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
[NUM_RCU_LVLS
];
65 #define RCU_STATE_INITIALIZER(structname) { \
66 .level = { &structname##_state.node[0] }, \
68 NUM_RCU_LVL_0, /* root of hierarchy. */ \
72 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
74 .fqs_state = RCU_GP_IDLE, \
77 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
78 .orphan_nxttail = &structname##_state.orphan_nxtlist, \
79 .orphan_donetail = &structname##_state.orphan_donelist, \
80 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
82 .n_force_qs_ngp = 0, \
83 .name = #structname, \
86 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched
);
87 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
89 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
);
90 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
92 static struct rcu_state
*rcu_state
;
95 * The rcu_scheduler_active variable transitions from zero to one just
96 * before the first task is spawned. So when this variable is zero, RCU
97 * can assume that there is but one task, allowing RCU to (for example)
98 * optimized synchronize_sched() to a simple barrier(). When this variable
99 * is one, RCU must actually do all the hard work required to detect real
100 * grace periods. This variable is also used to suppress boot-time false
101 * positives from lockdep-RCU error checking.
103 int rcu_scheduler_active __read_mostly
;
104 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
107 * The rcu_scheduler_fully_active variable transitions from zero to one
108 * during the early_initcall() processing, which is after the scheduler
109 * is capable of creating new tasks. So RCU processing (for example,
110 * creating tasks for RCU priority boosting) must be delayed until after
111 * rcu_scheduler_fully_active transitions from zero to one. We also
112 * currently delay invocation of any RCU callbacks until after this point.
114 * It might later prove better for people registering RCU callbacks during
115 * early boot to take responsibility for these callbacks, but one step at
118 static int rcu_scheduler_fully_active __read_mostly
;
120 #ifdef CONFIG_RCU_BOOST
123 * Control variables for per-CPU and per-rcu_node kthreads. These
124 * handle all flavors of RCU.
126 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
127 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
128 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
129 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
130 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
132 #endif /* #ifdef CONFIG_RCU_BOOST */
134 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
135 static void invoke_rcu_core(void);
136 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
139 * Track the rcutorture test sequence number and the update version
140 * number within a given test. The rcutorture_testseq is incremented
141 * on every rcutorture module load and unload, so has an odd value
142 * when a test is running. The rcutorture_vernum is set to zero
143 * when rcutorture starts and is incremented on each rcutorture update.
144 * These variables enable correlating rcutorture output with the
145 * RCU tracing information.
147 unsigned long rcutorture_testseq
;
148 unsigned long rcutorture_vernum
;
150 /* State information for rcu_barrier() and friends. */
152 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
153 static atomic_t rcu_barrier_cpu_count
;
154 static DEFINE_MUTEX(rcu_barrier_mutex
);
155 static struct completion rcu_barrier_completion
;
158 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
159 * permit this function to be invoked without holding the root rcu_node
160 * structure's ->lock, but of course results can be subject to change.
162 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
164 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
168 * Note a quiescent state. Because we do not need to know
169 * how many quiescent states passed, just if there was at least
170 * one since the start of the grace period, this just sets a flag.
171 * The caller must have disabled preemption.
173 void rcu_sched_qs(int cpu
)
175 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
177 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
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 rdp
->passed_quiesce_gpnum
= rdp
->gpnum
;
190 if (rdp
->passed_quiesce
== 0)
191 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
192 rdp
->passed_quiesce
= 1;
196 * Note a context switch. This is a quiescent state for RCU-sched,
197 * and requires special handling for preemptible RCU.
198 * The caller must have disabled preemption.
200 void rcu_note_context_switch(int cpu
)
202 trace_rcu_utilization("Start context switch");
204 rcu_preempt_note_context_switch(cpu
);
205 trace_rcu_utilization("End context switch");
207 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
209 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
210 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
211 .dynticks
= ATOMIC_INIT(1),
214 static int blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
215 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
216 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
218 module_param(blimit
, int, 0);
219 module_param(qhimark
, int, 0);
220 module_param(qlowmark
, int, 0);
222 int rcu_cpu_stall_suppress __read_mostly
; /* 1 = suppress stall warnings. */
223 int rcu_cpu_stall_timeout __read_mostly
= CONFIG_RCU_CPU_STALL_TIMEOUT
;
225 module_param(rcu_cpu_stall_suppress
, int, 0644);
226 module_param(rcu_cpu_stall_timeout
, int, 0644);
228 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
229 static int rcu_pending(int cpu
);
232 * Return the number of RCU-sched batches processed thus far for debug & stats.
234 long rcu_batches_completed_sched(void)
236 return rcu_sched_state
.completed
;
238 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
241 * Return the number of RCU BH batches processed thus far for debug & stats.
243 long rcu_batches_completed_bh(void)
245 return rcu_bh_state
.completed
;
247 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
250 * Force a quiescent state for RCU BH.
252 void rcu_bh_force_quiescent_state(void)
254 force_quiescent_state(&rcu_bh_state
, 0);
256 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
259 * Record the number of times rcutorture tests have been initiated and
260 * terminated. This information allows the debugfs tracing stats to be
261 * correlated to the rcutorture messages, even when the rcutorture module
262 * is being repeatedly loaded and unloaded. In other words, we cannot
263 * store this state in rcutorture itself.
265 void rcutorture_record_test_transition(void)
267 rcutorture_testseq
++;
268 rcutorture_vernum
= 0;
270 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
273 * Record the number of writer passes through the current rcutorture test.
274 * This is also used to correlate debugfs tracing stats with the rcutorture
277 void rcutorture_record_progress(unsigned long vernum
)
281 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
284 * Force a quiescent state for RCU-sched.
286 void rcu_sched_force_quiescent_state(void)
288 force_quiescent_state(&rcu_sched_state
, 0);
290 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
293 * Does the CPU have callbacks ready to be invoked?
296 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
298 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
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 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
311 * Return the root node of the specified rcu_state structure.
313 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
315 return &rsp
->node
[0];
319 * If the specified CPU is offline, tell the caller that it is in
320 * a quiescent state. Otherwise, whack it with a reschedule IPI.
321 * Grace periods can end up waiting on an offline CPU when that
322 * CPU is in the process of coming online -- it will be added to the
323 * rcu_node bitmasks before it actually makes it online. The same thing
324 * can happen while a CPU is in the process of coming online. Because this
325 * race is quite rare, we check for it after detecting that the grace
326 * period has been delayed rather than checking each and every CPU
327 * each and every time we start a new grace period.
329 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
332 * If the CPU is offline for more than a jiffy, it is in a quiescent
333 * state. We can trust its state not to change because interrupts
334 * are disabled. The reason for the jiffy's worth of slack is to
335 * handle CPUs initializing on the way up and finding their way
336 * to the idle loop on the way down.
338 if (cpu_is_offline(rdp
->cpu
) &&
339 ULONG_CMP_LT(rdp
->rsp
->gp_start
+ 2, jiffies
)) {
340 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
348 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
350 * If the new value of the ->dynticks_nesting counter now is zero,
351 * we really have entered idle, and must do the appropriate accounting.
352 * The caller must have disabled interrupts.
354 static void rcu_idle_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
)
356 trace_rcu_dyntick("Start", oldval
, 0);
357 if (!is_idle_task(current
)) {
358 struct task_struct
*idle
= idle_task(smp_processor_id());
360 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
361 ftrace_dump(DUMP_ALL
);
362 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
363 current
->pid
, current
->comm
,
364 idle
->pid
, idle
->comm
); /* must be idle task! */
366 rcu_prepare_for_idle(smp_processor_id());
367 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
368 smp_mb__before_atomic_inc(); /* See above. */
369 atomic_inc(&rdtp
->dynticks
);
370 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
371 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
374 * The idle task is not permitted to enter the idle loop while
375 * in an RCU read-side critical section.
377 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
378 "Illegal idle entry in RCU read-side critical section.");
379 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
380 "Illegal idle entry in RCU-bh read-side critical section.");
381 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
382 "Illegal idle entry in RCU-sched read-side critical section.");
386 * rcu_idle_enter - inform RCU that current CPU is entering idle
388 * Enter idle mode, in other words, -leave- the mode in which RCU
389 * read-side critical sections can occur. (Though RCU read-side
390 * critical sections can occur in irq handlers in idle, a possibility
391 * handled by irq_enter() and irq_exit().)
393 * We crowbar the ->dynticks_nesting field to zero to allow for
394 * the possibility of usermode upcalls having messed up our count
395 * of interrupt nesting level during the prior busy period.
397 void rcu_idle_enter(void)
401 struct rcu_dynticks
*rdtp
;
403 local_irq_save(flags
);
404 rdtp
= &__get_cpu_var(rcu_dynticks
);
405 oldval
= rdtp
->dynticks_nesting
;
406 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
407 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
408 rdtp
->dynticks_nesting
= 0;
410 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
411 rcu_idle_enter_common(rdtp
, oldval
);
412 local_irq_restore(flags
);
414 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
417 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
419 * Exit from an interrupt handler, which might possibly result in entering
420 * idle mode, in other words, leaving the mode in which read-side critical
421 * sections can occur.
423 * This code assumes that the idle loop never does anything that might
424 * result in unbalanced calls to irq_enter() and irq_exit(). If your
425 * architecture violates this assumption, RCU will give you what you
426 * deserve, good and hard. But very infrequently and irreproducibly.
428 * Use things like work queues to work around this limitation.
430 * You have been warned.
432 void rcu_irq_exit(void)
436 struct rcu_dynticks
*rdtp
;
438 local_irq_save(flags
);
439 rdtp
= &__get_cpu_var(rcu_dynticks
);
440 oldval
= rdtp
->dynticks_nesting
;
441 rdtp
->dynticks_nesting
--;
442 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
443 if (rdtp
->dynticks_nesting
)
444 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
446 rcu_idle_enter_common(rdtp
, oldval
);
447 local_irq_restore(flags
);
451 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
453 * If the new value of the ->dynticks_nesting counter was previously zero,
454 * we really have exited idle, and must do the appropriate accounting.
455 * The caller must have disabled interrupts.
457 static void rcu_idle_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
)
459 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
460 atomic_inc(&rdtp
->dynticks
);
461 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
462 smp_mb__after_atomic_inc(); /* See above. */
463 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
464 rcu_cleanup_after_idle(smp_processor_id());
465 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
466 if (!is_idle_task(current
)) {
467 struct task_struct
*idle
= idle_task(smp_processor_id());
469 trace_rcu_dyntick("Error on exit: not idle task",
470 oldval
, rdtp
->dynticks_nesting
);
471 ftrace_dump(DUMP_ALL
);
472 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
473 current
->pid
, current
->comm
,
474 idle
->pid
, idle
->comm
); /* must be idle task! */
479 * rcu_idle_exit - inform RCU that current CPU is leaving idle
481 * Exit idle mode, in other words, -enter- the mode in which RCU
482 * read-side critical sections can occur.
484 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
485 * allow for the possibility of usermode upcalls messing up our count
486 * of interrupt nesting level during the busy period that is just
489 void rcu_idle_exit(void)
492 struct rcu_dynticks
*rdtp
;
495 local_irq_save(flags
);
496 rdtp
= &__get_cpu_var(rcu_dynticks
);
497 oldval
= rdtp
->dynticks_nesting
;
498 WARN_ON_ONCE(oldval
< 0);
499 if (oldval
& DYNTICK_TASK_NEST_MASK
)
500 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
502 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
503 rcu_idle_exit_common(rdtp
, oldval
);
504 local_irq_restore(flags
);
506 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
509 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
511 * Enter an interrupt handler, which might possibly result in exiting
512 * idle mode, in other words, entering the mode in which read-side critical
513 * sections can occur.
515 * Note that the Linux kernel is fully capable of entering an interrupt
516 * handler that it never exits, for example when doing upcalls to
517 * user mode! This code assumes that the idle loop never does upcalls to
518 * user mode. If your architecture does do upcalls from the idle loop (or
519 * does anything else that results in unbalanced calls to the irq_enter()
520 * and irq_exit() functions), RCU will give you what you deserve, good
521 * and hard. But very infrequently and irreproducibly.
523 * Use things like work queues to work around this limitation.
525 * You have been warned.
527 void rcu_irq_enter(void)
530 struct rcu_dynticks
*rdtp
;
533 local_irq_save(flags
);
534 rdtp
= &__get_cpu_var(rcu_dynticks
);
535 oldval
= rdtp
->dynticks_nesting
;
536 rdtp
->dynticks_nesting
++;
537 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
539 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
541 rcu_idle_exit_common(rdtp
, oldval
);
542 local_irq_restore(flags
);
546 * rcu_nmi_enter - inform RCU of entry to NMI context
548 * If the CPU was idle with dynamic ticks active, and there is no
549 * irq handler running, this updates rdtp->dynticks_nmi to let the
550 * RCU grace-period handling know that the CPU is active.
552 void rcu_nmi_enter(void)
554 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
556 if (rdtp
->dynticks_nmi_nesting
== 0 &&
557 (atomic_read(&rdtp
->dynticks
) & 0x1))
559 rdtp
->dynticks_nmi_nesting
++;
560 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
561 atomic_inc(&rdtp
->dynticks
);
562 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
563 smp_mb__after_atomic_inc(); /* See above. */
564 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
568 * rcu_nmi_exit - inform RCU of exit from NMI context
570 * If the CPU was idle with dynamic ticks active, and there is no
571 * irq handler running, this updates rdtp->dynticks_nmi to let the
572 * RCU grace-period handling know that the CPU is no longer active.
574 void rcu_nmi_exit(void)
576 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
578 if (rdtp
->dynticks_nmi_nesting
== 0 ||
579 --rdtp
->dynticks_nmi_nesting
!= 0)
581 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
582 smp_mb__before_atomic_inc(); /* See above. */
583 atomic_inc(&rdtp
->dynticks
);
584 smp_mb__after_atomic_inc(); /* Force delay to next write. */
585 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
589 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
591 * If the current CPU is in its idle loop and is neither in an interrupt
592 * or NMI handler, return true.
594 int rcu_is_cpu_idle(void)
599 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
603 EXPORT_SYMBOL(rcu_is_cpu_idle
);
605 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
608 * Is the current CPU online? Disable preemption to avoid false positives
609 * that could otherwise happen due to the current CPU number being sampled,
610 * this task being preempted, its old CPU being taken offline, resuming
611 * on some other CPU, then determining that its old CPU is now offline.
612 * It is OK to use RCU on an offline processor during initial boot, hence
613 * the check for rcu_scheduler_fully_active. Note also that it is OK
614 * for a CPU coming online to use RCU for one jiffy prior to marking itself
615 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
616 * offline to continue to use RCU for one jiffy after marking itself
617 * offline in the cpu_online_mask. This leniency is necessary given the
618 * non-atomic nature of the online and offline processing, for example,
619 * the fact that a CPU enters the scheduler after completing the CPU_DYING
622 * This is also why RCU internally marks CPUs online during the
623 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
625 * Disable checking if in an NMI handler because we cannot safely report
626 * errors from NMI handlers anyway.
628 bool rcu_lockdep_current_cpu_online(void)
630 struct rcu_data
*rdp
;
631 struct rcu_node
*rnp
;
637 rdp
= &__get_cpu_var(rcu_sched_data
);
639 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
640 !rcu_scheduler_fully_active
;
644 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
646 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
649 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
651 * If the current CPU is idle or running at a first-level (not nested)
652 * interrupt from idle, return true. The caller must have at least
653 * disabled preemption.
655 int rcu_is_cpu_rrupt_from_idle(void)
657 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
661 * Snapshot the specified CPU's dynticks counter so that we can later
662 * credit them with an implicit quiescent state. Return 1 if this CPU
663 * is in dynticks idle mode, which is an extended quiescent state.
665 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
667 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
668 return (rdp
->dynticks_snap
& 0x1) == 0;
672 * Return true if the specified CPU has passed through a quiescent
673 * state by virtue of being in or having passed through an dynticks
674 * idle state since the last call to dyntick_save_progress_counter()
677 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
682 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
683 snap
= (unsigned int)rdp
->dynticks_snap
;
686 * If the CPU passed through or entered a dynticks idle phase with
687 * no active irq/NMI handlers, then we can safely pretend that the CPU
688 * already acknowledged the request to pass through a quiescent
689 * state. Either way, that CPU cannot possibly be in an RCU
690 * read-side critical section that started before the beginning
691 * of the current RCU grace period.
693 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
694 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
699 /* Go check for the CPU being offline. */
700 return rcu_implicit_offline_qs(rdp
);
703 static int jiffies_till_stall_check(void)
705 int till_stall_check
= ACCESS_ONCE(rcu_cpu_stall_timeout
);
708 * Limit check must be consistent with the Kconfig limits
709 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
711 if (till_stall_check
< 3) {
712 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 3;
713 till_stall_check
= 3;
714 } else if (till_stall_check
> 300) {
715 ACCESS_ONCE(rcu_cpu_stall_timeout
) = 300;
716 till_stall_check
= 300;
718 return till_stall_check
* HZ
+ RCU_STALL_DELAY_DELTA
;
721 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
723 rsp
->gp_start
= jiffies
;
724 rsp
->jiffies_stall
= jiffies
+ jiffies_till_stall_check();
727 static void print_other_cpu_stall(struct rcu_state
*rsp
)
733 struct rcu_node
*rnp
= rcu_get_root(rsp
);
735 /* Only let one CPU complain about others per time interval. */
737 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
738 delta
= jiffies
- rsp
->jiffies_stall
;
739 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
740 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
743 rsp
->jiffies_stall
= jiffies
+ 3 * jiffies_till_stall_check() + 3;
744 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
747 * OK, time to rat on our buddy...
748 * See Documentation/RCU/stallwarn.txt for info on how to debug
749 * RCU CPU stall warnings.
751 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
753 print_cpu_stall_info_begin();
754 rcu_for_each_leaf_node(rsp
, rnp
) {
755 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
756 ndetected
+= rcu_print_task_stall(rnp
);
757 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
758 if (rnp
->qsmask
== 0)
760 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
761 if (rnp
->qsmask
& (1UL << cpu
)) {
762 print_cpu_stall_info(rsp
, rnp
->grplo
+ cpu
);
768 * Now rat on any tasks that got kicked up to the root rcu_node
769 * due to CPU offlining.
771 rnp
= rcu_get_root(rsp
);
772 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
773 ndetected
+= rcu_print_task_stall(rnp
);
774 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
776 print_cpu_stall_info_end();
777 printk(KERN_CONT
"(detected by %d, t=%ld jiffies)\n",
778 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
780 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
781 else if (!trigger_all_cpu_backtrace())
784 /* If so configured, complain about tasks blocking the grace period. */
786 rcu_print_detail_task_stall(rsp
);
788 force_quiescent_state(rsp
, 0); /* Kick them all. */
791 static void print_cpu_stall(struct rcu_state
*rsp
)
794 struct rcu_node
*rnp
= rcu_get_root(rsp
);
797 * OK, time to rat on ourselves...
798 * See Documentation/RCU/stallwarn.txt for info on how to debug
799 * RCU CPU stall warnings.
801 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
802 print_cpu_stall_info_begin();
803 print_cpu_stall_info(rsp
, smp_processor_id());
804 print_cpu_stall_info_end();
805 printk(KERN_CONT
" (t=%lu jiffies)\n", jiffies
- rsp
->gp_start
);
806 if (!trigger_all_cpu_backtrace())
809 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
810 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
811 rsp
->jiffies_stall
= jiffies
+
812 3 * jiffies_till_stall_check() + 3;
813 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
815 set_need_resched(); /* kick ourselves to get things going. */
818 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
822 struct rcu_node
*rnp
;
824 if (rcu_cpu_stall_suppress
)
826 j
= ACCESS_ONCE(jiffies
);
827 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
829 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
831 /* We haven't checked in, so go dump stack. */
832 print_cpu_stall(rsp
);
834 } else if (rcu_gp_in_progress(rsp
) &&
835 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
837 /* They had a few time units to dump stack, so complain. */
838 print_other_cpu_stall(rsp
);
842 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
844 rcu_cpu_stall_suppress
= 1;
849 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
851 * Set the stall-warning timeout way off into the future, thus preventing
852 * any RCU CPU stall-warning messages from appearing in the current set of
855 * The caller must disable hard irqs.
857 void rcu_cpu_stall_reset(void)
859 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
860 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
861 rcu_preempt_stall_reset();
864 static struct notifier_block rcu_panic_block
= {
865 .notifier_call
= rcu_panic
,
868 static void __init
check_cpu_stall_init(void)
870 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
874 * Update CPU-local rcu_data state to record the newly noticed grace period.
875 * This is used both when we started the grace period and when we notice
876 * that someone else started the grace period. The caller must hold the
877 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
878 * and must have irqs disabled.
880 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
882 if (rdp
->gpnum
!= rnp
->gpnum
) {
884 * If the current grace period is waiting for this CPU,
885 * set up to detect a quiescent state, otherwise don't
886 * go looking for one.
888 rdp
->gpnum
= rnp
->gpnum
;
889 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
890 if (rnp
->qsmask
& rdp
->grpmask
) {
892 rdp
->passed_quiesce
= 0;
895 zero_cpu_stall_ticks(rdp
);
899 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
902 struct rcu_node
*rnp
;
904 local_irq_save(flags
);
906 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
907 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
908 local_irq_restore(flags
);
911 __note_new_gpnum(rsp
, rnp
, rdp
);
912 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
916 * Did someone else start a new RCU grace period start since we last
917 * checked? Update local state appropriately if so. Must be called
918 * on the CPU corresponding to rdp.
921 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
926 local_irq_save(flags
);
927 if (rdp
->gpnum
!= rsp
->gpnum
) {
928 note_new_gpnum(rsp
, rdp
);
931 local_irq_restore(flags
);
936 * Initialize the specified rcu_data structure's callback list to empty.
938 static void init_callback_list(struct rcu_data
*rdp
)
943 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
944 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
948 * Advance this CPU's callbacks, but only if the current grace period
949 * has ended. This may be called only from the CPU to whom the rdp
950 * belongs. In addition, the corresponding leaf rcu_node structure's
951 * ->lock must be held by the caller, with irqs disabled.
954 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
956 /* Did another grace period end? */
957 if (rdp
->completed
!= rnp
->completed
) {
959 /* Advance callbacks. No harm if list empty. */
960 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
961 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
962 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
964 /* Remember that we saw this grace-period completion. */
965 rdp
->completed
= rnp
->completed
;
966 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
969 * If we were in an extended quiescent state, we may have
970 * missed some grace periods that others CPUs handled on
971 * our behalf. Catch up with this state to avoid noting
972 * spurious new grace periods. If another grace period
973 * has started, then rnp->gpnum will have advanced, so
974 * we will detect this later on.
976 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
977 rdp
->gpnum
= rdp
->completed
;
980 * If RCU does not need a quiescent state from this CPU,
981 * then make sure that this CPU doesn't go looking for one.
983 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
989 * Advance this CPU's callbacks, but only if the current grace period
990 * has ended. This may be called only from the CPU to whom the rdp
994 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
997 struct rcu_node
*rnp
;
999 local_irq_save(flags
);
1001 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1002 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1003 local_irq_restore(flags
);
1006 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1007 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1011 * Do per-CPU grace-period initialization for running CPU. The caller
1012 * must hold the lock of the leaf rcu_node structure corresponding to
1016 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1018 /* Prior grace period ended, so advance callbacks for current CPU. */
1019 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1022 * Because this CPU just now started the new grace period, we know
1023 * that all of its callbacks will be covered by this upcoming grace
1024 * period, even the ones that were registered arbitrarily recently.
1025 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1027 * Other CPUs cannot be sure exactly when the grace period started.
1028 * Therefore, their recently registered callbacks must pass through
1029 * an additional RCU_NEXT_READY stage, so that they will be handled
1030 * by the next RCU grace period.
1032 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1033 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1035 /* Set state so that this CPU will detect the next quiescent state. */
1036 __note_new_gpnum(rsp
, rnp
, rdp
);
1040 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1041 * in preparation for detecting the next grace period. The caller must hold
1042 * the root node's ->lock, which is released before return. Hard irqs must
1045 * Note that it is legal for a dying CPU (which is marked as offline) to
1046 * invoke this function. This can happen when the dying CPU reports its
1050 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
1051 __releases(rcu_get_root(rsp
)->lock
)
1053 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1054 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1056 if (!rcu_scheduler_fully_active
||
1057 !cpu_needs_another_gp(rsp
, rdp
)) {
1059 * Either the scheduler hasn't yet spawned the first
1060 * non-idle task or this CPU does not need another
1061 * grace period. Either way, don't start a new grace
1064 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1068 if (rsp
->fqs_active
) {
1070 * This CPU needs a grace period, but force_quiescent_state()
1071 * is running. Tell it to start one on this CPU's behalf.
1073 rsp
->fqs_need_gp
= 1;
1074 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1078 /* Advance to a new grace period and initialize state. */
1080 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1081 WARN_ON_ONCE(rsp
->fqs_state
== RCU_GP_INIT
);
1082 rsp
->fqs_state
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
1083 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1084 record_gp_stall_check_time(rsp
);
1085 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
1087 /* Exclude any concurrent CPU-hotplug operations. */
1088 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
1091 * Set the quiescent-state-needed bits in all the rcu_node
1092 * structures for all currently online CPUs in breadth-first
1093 * order, starting from the root rcu_node structure. This
1094 * operation relies on the layout of the hierarchy within the
1095 * rsp->node[] array. Note that other CPUs will access only
1096 * the leaves of the hierarchy, which still indicate that no
1097 * grace period is in progress, at least until the corresponding
1098 * leaf node has been initialized. In addition, we have excluded
1099 * CPU-hotplug operations.
1101 * Note that the grace period cannot complete until we finish
1102 * the initialization process, as there will be at least one
1103 * qsmask bit set in the root node until that time, namely the
1104 * one corresponding to this CPU, due to the fact that we have
1107 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1108 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1109 rcu_preempt_check_blocked_tasks(rnp
);
1110 rnp
->qsmask
= rnp
->qsmaskinit
;
1111 rnp
->gpnum
= rsp
->gpnum
;
1112 rnp
->completed
= rsp
->completed
;
1113 if (rnp
== rdp
->mynode
)
1114 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1115 rcu_preempt_boost_start_gp(rnp
);
1116 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1117 rnp
->level
, rnp
->grplo
,
1118 rnp
->grphi
, rnp
->qsmask
);
1119 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1122 rnp
= rcu_get_root(rsp
);
1123 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1124 rsp
->fqs_state
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
1125 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1126 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
1130 * Report a full set of quiescent states to the specified rcu_state
1131 * data structure. This involves cleaning up after the prior grace
1132 * period and letting rcu_start_gp() start up the next grace period
1133 * if one is needed. Note that the caller must hold rnp->lock, as
1134 * required by rcu_start_gp(), which will release it.
1136 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1137 __releases(rcu_get_root(rsp
)->lock
)
1139 unsigned long gp_duration
;
1140 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1141 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1143 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1146 * Ensure that all grace-period and pre-grace-period activity
1147 * is seen before the assignment to rsp->completed.
1149 smp_mb(); /* See above block comment. */
1150 gp_duration
= jiffies
- rsp
->gp_start
;
1151 if (gp_duration
> rsp
->gp_max
)
1152 rsp
->gp_max
= gp_duration
;
1155 * We know the grace period is complete, but to everyone else
1156 * it appears to still be ongoing. But it is also the case
1157 * that to everyone else it looks like there is nothing that
1158 * they can do to advance the grace period. It is therefore
1159 * safe for us to drop the lock in order to mark the grace
1160 * period as completed in all of the rcu_node structures.
1162 * But if this CPU needs another grace period, it will take
1163 * care of this while initializing the next grace period.
1164 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1165 * because the callbacks have not yet been advanced: Those
1166 * callbacks are waiting on the grace period that just now
1169 if (*rdp
->nxttail
[RCU_WAIT_TAIL
] == NULL
) {
1170 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1173 * Propagate new ->completed value to rcu_node structures
1174 * so that other CPUs don't have to wait until the start
1175 * of the next grace period to process their callbacks.
1177 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1178 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1179 rnp
->completed
= rsp
->gpnum
;
1180 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1182 rnp
= rcu_get_root(rsp
);
1183 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1186 rsp
->completed
= rsp
->gpnum
; /* Declare the grace period complete. */
1187 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1188 rsp
->fqs_state
= RCU_GP_IDLE
;
1189 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
1193 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1194 * Allows quiescent states for a group of CPUs to be reported at one go
1195 * to the specified rcu_node structure, though all the CPUs in the group
1196 * must be represented by the same rcu_node structure (which need not be
1197 * a leaf rcu_node structure, though it often will be). That structure's
1198 * lock must be held upon entry, and it is released before return.
1201 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1202 struct rcu_node
*rnp
, unsigned long flags
)
1203 __releases(rnp
->lock
)
1205 struct rcu_node
*rnp_c
;
1207 /* Walk up the rcu_node hierarchy. */
1209 if (!(rnp
->qsmask
& mask
)) {
1211 /* Our bit has already been cleared, so done. */
1212 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1215 rnp
->qsmask
&= ~mask
;
1216 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1217 mask
, rnp
->qsmask
, rnp
->level
,
1218 rnp
->grplo
, rnp
->grphi
,
1220 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1222 /* Other bits still set at this level, so done. */
1223 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1226 mask
= rnp
->grpmask
;
1227 if (rnp
->parent
== NULL
) {
1229 /* No more levels. Exit loop holding root lock. */
1233 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1236 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1237 WARN_ON_ONCE(rnp_c
->qsmask
);
1241 * Get here if we are the last CPU to pass through a quiescent
1242 * state for this grace period. Invoke rcu_report_qs_rsp()
1243 * to clean up and start the next grace period if one is needed.
1245 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1249 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1250 * structure. This must be either called from the specified CPU, or
1251 * called when the specified CPU is known to be offline (and when it is
1252 * also known that no other CPU is concurrently trying to help the offline
1253 * CPU). The lastcomp argument is used to make sure we are still in the
1254 * grace period of interest. We don't want to end the current grace period
1255 * based on quiescent states detected in an earlier grace period!
1258 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastgp
)
1260 unsigned long flags
;
1262 struct rcu_node
*rnp
;
1265 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1266 if (lastgp
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
) {
1269 * The grace period in which this quiescent state was
1270 * recorded has ended, so don't report it upwards.
1271 * We will instead need a new quiescent state that lies
1272 * within the current grace period.
1274 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1275 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1278 mask
= rdp
->grpmask
;
1279 if ((rnp
->qsmask
& mask
) == 0) {
1280 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1282 rdp
->qs_pending
= 0;
1285 * This GP can't end until cpu checks in, so all of our
1286 * callbacks can be processed during the next GP.
1288 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1290 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1295 * Check to see if there is a new grace period of which this CPU
1296 * is not yet aware, and if so, set up local rcu_data state for it.
1297 * Otherwise, see if this CPU has just passed through its first
1298 * quiescent state for this grace period, and record that fact if so.
1301 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1303 /* If there is now a new grace period, record and return. */
1304 if (check_for_new_grace_period(rsp
, rdp
))
1308 * Does this CPU still need to do its part for current grace period?
1309 * If no, return and let the other CPUs do their part as well.
1311 if (!rdp
->qs_pending
)
1315 * Was there a quiescent state since the beginning of the grace
1316 * period? If no, then exit and wait for the next call.
1318 if (!rdp
->passed_quiesce
)
1322 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1325 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesce_gpnum
);
1328 #ifdef CONFIG_HOTPLUG_CPU
1331 * Send the specified CPU's RCU callbacks to the orphanage. The
1332 * specified CPU must be offline, and the caller must hold the
1336 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1337 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1340 * Orphan the callbacks. First adjust the counts. This is safe
1341 * because ->onofflock excludes _rcu_barrier()'s adoption of
1342 * the callbacks, thus no memory barrier is required.
1344 if (rdp
->nxtlist
!= NULL
) {
1345 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1346 rsp
->qlen
+= rdp
->qlen
;
1347 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1349 ACCESS_ONCE(rdp
->qlen
) = 0;
1353 * Next, move those callbacks still needing a grace period to
1354 * the orphanage, where some other CPU will pick them up.
1355 * Some of the callbacks might have gone partway through a grace
1356 * period, but that is too bad. They get to start over because we
1357 * cannot assume that grace periods are synchronized across CPUs.
1358 * We don't bother updating the ->nxttail[] array yet, instead
1359 * we just reset the whole thing later on.
1361 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1362 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1363 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1364 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1368 * Then move the ready-to-invoke callbacks to the orphanage,
1369 * where some other CPU will pick them up. These will not be
1370 * required to pass though another grace period: They are done.
1372 if (rdp
->nxtlist
!= NULL
) {
1373 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1374 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1377 /* Finally, initialize the rcu_data structure's list to empty. */
1378 init_callback_list(rdp
);
1382 * Adopt the RCU callbacks from the specified rcu_state structure's
1383 * orphanage. The caller must hold the ->onofflock.
1385 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1388 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1391 * If there is an rcu_barrier() operation in progress, then
1392 * only the task doing that operation is permitted to adopt
1393 * callbacks. To do otherwise breaks rcu_barrier() and friends
1394 * by causing them to fail to wait for the callbacks in the
1397 if (rsp
->rcu_barrier_in_progress
&&
1398 rsp
->rcu_barrier_in_progress
!= current
)
1401 /* Do the accounting first. */
1402 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1403 rdp
->qlen
+= rsp
->qlen
;
1404 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1405 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1406 rcu_idle_count_callbacks_posted();
1411 * We do not need a memory barrier here because the only way we
1412 * can get here if there is an rcu_barrier() in flight is if
1413 * we are the task doing the rcu_barrier().
1416 /* First adopt the ready-to-invoke callbacks. */
1417 if (rsp
->orphan_donelist
!= NULL
) {
1418 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1419 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1420 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1421 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1422 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1423 rsp
->orphan_donelist
= NULL
;
1424 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1427 /* And then adopt the callbacks that still need a grace period. */
1428 if (rsp
->orphan_nxtlist
!= NULL
) {
1429 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1430 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1431 rsp
->orphan_nxtlist
= NULL
;
1432 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1437 * Trace the fact that this CPU is going offline.
1439 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1441 RCU_TRACE(unsigned long mask
);
1442 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1443 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1445 RCU_TRACE(mask
= rdp
->grpmask
);
1446 trace_rcu_grace_period(rsp
->name
,
1447 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1452 * The CPU has been completely removed, and some other CPU is reporting
1453 * this fact from process context. Do the remainder of the cleanup,
1454 * including orphaning the outgoing CPU's RCU callbacks, and also
1455 * adopting them, if there is no _rcu_barrier() instance running.
1456 * There can only be one CPU hotplug operation at a time, so no other
1457 * CPU can be attempting to update rcu_cpu_kthread_task.
1459 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1461 unsigned long flags
;
1463 int need_report
= 0;
1464 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1465 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1467 /* Adjust any no-longer-needed kthreads. */
1468 rcu_stop_cpu_kthread(cpu
);
1469 rcu_node_kthread_setaffinity(rnp
, -1);
1471 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1473 /* Exclude any attempts to start a new grace period. */
1474 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1476 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1477 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1478 rcu_adopt_orphan_cbs(rsp
);
1480 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1481 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1483 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1484 rnp
->qsmaskinit
&= ~mask
;
1485 if (rnp
->qsmaskinit
!= 0) {
1486 if (rnp
!= rdp
->mynode
)
1487 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1490 if (rnp
== rdp
->mynode
)
1491 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1493 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1494 mask
= rnp
->grpmask
;
1496 } while (rnp
!= NULL
);
1499 * We still hold the leaf rcu_node structure lock here, and
1500 * irqs are still disabled. The reason for this subterfuge is
1501 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1502 * held leads to deadlock.
1504 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1506 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1507 rcu_report_unblock_qs_rnp(rnp
, flags
);
1509 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1510 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1511 rcu_report_exp_rnp(rsp
, rnp
, true);
1514 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1516 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1520 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1524 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1528 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1531 * Invoke any RCU callbacks that have made it to the end of their grace
1532 * period. Thottle as specified by rdp->blimit.
1534 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1536 unsigned long flags
;
1537 struct rcu_head
*next
, *list
, **tail
;
1538 int bl
, count
, count_lazy
, i
;
1540 /* If no callbacks are ready, just return.*/
1541 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1542 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1543 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1544 need_resched(), is_idle_task(current
),
1545 rcu_is_callbacks_kthread());
1550 * Extract the list of ready callbacks, disabling to prevent
1551 * races with call_rcu() from interrupt handlers.
1553 local_irq_save(flags
);
1554 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1556 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1557 list
= rdp
->nxtlist
;
1558 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1559 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1560 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1561 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1562 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1563 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1564 local_irq_restore(flags
);
1566 /* Invoke callbacks. */
1567 count
= count_lazy
= 0;
1571 debug_rcu_head_unqueue(list
);
1572 if (__rcu_reclaim(rsp
->name
, list
))
1575 /* Stop only if limit reached and CPU has something to do. */
1576 if (++count
>= bl
&&
1578 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
1582 local_irq_save(flags
);
1583 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
1584 is_idle_task(current
),
1585 rcu_is_callbacks_kthread());
1587 /* Update count, and requeue any remaining callbacks. */
1589 *tail
= rdp
->nxtlist
;
1590 rdp
->nxtlist
= list
;
1591 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1592 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
1593 rdp
->nxttail
[i
] = tail
;
1597 smp_mb(); /* List handling before counting for rcu_barrier(). */
1598 rdp
->qlen_lazy
-= count_lazy
;
1599 ACCESS_ONCE(rdp
->qlen
) -= count
;
1600 rdp
->n_cbs_invoked
+= count
;
1602 /* Reinstate batch limit if we have worked down the excess. */
1603 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1604 rdp
->blimit
= blimit
;
1606 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1607 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1608 rdp
->qlen_last_fqs_check
= 0;
1609 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1610 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1611 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1613 local_irq_restore(flags
);
1615 /* Re-invoke RCU core processing if there are callbacks remaining. */
1616 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1621 * Check to see if this CPU is in a non-context-switch quiescent state
1622 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1623 * Also schedule RCU core processing.
1625 * This function must be called from hardirq context. It is normally
1626 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1627 * false, there is no point in invoking rcu_check_callbacks().
1629 void rcu_check_callbacks(int cpu
, int user
)
1631 trace_rcu_utilization("Start scheduler-tick");
1632 increment_cpu_stall_ticks();
1633 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
1636 * Get here if this CPU took its interrupt from user
1637 * mode or from the idle loop, and if this is not a
1638 * nested interrupt. In this case, the CPU is in
1639 * a quiescent state, so note it.
1641 * No memory barrier is required here because both
1642 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1643 * variables that other CPUs neither access nor modify,
1644 * at least not while the corresponding CPU is online.
1650 } else if (!in_softirq()) {
1653 * Get here if this CPU did not take its interrupt from
1654 * softirq, in other words, if it is not interrupting
1655 * a rcu_bh read-side critical section. This is an _bh
1656 * critical section, so note it.
1661 rcu_preempt_check_callbacks(cpu
);
1662 if (rcu_pending(cpu
))
1664 trace_rcu_utilization("End scheduler-tick");
1668 * Scan the leaf rcu_node structures, processing dyntick state for any that
1669 * have not yet encountered a quiescent state, using the function specified.
1670 * Also initiate boosting for any threads blocked on the root rcu_node.
1672 * The caller must have suppressed start of new grace periods.
1674 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1678 unsigned long flags
;
1680 struct rcu_node
*rnp
;
1682 rcu_for_each_leaf_node(rsp
, rnp
) {
1684 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1685 if (!rcu_gp_in_progress(rsp
)) {
1686 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1689 if (rnp
->qsmask
== 0) {
1690 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1695 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1696 if ((rnp
->qsmask
& bit
) != 0 &&
1697 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1702 /* rcu_report_qs_rnp() releases rnp->lock. */
1703 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1706 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1708 rnp
= rcu_get_root(rsp
);
1709 if (rnp
->qsmask
== 0) {
1710 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1711 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1716 * Force quiescent states on reluctant CPUs, and also detect which
1717 * CPUs are in dyntick-idle mode.
1719 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1721 unsigned long flags
;
1722 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1724 trace_rcu_utilization("Start fqs");
1725 if (!rcu_gp_in_progress(rsp
)) {
1726 trace_rcu_utilization("End fqs");
1727 return; /* No grace period in progress, nothing to force. */
1729 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1730 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1731 trace_rcu_utilization("End fqs");
1732 return; /* Someone else is already on the job. */
1734 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1735 goto unlock_fqs_ret
; /* no emergency and done recently. */
1737 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1738 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1739 if(!rcu_gp_in_progress(rsp
)) {
1740 rsp
->n_force_qs_ngp
++;
1741 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1742 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1744 rsp
->fqs_active
= 1;
1745 switch (rsp
->fqs_state
) {
1749 break; /* grace period idle or initializing, ignore. */
1751 case RCU_SAVE_DYNTICK
:
1752 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1753 break; /* So gcc recognizes the dead code. */
1755 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1757 /* Record dyntick-idle state. */
1758 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1759 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1760 if (rcu_gp_in_progress(rsp
))
1761 rsp
->fqs_state
= RCU_FORCE_QS
;
1766 /* Check dyntick-idle state, send IPI to laggarts. */
1767 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1768 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1770 /* Leave state in case more forcing is required. */
1772 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1775 rsp
->fqs_active
= 0;
1776 if (rsp
->fqs_need_gp
) {
1777 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1778 rsp
->fqs_need_gp
= 0;
1779 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1780 trace_rcu_utilization("End fqs");
1783 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1785 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1786 trace_rcu_utilization("End fqs");
1790 * This does the RCU core processing work for the specified rcu_state
1791 * and rcu_data structures. This may be called only from the CPU to
1792 * whom the rdp belongs.
1795 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1797 unsigned long flags
;
1799 WARN_ON_ONCE(rdp
->beenonline
== 0);
1802 * If an RCU GP has gone long enough, go check for dyntick
1803 * idle CPUs and, if needed, send resched IPIs.
1805 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1806 force_quiescent_state(rsp
, 1);
1809 * Advance callbacks in response to end of earlier grace
1810 * period that some other CPU ended.
1812 rcu_process_gp_end(rsp
, rdp
);
1814 /* Update RCU state based on any recent quiescent states. */
1815 rcu_check_quiescent_state(rsp
, rdp
);
1817 /* Does this CPU require a not-yet-started grace period? */
1818 if (cpu_needs_another_gp(rsp
, rdp
)) {
1819 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1820 rcu_start_gp(rsp
, flags
); /* releases above lock */
1823 /* If there are callbacks ready, invoke them. */
1824 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1825 invoke_rcu_callbacks(rsp
, rdp
);
1829 * Do RCU core processing for the current CPU.
1831 static void rcu_process_callbacks(struct softirq_action
*unused
)
1833 trace_rcu_utilization("Start RCU core");
1834 __rcu_process_callbacks(&rcu_sched_state
,
1835 &__get_cpu_var(rcu_sched_data
));
1836 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1837 rcu_preempt_process_callbacks();
1838 trace_rcu_utilization("End RCU core");
1842 * Schedule RCU callback invocation. If the specified type of RCU
1843 * does not support RCU priority boosting, just do a direct call,
1844 * otherwise wake up the per-CPU kernel kthread. Note that because we
1845 * are running on the current CPU with interrupts disabled, the
1846 * rcu_cpu_kthread_task cannot disappear out from under us.
1848 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1850 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
1852 if (likely(!rsp
->boost
)) {
1853 rcu_do_batch(rsp
, rdp
);
1856 invoke_rcu_callbacks_kthread();
1859 static void invoke_rcu_core(void)
1861 raise_softirq(RCU_SOFTIRQ
);
1865 * Handle any core-RCU processing required by a call_rcu() invocation.
1867 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
1868 struct rcu_head
*head
, unsigned long flags
)
1871 * If called from an extended quiescent state, invoke the RCU
1872 * core in order to force a re-evaluation of RCU's idleness.
1874 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1877 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1878 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
1882 * Force the grace period if too many callbacks or too long waiting.
1883 * Enforce hysteresis, and don't invoke force_quiescent_state()
1884 * if some other CPU has recently done so. Also, don't bother
1885 * invoking force_quiescent_state() if the newly enqueued callback
1886 * is the only one waiting for a grace period to complete.
1888 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1890 /* Are we ignoring a completed grace period? */
1891 rcu_process_gp_end(rsp
, rdp
);
1892 check_for_new_grace_period(rsp
, rdp
);
1894 /* Start a new grace period if one not already started. */
1895 if (!rcu_gp_in_progress(rsp
)) {
1896 unsigned long nestflag
;
1897 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1899 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1900 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1902 /* Give the grace period a kick. */
1903 rdp
->blimit
= LONG_MAX
;
1904 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1905 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1906 force_quiescent_state(rsp
, 0);
1907 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1908 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1910 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1911 force_quiescent_state(rsp
, 1);
1915 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1916 struct rcu_state
*rsp
, bool lazy
)
1918 unsigned long flags
;
1919 struct rcu_data
*rdp
;
1921 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
1922 debug_rcu_head_queue(head
);
1926 smp_mb(); /* Ensure RCU update seen before callback registry. */
1929 * Opportunistically note grace-period endings and beginnings.
1930 * Note that we might see a beginning right after we see an
1931 * end, but never vice versa, since this CPU has to pass through
1932 * a quiescent state betweentimes.
1934 local_irq_save(flags
);
1935 rdp
= this_cpu_ptr(rsp
->rda
);
1937 /* Add the callback to our list. */
1938 ACCESS_ONCE(rdp
->qlen
)++;
1942 rcu_idle_count_callbacks_posted();
1943 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1944 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1945 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1947 if (__is_kfree_rcu_offset((unsigned long)func
))
1948 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
1949 rdp
->qlen_lazy
, rdp
->qlen
);
1951 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
1953 /* Go handle any RCU core processing required. */
1954 __call_rcu_core(rsp
, rdp
, head
, flags
);
1955 local_irq_restore(flags
);
1959 * Queue an RCU-sched callback for invocation after a grace period.
1961 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1963 __call_rcu(head
, func
, &rcu_sched_state
, 0);
1965 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1968 * Queue an RCU callback for invocation after a quicker grace period.
1970 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1972 __call_rcu(head
, func
, &rcu_bh_state
, 0);
1974 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1977 * Because a context switch is a grace period for RCU-sched and RCU-bh,
1978 * any blocking grace-period wait automatically implies a grace period
1979 * if there is only one CPU online at any point time during execution
1980 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
1981 * occasionally incorrectly indicate that there are multiple CPUs online
1982 * when there was in fact only one the whole time, as this just adds
1983 * some overhead: RCU still operates correctly.
1985 static inline int rcu_blocking_is_gp(void)
1989 might_sleep(); /* Check for RCU read-side critical section. */
1991 ret
= num_online_cpus() <= 1;
1997 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1999 * Control will return to the caller some time after a full rcu-sched
2000 * grace period has elapsed, in other words after all currently executing
2001 * rcu-sched read-side critical sections have completed. These read-side
2002 * critical sections are delimited by rcu_read_lock_sched() and
2003 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2004 * local_irq_disable(), and so on may be used in place of
2005 * rcu_read_lock_sched().
2007 * This means that all preempt_disable code sequences, including NMI and
2008 * hardware-interrupt handlers, in progress on entry will have completed
2009 * before this primitive returns. However, this does not guarantee that
2010 * softirq handlers will have completed, since in some kernels, these
2011 * handlers can run in process context, and can block.
2013 * This primitive provides the guarantees made by the (now removed)
2014 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2015 * guarantees that rcu_read_lock() sections will have completed.
2016 * In "classic RCU", these two guarantees happen to be one and
2017 * the same, but can differ in realtime RCU implementations.
2019 void synchronize_sched(void)
2021 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2022 !lock_is_held(&rcu_lock_map
) &&
2023 !lock_is_held(&rcu_sched_lock_map
),
2024 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2025 if (rcu_blocking_is_gp())
2027 wait_rcu_gp(call_rcu_sched
);
2029 EXPORT_SYMBOL_GPL(synchronize_sched
);
2032 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2034 * Control will return to the caller some time after a full rcu_bh grace
2035 * period has elapsed, in other words after all currently executing rcu_bh
2036 * read-side critical sections have completed. RCU read-side critical
2037 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2038 * and may be nested.
2040 void synchronize_rcu_bh(void)
2042 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2043 !lock_is_held(&rcu_lock_map
) &&
2044 !lock_is_held(&rcu_sched_lock_map
),
2045 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2046 if (rcu_blocking_is_gp())
2048 wait_rcu_gp(call_rcu_bh
);
2050 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2052 static atomic_t sync_sched_expedited_started
= ATOMIC_INIT(0);
2053 static atomic_t sync_sched_expedited_done
= ATOMIC_INIT(0);
2055 static int synchronize_sched_expedited_cpu_stop(void *data
)
2058 * There must be a full memory barrier on each affected CPU
2059 * between the time that try_stop_cpus() is called and the
2060 * time that it returns.
2062 * In the current initial implementation of cpu_stop, the
2063 * above condition is already met when the control reaches
2064 * this point and the following smp_mb() is not strictly
2065 * necessary. Do smp_mb() anyway for documentation and
2066 * robustness against future implementation changes.
2068 smp_mb(); /* See above comment block. */
2073 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2075 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2076 * approach to force the grace period to end quickly. This consumes
2077 * significant time on all CPUs and is unfriendly to real-time workloads,
2078 * so is thus not recommended for any sort of common-case code. In fact,
2079 * if you are using synchronize_sched_expedited() in a loop, please
2080 * restructure your code to batch your updates, and then use a single
2081 * synchronize_sched() instead.
2083 * Note that it is illegal to call this function while holding any lock
2084 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2085 * to call this function from a CPU-hotplug notifier. Failing to observe
2086 * these restriction will result in deadlock.
2088 * This implementation can be thought of as an application of ticket
2089 * locking to RCU, with sync_sched_expedited_started and
2090 * sync_sched_expedited_done taking on the roles of the halves
2091 * of the ticket-lock word. Each task atomically increments
2092 * sync_sched_expedited_started upon entry, snapshotting the old value,
2093 * then attempts to stop all the CPUs. If this succeeds, then each
2094 * CPU will have executed a context switch, resulting in an RCU-sched
2095 * grace period. We are then done, so we use atomic_cmpxchg() to
2096 * update sync_sched_expedited_done to match our snapshot -- but
2097 * only if someone else has not already advanced past our snapshot.
2099 * On the other hand, if try_stop_cpus() fails, we check the value
2100 * of sync_sched_expedited_done. If it has advanced past our
2101 * initial snapshot, then someone else must have forced a grace period
2102 * some time after we took our snapshot. In this case, our work is
2103 * done for us, and we can simply return. Otherwise, we try again,
2104 * but keep our initial snapshot for purposes of checking for someone
2105 * doing our work for us.
2107 * If we fail too many times in a row, we fall back to synchronize_sched().
2109 void synchronize_sched_expedited(void)
2111 int firstsnap
, s
, snap
, trycount
= 0;
2113 /* Note that atomic_inc_return() implies full memory barrier. */
2114 firstsnap
= snap
= atomic_inc_return(&sync_sched_expedited_started
);
2116 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2119 * Each pass through the following loop attempts to force a
2120 * context switch on each CPU.
2122 while (try_stop_cpus(cpu_online_mask
,
2123 synchronize_sched_expedited_cpu_stop
,
2127 /* No joy, try again later. Or just synchronize_sched(). */
2128 if (trycount
++ < 10)
2129 udelay(trycount
* num_online_cpus());
2131 synchronize_sched();
2135 /* Check to see if someone else did our work for us. */
2136 s
= atomic_read(&sync_sched_expedited_done
);
2137 if (UINT_CMP_GE((unsigned)s
, (unsigned)firstsnap
)) {
2138 smp_mb(); /* ensure test happens before caller kfree */
2143 * Refetching sync_sched_expedited_started allows later
2144 * callers to piggyback on our grace period. We subtract
2145 * 1 to get the same token that the last incrementer got.
2146 * We retry after they started, so our grace period works
2147 * for them, and they started after our first try, so their
2148 * grace period works for us.
2151 snap
= atomic_read(&sync_sched_expedited_started
);
2152 smp_mb(); /* ensure read is before try_stop_cpus(). */
2156 * Everyone up to our most recent fetch is covered by our grace
2157 * period. Update the counter, but only if our work is still
2158 * relevant -- which it won't be if someone who started later
2159 * than we did beat us to the punch.
2162 s
= atomic_read(&sync_sched_expedited_done
);
2163 if (UINT_CMP_GE((unsigned)s
, (unsigned)snap
)) {
2164 smp_mb(); /* ensure test happens before caller kfree */
2167 } while (atomic_cmpxchg(&sync_sched_expedited_done
, s
, snap
) != s
);
2171 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2174 * Check to see if there is any immediate RCU-related work to be done
2175 * by the current CPU, for the specified type of RCU, returning 1 if so.
2176 * The checks are in order of increasing expense: checks that can be
2177 * carried out against CPU-local state are performed first. However,
2178 * we must check for CPU stalls first, else we might not get a chance.
2180 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2182 struct rcu_node
*rnp
= rdp
->mynode
;
2184 rdp
->n_rcu_pending
++;
2186 /* Check for CPU stalls, if enabled. */
2187 check_cpu_stall(rsp
, rdp
);
2189 /* Is the RCU core waiting for a quiescent state from this CPU? */
2190 if (rcu_scheduler_fully_active
&&
2191 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2194 * If force_quiescent_state() coming soon and this CPU
2195 * needs a quiescent state, and this is either RCU-sched
2196 * or RCU-bh, force a local reschedule.
2198 rdp
->n_rp_qs_pending
++;
2199 if (!rdp
->preemptible
&&
2200 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
2203 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2204 rdp
->n_rp_report_qs
++;
2208 /* Does this CPU have callbacks ready to invoke? */
2209 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2210 rdp
->n_rp_cb_ready
++;
2214 /* Has RCU gone idle with this CPU needing another grace period? */
2215 if (cpu_needs_another_gp(rsp
, rdp
)) {
2216 rdp
->n_rp_cpu_needs_gp
++;
2220 /* Has another RCU grace period completed? */
2221 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2222 rdp
->n_rp_gp_completed
++;
2226 /* Has a new RCU grace period started? */
2227 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2228 rdp
->n_rp_gp_started
++;
2232 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2233 if (rcu_gp_in_progress(rsp
) &&
2234 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
2235 rdp
->n_rp_need_fqs
++;
2240 rdp
->n_rp_need_nothing
++;
2245 * Check to see if there is any immediate RCU-related work to be done
2246 * by the current CPU, returning 1 if so. This function is part of the
2247 * RCU implementation; it is -not- an exported member of the RCU API.
2249 static int rcu_pending(int cpu
)
2251 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
2252 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
2253 rcu_preempt_pending(cpu
);
2257 * Check to see if any future RCU-related work will need to be done
2258 * by the current CPU, even if none need be done immediately, returning
2261 static int rcu_cpu_has_callbacks(int cpu
)
2263 /* RCU callbacks either ready or pending? */
2264 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
2265 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
2266 rcu_preempt_cpu_has_callbacks(cpu
);
2270 * RCU callback function for _rcu_barrier(). If we are last, wake
2271 * up the task executing _rcu_barrier().
2273 static void rcu_barrier_callback(struct rcu_head
*notused
)
2275 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2276 complete(&rcu_barrier_completion
);
2280 * Called with preemption disabled, and from cross-cpu IRQ context.
2282 static void rcu_barrier_func(void *type
)
2284 int cpu
= smp_processor_id();
2285 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
2286 void (*call_rcu_func
)(struct rcu_head
*head
,
2287 void (*func
)(struct rcu_head
*head
));
2289 atomic_inc(&rcu_barrier_cpu_count
);
2290 call_rcu_func
= type
;
2291 call_rcu_func(head
, rcu_barrier_callback
);
2295 * Orchestrate the specified type of RCU barrier, waiting for all
2296 * RCU callbacks of the specified type to complete.
2298 static void _rcu_barrier(struct rcu_state
*rsp
,
2299 void (*call_rcu_func
)(struct rcu_head
*head
,
2300 void (*func
)(struct rcu_head
*head
)))
2303 unsigned long flags
;
2304 struct rcu_data
*rdp
;
2307 init_rcu_head_on_stack(&rh
);
2309 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2310 mutex_lock(&rcu_barrier_mutex
);
2312 smp_mb(); /* Prevent any prior operations from leaking in. */
2315 * Initialize the count to one rather than to zero in order to
2316 * avoid a too-soon return to zero in case of a short grace period
2317 * (or preemption of this task). Also flag this task as doing
2318 * an rcu_barrier(). This will prevent anyone else from adopting
2319 * orphaned callbacks, which could cause otherwise failure if a
2320 * CPU went offline and quickly came back online. To see this,
2321 * consider the following sequence of events:
2323 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2324 * 2. CPU 1 goes offline, orphaning its callbacks.
2325 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2326 * 4. CPU 1 comes back online.
2327 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2328 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2329 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2331 init_completion(&rcu_barrier_completion
);
2332 atomic_set(&rcu_barrier_cpu_count
, 1);
2333 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
2334 rsp
->rcu_barrier_in_progress
= current
;
2335 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2338 * Force every CPU with callbacks to register a new callback
2339 * that will tell us when all the preceding callbacks have
2340 * been invoked. If an offline CPU has callbacks, wait for
2341 * it to either come back online or to finish orphaning those
2344 for_each_possible_cpu(cpu
) {
2346 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2347 if (cpu_is_offline(cpu
)) {
2349 while (cpu_is_offline(cpu
) && ACCESS_ONCE(rdp
->qlen
))
2350 schedule_timeout_interruptible(1);
2351 } else if (ACCESS_ONCE(rdp
->qlen
)) {
2352 smp_call_function_single(cpu
, rcu_barrier_func
,
2353 (void *)call_rcu_func
, 1);
2361 * Now that all online CPUs have rcu_barrier_callback() callbacks
2362 * posted, we can adopt all of the orphaned callbacks and place
2363 * an rcu_barrier_callback() callback after them. When that is done,
2364 * we are guaranteed to have an rcu_barrier_callback() callback
2365 * following every callback that could possibly have been
2366 * registered before _rcu_barrier() was called.
2368 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
2369 rcu_adopt_orphan_cbs(rsp
);
2370 rsp
->rcu_barrier_in_progress
= NULL
;
2371 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2372 atomic_inc(&rcu_barrier_cpu_count
);
2373 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2374 call_rcu_func(&rh
, rcu_barrier_callback
);
2377 * Now that we have an rcu_barrier_callback() callback on each
2378 * CPU, and thus each counted, remove the initial count.
2380 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
2381 complete(&rcu_barrier_completion
);
2383 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2384 wait_for_completion(&rcu_barrier_completion
);
2386 /* Other rcu_barrier() invocations can now safely proceed. */
2387 mutex_unlock(&rcu_barrier_mutex
);
2389 destroy_rcu_head_on_stack(&rh
);
2393 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2395 void rcu_barrier_bh(void)
2397 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
2399 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2402 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2404 void rcu_barrier_sched(void)
2406 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
2408 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2411 * Do boot-time initialization of a CPU's per-CPU RCU data.
2414 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2416 unsigned long flags
;
2417 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2418 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2420 /* Set up local state, ensuring consistent view of global state. */
2421 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2422 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2423 init_callback_list(rdp
);
2425 ACCESS_ONCE(rdp
->qlen
) = 0;
2426 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2427 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2428 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2431 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2435 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2436 * offline event can be happening at a given time. Note also that we
2437 * can accept some slop in the rsp->completed access due to the fact
2438 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2440 static void __cpuinit
2441 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2443 unsigned long flags
;
2445 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2446 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2448 /* Set up local state, ensuring consistent view of global state. */
2449 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2450 rdp
->beenonline
= 1; /* We have now been online. */
2451 rdp
->preemptible
= preemptible
;
2452 rdp
->qlen_last_fqs_check
= 0;
2453 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2454 rdp
->blimit
= blimit
;
2455 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2456 atomic_set(&rdp
->dynticks
->dynticks
,
2457 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2458 rcu_prepare_for_idle_init(cpu
);
2459 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2462 * A new grace period might start here. If so, we won't be part
2463 * of it, but that is OK, as we are currently in a quiescent state.
2466 /* Exclude any attempts to start a new GP on large systems. */
2467 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
2469 /* Add CPU to rcu_node bitmasks. */
2471 mask
= rdp
->grpmask
;
2473 /* Exclude any attempts to start a new GP on small systems. */
2474 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2475 rnp
->qsmaskinit
|= mask
;
2476 mask
= rnp
->grpmask
;
2477 if (rnp
== rdp
->mynode
) {
2479 * If there is a grace period in progress, we will
2480 * set up to wait for it next time we run the
2483 rdp
->gpnum
= rnp
->completed
;
2484 rdp
->completed
= rnp
->completed
;
2485 rdp
->passed_quiesce
= 0;
2486 rdp
->qs_pending
= 0;
2487 rdp
->passed_quiesce_gpnum
= rnp
->gpnum
- 1;
2488 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
2490 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2492 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2494 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
2497 static void __cpuinit
rcu_prepare_cpu(int cpu
)
2499 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
2500 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
2501 rcu_preempt_init_percpu_data(cpu
);
2505 * Handle CPU online/offline notification events.
2507 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
2508 unsigned long action
, void *hcpu
)
2510 long cpu
= (long)hcpu
;
2511 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2512 struct rcu_node
*rnp
= rdp
->mynode
;
2514 trace_rcu_utilization("Start CPU hotplug");
2516 case CPU_UP_PREPARE
:
2517 case CPU_UP_PREPARE_FROZEN
:
2518 rcu_prepare_cpu(cpu
);
2519 rcu_prepare_kthreads(cpu
);
2522 case CPU_DOWN_FAILED
:
2523 rcu_node_kthread_setaffinity(rnp
, -1);
2524 rcu_cpu_kthread_setrt(cpu
, 1);
2526 case CPU_DOWN_PREPARE
:
2527 rcu_node_kthread_setaffinity(rnp
, cpu
);
2528 rcu_cpu_kthread_setrt(cpu
, 0);
2531 case CPU_DYING_FROZEN
:
2533 * The whole machine is "stopped" except this CPU, so we can
2534 * touch any data without introducing corruption. We send the
2535 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2537 rcu_cleanup_dying_cpu(&rcu_bh_state
);
2538 rcu_cleanup_dying_cpu(&rcu_sched_state
);
2539 rcu_preempt_cleanup_dying_cpu();
2540 rcu_cleanup_after_idle(cpu
);
2543 case CPU_DEAD_FROZEN
:
2544 case CPU_UP_CANCELED
:
2545 case CPU_UP_CANCELED_FROZEN
:
2546 rcu_cleanup_dead_cpu(cpu
, &rcu_bh_state
);
2547 rcu_cleanup_dead_cpu(cpu
, &rcu_sched_state
);
2548 rcu_preempt_cleanup_dead_cpu(cpu
);
2553 trace_rcu_utilization("End CPU hotplug");
2558 * This function is invoked towards the end of the scheduler's initialization
2559 * process. Before this is called, the idle task might contain
2560 * RCU read-side critical sections (during which time, this idle
2561 * task is booting the system). After this function is called, the
2562 * idle tasks are prohibited from containing RCU read-side critical
2563 * sections. This function also enables RCU lockdep checking.
2565 void rcu_scheduler_starting(void)
2567 WARN_ON(num_online_cpus() != 1);
2568 WARN_ON(nr_context_switches() > 0);
2569 rcu_scheduler_active
= 1;
2573 * Compute the per-level fanout, either using the exact fanout specified
2574 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2576 #ifdef CONFIG_RCU_FANOUT_EXACT
2577 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2581 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
2582 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
2583 rsp
->levelspread
[0] = CONFIG_RCU_FANOUT_LEAF
;
2585 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2586 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2593 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2594 ccur
= rsp
->levelcnt
[i
];
2595 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2599 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2602 * Helper function for rcu_init() that initializes one rcu_state structure.
2604 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2605 struct rcu_data __percpu
*rda
)
2607 static char *buf
[] = { "rcu_node_level_0",
2610 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2614 struct rcu_node
*rnp
;
2616 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2618 /* Initialize the level-tracking arrays. */
2620 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2621 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2622 rcu_init_levelspread(rsp
);
2624 /* Initialize the elements themselves, starting from the leaves. */
2626 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2627 cpustride
*= rsp
->levelspread
[i
];
2628 rnp
= rsp
->level
[i
];
2629 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2630 raw_spin_lock_init(&rnp
->lock
);
2631 lockdep_set_class_and_name(&rnp
->lock
,
2632 &rcu_node_class
[i
], buf
[i
]);
2635 rnp
->qsmaskinit
= 0;
2636 rnp
->grplo
= j
* cpustride
;
2637 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2638 if (rnp
->grphi
>= NR_CPUS
)
2639 rnp
->grphi
= NR_CPUS
- 1;
2645 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2646 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2647 rnp
->parent
= rsp
->level
[i
- 1] +
2648 j
/ rsp
->levelspread
[i
- 1];
2651 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2656 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2657 for_each_possible_cpu(i
) {
2658 while (i
> rnp
->grphi
)
2660 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2661 rcu_boot_init_percpu_data(i
, rsp
);
2665 void __init
rcu_init(void)
2669 rcu_bootup_announce();
2670 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2671 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2672 __rcu_init_preempt();
2673 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2676 * We don't need protection against CPU-hotplug here because
2677 * this is called early in boot, before either interrupts
2678 * or the scheduler are operational.
2680 cpu_notifier(rcu_cpu_notify
, 0);
2681 for_each_online_cpu(cpu
)
2682 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2683 check_cpu_stall_init();
2686 #include "rcutree_plugin.h"