2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
55 #include <trace/events/rcu.h>
59 /* Data structures. */
61 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
63 #define RCU_STATE_INITIALIZER(structname) { \
64 .level = { &structname##_state.node[0] }, \
66 NUM_RCU_LVL_0, /* root of hierarchy. */ \
70 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
72 .fqs_state = RCU_GP_IDLE, \
75 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
76 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
78 .n_force_qs_ngp = 0, \
79 .name = #structname, \
82 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched);
83 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
85 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh);
86 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
88 static struct rcu_state *rcu_state;
91 * The rcu_scheduler_active variable transitions from zero to one just
92 * before the first task is spawned. So when this variable is zero, RCU
93 * can assume that there is but one task, allowing RCU to (for example)
94 * optimized synchronize_sched() to a simple barrier(). When this variable
95 * is one, RCU must actually do all the hard work required to detect real
96 * grace periods. This variable is also used to suppress boot-time false
97 * positives from lockdep-RCU error checking.
99 int rcu_scheduler_active __read_mostly;
100 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
103 * The rcu_scheduler_fully_active variable transitions from zero to one
104 * during the early_initcall() processing, which is after the scheduler
105 * is capable of creating new tasks. So RCU processing (for example,
106 * creating tasks for RCU priority boosting) must be delayed until after
107 * rcu_scheduler_fully_active transitions from zero to one. We also
108 * currently delay invocation of any RCU callbacks until after this point.
110 * It might later prove better for people registering RCU callbacks during
111 * early boot to take responsibility for these callbacks, but one step at
114 static int rcu_scheduler_fully_active __read_mostly;
116 #ifdef CONFIG_RCU_BOOST
119 * Control variables for per-CPU and per-rcu_node kthreads. These
120 * handle all flavors of RCU.
122 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
123 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
124 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
125 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
126 DEFINE_PER_CPU(char, rcu_cpu_has_work);
128 #endif /* #ifdef CONFIG_RCU_BOOST */
130 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
131 static void invoke_rcu_core(void);
132 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
135 * Track the rcutorture test sequence number and the update version
136 * number within a given test. The rcutorture_testseq is incremented
137 * on every rcutorture module load and unload, so has an odd value
138 * when a test is running. The rcutorture_vernum is set to zero
139 * when rcutorture starts and is incremented on each rcutorture update.
140 * These variables enable correlating rcutorture output with the
141 * RCU tracing information.
143 unsigned long rcutorture_testseq;
144 unsigned long rcutorture_vernum;
147 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
148 * permit this function to be invoked without holding the root rcu_node
149 * structure's ->lock, but of course results can be subject to change.
151 static int rcu_gp_in_progress(struct rcu_state *rsp)
153 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
157 * Note a quiescent state. Because we do not need to know
158 * how many quiescent states passed, just if there was at least
159 * one since the start of the grace period, this just sets a flag.
160 * The caller must have disabled preemption.
162 void rcu_sched_qs(int cpu)
164 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
166 rdp->passed_quiesce_gpnum = rdp->gpnum;
168 if (rdp->passed_quiesce == 0)
169 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
170 rdp->passed_quiesce = 1;
173 void rcu_bh_qs(int cpu)
175 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
177 rdp->passed_quiesce_gpnum = rdp->gpnum;
179 if (rdp->passed_quiesce == 0)
180 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
181 rdp->passed_quiesce = 1;
185 * Note a context switch. This is a quiescent state for RCU-sched,
186 * and requires special handling for preemptible RCU.
187 * The caller must have disabled preemption.
189 void rcu_note_context_switch(int cpu)
191 trace_rcu_utilization("Start context switch");
193 rcu_preempt_note_context_switch(cpu);
194 trace_rcu_utilization("End context switch");
196 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
198 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
199 .dynticks_nesting = LLONG_MAX / 2,
200 .dynticks = ATOMIC_INIT(1),
203 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
204 static int qhimark = 10000; /* If this many pending, ignore blimit. */
205 static int qlowmark = 100; /* Once only this many pending, use blimit. */
207 module_param(blimit, int, 0);
208 module_param(qhimark, int, 0);
209 module_param(qlowmark, int, 0);
211 int rcu_cpu_stall_suppress __read_mostly;
212 module_param(rcu_cpu_stall_suppress, int, 0644);
214 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
215 static int rcu_pending(int cpu);
218 * Return the number of RCU-sched batches processed thus far for debug & stats.
220 long rcu_batches_completed_sched(void)
222 return rcu_sched_state.completed;
224 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
227 * Return the number of RCU BH batches processed thus far for debug & stats.
229 long rcu_batches_completed_bh(void)
231 return rcu_bh_state.completed;
233 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
236 * Force a quiescent state for RCU BH.
238 void rcu_bh_force_quiescent_state(void)
240 force_quiescent_state(&rcu_bh_state, 0);
242 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
245 * Record the number of times rcutorture tests have been initiated and
246 * terminated. This information allows the debugfs tracing stats to be
247 * correlated to the rcutorture messages, even when the rcutorture module
248 * is being repeatedly loaded and unloaded. In other words, we cannot
249 * store this state in rcutorture itself.
251 void rcutorture_record_test_transition(void)
253 rcutorture_testseq++;
254 rcutorture_vernum = 0;
256 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
259 * Record the number of writer passes through the current rcutorture test.
260 * This is also used to correlate debugfs tracing stats with the rcutorture
263 void rcutorture_record_progress(unsigned long vernum)
267 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
270 * Force a quiescent state for RCU-sched.
272 void rcu_sched_force_quiescent_state(void)
274 force_quiescent_state(&rcu_sched_state, 0);
276 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
279 * Does the CPU have callbacks ready to be invoked?
282 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
284 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
288 * Does the current CPU require a yet-as-unscheduled grace period?
291 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
293 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
297 * Return the root node of the specified rcu_state structure.
299 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
301 return &rsp->node[0];
307 * If the specified CPU is offline, tell the caller that it is in
308 * a quiescent state. Otherwise, whack it with a reschedule IPI.
309 * Grace periods can end up waiting on an offline CPU when that
310 * CPU is in the process of coming online -- it will be added to the
311 * rcu_node bitmasks before it actually makes it online. The same thing
312 * can happen while a CPU is in the process of coming online. Because this
313 * race is quite rare, we check for it after detecting that the grace
314 * period has been delayed rather than checking each and every CPU
315 * each and every time we start a new grace period.
317 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
320 * If the CPU is offline, it is in a quiescent state. We can
321 * trust its state not to change because interrupts are disabled.
323 if (cpu_is_offline(rdp->cpu)) {
324 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
330 * The CPU is online, so send it a reschedule IPI. This forces
331 * it through the scheduler, and (inefficiently) also handles cases
332 * where idle loops fail to inform RCU about the CPU being idle.
334 if (rdp->cpu != smp_processor_id())
335 smp_send_reschedule(rdp->cpu);
342 #endif /* #ifdef CONFIG_SMP */
345 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
347 * If the new value of the ->dynticks_nesting counter now is zero,
348 * we really have entered idle, and must do the appropriate accounting.
349 * The caller must have disabled interrupts.
351 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp)
353 if (rdtp->dynticks_nesting) {
354 trace_rcu_dyntick("--=", rdtp->dynticks_nesting);
357 trace_rcu_dyntick("Start", rdtp->dynticks_nesting);
358 if (!idle_cpu(smp_processor_id())) {
359 WARN_ON_ONCE(1); /* must be idle task! */
360 trace_rcu_dyntick("Error on entry: not idle task",
361 rdtp->dynticks_nesting);
362 ftrace_dump(DUMP_ALL);
364 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
365 smp_mb__before_atomic_inc(); /* See above. */
366 atomic_inc(&rdtp->dynticks);
367 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
368 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
372 * rcu_idle_enter - inform RCU that current CPU is entering idle
374 * Enter idle mode, in other words, -leave- the mode in which RCU
375 * read-side critical sections can occur. (Though RCU read-side
376 * critical sections can occur in irq handlers in idle, a possibility
377 * handled by irq_enter() and irq_exit().)
379 * We crowbar the ->dynticks_nesting field to zero to allow for
380 * the possibility of usermode upcalls having messed up our count
381 * of interrupt nesting level during the prior busy period.
383 void rcu_idle_enter(void)
386 struct rcu_dynticks *rdtp;
388 local_irq_save(flags);
389 rdtp = &__get_cpu_var(rcu_dynticks);
390 rdtp->dynticks_nesting = 0;
391 rcu_idle_enter_common(rdtp);
392 local_irq_restore(flags);
396 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
398 * Exit from an interrupt handler, which might possibly result in entering
399 * idle mode, in other words, leaving the mode in which read-side critical
400 * sections can occur.
402 * This code assumes that the idle loop never does anything that might
403 * result in unbalanced calls to irq_enter() and irq_exit(). If your
404 * architecture violates this assumption, RCU will give you what you
405 * deserve, good and hard. But very infrequently and irreproducibly.
407 * Use things like work queues to work around this limitation.
409 * You have been warned.
411 void rcu_irq_exit(void)
414 struct rcu_dynticks *rdtp;
416 local_irq_save(flags);
417 rdtp = &__get_cpu_var(rcu_dynticks);
418 rdtp->dynticks_nesting--;
419 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
420 rcu_idle_enter_common(rdtp);
421 local_irq_restore(flags);
425 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
427 * If the new value of the ->dynticks_nesting counter was previously zero,
428 * we really have exited idle, and must do the appropriate accounting.
429 * The caller must have disabled interrupts.
431 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
434 trace_rcu_dyntick("++=", rdtp->dynticks_nesting);
437 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
438 atomic_inc(&rdtp->dynticks);
439 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
440 smp_mb__after_atomic_inc(); /* See above. */
441 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
442 trace_rcu_dyntick("End", oldval);
443 if (!idle_cpu(smp_processor_id())) {
444 WARN_ON_ONCE(1); /* must be idle task! */
445 trace_rcu_dyntick("Error on exit: not idle task", oldval);
446 ftrace_dump(DUMP_ALL);
451 * rcu_idle_exit - inform RCU that current CPU is leaving idle
453 * Exit idle mode, in other words, -enter- the mode in which RCU
454 * read-side critical sections can occur.
456 * We crowbar the ->dynticks_nesting field to LLONG_MAX/2 to allow for
457 * the possibility of usermode upcalls messing up our count
458 * of interrupt nesting level during the busy period that is just
461 void rcu_idle_exit(void)
464 struct rcu_dynticks *rdtp;
467 local_irq_save(flags);
468 rdtp = &__get_cpu_var(rcu_dynticks);
469 oldval = rdtp->dynticks_nesting;
470 WARN_ON_ONCE(oldval != 0);
471 rdtp->dynticks_nesting = LLONG_MAX / 2;
472 rcu_idle_exit_common(rdtp, oldval);
473 local_irq_restore(flags);
477 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
479 * Enter an interrupt handler, which might possibly result in exiting
480 * idle mode, in other words, entering the mode in which read-side critical
481 * sections can occur.
483 * Note that the Linux kernel is fully capable of entering an interrupt
484 * handler that it never exits, for example when doing upcalls to
485 * user mode! This code assumes that the idle loop never does upcalls to
486 * user mode. If your architecture does do upcalls from the idle loop (or
487 * does anything else that results in unbalanced calls to the irq_enter()
488 * and irq_exit() functions), RCU will give you what you deserve, good
489 * and hard. But very infrequently and irreproducibly.
491 * Use things like work queues to work around this limitation.
493 * You have been warned.
495 void rcu_irq_enter(void)
498 struct rcu_dynticks *rdtp;
501 local_irq_save(flags);
502 rdtp = &__get_cpu_var(rcu_dynticks);
503 oldval = rdtp->dynticks_nesting;
504 rdtp->dynticks_nesting++;
505 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
506 rcu_idle_exit_common(rdtp, oldval);
507 local_irq_restore(flags);
511 * rcu_nmi_enter - inform RCU of entry to NMI context
513 * If the CPU was idle with dynamic ticks active, and there is no
514 * irq handler running, this updates rdtp->dynticks_nmi to let the
515 * RCU grace-period handling know that the CPU is active.
517 void rcu_nmi_enter(void)
519 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
521 if (rdtp->dynticks_nmi_nesting == 0 &&
522 (atomic_read(&rdtp->dynticks) & 0x1))
524 rdtp->dynticks_nmi_nesting++;
525 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
526 atomic_inc(&rdtp->dynticks);
527 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
528 smp_mb__after_atomic_inc(); /* See above. */
529 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
533 * rcu_nmi_exit - inform RCU of exit from NMI context
535 * If the CPU was idle with dynamic ticks active, and there is no
536 * irq handler running, this updates rdtp->dynticks_nmi to let the
537 * RCU grace-period handling know that the CPU is no longer active.
539 void rcu_nmi_exit(void)
541 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
543 if (rdtp->dynticks_nmi_nesting == 0 ||
544 --rdtp->dynticks_nmi_nesting != 0)
546 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
547 smp_mb__before_atomic_inc(); /* See above. */
548 atomic_inc(&rdtp->dynticks);
549 smp_mb__after_atomic_inc(); /* Force delay to next write. */
550 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
553 #ifdef CONFIG_PROVE_RCU
556 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
558 * If the current CPU is in its idle loop and is neither in an interrupt
559 * or NMI handler, return true. The caller must have at least disabled
562 int rcu_is_cpu_idle(void)
564 return (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
567 #endif /* #ifdef CONFIG_PROVE_RCU */
570 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
572 * If the current CPU is idle or running at a first-level (not nested)
573 * interrupt from idle, return true. The caller must have at least
574 * disabled preemption.
576 int rcu_is_cpu_rrupt_from_idle(void)
578 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
584 * Snapshot the specified CPU's dynticks counter so that we can later
585 * credit them with an implicit quiescent state. Return 1 if this CPU
586 * is in dynticks idle mode, which is an extended quiescent state.
588 static int dyntick_save_progress_counter(struct rcu_data *rdp)
590 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
595 * Return true if the specified CPU has passed through a quiescent
596 * state by virtue of being in or having passed through an dynticks
597 * idle state since the last call to dyntick_save_progress_counter()
600 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
605 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
606 snap = (unsigned int)rdp->dynticks_snap;
609 * If the CPU passed through or entered a dynticks idle phase with
610 * no active irq/NMI handlers, then we can safely pretend that the CPU
611 * already acknowledged the request to pass through a quiescent
612 * state. Either way, that CPU cannot possibly be in an RCU
613 * read-side critical section that started before the beginning
614 * of the current RCU grace period.
616 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
617 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
622 /* Go check for the CPU being offline. */
623 return rcu_implicit_offline_qs(rdp);
626 #endif /* #ifdef CONFIG_SMP */
628 int rcu_cpu_stall_suppress __read_mostly;
630 static void record_gp_stall_check_time(struct rcu_state *rsp)
632 rsp->gp_start = jiffies;
633 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
636 static void print_other_cpu_stall(struct rcu_state *rsp)
642 struct rcu_node *rnp = rcu_get_root(rsp);
644 /* Only let one CPU complain about others per time interval. */
646 raw_spin_lock_irqsave(&rnp->lock, flags);
647 delta = jiffies - rsp->jiffies_stall;
648 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
649 raw_spin_unlock_irqrestore(&rnp->lock, flags);
652 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
655 * Now rat on any tasks that got kicked up to the root rcu_node
656 * due to CPU offlining.
658 ndetected = rcu_print_task_stall(rnp);
659 raw_spin_unlock_irqrestore(&rnp->lock, flags);
662 * OK, time to rat on our buddy...
663 * See Documentation/RCU/stallwarn.txt for info on how to debug
664 * RCU CPU stall warnings.
666 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
668 rcu_for_each_leaf_node(rsp, rnp) {
669 raw_spin_lock_irqsave(&rnp->lock, flags);
670 ndetected += rcu_print_task_stall(rnp);
671 raw_spin_unlock_irqrestore(&rnp->lock, flags);
672 if (rnp->qsmask == 0)
674 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
675 if (rnp->qsmask & (1UL << cpu)) {
676 printk(" %d", rnp->grplo + cpu);
680 printk("} (detected by %d, t=%ld jiffies)\n",
681 smp_processor_id(), (long)(jiffies - rsp->gp_start));
683 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
684 else if (!trigger_all_cpu_backtrace())
687 /* If so configured, complain about tasks blocking the grace period. */
689 rcu_print_detail_task_stall(rsp);
691 force_quiescent_state(rsp, 0); /* Kick them all. */
694 static void print_cpu_stall(struct rcu_state *rsp)
697 struct rcu_node *rnp = rcu_get_root(rsp);
700 * OK, time to rat on ourselves...
701 * See Documentation/RCU/stallwarn.txt for info on how to debug
702 * RCU CPU stall warnings.
704 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
705 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
706 if (!trigger_all_cpu_backtrace())
709 raw_spin_lock_irqsave(&rnp->lock, flags);
710 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
712 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
713 raw_spin_unlock_irqrestore(&rnp->lock, flags);
715 set_need_resched(); /* kick ourselves to get things going. */
718 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
722 struct rcu_node *rnp;
724 if (rcu_cpu_stall_suppress)
726 j = ACCESS_ONCE(jiffies);
727 js = ACCESS_ONCE(rsp->jiffies_stall);
729 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
731 /* We haven't checked in, so go dump stack. */
732 print_cpu_stall(rsp);
734 } else if (rcu_gp_in_progress(rsp) &&
735 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
737 /* They had a few time units to dump stack, so complain. */
738 print_other_cpu_stall(rsp);
742 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
744 rcu_cpu_stall_suppress = 1;
749 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
751 * Set the stall-warning timeout way off into the future, thus preventing
752 * any RCU CPU stall-warning messages from appearing in the current set of
755 * The caller must disable hard irqs.
757 void rcu_cpu_stall_reset(void)
759 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
760 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
761 rcu_preempt_stall_reset();
764 static struct notifier_block rcu_panic_block = {
765 .notifier_call = rcu_panic,
768 static void __init check_cpu_stall_init(void)
770 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
774 * Update CPU-local rcu_data state to record the newly noticed grace period.
775 * This is used both when we started the grace period and when we notice
776 * that someone else started the grace period. The caller must hold the
777 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
778 * and must have irqs disabled.
780 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
782 if (rdp->gpnum != rnp->gpnum) {
784 * If the current grace period is waiting for this CPU,
785 * set up to detect a quiescent state, otherwise don't
786 * go looking for one.
788 rdp->gpnum = rnp->gpnum;
789 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
790 if (rnp->qsmask & rdp->grpmask) {
792 rdp->passed_quiesce = 0;
798 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
801 struct rcu_node *rnp;
803 local_irq_save(flags);
805 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
806 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
807 local_irq_restore(flags);
810 __note_new_gpnum(rsp, rnp, rdp);
811 raw_spin_unlock_irqrestore(&rnp->lock, flags);
815 * Did someone else start a new RCU grace period start since we last
816 * checked? Update local state appropriately if so. Must be called
817 * on the CPU corresponding to rdp.
820 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
825 local_irq_save(flags);
826 if (rdp->gpnum != rsp->gpnum) {
827 note_new_gpnum(rsp, rdp);
830 local_irq_restore(flags);
835 * Advance this CPU's callbacks, but only if the current grace period
836 * has ended. This may be called only from the CPU to whom the rdp
837 * belongs. In addition, the corresponding leaf rcu_node structure's
838 * ->lock must be held by the caller, with irqs disabled.
841 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
843 /* Did another grace period end? */
844 if (rdp->completed != rnp->completed) {
846 /* Advance callbacks. No harm if list empty. */
847 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
848 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
849 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
851 /* Remember that we saw this grace-period completion. */
852 rdp->completed = rnp->completed;
853 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
856 * If we were in an extended quiescent state, we may have
857 * missed some grace periods that others CPUs handled on
858 * our behalf. Catch up with this state to avoid noting
859 * spurious new grace periods. If another grace period
860 * has started, then rnp->gpnum will have advanced, so
861 * we will detect this later on.
863 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
864 rdp->gpnum = rdp->completed;
867 * If RCU does not need a quiescent state from this CPU,
868 * then make sure that this CPU doesn't go looking for one.
870 if ((rnp->qsmask & rdp->grpmask) == 0)
876 * Advance this CPU's callbacks, but only if the current grace period
877 * has ended. This may be called only from the CPU to whom the rdp
881 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
884 struct rcu_node *rnp;
886 local_irq_save(flags);
888 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
889 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
890 local_irq_restore(flags);
893 __rcu_process_gp_end(rsp, rnp, rdp);
894 raw_spin_unlock_irqrestore(&rnp->lock, flags);
898 * Do per-CPU grace-period initialization for running CPU. The caller
899 * must hold the lock of the leaf rcu_node structure corresponding to
903 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
905 /* Prior grace period ended, so advance callbacks for current CPU. */
906 __rcu_process_gp_end(rsp, rnp, rdp);
909 * Because this CPU just now started the new grace period, we know
910 * that all of its callbacks will be covered by this upcoming grace
911 * period, even the ones that were registered arbitrarily recently.
912 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
914 * Other CPUs cannot be sure exactly when the grace period started.
915 * Therefore, their recently registered callbacks must pass through
916 * an additional RCU_NEXT_READY stage, so that they will be handled
917 * by the next RCU grace period.
919 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
920 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
922 /* Set state so that this CPU will detect the next quiescent state. */
923 __note_new_gpnum(rsp, rnp, rdp);
927 * Start a new RCU grace period if warranted, re-initializing the hierarchy
928 * in preparation for detecting the next grace period. The caller must hold
929 * the root node's ->lock, which is released before return. Hard irqs must
933 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
934 __releases(rcu_get_root(rsp)->lock)
936 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
937 struct rcu_node *rnp = rcu_get_root(rsp);
939 if (!rcu_scheduler_fully_active ||
940 !cpu_needs_another_gp(rsp, rdp)) {
942 * Either the scheduler hasn't yet spawned the first
943 * non-idle task or this CPU does not need another
944 * grace period. Either way, don't start a new grace
947 raw_spin_unlock_irqrestore(&rnp->lock, flags);
951 if (rsp->fqs_active) {
953 * This CPU needs a grace period, but force_quiescent_state()
954 * is running. Tell it to start one on this CPU's behalf.
956 rsp->fqs_need_gp = 1;
957 raw_spin_unlock_irqrestore(&rnp->lock, flags);
961 /* Advance to a new grace period and initialize state. */
963 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
964 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
965 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
966 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
967 record_gp_stall_check_time(rsp);
969 /* Special-case the common single-level case. */
970 if (NUM_RCU_NODES == 1) {
971 rcu_preempt_check_blocked_tasks(rnp);
972 rnp->qsmask = rnp->qsmaskinit;
973 rnp->gpnum = rsp->gpnum;
974 rnp->completed = rsp->completed;
975 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state OK */
976 rcu_start_gp_per_cpu(rsp, rnp, rdp);
977 rcu_preempt_boost_start_gp(rnp);
978 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
979 rnp->level, rnp->grplo,
980 rnp->grphi, rnp->qsmask);
981 raw_spin_unlock_irqrestore(&rnp->lock, flags);
985 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
988 /* Exclude any concurrent CPU-hotplug operations. */
989 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
992 * Set the quiescent-state-needed bits in all the rcu_node
993 * structures for all currently online CPUs in breadth-first
994 * order, starting from the root rcu_node structure. This
995 * operation relies on the layout of the hierarchy within the
996 * rsp->node[] array. Note that other CPUs will access only
997 * the leaves of the hierarchy, which still indicate that no
998 * grace period is in progress, at least until the corresponding
999 * leaf node has been initialized. In addition, we have excluded
1000 * CPU-hotplug operations.
1002 * Note that the grace period cannot complete until we finish
1003 * the initialization process, as there will be at least one
1004 * qsmask bit set in the root node until that time, namely the
1005 * one corresponding to this CPU, due to the fact that we have
1008 rcu_for_each_node_breadth_first(rsp, rnp) {
1009 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1010 rcu_preempt_check_blocked_tasks(rnp);
1011 rnp->qsmask = rnp->qsmaskinit;
1012 rnp->gpnum = rsp->gpnum;
1013 rnp->completed = rsp->completed;
1014 if (rnp == rdp->mynode)
1015 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1016 rcu_preempt_boost_start_gp(rnp);
1017 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1018 rnp->level, rnp->grplo,
1019 rnp->grphi, rnp->qsmask);
1020 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1023 rnp = rcu_get_root(rsp);
1024 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1025 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1026 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1027 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1031 * Report a full set of quiescent states to the specified rcu_state
1032 * data structure. This involves cleaning up after the prior grace
1033 * period and letting rcu_start_gp() start up the next grace period
1034 * if one is needed. Note that the caller must hold rnp->lock, as
1035 * required by rcu_start_gp(), which will release it.
1037 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1038 __releases(rcu_get_root(rsp)->lock)
1040 unsigned long gp_duration;
1041 struct rcu_node *rnp = rcu_get_root(rsp);
1042 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1044 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1047 * Ensure that all grace-period and pre-grace-period activity
1048 * is seen before the assignment to rsp->completed.
1050 smp_mb(); /* See above block comment. */
1051 gp_duration = jiffies - rsp->gp_start;
1052 if (gp_duration > rsp->gp_max)
1053 rsp->gp_max = gp_duration;
1056 * We know the grace period is complete, but to everyone else
1057 * it appears to still be ongoing. But it is also the case
1058 * that to everyone else it looks like there is nothing that
1059 * they can do to advance the grace period. It is therefore
1060 * safe for us to drop the lock in order to mark the grace
1061 * period as completed in all of the rcu_node structures.
1063 * But if this CPU needs another grace period, it will take
1064 * care of this while initializing the next grace period.
1065 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1066 * because the callbacks have not yet been advanced: Those
1067 * callbacks are waiting on the grace period that just now
1070 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1071 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1074 * Propagate new ->completed value to rcu_node structures
1075 * so that other CPUs don't have to wait until the start
1076 * of the next grace period to process their callbacks.
1078 rcu_for_each_node_breadth_first(rsp, rnp) {
1079 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1080 rnp->completed = rsp->gpnum;
1081 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1083 rnp = rcu_get_root(rsp);
1084 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1087 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1088 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1089 rsp->fqs_state = RCU_GP_IDLE;
1090 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1094 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1095 * Allows quiescent states for a group of CPUs to be reported at one go
1096 * to the specified rcu_node structure, though all the CPUs in the group
1097 * must be represented by the same rcu_node structure (which need not be
1098 * a leaf rcu_node structure, though it often will be). That structure's
1099 * lock must be held upon entry, and it is released before return.
1102 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1103 struct rcu_node *rnp, unsigned long flags)
1104 __releases(rnp->lock)
1106 struct rcu_node *rnp_c;
1108 /* Walk up the rcu_node hierarchy. */
1110 if (!(rnp->qsmask & mask)) {
1112 /* Our bit has already been cleared, so done. */
1113 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1116 rnp->qsmask &= ~mask;
1117 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1118 mask, rnp->qsmask, rnp->level,
1119 rnp->grplo, rnp->grphi,
1121 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1123 /* Other bits still set at this level, so done. */
1124 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1127 mask = rnp->grpmask;
1128 if (rnp->parent == NULL) {
1130 /* No more levels. Exit loop holding root lock. */
1134 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1137 raw_spin_lock_irqsave(&rnp->lock, flags);
1138 WARN_ON_ONCE(rnp_c->qsmask);
1142 * Get here if we are the last CPU to pass through a quiescent
1143 * state for this grace period. Invoke rcu_report_qs_rsp()
1144 * to clean up and start the next grace period if one is needed.
1146 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1150 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1151 * structure. This must be either called from the specified CPU, or
1152 * called when the specified CPU is known to be offline (and when it is
1153 * also known that no other CPU is concurrently trying to help the offline
1154 * CPU). The lastcomp argument is used to make sure we are still in the
1155 * grace period of interest. We don't want to end the current grace period
1156 * based on quiescent states detected in an earlier grace period!
1159 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1161 unsigned long flags;
1163 struct rcu_node *rnp;
1166 raw_spin_lock_irqsave(&rnp->lock, flags);
1167 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1170 * The grace period in which this quiescent state was
1171 * recorded has ended, so don't report it upwards.
1172 * We will instead need a new quiescent state that lies
1173 * within the current grace period.
1175 rdp->passed_quiesce = 0; /* need qs for new gp. */
1176 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1179 mask = rdp->grpmask;
1180 if ((rnp->qsmask & mask) == 0) {
1181 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1183 rdp->qs_pending = 0;
1186 * This GP can't end until cpu checks in, so all of our
1187 * callbacks can be processed during the next GP.
1189 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1191 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1196 * Check to see if there is a new grace period of which this CPU
1197 * is not yet aware, and if so, set up local rcu_data state for it.
1198 * Otherwise, see if this CPU has just passed through its first
1199 * quiescent state for this grace period, and record that fact if so.
1202 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1204 /* If there is now a new grace period, record and return. */
1205 if (check_for_new_grace_period(rsp, rdp))
1209 * Does this CPU still need to do its part for current grace period?
1210 * If no, return and let the other CPUs do their part as well.
1212 if (!rdp->qs_pending)
1216 * Was there a quiescent state since the beginning of the grace
1217 * period? If no, then exit and wait for the next call.
1219 if (!rdp->passed_quiesce)
1223 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1226 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1229 #ifdef CONFIG_HOTPLUG_CPU
1232 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1233 * Synchronization is not required because this function executes
1234 * in stop_machine() context.
1236 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1239 /* current DYING CPU is cleared in the cpu_online_mask */
1240 int receive_cpu = cpumask_any(cpu_online_mask);
1241 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1242 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1244 if (rdp->nxtlist == NULL)
1245 return; /* irqs disabled, so comparison is stable. */
1247 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1248 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1249 receive_rdp->qlen += rdp->qlen;
1250 receive_rdp->n_cbs_adopted += rdp->qlen;
1251 rdp->n_cbs_orphaned += rdp->qlen;
1253 rdp->nxtlist = NULL;
1254 for (i = 0; i < RCU_NEXT_SIZE; i++)
1255 rdp->nxttail[i] = &rdp->nxtlist;
1260 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1261 * and move all callbacks from the outgoing CPU to the current one.
1262 * There can only be one CPU hotplug operation at a time, so no other
1263 * CPU can be attempting to update rcu_cpu_kthread_task.
1265 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1267 unsigned long flags;
1269 int need_report = 0;
1270 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1271 struct rcu_node *rnp;
1273 rcu_stop_cpu_kthread(cpu);
1275 /* Exclude any attempts to start a new grace period. */
1276 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1278 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1279 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1280 mask = rdp->grpmask; /* rnp->grplo is constant. */
1282 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1283 rnp->qsmaskinit &= ~mask;
1284 if (rnp->qsmaskinit != 0) {
1285 if (rnp != rdp->mynode)
1286 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1288 trace_rcu_grace_period(rsp->name,
1290 !!(rnp->qsmask & mask),
1294 if (rnp == rdp->mynode) {
1295 trace_rcu_grace_period(rsp->name,
1297 !!(rnp->qsmask & mask),
1299 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1301 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1302 mask = rnp->grpmask;
1304 } while (rnp != NULL);
1307 * We still hold the leaf rcu_node structure lock here, and
1308 * irqs are still disabled. The reason for this subterfuge is
1309 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1310 * held leads to deadlock.
1312 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1314 if (need_report & RCU_OFL_TASKS_NORM_GP)
1315 rcu_report_unblock_qs_rnp(rnp, flags);
1317 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1318 if (need_report & RCU_OFL_TASKS_EXP_GP)
1319 rcu_report_exp_rnp(rsp, rnp);
1320 rcu_node_kthread_setaffinity(rnp, -1);
1324 * Remove the specified CPU from the RCU hierarchy and move any pending
1325 * callbacks that it might have to the current CPU. This code assumes
1326 * that at least one CPU in the system will remain running at all times.
1327 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1329 static void rcu_offline_cpu(int cpu)
1331 __rcu_offline_cpu(cpu, &rcu_sched_state);
1332 __rcu_offline_cpu(cpu, &rcu_bh_state);
1333 rcu_preempt_offline_cpu(cpu);
1336 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1338 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1342 static void rcu_offline_cpu(int cpu)
1346 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1349 * Invoke any RCU callbacks that have made it to the end of their grace
1350 * period. Thottle as specified by rdp->blimit.
1352 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1354 unsigned long flags;
1355 struct rcu_head *next, *list, **tail;
1358 /* If no callbacks are ready, just return.*/
1359 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1360 trace_rcu_batch_start(rsp->name, 0, 0);
1361 trace_rcu_batch_end(rsp->name, 0);
1366 * Extract the list of ready callbacks, disabling to prevent
1367 * races with call_rcu() from interrupt handlers.
1369 local_irq_save(flags);
1371 trace_rcu_batch_start(rsp->name, rdp->qlen, bl);
1372 list = rdp->nxtlist;
1373 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1374 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1375 tail = rdp->nxttail[RCU_DONE_TAIL];
1376 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1377 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1378 rdp->nxttail[count] = &rdp->nxtlist;
1379 local_irq_restore(flags);
1381 /* Invoke callbacks. */
1386 debug_rcu_head_unqueue(list);
1387 __rcu_reclaim(rsp->name, list);
1393 local_irq_save(flags);
1394 trace_rcu_batch_end(rsp->name, count);
1396 /* Update count, and requeue any remaining callbacks. */
1398 rdp->n_cbs_invoked += count;
1400 *tail = rdp->nxtlist;
1401 rdp->nxtlist = list;
1402 for (count = 0; count < RCU_NEXT_SIZE; count++)
1403 if (&rdp->nxtlist == rdp->nxttail[count])
1404 rdp->nxttail[count] = tail;
1409 /* Reinstate batch limit if we have worked down the excess. */
1410 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1411 rdp->blimit = blimit;
1413 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1414 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1415 rdp->qlen_last_fqs_check = 0;
1416 rdp->n_force_qs_snap = rsp->n_force_qs;
1417 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1418 rdp->qlen_last_fqs_check = rdp->qlen;
1420 local_irq_restore(flags);
1422 /* Re-invoke RCU core processing if there are callbacks remaining. */
1423 if (cpu_has_callbacks_ready_to_invoke(rdp))
1428 * Check to see if this CPU is in a non-context-switch quiescent state
1429 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1430 * Also schedule RCU core processing.
1432 * This function must be called from hardirq context. It is normally
1433 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1434 * false, there is no point in invoking rcu_check_callbacks().
1436 void rcu_check_callbacks(int cpu, int user)
1438 trace_rcu_utilization("Start scheduler-tick");
1439 if (user || rcu_is_cpu_rrupt_from_idle()) {
1442 * Get here if this CPU took its interrupt from user
1443 * mode or from the idle loop, and if this is not a
1444 * nested interrupt. In this case, the CPU is in
1445 * a quiescent state, so note it.
1447 * No memory barrier is required here because both
1448 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1449 * variables that other CPUs neither access nor modify,
1450 * at least not while the corresponding CPU is online.
1456 } else if (!in_softirq()) {
1459 * Get here if this CPU did not take its interrupt from
1460 * softirq, in other words, if it is not interrupting
1461 * a rcu_bh read-side critical section. This is an _bh
1462 * critical section, so note it.
1467 rcu_preempt_check_callbacks(cpu);
1468 if (rcu_pending(cpu))
1470 trace_rcu_utilization("End scheduler-tick");
1476 * Scan the leaf rcu_node structures, processing dyntick state for any that
1477 * have not yet encountered a quiescent state, using the function specified.
1478 * Also initiate boosting for any threads blocked on the root rcu_node.
1480 * The caller must have suppressed start of new grace periods.
1482 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1486 unsigned long flags;
1488 struct rcu_node *rnp;
1490 rcu_for_each_leaf_node(rsp, rnp) {
1492 raw_spin_lock_irqsave(&rnp->lock, flags);
1493 if (!rcu_gp_in_progress(rsp)) {
1494 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1497 if (rnp->qsmask == 0) {
1498 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1503 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1504 if ((rnp->qsmask & bit) != 0 &&
1505 f(per_cpu_ptr(rsp->rda, cpu)))
1510 /* rcu_report_qs_rnp() releases rnp->lock. */
1511 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1514 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1516 rnp = rcu_get_root(rsp);
1517 if (rnp->qsmask == 0) {
1518 raw_spin_lock_irqsave(&rnp->lock, flags);
1519 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1524 * Force quiescent states on reluctant CPUs, and also detect which
1525 * CPUs are in dyntick-idle mode.
1527 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1529 unsigned long flags;
1530 struct rcu_node *rnp = rcu_get_root(rsp);
1532 trace_rcu_utilization("Start fqs");
1533 if (!rcu_gp_in_progress(rsp)) {
1534 trace_rcu_utilization("End fqs");
1535 return; /* No grace period in progress, nothing to force. */
1537 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1538 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1539 trace_rcu_utilization("End fqs");
1540 return; /* Someone else is already on the job. */
1542 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1543 goto unlock_fqs_ret; /* no emergency and done recently. */
1545 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1546 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1547 if(!rcu_gp_in_progress(rsp)) {
1548 rsp->n_force_qs_ngp++;
1549 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1550 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1552 rsp->fqs_active = 1;
1553 switch (rsp->fqs_state) {
1557 break; /* grace period idle or initializing, ignore. */
1559 case RCU_SAVE_DYNTICK:
1560 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1561 break; /* So gcc recognizes the dead code. */
1563 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1565 /* Record dyntick-idle state. */
1566 force_qs_rnp(rsp, dyntick_save_progress_counter);
1567 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1568 if (rcu_gp_in_progress(rsp))
1569 rsp->fqs_state = RCU_FORCE_QS;
1574 /* Check dyntick-idle state, send IPI to laggarts. */
1575 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1576 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1578 /* Leave state in case more forcing is required. */
1580 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1583 rsp->fqs_active = 0;
1584 if (rsp->fqs_need_gp) {
1585 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1586 rsp->fqs_need_gp = 0;
1587 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1588 trace_rcu_utilization("End fqs");
1591 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1593 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1594 trace_rcu_utilization("End fqs");
1597 #else /* #ifdef CONFIG_SMP */
1599 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1604 #endif /* #else #ifdef CONFIG_SMP */
1607 * This does the RCU core processing work for the specified rcu_state
1608 * and rcu_data structures. This may be called only from the CPU to
1609 * whom the rdp belongs.
1612 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1614 unsigned long flags;
1616 WARN_ON_ONCE(rdp->beenonline == 0);
1619 * If an RCU GP has gone long enough, go check for dyntick
1620 * idle CPUs and, if needed, send resched IPIs.
1622 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1623 force_quiescent_state(rsp, 1);
1626 * Advance callbacks in response to end of earlier grace
1627 * period that some other CPU ended.
1629 rcu_process_gp_end(rsp, rdp);
1631 /* Update RCU state based on any recent quiescent states. */
1632 rcu_check_quiescent_state(rsp, rdp);
1634 /* Does this CPU require a not-yet-started grace period? */
1635 if (cpu_needs_another_gp(rsp, rdp)) {
1636 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1637 rcu_start_gp(rsp, flags); /* releases above lock */
1640 /* If there are callbacks ready, invoke them. */
1641 if (cpu_has_callbacks_ready_to_invoke(rdp))
1642 invoke_rcu_callbacks(rsp, rdp);
1646 * Do RCU core processing for the current CPU.
1648 static void rcu_process_callbacks(struct softirq_action *unused)
1650 trace_rcu_utilization("Start RCU core");
1651 __rcu_process_callbacks(&rcu_sched_state,
1652 &__get_cpu_var(rcu_sched_data));
1653 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1654 rcu_preempt_process_callbacks();
1655 trace_rcu_utilization("End RCU core");
1659 * Schedule RCU callback invocation. If the specified type of RCU
1660 * does not support RCU priority boosting, just do a direct call,
1661 * otherwise wake up the per-CPU kernel kthread. Note that because we
1662 * are running on the current CPU with interrupts disabled, the
1663 * rcu_cpu_kthread_task cannot disappear out from under us.
1665 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1667 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1669 if (likely(!rsp->boost)) {
1670 rcu_do_batch(rsp, rdp);
1673 invoke_rcu_callbacks_kthread();
1676 static void invoke_rcu_core(void)
1678 raise_softirq(RCU_SOFTIRQ);
1682 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1683 struct rcu_state *rsp)
1685 unsigned long flags;
1686 struct rcu_data *rdp;
1688 debug_rcu_head_queue(head);
1692 smp_mb(); /* Ensure RCU update seen before callback registry. */
1695 * Opportunistically note grace-period endings and beginnings.
1696 * Note that we might see a beginning right after we see an
1697 * end, but never vice versa, since this CPU has to pass through
1698 * a quiescent state betweentimes.
1700 local_irq_save(flags);
1701 rdp = this_cpu_ptr(rsp->rda);
1703 /* Add the callback to our list. */
1704 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1705 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1708 if (__is_kfree_rcu_offset((unsigned long)func))
1709 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1712 trace_rcu_callback(rsp->name, head, rdp->qlen);
1714 /* If interrupts were disabled, don't dive into RCU core. */
1715 if (irqs_disabled_flags(flags)) {
1716 local_irq_restore(flags);
1721 * Force the grace period if too many callbacks or too long waiting.
1722 * Enforce hysteresis, and don't invoke force_quiescent_state()
1723 * if some other CPU has recently done so. Also, don't bother
1724 * invoking force_quiescent_state() if the newly enqueued callback
1725 * is the only one waiting for a grace period to complete.
1727 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1729 /* Are we ignoring a completed grace period? */
1730 rcu_process_gp_end(rsp, rdp);
1731 check_for_new_grace_period(rsp, rdp);
1733 /* Start a new grace period if one not already started. */
1734 if (!rcu_gp_in_progress(rsp)) {
1735 unsigned long nestflag;
1736 struct rcu_node *rnp_root = rcu_get_root(rsp);
1738 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1739 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1741 /* Give the grace period a kick. */
1742 rdp->blimit = LONG_MAX;
1743 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1744 *rdp->nxttail[RCU_DONE_TAIL] != head)
1745 force_quiescent_state(rsp, 0);
1746 rdp->n_force_qs_snap = rsp->n_force_qs;
1747 rdp->qlen_last_fqs_check = rdp->qlen;
1749 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1750 force_quiescent_state(rsp, 1);
1751 local_irq_restore(flags);
1755 * Queue an RCU-sched callback for invocation after a grace period.
1757 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1759 __call_rcu(head, func, &rcu_sched_state);
1761 EXPORT_SYMBOL_GPL(call_rcu_sched);
1764 * Queue an RCU for invocation after a quicker grace period.
1766 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1768 __call_rcu(head, func, &rcu_bh_state);
1770 EXPORT_SYMBOL_GPL(call_rcu_bh);
1773 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1775 * Control will return to the caller some time after a full rcu-sched
1776 * grace period has elapsed, in other words after all currently executing
1777 * rcu-sched read-side critical sections have completed. These read-side
1778 * critical sections are delimited by rcu_read_lock_sched() and
1779 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1780 * local_irq_disable(), and so on may be used in place of
1781 * rcu_read_lock_sched().
1783 * This means that all preempt_disable code sequences, including NMI and
1784 * hardware-interrupt handlers, in progress on entry will have completed
1785 * before this primitive returns. However, this does not guarantee that
1786 * softirq handlers will have completed, since in some kernels, these
1787 * handlers can run in process context, and can block.
1789 * This primitive provides the guarantees made by the (now removed)
1790 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1791 * guarantees that rcu_read_lock() sections will have completed.
1792 * In "classic RCU", these two guarantees happen to be one and
1793 * the same, but can differ in realtime RCU implementations.
1795 void synchronize_sched(void)
1797 if (rcu_blocking_is_gp())
1799 wait_rcu_gp(call_rcu_sched);
1801 EXPORT_SYMBOL_GPL(synchronize_sched);
1804 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1806 * Control will return to the caller some time after a full rcu_bh grace
1807 * period has elapsed, in other words after all currently executing rcu_bh
1808 * read-side critical sections have completed. RCU read-side critical
1809 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1810 * and may be nested.
1812 void synchronize_rcu_bh(void)
1814 if (rcu_blocking_is_gp())
1816 wait_rcu_gp(call_rcu_bh);
1818 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1821 * Check to see if there is any immediate RCU-related work to be done
1822 * by the current CPU, for the specified type of RCU, returning 1 if so.
1823 * The checks are in order of increasing expense: checks that can be
1824 * carried out against CPU-local state are performed first. However,
1825 * we must check for CPU stalls first, else we might not get a chance.
1827 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1829 struct rcu_node *rnp = rdp->mynode;
1831 rdp->n_rcu_pending++;
1833 /* Check for CPU stalls, if enabled. */
1834 check_cpu_stall(rsp, rdp);
1836 /* Is the RCU core waiting for a quiescent state from this CPU? */
1837 if (rcu_scheduler_fully_active &&
1838 rdp->qs_pending && !rdp->passed_quiesce) {
1841 * If force_quiescent_state() coming soon and this CPU
1842 * needs a quiescent state, and this is either RCU-sched
1843 * or RCU-bh, force a local reschedule.
1845 rdp->n_rp_qs_pending++;
1846 if (!rdp->preemptible &&
1847 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1850 } else if (rdp->qs_pending && rdp->passed_quiesce) {
1851 rdp->n_rp_report_qs++;
1855 /* Does this CPU have callbacks ready to invoke? */
1856 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1857 rdp->n_rp_cb_ready++;
1861 /* Has RCU gone idle with this CPU needing another grace period? */
1862 if (cpu_needs_another_gp(rsp, rdp)) {
1863 rdp->n_rp_cpu_needs_gp++;
1867 /* Has another RCU grace period completed? */
1868 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1869 rdp->n_rp_gp_completed++;
1873 /* Has a new RCU grace period started? */
1874 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1875 rdp->n_rp_gp_started++;
1879 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1880 if (rcu_gp_in_progress(rsp) &&
1881 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1882 rdp->n_rp_need_fqs++;
1887 rdp->n_rp_need_nothing++;
1892 * Check to see if there is any immediate RCU-related work to be done
1893 * by the current CPU, returning 1 if so. This function is part of the
1894 * RCU implementation; it is -not- an exported member of the RCU API.
1896 static int rcu_pending(int cpu)
1898 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1899 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1900 rcu_preempt_pending(cpu);
1904 * Check to see if any future RCU-related work will need to be done
1905 * by the current CPU, even if none need be done immediately, returning
1908 static int rcu_needs_cpu_quick_check(int cpu)
1910 /* RCU callbacks either ready or pending? */
1911 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1912 per_cpu(rcu_bh_data, cpu).nxtlist ||
1913 rcu_preempt_needs_cpu(cpu);
1916 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1917 static atomic_t rcu_barrier_cpu_count;
1918 static DEFINE_MUTEX(rcu_barrier_mutex);
1919 static struct completion rcu_barrier_completion;
1921 static void rcu_barrier_callback(struct rcu_head *notused)
1923 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1924 complete(&rcu_barrier_completion);
1928 * Called with preemption disabled, and from cross-cpu IRQ context.
1930 static void rcu_barrier_func(void *type)
1932 int cpu = smp_processor_id();
1933 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1934 void (*call_rcu_func)(struct rcu_head *head,
1935 void (*func)(struct rcu_head *head));
1937 atomic_inc(&rcu_barrier_cpu_count);
1938 call_rcu_func = type;
1939 call_rcu_func(head, rcu_barrier_callback);
1943 * Orchestrate the specified type of RCU barrier, waiting for all
1944 * RCU callbacks of the specified type to complete.
1946 static void _rcu_barrier(struct rcu_state *rsp,
1947 void (*call_rcu_func)(struct rcu_head *head,
1948 void (*func)(struct rcu_head *head)))
1950 BUG_ON(in_interrupt());
1951 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1952 mutex_lock(&rcu_barrier_mutex);
1953 init_completion(&rcu_barrier_completion);
1955 * Initialize rcu_barrier_cpu_count to 1, then invoke
1956 * rcu_barrier_func() on each CPU, so that each CPU also has
1957 * incremented rcu_barrier_cpu_count. Only then is it safe to
1958 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1959 * might complete its grace period before all of the other CPUs
1960 * did their increment, causing this function to return too
1961 * early. Note that on_each_cpu() disables irqs, which prevents
1962 * any CPUs from coming online or going offline until each online
1963 * CPU has queued its RCU-barrier callback.
1965 atomic_set(&rcu_barrier_cpu_count, 1);
1966 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1967 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1968 complete(&rcu_barrier_completion);
1969 wait_for_completion(&rcu_barrier_completion);
1970 mutex_unlock(&rcu_barrier_mutex);
1974 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1976 void rcu_barrier_bh(void)
1978 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1980 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1983 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1985 void rcu_barrier_sched(void)
1987 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1989 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1992 * Do boot-time initialization of a CPU's per-CPU RCU data.
1995 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1997 unsigned long flags;
1999 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2000 struct rcu_node *rnp = rcu_get_root(rsp);
2002 /* Set up local state, ensuring consistent view of global state. */
2003 raw_spin_lock_irqsave(&rnp->lock, flags);
2004 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2005 rdp->nxtlist = NULL;
2006 for (i = 0; i < RCU_NEXT_SIZE; i++)
2007 rdp->nxttail[i] = &rdp->nxtlist;
2009 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2010 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != LLONG_MAX / 2);
2011 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2014 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2018 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2019 * offline event can be happening at a given time. Note also that we
2020 * can accept some slop in the rsp->completed access due to the fact
2021 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2023 static void __cpuinit
2024 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2026 unsigned long flags;
2028 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2029 struct rcu_node *rnp = rcu_get_root(rsp);
2031 /* Set up local state, ensuring consistent view of global state. */
2032 raw_spin_lock_irqsave(&rnp->lock, flags);
2033 rdp->beenonline = 1; /* We have now been online. */
2034 rdp->preemptible = preemptible;
2035 rdp->qlen_last_fqs_check = 0;
2036 rdp->n_force_qs_snap = rsp->n_force_qs;
2037 rdp->blimit = blimit;
2038 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != LLONG_MAX / 2);
2039 WARN_ON_ONCE((atomic_read(&rdp->dynticks->dynticks) & 0x1) != 1);
2040 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2043 * A new grace period might start here. If so, we won't be part
2044 * of it, but that is OK, as we are currently in a quiescent state.
2047 /* Exclude any attempts to start a new GP on large systems. */
2048 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2050 /* Add CPU to rcu_node bitmasks. */
2052 mask = rdp->grpmask;
2054 /* Exclude any attempts to start a new GP on small systems. */
2055 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2056 rnp->qsmaskinit |= mask;
2057 mask = rnp->grpmask;
2058 if (rnp == rdp->mynode) {
2060 * If there is a grace period in progress, we will
2061 * set up to wait for it next time we run the
2064 rdp->gpnum = rnp->completed;
2065 rdp->completed = rnp->completed;
2066 rdp->passed_quiesce = 0;
2067 rdp->qs_pending = 0;
2068 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2069 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2071 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2073 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2075 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2078 static void __cpuinit rcu_prepare_cpu(int cpu)
2080 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2081 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2082 rcu_preempt_init_percpu_data(cpu);
2086 * Handle CPU online/offline notification events.
2088 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2089 unsigned long action, void *hcpu)
2091 long cpu = (long)hcpu;
2092 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2093 struct rcu_node *rnp = rdp->mynode;
2095 trace_rcu_utilization("Start CPU hotplug");
2097 case CPU_UP_PREPARE:
2098 case CPU_UP_PREPARE_FROZEN:
2099 rcu_prepare_cpu(cpu);
2100 rcu_prepare_kthreads(cpu);
2103 case CPU_DOWN_FAILED:
2104 rcu_node_kthread_setaffinity(rnp, -1);
2105 rcu_cpu_kthread_setrt(cpu, 1);
2107 case CPU_DOWN_PREPARE:
2108 rcu_node_kthread_setaffinity(rnp, cpu);
2109 rcu_cpu_kthread_setrt(cpu, 0);
2112 case CPU_DYING_FROZEN:
2114 * The whole machine is "stopped" except this CPU, so we can
2115 * touch any data without introducing corruption. We send the
2116 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2118 rcu_send_cbs_to_online(&rcu_bh_state);
2119 rcu_send_cbs_to_online(&rcu_sched_state);
2120 rcu_preempt_send_cbs_to_online();
2123 case CPU_DEAD_FROZEN:
2124 case CPU_UP_CANCELED:
2125 case CPU_UP_CANCELED_FROZEN:
2126 rcu_offline_cpu(cpu);
2131 trace_rcu_utilization("End CPU hotplug");
2136 * This function is invoked towards the end of the scheduler's initialization
2137 * process. Before this is called, the idle task might contain
2138 * RCU read-side critical sections (during which time, this idle
2139 * task is booting the system). After this function is called, the
2140 * idle tasks are prohibited from containing RCU read-side critical
2141 * sections. This function also enables RCU lockdep checking.
2143 void rcu_scheduler_starting(void)
2145 WARN_ON(num_online_cpus() != 1);
2146 WARN_ON(nr_context_switches() > 0);
2147 rcu_scheduler_active = 1;
2151 * Compute the per-level fanout, either using the exact fanout specified
2152 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2154 #ifdef CONFIG_RCU_FANOUT_EXACT
2155 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2159 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
2160 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2161 rsp->levelspread[0] = RCU_FANOUT_LEAF;
2163 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2164 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2171 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2172 ccur = rsp->levelcnt[i];
2173 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2177 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2180 * Helper function for rcu_init() that initializes one rcu_state structure.
2182 static void __init rcu_init_one(struct rcu_state *rsp,
2183 struct rcu_data __percpu *rda)
2185 static char *buf[] = { "rcu_node_level_0",
2188 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2192 struct rcu_node *rnp;
2194 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2196 /* Initialize the level-tracking arrays. */
2198 for (i = 1; i < NUM_RCU_LVLS; i++)
2199 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2200 rcu_init_levelspread(rsp);
2202 /* Initialize the elements themselves, starting from the leaves. */
2204 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2205 cpustride *= rsp->levelspread[i];
2206 rnp = rsp->level[i];
2207 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2208 raw_spin_lock_init(&rnp->lock);
2209 lockdep_set_class_and_name(&rnp->lock,
2210 &rcu_node_class[i], buf[i]);
2213 rnp->qsmaskinit = 0;
2214 rnp->grplo = j * cpustride;
2215 rnp->grphi = (j + 1) * cpustride - 1;
2216 if (rnp->grphi >= NR_CPUS)
2217 rnp->grphi = NR_CPUS - 1;
2223 rnp->grpnum = j % rsp->levelspread[i - 1];
2224 rnp->grpmask = 1UL << rnp->grpnum;
2225 rnp->parent = rsp->level[i - 1] +
2226 j / rsp->levelspread[i - 1];
2229 INIT_LIST_HEAD(&rnp->blkd_tasks);
2234 rnp = rsp->level[NUM_RCU_LVLS - 1];
2235 for_each_possible_cpu(i) {
2236 while (i > rnp->grphi)
2238 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2239 rcu_boot_init_percpu_data(i, rsp);
2243 void __init rcu_init(void)
2247 rcu_bootup_announce();
2248 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2249 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2250 __rcu_init_preempt();
2251 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2254 * We don't need protection against CPU-hotplug here because
2255 * this is called early in boot, before either interrupts
2256 * or the scheduler are operational.
2258 cpu_notifier(rcu_cpu_notify, 0);
2259 for_each_online_cpu(cpu)
2260 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2261 check_cpu_stall_init();
2264 #include "rcutree_plugin.h"