeout way off into the future, thus preventing
1035 * any RCU CPU stall-warning messages from appearing in the current set of
1036 * RCU grace periods.
1038 * The caller must disable hard irqs.
1040 void rcu_cpu_stall_reset(void)
1042 struct rcu_state
*rsp
;
1044 for_each_rcu_flavor(rsp
)
1045 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
1048 static struct notifier_block rcu_panic_block
= {
1049 .notifier_call
= rcu_panic
,
1052 static void __init
check_cpu_stall_init(void)
1054 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
1058 * Update CPU-local rcu_data state to record the newly noticed grace period.
1059 * This is used both when we started the grace period and when we notice
1060 * that someone else started the grace period. The caller must hold the
1061 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
1062 * and must have irqs disabled.
1064 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1066 if (rdp
->gpnum
!= rnp
->gpnum
) {
1068 * If the current grace period is waiting for this CPU,
1069 * set up to detect a quiescent state, otherwise don't
1070 * go looking for one.
1072 rdp
->gpnum
= rnp
->gpnum
;
1073 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
1074 rdp
->passed_quiesce
= 0;
1075 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1076 zero_cpu_stall_ticks(rdp
);
1080 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1082 unsigned long flags
;
1083 struct rcu_node
*rnp
;
1085 local_irq_save(flags
);
1087 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
1088 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1089 local_irq_restore(flags
);
1092 __note_new_gpnum(rsp
, rnp
, rdp
);
1093 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1097 * Did someone else start a new RCU grace period start since we last
1098 * checked? Update local state appropriately if so. Must be called
1099 * on the CPU corresponding to rdp.
1102 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1104 unsigned long flags
;
1107 local_irq_save(flags
);
1108 if (rdp
->gpnum
!= rsp
->gpnum
) {
1109 note_new_gpnum(rsp
, rdp
);
1112 local_irq_restore(flags
);
1117 * Initialize the specified rcu_data structure's callback list to empty.
1119 static void init_callback_list(struct rcu_data
*rdp
)
1123 rdp
->nxtlist
= NULL
;
1124 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1125 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1129 * Advance this CPU's callbacks, but only if the current grace period
1130 * has ended. This may be called only from the CPU to whom the rdp
1131 * belongs. In addition, the corresponding leaf rcu_node structure's
1132 * ->lock must be held by the caller, with irqs disabled.
1135 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1137 /* Did another grace period end? */
1138 if (rdp
->completed
!= rnp
->completed
) {
1140 /* Advance callbacks. No harm if list empty. */
1141 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
1142 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
1143 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1145 /* Remember that we saw this grace-period completion. */
1146 rdp
->completed
= rnp
->completed
;
1147 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
1150 * If we were in an extended quiescent state, we may have
1151 * missed some grace periods that others CPUs handled on
1152 * our behalf. Catch up with this state to avoid noting
1153 * spurious new grace periods. If another grace period
1154 * has started, then rnp->gpnum will have advanced, so
1155 * we will detect this later on. Of course, any quiescent
1156 * states we found for the old GP are now invalid.
1158 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
)) {
1159 rdp
->gpnum
= rdp
->completed
;
1160 rdp
->passed_quiesce
= 0;
1164 * If RCU does not need a quiescent state from this CPU,
1165 * then make sure that this CPU doesn't go looking for one.
1167 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1168 rdp
->qs_pending
= 0;
1173 * Advance this CPU's callbacks, but only if the current grace period
1174 * has ended. This may be called only from the CPU to whom the rdp
1178 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1180 unsigned long flags
;
1181 struct rcu_node
*rnp
;
1183 local_irq_save(flags
);
1185 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1186 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1187 local_irq_restore(flags
);
1190 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1191 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1195 * Do per-CPU grace-period initialization for running CPU. The caller
1196 * must hold the lock of the leaf rcu_node structure corresponding to
1200 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1202 /* Prior grace period ended, so advance callbacks for current CPU. */
1203 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1205 /* Set state so that this CPU will detect the next quiescent state. */
1206 __note_new_gpnum(rsp
, rnp
, rdp
);
1210 * Initialize a new grace period.
1212 static int rcu_gp_init(struct rcu_state
*rsp
)
1214 struct rcu_data
*rdp
;
1215 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1217 raw_spin_lock_irq(&rnp
->lock
);
1218 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1220 if (rcu_gp_in_progress(rsp
)) {
1221 /* Grace period already in progress, don't start another. */
1222 raw_spin_unlock_irq(&rnp
->lock
);
1226 /* Advance to a new grace period and initialize state. */
1228 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1229 record_gp_stall_check_time(rsp
);
1230 raw_spin_unlock_irq(&rnp
->lock
);
1232 /* Exclude any concurrent CPU-hotplug operations. */
1233 mutex_lock(&rsp
->onoff_mutex
);
1236 * Set the quiescent-state-needed bits in all the rcu_node
1237 * structures for all currently online CPUs in breadth-first order,
1238 * starting from the root rcu_node structure, relying on the layout
1239 * of the tree within the rsp->node[] array. Note that other CPUs
1240 * will access only the leaves of the hierarchy, thus seeing that no
1241 * grace period is in progress, at least until the corresponding
1242 * leaf node has been initialized. In addition, we have excluded
1243 * CPU-hotplug operations.
1245 * The grace period cannot complete until the initialization
1246 * process finishes, because this kthread handles both.
1248 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1249 raw_spin_lock_irq(&rnp
->lock
);
1250 rdp
= this_cpu_ptr(rsp
->rda
);
1251 rcu_preempt_check_blocked_tasks(rnp
);
1252 rnp
->qsmask
= rnp
->qsmaskinit
;
1253 rnp
->gpnum
= rsp
->gpnum
;
1254 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1255 rnp
->completed
= rsp
->completed
;
1256 if (rnp
== rdp
->mynode
)
1257 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1258 rcu_preempt_boost_start_gp(rnp
);
1259 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1260 rnp
->level
, rnp
->grplo
,
1261 rnp
->grphi
, rnp
->qsmask
);
1262 raw_spin_unlock_irq(&rnp
->lock
);
1263 #ifdef CONFIG_PROVE_RCU_DELAY
1264 if ((random32() % (rcu_num_nodes
* 8)) == 0)
1265 schedule_timeout_uninterruptible(2);
1266 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1270 mutex_unlock(&rsp
->onoff_mutex
);
1275 * Do one round of quiescent-state forcing.
1277 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1279 int fqs_state
= fqs_state_in
;
1280 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1283 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1284 /* Collect dyntick-idle snapshots. */
1285 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1286 fqs_state
= RCU_FORCE_QS
;
1288 /* Handle dyntick-idle and offline CPUs. */
1289 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1291 /* Clear flag to prevent immediate re-entry. */
1292 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1293 raw_spin_lock_irq(&rnp
->lock
);
1294 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1295 raw_spin_unlock_irq(&rnp
->lock
);
1301 * Clean up after the old grace period.
1303 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1305 unsigned long gp_duration
;
1306 struct rcu_data
*rdp
;
1307 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1309 raw_spin_lock_irq(&rnp
->lock
);
1310 gp_duration
= jiffies
- rsp
->gp_start
;
1311 if (gp_duration
> rsp
->gp_max
)
1312 rsp
->gp_max
= gp_duration
;
1315 * We know the grace period is complete, but to everyone else
1316 * it appears to still be ongoing. But it is also the case
1317 * that to everyone else it looks like there is nothing that
1318 * they can do to advance the grace period. It is therefore
1319 * safe for us to drop the lock in order to mark the grace
1320 * period as completed in all of the rcu_node structures.
1322 raw_spin_unlock_irq(&rnp
->lock
);
1325 * Propagate new ->completed value to rcu_node structures so
1326 * that other CPUs don't have to wait until the start of the next
1327 * grace period to process their callbacks. This also avoids
1328 * some nasty RCU grace-period initialization races by forcing
1329 * the end of the current grace period to be completely recorded in
1330 * all of the rcu_node structures before the beginning of the next
1331 * grace period is recorded in any of the rcu_node structures.
1333 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1334 raw_spin_lock_irq(&rnp
->lock
);
1335 rnp
->completed
= rsp
->gpnum
;
1336 raw_spin_unlock_irq(&rnp
->lock
);
1339 rnp
= rcu_get_root(rsp
);
1340 raw_spin_lock_irq(&rnp
->lock
);
1342 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1343 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1344 rsp
->fqs_state
= RCU_GP_IDLE
;
1345 rdp
= this_cpu_ptr(rsp
->rda
);
1346 if (cpu_needs_another_gp(rsp
, rdp
))
1348 raw_spin_unlock_irq(&rnp
->lock
);
1352 * Body of kthread that handles grace periods.
1354 static int __noreturn
rcu_gp_kthread(void *arg
)
1359 struct rcu_state
*rsp
= arg
;
1360 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1364 /* Handle grace-period start. */
1366 wait_event_interruptible(rsp
->gp_wq
,
1369 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1373 flush_signals(current
);
1376 /* Handle quiescent-state forcing. */
1377 fqs_state
= RCU_SAVE_DYNTICK
;
1378 j
= jiffies_till_first_fqs
;
1381 jiffies_till_first_fqs
= HZ
;
1384 rsp
->jiffies_force_qs
= jiffies
+ j
;
1385 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1386 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1387 (!ACCESS_ONCE(rnp
->qsmask
) &&
1388 !rcu_preempt_blocked_readers_cgp(rnp
)),
1390 /* If grace period done, leave loop. */
1391 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1392 !rcu_preempt_blocked_readers_cgp(rnp
))
1394 /* If time for quiescent-state forcing, do it. */
1395 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1396 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1399 /* Deal with stray signal. */
1401 flush_signals(current
);
1403 j
= jiffies_till_next_fqs
;
1406 jiffies_till_next_fqs
= HZ
;
1409 jiffies_till_next_fqs
= 1;
1413 /* Handle grace-period end. */
1414 rcu_gp_cleanup(rsp
);
1419 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1420 * in preparation for detecting the next grace period. The caller must hold
1421 * the root node's ->lock, which is released before return. Hard irqs must
1424 * Note that it is legal for a dying CPU (which is marked as offline) to
1425 * invoke this function. This can happen when the dying CPU reports its
1429 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1430 __releases(rcu_get_root(rsp
)->lock
)
1432 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1433 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1435 if (!rsp
->gp_kthread
||
1436 !cpu_needs_another_gp(rsp
, rdp
)) {
1438 * Either we have not yet spawned the grace-period
1439 * task, this CPU does not need another grace period,
1440 * or a grace period is already in progress.
1441 * Either way, don't start a new grace period.
1443 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1448 * Because there is no grace period in progress right now,
1449 * any callbacks we have up to this point will be satisfied
1450 * by the next grace period. So promote all callbacks to be
1451 * handled after the end of the next grace period. If the
1452 * CPU is not yet aware of the end of the previous grace period,
1453 * we need to allow for the callback advancement that will
1454 * occur when it does become aware. Deadlock prevents us from
1455 * making it aware at this point: We cannot acquire a leaf
1456 * rcu_node ->lock while holding the root rcu_node ->lock.
1458 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1459 if (rdp
->completed
== rsp
->completed
)
1460 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1462 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1463 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
1465 /* Ensure that CPU is aware of completion of last grace period. */
1466 rcu_process_gp_end(rsp
, rdp
);
1467 local_irq_restore(flags
);
1469 /* Wake up rcu_gp_kthread() to start the grace period. */
1470 wake_up(&rsp
->gp_wq
);
1474 * Report a full set of quiescent states to the specified rcu_state
1475 * data structure. This involves cleaning up after the prior grace
1476 * period and letting rcu_start_gp() start up the next grace period
1477 * if one is needed. Note that the caller must hold rnp->lock, as
1478 * required by rcu_start_gp(), which will release it.
1480 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1481 __releases(rcu_get_root(rsp
)->lock
)
1483 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1484 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1485 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1489 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1490 * Allows quiescent states for a group of CPUs to be reported at one go
1491 * to the specified rcu_node structure, though all the CPUs in the group
1492 * must be represented by the same rcu_node structure (which need not be
1493 * a leaf rcu_node structure, though it often will be). That structure's
1494 * lock must be held upon entry, and it is released before return.
1497 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1498 struct rcu_node
*rnp
, unsigned long flags
)
1499 __releases(rnp
->lock
)
1501 struct rcu_node
*rnp_c
;
1503 /* Walk up the rcu_node hierarchy. */
1505 if (!(rnp
->qsmask
& mask
)) {
1507 /* Our bit has already been cleared, so done. */
1508 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1511 rnp
->qsmask
&= ~mask
;
1512 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1513 mask
, rnp
->qsmask
, rnp
->level
,
1514 rnp
->grplo
, rnp
->grphi
,
1516 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1518 /* Other bits still set at this level, so done. */
1519 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1522 mask
= rnp
->grpmask
;
1523 if (rnp
->parent
== NULL
) {
1525 /* No more levels. Exit loop holding root lock. */
1529 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1532 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1533 WARN_ON_ONCE(rnp_c
->qsmask
);
1537 * Get here if we are the last CPU to pass through a quiescent
1538 * state for this grace period. Invoke rcu_report_qs_rsp()
1539 * to clean up and start the next grace period if one is needed.
1541 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1545 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1546 * structure. This must be either called from the specified CPU, or
1547 * called when the specified CPU is known to be offline (and when it is
1548 * also known that no other CPU is concurrently trying to help the offline
1549 * CPU). The lastcomp argument is used to make sure we are still in the
1550 * grace period of interest. We don't want to end the current grace period
1551 * based on quiescent states detected in an earlier grace period!
1554 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1556 unsigned long flags
;
1558 struct rcu_node
*rnp
;
1561 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1562 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1563 rnp
->completed
== rnp
->gpnum
) {
1566 * The grace period in which this quiescent state was
1567 * recorded has ended, so don't report it upwards.
1568 * We will instead need a new quiescent state that lies
1569 * within the current grace period.
1571 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1572 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1575 mask
= rdp
->grpmask
;
1576 if ((rnp
->qsmask
& mask
) == 0) {
1577 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1579 rdp
->qs_pending
= 0;
1582 * This GP can't end until cpu checks in, so all of our
1583 * callbacks can be processed during the next GP.
1585 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1587 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1592 * Check to see if there is a new grace period of which this CPU
1593 * is not yet aware, and if so, set up local rcu_data state for it.
1594 * Otherwise, see if this CPU has just passed through its first
1595 * quiescent state for this grace period, and record that fact if so.
1598 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1600 /* If there is now a new grace period, record and return. */
1601 if (check_for_new_grace_period(rsp
, rdp
))
1605 * Does this CPU still need to do its part for current grace period?
1606 * If no, return and let the other CPUs do their part as well.
1608 if (!rdp
->qs_pending
)
1612 * Was there a quiescent state since the beginning of the grace
1613 * period? If no, then exit and wait for the next call.
1615 if (!rdp
->passed_quiesce
)
1619 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1622 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1625 #ifdef CONFIG_HOTPLUG_CPU
1628 * Send the specified CPU's RCU callbacks to the orphanage. The
1629 * specified CPU must be offline, and the caller must hold the
1633 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1634 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1637 * Orphan the callbacks. First adjust the counts. This is safe
1638 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1639 * cannot be running now. Thus no memory barrier is required.
1641 if (rdp
->nxtlist
!= NULL
) {
1642 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1643 rsp
->qlen
+= rdp
->qlen
;
1644 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1646 ACCESS_ONCE(rdp
->qlen
) = 0;
1650 * Next, move those callbacks still needing a grace period to
1651 * the orphanage, where some other CPU will pick them up.
1652 * Some of the callbacks might have gone partway through a grace
1653 * period, but that is too bad. They get to start over because we
1654 * cannot assume that grace periods are synchronized across CPUs.
1655 * We don't bother updating the ->nxttail[] array yet, instead
1656 * we just reset the whole thing later on.
1658 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1659 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1660 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1661 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1665 * Then move the ready-to-invoke callbacks to the orphanage,
1666 * where some other CPU will pick them up. These will not be
1667 * required to pass though another grace period: They are done.
1669 if (rdp
->nxtlist
!= NULL
) {
1670 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1671 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1674 /* Finally, initialize the rcu_data structure's list to empty. */
1675 init_callback_list(rdp
);
1679 * Adopt the RCU callbacks from the specified rcu_state structure's
1680 * orphanage. The caller must hold the ->orphan_lock.
1682 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1685 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1687 /* Do the accounting first. */
1688 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1689 rdp
->qlen
+= rsp
->qlen
;
1690 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1691 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1692 rcu_idle_count_callbacks_posted();
1697 * We do not need a memory barrier here because the only way we
1698 * can get here if there is an rcu_barrier() in flight is if
1699 * we are the task doing the rcu_barrier().
1702 /* First adopt the ready-to-invoke callbacks. */
1703 if (rsp
->orphan_donelist
!= NULL
) {
1704 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1705 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1706 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1707 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1708 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1709 rsp
->orphan_donelist
= NULL
;
1710 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1713 /* And then adopt the callbacks that still need a grace period. */
1714 if (rsp
->orphan_nxtlist
!= NULL
) {
1715 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1716 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1717 rsp
->orphan_nxtlist
= NULL
;
1718 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1723 * Trace the fact that this CPU is going offline.
1725 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1727 RCU_TRACE(unsigned long mask
);
1728 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1729 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1731 RCU_TRACE(mask
= rdp
->grpmask
);
1732 trace_rcu_grace_period(rsp
->name
,
1733 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1738 * The CPU has been completely removed, and some other CPU is reporting
1739 * this fact from process context. Do the remainder of the cleanup,
1740 * including orphaning the outgoing CPU's RCU callbacks, and also
1741 * adopting them. There can only be one CPU hotplug operation at a time,
1742 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1744 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1746 unsigned long flags
;
1748 int need_report
= 0;
1749 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1750 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1752 /* Adjust any no-longer-needed kthreads. */
1753 rcu_boost_kthread_setaffinity(rnp
, -1);
1755 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1757 /* Exclude any attempts to start a new grace period. */
1758 mutex_lock(&rsp
->onoff_mutex
);
1759 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
1761 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1762 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1763 rcu_adopt_orphan_cbs(rsp
);
1765 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1766 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1768 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1769 rnp
->qsmaskinit
&= ~mask
;
1770 if (rnp
->qsmaskinit
!= 0) {
1771 if (rnp
!= rdp
->mynode
)
1772 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1775 if (rnp
== rdp
->mynode
)
1776 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1778 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1779 mask
= rnp
->grpmask
;
1781 } while (rnp
!= NULL
);
1784 * We still hold the leaf rcu_node structure lock here, and
1785 * irqs are still disabled. The reason for this subterfuge is
1786 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1787 * held leads to deadlock.
1789 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
1791 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1792 rcu_report_unblock_qs_rnp(rnp
, flags
);
1794 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1795 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1796 rcu_report_exp_rnp(rsp
, rnp
, true);
1797 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1798 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1799 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1800 init_callback_list(rdp
);
1801 /* Disallow further callbacks on this CPU. */
1802 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1803 mutex_unlock(&rsp
->onoff_mutex
);
1806 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1808 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1812 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1816 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1819 * Invoke any RCU callbacks that have made it to the end of their grace
1820 * period. Thottle as specified by rdp->blimit.
1822 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1824 unsigned long flags
;
1825 struct rcu_head
*next
, *list
, **tail
;
1826 long bl
, count
, count_lazy
;
1829 /* If no callbacks are ready, just return.*/
1830 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1831 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1832 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1833 need_resched(), is_idle_task(current
),
1834 rcu_is_callbacks_kthread());
1839 * Extract the list of ready callbacks, disabling to prevent
1840 * races with call_rcu() from interrupt handlers.
1842 local_irq_save(flags
);
1843 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1845 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1846 list
= rdp
->nxtlist
;
1847 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1848 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1849 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1850 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1851 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1852 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1853 local_irq_restore(flags
);
1855 /* Invoke callbacks. */
1856 count
= count_lazy
= 0;
1860 debug_rcu_head_unqueue(list
);
1861 if (__rcu_reclaim(rsp
->name
, list
))
1864 /* Stop only if limit reached and CPU has something to do. */
1865 if (++count
>= bl
&&
1867 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1871 local_irq_save(flags
);
1872 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1873 is_idle_task(current
),
1874 rcu_is_callbacks_kthread());
1876 /* Update count, and requeue any remaining callbacks. */
1878 *tail
= rdp
->nxtlist
;
1879 rdp
->nxtlist
= list
;
1880 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1881 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1882 rdp
->nxttail
[i
] = tail
;
1886 smp_mb(); /* List handling before counting for rcu_barrier(). */
1887 rdp
->qlen_lazy
-= count_lazy
;
1888 ACCESS_ONCE(rdp
->qlen
) -= count
;
1889 rdp
->n_cbs_invoked
+= count
;
1891 /* Reinstate batch limit if we have worked down the excess. */
1892 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1893 rdp
->blimit
= blimit
;
1895 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1896 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1897 rdp
->qlen_last_fqs_check
= 0;
1898 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1899 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1900 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1901 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
1903 local_irq_restore(flags
);
1905 /* Re-invoke RCU core processing if there are callbacks remaining. */
1906 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1911 * Check to see if this CPU is in a non-context-switch quiescent state
1912 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1913 * Also schedule RCU core processing.
1915 * This function must be called from hardirq context. It is normally
1916 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1917 * false, there is no point in invoking rcu_check_callbacks().
1919 void rcu_check_callbacks(int cpu
, int user
)
1921 trace_rcu_utilization("Start scheduler-tick");
1922 increment_cpu_stall_ticks();
1923 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1926 * Get here if this CPU took its interrupt from user
1927 * mode or from the idle loop, and if this is not a
1928 * nested interrupt. In this case, the CPU is in
1929 * a quiescent state, so note it.
1931 * No memory barrier is required here because both
1932 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1933 * variables that other CPUs neither access nor modify,
1934 * at least not while the corresponding CPU is online.
1940 } else if (!in_softirq()) {
1943 * Get here if this CPU did not take its interrupt from
1944 * softirq, in other words, if it is not interrupting
1945 * a rcu_bh read-side critical section. This is an _bh
1946 * critical section, so note it.
1951 rcu_preempt_check_callbacks(cpu
);
1952 if (rcu_pending(cpu
))
1954 trace_rcu_utilization("End scheduler-tick");
1958 * Scan the leaf rcu_node structures, processing dyntick state for any that
1959 * have not yet encountered a quiescent state, using the function specified.
1960 * Also initiate boosting for any threads blocked on the root rcu_node.
1962 * The caller must have suppressed start of new grace periods.
1964 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1968 unsigned long flags
;
1970 struct rcu_node
*rnp
;
1972 rcu_for_each_leaf_node(rsp
, rnp
) {
1975 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1976 if (!rcu_gp_in_progress(rsp
)) {
1977 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1980 if (rnp
->qsmask
== 0) {
1981 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1986 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1987 if ((rnp
->qsmask
& bit
) != 0 &&
1988 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1993 /* rcu_report_qs_rnp() releases rnp->lock. */
1994 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1997 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1999 rnp
= rcu_get_root(rsp
);
2000 if (rnp
->qsmask
== 0) {
2001 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2002 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
2007 * Force quiescent states on reluctant CPUs, and also detect which
2008 * CPUs are in dyntick-idle mode.
2010 static void force_quiescent_state(struct rcu_state
*rsp
)
2012 unsigned long flags
;
2014 struct rcu_node
*rnp
;
2015 struct rcu_node
*rnp_old
= NULL
;
2017 /* Funnel through hierarchy to reduce memory contention. */
2018 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
2019 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2020 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2021 !raw_spin_trylock(&rnp
->fqslock
);
2022 if (rnp_old
!= NULL
)
2023 raw_spin_unlock(&rnp_old
->fqslock
);
2025 rsp
->n_force_qs_lh
++;
2030 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2032 /* Reached the root of the rcu_node tree, acquire lock. */
2033 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2034 raw_spin_unlock(&rnp_old
->fqslock
);
2035 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2036 rsp
->n_force_qs_lh
++;
2037 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2038 return; /* Someone beat us to it. */
2040 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
2041 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2042 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
2046 * This does the RCU core processing work for the specified rcu_state
2047 * and rcu_data structures. This may be called only from the CPU to
2048 * whom the rdp belongs.
2051 __rcu_process_callbacks(struct rcu_state
*rsp
)
2053 unsigned long flags
;
2054 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2056 WARN_ON_ONCE(rdp
->beenonline
== 0);
2059 * Advance callbacks in response to end of earlier grace
2060 * period that some other CPU ended.
2062 rcu_process_gp_end(rsp
, rdp
);
2064 /* Update RCU state based on any recent quiescent states. */
2065 rcu_check_quiescent_state(rsp
, rdp
);
2067 /* Does this CPU require a not-yet-started grace period? */
2068 if (cpu_needs_another_gp(rsp
, rdp
)) {
2069 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
2070 rcu_start_gp(rsp
, flags
); /* releases above lock */
2073 /* If there are callbacks ready, invoke them. */
2074 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2075 invoke_rcu_callbacks(rsp
, rdp
);
2079 * Do RCU core processing for the current CPU.
2081 static void rcu_process_callbacks(struct softirq_action
*unused
)
2083 struct rcu_state
*rsp
;
2085 if (cpu_is_offline(smp_processor_id()))
2087 trace_rcu_utilization("Start RCU core");
2088 for_each_rcu_flavor(rsp
)
2089 __rcu_process_callbacks(rsp
);
2090 trace_rcu_utilization("End RCU core");
2094 * Schedule RCU callback invocation. If the specified type of RCU
2095 * does not support RCU priority boosting, just do a direct call,
2096 * otherwise wake up the per-CPU kernel kthread. Note that because we
2097 * are running on the current CPU with interrupts disabled, the
2098 * rcu_cpu_kthread_task cannot disappear out from under us.
2100 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2102 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2104 if (likely(!rsp
->boost
)) {
2105 rcu_do_batch(rsp
, rdp
);
2108 invoke_rcu_callbacks_kthread();
2111 static void invoke_rcu_core(void)
2113 raise_softirq(RCU_SOFTIRQ
);
2117 * Handle any core-RCU processing required by a call_rcu() invocation.
2119 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2120 struct rcu_head
*head
, unsigned long flags
)
2123 * If called from an extended quiescent state, invoke the RCU
2124 * core in order to force a re-evaluation of RCU's idleness.
2126 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2129 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2130 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2134 * Force the grace period if too many callbacks or too long waiting.
2135 * Enforce hysteresis, and don't invoke force_quiescent_state()
2136 * if some other CPU has recently done so. Also, don't bother
2137 * invoking force_quiescent_state() if the newly enqueued callback
2138 * is the only one waiting for a grace period to complete.
2140 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2142 /* Are we ignoring a completed grace period? */
2143 rcu_process_gp_end(rsp
, rdp
);
2144 check_for_new_grace_period(rsp
, rdp
);
2146 /* Start a new grace period if one not already started. */
2147 if (!rcu_gp_in_progress(rsp
)) {
2148 unsigned long nestflag
;
2149 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2151 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
2152 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
2154 /* Give the grace period a kick. */
2155 rdp
->blimit
= LONG_MAX
;
2156 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2157 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2158 force_quiescent_state(rsp
);
2159 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2160 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2166 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2167 struct rcu_state
*rsp
, bool lazy
)
2169 unsigned long flags
;
2170 struct rcu_data
*rdp
;
2172 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2173 debug_rcu_head_queue(head
);
2178 * Opportunistically note grace-period endings and beginnings.
2179 * Note that we might see a beginning right after we see an
2180 * end, but never vice versa, since this CPU has to pass through
2181 * a quiescent state betweentimes.
2183 local_irq_save(flags
);
2184 rdp
= this_cpu_ptr(rsp
->rda
);
2186 /* Add the callback to our list. */
2187 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
)) {
2188 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2190 local_irq_restore(flags
);
2193 ACCESS_ONCE(rdp
->qlen
)++;
2197 rcu_idle_count_callbacks_posted();
2198 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2199 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2200 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2202 if (__is_kfree_rcu_offset((unsigned long)func
))
2203 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2204 rdp
->qlen_lazy
, rdp
->qlen
);
2206 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2208 /* Go handle any RCU core processing required. */
2209 __call_rcu_core(rsp
, rdp
, head
, flags
);
2210 local_irq_restore(flags
);
2214 * Queue an RCU-sched callback for invocation after a grace period.
2216 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2218 __call_rcu(head
, func
, &rcu_sched_state
, 0);
2220 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2223 * Queue an RCU callback for invocation after a quicker grace period.
2225 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2227 __call_rcu(head
, func
, &rcu_bh_state
, 0);
2229 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2232 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2233 * any blocking grace-period wait automatically implies a grace period
2234 * if there is only one CPU online at any point time during execution
2235 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2236 * occasionally incorrectly indicate that there are multiple CPUs online
2237 * when there was in fact only one the whole time, as this just adds
2238 * some overhead: RCU still operates correctly.
2240 static inline int rcu_blocking_is_gp(void)
2244 might_sleep(); /* Check for RCU read-side critical section. */
2246 ret
= num_online_cpus() <= 1;
2252 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2254 * Control will return to the caller some time after a full rcu-sched
2255 * grace period has elapsed, in other words after all currently executing
2256 * rcu-sched read-side critical sections have completed. These read-side
2257 * critical sections are delimited by rcu_read_lock_sched() and
2258 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2259 * local_irq_disable(), and so on may be used in place of
2260 * rcu_read_lock_sched().
2262 * This means that all preempt_disable code sequences, including NMI and
2263 * non-threaded hardware-interrupt handlers, in progress on entry will
2264 * have completed before this primitive returns. However, this does not
2265 * guarantee that softirq handlers will have completed, since in some
2266 * kernels, these handlers can run in process context, and can block.
2268 * Note that this guarantee implies further memory-ordering guarantees.
2269 * On systems with more than one CPU, when synchronize_sched() returns,
2270 * each CPU is guaranteed to have executed a full memory barrier since the
2271 * end of its last RCU-sched read-side critical section whose beginning
2272 * preceded the call to synchronize_sched(). In addition, each CPU having
2273 * an RCU read-side critical section that extends beyond the return from
2274 * synchronize_sched() is guaranteed to have executed a full memory barrier
2275 * after the beginning of synchronize_sched() and before the beginning of
2276 * that RCU read-side critical section. Note that these guarantees include
2277 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2278 * that are executing in the kernel.
2280 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2281 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2282 * to have executed a full memory barrier during the execution of
2283 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2284 * again only if the system has more than one CPU).
2286 * This primitive provides the guarantees made by the (now removed)
2287 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2288 * guarantees that rcu_read_lock() sections will have completed.
2289 * In "classic RCU", these two guarantees happen to be one and
2290 * the same, but can differ in realtime RCU implementations.
2292 void synchronize_sched(void)
2294 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2295 !lock_is_held(&rcu_lock_map
) &&
2296 !lock_is_held(&rcu_sched_lock_map
),
2297 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2298 if (rcu_blocking_is_gp())
2301 synchronize_sched_expedited();
2303 wait_rcu_gp(call_rcu_sched
);
2305 EXPORT_SYMBOL_GPL(synchronize_sched
);
2308 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2310 * Control will return to the caller some time after a full rcu_bh grace
2311 * period has elapsed, in other words after all currently executing rcu_bh
2312 * read-side critical sections have completed. RCU read-side critical
2313 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2314 * and may be nested.
2316 * See the description of synchronize_sched() for more detailed information
2317 * on memory ordering guarantees.
2319 void synchronize_rcu_bh(void)
2321 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2322 !lock_is_held(&rcu_lock_map
) &&
2323 !lock_is_held(&rcu_sched_lock_map
),
2324 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2325 if (rcu_blocking_is_gp())
2328 synchronize_rcu_bh_expedited();
2330 wait_rcu_gp(call_rcu_bh
);
2332 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2334 static int synchronize_sched_expedited_cpu_stop(void *data
)
2337 * There must be a full memory barrier on each affected CPU
2338 * between the time that try_stop_cpus() is called and the
2339 * time that it returns.
2341 * In the current initial implementation of cpu_stop, the
2342 * above condition is already met when the control reaches
2343 * this point and the following smp_mb() is not strictly
2344 * necessary. Do smp_mb() anyway for documentation and
2345 * robustness against future implementation changes.
2347 smp_mb(); /* See above comment block. */
2352 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2354 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2355 * approach to force the grace period to end quickly. This consumes
2356 * significant time on all CPUs and is unfriendly to real-time workloads,
2357 * so is thus not recommended for any sort of common-case code. In fact,
2358 * if you are using synchronize_sched_expedited() in a loop, please
2359 * restructure your code to batch your updates, and then use a single
2360 * synchronize_sched() instead.
2362 * Note that it is illegal to call this function while holding any lock
2363 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2364 * to call this function from a CPU-hotplug notifier. Failing to observe
2365 * these restriction will result in deadlock.
2367 * This implementation can be thought of as an application of ticket
2368 * locking to RCU, with sync_sched_expedited_started and
2369 * sync_sched_expedited_done taking on the roles of the halves
2370 * of the ticket-lock word. Each task atomically increments
2371 * sync_sched_expedited_started upon entry, snapshotting the old value,
2372 * then attempts to stop all the CPUs. If this succeeds, then each
2373 * CPU will have executed a context switch, resulting in an RCU-sched
2374 * grace period. We are then done, so we use atomic_cmpxchg() to
2375 * update sync_sched_expedited_done to match our snapshot -- but
2376 * only if someone else has not already advanced past our snapshot.
2378 * On the other hand, if try_stop_cpus() fails, we check the value
2379 * of sync_sched_expedited_done. If it has advanced past our
2380 * initial snapshot, then someone else must have forced a grace period
2381 * some time after we took our snapshot. In this case, our work is
2382 * done for us, and we can simply return. Otherwise, we try again,
2383 * but keep our initial snapshot for purposes of checking for someone
2384 * doing our work for us.
2386 * If we fail too many times in a row, we fall back to synchronize_sched().
2388 void synchronize_sched_expedited(void)
2390 long firstsnap
, s
, snap
;
2392 struct rcu_state
*rsp
= &rcu_sched_state
;
2395 * If we are in danger of counter wrap, just do synchronize_sched().
2396 * By allowing sync_sched_expedited_started to advance no more than
2397 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2398 * that more than 3.5 billion CPUs would be required to force a
2399 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2400 * course be required on a 64-bit system.
2402 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
2403 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
2405 synchronize_sched();
2406 atomic_long_inc(&rsp
->expedited_wrap
);
2411 * Take a ticket. Note that atomic_inc_return() implies a
2412 * full memory barrier.
2414 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
2417 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2420 * Each pass through the following loop attempts to force a
2421 * context switch on each CPU.
2423 while (try_stop_cpus(cpu_online_mask
,
2424 synchronize_sched_expedited_cpu_stop
,
2427 atomic_long_inc(&rsp
->expedited_tryfail
);
2429 /* Check to see if someone else did our work for us. */
2430 s
= atomic_long_read(&rsp
->expedited_done
);
2431 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2432 /* ensure test happens before caller kfree */
2433 smp_mb__before_atomic_inc(); /* ^^^ */
2434 atomic_long_inc(&rsp
->expedited_workdone1
);
2438 /* No joy, try again later. Or just synchronize_sched(). */
2439 if (trycount
++ < 10) {
2440 udelay(trycount
* num_online_cpus());
2442 wait_rcu_gp(call_rcu_sched
);
2443 atomic_long_inc(&rsp
->expedited_normal
);
2447 /* Recheck to see if someone else did our work for us. */
2448 s
= atomic_long_read(&rsp
->expedited_done
);
2449 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2450 /* ensure test happens before caller kfree */
2451 smp_mb__before_atomic_inc(); /* ^^^ */
2452 atomic_long_inc(&rsp
->expedited_workdone2
);
2457 * Refetching sync_sched_expedited_started allows later
2458 * callers to piggyback on our grace period. We retry
2459 * after they started, so our grace period works for them,
2460 * and they started after our first try, so their grace
2461 * period works for us.
2464 snap
= atomic_long_read(&rsp
->expedited_start
);
2465 smp_mb(); /* ensure read is before try_stop_cpus(). */
2467 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
2470 * Everyone up to our most recent fetch is covered by our grace
2471 * period. Update the counter, but only if our work is still
2472 * relevant -- which it won't be if someone who started later
2473 * than we did already did their update.
2476 atomic_long_inc(&rsp
->expedited_done_tries
);
2477 s
= atomic_long_read(&rsp
->expedited_done
);
2478 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
2479 /* ensure test happens before caller kfree */
2480 smp_mb__before_atomic_inc(); /* ^^^ */
2481 atomic_long_inc(&rsp
->expedited_done_lost
);
2484 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
2485 atomic_long_inc(&rsp
->expedited_done_exit
);
2489 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2492 * Check to see if there is any immediate RCU-related work to be done
2493 * by the current CPU, for the specified type of RCU, returning 1 if so.
2494 * The checks are in order of increasing expense: checks that can be
2495 * carried out against CPU-local state are performed first. However,
2496 * we must check for CPU stalls first, else we might not get a chance.
2498 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2500 struct rcu_node
*rnp
= rdp
->mynode
;
2502 rdp
->n_rcu_pending
++;
2504 /* Check for CPU stalls, if enabled. */
2505 check_cpu_stall(rsp
, rdp
);
2507 /* Is the RCU core waiting for a quiescent state from this CPU? */
2508 if (rcu_scheduler_fully_active
&&
2509 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2510 rdp
->n_rp_qs_pending
++;
2511 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2512 rdp
->n_rp_report_qs
++;
2516 /* Does this CPU have callbacks ready to invoke? */
2517 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2518 rdp
->n_rp_cb_ready
++;
2522 /* Has RCU gone idle with this CPU needing another grace period? */
2523 if (cpu_needs_another_gp(rsp
, rdp
)) {
2524 rdp
->n_rp_cpu_needs_gp
++;
2528 /* Has another RCU grace period completed? */
2529 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2530 rdp
->n_rp_gp_completed
++;
2534 /* Has a new RCU grace period started? */
2535 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2536 rdp
->n_rp_gp_started
++;
2541 rdp
->n_rp_need_nothing
++;
2546 * Check to see if there is any immediate RCU-related work to be done
2547 * by the current CPU, returning 1 if so. This function is part of the
2548 * RCU implementation; it is -not- an exported member of the RCU API.
2550 static int rcu_pending(int cpu
)
2552 struct rcu_state
*rsp
;
2554 for_each_rcu_flavor(rsp
)
2555 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2561 * Check to see if any future RCU-related work will need to be done
2562 * by the current CPU, even if none need be done immediately, returning
2565 static int rcu_cpu_has_callbacks(int cpu
)
2567 struct rcu_state
*rsp
;
2569 /* RCU callbacks either ready or pending? */
2570 for_each_rcu_flavor(rsp
)
2571 if (per_cpu_ptr(rsp
->rda
, cpu
)->nxtlist
)
2577 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2578 * the compiler is expected to optimize this away.
2580 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2581 int cpu
, unsigned long done
)
2583 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2584 atomic_read(&rsp
->barrier_cpu_count
), done
);
2588 * RCU callback function for _rcu_barrier(). If we are last, wake
2589 * up the task executing _rcu_barrier().
2591 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2593 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2594 struct rcu_state
*rsp
= rdp
->rsp
;
2596 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2597 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2598 complete(&rsp
->barrier_completion
);
2600 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2605 * Called with preemption disabled, and from cross-cpu IRQ context.
2607 static void rcu_barrier_func(void *type
)
2609 struct rcu_state
*rsp
= type
;
2610 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2612 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2613 atomic_inc(&rsp
->barrier_cpu_count
);
2614 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2618 * Orchestrate the specified type of RCU barrier, waiting for all
2619 * RCU callbacks of the specified type to complete.
2621 static void _rcu_barrier(struct rcu_state
*rsp
)
2624 struct rcu_data
*rdp
;
2625 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2626 unsigned long snap_done
;
2628 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2630 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2631 mutex_lock(&rsp
->barrier_mutex
);
2634 * Ensure that all prior references, including to ->n_barrier_done,
2635 * are ordered before the _rcu_barrier() machinery.
2637 smp_mb(); /* See above block comment. */
2640 * Recheck ->n_barrier_done to see if others did our work for us.
2641 * This means checking ->n_barrier_done for an even-to-odd-to-even
2642 * transition. The "if" expression below therefore rounds the old
2643 * value up to the next even number and adds two before comparing.
2645 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2646 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2647 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2648 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2649 smp_mb(); /* caller's subsequent code after above check. */
2650 mutex_unlock(&rsp
->barrier_mutex
);
2655 * Increment ->n_barrier_done to avoid duplicate work. Use
2656 * ACCESS_ONCE() to prevent the compiler from speculating
2657 * the increment to precede the early-exit check.
2659 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2660 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2661 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2662 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2665 * Initialize the count to one rather than to zero in order to
2666 * avoid a too-soon return to zero in case of a short grace period
2667 * (or preemption of this task). Exclude CPU-hotplug operations
2668 * to ensure that no offline CPU has callbacks queued.
2670 init_completion(&rsp
->barrier_completion
);
2671 atomic_set(&rsp
->barrier_cpu_count
, 1);
2675 * Force each CPU with callbacks to register a new callback.
2676 * When that callback is invoked, we will know that all of the
2677 * corresponding CPU's preceding callbacks have been invoked.
2679 for_each_online_cpu(cpu
) {
2680 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2681 if (ACCESS_ONCE(rdp
->qlen
)) {
2682 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2683 rsp
->n_barrier_done
);
2684 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2686 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2687 rsp
->n_barrier_done
);
2693 * Now that we have an rcu_barrier_callback() callback on each
2694 * CPU, and thus each counted, remove the initial count.
2696 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2697 complete(&rsp
->barrier_completion
);
2699 /* Increment ->n_barrier_done to prevent duplicate work. */
2700 smp_mb(); /* Keep increment after above mechanism. */
2701 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2702 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2703 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2704 smp_mb(); /* Keep increment before caller's subsequent code. */
2706 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2707 wait_for_completion(&rsp
->barrier_completion
);
2709 /* Other rcu_barrier() invocations can now safely proceed. */
2710 mutex_unlock(&rsp
->barrier_mutex
);
2714 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2716 void rcu_barrier_bh(void)
2718 _rcu_barrier(&rcu_bh_state
);
2720 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2723 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2725 void rcu_barrier_sched(void)
2727 _rcu_barrier(&rcu_sched_state
);
2729 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2732 * Do boot-time initialization of a CPU's per-CPU RCU data.
2735 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2737 unsigned long flags
;
2738 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2739 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2741 /* Set up local state, ensuring consistent view of global state. */
2742 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2743 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2744 init_callback_list(rdp
);
2746 ACCESS_ONCE(rdp
->qlen
) = 0;
2747 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2748 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2749 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2750 #ifdef CONFIG_RCU_USER_QS
2751 WARN_ON_ONCE(rdp
->dynticks
->in_user
);
2755 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2759 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2760 * offline event can be happening at a given time. Note also that we
2761 * can accept some slop in the rsp->completed access due to the fact
2762 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2764 static void __cpuinit
2765 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2767 unsigned long flags
;
2769 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2770 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2772 /* Exclude new grace periods. */
2773 mutex_lock(&rsp
->onoff_mutex
);
2775 /* Set up local state, ensuring consistent view of global state. */
2776 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2777 rdp
->beenonline
= 1; /* We have now been online. */
2778 rdp
->preemptible
= preemptible
;
2779 rdp
->qlen_last_fqs_check
= 0;
2780 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2781 rdp
->blimit
= blimit
;
2782 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
2783 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2784 atomic_set(&rdp
->dynticks
->dynticks
,
2785 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2786 rcu_prepare_for_idle_init(cpu
);
2787 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2789 /* Add CPU to rcu_node bitmasks. */
2791 mask
= rdp
->grpmask
;
2793 /* Exclude any attempts to start a new GP on small systems. */
2794 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2795 rnp
->qsmaskinit
|= mask
;
2796 mask
= rnp
->grpmask
;
2797 if (rnp
== rdp
->mynode
) {
2799 * If there is a grace period in progress, we will
2800 * set up to wait for it next time we run the
2803 rdp
->gpnum
= rnp
->completed
;
2804 rdp
->completed
= rnp
->completed
;
2805 rdp
->passed_quiesce
= 0;
2806 rdp
->qs_pending
= 0;
2807 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2809 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2811 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2812 local_irq_restore(flags
);
2814 mutex_unlock(&rsp
->onoff_mutex
);
2817 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2819 struct rcu_state
*rsp
;
2821 for_each_rcu_flavor(rsp
)
2822 rcu_init_percpu_data(cpu
, rsp
,
2823 strcmp(rsp
->name
, "rcu_preempt") == 0);
2827 * Handle CPU online/offline notification events.
2829 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2830 unsigned long action
, void *hcpu
)
2832 long cpu
= (long)hcpu
;
2833 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2834 struct rcu_node
*rnp
= rdp
->mynode
;
2835 struct rcu_state
*rsp
;
2837 trace_rcu_utilization("Start CPU hotplug");
2839 case CPU_UP_PREPARE
:
2840 case CPU_UP_PREPARE_FROZEN
:
2841 rcu_prepare_cpu(cpu
);
2842 rcu_prepare_kthreads(cpu
);
2845 case CPU_DOWN_FAILED
:
2846 rcu_boost_kthread_setaffinity(rnp
, -1);
2848 case CPU_DOWN_PREPARE
:
2849 rcu_boost_kthread_setaffinity(rnp
, cpu
);
2852 case CPU_DYING_FROZEN
:
2854 * The whole machine is "stopped" except this CPU, so we can
2855 * touch any data without introducing corruption. We send the
2856 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2858 for_each_rcu_flavor(rsp
)
2859 rcu_cleanup_dying_cpu(rsp
);
2860 rcu_cleanup_after_idle(cpu
);
2863 case CPU_DEAD_FROZEN
:
2864 case CPU_UP_CANCELED
:
2865 case CPU_UP_CANCELED_FROZEN
:
2866 for_each_rcu_flavor(rsp
)
2867 rcu_cleanup_dead_cpu(cpu
, rsp
);
2872 trace_rcu_utilization("End CPU hotplug");
2877 * Spawn the kthread that handles this RCU flavor's grace periods.
2879 static int __init
rcu_spawn_gp_kthread(void)
2881 unsigned long flags
;
2882 struct rcu_node
*rnp
;
2883 struct rcu_state
*rsp
;
2884 struct task_struct
*t
;
2886 for_each_rcu_flavor(rsp
) {
2887 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
2889 rnp
= rcu_get_root(rsp
);
2890 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2891 rsp
->gp_kthread
= t
;
2892 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2896 early_initcall(rcu_spawn_gp_kthread
);
2899 * This function is invoked towards the end of the scheduler's initialization
2900 * process. Before this is called, the idle task might contain
2901 * RCU read-side critical sections (during which time, this idle
2902 * task is booting the system). After this function is called, the
2903 * idle tasks are prohibited from containing RCU read-side critical
2904 * sections. This function also enables RCU lockdep checking.
2906 void rcu_scheduler_starting(void)
2908 WARN_ON(num_online_cpus() != 1);
2909 WARN_ON(nr_context_switches() > 0);
2910 rcu_scheduler_active
= 1;
2914 * Compute the per-level fanout, either using the exact fanout specified
2915 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2917 #ifdef CONFIG_RCU_FANOUT_EXACT
2918 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2922 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
2923 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2924 rsp
->levelspread
[0] = rcu_fanout_leaf
;
2926 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2927 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2934 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2935 ccur
= rsp
->levelcnt
[i
];
2936 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2940 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2943 * Helper function for rcu_init() that initializes one rcu_state structure.
2945 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2946 struct rcu_data __percpu
*rda
)
2948 static char *buf
[] = { "rcu_node_0",
2951 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2952 static char *fqs
[] = { "rcu_node_fqs_0",
2955 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2959 struct rcu_node
*rnp
;
2961 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2963 /* Initialize the level-tracking arrays. */
2965 for (i
= 0; i
< rcu_num_lvls
; i
++)
2966 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
2967 for (i
= 1; i
< rcu_num_lvls
; i
++)
2968 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2969 rcu_init_levelspread(rsp
);
2971 /* Initialize the elements themselves, starting from the leaves. */
2973 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2974 cpustride
*= rsp
->levelspread
[i
];
2975 rnp
= rsp
->level
[i
];
2976 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2977 raw_spin_lock_init(&rnp
->lock
);
2978 lockdep_set_class_and_name(&rnp
->lock
,
2979 &rcu_node_class
[i
], buf
[i
]);
2980 raw_spin_lock_init(&rnp
->fqslock
);
2981 lockdep_set_class_and_name(&rnp
->fqslock
,
2982 &rcu_fqs_class
[i
], fqs
[i
]);
2983 rnp
->gpnum
= rsp
->gpnum
;
2984 rnp
->completed
= rsp
->completed
;
2986 rnp
->qsmaskinit
= 0;
2987 rnp
->grplo
= j
* cpustride
;
2988 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2989 if (rnp
->grphi
>= NR_CPUS
)
2990 rnp
->grphi
= NR_CPUS
- 1;
2996 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2997 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2998 rnp
->parent
= rsp
->level
[i
- 1] +
2999 j
/ rsp
->levelspread
[i
- 1];
3002 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3007 init_waitqueue_head(&rsp
->gp_wq
);
3008 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3009 for_each_possible_cpu(i
) {
3010 while (i
> rnp
->grphi
)
3012 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3013 rcu_boot_init_percpu_data(i
, rsp
);
3015 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3019 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3020 * replace the definitions in rcutree.h because those are needed to size
3021 * the ->node array in the rcu_state structure.
3023 static void __init
rcu_init_geometry(void)
3028 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3030 /* If the compile-time values are accurate, just leave. */
3031 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3032 nr_cpu_ids
== NR_CPUS
)
3036 * Compute number of nodes that can be handled an rcu_node tree
3037 * with the given number of levels. Setting rcu_capacity[0] makes
3038 * some of the arithmetic easier.
3040 rcu_capacity
[0] = 1;
3041 rcu_capacity
[1] = rcu_fanout_leaf
;
3042 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3043 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3046 * The boot-time rcu_fanout_leaf parameter is only permitted
3047 * to increase the leaf-level fanout, not decrease it. Of course,
3048 * the leaf-level fanout cannot exceed the number of bits in
3049 * the rcu_node masks. Finally, the tree must be able to accommodate
3050 * the configured number of CPUs. Complain and fall back to the
3051 * compile-time values if these limits are exceeded.
3053 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3054 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3055 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3060 /* Calculate the number of rcu_nodes at each level of the tree. */
3061 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3062 if (n
<= rcu_capacity
[i
]) {
3063 for (j
= 0; j
<= i
; j
++)
3065 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3067 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3072 /* Calculate the total number of rcu_node structures. */
3074 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3075 rcu_num_nodes
+= num_rcu_lvl
[i
];
3079 void __init
rcu_init(void)
3083 rcu_bootup_announce();
3084 rcu_init_geometry();
3085 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3086 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3087 __rcu_init_preempt();
3088 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3091 * We don't need protection against CPU-hotplug here because
3092 * this is called early in boot, before either interrupts
3093 * or the scheduler are operational.
3095 cpu_notifier(rcu_cpu_notify
, 0);
3096 for_each_online_cpu(cpu
)
3097 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3098 check_cpu_stall_init();
3101 #include "rcutree_plugin.h"