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>
57 #include <trace/events/rcu.h>
61 /* Data structures. */
63 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
64 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
66 #define RCU_STATE_INITIALIZER(sname, cr) { \
67 .level = { &sname##_state.node[0] }, \
69 .fqs_state = RCU_GP_IDLE, \
72 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
73 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
74 .orphan_donetail = &sname##_state.orphan_donelist, \
75 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
79 struct rcu_state rcu_sched_state
=
80 RCU_STATE_INITIALIZER(rcu_sched
, call_rcu_sched
);
81 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
83 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
, call_rcu_bh
);
84 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
86 static struct rcu_state
*rcu_state
;
87 LIST_HEAD(rcu_struct_flavors
);
89 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
90 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
91 module_param(rcu_fanout_leaf
, int, 0);
92 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
93 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
100 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
103 * The rcu_scheduler_active variable transitions from zero to one just
104 * before the first task is spawned. So when this variable is zero, RCU
105 * can assume that there is but one task, allowing RCU to (for example)
106 * optimized synchronize_sched() to a simple barrier(). When this variable
107 * is one, RCU must actually do all the hard work required to detect real
108 * grace periods. This variable is also used to suppress boot-time false
109 * positives from lockdep-RCU error checking.
111 int rcu_scheduler_active __read_mostly
;
112 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
115 * The rcu_scheduler_fully_active variable transitions from zero to one
116 * during the early_initcall() processing, which is after the scheduler
117 * is capable of creating new tasks. So RCU processing (for example,
118 * creating tasks for RCU priority boosting) must be delayed until after
119 * rcu_scheduler_fully_active transitions from zero to one. We also
120 * currently delay invocation of any RCU callbacks until after this point.
122 * It might later prove better for people registering RCU callbacks during
123 * early boot to take responsibility for these callbacks, but one step at
126 static int rcu_scheduler_fully_active __read_mostly
;
128 #ifdef CONFIG_RCU_BOOST
131 * Control variables for per-CPU and per-rcu_node kthreads. These
132 * handle all flavors of RCU.
134 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
135 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
136 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
138 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
140 #endif /* #ifdef CONFIG_RCU_BOOST */
142 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
143 static void invoke_rcu_core(void);
144 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
147 * Track the rcutorture test sequence number and the update version
148 * number within a given test. The rcutorture_testseq is incremented
149 * on every rcutorture module load and unload, so has an odd value
150 * when a test is running. The rcutorture_vernum is set to zero
151 * when rcutorture starts and is incremented on each rcutorture update.
152 * These variables enable correlating rcutorture output with the
153 * RCU tracing information.
155 unsigned long rcutorture_testseq
;
156 unsigned long rcutorture_vernum
;
159 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
160 * permit this function to be invoked without holding the root rcu_node
161 * structure's ->lock, but of course results can be subject to change.
163 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
165 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
169 * Note a quiescent state. Because we do not need to know
170 * how many quiescent states passed, just if there was at least
171 * one since the start of the grace period, this just sets a flag.
172 * The caller must have disabled preemption.
174 void rcu_sched_qs(int cpu
)
176 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
178 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
180 if (rdp
->passed_quiesce
== 0)
181 trace_rcu_grace_period("rcu_sched", rdp
->gpnum
, "cpuqs");
182 rdp
->passed_quiesce
= 1;
185 void rcu_bh_qs(int cpu
)
187 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
189 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
191 if (rdp
->passed_quiesce
== 0)
192 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
193 rdp
->passed_quiesce
= 1;
197 * Note a context switch. This is a quiescent state for RCU-sched,
198 * and requires special handling for preemptible RCU.
199 * The caller must have disabled preemption.
201 void rcu_note_context_switch(int cpu
)
203 trace_rcu_utilization("Start context switch");
205 rcu_preempt_note_context_switch(cpu
);
206 trace_rcu_utilization("End context switch");
208 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
210 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
211 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
212 .dynticks
= ATOMIC_INIT(1),
215 static int blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
216 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
217 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
219 module_param(blimit
, int, 0);
220 module_param(qhimark
, int, 0);
221 module_param(qlowmark
, int, 0);
223 int rcu_cpu_stall_suppress __read_mostly
; /* 1 = suppress stall warnings. */
224 int rcu_cpu_stall_timeout __read_mostly
= CONFIG_RCU_CPU_STALL_TIMEOUT
;
226 module_param(rcu_cpu_stall_suppress
, int, 0644);
227 module_param(rcu_cpu_stall_timeout
, int, 0644);
229 static ulong jiffies_till_first_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
230 static ulong jiffies_till_next_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
232 module_param(jiffies_till_first_fqs
, ulong
, 0644);
233 module_param(jiffies_till_next_fqs
, ulong
, 0644);
235 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*));
236 static void force_quiescent_state(struct rcu_state
*rsp
);
237 static int rcu_pending(int cpu
);
240 * Return the number of RCU-sched batches processed thus far for debug & stats.
242 long rcu_batches_completed_sched(void)
244 return rcu_sched_state
.completed
;
246 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
249 * Return the number of RCU BH batches processed thus far for debug & stats.
251 long rcu_batches_completed_bh(void)
253 return rcu_bh_state
.completed
;
255 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
258 * Force a quiescent state for RCU BH.
260 void rcu_bh_force_quiescent_state(void)
262 force_quiescent_state(&rcu_bh_state
);
264 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
267 * Record the number of times rcutorture tests have been initiated and
268 * terminated. This information allows the debugfs tracing stats to be
269 * correlated to the rcutorture messages, even when the rcutorture module
270 * is being repeatedly loaded and unloaded. In other words, we cannot
271 * store this state in rcutorture itself.
273 void rcutorture_record_test_transition(void)
275 rcutorture_testseq
++;
276 rcutorture_vernum
= 0;
278 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
281 * Record the number of writer passes through the current rcutorture test.
282 * This is also used to correlate debugfs tracing stats with the rcutorture
285 void rcutorture_record_progress(unsigned long vernum
)
289 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
292 * Force a quiescent state for RCU-sched.
294 void rcu_sched_force_quiescent_state(void)
296 force_quiescent_state(&rcu_sched_state
);
298 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
301 * Does the CPU have callbacks ready to be invoked?
304 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
306 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
310 * Does the current CPU require a yet-as-unscheduled grace period?
313 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
315 return *rdp
->nxttail
[RCU_DONE_TAIL
+
316 ACCESS_ONCE(rsp
->completed
) != rdp
->completed
] &&
317 !rcu_gp_in_progress(rsp
);
321 * Return the root node of the specified rcu_state structure.
323 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
325 return &rsp
->node
[0];
329 * If the specified CPU is offline, tell the caller that it is in
330 * a quiescent state. Otherwise, whack it with a reschedule IPI.
331 * Grace periods can end up waiting on an offline CPU when that
332 * CPU is in the process of coming online -- it will be added to the
333 * rcu_node bitmasks before it actually makes it online. The same thing
334 * can happen while a CPU is in the process of coming online. Because this
335 * race is quite rare, we check for it after detecting that the grace
336 * period has been delayed rather than checking each and every CPU
337 * each and every time we start a new grace period.
339 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
342 * If the CPU is offline for more than a jiffy, it is in a quiescent
343 * state. We can trust its state not to change because interrupts
344 * are disabled. The reason for the jiffy's worth of slack is to
345 * handle CPUs initializing on the way up and finding their way
346 * to the idle loop on the way down.
348 if (cpu_is_offline(rdp
->cpu
) &&
349 ULONG_CMP_LT(rdp
->rsp
->gp_start
+ 2, jiffies
)) {
350 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
358 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
360 * If the new value of the ->dynticks_nesting counter now is zero,
361 * we really have entered idle, and must do the appropriate accounting.
362 * The caller must have disabled interrupts.
364 static void rcu_idle_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
)
366 trace_rcu_dyntick("Start", oldval
, 0);
367 if (!is_idle_task(current
)) {
368 struct task_struct
*idle
= idle_task(smp_processor_id());
370 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
371 ftrace_dump(DUMP_ORIG
);
372 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
373 current
->pid
, current
->comm
,
374 idle
->pid
, idle
->comm
); /* must be idle task! */
376 rcu_prepare_for_idle(smp_processor_id());
377 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
378 smp_mb__before_atomic_inc(); /* See above. */
379 atomic_inc(&rdtp
->dynticks
);
380 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
381 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
384 * The idle task is not permitted to enter the idle loop while
385 * in an RCU read-side critical section.
387 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
388 "Illegal idle entry in RCU read-side critical section.");
389 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
390 "Illegal idle entry in RCU-bh read-side critical section.");
391 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
392 "Illegal idle entry in RCU-sched read-side critical section.");
396 * rcu_idle_enter - inform RCU that current CPU is entering idle
398 * Enter idle mode, in other words, -leave- the mode in which RCU
399 * read-side critical sections can occur. (Though RCU read-side
400 * critical sections can occur in irq handlers in idle, a possibility
401 * handled by irq_enter() and irq_exit().)
403 * We crowbar the ->dynticks_nesting field to zero to allow for
404 * the possibility of usermode upcalls having messed up our count
405 * of interrupt nesting level during the prior busy period.
407 void rcu_idle_enter(void)
411 struct rcu_dynticks
*rdtp
;
413 local_irq_save(flags
);
414 rdtp
= &__get_cpu_var(rcu_dynticks
);
415 oldval
= rdtp
->dynticks_nesting
;
416 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
417 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
418 rdtp
->dynticks_nesting
= 0;
420 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
421 rcu_idle_enter_common(rdtp
, oldval
);
422 local_irq_restore(flags
);
424 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
427 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
429 * Exit from an interrupt handler, which might possibly result in entering
430 * idle mode, in other words, leaving the mode in which read-side critical
431 * sections can occur.
433 * This code assumes that the idle loop never does anything that might
434 * result in unbalanced calls to irq_enter() and irq_exit(). If your
435 * architecture violates this assumption, RCU will give you what you
436 * deserve, good and hard. But very infrequently and irreproducibly.
438 * Use things like work queues to work around this limitation.
440 * You have been warned.
442 void rcu_irq_exit(void)
446 struct rcu_dynticks
*rdtp
;
448 local_irq_save(flags
);
449 rdtp
= &__get_cpu_var(rcu_dynticks
);
450 oldval
= rdtp
->dynticks_nesting
;
451 rdtp
->dynticks_nesting
--;
452 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
453 if (rdtp
->dynticks_nesting
)
454 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
456 rcu_idle_enter_common(rdtp
, oldval
);
457 local_irq_restore(flags
);
461 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
463 * If the new value of the ->dynticks_nesting counter was previously zero,
464 * we really have exited idle, and must do the appropriate accounting.
465 * The caller must have disabled interrupts.
467 static void rcu_idle_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
)
469 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
470 atomic_inc(&rdtp
->dynticks
);
471 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
472 smp_mb__after_atomic_inc(); /* See above. */
473 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
474 rcu_cleanup_after_idle(smp_processor_id());
475 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
476 if (!is_idle_task(current
)) {
477 struct task_struct
*idle
= idle_task(smp_processor_id());
479 trace_rcu_dyntick("Error on exit: not idle task",
480 oldval
, rdtp
->dynticks_nesting
);
481 ftrace_dump(DUMP_ORIG
);
482 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
483 current
->pid
, current
->comm
,
484 idle
->pid
, idle
->comm
); /* must be idle task! */
489 * rcu_idle_exit - inform RCU that current CPU is leaving idle
491 * Exit idle mode, in other words, -enter- the mode in which RCU
492 * read-side critical sections can occur.
494 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
495 * allow for the possibility of usermode upcalls messing up our count
496 * of interrupt nesting level during the busy period that is just
499 void rcu_idle_exit(void)
502 struct rcu_dynticks
*rdtp
;
505 local_irq_save(flags
);
506 rdtp
= &__get_cpu_var(rcu_dynticks
);
507 oldval
= rdtp
->dynticks_nesting
;
508 WARN_ON_ONCE(oldval
< 0);
509 if (oldval
& DYNTICK_TASK_NEST_MASK
)
510 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
512 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
513 rcu_idle_exit_common(rdtp
, oldval
);
514 local_irq_restore(flags
);
516 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
519 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
521 * Enter an interrupt handler, which might possibly result in exiting
522 * idle mode, in other words, entering the mode in which read-side critical
523 * sections can occur.
525 * Note that the Linux kernel is fully capable of entering an interrupt
526 * handler that it never exits, for example when doing upcalls to
527 * user mode! This code assumes that the idle loop never does upcalls to
528 * user mode. If your architecture does do upcalls from the idle loop (or
529 * does anything else that results in unbalanced calls to the irq_enter()
530 * and irq_exit() functions), RCU will give you what you deserve, good
531 * and hard. But very infrequently and irreproducibly.
533 * Use things like work queues to work around this limitation.
535 * You have been warned.
537 void rcu_irq_enter(void)
540 struct rcu_dynticks
*rdtp
;
543 local_irq_save(flags
);
544 rdtp
= &__get_cpu_var(rcu_dynticks
);
545 oldval
= rdtp
->dynticks_nesting
;
546 rdtp
->dynticks_nesting
++;
547 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
549 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
551 rcu_idle_exit_common(rdtp
, oldval
);
552 local_irq_restore(flags
);
556 * rcu_nmi_enter - inform RCU of entry to NMI context
558 * If the CPU was idle with dynamic ticks active, and there is no
559 * irq handler running, this updates rdtp->dynticks_nmi to let the
560 * RCU grace-period handling know that the CPU is active.
562 void rcu_nmi_enter(void)
564 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
566 if (rdtp
->dynticks_nmi_nesting
== 0 &&
567 (atomic_read(&rdtp
->dynticks
) & 0x1))
569 rdtp
->dynticks_nmi_nesting
++;
570 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
571 atomic_inc(&rdtp
->dynticks
);
572 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
573 smp_mb__after_atomic_inc(); /* See above. */
574 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
578 * rcu_nmi_exit - inform RCU of exit from NMI context
580 * If the CPU was idle with dynamic ticks active, and there is no
581 * irq handler running, this updates rdtp->dynticks_nmi to let the
582 * RCU grace-period handling know that the CPU is no longer active.
584 void rcu_nmi_exit(void)
586 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
588 if (rdtp
->dynticks_nmi_nesting
== 0 ||
589 --rdtp
->dynticks_nmi_nesting
!= 0)
591 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
592 smp_mb__before_atomic_inc(); /* See above. */
593 atomic_inc(&rdtp
->dynticks
);
594 smp_mb__after_atomic_inc(); /* Force delay to next write. */
595 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
599 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
601 * If the current CPU is in its idle loop and is neither in an interrupt
602 * or NMI handler, return true.
604 int rcu_is_cpu_idle(void)
609 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
613 EXPORT_SYMBOL(rcu_is_cpu_idle
);
615 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
618 * Is the current CPU online? Disable preemption to avoid false positives
619 * that could otherwise happen due to the current CPU number being sampled,
620 * this task being preempted, its old CPU being taken offline, resuming
621 * on some other CPU, then determining that its old CPU is now offline.
622 * It is OK to use RCU on an offline processor during initial boot, hence
623 * the check for rcu_scheduler_fully_active. Note also that it is OK
624 * for a CPU coming online to use RCU for one jiffy prior to marking itself
625 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
626 * offline to continue to use RCU for one jiffy after marking itself
627 * offline in the cpu_online_mask. This leniency is necessary given the
628 * non-atomic nature of the online and offline processing, for example,
629 * the fact that a CPU enters the scheduler after completing the CPU_DYING
632 * This is also why RCU internally marks CPUs online during the
633 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
635 * Disable checking if in an NMI handler because we cannot safely report
636 * errors from NMI handlers anyway.
638 bool rcu_lockdep_current_cpu_online(void)
640 struct rcu_data
*rdp
;
641 struct rcu_node
*rnp
;
647 rdp
= &__get_cpu_var(rcu_sched_data
);
649 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
650 !rcu_scheduler_fully_active
;
654 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
656 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
659 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
661 * If the current CPU is idle or running at a first-level (not nested)
662 * interrupt from idle, return true. The caller must have at least
663 * disabled preemption.
665 int rcu_is_cpu_rrupt_from_idle(void)
667 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
671 * Snapshot the specified CPU's dynticks counter so that we can later
672 * credit them with an implicit quiescent state. Return 1 if this CPU
673 * is in dynticks idle mode, which is an extended quiescent state.
675 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
677 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
678 return (rdp
->dynticks_snap
& 0x1) == 0;
682 * Return true if the specified CPU has passed through a quiescent
683 * state by virtue of being in or having passed through an dynticks
684 * idle state since the last call to dyntick_save_progress_counter()
687 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
692 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
693 snap
= (unsigned int)rdp
->dynticks_snap
;
696 * If the CPU passed through or entered a dynticks idle phase with
697 * no active irq/NMI handlers, then we can safely pretend that the CPU
698 * already acknowledged the request to pass through a quiescent
699 * state. Either way, that CPU cannot possibly be in an RCU
700 * read-side critical section that started before the beginning
701 * of the current RCU grace period.
703 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
704 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
709 /* Go check for the CPU being offline. */
710 return rcu_implicit_offline_qs(rdp
);
713 static int jiffies_till_stall_check(void)
715 int till_stall_check
= ACCESS_ONCE(rcu_cpu_stall_timeout
);
718 * Limit check must be consistent with the Kconfig limits
719 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
721 if (till_stall_check
< 3) {
722 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 3;
723 till_stall_check
= 3;
724 } else if (till_stall_check
> 300) {
725 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 300;
726 till_stall_check
= 300;
728 return till_stall_check
* HZ
+ RCU_STALL_DELAY_DELTA
;
731 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
733 rsp
->gp_start
= jiffies
;
734 rsp
->jiffies_stall
= jiffies
+ jiffies_till_stall_check();
737 static void print_other_cpu_stall(struct rcu_state
*rsp
)
743 struct rcu_node
*rnp
= rcu_get_root(rsp
);
745 /* Only let one CPU complain about others per time interval. */
747 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
748 delta
= jiffies
- rsp
->jiffies_stall
;
749 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
750 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
753 rsp
->jiffies_stall
= jiffies
+ 3 * jiffies_till_stall_check() + 3;
754 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
757 * OK, time to rat on our buddy...
758 * See Documentation/RCU/stallwarn.txt for info on how to debug
759 * RCU CPU stall warnings.
761 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
763 print_cpu_stall_info_begin();
764 rcu_for_each_leaf_node(rsp
, rnp
) {
765 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
766 ndetected
+= rcu_print_task_stall(rnp
);
767 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
768 if (rnp
->qsmask
== 0)
770 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
771 if (rnp
->qsmask
& (1UL << cpu
)) {
772 print_cpu_stall_info(rsp
, rnp
->grplo
+ cpu
);
778 * Now rat on any tasks that got kicked up to the root rcu_node
779 * due to CPU offlining.
781 rnp
= rcu_get_root(rsp
);
782 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
783 ndetected
+= rcu_print_task_stall(rnp
);
784 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
786 print_cpu_stall_info_end();
787 printk(KERN_CONT
"(detected by %d, t=%ld jiffies)\n",
788 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
790 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
791 else if (!trigger_all_cpu_backtrace())
794 /* Complain about tasks blocking the grace period. */
796 rcu_print_detail_task_stall(rsp
);
798 force_quiescent_state(rsp
); /* Kick them all. */
801 static void print_cpu_stall(struct rcu_state
*rsp
)
804 struct rcu_node
*rnp
= rcu_get_root(rsp
);
807 * OK, time to rat on ourselves...
808 * See Documentation/RCU/stallwarn.txt for info on how to debug
809 * RCU CPU stall warnings.
811 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
812 print_cpu_stall_info_begin();
813 print_cpu_stall_info(rsp
, smp_processor_id());
814 print_cpu_stall_info_end();
815 printk(KERN_CONT
" (t=%lu jiffies)\n", jiffies
- rsp
->gp_start
);
816 if (!trigger_all_cpu_backtrace())
819 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
820 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
821 rsp
->jiffies_stall
= jiffies
+
822 3 * jiffies_till_stall_check() + 3;
823 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
825 set_need_resched(); /* kick ourselves to get things going. */
828 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
832 struct rcu_node
*rnp
;
834 if (rcu_cpu_stall_suppress
)
836 j
= ACCESS_ONCE(jiffies
);
837 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
839 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
841 /* We haven't checked in, so go dump stack. */
842 print_cpu_stall(rsp
);
844 } else if (rcu_gp_in_progress(rsp
) &&
845 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
847 /* They had a few time units to dump stack, so complain. */
848 print_other_cpu_stall(rsp
);
852 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
854 rcu_cpu_stall_suppress
= 1;
859 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
861 * Set the stall-warning timeout way off into the future, thus preventing
862 * any RCU CPU stall-warning messages from appearing in the current set of
865 * The caller must disable hard irqs.
867 void rcu_cpu_stall_reset(void)
869 struct rcu_state
*rsp
;
871 for_each_rcu_flavor(rsp
)
872 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
875 static struct notifier_block rcu_panic_block
= {
876 .notifier_call
= rcu_panic
,
879 static void __init
check_cpu_stall_init(void)
881 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
885 * Update CPU-local rcu_data state to record the newly noticed grace period.
886 * This is used both when we started the grace period and when we notice
887 * that someone else started the grace period. The caller must hold the
888 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
889 * and must have irqs disabled.
891 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
893 if (rdp
->gpnum
!= rnp
->gpnum
) {
895 * If the current grace period is waiting for this CPU,
896 * set up to detect a quiescent state, otherwise don't
897 * go looking for one.
899 rdp
->gpnum
= rnp
->gpnum
;
900 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
901 if (rnp
->qsmask
& rdp
->grpmask
) {
903 rdp
->passed_quiesce
= 0;
907 zero_cpu_stall_ticks(rdp
);
911 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
914 struct rcu_node
*rnp
;
916 local_irq_save(flags
);
918 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
919 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
920 local_irq_restore(flags
);
923 __note_new_gpnum(rsp
, rnp
, rdp
);
924 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
928 * Did someone else start a new RCU grace period start since we last
929 * checked? Update local state appropriately if so. Must be called
930 * on the CPU corresponding to rdp.
933 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
938 local_irq_save(flags
);
939 if (rdp
->gpnum
!= rsp
->gpnum
) {
940 note_new_gpnum(rsp
, rdp
);
943 local_irq_restore(flags
);
948 * Initialize the specified rcu_data structure's callback list to empty.
950 static void init_callback_list(struct rcu_data
*rdp
)
955 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
956 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
960 * Advance this CPU's callbacks, but only if the current grace period
961 * has ended. This may be called only from the CPU to whom the rdp
962 * belongs. In addition, the corresponding leaf rcu_node structure's
963 * ->lock must be held by the caller, with irqs disabled.
966 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
968 /* Did another grace period end? */
969 if (rdp
->completed
!= rnp
->completed
) {
971 /* Advance callbacks. No harm if list empty. */
972 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
973 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
974 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
976 /* Remember that we saw this grace-period completion. */
977 rdp
->completed
= rnp
->completed
;
978 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
981 * If we were in an extended quiescent state, we may have
982 * missed some grace periods that others CPUs handled on
983 * our behalf. Catch up with this state to avoid noting
984 * spurious new grace periods. If another grace period
985 * has started, then rnp->gpnum will have advanced, so
986 * we will detect this later on.
988 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
989 rdp
->gpnum
= rdp
->completed
;
992 * If RCU does not need a quiescent state from this CPU,
993 * then make sure that this CPU doesn't go looking for one.
995 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1001 * Advance this CPU's callbacks, but only if the current grace period
1002 * has ended. This may be called only from the CPU to whom the rdp
1006 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1008 unsigned long flags
;
1009 struct rcu_node
*rnp
;
1011 local_irq_save(flags
);
1013 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1014 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1015 local_irq_restore(flags
);
1018 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1019 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1023 * Do per-CPU grace-period initialization for running CPU. The caller
1024 * must hold the lock of the leaf rcu_node structure corresponding to
1028 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1030 /* Prior grace period ended, so advance callbacks for current CPU. */
1031 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1033 /* Set state so that this CPU will detect the next quiescent state. */
1034 __note_new_gpnum(rsp
, rnp
, rdp
);
1038 * Initialize a new grace period.
1040 static int rcu_gp_init(struct rcu_state
*rsp
)
1042 struct rcu_data
*rdp
;
1043 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1045 raw_spin_lock_irq(&rnp
->lock
);
1046 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1048 if (rcu_gp_in_progress(rsp
)) {
1049 /* Grace period already in progress, don't start another. */
1050 raw_spin_unlock_irq(&rnp
->lock
);
1054 /* Advance to a new grace period and initialize state. */
1056 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1057 record_gp_stall_check_time(rsp
);
1058 raw_spin_unlock_irq(&rnp
->lock
);
1060 /* Exclude any concurrent CPU-hotplug operations. */
1064 * Set the quiescent-state-needed bits in all the rcu_node
1065 * structures for all currently online CPUs in breadth-first order,
1066 * starting from the root rcu_node structure, relying on the layout
1067 * of the tree within the rsp->node[] array. Note that other CPUs
1068 * will access only the leaves of the hierarchy, thus seeing that no
1069 * grace period is in progress, at least until the corresponding
1070 * leaf node has been initialized. In addition, we have excluded
1071 * CPU-hotplug operations.
1073 * The grace period cannot complete until the initialization
1074 * process finishes, because this kthread handles both.
1076 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1077 raw_spin_lock_irq(&rnp
->lock
);
1078 rdp
= this_cpu_ptr(rsp
->rda
);
1079 rcu_preempt_check_blocked_tasks(rnp
);
1080 rnp
->qsmask
= rnp
->qsmaskinit
;
1081 rnp
->gpnum
= rsp
->gpnum
;
1082 rnp
->completed
= rsp
->completed
;
1083 if (rnp
== rdp
->mynode
)
1084 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1085 rcu_preempt_boost_start_gp(rnp
);
1086 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1087 rnp
->level
, rnp
->grplo
,
1088 rnp
->grphi
, rnp
->qsmask
);
1089 raw_spin_unlock_irq(&rnp
->lock
);
1098 * Do one round of quiescent-state forcing.
1100 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1102 int fqs_state
= fqs_state_in
;
1103 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1106 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1107 /* Collect dyntick-idle snapshots. */
1108 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1109 fqs_state
= RCU_FORCE_QS
;
1111 /* Handle dyntick-idle and offline CPUs. */
1112 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1114 /* Clear flag to prevent immediate re-entry. */
1115 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1116 raw_spin_lock_irq(&rnp
->lock
);
1117 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1118 raw_spin_unlock_irq(&rnp
->lock
);
1124 * Clean up after the old grace period.
1126 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1128 unsigned long gp_duration
;
1129 struct rcu_data
*rdp
;
1130 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1132 raw_spin_lock_irq(&rnp
->lock
);
1133 gp_duration
= jiffies
- rsp
->gp_start
;
1134 if (gp_duration
> rsp
->gp_max
)
1135 rsp
->gp_max
= gp_duration
;
1138 * We know the grace period is complete, but to everyone else
1139 * it appears to still be ongoing. But it is also the case
1140 * that to everyone else it looks like there is nothing that
1141 * they can do to advance the grace period. It is therefore
1142 * safe for us to drop the lock in order to mark the grace
1143 * period as completed in all of the rcu_node structures.
1145 * But if this CPU needs another grace period, it will take
1146 * care of this while initializing the next grace period.
1147 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1148 * because the callbacks have not yet been advanced: Those
1149 * callbacks are waiting on the grace period that just now
1152 rdp
= this_cpu_ptr(rsp
->rda
);
1153 if (*rdp
->nxttail
[RCU_WAIT_TAIL
] == NULL
) {
1154 raw_spin_unlock_irq(&rnp
->lock
);
1157 * Propagate new ->completed value to rcu_node
1158 * structures so that other CPUs don't have to
1159 * wait until the start of the next grace period
1160 * to process their callbacks.
1162 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1163 raw_spin_lock_irq(&rnp
->lock
);
1164 rnp
->completed
= rsp
->gpnum
;
1165 raw_spin_unlock_irq(&rnp
->lock
);
1168 rnp
= rcu_get_root(rsp
);
1169 raw_spin_lock_irq(&rnp
->lock
);
1172 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1173 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1174 rsp
->fqs_state
= RCU_GP_IDLE
;
1175 if (cpu_needs_another_gp(rsp
, rdp
))
1177 raw_spin_unlock_irq(&rnp
->lock
);
1181 * Body of kthread that handles grace periods.
1183 static int __noreturn
rcu_gp_kthread(void *arg
)
1188 struct rcu_state
*rsp
= arg
;
1189 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1193 /* Handle grace-period start. */
1195 wait_event_interruptible(rsp
->gp_wq
,
1198 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1202 flush_signals(current
);
1205 /* Handle quiescent-state forcing. */
1206 fqs_state
= RCU_SAVE_DYNTICK
;
1207 j
= jiffies_till_first_fqs
;
1210 jiffies_till_first_fqs
= HZ
;
1213 rsp
->jiffies_force_qs
= jiffies
+ j
;
1214 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1215 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1216 (!ACCESS_ONCE(rnp
->qsmask
) &&
1217 !rcu_preempt_blocked_readers_cgp(rnp
)),
1219 /* If grace period done, leave loop. */
1220 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1221 !rcu_preempt_blocked_readers_cgp(rnp
))
1223 /* If time for quiescent-state forcing, do it. */
1224 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1225 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1228 /* Deal with stray signal. */
1230 flush_signals(current
);
1232 j
= jiffies_till_next_fqs
;
1235 jiffies_till_next_fqs
= HZ
;
1238 jiffies_till_next_fqs
= 1;
1242 /* Handle grace-period end. */
1243 rcu_gp_cleanup(rsp
);
1248 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1249 * in preparation for detecting the next grace period. The caller must hold
1250 * the root node's ->lock, which is released before return. Hard irqs must
1253 * Note that it is legal for a dying CPU (which is marked as offline) to
1254 * invoke this function. This can happen when the dying CPU reports its
1258 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1259 __releases(rcu_get_root(rsp
)->lock
)
1261 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1262 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1264 if (!rsp
->gp_kthread
||
1265 !cpu_needs_another_gp(rsp
, rdp
)) {
1267 * Either we have not yet spawned the grace-period
1268 * task or this CPU does not need another grace period.
1269 * Either way, don't start a new grace period.
1271 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1275 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1276 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1277 wake_up(&rsp
->gp_wq
);
1281 * Report a full set of quiescent states to the specified rcu_state
1282 * data structure. This involves cleaning up after the prior grace
1283 * period and letting rcu_start_gp() start up the next grace period
1284 * if one is needed. Note that the caller must hold rnp->lock, as
1285 * required by rcu_start_gp(), which will release it.
1287 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1288 __releases(rcu_get_root(rsp
)->lock
)
1290 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1291 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1292 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1296 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1297 * Allows quiescent states for a group of CPUs to be reported at one go
1298 * to the specified rcu_node structure, though all the CPUs in the group
1299 * must be represented by the same rcu_node structure (which need not be
1300 * a leaf rcu_node structure, though it often will be). That structure's
1301 * lock must be held upon entry, and it is released before return.
1304 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1305 struct rcu_node
*rnp
, unsigned long flags
)
1306 __releases(rnp
->lock
)
1308 struct rcu_node
*rnp_c
;
1310 /* Walk up the rcu_node hierarchy. */
1312 if (!(rnp
->qsmask
& mask
)) {
1314 /* Our bit has already been cleared, so done. */
1315 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1318 rnp
->qsmask
&= ~mask
;
1319 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1320 mask
, rnp
->qsmask
, rnp
->level
,
1321 rnp
->grplo
, rnp
->grphi
,
1323 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1325 /* Other bits still set at this level, so done. */
1326 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1329 mask
= rnp
->grpmask
;
1330 if (rnp
->parent
== NULL
) {
1332 /* No more levels. Exit loop holding root lock. */
1336 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1339 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1340 WARN_ON_ONCE(rnp_c
->qsmask
);
1344 * Get here if we are the last CPU to pass through a quiescent
1345 * state for this grace period. Invoke rcu_report_qs_rsp()
1346 * to clean up and start the next grace period if one is needed.
1348 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1352 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1353 * structure. This must be either called from the specified CPU, or
1354 * called when the specified CPU is known to be offline (and when it is
1355 * also known that no other CPU is concurrently trying to help the offline
1356 * CPU). The lastcomp argument is used to make sure we are still in the
1357 * grace period of interest. We don't want to end the current grace period
1358 * based on quiescent states detected in an earlier grace period!
1361 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastgp
)
1363 unsigned long flags
;
1365 struct rcu_node
*rnp
;
1368 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1369 if (lastgp
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
) {
1372 * The grace period in which this quiescent state was
1373 * recorded has ended, so don't report it upwards.
1374 * We will instead need a new quiescent state that lies
1375 * within the current grace period.
1377 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1378 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1381 mask
= rdp
->grpmask
;
1382 if ((rnp
->qsmask
& mask
) == 0) {
1383 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1385 rdp
->qs_pending
= 0;
1388 * This GP can't end until cpu checks in, so all of our
1389 * callbacks can be processed during the next GP.
1391 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1393 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1398 * Check to see if there is a new grace period of which this CPU
1399 * is not yet aware, and if so, set up local rcu_data state for it.
1400 * Otherwise, see if this CPU has just passed through its first
1401 * quiescent state for this grace period, and record that fact if so.
1404 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1406 /* If there is now a new grace period, record and return. */
1407 if (check_for_new_grace_period(rsp
, rdp
))
1411 * Does this CPU still need to do its part for current grace period?
1412 * If no, return and let the other CPUs do their part as well.
1414 if (!rdp
->qs_pending
)
1418 * Was there a quiescent state since the beginning of the grace
1419 * period? If no, then exit and wait for the next call.
1421 if (!rdp
->passed_quiesce
)
1425 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1428 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesce_gpnum
);
1431 #ifdef CONFIG_HOTPLUG_CPU
1434 * Send the specified CPU's RCU callbacks to the orphanage. The
1435 * specified CPU must be offline, and the caller must hold the
1439 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1440 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1443 * Orphan the callbacks. First adjust the counts. This is safe
1444 * because ->onofflock excludes _rcu_barrier()'s adoption of
1445 * the callbacks, thus no memory barrier is required.
1447 if (rdp
->nxtlist
!= NULL
) {
1448 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1449 rsp
->qlen
+= rdp
->qlen
;
1450 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1452 ACCESS_ONCE(rdp
->qlen
) = 0;
1456 * Next, move those callbacks still needing a grace period to
1457 * the orphanage, where some other CPU will pick them up.
1458 * Some of the callbacks might have gone partway through a grace
1459 * period, but that is too bad. They get to start over because we
1460 * cannot assume that grace periods are synchronized across CPUs.
1461 * We don't bother updating the ->nxttail[] array yet, instead
1462 * we just reset the whole thing later on.
1464 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1465 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1466 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1467 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1471 * Then move the ready-to-invoke callbacks to the orphanage,
1472 * where some other CPU will pick them up. These will not be
1473 * required to pass though another grace period: They are done.
1475 if (rdp
->nxtlist
!= NULL
) {
1476 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1477 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1480 /* Finally, initialize the rcu_data structure's list to empty. */
1481 init_callback_list(rdp
);
1485 * Adopt the RCU callbacks from the specified rcu_state structure's
1486 * orphanage. The caller must hold the ->onofflock.
1488 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1491 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1494 * If there is an rcu_barrier() operation in progress, then
1495 * only the task doing that operation is permitted to adopt
1496 * callbacks. To do otherwise breaks rcu_barrier() and friends
1497 * by causing them to fail to wait for the callbacks in the
1500 if (rsp
->rcu_barrier_in_progress
&&
1501 rsp
->rcu_barrier_in_progress
!= current
)
1504 /* Do the accounting first. */
1505 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1506 rdp
->qlen
+= rsp
->qlen
;
1507 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1508 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1509 rcu_idle_count_callbacks_posted();
1514 * We do not need a memory barrier here because the only way we
1515 * can get here if there is an rcu_barrier() in flight is if
1516 * we are the task doing the rcu_barrier().
1519 /* First adopt the ready-to-invoke callbacks. */
1520 if (rsp
->orphan_donelist
!= NULL
) {
1521 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1522 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1523 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1524 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1525 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1526 rsp
->orphan_donelist
= NULL
;
1527 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1530 /* And then adopt the callbacks that still need a grace period. */
1531 if (rsp
->orphan_nxtlist
!= NULL
) {
1532 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1533 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1534 rsp
->orphan_nxtlist
= NULL
;
1535 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1540 * Trace the fact that this CPU is going offline.
1542 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1544 RCU_TRACE(unsigned long mask
);
1545 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1546 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1548 RCU_TRACE(mask
= rdp
->grpmask
);
1549 trace_rcu_grace_period(rsp
->name
,
1550 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1555 * The CPU has been completely removed, and some other CPU is reporting
1556 * this fact from process context. Do the remainder of the cleanup,
1557 * including orphaning the outgoing CPU's RCU callbacks, and also
1558 * adopting them, if there is no _rcu_barrier() instance running.
1559 * There can only be one CPU hotplug operation at a time, so no other
1560 * CPU can be attempting to update rcu_cpu_kthread_task.
1562 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1564 unsigned long flags
;
1566 int need_report
= 0;
1567 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1568 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1570 /* Adjust any no-longer-needed kthreads. */
1571 rcu_stop_cpu_kthread(cpu
);
1572 rcu_node_kthread_setaffinity(rnp
, -1);
1574 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1576 /* Exclude any attempts to start a new grace period. */
1577 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1579 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1580 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1581 rcu_adopt_orphan_cbs(rsp
);
1583 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1584 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1586 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1587 rnp
->qsmaskinit
&= ~mask
;
1588 if (rnp
->qsmaskinit
!= 0) {
1589 if (rnp
!= rdp
->mynode
)
1590 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1593 if (rnp
== rdp
->mynode
)
1594 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1596 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1597 mask
= rnp
->grpmask
;
1599 } while (rnp
!= NULL
);
1602 * We still hold the leaf rcu_node structure lock here, and
1603 * irqs are still disabled. The reason for this subterfuge is
1604 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1605 * held leads to deadlock.
1607 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1609 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1610 rcu_report_unblock_qs_rnp(rnp
, flags
);
1612 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1613 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1614 rcu_report_exp_rnp(rsp
, rnp
, true);
1615 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1616 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1617 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1620 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1622 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1626 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1630 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1634 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1637 * Invoke any RCU callbacks that have made it to the end of their grace
1638 * period. Thottle as specified by rdp->blimit.
1640 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1642 unsigned long flags
;
1643 struct rcu_head
*next
, *list
, **tail
;
1644 int bl
, count
, count_lazy
, i
;
1646 /* If no callbacks are ready, just return.*/
1647 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1648 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1649 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1650 need_resched(), is_idle_task(current
),
1651 rcu_is_callbacks_kthread());
1656 * Extract the list of ready callbacks, disabling to prevent
1657 * races with call_rcu() from interrupt handlers.
1659 local_irq_save(flags
);
1660 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1662 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1663 list
= rdp
->nxtlist
;
1664 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1665 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1666 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1667 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1668 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1669 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1670 local_irq_restore(flags
);
1672 /* Invoke callbacks. */
1673 count
= count_lazy
= 0;
1677 debug_rcu_head_unqueue(list
);
1678 if (__rcu_reclaim(rsp
->name
, list
))
1681 /* Stop only if limit reached and CPU has something to do. */
1682 if (++count
>= bl
&&
1684 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1688 local_irq_save(flags
);
1689 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1690 is_idle_task(current
),
1691 rcu_is_callbacks_kthread());
1693 /* Update count, and requeue any remaining callbacks. */
1695 *tail
= rdp
->nxtlist
;
1696 rdp
->nxtlist
= list
;
1697 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1698 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1699 rdp
->nxttail
[i
] = tail
;
1703 smp_mb(); /* List handling before counting for rcu_barrier(). */
1704 rdp
->qlen_lazy
-= count_lazy
;
1705 ACCESS_ONCE(rdp
->qlen
) -= count
;
1706 rdp
->n_cbs_invoked
+= count
;
1708 /* Reinstate batch limit if we have worked down the excess. */
1709 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1710 rdp
->blimit
= blimit
;
1712 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1713 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1714 rdp
->qlen_last_fqs_check
= 0;
1715 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1716 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1717 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1718 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
1720 local_irq_restore(flags
);
1722 /* Re-invoke RCU core processing if there are callbacks remaining. */
1723 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1728 * Check to see if this CPU is in a non-context-switch quiescent state
1729 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1730 * Also schedule RCU core processing.
1732 * This function must be called from hardirq context. It is normally
1733 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1734 * false, there is no point in invoking rcu_check_callbacks().
1736 void rcu_check_callbacks(int cpu
, int user
)
1738 trace_rcu_utilization("Start scheduler-tick");
1739 increment_cpu_stall_ticks();
1740 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1743 * Get here if this CPU took its interrupt from user
1744 * mode or from the idle loop, and if this is not a
1745 * nested interrupt. In this case, the CPU is in
1746 * a quiescent state, so note it.
1748 * No memory barrier is required here because both
1749 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1750 * variables that other CPUs neither access nor modify,
1751 * at least not while the corresponding CPU is online.
1757 } else if (!in_softirq()) {
1760 * Get here if this CPU did not take its interrupt from
1761 * softirq, in other words, if it is not interrupting
1762 * a rcu_bh read-side critical section. This is an _bh
1763 * critical section, so note it.
1768 rcu_preempt_check_callbacks(cpu
);
1769 if (rcu_pending(cpu
))
1771 trace_rcu_utilization("End scheduler-tick");
1775 * Scan the leaf rcu_node structures, processing dyntick state for any that
1776 * have not yet encountered a quiescent state, using the function specified.
1777 * Also initiate boosting for any threads blocked on the root rcu_node.
1779 * The caller must have suppressed start of new grace periods.
1781 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1785 unsigned long flags
;
1787 struct rcu_node
*rnp
;
1789 rcu_for_each_leaf_node(rsp
, rnp
) {
1792 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1793 if (!rcu_gp_in_progress(rsp
)) {
1794 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1797 if (rnp
->qsmask
== 0) {
1798 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1803 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1804 if ((rnp
->qsmask
& bit
) != 0 &&
1805 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1810 /* rcu_report_qs_rnp() releases rnp->lock. */
1811 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1814 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1816 rnp
= rcu_get_root(rsp
);
1817 if (rnp
->qsmask
== 0) {
1818 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1819 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1824 * Force quiescent states on reluctant CPUs, and also detect which
1825 * CPUs are in dyntick-idle mode.
1827 static void force_quiescent_state(struct rcu_state
*rsp
)
1829 unsigned long flags
;
1831 struct rcu_node
*rnp
;
1832 struct rcu_node
*rnp_old
= NULL
;
1834 /* Funnel through hierarchy to reduce memory contention. */
1835 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
1836 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
1837 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
1838 !raw_spin_trylock(&rnp
->fqslock
);
1839 if (rnp_old
!= NULL
)
1840 raw_spin_unlock(&rnp_old
->fqslock
);
1842 rsp
->n_force_qs_lh
++;
1847 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1849 /* Reached the root of the rcu_node tree, acquire lock. */
1850 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
1851 raw_spin_unlock(&rnp_old
->fqslock
);
1852 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1853 rsp
->n_force_qs_lh
++;
1854 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1855 return; /* Someone beat us to it. */
1857 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
1858 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
1859 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1863 * This does the RCU core processing work for the specified rcu_state
1864 * and rcu_data structures. This may be called only from the CPU to
1865 * whom the rdp belongs.
1868 __rcu_process_callbacks(struct rcu_state
*rsp
)
1870 unsigned long flags
;
1871 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1873 WARN_ON_ONCE(rdp
->beenonline
== 0);
1876 * Advance callbacks in response to end of earlier grace
1877 * period that some other CPU ended.
1879 rcu_process_gp_end(rsp
, rdp
);
1881 /* Update RCU state based on any recent quiescent states. */
1882 rcu_check_quiescent_state(rsp
, rdp
);
1884 /* Does this CPU require a not-yet-started grace period? */
1885 if (cpu_needs_another_gp(rsp
, rdp
)) {
1886 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1887 rcu_start_gp(rsp
, flags
); /* releases above lock */
1890 /* If there are callbacks ready, invoke them. */
1891 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1892 invoke_rcu_callbacks(rsp
, rdp
);
1896 * Do RCU core processing for the current CPU.
1898 static void rcu_process_callbacks(struct softirq_action
*unused
)
1900 struct rcu_state
*rsp
;
1902 if (cpu_is_offline(smp_processor_id()))
1904 trace_rcu_utilization("Start RCU core");
1905 for_each_rcu_flavor(rsp
)
1906 __rcu_process_callbacks(rsp
);
1907 trace_rcu_utilization("End RCU core");
1911 * Schedule RCU callback invocation. If the specified type of RCU
1912 * does not support RCU priority boosting, just do a direct call,
1913 * otherwise wake up the per-CPU kernel kthread. Note that because we
1914 * are running on the current CPU with interrupts disabled, the
1915 * rcu_cpu_kthread_task cannot disappear out from under us.
1917 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1919 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
1921 if (likely(!rsp
->boost
)) {
1922 rcu_do_batch(rsp
, rdp
);
1925 invoke_rcu_callbacks_kthread();
1928 static void invoke_rcu_core(void)
1930 raise_softirq(RCU_SOFTIRQ
);
1934 * Handle any core-RCU processing required by a call_rcu() invocation.
1936 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
1937 struct rcu_head
*head
, unsigned long flags
)
1940 * If called from an extended quiescent state, invoke the RCU
1941 * core in order to force a re-evaluation of RCU's idleness.
1943 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1946 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1947 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
1951 * Force the grace period if too many callbacks or too long waiting.
1952 * Enforce hysteresis, and don't invoke force_quiescent_state()
1953 * if some other CPU has recently done so. Also, don't bother
1954 * invoking force_quiescent_state() if the newly enqueued callback
1955 * is the only one waiting for a grace period to complete.
1957 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1959 /* Are we ignoring a completed grace period? */
1960 rcu_process_gp_end(rsp
, rdp
);
1961 check_for_new_grace_period(rsp
, rdp
);
1963 /* Start a new grace period if one not already started. */
1964 if (!rcu_gp_in_progress(rsp
)) {
1965 unsigned long nestflag
;
1966 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1968 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1969 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1971 /* Give the grace period a kick. */
1972 rdp
->blimit
= LONG_MAX
;
1973 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1974 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1975 force_quiescent_state(rsp
);
1976 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1977 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1983 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1984 struct rcu_state
*rsp
, bool lazy
)
1986 unsigned long flags
;
1987 struct rcu_data
*rdp
;
1989 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
1990 debug_rcu_head_queue(head
);
1994 smp_mb(); /* Ensure RCU update seen before callback registry. */
1997 * Opportunistically note grace-period endings and beginnings.
1998 * Note that we might see a beginning right after we see an
1999 * end, but never vice versa, since this CPU has to pass through
2000 * a quiescent state betweentimes.
2002 local_irq_save(flags
);
2003 rdp
= this_cpu_ptr(rsp
->rda
);
2005 /* Add the callback to our list. */
2006 ACCESS_ONCE(rdp
->qlen
)++;
2010 rcu_idle_count_callbacks_posted();
2011 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2012 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2013 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2015 if (__is_kfree_rcu_offset((unsigned long)func
))
2016 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2017 rdp
->qlen_lazy
, rdp
->qlen
);
2019 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2021 /* Go handle any RCU core processing required. */
2022 __call_rcu_core(rsp
, rdp
, head
, flags
);
2023 local_irq_restore(flags
);
2027 * Queue an RCU-sched callback for invocation after a grace period.
2029 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2031 __call_rcu(head
, func
, &rcu_sched_state
, 0);
2033 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2036 * Queue an RCU callback for invocation after a quicker grace period.
2038 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2040 __call_rcu(head
, func
, &rcu_bh_state
, 0);
2042 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2045 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2046 * any blocking grace-period wait automatically implies a grace period
2047 * if there is only one CPU online at any point time during execution
2048 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2049 * occasionally incorrectly indicate that there are multiple CPUs online
2050 * when there was in fact only one the whole time, as this just adds
2051 * some overhead: RCU still operates correctly.
2053 static inline int rcu_blocking_is_gp(void)
2057 might_sleep(); /* Check for RCU read-side critical section. */
2059 ret
= num_online_cpus() <= 1;
2065 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2067 * Control will return to the caller some time after a full rcu-sched
2068 * grace period has elapsed, in other words after all currently executing
2069 * rcu-sched read-side critical sections have completed. These read-side
2070 * critical sections are delimited by rcu_read_lock_sched() and
2071 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2072 * local_irq_disable(), and so on may be used in place of
2073 * rcu_read_lock_sched().
2075 * This means that all preempt_disable code sequences, including NMI and
2076 * hardware-interrupt handlers, in progress on entry will have completed
2077 * before this primitive returns. However, this does not guarantee that
2078 * softirq handlers will have completed, since in some kernels, these
2079 * handlers can run in process context, and can block.
2081 * This primitive provides the guarantees made by the (now removed)
2082 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2083 * guarantees that rcu_read_lock() sections will have completed.
2084 * In "classic RCU", these two guarantees happen to be one and
2085 * the same, but can differ in realtime RCU implementations.
2087 void synchronize_sched(void)
2089 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2090 !lock_is_held(&rcu_lock_map
) &&
2091 !lock_is_held(&rcu_sched_lock_map
),
2092 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2093 if (rcu_blocking_is_gp())
2095 wait_rcu_gp(call_rcu_sched
);
2097 EXPORT_SYMBOL_GPL(synchronize_sched
);
2100 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2102 * Control will return to the caller some time after a full rcu_bh grace
2103 * period has elapsed, in other words after all currently executing rcu_bh
2104 * read-side critical sections have completed. RCU read-side critical
2105 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2106 * and may be nested.
2108 void synchronize_rcu_bh(void)
2110 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2111 !lock_is_held(&rcu_lock_map
) &&
2112 !lock_is_held(&rcu_sched_lock_map
),
2113 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2114 if (rcu_blocking_is_gp())
2116 wait_rcu_gp(call_rcu_bh
);
2118 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2120 static atomic_t sync_sched_expedited_started
= ATOMIC_INIT(0);
2121 static atomic_t sync_sched_expedited_done
= ATOMIC_INIT(0);
2123 static int synchronize_sched_expedited_cpu_stop(void *data
)
2126 * There must be a full memory barrier on each affected CPU
2127 * between the time that try_stop_cpus() is called and the
2128 * time that it returns.
2130 * In the current initial implementation of cpu_stop, the
2131 * above condition is already met when the control reaches
2132 * this point and the following smp_mb() is not strictly
2133 * necessary. Do smp_mb() anyway for documentation and
2134 * robustness against future implementation changes.
2136 smp_mb(); /* See above comment block. */
2141 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2143 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2144 * approach to force the grace period to end quickly. This consumes
2145 * significant time on all CPUs and is unfriendly to real-time workloads,
2146 * so is thus not recommended for any sort of common-case code. In fact,
2147 * if you are using synchronize_sched_expedited() in a loop, please
2148 * restructure your code to batch your updates, and then use a single
2149 * synchronize_sched() instead.
2151 * Note that it is illegal to call this function while holding any lock
2152 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2153 * to call this function from a CPU-hotplug notifier. Failing to observe
2154 * these restriction will result in deadlock.
2156 * This implementation can be thought of as an application of ticket
2157 * locking to RCU, with sync_sched_expedited_started and
2158 * sync_sched_expedited_done taking on the roles of the halves
2159 * of the ticket-lock word. Each task atomically increments
2160 * sync_sched_expedited_started upon entry, snapshotting the old value,
2161 * then attempts to stop all the CPUs. If this succeeds, then each
2162 * CPU will have executed a context switch, resulting in an RCU-sched
2163 * grace period. We are then done, so we use atomic_cmpxchg() to
2164 * update sync_sched_expedited_done to match our snapshot -- but
2165 * only if someone else has not already advanced past our snapshot.
2167 * On the other hand, if try_stop_cpus() fails, we check the value
2168 * of sync_sched_expedited_done. If it has advanced past our
2169 * initial snapshot, then someone else must have forced a grace period
2170 * some time after we took our snapshot. In this case, our work is
2171 * done for us, and we can simply return. Otherwise, we try again,
2172 * but keep our initial snapshot for purposes of checking for someone
2173 * doing our work for us.
2175 * If we fail too many times in a row, we fall back to synchronize_sched().
2177 void synchronize_sched_expedited(void)
2179 int firstsnap
, s
, snap
, trycount
= 0;
2181 /* Note that atomic_inc_return() implies full memory barrier. */
2182 firstsnap
= snap
= atomic_inc_return(&sync_sched_expedited_started
);
2184 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2187 * Each pass through the following loop attempts to force a
2188 * context switch on each CPU.
2190 while (try_stop_cpus(cpu_online_mask
,
2191 synchronize_sched_expedited_cpu_stop
,
2195 /* No joy, try again later. Or just synchronize_sched(). */
2196 if (trycount
++ < 10) {
2197 udelay(trycount
* num_online_cpus());
2199 synchronize_sched();
2203 /* Check to see if someone else did our work for us. */
2204 s
= atomic_read(&sync_sched_expedited_done
);
2205 if (UINT_CMP_GE((unsigned)s
, (unsigned)firstsnap
)) {
2206 smp_mb(); /* ensure test happens before caller kfree */
2211 * Refetching sync_sched_expedited_started allows later
2212 * callers to piggyback on our grace period. We subtract
2213 * 1 to get the same token that the last incrementer got.
2214 * We retry after they started, so our grace period works
2215 * for them, and they started after our first try, so their
2216 * grace period works for us.
2219 snap
= atomic_read(&sync_sched_expedited_started
);
2220 smp_mb(); /* ensure read is before try_stop_cpus(). */
2224 * Everyone up to our most recent fetch is covered by our grace
2225 * period. Update the counter, but only if our work is still
2226 * relevant -- which it won't be if someone who started later
2227 * than we did beat us to the punch.
2230 s
= atomic_read(&sync_sched_expedited_done
);
2231 if (UINT_CMP_GE((unsigned)s
, (unsigned)snap
)) {
2232 smp_mb(); /* ensure test happens before caller kfree */
2235 } while (atomic_cmpxchg(&sync_sched_expedited_done
, s
, snap
) != s
);
2239 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2242 * Check to see if there is any immediate RCU-related work to be done
2243 * by the current CPU, for the specified type of RCU, returning 1 if so.
2244 * The checks are in order of increasing expense: checks that can be
2245 * carried out against CPU-local state are performed first. However,
2246 * we must check for CPU stalls first, else we might not get a chance.
2248 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2250 struct rcu_node
*rnp
= rdp
->mynode
;
2252 rdp
->n_rcu_pending
++;
2254 /* Check for CPU stalls, if enabled. */
2255 check_cpu_stall(rsp
, rdp
);
2257 /* Is the RCU core waiting for a quiescent state from this CPU? */
2258 if (rcu_scheduler_fully_active
&&
2259 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2260 rdp
->n_rp_qs_pending
++;
2261 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2262 rdp
->n_rp_report_qs
++;
2266 /* Does this CPU have callbacks ready to invoke? */
2267 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2268 rdp
->n_rp_cb_ready
++;
2272 /* Has RCU gone idle with this CPU needing another grace period? */
2273 if (cpu_needs_another_gp(rsp
, rdp
)) {
2274 rdp
->n_rp_cpu_needs_gp
++;
2278 /* Has another RCU grace period completed? */
2279 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2280 rdp
->n_rp_gp_completed
++;
2284 /* Has a new RCU grace period started? */
2285 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2286 rdp
->n_rp_gp_started
++;
2291 rdp
->n_rp_need_nothing
++;
2296 * Check to see if there is any immediate RCU-related work to be done
2297 * by the current CPU, returning 1 if so. This function is part of the
2298 * RCU implementation; it is -not- an exported member of the RCU API.
2300 static int rcu_pending(int cpu
)
2302 struct rcu_state
*rsp
;
2304 for_each_rcu_flavor(rsp
)
2305 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2311 * Check to see if any future RCU-related work will need to be done
2312 * by the current CPU, even if none need be done immediately, returning
2315 static int rcu_cpu_has_callbacks(int cpu
)
2317 struct rcu_state
*rsp
;
2319 /* RCU callbacks either ready or pending? */
2320 for_each_rcu_flavor(rsp
)
2321 if (per_cpu_ptr(rsp
->rda
, cpu
)->nxtlist
)
2327 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2328 * the compiler is expected to optimize this away.
2330 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2331 int cpu
, unsigned long done
)
2333 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2334 atomic_read(&rsp
->barrier_cpu_count
), done
);
2338 * RCU callback function for _rcu_barrier(). If we are last, wake
2339 * up the task executing _rcu_barrier().
2341 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2343 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2344 struct rcu_state
*rsp
= rdp
->rsp
;
2346 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2347 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2348 complete(&rsp
->barrier_completion
);
2350 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2355 * Called with preemption disabled, and from cross-cpu IRQ context.
2357 static void rcu_barrier_func(void *type
)
2359 struct rcu_state
*rsp
= type
;
2360 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2362 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2363 atomic_inc(&rsp
->barrier_cpu_count
);
2364 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2368 * Orchestrate the specified type of RCU barrier, waiting for all
2369 * RCU callbacks of the specified type to complete.
2371 static void _rcu_barrier(struct rcu_state
*rsp
)
2374 unsigned long flags
;
2375 struct rcu_data
*rdp
;
2377 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2378 unsigned long snap_done
;
2380 init_rcu_head_on_stack(&rd
.barrier_head
);
2381 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2383 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2384 mutex_lock(&rsp
->barrier_mutex
);
2387 * Ensure that all prior references, including to ->n_barrier_done,
2388 * are ordered before the _rcu_barrier() machinery.
2390 smp_mb(); /* See above block comment. */
2393 * Recheck ->n_barrier_done to see if others did our work for us.
2394 * This means checking ->n_barrier_done for an even-to-odd-to-even
2395 * transition. The "if" expression below therefore rounds the old
2396 * value up to the next even number and adds two before comparing.
2398 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2399 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2400 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2401 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2402 smp_mb(); /* caller's subsequent code after above check. */
2403 mutex_unlock(&rsp
->barrier_mutex
);
2408 * Increment ->n_barrier_done to avoid duplicate work. Use
2409 * ACCESS_ONCE() to prevent the compiler from speculating
2410 * the increment to precede the early-exit check.
2412 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2413 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2414 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2415 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2418 * Initialize the count to one rather than to zero in order to
2419 * avoid a too-soon return to zero in case of a short grace period
2420 * (or preemption of this task). Also flag this task as doing
2421 * an rcu_barrier(). This will prevent anyone else from adopting
2422 * orphaned callbacks, which could cause otherwise failure if a
2423 * CPU went offline and quickly came back online. To see this,
2424 * consider the following sequence of events:
2426 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2427 * 2. CPU 1 goes offline, orphaning its callbacks.
2428 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2429 * 4. CPU 1 comes back online.
2430 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2431 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2432 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2434 init_completion(&rsp
->barrier_completion
);
2435 atomic_set(&rsp
->barrier_cpu_count
, 1);
2436 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
2437 rsp
->rcu_barrier_in_progress
= current
;
2438 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2441 * Force every CPU with callbacks to register a new callback
2442 * that will tell us when all the preceding callbacks have
2443 * been invoked. If an offline CPU has callbacks, wait for
2444 * it to either come back online or to finish orphaning those
2447 for_each_possible_cpu(cpu
) {
2449 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2450 if (cpu_is_offline(cpu
)) {
2451 _rcu_barrier_trace(rsp
, "Offline", cpu
,
2452 rsp
->n_barrier_done
);
2454 while (cpu_is_offline(cpu
) && ACCESS_ONCE(rdp
->qlen
))
2455 schedule_timeout_interruptible(1);
2456 } else if (ACCESS_ONCE(rdp
->qlen
)) {
2457 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2458 rsp
->n_barrier_done
);
2459 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2462 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2463 rsp
->n_barrier_done
);
2469 * Now that all online CPUs have rcu_barrier_callback() callbacks
2470 * posted, we can adopt all of the orphaned callbacks and place
2471 * an rcu_barrier_callback() callback after them. When that is done,
2472 * we are guaranteed to have an rcu_barrier_callback() callback
2473 * following every callback that could possibly have been
2474 * registered before _rcu_barrier() was called.
2476 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
2477 rcu_adopt_orphan_cbs(rsp
);
2478 rsp
->rcu_barrier_in_progress
= NULL
;
2479 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2480 atomic_inc(&rsp
->barrier_cpu_count
);
2481 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2483 rsp
->call(&rd
.barrier_head
, rcu_barrier_callback
);
2486 * Now that we have an rcu_barrier_callback() callback on each
2487 * CPU, and thus each counted, remove the initial count.
2489 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2490 complete(&rsp
->barrier_completion
);
2492 /* Increment ->n_barrier_done to prevent duplicate work. */
2493 smp_mb(); /* Keep increment after above mechanism. */
2494 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2495 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2496 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2497 smp_mb(); /* Keep increment before caller's subsequent code. */
2499 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2500 wait_for_completion(&rsp
->barrier_completion
);
2502 /* Other rcu_barrier() invocations can now safely proceed. */
2503 mutex_unlock(&rsp
->barrier_mutex
);
2505 destroy_rcu_head_on_stack(&rd
.barrier_head
);
2509 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2511 void rcu_barrier_bh(void)
2513 _rcu_barrier(&rcu_bh_state
);
2515 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2518 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2520 void rcu_barrier_sched(void)
2522 _rcu_barrier(&rcu_sched_state
);
2524 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2527 * Do boot-time initialization of a CPU's per-CPU RCU data.
2530 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2532 unsigned long flags
;
2533 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2534 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2536 /* Set up local state, ensuring consistent view of global state. */
2537 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2538 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2539 init_callback_list(rdp
);
2541 ACCESS_ONCE(rdp
->qlen
) = 0;
2542 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2543 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2544 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2547 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2551 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2552 * offline event can be happening at a given time. Note also that we
2553 * can accept some slop in the rsp->completed access due to the fact
2554 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2556 static void __cpuinit
2557 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2559 unsigned long flags
;
2561 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2562 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2564 /* Set up local state, ensuring consistent view of global state. */
2565 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2566 rdp
->beenonline
= 1; /* We have now been online. */
2567 rdp
->preemptible
= preemptible
;
2568 rdp
->qlen_last_fqs_check
= 0;
2569 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2570 rdp
->blimit
= blimit
;
2571 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2572 atomic_set(&rdp
->dynticks
->dynticks
,
2573 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2574 rcu_prepare_for_idle_init(cpu
);
2575 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2578 * A new grace period might start here. If so, we won't be part
2579 * of it, but that is OK, as we are currently in a quiescent state.
2582 /* Exclude any attempts to start a new GP on large systems. */
2583 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2585 /* Add CPU to rcu_node bitmasks. */
2587 mask
= rdp
->grpmask
;
2589 /* Exclude any attempts to start a new GP on small systems. */
2590 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2591 rnp
->qsmaskinit
|= mask
;
2592 mask
= rnp
->grpmask
;
2593 if (rnp
== rdp
->mynode
) {
2595 * If there is a grace period in progress, we will
2596 * set up to wait for it next time we run the
2599 rdp
->gpnum
= rnp
->completed
;
2600 rdp
->completed
= rnp
->completed
;
2601 rdp
->passed_quiesce
= 0;
2602 rdp
->qs_pending
= 0;
2603 rdp
->passed_quiesce_gpnum
= rnp
->gpnum
- 1;
2604 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2606 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2608 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2610 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2613 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2615 struct rcu_state
*rsp
;
2617 for_each_rcu_flavor(rsp
)
2618 rcu_init_percpu_data(cpu
, rsp
,
2619 strcmp(rsp
->name
, "rcu_preempt") == 0);
2623 * Handle CPU online/offline notification events.
2625 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2626 unsigned long action
, void *hcpu
)
2628 long cpu
= (long)hcpu
;
2629 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2630 struct rcu_node
*rnp
= rdp
->mynode
;
2631 struct rcu_state
*rsp
;
2633 trace_rcu_utilization("Start CPU hotplug");
2635 case CPU_UP_PREPARE
:
2636 case CPU_UP_PREPARE_FROZEN
:
2637 rcu_prepare_cpu(cpu
);
2638 rcu_prepare_kthreads(cpu
);
2641 case CPU_DOWN_FAILED
:
2642 rcu_node_kthread_setaffinity(rnp
, -1);
2643 rcu_cpu_kthread_setrt(cpu
, 1);
2645 case CPU_DOWN_PREPARE
:
2646 rcu_node_kthread_setaffinity(rnp
, cpu
);
2647 rcu_cpu_kthread_setrt(cpu
, 0);
2650 case CPU_DYING_FROZEN
:
2652 * The whole machine is "stopped" except this CPU, so we can
2653 * touch any data without introducing corruption. We send the
2654 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2656 for_each_rcu_flavor(rsp
)
2657 rcu_cleanup_dying_cpu(rsp
);
2658 rcu_cleanup_after_idle(cpu
);
2661 case CPU_DEAD_FROZEN
:
2662 case CPU_UP_CANCELED
:
2663 case CPU_UP_CANCELED_FROZEN
:
2664 for_each_rcu_flavor(rsp
)
2665 rcu_cleanup_dead_cpu(cpu
, rsp
);
2670 trace_rcu_utilization("End CPU hotplug");
2675 * Spawn the kthread that handles this RCU flavor's grace periods.
2677 static int __init
rcu_spawn_gp_kthread(void)
2679 unsigned long flags
;
2680 struct rcu_node
*rnp
;
2681 struct rcu_state
*rsp
;
2682 struct task_struct
*t
;
2684 for_each_rcu_flavor(rsp
) {
2685 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
2687 rnp
= rcu_get_root(rsp
);
2688 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2689 rsp
->gp_kthread
= t
;
2690 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2694 early_initcall(rcu_spawn_gp_kthread
);
2697 * This function is invoked towards the end of the scheduler's initialization
2698 * process. Before this is called, the idle task might contain
2699 * RCU read-side critical sections (during which time, this idle
2700 * task is booting the system). After this function is called, the
2701 * idle tasks are prohibited from containing RCU read-side critical
2702 * sections. This function also enables RCU lockdep checking.
2704 void rcu_scheduler_starting(void)
2706 WARN_ON(num_online_cpus() != 1);
2707 WARN_ON(nr_context_switches() > 0);
2708 rcu_scheduler_active
= 1;
2712 * Compute the per-level fanout, either using the exact fanout specified
2713 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2715 #ifdef CONFIG_RCU_FANOUT_EXACT
2716 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2720 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
2721 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2722 rsp
->levelspread
[0] = rcu_fanout_leaf
;
2724 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2725 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2732 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2733 ccur
= rsp
->levelcnt
[i
];
2734 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2738 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2741 * Helper function for rcu_init() that initializes one rcu_state structure.
2743 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2744 struct rcu_data __percpu
*rda
)
2746 static char *buf
[] = { "rcu_node_0",
2749 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2750 static char *fqs
[] = { "rcu_node_fqs_0",
2753 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2757 struct rcu_node
*rnp
;
2759 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2761 /* Initialize the level-tracking arrays. */
2763 for (i
= 0; i
< rcu_num_lvls
; i
++)
2764 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
2765 for (i
= 1; i
< rcu_num_lvls
; i
++)
2766 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2767 rcu_init_levelspread(rsp
);
2769 /* Initialize the elements themselves, starting from the leaves. */
2771 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
2772 cpustride
*= rsp
->levelspread
[i
];
2773 rnp
= rsp
->level
[i
];
2774 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2775 raw_spin_lock_init(&rnp
->lock
);
2776 lockdep_set_class_and_name(&rnp
->lock
,
2777 &rcu_node_class
[i
], buf
[i
]);
2778 raw_spin_lock_init(&rnp
->fqslock
);
2779 lockdep_set_class_and_name(&rnp
->fqslock
,
2780 &rcu_fqs_class
[i
], fqs
[i
]);
2783 rnp
->qsmaskinit
= 0;
2784 rnp
->grplo
= j
* cpustride
;
2785 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2786 if (rnp
->grphi
>= NR_CPUS
)
2787 rnp
->grphi
= NR_CPUS
- 1;
2793 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2794 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2795 rnp
->parent
= rsp
->level
[i
- 1] +
2796 j
/ rsp
->levelspread
[i
- 1];
2799 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2804 init_waitqueue_head(&rsp
->gp_wq
);
2805 rnp
= rsp
->level
[rcu_num_lvls
- 1];
2806 for_each_possible_cpu(i
) {
2807 while (i
> rnp
->grphi
)
2809 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2810 rcu_boot_init_percpu_data(i
, rsp
);
2812 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
2816 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2817 * replace the definitions in rcutree.h because those are needed to size
2818 * the ->node array in the rcu_state structure.
2820 static void __init
rcu_init_geometry(void)
2825 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
2827 /* If the compile-time values are accurate, just leave. */
2828 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
)
2832 * Compute number of nodes that can be handled an rcu_node tree
2833 * with the given number of levels. Setting rcu_capacity[0] makes
2834 * some of the arithmetic easier.
2836 rcu_capacity
[0] = 1;
2837 rcu_capacity
[1] = rcu_fanout_leaf
;
2838 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
2839 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
2842 * The boot-time rcu_fanout_leaf parameter is only permitted
2843 * to increase the leaf-level fanout, not decrease it. Of course,
2844 * the leaf-level fanout cannot exceed the number of bits in
2845 * the rcu_node masks. Finally, the tree must be able to accommodate
2846 * the configured number of CPUs. Complain and fall back to the
2847 * compile-time values if these limits are exceeded.
2849 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
2850 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
2851 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
2856 /* Calculate the number of rcu_nodes at each level of the tree. */
2857 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
2858 if (n
<= rcu_capacity
[i
]) {
2859 for (j
= 0; j
<= i
; j
++)
2861 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
2863 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
2868 /* Calculate the total number of rcu_node structures. */
2870 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
2871 rcu_num_nodes
+= num_rcu_lvl
[i
];
2875 void __init
rcu_init(void)
2879 rcu_bootup_announce();
2880 rcu_init_geometry();
2881 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2882 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2883 __rcu_init_preempt();
2884 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2887 * We don't need protection against CPU-hotplug here because
2888 * this is called early in boot, before either interrupts
2889 * or the scheduler are operational.
2891 cpu_notifier(rcu_cpu_notify
, 0);
2892 for_each_online_cpu(cpu
)
2893 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2894 check_cpu_stall_init();
2897 #include "rcutree_plugin.h"