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/export.h>
42#include <linux/completion.h>
43#include <linux/moduleparam.h>
44#include <linux/percpu.h>
45#include <linux/notifier.h>
46#include <linux/cpu.h>
47#include <linux/mutex.h>
48#include <linux/time.h>
49#include <linux/kernel_stat.h>
50#include <linux/wait.h>
51#include <linux/kthread.h>
52#include <linux/prefetch.h>
53#include <linux/delay.h>
54#include <linux/stop_machine.h>
55
56#include "rcutree.h"
57#include <trace/events/rcu.h>
58
59#include "rcu.h"
60
61/* Data structures. */
62
63static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
64
65#define RCU_STATE_INITIALIZER(structname) { \
66    .level = { &structname##_state.node[0] }, \
67    .levelcnt = { \
68        NUM_RCU_LVL_0, /* root of hierarchy. */ \
69        NUM_RCU_LVL_1, \
70        NUM_RCU_LVL_2, \
71        NUM_RCU_LVL_3, \
72        NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
73    }, \
74    .fqs_state = RCU_GP_IDLE, \
75    .gpnum = -300, \
76    .completed = -300, \
77    .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
78    .orphan_nxttail = &structname##_state.orphan_nxtlist, \
79    .orphan_donetail = &structname##_state.orphan_donelist, \
80    .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
81    .n_force_qs = 0, \
82    .n_force_qs_ngp = 0, \
83    .name = #structname, \
84}
85
86struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched);
87DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
88
89struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh);
90DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
91
92static struct rcu_state *rcu_state;
93
94/*
95 * The rcu_scheduler_active variable transitions from zero to one just
96 * before the first task is spawned. So when this variable is zero, RCU
97 * can assume that there is but one task, allowing RCU to (for example)
98 * optimized synchronize_sched() to a simple barrier(). When this variable
99 * is one, RCU must actually do all the hard work required to detect real
100 * grace periods. This variable is also used to suppress boot-time false
101 * positives from lockdep-RCU error checking.
102 */
103int rcu_scheduler_active __read_mostly;
104EXPORT_SYMBOL_GPL(rcu_scheduler_active);
105
106/*
107 * The rcu_scheduler_fully_active variable transitions from zero to one
108 * during the early_initcall() processing, which is after the scheduler
109 * is capable of creating new tasks. So RCU processing (for example,
110 * creating tasks for RCU priority boosting) must be delayed until after
111 * rcu_scheduler_fully_active transitions from zero to one. We also
112 * currently delay invocation of any RCU callbacks until after this point.
113 *
114 * It might later prove better for people registering RCU callbacks during
115 * early boot to take responsibility for these callbacks, but one step at
116 * a time.
117 */
118static int rcu_scheduler_fully_active __read_mostly;
119
120#ifdef CONFIG_RCU_BOOST
121
122/*
123 * Control variables for per-CPU and per-rcu_node kthreads. These
124 * handle all flavors of RCU.
125 */
126static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
127DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
128DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
129DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
130DEFINE_PER_CPU(char, rcu_cpu_has_work);
131
132#endif /* #ifdef CONFIG_RCU_BOOST */
133
134static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
135static void invoke_rcu_core(void);
136static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
137
138/*
139 * Track the rcutorture test sequence number and the update version
140 * number within a given test. The rcutorture_testseq is incremented
141 * on every rcutorture module load and unload, so has an odd value
142 * when a test is running. The rcutorture_vernum is set to zero
143 * when rcutorture starts and is incremented on each rcutorture update.
144 * These variables enable correlating rcutorture output with the
145 * RCU tracing information.
146 */
147unsigned long rcutorture_testseq;
148unsigned long rcutorture_vernum;
149
150/* State information for rcu_barrier() and friends. */
151
152static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
153static atomic_t rcu_barrier_cpu_count;
154static DEFINE_MUTEX(rcu_barrier_mutex);
155static struct completion rcu_barrier_completion;
156
157/*
158 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
159 * permit this function to be invoked without holding the root rcu_node
160 * structure's ->lock, but of course results can be subject to change.
161 */
162static int rcu_gp_in_progress(struct rcu_state *rsp)
163{
164    return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
165}
166
167/*
168 * Note a quiescent state. Because we do not need to know
169 * how many quiescent states passed, just if there was at least
170 * one since the start of the grace period, this just sets a flag.
171 * The caller must have disabled preemption.
172 */
173void rcu_sched_qs(int cpu)
174{
175    struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
176
177    rdp->passed_quiesce_gpnum = rdp->gpnum;
178    barrier();
179    if (rdp->passed_quiesce == 0)
180        trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
181    rdp->passed_quiesce = 1;
182}
183
184void rcu_bh_qs(int cpu)
185{
186    struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
187
188    rdp->passed_quiesce_gpnum = rdp->gpnum;
189    barrier();
190    if (rdp->passed_quiesce == 0)
191        trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
192    rdp->passed_quiesce = 1;
193}
194
195/*
196 * Note a context switch. This is a quiescent state for RCU-sched,
197 * and requires special handling for preemptible RCU.
198 * The caller must have disabled preemption.
199 */
200void rcu_note_context_switch(int cpu)
201{
202    trace_rcu_utilization("Start context switch");
203    rcu_sched_qs(cpu);
204    rcu_preempt_note_context_switch(cpu);
205    trace_rcu_utilization("End context switch");
206}
207EXPORT_SYMBOL_GPL(rcu_note_context_switch);
208
209DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
210    .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
211    .dynticks = ATOMIC_INIT(1),
212};
213
214static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
215static int qhimark = 10000; /* If this many pending, ignore blimit. */
216static int qlowmark = 100; /* Once only this many pending, use blimit. */
217
218module_param(blimit, int, 0);
219module_param(qhimark, int, 0);
220module_param(qlowmark, int, 0);
221
222int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
223int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
224
225module_param(rcu_cpu_stall_suppress, int, 0644);
226module_param(rcu_cpu_stall_timeout, int, 0644);
227
228static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
229static int rcu_pending(int cpu);
230
231/*
232 * Return the number of RCU-sched batches processed thus far for debug & stats.
233 */
234long rcu_batches_completed_sched(void)
235{
236    return rcu_sched_state.completed;
237}
238EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
239
240/*
241 * Return the number of RCU BH batches processed thus far for debug & stats.
242 */
243long rcu_batches_completed_bh(void)
244{
245    return rcu_bh_state.completed;
246}
247EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
248
249/*
250 * Force a quiescent state for RCU BH.
251 */
252void rcu_bh_force_quiescent_state(void)
253{
254    force_quiescent_state(&rcu_bh_state, 0);
255}
256EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
257
258/*
259 * Record the number of times rcutorture tests have been initiated and
260 * terminated. This information allows the debugfs tracing stats to be
261 * correlated to the rcutorture messages, even when the rcutorture module
262 * is being repeatedly loaded and unloaded. In other words, we cannot
263 * store this state in rcutorture itself.
264 */
265void rcutorture_record_test_transition(void)
266{
267    rcutorture_testseq++;
268    rcutorture_vernum = 0;
269}
270EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
271
272/*
273 * Record the number of writer passes through the current rcutorture test.
274 * This is also used to correlate debugfs tracing stats with the rcutorture
275 * messages.
276 */
277void rcutorture_record_progress(unsigned long vernum)
278{
279    rcutorture_vernum++;
280}
281EXPORT_SYMBOL_GPL(rcutorture_record_progress);
282
283/*
284 * Force a quiescent state for RCU-sched.
285 */
286void rcu_sched_force_quiescent_state(void)
287{
288    force_quiescent_state(&rcu_sched_state, 0);
289}
290EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
291
292/*
293 * Does the CPU have callbacks ready to be invoked?
294 */
295static int
296cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
297{
298    return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
299}
300
301/*
302 * Does the current CPU require a yet-as-unscheduled grace period?
303 */
304static int
305cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
306{
307    return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
308}
309
310/*
311 * Return the root node of the specified rcu_state structure.
312 */
313static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
314{
315    return &rsp->node[0];
316}
317
318/*
319 * If the specified CPU is offline, tell the caller that it is in
320 * a quiescent state. Otherwise, whack it with a reschedule IPI.
321 * Grace periods can end up waiting on an offline CPU when that
322 * CPU is in the process of coming online -- it will be added to the
323 * rcu_node bitmasks before it actually makes it online. The same thing
324 * can happen while a CPU is in the process of coming online. Because this
325 * race is quite rare, we check for it after detecting that the grace
326 * period has been delayed rather than checking each and every CPU
327 * each and every time we start a new grace period.
328 */
329static int rcu_implicit_offline_qs(struct rcu_data *rdp)
330{
331    /*
332     * If the CPU is offline for more than a jiffy, it is in a quiescent
333     * state. We can trust its state not to change because interrupts
334     * are disabled. The reason for the jiffy's worth of slack is to
335     * handle CPUs initializing on the way up and finding their way
336     * to the idle loop on the way down.
337     */
338    if (cpu_is_offline(rdp->cpu) &&
339        ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
340        trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
341        rdp->offline_fqs++;
342        return 1;
343    }
344    return 0;
345}
346
347/*
348 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
349 *
350 * If the new value of the ->dynticks_nesting counter now is zero,
351 * we really have entered idle, and must do the appropriate accounting.
352 * The caller must have disabled interrupts.
353 */
354static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
355{
356    trace_rcu_dyntick("Start", oldval, 0);
357    if (!is_idle_task(current)) {
358        struct task_struct *idle = idle_task(smp_processor_id());
359
360        trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
361        ftrace_dump(DUMP_ALL);
362        WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
363              current->pid, current->comm,
364              idle->pid, idle->comm); /* must be idle task! */
365    }
366    rcu_prepare_for_idle(smp_processor_id());
367    /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
368    smp_mb__before_atomic_inc(); /* See above. */
369    atomic_inc(&rdtp->dynticks);
370    smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
371    WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
372
373    /*
374     * The idle task is not permitted to enter the idle loop while
375     * in an RCU read-side critical section.
376     */
377    rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
378               "Illegal idle entry in RCU read-side critical section.");
379    rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
380               "Illegal idle entry in RCU-bh read-side critical section.");
381    rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
382               "Illegal idle entry in RCU-sched read-side critical section.");
383}
384
385/**
386 * rcu_idle_enter - inform RCU that current CPU is entering idle
387 *
388 * Enter idle mode, in other words, -leave- the mode in which RCU
389 * read-side critical sections can occur. (Though RCU read-side
390 * critical sections can occur in irq handlers in idle, a possibility
391 * handled by irq_enter() and irq_exit().)
392 *
393 * We crowbar the ->dynticks_nesting field to zero to allow for
394 * the possibility of usermode upcalls having messed up our count
395 * of interrupt nesting level during the prior busy period.
396 */
397void rcu_idle_enter(void)
398{
399    unsigned long flags;
400    long long oldval;
401    struct rcu_dynticks *rdtp;
402
403    local_irq_save(flags);
404    rdtp = &__get_cpu_var(rcu_dynticks);
405    oldval = rdtp->dynticks_nesting;
406    WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
407    if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
408        rdtp->dynticks_nesting = 0;
409    else
410        rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
411    rcu_idle_enter_common(rdtp, oldval);
412    local_irq_restore(flags);
413}
414EXPORT_SYMBOL_GPL(rcu_idle_enter);
415
416/**
417 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
418 *
419 * Exit from an interrupt handler, which might possibly result in entering
420 * idle mode, in other words, leaving the mode in which read-side critical
421 * sections can occur.
422 *
423 * This code assumes that the idle loop never does anything that might
424 * result in unbalanced calls to irq_enter() and irq_exit(). If your
425 * architecture violates this assumption, RCU will give you what you
426 * deserve, good and hard. But very infrequently and irreproducibly.
427 *
428 * Use things like work queues to work around this limitation.
429 *
430 * You have been warned.
431 */
432void rcu_irq_exit(void)
433{
434    unsigned long flags;
435    long long oldval;
436    struct rcu_dynticks *rdtp;
437
438    local_irq_save(flags);
439    rdtp = &__get_cpu_var(rcu_dynticks);
440    oldval = rdtp->dynticks_nesting;
441    rdtp->dynticks_nesting--;
442    WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
443    if (rdtp->dynticks_nesting)
444        trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
445    else
446        rcu_idle_enter_common(rdtp, oldval);
447    local_irq_restore(flags);
448}
449
450/*
451 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
452 *
453 * If the new value of the ->dynticks_nesting counter was previously zero,
454 * we really have exited idle, and must do the appropriate accounting.
455 * The caller must have disabled interrupts.
456 */
457static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
458{
459    smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
460    atomic_inc(&rdtp->dynticks);
461    /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
462    smp_mb__after_atomic_inc(); /* See above. */
463    WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
464    rcu_cleanup_after_idle(smp_processor_id());
465    trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
466    if (!is_idle_task(current)) {
467        struct task_struct *idle = idle_task(smp_processor_id());
468
469        trace_rcu_dyntick("Error on exit: not idle task",
470                  oldval, rdtp->dynticks_nesting);
471        ftrace_dump(DUMP_ALL);
472        WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
473              current->pid, current->comm,
474              idle->pid, idle->comm); /* must be idle task! */
475    }
476}
477
478/**
479 * rcu_idle_exit - inform RCU that current CPU is leaving idle
480 *
481 * Exit idle mode, in other words, -enter- the mode in which RCU
482 * read-side critical sections can occur.
483 *
484 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
485 * allow for the possibility of usermode upcalls messing up our count
486 * of interrupt nesting level during the busy period that is just
487 * now starting.
488 */
489void rcu_idle_exit(void)
490{
491    unsigned long flags;
492    struct rcu_dynticks *rdtp;
493    long long oldval;
494
495    local_irq_save(flags);
496    rdtp = &__get_cpu_var(rcu_dynticks);
497    oldval = rdtp->dynticks_nesting;
498    WARN_ON_ONCE(oldval < 0);
499    if (oldval & DYNTICK_TASK_NEST_MASK)
500        rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
501    else
502        rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
503    rcu_idle_exit_common(rdtp, oldval);
504    local_irq_restore(flags);
505}
506EXPORT_SYMBOL_GPL(rcu_idle_exit);
507
508/**
509 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
510 *
511 * Enter an interrupt handler, which might possibly result in exiting
512 * idle mode, in other words, entering the mode in which read-side critical
513 * sections can occur.
514 *
515 * Note that the Linux kernel is fully capable of entering an interrupt
516 * handler that it never exits, for example when doing upcalls to
517 * user mode! This code assumes that the idle loop never does upcalls to
518 * user mode. If your architecture does do upcalls from the idle loop (or
519 * does anything else that results in unbalanced calls to the irq_enter()
520 * and irq_exit() functions), RCU will give you what you deserve, good
521 * and hard. But very infrequently and irreproducibly.
522 *
523 * Use things like work queues to work around this limitation.
524 *
525 * You have been warned.
526 */
527void rcu_irq_enter(void)
528{
529    unsigned long flags;
530    struct rcu_dynticks *rdtp;
531    long long oldval;
532
533    local_irq_save(flags);
534    rdtp = &__get_cpu_var(rcu_dynticks);
535    oldval = rdtp->dynticks_nesting;
536    rdtp->dynticks_nesting++;
537    WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
538    if (oldval)
539        trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
540    else
541        rcu_idle_exit_common(rdtp, oldval);
542    local_irq_restore(flags);
543}
544
545/**
546 * rcu_nmi_enter - inform RCU of entry to NMI context
547 *
548 * If the CPU was idle with dynamic ticks active, and there is no
549 * irq handler running, this updates rdtp->dynticks_nmi to let the
550 * RCU grace-period handling know that the CPU is active.
551 */
552void rcu_nmi_enter(void)
553{
554    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
555
556    if (rdtp->dynticks_nmi_nesting == 0 &&
557        (atomic_read(&rdtp->dynticks) & 0x1))
558        return;
559    rdtp->dynticks_nmi_nesting++;
560    smp_mb__before_atomic_inc(); /* Force delay from prior write. */
561    atomic_inc(&rdtp->dynticks);
562    /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
563    smp_mb__after_atomic_inc(); /* See above. */
564    WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
565}
566
567/**
568 * rcu_nmi_exit - inform RCU of exit from NMI context
569 *
570 * If the CPU was idle with dynamic ticks active, and there is no
571 * irq handler running, this updates rdtp->dynticks_nmi to let the
572 * RCU grace-period handling know that the CPU is no longer active.
573 */
574void rcu_nmi_exit(void)
575{
576    struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
577
578    if (rdtp->dynticks_nmi_nesting == 0 ||
579        --rdtp->dynticks_nmi_nesting != 0)
580        return;
581    /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
582    smp_mb__before_atomic_inc(); /* See above. */
583    atomic_inc(&rdtp->dynticks);
584    smp_mb__after_atomic_inc(); /* Force delay to next write. */
585    WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
586}
587
588#ifdef CONFIG_PROVE_RCU
589
590/**
591 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
592 *
593 * If the current CPU is in its idle loop and is neither in an interrupt
594 * or NMI handler, return true.
595 */
596int rcu_is_cpu_idle(void)
597{
598    int ret;
599
600    preempt_disable();
601    ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
602    preempt_enable();
603    return ret;
604}
605EXPORT_SYMBOL(rcu_is_cpu_idle);
606
607#ifdef CONFIG_HOTPLUG_CPU
608
609/*
610 * Is the current CPU online? Disable preemption to avoid false positives
611 * that could otherwise happen due to the current CPU number being sampled,
612 * this task being preempted, its old CPU being taken offline, resuming
613 * on some other CPU, then determining that its old CPU is now offline.
614 * It is OK to use RCU on an offline processor during initial boot, hence
615 * the check for rcu_scheduler_fully_active. Note also that it is OK
616 * for a CPU coming online to use RCU for one jiffy prior to marking itself
617 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
618 * offline to continue to use RCU for one jiffy after marking itself
619 * offline in the cpu_online_mask. This leniency is necessary given the
620 * non-atomic nature of the online and offline processing, for example,
621 * the fact that a CPU enters the scheduler after completing the CPU_DYING
622 * notifiers.
623 *
624 * This is also why RCU internally marks CPUs online during the
625 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
626 *
627 * Disable checking if in an NMI handler because we cannot safely report
628 * errors from NMI handlers anyway.
629 */
630bool rcu_lockdep_current_cpu_online(void)
631{
632    struct rcu_data *rdp;
633    struct rcu_node *rnp;
634    bool ret;
635
636    if (in_nmi())
637        return 1;
638    preempt_disable();
639    rdp = &__get_cpu_var(rcu_sched_data);
640    rnp = rdp->mynode;
641    ret = (rdp->grpmask & rnp->qsmaskinit) ||
642          !rcu_scheduler_fully_active;
643    preempt_enable();
644    return ret;
645}
646EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
647
648#endif /* #ifdef CONFIG_HOTPLUG_CPU */
649
650#endif /* #ifdef CONFIG_PROVE_RCU */
651
652/**
653 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
654 *
655 * If the current CPU is idle or running at a first-level (not nested)
656 * interrupt from idle, return true. The caller must have at least
657 * disabled preemption.
658 */
659int rcu_is_cpu_rrupt_from_idle(void)
660{
661    return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
662}
663
664/*
665 * Snapshot the specified CPU's dynticks counter so that we can later
666 * credit them with an implicit quiescent state. Return 1 if this CPU
667 * is in dynticks idle mode, which is an extended quiescent state.
668 */
669static int dyntick_save_progress_counter(struct rcu_data *rdp)
670{
671    rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
672    return (rdp->dynticks_snap & 0x1) == 0;
673}
674
675/*
676 * Return true if the specified CPU has passed through a quiescent
677 * state by virtue of being in or having passed through an dynticks
678 * idle state since the last call to dyntick_save_progress_counter()
679 * for this same CPU.
680 */
681static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
682{
683    unsigned int curr;
684    unsigned int snap;
685
686    curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
687    snap = (unsigned int)rdp->dynticks_snap;
688
689    /*
690     * If the CPU passed through or entered a dynticks idle phase with
691     * no active irq/NMI handlers, then we can safely pretend that the CPU
692     * already acknowledged the request to pass through a quiescent
693     * state. Either way, that CPU cannot possibly be in an RCU
694     * read-side critical section that started before the beginning
695     * of the current RCU grace period.
696     */
697    if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
698        trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
699        rdp->dynticks_fqs++;
700        return 1;
701    }
702
703    /* Go check for the CPU being offline. */
704    return rcu_implicit_offline_qs(rdp);
705}
706
707static int jiffies_till_stall_check(void)
708{
709    int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
710
711    /*
712     * Limit check must be consistent with the Kconfig limits
713     * for CONFIG_RCU_CPU_STALL_TIMEOUT.
714     */
715    if (till_stall_check < 3) {
716        ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
717        till_stall_check = 3;
718    } else if (till_stall_check > 300) {
719        ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
720        till_stall_check = 300;
721    }
722    return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
723}
724
725static void record_gp_stall_check_time(struct rcu_state *rsp)
726{
727    rsp->gp_start = jiffies;
728    rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
729}
730
731static void print_other_cpu_stall(struct rcu_state *rsp)
732{
733    int cpu;
734    long delta;
735    unsigned long flags;
736    int ndetected;
737    struct rcu_node *rnp = rcu_get_root(rsp);
738
739    /* Only let one CPU complain about others per time interval. */
740
741    raw_spin_lock_irqsave(&rnp->lock, flags);
742    delta = jiffies - rsp->jiffies_stall;
743    if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
744        raw_spin_unlock_irqrestore(&rnp->lock, flags);
745        return;
746    }
747    rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
748    raw_spin_unlock_irqrestore(&rnp->lock, flags);
749
750    /*
751     * OK, time to rat on our buddy...
752     * See Documentation/RCU/stallwarn.txt for info on how to debug
753     * RCU CPU stall warnings.
754     */
755    printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
756           rsp->name);
757    print_cpu_stall_info_begin();
758    rcu_for_each_leaf_node(rsp, rnp) {
759        raw_spin_lock_irqsave(&rnp->lock, flags);
760        ndetected += rcu_print_task_stall(rnp);
761        raw_spin_unlock_irqrestore(&rnp->lock, flags);
762        if (rnp->qsmask == 0)
763            continue;
764        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
765            if (rnp->qsmask & (1UL << cpu)) {
766                print_cpu_stall_info(rsp, rnp->grplo + cpu);
767                ndetected++;
768            }
769    }
770
771    /*
772     * Now rat on any tasks that got kicked up to the root rcu_node
773     * due to CPU offlining.
774     */
775    rnp = rcu_get_root(rsp);
776    raw_spin_lock_irqsave(&rnp->lock, flags);
777    ndetected = rcu_print_task_stall(rnp);
778    raw_spin_unlock_irqrestore(&rnp->lock, flags);
779
780    print_cpu_stall_info_end();
781    printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
782           smp_processor_id(), (long)(jiffies - rsp->gp_start));
783    if (ndetected == 0)
784        printk(KERN_ERR "INFO: Stall ended before state dump start\n");
785    else if (!trigger_all_cpu_backtrace())
786        dump_stack();
787
788    /* If so configured, complain about tasks blocking the grace period. */
789
790    rcu_print_detail_task_stall(rsp);
791
792    force_quiescent_state(rsp, 0); /* Kick them all. */
793}
794
795static void print_cpu_stall(struct rcu_state *rsp)
796{
797    unsigned long flags;
798    struct rcu_node *rnp = rcu_get_root(rsp);
799
800    /*
801     * OK, time to rat on ourselves...
802     * See Documentation/RCU/stallwarn.txt for info on how to debug
803     * RCU CPU stall warnings.
804     */
805    printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
806    print_cpu_stall_info_begin();
807    print_cpu_stall_info(rsp, smp_processor_id());
808    print_cpu_stall_info_end();
809    printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
810    if (!trigger_all_cpu_backtrace())
811        dump_stack();
812
813    raw_spin_lock_irqsave(&rnp->lock, flags);
814    if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
815        rsp->jiffies_stall = jiffies +
816                     3 * jiffies_till_stall_check() + 3;
817    raw_spin_unlock_irqrestore(&rnp->lock, flags);
818
819    set_need_resched(); /* kick ourselves to get things going. */
820}
821
822static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
823{
824    unsigned long j;
825    unsigned long js;
826    struct rcu_node *rnp;
827
828    if (rcu_cpu_stall_suppress)
829        return;
830    j = ACCESS_ONCE(jiffies);
831    js = ACCESS_ONCE(rsp->jiffies_stall);
832    rnp = rdp->mynode;
833    if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
834
835        /* We haven't checked in, so go dump stack. */
836        print_cpu_stall(rsp);
837
838    } else if (rcu_gp_in_progress(rsp) &&
839           ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
840
841        /* They had a few time units to dump stack, so complain. */
842        print_other_cpu_stall(rsp);
843    }
844}
845
846static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
847{
848    rcu_cpu_stall_suppress = 1;
849    return NOTIFY_DONE;
850}
851
852/**
853 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
854 *
855 * Set the stall-warning timeout way off into the future, thus preventing
856 * any RCU CPU stall-warning messages from appearing in the current set of
857 * RCU grace periods.
858 *
859 * The caller must disable hard irqs.
860 */
861void rcu_cpu_stall_reset(void)
862{
863    rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
864    rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
865    rcu_preempt_stall_reset();
866}
867
868static struct notifier_block rcu_panic_block = {
869    .notifier_call = rcu_panic,
870};
871
872static void __init check_cpu_stall_init(void)
873{
874    atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
875}
876
877/*
878 * Update CPU-local rcu_data state to record the newly noticed grace period.
879 * This is used both when we started the grace period and when we notice
880 * that someone else started the grace period. The caller must hold the
881 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
882 * and must have irqs disabled.
883 */
884static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
885{
886    if (rdp->gpnum != rnp->gpnum) {
887        /*
888         * If the current grace period is waiting for this CPU,
889         * set up to detect a quiescent state, otherwise don't
890         * go looking for one.
891         */
892        rdp->gpnum = rnp->gpnum;
893        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
894        if (rnp->qsmask & rdp->grpmask) {
895            rdp->qs_pending = 1;
896            rdp->passed_quiesce = 0;
897        } else
898            rdp->qs_pending = 0;
899        zero_cpu_stall_ticks(rdp);
900    }
901}
902
903static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
904{
905    unsigned long flags;
906    struct rcu_node *rnp;
907
908    local_irq_save(flags);
909    rnp = rdp->mynode;
910    if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
911        !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
912        local_irq_restore(flags);
913        return;
914    }
915    __note_new_gpnum(rsp, rnp, rdp);
916    raw_spin_unlock_irqrestore(&rnp->lock, flags);
917}
918
919/*
920 * Did someone else start a new RCU grace period start since we last
921 * checked? Update local state appropriately if so. Must be called
922 * on the CPU corresponding to rdp.
923 */
924static int
925check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
926{
927    unsigned long flags;
928    int ret = 0;
929
930    local_irq_save(flags);
931    if (rdp->gpnum != rsp->gpnum) {
932        note_new_gpnum(rsp, rdp);
933        ret = 1;
934    }
935    local_irq_restore(flags);
936    return ret;
937}
938
939/*
940 * Advance this CPU's callbacks, but only if the current grace period
941 * has ended. This may be called only from the CPU to whom the rdp
942 * belongs. In addition, the corresponding leaf rcu_node structure's
943 * ->lock must be held by the caller, with irqs disabled.
944 */
945static void
946__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
947{
948    /* Did another grace period end? */
949    if (rdp->completed != rnp->completed) {
950
951        /* Advance callbacks. No harm if list empty. */
952        rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
953        rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
954        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
955
956        /* Remember that we saw this grace-period completion. */
957        rdp->completed = rnp->completed;
958        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
959
960        /*
961         * If we were in an extended quiescent state, we may have
962         * missed some grace periods that others CPUs handled on
963         * our behalf. Catch up with this state to avoid noting
964         * spurious new grace periods. If another grace period
965         * has started, then rnp->gpnum will have advanced, so
966         * we will detect this later on.
967         */
968        if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
969            rdp->gpnum = rdp->completed;
970
971        /*
972         * If RCU does not need a quiescent state from this CPU,
973         * then make sure that this CPU doesn't go looking for one.
974         */
975        if ((rnp->qsmask & rdp->grpmask) == 0)
976            rdp->qs_pending = 0;
977    }
978}
979
980/*
981 * Advance this CPU's callbacks, but only if the current grace period
982 * has ended. This may be called only from the CPU to whom the rdp
983 * belongs.
984 */
985static void
986rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
987{
988    unsigned long flags;
989    struct rcu_node *rnp;
990
991    local_irq_save(flags);
992    rnp = rdp->mynode;
993    if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
994        !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
995        local_irq_restore(flags);
996        return;
997    }
998    __rcu_process_gp_end(rsp, rnp, rdp);
999    raw_spin_unlock_irqrestore(&rnp->lock, flags);
1000}
1001
1002/*
1003 * Do per-CPU grace-period initialization for running CPU. The caller
1004 * must hold the lock of the leaf rcu_node structure corresponding to
1005 * this CPU.
1006 */
1007static void
1008rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1009{
1010    /* Prior grace period ended, so advance callbacks for current CPU. */
1011    __rcu_process_gp_end(rsp, rnp, rdp);
1012
1013    /*
1014     * Because this CPU just now started the new grace period, we know
1015     * that all of its callbacks will be covered by this upcoming grace
1016     * period, even the ones that were registered arbitrarily recently.
1017     * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1018     *
1019     * Other CPUs cannot be sure exactly when the grace period started.
1020     * Therefore, their recently registered callbacks must pass through
1021     * an additional RCU_NEXT_READY stage, so that they will be handled
1022     * by the next RCU grace period.
1023     */
1024    rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1025    rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1026
1027    /* Set state so that this CPU will detect the next quiescent state. */
1028    __note_new_gpnum(rsp, rnp, rdp);
1029}
1030
1031/*
1032 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1033 * in preparation for detecting the next grace period. The caller must hold
1034 * the root node's ->lock, which is released before return. Hard irqs must
1035 * be disabled.
1036 *
1037 * Note that it is legal for a dying CPU (which is marked as offline) to
1038 * invoke this function. This can happen when the dying CPU reports its
1039 * quiescent state.
1040 */
1041static void
1042rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1043    __releases(rcu_get_root(rsp)->lock)
1044{
1045    struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1046    struct rcu_node *rnp = rcu_get_root(rsp);
1047
1048    if (!rcu_scheduler_fully_active ||
1049        !cpu_needs_another_gp(rsp, rdp)) {
1050        /*
1051         * Either the scheduler hasn't yet spawned the first
1052         * non-idle task or this CPU does not need another
1053         * grace period. Either way, don't start a new grace
1054         * period.
1055         */
1056        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1057        return;
1058    }
1059
1060    if (rsp->fqs_active) {
1061        /*
1062         * This CPU needs a grace period, but force_quiescent_state()
1063         * is running. Tell it to start one on this CPU's behalf.
1064         */
1065        rsp->fqs_need_gp = 1;
1066        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1067        return;
1068    }
1069
1070    /* Advance to a new grace period and initialize state. */
1071    rsp->gpnum++;
1072    trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1073    WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1074    rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
1075    rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1076    record_gp_stall_check_time(rsp);
1077    raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
1078
1079    /* Exclude any concurrent CPU-hotplug operations. */
1080    raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1081
1082    /*
1083     * Set the quiescent-state-needed bits in all the rcu_node
1084     * structures for all currently online CPUs in breadth-first
1085     * order, starting from the root rcu_node structure. This
1086     * operation relies on the layout of the hierarchy within the
1087     * rsp->node[] array. Note that other CPUs will access only
1088     * the leaves of the hierarchy, which still indicate that no
1089     * grace period is in progress, at least until the corresponding
1090     * leaf node has been initialized. In addition, we have excluded
1091     * CPU-hotplug operations.
1092     *
1093     * Note that the grace period cannot complete until we finish
1094     * the initialization process, as there will be at least one
1095     * qsmask bit set in the root node until that time, namely the
1096     * one corresponding to this CPU, due to the fact that we have
1097     * irqs disabled.
1098     */
1099    rcu_for_each_node_breadth_first(rsp, rnp) {
1100        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1101        rcu_preempt_check_blocked_tasks(rnp);
1102        rnp->qsmask = rnp->qsmaskinit;
1103        rnp->gpnum = rsp->gpnum;
1104        rnp->completed = rsp->completed;
1105        if (rnp == rdp->mynode)
1106            rcu_start_gp_per_cpu(rsp, rnp, rdp);
1107        rcu_preempt_boost_start_gp(rnp);
1108        trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1109                        rnp->level, rnp->grplo,
1110                        rnp->grphi, rnp->qsmask);
1111        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1112    }
1113
1114    rnp = rcu_get_root(rsp);
1115    raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1116    rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1117    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1118    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1119}
1120
1121/*
1122 * Report a full set of quiescent states to the specified rcu_state
1123 * data structure. This involves cleaning up after the prior grace
1124 * period and letting rcu_start_gp() start up the next grace period
1125 * if one is needed. Note that the caller must hold rnp->lock, as
1126 * required by rcu_start_gp(), which will release it.
1127 */
1128static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1129    __releases(rcu_get_root(rsp)->lock)
1130{
1131    unsigned long gp_duration;
1132    struct rcu_node *rnp = rcu_get_root(rsp);
1133    struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1134
1135    WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1136
1137    /*
1138     * Ensure that all grace-period and pre-grace-period activity
1139     * is seen before the assignment to rsp->completed.
1140     */
1141    smp_mb(); /* See above block comment. */
1142    gp_duration = jiffies - rsp->gp_start;
1143    if (gp_duration > rsp->gp_max)
1144        rsp->gp_max = gp_duration;
1145
1146    /*
1147     * We know the grace period is complete, but to everyone else
1148     * it appears to still be ongoing. But it is also the case
1149     * that to everyone else it looks like there is nothing that
1150     * they can do to advance the grace period. It is therefore
1151     * safe for us to drop the lock in order to mark the grace
1152     * period as completed in all of the rcu_node structures.
1153     *
1154     * But if this CPU needs another grace period, it will take
1155     * care of this while initializing the next grace period.
1156     * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1157     * because the callbacks have not yet been advanced: Those
1158     * callbacks are waiting on the grace period that just now
1159     * completed.
1160     */
1161    if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1162        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1163
1164        /*
1165         * Propagate new ->completed value to rcu_node structures
1166         * so that other CPUs don't have to wait until the start
1167         * of the next grace period to process their callbacks.
1168         */
1169        rcu_for_each_node_breadth_first(rsp, rnp) {
1170            raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1171            rnp->completed = rsp->gpnum;
1172            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1173        }
1174        rnp = rcu_get_root(rsp);
1175        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1176    }
1177
1178    rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1179    trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1180    rsp->fqs_state = RCU_GP_IDLE;
1181    rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1182}
1183
1184/*
1185 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1186 * Allows quiescent states for a group of CPUs to be reported at one go
1187 * to the specified rcu_node structure, though all the CPUs in the group
1188 * must be represented by the same rcu_node structure (which need not be
1189 * a leaf rcu_node structure, though it often will be). That structure's
1190 * lock must be held upon entry, and it is released before return.
1191 */
1192static void
1193rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1194          struct rcu_node *rnp, unsigned long flags)
1195    __releases(rnp->lock)
1196{
1197    struct rcu_node *rnp_c;
1198
1199    /* Walk up the rcu_node hierarchy. */
1200    for (;;) {
1201        if (!(rnp->qsmask & mask)) {
1202
1203            /* Our bit has already been cleared, so done. */
1204            raw_spin_unlock_irqrestore(&rnp->lock, flags);
1205            return;
1206        }
1207        rnp->qsmask &= ~mask;
1208        trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1209                         mask, rnp->qsmask, rnp->level,
1210                         rnp->grplo, rnp->grphi,
1211                         !!rnp->gp_tasks);
1212        if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1213
1214            /* Other bits still set at this level, so done. */
1215            raw_spin_unlock_irqrestore(&rnp->lock, flags);
1216            return;
1217        }
1218        mask = rnp->grpmask;
1219        if (rnp->parent == NULL) {
1220
1221            /* No more levels. Exit loop holding root lock. */
1222
1223            break;
1224        }
1225        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1226        rnp_c = rnp;
1227        rnp = rnp->parent;
1228        raw_spin_lock_irqsave(&rnp->lock, flags);
1229        WARN_ON_ONCE(rnp_c->qsmask);
1230    }
1231
1232    /*
1233     * Get here if we are the last CPU to pass through a quiescent
1234     * state for this grace period. Invoke rcu_report_qs_rsp()
1235     * to clean up and start the next grace period if one is needed.
1236     */
1237    rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1238}
1239
1240/*
1241 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1242 * structure. This must be either called from the specified CPU, or
1243 * called when the specified CPU is known to be offline (and when it is
1244 * also known that no other CPU is concurrently trying to help the offline
1245 * CPU). The lastcomp argument is used to make sure we are still in the
1246 * grace period of interest. We don't want to end the current grace period
1247 * based on quiescent states detected in an earlier grace period!
1248 */
1249static void
1250rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1251{
1252    unsigned long flags;
1253    unsigned long mask;
1254    struct rcu_node *rnp;
1255
1256    rnp = rdp->mynode;
1257    raw_spin_lock_irqsave(&rnp->lock, flags);
1258    if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1259
1260        /*
1261         * The grace period in which this quiescent state was
1262         * recorded has ended, so don't report it upwards.
1263         * We will instead need a new quiescent state that lies
1264         * within the current grace period.
1265         */
1266        rdp->passed_quiesce = 0; /* need qs for new gp. */
1267        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1268        return;
1269    }
1270    mask = rdp->grpmask;
1271    if ((rnp->qsmask & mask) == 0) {
1272        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1273    } else {
1274        rdp->qs_pending = 0;
1275
1276        /*
1277         * This GP can't end until cpu checks in, so all of our
1278         * callbacks can be processed during the next GP.
1279         */
1280        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1281
1282        rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1283    }
1284}
1285
1286/*
1287 * Check to see if there is a new grace period of which this CPU
1288 * is not yet aware, and if so, set up local rcu_data state for it.
1289 * Otherwise, see if this CPU has just passed through its first
1290 * quiescent state for this grace period, and record that fact if so.
1291 */
1292static void
1293rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1294{
1295    /* If there is now a new grace period, record and return. */
1296    if (check_for_new_grace_period(rsp, rdp))
1297        return;
1298
1299    /*
1300     * Does this CPU still need to do its part for current grace period?
1301     * If no, return and let the other CPUs do their part as well.
1302     */
1303    if (!rdp->qs_pending)
1304        return;
1305
1306    /*
1307     * Was there a quiescent state since the beginning of the grace
1308     * period? If no, then exit and wait for the next call.
1309     */
1310    if (!rdp->passed_quiesce)
1311        return;
1312
1313    /*
1314     * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1315     * judge of that).
1316     */
1317    rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1318}
1319
1320#ifdef CONFIG_HOTPLUG_CPU
1321
1322/*
1323 * Send the specified CPU's RCU callbacks to the orphanage. The
1324 * specified CPU must be offline, and the caller must hold the
1325 * ->onofflock.
1326 */
1327static void
1328rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1329              struct rcu_node *rnp, struct rcu_data *rdp)
1330{
1331    int i;
1332
1333    /*
1334     * Orphan the callbacks. First adjust the counts. This is safe
1335     * because ->onofflock excludes _rcu_barrier()'s adoption of
1336     * the callbacks, thus no memory barrier is required.
1337     */
1338    if (rdp->nxtlist != NULL) {
1339        rsp->qlen_lazy += rdp->qlen_lazy;
1340        rsp->qlen += rdp->qlen;
1341        rdp->n_cbs_orphaned += rdp->qlen;
1342        rdp->qlen_lazy = 0;
1343        rdp->qlen = 0;
1344    }
1345
1346    /*
1347     * Next, move those callbacks still needing a grace period to
1348     * the orphanage, where some other CPU will pick them up.
1349     * Some of the callbacks might have gone partway through a grace
1350     * period, but that is too bad. They get to start over because we
1351     * cannot assume that grace periods are synchronized across CPUs.
1352     * We don't bother updating the ->nxttail[] array yet, instead
1353     * we just reset the whole thing later on.
1354     */
1355    if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1356        *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1357        rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1358        *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1359    }
1360
1361    /*
1362     * Then move the ready-to-invoke callbacks to the orphanage,
1363     * where some other CPU will pick them up. These will not be
1364     * required to pass though another grace period: They are done.
1365     */
1366    if (rdp->nxtlist != NULL) {
1367        *rsp->orphan_donetail = rdp->nxtlist;
1368        rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1369    }
1370
1371    /* Finally, initialize the rcu_data structure's list to empty. */
1372    rdp->nxtlist = NULL;
1373    for (i = 0; i < RCU_NEXT_SIZE; i++)
1374        rdp->nxttail[i] = &rdp->nxtlist;
1375}
1376
1377/*
1378 * Adopt the RCU callbacks from the specified rcu_state structure's
1379 * orphanage. The caller must hold the ->onofflock.
1380 */
1381static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1382{
1383    int i;
1384    struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1385
1386    /*
1387     * If there is an rcu_barrier() operation in progress, then
1388     * only the task doing that operation is permitted to adopt
1389     * callbacks. To do otherwise breaks rcu_barrier() and friends
1390     * by causing them to fail to wait for the callbacks in the
1391     * orphanage.
1392     */
1393    if (rsp->rcu_barrier_in_progress &&
1394        rsp->rcu_barrier_in_progress != current)
1395        return;
1396
1397    /* Do the accounting first. */
1398    rdp->qlen_lazy += rsp->qlen_lazy;
1399    rdp->qlen += rsp->qlen;
1400    rdp->n_cbs_adopted += rsp->qlen;
1401    if (rsp->qlen_lazy != rsp->qlen)
1402        rcu_idle_count_callbacks_posted();
1403    rsp->qlen_lazy = 0;
1404    rsp->qlen = 0;
1405
1406    /*
1407     * We do not need a memory barrier here because the only way we
1408     * can get here if there is an rcu_barrier() in flight is if
1409     * we are the task doing the rcu_barrier().
1410     */
1411
1412    /* First adopt the ready-to-invoke callbacks. */
1413    if (rsp->orphan_donelist != NULL) {
1414        *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1415        *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1416        for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1417            if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1418                rdp->nxttail[i] = rsp->orphan_donetail;
1419        rsp->orphan_donelist = NULL;
1420        rsp->orphan_donetail = &rsp->orphan_donelist;
1421    }
1422
1423    /* And then adopt the callbacks that still need a grace period. */
1424    if (rsp->orphan_nxtlist != NULL) {
1425        *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1426        rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1427        rsp->orphan_nxtlist = NULL;
1428        rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1429    }
1430}
1431
1432/*
1433 * Trace the fact that this CPU is going offline.
1434 */
1435static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1436{
1437    RCU_TRACE(unsigned long mask);
1438    RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1439    RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1440
1441    RCU_TRACE(mask = rdp->grpmask);
1442    trace_rcu_grace_period(rsp->name,
1443                   rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1444                   "cpuofl");
1445}
1446
1447/*
1448 * The CPU has been completely removed, and some other CPU is reporting
1449 * this fact from process context. Do the remainder of the cleanup,
1450 * including orphaning the outgoing CPU's RCU callbacks, and also
1451 * adopting them, if there is no _rcu_barrier() instance running.
1452 * There can only be one CPU hotplug operation at a time, so no other
1453 * CPU can be attempting to update rcu_cpu_kthread_task.
1454 */
1455static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1456{
1457    unsigned long flags;
1458    unsigned long mask;
1459    int need_report = 0;
1460    struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1461    struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1462
1463    /* Adjust any no-longer-needed kthreads. */
1464    rcu_stop_cpu_kthread(cpu);
1465    rcu_node_kthread_setaffinity(rnp, -1);
1466
1467    /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1468
1469    /* Exclude any attempts to start a new grace period. */
1470    raw_spin_lock_irqsave(&rsp->onofflock, flags);
1471
1472    /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1473    rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1474    rcu_adopt_orphan_cbs(rsp);
1475
1476    /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1477    mask = rdp->grpmask; /* rnp->grplo is constant. */
1478    do {
1479        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1480        rnp->qsmaskinit &= ~mask;
1481        if (rnp->qsmaskinit != 0) {
1482            if (rnp != rdp->mynode)
1483                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1484            break;
1485        }
1486        if (rnp == rdp->mynode)
1487            need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1488        else
1489            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1490        mask = rnp->grpmask;
1491        rnp = rnp->parent;
1492    } while (rnp != NULL);
1493
1494    /*
1495     * We still hold the leaf rcu_node structure lock here, and
1496     * irqs are still disabled. The reason for this subterfuge is
1497     * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1498     * held leads to deadlock.
1499     */
1500    raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1501    rnp = rdp->mynode;
1502    if (need_report & RCU_OFL_TASKS_NORM_GP)
1503        rcu_report_unblock_qs_rnp(rnp, flags);
1504    else
1505        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1506    if (need_report & RCU_OFL_TASKS_EXP_GP)
1507        rcu_report_exp_rnp(rsp, rnp, true);
1508}
1509
1510#else /* #ifdef CONFIG_HOTPLUG_CPU */
1511
1512static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1513{
1514}
1515
1516static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1517{
1518}
1519
1520static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1521{
1522}
1523
1524#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1525
1526/*
1527 * Invoke any RCU callbacks that have made it to the end of their grace
1528 * period. Thottle as specified by rdp->blimit.
1529 */
1530static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1531{
1532    unsigned long flags;
1533    struct rcu_head *next, *list, **tail;
1534    int bl, count, count_lazy, i;
1535
1536    /* If no callbacks are ready, just return.*/
1537    if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1538        trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1539        trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1540                    need_resched(), is_idle_task(current),
1541                    rcu_is_callbacks_kthread());
1542        return;
1543    }
1544
1545    /*
1546     * Extract the list of ready callbacks, disabling to prevent
1547     * races with call_rcu() from interrupt handlers.
1548     */
1549    local_irq_save(flags);
1550    WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1551    bl = rdp->blimit;
1552    trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1553    list = rdp->nxtlist;
1554    rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1555    *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1556    tail = rdp->nxttail[RCU_DONE_TAIL];
1557    for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1558        if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1559            rdp->nxttail[i] = &rdp->nxtlist;
1560    local_irq_restore(flags);
1561
1562    /* Invoke callbacks. */
1563    count = count_lazy = 0;
1564    while (list) {
1565        next = list->next;
1566        prefetch(next);
1567        debug_rcu_head_unqueue(list);
1568        if (__rcu_reclaim(rsp->name, list))
1569            count_lazy++;
1570        list = next;
1571        /* Stop only if limit reached and CPU has something to do. */
1572        if (++count >= bl &&
1573            (need_resched() ||
1574             (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1575            break;
1576    }
1577
1578    local_irq_save(flags);
1579    trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1580                is_idle_task(current),
1581                rcu_is_callbacks_kthread());
1582
1583    /* Update count, and requeue any remaining callbacks. */
1584    if (list != NULL) {
1585        *tail = rdp->nxtlist;
1586        rdp->nxtlist = list;
1587        for (i = 0; i < RCU_NEXT_SIZE; i++)
1588            if (&rdp->nxtlist == rdp->nxttail[i])
1589                rdp->nxttail[i] = tail;
1590            else
1591                break;
1592    }
1593    smp_mb(); /* List handling before counting for rcu_barrier(). */
1594    rdp->qlen_lazy -= count_lazy;
1595    rdp->qlen -= count;
1596    rdp->n_cbs_invoked += count;
1597
1598    /* Reinstate batch limit if we have worked down the excess. */
1599    if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1600        rdp->blimit = blimit;
1601
1602    /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1603    if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1604        rdp->qlen_last_fqs_check = 0;
1605        rdp->n_force_qs_snap = rsp->n_force_qs;
1606    } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1607        rdp->qlen_last_fqs_check = rdp->qlen;
1608
1609    local_irq_restore(flags);
1610
1611    /* Re-invoke RCU core processing if there are callbacks remaining. */
1612    if (cpu_has_callbacks_ready_to_invoke(rdp))
1613        invoke_rcu_core();
1614}
1615
1616/*
1617 * Check to see if this CPU is in a non-context-switch quiescent state
1618 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1619 * Also schedule RCU core processing.
1620 *
1621 * This function must be called from hardirq context. It is normally
1622 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1623 * false, there is no point in invoking rcu_check_callbacks().
1624 */
1625void rcu_check_callbacks(int cpu, int user)
1626{
1627    trace_rcu_utilization("Start scheduler-tick");
1628    increment_cpu_stall_ticks();
1629    if (user || rcu_is_cpu_rrupt_from_idle()) {
1630
1631        /*
1632         * Get here if this CPU took its interrupt from user
1633         * mode or from the idle loop, and if this is not a
1634         * nested interrupt. In this case, the CPU is in
1635         * a quiescent state, so note it.
1636         *
1637         * No memory barrier is required here because both
1638         * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1639         * variables that other CPUs neither access nor modify,
1640         * at least not while the corresponding CPU is online.
1641         */
1642
1643        rcu_sched_qs(cpu);
1644        rcu_bh_qs(cpu);
1645
1646    } else if (!in_softirq()) {
1647
1648        /*
1649         * Get here if this CPU did not take its interrupt from
1650         * softirq, in other words, if it is not interrupting
1651         * a rcu_bh read-side critical section. This is an _bh
1652         * critical section, so note it.
1653         */
1654
1655        rcu_bh_qs(cpu);
1656    }
1657    rcu_preempt_check_callbacks(cpu);
1658    if (rcu_pending(cpu))
1659        invoke_rcu_core();
1660    trace_rcu_utilization("End scheduler-tick");
1661}
1662
1663/*
1664 * Scan the leaf rcu_node structures, processing dyntick state for any that
1665 * have not yet encountered a quiescent state, using the function specified.
1666 * Also initiate boosting for any threads blocked on the root rcu_node.
1667 *
1668 * The caller must have suppressed start of new grace periods.
1669 */
1670static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1671{
1672    unsigned long bit;
1673    int cpu;
1674    unsigned long flags;
1675    unsigned long mask;
1676    struct rcu_node *rnp;
1677
1678    rcu_for_each_leaf_node(rsp, rnp) {
1679        mask = 0;
1680        raw_spin_lock_irqsave(&rnp->lock, flags);
1681        if (!rcu_gp_in_progress(rsp)) {
1682            raw_spin_unlock_irqrestore(&rnp->lock, flags);
1683            return;
1684        }
1685        if (rnp->qsmask == 0) {
1686            rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1687            continue;
1688        }
1689        cpu = rnp->grplo;
1690        bit = 1;
1691        for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1692            if ((rnp->qsmask & bit) != 0 &&
1693                f(per_cpu_ptr(rsp->rda, cpu)))
1694                mask |= bit;
1695        }
1696        if (mask != 0) {
1697
1698            /* rcu_report_qs_rnp() releases rnp->lock. */
1699            rcu_report_qs_rnp(mask, rsp, rnp, flags);
1700            continue;
1701        }
1702        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1703    }
1704    rnp = rcu_get_root(rsp);
1705    if (rnp->qsmask == 0) {
1706        raw_spin_lock_irqsave(&rnp->lock, flags);
1707        rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1708    }
1709}
1710
1711/*
1712 * Force quiescent states on reluctant CPUs, and also detect which
1713 * CPUs are in dyntick-idle mode.
1714 */
1715static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1716{
1717    unsigned long flags;
1718    struct rcu_node *rnp = rcu_get_root(rsp);
1719
1720    trace_rcu_utilization("Start fqs");
1721    if (!rcu_gp_in_progress(rsp)) {
1722        trace_rcu_utilization("End fqs");
1723        return; /* No grace period in progress, nothing to force. */
1724    }
1725    if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1726        rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1727        trace_rcu_utilization("End fqs");
1728        return; /* Someone else is already on the job. */
1729    }
1730    if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1731        goto unlock_fqs_ret; /* no emergency and done recently. */
1732    rsp->n_force_qs++;
1733    raw_spin_lock(&rnp->lock); /* irqs already disabled */
1734    rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1735    if(!rcu_gp_in_progress(rsp)) {
1736        rsp->n_force_qs_ngp++;
1737        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1738        goto unlock_fqs_ret; /* no GP in progress, time updated. */
1739    }
1740    rsp->fqs_active = 1;
1741    switch (rsp->fqs_state) {
1742    case RCU_GP_IDLE:
1743    case RCU_GP_INIT:
1744
1745        break; /* grace period idle or initializing, ignore. */
1746
1747    case RCU_SAVE_DYNTICK:
1748        if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1749            break; /* So gcc recognizes the dead code. */
1750
1751        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1752
1753        /* Record dyntick-idle state. */
1754        force_qs_rnp(rsp, dyntick_save_progress_counter);
1755        raw_spin_lock(&rnp->lock); /* irqs already disabled */
1756        if (rcu_gp_in_progress(rsp))
1757            rsp->fqs_state = RCU_FORCE_QS;
1758        break;
1759
1760    case RCU_FORCE_QS:
1761
1762        /* Check dyntick-idle state, send IPI to laggarts. */
1763        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1764        force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1765
1766        /* Leave state in case more forcing is required. */
1767
1768        raw_spin_lock(&rnp->lock); /* irqs already disabled */
1769        break;
1770    }
1771    rsp->fqs_active = 0;
1772    if (rsp->fqs_need_gp) {
1773        raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1774        rsp->fqs_need_gp = 0;
1775        rcu_start_gp(rsp, flags); /* releases rnp->lock */
1776        trace_rcu_utilization("End fqs");
1777        return;
1778    }
1779    raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1780unlock_fqs_ret:
1781    raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1782    trace_rcu_utilization("End fqs");
1783}
1784
1785/*
1786 * This does the RCU core processing work for the specified rcu_state
1787 * and rcu_data structures. This may be called only from the CPU to
1788 * whom the rdp belongs.
1789 */
1790static void
1791__rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1792{
1793    unsigned long flags;
1794
1795    WARN_ON_ONCE(rdp->beenonline == 0);
1796
1797    /*
1798     * If an RCU GP has gone long enough, go check for dyntick
1799     * idle CPUs and, if needed, send resched IPIs.
1800     */
1801    if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1802        force_quiescent_state(rsp, 1);
1803
1804    /*
1805     * Advance callbacks in response to end of earlier grace
1806     * period that some other CPU ended.
1807     */
1808    rcu_process_gp_end(rsp, rdp);
1809
1810    /* Update RCU state based on any recent quiescent states. */
1811    rcu_check_quiescent_state(rsp, rdp);
1812
1813    /* Does this CPU require a not-yet-started grace period? */
1814    if (cpu_needs_another_gp(rsp, rdp)) {
1815        raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1816        rcu_start_gp(rsp, flags); /* releases above lock */
1817    }
1818
1819    /* If there are callbacks ready, invoke them. */
1820    if (cpu_has_callbacks_ready_to_invoke(rdp))
1821        invoke_rcu_callbacks(rsp, rdp);
1822}
1823
1824/*
1825 * Do RCU core processing for the current CPU.
1826 */
1827static void rcu_process_callbacks(struct softirq_action *unused)
1828{
1829    trace_rcu_utilization("Start RCU core");
1830    __rcu_process_callbacks(&rcu_sched_state,
1831                &__get_cpu_var(rcu_sched_data));
1832    __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1833    rcu_preempt_process_callbacks();
1834    trace_rcu_utilization("End RCU core");
1835}
1836
1837/*
1838 * Schedule RCU callback invocation. If the specified type of RCU
1839 * does not support RCU priority boosting, just do a direct call,
1840 * otherwise wake up the per-CPU kernel kthread. Note that because we
1841 * are running on the current CPU with interrupts disabled, the
1842 * rcu_cpu_kthread_task cannot disappear out from under us.
1843 */
1844static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1845{
1846    if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1847        return;
1848    if (likely(!rsp->boost)) {
1849        rcu_do_batch(rsp, rdp);
1850        return;
1851    }
1852    invoke_rcu_callbacks_kthread();
1853}
1854
1855static void invoke_rcu_core(void)
1856{
1857    raise_softirq(RCU_SOFTIRQ);
1858}
1859
1860static void
1861__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1862       struct rcu_state *rsp, bool lazy)
1863{
1864    unsigned long flags;
1865    struct rcu_data *rdp;
1866
1867    WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1868    debug_rcu_head_queue(head);
1869    head->func = func;
1870    head->next = NULL;
1871
1872    smp_mb(); /* Ensure RCU update seen before callback registry. */
1873
1874    /*
1875     * Opportunistically note grace-period endings and beginnings.
1876     * Note that we might see a beginning right after we see an
1877     * end, but never vice versa, since this CPU has to pass through
1878     * a quiescent state betweentimes.
1879     */
1880    local_irq_save(flags);
1881    rdp = this_cpu_ptr(rsp->rda);
1882
1883    /* Add the callback to our list. */
1884    rdp->qlen++;
1885    if (lazy)
1886        rdp->qlen_lazy++;
1887    else
1888        rcu_idle_count_callbacks_posted();
1889    smp_mb(); /* Count before adding callback for rcu_barrier(). */
1890    *rdp->nxttail[RCU_NEXT_TAIL] = head;
1891    rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1892
1893    if (__is_kfree_rcu_offset((unsigned long)func))
1894        trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1895                     rdp->qlen_lazy, rdp->qlen);
1896    else
1897        trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1898
1899    /* If interrupts were disabled, don't dive into RCU core. */
1900    if (irqs_disabled_flags(flags)) {
1901        local_irq_restore(flags);
1902        return;
1903    }
1904
1905    /*
1906     * Force the grace period if too many callbacks or too long waiting.
1907     * Enforce hysteresis, and don't invoke force_quiescent_state()
1908     * if some other CPU has recently done so. Also, don't bother
1909     * invoking force_quiescent_state() if the newly enqueued callback
1910     * is the only one waiting for a grace period to complete.
1911     */
1912    if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1913
1914        /* Are we ignoring a completed grace period? */
1915        rcu_process_gp_end(rsp, rdp);
1916        check_for_new_grace_period(rsp, rdp);
1917
1918        /* Start a new grace period if one not already started. */
1919        if (!rcu_gp_in_progress(rsp)) {
1920            unsigned long nestflag;
1921            struct rcu_node *rnp_root = rcu_get_root(rsp);
1922
1923            raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1924            rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1925        } else {
1926            /* Give the grace period a kick. */
1927            rdp->blimit = LONG_MAX;
1928            if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1929                *rdp->nxttail[RCU_DONE_TAIL] != head)
1930                force_quiescent_state(rsp, 0);
1931            rdp->n_force_qs_snap = rsp->n_force_qs;
1932            rdp->qlen_last_fqs_check = rdp->qlen;
1933        }
1934    } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1935        force_quiescent_state(rsp, 1);
1936    local_irq_restore(flags);
1937}
1938
1939/*
1940 * Queue an RCU-sched callback for invocation after a grace period.
1941 */
1942void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1943{
1944    __call_rcu(head, func, &rcu_sched_state, 0);
1945}
1946EXPORT_SYMBOL_GPL(call_rcu_sched);
1947
1948/*
1949 * Queue an RCU callback for invocation after a quicker grace period.
1950 */
1951void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1952{
1953    __call_rcu(head, func, &rcu_bh_state, 0);
1954}
1955EXPORT_SYMBOL_GPL(call_rcu_bh);
1956
1957/*
1958 * Because a context switch is a grace period for RCU-sched and RCU-bh,
1959 * any blocking grace-period wait automatically implies a grace period
1960 * if there is only one CPU online at any point time during execution
1961 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
1962 * occasionally incorrectly indicate that there are multiple CPUs online
1963 * when there was in fact only one the whole time, as this just adds
1964 * some overhead: RCU still operates correctly.
1965 *
1966 * Of course, sampling num_online_cpus() with preemption enabled can
1967 * give erroneous results if there are concurrent CPU-hotplug operations.
1968 * For example, given a demonic sequence of preemptions in num_online_cpus()
1969 * and CPU-hotplug operations, there could be two or more CPUs online at
1970 * all times, but num_online_cpus() might well return one (or even zero).
1971 *
1972 * However, all such demonic sequences require at least one CPU-offline
1973 * operation. Furthermore, rcu_blocking_is_gp() giving the wrong answer
1974 * is only a problem if there is an RCU read-side critical section executing
1975 * throughout. But RCU-sched and RCU-bh read-side critical sections
1976 * disable either preemption or bh, which prevents a CPU from going offline.
1977 * Therefore, the only way that rcu_blocking_is_gp() can incorrectly return
1978 * that there is only one CPU when in fact there was more than one throughout
1979 * is when there were no RCU readers in the system. If there are no
1980 * RCU readers, the grace period by definition can be of zero length,
1981 * regardless of the number of online CPUs.
1982 */
1983static inline int rcu_blocking_is_gp(void)
1984{
1985    might_sleep(); /* Check for RCU read-side critical section. */
1986    return num_online_cpus() <= 1;
1987}
1988
1989/**
1990 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1991 *
1992 * Control will return to the caller some time after a full rcu-sched
1993 * grace period has elapsed, in other words after all currently executing
1994 * rcu-sched read-side critical sections have completed. These read-side
1995 * critical sections are delimited by rcu_read_lock_sched() and
1996 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1997 * local_irq_disable(), and so on may be used in place of
1998 * rcu_read_lock_sched().
1999 *
2000 * This means that all preempt_disable code sequences, including NMI and
2001 * hardware-interrupt handlers, in progress on entry will have completed
2002 * before this primitive returns. However, this does not guarantee that
2003 * softirq handlers will have completed, since in some kernels, these
2004 * handlers can run in process context, and can block.
2005 *
2006 * This primitive provides the guarantees made by the (now removed)
2007 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2008 * guarantees that rcu_read_lock() sections will have completed.
2009 * In "classic RCU", these two guarantees happen to be one and
2010 * the same, but can differ in realtime RCU implementations.
2011 */
2012void synchronize_sched(void)
2013{
2014    rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2015               !lock_is_held(&rcu_lock_map) &&
2016               !lock_is_held(&rcu_sched_lock_map),
2017               "Illegal synchronize_sched() in RCU-sched read-side critical section");
2018    if (rcu_blocking_is_gp())
2019        return;
2020    wait_rcu_gp(call_rcu_sched);
2021}
2022EXPORT_SYMBOL_GPL(synchronize_sched);
2023
2024/**
2025 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2026 *
2027 * Control will return to the caller some time after a full rcu_bh grace
2028 * period has elapsed, in other words after all currently executing rcu_bh
2029 * read-side critical sections have completed. RCU read-side critical
2030 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2031 * and may be nested.
2032 */
2033void synchronize_rcu_bh(void)
2034{
2035    rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2036               !lock_is_held(&rcu_lock_map) &&
2037               !lock_is_held(&rcu_sched_lock_map),
2038               "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2039    if (rcu_blocking_is_gp())
2040        return;
2041    wait_rcu_gp(call_rcu_bh);
2042}
2043EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2044
2045static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2046static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2047
2048static int synchronize_sched_expedited_cpu_stop(void *data)
2049{
2050    /*
2051     * There must be a full memory barrier on each affected CPU
2052     * between the time that try_stop_cpus() is called and the
2053     * time that it returns.
2054     *
2055     * In the current initial implementation of cpu_stop, the
2056     * above condition is already met when the control reaches
2057     * this point and the following smp_mb() is not strictly
2058     * necessary. Do smp_mb() anyway for documentation and
2059     * robustness against future implementation changes.
2060     */
2061    smp_mb(); /* See above comment block. */
2062    return 0;
2063}
2064
2065/**
2066 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2067 *
2068 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2069 * approach to force the grace period to end quickly. This consumes
2070 * significant time on all CPUs and is unfriendly to real-time workloads,
2071 * so is thus not recommended for any sort of common-case code. In fact,
2072 * if you are using synchronize_sched_expedited() in a loop, please
2073 * restructure your code to batch your updates, and then use a single
2074 * synchronize_sched() instead.
2075 *
2076 * Note that it is illegal to call this function while holding any lock
2077 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2078 * to call this function from a CPU-hotplug notifier. Failing to observe
2079 * these restriction will result in deadlock.
2080 *
2081 * This implementation can be thought of as an application of ticket
2082 * locking to RCU, with sync_sched_expedited_started and
2083 * sync_sched_expedited_done taking on the roles of the halves
2084 * of the ticket-lock word. Each task atomically increments
2085 * sync_sched_expedited_started upon entry, snapshotting the old value,
2086 * then attempts to stop all the CPUs. If this succeeds, then each
2087 * CPU will have executed a context switch, resulting in an RCU-sched
2088 * grace period. We are then done, so we use atomic_cmpxchg() to
2089 * update sync_sched_expedited_done to match our snapshot -- but
2090 * only if someone else has not already advanced past our snapshot.
2091 *
2092 * On the other hand, if try_stop_cpus() fails, we check the value
2093 * of sync_sched_expedited_done. If it has advanced past our
2094 * initial snapshot, then someone else must have forced a grace period
2095 * some time after we took our snapshot. In this case, our work is
2096 * done for us, and we can simply return. Otherwise, we try again,
2097 * but keep our initial snapshot for purposes of checking for someone
2098 * doing our work for us.
2099 *
2100 * If we fail too many times in a row, we fall back to synchronize_sched().
2101 */
2102void synchronize_sched_expedited(void)
2103{
2104    int firstsnap, s, snap, trycount = 0;
2105
2106    /* Note that atomic_inc_return() implies full memory barrier. */
2107    firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2108    get_online_cpus();
2109    WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2110
2111    /*
2112     * Each pass through the following loop attempts to force a
2113     * context switch on each CPU.
2114     */
2115    while (try_stop_cpus(cpu_online_mask,
2116                 synchronize_sched_expedited_cpu_stop,
2117                 NULL) == -EAGAIN) {
2118        put_online_cpus();
2119
2120        /* No joy, try again later. Or just synchronize_sched(). */
2121        if (trycount++ < 10)
2122            udelay(trycount * num_online_cpus());
2123        else {
2124            synchronize_sched();
2125            return;
2126        }
2127
2128        /* Check to see if someone else did our work for us. */
2129        s = atomic_read(&sync_sched_expedited_done);
2130        if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2131            smp_mb(); /* ensure test happens before caller kfree */
2132            return;
2133        }
2134
2135        /*
2136         * Refetching sync_sched_expedited_started allows later
2137         * callers to piggyback on our grace period. We subtract
2138         * 1 to get the same token that the last incrementer got.
2139         * We retry after they started, so our grace period works
2140         * for them, and they started after our first try, so their
2141         * grace period works for us.
2142         */
2143        get_online_cpus();
2144        snap = atomic_read(&sync_sched_expedited_started);
2145        smp_mb(); /* ensure read is before try_stop_cpus(). */
2146    }
2147
2148    /*
2149     * Everyone up to our most recent fetch is covered by our grace
2150     * period. Update the counter, but only if our work is still
2151     * relevant -- which it won't be if someone who started later
2152     * than we did beat us to the punch.
2153     */
2154    do {
2155        s = atomic_read(&sync_sched_expedited_done);
2156        if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2157            smp_mb(); /* ensure test happens before caller kfree */
2158            break;
2159        }
2160    } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2161
2162    put_online_cpus();
2163}
2164EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2165
2166/*
2167 * Check to see if there is any immediate RCU-related work to be done
2168 * by the current CPU, for the specified type of RCU, returning 1 if so.
2169 * The checks are in order of increasing expense: checks that can be
2170 * carried out against CPU-local state are performed first. However,
2171 * we must check for CPU stalls first, else we might not get a chance.
2172 */
2173static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2174{
2175    struct rcu_node *rnp = rdp->mynode;
2176
2177    rdp->n_rcu_pending++;
2178
2179    /* Check for CPU stalls, if enabled. */
2180    check_cpu_stall(rsp, rdp);
2181
2182    /* Is the RCU core waiting for a quiescent state from this CPU? */
2183    if (rcu_scheduler_fully_active &&
2184        rdp->qs_pending && !rdp->passed_quiesce) {
2185
2186        /*
2187         * If force_quiescent_state() coming soon and this CPU
2188         * needs a quiescent state, and this is either RCU-sched
2189         * or RCU-bh, force a local reschedule.
2190         */
2191        rdp->n_rp_qs_pending++;
2192        if (!rdp->preemptible &&
2193            ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2194                 jiffies))
2195            set_need_resched();
2196    } else if (rdp->qs_pending && rdp->passed_quiesce) {
2197        rdp->n_rp_report_qs++;
2198        return 1;
2199    }
2200
2201    /* Does this CPU have callbacks ready to invoke? */
2202    if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2203        rdp->n_rp_cb_ready++;
2204        return 1;
2205    }
2206
2207    /* Has RCU gone idle with this CPU needing another grace period? */
2208    if (cpu_needs_another_gp(rsp, rdp)) {
2209        rdp->n_rp_cpu_needs_gp++;
2210        return 1;
2211    }
2212
2213    /* Has another RCU grace period completed? */
2214    if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2215        rdp->n_rp_gp_completed++;
2216        return 1;
2217    }
2218
2219    /* Has a new RCU grace period started? */
2220    if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2221        rdp->n_rp_gp_started++;
2222        return 1;
2223    }
2224
2225    /* Has an RCU GP gone long enough to send resched IPIs &c? */
2226    if (rcu_gp_in_progress(rsp) &&
2227        ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2228        rdp->n_rp_need_fqs++;
2229        return 1;
2230    }
2231
2232    /* nothing to do */
2233    rdp->n_rp_need_nothing++;
2234    return 0;
2235}
2236
2237/*
2238 * Check to see if there is any immediate RCU-related work to be done
2239 * by the current CPU, returning 1 if so. This function is part of the
2240 * RCU implementation; it is -not- an exported member of the RCU API.
2241 */
2242static int rcu_pending(int cpu)
2243{
2244    return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
2245           __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
2246           rcu_preempt_pending(cpu);
2247}
2248
2249/*
2250 * Check to see if any future RCU-related work will need to be done
2251 * by the current CPU, even if none need be done immediately, returning
2252 * 1 if so.
2253 */
2254static int rcu_cpu_has_callbacks(int cpu)
2255{
2256    /* RCU callbacks either ready or pending? */
2257    return per_cpu(rcu_sched_data, cpu).nxtlist ||
2258           per_cpu(rcu_bh_data, cpu).nxtlist ||
2259           rcu_preempt_cpu_has_callbacks(cpu);
2260}
2261
2262/*
2263 * RCU callback function for _rcu_barrier(). If we are last, wake
2264 * up the task executing _rcu_barrier().
2265 */
2266static void rcu_barrier_callback(struct rcu_head *notused)
2267{
2268    if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2269        complete(&rcu_barrier_completion);
2270}
2271
2272/*
2273 * Called with preemption disabled, and from cross-cpu IRQ context.
2274 */
2275static void rcu_barrier_func(void *type)
2276{
2277    int cpu = smp_processor_id();
2278    struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
2279    void (*call_rcu_func)(struct rcu_head *head,
2280                  void (*func)(struct rcu_head *head));
2281
2282    atomic_inc(&rcu_barrier_cpu_count);
2283    call_rcu_func = type;
2284    call_rcu_func(head, rcu_barrier_callback);
2285}
2286
2287/*
2288 * Orchestrate the specified type of RCU barrier, waiting for all
2289 * RCU callbacks of the specified type to complete.
2290 */
2291static void _rcu_barrier(struct rcu_state *rsp,
2292             void (*call_rcu_func)(struct rcu_head *head,
2293                           void (*func)(struct rcu_head *head)))
2294{
2295    int cpu;
2296    unsigned long flags;
2297    struct rcu_data *rdp;
2298    struct rcu_head rh;
2299
2300    init_rcu_head_on_stack(&rh);
2301
2302    /* Take mutex to serialize concurrent rcu_barrier() requests. */
2303    mutex_lock(&rcu_barrier_mutex);
2304
2305    smp_mb(); /* Prevent any prior operations from leaking in. */
2306
2307    /*
2308     * Initialize the count to one rather than to zero in order to
2309     * avoid a too-soon return to zero in case of a short grace period
2310     * (or preemption of this task). Also flag this task as doing
2311     * an rcu_barrier(). This will prevent anyone else from adopting
2312     * orphaned callbacks, which could cause otherwise failure if a
2313     * CPU went offline and quickly came back online. To see this,
2314     * consider the following sequence of events:
2315     *
2316     * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2317     * 2. CPU 1 goes offline, orphaning its callbacks.
2318     * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2319     * 4. CPU 1 comes back online.
2320     * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2321     * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2322     * us -- but before CPU 1's orphaned callbacks are invoked!!!
2323     */
2324    init_completion(&rcu_barrier_completion);
2325    atomic_set(&rcu_barrier_cpu_count, 1);
2326    raw_spin_lock_irqsave(&rsp->onofflock, flags);
2327    rsp->rcu_barrier_in_progress = current;
2328    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2329
2330    /*
2331     * Force every CPU with callbacks to register a new callback
2332     * that will tell us when all the preceding callbacks have
2333     * been invoked. If an offline CPU has callbacks, wait for
2334     * it to either come back online or to finish orphaning those
2335     * callbacks.
2336     */
2337    for_each_possible_cpu(cpu) {
2338        preempt_disable();
2339        rdp = per_cpu_ptr(rsp->rda, cpu);
2340        if (cpu_is_offline(cpu)) {
2341            preempt_enable();
2342            while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2343                schedule_timeout_interruptible(1);
2344        } else if (ACCESS_ONCE(rdp->qlen)) {
2345            smp_call_function_single(cpu, rcu_barrier_func,
2346                         (void *)call_rcu_func, 1);
2347            preempt_enable();
2348        } else {
2349            preempt_enable();
2350        }
2351    }
2352
2353    /*
2354     * Now that all online CPUs have rcu_barrier_callback() callbacks
2355     * posted, we can adopt all of the orphaned callbacks and place
2356     * an rcu_barrier_callback() callback after them. When that is done,
2357     * we are guaranteed to have an rcu_barrier_callback() callback
2358     * following every callback that could possibly have been
2359     * registered before _rcu_barrier() was called.
2360     */
2361    raw_spin_lock_irqsave(&rsp->onofflock, flags);
2362    rcu_adopt_orphan_cbs(rsp);
2363    rsp->rcu_barrier_in_progress = NULL;
2364    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2365    atomic_inc(&rcu_barrier_cpu_count);
2366    smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2367    call_rcu_func(&rh, rcu_barrier_callback);
2368
2369    /*
2370     * Now that we have an rcu_barrier_callback() callback on each
2371     * CPU, and thus each counted, remove the initial count.
2372     */
2373    if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2374        complete(&rcu_barrier_completion);
2375
2376    /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2377    wait_for_completion(&rcu_barrier_completion);
2378
2379    /* Other rcu_barrier() invocations can now safely proceed. */
2380    mutex_unlock(&rcu_barrier_mutex);
2381
2382    destroy_rcu_head_on_stack(&rh);
2383}
2384
2385/**
2386 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2387 */
2388void rcu_barrier_bh(void)
2389{
2390    _rcu_barrier(&rcu_bh_state, call_rcu_bh);
2391}
2392EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2393
2394/**
2395 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2396 */
2397void rcu_barrier_sched(void)
2398{
2399    _rcu_barrier(&rcu_sched_state, call_rcu_sched);
2400}
2401EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2402
2403/*
2404 * Do boot-time initialization of a CPU's per-CPU RCU data.
2405 */
2406static void __init
2407rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2408{
2409    unsigned long flags;
2410    int i;
2411    struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2412    struct rcu_node *rnp = rcu_get_root(rsp);
2413
2414    /* Set up local state, ensuring consistent view of global state. */
2415    raw_spin_lock_irqsave(&rnp->lock, flags);
2416    rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2417    rdp->nxtlist = NULL;
2418    for (i = 0; i < RCU_NEXT_SIZE; i++)
2419        rdp->nxttail[i] = &rdp->nxtlist;
2420    rdp->qlen_lazy = 0;
2421    rdp->qlen = 0;
2422    rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2423    WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2424    WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2425    rdp->cpu = cpu;
2426    rdp->rsp = rsp;
2427    raw_spin_unlock_irqrestore(&rnp->lock, flags);
2428}
2429
2430/*
2431 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2432 * offline event can be happening at a given time. Note also that we
2433 * can accept some slop in the rsp->completed access due to the fact
2434 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2435 */
2436static void __cpuinit
2437rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2438{
2439    unsigned long flags;
2440    unsigned long mask;
2441    struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2442    struct rcu_node *rnp = rcu_get_root(rsp);
2443
2444    /* Set up local state, ensuring consistent view of global state. */
2445    raw_spin_lock_irqsave(&rnp->lock, flags);
2446    rdp->beenonline = 1; /* We have now been online. */
2447    rdp->preemptible = preemptible;
2448    rdp->qlen_last_fqs_check = 0;
2449    rdp->n_force_qs_snap = rsp->n_force_qs;
2450    rdp->blimit = blimit;
2451    rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2452    atomic_set(&rdp->dynticks->dynticks,
2453           (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2454    rcu_prepare_for_idle_init(cpu);
2455    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2456
2457    /*
2458     * A new grace period might start here. If so, we won't be part
2459     * of it, but that is OK, as we are currently in a quiescent state.
2460     */
2461
2462    /* Exclude any attempts to start a new GP on large systems. */
2463    raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2464
2465    /* Add CPU to rcu_node bitmasks. */
2466    rnp = rdp->mynode;
2467    mask = rdp->grpmask;
2468    do {
2469        /* Exclude any attempts to start a new GP on small systems. */
2470        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2471        rnp->qsmaskinit |= mask;
2472        mask = rnp->grpmask;
2473        if (rnp == rdp->mynode) {
2474            /*
2475             * If there is a grace period in progress, we will
2476             * set up to wait for it next time we run the
2477             * RCU core code.
2478             */
2479            rdp->gpnum = rnp->completed;
2480            rdp->completed = rnp->completed;
2481            rdp->passed_quiesce = 0;
2482            rdp->qs_pending = 0;
2483            rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2484            trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2485        }
2486        raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2487        rnp = rnp->parent;
2488    } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2489
2490    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2491}
2492
2493static void __cpuinit rcu_prepare_cpu(int cpu)
2494{
2495    rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2496    rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2497    rcu_preempt_init_percpu_data(cpu);
2498}
2499
2500/*
2501 * Handle CPU online/offline notification events.
2502 */
2503static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2504                    unsigned long action, void *hcpu)
2505{
2506    long cpu = (long)hcpu;
2507    struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2508    struct rcu_node *rnp = rdp->mynode;
2509
2510    trace_rcu_utilization("Start CPU hotplug");
2511    switch (action) {
2512    case CPU_UP_PREPARE:
2513    case CPU_UP_PREPARE_FROZEN:
2514        rcu_prepare_cpu(cpu);
2515        rcu_prepare_kthreads(cpu);
2516        break;
2517    case CPU_ONLINE:
2518    case CPU_DOWN_FAILED:
2519        rcu_node_kthread_setaffinity(rnp, -1);
2520        rcu_cpu_kthread_setrt(cpu, 1);
2521        break;
2522    case CPU_DOWN_PREPARE:
2523        rcu_node_kthread_setaffinity(rnp, cpu);
2524        rcu_cpu_kthread_setrt(cpu, 0);
2525        break;
2526    case CPU_DYING:
2527    case CPU_DYING_FROZEN:
2528        /*
2529         * The whole machine is "stopped" except this CPU, so we can
2530         * touch any data without introducing corruption. We send the
2531         * dying CPU's callbacks to an arbitrarily chosen online CPU.
2532         */
2533        rcu_cleanup_dying_cpu(&rcu_bh_state);
2534        rcu_cleanup_dying_cpu(&rcu_sched_state);
2535        rcu_preempt_cleanup_dying_cpu();
2536        rcu_cleanup_after_idle(cpu);
2537        break;
2538    case CPU_DEAD:
2539    case CPU_DEAD_FROZEN:
2540    case CPU_UP_CANCELED:
2541    case CPU_UP_CANCELED_FROZEN:
2542        rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
2543        rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
2544        rcu_preempt_cleanup_dead_cpu(cpu);
2545        break;
2546    default:
2547        break;
2548    }
2549    trace_rcu_utilization("End CPU hotplug");
2550    return NOTIFY_OK;
2551}
2552
2553/*
2554 * This function is invoked towards the end of the scheduler's initialization
2555 * process. Before this is called, the idle task might contain
2556 * RCU read-side critical sections (during which time, this idle
2557 * task is booting the system). After this function is called, the
2558 * idle tasks are prohibited from containing RCU read-side critical
2559 * sections. This function also enables RCU lockdep checking.
2560 */
2561void rcu_scheduler_starting(void)
2562{
2563    WARN_ON(num_online_cpus() != 1);
2564    WARN_ON(nr_context_switches() > 0);
2565    rcu_scheduler_active = 1;
2566}
2567
2568/*
2569 * Compute the per-level fanout, either using the exact fanout specified
2570 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2571 */
2572#ifdef CONFIG_RCU_FANOUT_EXACT
2573static void __init rcu_init_levelspread(struct rcu_state *rsp)
2574{
2575    int i;
2576
2577    for (i = NUM_RCU_LVLS - 1; i > 0; i--)
2578        rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2579    rsp->levelspread[0] = CONFIG_RCU_FANOUT_LEAF;
2580}
2581#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2582static void __init rcu_init_levelspread(struct rcu_state *rsp)
2583{
2584    int ccur;
2585    int cprv;
2586    int i;
2587
2588    cprv = NR_CPUS;
2589    for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2590        ccur = rsp->levelcnt[i];
2591        rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2592        cprv = ccur;
2593    }
2594}
2595#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2596
2597/*
2598 * Helper function for rcu_init() that initializes one rcu_state structure.
2599 */
2600static void __init rcu_init_one(struct rcu_state *rsp,
2601        struct rcu_data __percpu *rda)
2602{
2603    static char *buf[] = { "rcu_node_level_0",
2604                   "rcu_node_level_1",
2605                   "rcu_node_level_2",
2606                   "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2607    int cpustride = 1;
2608    int i;
2609    int j;
2610    struct rcu_node *rnp;
2611
2612    BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2613
2614    /* Initialize the level-tracking arrays. */
2615
2616    for (i = 1; i < NUM_RCU_LVLS; i++)
2617        rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2618    rcu_init_levelspread(rsp);
2619
2620    /* Initialize the elements themselves, starting from the leaves. */
2621
2622    for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2623        cpustride *= rsp->levelspread[i];
2624        rnp = rsp->level[i];
2625        for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2626            raw_spin_lock_init(&rnp->lock);
2627            lockdep_set_class_and_name(&rnp->lock,
2628                           &rcu_node_class[i], buf[i]);
2629            rnp->gpnum = 0;
2630            rnp->qsmask = 0;
2631            rnp->qsmaskinit = 0;
2632            rnp->grplo = j * cpustride;
2633            rnp->grphi = (j + 1) * cpustride - 1;
2634            if (rnp->grphi >= NR_CPUS)
2635                rnp->grphi = NR_CPUS - 1;
2636            if (i == 0) {
2637                rnp->grpnum = 0;
2638                rnp->grpmask = 0;
2639                rnp->parent = NULL;
2640            } else {
2641                rnp->grpnum = j % rsp->levelspread[i - 1];
2642                rnp->grpmask = 1UL << rnp->grpnum;
2643                rnp->parent = rsp->level[i - 1] +
2644                          j / rsp->levelspread[i - 1];
2645            }
2646            rnp->level = i;
2647            INIT_LIST_HEAD(&rnp->blkd_tasks);
2648        }
2649    }
2650
2651    rsp->rda = rda;
2652    rnp = rsp->level[NUM_RCU_LVLS - 1];
2653    for_each_possible_cpu(i) {
2654        while (i > rnp->grphi)
2655            rnp++;
2656        per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2657        rcu_boot_init_percpu_data(i, rsp);
2658    }
2659}
2660
2661void __init rcu_init(void)
2662{
2663    int cpu;
2664
2665    rcu_bootup_announce();
2666    rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2667    rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2668    __rcu_init_preempt();
2669     open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2670
2671    /*
2672     * We don't need protection against CPU-hotplug here because
2673     * this is called early in boot, before either interrupts
2674     * or the scheduler are operational.
2675     */
2676    cpu_notifier(rcu_cpu_notify, 0);
2677    for_each_online_cpu(cpu)
2678        rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2679    check_cpu_stall_init();
2680}
2681
2682#include "rcutree_plugin.h"
2683

Archive Download this file



interactive