2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
58 #include <trace/events/rcu.h>
62 /* Data structures. */
64 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
65 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
67 #define RCU_STATE_INITIALIZER(sname, cr) { \
68 .level = { &sname##_state.node[0] }, \
70 .fqs_state = RCU_GP_IDLE, \
71 .gpnum = 0UL - 300UL, \
72 .completed = 0UL - 300UL, \
73 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
77 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
81 struct rcu_state rcu_sched_state
=
82 RCU_STATE_INITIALIZER(rcu_sched
, call_rcu_sched
);
83 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
85 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
, call_rcu_bh
);
86 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
88 static struct rcu_state
*rcu_state
;
89 LIST_HEAD(rcu_struct_flavors
);
91 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
92 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
93 module_param(rcu_fanout_leaf
, int, 0444);
94 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
95 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
102 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
105 * The rcu_scheduler_active variable transitions from zero to one just
106 * before the first task is spawned. So when this variable is zero, RCU
107 * can assume that there is but one task, allowing RCU to (for example)
108 * optimized synchronize_sched() to a simple barrier(). When this variable
109 * is one, RCU must actually do all the hard work required to detect real
110 * grace periods. This variable is also used to suppress boot-time false
111 * positives from lockdep-RCU error checking.
113 int rcu_scheduler_active __read_mostly
;
114 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
117 * The rcu_scheduler_fully_active variable transitions from zero to one
118 * during the early_initcall() processing, which is after the scheduler
119 * is capable of creating new tasks. So RCU processing (for example,
120 * creating tasks for RCU priority boosting) must be delayed until after
121 * rcu_scheduler_fully_active transitions from zero to one. We also
122 * currently delay invocation of any RCU callbacks until after this point.
124 * It might later prove better for people registering RCU callbacks during
125 * early boot to take responsibility for these callbacks, but one step at
128 static int rcu_scheduler_fully_active __read_mostly
;
130 #ifdef CONFIG_RCU_BOOST
133 * Control variables for per-CPU and per-rcu_node kthreads. These
134 * handle all flavors of RCU.
136 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
139 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
141 #endif /* #ifdef CONFIG_RCU_BOOST */
143 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
144 static void invoke_rcu_core(void);
145 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
148 * Track the rcutorture test sequence number and the update version
149 * number within a given test. The rcutorture_testseq is incremented
150 * on every rcutorture module load and unload, so has an odd value
151 * when a test is running. The rcutorture_vernum is set to zero
152 * when rcutorture starts and is incremented on each rcutorture update.
153 * These variables enable correlating rcutorture output with the
154 * RCU tracing information.
156 unsigned long rcutorture_testseq
;
157 unsigned long rcutorture_vernum
;
160 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
161 * permit this function to be invoked without holding the root rcu_node
162 * structure's ->lock, but of course results can be subject to change.
164 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
166 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
170 * Note a quiescent state. Because we do not need to know
171 * how many quiescent states passed, just if there was at least
172 * one since the start of the grace period, this just sets a flag.
173 * The caller must have disabled preemption.
175 void rcu_sched_qs(int cpu
)
177 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
179 if (rdp
->passed_quiesce
== 0)
180 trace_rcu_grace_period("rcu_sched", rdp
->gpnum
, "cpuqs");
181 rdp
->passed_quiesce
= 1;
184 void rcu_bh_qs(int cpu
)
186 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
188 if (rdp
->passed_quiesce
== 0)
189 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
190 rdp
->passed_quiesce
= 1;
194 * Note a context switch. This is a quiescent state for RCU-sched,
195 * and requires special handling for preemptible RCU.
196 * The caller must have disabled preemption.
198 void rcu_note_context_switch(int cpu
)
200 trace_rcu_utilization("Start context switch");
202 rcu_preempt_note_context_switch(cpu
);
203 trace_rcu_utilization("End context switch");
205 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
207 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
208 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
209 .dynticks
= ATOMIC_INIT(1),
212 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
213 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
214 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
216 module_param(blimit
, long, 0444);
217 module_param(qhimark
, long, 0444);
218 module_param(qlowmark
, long, 0444);
220 static ulong jiffies_till_first_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
221 static ulong jiffies_till_next_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
223 module_param(jiffies_till_first_fqs
, ulong
, 0644);
224 module_param(jiffies_till_next_fqs
, ulong
, 0644);
226 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*));
227 static void force_quiescent_state(struct rcu_state
*rsp
);
228 static int rcu_pending(int cpu
);
231 * Return the number of RCU-sched batches processed thus far for debug & stats.
233 long rcu_batches_completed_sched(void)
235 return rcu_sched_state
.completed
;
237 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
240 * Return the number of RCU BH batches processed thus far for debug & stats.
242 long rcu_batches_completed_bh(void)
244 return rcu_bh_state
.completed
;
246 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
249 * Force a quiescent state for RCU BH.
251 void rcu_bh_force_quiescent_state(void)
253 force_quiescent_state(&rcu_bh_state
);
255 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
258 * Record the number of times rcutorture tests have been initiated and
259 * terminated. This information allows the debugfs tracing stats to be
260 * correlated to the rcutorture messages, even when the rcutorture module
261 * is being repeatedly loaded and unloaded. In other words, we cannot
262 * store this state in rcutorture itself.
264 void rcutorture_record_test_transition(void)
266 rcutorture_testseq
++;
267 rcutorture_vernum
= 0;
269 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
272 * Record the number of writer passes through the current rcutorture test.
273 * This is also used to correlate debugfs tracing stats with the rcutorture
276 void rcutorture_record_progress(unsigned long vernum
)
280 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
283 * Force a quiescent state for RCU-sched.
285 void rcu_sched_force_quiescent_state(void)
287 force_quiescent_state(&rcu_sched_state
);
289 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
292 * Does the CPU have callbacks ready to be invoked?
295 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
297 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
298 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
302 * Does the current CPU require a yet-as-unscheduled grace period?
305 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
307 struct rcu_head
**ntp
;
309 ntp
= rdp
->nxttail
[RCU_DONE_TAIL
+
310 (ACCESS_ONCE(rsp
->completed
) != rdp
->completed
)];
311 return rdp
->nxttail
[RCU_DONE_TAIL
] && ntp
&& *ntp
&&
312 !rcu_gp_in_progress(rsp
);
316 * Return the root node of the specified rcu_state structure.
318 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
320 return &rsp
->node
[0];
324 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
326 * If the new value of the ->dynticks_nesting counter now is zero,
327 * we really have entered idle, and must do the appropriate accounting.
328 * The caller must have disabled interrupts.
330 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
333 trace_rcu_dyntick("Start", oldval
, 0);
334 if (!user
&& !is_idle_task(current
)) {
335 struct task_struct
*idle
= idle_task(smp_processor_id());
337 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
338 ftrace_dump(DUMP_ORIG
);
339 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
340 current
->pid
, current
->comm
,
341 idle
->pid
, idle
->comm
); /* must be idle task! */
343 rcu_prepare_for_idle(smp_processor_id());
344 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
345 smp_mb__before_atomic_inc(); /* See above. */
346 atomic_inc(&rdtp
->dynticks
);
347 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
348 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
351 * It is illegal to enter an extended quiescent state while
352 * in an RCU read-side critical section.
354 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
355 "Illegal idle entry in RCU read-side critical section.");
356 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
357 "Illegal idle entry in RCU-bh read-side critical section.");
358 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
359 "Illegal idle entry in RCU-sched read-side critical section.");
363 * Enter an RCU extended quiescent state, which can be either the
364 * idle loop or adaptive-tickless usermode execution.
366 static void rcu_eqs_enter(bool user
)
369 struct rcu_dynticks
*rdtp
;
371 rdtp
= &__get_cpu_var(rcu_dynticks
);
372 oldval
= rdtp
->dynticks_nesting
;
373 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
374 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
375 rdtp
->dynticks_nesting
= 0;
377 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
378 rcu_eqs_enter_common(rdtp
, oldval
, user
);
382 * rcu_idle_enter - inform RCU that current CPU is entering idle
384 * Enter idle mode, in other words, -leave- the mode in which RCU
385 * read-side critical sections can occur. (Though RCU read-side
386 * critical sections can occur in irq handlers in idle, a possibility
387 * handled by irq_enter() and irq_exit().)
389 * We crowbar the ->dynticks_nesting field to zero to allow for
390 * the possibility of usermode upcalls having messed up our count
391 * of interrupt nesting level during the prior busy period.
393 void rcu_idle_enter(void)
397 local_irq_save(flags
);
398 rcu_eqs_enter(false);
399 local_irq_restore(flags
);
401 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
403 #ifdef CONFIG_RCU_USER_QS
405 * rcu_user_enter - inform RCU that we are resuming userspace.
407 * Enter RCU idle mode right before resuming userspace. No use of RCU
408 * is permitted between this call and rcu_user_exit(). This way the
409 * CPU doesn't need to maintain the tick for RCU maintenance purposes
410 * when the CPU runs in userspace.
412 void rcu_user_enter(void)
418 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
419 * after the current irq returns.
421 * This is similar to rcu_user_enter() but in the context of a non-nesting
422 * irq. After this call, RCU enters into idle mode when the interrupt
425 void rcu_user_enter_after_irq(void)
428 struct rcu_dynticks
*rdtp
;
430 local_irq_save(flags
);
431 rdtp
= &__get_cpu_var(rcu_dynticks
);
432 /* Ensure this irq is interrupting a non-idle RCU state. */
433 WARN_ON_ONCE(!(rdtp
->dynticks_nesting
& DYNTICK_TASK_MASK
));
434 rdtp
->dynticks_nesting
= 1;
435 local_irq_restore(flags
);
437 #endif /* CONFIG_RCU_USER_QS */
440 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
442 * Exit from an interrupt handler, which might possibly result in entering
443 * idle mode, in other words, leaving the mode in which read-side critical
444 * sections can occur.
446 * This code assumes that the idle loop never does anything that might
447 * result in unbalanced calls to irq_enter() and irq_exit(). If your
448 * architecture violates this assumption, RCU will give you what you
449 * deserve, good and hard. But very infrequently and irreproducibly.
451 * Use things like work queues to work around this limitation.
453 * You have been warned.
455 void rcu_irq_exit(void)
459 struct rcu_dynticks
*rdtp
;
461 local_irq_save(flags
);
462 rdtp
= &__get_cpu_var(rcu_dynticks
);
463 oldval
= rdtp
->dynticks_nesting
;
464 rdtp
->dynticks_nesting
--;
465 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
466 if (rdtp
->dynticks_nesting
)
467 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
469 rcu_eqs_enter_common(rdtp
, oldval
, true);
470 local_irq_restore(flags
);
474 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
476 * If the new value of the ->dynticks_nesting counter was previously zero,
477 * we really have exited idle, and must do the appropriate accounting.
478 * The caller must have disabled interrupts.
480 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
483 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
484 atomic_inc(&rdtp
->dynticks
);
485 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
486 smp_mb__after_atomic_inc(); /* See above. */
487 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
488 rcu_cleanup_after_idle(smp_processor_id());
489 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
490 if (!user
&& !is_idle_task(current
)) {
491 struct task_struct
*idle
= idle_task(smp_processor_id());
493 trace_rcu_dyntick("Error on exit: not idle task",
494 oldval
, rdtp
->dynticks_nesting
);
495 ftrace_dump(DUMP_ORIG
);
496 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
497 current
->pid
, current
->comm
,
498 idle
->pid
, idle
->comm
); /* must be idle task! */
503 * Exit an RCU extended quiescent state, which can be either the
504 * idle loop or adaptive-tickless usermode execution.
506 static void rcu_eqs_exit(bool user
)
508 struct rcu_dynticks
*rdtp
;
511 rdtp
= &__get_cpu_var(rcu_dynticks
);
512 oldval
= rdtp
->dynticks_nesting
;
513 WARN_ON_ONCE(oldval
< 0);
514 if (oldval
& DYNTICK_TASK_NEST_MASK
)
515 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
517 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
518 rcu_eqs_exit_common(rdtp
, oldval
, user
);
522 * rcu_idle_exit - inform RCU that current CPU is leaving idle
524 * Exit idle mode, in other words, -enter- the mode in which RCU
525 * read-side critical sections can occur.
527 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
528 * allow for the possibility of usermode upcalls messing up our count
529 * of interrupt nesting level during the busy period that is just
532 void rcu_idle_exit(void)
536 local_irq_save(flags
);
538 local_irq_restore(flags
);
540 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
542 #ifdef CONFIG_RCU_USER_QS
544 * rcu_user_exit - inform RCU that we are exiting userspace.
546 * Exit RCU idle mode while entering the kernel because it can
547 * run a RCU read side critical section anytime.
549 void rcu_user_exit(void)
555 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
556 * idle mode after the current non-nesting irq returns.
558 * This is similar to rcu_user_exit() but in the context of an irq.
559 * This is called when the irq has interrupted a userspace RCU idle mode
560 * context. When the current non-nesting interrupt returns after this call,
561 * the CPU won't restore the RCU idle mode.
563 void rcu_user_exit_after_irq(void)
566 struct rcu_dynticks
*rdtp
;
568 local_irq_save(flags
);
569 rdtp
= &__get_cpu_var(rcu_dynticks
);
570 /* Ensure we are interrupting an RCU idle mode. */
571 WARN_ON_ONCE(rdtp
->dynticks_nesting
& DYNTICK_TASK_NEST_MASK
);
572 rdtp
->dynticks_nesting
+= DYNTICK_TASK_EXIT_IDLE
;
573 local_irq_restore(flags
);
575 #endif /* CONFIG_RCU_USER_QS */
578 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
580 * Enter an interrupt handler, which might possibly result in exiting
581 * idle mode, in other words, entering the mode in which read-side critical
582 * sections can occur.
584 * Note that the Linux kernel is fully capable of entering an interrupt
585 * handler that it never exits, for example when doing upcalls to
586 * user mode! This code assumes that the idle loop never does upcalls to
587 * user mode. If your architecture does do upcalls from the idle loop (or
588 * does anything else that results in unbalanced calls to the irq_enter()
589 * and irq_exit() functions), RCU will give you what you deserve, good
590 * and hard. But very infrequently and irreproducibly.
592 * Use things like work queues to work around this limitation.
594 * You have been warned.
596 void rcu_irq_enter(void)
599 struct rcu_dynticks
*rdtp
;
602 local_irq_save(flags
);
603 rdtp
= &__get_cpu_var(rcu_dynticks
);
604 oldval
= rdtp
->dynticks_nesting
;
605 rdtp
->dynticks_nesting
++;
606 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
608 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
610 rcu_eqs_exit_common(rdtp
, oldval
, true);
611 local_irq_restore(flags
);
615 * rcu_nmi_enter - inform RCU of entry to NMI context
617 * If the CPU was idle with dynamic ticks active, and there is no
618 * irq handler running, this updates rdtp->dynticks_nmi to let the
619 * RCU grace-period handling know that the CPU is active.
621 void rcu_nmi_enter(void)
623 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
625 if (rdtp
->dynticks_nmi_nesting
== 0 &&
626 (atomic_read(&rdtp
->dynticks
) & 0x1))
628 rdtp
->dynticks_nmi_nesting
++;
629 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
630 atomic_inc(&rdtp
->dynticks
);
631 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
632 smp_mb__after_atomic_inc(); /* See above. */
633 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
637 * rcu_nmi_exit - inform RCU of exit from NMI context
639 * If the CPU was idle with dynamic ticks active, and there is no
640 * irq handler running, this updates rdtp->dynticks_nmi to let the
641 * RCU grace-period handling know that the CPU is no longer active.
643 void rcu_nmi_exit(void)
645 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
647 if (rdtp
->dynticks_nmi_nesting
== 0 ||
648 --rdtp
->dynticks_nmi_nesting
!= 0)
650 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
651 smp_mb__before_atomic_inc(); /* See above. */
652 atomic_inc(&rdtp
->dynticks
);
653 smp_mb__after_atomic_inc(); /* Force delay to next write. */
654 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
658 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
660 * If the current CPU is in its idle loop and is neither in an interrupt
661 * or NMI handler, return true.
663 int rcu_is_cpu_idle(void)
668 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
672 EXPORT_SYMBOL(rcu_is_cpu_idle
);
674 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
677 * Is the current CPU online? Disable preemption to avoid false positives
678 * that could otherwise happen due to the current CPU number being sampled,
679 * this task being preempted, its old CPU being taken offline, resuming
680 * on some other CPU, then determining that its old CPU is now offline.
681 * It is OK to use RCU on an offline processor during initial boot, hence
682 * the check for rcu_scheduler_fully_active. Note also that it is OK
683 * for a CPU coming online to use RCU for one jiffy prior to marking itself
684 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
685 * offline to continue to use RCU for one jiffy after marking itself
686 * offline in the cpu_online_mask. This leniency is necessary given the
687 * non-atomic nature of the online and offline processing, for example,
688 * the fact that a CPU enters the scheduler after completing the CPU_DYING
691 * This is also why RCU internally marks CPUs online during the
692 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
694 * Disable checking if in an NMI handler because we cannot safely report
695 * errors from NMI handlers anyway.
697 bool rcu_lockdep_current_cpu_online(void)
699 struct rcu_data
*rdp
;
700 struct rcu_node
*rnp
;
706 rdp
= &__get_cpu_var(rcu_sched_data
);
708 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
709 !rcu_scheduler_fully_active
;
713 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
715 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
718 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
720 * If the current CPU is idle or running at a first-level (not nested)
721 * interrupt from idle, return true. The caller must have at least
722 * disabled preemption.
724 int rcu_is_cpu_rrupt_from_idle(void)
726 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
730 * Snapshot the specified CPU's dynticks counter so that we can later
731 * credit them with an implicit quiescent state. Return 1 if this CPU
732 * is in dynticks idle mode, which is an extended quiescent state.
734 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
736 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
737 return (rdp
->dynticks_snap
& 0x1) == 0;
741 * Return true if the specified CPU has passed through a quiescent
742 * state by virtue of being in or having passed through an dynticks
743 * idle state since the last call to dyntick_save_progress_counter()
744 * for this same CPU, or by virtue of having been offline.
746 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
751 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
752 snap
= (unsigned int)rdp
->dynticks_snap
;
755 * If the CPU passed through or entered a dynticks idle phase with
756 * no active irq/NMI handlers, then we can safely pretend that the CPU
757 * already acknowledged the request to pass through a quiescent
758 * state. Either way, that CPU cannot possibly be in an RCU
759 * read-side critical section that started before the beginning
760 * of the current RCU grace period.
762 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
763 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
769 * Check for the CPU being offline, but only if the grace period
770 * is old enough. We don't need to worry about the CPU changing
771 * state: If we see it offline even once, it has been through a
774 * The reason for insisting that the grace period be at least
775 * one jiffy old is that CPUs that are not quite online and that
776 * have just gone offline can still execute RCU read-side critical
779 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
780 return 0; /* Grace period is not old enough. */
782 if (cpu_is_offline(rdp
->cpu
)) {
783 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
790 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
792 rsp
->gp_start
= jiffies
;
793 rsp
->jiffies_stall
= jiffies
+ rcu_jiffies_till_stall_check();
797 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
798 * for architectures that do not implement trigger_all_cpu_backtrace().
799 * The NMI-triggered stack traces are more accurate because they are
800 * printed by the target CPU.
802 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
806 struct rcu_node
*rnp
;
808 rcu_for_each_leaf_node(rsp
, rnp
) {
809 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
810 if (rnp
->qsmask
!= 0) {
811 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
812 if (rnp
->qsmask
& (1UL << cpu
))
813 dump_cpu_task(rnp
->grplo
+ cpu
);
815 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
819 static void print_other_cpu_stall(struct rcu_state
*rsp
)
825 struct rcu_node
*rnp
= rcu_get_root(rsp
);
828 /* Only let one CPU complain about others per time interval. */
830 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
831 delta
= jiffies
- rsp
->jiffies_stall
;
832 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
833 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
836 rsp
->jiffies_stall
= jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
837 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
840 * OK, time to rat on our buddy...
841 * See Documentation/RCU/stallwarn.txt for info on how to debug
842 * RCU CPU stall warnings.
844 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
846 print_cpu_stall_info_begin();
847 rcu_for_each_leaf_node(rsp
, rnp
) {
848 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
849 ndetected
+= rcu_print_task_stall(rnp
);
850 if (rnp
->qsmask
!= 0) {
851 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
852 if (rnp
->qsmask
& (1UL << cpu
)) {
853 print_cpu_stall_info(rsp
,
858 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
862 * Now rat on any tasks that got kicked up to the root rcu_node
863 * due to CPU offlining.
865 rnp
= rcu_get_root(rsp
);
866 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
867 ndetected
+= rcu_print_task_stall(rnp
);
868 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
870 print_cpu_stall_info_end();
871 for_each_possible_cpu(cpu
)
872 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
873 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
874 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
875 rsp
->gpnum
, rsp
->completed
, totqlen
);
877 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
878 else if (!trigger_all_cpu_backtrace())
879 rcu_dump_cpu_stacks(rsp
);
881 /* Complain about tasks blocking the grace period. */
883 rcu_print_detail_task_stall(rsp
);
885 force_quiescent_state(rsp
); /* Kick them all. */
888 static void print_cpu_stall(struct rcu_state
*rsp
)
892 struct rcu_node
*rnp
= rcu_get_root(rsp
);
896 * OK, time to rat on ourselves...
897 * See Documentation/RCU/stallwarn.txt for info on how to debug
898 * RCU CPU stall warnings.
900 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
901 print_cpu_stall_info_begin();
902 print_cpu_stall_info(rsp
, smp_processor_id());
903 print_cpu_stall_info_end();
904 for_each_possible_cpu(cpu
)
905 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
906 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
907 jiffies
- rsp
->gp_start
, rsp
->gpnum
, rsp
->completed
, totqlen
);
908 if (!trigger_all_cpu_backtrace())
911 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
912 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
913 rsp
->jiffies_stall
= jiffies
+
914 3 * rcu_jiffies_till_stall_check() + 3;
915 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
917 set_need_resched(); /* kick ourselves to get things going. */
920 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
924 struct rcu_node
*rnp
;
926 if (rcu_cpu_stall_suppress
)
928 j
= ACCESS_ONCE(jiffies
);
929 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
931 if (rcu_gp_in_progress(rsp
) &&
932 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
934 /* We haven't checked in, so go dump stack. */
935 print_cpu_stall(rsp
);
937 } else if (rcu_gp_in_progress(rsp
) &&
938 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
940 /* They had a few time units to dump stack, so complain. */
941 print_other_cpu_stall(rsp
);
946 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
948 * Set the stall-warning timeout way off into the future, thus preventing
949 * any RCU CPU stall-warning messages from appearing in the current set of
952 * The caller must disable hard irqs.
954 void rcu_cpu_stall_reset(void)
956 struct rcu_state
*rsp
;
958 for_each_rcu_flavor(rsp
)
959 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
963 * Update CPU-local rcu_data state to record the newly noticed grace period.
964 * This is used both when we started the grace period and when we notice
965 * that someone else started the grace period. The caller must hold the
966 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
967 * and must have irqs disabled.
969 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
971 if (rdp
->gpnum
!= rnp
->gpnum
) {
973 * If the current grace period is waiting for this CPU,
974 * set up to detect a quiescent state, otherwise don't
975 * go looking for one.
977 rdp
->gpnum
= rnp
->gpnum
;
978 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
979 rdp
->passed_quiesce
= 0;
980 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
981 zero_cpu_stall_ticks(rdp
);
985 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
988 struct rcu_node
*rnp
;
990 local_irq_save(flags
);
992 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
993 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
994 local_irq_restore(flags
);
997 __note_new_gpnum(rsp
, rnp
, rdp
);
998 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1002 * Did someone else start a new RCU grace period start since we last
1003 * checked? Update local state appropriately if so. Must be called
1004 * on the CPU corresponding to rdp.
1007 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1009 unsigned long flags
;
1012 local_irq_save(flags
);
1013 if (rdp
->gpnum
!= rsp
->gpnum
) {
1014 note_new_gpnum(rsp
, rdp
);
1017 local_irq_restore(flags
);
1022 * Initialize the specified rcu_data structure's callback list to empty.
1024 static void init_callback_list(struct rcu_data
*rdp
)
1028 rdp
->nxtlist
= NULL
;
1029 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1030 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1031 init_nocb_callback_list(rdp
);
1035 * Advance this CPU's callbacks, but only if the current grace period
1036 * has ended. This may be called only from the CPU to whom the rdp
1037 * belongs. In addition, the corresponding leaf rcu_node structure's
1038 * ->lock must be held by the caller, with irqs disabled.
1041 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1043 /* Did another grace period end? */
1044 if (rdp
->completed
!= rnp
->completed
) {
1046 /* Advance callbacks. No harm if list empty. */
1047 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
1048 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
1049 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1051 /* Remember that we saw this grace-period completion. */
1052 rdp
->completed
= rnp
->completed
;
1053 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
1056 * If we were in an extended quiescent state, we may have
1057 * missed some grace periods that others CPUs handled on
1058 * our behalf. Catch up with this state to avoid noting
1059 * spurious new grace periods. If another grace period
1060 * has started, then rnp->gpnum will have advanced, so
1061 * we will detect this later on. Of course, any quiescent
1062 * states we found for the old GP are now invalid.
1064 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
)) {
1065 rdp
->gpnum
= rdp
->completed
;
1066 rdp
->passed_quiesce
= 0;
1070 * If RCU does not need a quiescent state from this CPU,
1071 * then make sure that this CPU doesn't go looking for one.
1073 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1074 rdp
->qs_pending
= 0;
1079 * Advance this CPU's callbacks, but only if the current grace period
1080 * has ended. This may be called only from the CPU to whom the rdp
1084 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1086 unsigned long flags
;
1087 struct rcu_node
*rnp
;
1089 local_irq_save(flags
);
1091 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1092 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1093 local_irq_restore(flags
);
1096 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1097 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1101 * Do per-CPU grace-period initialization for running CPU. The caller
1102 * must hold the lock of the leaf rcu_node structure corresponding to
1106 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1108 /* Prior grace period ended, so advance callbacks for current CPU. */
1109 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1111 /* Set state so that this CPU will detect the next quiescent state. */
1112 __note_new_gpnum(rsp
, rnp
, rdp
);
1116 * Initialize a new grace period.
1118 static int rcu_gp_init(struct rcu_state
*rsp
)
1120 struct rcu_data
*rdp
;
1121 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1123 raw_spin_lock_irq(&rnp
->lock
);
1124 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1126 if (rcu_gp_in_progress(rsp
)) {
1127 /* Grace period already in progress, don't start another. */
1128 raw_spin_unlock_irq(&rnp
->lock
);
1132 /* Advance to a new grace period and initialize state. */
1134 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1135 record_gp_stall_check_time(rsp
);
1136 raw_spin_unlock_irq(&rnp
->lock
);
1138 /* Exclude any concurrent CPU-hotplug operations. */
1139 mutex_lock(&rsp
->onoff_mutex
);
1142 * Set the quiescent-state-needed bits in all the rcu_node
1143 * structures for all currently online CPUs in breadth-first order,
1144 * starting from the root rcu_node structure, relying on the layout
1145 * of the tree within the rsp->node[] array. Note that other CPUs
1146 * will access only the leaves of the hierarchy, thus seeing that no
1147 * grace period is in progress, at least until the corresponding
1148 * leaf node has been initialized. In addition, we have excluded
1149 * CPU-hotplug operations.
1151 * The grace period cannot complete until the initialization
1152 * process finishes, because this kthread handles both.
1154 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1155 raw_spin_lock_irq(&rnp
->lock
);
1156 rdp
= this_cpu_ptr(rsp
->rda
);
1157 rcu_preempt_check_blocked_tasks(rnp
);
1158 rnp
->qsmask
= rnp
->qsmaskinit
;
1159 rnp
->gpnum
= rsp
->gpnum
;
1160 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1161 rnp
->completed
= rsp
->completed
;
1162 if (rnp
== rdp
->mynode
)
1163 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1164 rcu_preempt_boost_start_gp(rnp
);
1165 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1166 rnp
->level
, rnp
->grplo
,
1167 rnp
->grphi
, rnp
->qsmask
);
1168 raw_spin_unlock_irq(&rnp
->lock
);
1169 #ifdef CONFIG_PROVE_RCU_DELAY
1170 if ((random32() % (rcu_num_nodes
* 8)) == 0)
1171 schedule_timeout_uninterruptible(2);
1172 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1176 mutex_unlock(&rsp
->onoff_mutex
);
1181 * Do one round of quiescent-state forcing.
1183 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1185 int fqs_state
= fqs_state_in
;
1186 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1189 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1190 /* Collect dyntick-idle snapshots. */
1191 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1192 fqs_state
= RCU_FORCE_QS
;
1194 /* Handle dyntick-idle and offline CPUs. */
1195 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1197 /* Clear flag to prevent immediate re-entry. */
1198 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1199 raw_spin_lock_irq(&rnp
->lock
);
1200 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1201 raw_spin_unlock_irq(&rnp
->lock
);
1207 * Clean up after the old grace period.
1209 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1211 unsigned long gp_duration
;
1212 struct rcu_data
*rdp
;
1213 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1215 raw_spin_lock_irq(&rnp
->lock
);
1216 gp_duration
= jiffies
- rsp
->gp_start
;
1217 if (gp_duration
> rsp
->gp_max
)
1218 rsp
->gp_max
= gp_duration
;
1221 * We know the grace period is complete, but to everyone else
1222 * it appears to still be ongoing. But it is also the case
1223 * that to everyone else it looks like there is nothing that
1224 * they can do to advance the grace period. It is therefore
1225 * safe for us to drop the lock in order to mark the grace
1226 * period as completed in all of the rcu_node structures.
1228 raw_spin_unlock_irq(&rnp
->lock
);
1231 * Propagate new ->completed value to rcu_node structures so
1232 * that other CPUs don't have to wait until the start of the next
1233 * grace period to process their callbacks. This also avoids
1234 * some nasty RCU grace-period initialization races by forcing
1235 * the end of the current grace period to be completely recorded in
1236 * all of the rcu_node structures before the beginning of the next
1237 * grace period is recorded in any of the rcu_node structures.
1239 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1240 raw_spin_lock_irq(&rnp
->lock
);
1241 rnp
->completed
= rsp
->gpnum
;
1242 raw_spin_unlock_irq(&rnp
->lock
);
1245 rnp
= rcu_get_root(rsp
);
1246 raw_spin_lock_irq(&rnp
->lock
);
1248 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1249 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1250 rsp
->fqs_state
= RCU_GP_IDLE
;
1251 rdp
= this_cpu_ptr(rsp
->rda
);
1252 if (cpu_needs_another_gp(rsp
, rdp
))
1254 raw_spin_unlock_irq(&rnp
->lock
);
1258 * Body of kthread that handles grace periods.
1260 static int __noreturn
rcu_gp_kthread(void *arg
)
1265 struct rcu_state
*rsp
= arg
;
1266 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1270 /* Handle grace-period start. */
1272 wait_event_interruptible(rsp
->gp_wq
,
1275 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1279 flush_signals(current
);
1282 /* Handle quiescent-state forcing. */
1283 fqs_state
= RCU_SAVE_DYNTICK
;
1284 j
= jiffies_till_first_fqs
;
1287 jiffies_till_first_fqs
= HZ
;
1290 rsp
->jiffies_force_qs
= jiffies
+ j
;
1291 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1292 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1293 (!ACCESS_ONCE(rnp
->qsmask
) &&
1294 !rcu_preempt_blocked_readers_cgp(rnp
)),
1296 /* If grace period done, leave loop. */
1297 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1298 !rcu_preempt_blocked_readers_cgp(rnp
))
1300 /* If time for quiescent-state forcing, do it. */
1301 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1302 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1305 /* Deal with stray signal. */
1307 flush_signals(current
);
1309 j
= jiffies_till_next_fqs
;
1312 jiffies_till_next_fqs
= HZ
;
1315 jiffies_till_next_fqs
= 1;
1319 /* Handle grace-period end. */
1320 rcu_gp_cleanup(rsp
);
1325 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1326 * in preparation for detecting the next grace period. The caller must hold
1327 * the root node's ->lock, which is released before return. Hard irqs must
1330 * Note that it is legal for a dying CPU (which is marked as offline) to
1331 * invoke this function. This can happen when the dying CPU reports its
1335 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1336 __releases(rcu_get_root(rsp
)->lock
)
1338 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1339 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1341 if (!rsp
->gp_kthread
||
1342 !cpu_needs_another_gp(rsp
, rdp
)) {
1344 * Either we have not yet spawned the grace-period
1345 * task, this CPU does not need another grace period,
1346 * or a grace period is already in progress.
1347 * Either way, don't start a new grace period.
1349 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1354 * Because there is no grace period in progress right now,
1355 * any callbacks we have up to this point will be satisfied
1356 * by the next grace period. So promote all callbacks to be
1357 * handled after the end of the next grace period. If the
1358 * CPU is not yet aware of the end of the previous grace period,
1359 * we need to allow for the callback advancement that will
1360 * occur when it does become aware. Deadlock prevents us from
1361 * making it aware at this point: We cannot acquire a leaf
1362 * rcu_node ->lock while holding the root rcu_node ->lock.
1364 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1365 if (rdp
->completed
== rsp
->completed
)
1366 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1368 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1369 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
1371 /* Ensure that CPU is aware of completion of last grace period. */
1372 rcu_process_gp_end(rsp
, rdp
);
1373 local_irq_restore(flags
);
1375 /* Wake up rcu_gp_kthread() to start the grace period. */
1376 wake_up(&rsp
->gp_wq
);
1380 * Report a full set of quiescent states to the specified rcu_state
1381 * data structure. This involves cleaning up after the prior grace
1382 * period and letting rcu_start_gp() start up the next grace period
1383 * if one is needed. Note that the caller must hold rnp->lock, as
1384 * required by rcu_start_gp(), which will release it.
1386 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1387 __releases(rcu_get_root(rsp
)->lock
)
1389 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1390 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1391 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1395 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1396 * Allows quiescent states for a group of CPUs to be reported at one go
1397 * to the specified rcu_node structure, though all the CPUs in the group
1398 * must be represented by the same rcu_node structure (which need not be
1399 * a leaf rcu_node structure, though it often will be). That structure's
1400 * lock must be held upon entry, and it is released before return.
1403 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1404 struct rcu_node
*rnp
, unsigned long flags
)
1405 __releases(rnp
->lock
)
1407 struct rcu_node
*rnp_c
;
1409 /* Walk up the rcu_node hierarchy. */
1411 if (!(rnp
->qsmask
& mask
)) {
1413 /* Our bit has already been cleared, so done. */
1414 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1417 rnp
->qsmask
&= ~mask
;
1418 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1419 mask
, rnp
->qsmask
, rnp
->level
,
1420 rnp
->grplo
, rnp
->grphi
,
1422 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1424 /* Other bits still set at this level, so done. */
1425 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1428 mask
= rnp
->grpmask
;
1429 if (rnp
->parent
== NULL
) {
1431 /* No more levels. Exit loop holding root lock. */
1435 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1438 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1439 WARN_ON_ONCE(rnp_c
->qsmask
);
1443 * Get here if we are the last CPU to pass through a quiescent
1444 * state for this grace period. Invoke rcu_report_qs_rsp()
1445 * to clean up and start the next grace period if one is needed.
1447 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1451 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1452 * structure. This must be either called from the specified CPU, or
1453 * called when the specified CPU is known to be offline (and when it is
1454 * also known that no other CPU is concurrently trying to help the offline
1455 * CPU). The lastcomp argument is used to make sure we are still in the
1456 * grace period of interest. We don't want to end the current grace period
1457 * based on quiescent states detected in an earlier grace period!
1460 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1462 unsigned long flags
;
1464 struct rcu_node
*rnp
;
1467 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1468 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1469 rnp
->completed
== rnp
->gpnum
) {
1472 * The grace period in which this quiescent state was
1473 * recorded has ended, so don't report it upwards.
1474 * We will instead need a new quiescent state that lies
1475 * within the current grace period.
1477 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1478 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1481 mask
= rdp
->grpmask
;
1482 if ((rnp
->qsmask
& mask
) == 0) {
1483 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1485 rdp
->qs_pending
= 0;
1488 * This GP can't end until cpu checks in, so all of our
1489 * callbacks can be processed during the next GP.
1491 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1493 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1498 * Check to see if there is a new grace period of which this CPU
1499 * is not yet aware, and if so, set up local rcu_data state for it.
1500 * Otherwise, see if this CPU has just passed through its first
1501 * quiescent state for this grace period, and record that fact if so.
1504 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1506 /* If there is now a new grace period, record and return. */
1507 if (check_for_new_grace_period(rsp
, rdp
))
1511 * Does this CPU still need to do its part for current grace period?
1512 * If no, return and let the other CPUs do their part as well.
1514 if (!rdp
->qs_pending
)
1518 * Was there a quiescent state since the beginning of the grace
1519 * period? If no, then exit and wait for the next call.
1521 if (!rdp
->passed_quiesce
)
1525 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1528 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1531 #ifdef CONFIG_HOTPLUG_CPU
1534 * Send the specified CPU's RCU callbacks to the orphanage. The
1535 * specified CPU must be offline, and the caller must hold the
1539 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1540 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1542 /* No-CBs CPUs do not have orphanable callbacks. */
1543 if (is_nocb_cpu(rdp
->cpu
))
1547 * Orphan the callbacks. First adjust the counts. This is safe
1548 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1549 * cannot be running now. Thus no memory barrier is required.
1551 if (rdp
->nxtlist
!= NULL
) {
1552 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1553 rsp
->qlen
+= rdp
->qlen
;
1554 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1556 ACCESS_ONCE(rdp
->qlen
) = 0;
1560 * Next, move those callbacks still needing a grace period to
1561 * the orphanage, where some other CPU will pick them up.
1562 * Some of the callbacks might have gone partway through a grace
1563 * period, but that is too bad. They get to start over because we
1564 * cannot assume that grace periods are synchronized across CPUs.
1565 * We don't bother updating the ->nxttail[] array yet, instead
1566 * we just reset the whole thing later on.
1568 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1569 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1570 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1571 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1575 * Then move the ready-to-invoke callbacks to the orphanage,
1576 * where some other CPU will pick them up. These will not be
1577 * required to pass though another grace period: They are done.
1579 if (rdp
->nxtlist
!= NULL
) {
1580 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1581 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1584 /* Finally, initialize the rcu_data structure's list to empty. */
1585 init_callback_list(rdp
);
1589 * Adopt the RCU callbacks from the specified rcu_state structure's
1590 * orphanage. The caller must hold the ->orphan_lock.
1592 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1595 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1597 /* No-CBs CPUs are handled specially. */
1598 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
))
1601 /* Do the accounting first. */
1602 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1603 rdp
->qlen
+= rsp
->qlen
;
1604 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1605 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1606 rcu_idle_count_callbacks_posted();
1611 * We do not need a memory barrier here because the only way we
1612 * can get here if there is an rcu_barrier() in flight is if
1613 * we are the task doing the rcu_barrier().
1616 /* First adopt the ready-to-invoke callbacks. */
1617 if (rsp
->orphan_donelist
!= NULL
) {
1618 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1619 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1620 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1621 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1622 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1623 rsp
->orphan_donelist
= NULL
;
1624 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1627 /* And then adopt the callbacks that still need a grace period. */
1628 if (rsp
->orphan_nxtlist
!= NULL
) {
1629 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1630 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1631 rsp
->orphan_nxtlist
= NULL
;
1632 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1637 * Trace the fact that this CPU is going offline.
1639 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1641 RCU_TRACE(unsigned long mask
);
1642 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1643 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1645 RCU_TRACE(mask
= rdp
->grpmask
);
1646 trace_rcu_grace_period(rsp
->name
,
1647 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1652 * The CPU has been completely removed, and some other CPU is reporting
1653 * this fact from process context. Do the remainder of the cleanup,
1654 * including orphaning the outgoing CPU's RCU callbacks, and also
1655 * adopting them. There can only be one CPU hotplug operation at a time,
1656 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1658 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1660 unsigned long flags
;
1662 int need_report
= 0;
1663 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1664 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1666 /* Adjust any no-longer-needed kthreads. */
1667 rcu_boost_kthread_setaffinity(rnp
, -1);
1669 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1671 /* Exclude any attempts to start a new grace period. */
1672 mutex_lock(&rsp
->onoff_mutex
);
1673 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
1675 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1676 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1677 rcu_adopt_orphan_cbs(rsp
);
1679 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1680 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1682 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1683 rnp
->qsmaskinit
&= ~mask
;
1684 if (rnp
->qsmaskinit
!= 0) {
1685 if (rnp
!= rdp
->mynode
)
1686 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1689 if (rnp
== rdp
->mynode
)
1690 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1692 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1693 mask
= rnp
->grpmask
;
1695 } while (rnp
!= NULL
);
1698 * We still hold the leaf rcu_node structure lock here, and
1699 * irqs are still disabled. The reason for this subterfuge is
1700 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1701 * held leads to deadlock.
1703 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
1705 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1706 rcu_report_unblock_qs_rnp(rnp
, flags
);
1708 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1709 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1710 rcu_report_exp_rnp(rsp
, rnp
, true);
1711 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1712 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1713 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1714 init_callback_list(rdp
);
1715 /* Disallow further callbacks on this CPU. */
1716 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1717 mutex_unlock(&rsp
->onoff_mutex
);
1720 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1722 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1726 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1730 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1733 * Invoke any RCU callbacks that have made it to the end of their grace
1734 * period. Thottle as specified by rdp->blimit.
1736 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1738 unsigned long flags
;
1739 struct rcu_head
*next
, *list
, **tail
;
1740 long bl
, count
, count_lazy
;
1743 /* If no callbacks are ready, just return.*/
1744 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1745 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1746 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1747 need_resched(), is_idle_task(current
),
1748 rcu_is_callbacks_kthread());
1753 * Extract the list of ready callbacks, disabling to prevent
1754 * races with call_rcu() from interrupt handlers.
1756 local_irq_save(flags
);
1757 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1759 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1760 list
= rdp
->nxtlist
;
1761 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1762 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1763 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1764 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1765 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1766 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1767 local_irq_restore(flags
);
1769 /* Invoke callbacks. */
1770 count
= count_lazy
= 0;
1774 debug_rcu_head_unqueue(list
);
1775 if (__rcu_reclaim(rsp
->name
, list
))
1778 /* Stop only if limit reached and CPU has something to do. */
1779 if (++count
>= bl
&&
1781 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1785 local_irq_save(flags
);
1786 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1787 is_idle_task(current
),
1788 rcu_is_callbacks_kthread());
1790 /* Update count, and requeue any remaining callbacks. */
1792 *tail
= rdp
->nxtlist
;
1793 rdp
->nxtlist
= list
;
1794 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1795 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1796 rdp
->nxttail
[i
] = tail
;
1800 smp_mb(); /* List handling before counting for rcu_barrier(). */
1801 rdp
->qlen_lazy
-= count_lazy
;
1802 ACCESS_ONCE(rdp
->qlen
) -= count
;
1803 rdp
->n_cbs_invoked
+= count
;
1805 /* Reinstate batch limit if we have worked down the excess. */
1806 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1807 rdp
->blimit
= blimit
;
1809 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1810 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1811 rdp
->qlen_last_fqs_check
= 0;
1812 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1813 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1814 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1815 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
1817 local_irq_restore(flags
);
1819 /* Re-invoke RCU core processing if there are callbacks remaining. */
1820 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1825 * Check to see if this CPU is in a non-context-switch quiescent state
1826 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1827 * Also schedule RCU core processing.
1829 * This function must be called from hardirq context. It is normally
1830 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1831 * false, there is no point in invoking rcu_check_callbacks().
1833 void rcu_check_callbacks(int cpu
, int user
)
1835 trace_rcu_utilization("Start scheduler-tick");
1836 increment_cpu_stall_ticks();
1837 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1840 * Get here if this CPU took its interrupt from user
1841 * mode or from the idle loop, and if this is not a
1842 * nested interrupt. In this case, the CPU is in
1843 * a quiescent state, so note it.
1845 * No memory barrier is required here because both
1846 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1847 * variables that other CPUs neither access nor modify,
1848 * at least not while the corresponding CPU is online.
1854 } else if (!in_softirq()) {
1857 * Get here if this CPU did not take its interrupt from
1858 * softirq, in other words, if it is not interrupting
1859 * a rcu_bh read-side critical section. This is an _bh
1860 * critical section, so note it.
1865 rcu_preempt_check_callbacks(cpu
);
1866 if (rcu_pending(cpu
))
1868 trace_rcu_utilization("End scheduler-tick");
1872 * Scan the leaf rcu_node structures, processing dyntick state for any that
1873 * have not yet encountered a quiescent state, using the function specified.
1874 * Also initiate boosting for any threads blocked on the root rcu_node.
1876 * The caller must have suppressed start of new grace periods.
1878 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1882 unsigned long flags
;
1884 struct rcu_node
*rnp
;
1886 rcu_for_each_leaf_node(rsp
, rnp
) {
1889 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1890 if (!rcu_gp_in_progress(rsp
)) {
1891 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1894 if (rnp
->qsmask
== 0) {
1895 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1900 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1901 if ((rnp
->qsmask
& bit
) != 0 &&
1902 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1907 /* rcu_report_qs_rnp() releases rnp->lock. */
1908 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1911 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1913 rnp
= rcu_get_root(rsp
);
1914 if (rnp
->qsmask
== 0) {
1915 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1916 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1921 * Force quiescent states on reluctant CPUs, and also detect which
1922 * CPUs are in dyntick-idle mode.
1924 static void force_quiescent_state(struct rcu_state
*rsp
)
1926 unsigned long flags
;
1928 struct rcu_node
*rnp
;
1929 struct rcu_node
*rnp_old
= NULL
;
1931 /* Funnel through hierarchy to reduce memory contention. */
1932 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
1933 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
1934 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
1935 !raw_spin_trylock(&rnp
->fqslock
);
1936 if (rnp_old
!= NULL
)
1937 raw_spin_unlock(&rnp_old
->fqslock
);
1939 rsp
->n_force_qs_lh
++;
1944 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1946 /* Reached the root of the rcu_node tree, acquire lock. */
1947 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
1948 raw_spin_unlock(&rnp_old
->fqslock
);
1949 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1950 rsp
->n_force_qs_lh
++;
1951 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1952 return; /* Someone beat us to it. */
1954 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
1955 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1956 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1960 * This does the RCU core processing work for the specified rcu_state
1961 * and rcu_data structures. This may be called only from the CPU to
1962 * whom the rdp belongs.
1965 __rcu_process_callbacks(struct rcu_state
*rsp
)
1967 unsigned long flags
;
1968 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1970 WARN_ON_ONCE(rdp
->beenonline
== 0);
1973 * Advance callbacks in response to end of earlier grace
1974 * period that some other CPU ended.
1976 rcu_process_gp_end(rsp
, rdp
);
1978 /* Update RCU state based on any recent quiescent states. */
1979 rcu_check_quiescent_state(rsp
, rdp
);
1981 /* Does this CPU require a not-yet-started grace period? */
1982 if (cpu_needs_another_gp(rsp
, rdp
)) {
1983 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1984 rcu_start_gp(rsp
, flags
); /* releases above lock */
1987 /* If there are callbacks ready, invoke them. */
1988 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1989 invoke_rcu_callbacks(rsp
, rdp
);
1993 * Do RCU core processing for the current CPU.
1995 static void rcu_process_callbacks(struct softirq_action
*unused
)
1997 struct rcu_state
*rsp
;
1999 if (cpu_is_offline(smp_processor_id()))
2001 trace_rcu_utilization("Start RCU core");
2002 for_each_rcu_flavor(rsp
)
2003 __rcu_process_callbacks(rsp
);
2004 trace_rcu_utilization("End RCU core");
2008 * Schedule RCU callback invocation. If the specified type of RCU
2009 * does not support RCU priority boosting, just do a direct call,
2010 * otherwise wake up the per-CPU kernel kthread. Note that because we
2011 * are running on the current CPU with interrupts disabled, the
2012 * rcu_cpu_kthread_task cannot disappear out from under us.
2014 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2016 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2018 if (likely(!rsp
->boost
)) {
2019 rcu_do_batch(rsp
, rdp
);
2022 invoke_rcu_callbacks_kthread();
2025 static void invoke_rcu_core(void)
2027 raise_softirq(RCU_SOFTIRQ
);
2031 * Handle any core-RCU processing required by a call_rcu() invocation.
2033 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2034 struct rcu_head
*head
, unsigned long flags
)
2037 * If called from an extended quiescent state, invoke the RCU
2038 * core in order to force a re-evaluation of RCU's idleness.
2040 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2043 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2044 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2048 * Force the grace period if too many callbacks or too long waiting.
2049 * Enforce hysteresis, and don't invoke force_quiescent_state()
2050 * if some other CPU has recently done so. Also, don't bother
2051 * invoking force_quiescent_state() if the newly enqueued callback
2052 * is the only one waiting for a grace period to complete.
2054 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2056 /* Are we ignoring a completed grace period? */
2057 rcu_process_gp_end(rsp
, rdp
);
2058 check_for_new_grace_period(rsp
, rdp
);
2060 /* Start a new grace period if one not already started. */
2061 if (!rcu_gp_in_progress(rsp
)) {
2062 unsigned long nestflag
;
2063 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2065 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
2066 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
2068 /* Give the grace period a kick. */
2069 rdp
->blimit
= LONG_MAX
;
2070 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2071 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2072 force_quiescent_state(rsp
);
2073 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2074 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2080 * Helper function for call_rcu() and friends. The cpu argument will
2081 * normally be -1, indicating "currently running CPU". It may specify
2082 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2083 * is expected to specify a CPU.
2086 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2087 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2089 unsigned long flags
;
2090 struct rcu_data
*rdp
;
2092 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2093 debug_rcu_head_queue(head
);
2098 * Opportunistically note grace-period endings and beginnings.
2099 * Note that we might see a beginning right after we see an
2100 * end, but never vice versa, since this CPU has to pass through
2101 * a quiescent state betweentimes.
2103 local_irq_save(flags
);
2104 rdp
= this_cpu_ptr(rsp
->rda
);
2106 /* Add the callback to our list. */
2107 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2111 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2112 offline
= !__call_rcu_nocb(rdp
, head
, lazy
);
2113 WARN_ON_ONCE(offline
);
2114 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2115 local_irq_restore(flags
);
2118 ACCESS_ONCE(rdp
->qlen
)++;
2122 rcu_idle_count_callbacks_posted();
2123 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2124 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2125 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2127 if (__is_kfree_rcu_offset((unsigned long)func
))
2128 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2129 rdp
->qlen_lazy
, rdp
->qlen
);
2131 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2133 /* Go handle any RCU core processing required. */
2134 __call_rcu_core(rsp
, rdp
, head
, flags
);
2135 local_irq_restore(flags
);
2139 * Queue an RCU-sched callback for invocation after a grace period.
2141 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2143 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2145 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2148 * Queue an RCU callback for invocation after a quicker grace period.
2150 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2152 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2154 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2157 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2158 * any blocking grace-period wait automatically implies a grace period
2159 * if there is only one CPU online at any point time during execution
2160 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2161 * occasionally incorrectly indicate that there are multiple CPUs online
2162 * when there was in fact only one the whole time, as this just adds
2163 * some overhead: RCU still operates correctly.
2165 static inline int rcu_blocking_is_gp(void)
2169 might_sleep(); /* Check for RCU read-side critical section. */
2171 ret
= num_online_cpus() <= 1;
2177 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2179 * Control will return to the caller some time after a full rcu-sched
2180 * grace period has elapsed, in other words after all currently executing
2181 * rcu-sched read-side critical sections have completed. These read-side
2182 * critical sections are delimited by rcu_read_lock_sched() and
2183 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2184 * local_irq_disable(), and so on may be used in place of
2185 * rcu_read_lock_sched().
2187 * This means that all preempt_disable code sequences, including NMI and
2188 * non-threaded hardware-interrupt handlers, in progress on entry will
2189 * have completed before this primitive returns. However, this does not
2190 * guarantee that softirq handlers will have completed, since in some
2191 * kernels, these handlers can run in process context, and can block.
2193 * Note that this guarantee implies further memory-ordering guarantees.
2194 * On systems with more than one CPU, when synchronize_sched() returns,
2195 * each CPU is guaranteed to have executed a full memory barrier since the
2196 * end of its last RCU-sched read-side critical section whose beginning
2197 * preceded the call to synchronize_sched(). In addition, each CPU having
2198 * an RCU read-side critical section that extends beyond the return from
2199 * synchronize_sched() is guaranteed to have executed a full memory barrier
2200 * after the beginning of synchronize_sched() and before the beginning of
2201 * that RCU read-side critical section. Note that these guarantees include
2202 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2203 * that are executing in the kernel.
2205 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2206 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2207 * to have executed a full memory barrier during the execution of
2208 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2209 * again only if the system has more than one CPU).
2211 * This primitive provides the guarantees made by the (now removed)
2212 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2213 * guarantees that rcu_read_lock() sections will have completed.
2214 * In "classic RCU", these two guarantees happen to be one and
2215 * the same, but can differ in realtime RCU implementations.
2217 void synchronize_sched(void)
2219 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2220 !lock_is_held(&rcu_lock_map
) &&
2221 !lock_is_held(&rcu_sched_lock_map
),
2222 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2223 if (rcu_blocking_is_gp())
2226 synchronize_sched_expedited();
2228 wait_rcu_gp(call_rcu_sched
);
2230 EXPORT_SYMBOL_GPL(synchronize_sched
);
2233 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2235 * Control will return to the caller some time after a full rcu_bh grace
2236 * period has elapsed, in other words after all currently executing rcu_bh
2237 * read-side critical sections have completed. RCU read-side critical
2238 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2239 * and may be nested.
2241 * See the description of synchronize_sched() for more detailed information
2242 * on memory ordering guarantees.
2244 void synchronize_rcu_bh(void)
2246 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2247 !lock_is_held(&rcu_lock_map
) &&
2248 !lock_is_held(&rcu_sched_lock_map
),
2249 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2250 if (rcu_blocking_is_gp())
2253 synchronize_rcu_bh_expedited();
2255 wait_rcu_gp(call_rcu_bh
);
2257 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2259 static int synchronize_sched_expedited_cpu_stop(void *data
)
2262 * There must be a full memory barrier on each affected CPU
2263 * between the time that try_stop_cpus() is called and the
2264 * time that it returns.
2266 * In the current initial implementation of cpu_stop, the
2267 * above condition is already met when the control reaches
2268 * this point and the following smp_mb() is not strictly
2269 * necessary. Do smp_mb() anyway for documentation and
2270 * robustness against future implementation changes.
2272 smp_mb(); /* See above comment block. */
2277 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2279 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2280 * approach to force the grace period to end quickly. This consumes
2281 * significant time on all CPUs and is unfriendly to real-time workloads,
2282 * so is thus not recommended for any sort of common-case code. In fact,
2283 * if you are using synchronize_sched_expedited() in a loop, please
2284 * restructure your code to batch your updates, and then use a single
2285 * synchronize_sched() instead.
2287 * Note that it is illegal to call this function while holding any lock
2288 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2289 * to call this function from a CPU-hotplug notifier. Failing to observe
2290 * these restriction will result in deadlock.
2292 * This implementation can be thought of as an application of ticket
2293 * locking to RCU, with sync_sched_expedited_started and
2294 * sync_sched_expedited_done taking on the roles of the halves
2295 * of the ticket-lock word. Each task atomically increments
2296 * sync_sched_expedited_started upon entry, snapshotting the old value,
2297 * then attempts to stop all the CPUs. If this succeeds, then each
2298 * CPU will have executed a context switch, resulting in an RCU-sched
2299 * grace period. We are then done, so we use atomic_cmpxchg() to
2300 * update sync_sched_expedited_done to match our snapshot -- but
2301 * only if someone else has not already advanced past our snapshot.
2303 * On the other hand, if try_stop_cpus() fails, we check the value
2304 * of sync_sched_expedited_done. If it has advanced past our
2305 * initial snapshot, then someone else must have forced a grace period
2306 * some time after we took our snapshot. In this case, our work is
2307 * done for us, and we can simply return. Otherwise, we try again,
2308 * but keep our initial snapshot for purposes of checking for someone
2309 * doing our work for us.
2311 * If we fail too many times in a row, we fall back to synchronize_sched().
2313 void synchronize_sched_expedited(void)
2315 long firstsnap
, s
, snap
;
2317 struct rcu_state
*rsp
= &rcu_sched_state
;
2320 * If we are in danger of counter wrap, just do synchronize_sched().
2321 * By allowing sync_sched_expedited_started to advance no more than
2322 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2323 * that more than 3.5 billion CPUs would be required to force a
2324 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2325 * course be required on a 64-bit system.
2327 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
2328 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
2330 synchronize_sched();
2331 atomic_long_inc(&rsp
->expedited_wrap
);
2336 * Take a ticket. Note that atomic_inc_return() implies a
2337 * full memory barrier.
2339 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
2342 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2345 * Each pass through the following loop attempts to force a
2346 * context switch on each CPU.
2348 while (try_stop_cpus(cpu_online_mask
,
2349 synchronize_sched_expedited_cpu_stop
,
2352 atomic_long_inc(&rsp
->expedited_tryfail
);
2354 /* Check to see if someone else did our work for us. */
2355 s
= atomic_long_read(&rsp
->expedited_done
);
2356 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2357 /* ensure test happens before caller kfree */
2358 smp_mb__before_atomic_inc(); /* ^^^ */
2359 atomic_long_inc(&rsp
->expedited_workdone1
);
2363 /* No joy, try again later. Or just synchronize_sched(). */
2364 if (trycount
++ < 10) {
2365 udelay(trycount
* num_online_cpus());
2367 wait_rcu_gp(call_rcu_sched
);
2368 atomic_long_inc(&rsp
->expedited_normal
);
2372 /* Recheck to see if someone else did our work for us. */
2373 s
= atomic_long_read(&rsp
->expedited_done
);
2374 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2375 /* ensure test happens before caller kfree */
2376 smp_mb__before_atomic_inc(); /* ^^^ */
2377 atomic_long_inc(&rsp
->expedited_workdone2
);
2382 * Refetching sync_sched_expedited_started allows later
2383 * callers to piggyback on our grace period. We retry
2384 * after they started, so our grace period works for them,
2385 * and they started after our first try, so their grace
2386 * period works for us.
2389 snap
= atomic_long_read(&rsp
->expedited_start
);
2390 smp_mb(); /* ensure read is before try_stop_cpus(). */
2392 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
2395 * Everyone up to our most recent fetch is covered by our grace
2396 * period. Update the counter, but only if our work is still
2397 * relevant -- which it won't be if someone who started later
2398 * than we did already did their update.
2401 atomic_long_inc(&rsp
->expedited_done_tries
);
2402 s
= atomic_long_read(&rsp
->expedited_done
);
2403 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
2404 /* ensure test happens before caller kfree */
2405 smp_mb__before_atomic_inc(); /* ^^^ */
2406 atomic_long_inc(&rsp
->expedited_done_lost
);
2409 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
2410 atomic_long_inc(&rsp
->expedited_done_exit
);
2414 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2417 * Check to see if there is any immediate RCU-related work to be done
2418 * by the current CPU, for the specified type of RCU, returning 1 if so.
2419 * The checks are in order of increasing expense: checks that can be
2420 * carried out against CPU-local state are performed first. However,
2421 * we must check for CPU stalls first, else we might not get a chance.
2423 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2425 struct rcu_node
*rnp
= rdp
->mynode
;
2427 rdp
->n_rcu_pending
++;
2429 /* Check for CPU stalls, if enabled. */
2430 check_cpu_stall(rsp
, rdp
);
2432 /* Is the RCU core waiting for a quiescent state from this CPU? */
2433 if (rcu_scheduler_fully_active
&&
2434 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2435 rdp
->n_rp_qs_pending
++;
2436 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2437 rdp
->n_rp_report_qs
++;
2441 /* Does this CPU have callbacks ready to invoke? */
2442 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2443 rdp
->n_rp_cb_ready
++;
2447 /* Has RCU gone idle with this CPU needing another grace period? */
2448 if (cpu_needs_another_gp(rsp
, rdp
)) {
2449 rdp
->n_rp_cpu_needs_gp
++;
2453 /* Has another RCU grace period completed? */
2454 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2455 rdp
->n_rp_gp_completed
++;
2459 /* Has a new RCU grace period started? */
2460 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2461 rdp
->n_rp_gp_started
++;
2466 rdp
->n_rp_need_nothing
++;
2471 * Check to see if there is any immediate RCU-related work to be done
2472 * by the current CPU, returning 1 if so. This function is part of the
2473 * RCU implementation; it is -not- an exported member of the RCU API.
2475 static int rcu_pending(int cpu
)
2477 struct rcu_state
*rsp
;
2479 for_each_rcu_flavor(rsp
)
2480 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2486 * Check to see if any future RCU-related work will need to be done
2487 * by the current CPU, even if none need be done immediately, returning
2490 static int rcu_cpu_has_callbacks(int cpu
)
2492 struct rcu_state
*rsp
;
2494 /* RCU callbacks either ready or pending? */
2495 for_each_rcu_flavor(rsp
)
2496 if (per_cpu_ptr(rsp
->rda
, cpu
)->nxtlist
)
2502 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2503 * the compiler is expected to optimize this away.
2505 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2506 int cpu
, unsigned long done
)
2508 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2509 atomic_read(&rsp
->barrier_cpu_count
), done
);
2513 * RCU callback function for _rcu_barrier(). If we are last, wake
2514 * up the task executing _rcu_barrier().
2516 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2518 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2519 struct rcu_state
*rsp
= rdp
->rsp
;
2521 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2522 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2523 complete(&rsp
->barrier_completion
);
2525 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2530 * Called with preemption disabled, and from cross-cpu IRQ context.
2532 static void rcu_barrier_func(void *type
)
2534 struct rcu_state
*rsp
= type
;
2535 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2537 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2538 atomic_inc(&rsp
->barrier_cpu_count
);
2539 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2543 * Orchestrate the specified type of RCU barrier, waiting for all
2544 * RCU callbacks of the specified type to complete.
2546 static void _rcu_barrier(struct rcu_state
*rsp
)
2549 struct rcu_data
*rdp
;
2550 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2551 unsigned long snap_done
;
2553 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2555 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2556 mutex_lock(&rsp
->barrier_mutex
);
2559 * Ensure that all prior references, including to ->n_barrier_done,
2560 * are ordered before the _rcu_barrier() machinery.
2562 smp_mb(); /* See above block comment. */
2565 * Recheck ->n_barrier_done to see if others did our work for us.
2566 * This means checking ->n_barrier_done for an even-to-odd-to-even
2567 * transition. The "if" expression below therefore rounds the old
2568 * value up to the next even number and adds two before comparing.
2570 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2571 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2572 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2573 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2574 smp_mb(); /* caller's subsequent code after above check. */
2575 mutex_unlock(&rsp
->barrier_mutex
);
2580 * Increment ->n_barrier_done to avoid duplicate work. Use
2581 * ACCESS_ONCE() to prevent the compiler from speculating
2582 * the increment to precede the early-exit check.
2584 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2585 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2586 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2587 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2590 * Initialize the count to one rather than to zero in order to
2591 * avoid a too-soon return to zero in case of a short grace period
2592 * (or preemption of this task). Exclude CPU-hotplug operations
2593 * to ensure that no offline CPU has callbacks queued.
2595 init_completion(&rsp
->barrier_completion
);
2596 atomic_set(&rsp
->barrier_cpu_count
, 1);
2600 * Force each CPU with callbacks to register a new callback.
2601 * When that callback is invoked, we will know that all of the
2602 * corresponding CPU's preceding callbacks have been invoked.
2604 for_each_possible_cpu(cpu
) {
2605 if (!cpu_online(cpu
) && !is_nocb_cpu(cpu
))
2607 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2608 if (is_nocb_cpu(cpu
)) {
2609 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
2610 rsp
->n_barrier_done
);
2611 atomic_inc(&rsp
->barrier_cpu_count
);
2612 __call_rcu(&rdp
->barrier_head
, rcu_barrier_callback
,
2614 } else if (ACCESS_ONCE(rdp
->qlen
)) {
2615 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2616 rsp
->n_barrier_done
);
2617 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2619 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2620 rsp
->n_barrier_done
);
2626 * Now that we have an rcu_barrier_callback() callback on each
2627 * CPU, and thus each counted, remove the initial count.
2629 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2630 complete(&rsp
->barrier_completion
);
2632 /* Increment ->n_barrier_done to prevent duplicate work. */
2633 smp_mb(); /* Keep increment after above mechanism. */
2634 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2635 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2636 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2637 smp_mb(); /* Keep increment before caller's subsequent code. */
2639 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2640 wait_for_completion(&rsp
->barrier_completion
);
2642 /* Other rcu_barrier() invocations can now safely proceed. */
2643 mutex_unlock(&rsp
->barrier_mutex
);
2647 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2649 void rcu_barrier_bh(void)
2651 _rcu_barrier(&rcu_bh_state
);
2653 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2656 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2658 void rcu_barrier_sched(void)
2660 _rcu_barrier(&rcu_sched_state
);
2662 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2665 * Do boot-time initialization of a CPU's per-CPU RCU data.
2668 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2670 unsigned long flags
;
2671 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2672 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2674 /* Set up local state, ensuring consistent view of global state. */
2675 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2676 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2677 init_callback_list(rdp
);
2679 ACCESS_ONCE(rdp
->qlen
) = 0;
2680 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2681 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2682 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2683 #ifdef CONFIG_RCU_USER_QS
2684 WARN_ON_ONCE(rdp
->dynticks
->in_user
);
2688 rcu_boot_init_nocb_percpu_data(rdp
);
2689 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2693 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2694 * offline event can be happening at a given time. Note also that we
2695 * can accept some slop in the rsp->completed access due to the fact
2696 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2698 static void __cpuinit
2699 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2701 unsigned long flags
;
2703 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2704 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2706 /* Exclude new grace periods. */
2707 mutex_lock(&rsp
->onoff_mutex
);
2709 /* Set up local state, ensuring consistent view of global state. */
2710 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2711 rdp
->beenonline
= 1; /* We have now been online. */
2712 rdp
->preemptible
= preemptible
;
2713 rdp
->qlen_last_fqs_check
= 0;
2714 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2715 rdp
->blimit
= blimit
;
2716 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
2717 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2718 atomic_set(&rdp
->dynticks
->dynticks
,
2719 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2720 rcu_prepare_for_idle_init(cpu
);
2721 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2723 /* Add CPU to rcu_node bitmasks. */
2725 mask
= rdp
->grpmask
;
2727 /* Exclude any attempts to start a new GP on small systems. */
2728 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2729 rnp
->qsmaskinit
|= mask
;
2730 mask
= rnp
->grpmask
;
2731 if (rnp
== rdp
->mynode
) {
2733 * If there is a grace period in progress, we will
2734 * set up to wait for it next time we run the
2737 rdp
->gpnum
= rnp
->completed
;
2738 rdp
->completed
= rnp
->completed
;
2739 rdp
->passed_quiesce
= 0;
2740 rdp
->qs_pending
= 0;
2741 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2743 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2745 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2746 local_irq_restore(flags
);
2748 mutex_unlock(&rsp
->onoff_mutex
);
2751 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2753 struct rcu_state
*rsp
;
2755 for_each_rcu_flavor(rsp
)
2756 rcu_init_percpu_data(cpu
, rsp
,
2757 strcmp(rsp
->name
, "rcu_preempt") == 0);
2761 * Handle CPU online/offline notification events.
2763 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2764 unsigned long action
, void *hcpu
)
2766 long cpu
= (long)hcpu
;
2767 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2768 struct rcu_node
*rnp
= rdp
->mynode
;
2769 struct rcu_state
*rsp
;
2770 int ret
= NOTIFY_OK
;
2772 trace_rcu_utilization("Start CPU hotplug");
2774 case CPU_UP_PREPARE
:
2775 case CPU_UP_PREPARE_FROZEN
:
2776 rcu_prepare_cpu(cpu
);
2777 rcu_prepare_kthreads(cpu
);
2780 case CPU_DOWN_FAILED
:
2781 rcu_boost_kthread_setaffinity(rnp
, -1);
2783 case CPU_DOWN_PREPARE
:
2784 if (nocb_cpu_expendable(cpu
))
2785 rcu_boost_kthread_setaffinity(rnp
, cpu
);
2790 case CPU_DYING_FROZEN
:
2792 * The whole machine is "stopped" except this CPU, so we can
2793 * touch any data without introducing corruption. We send the
2794 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2796 for_each_rcu_flavor(rsp
)
2797 rcu_cleanup_dying_cpu(rsp
);
2798 rcu_cleanup_after_idle(cpu
);
2801 case CPU_DEAD_FROZEN
:
2802 case CPU_UP_CANCELED
:
2803 case CPU_UP_CANCELED_FROZEN
:
2804 for_each_rcu_flavor(rsp
)
2805 rcu_cleanup_dead_cpu(cpu
, rsp
);
2810 trace_rcu_utilization("End CPU hotplug");
2815 * Spawn the kthread that handles this RCU flavor's grace periods.
2817 static int __init
rcu_spawn_gp_kthread(void)
2819 unsigned long flags
;
2820 struct rcu_node
*rnp
;
2821 struct rcu_state
*rsp
;
2822 struct task_struct
*t
;
2824 for_each_rcu_flavor(rsp
) {
2825 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
2827 rnp
= rcu_get_root(rsp
);
2828 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2829 rsp
->gp_kthread
= t
;
2830 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2831 rcu_spawn_nocb_kthreads(rsp
);
2835 early_initcall(rcu_spawn_gp_kthread
);
2838 * This function is invoked towards the end of the scheduler's initialization
2839 * process. Before this is called, the idle task might contain
2840 * RCU read-side critical sections (during which time, this idle
2841 * task is booting the system). After this function is called, the
2842 * idle tasks are prohibited from containing RCU read-side critical
2843 * sections. This function also enables RCU lockdep checking.
2845 void rcu_scheduler_starting(void)
2847 WARN_ON(num_online_cpus() != 1);
2848 WARN_ON(nr_context_switches() > 0);
2849 rcu_scheduler_active
= 1;
2853 * Compute the per-level fanout, either using the exact fanout specified
2854 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2856 #ifdef CONFIG_RCU_FANOUT_EXACT
2857 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2861 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
2862 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2863 rsp
->levelspread
[0] = rcu_fanout_leaf
;
2865 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2866 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2873 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2874 ccur
= rsp
->levelcnt
[i
];
2875 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2879 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2882 * Helper function for rcu_init() that initializes one rcu_state structure.
2884 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2885 struct rcu_data __percpu
*rda
)
2887 static char *buf
[] = { "rcu_node_0",
2890 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2891 static char *fqs
[] = { "rcu_node_fqs_0",
2894 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2898 struct rcu_node
*rnp
;
2900 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2902 /* Initialize the level-tracking arrays. */
2904 for (i
= 0; i
< rcu_num_lvls
; i
++)
2905 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
2906 for (i
= 1; i
< rcu_num_lvls
; i
++)
2907 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2908 rcu_init_levelspread(rsp
);
2910 /* Initialize the elements themselves, starting from the leaves. */
2912 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2913 cpustride
*= rsp
->levelspread
[i
];
2914 rnp
= rsp
->level
[i
];
2915 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2916 raw_spin_lock_init(&rnp
->lock
);
2917 lockdep_set_class_and_name(&rnp
->lock
,
2918 &rcu_node_class
[i
], buf
[i
]);
2919 raw_spin_lock_init(&rnp
->fqslock
);
2920 lockdep_set_class_and_name(&rnp
->fqslock
,
2921 &rcu_fqs_class
[i
], fqs
[i
]);
2922 rnp
->gpnum
= rsp
->gpnum
;
2923 rnp
->completed
= rsp
->completed
;
2925 rnp
->qsmaskinit
= 0;
2926 rnp
->grplo
= j
* cpustride
;
2927 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2928 if (rnp
->grphi
>= NR_CPUS
)
2929 rnp
->grphi
= NR_CPUS
- 1;
2935 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2936 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2937 rnp
->parent
= rsp
->level
[i
- 1] +
2938 j
/ rsp
->levelspread
[i
- 1];
2941 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2946 init_waitqueue_head(&rsp
->gp_wq
);
2947 rnp
= rsp
->level
[rcu_num_lvls
- 1];
2948 for_each_possible_cpu(i
) {
2949 while (i
> rnp
->grphi
)
2951 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2952 rcu_boot_init_percpu_data(i
, rsp
);
2954 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
2958 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2959 * replace the definitions in rcutree.h because those are needed to size
2960 * the ->node array in the rcu_state structure.
2962 static void __init
rcu_init_geometry(void)
2967 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
2969 /* If the compile-time values are accurate, just leave. */
2970 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
2971 nr_cpu_ids
== NR_CPUS
)
2975 * Compute number of nodes that can be handled an rcu_node tree
2976 * with the given number of levels. Setting rcu_capacity[0] makes
2977 * some of the arithmetic easier.
2979 rcu_capacity
[0] = 1;
2980 rcu_capacity
[1] = rcu_fanout_leaf
;
2981 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
2982 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
2985 * The boot-time rcu_fanout_leaf parameter is only permitted
2986 * to increase the leaf-level fanout, not decrease it. Of course,
2987 * the leaf-level fanout cannot exceed the number of bits in
2988 * the rcu_node masks. Finally, the tree must be able to accommodate
2989 * the configured number of CPUs. Complain and fall back to the
2990 * compile-time values if these limits are exceeded.
2992 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
2993 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
2994 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
2999 /* Calculate the number of rcu_nodes at each level of the tree. */
3000 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3001 if (n
<= rcu_capacity
[i
]) {
3002 for (j
= 0; j
<= i
; j
++)
3004 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3006 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3011 /* Calculate the total number of rcu_node structures. */
3013 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3014 rcu_num_nodes
+= num_rcu_lvl
[i
];
3018 void __init
rcu_init(void)
3022 rcu_bootup_announce();
3023 rcu_init_geometry();
3024 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3025 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3026 __rcu_init_preempt();
3028 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3031 * We don't need protection against CPU-hotplug here because
3032 * this is called early in boot, before either interrupts
3033 * or the scheduler are operational.
3035 cpu_notifier(rcu_cpu_notify
, 0);
3036 for_each_online_cpu(cpu
)
3037 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3040 #include "rcutree_plugin.h"