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 <linux/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#include <linux/wait.h>
51#include <linux/kthread.h>
52#include <linux/prefetch.h>
53
54#include "rcutree.h"
55
56/* Data structures. */
57
58static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
59
60#define RCU_STATE_INITIALIZER(structname) { \
61    .level = { &structname.node[0] }, \
62    .levelcnt = { \
63        NUM_RCU_LVL_0, /* root of hierarchy. */ \
64        NUM_RCU_LVL_1, \
65        NUM_RCU_LVL_2, \
66        NUM_RCU_LVL_3, \
67        NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
68    }, \
69    .signaled = RCU_GP_IDLE, \
70    .gpnum = -300, \
71    .completed = -300, \
72    .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
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
85static struct rcu_state *rcu_state;
86
87/*
88 * The rcu_scheduler_active variable transitions from zero to one just
89 * before the first task is spawned. So when this variable is zero, RCU
90 * can assume that there is but one task, allowing RCU to (for example)
91 * optimized synchronize_sched() to a simple barrier(). When this variable
92 * is one, RCU must actually do all the hard work required to detect real
93 * grace periods. This variable is also used to suppress boot-time false
94 * positives from lockdep-RCU error checking.
95 */
96int rcu_scheduler_active __read_mostly;
97EXPORT_SYMBOL_GPL(rcu_scheduler_active);
98
99/*
100 * The rcu_scheduler_fully_active variable transitions from zero to one
101 * during the early_initcall() processing, which is after the scheduler
102 * is capable of creating new tasks. So RCU processing (for example,
103 * creating tasks for RCU priority boosting) must be delayed until after
104 * rcu_scheduler_fully_active transitions from zero to one. We also
105 * currently delay invocation of any RCU callbacks until after this point.
106 *
107 * It might later prove better for people registering RCU callbacks during
108 * early boot to take responsibility for these callbacks, but one step at
109 * a time.
110 */
111static int rcu_scheduler_fully_active __read_mostly;
112
113#ifdef CONFIG_RCU_BOOST
114
115/*
116 * Control variables for per-CPU and per-rcu_node kthreads. These
117 * handle all flavors of RCU.
118 */
119static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
120DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
121DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
122DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
123DEFINE_PER_CPU(char, rcu_cpu_has_work);
124
125#endif /* #ifdef CONFIG_RCU_BOOST */
126
127static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
128static void invoke_rcu_core(void);
129static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
130
131#define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
132
133/*
134 * Track the rcutorture test sequence number and the update version
135 * number within a given test. The rcutorture_testseq is incremented
136 * on every rcutorture module load and unload, so has an odd value
137 * when a test is running. The rcutorture_vernum is set to zero
138 * when rcutorture starts and is incremented on each rcutorture update.
139 * These variables enable correlating rcutorture output with the
140 * RCU tracing information.
141 */
142unsigned long rcutorture_testseq;
143unsigned long rcutorture_vernum;
144
145/*
146 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
147 * permit this function to be invoked without holding the root rcu_node
148 * structure's ->lock, but of course results can be subject to change.
149 */
150static int rcu_gp_in_progress(struct rcu_state *rsp)
151{
152    return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
153}
154
155/*
156 * Note a quiescent state. Because we do not need to know
157 * how many quiescent states passed, just if there was at least
158 * one since the start of the grace period, this just sets a flag.
159 */
160void rcu_sched_qs(int cpu)
161{
162    struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
163
164    rdp->passed_quiesc_completed = rdp->gpnum - 1;
165    barrier();
166    rdp->passed_quiesc = 1;
167}
168
169void rcu_bh_qs(int cpu)
170{
171    struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
172
173    rdp->passed_quiesc_completed = rdp->gpnum - 1;
174    barrier();
175    rdp->passed_quiesc = 1;
176}
177
178/*
179 * Note a context switch. This is a quiescent state for RCU-sched,
180 * and requires special handling for preemptible RCU.
181 */
182void rcu_note_context_switch(int cpu)
183{
184    rcu_sched_qs(cpu);
185    rcu_preempt_note_context_switch(cpu);
186}
187EXPORT_SYMBOL_GPL(rcu_note_context_switch);
188
189#ifdef CONFIG_NO_HZ
190DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
191    .dynticks_nesting = 1,
192    .dynticks = ATOMIC_INIT(1),
193};
194#endif /* #ifdef CONFIG_NO_HZ */
195
196static int blimit = 10; /* Maximum callbacks per softirq. */
197static int qhimark = 10000; /* If this many pending, ignore blimit. */
198static int qlowmark = 100; /* Once only this many pending, use blimit. */
199
200module_param(blimit, int, 0);
201module_param(qhimark, int, 0);
202module_param(qlowmark, int, 0);
203
204int rcu_cpu_stall_suppress __read_mostly;
205module_param(rcu_cpu_stall_suppress, int, 0644);
206
207static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
208static int rcu_pending(int cpu);
209
210/*
211 * Return the number of RCU-sched batches processed thus far for debug & stats.
212 */
213long rcu_batches_completed_sched(void)
214{
215    return rcu_sched_state.completed;
216}
217EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
218
219/*
220 * Return the number of RCU BH batches processed thus far for debug & stats.
221 */
222long rcu_batches_completed_bh(void)
223{
224    return rcu_bh_state.completed;
225}
226EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
227
228/*
229 * Force a quiescent state for RCU BH.
230 */
231void rcu_bh_force_quiescent_state(void)
232{
233    force_quiescent_state(&rcu_bh_state, 0);
234}
235EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
236
237/*
238 * Record the number of times rcutorture tests have been initiated and
239 * terminated. This information allows the debugfs tracing stats to be
240 * correlated to the rcutorture messages, even when the rcutorture module
241 * is being repeatedly loaded and unloaded. In other words, we cannot
242 * store this state in rcutorture itself.
243 */
244void rcutorture_record_test_transition(void)
245{
246    rcutorture_testseq++;
247    rcutorture_vernum = 0;
248}
249EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
250
251/*
252 * Record the number of writer passes through the current rcutorture test.
253 * This is also used to correlate debugfs tracing stats with the rcutorture
254 * messages.
255 */
256void rcutorture_record_progress(unsigned long vernum)
257{
258    rcutorture_vernum++;
259}
260EXPORT_SYMBOL_GPL(rcutorture_record_progress);
261
262/*
263 * Force a quiescent state for RCU-sched.
264 */
265void rcu_sched_force_quiescent_state(void)
266{
267    force_quiescent_state(&rcu_sched_state, 0);
268}
269EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
270
271/*
272 * Does the CPU have callbacks ready to be invoked?
273 */
274static int
275cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
276{
277    return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
278}
279
280/*
281 * Does the current CPU require a yet-as-unscheduled grace period?
282 */
283static int
284cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
285{
286    return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
287}
288
289/*
290 * Return the root node of the specified rcu_state structure.
291 */
292static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
293{
294    return &rsp->node[0];
295}
296
297#ifdef CONFIG_SMP
298
299/*
300 * If the specified CPU is offline, tell the caller that it is in
301 * a quiescent state. Otherwise, whack it with a reschedule IPI.
302 * Grace periods can end up waiting on an offline CPU when that
303 * CPU is in the process of coming online -- it will be added to the
304 * rcu_node bitmasks before it actually makes it online. The same thing
305 * can happen while a CPU is in the process of coming online. Because this
306 * race is quite rare, we check for it after detecting that the grace
307 * period has been delayed rather than checking each and every CPU
308 * each and every time we start a new grace period.
309 */
310static int rcu_implicit_offline_qs(struct rcu_data *rdp)
311{
312    /*
313     * If the CPU is offline, it is in a quiescent state. We can
314     * trust its state not to change because interrupts are disabled.
315     */
316    if (cpu_is_offline(rdp->cpu)) {
317        rdp->offline_fqs++;
318        return 1;
319    }
320
321    /* If preemptible RCU, no point in sending reschedule IPI. */
322    if (rdp->preemptible)
323        return 0;
324
325    /* The CPU is online, so send it a reschedule IPI. */
326    if (rdp->cpu != smp_processor_id())
327        smp_send_reschedule(rdp->cpu);
328    else
329        set_need_resched();
330    rdp->resched_ipi++;
331    return 0;
332}
333
334#endif /* #ifdef CONFIG_SMP */
335
336#ifdef CONFIG_NO_HZ
337
338/**
339 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
340 *
341 * Enter nohz mode, in other words, -leave- the mode in which RCU
342 * read-side critical sections can occur. (Though RCU read-side
343 * critical sections can occur in irq handlers in nohz mode, a possibility
344 * handled by rcu_irq_enter() and rcu_irq_exit()).
345 */
346void rcu_enter_nohz(void)
347{
348    unsigned long flags;
349    struct rcu_dynticks *rdtp;
350
351    local_irq_save(flags);
352    rdtp = &__get_cpu_var(rcu_dynticks);
353    if (--rdtp->dynticks_nesting) {
354        local_irq_restore(flags);
355        return;
356    }
357    /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
358    smp_mb__before_atomic_inc(); /* See above. */
359    atomic_inc(&rdtp->dynticks);
360    smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
361    WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
362    local_irq_restore(flags);
363
364    /* If the interrupt queued a callback, get out of dyntick mode. */
365    if (in_irq() &&
366        (__get_cpu_var(rcu_sched_data).nxtlist ||
367         __get_cpu_var(rcu_bh_data).nxtlist ||
368         rcu_preempt_needs_cpu(smp_processor_id())))
369        set_need_resched();
370}
371
372/*
373 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
374 *
375 * Exit nohz mode, in other words, -enter- the mode in which RCU
376 * read-side critical sections normally occur.
377 */
378void rcu_exit_nohz(void)
379{
380    unsigned long flags;
381    struct rcu_dynticks *rdtp;
382
383    local_irq_save(flags);
384    rdtp = &__get_cpu_var(rcu_dynticks);
385    if (rdtp->dynticks_nesting++) {
386        local_irq_restore(flags);
387        return;
388    }
389    smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
390    atomic_inc(&rdtp->dynticks);
391    /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
392    smp_mb__after_atomic_inc(); /* See above. */
393    WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
394    local_irq_restore(flags);
395}
396
397/**
398 * rcu_nmi_enter - inform RCU of entry to NMI context
399 *
400 * If the CPU was idle with dynamic ticks active, and there is no
401 * irq handler running, this updates rdtp->dynticks_nmi to let the
402 * RCU grace-period handling know that the CPU is active.
403 */
404void rcu_nmi_enter(void)
405{
406    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
407
408    if (rdtp->dynticks_nmi_nesting == 0 &&
409        (atomic_read(&rdtp->dynticks) & 0x1))
410        return;
411    rdtp->dynticks_nmi_nesting++;
412    smp_mb__before_atomic_inc(); /* Force delay from prior write. */
413    atomic_inc(&rdtp->dynticks);
414    /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
415    smp_mb__after_atomic_inc(); /* See above. */
416    WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
417}
418
419/**
420 * rcu_nmi_exit - inform RCU of exit from NMI context
421 *
422 * If the CPU was idle with dynamic ticks active, and there is no
423 * irq handler running, this updates rdtp->dynticks_nmi to let the
424 * RCU grace-period handling know that the CPU is no longer active.
425 */
426void rcu_nmi_exit(void)
427{
428    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
429
430    if (rdtp->dynticks_nmi_nesting == 0 ||
431        --rdtp->dynticks_nmi_nesting != 0)
432        return;
433    /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
434    smp_mb__before_atomic_inc(); /* See above. */
435    atomic_inc(&rdtp->dynticks);
436    smp_mb__after_atomic_inc(); /* Force delay to next write. */
437    WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
438}
439
440/**
441 * rcu_irq_enter - inform RCU of entry to hard irq context
442 *
443 * If the CPU was idle with dynamic ticks active, this updates the
444 * rdtp->dynticks to let the RCU handling know that the CPU is active.
445 */
446void rcu_irq_enter(void)
447{
448    rcu_exit_nohz();
449}
450
451/**
452 * rcu_irq_exit - inform RCU of exit from hard irq context
453 *
454 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
455 * to put let the RCU handling be aware that the CPU is going back to idle
456 * with no ticks.
457 */
458void rcu_irq_exit(void)
459{
460    rcu_enter_nohz();
461}
462
463#ifdef CONFIG_SMP
464
465/*
466 * Snapshot the specified CPU's dynticks counter so that we can later
467 * credit them with an implicit quiescent state. Return 1 if this CPU
468 * is in dynticks idle mode, which is an extended quiescent state.
469 */
470static int dyntick_save_progress_counter(struct rcu_data *rdp)
471{
472    rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
473    return 0;
474}
475
476/*
477 * Return true if the specified CPU has passed through a quiescent
478 * state by virtue of being in or having passed through an dynticks
479 * idle state since the last call to dyntick_save_progress_counter()
480 * for this same CPU.
481 */
482static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
483{
484    unsigned long curr;
485    unsigned long snap;
486
487    curr = (unsigned long)atomic_add_return(0, &rdp->dynticks->dynticks);
488    snap = (unsigned long)rdp->dynticks_snap;
489
490    /*
491     * If the CPU passed through or entered a dynticks idle phase with
492     * no active irq/NMI handlers, then we can safely pretend that the CPU
493     * already acknowledged the request to pass through a quiescent
494     * state. Either way, that CPU cannot possibly be in an RCU
495     * read-side critical section that started before the beginning
496     * of the current RCU grace period.
497     */
498    if ((curr & 0x1) == 0 || ULONG_CMP_GE(curr, snap + 2)) {
499        rdp->dynticks_fqs++;
500        return 1;
501    }
502
503    /* Go check for the CPU being offline. */
504    return rcu_implicit_offline_qs(rdp);
505}
506
507#endif /* #ifdef CONFIG_SMP */
508
509#else /* #ifdef CONFIG_NO_HZ */
510
511#ifdef CONFIG_SMP
512
513static int dyntick_save_progress_counter(struct rcu_data *rdp)
514{
515    return 0;
516}
517
518static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
519{
520    return rcu_implicit_offline_qs(rdp);
521}
522
523#endif /* #ifdef CONFIG_SMP */
524
525#endif /* #else #ifdef CONFIG_NO_HZ */
526
527int rcu_cpu_stall_suppress __read_mostly;
528
529static void record_gp_stall_check_time(struct rcu_state *rsp)
530{
531    rsp->gp_start = jiffies;
532    rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
533}
534
535static void print_other_cpu_stall(struct rcu_state *rsp)
536{
537    int cpu;
538    long delta;
539    unsigned long flags;
540    struct rcu_node *rnp = rcu_get_root(rsp);
541
542    /* Only let one CPU complain about others per time interval. */
543
544    raw_spin_lock_irqsave(&rnp->lock, flags);
545    delta = jiffies - rsp->jiffies_stall;
546    if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
547        raw_spin_unlock_irqrestore(&rnp->lock, flags);
548        return;
549    }
550    rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
551
552    /*
553     * Now rat on any tasks that got kicked up to the root rcu_node
554     * due to CPU offlining.
555     */
556    rcu_print_task_stall(rnp);
557    raw_spin_unlock_irqrestore(&rnp->lock, flags);
558
559    /*
560     * OK, time to rat on our buddy...
561     * See Documentation/RCU/stallwarn.txt for info on how to debug
562     * RCU CPU stall warnings.
563     */
564    printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
565           rsp->name);
566    rcu_for_each_leaf_node(rsp, rnp) {
567        raw_spin_lock_irqsave(&rnp->lock, flags);
568        rcu_print_task_stall(rnp);
569        raw_spin_unlock_irqrestore(&rnp->lock, flags);
570        if (rnp->qsmask == 0)
571            continue;
572        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
573            if (rnp->qsmask & (1UL << cpu))
574                printk(" %d", rnp->grplo + cpu);
575    }
576    printk("} (detected by %d, t=%ld jiffies)\n",
577           smp_processor_id(), (long)(jiffies - rsp->gp_start));
578    trigger_all_cpu_backtrace();
579
580    /* If so configured, complain about tasks blocking the grace period. */
581
582    rcu_print_detail_task_stall(rsp);
583
584    force_quiescent_state(rsp, 0); /* Kick them all. */
585}
586
587static void print_cpu_stall(struct rcu_state *rsp)
588{
589    unsigned long flags;
590    struct rcu_node *rnp = rcu_get_root(rsp);
591
592    /*
593     * OK, time to rat on ourselves...
594     * See Documentation/RCU/stallwarn.txt for info on how to debug
595     * RCU CPU stall warnings.
596     */
597    printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
598           rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
599    trigger_all_cpu_backtrace();
600
601    raw_spin_lock_irqsave(&rnp->lock, flags);
602    if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
603        rsp->jiffies_stall =
604            jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
605    raw_spin_unlock_irqrestore(&rnp->lock, flags);
606
607    set_need_resched(); /* kick ourselves to get things going. */
608}
609
610static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
611{
612    unsigned long j;
613    unsigned long js;
614    struct rcu_node *rnp;
615
616    if (rcu_cpu_stall_suppress)
617        return;
618    j = ACCESS_ONCE(jiffies);
619    js = ACCESS_ONCE(rsp->jiffies_stall);
620    rnp = rdp->mynode;
621    if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
622
623        /* We haven't checked in, so go dump stack. */
624        print_cpu_stall(rsp);
625
626    } else if (rcu_gp_in_progress(rsp) &&
627           ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
628
629        /* They had a few time units to dump stack, so complain. */
630        print_other_cpu_stall(rsp);
631    }
632}
633
634static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
635{
636    rcu_cpu_stall_suppress = 1;
637    return NOTIFY_DONE;
638}
639
640/**
641 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
642 *
643 * Set the stall-warning timeout way off into the future, thus preventing
644 * any RCU CPU stall-warning messages from appearing in the current set of
645 * RCU grace periods.
646 *
647 * The caller must disable hard irqs.
648 */
649void rcu_cpu_stall_reset(void)
650{
651    rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
652    rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
653    rcu_preempt_stall_reset();
654}
655
656static struct notifier_block rcu_panic_block = {
657    .notifier_call = rcu_panic,
658};
659
660static void __init check_cpu_stall_init(void)
661{
662    atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
663}
664
665/*
666 * Update CPU-local rcu_data state to record the newly noticed grace period.
667 * This is used both when we started the grace period and when we notice
668 * that someone else started the grace period. The caller must hold the
669 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
670 * and must have irqs disabled.
671 */
672static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
673{
674    if (rdp->gpnum != rnp->gpnum) {
675        /*
676         * If the current grace period is waiting for this CPU,
677         * set up to detect a quiescent state, otherwise don't
678         * go looking for one.
679         */
680        rdp->gpnum = rnp->gpnum;
681        if (rnp->qsmask & rdp->grpmask) {
682            rdp->qs_pending = 1;
683            rdp->passed_quiesc = 0;
684        } else
685            rdp->qs_pending = 0;
686    }
687}
688
689static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
690{
691    unsigned long flags;
692    struct rcu_node *rnp;
693
694    local_irq_save(flags);
695    rnp = rdp->mynode;
696    if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
697        !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
698        local_irq_restore(flags);
699        return;
700    }
701    __note_new_gpnum(rsp, rnp, rdp);
702    raw_spin_unlock_irqrestore(&rnp->lock, flags);
703}
704
705/*
706 * Did someone else start a new RCU grace period start since we last
707 * checked? Update local state appropriately if so. Must be called
708 * on the CPU corresponding to rdp.
709 */
710static int
711check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
712{
713    unsigned long flags;
714    int ret = 0;
715
716    local_irq_save(flags);
717    if (rdp->gpnum != rsp->gpnum) {
718        note_new_gpnum(rsp, rdp);
719        ret = 1;
720    }
721    local_irq_restore(flags);
722    return ret;
723}
724
725/*
726 * Advance this CPU's callbacks, but only if the current grace period
727 * has ended. This may be called only from the CPU to whom the rdp
728 * belongs. In addition, the corresponding leaf rcu_node structure's
729 * ->lock must be held by the caller, with irqs disabled.
730 */
731static void
732__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
733{
734    /* Did another grace period end? */
735    if (rdp->completed != rnp->completed) {
736
737        /* Advance callbacks. No harm if list empty. */
738        rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
739        rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
740        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
741
742        /* Remember that we saw this grace-period completion. */
743        rdp->completed = rnp->completed;
744
745        /*
746         * If we were in an extended quiescent state, we may have
747         * missed some grace periods that others CPUs handled on
748         * our behalf. Catch up with this state to avoid noting
749         * spurious new grace periods. If another grace period
750         * has started, then rnp->gpnum will have advanced, so
751         * we will detect this later on.
752         */
753        if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
754            rdp->gpnum = rdp->completed;
755
756        /*
757         * If RCU does not need a quiescent state from this CPU,
758         * then make sure that this CPU doesn't go looking for one.
759         */
760        if ((rnp->qsmask & rdp->grpmask) == 0)
761            rdp->qs_pending = 0;
762    }
763}
764
765/*
766 * Advance this CPU's callbacks, but only if the current grace period
767 * has ended. This may be called only from the CPU to whom the rdp
768 * belongs.
769 */
770static void
771rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
772{
773    unsigned long flags;
774    struct rcu_node *rnp;
775
776    local_irq_save(flags);
777    rnp = rdp->mynode;
778    if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
779        !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
780        local_irq_restore(flags);
781        return;
782    }
783    __rcu_process_gp_end(rsp, rnp, rdp);
784    raw_spin_unlock_irqrestore(&rnp->lock, flags);
785}
786
787/*
788 * Do per-CPU grace-period initialization for running CPU. The caller
789 * must hold the lock of the leaf rcu_node structure corresponding to
790 * this CPU.
791 */
792static void
793rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
794{
795    /* Prior grace period ended, so advance callbacks for current CPU. */
796    __rcu_process_gp_end(rsp, rnp, rdp);
797
798    /*
799     * Because this CPU just now started the new grace period, we know
800     * that all of its callbacks will be covered by this upcoming grace
801     * period, even the ones that were registered arbitrarily recently.
802     * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
803     *
804     * Other CPUs cannot be sure exactly when the grace period started.
805     * Therefore, their recently registered callbacks must pass through
806     * an additional RCU_NEXT_READY stage, so that they will be handled
807     * by the next RCU grace period.
808     */
809    rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
810    rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
811
812    /* Set state so that this CPU will detect the next quiescent state. */
813    __note_new_gpnum(rsp, rnp, rdp);
814}
815
816/*
817 * Start a new RCU grace period if warranted, re-initializing the hierarchy
818 * in preparation for detecting the next grace period. The caller must hold
819 * the root node's ->lock, which is released before return. Hard irqs must
820 * be disabled.
821 */
822static void
823rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
824    __releases(rcu_get_root(rsp)->lock)
825{
826    struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
827    struct rcu_node *rnp = rcu_get_root(rsp);
828
829    if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
830        if (cpu_needs_another_gp(rsp, rdp))
831            rsp->fqs_need_gp = 1;
832        if (rnp->completed == rsp->completed) {
833            raw_spin_unlock_irqrestore(&rnp->lock, flags);
834            return;
835        }
836        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
837
838        /*
839         * Propagate new ->completed value to rcu_node structures
840         * so that other CPUs don't have to wait until the start
841         * of the next grace period to process their callbacks.
842         */
843        rcu_for_each_node_breadth_first(rsp, rnp) {
844            raw_spin_lock(&rnp->lock); /* irqs already disabled. */
845            rnp->completed = rsp->completed;
846            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
847        }
848        local_irq_restore(flags);
849        return;
850    }
851
852    /* Advance to a new grace period and initialize state. */
853    rsp->gpnum++;
854    WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
855    rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
856    rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
857    record_gp_stall_check_time(rsp);
858
859    /* Special-case the common single-level case. */
860    if (NUM_RCU_NODES == 1) {
861        rcu_preempt_check_blocked_tasks(rnp);
862        rnp->qsmask = rnp->qsmaskinit;
863        rnp->gpnum = rsp->gpnum;
864        rnp->completed = rsp->completed;
865        rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
866        rcu_start_gp_per_cpu(rsp, rnp, rdp);
867        rcu_preempt_boost_start_gp(rnp);
868        raw_spin_unlock_irqrestore(&rnp->lock, flags);
869        return;
870    }
871
872    raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
873
874
875    /* Exclude any concurrent CPU-hotplug operations. */
876    raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
877
878    /*
879     * Set the quiescent-state-needed bits in all the rcu_node
880     * structures for all currently online CPUs in breadth-first
881     * order, starting from the root rcu_node structure. This
882     * operation relies on the layout of the hierarchy within the
883     * rsp->node[] array. Note that other CPUs will access only
884     * the leaves of the hierarchy, which still indicate that no
885     * grace period is in progress, at least until the corresponding
886     * leaf node has been initialized. In addition, we have excluded
887     * CPU-hotplug operations.
888     *
889     * Note that the grace period cannot complete until we finish
890     * the initialization process, as there will be at least one
891     * qsmask bit set in the root node until that time, namely the
892     * one corresponding to this CPU, due to the fact that we have
893     * irqs disabled.
894     */
895    rcu_for_each_node_breadth_first(rsp, rnp) {
896        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
897        rcu_preempt_check_blocked_tasks(rnp);
898        rnp->qsmask = rnp->qsmaskinit;
899        rnp->gpnum = rsp->gpnum;
900        rnp->completed = rsp->completed;
901        if (rnp == rdp->mynode)
902            rcu_start_gp_per_cpu(rsp, rnp, rdp);
903        rcu_preempt_boost_start_gp(rnp);
904        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
905    }
906
907    rnp = rcu_get_root(rsp);
908    raw_spin_lock(&rnp->lock); /* irqs already disabled. */
909    rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
910    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
911    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
912}
913
914/*
915 * Report a full set of quiescent states to the specified rcu_state
916 * data structure. This involves cleaning up after the prior grace
917 * period and letting rcu_start_gp() start up the next grace period
918 * if one is needed. Note that the caller must hold rnp->lock, as
919 * required by rcu_start_gp(), which will release it.
920 */
921static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
922    __releases(rcu_get_root(rsp)->lock)
923{
924    unsigned long gp_duration;
925
926    WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
927
928    /*
929     * Ensure that all grace-period and pre-grace-period activity
930     * is seen before the assignment to rsp->completed.
931     */
932    smp_mb(); /* See above block comment. */
933    gp_duration = jiffies - rsp->gp_start;
934    if (gp_duration > rsp->gp_max)
935        rsp->gp_max = gp_duration;
936    rsp->completed = rsp->gpnum;
937    rsp->signaled = RCU_GP_IDLE;
938    rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
939}
940
941/*
942 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
943 * Allows quiescent states for a group of CPUs to be reported at one go
944 * to the specified rcu_node structure, though all the CPUs in the group
945 * must be represented by the same rcu_node structure (which need not be
946 * a leaf rcu_node structure, though it often will be). That structure's
947 * lock must be held upon entry, and it is released before return.
948 */
949static void
950rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
951          struct rcu_node *rnp, unsigned long flags)
952    __releases(rnp->lock)
953{
954    struct rcu_node *rnp_c;
955
956    /* Walk up the rcu_node hierarchy. */
957    for (;;) {
958        if (!(rnp->qsmask & mask)) {
959
960            /* Our bit has already been cleared, so done. */
961            raw_spin_unlock_irqrestore(&rnp->lock, flags);
962            return;
963        }
964        rnp->qsmask &= ~mask;
965        if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
966
967            /* Other bits still set at this level, so done. */
968            raw_spin_unlock_irqrestore(&rnp->lock, flags);
969            return;
970        }
971        mask = rnp->grpmask;
972        if (rnp->parent == NULL) {
973
974            /* No more levels. Exit loop holding root lock. */
975
976            break;
977        }
978        raw_spin_unlock_irqrestore(&rnp->lock, flags);
979        rnp_c = rnp;
980        rnp = rnp->parent;
981        raw_spin_lock_irqsave(&rnp->lock, flags);
982        WARN_ON_ONCE(rnp_c->qsmask);
983    }
984
985    /*
986     * Get here if we are the last CPU to pass through a quiescent
987     * state for this grace period. Invoke rcu_report_qs_rsp()
988     * to clean up and start the next grace period if one is needed.
989     */
990    rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
991}
992
993/*
994 * Record a quiescent state for the specified CPU to that CPU's rcu_data
995 * structure. This must be either called from the specified CPU, or
996 * called when the specified CPU is known to be offline (and when it is
997 * also known that no other CPU is concurrently trying to help the offline
998 * CPU). The lastcomp argument is used to make sure we are still in the
999 * grace period of interest. We don't want to end the current grace period
1000 * based on quiescent states detected in an earlier grace period!
1001 */
1002static void
1003rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
1004{
1005    unsigned long flags;
1006    unsigned long mask;
1007    struct rcu_node *rnp;
1008
1009    rnp = rdp->mynode;
1010    raw_spin_lock_irqsave(&rnp->lock, flags);
1011    if (lastcomp != rnp->completed) {
1012
1013        /*
1014         * Someone beat us to it for this grace period, so leave.
1015         * The race with GP start is resolved by the fact that we
1016         * hold the leaf rcu_node lock, so that the per-CPU bits
1017         * cannot yet be initialized -- so we would simply find our
1018         * CPU's bit already cleared in rcu_report_qs_rnp() if this
1019         * race occurred.
1020         */
1021        rdp->passed_quiesc = 0; /* try again later! */
1022        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1023        return;
1024    }
1025    mask = rdp->grpmask;
1026    if ((rnp->qsmask & mask) == 0) {
1027        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1028    } else {
1029        rdp->qs_pending = 0;
1030
1031        /*
1032         * This GP can't end until cpu checks in, so all of our
1033         * callbacks can be processed during the next GP.
1034         */
1035        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1036
1037        rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1038    }
1039}
1040
1041/*
1042 * Check to see if there is a new grace period of which this CPU
1043 * is not yet aware, and if so, set up local rcu_data state for it.
1044 * Otherwise, see if this CPU has just passed through its first
1045 * quiescent state for this grace period, and record that fact if so.
1046 */
1047static void
1048rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1049{
1050    /* If there is now a new grace period, record and return. */
1051    if (check_for_new_grace_period(rsp, rdp))
1052        return;
1053
1054    /*
1055     * Does this CPU still need to do its part for current grace period?
1056     * If no, return and let the other CPUs do their part as well.
1057     */
1058    if (!rdp->qs_pending)
1059        return;
1060
1061    /*
1062     * Was there a quiescent state since the beginning of the grace
1063     * period? If no, then exit and wait for the next call.
1064     */
1065    if (!rdp->passed_quiesc)
1066        return;
1067
1068    /*
1069     * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1070     * judge of that).
1071     */
1072    rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
1073}
1074
1075#ifdef CONFIG_HOTPLUG_CPU
1076
1077/*
1078 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1079 * Synchronization is not required because this function executes
1080 * in stop_machine() context.
1081 */
1082static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1083{
1084    int i;
1085    /* current DYING CPU is cleared in the cpu_online_mask */
1086    int receive_cpu = cpumask_any(cpu_online_mask);
1087    struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1088    struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1089
1090    if (rdp->nxtlist == NULL)
1091        return; /* irqs disabled, so comparison is stable. */
1092
1093    *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1094    receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1095    receive_rdp->qlen += rdp->qlen;
1096    receive_rdp->n_cbs_adopted += rdp->qlen;
1097    rdp->n_cbs_orphaned += rdp->qlen;
1098
1099    rdp->nxtlist = NULL;
1100    for (i = 0; i < RCU_NEXT_SIZE; i++)
1101        rdp->nxttail[i] = &rdp->nxtlist;
1102    rdp->qlen = 0;
1103}
1104
1105/*
1106 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1107 * and move all callbacks from the outgoing CPU to the current one.
1108 * There can only be one CPU hotplug operation at a time, so no other
1109 * CPU can be attempting to update rcu_cpu_kthread_task.
1110 */
1111static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1112{
1113    unsigned long flags;
1114    unsigned long mask;
1115    int need_report = 0;
1116    struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1117    struct rcu_node *rnp;
1118
1119    rcu_stop_cpu_kthread(cpu);
1120
1121    /* Exclude any attempts to start a new grace period. */
1122    raw_spin_lock_irqsave(&rsp->onofflock, flags);
1123
1124    /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1125    rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1126    mask = rdp->grpmask; /* rnp->grplo is constant. */
1127    do {
1128        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1129        rnp->qsmaskinit &= ~mask;
1130        if (rnp->qsmaskinit != 0) {
1131            if (rnp != rdp->mynode)
1132                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1133            break;
1134        }
1135        if (rnp == rdp->mynode)
1136            need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1137        else
1138            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1139        mask = rnp->grpmask;
1140        rnp = rnp->parent;
1141    } while (rnp != NULL);
1142
1143    /*
1144     * We still hold the leaf rcu_node structure lock here, and
1145     * irqs are still disabled. The reason for this subterfuge is
1146     * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1147     * held leads to deadlock.
1148     */
1149    raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1150    rnp = rdp->mynode;
1151    if (need_report & RCU_OFL_TASKS_NORM_GP)
1152        rcu_report_unblock_qs_rnp(rnp, flags);
1153    else
1154        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1155    if (need_report & RCU_OFL_TASKS_EXP_GP)
1156        rcu_report_exp_rnp(rsp, rnp);
1157    rcu_node_kthread_setaffinity(rnp, -1);
1158}
1159
1160/*
1161 * Remove the specified CPU from the RCU hierarchy and move any pending
1162 * callbacks that it might have to the current CPU. This code assumes
1163 * that at least one CPU in the system will remain running at all times.
1164 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1165 */
1166static void rcu_offline_cpu(int cpu)
1167{
1168    __rcu_offline_cpu(cpu, &rcu_sched_state);
1169    __rcu_offline_cpu(cpu, &rcu_bh_state);
1170    rcu_preempt_offline_cpu(cpu);
1171}
1172
1173#else /* #ifdef CONFIG_HOTPLUG_CPU */
1174
1175static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1176{
1177}
1178
1179static void rcu_offline_cpu(int cpu)
1180{
1181}
1182
1183#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1184
1185/*
1186 * Invoke any RCU callbacks that have made it to the end of their grace
1187 * period. Thottle as specified by rdp->blimit.
1188 */
1189static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1190{
1191    unsigned long flags;
1192    struct rcu_head *next, *list, **tail;
1193    int count;
1194
1195    /* If no callbacks are ready, just return.*/
1196    if (!cpu_has_callbacks_ready_to_invoke(rdp))
1197        return;
1198
1199    /*
1200     * Extract the list of ready callbacks, disabling to prevent
1201     * races with call_rcu() from interrupt handlers.
1202     */
1203    local_irq_save(flags);
1204    list = rdp->nxtlist;
1205    rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1206    *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1207    tail = rdp->nxttail[RCU_DONE_TAIL];
1208    for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1209        if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1210            rdp->nxttail[count] = &rdp->nxtlist;
1211    local_irq_restore(flags);
1212
1213    /* Invoke callbacks. */
1214    count = 0;
1215    while (list) {
1216        next = list->next;
1217        prefetch(next);
1218        debug_rcu_head_unqueue(list);
1219        __rcu_reclaim(list);
1220        list = next;
1221        if (++count >= rdp->blimit)
1222            break;
1223    }
1224
1225    local_irq_save(flags);
1226
1227    /* Update count, and requeue any remaining callbacks. */
1228    rdp->qlen -= count;
1229    rdp->n_cbs_invoked += count;
1230    if (list != NULL) {
1231        *tail = rdp->nxtlist;
1232        rdp->nxtlist = list;
1233        for (count = 0; count < RCU_NEXT_SIZE; count++)
1234            if (&rdp->nxtlist == rdp->nxttail[count])
1235                rdp->nxttail[count] = tail;
1236            else
1237                break;
1238    }
1239
1240    /* Reinstate batch limit if we have worked down the excess. */
1241    if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1242        rdp->blimit = blimit;
1243
1244    /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1245    if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1246        rdp->qlen_last_fqs_check = 0;
1247        rdp->n_force_qs_snap = rsp->n_force_qs;
1248    } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1249        rdp->qlen_last_fqs_check = rdp->qlen;
1250
1251    local_irq_restore(flags);
1252
1253    /* Re-raise the RCU softirq if there are callbacks remaining. */
1254    if (cpu_has_callbacks_ready_to_invoke(rdp))
1255        invoke_rcu_core();
1256}
1257
1258/*
1259 * Check to see if this CPU is in a non-context-switch quiescent state
1260 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1261 * Also schedule the RCU softirq handler.
1262 *
1263 * This function must be called with hardirqs disabled. It is normally
1264 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1265 * false, there is no point in invoking rcu_check_callbacks().
1266 */
1267void rcu_check_callbacks(int cpu, int user)
1268{
1269    if (user ||
1270        (idle_cpu(cpu) && rcu_scheduler_active &&
1271         !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1272
1273        /*
1274         * Get here if this CPU took its interrupt from user
1275         * mode or from the idle loop, and if this is not a
1276         * nested interrupt. In this case, the CPU is in
1277         * a quiescent state, so note it.
1278         *
1279         * No memory barrier is required here because both
1280         * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1281         * variables that other CPUs neither access nor modify,
1282         * at least not while the corresponding CPU is online.
1283         */
1284
1285        rcu_sched_qs(cpu);
1286        rcu_bh_qs(cpu);
1287
1288    } else if (!in_softirq()) {
1289
1290        /*
1291         * Get here if this CPU did not take its interrupt from
1292         * softirq, in other words, if it is not interrupting
1293         * a rcu_bh read-side critical section. This is an _bh
1294         * critical section, so note it.
1295         */
1296
1297        rcu_bh_qs(cpu);
1298    }
1299    rcu_preempt_check_callbacks(cpu);
1300    if (rcu_pending(cpu))
1301        invoke_rcu_core();
1302}
1303
1304#ifdef CONFIG_SMP
1305
1306/*
1307 * Scan the leaf rcu_node structures, processing dyntick state for any that
1308 * have not yet encountered a quiescent state, using the function specified.
1309 * Also initiate boosting for any threads blocked on the root rcu_node.
1310 *
1311 * The caller must have suppressed start of new grace periods.
1312 */
1313static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1314{
1315    unsigned long bit;
1316    int cpu;
1317    unsigned long flags;
1318    unsigned long mask;
1319    struct rcu_node *rnp;
1320
1321    rcu_for_each_leaf_node(rsp, rnp) {
1322        mask = 0;
1323        raw_spin_lock_irqsave(&rnp->lock, flags);
1324        if (!rcu_gp_in_progress(rsp)) {
1325            raw_spin_unlock_irqrestore(&rnp->lock, flags);
1326            return;
1327        }
1328        if (rnp->qsmask == 0) {
1329            rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1330            continue;
1331        }
1332        cpu = rnp->grplo;
1333        bit = 1;
1334        for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1335            if ((rnp->qsmask & bit) != 0 &&
1336                f(per_cpu_ptr(rsp->rda, cpu)))
1337                mask |= bit;
1338        }
1339        if (mask != 0) {
1340
1341            /* rcu_report_qs_rnp() releases rnp->lock. */
1342            rcu_report_qs_rnp(mask, rsp, rnp, flags);
1343            continue;
1344        }
1345        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1346    }
1347    rnp = rcu_get_root(rsp);
1348    if (rnp->qsmask == 0) {
1349        raw_spin_lock_irqsave(&rnp->lock, flags);
1350        rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1351    }
1352}
1353
1354/*
1355 * Force quiescent states on reluctant CPUs, and also detect which
1356 * CPUs are in dyntick-idle mode.
1357 */
1358static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1359{
1360    unsigned long flags;
1361    struct rcu_node *rnp = rcu_get_root(rsp);
1362
1363    if (!rcu_gp_in_progress(rsp))
1364        return; /* No grace period in progress, nothing to force. */
1365    if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1366        rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1367        return; /* Someone else is already on the job. */
1368    }
1369    if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1370        goto unlock_fqs_ret; /* no emergency and done recently. */
1371    rsp->n_force_qs++;
1372    raw_spin_lock(&rnp->lock); /* irqs already disabled */
1373    rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1374    if(!rcu_gp_in_progress(rsp)) {
1375        rsp->n_force_qs_ngp++;
1376        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1377        goto unlock_fqs_ret; /* no GP in progress, time updated. */
1378    }
1379    rsp->fqs_active = 1;
1380    switch (rsp->signaled) {
1381    case RCU_GP_IDLE:
1382    case RCU_GP_INIT:
1383
1384        break; /* grace period idle or initializing, ignore. */
1385
1386    case RCU_SAVE_DYNTICK:
1387        if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1388            break; /* So gcc recognizes the dead code. */
1389
1390        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1391
1392        /* Record dyntick-idle state. */
1393        force_qs_rnp(rsp, dyntick_save_progress_counter);
1394        raw_spin_lock(&rnp->lock); /* irqs already disabled */
1395        if (rcu_gp_in_progress(rsp))
1396            rsp->signaled = RCU_FORCE_QS;
1397        break;
1398
1399    case RCU_FORCE_QS:
1400
1401        /* Check dyntick-idle state, send IPI to laggarts. */
1402        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1403        force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1404
1405        /* Leave state in case more forcing is required. */
1406
1407        raw_spin_lock(&rnp->lock); /* irqs already disabled */
1408        break;
1409    }
1410    rsp->fqs_active = 0;
1411    if (rsp->fqs_need_gp) {
1412        raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1413        rsp->fqs_need_gp = 0;
1414        rcu_start_gp(rsp, flags); /* releases rnp->lock */
1415        return;
1416    }
1417    raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1418unlock_fqs_ret:
1419    raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1420}
1421
1422#else /* #ifdef CONFIG_SMP */
1423
1424static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1425{
1426    set_need_resched();
1427}
1428
1429#endif /* #else #ifdef CONFIG_SMP */
1430
1431/*
1432 * This does the RCU processing work from softirq context for the
1433 * specified rcu_state and rcu_data structures. This may be called
1434 * only from the CPU to whom the rdp belongs.
1435 */
1436static void
1437__rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1438{
1439    unsigned long flags;
1440
1441    WARN_ON_ONCE(rdp->beenonline == 0);
1442
1443    /*
1444     * If an RCU GP has gone long enough, go check for dyntick
1445     * idle CPUs and, if needed, send resched IPIs.
1446     */
1447    if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1448        force_quiescent_state(rsp, 1);
1449
1450    /*
1451     * Advance callbacks in response to end of earlier grace
1452     * period that some other CPU ended.
1453     */
1454    rcu_process_gp_end(rsp, rdp);
1455
1456    /* Update RCU state based on any recent quiescent states. */
1457    rcu_check_quiescent_state(rsp, rdp);
1458
1459    /* Does this CPU require a not-yet-started grace period? */
1460    if (cpu_needs_another_gp(rsp, rdp)) {
1461        raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1462        rcu_start_gp(rsp, flags); /* releases above lock */
1463    }
1464
1465    /* If there are callbacks ready, invoke them. */
1466    if (cpu_has_callbacks_ready_to_invoke(rdp))
1467        invoke_rcu_callbacks(rsp, rdp);
1468}
1469
1470/*
1471 * Do softirq processing for the current CPU.
1472 */
1473static void rcu_process_callbacks(struct softirq_action *unused)
1474{
1475    __rcu_process_callbacks(&rcu_sched_state,
1476                &__get_cpu_var(rcu_sched_data));
1477    __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1478    rcu_preempt_process_callbacks();
1479
1480    /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1481    rcu_needs_cpu_flush();
1482}
1483
1484/*
1485 * Wake up the current CPU's kthread. This replaces raise_softirq()
1486 * in earlier versions of RCU. Note that because we are running on
1487 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1488 * cannot disappear out from under us.
1489 */
1490static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1491{
1492    if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1493        return;
1494    if (likely(!rsp->boost)) {
1495        rcu_do_batch(rsp, rdp);
1496        return;
1497    }
1498    invoke_rcu_callbacks_kthread();
1499}
1500
1501static void invoke_rcu_core(void)
1502{
1503    raise_softirq(RCU_SOFTIRQ);
1504}
1505
1506static void
1507__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1508       struct rcu_state *rsp)
1509{
1510    unsigned long flags;
1511    struct rcu_data *rdp;
1512
1513    debug_rcu_head_queue(head);
1514    head->func = func;
1515    head->next = NULL;
1516
1517    smp_mb(); /* Ensure RCU update seen before callback registry. */
1518
1519    /*
1520     * Opportunistically note grace-period endings and beginnings.
1521     * Note that we might see a beginning right after we see an
1522     * end, but never vice versa, since this CPU has to pass through
1523     * a quiescent state betweentimes.
1524     */
1525    local_irq_save(flags);
1526    rdp = this_cpu_ptr(rsp->rda);
1527
1528    /* Add the callback to our list. */
1529    *rdp->nxttail[RCU_NEXT_TAIL] = head;
1530    rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1531    rdp->qlen++;
1532
1533    /* If interrupts were disabled, don't dive into RCU core. */
1534    if (irqs_disabled_flags(flags)) {
1535        local_irq_restore(flags);
1536        return;
1537    }
1538
1539    /*
1540     * Force the grace period if too many callbacks or too long waiting.
1541     * Enforce hysteresis, and don't invoke force_quiescent_state()
1542     * if some other CPU has recently done so. Also, don't bother
1543     * invoking force_quiescent_state() if the newly enqueued callback
1544     * is the only one waiting for a grace period to complete.
1545     */
1546    if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1547
1548        /* Are we ignoring a completed grace period? */
1549        rcu_process_gp_end(rsp, rdp);
1550        check_for_new_grace_period(rsp, rdp);
1551
1552        /* Start a new grace period if one not already started. */
1553        if (!rcu_gp_in_progress(rsp)) {
1554            unsigned long nestflag;
1555            struct rcu_node *rnp_root = rcu_get_root(rsp);
1556
1557            raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1558            rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1559        } else {
1560            /* Give the grace period a kick. */
1561            rdp->blimit = LONG_MAX;
1562            if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1563                *rdp->nxttail[RCU_DONE_TAIL] != head)
1564                force_quiescent_state(rsp, 0);
1565            rdp->n_force_qs_snap = rsp->n_force_qs;
1566            rdp->qlen_last_fqs_check = rdp->qlen;
1567        }
1568    } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1569        force_quiescent_state(rsp, 1);
1570    local_irq_restore(flags);
1571}
1572
1573/*
1574 * Queue an RCU-sched callback for invocation after a grace period.
1575 */
1576void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1577{
1578    __call_rcu(head, func, &rcu_sched_state);
1579}
1580EXPORT_SYMBOL_GPL(call_rcu_sched);
1581
1582/*
1583 * Queue an RCU for invocation after a quicker grace period.
1584 */
1585void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1586{
1587    __call_rcu(head, func, &rcu_bh_state);
1588}
1589EXPORT_SYMBOL_GPL(call_rcu_bh);
1590
1591/**
1592 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1593 *
1594 * Control will return to the caller some time after a full rcu-sched
1595 * grace period has elapsed, in other words after all currently executing
1596 * rcu-sched read-side critical sections have completed. These read-side
1597 * critical sections are delimited by rcu_read_lock_sched() and
1598 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1599 * local_irq_disable(), and so on may be used in place of
1600 * rcu_read_lock_sched().
1601 *
1602 * This means that all preempt_disable code sequences, including NMI and
1603 * hardware-interrupt handlers, in progress on entry will have completed
1604 * before this primitive returns. However, this does not guarantee that
1605 * softirq handlers will have completed, since in some kernels, these
1606 * handlers can run in process context, and can block.
1607 *
1608 * This primitive provides the guarantees made by the (now removed)
1609 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1610 * guarantees that rcu_read_lock() sections will have completed.
1611 * In "classic RCU", these two guarantees happen to be one and
1612 * the same, but can differ in realtime RCU implementations.
1613 */
1614void synchronize_sched(void)
1615{
1616    struct rcu_synchronize rcu;
1617
1618    if (rcu_blocking_is_gp())
1619        return;
1620
1621    init_rcu_head_on_stack(&rcu.head);
1622    init_completion(&rcu.completion);
1623    /* Will wake me after RCU finished. */
1624    call_rcu_sched(&rcu.head, wakeme_after_rcu);
1625    /* Wait for it. */
1626    wait_for_completion(&rcu.completion);
1627    destroy_rcu_head_on_stack(&rcu.head);
1628}
1629EXPORT_SYMBOL_GPL(synchronize_sched);
1630
1631/**
1632 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1633 *
1634 * Control will return to the caller some time after a full rcu_bh grace
1635 * period has elapsed, in other words after all currently executing rcu_bh
1636 * read-side critical sections have completed. RCU read-side critical
1637 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1638 * and may be nested.
1639 */
1640void synchronize_rcu_bh(void)
1641{
1642    struct rcu_synchronize rcu;
1643
1644    if (rcu_blocking_is_gp())
1645        return;
1646
1647    init_rcu_head_on_stack(&rcu.head);
1648    init_completion(&rcu.completion);
1649    /* Will wake me after RCU finished. */
1650    call_rcu_bh(&rcu.head, wakeme_after_rcu);
1651    /* Wait for it. */
1652    wait_for_completion(&rcu.completion);
1653    destroy_rcu_head_on_stack(&rcu.head);
1654}
1655EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1656
1657/*
1658 * Check to see if there is any immediate RCU-related work to be done
1659 * by the current CPU, for the specified type of RCU, returning 1 if so.
1660 * The checks are in order of increasing expense: checks that can be
1661 * carried out against CPU-local state are performed first. However,
1662 * we must check for CPU stalls first, else we might not get a chance.
1663 */
1664static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1665{
1666    struct rcu_node *rnp = rdp->mynode;
1667
1668    rdp->n_rcu_pending++;
1669
1670    /* Check for CPU stalls, if enabled. */
1671    check_cpu_stall(rsp, rdp);
1672
1673    /* Is the RCU core waiting for a quiescent state from this CPU? */
1674    if (rdp->qs_pending && !rdp->passed_quiesc) {
1675
1676        /*
1677         * If force_quiescent_state() coming soon and this CPU
1678         * needs a quiescent state, and this is either RCU-sched
1679         * or RCU-bh, force a local reschedule.
1680         */
1681        rdp->n_rp_qs_pending++;
1682        if (!rdp->preemptible &&
1683            ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1684                 jiffies))
1685            set_need_resched();
1686    } else if (rdp->qs_pending && rdp->passed_quiesc) {
1687        rdp->n_rp_report_qs++;
1688        return 1;
1689    }
1690
1691    /* Does this CPU have callbacks ready to invoke? */
1692    if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1693        rdp->n_rp_cb_ready++;
1694        return 1;
1695    }
1696
1697    /* Has RCU gone idle with this CPU needing another grace period? */
1698    if (cpu_needs_another_gp(rsp, rdp)) {
1699        rdp->n_rp_cpu_needs_gp++;
1700        return 1;
1701    }
1702
1703    /* Has another RCU grace period completed? */
1704    if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1705        rdp->n_rp_gp_completed++;
1706        return 1;
1707    }
1708
1709    /* Has a new RCU grace period started? */
1710    if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1711        rdp->n_rp_gp_started++;
1712        return 1;
1713    }
1714
1715    /* Has an RCU GP gone long enough to send resched IPIs &c? */
1716    if (rcu_gp_in_progress(rsp) &&
1717        ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1718        rdp->n_rp_need_fqs++;
1719        return 1;
1720    }
1721
1722    /* nothing to do */
1723    rdp->n_rp_need_nothing++;
1724    return 0;
1725}
1726
1727/*
1728 * Check to see if there is any immediate RCU-related work to be done
1729 * by the current CPU, returning 1 if so. This function is part of the
1730 * RCU implementation; it is -not- an exported member of the RCU API.
1731 */
1732static int rcu_pending(int cpu)
1733{
1734    return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1735           __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1736           rcu_preempt_pending(cpu);
1737}
1738
1739/*
1740 * Check to see if any future RCU-related work will need to be done
1741 * by the current CPU, even if none need be done immediately, returning
1742 * 1 if so.
1743 */
1744static int rcu_needs_cpu_quick_check(int cpu)
1745{
1746    /* RCU callbacks either ready or pending? */
1747    return per_cpu(rcu_sched_data, cpu).nxtlist ||
1748           per_cpu(rcu_bh_data, cpu).nxtlist ||
1749           rcu_preempt_needs_cpu(cpu);
1750}
1751
1752static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1753static atomic_t rcu_barrier_cpu_count;
1754static DEFINE_MUTEX(rcu_barrier_mutex);
1755static struct completion rcu_barrier_completion;
1756
1757static void rcu_barrier_callback(struct rcu_head *notused)
1758{
1759    if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1760        complete(&rcu_barrier_completion);
1761}
1762
1763/*
1764 * Called with preemption disabled, and from cross-cpu IRQ context.
1765 */
1766static void rcu_barrier_func(void *type)
1767{
1768    int cpu = smp_processor_id();
1769    struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1770    void (*call_rcu_func)(struct rcu_head *head,
1771                  void (*func)(struct rcu_head *head));
1772
1773    atomic_inc(&rcu_barrier_cpu_count);
1774    call_rcu_func = type;
1775    call_rcu_func(head, rcu_barrier_callback);
1776}
1777
1778/*
1779 * Orchestrate the specified type of RCU barrier, waiting for all
1780 * RCU callbacks of the specified type to complete.
1781 */
1782static void _rcu_barrier(struct rcu_state *rsp,
1783             void (*call_rcu_func)(struct rcu_head *head,
1784                           void (*func)(struct rcu_head *head)))
1785{
1786    BUG_ON(in_interrupt());
1787    /* Take mutex to serialize concurrent rcu_barrier() requests. */
1788    mutex_lock(&rcu_barrier_mutex);
1789    init_completion(&rcu_barrier_completion);
1790    /*
1791     * Initialize rcu_barrier_cpu_count to 1, then invoke
1792     * rcu_barrier_func() on each CPU, so that each CPU also has
1793     * incremented rcu_barrier_cpu_count. Only then is it safe to
1794     * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1795     * might complete its grace period before all of the other CPUs
1796     * did their increment, causing this function to return too
1797     * early. Note that on_each_cpu() disables irqs, which prevents
1798     * any CPUs from coming online or going offline until each online
1799     * CPU has queued its RCU-barrier callback.
1800     */
1801    atomic_set(&rcu_barrier_cpu_count, 1);
1802    on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1803    if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1804        complete(&rcu_barrier_completion);
1805    wait_for_completion(&rcu_barrier_completion);
1806    mutex_unlock(&rcu_barrier_mutex);
1807}
1808
1809/**
1810 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1811 */
1812void rcu_barrier_bh(void)
1813{
1814    _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1815}
1816EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1817
1818/**
1819 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1820 */
1821void rcu_barrier_sched(void)
1822{
1823    _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1824}
1825EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1826
1827/*
1828 * Do boot-time initialization of a CPU's per-CPU RCU data.
1829 */
1830static void __init
1831rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1832{
1833    unsigned long flags;
1834    int i;
1835    struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1836    struct rcu_node *rnp = rcu_get_root(rsp);
1837
1838    /* Set up local state, ensuring consistent view of global state. */
1839    raw_spin_lock_irqsave(&rnp->lock, flags);
1840    rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1841    rdp->nxtlist = NULL;
1842    for (i = 0; i < RCU_NEXT_SIZE; i++)
1843        rdp->nxttail[i] = &rdp->nxtlist;
1844    rdp->qlen = 0;
1845#ifdef CONFIG_NO_HZ
1846    rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1847#endif /* #ifdef CONFIG_NO_HZ */
1848    rdp->cpu = cpu;
1849    raw_spin_unlock_irqrestore(&rnp->lock, flags);
1850}
1851
1852/*
1853 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1854 * offline event can be happening at a given time. Note also that we
1855 * can accept some slop in the rsp->completed access due to the fact
1856 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1857 */
1858static void __cpuinit
1859rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
1860{
1861    unsigned long flags;
1862    unsigned long mask;
1863    struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1864    struct rcu_node *rnp = rcu_get_root(rsp);
1865
1866    /* Set up local state, ensuring consistent view of global state. */
1867    raw_spin_lock_irqsave(&rnp->lock, flags);
1868    rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1869    rdp->qs_pending = 1; /* so set up to respond to current GP. */
1870    rdp->beenonline = 1; /* We have now been online. */
1871    rdp->preemptible = preemptible;
1872    rdp->qlen_last_fqs_check = 0;
1873    rdp->n_force_qs_snap = rsp->n_force_qs;
1874    rdp->blimit = blimit;
1875    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1876
1877    /*
1878     * A new grace period might start here. If so, we won't be part
1879     * of it, but that is OK, as we are currently in a quiescent state.
1880     */
1881
1882    /* Exclude any attempts to start a new GP on large systems. */
1883    raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1884
1885    /* Add CPU to rcu_node bitmasks. */
1886    rnp = rdp->mynode;
1887    mask = rdp->grpmask;
1888    do {
1889        /* Exclude any attempts to start a new GP on small systems. */
1890        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1891        rnp->qsmaskinit |= mask;
1892        mask = rnp->grpmask;
1893        if (rnp == rdp->mynode) {
1894            rdp->gpnum = rnp->completed; /* if GP in progress... */
1895            rdp->completed = rnp->completed;
1896            rdp->passed_quiesc_completed = rnp->completed - 1;
1897        }
1898        raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1899        rnp = rnp->parent;
1900    } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1901
1902    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1903}
1904
1905static void __cpuinit rcu_prepare_cpu(int cpu)
1906{
1907    rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1908    rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1909    rcu_preempt_init_percpu_data(cpu);
1910}
1911
1912/*
1913 * Handle CPU online/offline notification events.
1914 */
1915static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1916                    unsigned long action, void *hcpu)
1917{
1918    long cpu = (long)hcpu;
1919    struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1920    struct rcu_node *rnp = rdp->mynode;
1921
1922    switch (action) {
1923    case CPU_UP_PREPARE:
1924    case CPU_UP_PREPARE_FROZEN:
1925        rcu_prepare_cpu(cpu);
1926        rcu_prepare_kthreads(cpu);
1927        break;
1928    case CPU_ONLINE:
1929    case CPU_DOWN_FAILED:
1930        rcu_node_kthread_setaffinity(rnp, -1);
1931        rcu_cpu_kthread_setrt(cpu, 1);
1932        break;
1933    case CPU_DOWN_PREPARE:
1934        rcu_node_kthread_setaffinity(rnp, cpu);
1935        rcu_cpu_kthread_setrt(cpu, 0);
1936        break;
1937    case CPU_DYING:
1938    case CPU_DYING_FROZEN:
1939        /*
1940         * The whole machine is "stopped" except this CPU, so we can
1941         * touch any data without introducing corruption. We send the
1942         * dying CPU's callbacks to an arbitrarily chosen online CPU.
1943         */
1944        rcu_send_cbs_to_online(&rcu_bh_state);
1945        rcu_send_cbs_to_online(&rcu_sched_state);
1946        rcu_preempt_send_cbs_to_online();
1947        break;
1948    case CPU_DEAD:
1949    case CPU_DEAD_FROZEN:
1950    case CPU_UP_CANCELED:
1951    case CPU_UP_CANCELED_FROZEN:
1952        rcu_offline_cpu(cpu);
1953        break;
1954    default:
1955        break;
1956    }
1957    return NOTIFY_OK;
1958}
1959
1960/*
1961 * This function is invoked towards the end of the scheduler's initialization
1962 * process. Before this is called, the idle task might contain
1963 * RCU read-side critical sections (during which time, this idle
1964 * task is booting the system). After this function is called, the
1965 * idle tasks are prohibited from containing RCU read-side critical
1966 * sections. This function also enables RCU lockdep checking.
1967 */
1968void rcu_scheduler_starting(void)
1969{
1970    WARN_ON(num_online_cpus() != 1);
1971    WARN_ON(nr_context_switches() > 0);
1972    rcu_scheduler_active = 1;
1973}
1974
1975/*
1976 * Compute the per-level fanout, either using the exact fanout specified
1977 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1978 */
1979#ifdef CONFIG_RCU_FANOUT_EXACT
1980static void __init rcu_init_levelspread(struct rcu_state *rsp)
1981{
1982    int i;
1983
1984    for (i = NUM_RCU_LVLS - 1; i > 0; i--)
1985        rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1986    rsp->levelspread[0] = RCU_FANOUT_LEAF;
1987}
1988#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1989static void __init rcu_init_levelspread(struct rcu_state *rsp)
1990{
1991    int ccur;
1992    int cprv;
1993    int i;
1994
1995    cprv = NR_CPUS;
1996    for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1997        ccur = rsp->levelcnt[i];
1998        rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1999        cprv = ccur;
2000    }
2001}
2002#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2003
2004/*
2005 * Helper function for rcu_init() that initializes one rcu_state structure.
2006 */
2007static void __init rcu_init_one(struct rcu_state *rsp,
2008        struct rcu_data __percpu *rda)
2009{
2010    static char *buf[] = { "rcu_node_level_0",
2011                   "rcu_node_level_1",
2012                   "rcu_node_level_2",
2013                   "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2014    int cpustride = 1;
2015    int i;
2016    int j;
2017    struct rcu_node *rnp;
2018
2019    BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2020
2021    /* Initialize the level-tracking arrays. */
2022
2023    for (i = 1; i < NUM_RCU_LVLS; i++)
2024        rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2025    rcu_init_levelspread(rsp);
2026
2027    /* Initialize the elements themselves, starting from the leaves. */
2028
2029    for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2030        cpustride *= rsp->levelspread[i];
2031        rnp = rsp->level[i];
2032        for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2033            raw_spin_lock_init(&rnp->lock);
2034            lockdep_set_class_and_name(&rnp->lock,
2035                           &rcu_node_class[i], buf[i]);
2036            rnp->gpnum = 0;
2037            rnp->qsmask = 0;
2038            rnp->qsmaskinit = 0;
2039            rnp->grplo = j * cpustride;
2040            rnp->grphi = (j + 1) * cpustride - 1;
2041            if (rnp->grphi >= NR_CPUS)
2042                rnp->grphi = NR_CPUS - 1;
2043            if (i == 0) {
2044                rnp->grpnum = 0;
2045                rnp->grpmask = 0;
2046                rnp->parent = NULL;
2047            } else {
2048                rnp->grpnum = j % rsp->levelspread[i - 1];
2049                rnp->grpmask = 1UL << rnp->grpnum;
2050                rnp->parent = rsp->level[i - 1] +
2051                          j / rsp->levelspread[i - 1];
2052            }
2053            rnp->level = i;
2054            INIT_LIST_HEAD(&rnp->blkd_tasks);
2055        }
2056    }
2057
2058    rsp->rda = rda;
2059    rnp = rsp->level[NUM_RCU_LVLS - 1];
2060    for_each_possible_cpu(i) {
2061        while (i > rnp->grphi)
2062            rnp++;
2063        per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2064        rcu_boot_init_percpu_data(i, rsp);
2065    }
2066}
2067
2068void __init rcu_init(void)
2069{
2070    int cpu;
2071
2072    rcu_bootup_announce();
2073    rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2074    rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2075    __rcu_init_preempt();
2076     open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2077
2078    /*
2079     * We don't need protection against CPU-hotplug here because
2080     * this is called early in boot, before either interrupts
2081     * or the scheduler are operational.
2082     */
2083    cpu_notifier(rcu_cpu_notify, 0);
2084    for_each_online_cpu(cpu)
2085        rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2086    check_cpu_stall_init();
2087}
2088
2089#include "rcutree_plugin.h"
2090

Archive Download this file



interactive