Root/kernel/rcutree.c

1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
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.
8 *
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.
13 *
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.
17 *
18 * Copyright IBM Corporation, 2008
19 *
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
23 *
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
26 *
27 * For detailed explanation of Read-Copy Update mechanism see -
28 * Documentation/RCU
29 */
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 <asm/atomic.h>
40#include <linux/bitops.h>
41#include <linux/module.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
51#include "rcutree.h"
52
53/* Data structures. */
54
55static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
56
57#define RCU_STATE_INITIALIZER(structname) { \
58    .level = { &structname.node[0] }, \
59    .levelcnt = { \
60        NUM_RCU_LVL_0, /* root of hierarchy. */ \
61        NUM_RCU_LVL_1, \
62        NUM_RCU_LVL_2, \
63        NUM_RCU_LVL_3, \
64        NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
65    }, \
66    .signaled = RCU_GP_IDLE, \
67    .gpnum = -300, \
68    .completed = -300, \
69    .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
70    .orphan_cbs_list = NULL, \
71    .orphan_cbs_tail = &structname.orphan_cbs_list, \
72    .orphan_qlen = 0, \
73    .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
74    .n_force_qs = 0, \
75    .n_force_qs_ngp = 0, \
76    .name = #structname, \
77}
78
79struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
80DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
81
82struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
83DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
84
85int rcu_scheduler_active __read_mostly;
86EXPORT_SYMBOL_GPL(rcu_scheduler_active);
87
88/*
89 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
90 * permit this function to be invoked without holding the root rcu_node
91 * structure's ->lock, but of course results can be subject to change.
92 */
93static int rcu_gp_in_progress(struct rcu_state *rsp)
94{
95    return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
96}
97
98/*
99 * Note a quiescent state. Because we do not need to know
100 * how many quiescent states passed, just if there was at least
101 * one since the start of the grace period, this just sets a flag.
102 */
103void rcu_sched_qs(int cpu)
104{
105    struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
106
107    rdp->passed_quiesc_completed = rdp->gpnum - 1;
108    barrier();
109    rdp->passed_quiesc = 1;
110}
111
112void rcu_bh_qs(int cpu)
113{
114    struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
115
116    rdp->passed_quiesc_completed = rdp->gpnum - 1;
117    barrier();
118    rdp->passed_quiesc = 1;
119}
120
121/*
122 * Note a context switch. This is a quiescent state for RCU-sched,
123 * and requires special handling for preemptible RCU.
124 */
125void rcu_note_context_switch(int cpu)
126{
127    rcu_sched_qs(cpu);
128    rcu_preempt_note_context_switch(cpu);
129}
130
131#ifdef CONFIG_NO_HZ
132DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
133    .dynticks_nesting = 1,
134    .dynticks = 1,
135};
136#endif /* #ifdef CONFIG_NO_HZ */
137
138static int blimit = 10; /* Maximum callbacks per softirq. */
139static int qhimark = 10000; /* If this many pending, ignore blimit. */
140static int qlowmark = 100; /* Once only this many pending, use blimit. */
141
142module_param(blimit, int, 0);
143module_param(qhimark, int, 0);
144module_param(qlowmark, int, 0);
145
146static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
147static int rcu_pending(int cpu);
148
149/*
150 * Return the number of RCU-sched batches processed thus far for debug & stats.
151 */
152long rcu_batches_completed_sched(void)
153{
154    return rcu_sched_state.completed;
155}
156EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
157
158/*
159 * Return the number of RCU BH batches processed thus far for debug & stats.
160 */
161long rcu_batches_completed_bh(void)
162{
163    return rcu_bh_state.completed;
164}
165EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
166
167/*
168 * Force a quiescent state for RCU BH.
169 */
170void rcu_bh_force_quiescent_state(void)
171{
172    force_quiescent_state(&rcu_bh_state, 0);
173}
174EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
175
176/*
177 * Force a quiescent state for RCU-sched.
178 */
179void rcu_sched_force_quiescent_state(void)
180{
181    force_quiescent_state(&rcu_sched_state, 0);
182}
183EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
184
185/*
186 * Does the CPU have callbacks ready to be invoked?
187 */
188static int
189cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
190{
191    return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
192}
193
194/*
195 * Does the current CPU require a yet-as-unscheduled grace period?
196 */
197static int
198cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
199{
200    return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
201}
202
203/*
204 * Return the root node of the specified rcu_state structure.
205 */
206static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
207{
208    return &rsp->node[0];
209}
210
211#ifdef CONFIG_SMP
212
213/*
214 * If the specified CPU is offline, tell the caller that it is in
215 * a quiescent state. Otherwise, whack it with a reschedule IPI.
216 * Grace periods can end up waiting on an offline CPU when that
217 * CPU is in the process of coming online -- it will be added to the
218 * rcu_node bitmasks before it actually makes it online. The same thing
219 * can happen while a CPU is in the process of coming online. Because this
220 * race is quite rare, we check for it after detecting that the grace
221 * period has been delayed rather than checking each and every CPU
222 * each and every time we start a new grace period.
223 */
224static int rcu_implicit_offline_qs(struct rcu_data *rdp)
225{
226    /*
227     * If the CPU is offline, it is in a quiescent state. We can
228     * trust its state not to change because interrupts are disabled.
229     */
230    if (cpu_is_offline(rdp->cpu)) {
231        rdp->offline_fqs++;
232        return 1;
233    }
234
235    /* If preemptable RCU, no point in sending reschedule IPI. */
236    if (rdp->preemptable)
237        return 0;
238
239    /* The CPU is online, so send it a reschedule IPI. */
240    if (rdp->cpu != smp_processor_id())
241        smp_send_reschedule(rdp->cpu);
242    else
243        set_need_resched();
244    rdp->resched_ipi++;
245    return 0;
246}
247
248#endif /* #ifdef CONFIG_SMP */
249
250#ifdef CONFIG_NO_HZ
251
252/**
253 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
254 *
255 * Enter nohz mode, in other words, -leave- the mode in which RCU
256 * read-side critical sections can occur. (Though RCU read-side
257 * critical sections can occur in irq handlers in nohz mode, a possibility
258 * handled by rcu_irq_enter() and rcu_irq_exit()).
259 */
260void rcu_enter_nohz(void)
261{
262    unsigned long flags;
263    struct rcu_dynticks *rdtp;
264
265    smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
266    local_irq_save(flags);
267    rdtp = &__get_cpu_var(rcu_dynticks);
268    rdtp->dynticks++;
269    rdtp->dynticks_nesting--;
270    WARN_ON_ONCE(rdtp->dynticks & 0x1);
271    local_irq_restore(flags);
272}
273
274/*
275 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
276 *
277 * Exit nohz mode, in other words, -enter- the mode in which RCU
278 * read-side critical sections normally occur.
279 */
280void rcu_exit_nohz(void)
281{
282    unsigned long flags;
283    struct rcu_dynticks *rdtp;
284
285    local_irq_save(flags);
286    rdtp = &__get_cpu_var(rcu_dynticks);
287    rdtp->dynticks++;
288    rdtp->dynticks_nesting++;
289    WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
290    local_irq_restore(flags);
291    smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
292}
293
294/**
295 * rcu_nmi_enter - inform RCU of entry to NMI context
296 *
297 * If the CPU was idle with dynamic ticks active, and there is no
298 * irq handler running, this updates rdtp->dynticks_nmi to let the
299 * RCU grace-period handling know that the CPU is active.
300 */
301void rcu_nmi_enter(void)
302{
303    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
304
305    if (rdtp->dynticks & 0x1)
306        return;
307    rdtp->dynticks_nmi++;
308    WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
309    smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
310}
311
312/**
313 * rcu_nmi_exit - inform RCU of exit from NMI context
314 *
315 * If the CPU was idle with dynamic ticks active, and there is no
316 * irq handler running, this updates rdtp->dynticks_nmi to let the
317 * RCU grace-period handling know that the CPU is no longer active.
318 */
319void rcu_nmi_exit(void)
320{
321    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
322
323    if (rdtp->dynticks & 0x1)
324        return;
325    smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
326    rdtp->dynticks_nmi++;
327    WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
328}
329
330/**
331 * rcu_irq_enter - inform RCU of entry to hard irq context
332 *
333 * If the CPU was idle with dynamic ticks active, this updates the
334 * rdtp->dynticks to let the RCU handling know that the CPU is active.
335 */
336void rcu_irq_enter(void)
337{
338    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
339
340    if (rdtp->dynticks_nesting++)
341        return;
342    rdtp->dynticks++;
343    WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
344    smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
345}
346
347/**
348 * rcu_irq_exit - inform RCU of exit from hard irq context
349 *
350 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
351 * to put let the RCU handling be aware that the CPU is going back to idle
352 * with no ticks.
353 */
354void rcu_irq_exit(void)
355{
356    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
357
358    if (--rdtp->dynticks_nesting)
359        return;
360    smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
361    rdtp->dynticks++;
362    WARN_ON_ONCE(rdtp->dynticks & 0x1);
363
364    /* If the interrupt queued a callback, get out of dyntick mode. */
365    if (__get_cpu_var(rcu_sched_data).nxtlist ||
366        __get_cpu_var(rcu_bh_data).nxtlist)
367        set_need_resched();
368}
369
370#ifdef CONFIG_SMP
371
372/*
373 * Snapshot the specified CPU's dynticks counter so that we can later
374 * credit them with an implicit quiescent state. Return 1 if this CPU
375 * is in dynticks idle mode, which is an extended quiescent state.
376 */
377static int dyntick_save_progress_counter(struct rcu_data *rdp)
378{
379    int ret;
380    int snap;
381    int snap_nmi;
382
383    snap = rdp->dynticks->dynticks;
384    snap_nmi = rdp->dynticks->dynticks_nmi;
385    smp_mb(); /* Order sampling of snap with end of grace period. */
386    rdp->dynticks_snap = snap;
387    rdp->dynticks_nmi_snap = snap_nmi;
388    ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
389    if (ret)
390        rdp->dynticks_fqs++;
391    return ret;
392}
393
394/*
395 * Return true if the specified CPU has passed through a quiescent
396 * state by virtue of being in or having passed through an dynticks
397 * idle state since the last call to dyntick_save_progress_counter()
398 * for this same CPU.
399 */
400static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
401{
402    long curr;
403    long curr_nmi;
404    long snap;
405    long snap_nmi;
406
407    curr = rdp->dynticks->dynticks;
408    snap = rdp->dynticks_snap;
409    curr_nmi = rdp->dynticks->dynticks_nmi;
410    snap_nmi = rdp->dynticks_nmi_snap;
411    smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
412
413    /*
414     * If the CPU passed through or entered a dynticks idle phase with
415     * no active irq/NMI handlers, then we can safely pretend that the CPU
416     * already acknowledged the request to pass through a quiescent
417     * state. Either way, that CPU cannot possibly be in an RCU
418     * read-side critical section that started before the beginning
419     * of the current RCU grace period.
420     */
421    if ((curr != snap || (curr & 0x1) == 0) &&
422        (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
423        rdp->dynticks_fqs++;
424        return 1;
425    }
426
427    /* Go check for the CPU being offline. */
428    return rcu_implicit_offline_qs(rdp);
429}
430
431#endif /* #ifdef CONFIG_SMP */
432
433#else /* #ifdef CONFIG_NO_HZ */
434
435#ifdef CONFIG_SMP
436
437static int dyntick_save_progress_counter(struct rcu_data *rdp)
438{
439    return 0;
440}
441
442static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
443{
444    return rcu_implicit_offline_qs(rdp);
445}
446
447#endif /* #ifdef CONFIG_SMP */
448
449#endif /* #else #ifdef CONFIG_NO_HZ */
450
451#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
452
453int rcu_cpu_stall_panicking __read_mostly;
454
455static void record_gp_stall_check_time(struct rcu_state *rsp)
456{
457    rsp->gp_start = jiffies;
458    rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
459}
460
461static void print_other_cpu_stall(struct rcu_state *rsp)
462{
463    int cpu;
464    long delta;
465    unsigned long flags;
466    struct rcu_node *rnp = rcu_get_root(rsp);
467
468    /* Only let one CPU complain about others per time interval. */
469
470    raw_spin_lock_irqsave(&rnp->lock, flags);
471    delta = jiffies - rsp->jiffies_stall;
472    if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
473        raw_spin_unlock_irqrestore(&rnp->lock, flags);
474        return;
475    }
476    rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
477
478    /*
479     * Now rat on any tasks that got kicked up to the root rcu_node
480     * due to CPU offlining.
481     */
482    rcu_print_task_stall(rnp);
483    raw_spin_unlock_irqrestore(&rnp->lock, flags);
484
485    /* OK, time to rat on our buddy... */
486
487    printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
488           rsp->name);
489    rcu_for_each_leaf_node(rsp, rnp) {
490        raw_spin_lock_irqsave(&rnp->lock, flags);
491        rcu_print_task_stall(rnp);
492        raw_spin_unlock_irqrestore(&rnp->lock, flags);
493        if (rnp->qsmask == 0)
494            continue;
495        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
496            if (rnp->qsmask & (1UL << cpu))
497                printk(" %d", rnp->grplo + cpu);
498    }
499    printk("} (detected by %d, t=%ld jiffies)\n",
500           smp_processor_id(), (long)(jiffies - rsp->gp_start));
501    trigger_all_cpu_backtrace();
502
503    /* If so configured, complain about tasks blocking the grace period. */
504
505    rcu_print_detail_task_stall(rsp);
506
507    force_quiescent_state(rsp, 0); /* Kick them all. */
508}
509
510static void print_cpu_stall(struct rcu_state *rsp)
511{
512    unsigned long flags;
513    struct rcu_node *rnp = rcu_get_root(rsp);
514
515    printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
516           rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
517    trigger_all_cpu_backtrace();
518
519    raw_spin_lock_irqsave(&rnp->lock, flags);
520    if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
521        rsp->jiffies_stall =
522            jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
523    raw_spin_unlock_irqrestore(&rnp->lock, flags);
524
525    set_need_resched(); /* kick ourselves to get things going. */
526}
527
528static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
529{
530    long delta;
531    struct rcu_node *rnp;
532
533    if (rcu_cpu_stall_panicking)
534        return;
535    delta = jiffies - rsp->jiffies_stall;
536    rnp = rdp->mynode;
537    if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
538
539        /* We haven't checked in, so go dump stack. */
540        print_cpu_stall(rsp);
541
542    } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
543
544        /* They had two time units to dump stack, so complain. */
545        print_other_cpu_stall(rsp);
546    }
547}
548
549static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
550{
551    rcu_cpu_stall_panicking = 1;
552    return NOTIFY_DONE;
553}
554
555static struct notifier_block rcu_panic_block = {
556    .notifier_call = rcu_panic,
557};
558
559static void __init check_cpu_stall_init(void)
560{
561    atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
562}
563
564#else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
565
566static void record_gp_stall_check_time(struct rcu_state *rsp)
567{
568}
569
570static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
571{
572}
573
574static void __init check_cpu_stall_init(void)
575{
576}
577
578#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
579
580/*
581 * Update CPU-local rcu_data state to record the newly noticed grace period.
582 * This is used both when we started the grace period and when we notice
583 * that someone else started the grace period. The caller must hold the
584 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
585 * and must have irqs disabled.
586 */
587static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
588{
589    if (rdp->gpnum != rnp->gpnum) {
590        rdp->qs_pending = 1;
591        rdp->passed_quiesc = 0;
592        rdp->gpnum = rnp->gpnum;
593    }
594}
595
596static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
597{
598    unsigned long flags;
599    struct rcu_node *rnp;
600
601    local_irq_save(flags);
602    rnp = rdp->mynode;
603    if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
604        !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
605        local_irq_restore(flags);
606        return;
607    }
608    __note_new_gpnum(rsp, rnp, rdp);
609    raw_spin_unlock_irqrestore(&rnp->lock, flags);
610}
611
612/*
613 * Did someone else start a new RCU grace period start since we last
614 * checked? Update local state appropriately if so. Must be called
615 * on the CPU corresponding to rdp.
616 */
617static int
618check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
619{
620    unsigned long flags;
621    int ret = 0;
622
623    local_irq_save(flags);
624    if (rdp->gpnum != rsp->gpnum) {
625        note_new_gpnum(rsp, rdp);
626        ret = 1;
627    }
628    local_irq_restore(flags);
629    return ret;
630}
631
632/*
633 * Advance this CPU's callbacks, but only if the current grace period
634 * has ended. This may be called only from the CPU to whom the rdp
635 * belongs. In addition, the corresponding leaf rcu_node structure's
636 * ->lock must be held by the caller, with irqs disabled.
637 */
638static void
639__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
640{
641    /* Did another grace period end? */
642    if (rdp->completed != rnp->completed) {
643
644        /* Advance callbacks. No harm if list empty. */
645        rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
646        rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
647        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
648
649        /* Remember that we saw this grace-period completion. */
650        rdp->completed = rnp->completed;
651    }
652}
653
654/*
655 * Advance this CPU's callbacks, but only if the current grace period
656 * has ended. This may be called only from the CPU to whom the rdp
657 * belongs.
658 */
659static void
660rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
661{
662    unsigned long flags;
663    struct rcu_node *rnp;
664
665    local_irq_save(flags);
666    rnp = rdp->mynode;
667    if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
668        !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
669        local_irq_restore(flags);
670        return;
671    }
672    __rcu_process_gp_end(rsp, rnp, rdp);
673    raw_spin_unlock_irqrestore(&rnp->lock, flags);
674}
675
676/*
677 * Do per-CPU grace-period initialization for running CPU. The caller
678 * must hold the lock of the leaf rcu_node structure corresponding to
679 * this CPU.
680 */
681static void
682rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
683{
684    /* Prior grace period ended, so advance callbacks for current CPU. */
685    __rcu_process_gp_end(rsp, rnp, rdp);
686
687    /*
688     * Because this CPU just now started the new grace period, we know
689     * that all of its callbacks will be covered by this upcoming grace
690     * period, even the ones that were registered arbitrarily recently.
691     * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
692     *
693     * Other CPUs cannot be sure exactly when the grace period started.
694     * Therefore, their recently registered callbacks must pass through
695     * an additional RCU_NEXT_READY stage, so that they will be handled
696     * by the next RCU grace period.
697     */
698    rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
699    rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
700
701    /* Set state so that this CPU will detect the next quiescent state. */
702    __note_new_gpnum(rsp, rnp, rdp);
703}
704
705/*
706 * Start a new RCU grace period if warranted, re-initializing the hierarchy
707 * in preparation for detecting the next grace period. The caller must hold
708 * the root node's ->lock, which is released before return. Hard irqs must
709 * be disabled.
710 */
711static void
712rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
713    __releases(rcu_get_root(rsp)->lock)
714{
715    struct rcu_data *rdp = rsp->rda[smp_processor_id()];
716    struct rcu_node *rnp = rcu_get_root(rsp);
717
718    if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
719        if (cpu_needs_another_gp(rsp, rdp))
720            rsp->fqs_need_gp = 1;
721        if (rnp->completed == rsp->completed) {
722            raw_spin_unlock_irqrestore(&rnp->lock, flags);
723            return;
724        }
725        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
726
727        /*
728         * Propagate new ->completed value to rcu_node structures
729         * so that other CPUs don't have to wait until the start
730         * of the next grace period to process their callbacks.
731         */
732        rcu_for_each_node_breadth_first(rsp, rnp) {
733            raw_spin_lock(&rnp->lock); /* irqs already disabled. */
734            rnp->completed = rsp->completed;
735            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
736        }
737        local_irq_restore(flags);
738        return;
739    }
740
741    /* Advance to a new grace period and initialize state. */
742    rsp->gpnum++;
743    WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
744    rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
745    rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
746    record_gp_stall_check_time(rsp);
747
748    /* Special-case the common single-level case. */
749    if (NUM_RCU_NODES == 1) {
750        rcu_preempt_check_blocked_tasks(rnp);
751        rnp->qsmask = rnp->qsmaskinit;
752        rnp->gpnum = rsp->gpnum;
753        rnp->completed = rsp->completed;
754        rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
755        rcu_start_gp_per_cpu(rsp, rnp, rdp);
756        raw_spin_unlock_irqrestore(&rnp->lock, flags);
757        return;
758    }
759
760    raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
761
762
763    /* Exclude any concurrent CPU-hotplug operations. */
764    raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
765
766    /*
767     * Set the quiescent-state-needed bits in all the rcu_node
768     * structures for all currently online CPUs in breadth-first
769     * order, starting from the root rcu_node structure. This
770     * operation relies on the layout of the hierarchy within the
771     * rsp->node[] array. Note that other CPUs will access only
772     * the leaves of the hierarchy, which still indicate that no
773     * grace period is in progress, at least until the corresponding
774     * leaf node has been initialized. In addition, we have excluded
775     * CPU-hotplug operations.
776     *
777     * Note that the grace period cannot complete until we finish
778     * the initialization process, as there will be at least one
779     * qsmask bit set in the root node until that time, namely the
780     * one corresponding to this CPU, due to the fact that we have
781     * irqs disabled.
782     */
783    rcu_for_each_node_breadth_first(rsp, rnp) {
784        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
785        rcu_preempt_check_blocked_tasks(rnp);
786        rnp->qsmask = rnp->qsmaskinit;
787        rnp->gpnum = rsp->gpnum;
788        rnp->completed = rsp->completed;
789        if (rnp == rdp->mynode)
790            rcu_start_gp_per_cpu(rsp, rnp, rdp);
791        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
792    }
793
794    rnp = rcu_get_root(rsp);
795    raw_spin_lock(&rnp->lock); /* irqs already disabled. */
796    rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
797    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
798    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
799}
800
801/*
802 * Report a full set of quiescent states to the specified rcu_state
803 * data structure. This involves cleaning up after the prior grace
804 * period and letting rcu_start_gp() start up the next grace period
805 * if one is needed. Note that the caller must hold rnp->lock, as
806 * required by rcu_start_gp(), which will release it.
807 */
808static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
809    __releases(rcu_get_root(rsp)->lock)
810{
811    WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
812    rsp->completed = rsp->gpnum;
813    rsp->signaled = RCU_GP_IDLE;
814    rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
815}
816
817/*
818 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
819 * Allows quiescent states for a group of CPUs to be reported at one go
820 * to the specified rcu_node structure, though all the CPUs in the group
821 * must be represented by the same rcu_node structure (which need not be
822 * a leaf rcu_node structure, though it often will be). That structure's
823 * lock must be held upon entry, and it is released before return.
824 */
825static void
826rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
827          struct rcu_node *rnp, unsigned long flags)
828    __releases(rnp->lock)
829{
830    struct rcu_node *rnp_c;
831
832    /* Walk up the rcu_node hierarchy. */
833    for (;;) {
834        if (!(rnp->qsmask & mask)) {
835
836            /* Our bit has already been cleared, so done. */
837            raw_spin_unlock_irqrestore(&rnp->lock, flags);
838            return;
839        }
840        rnp->qsmask &= ~mask;
841        if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
842
843            /* Other bits still set at this level, so done. */
844            raw_spin_unlock_irqrestore(&rnp->lock, flags);
845            return;
846        }
847        mask = rnp->grpmask;
848        if (rnp->parent == NULL) {
849
850            /* No more levels. Exit loop holding root lock. */
851
852            break;
853        }
854        raw_spin_unlock_irqrestore(&rnp->lock, flags);
855        rnp_c = rnp;
856        rnp = rnp->parent;
857        raw_spin_lock_irqsave(&rnp->lock, flags);
858        WARN_ON_ONCE(rnp_c->qsmask);
859    }
860
861    /*
862     * Get here if we are the last CPU to pass through a quiescent
863     * state for this grace period. Invoke rcu_report_qs_rsp()
864     * to clean up and start the next grace period if one is needed.
865     */
866    rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
867}
868
869/*
870 * Record a quiescent state for the specified CPU to that CPU's rcu_data
871 * structure. This must be either called from the specified CPU, or
872 * called when the specified CPU is known to be offline (and when it is
873 * also known that no other CPU is concurrently trying to help the offline
874 * CPU). The lastcomp argument is used to make sure we are still in the
875 * grace period of interest. We don't want to end the current grace period
876 * based on quiescent states detected in an earlier grace period!
877 */
878static void
879rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
880{
881    unsigned long flags;
882    unsigned long mask;
883    struct rcu_node *rnp;
884
885    rnp = rdp->mynode;
886    raw_spin_lock_irqsave(&rnp->lock, flags);
887    if (lastcomp != rnp->completed) {
888
889        /*
890         * Someone beat us to it for this grace period, so leave.
891         * The race with GP start is resolved by the fact that we
892         * hold the leaf rcu_node lock, so that the per-CPU bits
893         * cannot yet be initialized -- so we would simply find our
894         * CPU's bit already cleared in rcu_report_qs_rnp() if this
895         * race occurred.
896         */
897        rdp->passed_quiesc = 0; /* try again later! */
898        raw_spin_unlock_irqrestore(&rnp->lock, flags);
899        return;
900    }
901    mask = rdp->grpmask;
902    if ((rnp->qsmask & mask) == 0) {
903        raw_spin_unlock_irqrestore(&rnp->lock, flags);
904    } else {
905        rdp->qs_pending = 0;
906
907        /*
908         * This GP can't end until cpu checks in, so all of our
909         * callbacks can be processed during the next GP.
910         */
911        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
912
913        rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
914    }
915}
916
917/*
918 * Check to see if there is a new grace period of which this CPU
919 * is not yet aware, and if so, set up local rcu_data state for it.
920 * Otherwise, see if this CPU has just passed through its first
921 * quiescent state for this grace period, and record that fact if so.
922 */
923static void
924rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
925{
926    /* If there is now a new grace period, record and return. */
927    if (check_for_new_grace_period(rsp, rdp))
928        return;
929
930    /*
931     * Does this CPU still need to do its part for current grace period?
932     * If no, return and let the other CPUs do their part as well.
933     */
934    if (!rdp->qs_pending)
935        return;
936
937    /*
938     * Was there a quiescent state since the beginning of the grace
939     * period? If no, then exit and wait for the next call.
940     */
941    if (!rdp->passed_quiesc)
942        return;
943
944    /*
945     * Tell RCU we are done (but rcu_report_qs_rdp() will be the
946     * judge of that).
947     */
948    rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
949}
950
951#ifdef CONFIG_HOTPLUG_CPU
952
953/*
954 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
955 * specified flavor of RCU. The callbacks will be adopted by the next
956 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
957 * comes first. Because this is invoked from the CPU_DYING notifier,
958 * irqs are already disabled.
959 */
960static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
961{
962    int i;
963    struct rcu_data *rdp = rsp->rda[smp_processor_id()];
964
965    if (rdp->nxtlist == NULL)
966        return; /* irqs disabled, so comparison is stable. */
967    raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
968    *rsp->orphan_cbs_tail = rdp->nxtlist;
969    rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
970    rdp->nxtlist = NULL;
971    for (i = 0; i < RCU_NEXT_SIZE; i++)
972        rdp->nxttail[i] = &rdp->nxtlist;
973    rsp->orphan_qlen += rdp->qlen;
974    rdp->qlen = 0;
975    raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
976}
977
978/*
979 * Adopt previously orphaned RCU callbacks.
980 */
981static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
982{
983    unsigned long flags;
984    struct rcu_data *rdp;
985
986    raw_spin_lock_irqsave(&rsp->onofflock, flags);
987    rdp = rsp->rda[smp_processor_id()];
988    if (rsp->orphan_cbs_list == NULL) {
989        raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
990        return;
991    }
992    *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
993    rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
994    rdp->qlen += rsp->orphan_qlen;
995    rsp->orphan_cbs_list = NULL;
996    rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
997    rsp->orphan_qlen = 0;
998    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
999}
1000
1001/*
1002 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1003 * and move all callbacks from the outgoing CPU to the current one.
1004 */
1005static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1006{
1007    unsigned long flags;
1008    unsigned long mask;
1009    int need_report = 0;
1010    struct rcu_data *rdp = rsp->rda[cpu];
1011    struct rcu_node *rnp;
1012
1013    /* Exclude any attempts to start a new grace period. */
1014    raw_spin_lock_irqsave(&rsp->onofflock, flags);
1015
1016    /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1017    rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1018    mask = rdp->grpmask; /* rnp->grplo is constant. */
1019    do {
1020        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1021        rnp->qsmaskinit &= ~mask;
1022        if (rnp->qsmaskinit != 0) {
1023            if (rnp != rdp->mynode)
1024                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1025            break;
1026        }
1027        if (rnp == rdp->mynode)
1028            need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1029        else
1030            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1031        mask = rnp->grpmask;
1032        rnp = rnp->parent;
1033    } while (rnp != NULL);
1034
1035    /*
1036     * We still hold the leaf rcu_node structure lock here, and
1037     * irqs are still disabled. The reason for this subterfuge is
1038     * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1039     * held leads to deadlock.
1040     */
1041    raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1042    rnp = rdp->mynode;
1043    if (need_report & RCU_OFL_TASKS_NORM_GP)
1044        rcu_report_unblock_qs_rnp(rnp, flags);
1045    else
1046        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1047    if (need_report & RCU_OFL_TASKS_EXP_GP)
1048        rcu_report_exp_rnp(rsp, rnp);
1049
1050    rcu_adopt_orphan_cbs(rsp);
1051}
1052
1053/*
1054 * Remove the specified CPU from the RCU hierarchy and move any pending
1055 * callbacks that it might have to the current CPU. This code assumes
1056 * that at least one CPU in the system will remain running at all times.
1057 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1058 */
1059static void rcu_offline_cpu(int cpu)
1060{
1061    __rcu_offline_cpu(cpu, &rcu_sched_state);
1062    __rcu_offline_cpu(cpu, &rcu_bh_state);
1063    rcu_preempt_offline_cpu(cpu);
1064}
1065
1066#else /* #ifdef CONFIG_HOTPLUG_CPU */
1067
1068static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1069{
1070}
1071
1072static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1073{
1074}
1075
1076static void rcu_offline_cpu(int cpu)
1077{
1078}
1079
1080#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1081
1082/*
1083 * Invoke any RCU callbacks that have made it to the end of their grace
1084 * period. Thottle as specified by rdp->blimit.
1085 */
1086static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1087{
1088    unsigned long flags;
1089    struct rcu_head *next, *list, **tail;
1090    int count;
1091
1092    /* If no callbacks are ready, just return.*/
1093    if (!cpu_has_callbacks_ready_to_invoke(rdp))
1094        return;
1095
1096    /*
1097     * Extract the list of ready callbacks, disabling to prevent
1098     * races with call_rcu() from interrupt handlers.
1099     */
1100    local_irq_save(flags);
1101    list = rdp->nxtlist;
1102    rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1103    *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1104    tail = rdp->nxttail[RCU_DONE_TAIL];
1105    for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1106        if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1107            rdp->nxttail[count] = &rdp->nxtlist;
1108    local_irq_restore(flags);
1109
1110    /* Invoke callbacks. */
1111    count = 0;
1112    while (list) {
1113        next = list->next;
1114        prefetch(next);
1115        debug_rcu_head_unqueue(list);
1116        list->func(list);
1117        list = next;
1118        if (++count >= rdp->blimit)
1119            break;
1120    }
1121
1122    local_irq_save(flags);
1123
1124    /* Update count, and requeue any remaining callbacks. */
1125    rdp->qlen -= count;
1126    if (list != NULL) {
1127        *tail = rdp->nxtlist;
1128        rdp->nxtlist = list;
1129        for (count = 0; count < RCU_NEXT_SIZE; count++)
1130            if (&rdp->nxtlist == rdp->nxttail[count])
1131                rdp->nxttail[count] = tail;
1132            else
1133                break;
1134    }
1135
1136    /* Reinstate batch limit if we have worked down the excess. */
1137    if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1138        rdp->blimit = blimit;
1139
1140    /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1141    if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1142        rdp->qlen_last_fqs_check = 0;
1143        rdp->n_force_qs_snap = rsp->n_force_qs;
1144    } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1145        rdp->qlen_last_fqs_check = rdp->qlen;
1146
1147    local_irq_restore(flags);
1148
1149    /* Re-raise the RCU softirq if there are callbacks remaining. */
1150    if (cpu_has_callbacks_ready_to_invoke(rdp))
1151        raise_softirq(RCU_SOFTIRQ);
1152}
1153
1154/*
1155 * Check to see if this CPU is in a non-context-switch quiescent state
1156 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1157 * Also schedule the RCU softirq handler.
1158 *
1159 * This function must be called with hardirqs disabled. It is normally
1160 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1161 * false, there is no point in invoking rcu_check_callbacks().
1162 */
1163void rcu_check_callbacks(int cpu, int user)
1164{
1165    if (user ||
1166        (idle_cpu(cpu) && rcu_scheduler_active &&
1167         !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1168
1169        /*
1170         * Get here if this CPU took its interrupt from user
1171         * mode or from the idle loop, and if this is not a
1172         * nested interrupt. In this case, the CPU is in
1173         * a quiescent state, so note it.
1174         *
1175         * No memory barrier is required here because both
1176         * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1177         * variables that other CPUs neither access nor modify,
1178         * at least not while the corresponding CPU is online.
1179         */
1180
1181        rcu_sched_qs(cpu);
1182        rcu_bh_qs(cpu);
1183
1184    } else if (!in_softirq()) {
1185
1186        /*
1187         * Get here if this CPU did not take its interrupt from
1188         * softirq, in other words, if it is not interrupting
1189         * a rcu_bh read-side critical section. This is an _bh
1190         * critical section, so note it.
1191         */
1192
1193        rcu_bh_qs(cpu);
1194    }
1195    rcu_preempt_check_callbacks(cpu);
1196    if (rcu_pending(cpu))
1197        raise_softirq(RCU_SOFTIRQ);
1198}
1199
1200#ifdef CONFIG_SMP
1201
1202/*
1203 * Scan the leaf rcu_node structures, processing dyntick state for any that
1204 * have not yet encountered a quiescent state, using the function specified.
1205 * The caller must have suppressed start of new grace periods.
1206 */
1207static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1208{
1209    unsigned long bit;
1210    int cpu;
1211    unsigned long flags;
1212    unsigned long mask;
1213    struct rcu_node *rnp;
1214
1215    rcu_for_each_leaf_node(rsp, rnp) {
1216        mask = 0;
1217        raw_spin_lock_irqsave(&rnp->lock, flags);
1218        if (!rcu_gp_in_progress(rsp)) {
1219            raw_spin_unlock_irqrestore(&rnp->lock, flags);
1220            return;
1221        }
1222        if (rnp->qsmask == 0) {
1223            raw_spin_unlock_irqrestore(&rnp->lock, flags);
1224            continue;
1225        }
1226        cpu = rnp->grplo;
1227        bit = 1;
1228        for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1229            if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1230                mask |= bit;
1231        }
1232        if (mask != 0) {
1233
1234            /* rcu_report_qs_rnp() releases rnp->lock. */
1235            rcu_report_qs_rnp(mask, rsp, rnp, flags);
1236            continue;
1237        }
1238        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1239    }
1240}
1241
1242/*
1243 * Force quiescent states on reluctant CPUs, and also detect which
1244 * CPUs are in dyntick-idle mode.
1245 */
1246static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1247{
1248    unsigned long flags;
1249    struct rcu_node *rnp = rcu_get_root(rsp);
1250
1251    if (!rcu_gp_in_progress(rsp))
1252        return; /* No grace period in progress, nothing to force. */
1253    if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1254        rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1255        return; /* Someone else is already on the job. */
1256    }
1257    if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1258        goto unlock_fqs_ret; /* no emergency and done recently. */
1259    rsp->n_force_qs++;
1260    raw_spin_lock(&rnp->lock); /* irqs already disabled */
1261    rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1262    if(!rcu_gp_in_progress(rsp)) {
1263        rsp->n_force_qs_ngp++;
1264        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1265        goto unlock_fqs_ret; /* no GP in progress, time updated. */
1266    }
1267    rsp->fqs_active = 1;
1268    switch (rsp->signaled) {
1269    case RCU_GP_IDLE:
1270    case RCU_GP_INIT:
1271
1272        break; /* grace period idle or initializing, ignore. */
1273
1274    case RCU_SAVE_DYNTICK:
1275        if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1276            break; /* So gcc recognizes the dead code. */
1277
1278        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1279
1280        /* Record dyntick-idle state. */
1281        force_qs_rnp(rsp, dyntick_save_progress_counter);
1282        raw_spin_lock(&rnp->lock); /* irqs already disabled */
1283        if (rcu_gp_in_progress(rsp))
1284            rsp->signaled = RCU_FORCE_QS;
1285        break;
1286
1287    case RCU_FORCE_QS:
1288
1289        /* Check dyntick-idle state, send IPI to laggarts. */
1290        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1291        force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1292
1293        /* Leave state in case more forcing is required. */
1294
1295        raw_spin_lock(&rnp->lock); /* irqs already disabled */
1296        break;
1297    }
1298    rsp->fqs_active = 0;
1299    if (rsp->fqs_need_gp) {
1300        raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1301        rsp->fqs_need_gp = 0;
1302        rcu_start_gp(rsp, flags); /* releases rnp->lock */
1303        return;
1304    }
1305    raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1306unlock_fqs_ret:
1307    raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1308}
1309
1310#else /* #ifdef CONFIG_SMP */
1311
1312static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1313{
1314    set_need_resched();
1315}
1316
1317#endif /* #else #ifdef CONFIG_SMP */
1318
1319/*
1320 * This does the RCU processing work from softirq context for the
1321 * specified rcu_state and rcu_data structures. This may be called
1322 * only from the CPU to whom the rdp belongs.
1323 */
1324static void
1325__rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1326{
1327    unsigned long flags;
1328
1329    WARN_ON_ONCE(rdp->beenonline == 0);
1330
1331    /*
1332     * If an RCU GP has gone long enough, go check for dyntick
1333     * idle CPUs and, if needed, send resched IPIs.
1334     */
1335    if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1336        force_quiescent_state(rsp, 1);
1337
1338    /*
1339     * Advance callbacks in response to end of earlier grace
1340     * period that some other CPU ended.
1341     */
1342    rcu_process_gp_end(rsp, rdp);
1343
1344    /* Update RCU state based on any recent quiescent states. */
1345    rcu_check_quiescent_state(rsp, rdp);
1346
1347    /* Does this CPU require a not-yet-started grace period? */
1348    if (cpu_needs_another_gp(rsp, rdp)) {
1349        raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1350        rcu_start_gp(rsp, flags); /* releases above lock */
1351    }
1352
1353    /* If there are callbacks ready, invoke them. */
1354    rcu_do_batch(rsp, rdp);
1355}
1356
1357/*
1358 * Do softirq processing for the current CPU.
1359 */
1360static void rcu_process_callbacks(struct softirq_action *unused)
1361{
1362    /*
1363     * Memory references from any prior RCU read-side critical sections
1364     * executed by the interrupted code must be seen before any RCU
1365     * grace-period manipulations below.
1366     */
1367    smp_mb(); /* See above block comment. */
1368
1369    __rcu_process_callbacks(&rcu_sched_state,
1370                &__get_cpu_var(rcu_sched_data));
1371    __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1372    rcu_preempt_process_callbacks();
1373
1374    /*
1375     * Memory references from any later RCU read-side critical sections
1376     * executed by the interrupted code must be seen after any RCU
1377     * grace-period manipulations above.
1378     */
1379    smp_mb(); /* See above block comment. */
1380
1381    /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1382    rcu_needs_cpu_flush();
1383}
1384
1385static void
1386__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1387       struct rcu_state *rsp)
1388{
1389    unsigned long flags;
1390    struct rcu_data *rdp;
1391
1392    debug_rcu_head_queue(head);
1393    head->func = func;
1394    head->next = NULL;
1395
1396    smp_mb(); /* Ensure RCU update seen before callback registry. */
1397
1398    /*
1399     * Opportunistically note grace-period endings and beginnings.
1400     * Note that we might see a beginning right after we see an
1401     * end, but never vice versa, since this CPU has to pass through
1402     * a quiescent state betweentimes.
1403     */
1404    local_irq_save(flags);
1405    rdp = rsp->rda[smp_processor_id()];
1406    rcu_process_gp_end(rsp, rdp);
1407    check_for_new_grace_period(rsp, rdp);
1408
1409    /* Add the callback to our list. */
1410    *rdp->nxttail[RCU_NEXT_TAIL] = head;
1411    rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1412
1413    /* Start a new grace period if one not already started. */
1414    if (!rcu_gp_in_progress(rsp)) {
1415        unsigned long nestflag;
1416        struct rcu_node *rnp_root = rcu_get_root(rsp);
1417
1418        raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1419        rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1420    }
1421
1422    /*
1423     * Force the grace period if too many callbacks or too long waiting.
1424     * Enforce hysteresis, and don't invoke force_quiescent_state()
1425     * if some other CPU has recently done so. Also, don't bother
1426     * invoking force_quiescent_state() if the newly enqueued callback
1427     * is the only one waiting for a grace period to complete.
1428     */
1429    if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1430        rdp->blimit = LONG_MAX;
1431        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1432            *rdp->nxttail[RCU_DONE_TAIL] != head)
1433            force_quiescent_state(rsp, 0);
1434        rdp->n_force_qs_snap = rsp->n_force_qs;
1435        rdp->qlen_last_fqs_check = rdp->qlen;
1436    } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1437        force_quiescent_state(rsp, 1);
1438    local_irq_restore(flags);
1439}
1440
1441/*
1442 * Queue an RCU-sched callback for invocation after a grace period.
1443 */
1444void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1445{
1446    __call_rcu(head, func, &rcu_sched_state);
1447}
1448EXPORT_SYMBOL_GPL(call_rcu_sched);
1449
1450/*
1451 * Queue an RCU for invocation after a quicker grace period.
1452 */
1453void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1454{
1455    __call_rcu(head, func, &rcu_bh_state);
1456}
1457EXPORT_SYMBOL_GPL(call_rcu_bh);
1458
1459/**
1460 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1461 *
1462 * Control will return to the caller some time after a full rcu-sched
1463 * grace period has elapsed, in other words after all currently executing
1464 * rcu-sched read-side critical sections have completed. These read-side
1465 * critical sections are delimited by rcu_read_lock_sched() and
1466 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1467 * local_irq_disable(), and so on may be used in place of
1468 * rcu_read_lock_sched().
1469 *
1470 * This means that all preempt_disable code sequences, including NMI and
1471 * hardware-interrupt handlers, in progress on entry will have completed
1472 * before this primitive returns. However, this does not guarantee that
1473 * softirq handlers will have completed, since in some kernels, these
1474 * handlers can run in process context, and can block.
1475 *
1476 * This primitive provides the guarantees made by the (now removed)
1477 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1478 * guarantees that rcu_read_lock() sections will have completed.
1479 * In "classic RCU", these two guarantees happen to be one and
1480 * the same, but can differ in realtime RCU implementations.
1481 */
1482void synchronize_sched(void)
1483{
1484    struct rcu_synchronize rcu;
1485
1486    if (rcu_blocking_is_gp())
1487        return;
1488
1489    init_rcu_head_on_stack(&rcu.head);
1490    init_completion(&rcu.completion);
1491    /* Will wake me after RCU finished. */
1492    call_rcu_sched(&rcu.head, wakeme_after_rcu);
1493    /* Wait for it. */
1494    wait_for_completion(&rcu.completion);
1495    destroy_rcu_head_on_stack(&rcu.head);
1496}
1497EXPORT_SYMBOL_GPL(synchronize_sched);
1498
1499/**
1500 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1501 *
1502 * Control will return to the caller some time after a full rcu_bh grace
1503 * period has elapsed, in other words after all currently executing rcu_bh
1504 * read-side critical sections have completed. RCU read-side critical
1505 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1506 * and may be nested.
1507 */
1508void synchronize_rcu_bh(void)
1509{
1510    struct rcu_synchronize rcu;
1511
1512    if (rcu_blocking_is_gp())
1513        return;
1514
1515    init_rcu_head_on_stack(&rcu.head);
1516    init_completion(&rcu.completion);
1517    /* Will wake me after RCU finished. */
1518    call_rcu_bh(&rcu.head, wakeme_after_rcu);
1519    /* Wait for it. */
1520    wait_for_completion(&rcu.completion);
1521    destroy_rcu_head_on_stack(&rcu.head);
1522}
1523EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1524
1525/*
1526 * Check to see if there is any immediate RCU-related work to be done
1527 * by the current CPU, for the specified type of RCU, returning 1 if so.
1528 * The checks are in order of increasing expense: checks that can be
1529 * carried out against CPU-local state are performed first. However,
1530 * we must check for CPU stalls first, else we might not get a chance.
1531 */
1532static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1533{
1534    struct rcu_node *rnp = rdp->mynode;
1535
1536    rdp->n_rcu_pending++;
1537
1538    /* Check for CPU stalls, if enabled. */
1539    check_cpu_stall(rsp, rdp);
1540
1541    /* Is the RCU core waiting for a quiescent state from this CPU? */
1542    if (rdp->qs_pending && !rdp->passed_quiesc) {
1543
1544        /*
1545         * If force_quiescent_state() coming soon and this CPU
1546         * needs a quiescent state, and this is either RCU-sched
1547         * or RCU-bh, force a local reschedule.
1548         */
1549        rdp->n_rp_qs_pending++;
1550        if (!rdp->preemptable &&
1551            ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1552                 jiffies))
1553            set_need_resched();
1554    } else if (rdp->qs_pending && rdp->passed_quiesc) {
1555        rdp->n_rp_report_qs++;
1556        return 1;
1557    }
1558
1559    /* Does this CPU have callbacks ready to invoke? */
1560    if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1561        rdp->n_rp_cb_ready++;
1562        return 1;
1563    }
1564
1565    /* Has RCU gone idle with this CPU needing another grace period? */
1566    if (cpu_needs_another_gp(rsp, rdp)) {
1567        rdp->n_rp_cpu_needs_gp++;
1568        return 1;
1569    }
1570
1571    /* Has another RCU grace period completed? */
1572    if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1573        rdp->n_rp_gp_completed++;
1574        return 1;
1575    }
1576
1577    /* Has a new RCU grace period started? */
1578    if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1579        rdp->n_rp_gp_started++;
1580        return 1;
1581    }
1582
1583    /* Has an RCU GP gone long enough to send resched IPIs &c? */
1584    if (rcu_gp_in_progress(rsp) &&
1585        ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1586        rdp->n_rp_need_fqs++;
1587        return 1;
1588    }
1589
1590    /* nothing to do */
1591    rdp->n_rp_need_nothing++;
1592    return 0;
1593}
1594
1595/*
1596 * Check to see if there is any immediate RCU-related work to be done
1597 * by the current CPU, returning 1 if so. This function is part of the
1598 * RCU implementation; it is -not- an exported member of the RCU API.
1599 */
1600static int rcu_pending(int cpu)
1601{
1602    return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1603           __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1604           rcu_preempt_pending(cpu);
1605}
1606
1607/*
1608 * Check to see if any future RCU-related work will need to be done
1609 * by the current CPU, even if none need be done immediately, returning
1610 * 1 if so.
1611 */
1612static int rcu_needs_cpu_quick_check(int cpu)
1613{
1614    /* RCU callbacks either ready or pending? */
1615    return per_cpu(rcu_sched_data, cpu).nxtlist ||
1616           per_cpu(rcu_bh_data, cpu).nxtlist ||
1617           rcu_preempt_needs_cpu(cpu);
1618}
1619
1620static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1621static atomic_t rcu_barrier_cpu_count;
1622static DEFINE_MUTEX(rcu_barrier_mutex);
1623static struct completion rcu_barrier_completion;
1624
1625static void rcu_barrier_callback(struct rcu_head *notused)
1626{
1627    if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1628        complete(&rcu_barrier_completion);
1629}
1630
1631/*
1632 * Called with preemption disabled, and from cross-cpu IRQ context.
1633 */
1634static void rcu_barrier_func(void *type)
1635{
1636    int cpu = smp_processor_id();
1637    struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1638    void (*call_rcu_func)(struct rcu_head *head,
1639                  void (*func)(struct rcu_head *head));
1640
1641    atomic_inc(&rcu_barrier_cpu_count);
1642    call_rcu_func = type;
1643    call_rcu_func(head, rcu_barrier_callback);
1644}
1645
1646/*
1647 * Orchestrate the specified type of RCU barrier, waiting for all
1648 * RCU callbacks of the specified type to complete.
1649 */
1650static void _rcu_barrier(struct rcu_state *rsp,
1651             void (*call_rcu_func)(struct rcu_head *head,
1652                           void (*func)(struct rcu_head *head)))
1653{
1654    BUG_ON(in_interrupt());
1655    /* Take mutex to serialize concurrent rcu_barrier() requests. */
1656    mutex_lock(&rcu_barrier_mutex);
1657    init_completion(&rcu_barrier_completion);
1658    /*
1659     * Initialize rcu_barrier_cpu_count to 1, then invoke
1660     * rcu_barrier_func() on each CPU, so that each CPU also has
1661     * incremented rcu_barrier_cpu_count. Only then is it safe to
1662     * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1663     * might complete its grace period before all of the other CPUs
1664     * did their increment, causing this function to return too
1665     * early.
1666     */
1667    atomic_set(&rcu_barrier_cpu_count, 1);
1668    preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1669    rcu_adopt_orphan_cbs(rsp);
1670    on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1671    preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1672    if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1673        complete(&rcu_barrier_completion);
1674    wait_for_completion(&rcu_barrier_completion);
1675    mutex_unlock(&rcu_barrier_mutex);
1676}
1677
1678/**
1679 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1680 */
1681void rcu_barrier_bh(void)
1682{
1683    _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1684}
1685EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1686
1687/**
1688 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1689 */
1690void rcu_barrier_sched(void)
1691{
1692    _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1693}
1694EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1695
1696/*
1697 * Do boot-time initialization of a CPU's per-CPU RCU data.
1698 */
1699static void __init
1700rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1701{
1702    unsigned long flags;
1703    int i;
1704    struct rcu_data *rdp = rsp->rda[cpu];
1705    struct rcu_node *rnp = rcu_get_root(rsp);
1706
1707    /* Set up local state, ensuring consistent view of global state. */
1708    raw_spin_lock_irqsave(&rnp->lock, flags);
1709    rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1710    rdp->nxtlist = NULL;
1711    for (i = 0; i < RCU_NEXT_SIZE; i++)
1712        rdp->nxttail[i] = &rdp->nxtlist;
1713    rdp->qlen = 0;
1714#ifdef CONFIG_NO_HZ
1715    rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1716#endif /* #ifdef CONFIG_NO_HZ */
1717    rdp->cpu = cpu;
1718    raw_spin_unlock_irqrestore(&rnp->lock, flags);
1719}
1720
1721/*
1722 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1723 * offline event can be happening at a given time. Note also that we
1724 * can accept some slop in the rsp->completed access due to the fact
1725 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1726 */
1727static void __cpuinit
1728rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1729{
1730    unsigned long flags;
1731    unsigned long mask;
1732    struct rcu_data *rdp = rsp->rda[cpu];
1733    struct rcu_node *rnp = rcu_get_root(rsp);
1734
1735    /* Set up local state, ensuring consistent view of global state. */
1736    raw_spin_lock_irqsave(&rnp->lock, flags);
1737    rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1738    rdp->qs_pending = 1; /* so set up to respond to current GP. */
1739    rdp->beenonline = 1; /* We have now been online. */
1740    rdp->preemptable = preemptable;
1741    rdp->qlen_last_fqs_check = 0;
1742    rdp->n_force_qs_snap = rsp->n_force_qs;
1743    rdp->blimit = blimit;
1744    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1745
1746    /*
1747     * A new grace period might start here. If so, we won't be part
1748     * of it, but that is OK, as we are currently in a quiescent state.
1749     */
1750
1751    /* Exclude any attempts to start a new GP on large systems. */
1752    raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1753
1754    /* Add CPU to rcu_node bitmasks. */
1755    rnp = rdp->mynode;
1756    mask = rdp->grpmask;
1757    do {
1758        /* Exclude any attempts to start a new GP on small systems. */
1759        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1760        rnp->qsmaskinit |= mask;
1761        mask = rnp->grpmask;
1762        if (rnp == rdp->mynode) {
1763            rdp->gpnum = rnp->completed; /* if GP in progress... */
1764            rdp->completed = rnp->completed;
1765            rdp->passed_quiesc_completed = rnp->completed - 1;
1766        }
1767        raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1768        rnp = rnp->parent;
1769    } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1770
1771    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1772}
1773
1774static void __cpuinit rcu_online_cpu(int cpu)
1775{
1776    rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1777    rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1778    rcu_preempt_init_percpu_data(cpu);
1779}
1780
1781/*
1782 * Handle CPU online/offline notification events.
1783 */
1784static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1785                    unsigned long action, void *hcpu)
1786{
1787    long cpu = (long)hcpu;
1788
1789    switch (action) {
1790    case CPU_UP_PREPARE:
1791    case CPU_UP_PREPARE_FROZEN:
1792        rcu_online_cpu(cpu);
1793        break;
1794    case CPU_DYING:
1795    case CPU_DYING_FROZEN:
1796        /*
1797         * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1798         * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1799         * returns, all online cpus have queued rcu_barrier_func().
1800         * The dying CPU clears its cpu_online_mask bit and
1801         * moves all of its RCU callbacks to ->orphan_cbs_list
1802         * in the context of stop_machine(), so subsequent calls
1803         * to _rcu_barrier() will adopt these callbacks and only
1804         * then queue rcu_barrier_func() on all remaining CPUs.
1805         */
1806        rcu_send_cbs_to_orphanage(&rcu_bh_state);
1807        rcu_send_cbs_to_orphanage(&rcu_sched_state);
1808        rcu_preempt_send_cbs_to_orphanage();
1809        break;
1810    case CPU_DEAD:
1811    case CPU_DEAD_FROZEN:
1812    case CPU_UP_CANCELED:
1813    case CPU_UP_CANCELED_FROZEN:
1814        rcu_offline_cpu(cpu);
1815        break;
1816    default:
1817        break;
1818    }
1819    return NOTIFY_OK;
1820}
1821
1822/*
1823 * This function is invoked towards the end of the scheduler's initialization
1824 * process. Before this is called, the idle task might contain
1825 * RCU read-side critical sections (during which time, this idle
1826 * task is booting the system). After this function is called, the
1827 * idle tasks are prohibited from containing RCU read-side critical
1828 * sections. This function also enables RCU lockdep checking.
1829 */
1830void rcu_scheduler_starting(void)
1831{
1832    WARN_ON(num_online_cpus() != 1);
1833    WARN_ON(nr_context_switches() > 0);
1834    rcu_scheduler_active = 1;
1835}
1836
1837/*
1838 * Compute the per-level fanout, either using the exact fanout specified
1839 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1840 */
1841#ifdef CONFIG_RCU_FANOUT_EXACT
1842static void __init rcu_init_levelspread(struct rcu_state *rsp)
1843{
1844    int i;
1845
1846    for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1847        rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1848}
1849#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1850static void __init rcu_init_levelspread(struct rcu_state *rsp)
1851{
1852    int ccur;
1853    int cprv;
1854    int i;
1855
1856    cprv = NR_CPUS;
1857    for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1858        ccur = rsp->levelcnt[i];
1859        rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1860        cprv = ccur;
1861    }
1862}
1863#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1864
1865/*
1866 * Helper function for rcu_init() that initializes one rcu_state structure.
1867 */
1868static void __init rcu_init_one(struct rcu_state *rsp)
1869{
1870    static char *buf[] = { "rcu_node_level_0",
1871                   "rcu_node_level_1",
1872                   "rcu_node_level_2",
1873                   "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1874    int cpustride = 1;
1875    int i;
1876    int j;
1877    struct rcu_node *rnp;
1878
1879    BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1880
1881    /* Initialize the level-tracking arrays. */
1882
1883    for (i = 1; i < NUM_RCU_LVLS; i++)
1884        rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1885    rcu_init_levelspread(rsp);
1886
1887    /* Initialize the elements themselves, starting from the leaves. */
1888
1889    for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1890        cpustride *= rsp->levelspread[i];
1891        rnp = rsp->level[i];
1892        for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1893            raw_spin_lock_init(&rnp->lock);
1894            lockdep_set_class_and_name(&rnp->lock,
1895                           &rcu_node_class[i], buf[i]);
1896            rnp->gpnum = 0;
1897            rnp->qsmask = 0;
1898            rnp->qsmaskinit = 0;
1899            rnp->grplo = j * cpustride;
1900            rnp->grphi = (j + 1) * cpustride - 1;
1901            if (rnp->grphi >= NR_CPUS)
1902                rnp->grphi = NR_CPUS - 1;
1903            if (i == 0) {
1904                rnp->grpnum = 0;
1905                rnp->grpmask = 0;
1906                rnp->parent = NULL;
1907            } else {
1908                rnp->grpnum = j % rsp->levelspread[i - 1];
1909                rnp->grpmask = 1UL << rnp->grpnum;
1910                rnp->parent = rsp->level[i - 1] +
1911                          j / rsp->levelspread[i - 1];
1912            }
1913            rnp->level = i;
1914            INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1915            INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1916            INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1917            INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1918        }
1919    }
1920
1921    rnp = rsp->level[NUM_RCU_LVLS - 1];
1922    for_each_possible_cpu(i) {
1923        while (i > rnp->grphi)
1924            rnp++;
1925        rsp->rda[i]->mynode = rnp;
1926        rcu_boot_init_percpu_data(i, rsp);
1927    }
1928}
1929
1930/*
1931 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1932 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1933 * structure.
1934 */
1935#define RCU_INIT_FLAVOR(rsp, rcu_data) \
1936do { \
1937    int i; \
1938    \
1939    for_each_possible_cpu(i) { \
1940        (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1941    } \
1942    rcu_init_one(rsp); \
1943} while (0)
1944
1945void __init rcu_init(void)
1946{
1947    int cpu;
1948
1949    rcu_bootup_announce();
1950    RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1951    RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1952    __rcu_init_preempt();
1953    open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1954
1955    /*
1956     * We don't need protection against CPU-hotplug here because
1957     * this is called early in boot, before either interrupts
1958     * or the scheduler are operational.
1959     */
1960    cpu_notifier(rcu_cpu_notify, 0);
1961    for_each_online_cpu(cpu)
1962        rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1963    check_cpu_stall_init();
1964}
1965
1966#include "rcutree_plugin.h"
1967

Archive Download this file



interactive