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, \
73 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
80 struct rcu_state rcu_sched_state
=
81 RCU_STATE_INITIALIZER(rcu_sched
, call_rcu_sched
);
82 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
84 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
, call_rcu_bh
);
85 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
87 static struct rcu_state
*rcu_state
;
88 LIST_HEAD(rcu_struct_flavors
);
90 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
91 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
92 module_param(rcu_fanout_leaf
, int, 0444);
93 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
94 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
101 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
104 * The rcu_scheduler_active variable transitions from zero to one just
105 * before the first task is spawned. So when this variable is zero, RCU
106 * can assume that there is but one task, allowing RCU to (for example)
107 * optimized synchronize_sched() to a simple barrier(). When this variable
108 * is one, RCU must actually do all the hard work required to detect real
109 * grace periods. This variable is also used to suppress boot-time false
110 * positives from lockdep-RCU error checking.
112 int rcu_scheduler_active __read_mostly
;
113 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
116 * The rcu_scheduler_fully_active variable transitions from zero to one
117 * during the early_initcall() processing, which is after the scheduler
118 * is capable of creating new tasks. So RCU processing (for example,
119 * creating tasks for RCU priority boosting) must be delayed until after
120 * rcu_scheduler_fully_active transitions from zero to one. We also
121 * currently delay invocation of any RCU callbacks until after this point.
123 * It might later prove better for people registering RCU callbacks during
124 * early boot to take responsibility for these callbacks, but one step at
127 static int rcu_scheduler_fully_active __read_mostly
;
129 #ifdef CONFIG_RCU_BOOST
132 * Control variables for per-CPU and per-rcu_node kthreads. These
133 * handle all flavors of RCU.
135 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
136 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
138 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
140 #endif /* #ifdef CONFIG_RCU_BOOST */
142 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
143 static void invoke_rcu_core(void);
144 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
147 * Track the rcutorture test sequence number and the update version
148 * number within a given test. The rcutorture_testseq is incremented
149 * on every rcutorture module load and unload, so has an odd value
150 * when a test is running. The rcutorture_vernum is set to zero
151 * when rcutorture starts and is incremented on each rcutorture update.
152 * These variables enable correlating rcutorture output with the
153 * RCU tracing information.
155 unsigned long rcutorture_testseq
;
156 unsigned long rcutorture_vernum
;
159 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
160 * permit this function to be invoked without holding the root rcu_node
161 * structure's ->lock, but of course results can be subject to change.
163 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
165 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
169 * Note a quiescent state. Because we do not need to know
170 * how many quiescent states passed, just if there was at least
171 * one since the start of the grace period, this just sets a flag.
172 * The caller must have disabled preemption.
174 void rcu_sched_qs(int cpu
)
176 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
178 if (rdp
->passed_quiesce
== 0)
179 trace_rcu_grace_period("rcu_sched", rdp
->gpnum
, "cpuqs");
180 rdp
->passed_quiesce
= 1;
183 void rcu_bh_qs(int cpu
)
185 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
187 if (rdp
->passed_quiesce
== 0)
188 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
189 rdp
->passed_quiesce
= 1;
193 * Note a context switch. This is a quiescent state for RCU-sched,
194 * and requires special handling for preemptible RCU.
195 * The caller must have disabled preemption.
197 void rcu_note_context_switch(int cpu
)
199 trace_rcu_utilization("Start context switch");
201 rcu_preempt_note_context_switch(cpu
);
202 trace_rcu_utilization("End context switch");
204 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
206 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
207 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
208 .dynticks
= ATOMIC_INIT(1),
211 static int blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
212 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
213 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
215 module_param(blimit
, int, 0444);
216 module_param(qhimark
, int, 0444);
217 module_param(qlowmark
, int, 0444);
219 int rcu_cpu_stall_suppress __read_mostly
; /* 1 = suppress stall warnings. */
220 int rcu_cpu_stall_timeout __read_mostly
= CONFIG_RCU_CPU_STALL_TIMEOUT
;
222 module_param(rcu_cpu_stall_suppress
, int, 0644);
223 module_param(rcu_cpu_stall_timeout
, int, 0644);
225 static ulong jiffies_till_first_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
226 static ulong jiffies_till_next_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
228 module_param(jiffies_till_first_fqs
, ulong
, 0644);
229 module_param(jiffies_till_next_fqs
, ulong
, 0644);
231 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*));
232 static void force_quiescent_state(struct rcu_state
*rsp
);
233 static int rcu_pending(int cpu
);
236 * Return the number of RCU-sched batches processed thus far for debug & stats.
238 long rcu_batches_completed_sched(void)
240 return rcu_sched_state
.completed
;
242 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
245 * Return the number of RCU BH batches processed thus far for debug & stats.
247 long rcu_batches_completed_bh(void)
249 return rcu_bh_state
.completed
;
251 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
254 * Force a quiescent state for RCU BH.
256 void rcu_bh_force_quiescent_state(void)
258 force_quiescent_state(&rcu_bh_state
);
260 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
263 * Record the number of times rcutorture tests have been initiated and
264 * terminated. This information allows the debugfs tracing stats to be
265 * correlated to the rcutorture messages, even when the rcutorture module
266 * is being repeatedly loaded and unloaded. In other words, we cannot
267 * store this state in rcutorture itself.
269 void rcutorture_record_test_transition(void)
271 rcutorture_testseq
++;
272 rcutorture_vernum
= 0;
274 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
277 * Record the number of writer passes through the current rcutorture test.
278 * This is also used to correlate debugfs tracing stats with the rcutorture
281 void rcutorture_record_progress(unsigned long vernum
)
285 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
288 * Force a quiescent state for RCU-sched.
290 void rcu_sched_force_quiescent_state(void)
292 force_quiescent_state(&rcu_sched_state
);
294 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
297 * Does the CPU have callbacks ready to be invoked?
300 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
302 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
306 * Does the current CPU require a yet-as-unscheduled grace period?
309 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
311 return *rdp
->nxttail
[RCU_DONE_TAIL
+
312 ACCESS_ONCE(rsp
->completed
) != rdp
->completed
] &&
313 !rcu_gp_in_progress(rsp
);
317 * Return the root node of the specified rcu_state structure.
319 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
321 return &rsp
->node
[0];
325 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
327 * If the new value of the ->dynticks_nesting counter now is zero,
328 * we really have entered idle, and must do the appropriate accounting.
329 * The caller must have disabled interrupts.
331 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
334 trace_rcu_dyntick("Start", oldval
, 0);
335 if (!is_idle_task(current
) && !user
) {
336 struct task_struct
*idle
= idle_task(smp_processor_id());
338 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
339 ftrace_dump(DUMP_ORIG
);
340 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
341 current
->pid
, current
->comm
,
342 idle
->pid
, idle
->comm
); /* must be idle task! */
344 rcu_prepare_for_idle(smp_processor_id());
345 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
346 smp_mb__before_atomic_inc(); /* See above. */
347 atomic_inc(&rdtp
->dynticks
);
348 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
349 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
352 * It is illegal to enter an extended quiescent state while
353 * in an RCU read-side critical section.
355 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
356 "Illegal idle entry in RCU read-side critical section.");
357 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
358 "Illegal idle entry in RCU-bh read-side critical section.");
359 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
360 "Illegal idle entry in RCU-sched read-side critical section.");
364 * Enter an RCU extended quiescent state, which can be either the
365 * idle loop or adaptive-tickless usermode execution.
367 static void rcu_eqs_enter(bool user
)
371 struct rcu_dynticks
*rdtp
;
373 local_irq_save(flags
);
374 rdtp
= &__get_cpu_var(rcu_dynticks
);
375 oldval
= rdtp
->dynticks_nesting
;
376 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
377 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
378 rdtp
->dynticks_nesting
= 0;
380 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
381 rcu_eqs_enter_common(rdtp
, oldval
, user
);
382 local_irq_restore(flags
);
386 * rcu_idle_enter - inform RCU that current CPU is entering idle
388 * Enter idle mode, in other words, -leave- the mode in which RCU
389 * read-side critical sections can occur. (Though RCU read-side
390 * critical sections can occur in irq handlers in idle, a possibility
391 * handled by irq_enter() and irq_exit().)
393 * We crowbar the ->dynticks_nesting field to zero to allow for
394 * the possibility of usermode upcalls having messed up our count
395 * of interrupt nesting level during the prior busy period.
397 void rcu_idle_enter(void)
401 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
404 * rcu_user_enter - inform RCU that we are resuming userspace.
406 * Enter RCU idle mode right before resuming userspace. No use of RCU
407 * is permitted between this call and rcu_user_exit(). This way the
408 * CPU doesn't need to maintain the tick for RCU maintenance purposes
409 * when the CPU runs in userspace.
411 void rcu_user_enter(void)
414 * Some contexts may involve an exception occuring in an irq,
415 * leading to that nesting:
416 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
417 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
418 * helpers are enough to protect RCU uses inside the exception. So
419 * just return immediately if we detect we are in an IRQ.
429 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
431 * Exit from an interrupt handler, which might possibly result in entering
432 * idle mode, in other words, leaving the mode in which read-side critical
433 * sections can occur.
435 * This code assumes that the idle loop never does anything that might
436 * result in unbalanced calls to irq_enter() and irq_exit(). If your
437 * architecture violates this assumption, RCU will give you what you
438 * deserve, good and hard. But very infrequently and irreproducibly.
440 * Use things like work queues to work around this limitation.
442 * You have been warned.
444 void rcu_irq_exit(void)
448 struct rcu_dynticks
*rdtp
;
450 local_irq_save(flags
);
451 rdtp
= &__get_cpu_var(rcu_dynticks
);
452 oldval
= rdtp
->dynticks_nesting
;
453 rdtp
->dynticks_nesting
--;
454 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
455 if (rdtp
->dynticks_nesting
)
456 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
458 rcu_eqs_enter_common(rdtp
, oldval
, 1);
459 local_irq_restore(flags
);
463 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
465 * If the new value of the ->dynticks_nesting counter was previously zero,
466 * we really have exited idle, and must do the appropriate accounting.
467 * The caller must have disabled interrupts.
469 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
472 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
473 atomic_inc(&rdtp
->dynticks
);
474 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
475 smp_mb__after_atomic_inc(); /* See above. */
476 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
477 rcu_cleanup_after_idle(smp_processor_id());
478 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
479 if (!is_idle_task(current
) && !user
) {
480 struct task_struct
*idle
= idle_task(smp_processor_id());
482 trace_rcu_dyntick("Error on exit: not idle task",
483 oldval
, rdtp
->dynticks_nesting
);
484 ftrace_dump(DUMP_ORIG
);
485 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
486 current
->pid
, current
->comm
,
487 idle
->pid
, idle
->comm
); /* must be idle task! */
492 * Exit an RCU extended quiescent state, which can be either the
493 * idle loop or adaptive-tickless usermode execution.
495 static void rcu_eqs_exit(bool user
)
498 struct rcu_dynticks
*rdtp
;
501 local_irq_save(flags
);
502 rdtp
= &__get_cpu_var(rcu_dynticks
);
503 oldval
= rdtp
->dynticks_nesting
;
504 WARN_ON_ONCE(oldval
< 0);
505 if (oldval
& DYNTICK_TASK_NEST_MASK
)
506 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
508 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
509 rcu_eqs_exit_common(rdtp
, oldval
, user
);
510 local_irq_restore(flags
);
514 * rcu_idle_exit - inform RCU that current CPU is leaving idle
516 * Exit idle mode, in other words, -enter- the mode in which RCU
517 * read-side critical sections can occur.
519 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
520 * allow for the possibility of usermode upcalls messing up our count
521 * of interrupt nesting level during the busy period that is just
524 void rcu_idle_exit(void)
528 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
531 * rcu_user_exit - inform RCU that we are exiting userspace.
533 * Exit RCU idle mode while entering the kernel because it can
534 * run a RCU read side critical section anytime.
536 void rcu_user_exit(void)
539 * Some contexts may involve an exception occuring in an irq,
540 * leading to that nesting:
541 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
542 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
543 * helpers are enough to protect RCU uses inside the exception. So
544 * just return immediately if we detect we are in an IRQ.
553 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
555 * Enter an interrupt handler, which might possibly result in exiting
556 * idle mode, in other words, entering the mode in which read-side critical
557 * sections can occur.
559 * Note that the Linux kernel is fully capable of entering an interrupt
560 * handler that it never exits, for example when doing upcalls to
561 * user mode! This code assumes that the idle loop never does upcalls to
562 * user mode. If your architecture does do upcalls from the idle loop (or
563 * does anything else that results in unbalanced calls to the irq_enter()
564 * and irq_exit() functions), RCU will give you what you deserve, good
565 * and hard. But very infrequently and irreproducibly.
567 * Use things like work queues to work around this limitation.
569 * You have been warned.
571 void rcu_irq_enter(void)
574 struct rcu_dynticks
*rdtp
;
577 local_irq_save(flags
);
578 rdtp
= &__get_cpu_var(rcu_dynticks
);
579 oldval
= rdtp
->dynticks_nesting
;
580 rdtp
->dynticks_nesting
++;
581 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
583 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
585 rcu_eqs_exit_common(rdtp
, oldval
, 1);
586 local_irq_restore(flags
);
590 * rcu_nmi_enter - inform RCU of entry to NMI context
592 * If the CPU was idle with dynamic ticks active, and there is no
593 * irq handler running, this updates rdtp->dynticks_nmi to let the
594 * RCU grace-period handling know that the CPU is active.
596 void rcu_nmi_enter(void)
598 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
600 if (rdtp
->dynticks_nmi_nesting
== 0 &&
601 (atomic_read(&rdtp
->dynticks
) & 0x1))
603 rdtp
->dynticks_nmi_nesting
++;
604 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
605 atomic_inc(&rdtp
->dynticks
);
606 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
607 smp_mb__after_atomic_inc(); /* See above. */
608 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
612 * rcu_nmi_exit - inform RCU of exit from NMI context
614 * If the CPU was idle with dynamic ticks active, and there is no
615 * irq handler running, this updates rdtp->dynticks_nmi to let the
616 * RCU grace-period handling know that the CPU is no longer active.
618 void rcu_nmi_exit(void)
620 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
622 if (rdtp
->dynticks_nmi_nesting
== 0 ||
623 --rdtp
->dynticks_nmi_nesting
!= 0)
625 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
626 smp_mb__before_atomic_inc(); /* See above. */
627 atomic_inc(&rdtp
->dynticks
);
628 smp_mb__after_atomic_inc(); /* Force delay to next write. */
629 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
633 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
635 * If the current CPU is in its idle loop and is neither in an interrupt
636 * or NMI handler, return true.
638 int rcu_is_cpu_idle(void)
643 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
647 EXPORT_SYMBOL(rcu_is_cpu_idle
);
649 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
652 * Is the current CPU online? Disable preemption to avoid false positives
653 * that could otherwise happen due to the current CPU number being sampled,
654 * this task being preempted, its old CPU being taken offline, resuming
655 * on some other CPU, then determining that its old CPU is now offline.
656 * It is OK to use RCU on an offline processor during initial boot, hence
657 * the check for rcu_scheduler_fully_active. Note also that it is OK
658 * for a CPU coming online to use RCU for one jiffy prior to marking itself
659 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
660 * offline to continue to use RCU for one jiffy after marking itself
661 * offline in the cpu_online_mask. This leniency is necessary given the
662 * non-atomic nature of the online and offline processing, for example,
663 * the fact that a CPU enters the scheduler after completing the CPU_DYING
666 * This is also why RCU internally marks CPUs online during the
667 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
669 * Disable checking if in an NMI handler because we cannot safely report
670 * errors from NMI handlers anyway.
672 bool rcu_lockdep_current_cpu_online(void)
674 struct rcu_data
*rdp
;
675 struct rcu_node
*rnp
;
681 rdp
= &__get_cpu_var(rcu_sched_data
);
683 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
684 !rcu_scheduler_fully_active
;
688 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
690 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
693 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
695 * If the current CPU is idle or running at a first-level (not nested)
696 * interrupt from idle, return true. The caller must have at least
697 * disabled preemption.
699 int rcu_is_cpu_rrupt_from_idle(void)
701 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
705 * Snapshot the specified CPU's dynticks counter so that we can later
706 * credit them with an implicit quiescent state. Return 1 if this CPU
707 * is in dynticks idle mode, which is an extended quiescent state.
709 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
711 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
712 return (rdp
->dynticks_snap
& 0x1) == 0;
716 * Return true if the specified CPU has passed through a quiescent
717 * state by virtue of being in or having passed through an dynticks
718 * idle state since the last call to dyntick_save_progress_counter()
719 * for this same CPU, or by virtue of having been offline.
721 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
726 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
727 snap
= (unsigned int)rdp
->dynticks_snap
;
730 * If the CPU passed through or entered a dynticks idle phase with
731 * no active irq/NMI handlers, then we can safely pretend that the CPU
732 * already acknowledged the request to pass through a quiescent
733 * state. Either way, that CPU cannot possibly be in an RCU
734 * read-side critical section that started before the beginning
735 * of the current RCU grace period.
737 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
738 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
744 * Check for the CPU being offline, but only if the grace period
745 * is old enough. We don't need to worry about the CPU changing
746 * state: If we see it offline even once, it has been through a
749 * The reason for insisting that the grace period be at least
750 * one jiffy old is that CPUs that are not quite online and that
751 * have just gone offline can still execute RCU read-side critical
754 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
755 return 0; /* Grace period is not old enough. */
757 if (cpu_is_offline(rdp
->cpu
)) {
758 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
765 static int jiffies_till_stall_check(void)
767 int till_stall_check
= ACCESS_ONCE(rcu_cpu_stall_timeout
);
770 * Limit check must be consistent with the Kconfig limits
771 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
773 if (till_stall_check
< 3) {
774 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 3;
775 till_stall_check
= 3;
776 } else if (till_stall_check
> 300) {
777 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 300;
778 till_stall_check
= 300;
780 return till_stall_check
* HZ
+ RCU_STALL_DELAY_DELTA
;
783 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
785 rsp
->gp_start
= jiffies
;
786 rsp
->jiffies_stall
= jiffies
+ jiffies_till_stall_check();
789 static void print_other_cpu_stall(struct rcu_state
*rsp
)
795 struct rcu_node
*rnp
= rcu_get_root(rsp
);
797 /* Only let one CPU complain about others per time interval. */
799 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
800 delta
= jiffies
- rsp
->jiffies_stall
;
801 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
802 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
805 rsp
->jiffies_stall
= jiffies
+ 3 * jiffies_till_stall_check() + 3;
806 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
809 * OK, time to rat on our buddy...
810 * See Documentation/RCU/stallwarn.txt for info on how to debug
811 * RCU CPU stall warnings.
813 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
815 print_cpu_stall_info_begin();
816 rcu_for_each_leaf_node(rsp
, rnp
) {
817 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
818 ndetected
+= rcu_print_task_stall(rnp
);
819 if (rnp
->qsmask
!= 0) {
820 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
821 if (rnp
->qsmask
& (1UL << cpu
)) {
822 print_cpu_stall_info(rsp
,
827 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
831 * Now rat on any tasks that got kicked up to the root rcu_node
832 * due to CPU offlining.
834 rnp
= rcu_get_root(rsp
);
835 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
836 ndetected
+= rcu_print_task_stall(rnp
);
837 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
839 print_cpu_stall_info_end();
840 printk(KERN_CONT
"(detected by %d, t=%ld jiffies)\n",
841 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
843 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
844 else if (!trigger_all_cpu_backtrace())
847 /* Complain about tasks blocking the grace period. */
849 rcu_print_detail_task_stall(rsp
);
851 force_quiescent_state(rsp
); /* Kick them all. */
854 static void print_cpu_stall(struct rcu_state
*rsp
)
857 struct rcu_node
*rnp
= rcu_get_root(rsp
);
860 * OK, time to rat on ourselves...
861 * See Documentation/RCU/stallwarn.txt for info on how to debug
862 * RCU CPU stall warnings.
864 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
865 print_cpu_stall_info_begin();
866 print_cpu_stall_info(rsp
, smp_processor_id());
867 print_cpu_stall_info_end();
868 printk(KERN_CONT
" (t=%lu jiffies)\n", jiffies
- rsp
->gp_start
);
869 if (!trigger_all_cpu_backtrace())
872 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
873 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
874 rsp
->jiffies_stall
= jiffies
+
875 3 * jiffies_till_stall_check() + 3;
876 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
878 set_need_resched(); /* kick ourselves to get things going. */
881 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
885 struct rcu_node
*rnp
;
887 if (rcu_cpu_stall_suppress
)
889 j
= ACCESS_ONCE(jiffies
);
890 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
892 if (rcu_gp_in_progress(rsp
) &&
893 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
895 /* We haven't checked in, so go dump stack. */
896 print_cpu_stall(rsp
);
898 } else if (rcu_gp_in_progress(rsp
) &&
899 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
901 /* They had a few time units to dump stack, so complain. */
902 print_other_cpu_stall(rsp
);
906 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
908 rcu_cpu_stall_suppress
= 1;
913 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
915 * Set the stall-warning timeout way off into the future, thus preventing
916 * any RCU CPU stall-warning messages from appearing in the current set of
919 * The caller must disable hard irqs.
921 void rcu_cpu_stall_reset(void)
923 struct rcu_state
*rsp
;
925 for_each_rcu_flavor(rsp
)
926 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
929 static struct notifier_block rcu_panic_block
= {
930 .notifier_call
= rcu_panic
,
933 static void __init
check_cpu_stall_init(void)
935 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
939 * Update CPU-local rcu_data state to record the newly noticed grace period.
940 * This is used both when we started the grace period and when we notice
941 * that someone else started the grace period. The caller must hold the
942 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
943 * and must have irqs disabled.
945 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
947 if (rdp
->gpnum
!= rnp
->gpnum
) {
949 * If the current grace period is waiting for this CPU,
950 * set up to detect a quiescent state, otherwise don't
951 * go looking for one.
953 rdp
->gpnum
= rnp
->gpnum
;
954 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
955 rdp
->passed_quiesce
= 0;
956 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
957 zero_cpu_stall_ticks(rdp
);
961 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
964 struct rcu_node
*rnp
;
966 local_irq_save(flags
);
968 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
969 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
970 local_irq_restore(flags
);
973 __note_new_gpnum(rsp
, rnp
, rdp
);
974 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
978 * Did someone else start a new RCU grace period start since we last
979 * checked? Update local state appropriately if so. Must be called
980 * on the CPU corresponding to rdp.
983 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
988 local_irq_save(flags
);
989 if (rdp
->gpnum
!= rsp
->gpnum
) {
990 note_new_gpnum(rsp
, rdp
);
993 local_irq_restore(flags
);
998 * Initialize the specified rcu_data structure's callback list to empty.
1000 static void init_callback_list(struct rcu_data
*rdp
)
1004 rdp
->nxtlist
= NULL
;
1005 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1006 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1010 * Advance this CPU's callbacks, but only if the current grace period
1011 * has ended. This may be called only from the CPU to whom the rdp
1012 * belongs. In addition, the corresponding leaf rcu_node structure's
1013 * ->lock must be held by the caller, with irqs disabled.
1016 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1018 /* Did another grace period end? */
1019 if (rdp
->completed
!= rnp
->completed
) {
1021 /* Advance callbacks. No harm if list empty. */
1022 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
1023 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
1024 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1026 /* Remember that we saw this grace-period completion. */
1027 rdp
->completed
= rnp
->completed
;
1028 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
1031 * If we were in an extended quiescent state, we may have
1032 * missed some grace periods that others CPUs handled on
1033 * our behalf. Catch up with this state to avoid noting
1034 * spurious new grace periods. If another grace period
1035 * has started, then rnp->gpnum will have advanced, so
1036 * we will detect this later on. Of course, any quiescent
1037 * states we found for the old GP are now invalid.
1039 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
)) {
1040 rdp
->gpnum
= rdp
->completed
;
1041 rdp
->passed_quiesce
= 0;
1045 * If RCU does not need a quiescent state from this CPU,
1046 * then make sure that this CPU doesn't go looking for one.
1048 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1049 rdp
->qs_pending
= 0;
1054 * Advance this CPU's callbacks, but only if the current grace period
1055 * has ended. This may be called only from the CPU to whom the rdp
1059 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1061 unsigned long flags
;
1062 struct rcu_node
*rnp
;
1064 local_irq_save(flags
);
1066 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1067 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1068 local_irq_restore(flags
);
1071 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1072 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1076 * Do per-CPU grace-period initialization for running CPU. The caller
1077 * must hold the lock of the leaf rcu_node structure corresponding to
1081 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1083 /* Prior grace period ended, so advance callbacks for current CPU. */
1084 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1086 /* Set state so that this CPU will detect the next quiescent state. */
1087 __note_new_gpnum(rsp
, rnp
, rdp
);
1091 * Initialize a new grace period.
1093 static int rcu_gp_init(struct rcu_state
*rsp
)
1095 struct rcu_data
*rdp
;
1096 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1098 raw_spin_lock_irq(&rnp
->lock
);
1099 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1101 if (rcu_gp_in_progress(rsp
)) {
1102 /* Grace period already in progress, don't start another. */
1103 raw_spin_unlock_irq(&rnp
->lock
);
1107 /* Advance to a new grace period and initialize state. */
1109 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1110 record_gp_stall_check_time(rsp
);
1111 raw_spin_unlock_irq(&rnp
->lock
);
1113 /* Exclude any concurrent CPU-hotplug operations. */
1117 * Set the quiescent-state-needed bits in all the rcu_node
1118 * structures for all currently online CPUs in breadth-first order,
1119 * starting from the root rcu_node structure, relying on the layout
1120 * of the tree within the rsp->node[] array. Note that other CPUs
1121 * will access only the leaves of the hierarchy, thus seeing that no
1122 * grace period is in progress, at least until the corresponding
1123 * leaf node has been initialized. In addition, we have excluded
1124 * CPU-hotplug operations.
1126 * The grace period cannot complete until the initialization
1127 * process finishes, because this kthread handles both.
1129 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1130 raw_spin_lock_irq(&rnp
->lock
);
1131 rdp
= this_cpu_ptr(rsp
->rda
);
1132 rcu_preempt_check_blocked_tasks(rnp
);
1133 rnp
->qsmask
= rnp
->qsmaskinit
;
1134 rnp
->gpnum
= rsp
->gpnum
;
1135 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1136 rnp
->completed
= rsp
->completed
;
1137 if (rnp
== rdp
->mynode
)
1138 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1139 rcu_preempt_boost_start_gp(rnp
);
1140 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1141 rnp
->level
, rnp
->grplo
,
1142 rnp
->grphi
, rnp
->qsmask
);
1143 raw_spin_unlock_irq(&rnp
->lock
);
1144 #ifdef CONFIG_PROVE_RCU_DELAY
1145 if ((random32() % (rcu_num_nodes
* 8)) == 0)
1146 schedule_timeout_uninterruptible(2);
1147 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1156 * Do one round of quiescent-state forcing.
1158 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1160 int fqs_state
= fqs_state_in
;
1161 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1164 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1165 /* Collect dyntick-idle snapshots. */
1166 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1167 fqs_state
= RCU_FORCE_QS
;
1169 /* Handle dyntick-idle and offline CPUs. */
1170 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1172 /* Clear flag to prevent immediate re-entry. */
1173 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1174 raw_spin_lock_irq(&rnp
->lock
);
1175 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1176 raw_spin_unlock_irq(&rnp
->lock
);
1182 * Clean up after the old grace period.
1184 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1186 unsigned long gp_duration
;
1187 struct rcu_data
*rdp
;
1188 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1190 raw_spin_lock_irq(&rnp
->lock
);
1191 gp_duration
= jiffies
- rsp
->gp_start
;
1192 if (gp_duration
> rsp
->gp_max
)
1193 rsp
->gp_max
= gp_duration
;
1196 * We know the grace period is complete, but to everyone else
1197 * it appears to still be ongoing. But it is also the case
1198 * that to everyone else it looks like there is nothing that
1199 * they can do to advance the grace period. It is therefore
1200 * safe for us to drop the lock in order to mark the grace
1201 * period as completed in all of the rcu_node structures.
1203 raw_spin_unlock_irq(&rnp
->lock
);
1206 * Propagate new ->completed value to rcu_node structures so
1207 * that other CPUs don't have to wait until the start of the next
1208 * grace period to process their callbacks. This also avoids
1209 * some nasty RCU grace-period initialization races by forcing
1210 * the end of the current grace period to be completely recorded in
1211 * all of the rcu_node structures before the beginning of the next
1212 * grace period is recorded in any of the rcu_node structures.
1214 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1215 raw_spin_lock_irq(&rnp
->lock
);
1216 rnp
->completed
= rsp
->gpnum
;
1217 raw_spin_unlock_irq(&rnp
->lock
);
1220 rnp
= rcu_get_root(rsp
);
1221 raw_spin_lock_irq(&rnp
->lock
);
1223 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1224 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1225 rsp
->fqs_state
= RCU_GP_IDLE
;
1226 rdp
= this_cpu_ptr(rsp
->rda
);
1227 if (cpu_needs_another_gp(rsp
, rdp
))
1229 raw_spin_unlock_irq(&rnp
->lock
);
1233 * Body of kthread that handles grace periods.
1235 static int __noreturn
rcu_gp_kthread(void *arg
)
1240 struct rcu_state
*rsp
= arg
;
1241 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1245 /* Handle grace-period start. */
1247 wait_event_interruptible(rsp
->gp_wq
,
1250 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1254 flush_signals(current
);
1257 /* Handle quiescent-state forcing. */
1258 fqs_state
= RCU_SAVE_DYNTICK
;
1259 j
= jiffies_till_first_fqs
;
1262 jiffies_till_first_fqs
= HZ
;
1265 rsp
->jiffies_force_qs
= jiffies
+ j
;
1266 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1267 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1268 (!ACCESS_ONCE(rnp
->qsmask
) &&
1269 !rcu_preempt_blocked_readers_cgp(rnp
)),
1271 /* If grace period done, leave loop. */
1272 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1273 !rcu_preempt_blocked_readers_cgp(rnp
))
1275 /* If time for quiescent-state forcing, do it. */
1276 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1277 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1280 /* Deal with stray signal. */
1282 flush_signals(current
);
1284 j
= jiffies_till_next_fqs
;
1287 jiffies_till_next_fqs
= HZ
;
1290 jiffies_till_next_fqs
= 1;
1294 /* Handle grace-period end. */
1295 rcu_gp_cleanup(rsp
);
1300 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1301 * in preparation for detecting the next grace period. The caller must hold
1302 * the root node's ->lock, which is released before return. Hard irqs must
1305 * Note that it is legal for a dying CPU (which is marked as offline) to
1306 * invoke this function. This can happen when the dying CPU reports its
1310 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1311 __releases(rcu_get_root(rsp
)->lock
)
1313 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1314 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1316 if (!rsp
->gp_kthread
||
1317 !cpu_needs_another_gp(rsp
, rdp
)) {
1319 * Either we have not yet spawned the grace-period
1320 * task or this CPU does not need another grace period.
1321 * Either way, don't start a new grace period.
1323 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1327 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1328 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1329 wake_up(&rsp
->gp_wq
);
1333 * Report a full set of quiescent states to the specified rcu_state
1334 * data structure. This involves cleaning up after the prior grace
1335 * period and letting rcu_start_gp() start up the next grace period
1336 * if one is needed. Note that the caller must hold rnp->lock, as
1337 * required by rcu_start_gp(), which will release it.
1339 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1340 __releases(rcu_get_root(rsp
)->lock
)
1342 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1343 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1344 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1348 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1349 * Allows quiescent states for a group of CPUs to be reported at one go
1350 * to the specified rcu_node structure, though all the CPUs in the group
1351 * must be represented by the same rcu_node structure (which need not be
1352 * a leaf rcu_node structure, though it often will be). That structure's
1353 * lock must be held upon entry, and it is released before return.
1356 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1357 struct rcu_node
*rnp
, unsigned long flags
)
1358 __releases(rnp
->lock
)
1360 struct rcu_node
*rnp_c
;
1362 /* Walk up the rcu_node hierarchy. */
1364 if (!(rnp
->qsmask
& mask
)) {
1366 /* Our bit has already been cleared, so done. */
1367 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1370 rnp
->qsmask
&= ~mask
;
1371 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1372 mask
, rnp
->qsmask
, rnp
->level
,
1373 rnp
->grplo
, rnp
->grphi
,
1375 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1377 /* Other bits still set at this level, so done. */
1378 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1381 mask
= rnp
->grpmask
;
1382 if (rnp
->parent
== NULL
) {
1384 /* No more levels. Exit loop holding root lock. */
1388 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1391 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1392 WARN_ON_ONCE(rnp_c
->qsmask
);
1396 * Get here if we are the last CPU to pass through a quiescent
1397 * state for this grace period. Invoke rcu_report_qs_rsp()
1398 * to clean up and start the next grace period if one is needed.
1400 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1404 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1405 * structure. This must be either called from the specified CPU, or
1406 * called when the specified CPU is known to be offline (and when it is
1407 * also known that no other CPU is concurrently trying to help the offline
1408 * CPU). The lastcomp argument is used to make sure we are still in the
1409 * grace period of interest. We don't want to end the current grace period
1410 * based on quiescent states detected in an earlier grace period!
1413 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1415 unsigned long flags
;
1417 struct rcu_node
*rnp
;
1420 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1421 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1422 rnp
->completed
== rnp
->gpnum
) {
1425 * The grace period in which this quiescent state was
1426 * recorded has ended, so don't report it upwards.
1427 * We will instead need a new quiescent state that lies
1428 * within the current grace period.
1430 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1431 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1434 mask
= rdp
->grpmask
;
1435 if ((rnp
->qsmask
& mask
) == 0) {
1436 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1438 rdp
->qs_pending
= 0;
1441 * This GP can't end until cpu checks in, so all of our
1442 * callbacks can be processed during the next GP.
1444 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1446 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1451 * Check to see if there is a new grace period of which this CPU
1452 * is not yet aware, and if so, set up local rcu_data state for it.
1453 * Otherwise, see if this CPU has just passed through its first
1454 * quiescent state for this grace period, and record that fact if so.
1457 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1459 /* If there is now a new grace period, record and return. */
1460 if (check_for_new_grace_period(rsp
, rdp
))
1464 * Does this CPU still need to do its part for current grace period?
1465 * If no, return and let the other CPUs do their part as well.
1467 if (!rdp
->qs_pending
)
1471 * Was there a quiescent state since the beginning of the grace
1472 * period? If no, then exit and wait for the next call.
1474 if (!rdp
->passed_quiesce
)
1478 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1481 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1484 #ifdef CONFIG_HOTPLUG_CPU
1487 * Send the specified CPU's RCU callbacks to the orphanage. The
1488 * specified CPU must be offline, and the caller must hold the
1492 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1493 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1496 * Orphan the callbacks. First adjust the counts. This is safe
1497 * because ->onofflock excludes _rcu_barrier()'s adoption of
1498 * the callbacks, thus no memory barrier is required.
1500 if (rdp
->nxtlist
!= NULL
) {
1501 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1502 rsp
->qlen
+= rdp
->qlen
;
1503 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1505 ACCESS_ONCE(rdp
->qlen
) = 0;
1509 * Next, move those callbacks still needing a grace period to
1510 * the orphanage, where some other CPU will pick them up.
1511 * Some of the callbacks might have gone partway through a grace
1512 * period, but that is too bad. They get to start over because we
1513 * cannot assume that grace periods are synchronized across CPUs.
1514 * We don't bother updating the ->nxttail[] array yet, instead
1515 * we just reset the whole thing later on.
1517 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1518 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1519 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1520 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1524 * Then move the ready-to-invoke callbacks to the orphanage,
1525 * where some other CPU will pick them up. These will not be
1526 * required to pass though another grace period: They are done.
1528 if (rdp
->nxtlist
!= NULL
) {
1529 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1530 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1533 /* Finally, initialize the rcu_data structure's list to empty. */
1534 init_callback_list(rdp
);
1538 * Adopt the RCU callbacks from the specified rcu_state structure's
1539 * orphanage. The caller must hold the ->onofflock.
1541 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1544 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1546 /* Do the accounting first. */
1547 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1548 rdp
->qlen
+= rsp
->qlen
;
1549 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1550 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1551 rcu_idle_count_callbacks_posted();
1556 * We do not need a memory barrier here because the only way we
1557 * can get here if there is an rcu_barrier() in flight is if
1558 * we are the task doing the rcu_barrier().
1561 /* First adopt the ready-to-invoke callbacks. */
1562 if (rsp
->orphan_donelist
!= NULL
) {
1563 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1564 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1565 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1566 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1567 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1568 rsp
->orphan_donelist
= NULL
;
1569 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1572 /* And then adopt the callbacks that still need a grace period. */
1573 if (rsp
->orphan_nxtlist
!= NULL
) {
1574 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1575 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1576 rsp
->orphan_nxtlist
= NULL
;
1577 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1582 * Trace the fact that this CPU is going offline.
1584 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1586 RCU_TRACE(unsigned long mask
);
1587 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1588 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1590 RCU_TRACE(mask
= rdp
->grpmask
);
1591 trace_rcu_grace_period(rsp
->name
,
1592 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1597 * The CPU has been completely removed, and some other CPU is reporting
1598 * this fact from process context. Do the remainder of the cleanup,
1599 * including orphaning the outgoing CPU's RCU callbacks, and also
1600 * adopting them. There can only be one CPU hotplug operation at a time,
1601 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1603 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1605 unsigned long flags
;
1607 int need_report
= 0;
1608 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1609 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1611 /* Adjust any no-longer-needed kthreads. */
1612 rcu_boost_kthread_setaffinity(rnp
, -1);
1614 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1616 /* Exclude any attempts to start a new grace period. */
1617 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1619 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1620 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1621 rcu_adopt_orphan_cbs(rsp
);
1623 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1624 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1626 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1627 rnp
->qsmaskinit
&= ~mask
;
1628 if (rnp
->qsmaskinit
!= 0) {
1629 if (rnp
!= rdp
->mynode
)
1630 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1633 if (rnp
== rdp
->mynode
)
1634 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1636 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1637 mask
= rnp
->grpmask
;
1639 } while (rnp
!= NULL
);
1642 * We still hold the leaf rcu_node structure lock here, and
1643 * irqs are still disabled. The reason for this subterfuge is
1644 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1645 * held leads to deadlock.
1647 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1649 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1650 rcu_report_unblock_qs_rnp(rnp
, flags
);
1652 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1653 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1654 rcu_report_exp_rnp(rsp
, rnp
, true);
1655 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1656 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1657 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1658 init_callback_list(rdp
);
1659 /* Disallow further callbacks on this CPU. */
1660 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1663 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1665 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1669 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1673 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1676 * Invoke any RCU callbacks that have made it to the end of their grace
1677 * period. Thottle as specified by rdp->blimit.
1679 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1681 unsigned long flags
;
1682 struct rcu_head
*next
, *list
, **tail
;
1683 int bl
, count
, count_lazy
, i
;
1685 /* If no callbacks are ready, just return.*/
1686 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1687 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1688 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1689 need_resched(), is_idle_task(current
),
1690 rcu_is_callbacks_kthread());
1695 * Extract the list of ready callbacks, disabling to prevent
1696 * races with call_rcu() from interrupt handlers.
1698 local_irq_save(flags
);
1699 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1701 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1702 list
= rdp
->nxtlist
;
1703 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1704 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1705 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1706 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1707 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1708 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1709 local_irq_restore(flags
);
1711 /* Invoke callbacks. */
1712 count
= count_lazy
= 0;
1716 debug_rcu_head_unqueue(list
);
1717 if (__rcu_reclaim(rsp
->name
, list
))
1720 /* Stop only if limit reached and CPU has something to do. */
1721 if (++count
>= bl
&&
1723 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1727 local_irq_save(flags
);
1728 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1729 is_idle_task(current
),
1730 rcu_is_callbacks_kthread());
1732 /* Update count, and requeue any remaining callbacks. */
1734 *tail
= rdp
->nxtlist
;
1735 rdp
->nxtlist
= list
;
1736 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1737 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1738 rdp
->nxttail
[i
] = tail
;
1742 smp_mb(); /* List handling before counting for rcu_barrier(). */
1743 rdp
->qlen_lazy
-= count_lazy
;
1744 ACCESS_ONCE(rdp
->qlen
) -= count
;
1745 rdp
->n_cbs_invoked
+= count
;
1747 /* Reinstate batch limit if we have worked down the excess. */
1748 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1749 rdp
->blimit
= blimit
;
1751 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1752 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1753 rdp
->qlen_last_fqs_check
= 0;
1754 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1755 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1756 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1757 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
1759 local_irq_restore(flags
);
1761 /* Re-invoke RCU core processing if there are callbacks remaining. */
1762 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1767 * Check to see if this CPU is in a non-context-switch quiescent state
1768 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1769 * Also schedule RCU core processing.
1771 * This function must be called from hardirq context. It is normally
1772 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1773 * false, there is no point in invoking rcu_check_callbacks().
1775 void rcu_check_callbacks(int cpu
, int user
)
1777 trace_rcu_utilization("Start scheduler-tick");
1778 increment_cpu_stall_ticks();
1779 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1782 * Get here if this CPU took its interrupt from user
1783 * mode or from the idle loop, and if this is not a
1784 * nested interrupt. In this case, the CPU is in
1785 * a quiescent state, so note it.
1787 * No memory barrier is required here because both
1788 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1789 * variables that other CPUs neither access nor modify,
1790 * at least not while the corresponding CPU is online.
1796 } else if (!in_softirq()) {
1799 * Get here if this CPU did not take its interrupt from
1800 * softirq, in other words, if it is not interrupting
1801 * a rcu_bh read-side critical section. This is an _bh
1802 * critical section, so note it.
1807 rcu_preempt_check_callbacks(cpu
);
1808 if (rcu_pending(cpu
))
1810 trace_rcu_utilization("End scheduler-tick");
1814 * Scan the leaf rcu_node structures, processing dyntick state for any that
1815 * have not yet encountered a quiescent state, using the function specified.
1816 * Also initiate boosting for any threads blocked on the root rcu_node.
1818 * The caller must have suppressed start of new grace periods.
1820 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1824 unsigned long flags
;
1826 struct rcu_node
*rnp
;
1828 rcu_for_each_leaf_node(rsp
, rnp
) {
1831 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1832 if (!rcu_gp_in_progress(rsp
)) {
1833 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1836 if (rnp
->qsmask
== 0) {
1837 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1842 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1843 if ((rnp
->qsmask
& bit
) != 0 &&
1844 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1849 /* rcu_report_qs_rnp() releases rnp->lock. */
1850 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1853 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1855 rnp
= rcu_get_root(rsp
);
1856 if (rnp
->qsmask
== 0) {
1857 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1858 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1863 * Force quiescent states on reluctant CPUs, and also detect which
1864 * CPUs are in dyntick-idle mode.
1866 static void force_quiescent_state(struct rcu_state
*rsp
)
1868 unsigned long flags
;
1870 struct rcu_node
*rnp
;
1871 struct rcu_node
*rnp_old
= NULL
;
1873 /* Funnel through hierarchy to reduce memory contention. */
1874 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
1875 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
1876 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
1877 !raw_spin_trylock(&rnp
->fqslock
);
1878 if (rnp_old
!= NULL
)
1879 raw_spin_unlock(&rnp_old
->fqslock
);
1881 rsp
->n_force_qs_lh
++;
1886 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1888 /* Reached the root of the rcu_node tree, acquire lock. */
1889 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
1890 raw_spin_unlock(&rnp_old
->fqslock
);
1891 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1892 rsp
->n_force_qs_lh
++;
1893 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1894 return; /* Someone beat us to it. */
1896 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
1897 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1898 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1902 * This does the RCU core processing work for the specified rcu_state
1903 * and rcu_data structures. This may be called only from the CPU to
1904 * whom the rdp belongs.
1907 __rcu_process_callbacks(struct rcu_state
*rsp
)
1909 unsigned long flags
;
1910 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1912 WARN_ON_ONCE(rdp
->beenonline
== 0);
1915 * Advance callbacks in response to end of earlier grace
1916 * period that some other CPU ended.
1918 rcu_process_gp_end(rsp
, rdp
);
1920 /* Update RCU state based on any recent quiescent states. */
1921 rcu_check_quiescent_state(rsp
, rdp
);
1923 /* Does this CPU require a not-yet-started grace period? */
1924 if (cpu_needs_another_gp(rsp
, rdp
)) {
1925 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1926 rcu_start_gp(rsp
, flags
); /* releases above lock */
1929 /* If there are callbacks ready, invoke them. */
1930 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1931 invoke_rcu_callbacks(rsp
, rdp
);
1935 * Do RCU core processing for the current CPU.
1937 static void rcu_process_callbacks(struct softirq_action
*unused
)
1939 struct rcu_state
*rsp
;
1941 if (cpu_is_offline(smp_processor_id()))
1943 trace_rcu_utilization("Start RCU core");
1944 for_each_rcu_flavor(rsp
)
1945 __rcu_process_callbacks(rsp
);
1946 trace_rcu_utilization("End RCU core");
1950 * Schedule RCU callback invocation. If the specified type of RCU
1951 * does not support RCU priority boosting, just do a direct call,
1952 * otherwise wake up the per-CPU kernel kthread. Note that because we
1953 * are running on the current CPU with interrupts disabled, the
1954 * rcu_cpu_kthread_task cannot disappear out from under us.
1956 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1958 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
1960 if (likely(!rsp
->boost
)) {
1961 rcu_do_batch(rsp
, rdp
);
1964 invoke_rcu_callbacks_kthread();
1967 static void invoke_rcu_core(void)
1969 raise_softirq(RCU_SOFTIRQ
);
1973 * Handle any core-RCU processing required by a call_rcu() invocation.
1975 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
1976 struct rcu_head
*head
, unsigned long flags
)
1979 * If called from an extended quiescent state, invoke the RCU
1980 * core in order to force a re-evaluation of RCU's idleness.
1982 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1985 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1986 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
1990 * Force the grace period if too many callbacks or too long waiting.
1991 * Enforce hysteresis, and don't invoke force_quiescent_state()
1992 * if some other CPU has recently done so. Also, don't bother
1993 * invoking force_quiescent_state() if the newly enqueued callback
1994 * is the only one waiting for a grace period to complete.
1996 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1998 /* Are we ignoring a completed grace period? */
1999 rcu_process_gp_end(rsp
, rdp
);
2000 check_for_new_grace_period(rsp
, rdp
);
2002 /* Start a new grace period if one not already started. */
2003 if (!rcu_gp_in_progress(rsp
)) {
2004 unsigned long nestflag
;
2005 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2007 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
2008 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
2010 /* Give the grace period a kick. */
2011 rdp
->blimit
= LONG_MAX
;
2012 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2013 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2014 force_quiescent_state(rsp
);
2015 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2016 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2022 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2023 struct rcu_state
*rsp
, bool lazy
)
2025 unsigned long flags
;
2026 struct rcu_data
*rdp
;
2028 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2029 debug_rcu_head_queue(head
);
2034 * Opportunistically note grace-period endings and beginnings.
2035 * Note that we might see a beginning right after we see an
2036 * end, but never vice versa, since this CPU has to pass through
2037 * a quiescent state betweentimes.
2039 local_irq_save(flags
);
2040 rdp
= this_cpu_ptr(rsp
->rda
);
2042 /* Add the callback to our list. */
2043 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
)) {
2044 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2046 local_irq_restore(flags
);
2049 ACCESS_ONCE(rdp
->qlen
)++;
2053 rcu_idle_count_callbacks_posted();
2054 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2055 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2056 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2058 if (__is_kfree_rcu_offset((unsigned long)func
))
2059 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2060 rdp
->qlen_lazy
, rdp
->qlen
);
2062 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2064 /* Go handle any RCU core processing required. */
2065 __call_rcu_core(rsp
, rdp
, head
, flags
);
2066 local_irq_restore(flags
);
2070 * Queue an RCU-sched callback for invocation after a grace period.
2072 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2074 __call_rcu(head
, func
, &rcu_sched_state
, 0);
2076 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2079 * Queue an RCU callback for invocation after a quicker grace period.
2081 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2083 __call_rcu(head
, func
, &rcu_bh_state
, 0);
2085 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2088 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2089 * any blocking grace-period wait automatically implies a grace period
2090 * if there is only one CPU online at any point time during execution
2091 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2092 * occasionally incorrectly indicate that there are multiple CPUs online
2093 * when there was in fact only one the whole time, as this just adds
2094 * some overhead: RCU still operates correctly.
2096 static inline int rcu_blocking_is_gp(void)
2100 might_sleep(); /* Check for RCU read-side critical section. */
2102 ret
= num_online_cpus() <= 1;
2108 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2110 * Control will return to the caller some time after a full rcu-sched
2111 * grace period has elapsed, in other words after all currently executing
2112 * rcu-sched read-side critical sections have completed. These read-side
2113 * critical sections are delimited by rcu_read_lock_sched() and
2114 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2115 * local_irq_disable(), and so on may be used in place of
2116 * rcu_read_lock_sched().
2118 * This means that all preempt_disable code sequences, including NMI and
2119 * hardware-interrupt handlers, in progress on entry will have completed
2120 * before this primitive returns. However, this does not guarantee that
2121 * softirq handlers will have completed, since in some kernels, these
2122 * handlers can run in process context, and can block.
2124 * This primitive provides the guarantees made by the (now removed)
2125 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2126 * guarantees that rcu_read_lock() sections will have completed.
2127 * In "classic RCU", these two guarantees happen to be one and
2128 * the same, but can differ in realtime RCU implementations.
2130 void synchronize_sched(void)
2132 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2133 !lock_is_held(&rcu_lock_map
) &&
2134 !lock_is_held(&rcu_sched_lock_map
),
2135 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2136 if (rcu_blocking_is_gp())
2138 wait_rcu_gp(call_rcu_sched
);
2140 EXPORT_SYMBOL_GPL(synchronize_sched
);
2143 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2145 * Control will return to the caller some time after a full rcu_bh grace
2146 * period has elapsed, in other words after all currently executing rcu_bh
2147 * read-side critical sections have completed. RCU read-side critical
2148 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2149 * and may be nested.
2151 void synchronize_rcu_bh(void)
2153 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2154 !lock_is_held(&rcu_lock_map
) &&
2155 !lock_is_held(&rcu_sched_lock_map
),
2156 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2157 if (rcu_blocking_is_gp())
2159 wait_rcu_gp(call_rcu_bh
);
2161 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2163 static atomic_t sync_sched_expedited_started
= ATOMIC_INIT(0);
2164 static atomic_t sync_sched_expedited_done
= ATOMIC_INIT(0);
2166 static int synchronize_sched_expedited_cpu_stop(void *data
)
2169 * There must be a full memory barrier on each affected CPU
2170 * between the time that try_stop_cpus() is called and the
2171 * time that it returns.
2173 * In the current initial implementation of cpu_stop, the
2174 * above condition is already met when the control reaches
2175 * this point and the following smp_mb() is not strictly
2176 * necessary. Do smp_mb() anyway for documentation and
2177 * robustness against future implementation changes.
2179 smp_mb(); /* See above comment block. */
2184 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2186 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2187 * approach to force the grace period to end quickly. This consumes
2188 * significant time on all CPUs and is unfriendly to real-time workloads,
2189 * so is thus not recommended for any sort of common-case code. In fact,
2190 * if you are using synchronize_sched_expedited() in a loop, please
2191 * restructure your code to batch your updates, and then use a single
2192 * synchronize_sched() instead.
2194 * Note that it is illegal to call this function while holding any lock
2195 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2196 * to call this function from a CPU-hotplug notifier. Failing to observe
2197 * these restriction will result in deadlock.
2199 * This implementation can be thought of as an application of ticket
2200 * locking to RCU, with sync_sched_expedited_started and
2201 * sync_sched_expedited_done taking on the roles of the halves
2202 * of the ticket-lock word. Each task atomically increments
2203 * sync_sched_expedited_started upon entry, snapshotting the old value,
2204 * then attempts to stop all the CPUs. If this succeeds, then each
2205 * CPU will have executed a context switch, resulting in an RCU-sched
2206 * grace period. We are then done, so we use atomic_cmpxchg() to
2207 * update sync_sched_expedited_done to match our snapshot -- but
2208 * only if someone else has not already advanced past our snapshot.
2210 * On the other hand, if try_stop_cpus() fails, we check the value
2211 * of sync_sched_expedited_done. If it has advanced past our
2212 * initial snapshot, then someone else must have forced a grace period
2213 * some time after we took our snapshot. In this case, our work is
2214 * done for us, and we can simply return. Otherwise, we try again,
2215 * but keep our initial snapshot for purposes of checking for someone
2216 * doing our work for us.
2218 * If we fail too many times in a row, we fall back to synchronize_sched().
2220 void synchronize_sched_expedited(void)
2222 int firstsnap
, s
, snap
, trycount
= 0;
2224 /* Note that atomic_inc_return() implies full memory barrier. */
2225 firstsnap
= snap
= atomic_inc_return(&sync_sched_expedited_started
);
2227 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2230 * Each pass through the following loop attempts to force a
2231 * context switch on each CPU.
2233 while (try_stop_cpus(cpu_online_mask
,
2234 synchronize_sched_expedited_cpu_stop
,
2238 /* No joy, try again later. Or just synchronize_sched(). */
2239 if (trycount
++ < 10) {
2240 udelay(trycount
* num_online_cpus());
2242 synchronize_sched();
2246 /* Check to see if someone else did our work for us. */
2247 s
= atomic_read(&sync_sched_expedited_done
);
2248 if (UINT_CMP_GE((unsigned)s
, (unsigned)firstsnap
)) {
2249 smp_mb(); /* ensure test happens before caller kfree */
2254 * Refetching sync_sched_expedited_started allows later
2255 * callers to piggyback on our grace period. We subtract
2256 * 1 to get the same token that the last incrementer got.
2257 * We retry after they started, so our grace period works
2258 * for them, and they started after our first try, so their
2259 * grace period works for us.
2262 snap
= atomic_read(&sync_sched_expedited_started
);
2263 smp_mb(); /* ensure read is before try_stop_cpus(). */
2267 * Everyone up to our most recent fetch is covered by our grace
2268 * period. Update the counter, but only if our work is still
2269 * relevant -- which it won't be if someone who started later
2270 * than we did beat us to the punch.
2273 s
= atomic_read(&sync_sched_expedited_done
);
2274 if (UINT_CMP_GE((unsigned)s
, (unsigned)snap
)) {
2275 smp_mb(); /* ensure test happens before caller kfree */
2278 } while (atomic_cmpxchg(&sync_sched_expedited_done
, s
, snap
) != s
);
2282 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2285 * Check to see if there is any immediate RCU-related work to be done
2286 * by the current CPU, for the specified type of RCU, returning 1 if so.
2287 * The checks are in order of increasing expense: checks that can be
2288 * carried out against CPU-local state are performed first. However,
2289 * we must check for CPU stalls first, else we might not get a chance.
2291 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2293 struct rcu_node
*rnp
= rdp
->mynode
;
2295 rdp
->n_rcu_pending
++;
2297 /* Check for CPU stalls, if enabled. */
2298 check_cpu_stall(rsp
, rdp
);
2300 /* Is the RCU core waiting for a quiescent state from this CPU? */
2301 if (rcu_scheduler_fully_active
&&
2302 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2303 rdp
->n_rp_qs_pending
++;
2304 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2305 rdp
->n_rp_report_qs
++;
2309 /* Does this CPU have callbacks ready to invoke? */
2310 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2311 rdp
->n_rp_cb_ready
++;
2315 /* Has RCU gone idle with this CPU needing another grace period? */
2316 if (cpu_needs_another_gp(rsp
, rdp
)) {
2317 rdp
->n_rp_cpu_needs_gp
++;
2321 /* Has another RCU grace period completed? */
2322 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2323 rdp
->n_rp_gp_completed
++;
2327 /* Has a new RCU grace period started? */
2328 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2329 rdp
->n_rp_gp_started
++;
2334 rdp
->n_rp_need_nothing
++;
2339 * Check to see if there is any immediate RCU-related work to be done
2340 * by the current CPU, returning 1 if so. This function is part of the
2341 * RCU implementation; it is -not- an exported member of the RCU API.
2343 static int rcu_pending(int cpu
)
2345 struct rcu_state
*rsp
;
2347 for_each_rcu_flavor(rsp
)
2348 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2354 * Check to see if any future RCU-related work will need to be done
2355 * by the current CPU, even if none need be done immediately, returning
2358 static int rcu_cpu_has_callbacks(int cpu
)
2360 struct rcu_state
*rsp
;
2362 /* RCU callbacks either ready or pending? */
2363 for_each_rcu_flavor(rsp
)
2364 if (per_cpu_ptr(rsp
->rda
, cpu
)->nxtlist
)
2370 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2371 * the compiler is expected to optimize this away.
2373 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2374 int cpu
, unsigned long done
)
2376 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2377 atomic_read(&rsp
->barrier_cpu_count
), done
);
2381 * RCU callback function for _rcu_barrier(). If we are last, wake
2382 * up the task executing _rcu_barrier().
2384 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2386 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2387 struct rcu_state
*rsp
= rdp
->rsp
;
2389 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2390 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2391 complete(&rsp
->barrier_completion
);
2393 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2398 * Called with preemption disabled, and from cross-cpu IRQ context.
2400 static void rcu_barrier_func(void *type
)
2402 struct rcu_state
*rsp
= type
;
2403 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2405 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2406 atomic_inc(&rsp
->barrier_cpu_count
);
2407 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2411 * Orchestrate the specified type of RCU barrier, waiting for all
2412 * RCU callbacks of the specified type to complete.
2414 static void _rcu_barrier(struct rcu_state
*rsp
)
2417 struct rcu_data
*rdp
;
2418 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2419 unsigned long snap_done
;
2421 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2423 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2424 mutex_lock(&rsp
->barrier_mutex
);
2427 * Ensure that all prior references, including to ->n_barrier_done,
2428 * are ordered before the _rcu_barrier() machinery.
2430 smp_mb(); /* See above block comment. */
2433 * Recheck ->n_barrier_done to see if others did our work for us.
2434 * This means checking ->n_barrier_done for an even-to-odd-to-even
2435 * transition. The "if" expression below therefore rounds the old
2436 * value up to the next even number and adds two before comparing.
2438 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2439 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2440 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2441 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2442 smp_mb(); /* caller's subsequent code after above check. */
2443 mutex_unlock(&rsp
->barrier_mutex
);
2448 * Increment ->n_barrier_done to avoid duplicate work. Use
2449 * ACCESS_ONCE() to prevent the compiler from speculating
2450 * the increment to precede the early-exit check.
2452 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2453 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2454 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2455 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2458 * Initialize the count to one rather than to zero in order to
2459 * avoid a too-soon return to zero in case of a short grace period
2460 * (or preemption of this task). Exclude CPU-hotplug operations
2461 * to ensure that no offline CPU has callbacks queued.
2463 init_completion(&rsp
->barrier_completion
);
2464 atomic_set(&rsp
->barrier_cpu_count
, 1);
2468 * Force each CPU with callbacks to register a new callback.
2469 * When that callback is invoked, we will know that all of the
2470 * corresponding CPU's preceding callbacks have been invoked.
2472 for_each_online_cpu(cpu
) {
2473 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2474 if (ACCESS_ONCE(rdp
->qlen
)) {
2475 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2476 rsp
->n_barrier_done
);
2477 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2479 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2480 rsp
->n_barrier_done
);
2486 * Now that we have an rcu_barrier_callback() callback on each
2487 * CPU, and thus each counted, remove the initial count.
2489 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2490 complete(&rsp
->barrier_completion
);
2492 /* Increment ->n_barrier_done to prevent duplicate work. */
2493 smp_mb(); /* Keep increment after above mechanism. */
2494 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2495 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2496 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2497 smp_mb(); /* Keep increment before caller's subsequent code. */
2499 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2500 wait_for_completion(&rsp
->barrier_completion
);
2502 /* Other rcu_barrier() invocations can now safely proceed. */
2503 mutex_unlock(&rsp
->barrier_mutex
);
2507 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2509 void rcu_barrier_bh(void)
2511 _rcu_barrier(&rcu_bh_state
);
2513 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2516 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2518 void rcu_barrier_sched(void)
2520 _rcu_barrier(&rcu_sched_state
);
2522 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2525 * Do boot-time initialization of a CPU's per-CPU RCU data.
2528 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2530 unsigned long flags
;
2531 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2532 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2534 /* Set up local state, ensuring consistent view of global state. */
2535 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2536 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2537 init_callback_list(rdp
);
2539 ACCESS_ONCE(rdp
->qlen
) = 0;
2540 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2541 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2542 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2545 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2549 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2550 * offline event can be happening at a given time. Note also that we
2551 * can accept some slop in the rsp->completed access due to the fact
2552 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2554 static void __cpuinit
2555 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2557 unsigned long flags
;
2559 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2560 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2562 /* Set up local state, ensuring consistent view of global state. */
2563 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2564 rdp
->beenonline
= 1; /* We have now been online. */
2565 rdp
->preemptible
= preemptible
;
2566 rdp
->qlen_last_fqs_check
= 0;
2567 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2568 rdp
->blimit
= blimit
;
2569 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
2570 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2571 atomic_set(&rdp
->dynticks
->dynticks
,
2572 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2573 rcu_prepare_for_idle_init(cpu
);
2574 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2577 * A new grace period might start here. If so, we won't be part
2578 * of it, but that is OK, as we are currently in a quiescent state.
2581 /* Exclude any attempts to start a new GP on large systems. */
2582 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2584 /* Add CPU to rcu_node bitmasks. */
2586 mask
= rdp
->grpmask
;
2588 /* Exclude any attempts to start a new GP on small systems. */
2589 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2590 rnp
->qsmaskinit
|= mask
;
2591 mask
= rnp
->grpmask
;
2592 if (rnp
== rdp
->mynode
) {
2594 * If there is a grace period in progress, we will
2595 * set up to wait for it next time we run the
2598 rdp
->gpnum
= rnp
->completed
;
2599 rdp
->completed
= rnp
->completed
;
2600 rdp
->passed_quiesce
= 0;
2601 rdp
->qs_pending
= 0;
2602 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2604 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2606 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2608 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2611 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2613 struct rcu_state
*rsp
;
2615 for_each_rcu_flavor(rsp
)
2616 rcu_init_percpu_data(cpu
, rsp
,
2617 strcmp(rsp
->name
, "rcu_preempt") == 0);
2621 * Handle CPU online/offline notification events.
2623 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2624 unsigned long action
, void *hcpu
)
2626 long cpu
= (long)hcpu
;
2627 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2628 struct rcu_node
*rnp
= rdp
->mynode
;
2629 struct rcu_state
*rsp
;
2631 trace_rcu_utilization("Start CPU hotplug");
2633 case CPU_UP_PREPARE
:
2634 case CPU_UP_PREPARE_FROZEN
:
2635 rcu_prepare_cpu(cpu
);
2636 rcu_prepare_kthreads(cpu
);
2639 case CPU_DOWN_FAILED
:
2640 rcu_boost_kthread_setaffinity(rnp
, -1);
2642 case CPU_DOWN_PREPARE
:
2643 rcu_boost_kthread_setaffinity(rnp
, cpu
);
2646 case CPU_DYING_FROZEN
:
2648 * The whole machine is "stopped" except this CPU, so we can
2649 * touch any data without introducing corruption. We send the
2650 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2652 for_each_rcu_flavor(rsp
)
2653 rcu_cleanup_dying_cpu(rsp
);
2654 rcu_cleanup_after_idle(cpu
);
2657 case CPU_DEAD_FROZEN
:
2658 case CPU_UP_CANCELED
:
2659 case CPU_UP_CANCELED_FROZEN
:
2660 for_each_rcu_flavor(rsp
)
2661 rcu_cleanup_dead_cpu(cpu
, rsp
);
2666 trace_rcu_utilization("End CPU hotplug");
2671 * Spawn the kthread that handles this RCU flavor's grace periods.
2673 static int __init
rcu_spawn_gp_kthread(void)
2675 unsigned long flags
;
2676 struct rcu_node
*rnp
;
2677 struct rcu_state
*rsp
;
2678 struct task_struct
*t
;
2680 for_each_rcu_flavor(rsp
) {
2681 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
2683 rnp
= rcu_get_root(rsp
);
2684 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2685 rsp
->gp_kthread
= t
;
2686 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2690 early_initcall(rcu_spawn_gp_kthread
);
2693 * This function is invoked towards the end of the scheduler's initialization
2694 * process. Before this is called, the idle task might contain
2695 * RCU read-side critical sections (during which time, this idle
2696 * task is booting the system). After this function is called, the
2697 * idle tasks are prohibited from containing RCU read-side critical
2698 * sections. This function also enables RCU lockdep checking.
2700 void rcu_scheduler_starting(void)
2702 WARN_ON(num_online_cpus() != 1);
2703 WARN_ON(nr_context_switches() > 0);
2704 rcu_scheduler_active
= 1;
2708 * Compute the per-level fanout, either using the exact fanout specified
2709 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2711 #ifdef CONFIG_RCU_FANOUT_EXACT
2712 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2716 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
2717 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2718 rsp
->levelspread
[0] = rcu_fanout_leaf
;
2720 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2721 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2728 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2729 ccur
= rsp
->levelcnt
[i
];
2730 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2734 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2737 * Helper function for rcu_init() that initializes one rcu_state structure.
2739 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2740 struct rcu_data __percpu
*rda
)
2742 static char *buf
[] = { "rcu_node_0",
2745 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2746 static char *fqs
[] = { "rcu_node_fqs_0",
2749 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2753 struct rcu_node
*rnp
;
2755 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2757 /* Initialize the level-tracking arrays. */
2759 for (i
= 0; i
< rcu_num_lvls
; i
++)
2760 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
2761 for (i
= 1; i
< rcu_num_lvls
; i
++)
2762 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2763 rcu_init_levelspread(rsp
);
2765 /* Initialize the elements themselves, starting from the leaves. */
2767 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2768 cpustride
*= rsp
->levelspread
[i
];
2769 rnp
= rsp
->level
[i
];
2770 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2771 raw_spin_lock_init(&rnp
->lock
);
2772 lockdep_set_class_and_name(&rnp
->lock
,
2773 &rcu_node_class
[i
], buf
[i
]);
2774 raw_spin_lock_init(&rnp
->fqslock
);
2775 lockdep_set_class_and_name(&rnp
->fqslock
,
2776 &rcu_fqs_class
[i
], fqs
[i
]);
2777 rnp
->gpnum
= rsp
->gpnum
;
2778 rnp
->completed
= rsp
->completed
;
2780 rnp
->qsmaskinit
= 0;
2781 rnp
->grplo
= j
* cpustride
;
2782 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2783 if (rnp
->grphi
>= NR_CPUS
)
2784 rnp
->grphi
= NR_CPUS
- 1;
2790 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2791 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2792 rnp
->parent
= rsp
->level
[i
- 1] +
2793 j
/ rsp
->levelspread
[i
- 1];
2796 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2801 init_waitqueue_head(&rsp
->gp_wq
);
2802 rnp
= rsp
->level
[rcu_num_lvls
- 1];
2803 for_each_possible_cpu(i
) {
2804 while (i
> rnp
->grphi
)
2806 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2807 rcu_boot_init_percpu_data(i
, rsp
);
2809 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
2813 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2814 * replace the definitions in rcutree.h because those are needed to size
2815 * the ->node array in the rcu_state structure.
2817 static void __init
rcu_init_geometry(void)
2822 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
2824 /* If the compile-time values are accurate, just leave. */
2825 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
2826 nr_cpu_ids
== NR_CPUS
)
2830 * Compute number of nodes that can be handled an rcu_node tree
2831 * with the given number of levels. Setting rcu_capacity[0] makes
2832 * some of the arithmetic easier.
2834 rcu_capacity
[0] = 1;
2835 rcu_capacity
[1] = rcu_fanout_leaf
;
2836 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
2837 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
2840 * The boot-time rcu_fanout_leaf parameter is only permitted
2841 * to increase the leaf-level fanout, not decrease it. Of course,
2842 * the leaf-level fanout cannot exceed the number of bits in
2843 * the rcu_node masks. Finally, the tree must be able to accommodate
2844 * the configured number of CPUs. Complain and fall back to the
2845 * compile-time values if these limits are exceeded.
2847 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
2848 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
2849 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
2854 /* Calculate the number of rcu_nodes at each level of the tree. */
2855 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
2856 if (n
<= rcu_capacity
[i
]) {
2857 for (j
= 0; j
<= i
; j
++)
2859 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
2861 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
2866 /* Calculate the total number of rcu_node structures. */
2868 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
2869 rcu_num_nodes
+= num_rcu_lvl
[i
];
2873 void __init
rcu_init(void)
2877 rcu_bootup_announce();
2878 rcu_init_geometry();
2879 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2880 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2881 __rcu_init_preempt();
2882 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2885 * We don't need protection against CPU-hotplug here because
2886 * this is called early in boot, before either interrupts
2887 * or the scheduler are operational.
2889 cpu_notifier(rcu_cpu_notify
, 0);
2890 for_each_online_cpu(cpu
)
2891 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2892 check_cpu_stall_init();
2895 #include "rcutree_plugin.h"