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_user_enter_after_irq - inform RCU that we are going to resume userspace
430 * after the current irq returns.
432 * This is similar to rcu_user_enter() but in the context of a non-nesting
433 * irq. After this call, RCU enters into idle mode when the interrupt
436 void rcu_user_enter_after_irq(void)
439 struct rcu_dynticks
*rdtp
;
441 local_irq_save(flags
);
442 rdtp
= &__get_cpu_var(rcu_dynticks
);
443 /* Ensure this irq is interrupting a non-idle RCU state. */
444 WARN_ON_ONCE(!(rdtp
->dynticks_nesting
& DYNTICK_TASK_MASK
));
445 rdtp
->dynticks_nesting
= 1;
446 local_irq_restore(flags
);
450 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
452 * Exit from an interrupt handler, which might possibly result in entering
453 * idle mode, in other words, leaving the mode in which read-side critical
454 * sections can occur.
456 * This code assumes that the idle loop never does anything that might
457 * result in unbalanced calls to irq_enter() and irq_exit(). If your
458 * architecture violates this assumption, RCU will give you what you
459 * deserve, good and hard. But very infrequently and irreproducibly.
461 * Use things like work queues to work around this limitation.
463 * You have been warned.
465 void rcu_irq_exit(void)
469 struct rcu_dynticks
*rdtp
;
471 local_irq_save(flags
);
472 rdtp
= &__get_cpu_var(rcu_dynticks
);
473 oldval
= rdtp
->dynticks_nesting
;
474 rdtp
->dynticks_nesting
--;
475 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
476 if (rdtp
->dynticks_nesting
)
477 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
479 rcu_eqs_enter_common(rdtp
, oldval
, 1);
480 local_irq_restore(flags
);
484 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
486 * If the new value of the ->dynticks_nesting counter was previously zero,
487 * we really have exited idle, and must do the appropriate accounting.
488 * The caller must have disabled interrupts.
490 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
493 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
494 atomic_inc(&rdtp
->dynticks
);
495 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
496 smp_mb__after_atomic_inc(); /* See above. */
497 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
498 rcu_cleanup_after_idle(smp_processor_id());
499 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
500 if (!is_idle_task(current
) && !user
) {
501 struct task_struct
*idle
= idle_task(smp_processor_id());
503 trace_rcu_dyntick("Error on exit: not idle task",
504 oldval
, rdtp
->dynticks_nesting
);
505 ftrace_dump(DUMP_ORIG
);
506 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
507 current
->pid
, current
->comm
,
508 idle
->pid
, idle
->comm
); /* must be idle task! */
513 * Exit an RCU extended quiescent state, which can be either the
514 * idle loop or adaptive-tickless usermode execution.
516 static void rcu_eqs_exit(bool user
)
519 struct rcu_dynticks
*rdtp
;
522 local_irq_save(flags
);
523 rdtp
= &__get_cpu_var(rcu_dynticks
);
524 oldval
= rdtp
->dynticks_nesting
;
525 WARN_ON_ONCE(oldval
< 0);
526 if (oldval
& DYNTICK_TASK_NEST_MASK
)
527 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
529 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
530 rcu_eqs_exit_common(rdtp
, oldval
, user
);
531 local_irq_restore(flags
);
535 * rcu_idle_exit - inform RCU that current CPU is leaving idle
537 * Exit idle mode, in other words, -enter- the mode in which RCU
538 * read-side critical sections can occur.
540 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
541 * allow for the possibility of usermode upcalls messing up our count
542 * of interrupt nesting level during the busy period that is just
545 void rcu_idle_exit(void)
549 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
552 * rcu_user_exit - inform RCU that we are exiting userspace.
554 * Exit RCU idle mode while entering the kernel because it can
555 * run a RCU read side critical section anytime.
557 void rcu_user_exit(void)
560 * Some contexts may involve an exception occuring in an irq,
561 * leading to that nesting:
562 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
563 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
564 * helpers are enough to protect RCU uses inside the exception. So
565 * just return immediately if we detect we are in an IRQ.
574 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
575 * idle mode after the current non-nesting irq returns.
577 * This is similar to rcu_user_exit() but in the context of an irq.
578 * This is called when the irq has interrupted a userspace RCU idle mode
579 * context. When the current non-nesting interrupt returns after this call,
580 * the CPU won't restore the RCU idle mode.
582 void rcu_user_exit_after_irq(void)
585 struct rcu_dynticks
*rdtp
;
587 local_irq_save(flags
);
588 rdtp
= &__get_cpu_var(rcu_dynticks
);
589 /* Ensure we are interrupting an RCU idle mode. */
590 WARN_ON_ONCE(rdtp
->dynticks_nesting
& DYNTICK_TASK_NEST_MASK
);
591 rdtp
->dynticks_nesting
+= DYNTICK_TASK_EXIT_IDLE
;
592 local_irq_restore(flags
);
596 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
598 * Enter an interrupt handler, which might possibly result in exiting
599 * idle mode, in other words, entering the mode in which read-side critical
600 * sections can occur.
602 * Note that the Linux kernel is fully capable of entering an interrupt
603 * handler that it never exits, for example when doing upcalls to
604 * user mode! This code assumes that the idle loop never does upcalls to
605 * user mode. If your architecture does do upcalls from the idle loop (or
606 * does anything else that results in unbalanced calls to the irq_enter()
607 * and irq_exit() functions), RCU will give you what you deserve, good
608 * and hard. But very infrequently and irreproducibly.
610 * Use things like work queues to work around this limitation.
612 * You have been warned.
614 void rcu_irq_enter(void)
617 struct rcu_dynticks
*rdtp
;
620 local_irq_save(flags
);
621 rdtp
= &__get_cpu_var(rcu_dynticks
);
622 oldval
= rdtp
->dynticks_nesting
;
623 rdtp
->dynticks_nesting
++;
624 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
626 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
628 rcu_eqs_exit_common(rdtp
, oldval
, 1);
629 local_irq_restore(flags
);
633 * rcu_nmi_enter - inform RCU of entry to NMI context
635 * If the CPU was idle with dynamic ticks active, and there is no
636 * irq handler running, this updates rdtp->dynticks_nmi to let the
637 * RCU grace-period handling know that the CPU is active.
639 void rcu_nmi_enter(void)
641 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
643 if (rdtp
->dynticks_nmi_nesting
== 0 &&
644 (atomic_read(&rdtp
->dynticks
) & 0x1))
646 rdtp
->dynticks_nmi_nesting
++;
647 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
648 atomic_inc(&rdtp
->dynticks
);
649 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
650 smp_mb__after_atomic_inc(); /* See above. */
651 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
655 * rcu_nmi_exit - inform RCU of exit from NMI context
657 * If the CPU was idle with dynamic ticks active, and there is no
658 * irq handler running, this updates rdtp->dynticks_nmi to let the
659 * RCU grace-period handling know that the CPU is no longer active.
661 void rcu_nmi_exit(void)
663 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
665 if (rdtp
->dynticks_nmi_nesting
== 0 ||
666 --rdtp
->dynticks_nmi_nesting
!= 0)
668 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
669 smp_mb__before_atomic_inc(); /* See above. */
670 atomic_inc(&rdtp
->dynticks
);
671 smp_mb__after_atomic_inc(); /* Force delay to next write. */
672 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
676 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
678 * If the current CPU is in its idle loop and is neither in an interrupt
679 * or NMI handler, return true.
681 int rcu_is_cpu_idle(void)
686 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
690 EXPORT_SYMBOL(rcu_is_cpu_idle
);
692 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
695 * Is the current CPU online? Disable preemption to avoid false positives
696 * that could otherwise happen due to the current CPU number being sampled,
697 * this task being preempted, its old CPU being taken offline, resuming
698 * on some other CPU, then determining that its old CPU is now offline.
699 * It is OK to use RCU on an offline processor during initial boot, hence
700 * the check for rcu_scheduler_fully_active. Note also that it is OK
701 * for a CPU coming online to use RCU for one jiffy prior to marking itself
702 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
703 * offline to continue to use RCU for one jiffy after marking itself
704 * offline in the cpu_online_mask. This leniency is necessary given the
705 * non-atomic nature of the online and offline processing, for example,
706 * the fact that a CPU enters the scheduler after completing the CPU_DYING
709 * This is also why RCU internally marks CPUs online during the
710 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
712 * Disable checking if in an NMI handler because we cannot safely report
713 * errors from NMI handlers anyway.
715 bool rcu_lockdep_current_cpu_online(void)
717 struct rcu_data
*rdp
;
718 struct rcu_node
*rnp
;
724 rdp
= &__get_cpu_var(rcu_sched_data
);
726 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
727 !rcu_scheduler_fully_active
;
731 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
733 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
736 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
738 * If the current CPU is idle or running at a first-level (not nested)
739 * interrupt from idle, return true. The caller must have at least
740 * disabled preemption.
742 int rcu_is_cpu_rrupt_from_idle(void)
744 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
748 * Snapshot the specified CPU's dynticks counter so that we can later
749 * credit them with an implicit quiescent state. Return 1 if this CPU
750 * is in dynticks idle mode, which is an extended quiescent state.
752 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
754 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
755 return (rdp
->dynticks_snap
& 0x1) == 0;
759 * Return true if the specified CPU has passed through a quiescent
760 * state by virtue of being in or having passed through an dynticks
761 * idle state since the last call to dyntick_save_progress_counter()
762 * for this same CPU, or by virtue of having been offline.
764 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
769 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
770 snap
= (unsigned int)rdp
->dynticks_snap
;
773 * If the CPU passed through or entered a dynticks idle phase with
774 * no active irq/NMI handlers, then we can safely pretend that the CPU
775 * already acknowledged the request to pass through a quiescent
776 * state. Either way, that CPU cannot possibly be in an RCU
777 * read-side critical section that started before the beginning
778 * of the current RCU grace period.
780 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
781 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
787 * Check for the CPU being offline, but only if the grace period
788 * is old enough. We don't need to worry about the CPU changing
789 * state: If we see it offline even once, it has been through a
792 * The reason for insisting that the grace period be at least
793 * one jiffy old is that CPUs that are not quite online and that
794 * have just gone offline can still execute RCU read-side critical
797 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
798 return 0; /* Grace period is not old enough. */
800 if (cpu_is_offline(rdp
->cpu
)) {
801 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
808 static int jiffies_till_stall_check(void)
810 int till_stall_check
= ACCESS_ONCE(rcu_cpu_stall_timeout
);
813 * Limit check must be consistent with the Kconfig limits
814 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
816 if (till_stall_check
< 3) {
817 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 3;
818 till_stall_check
= 3;
819 } else if (till_stall_check
> 300) {
820 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 300;
821 till_stall_check
= 300;
823 return till_stall_check
* HZ
+ RCU_STALL_DELAY_DELTA
;
826 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
828 rsp
->gp_start
= jiffies
;
829 rsp
->jiffies_stall
= jiffies
+ jiffies_till_stall_check();
832 static void print_other_cpu_stall(struct rcu_state
*rsp
)
838 struct rcu_node
*rnp
= rcu_get_root(rsp
);
840 /* Only let one CPU complain about others per time interval. */
842 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
843 delta
= jiffies
- rsp
->jiffies_stall
;
844 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
845 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
848 rsp
->jiffies_stall
= jiffies
+ 3 * jiffies_till_stall_check() + 3;
849 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
852 * OK, time to rat on our buddy...
853 * See Documentation/RCU/stallwarn.txt for info on how to debug
854 * RCU CPU stall warnings.
856 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
858 print_cpu_stall_info_begin();
859 rcu_for_each_leaf_node(rsp
, rnp
) {
860 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
861 ndetected
+= rcu_print_task_stall(rnp
);
862 if (rnp
->qsmask
!= 0) {
863 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
864 if (rnp
->qsmask
& (1UL << cpu
)) {
865 print_cpu_stall_info(rsp
,
870 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
874 * Now rat on any tasks that got kicked up to the root rcu_node
875 * due to CPU offlining.
877 rnp
= rcu_get_root(rsp
);
878 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
879 ndetected
+= rcu_print_task_stall(rnp
);
880 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
882 print_cpu_stall_info_end();
883 printk(KERN_CONT
"(detected by %d, t=%ld jiffies)\n",
884 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
886 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
887 else if (!trigger_all_cpu_backtrace())
890 /* Complain about tasks blocking the grace period. */
892 rcu_print_detail_task_stall(rsp
);
894 force_quiescent_state(rsp
); /* Kick them all. */
897 static void print_cpu_stall(struct rcu_state
*rsp
)
900 struct rcu_node
*rnp
= rcu_get_root(rsp
);
903 * OK, time to rat on ourselves...
904 * See Documentation/RCU/stallwarn.txt for info on how to debug
905 * RCU CPU stall warnings.
907 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
908 print_cpu_stall_info_begin();
909 print_cpu_stall_info(rsp
, smp_processor_id());
910 print_cpu_stall_info_end();
911 printk(KERN_CONT
" (t=%lu jiffies)\n", jiffies
- rsp
->gp_start
);
912 if (!trigger_all_cpu_backtrace())
915 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
916 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
917 rsp
->jiffies_stall
= jiffies
+
918 3 * jiffies_till_stall_check() + 3;
919 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
921 set_need_resched(); /* kick ourselves to get things going. */
924 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
928 struct rcu_node
*rnp
;
930 if (rcu_cpu_stall_suppress
)
932 j
= ACCESS_ONCE(jiffies
);
933 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
935 if (rcu_gp_in_progress(rsp
) &&
936 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
938 /* We haven't checked in, so go dump stack. */
939 print_cpu_stall(rsp
);
941 } else if (rcu_gp_in_progress(rsp
) &&
942 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
944 /* They had a few time units to dump stack, so complain. */
945 print_other_cpu_stall(rsp
);
949 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
951 rcu_cpu_stall_suppress
= 1;
956 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
958 * Set the stall-warning timeout way off into the future, thus preventing
959 * any RCU CPU stall-warning messages from appearing in the current set of
962 * The caller must disable hard irqs.
964 void rcu_cpu_stall_reset(void)
966 struct rcu_state
*rsp
;
968 for_each_rcu_flavor(rsp
)
969 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
972 static struct notifier_block rcu_panic_block
= {
973 .notifier_call
= rcu_panic
,
976 static void __init
check_cpu_stall_init(void)
978 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
982 * Update CPU-local rcu_data state to record the newly noticed grace period.
983 * This is used both when we started the grace period and when we notice
984 * that someone else started the grace period. The caller must hold the
985 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
986 * and must have irqs disabled.
988 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
990 if (rdp
->gpnum
!= rnp
->gpnum
) {
992 * If the current grace period is waiting for this CPU,
993 * set up to detect a quiescent state, otherwise don't
994 * go looking for one.
996 rdp
->gpnum
= rnp
->gpnum
;
997 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
998 rdp
->passed_quiesce
= 0;
999 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1000 zero_cpu_stall_ticks(rdp
);
1004 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1006 unsigned long flags
;
1007 struct rcu_node
*rnp
;
1009 local_irq_save(flags
);
1011 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
1012 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1013 local_irq_restore(flags
);
1016 __note_new_gpnum(rsp
, rnp
, rdp
);
1017 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1021 * Did someone else start a new RCU grace period start since we last
1022 * checked? Update local state appropriately if so. Must be called
1023 * on the CPU corresponding to rdp.
1026 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1028 unsigned long flags
;
1031 local_irq_save(flags
);
1032 if (rdp
->gpnum
!= rsp
->gpnum
) {
1033 note_new_gpnum(rsp
, rdp
);
1036 local_irq_restore(flags
);
1041 * Initialize the specified rcu_data structure's callback list to empty.
1043 static void init_callback_list(struct rcu_data
*rdp
)
1047 rdp
->nxtlist
= NULL
;
1048 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1049 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1053 * Advance this CPU's callbacks, but only if the current grace period
1054 * has ended. This may be called only from the CPU to whom the rdp
1055 * belongs. In addition, the corresponding leaf rcu_node structure's
1056 * ->lock must be held by the caller, with irqs disabled.
1059 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1061 /* Did another grace period end? */
1062 if (rdp
->completed
!= rnp
->completed
) {
1064 /* Advance callbacks. No harm if list empty. */
1065 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
1066 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
1067 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1069 /* Remember that we saw this grace-period completion. */
1070 rdp
->completed
= rnp
->completed
;
1071 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
1074 * If we were in an extended quiescent state, we may have
1075 * missed some grace periods that others CPUs handled on
1076 * our behalf. Catch up with this state to avoid noting
1077 * spurious new grace periods. If another grace period
1078 * has started, then rnp->gpnum will have advanced, so
1079 * we will detect this later on. Of course, any quiescent
1080 * states we found for the old GP are now invalid.
1082 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
)) {
1083 rdp
->gpnum
= rdp
->completed
;
1084 rdp
->passed_quiesce
= 0;
1088 * If RCU does not need a quiescent state from this CPU,
1089 * then make sure that this CPU doesn't go looking for one.
1091 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1092 rdp
->qs_pending
= 0;
1097 * Advance this CPU's callbacks, but only if the current grace period
1098 * has ended. This may be called only from the CPU to whom the rdp
1102 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1104 unsigned long flags
;
1105 struct rcu_node
*rnp
;
1107 local_irq_save(flags
);
1109 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1110 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1111 local_irq_restore(flags
);
1114 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1115 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1119 * Do per-CPU grace-period initialization for running CPU. The caller
1120 * must hold the lock of the leaf rcu_node structure corresponding to
1124 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1126 /* Prior grace period ended, so advance callbacks for current CPU. */
1127 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1129 /* Set state so that this CPU will detect the next quiescent state. */
1130 __note_new_gpnum(rsp
, rnp
, rdp
);
1134 * Initialize a new grace period.
1136 static int rcu_gp_init(struct rcu_state
*rsp
)
1138 struct rcu_data
*rdp
;
1139 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1141 raw_spin_lock_irq(&rnp
->lock
);
1142 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1144 if (rcu_gp_in_progress(rsp
)) {
1145 /* Grace period already in progress, don't start another. */
1146 raw_spin_unlock_irq(&rnp
->lock
);
1150 /* Advance to a new grace period and initialize state. */
1152 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1153 record_gp_stall_check_time(rsp
);
1154 raw_spin_unlock_irq(&rnp
->lock
);
1156 /* Exclude any concurrent CPU-hotplug operations. */
1160 * Set the quiescent-state-needed bits in all the rcu_node
1161 * structures for all currently online CPUs in breadth-first order,
1162 * starting from the root rcu_node structure, relying on the layout
1163 * of the tree within the rsp->node[] array. Note that other CPUs
1164 * will access only the leaves of the hierarchy, thus seeing that no
1165 * grace period is in progress, at least until the corresponding
1166 * leaf node has been initialized. In addition, we have excluded
1167 * CPU-hotplug operations.
1169 * The grace period cannot complete until the initialization
1170 * process finishes, because this kthread handles both.
1172 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1173 raw_spin_lock_irq(&rnp
->lock
);
1174 rdp
= this_cpu_ptr(rsp
->rda
);
1175 rcu_preempt_check_blocked_tasks(rnp
);
1176 rnp
->qsmask
= rnp
->qsmaskinit
;
1177 rnp
->gpnum
= rsp
->gpnum
;
1178 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1179 rnp
->completed
= rsp
->completed
;
1180 if (rnp
== rdp
->mynode
)
1181 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1182 rcu_preempt_boost_start_gp(rnp
);
1183 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1184 rnp
->level
, rnp
->grplo
,
1185 rnp
->grphi
, rnp
->qsmask
);
1186 raw_spin_unlock_irq(&rnp
->lock
);
1187 #ifdef CONFIG_PROVE_RCU_DELAY
1188 if ((random32() % (rcu_num_nodes
* 8)) == 0)
1189 schedule_timeout_uninterruptible(2);
1190 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1199 * Do one round of quiescent-state forcing.
1201 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1203 int fqs_state
= fqs_state_in
;
1204 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1207 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1208 /* Collect dyntick-idle snapshots. */
1209 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1210 fqs_state
= RCU_FORCE_QS
;
1212 /* Handle dyntick-idle and offline CPUs. */
1213 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1215 /* Clear flag to prevent immediate re-entry. */
1216 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1217 raw_spin_lock_irq(&rnp
->lock
);
1218 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1219 raw_spin_unlock_irq(&rnp
->lock
);
1225 * Clean up after the old grace period.
1227 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1229 unsigned long gp_duration
;
1230 struct rcu_data
*rdp
;
1231 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1233 raw_spin_lock_irq(&rnp
->lock
);
1234 gp_duration
= jiffies
- rsp
->gp_start
;
1235 if (gp_duration
> rsp
->gp_max
)
1236 rsp
->gp_max
= gp_duration
;
1239 * We know the grace period is complete, but to everyone else
1240 * it appears to still be ongoing. But it is also the case
1241 * that to everyone else it looks like there is nothing that
1242 * they can do to advance the grace period. It is therefore
1243 * safe for us to drop the lock in order to mark the grace
1244 * period as completed in all of the rcu_node structures.
1246 raw_spin_unlock_irq(&rnp
->lock
);
1249 * Propagate new ->completed value to rcu_node structures so
1250 * that other CPUs don't have to wait until the start of the next
1251 * grace period to process their callbacks. This also avoids
1252 * some nasty RCU grace-period initialization races by forcing
1253 * the end of the current grace period to be completely recorded in
1254 * all of the rcu_node structures before the beginning of the next
1255 * grace period is recorded in any of the rcu_node structures.
1257 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1258 raw_spin_lock_irq(&rnp
->lock
);
1259 rnp
->completed
= rsp
->gpnum
;
1260 raw_spin_unlock_irq(&rnp
->lock
);
1263 rnp
= rcu_get_root(rsp
);
1264 raw_spin_lock_irq(&rnp
->lock
);
1266 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1267 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1268 rsp
->fqs_state
= RCU_GP_IDLE
;
1269 rdp
= this_cpu_ptr(rsp
->rda
);
1270 if (cpu_needs_another_gp(rsp
, rdp
))
1272 raw_spin_unlock_irq(&rnp
->lock
);
1276 * Body of kthread that handles grace periods.
1278 static int __noreturn
rcu_gp_kthread(void *arg
)
1283 struct rcu_state
*rsp
= arg
;
1284 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1288 /* Handle grace-period start. */
1290 wait_event_interruptible(rsp
->gp_wq
,
1293 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1297 flush_signals(current
);
1300 /* Handle quiescent-state forcing. */
1301 fqs_state
= RCU_SAVE_DYNTICK
;
1302 j
= jiffies_till_first_fqs
;
1305 jiffies_till_first_fqs
= HZ
;
1308 rsp
->jiffies_force_qs
= jiffies
+ j
;
1309 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1310 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1311 (!ACCESS_ONCE(rnp
->qsmask
) &&
1312 !rcu_preempt_blocked_readers_cgp(rnp
)),
1314 /* If grace period done, leave loop. */
1315 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1316 !rcu_preempt_blocked_readers_cgp(rnp
))
1318 /* If time for quiescent-state forcing, do it. */
1319 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1320 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1323 /* Deal with stray signal. */
1325 flush_signals(current
);
1327 j
= jiffies_till_next_fqs
;
1330 jiffies_till_next_fqs
= HZ
;
1333 jiffies_till_next_fqs
= 1;
1337 /* Handle grace-period end. */
1338 rcu_gp_cleanup(rsp
);
1343 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1344 * in preparation for detecting the next grace period. The caller must hold
1345 * the root node's ->lock, which is released before return. Hard irqs must
1348 * Note that it is legal for a dying CPU (which is marked as offline) to
1349 * invoke this function. This can happen when the dying CPU reports its
1353 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1354 __releases(rcu_get_root(rsp
)->lock
)
1356 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1357 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1359 if (!rsp
->gp_kthread
||
1360 !cpu_needs_another_gp(rsp
, rdp
)) {
1362 * Either we have not yet spawned the grace-period
1363 * task or this CPU does not need another grace period.
1364 * Either way, don't start a new grace period.
1366 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1370 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1371 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1372 wake_up(&rsp
->gp_wq
);
1376 * Report a full set of quiescent states to the specified rcu_state
1377 * data structure. This involves cleaning up after the prior grace
1378 * period and letting rcu_start_gp() start up the next grace period
1379 * if one is needed. Note that the caller must hold rnp->lock, as
1380 * required by rcu_start_gp(), which will release it.
1382 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1383 __releases(rcu_get_root(rsp
)->lock
)
1385 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1386 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1387 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1391 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1392 * Allows quiescent states for a group of CPUs to be reported at one go
1393 * to the specified rcu_node structure, though all the CPUs in the group
1394 * must be represented by the same rcu_node structure (which need not be
1395 * a leaf rcu_node structure, though it often will be). That structure's
1396 * lock must be held upon entry, and it is released before return.
1399 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1400 struct rcu_node
*rnp
, unsigned long flags
)
1401 __releases(rnp
->lock
)
1403 struct rcu_node
*rnp_c
;
1405 /* Walk up the rcu_node hierarchy. */
1407 if (!(rnp
->qsmask
& mask
)) {
1409 /* Our bit has already been cleared, so done. */
1410 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1413 rnp
->qsmask
&= ~mask
;
1414 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1415 mask
, rnp
->qsmask
, rnp
->level
,
1416 rnp
->grplo
, rnp
->grphi
,
1418 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1420 /* Other bits still set at this level, so done. */
1421 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1424 mask
= rnp
->grpmask
;
1425 if (rnp
->parent
== NULL
) {
1427 /* No more levels. Exit loop holding root lock. */
1431 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1434 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1435 WARN_ON_ONCE(rnp_c
->qsmask
);
1439 * Get here if we are the last CPU to pass through a quiescent
1440 * state for this grace period. Invoke rcu_report_qs_rsp()
1441 * to clean up and start the next grace period if one is needed.
1443 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1447 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1448 * structure. This must be either called from the specified CPU, or
1449 * called when the specified CPU is known to be offline (and when it is
1450 * also known that no other CPU is concurrently trying to help the offline
1451 * CPU). The lastcomp argument is used to make sure we are still in the
1452 * grace period of interest. We don't want to end the current grace period
1453 * based on quiescent states detected in an earlier grace period!
1456 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1458 unsigned long flags
;
1460 struct rcu_node
*rnp
;
1463 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1464 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1465 rnp
->completed
== rnp
->gpnum
) {
1468 * The grace period in which this quiescent state was
1469 * recorded has ended, so don't report it upwards.
1470 * We will instead need a new quiescent state that lies
1471 * within the current grace period.
1473 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1474 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1477 mask
= rdp
->grpmask
;
1478 if ((rnp
->qsmask
& mask
) == 0) {
1479 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1481 rdp
->qs_pending
= 0;
1484 * This GP can't end until cpu checks in, so all of our
1485 * callbacks can be processed during the next GP.
1487 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1489 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1494 * Check to see if there is a new grace period of which this CPU
1495 * is not yet aware, and if so, set up local rcu_data state for it.
1496 * Otherwise, see if this CPU has just passed through its first
1497 * quiescent state for this grace period, and record that fact if so.
1500 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1502 /* If there is now a new grace period, record and return. */
1503 if (check_for_new_grace_period(rsp
, rdp
))
1507 * Does this CPU still need to do its part for current grace period?
1508 * If no, return and let the other CPUs do their part as well.
1510 if (!rdp
->qs_pending
)
1514 * Was there a quiescent state since the beginning of the grace
1515 * period? If no, then exit and wait for the next call.
1517 if (!rdp
->passed_quiesce
)
1521 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1524 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1527 #ifdef CONFIG_HOTPLUG_CPU
1530 * Send the specified CPU's RCU callbacks to the orphanage. The
1531 * specified CPU must be offline, and the caller must hold the
1535 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1536 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1539 * Orphan the callbacks. First adjust the counts. This is safe
1540 * because ->onofflock excludes _rcu_barrier()'s adoption of
1541 * the callbacks, thus no memory barrier is required.
1543 if (rdp
->nxtlist
!= NULL
) {
1544 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1545 rsp
->qlen
+= rdp
->qlen
;
1546 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1548 ACCESS_ONCE(rdp
->qlen
) = 0;
1552 * Next, move those callbacks still needing a grace period to
1553 * the orphanage, where some other CPU will pick them up.
1554 * Some of the callbacks might have gone partway through a grace
1555 * period, but that is too bad. They get to start over because we
1556 * cannot assume that grace periods are synchronized across CPUs.
1557 * We don't bother updating the ->nxttail[] array yet, instead
1558 * we just reset the whole thing later on.
1560 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1561 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1562 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1563 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1567 * Then move the ready-to-invoke callbacks to the orphanage,
1568 * where some other CPU will pick them up. These will not be
1569 * required to pass though another grace period: They are done.
1571 if (rdp
->nxtlist
!= NULL
) {
1572 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1573 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1576 /* Finally, initialize the rcu_data structure's list to empty. */
1577 init_callback_list(rdp
);
1581 * Adopt the RCU callbacks from the specified rcu_state structure's
1582 * orphanage. The caller must hold the ->onofflock.
1584 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1587 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1589 /* Do the accounting first. */
1590 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1591 rdp
->qlen
+= rsp
->qlen
;
1592 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1593 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1594 rcu_idle_count_callbacks_posted();
1599 * We do not need a memory barrier here because the only way we
1600 * can get here if there is an rcu_barrier() in flight is if
1601 * we are the task doing the rcu_barrier().
1604 /* First adopt the ready-to-invoke callbacks. */
1605 if (rsp
->orphan_donelist
!= NULL
) {
1606 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1607 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1608 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1609 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1610 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1611 rsp
->orphan_donelist
= NULL
;
1612 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1615 /* And then adopt the callbacks that still need a grace period. */
1616 if (rsp
->orphan_nxtlist
!= NULL
) {
1617 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1618 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1619 rsp
->orphan_nxtlist
= NULL
;
1620 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1625 * Trace the fact that this CPU is going offline.
1627 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1629 RCU_TRACE(unsigned long mask
);
1630 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1631 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1633 RCU_TRACE(mask
= rdp
->grpmask
);
1634 trace_rcu_grace_period(rsp
->name
,
1635 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1640 * The CPU has been completely removed, and some other CPU is reporting
1641 * this fact from process context. Do the remainder of the cleanup,
1642 * including orphaning the outgoing CPU's RCU callbacks, and also
1643 * adopting them. There can only be one CPU hotplug operation at a time,
1644 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1646 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1648 unsigned long flags
;
1650 int need_report
= 0;
1651 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1652 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1654 /* Adjust any no-longer-needed kthreads. */
1655 rcu_boost_kthread_setaffinity(rnp
, -1);
1657 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1659 /* Exclude any attempts to start a new grace period. */
1660 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1662 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1663 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1664 rcu_adopt_orphan_cbs(rsp
);
1666 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1667 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1669 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1670 rnp
->qsmaskinit
&= ~mask
;
1671 if (rnp
->qsmaskinit
!= 0) {
1672 if (rnp
!= rdp
->mynode
)
1673 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1676 if (rnp
== rdp
->mynode
)
1677 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1679 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1680 mask
= rnp
->grpmask
;
1682 } while (rnp
!= NULL
);
1685 * We still hold the leaf rcu_node structure lock here, and
1686 * irqs are still disabled. The reason for this subterfuge is
1687 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1688 * held leads to deadlock.
1690 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1692 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1693 rcu_report_unblock_qs_rnp(rnp
, flags
);
1695 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1696 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1697 rcu_report_exp_rnp(rsp
, rnp
, true);
1698 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1699 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1700 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1701 init_callback_list(rdp
);
1702 /* Disallow further callbacks on this CPU. */
1703 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1706 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1708 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1712 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1716 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1719 * Invoke any RCU callbacks that have made it to the end of their grace
1720 * period. Thottle as specified by rdp->blimit.
1722 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1724 unsigned long flags
;
1725 struct rcu_head
*next
, *list
, **tail
;
1726 int bl
, count
, count_lazy
, i
;
1728 /* If no callbacks are ready, just return.*/
1729 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1730 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1731 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1732 need_resched(), is_idle_task(current
),
1733 rcu_is_callbacks_kthread());
1738 * Extract the list of ready callbacks, disabling to prevent
1739 * races with call_rcu() from interrupt handlers.
1741 local_irq_save(flags
);
1742 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1744 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1745 list
= rdp
->nxtlist
;
1746 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1747 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1748 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1749 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1750 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1751 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1752 local_irq_restore(flags
);
1754 /* Invoke callbacks. */
1755 count
= count_lazy
= 0;
1759 debug_rcu_head_unqueue(list
);
1760 if (__rcu_reclaim(rsp
->name
, list
))
1763 /* Stop only if limit reached and CPU has something to do. */
1764 if (++count
>= bl
&&
1766 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1770 local_irq_save(flags
);
1771 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1772 is_idle_task(current
),
1773 rcu_is_callbacks_kthread());
1775 /* Update count, and requeue any remaining callbacks. */
1777 *tail
= rdp
->nxtlist
;
1778 rdp
->nxtlist
= list
;
1779 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1780 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1781 rdp
->nxttail
[i
] = tail
;
1785 smp_mb(); /* List handling before counting for rcu_barrier(). */
1786 rdp
->qlen_lazy
-= count_lazy
;
1787 ACCESS_ONCE(rdp
->qlen
) -= count
;
1788 rdp
->n_cbs_invoked
+= count
;
1790 /* Reinstate batch limit if we have worked down the excess. */
1791 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1792 rdp
->blimit
= blimit
;
1794 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1795 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1796 rdp
->qlen_last_fqs_check
= 0;
1797 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1798 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1799 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1800 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
1802 local_irq_restore(flags
);
1804 /* Re-invoke RCU core processing if there are callbacks remaining. */
1805 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1810 * Check to see if this CPU is in a non-context-switch quiescent state
1811 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1812 * Also schedule RCU core processing.
1814 * This function must be called from hardirq context. It is normally
1815 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1816 * false, there is no point in invoking rcu_check_callbacks().
1818 void rcu_check_callbacks(int cpu
, int user
)
1820 trace_rcu_utilization("Start scheduler-tick");
1821 increment_cpu_stall_ticks();
1822 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1825 * Get here if this CPU took its interrupt from user
1826 * mode or from the idle loop, and if this is not a
1827 * nested interrupt. In this case, the CPU is in
1828 * a quiescent state, so note it.
1830 * No memory barrier is required here because both
1831 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1832 * variables that other CPUs neither access nor modify,
1833 * at least not while the corresponding CPU is online.
1839 } else if (!in_softirq()) {
1842 * Get here if this CPU did not take its interrupt from
1843 * softirq, in other words, if it is not interrupting
1844 * a rcu_bh read-side critical section. This is an _bh
1845 * critical section, so note it.
1850 rcu_preempt_check_callbacks(cpu
);
1851 if (rcu_pending(cpu
))
1853 trace_rcu_utilization("End scheduler-tick");
1857 * Scan the leaf rcu_node structures, processing dyntick state for any that
1858 * have not yet encountered a quiescent state, using the function specified.
1859 * Also initiate boosting for any threads blocked on the root rcu_node.
1861 * The caller must have suppressed start of new grace periods.
1863 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1867 unsigned long flags
;
1869 struct rcu_node
*rnp
;
1871 rcu_for_each_leaf_node(rsp
, rnp
) {
1874 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1875 if (!rcu_gp_in_progress(rsp
)) {
1876 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1879 if (rnp
->qsmask
== 0) {
1880 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1885 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1886 if ((rnp
->qsmask
& bit
) != 0 &&
1887 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1892 /* rcu_report_qs_rnp() releases rnp->lock. */
1893 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1896 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1898 rnp
= rcu_get_root(rsp
);
1899 if (rnp
->qsmask
== 0) {
1900 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1901 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1906 * Force quiescent states on reluctant CPUs, and also detect which
1907 * CPUs are in dyntick-idle mode.
1909 static void force_quiescent_state(struct rcu_state
*rsp
)
1911 unsigned long flags
;
1913 struct rcu_node
*rnp
;
1914 struct rcu_node
*rnp_old
= NULL
;
1916 /* Funnel through hierarchy to reduce memory contention. */
1917 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
1918 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
1919 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
1920 !raw_spin_trylock(&rnp
->fqslock
);
1921 if (rnp_old
!= NULL
)
1922 raw_spin_unlock(&rnp_old
->fqslock
);
1924 rsp
->n_force_qs_lh
++;
1929 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1931 /* Reached the root of the rcu_node tree, acquire lock. */
1932 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
1933 raw_spin_unlock(&rnp_old
->fqslock
);
1934 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1935 rsp
->n_force_qs_lh
++;
1936 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1937 return; /* Someone beat us to it. */
1939 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
1940 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1941 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1945 * This does the RCU core processing work for the specified rcu_state
1946 * and rcu_data structures. This may be called only from the CPU to
1947 * whom the rdp belongs.
1950 __rcu_process_callbacks(struct rcu_state
*rsp
)
1952 unsigned long flags
;
1953 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1955 WARN_ON_ONCE(rdp
->beenonline
== 0);
1958 * Advance callbacks in response to end of earlier grace
1959 * period that some other CPU ended.
1961 rcu_process_gp_end(rsp
, rdp
);
1963 /* Update RCU state based on any recent quiescent states. */
1964 rcu_check_quiescent_state(rsp
, rdp
);
1966 /* Does this CPU require a not-yet-started grace period? */
1967 if (cpu_needs_another_gp(rsp
, rdp
)) {
1968 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1969 rcu_start_gp(rsp
, flags
); /* releases above lock */
1972 /* If there are callbacks ready, invoke them. */
1973 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1974 invoke_rcu_callbacks(rsp
, rdp
);
1978 * Do RCU core processing for the current CPU.
1980 static void rcu_process_callbacks(struct softirq_action
*unused
)
1982 struct rcu_state
*rsp
;
1984 if (cpu_is_offline(smp_processor_id()))
1986 trace_rcu_utilization("Start RCU core");
1987 for_each_rcu_flavor(rsp
)
1988 __rcu_process_callbacks(rsp
);
1989 trace_rcu_utilization("End RCU core");
1993 * Schedule RCU callback invocation. If the specified type of RCU
1994 * does not support RCU priority boosting, just do a direct call,
1995 * otherwise wake up the per-CPU kernel kthread. Note that because we
1996 * are running on the current CPU with interrupts disabled, the
1997 * rcu_cpu_kthread_task cannot disappear out from under us.
1999 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2001 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2003 if (likely(!rsp
->boost
)) {
2004 rcu_do_batch(rsp
, rdp
);
2007 invoke_rcu_callbacks_kthread();
2010 static void invoke_rcu_core(void)
2012 raise_softirq(RCU_SOFTIRQ
);
2016 * Handle any core-RCU processing required by a call_rcu() invocation.
2018 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2019 struct rcu_head
*head
, unsigned long flags
)
2022 * If called from an extended quiescent state, invoke the RCU
2023 * core in order to force a re-evaluation of RCU's idleness.
2025 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2028 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2029 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2033 * Force the grace period if too many callbacks or too long waiting.
2034 * Enforce hysteresis, and don't invoke force_quiescent_state()
2035 * if some other CPU has recently done so. Also, don't bother
2036 * invoking force_quiescent_state() if the newly enqueued callback
2037 * is the only one waiting for a grace period to complete.
2039 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2041 /* Are we ignoring a completed grace period? */
2042 rcu_process_gp_end(rsp
, rdp
);
2043 check_for_new_grace_period(rsp
, rdp
);
2045 /* Start a new grace period if one not already started. */
2046 if (!rcu_gp_in_progress(rsp
)) {
2047 unsigned long nestflag
;
2048 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2050 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
2051 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
2053 /* Give the grace period a kick. */
2054 rdp
->blimit
= LONG_MAX
;
2055 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2056 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2057 force_quiescent_state(rsp
);
2058 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2059 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2065 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2066 struct rcu_state
*rsp
, bool lazy
)
2068 unsigned long flags
;
2069 struct rcu_data
*rdp
;
2071 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2072 debug_rcu_head_queue(head
);
2077 * Opportunistically note grace-period endings and beginnings.
2078 * Note that we might see a beginning right after we see an
2079 * end, but never vice versa, since this CPU has to pass through
2080 * a quiescent state betweentimes.
2082 local_irq_save(flags
);
2083 rdp
= this_cpu_ptr(rsp
->rda
);
2085 /* Add the callback to our list. */
2086 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
)) {
2087 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2089 local_irq_restore(flags
);
2092 ACCESS_ONCE(rdp
->qlen
)++;
2096 rcu_idle_count_callbacks_posted();
2097 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2098 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2099 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2101 if (__is_kfree_rcu_offset((unsigned long)func
))
2102 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2103 rdp
->qlen_lazy
, rdp
->qlen
);
2105 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2107 /* Go handle any RCU core processing required. */
2108 __call_rcu_core(rsp
, rdp
, head
, flags
);
2109 local_irq_restore(flags
);
2113 * Queue an RCU-sched callback for invocation after a grace period.
2115 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2117 __call_rcu(head
, func
, &rcu_sched_state
, 0);
2119 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2122 * Queue an RCU callback for invocation after a quicker grace period.
2124 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2126 __call_rcu(head
, func
, &rcu_bh_state
, 0);
2128 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2131 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2132 * any blocking grace-period wait automatically implies a grace period
2133 * if there is only one CPU online at any point time during execution
2134 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2135 * occasionally incorrectly indicate that there are multiple CPUs online
2136 * when there was in fact only one the whole time, as this just adds
2137 * some overhead: RCU still operates correctly.
2139 static inline int rcu_blocking_is_gp(void)
2143 might_sleep(); /* Check for RCU read-side critical section. */
2145 ret
= num_online_cpus() <= 1;
2151 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2153 * Control will return to the caller some time after a full rcu-sched
2154 * grace period has elapsed, in other words after all currently executing
2155 * rcu-sched read-side critical sections have completed. These read-side
2156 * critical sections are delimited by rcu_read_lock_sched() and
2157 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2158 * local_irq_disable(), and so on may be used in place of
2159 * rcu_read_lock_sched().
2161 * This means that all preempt_disable code sequences, including NMI and
2162 * hardware-interrupt handlers, in progress on entry will have completed
2163 * before this primitive returns. However, this does not guarantee that
2164 * softirq handlers will have completed, since in some kernels, these
2165 * handlers can run in process context, and can block.
2167 * This primitive provides the guarantees made by the (now removed)
2168 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2169 * guarantees that rcu_read_lock() sections will have completed.
2170 * In "classic RCU", these two guarantees happen to be one and
2171 * the same, but can differ in realtime RCU implementations.
2173 void synchronize_sched(void)
2175 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2176 !lock_is_held(&rcu_lock_map
) &&
2177 !lock_is_held(&rcu_sched_lock_map
),
2178 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2179 if (rcu_blocking_is_gp())
2181 wait_rcu_gp(call_rcu_sched
);
2183 EXPORT_SYMBOL_GPL(synchronize_sched
);
2186 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2188 * Control will return to the caller some time after a full rcu_bh grace
2189 * period has elapsed, in other words after all currently executing rcu_bh
2190 * read-side critical sections have completed. RCU read-side critical
2191 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2192 * and may be nested.
2194 void synchronize_rcu_bh(void)
2196 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2197 !lock_is_held(&rcu_lock_map
) &&
2198 !lock_is_held(&rcu_sched_lock_map
),
2199 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2200 if (rcu_blocking_is_gp())
2202 wait_rcu_gp(call_rcu_bh
);
2204 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2206 static atomic_t sync_sched_expedited_started
= ATOMIC_INIT(0);
2207 static atomic_t sync_sched_expedited_done
= ATOMIC_INIT(0);
2209 static int synchronize_sched_expedited_cpu_stop(void *data
)
2212 * There must be a full memory barrier on each affected CPU
2213 * between the time that try_stop_cpus() is called and the
2214 * time that it returns.
2216 * In the current initial implementation of cpu_stop, the
2217 * above condition is already met when the control reaches
2218 * this point and the following smp_mb() is not strictly
2219 * necessary. Do smp_mb() anyway for documentation and
2220 * robustness against future implementation changes.
2222 smp_mb(); /* See above comment block. */
2227 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2229 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2230 * approach to force the grace period to end quickly. This consumes
2231 * significant time on all CPUs and is unfriendly to real-time workloads,
2232 * so is thus not recommended for any sort of common-case code. In fact,
2233 * if you are using synchronize_sched_expedited() in a loop, please
2234 * restructure your code to batch your updates, and then use a single
2235 * synchronize_sched() instead.
2237 * Note that it is illegal to call this function while holding any lock
2238 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2239 * to call this function from a CPU-hotplug notifier. Failing to observe
2240 * these restriction will result in deadlock.
2242 * This implementation can be thought of as an application of ticket
2243 * locking to RCU, with sync_sched_expedited_started and
2244 * sync_sched_expedited_done taking on the roles of the halves
2245 * of the ticket-lock word. Each task atomically increments
2246 * sync_sched_expedited_started upon entry, snapshotting the old value,
2247 * then attempts to stop all the CPUs. If this succeeds, then each
2248 * CPU will have executed a context switch, resulting in an RCU-sched
2249 * grace period. We are then done, so we use atomic_cmpxchg() to
2250 * update sync_sched_expedited_done to match our snapshot -- but
2251 * only if someone else has not already advanced past our snapshot.
2253 * On the other hand, if try_stop_cpus() fails, we check the value
2254 * of sync_sched_expedited_done. If it has advanced past our
2255 * initial snapshot, then someone else must have forced a grace period
2256 * some time after we took our snapshot. In this case, our work is
2257 * done for us, and we can simply return. Otherwise, we try again,
2258 * but keep our initial snapshot for purposes of checking for someone
2259 * doing our work for us.
2261 * If we fail too many times in a row, we fall back to synchronize_sched().
2263 void synchronize_sched_expedited(void)
2265 int firstsnap
, s
, snap
, trycount
= 0;
2267 /* Note that atomic_inc_return() implies full memory barrier. */
2268 firstsnap
= snap
= atomic_inc_return(&sync_sched_expedited_started
);
2270 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2273 * Each pass through the following loop attempts to force a
2274 * context switch on each CPU.
2276 while (try_stop_cpus(cpu_online_mask
,
2277 synchronize_sched_expedited_cpu_stop
,
2281 /* No joy, try again later. Or just synchronize_sched(). */
2282 if (trycount
++ < 10) {
2283 udelay(trycount
* num_online_cpus());
2285 synchronize_sched();
2289 /* Check to see if someone else did our work for us. */
2290 s
= atomic_read(&sync_sched_expedited_done
);
2291 if (UINT_CMP_GE((unsigned)s
, (unsigned)firstsnap
)) {
2292 smp_mb(); /* ensure test happens before caller kfree */
2297 * Refetching sync_sched_expedited_started allows later
2298 * callers to piggyback on our grace period. We subtract
2299 * 1 to get the same token that the last incrementer got.
2300 * We retry after they started, so our grace period works
2301 * for them, and they started after our first try, so their
2302 * grace period works for us.
2305 snap
= atomic_read(&sync_sched_expedited_started
);
2306 smp_mb(); /* ensure read is before try_stop_cpus(). */
2310 * Everyone up to our most recent fetch is covered by our grace
2311 * period. Update the counter, but only if our work is still
2312 * relevant -- which it won't be if someone who started later
2313 * than we did beat us to the punch.
2316 s
= atomic_read(&sync_sched_expedited_done
);
2317 if (UINT_CMP_GE((unsigned)s
, (unsigned)snap
)) {
2318 smp_mb(); /* ensure test happens before caller kfree */
2321 } while (atomic_cmpxchg(&sync_sched_expedited_done
, s
, snap
) != s
);
2325 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2328 * Check to see if there is any immediate RCU-related work to be done
2329 * by the current CPU, for the specified type of RCU, returning 1 if so.
2330 * The checks are in order of increasing expense: checks that can be
2331 * carried out against CPU-local state are performed first. However,
2332 * we must check for CPU stalls first, else we might not get a chance.
2334 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2336 struct rcu_node
*rnp
= rdp
->mynode
;
2338 rdp
->n_rcu_pending
++;
2340 /* Check for CPU stalls, if enabled. */
2341 check_cpu_stall(rsp
, rdp
);
2343 /* Is the RCU core waiting for a quiescent state from this CPU? */
2344 if (rcu_scheduler_fully_active
&&
2345 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2346 rdp
->n_rp_qs_pending
++;
2347 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2348 rdp
->n_rp_report_qs
++;
2352 /* Does this CPU have callbacks ready to invoke? */
2353 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2354 rdp
->n_rp_cb_ready
++;
2358 /* Has RCU gone idle with this CPU needing another grace period? */
2359 if (cpu_needs_another_gp(rsp
, rdp
)) {
2360 rdp
->n_rp_cpu_needs_gp
++;
2364 /* Has another RCU grace period completed? */
2365 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2366 rdp
->n_rp_gp_completed
++;
2370 /* Has a new RCU grace period started? */
2371 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2372 rdp
->n_rp_gp_started
++;
2377 rdp
->n_rp_need_nothing
++;
2382 * Check to see if there is any immediate RCU-related work to be done
2383 * by the current CPU, returning 1 if so. This function is part of the
2384 * RCU implementation; it is -not- an exported member of the RCU API.
2386 static int rcu_pending(int cpu
)
2388 struct rcu_state
*rsp
;
2390 for_each_rcu_flavor(rsp
)
2391 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2397 * Check to see if any future RCU-related work will need to be done
2398 * by the current CPU, even if none need be done immediately, returning
2401 static int rcu_cpu_has_callbacks(int cpu
)
2403 struct rcu_state
*rsp
;
2405 /* RCU callbacks either ready or pending? */
2406 for_each_rcu_flavor(rsp
)
2407 if (per_cpu_ptr(rsp
->rda
, cpu
)->nxtlist
)
2413 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2414 * the compiler is expected to optimize this away.
2416 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2417 int cpu
, unsigned long done
)
2419 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2420 atomic_read(&rsp
->barrier_cpu_count
), done
);
2424 * RCU callback function for _rcu_barrier(). If we are last, wake
2425 * up the task executing _rcu_barrier().
2427 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2429 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2430 struct rcu_state
*rsp
= rdp
->rsp
;
2432 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2433 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2434 complete(&rsp
->barrier_completion
);
2436 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2441 * Called with preemption disabled, and from cross-cpu IRQ context.
2443 static void rcu_barrier_func(void *type
)
2445 struct rcu_state
*rsp
= type
;
2446 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2448 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2449 atomic_inc(&rsp
->barrier_cpu_count
);
2450 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2454 * Orchestrate the specified type of RCU barrier, waiting for all
2455 * RCU callbacks of the specified type to complete.
2457 static void _rcu_barrier(struct rcu_state
*rsp
)
2460 struct rcu_data
*rdp
;
2461 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2462 unsigned long snap_done
;
2464 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2466 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2467 mutex_lock(&rsp
->barrier_mutex
);
2470 * Ensure that all prior references, including to ->n_barrier_done,
2471 * are ordered before the _rcu_barrier() machinery.
2473 smp_mb(); /* See above block comment. */
2476 * Recheck ->n_barrier_done to see if others did our work for us.
2477 * This means checking ->n_barrier_done for an even-to-odd-to-even
2478 * transition. The "if" expression below therefore rounds the old
2479 * value up to the next even number and adds two before comparing.
2481 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2482 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2483 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2484 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2485 smp_mb(); /* caller's subsequent code after above check. */
2486 mutex_unlock(&rsp
->barrier_mutex
);
2491 * Increment ->n_barrier_done to avoid duplicate work. Use
2492 * ACCESS_ONCE() to prevent the compiler from speculating
2493 * the increment to precede the early-exit check.
2495 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2496 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2497 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2498 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2501 * Initialize the count to one rather than to zero in order to
2502 * avoid a too-soon return to zero in case of a short grace period
2503 * (or preemption of this task). Exclude CPU-hotplug operations
2504 * to ensure that no offline CPU has callbacks queued.
2506 init_completion(&rsp
->barrier_completion
);
2507 atomic_set(&rsp
->barrier_cpu_count
, 1);
2511 * Force each CPU with callbacks to register a new callback.
2512 * When that callback is invoked, we will know that all of the
2513 * corresponding CPU's preceding callbacks have been invoked.
2515 for_each_online_cpu(cpu
) {
2516 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2517 if (ACCESS_ONCE(rdp
->qlen
)) {
2518 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2519 rsp
->n_barrier_done
);
2520 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2522 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2523 rsp
->n_barrier_done
);
2529 * Now that we have an rcu_barrier_callback() callback on each
2530 * CPU, and thus each counted, remove the initial count.
2532 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2533 complete(&rsp
->barrier_completion
);
2535 /* Increment ->n_barrier_done to prevent duplicate work. */
2536 smp_mb(); /* Keep increment after above mechanism. */
2537 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2538 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2539 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2540 smp_mb(); /* Keep increment before caller's subsequent code. */
2542 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2543 wait_for_completion(&rsp
->barrier_completion
);
2545 /* Other rcu_barrier() invocations can now safely proceed. */
2546 mutex_unlock(&rsp
->barrier_mutex
);
2550 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2552 void rcu_barrier_bh(void)
2554 _rcu_barrier(&rcu_bh_state
);
2556 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2559 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2561 void rcu_barrier_sched(void)
2563 _rcu_barrier(&rcu_sched_state
);
2565 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2568 * Do boot-time initialization of a CPU's per-CPU RCU data.
2571 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2573 unsigned long flags
;
2574 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2575 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2577 /* Set up local state, ensuring consistent view of global state. */
2578 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2579 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2580 init_callback_list(rdp
);
2582 ACCESS_ONCE(rdp
->qlen
) = 0;
2583 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2584 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2585 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2588 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2592 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2593 * offline event can be happening at a given time. Note also that we
2594 * can accept some slop in the rsp->completed access due to the fact
2595 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2597 static void __cpuinit
2598 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2600 unsigned long flags
;
2602 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2603 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2605 /* Set up local state, ensuring consistent view of global state. */
2606 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2607 rdp
->beenonline
= 1; /* We have now been online. */
2608 rdp
->preemptible
= preemptible
;
2609 rdp
->qlen_last_fqs_check
= 0;
2610 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2611 rdp
->blimit
= blimit
;
2612 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
2613 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2614 atomic_set(&rdp
->dynticks
->dynticks
,
2615 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2616 rcu_prepare_for_idle_init(cpu
);
2617 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2620 * A new grace period might start here. If so, we won't be part
2621 * of it, but that is OK, as we are currently in a quiescent state.
2624 /* Exclude any attempts to start a new GP on large systems. */
2625 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2627 /* Add CPU to rcu_node bitmasks. */
2629 mask
= rdp
->grpmask
;
2631 /* Exclude any attempts to start a new GP on small systems. */
2632 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2633 rnp
->qsmaskinit
|= mask
;
2634 mask
= rnp
->grpmask
;
2635 if (rnp
== rdp
->mynode
) {
2637 * If there is a grace period in progress, we will
2638 * set up to wait for it next time we run the
2641 rdp
->gpnum
= rnp
->completed
;
2642 rdp
->completed
= rnp
->completed
;
2643 rdp
->passed_quiesce
= 0;
2644 rdp
->qs_pending
= 0;
2645 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2647 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2649 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2651 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2654 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2656 struct rcu_state
*rsp
;
2658 for_each_rcu_flavor(rsp
)
2659 rcu_init_percpu_data(cpu
, rsp
,
2660 strcmp(rsp
->name
, "rcu_preempt") == 0);
2664 * Handle CPU online/offline notification events.
2666 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2667 unsigned long action
, void *hcpu
)
2669 long cpu
= (long)hcpu
;
2670 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2671 struct rcu_node
*rnp
= rdp
->mynode
;
2672 struct rcu_state
*rsp
;
2674 trace_rcu_utilization("Start CPU hotplug");
2676 case CPU_UP_PREPARE
:
2677 case CPU_UP_PREPARE_FROZEN
:
2678 rcu_prepare_cpu(cpu
);
2679 rcu_prepare_kthreads(cpu
);
2682 case CPU_DOWN_FAILED
:
2683 rcu_boost_kthread_setaffinity(rnp
, -1);
2685 case CPU_DOWN_PREPARE
:
2686 rcu_boost_kthread_setaffinity(rnp
, cpu
);
2689 case CPU_DYING_FROZEN
:
2691 * The whole machine is "stopped" except this CPU, so we can
2692 * touch any data without introducing corruption. We send the
2693 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2695 for_each_rcu_flavor(rsp
)
2696 rcu_cleanup_dying_cpu(rsp
);
2697 rcu_cleanup_after_idle(cpu
);
2700 case CPU_DEAD_FROZEN
:
2701 case CPU_UP_CANCELED
:
2702 case CPU_UP_CANCELED_FROZEN
:
2703 for_each_rcu_flavor(rsp
)
2704 rcu_cleanup_dead_cpu(cpu
, rsp
);
2709 trace_rcu_utilization("End CPU hotplug");
2714 * Spawn the kthread that handles this RCU flavor's grace periods.
2716 static int __init
rcu_spawn_gp_kthread(void)
2718 unsigned long flags
;
2719 struct rcu_node
*rnp
;
2720 struct rcu_state
*rsp
;
2721 struct task_struct
*t
;
2723 for_each_rcu_flavor(rsp
) {
2724 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
2726 rnp
= rcu_get_root(rsp
);
2727 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2728 rsp
->gp_kthread
= t
;
2729 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2733 early_initcall(rcu_spawn_gp_kthread
);
2736 * This function is invoked towards the end of the scheduler's initialization
2737 * process. Before this is called, the idle task might contain
2738 * RCU read-side critical sections (during which time, this idle
2739 * task is booting the system). After this function is called, the
2740 * idle tasks are prohibited from containing RCU read-side critical
2741 * sections. This function also enables RCU lockdep checking.
2743 void rcu_scheduler_starting(void)
2745 WARN_ON(num_online_cpus() != 1);
2746 WARN_ON(nr_context_switches() > 0);
2747 rcu_scheduler_active
= 1;
2751 * Compute the per-level fanout, either using the exact fanout specified
2752 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2754 #ifdef CONFIG_RCU_FANOUT_EXACT
2755 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2759 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
2760 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2761 rsp
->levelspread
[0] = rcu_fanout_leaf
;
2763 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2764 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2771 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2772 ccur
= rsp
->levelcnt
[i
];
2773 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2777 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2780 * Helper function for rcu_init() that initializes one rcu_state structure.
2782 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2783 struct rcu_data __percpu
*rda
)
2785 static char *buf
[] = { "rcu_node_0",
2788 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2789 static char *fqs
[] = { "rcu_node_fqs_0",
2792 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2796 struct rcu_node
*rnp
;
2798 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2800 /* Initialize the level-tracking arrays. */
2802 for (i
= 0; i
< rcu_num_lvls
; i
++)
2803 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
2804 for (i
= 1; i
< rcu_num_lvls
; i
++)
2805 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2806 rcu_init_levelspread(rsp
);
2808 /* Initialize the elements themselves, starting from the leaves. */
2810 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2811 cpustride
*= rsp
->levelspread
[i
];
2812 rnp
= rsp
->level
[i
];
2813 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2814 raw_spin_lock_init(&rnp
->lock
);
2815 lockdep_set_class_and_name(&rnp
->lock
,
2816 &rcu_node_class
[i
], buf
[i
]);
2817 raw_spin_lock_init(&rnp
->fqslock
);
2818 lockdep_set_class_and_name(&rnp
->fqslock
,
2819 &rcu_fqs_class
[i
], fqs
[i
]);
2820 rnp
->gpnum
= rsp
->gpnum
;
2821 rnp
->completed
= rsp
->completed
;
2823 rnp
->qsmaskinit
= 0;
2824 rnp
->grplo
= j
* cpustride
;
2825 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2826 if (rnp
->grphi
>= NR_CPUS
)
2827 rnp
->grphi
= NR_CPUS
- 1;
2833 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2834 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2835 rnp
->parent
= rsp
->level
[i
- 1] +
2836 j
/ rsp
->levelspread
[i
- 1];
2839 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2844 init_waitqueue_head(&rsp
->gp_wq
);
2845 rnp
= rsp
->level
[rcu_num_lvls
- 1];
2846 for_each_possible_cpu(i
) {
2847 while (i
> rnp
->grphi
)
2849 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2850 rcu_boot_init_percpu_data(i
, rsp
);
2852 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
2856 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2857 * replace the definitions in rcutree.h because those are needed to size
2858 * the ->node array in the rcu_state structure.
2860 static void __init
rcu_init_geometry(void)
2865 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
2867 /* If the compile-time values are accurate, just leave. */
2868 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
2869 nr_cpu_ids
== NR_CPUS
)
2873 * Compute number of nodes that can be handled an rcu_node tree
2874 * with the given number of levels. Setting rcu_capacity[0] makes
2875 * some of the arithmetic easier.
2877 rcu_capacity
[0] = 1;
2878 rcu_capacity
[1] = rcu_fanout_leaf
;
2879 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
2880 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
2883 * The boot-time rcu_fanout_leaf parameter is only permitted
2884 * to increase the leaf-level fanout, not decrease it. Of course,
2885 * the leaf-level fanout cannot exceed the number of bits in
2886 * the rcu_node masks. Finally, the tree must be able to accommodate
2887 * the configured number of CPUs. Complain and fall back to the
2888 * compile-time values if these limits are exceeded.
2890 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
2891 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
2892 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
2897 /* Calculate the number of rcu_nodes at each level of the tree. */
2898 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
2899 if (n
<= rcu_capacity
[i
]) {
2900 for (j
= 0; j
<= i
; j
++)
2902 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
2904 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
2909 /* Calculate the total number of rcu_node structures. */
2911 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
2912 rcu_num_nodes
+= num_rcu_lvl
[i
];
2916 void __init
rcu_init(void)
2920 rcu_bootup_announce();
2921 rcu_init_geometry();
2922 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2923 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2924 __rcu_init_preempt();
2925 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2928 * We don't need protection against CPU-hotplug here because
2929 * this is called early in boot, before either interrupts
2930 * or the scheduler are operational.
2932 cpu_notifier(rcu_cpu_notify
, 0);
2933 for_each_online_cpu(cpu
)
2934 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2935 check_cpu_stall_init();
2938 #include "rcutree_plugin.h"