2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
55 #include <trace/events/rcu.h>
59 /* Data structures. */
61 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
63 #define RCU_STATE_INITIALIZER(structname) { \
64 .level = { &structname##_state.node[0] }, \
66 NUM_RCU_LVL_0, /* root of hierarchy. */ \
70 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
72 .fqs_state = RCU_GP_IDLE, \
75 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
76 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
78 .n_force_qs_ngp = 0, \
79 .name = #structname, \
82 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched
);
83 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
85 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
);
86 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
88 static struct rcu_state
*rcu_state
;
91 * The rcu_scheduler_active variable transitions from zero to one just
92 * before the first task is spawned. So when this variable is zero, RCU
93 * can assume that there is but one task, allowing RCU to (for example)
94 * optimized synchronize_sched() to a simple barrier(). When this variable
95 * is one, RCU must actually do all the hard work required to detect real
96 * grace periods. This variable is also used to suppress boot-time false
97 * positives from lockdep-RCU error checking.
99 int rcu_scheduler_active __read_mostly
;
100 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
103 * The rcu_scheduler_fully_active variable transitions from zero to one
104 * during the early_initcall() processing, which is after the scheduler
105 * is capable of creating new tasks. So RCU processing (for example,
106 * creating tasks for RCU priority boosting) must be delayed until after
107 * rcu_scheduler_fully_active transitions from zero to one. We also
108 * currently delay invocation of any RCU callbacks until after this point.
110 * It might later prove better for people registering RCU callbacks during
111 * early boot to take responsibility for these callbacks, but one step at
114 static int rcu_scheduler_fully_active __read_mostly
;
116 #ifdef CONFIG_RCU_BOOST
119 * Control variables for per-CPU and per-rcu_node kthreads. These
120 * handle all flavors of RCU.
122 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
123 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
124 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
125 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
126 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
128 #endif /* #ifdef CONFIG_RCU_BOOST */
130 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
131 static void invoke_rcu_core(void);
132 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
135 * Track the rcutorture test sequence number and the update version
136 * number within a given test. The rcutorture_testseq is incremented
137 * on every rcutorture module load and unload, so has an odd value
138 * when a test is running. The rcutorture_vernum is set to zero
139 * when rcutorture starts and is incremented on each rcutorture update.
140 * These variables enable correlating rcutorture output with the
141 * RCU tracing information.
143 unsigned long rcutorture_testseq
;
144 unsigned long rcutorture_vernum
;
147 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
148 * permit this function to be invoked without holding the root rcu_node
149 * structure's ->lock, but of course results can be subject to change.
151 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
153 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
157 * Note a quiescent state. Because we do not need to know
158 * how many quiescent states passed, just if there was at least
159 * one since the start of the grace period, this just sets a flag.
160 * The caller must have disabled preemption.
162 void rcu_sched_qs(int cpu
)
164 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
166 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
168 if (rdp
->passed_quiesce
== 0)
169 trace_rcu_grace_period("rcu_sched", rdp
->gpnum
, "cpuqs");
170 rdp
->passed_quiesce
= 1;
173 void rcu_bh_qs(int cpu
)
175 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
177 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
179 if (rdp
->passed_quiesce
== 0)
180 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
181 rdp
->passed_quiesce
= 1;
185 * Note a context switch. This is a quiescent state for RCU-sched,
186 * and requires special handling for preemptible RCU.
187 * The caller must have disabled preemption.
189 void rcu_note_context_switch(int cpu
)
191 trace_rcu_utilization("Start context switch");
193 rcu_preempt_note_context_switch(cpu
);
194 trace_rcu_utilization("End context switch");
196 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
198 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
199 .dynticks_nesting
= DYNTICK_TASK_NESTING
,
200 .dynticks
= ATOMIC_INIT(1),
203 static int blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
204 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
205 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
207 module_param(blimit
, int, 0);
208 module_param(qhimark
, int, 0);
209 module_param(qlowmark
, int, 0);
211 int rcu_cpu_stall_suppress __read_mostly
;
212 module_param(rcu_cpu_stall_suppress
, int, 0644);
214 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
215 static int rcu_pending(int cpu
);
218 * Return the number of RCU-sched batches processed thus far for debug & stats.
220 long rcu_batches_completed_sched(void)
222 return rcu_sched_state
.completed
;
224 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
227 * Return the number of RCU BH batches processed thus far for debug & stats.
229 long rcu_batches_completed_bh(void)
231 return rcu_bh_state
.completed
;
233 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
236 * Force a quiescent state for RCU BH.
238 void rcu_bh_force_quiescent_state(void)
240 force_quiescent_state(&rcu_bh_state
, 0);
242 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
245 * Record the number of times rcutorture tests have been initiated and
246 * terminated. This information allows the debugfs tracing stats to be
247 * correlated to the rcutorture messages, even when the rcutorture module
248 * is being repeatedly loaded and unloaded. In other words, we cannot
249 * store this state in rcutorture itself.
251 void rcutorture_record_test_transition(void)
253 rcutorture_testseq
++;
254 rcutorture_vernum
= 0;
256 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
259 * Record the number of writer passes through the current rcutorture test.
260 * This is also used to correlate debugfs tracing stats with the rcutorture
263 void rcutorture_record_progress(unsigned long vernum
)
267 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
270 * Force a quiescent state for RCU-sched.
272 void rcu_sched_force_quiescent_state(void)
274 force_quiescent_state(&rcu_sched_state
, 0);
276 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
279 * Does the CPU have callbacks ready to be invoked?
282 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
284 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
288 * Does the current CPU require a yet-as-unscheduled grace period?
291 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
293 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
297 * Return the root node of the specified rcu_state structure.
299 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
301 return &rsp
->node
[0];
305 * If the specified CPU is offline, tell the caller that it is in
306 * a quiescent state. Otherwise, whack it with a reschedule IPI.
307 * Grace periods can end up waiting on an offline CPU when that
308 * CPU is in the process of coming online -- it will be added to the
309 * rcu_node bitmasks before it actually makes it online. The same thing
310 * can happen while a CPU is in the process of coming online. Because this
311 * race is quite rare, we check for it after detecting that the grace
312 * period has been delayed rather than checking each and every CPU
313 * each and every time we start a new grace period.
315 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
318 * If the CPU is offline, it is in a quiescent state. We can
319 * trust its state not to change because interrupts are disabled.
321 if (cpu_is_offline(rdp
->cpu
)) {
322 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
328 * The CPU is online, so send it a reschedule IPI. This forces
329 * it through the scheduler, and (inefficiently) also handles cases
330 * where idle loops fail to inform RCU about the CPU being idle.
332 if (rdp
->cpu
!= smp_processor_id())
333 smp_send_reschedule(rdp
->cpu
);
341 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
343 * If the new value of the ->dynticks_nesting counter now is zero,
344 * we really have entered idle, and must do the appropriate accounting.
345 * The caller must have disabled interrupts.
347 static void rcu_idle_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
)
349 trace_rcu_dyntick("Start", oldval
, 0);
350 if (!is_idle_task(current
)) {
351 struct task_struct
*idle
= idle_task(smp_processor_id());
353 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
354 ftrace_dump(DUMP_ALL
);
355 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
356 current
->pid
, current
->comm
,
357 idle
->pid
, idle
->comm
); /* must be idle task! */
359 rcu_prepare_for_idle(smp_processor_id());
360 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
361 smp_mb__before_atomic_inc(); /* See above. */
362 atomic_inc(&rdtp
->dynticks
);
363 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
364 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
367 * The idle task is not permitted to enter the idle loop while
368 * in an RCU read-side critical section.
370 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
371 "Illegal idle entry in RCU read-side critical section.");
372 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
373 "Illegal idle entry in RCU-bh read-side critical section.");
374 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
375 "Illegal idle entry in RCU-sched read-side critical section.");
379 * rcu_idle_enter - inform RCU that current CPU is entering idle
381 * Enter idle mode, in other words, -leave- the mode in which RCU
382 * read-side critical sections can occur. (Though RCU read-side
383 * critical sections can occur in irq handlers in idle, a possibility
384 * handled by irq_enter() and irq_exit().)
386 * We crowbar the ->dynticks_nesting field to zero to allow for
387 * the possibility of usermode upcalls having messed up our count
388 * of interrupt nesting level during the prior busy period.
390 void rcu_idle_enter(void)
394 struct rcu_dynticks
*rdtp
;
396 local_irq_save(flags
);
397 rdtp
= &__get_cpu_var(rcu_dynticks
);
398 oldval
= rdtp
->dynticks_nesting
;
399 rdtp
->dynticks_nesting
= 0;
400 rcu_idle_enter_common(rdtp
, oldval
);
401 local_irq_restore(flags
);
405 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
407 * Exit from an interrupt handler, which might possibly result in entering
408 * idle mode, in other words, leaving the mode in which read-side critical
409 * sections can occur.
411 * This code assumes that the idle loop never does anything that might
412 * result in unbalanced calls to irq_enter() and irq_exit(). If your
413 * architecture violates this assumption, RCU will give you what you
414 * deserve, good and hard. But very infrequently and irreproducibly.
416 * Use things like work queues to work around this limitation.
418 * You have been warned.
420 void rcu_irq_exit(void)
424 struct rcu_dynticks
*rdtp
;
426 local_irq_save(flags
);
427 rdtp
= &__get_cpu_var(rcu_dynticks
);
428 oldval
= rdtp
->dynticks_nesting
;
429 rdtp
->dynticks_nesting
--;
430 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
431 if (rdtp
->dynticks_nesting
)
432 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
434 rcu_idle_enter_common(rdtp
, oldval
);
435 local_irq_restore(flags
);
439 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
441 * If the new value of the ->dynticks_nesting counter was previously zero,
442 * we really have exited idle, and must do the appropriate accounting.
443 * The caller must have disabled interrupts.
445 static void rcu_idle_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
)
447 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
448 atomic_inc(&rdtp
->dynticks
);
449 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
450 smp_mb__after_atomic_inc(); /* See above. */
451 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
452 rcu_cleanup_after_idle(smp_processor_id());
453 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
454 if (!is_idle_task(current
)) {
455 struct task_struct
*idle
= idle_task(smp_processor_id());
457 trace_rcu_dyntick("Error on exit: not idle task",
458 oldval
, rdtp
->dynticks_nesting
);
459 ftrace_dump(DUMP_ALL
);
460 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
461 current
->pid
, current
->comm
,
462 idle
->pid
, idle
->comm
); /* must be idle task! */
467 * rcu_idle_exit - inform RCU that current CPU is leaving idle
469 * Exit idle mode, in other words, -enter- the mode in which RCU
470 * read-side critical sections can occur.
472 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NESTING to
473 * allow for the possibility of usermode upcalls messing up our count
474 * of interrupt nesting level during the busy period that is just
477 void rcu_idle_exit(void)
480 struct rcu_dynticks
*rdtp
;
483 local_irq_save(flags
);
484 rdtp
= &__get_cpu_var(rcu_dynticks
);
485 oldval
= rdtp
->dynticks_nesting
;
486 WARN_ON_ONCE(oldval
!= 0);
487 rdtp
->dynticks_nesting
= DYNTICK_TASK_NESTING
;
488 rcu_idle_exit_common(rdtp
, oldval
);
489 local_irq_restore(flags
);
493 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
495 * Enter an interrupt handler, which might possibly result in exiting
496 * idle mode, in other words, entering the mode in which read-side critical
497 * sections can occur.
499 * Note that the Linux kernel is fully capable of entering an interrupt
500 * handler that it never exits, for example when doing upcalls to
501 * user mode! This code assumes that the idle loop never does upcalls to
502 * user mode. If your architecture does do upcalls from the idle loop (or
503 * does anything else that results in unbalanced calls to the irq_enter()
504 * and irq_exit() functions), RCU will give you what you deserve, good
505 * and hard. But very infrequently and irreproducibly.
507 * Use things like work queues to work around this limitation.
509 * You have been warned.
511 void rcu_irq_enter(void)
514 struct rcu_dynticks
*rdtp
;
517 local_irq_save(flags
);
518 rdtp
= &__get_cpu_var(rcu_dynticks
);
519 oldval
= rdtp
->dynticks_nesting
;
520 rdtp
->dynticks_nesting
++;
521 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
523 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
525 rcu_idle_exit_common(rdtp
, oldval
);
526 local_irq_restore(flags
);
530 * rcu_nmi_enter - inform RCU of entry to NMI context
532 * If the CPU was idle with dynamic ticks active, and there is no
533 * irq handler running, this updates rdtp->dynticks_nmi to let the
534 * RCU grace-period handling know that the CPU is active.
536 void rcu_nmi_enter(void)
538 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
540 if (rdtp
->dynticks_nmi_nesting
== 0 &&
541 (atomic_read(&rdtp
->dynticks
) & 0x1))
543 rdtp
->dynticks_nmi_nesting
++;
544 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
545 atomic_inc(&rdtp
->dynticks
);
546 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
547 smp_mb__after_atomic_inc(); /* See above. */
548 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
552 * rcu_nmi_exit - inform RCU of exit from NMI context
554 * If the CPU was idle with dynamic ticks active, and there is no
555 * irq handler running, this updates rdtp->dynticks_nmi to let the
556 * RCU grace-period handling know that the CPU is no longer active.
558 void rcu_nmi_exit(void)
560 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
562 if (rdtp
->dynticks_nmi_nesting
== 0 ||
563 --rdtp
->dynticks_nmi_nesting
!= 0)
565 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
566 smp_mb__before_atomic_inc(); /* See above. */
567 atomic_inc(&rdtp
->dynticks
);
568 smp_mb__after_atomic_inc(); /* Force delay to next write. */
569 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
572 #ifdef CONFIG_PROVE_RCU
575 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
577 * If the current CPU is in its idle loop and is neither in an interrupt
578 * or NMI handler, return true.
580 int rcu_is_cpu_idle(void)
585 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
589 EXPORT_SYMBOL(rcu_is_cpu_idle
);
591 #endif /* #ifdef CONFIG_PROVE_RCU */
594 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
596 * If the current CPU is idle or running at a first-level (not nested)
597 * interrupt from idle, return true. The caller must have at least
598 * disabled preemption.
600 int rcu_is_cpu_rrupt_from_idle(void)
602 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
606 * Snapshot the specified CPU's dynticks counter so that we can later
607 * credit them with an implicit quiescent state. Return 1 if this CPU
608 * is in dynticks idle mode, which is an extended quiescent state.
610 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
612 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
613 return (rdp
->dynticks_snap
& 0x1) == 0;
617 * Return true if the specified CPU has passed through a quiescent
618 * state by virtue of being in or having passed through an dynticks
619 * idle state since the last call to dyntick_save_progress_counter()
622 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
627 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
628 snap
= (unsigned int)rdp
->dynticks_snap
;
631 * If the CPU passed through or entered a dynticks idle phase with
632 * no active irq/NMI handlers, then we can safely pretend that the CPU
633 * already acknowledged the request to pass through a quiescent
634 * state. Either way, that CPU cannot possibly be in an RCU
635 * read-side critical section that started before the beginning
636 * of the current RCU grace period.
638 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
639 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
644 /* Go check for the CPU being offline. */
645 return rcu_implicit_offline_qs(rdp
);
648 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
650 rsp
->gp_start
= jiffies
;
651 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
654 static void print_other_cpu_stall(struct rcu_state
*rsp
)
660 struct rcu_node
*rnp
= rcu_get_root(rsp
);
662 /* Only let one CPU complain about others per time interval. */
664 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
665 delta
= jiffies
- rsp
->jiffies_stall
;
666 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
667 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
670 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
673 * Now rat on any tasks that got kicked up to the root rcu_node
674 * due to CPU offlining.
676 ndetected
= rcu_print_task_stall(rnp
);
677 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
680 * OK, time to rat on our buddy...
681 * See Documentation/RCU/stallwarn.txt for info on how to debug
682 * RCU CPU stall warnings.
684 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
686 rcu_for_each_leaf_node(rsp
, rnp
) {
687 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
688 ndetected
+= rcu_print_task_stall(rnp
);
689 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
690 if (rnp
->qsmask
== 0)
692 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
693 if (rnp
->qsmask
& (1UL << cpu
)) {
694 printk(" %d", rnp
->grplo
+ cpu
);
698 printk("} (detected by %d, t=%ld jiffies)\n",
699 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
701 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
702 else if (!trigger_all_cpu_backtrace())
705 /* If so configured, complain about tasks blocking the grace period. */
707 rcu_print_detail_task_stall(rsp
);
709 force_quiescent_state(rsp
, 0); /* Kick them all. */
712 static void print_cpu_stall(struct rcu_state
*rsp
)
715 struct rcu_node
*rnp
= rcu_get_root(rsp
);
718 * OK, time to rat on ourselves...
719 * See Documentation/RCU/stallwarn.txt for info on how to debug
720 * RCU CPU stall warnings.
722 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
723 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
724 if (!trigger_all_cpu_backtrace())
727 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
728 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
730 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
731 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
733 set_need_resched(); /* kick ourselves to get things going. */
736 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
740 struct rcu_node
*rnp
;
742 if (rcu_cpu_stall_suppress
)
744 j
= ACCESS_ONCE(jiffies
);
745 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
747 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
749 /* We haven't checked in, so go dump stack. */
750 print_cpu_stall(rsp
);
752 } else if (rcu_gp_in_progress(rsp
) &&
753 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
755 /* They had a few time units to dump stack, so complain. */
756 print_other_cpu_stall(rsp
);
760 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
762 rcu_cpu_stall_suppress
= 1;
767 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
769 * Set the stall-warning timeout way off into the future, thus preventing
770 * any RCU CPU stall-warning messages from appearing in the current set of
773 * The caller must disable hard irqs.
775 void rcu_cpu_stall_reset(void)
777 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
778 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
779 rcu_preempt_stall_reset();
782 static struct notifier_block rcu_panic_block
= {
783 .notifier_call
= rcu_panic
,
786 static void __init
check_cpu_stall_init(void)
788 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
792 * Update CPU-local rcu_data state to record the newly noticed grace period.
793 * This is used both when we started the grace period and when we notice
794 * that someone else started the grace period. The caller must hold the
795 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
796 * and must have irqs disabled.
798 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
800 if (rdp
->gpnum
!= rnp
->gpnum
) {
802 * If the current grace period is waiting for this CPU,
803 * set up to detect a quiescent state, otherwise don't
804 * go looking for one.
806 rdp
->gpnum
= rnp
->gpnum
;
807 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
808 if (rnp
->qsmask
& rdp
->grpmask
) {
810 rdp
->passed_quiesce
= 0;
816 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
819 struct rcu_node
*rnp
;
821 local_irq_save(flags
);
823 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
824 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
825 local_irq_restore(flags
);
828 __note_new_gpnum(rsp
, rnp
, rdp
);
829 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
833 * Did someone else start a new RCU grace period start since we last
834 * checked? Update local state appropriately if so. Must be called
835 * on the CPU corresponding to rdp.
838 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
843 local_irq_save(flags
);
844 if (rdp
->gpnum
!= rsp
->gpnum
) {
845 note_new_gpnum(rsp
, rdp
);
848 local_irq_restore(flags
);
853 * Advance this CPU's callbacks, but only if the current grace period
854 * has ended. This may be called only from the CPU to whom the rdp
855 * belongs. In addition, the corresponding leaf rcu_node structure's
856 * ->lock must be held by the caller, with irqs disabled.
859 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
861 /* Did another grace period end? */
862 if (rdp
->completed
!= rnp
->completed
) {
864 /* Advance callbacks. No harm if list empty. */
865 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
866 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
867 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
869 /* Remember that we saw this grace-period completion. */
870 rdp
->completed
= rnp
->completed
;
871 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
874 * If we were in an extended quiescent state, we may have
875 * missed some grace periods that others CPUs handled on
876 * our behalf. Catch up with this state to avoid noting
877 * spurious new grace periods. If another grace period
878 * has started, then rnp->gpnum will have advanced, so
879 * we will detect this later on.
881 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
882 rdp
->gpnum
= rdp
->completed
;
885 * If RCU does not need a quiescent state from this CPU,
886 * then make sure that this CPU doesn't go looking for one.
888 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
894 * Advance this CPU's callbacks, but only if the current grace period
895 * has ended. This may be called only from the CPU to whom the rdp
899 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
902 struct rcu_node
*rnp
;
904 local_irq_save(flags
);
906 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
907 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
908 local_irq_restore(flags
);
911 __rcu_process_gp_end(rsp
, rnp
, rdp
);
912 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
916 * Do per-CPU grace-period initialization for running CPU. The caller
917 * must hold the lock of the leaf rcu_node structure corresponding to
921 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
923 /* Prior grace period ended, so advance callbacks for current CPU. */
924 __rcu_process_gp_end(rsp
, rnp
, rdp
);
927 * Because this CPU just now started the new grace period, we know
928 * that all of its callbacks will be covered by this upcoming grace
929 * period, even the ones that were registered arbitrarily recently.
930 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
932 * Other CPUs cannot be sure exactly when the grace period started.
933 * Therefore, their recently registered callbacks must pass through
934 * an additional RCU_NEXT_READY stage, so that they will be handled
935 * by the next RCU grace period.
937 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
938 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
940 /* Set state so that this CPU will detect the next quiescent state. */
941 __note_new_gpnum(rsp
, rnp
, rdp
);
945 * Start a new RCU grace period if warranted, re-initializing the hierarchy
946 * in preparation for detecting the next grace period. The caller must hold
947 * the root node's ->lock, which is released before return. Hard irqs must
950 * Note that it is legal for a dying CPU (which is marked as offline) to
951 * invoke this function. This can happen when the dying CPU reports its
955 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
956 __releases(rcu_get_root(rsp
)->lock
)
958 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
959 struct rcu_node
*rnp
= rcu_get_root(rsp
);
961 if (!rcu_scheduler_fully_active
||
962 !cpu_needs_another_gp(rsp
, rdp
)) {
964 * Either the scheduler hasn't yet spawned the first
965 * non-idle task or this CPU does not need another
966 * grace period. Either way, don't start a new grace
969 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
973 if (rsp
->fqs_active
) {
975 * This CPU needs a grace period, but force_quiescent_state()
976 * is running. Tell it to start one on this CPU's behalf.
978 rsp
->fqs_need_gp
= 1;
979 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
983 /* Advance to a new grace period and initialize state. */
985 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
986 WARN_ON_ONCE(rsp
->fqs_state
== RCU_GP_INIT
);
987 rsp
->fqs_state
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
988 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
989 record_gp_stall_check_time(rsp
);
990 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
992 /* Exclude any concurrent CPU-hotplug operations. */
993 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
996 * Set the quiescent-state-needed bits in all the rcu_node
997 * structures for all currently online CPUs in breadth-first
998 * order, starting from the root rcu_node structure. This
999 * operation relies on the layout of the hierarchy within the
1000 * rsp->node[] array. Note that other CPUs will access only
1001 * the leaves of the hierarchy, which still indicate that no
1002 * grace period is in progress, at least until the corresponding
1003 * leaf node has been initialized. In addition, we have excluded
1004 * CPU-hotplug operations.
1006 * Note that the grace period cannot complete until we finish
1007 * the initialization process, as there will be at least one
1008 * qsmask bit set in the root node until that time, namely the
1009 * one corresponding to this CPU, due to the fact that we have
1012 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1013 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1014 rcu_preempt_check_blocked_tasks(rnp
);
1015 rnp
->qsmask
= rnp
->qsmaskinit
;
1016 rnp
->gpnum
= rsp
->gpnum
;
1017 rnp
->completed
= rsp
->completed
;
1018 if (rnp
== rdp
->mynode
)
1019 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1020 rcu_preempt_boost_start_gp(rnp
);
1021 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1022 rnp
->level
, rnp
->grplo
,
1023 rnp
->grphi
, rnp
->qsmask
);
1024 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1027 rnp
= rcu_get_root(rsp
);
1028 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1029 rsp
->fqs_state
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
1030 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1031 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
1035 * Report a full set of quiescent states to the specified rcu_state
1036 * data structure. This involves cleaning up after the prior grace
1037 * period and letting rcu_start_gp() start up the next grace period
1038 * if one is needed. Note that the caller must hold rnp->lock, as
1039 * required by rcu_start_gp(), which will release it.
1041 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1042 __releases(rcu_get_root(rsp
)->lock
)
1044 unsigned long gp_duration
;
1045 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1046 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1048 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1051 * Ensure that all grace-period and pre-grace-period activity
1052 * is seen before the assignment to rsp->completed.
1054 smp_mb(); /* See above block comment. */
1055 gp_duration
= jiffies
- rsp
->gp_start
;
1056 if (gp_duration
> rsp
->gp_max
)
1057 rsp
->gp_max
= gp_duration
;
1060 * We know the grace period is complete, but to everyone else
1061 * it appears to still be ongoing. But it is also the case
1062 * that to everyone else it looks like there is nothing that
1063 * they can do to advance the grace period. It is therefore
1064 * safe for us to drop the lock in order to mark the grace
1065 * period as completed in all of the rcu_node structures.
1067 * But if this CPU needs another grace period, it will take
1068 * care of this while initializing the next grace period.
1069 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1070 * because the callbacks have not yet been advanced: Those
1071 * callbacks are waiting on the grace period that just now
1074 if (*rdp
->nxttail
[RCU_WAIT_TAIL
] == NULL
) {
1075 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1078 * Propagate new ->completed value to rcu_node structures
1079 * so that other CPUs don't have to wait until the start
1080 * of the next grace period to process their callbacks.
1082 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1083 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1084 rnp
->completed
= rsp
->gpnum
;
1085 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1087 rnp
= rcu_get_root(rsp
);
1088 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1091 rsp
->completed
= rsp
->gpnum
; /* Declare the grace period complete. */
1092 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1093 rsp
->fqs_state
= RCU_GP_IDLE
;
1094 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
1098 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1099 * Allows quiescent states for a group of CPUs to be reported at one go
1100 * to the specified rcu_node structure, though all the CPUs in the group
1101 * must be represented by the same rcu_node structure (which need not be
1102 * a leaf rcu_node structure, though it often will be). That structure's
1103 * lock must be held upon entry, and it is released before return.
1106 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1107 struct rcu_node
*rnp
, unsigned long flags
)
1108 __releases(rnp
->lock
)
1110 struct rcu_node
*rnp_c
;
1112 /* Walk up the rcu_node hierarchy. */
1114 if (!(rnp
->qsmask
& mask
)) {
1116 /* Our bit has already been cleared, so done. */
1117 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1120 rnp
->qsmask
&= ~mask
;
1121 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1122 mask
, rnp
->qsmask
, rnp
->level
,
1123 rnp
->grplo
, rnp
->grphi
,
1125 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1127 /* Other bits still set at this level, so done. */
1128 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1131 mask
= rnp
->grpmask
;
1132 if (rnp
->parent
== NULL
) {
1134 /* No more levels. Exit loop holding root lock. */
1138 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1141 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1142 WARN_ON_ONCE(rnp_c
->qsmask
);
1146 * Get here if we are the last CPU to pass through a quiescent
1147 * state for this grace period. Invoke rcu_report_qs_rsp()
1148 * to clean up and start the next grace period if one is needed.
1150 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1154 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1155 * structure. This must be either called from the specified CPU, or
1156 * called when the specified CPU is known to be offline (and when it is
1157 * also known that no other CPU is concurrently trying to help the offline
1158 * CPU). The lastcomp argument is used to make sure we are still in the
1159 * grace period of interest. We don't want to end the current grace period
1160 * based on quiescent states detected in an earlier grace period!
1163 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastgp
)
1165 unsigned long flags
;
1167 struct rcu_node
*rnp
;
1170 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1171 if (lastgp
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
) {
1174 * The grace period in which this quiescent state was
1175 * recorded has ended, so don't report it upwards.
1176 * We will instead need a new quiescent state that lies
1177 * within the current grace period.
1179 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1180 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1183 mask
= rdp
->grpmask
;
1184 if ((rnp
->qsmask
& mask
) == 0) {
1185 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1187 rdp
->qs_pending
= 0;
1190 * This GP can't end until cpu checks in, so all of our
1191 * callbacks can be processed during the next GP.
1193 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1195 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1200 * Check to see if there is a new grace period of which this CPU
1201 * is not yet aware, and if so, set up local rcu_data state for it.
1202 * Otherwise, see if this CPU has just passed through its first
1203 * quiescent state for this grace period, and record that fact if so.
1206 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1208 /* If there is now a new grace period, record and return. */
1209 if (check_for_new_grace_period(rsp
, rdp
))
1213 * Does this CPU still need to do its part for current grace period?
1214 * If no, return and let the other CPUs do their part as well.
1216 if (!rdp
->qs_pending
)
1220 * Was there a quiescent state since the beginning of the grace
1221 * period? If no, then exit and wait for the next call.
1223 if (!rdp
->passed_quiesce
)
1227 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1230 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesce_gpnum
);
1233 #ifdef CONFIG_HOTPLUG_CPU
1236 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1237 * Also record a quiescent state for this CPU for the current grace period.
1238 * Synchronization and interrupt disabling are not required because
1239 * this function executes in stop_machine() context. Therefore, cleanup
1240 * operations that might block must be done later from the CPU_DEAD
1243 * Note that the outgoing CPU's bit has already been cleared in the
1244 * cpu_online_mask. This allows us to randomly pick a callback
1245 * destination from the bits set in that mask.
1247 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1249 unsigned long flags
;
1253 int receive_cpu
= cpumask_any(cpu_online_mask
);
1254 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1255 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1256 struct rcu_node
*rnp
= rdp
->mynode
; /* For dying CPU. */
1258 /* First, adjust the counts. */
1259 if (rdp
->nxtlist
!= NULL
) {
1260 receive_rdp
->qlen_lazy
+= rdp
->qlen_lazy
;
1261 receive_rdp
->qlen
+= rdp
->qlen
;
1267 * Next, move ready-to-invoke callbacks to be invoked on some
1268 * other CPU. These will not be required to pass through another
1269 * grace period: They are done, regardless of CPU.
1271 if (rdp
->nxtlist
!= NULL
&&
1272 rdp
->nxttail
[RCU_DONE_TAIL
] != &rdp
->nxtlist
) {
1273 struct rcu_head
*oldhead
;
1274 struct rcu_head
**oldtail
;
1275 struct rcu_head
**newtail
;
1277 oldhead
= rdp
->nxtlist
;
1278 oldtail
= receive_rdp
->nxttail
[RCU_DONE_TAIL
];
1279 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1280 *rdp
->nxttail
[RCU_DONE_TAIL
] = *oldtail
;
1281 *receive_rdp
->nxttail
[RCU_DONE_TAIL
] = oldhead
;
1282 newtail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1283 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_SIZE
; i
++) {
1284 if (receive_rdp
->nxttail
[i
] == oldtail
)
1285 receive_rdp
->nxttail
[i
] = newtail
;
1286 if (rdp
->nxttail
[i
] == newtail
)
1287 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1292 * Finally, put the rest of the callbacks at the end of the list.
1293 * The ones that made it partway through get to start over: We
1294 * cannot assume that grace periods are synchronized across CPUs.
1295 * (We could splice RCU_WAIT_TAIL into RCU_NEXT_READY_TAIL, but
1296 * this does not seem compelling. Not yet, anyway.)
1298 if (rdp
->nxtlist
!= NULL
) {
1299 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1300 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] =
1301 rdp
->nxttail
[RCU_NEXT_TAIL
];
1302 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1303 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1305 rdp
->nxtlist
= NULL
;
1306 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1307 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1311 * Record a quiescent state for the dying CPU. This is safe
1312 * only because we have already cleared out the callbacks.
1313 * (Otherwise, the RCU core might try to schedule the invocation
1314 * of callbacks on this now-offline CPU, which would be bad.)
1316 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1317 trace_rcu_grace_period(rsp
->name
,
1318 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1320 rcu_report_qs_rdp(smp_processor_id(), rsp
, rdp
, rsp
->gpnum
);
1321 /* Note that rcu_report_qs_rdp() might call trace_rcu_grace_period(). */
1324 * Remove the dying CPU from the bitmasks in the rcu_node
1325 * hierarchy. Because we are in stop_machine() context, we
1326 * automatically exclude ->onofflock critical sections.
1329 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1330 rnp
->qsmaskinit
&= ~mask
;
1331 if (rnp
->qsmaskinit
!= 0) {
1332 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1335 if (rnp
== rdp
->mynode
) {
1336 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1337 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1338 rcu_report_unblock_qs_rnp(rnp
, flags
);
1340 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1341 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1342 rcu_report_exp_rnp(rsp
, rnp
, true);
1344 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1345 mask
= rnp
->grpmask
;
1347 } while (rnp
!= NULL
);
1351 * The CPU has been completely removed, and some other CPU is reporting
1352 * this fact from process context. Do the remainder of the cleanup.
1353 * There can only be one CPU hotplug operation at a time, so no other
1354 * CPU can be attempting to update rcu_cpu_kthread_task.
1356 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1358 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1359 struct rcu_node
*rnp
= rdp
->mynode
;
1361 rcu_stop_cpu_kthread(cpu
);
1362 rcu_node_kthread_setaffinity(rnp
, -1);
1365 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1367 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1371 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1375 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1378 * Invoke any RCU callbacks that have made it to the end of their grace
1379 * period. Thottle as specified by rdp->blimit.
1381 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1383 unsigned long flags
;
1384 struct rcu_head
*next
, *list
, **tail
;
1385 int bl
, count
, count_lazy
;
1387 /* If no callbacks are ready, just return.*/
1388 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1389 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1390 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1391 need_resched(), is_idle_task(current
),
1392 rcu_is_callbacks_kthread());
1397 * Extract the list of ready callbacks, disabling to prevent
1398 * races with call_rcu() from interrupt handlers.
1400 local_irq_save(flags
);
1401 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1403 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1404 list
= rdp
->nxtlist
;
1405 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1406 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1407 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1408 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1409 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1410 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1411 local_irq_restore(flags
);
1413 /* Invoke callbacks. */
1414 count
= count_lazy
= 0;
1418 debug_rcu_head_unqueue(list
);
1419 if (__rcu_reclaim(rsp
->name
, list
))
1422 /* Stop only if limit reached and CPU has something to do. */
1423 if (++count
>= bl
&&
1425 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1429 local_irq_save(flags
);
1430 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1431 is_idle_task(current
),
1432 rcu_is_callbacks_kthread());
1434 /* Update count, and requeue any remaining callbacks. */
1435 rdp
->qlen_lazy
-= count_lazy
;
1437 rdp
->n_cbs_invoked
+= count
;
1439 *tail
= rdp
->nxtlist
;
1440 rdp
->nxtlist
= list
;
1441 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1442 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1443 rdp
->nxttail
[count
] = tail
;
1448 /* Reinstate batch limit if we have worked down the excess. */
1449 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1450 rdp
->blimit
= blimit
;
1452 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1453 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1454 rdp
->qlen_last_fqs_check
= 0;
1455 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1456 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1457 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1459 local_irq_restore(flags
);
1461 /* Re-invoke RCU core processing if there are callbacks remaining. */
1462 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1467 * Check to see if this CPU is in a non-context-switch quiescent state
1468 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1469 * Also schedule RCU core processing.
1471 * This function must be called from hardirq context. It is normally
1472 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1473 * false, there is no point in invoking rcu_check_callbacks().
1475 void rcu_check_callbacks(int cpu
, int user
)
1477 trace_rcu_utilization("Start scheduler-tick");
1478 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1481 * Get here if this CPU took its interrupt from user
1482 * mode or from the idle loop, and if this is not a
1483 * nested interrupt. In this case, the CPU is in
1484 * a quiescent state, so note it.
1486 * No memory barrier is required here because both
1487 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1488 * variables that other CPUs neither access nor modify,
1489 * at least not while the corresponding CPU is online.
1495 } else if (!in_softirq()) {
1498 * Get here if this CPU did not take its interrupt from
1499 * softirq, in other words, if it is not interrupting
1500 * a rcu_bh read-side critical section. This is an _bh
1501 * critical section, so note it.
1506 rcu_preempt_check_callbacks(cpu
);
1507 if (rcu_pending(cpu
))
1509 trace_rcu_utilization("End scheduler-tick");
1513 * Scan the leaf rcu_node structures, processing dyntick state for any that
1514 * have not yet encountered a quiescent state, using the function specified.
1515 * Also initiate boosting for any threads blocked on the root rcu_node.
1517 * The caller must have suppressed start of new grace periods.
1519 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1523 unsigned long flags
;
1525 struct rcu_node
*rnp
;
1527 rcu_for_each_leaf_node(rsp
, rnp
) {
1529 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1530 if (!rcu_gp_in_progress(rsp
)) {
1531 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1534 if (rnp
->qsmask
== 0) {
1535 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1540 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1541 if ((rnp
->qsmask
& bit
) != 0 &&
1542 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1547 /* rcu_report_qs_rnp() releases rnp->lock. */
1548 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1551 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1553 rnp
= rcu_get_root(rsp
);
1554 if (rnp
->qsmask
== 0) {
1555 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1556 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1561 * Force quiescent states on reluctant CPUs, and also detect which
1562 * CPUs are in dyntick-idle mode.
1564 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1566 unsigned long flags
;
1567 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1569 trace_rcu_utilization("Start fqs");
1570 if (!rcu_gp_in_progress(rsp
)) {
1571 trace_rcu_utilization("End fqs");
1572 return; /* No grace period in progress, nothing to force. */
1574 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1575 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1576 trace_rcu_utilization("End fqs");
1577 return; /* Someone else is already on the job. */
1579 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1580 goto unlock_fqs_ret
; /* no emergency and done recently. */
1582 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1583 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1584 if(!rcu_gp_in_progress(rsp
)) {
1585 rsp
->n_force_qs_ngp
++;
1586 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1587 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1589 rsp
->fqs_active
= 1;
1590 switch (rsp
->fqs_state
) {
1594 break; /* grace period idle or initializing, ignore. */
1596 case RCU_SAVE_DYNTICK
:
1597 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1598 break; /* So gcc recognizes the dead code. */
1600 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1602 /* Record dyntick-idle state. */
1603 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1604 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1605 if (rcu_gp_in_progress(rsp
))
1606 rsp
->fqs_state
= RCU_FORCE_QS
;
1611 /* Check dyntick-idle state, send IPI to laggarts. */
1612 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1613 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1615 /* Leave state in case more forcing is required. */
1617 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1620 rsp
->fqs_active
= 0;
1621 if (rsp
->fqs_need_gp
) {
1622 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1623 rsp
->fqs_need_gp
= 0;
1624 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1625 trace_rcu_utilization("End fqs");
1628 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1630 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1631 trace_rcu_utilization("End fqs");
1635 * This does the RCU core processing work for the specified rcu_state
1636 * and rcu_data structures. This may be called only from the CPU to
1637 * whom the rdp belongs.
1640 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1642 unsigned long flags
;
1644 WARN_ON_ONCE(rdp
->beenonline
== 0);
1647 * If an RCU GP has gone long enough, go check for dyntick
1648 * idle CPUs and, if needed, send resched IPIs.
1650 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1651 force_quiescent_state(rsp
, 1);
1654 * Advance callbacks in response to end of earlier grace
1655 * period that some other CPU ended.
1657 rcu_process_gp_end(rsp
, rdp
);
1659 /* Update RCU state based on any recent quiescent states. */
1660 rcu_check_quiescent_state(rsp
, rdp
);
1662 /* Does this CPU require a not-yet-started grace period? */
1663 if (cpu_needs_another_gp(rsp
, rdp
)) {
1664 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1665 rcu_start_gp(rsp
, flags
); /* releases above lock */
1668 /* If there are callbacks ready, invoke them. */
1669 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1670 invoke_rcu_callbacks(rsp
, rdp
);
1674 * Do RCU core processing for the current CPU.
1676 static void rcu_process_callbacks(struct softirq_action
*unused
)
1678 trace_rcu_utilization("Start RCU core");
1679 __rcu_process_callbacks(&rcu_sched_state
,
1680 &__get_cpu_var(rcu_sched_data
));
1681 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1682 rcu_preempt_process_callbacks();
1683 trace_rcu_utilization("End RCU core");
1687 * Schedule RCU callback invocation. If the specified type of RCU
1688 * does not support RCU priority boosting, just do a direct call,
1689 * otherwise wake up the per-CPU kernel kthread. Note that because we
1690 * are running on the current CPU with interrupts disabled, the
1691 * rcu_cpu_kthread_task cannot disappear out from under us.
1693 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1695 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
1697 if (likely(!rsp
->boost
)) {
1698 rcu_do_batch(rsp
, rdp
);
1701 invoke_rcu_callbacks_kthread();
1704 static void invoke_rcu_core(void)
1706 raise_softirq(RCU_SOFTIRQ
);
1710 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1711 struct rcu_state
*rsp
, bool lazy
)
1713 unsigned long flags
;
1714 struct rcu_data
*rdp
;
1716 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
1717 debug_rcu_head_queue(head
);
1721 smp_mb(); /* Ensure RCU update seen before callback registry. */
1724 * Opportunistically note grace-period endings and beginnings.
1725 * Note that we might see a beginning right after we see an
1726 * end, but never vice versa, since this CPU has to pass through
1727 * a quiescent state betweentimes.
1729 local_irq_save(flags
);
1730 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1731 rdp
= this_cpu_ptr(rsp
->rda
);
1733 /* Add the callback to our list. */
1734 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1735 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1740 if (__is_kfree_rcu_offset((unsigned long)func
))
1741 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
1742 rdp
->qlen_lazy
, rdp
->qlen
);
1744 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
1746 /* If interrupts were disabled, don't dive into RCU core. */
1747 if (irqs_disabled_flags(flags
)) {
1748 local_irq_restore(flags
);
1753 * Force the grace period if too many callbacks or too long waiting.
1754 * Enforce hysteresis, and don't invoke force_quiescent_state()
1755 * if some other CPU has recently done so. Also, don't bother
1756 * invoking force_quiescent_state() if the newly enqueued callback
1757 * is the only one waiting for a grace period to complete.
1759 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1761 /* Are we ignoring a completed grace period? */
1762 rcu_process_gp_end(rsp
, rdp
);
1763 check_for_new_grace_period(rsp
, rdp
);
1765 /* Start a new grace period if one not already started. */
1766 if (!rcu_gp_in_progress(rsp
)) {
1767 unsigned long nestflag
;
1768 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1770 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1771 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1773 /* Give the grace period a kick. */
1774 rdp
->blimit
= LONG_MAX
;
1775 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1776 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1777 force_quiescent_state(rsp
, 0);
1778 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1779 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1781 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1782 force_quiescent_state(rsp
, 1);
1783 local_irq_restore(flags
);
1787 * Queue an RCU-sched callback for invocation after a grace period.
1789 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1791 __call_rcu(head
, func
, &rcu_sched_state
, 0);
1793 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1796 * Queue an RCU callback for invocation after a quicker grace period.
1798 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1800 __call_rcu(head
, func
, &rcu_bh_state
, 0);
1802 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1805 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1807 * Control will return to the caller some time after a full rcu-sched
1808 * grace period has elapsed, in other words after all currently executing
1809 * rcu-sched read-side critical sections have completed. These read-side
1810 * critical sections are delimited by rcu_read_lock_sched() and
1811 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1812 * local_irq_disable(), and so on may be used in place of
1813 * rcu_read_lock_sched().
1815 * This means that all preempt_disable code sequences, including NMI and
1816 * hardware-interrupt handlers, in progress on entry will have completed
1817 * before this primitive returns. However, this does not guarantee that
1818 * softirq handlers will have completed, since in some kernels, these
1819 * handlers can run in process context, and can block.
1821 * This primitive provides the guarantees made by the (now removed)
1822 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1823 * guarantees that rcu_read_lock() sections will have completed.
1824 * In "classic RCU", these two guarantees happen to be one and
1825 * the same, but can differ in realtime RCU implementations.
1827 void synchronize_sched(void)
1829 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
1830 !lock_is_held(&rcu_lock_map
) &&
1831 !lock_is_held(&rcu_sched_lock_map
),
1832 "Illegal synchronize_sched() in RCU-sched read-side critical section");
1833 if (rcu_blocking_is_gp())
1835 wait_rcu_gp(call_rcu_sched
);
1837 EXPORT_SYMBOL_GPL(synchronize_sched
);
1840 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1842 * Control will return to the caller some time after a full rcu_bh grace
1843 * period has elapsed, in other words after all currently executing rcu_bh
1844 * read-side critical sections have completed. RCU read-side critical
1845 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1846 * and may be nested.
1848 void synchronize_rcu_bh(void)
1850 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
1851 !lock_is_held(&rcu_lock_map
) &&
1852 !lock_is_held(&rcu_sched_lock_map
),
1853 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
1854 if (rcu_blocking_is_gp())
1856 wait_rcu_gp(call_rcu_bh
);
1858 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1861 * Check to see if there is any immediate RCU-related work to be done
1862 * by the current CPU, for the specified type of RCU, returning 1 if so.
1863 * The checks are in order of increasing expense: checks that can be
1864 * carried out against CPU-local state are performed first. However,
1865 * we must check for CPU stalls first, else we might not get a chance.
1867 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1869 struct rcu_node
*rnp
= rdp
->mynode
;
1871 rdp
->n_rcu_pending
++;
1873 /* Check for CPU stalls, if enabled. */
1874 check_cpu_stall(rsp
, rdp
);
1876 /* Is the RCU core waiting for a quiescent state from this CPU? */
1877 if (rcu_scheduler_fully_active
&&
1878 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
1881 * If force_quiescent_state() coming soon and this CPU
1882 * needs a quiescent state, and this is either RCU-sched
1883 * or RCU-bh, force a local reschedule.
1885 rdp
->n_rp_qs_pending
++;
1886 if (!rdp
->preemptible
&&
1887 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
1890 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
1891 rdp
->n_rp_report_qs
++;
1895 /* Does this CPU have callbacks ready to invoke? */
1896 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1897 rdp
->n_rp_cb_ready
++;
1901 /* Has RCU gone idle with this CPU needing another grace period? */
1902 if (cpu_needs_another_gp(rsp
, rdp
)) {
1903 rdp
->n_rp_cpu_needs_gp
++;
1907 /* Has another RCU grace period completed? */
1908 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
1909 rdp
->n_rp_gp_completed
++;
1913 /* Has a new RCU grace period started? */
1914 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
1915 rdp
->n_rp_gp_started
++;
1919 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1920 if (rcu_gp_in_progress(rsp
) &&
1921 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
1922 rdp
->n_rp_need_fqs
++;
1927 rdp
->n_rp_need_nothing
++;
1932 * Check to see if there is any immediate RCU-related work to be done
1933 * by the current CPU, returning 1 if so. This function is part of the
1934 * RCU implementation; it is -not- an exported member of the RCU API.
1936 static int rcu_pending(int cpu
)
1938 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
1939 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
1940 rcu_preempt_pending(cpu
);
1944 * Check to see if any future RCU-related work will need to be done
1945 * by the current CPU, even if none need be done immediately, returning
1948 static int rcu_cpu_has_callbacks(int cpu
)
1950 /* RCU callbacks either ready or pending? */
1951 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
1952 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
1953 rcu_preempt_cpu_has_callbacks(cpu
);
1956 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
1957 static atomic_t rcu_barrier_cpu_count
;
1958 static DEFINE_MUTEX(rcu_barrier_mutex
);
1959 static struct completion rcu_barrier_completion
;
1961 static void rcu_barrier_callback(struct rcu_head
*notused
)
1963 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1964 complete(&rcu_barrier_completion
);
1968 * Called with preemption disabled, and from cross-cpu IRQ context.
1970 static void rcu_barrier_func(void *type
)
1972 int cpu
= smp_processor_id();
1973 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
1974 void (*call_rcu_func
)(struct rcu_head
*head
,
1975 void (*func
)(struct rcu_head
*head
));
1977 atomic_inc(&rcu_barrier_cpu_count
);
1978 call_rcu_func
= type
;
1979 call_rcu_func(head
, rcu_barrier_callback
);
1983 * Orchestrate the specified type of RCU barrier, waiting for all
1984 * RCU callbacks of the specified type to complete.
1986 static void _rcu_barrier(struct rcu_state
*rsp
,
1987 void (*call_rcu_func
)(struct rcu_head
*head
,
1988 void (*func
)(struct rcu_head
*head
)))
1990 BUG_ON(in_interrupt());
1991 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1992 mutex_lock(&rcu_barrier_mutex
);
1993 init_completion(&rcu_barrier_completion
);
1995 * Initialize rcu_barrier_cpu_count to 1, then invoke
1996 * rcu_barrier_func() on each CPU, so that each CPU also has
1997 * incremented rcu_barrier_cpu_count. Only then is it safe to
1998 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1999 * might complete its grace period before all of the other CPUs
2000 * did their increment, causing this function to return too
2001 * early. Note that on_each_cpu() disables irqs, which prevents
2002 * any CPUs from coming online or going offline until each online
2003 * CPU has queued its RCU-barrier callback.
2005 atomic_set(&rcu_barrier_cpu_count
, 1);
2006 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
2007 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2008 complete(&rcu_barrier_completion
);
2009 wait_for_completion(&rcu_barrier_completion
);
2010 mutex_unlock(&rcu_barrier_mutex
);
2014 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2016 void rcu_barrier_bh(void)
2018 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
2020 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2023 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2025 void rcu_barrier_sched(void)
2027 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2029 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2032 * Do boot-time initialization of a CPU's per-CPU RCU data.
2035 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2037 unsigned long flags
;
2039 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2040 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2042 /* Set up local state, ensuring consistent view of global state. */
2043 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2044 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2045 rdp
->nxtlist
= NULL
;
2046 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2047 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2050 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2051 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_NESTING
);
2052 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2055 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2059 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2060 * offline event can be happening at a given time. Note also that we
2061 * can accept some slop in the rsp->completed access due to the fact
2062 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2064 static void __cpuinit
2065 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2067 unsigned long flags
;
2069 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2070 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2072 /* Set up local state, ensuring consistent view of global state. */
2073 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2074 rdp
->beenonline
= 1; /* We have now been online. */
2075 rdp
->preemptible
= preemptible
;
2076 rdp
->qlen_last_fqs_check
= 0;
2077 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2078 rdp
->blimit
= blimit
;
2079 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_NESTING
;
2080 atomic_set(&rdp
->dynticks
->dynticks
,
2081 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2082 rcu_prepare_for_idle_init(cpu
);
2083 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2086 * A new grace period might start here. If so, we won't be part
2087 * of it, but that is OK, as we are currently in a quiescent state.
2090 /* Exclude any attempts to start a new GP on large systems. */
2091 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2093 /* Add CPU to rcu_node bitmasks. */
2095 mask
= rdp
->grpmask
;
2097 /* Exclude any attempts to start a new GP on small systems. */
2098 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2099 rnp
->qsmaskinit
|= mask
;
2100 mask
= rnp
->grpmask
;
2101 if (rnp
== rdp
->mynode
) {
2103 * If there is a grace period in progress, we will
2104 * set up to wait for it next time we run the
2107 rdp
->gpnum
= rnp
->completed
;
2108 rdp
->completed
= rnp
->completed
;
2109 rdp
->passed_quiesce
= 0;
2110 rdp
->qs_pending
= 0;
2111 rdp
->passed_quiesce_gpnum
= rnp
->gpnum
- 1;
2112 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2114 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2116 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2118 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2121 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2123 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2124 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2125 rcu_preempt_init_percpu_data(cpu
);
2129 * Handle CPU online/offline notification events.
2131 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2132 unsigned long action
, void *hcpu
)
2134 long cpu
= (long)hcpu
;
2135 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2136 struct rcu_node
*rnp
= rdp
->mynode
;
2138 trace_rcu_utilization("Start CPU hotplug");
2140 case CPU_UP_PREPARE
:
2141 case CPU_UP_PREPARE_FROZEN
:
2142 rcu_prepare_cpu(cpu
);
2143 rcu_prepare_kthreads(cpu
);
2146 case CPU_DOWN_FAILED
:
2147 rcu_node_kthread_setaffinity(rnp
, -1);
2148 rcu_cpu_kthread_setrt(cpu
, 1);
2150 case CPU_DOWN_PREPARE
:
2151 rcu_node_kthread_setaffinity(rnp
, cpu
);
2152 rcu_cpu_kthread_setrt(cpu
, 0);
2155 case CPU_DYING_FROZEN
:
2157 * The whole machine is "stopped" except this CPU, so we can
2158 * touch any data without introducing corruption. We send the
2159 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2161 rcu_cleanup_dying_cpu(&rcu_bh_state
);
2162 rcu_cleanup_dying_cpu(&rcu_sched_state
);
2163 rcu_preempt_cleanup_dying_cpu();
2164 rcu_cleanup_after_idle(cpu
);
2167 case CPU_DEAD_FROZEN
:
2168 case CPU_UP_CANCELED
:
2169 case CPU_UP_CANCELED_FROZEN
:
2170 rcu_cleanup_dead_cpu(cpu
, &rcu_bh_state
);
2171 rcu_cleanup_dead_cpu(cpu
, &rcu_sched_state
);
2172 rcu_preempt_cleanup_dead_cpu(cpu
);
2177 trace_rcu_utilization("End CPU hotplug");
2182 * This function is invoked towards the end of the scheduler's initialization
2183 * process. Before this is called, the idle task might contain
2184 * RCU read-side critical sections (during which time, this idle
2185 * task is booting the system). After this function is called, the
2186 * idle tasks are prohibited from containing RCU read-side critical
2187 * sections. This function also enables RCU lockdep checking.
2189 void rcu_scheduler_starting(void)
2191 WARN_ON(num_online_cpus() != 1);
2192 WARN_ON(nr_context_switches() > 0);
2193 rcu_scheduler_active
= 1;
2197 * Compute the per-level fanout, either using the exact fanout specified
2198 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2200 #ifdef CONFIG_RCU_FANOUT_EXACT
2201 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2205 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2206 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2207 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
2209 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2210 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2217 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2218 ccur
= rsp
->levelcnt
[i
];
2219 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2223 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2226 * Helper function for rcu_init() that initializes one rcu_state structure.
2228 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2229 struct rcu_data __percpu
*rda
)
2231 static char *buf
[] = { "rcu_node_level_0",
2234 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2238 struct rcu_node
*rnp
;
2240 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2242 /* Initialize the level-tracking arrays. */
2244 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2245 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2246 rcu_init_levelspread(rsp
);
2248 /* Initialize the elements themselves, starting from the leaves. */
2250 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2251 cpustride
*= rsp
->levelspread
[i
];
2252 rnp
= rsp
->level
[i
];
2253 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2254 raw_spin_lock_init(&rnp
->lock
);
2255 lockdep_set_class_and_name(&rnp
->lock
,
2256 &rcu_node_class
[i
], buf
[i
]);
2259 rnp
->qsmaskinit
= 0;
2260 rnp
->grplo
= j
* cpustride
;
2261 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2262 if (rnp
->grphi
>= NR_CPUS
)
2263 rnp
->grphi
= NR_CPUS
- 1;
2269 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2270 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2271 rnp
->parent
= rsp
->level
[i
- 1] +
2272 j
/ rsp
->levelspread
[i
- 1];
2275 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2280 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2281 for_each_possible_cpu(i
) {
2282 while (i
> rnp
->grphi
)
2284 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2285 rcu_boot_init_percpu_data(i
, rsp
);
2289 void __init
rcu_init(void)
2293 rcu_bootup_announce();
2294 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2295 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2296 __rcu_init_preempt();
2297 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2300 * We don't need protection against CPU-hotplug here because
2301 * this is called early in boot, before either interrupts
2302 * or the scheduler are operational.
2304 cpu_notifier(rcu_cpu_notify
, 0);
2305 for_each_online_cpu(cpu
)
2306 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2307 check_cpu_stall_init();
2310 #include "rcutree_plugin.h"