Root/kernel/rcutree_plugin.h

1/*
2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
19 *
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
22 *
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
25 */
26
27#include <linux/delay.h>
28#include <linux/stop_machine.h>
29
30/*
31 * Check the RCU kernel configuration parameters and print informative
32 * messages about anything out of the ordinary. If you like #ifdef, you
33 * will love this function.
34 */
35static void __init rcu_bootup_announce_oddness(void)
36{
37#ifdef CONFIG_RCU_TRACE
38    printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n");
39#endif
40#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
41    printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
42           CONFIG_RCU_FANOUT);
43#endif
44#ifdef CONFIG_RCU_FANOUT_EXACT
45    printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n");
46#endif
47#ifdef CONFIG_RCU_FAST_NO_HZ
48    printk(KERN_INFO
49           "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
50#endif
51#ifdef CONFIG_PROVE_RCU
52    printk(KERN_INFO "\tRCU lockdep checking is enabled.\n");
53#endif
54#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
55    printk(KERN_INFO "\tRCU torture testing starts during boot.\n");
56#endif
57#if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
58    printk(KERN_INFO "\tVerbose stalled-CPUs detection is disabled.\n");
59#endif
60#if NUM_RCU_LVL_4 != 0
61    printk(KERN_INFO "\tExperimental four-level hierarchy is enabled.\n");
62#endif
63}
64
65#ifdef CONFIG_TREE_PREEMPT_RCU
66
67struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state);
68DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
69static struct rcu_state *rcu_state = &rcu_preempt_state;
70
71static void rcu_read_unlock_special(struct task_struct *t);
72static int rcu_preempted_readers_exp(struct rcu_node *rnp);
73
74/*
75 * Tell them what RCU they are running.
76 */
77static void __init rcu_bootup_announce(void)
78{
79    printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n");
80    rcu_bootup_announce_oddness();
81}
82
83/*
84 * Return the number of RCU-preempt batches processed thus far
85 * for debug and statistics.
86 */
87long rcu_batches_completed_preempt(void)
88{
89    return rcu_preempt_state.completed;
90}
91EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
92
93/*
94 * Return the number of RCU batches processed thus far for debug & stats.
95 */
96long rcu_batches_completed(void)
97{
98    return rcu_batches_completed_preempt();
99}
100EXPORT_SYMBOL_GPL(rcu_batches_completed);
101
102/*
103 * Force a quiescent state for preemptible RCU.
104 */
105void rcu_force_quiescent_state(void)
106{
107    force_quiescent_state(&rcu_preempt_state, 0);
108}
109EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
110
111/*
112 * Record a preemptible-RCU quiescent state for the specified CPU. Note
113 * that this just means that the task currently running on the CPU is
114 * not in a quiescent state. There might be any number of tasks blocked
115 * while in an RCU read-side critical section.
116 *
117 * Unlike the other rcu_*_qs() functions, callers to this function
118 * must disable irqs in order to protect the assignment to
119 * ->rcu_read_unlock_special.
120 */
121static void rcu_preempt_qs(int cpu)
122{
123    struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
124
125    rdp->passed_quiesc_completed = rdp->gpnum - 1;
126    barrier();
127    rdp->passed_quiesc = 1;
128    current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
129}
130
131/*
132 * We have entered the scheduler, and the current task might soon be
133 * context-switched away from. If this task is in an RCU read-side
134 * critical section, we will no longer be able to rely on the CPU to
135 * record that fact, so we enqueue the task on the blkd_tasks list.
136 * The task will dequeue itself when it exits the outermost enclosing
137 * RCU read-side critical section. Therefore, the current grace period
138 * cannot be permitted to complete until the blkd_tasks list entries
139 * predating the current grace period drain, in other words, until
140 * rnp->gp_tasks becomes NULL.
141 *
142 * Caller must disable preemption.
143 */
144static void rcu_preempt_note_context_switch(int cpu)
145{
146    struct task_struct *t = current;
147    unsigned long flags;
148    struct rcu_data *rdp;
149    struct rcu_node *rnp;
150
151    if (t->rcu_read_lock_nesting > 0 &&
152        (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
153
154        /* Possibly blocking in an RCU read-side critical section. */
155        rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
156        rnp = rdp->mynode;
157        raw_spin_lock_irqsave(&rnp->lock, flags);
158        t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
159        t->rcu_blocked_node = rnp;
160
161        /*
162         * If this CPU has already checked in, then this task
163         * will hold up the next grace period rather than the
164         * current grace period. Queue the task accordingly.
165         * If the task is queued for the current grace period
166         * (i.e., this CPU has not yet passed through a quiescent
167         * state for the current grace period), then as long
168         * as that task remains queued, the current grace period
169         * cannot end. Note that there is some uncertainty as
170         * to exactly when the current grace period started.
171         * We take a conservative approach, which can result
172         * in unnecessarily waiting on tasks that started very
173         * slightly after the current grace period began. C'est
174         * la vie!!!
175         *
176         * But first, note that the current CPU must still be
177         * on line!
178         */
179        WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
180        WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
181        if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
182            list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
183            rnp->gp_tasks = &t->rcu_node_entry;
184#ifdef CONFIG_RCU_BOOST
185            if (rnp->boost_tasks != NULL)
186                rnp->boost_tasks = rnp->gp_tasks;
187#endif /* #ifdef CONFIG_RCU_BOOST */
188        } else {
189            list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
190            if (rnp->qsmask & rdp->grpmask)
191                rnp->gp_tasks = &t->rcu_node_entry;
192        }
193        raw_spin_unlock_irqrestore(&rnp->lock, flags);
194    } else if (t->rcu_read_lock_nesting < 0 &&
195           t->rcu_read_unlock_special) {
196
197        /*
198         * Complete exit from RCU read-side critical section on
199         * behalf of preempted instance of __rcu_read_unlock().
200         */
201        rcu_read_unlock_special(t);
202    }
203
204    /*
205     * Either we were not in an RCU read-side critical section to
206     * begin with, or we have now recorded that critical section
207     * globally. Either way, we can now note a quiescent state
208     * for this CPU. Again, if we were in an RCU read-side critical
209     * section, and if that critical section was blocking the current
210     * grace period, then the fact that the task has been enqueued
211     * means that we continue to block the current grace period.
212     */
213    local_irq_save(flags);
214    rcu_preempt_qs(cpu);
215    local_irq_restore(flags);
216}
217
218/*
219 * Tree-preemptible RCU implementation for rcu_read_lock().
220 * Just increment ->rcu_read_lock_nesting, shared state will be updated
221 * if we block.
222 */
223void __rcu_read_lock(void)
224{
225    current->rcu_read_lock_nesting++;
226    barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
227}
228EXPORT_SYMBOL_GPL(__rcu_read_lock);
229
230/*
231 * Check for preempted RCU readers blocking the current grace period
232 * for the specified rcu_node structure. If the caller needs a reliable
233 * answer, it must hold the rcu_node's ->lock.
234 */
235static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
236{
237    return rnp->gp_tasks != NULL;
238}
239
240/*
241 * Record a quiescent state for all tasks that were previously queued
242 * on the specified rcu_node structure and that were blocking the current
243 * RCU grace period. The caller must hold the specified rnp->lock with
244 * irqs disabled, and this lock is released upon return, but irqs remain
245 * disabled.
246 */
247static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
248    __releases(rnp->lock)
249{
250    unsigned long mask;
251    struct rcu_node *rnp_p;
252
253    if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
254        raw_spin_unlock_irqrestore(&rnp->lock, flags);
255        return; /* Still need more quiescent states! */
256    }
257
258    rnp_p = rnp->parent;
259    if (rnp_p == NULL) {
260        /*
261         * Either there is only one rcu_node in the tree,
262         * or tasks were kicked up to root rcu_node due to
263         * CPUs going offline.
264         */
265        rcu_report_qs_rsp(&rcu_preempt_state, flags);
266        return;
267    }
268
269    /* Report up the rest of the hierarchy. */
270    mask = rnp->grpmask;
271    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
272    raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
273    rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
274}
275
276/*
277 * Advance a ->blkd_tasks-list pointer to the next entry, instead
278 * returning NULL if at the end of the list.
279 */
280static struct list_head *rcu_next_node_entry(struct task_struct *t,
281                         struct rcu_node *rnp)
282{
283    struct list_head *np;
284
285    np = t->rcu_node_entry.next;
286    if (np == &rnp->blkd_tasks)
287        np = NULL;
288    return np;
289}
290
291/*
292 * Handle special cases during rcu_read_unlock(), such as needing to
293 * notify RCU core processing or task having blocked during the RCU
294 * read-side critical section.
295 */
296static noinline void rcu_read_unlock_special(struct task_struct *t)
297{
298    int empty;
299    int empty_exp;
300    unsigned long flags;
301    struct list_head *np;
302    struct rcu_node *rnp;
303    int special;
304
305    /* NMI handlers cannot block and cannot safely manipulate state. */
306    if (in_nmi())
307        return;
308
309    local_irq_save(flags);
310
311    /*
312     * If RCU core is waiting for this CPU to exit critical section,
313     * let it know that we have done so.
314     */
315    special = t->rcu_read_unlock_special;
316    if (special & RCU_READ_UNLOCK_NEED_QS) {
317        rcu_preempt_qs(smp_processor_id());
318    }
319
320    /* Hardware IRQ handlers cannot block. */
321    if (in_irq() || in_serving_softirq()) {
322        local_irq_restore(flags);
323        return;
324    }
325
326    /* Clean up if blocked during RCU read-side critical section. */
327    if (special & RCU_READ_UNLOCK_BLOCKED) {
328        t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
329
330        /*
331         * Remove this task from the list it blocked on. The
332         * task can migrate while we acquire the lock, but at
333         * most one time. So at most two passes through loop.
334         */
335        for (;;) {
336            rnp = t->rcu_blocked_node;
337            raw_spin_lock(&rnp->lock); /* irqs already disabled. */
338            if (rnp == t->rcu_blocked_node)
339                break;
340            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
341        }
342        empty = !rcu_preempt_blocked_readers_cgp(rnp);
343        empty_exp = !rcu_preempted_readers_exp(rnp);
344        smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
345        np = rcu_next_node_entry(t, rnp);
346        list_del_init(&t->rcu_node_entry);
347        if (&t->rcu_node_entry == rnp->gp_tasks)
348            rnp->gp_tasks = np;
349        if (&t->rcu_node_entry == rnp->exp_tasks)
350            rnp->exp_tasks = np;
351#ifdef CONFIG_RCU_BOOST
352        if (&t->rcu_node_entry == rnp->boost_tasks)
353            rnp->boost_tasks = np;
354        /* Snapshot and clear ->rcu_boosted with rcu_node lock held. */
355        if (t->rcu_boosted) {
356            special |= RCU_READ_UNLOCK_BOOSTED;
357            t->rcu_boosted = 0;
358        }
359#endif /* #ifdef CONFIG_RCU_BOOST */
360        t->rcu_blocked_node = NULL;
361
362        /*
363         * If this was the last task on the current list, and if
364         * we aren't waiting on any CPUs, report the quiescent state.
365         * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
366         */
367        if (empty)
368            raw_spin_unlock_irqrestore(&rnp->lock, flags);
369        else
370            rcu_report_unblock_qs_rnp(rnp, flags);
371
372#ifdef CONFIG_RCU_BOOST
373        /* Unboost if we were boosted. */
374        if (special & RCU_READ_UNLOCK_BOOSTED) {
375            rt_mutex_unlock(t->rcu_boost_mutex);
376            t->rcu_boost_mutex = NULL;
377        }
378#endif /* #ifdef CONFIG_RCU_BOOST */
379
380        /*
381         * If this was the last task on the expedited lists,
382         * then we need to report up the rcu_node hierarchy.
383         */
384        if (!empty_exp && !rcu_preempted_readers_exp(rnp))
385            rcu_report_exp_rnp(&rcu_preempt_state, rnp);
386    } else {
387        local_irq_restore(flags);
388    }
389}
390
391/*
392 * Tree-preemptible RCU implementation for rcu_read_unlock().
393 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
394 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
395 * invoke rcu_read_unlock_special() to clean up after a context switch
396 * in an RCU read-side critical section and other special cases.
397 */
398void __rcu_read_unlock(void)
399{
400    struct task_struct *t = current;
401
402    barrier(); /* needed if we ever invoke rcu_read_unlock in rcutree.c */
403    if (t->rcu_read_lock_nesting != 1)
404        --t->rcu_read_lock_nesting;
405    else {
406        t->rcu_read_lock_nesting = INT_MIN;
407        barrier(); /* assign before ->rcu_read_unlock_special load */
408        if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
409            rcu_read_unlock_special(t);
410        barrier(); /* ->rcu_read_unlock_special load before assign */
411        t->rcu_read_lock_nesting = 0;
412    }
413#ifdef CONFIG_PROVE_LOCKING
414    {
415        int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting);
416
417        WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
418    }
419#endif /* #ifdef CONFIG_PROVE_LOCKING */
420}
421EXPORT_SYMBOL_GPL(__rcu_read_unlock);
422
423#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
424
425/*
426 * Dump detailed information for all tasks blocking the current RCU
427 * grace period on the specified rcu_node structure.
428 */
429static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
430{
431    unsigned long flags;
432    struct task_struct *t;
433
434    if (!rcu_preempt_blocked_readers_cgp(rnp))
435        return;
436    raw_spin_lock_irqsave(&rnp->lock, flags);
437    t = list_entry(rnp->gp_tasks,
438               struct task_struct, rcu_node_entry);
439    list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
440        sched_show_task(t);
441    raw_spin_unlock_irqrestore(&rnp->lock, flags);
442}
443
444/*
445 * Dump detailed information for all tasks blocking the current RCU
446 * grace period.
447 */
448static void rcu_print_detail_task_stall(struct rcu_state *rsp)
449{
450    struct rcu_node *rnp = rcu_get_root(rsp);
451
452    rcu_print_detail_task_stall_rnp(rnp);
453    rcu_for_each_leaf_node(rsp, rnp)
454        rcu_print_detail_task_stall_rnp(rnp);
455}
456
457#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
458
459static void rcu_print_detail_task_stall(struct rcu_state *rsp)
460{
461}
462
463#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
464
465/*
466 * Scan the current list of tasks blocked within RCU read-side critical
467 * sections, printing out the tid of each.
468 */
469static void rcu_print_task_stall(struct rcu_node *rnp)
470{
471    struct task_struct *t;
472
473    if (!rcu_preempt_blocked_readers_cgp(rnp))
474        return;
475    t = list_entry(rnp->gp_tasks,
476               struct task_struct, rcu_node_entry);
477    list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
478        printk(" P%d", t->pid);
479}
480
481/*
482 * Suppress preemptible RCU's CPU stall warnings by pushing the
483 * time of the next stall-warning message comfortably far into the
484 * future.
485 */
486static void rcu_preempt_stall_reset(void)
487{
488    rcu_preempt_state.jiffies_stall = jiffies + ULONG_MAX / 2;
489}
490
491/*
492 * Check that the list of blocked tasks for the newly completed grace
493 * period is in fact empty. It is a serious bug to complete a grace
494 * period that still has RCU readers blocked! This function must be
495 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
496 * must be held by the caller.
497 *
498 * Also, if there are blocked tasks on the list, they automatically
499 * block the newly created grace period, so set up ->gp_tasks accordingly.
500 */
501static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
502{
503    WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
504    if (!list_empty(&rnp->blkd_tasks))
505        rnp->gp_tasks = rnp->blkd_tasks.next;
506    WARN_ON_ONCE(rnp->qsmask);
507}
508
509#ifdef CONFIG_HOTPLUG_CPU
510
511/*
512 * Handle tasklist migration for case in which all CPUs covered by the
513 * specified rcu_node have gone offline. Move them up to the root
514 * rcu_node. The reason for not just moving them to the immediate
515 * parent is to remove the need for rcu_read_unlock_special() to
516 * make more than two attempts to acquire the target rcu_node's lock.
517 * Returns true if there were tasks blocking the current RCU grace
518 * period.
519 *
520 * Returns 1 if there was previously a task blocking the current grace
521 * period on the specified rcu_node structure.
522 *
523 * The caller must hold rnp->lock with irqs disabled.
524 */
525static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
526                     struct rcu_node *rnp,
527                     struct rcu_data *rdp)
528{
529    struct list_head *lp;
530    struct list_head *lp_root;
531    int retval = 0;
532    struct rcu_node *rnp_root = rcu_get_root(rsp);
533    struct task_struct *t;
534
535    if (rnp == rnp_root) {
536        WARN_ONCE(1, "Last CPU thought to be offlined?");
537        return 0; /* Shouldn't happen: at least one CPU online. */
538    }
539
540    /* If we are on an internal node, complain bitterly. */
541    WARN_ON_ONCE(rnp != rdp->mynode);
542
543    /*
544     * Move tasks up to root rcu_node. Don't try to get fancy for
545     * this corner-case operation -- just put this node's tasks
546     * at the head of the root node's list, and update the root node's
547     * ->gp_tasks and ->exp_tasks pointers to those of this node's,
548     * if non-NULL. This might result in waiting for more tasks than
549     * absolutely necessary, but this is a good performance/complexity
550     * tradeoff.
551     */
552    if (rcu_preempt_blocked_readers_cgp(rnp))
553        retval |= RCU_OFL_TASKS_NORM_GP;
554    if (rcu_preempted_readers_exp(rnp))
555        retval |= RCU_OFL_TASKS_EXP_GP;
556    lp = &rnp->blkd_tasks;
557    lp_root = &rnp_root->blkd_tasks;
558    while (!list_empty(lp)) {
559        t = list_entry(lp->next, typeof(*t), rcu_node_entry);
560        raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
561        list_del(&t->rcu_node_entry);
562        t->rcu_blocked_node = rnp_root;
563        list_add(&t->rcu_node_entry, lp_root);
564        if (&t->rcu_node_entry == rnp->gp_tasks)
565            rnp_root->gp_tasks = rnp->gp_tasks;
566        if (&t->rcu_node_entry == rnp->exp_tasks)
567            rnp_root->exp_tasks = rnp->exp_tasks;
568#ifdef CONFIG_RCU_BOOST
569        if (&t->rcu_node_entry == rnp->boost_tasks)
570            rnp_root->boost_tasks = rnp->boost_tasks;
571#endif /* #ifdef CONFIG_RCU_BOOST */
572        raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
573    }
574
575#ifdef CONFIG_RCU_BOOST
576    /* In case root is being boosted and leaf is not. */
577    raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
578    if (rnp_root->boost_tasks != NULL &&
579        rnp_root->boost_tasks != rnp_root->gp_tasks)
580        rnp_root->boost_tasks = rnp_root->gp_tasks;
581    raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
582#endif /* #ifdef CONFIG_RCU_BOOST */
583
584    rnp->gp_tasks = NULL;
585    rnp->exp_tasks = NULL;
586    return retval;
587}
588
589/*
590 * Do CPU-offline processing for preemptible RCU.
591 */
592static void rcu_preempt_offline_cpu(int cpu)
593{
594    __rcu_offline_cpu(cpu, &rcu_preempt_state);
595}
596
597#endif /* #ifdef CONFIG_HOTPLUG_CPU */
598
599/*
600 * Check for a quiescent state from the current CPU. When a task blocks,
601 * the task is recorded in the corresponding CPU's rcu_node structure,
602 * which is checked elsewhere.
603 *
604 * Caller must disable hard irqs.
605 */
606static void rcu_preempt_check_callbacks(int cpu)
607{
608    struct task_struct *t = current;
609
610    if (t->rcu_read_lock_nesting == 0) {
611        rcu_preempt_qs(cpu);
612        return;
613    }
614    if (t->rcu_read_lock_nesting > 0 &&
615        per_cpu(rcu_preempt_data, cpu).qs_pending)
616        t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
617}
618
619/*
620 * Process callbacks for preemptible RCU.
621 */
622static void rcu_preempt_process_callbacks(void)
623{
624    __rcu_process_callbacks(&rcu_preempt_state,
625                &__get_cpu_var(rcu_preempt_data));
626}
627
628#ifdef CONFIG_RCU_BOOST
629
630static void rcu_preempt_do_callbacks(void)
631{
632    rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
633}
634
635#endif /* #ifdef CONFIG_RCU_BOOST */
636
637/*
638 * Queue a preemptible-RCU callback for invocation after a grace period.
639 */
640void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
641{
642    __call_rcu(head, func, &rcu_preempt_state);
643}
644EXPORT_SYMBOL_GPL(call_rcu);
645
646/**
647 * synchronize_rcu - wait until a grace period has elapsed.
648 *
649 * Control will return to the caller some time after a full grace
650 * period has elapsed, in other words after all currently executing RCU
651 * read-side critical sections have completed. Note, however, that
652 * upon return from synchronize_rcu(), the caller might well be executing
653 * concurrently with new RCU read-side critical sections that began while
654 * synchronize_rcu() was waiting. RCU read-side critical sections are
655 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
656 */
657void synchronize_rcu(void)
658{
659    struct rcu_synchronize rcu;
660
661    if (!rcu_scheduler_active)
662        return;
663
664    init_rcu_head_on_stack(&rcu.head);
665    init_completion(&rcu.completion);
666    /* Will wake me after RCU finished. */
667    call_rcu(&rcu.head, wakeme_after_rcu);
668    /* Wait for it. */
669    wait_for_completion(&rcu.completion);
670    destroy_rcu_head_on_stack(&rcu.head);
671}
672EXPORT_SYMBOL_GPL(synchronize_rcu);
673
674static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
675static long sync_rcu_preempt_exp_count;
676static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
677
678/*
679 * Return non-zero if there are any tasks in RCU read-side critical
680 * sections blocking the current preemptible-RCU expedited grace period.
681 * If there is no preemptible-RCU expedited grace period currently in
682 * progress, returns zero unconditionally.
683 */
684static int rcu_preempted_readers_exp(struct rcu_node *rnp)
685{
686    return rnp->exp_tasks != NULL;
687}
688
689/*
690 * return non-zero if there is no RCU expedited grace period in progress
691 * for the specified rcu_node structure, in other words, if all CPUs and
692 * tasks covered by the specified rcu_node structure have done their bit
693 * for the current expedited grace period. Works only for preemptible
694 * RCU -- other RCU implementation use other means.
695 *
696 * Caller must hold sync_rcu_preempt_exp_mutex.
697 */
698static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
699{
700    return !rcu_preempted_readers_exp(rnp) &&
701           ACCESS_ONCE(rnp->expmask) == 0;
702}
703
704/*
705 * Report the exit from RCU read-side critical section for the last task
706 * that queued itself during or before the current expedited preemptible-RCU
707 * grace period. This event is reported either to the rcu_node structure on
708 * which the task was queued or to one of that rcu_node structure's ancestors,
709 * recursively up the tree. (Calm down, calm down, we do the recursion
710 * iteratively!)
711 *
712 * Caller must hold sync_rcu_preempt_exp_mutex.
713 */
714static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
715{
716    unsigned long flags;
717    unsigned long mask;
718
719    raw_spin_lock_irqsave(&rnp->lock, flags);
720    for (;;) {
721        if (!sync_rcu_preempt_exp_done(rnp)) {
722            raw_spin_unlock_irqrestore(&rnp->lock, flags);
723            break;
724        }
725        if (rnp->parent == NULL) {
726            raw_spin_unlock_irqrestore(&rnp->lock, flags);
727            wake_up(&sync_rcu_preempt_exp_wq);
728            break;
729        }
730        mask = rnp->grpmask;
731        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
732        rnp = rnp->parent;
733        raw_spin_lock(&rnp->lock); /* irqs already disabled */
734        rnp->expmask &= ~mask;
735    }
736}
737
738/*
739 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
740 * grace period for the specified rcu_node structure. If there are no such
741 * tasks, report it up the rcu_node hierarchy.
742 *
743 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
744 */
745static void
746sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
747{
748    unsigned long flags;
749    int must_wait = 0;
750
751    raw_spin_lock_irqsave(&rnp->lock, flags);
752    if (list_empty(&rnp->blkd_tasks))
753        raw_spin_unlock_irqrestore(&rnp->lock, flags);
754    else {
755        rnp->exp_tasks = rnp->blkd_tasks.next;
756        rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
757        must_wait = 1;
758    }
759    if (!must_wait)
760        rcu_report_exp_rnp(rsp, rnp);
761}
762
763/*
764 * Wait for an rcu-preempt grace period, but expedite it. The basic idea
765 * is to invoke synchronize_sched_expedited() to push all the tasks to
766 * the ->blkd_tasks lists and wait for this list to drain.
767 */
768void synchronize_rcu_expedited(void)
769{
770    unsigned long flags;
771    struct rcu_node *rnp;
772    struct rcu_state *rsp = &rcu_preempt_state;
773    long snap;
774    int trycount = 0;
775
776    smp_mb(); /* Caller's modifications seen first by other CPUs. */
777    snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
778    smp_mb(); /* Above access cannot bleed into critical section. */
779
780    /*
781     * Acquire lock, falling back to synchronize_rcu() if too many
782     * lock-acquisition failures. Of course, if someone does the
783     * expedited grace period for us, just leave.
784     */
785    while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
786        if (trycount++ < 10)
787            udelay(trycount * num_online_cpus());
788        else {
789            synchronize_rcu();
790            return;
791        }
792        if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
793            goto mb_ret; /* Others did our work for us. */
794    }
795    if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
796        goto unlock_mb_ret; /* Others did our work for us. */
797
798    /* force all RCU readers onto ->blkd_tasks lists. */
799    synchronize_sched_expedited();
800
801    raw_spin_lock_irqsave(&rsp->onofflock, flags);
802
803    /* Initialize ->expmask for all non-leaf rcu_node structures. */
804    rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
805        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
806        rnp->expmask = rnp->qsmaskinit;
807        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
808    }
809
810    /* Snapshot current state of ->blkd_tasks lists. */
811    rcu_for_each_leaf_node(rsp, rnp)
812        sync_rcu_preempt_exp_init(rsp, rnp);
813    if (NUM_RCU_NODES > 1)
814        sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
815
816    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
817
818    /* Wait for snapshotted ->blkd_tasks lists to drain. */
819    rnp = rcu_get_root(rsp);
820    wait_event(sync_rcu_preempt_exp_wq,
821           sync_rcu_preempt_exp_done(rnp));
822
823    /* Clean up and exit. */
824    smp_mb(); /* ensure expedited GP seen before counter increment. */
825    ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
826unlock_mb_ret:
827    mutex_unlock(&sync_rcu_preempt_exp_mutex);
828mb_ret:
829    smp_mb(); /* ensure subsequent action seen after grace period. */
830}
831EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
832
833/*
834 * Check to see if there is any immediate preemptible-RCU-related work
835 * to be done.
836 */
837static int rcu_preempt_pending(int cpu)
838{
839    return __rcu_pending(&rcu_preempt_state,
840                 &per_cpu(rcu_preempt_data, cpu));
841}
842
843/*
844 * Does preemptible RCU need the CPU to stay out of dynticks mode?
845 */
846static int rcu_preempt_needs_cpu(int cpu)
847{
848    return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
849}
850
851/**
852 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
853 */
854void rcu_barrier(void)
855{
856    _rcu_barrier(&rcu_preempt_state, call_rcu);
857}
858EXPORT_SYMBOL_GPL(rcu_barrier);
859
860/*
861 * Initialize preemptible RCU's per-CPU data.
862 */
863static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
864{
865    rcu_init_percpu_data(cpu, &rcu_preempt_state, 1);
866}
867
868/*
869 * Move preemptible RCU's callbacks from dying CPU to other online CPU.
870 */
871static void rcu_preempt_send_cbs_to_online(void)
872{
873    rcu_send_cbs_to_online(&rcu_preempt_state);
874}
875
876/*
877 * Initialize preemptible RCU's state structures.
878 */
879static void __init __rcu_init_preempt(void)
880{
881    rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
882}
883
884/*
885 * Check for a task exiting while in a preemptible-RCU read-side
886 * critical section, clean up if so. No need to issue warnings,
887 * as debug_check_no_locks_held() already does this if lockdep
888 * is enabled.
889 */
890void exit_rcu(void)
891{
892    struct task_struct *t = current;
893
894    if (t->rcu_read_lock_nesting == 0)
895        return;
896    t->rcu_read_lock_nesting = 1;
897    __rcu_read_unlock();
898}
899
900#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
901
902static struct rcu_state *rcu_state = &rcu_sched_state;
903
904/*
905 * Tell them what RCU they are running.
906 */
907static void __init rcu_bootup_announce(void)
908{
909    printk(KERN_INFO "Hierarchical RCU implementation.\n");
910    rcu_bootup_announce_oddness();
911}
912
913/*
914 * Return the number of RCU batches processed thus far for debug & stats.
915 */
916long rcu_batches_completed(void)
917{
918    return rcu_batches_completed_sched();
919}
920EXPORT_SYMBOL_GPL(rcu_batches_completed);
921
922/*
923 * Force a quiescent state for RCU, which, because there is no preemptible
924 * RCU, becomes the same as rcu-sched.
925 */
926void rcu_force_quiescent_state(void)
927{
928    rcu_sched_force_quiescent_state();
929}
930EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
931
932/*
933 * Because preemptible RCU does not exist, we never have to check for
934 * CPUs being in quiescent states.
935 */
936static void rcu_preempt_note_context_switch(int cpu)
937{
938}
939
940/*
941 * Because preemptible RCU does not exist, there are never any preempted
942 * RCU readers.
943 */
944static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
945{
946    return 0;
947}
948
949#ifdef CONFIG_HOTPLUG_CPU
950
951/* Because preemptible RCU does not exist, no quieting of tasks. */
952static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
953{
954    raw_spin_unlock_irqrestore(&rnp->lock, flags);
955}
956
957#endif /* #ifdef CONFIG_HOTPLUG_CPU */
958
959/*
960 * Because preemptible RCU does not exist, we never have to check for
961 * tasks blocked within RCU read-side critical sections.
962 */
963static void rcu_print_detail_task_stall(struct rcu_state *rsp)
964{
965}
966
967/*
968 * Because preemptible RCU does not exist, we never have to check for
969 * tasks blocked within RCU read-side critical sections.
970 */
971static void rcu_print_task_stall(struct rcu_node *rnp)
972{
973}
974
975/*
976 * Because preemptible RCU does not exist, there is no need to suppress
977 * its CPU stall warnings.
978 */
979static void rcu_preempt_stall_reset(void)
980{
981}
982
983/*
984 * Because there is no preemptible RCU, there can be no readers blocked,
985 * so there is no need to check for blocked tasks. So check only for
986 * bogus qsmask values.
987 */
988static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
989{
990    WARN_ON_ONCE(rnp->qsmask);
991}
992
993#ifdef CONFIG_HOTPLUG_CPU
994
995/*
996 * Because preemptible RCU does not exist, it never needs to migrate
997 * tasks that were blocked within RCU read-side critical sections, and
998 * such non-existent tasks cannot possibly have been blocking the current
999 * grace period.
1000 */
1001static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1002                     struct rcu_node *rnp,
1003                     struct rcu_data *rdp)
1004{
1005    return 0;
1006}
1007
1008/*
1009 * Because preemptible RCU does not exist, it never needs CPU-offline
1010 * processing.
1011 */
1012static void rcu_preempt_offline_cpu(int cpu)
1013{
1014}
1015
1016#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1017
1018/*
1019 * Because preemptible RCU does not exist, it never has any callbacks
1020 * to check.
1021 */
1022static void rcu_preempt_check_callbacks(int cpu)
1023{
1024}
1025
1026/*
1027 * Because preemptible RCU does not exist, it never has any callbacks
1028 * to process.
1029 */
1030static void rcu_preempt_process_callbacks(void)
1031{
1032}
1033
1034/*
1035 * Wait for an rcu-preempt grace period, but make it happen quickly.
1036 * But because preemptible RCU does not exist, map to rcu-sched.
1037 */
1038void synchronize_rcu_expedited(void)
1039{
1040    synchronize_sched_expedited();
1041}
1042EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1043
1044#ifdef CONFIG_HOTPLUG_CPU
1045
1046/*
1047 * Because preemptible RCU does not exist, there is never any need to
1048 * report on tasks preempted in RCU read-side critical sections during
1049 * expedited RCU grace periods.
1050 */
1051static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
1052{
1053    return;
1054}
1055
1056#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1057
1058/*
1059 * Because preemptible RCU does not exist, it never has any work to do.
1060 */
1061static int rcu_preempt_pending(int cpu)
1062{
1063    return 0;
1064}
1065
1066/*
1067 * Because preemptible RCU does not exist, it never needs any CPU.
1068 */
1069static int rcu_preempt_needs_cpu(int cpu)
1070{
1071    return 0;
1072}
1073
1074/*
1075 * Because preemptible RCU does not exist, rcu_barrier() is just
1076 * another name for rcu_barrier_sched().
1077 */
1078void rcu_barrier(void)
1079{
1080    rcu_barrier_sched();
1081}
1082EXPORT_SYMBOL_GPL(rcu_barrier);
1083
1084/*
1085 * Because preemptible RCU does not exist, there is no per-CPU
1086 * data to initialize.
1087 */
1088static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
1089{
1090}
1091
1092/*
1093 * Because there is no preemptible RCU, there are no callbacks to move.
1094 */
1095static void rcu_preempt_send_cbs_to_online(void)
1096{
1097}
1098
1099/*
1100 * Because preemptible RCU does not exist, it need not be initialized.
1101 */
1102static void __init __rcu_init_preempt(void)
1103{
1104}
1105
1106#endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1107
1108#ifdef CONFIG_RCU_BOOST
1109
1110#include "rtmutex_common.h"
1111
1112#ifdef CONFIG_RCU_TRACE
1113
1114static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1115{
1116    if (list_empty(&rnp->blkd_tasks))
1117        rnp->n_balk_blkd_tasks++;
1118    else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1119        rnp->n_balk_exp_gp_tasks++;
1120    else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1121        rnp->n_balk_boost_tasks++;
1122    else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1123        rnp->n_balk_notblocked++;
1124    else if (rnp->gp_tasks != NULL &&
1125         ULONG_CMP_LT(jiffies, rnp->boost_time))
1126        rnp->n_balk_notyet++;
1127    else
1128        rnp->n_balk_nos++;
1129}
1130
1131#else /* #ifdef CONFIG_RCU_TRACE */
1132
1133static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1134{
1135}
1136
1137#endif /* #else #ifdef CONFIG_RCU_TRACE */
1138
1139/*
1140 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1141 * or ->boost_tasks, advancing the pointer to the next task in the
1142 * ->blkd_tasks list.
1143 *
1144 * Note that irqs must be enabled: boosting the task can block.
1145 * Returns 1 if there are more tasks needing to be boosted.
1146 */
1147static int rcu_boost(struct rcu_node *rnp)
1148{
1149    unsigned long flags;
1150    struct rt_mutex mtx;
1151    struct task_struct *t;
1152    struct list_head *tb;
1153
1154    if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1155        return 0; /* Nothing left to boost. */
1156
1157    raw_spin_lock_irqsave(&rnp->lock, flags);
1158
1159    /*
1160     * Recheck under the lock: all tasks in need of boosting
1161     * might exit their RCU read-side critical sections on their own.
1162     */
1163    if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1164        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1165        return 0;
1166    }
1167
1168    /*
1169     * Preferentially boost tasks blocking expedited grace periods.
1170     * This cannot starve the normal grace periods because a second
1171     * expedited grace period must boost all blocked tasks, including
1172     * those blocking the pre-existing normal grace period.
1173     */
1174    if (rnp->exp_tasks != NULL) {
1175        tb = rnp->exp_tasks;
1176        rnp->n_exp_boosts++;
1177    } else {
1178        tb = rnp->boost_tasks;
1179        rnp->n_normal_boosts++;
1180    }
1181    rnp->n_tasks_boosted++;
1182
1183    /*
1184     * We boost task t by manufacturing an rt_mutex that appears to
1185     * be held by task t. We leave a pointer to that rt_mutex where
1186     * task t can find it, and task t will release the mutex when it
1187     * exits its outermost RCU read-side critical section. Then
1188     * simply acquiring this artificial rt_mutex will boost task
1189     * t's priority. (Thanks to tglx for suggesting this approach!)
1190     *
1191     * Note that task t must acquire rnp->lock to remove itself from
1192     * the ->blkd_tasks list, which it will do from exit() if from
1193     * nowhere else. We therefore are guaranteed that task t will
1194     * stay around at least until we drop rnp->lock. Note that
1195     * rnp->lock also resolves races between our priority boosting
1196     * and task t's exiting its outermost RCU read-side critical
1197     * section.
1198     */
1199    t = container_of(tb, struct task_struct, rcu_node_entry);
1200    rt_mutex_init_proxy_locked(&mtx, t);
1201    t->rcu_boost_mutex = &mtx;
1202    t->rcu_boosted = 1;
1203    raw_spin_unlock_irqrestore(&rnp->lock, flags);
1204    rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1205    rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1206
1207    return rnp->exp_tasks != NULL || rnp->boost_tasks != NULL;
1208}
1209
1210/*
1211 * Timer handler to initiate waking up of boost kthreads that
1212 * have yielded the CPU due to excessive numbers of tasks to
1213 * boost. We wake up the per-rcu_node kthread, which in turn
1214 * will wake up the booster kthread.
1215 */
1216static void rcu_boost_kthread_timer(unsigned long arg)
1217{
1218    invoke_rcu_node_kthread((struct rcu_node *)arg);
1219}
1220
1221/*
1222 * Priority-boosting kthread. One per leaf rcu_node and one for the
1223 * root rcu_node.
1224 */
1225static int rcu_boost_kthread(void *arg)
1226{
1227    struct rcu_node *rnp = (struct rcu_node *)arg;
1228    int spincnt = 0;
1229    int more2boost;
1230
1231    for (;;) {
1232        rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1233        rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1234        rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1235        more2boost = rcu_boost(rnp);
1236        if (more2boost)
1237            spincnt++;
1238        else
1239            spincnt = 0;
1240        if (spincnt > 10) {
1241            rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp);
1242            spincnt = 0;
1243        }
1244    }
1245    /* NOTREACHED */
1246    return 0;
1247}
1248
1249/*
1250 * Check to see if it is time to start boosting RCU readers that are
1251 * blocking the current grace period, and, if so, tell the per-rcu_node
1252 * kthread to start boosting them. If there is an expedited grace
1253 * period in progress, it is always time to boost.
1254 *
1255 * The caller must hold rnp->lock, which this function releases,
1256 * but irqs remain disabled. The ->boost_kthread_task is immortal,
1257 * so we don't need to worry about it going away.
1258 */
1259static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1260{
1261    struct task_struct *t;
1262
1263    if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1264        rnp->n_balk_exp_gp_tasks++;
1265        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1266        return;
1267    }
1268    if (rnp->exp_tasks != NULL ||
1269        (rnp->gp_tasks != NULL &&
1270         rnp->boost_tasks == NULL &&
1271         rnp->qsmask == 0 &&
1272         ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1273        if (rnp->exp_tasks == NULL)
1274            rnp->boost_tasks = rnp->gp_tasks;
1275        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1276        t = rnp->boost_kthread_task;
1277        if (t != NULL)
1278            wake_up_process(t);
1279    } else {
1280        rcu_initiate_boost_trace(rnp);
1281        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1282    }
1283}
1284
1285/*
1286 * Wake up the per-CPU kthread to invoke RCU callbacks.
1287 */
1288static void invoke_rcu_callbacks_kthread(void)
1289{
1290    unsigned long flags;
1291
1292    local_irq_save(flags);
1293    __this_cpu_write(rcu_cpu_has_work, 1);
1294    if (__this_cpu_read(rcu_cpu_kthread_task) == NULL) {
1295        local_irq_restore(flags);
1296        return;
1297    }
1298    wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
1299    local_irq_restore(flags);
1300}
1301
1302/*
1303 * Set the affinity of the boost kthread. The CPU-hotplug locks are
1304 * held, so no one should be messing with the existence of the boost
1305 * kthread.
1306 */
1307static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
1308                      cpumask_var_t cm)
1309{
1310    struct task_struct *t;
1311
1312    t = rnp->boost_kthread_task;
1313    if (t != NULL)
1314        set_cpus_allowed_ptr(rnp->boost_kthread_task, cm);
1315}
1316
1317#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1318
1319/*
1320 * Do priority-boost accounting for the start of a new grace period.
1321 */
1322static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1323{
1324    rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1325}
1326
1327/*
1328 * Create an RCU-boost kthread for the specified node if one does not
1329 * already exist. We only create this kthread for preemptible RCU.
1330 * Returns zero if all is well, a negated errno otherwise.
1331 */
1332static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1333                         struct rcu_node *rnp,
1334                         int rnp_index)
1335{
1336    unsigned long flags;
1337    struct sched_param sp;
1338    struct task_struct *t;
1339
1340    if (&rcu_preempt_state != rsp)
1341        return 0;
1342    rsp->boost = 1;
1343    if (rnp->boost_kthread_task != NULL)
1344        return 0;
1345    t = kthread_create(rcu_boost_kthread, (void *)rnp,
1346               "rcub%d", rnp_index);
1347    if (IS_ERR(t))
1348        return PTR_ERR(t);
1349    raw_spin_lock_irqsave(&rnp->lock, flags);
1350    rnp->boost_kthread_task = t;
1351    raw_spin_unlock_irqrestore(&rnp->lock, flags);
1352    sp.sched_priority = RCU_KTHREAD_PRIO;
1353    sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1354    wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1355    return 0;
1356}
1357
1358#ifdef CONFIG_HOTPLUG_CPU
1359
1360/*
1361 * Stop the RCU's per-CPU kthread when its CPU goes offline,.
1362 */
1363static void rcu_stop_cpu_kthread(int cpu)
1364{
1365    struct task_struct *t;
1366
1367    /* Stop the CPU's kthread. */
1368    t = per_cpu(rcu_cpu_kthread_task, cpu);
1369    if (t != NULL) {
1370        per_cpu(rcu_cpu_kthread_task, cpu) = NULL;
1371        kthread_stop(t);
1372    }
1373}
1374
1375#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1376
1377static void rcu_kthread_do_work(void)
1378{
1379    rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1380    rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1381    rcu_preempt_do_callbacks();
1382}
1383
1384/*
1385 * Wake up the specified per-rcu_node-structure kthread.
1386 * Because the per-rcu_node kthreads are immortal, we don't need
1387 * to do anything to keep them alive.
1388 */
1389static void invoke_rcu_node_kthread(struct rcu_node *rnp)
1390{
1391    struct task_struct *t;
1392
1393    t = rnp->node_kthread_task;
1394    if (t != NULL)
1395        wake_up_process(t);
1396}
1397
1398/*
1399 * Set the specified CPU's kthread to run RT or not, as specified by
1400 * the to_rt argument. The CPU-hotplug locks are held, so the task
1401 * is not going away.
1402 */
1403static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1404{
1405    int policy;
1406    struct sched_param sp;
1407    struct task_struct *t;
1408
1409    t = per_cpu(rcu_cpu_kthread_task, cpu);
1410    if (t == NULL)
1411        return;
1412    if (to_rt) {
1413        policy = SCHED_FIFO;
1414        sp.sched_priority = RCU_KTHREAD_PRIO;
1415    } else {
1416        policy = SCHED_NORMAL;
1417        sp.sched_priority = 0;
1418    }
1419    sched_setscheduler_nocheck(t, policy, &sp);
1420}
1421
1422/*
1423 * Timer handler to initiate the waking up of per-CPU kthreads that
1424 * have yielded the CPU due to excess numbers of RCU callbacks.
1425 * We wake up the per-rcu_node kthread, which in turn will wake up
1426 * the booster kthread.
1427 */
1428static void rcu_cpu_kthread_timer(unsigned long arg)
1429{
1430    struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
1431    struct rcu_node *rnp = rdp->mynode;
1432
1433    atomic_or(rdp->grpmask, &rnp->wakemask);
1434    invoke_rcu_node_kthread(rnp);
1435}
1436
1437/*
1438 * Drop to non-real-time priority and yield, but only after posting a
1439 * timer that will cause us to regain our real-time priority if we
1440 * remain preempted. Either way, we restore our real-time priority
1441 * before returning.
1442 */
1443static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
1444{
1445    struct sched_param sp;
1446    struct timer_list yield_timer;
1447
1448    setup_timer_on_stack(&yield_timer, f, arg);
1449    mod_timer(&yield_timer, jiffies + 2);
1450    sp.sched_priority = 0;
1451    sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
1452    set_user_nice(current, 19);
1453    schedule();
1454    sp.sched_priority = RCU_KTHREAD_PRIO;
1455    sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1456    del_timer(&yield_timer);
1457}
1458
1459/*
1460 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1461 * This can happen while the corresponding CPU is either coming online
1462 * or going offline. We cannot wait until the CPU is fully online
1463 * before starting the kthread, because the various notifier functions
1464 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1465 * the corresponding CPU is online.
1466 *
1467 * Return 1 if the kthread needs to stop, 0 otherwise.
1468 *
1469 * Caller must disable bh. This function can momentarily enable it.
1470 */
1471static int rcu_cpu_kthread_should_stop(int cpu)
1472{
1473    while (cpu_is_offline(cpu) ||
1474           !cpumask_equal(&current->cpus_allowed, cpumask_of(cpu)) ||
1475           smp_processor_id() != cpu) {
1476        if (kthread_should_stop())
1477            return 1;
1478        per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1479        per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id();
1480        local_bh_enable();
1481        schedule_timeout_uninterruptible(1);
1482        if (!cpumask_equal(&current->cpus_allowed, cpumask_of(cpu)))
1483            set_cpus_allowed_ptr(current, cpumask_of(cpu));
1484        local_bh_disable();
1485    }
1486    per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1487    return 0;
1488}
1489
1490/*
1491 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1492 * earlier RCU softirq.
1493 */
1494static int rcu_cpu_kthread(void *arg)
1495{
1496    int cpu = (int)(long)arg;
1497    unsigned long flags;
1498    int spincnt = 0;
1499    unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu);
1500    char work;
1501    char *workp = &per_cpu(rcu_cpu_has_work, cpu);
1502
1503    for (;;) {
1504        *statusp = RCU_KTHREAD_WAITING;
1505        rcu_wait(*workp != 0 || kthread_should_stop());
1506        local_bh_disable();
1507        if (rcu_cpu_kthread_should_stop(cpu)) {
1508            local_bh_enable();
1509            break;
1510        }
1511        *statusp = RCU_KTHREAD_RUNNING;
1512        per_cpu(rcu_cpu_kthread_loops, cpu)++;
1513        local_irq_save(flags);
1514        work = *workp;
1515        *workp = 0;
1516        local_irq_restore(flags);
1517        if (work)
1518            rcu_kthread_do_work();
1519        local_bh_enable();
1520        if (*workp != 0)
1521            spincnt++;
1522        else
1523            spincnt = 0;
1524        if (spincnt > 10) {
1525            *statusp = RCU_KTHREAD_YIELDING;
1526            rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
1527            spincnt = 0;
1528        }
1529    }
1530    *statusp = RCU_KTHREAD_STOPPED;
1531    return 0;
1532}
1533
1534/*
1535 * Spawn a per-CPU kthread, setting up affinity and priority.
1536 * Because the CPU hotplug lock is held, no other CPU will be attempting
1537 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1538 * attempting to access it during boot, but the locking in kthread_bind()
1539 * will enforce sufficient ordering.
1540 *
1541 * Please note that we cannot simply refuse to wake up the per-CPU
1542 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1543 * which can result in softlockup complaints if the task ends up being
1544 * idle for more than a couple of minutes.
1545 *
1546 * However, please note also that we cannot bind the per-CPU kthread to its
1547 * CPU until that CPU is fully online. We also cannot wait until the
1548 * CPU is fully online before we create its per-CPU kthread, as this would
1549 * deadlock the system when CPU notifiers tried waiting for grace
1550 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1551 * is online. If its CPU is not yet fully online, then the code in
1552 * rcu_cpu_kthread() will wait until it is fully online, and then do
1553 * the binding.
1554 */
1555static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu)
1556{
1557    struct sched_param sp;
1558    struct task_struct *t;
1559
1560    if (!rcu_scheduler_fully_active ||
1561        per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
1562        return 0;
1563    t = kthread_create(rcu_cpu_kthread, (void *)(long)cpu, "rcuc%d", cpu);
1564    if (IS_ERR(t))
1565        return PTR_ERR(t);
1566    if (cpu_online(cpu))
1567        kthread_bind(t, cpu);
1568    per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1569    WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL);
1570    sp.sched_priority = RCU_KTHREAD_PRIO;
1571    sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1572    per_cpu(rcu_cpu_kthread_task, cpu) = t;
1573    wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */
1574    return 0;
1575}
1576
1577/*
1578 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1579 * kthreads when needed. We ignore requests to wake up kthreads
1580 * for offline CPUs, which is OK because force_quiescent_state()
1581 * takes care of this case.
1582 */
1583static int rcu_node_kthread(void *arg)
1584{
1585    int cpu;
1586    unsigned long flags;
1587    unsigned long mask;
1588    struct rcu_node *rnp = (struct rcu_node *)arg;
1589    struct sched_param sp;
1590    struct task_struct *t;
1591
1592    for (;;) {
1593        rnp->node_kthread_status = RCU_KTHREAD_WAITING;
1594        rcu_wait(atomic_read(&rnp->wakemask) != 0);
1595        rnp->node_kthread_status = RCU_KTHREAD_RUNNING;
1596        raw_spin_lock_irqsave(&rnp->lock, flags);
1597        mask = atomic_xchg(&rnp->wakemask, 0);
1598        rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1599        for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
1600            if ((mask & 0x1) == 0)
1601                continue;
1602            preempt_disable();
1603            t = per_cpu(rcu_cpu_kthread_task, cpu);
1604            if (!cpu_online(cpu) || t == NULL) {
1605                preempt_enable();
1606                continue;
1607            }
1608            per_cpu(rcu_cpu_has_work, cpu) = 1;
1609            sp.sched_priority = RCU_KTHREAD_PRIO;
1610            sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1611            preempt_enable();
1612        }
1613    }
1614    /* NOTREACHED */
1615    rnp->node_kthread_status = RCU_KTHREAD_STOPPED;
1616    return 0;
1617}
1618
1619/*
1620 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1621 * served by the rcu_node in question. The CPU hotplug lock is still
1622 * held, so the value of rnp->qsmaskinit will be stable.
1623 *
1624 * We don't include outgoingcpu in the affinity set, use -1 if there is
1625 * no outgoing CPU. If there are no CPUs left in the affinity set,
1626 * this function allows the kthread to execute on any CPU.
1627 */
1628static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1629{
1630    cpumask_var_t cm;
1631    int cpu;
1632    unsigned long mask = rnp->qsmaskinit;
1633
1634    if (rnp->node_kthread_task == NULL)
1635        return;
1636    if (!alloc_cpumask_var(&cm, GFP_KERNEL))
1637        return;
1638    cpumask_clear(cm);
1639    for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1640        if ((mask & 0x1) && cpu != outgoingcpu)
1641            cpumask_set_cpu(cpu, cm);
1642    if (cpumask_weight(cm) == 0) {
1643        cpumask_setall(cm);
1644        for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1645            cpumask_clear_cpu(cpu, cm);
1646        WARN_ON_ONCE(cpumask_weight(cm) == 0);
1647    }
1648    set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
1649    rcu_boost_kthread_setaffinity(rnp, cm);
1650    free_cpumask_var(cm);
1651}
1652
1653/*
1654 * Spawn a per-rcu_node kthread, setting priority and affinity.
1655 * Called during boot before online/offline can happen, or, if
1656 * during runtime, with the main CPU-hotplug locks held. So only
1657 * one of these can be executing at a time.
1658 */
1659static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
1660                        struct rcu_node *rnp)
1661{
1662    unsigned long flags;
1663    int rnp_index = rnp - &rsp->node[0];
1664    struct sched_param sp;
1665    struct task_struct *t;
1666
1667    if (!rcu_scheduler_fully_active ||
1668        rnp->qsmaskinit == 0)
1669        return 0;
1670    if (rnp->node_kthread_task == NULL) {
1671        t = kthread_create(rcu_node_kthread, (void *)rnp,
1672                   "rcun%d", rnp_index);
1673        if (IS_ERR(t))
1674            return PTR_ERR(t);
1675        raw_spin_lock_irqsave(&rnp->lock, flags);
1676        rnp->node_kthread_task = t;
1677        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1678        sp.sched_priority = 99;
1679        sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1680        wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1681    }
1682    return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
1683}
1684
1685/*
1686 * Spawn all kthreads -- called as soon as the scheduler is running.
1687 */
1688static int __init rcu_spawn_kthreads(void)
1689{
1690    int cpu;
1691    struct rcu_node *rnp;
1692
1693    rcu_scheduler_fully_active = 1;
1694    for_each_possible_cpu(cpu) {
1695        per_cpu(rcu_cpu_has_work, cpu) = 0;
1696        if (cpu_online(cpu))
1697            (void)rcu_spawn_one_cpu_kthread(cpu);
1698    }
1699    rnp = rcu_get_root(rcu_state);
1700    (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1701    if (NUM_RCU_NODES > 1) {
1702        rcu_for_each_leaf_node(rcu_state, rnp)
1703            (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1704    }
1705    return 0;
1706}
1707early_initcall(rcu_spawn_kthreads);
1708
1709static void __cpuinit rcu_prepare_kthreads(int cpu)
1710{
1711    struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1712    struct rcu_node *rnp = rdp->mynode;
1713
1714    /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1715    if (rcu_scheduler_fully_active) {
1716        (void)rcu_spawn_one_cpu_kthread(cpu);
1717        if (rnp->node_kthread_task == NULL)
1718            (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1719    }
1720}
1721
1722#else /* #ifdef CONFIG_RCU_BOOST */
1723
1724static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1725{
1726    raw_spin_unlock_irqrestore(&rnp->lock, flags);
1727}
1728
1729static void invoke_rcu_callbacks_kthread(void)
1730{
1731    WARN_ON_ONCE(1);
1732}
1733
1734static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1735{
1736}
1737
1738#ifdef CONFIG_HOTPLUG_CPU
1739
1740static void rcu_stop_cpu_kthread(int cpu)
1741{
1742}
1743
1744#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1745
1746static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1747{
1748}
1749
1750static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1751{
1752}
1753
1754static int __init rcu_scheduler_really_started(void)
1755{
1756    rcu_scheduler_fully_active = 1;
1757    return 0;
1758}
1759early_initcall(rcu_scheduler_really_started);
1760
1761static void __cpuinit rcu_prepare_kthreads(int cpu)
1762{
1763}
1764
1765#endif /* #else #ifdef CONFIG_RCU_BOOST */
1766
1767#ifndef CONFIG_SMP
1768
1769void synchronize_sched_expedited(void)
1770{
1771    cond_resched();
1772}
1773EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
1774
1775#else /* #ifndef CONFIG_SMP */
1776
1777static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
1778static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
1779
1780static int synchronize_sched_expedited_cpu_stop(void *data)
1781{
1782    /*
1783     * There must be a full memory barrier on each affected CPU
1784     * between the time that try_stop_cpus() is called and the
1785     * time that it returns.
1786     *
1787     * In the current initial implementation of cpu_stop, the
1788     * above condition is already met when the control reaches
1789     * this point and the following smp_mb() is not strictly
1790     * necessary. Do smp_mb() anyway for documentation and
1791     * robustness against future implementation changes.
1792     */
1793    smp_mb(); /* See above comment block. */
1794    return 0;
1795}
1796
1797/*
1798 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
1799 * approach to force grace period to end quickly. This consumes
1800 * significant time on all CPUs, and is thus not recommended for
1801 * any sort of common-case code.
1802 *
1803 * Note that it is illegal to call this function while holding any
1804 * lock that is acquired by a CPU-hotplug notifier. Failing to
1805 * observe this restriction will result in deadlock.
1806 *
1807 * This implementation can be thought of as an application of ticket
1808 * locking to RCU, with sync_sched_expedited_started and
1809 * sync_sched_expedited_done taking on the roles of the halves
1810 * of the ticket-lock word. Each task atomically increments
1811 * sync_sched_expedited_started upon entry, snapshotting the old value,
1812 * then attempts to stop all the CPUs. If this succeeds, then each
1813 * CPU will have executed a context switch, resulting in an RCU-sched
1814 * grace period. We are then done, so we use atomic_cmpxchg() to
1815 * update sync_sched_expedited_done to match our snapshot -- but
1816 * only if someone else has not already advanced past our snapshot.
1817 *
1818 * On the other hand, if try_stop_cpus() fails, we check the value
1819 * of sync_sched_expedited_done. If it has advanced past our
1820 * initial snapshot, then someone else must have forced a grace period
1821 * some time after we took our snapshot. In this case, our work is
1822 * done for us, and we can simply return. Otherwise, we try again,
1823 * but keep our initial snapshot for purposes of checking for someone
1824 * doing our work for us.
1825 *
1826 * If we fail too many times in a row, we fall back to synchronize_sched().
1827 */
1828void synchronize_sched_expedited(void)
1829{
1830    int firstsnap, s, snap, trycount = 0;
1831
1832    /* Note that atomic_inc_return() implies full memory barrier. */
1833    firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
1834    get_online_cpus();
1835
1836    /*
1837     * Each pass through the following loop attempts to force a
1838     * context switch on each CPU.
1839     */
1840    while (try_stop_cpus(cpu_online_mask,
1841                 synchronize_sched_expedited_cpu_stop,
1842                 NULL) == -EAGAIN) {
1843        put_online_cpus();
1844
1845        /* No joy, try again later. Or just synchronize_sched(). */
1846        if (trycount++ < 10)
1847            udelay(trycount * num_online_cpus());
1848        else {
1849            synchronize_sched();
1850            return;
1851        }
1852
1853        /* Check to see if someone else did our work for us. */
1854        s = atomic_read(&sync_sched_expedited_done);
1855        if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
1856            smp_mb(); /* ensure test happens before caller kfree */
1857            return;
1858        }
1859
1860        /*
1861         * Refetching sync_sched_expedited_started allows later
1862         * callers to piggyback on our grace period. We subtract
1863         * 1 to get the same token that the last incrementer got.
1864         * We retry after they started, so our grace period works
1865         * for them, and they started after our first try, so their
1866         * grace period works for us.
1867         */
1868        get_online_cpus();
1869        snap = atomic_read(&sync_sched_expedited_started) - 1;
1870        smp_mb(); /* ensure read is before try_stop_cpus(). */
1871    }
1872
1873    /*
1874     * Everyone up to our most recent fetch is covered by our grace
1875     * period. Update the counter, but only if our work is still
1876     * relevant -- which it won't be if someone who started later
1877     * than we did beat us to the punch.
1878     */
1879    do {
1880        s = atomic_read(&sync_sched_expedited_done);
1881        if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
1882            smp_mb(); /* ensure test happens before caller kfree */
1883            break;
1884        }
1885    } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
1886
1887    put_online_cpus();
1888}
1889EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
1890
1891#endif /* #else #ifndef CONFIG_SMP */
1892
1893#if !defined(CONFIG_RCU_FAST_NO_HZ)
1894
1895/*
1896 * Check to see if any future RCU-related work will need to be done
1897 * by the current CPU, even if none need be done immediately, returning
1898 * 1 if so. This function is part of the RCU implementation; it is -not-
1899 * an exported member of the RCU API.
1900 *
1901 * Because we have preemptible RCU, just check whether this CPU needs
1902 * any flavor of RCU. Do not chew up lots of CPU cycles with preemption
1903 * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
1904 */
1905int rcu_needs_cpu(int cpu)
1906{
1907    return rcu_needs_cpu_quick_check(cpu);
1908}
1909
1910/*
1911 * Check to see if we need to continue a callback-flush operations to
1912 * allow the last CPU to enter dyntick-idle mode. But fast dyntick-idle
1913 * entry is not configured, so we never do need to.
1914 */
1915static void rcu_needs_cpu_flush(void)
1916{
1917}
1918
1919#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1920
1921#define RCU_NEEDS_CPU_FLUSHES 5
1922static DEFINE_PER_CPU(int, rcu_dyntick_drain);
1923static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff);
1924
1925/*
1926 * Check to see if any future RCU-related work will need to be done
1927 * by the current CPU, even if none need be done immediately, returning
1928 * 1 if so. This function is part of the RCU implementation; it is -not-
1929 * an exported member of the RCU API.
1930 *
1931 * Because we are not supporting preemptible RCU, attempt to accelerate
1932 * any current grace periods so that RCU no longer needs this CPU, but
1933 * only if all other CPUs are already in dynticks-idle mode. This will
1934 * allow the CPU cores to be powered down immediately, as opposed to after
1935 * waiting many milliseconds for grace periods to elapse.
1936 *
1937 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1938 * disabled, we do one pass of force_quiescent_state(), then do a
1939 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
1940 * later. The per-cpu rcu_dyntick_drain variable controls the sequencing.
1941 */
1942int rcu_needs_cpu(int cpu)
1943{
1944    int c = 0;
1945    int snap;
1946    int thatcpu;
1947
1948    /* Check for being in the holdoff period. */
1949    if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies)
1950        return rcu_needs_cpu_quick_check(cpu);
1951
1952    /* Don't bother unless we are the last non-dyntick-idle CPU. */
1953    for_each_online_cpu(thatcpu) {
1954        if (thatcpu == cpu)
1955            continue;
1956        snap = atomic_add_return(0, &per_cpu(rcu_dynticks,
1957                             thatcpu).dynticks);
1958        smp_mb(); /* Order sampling of snap with end of grace period. */
1959        if ((snap & 0x1) != 0) {
1960            per_cpu(rcu_dyntick_drain, cpu) = 0;
1961            per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1;
1962            return rcu_needs_cpu_quick_check(cpu);
1963        }
1964    }
1965
1966    /* Check and update the rcu_dyntick_drain sequencing. */
1967    if (per_cpu(rcu_dyntick_drain, cpu) <= 0) {
1968        /* First time through, initialize the counter. */
1969        per_cpu(rcu_dyntick_drain, cpu) = RCU_NEEDS_CPU_FLUSHES;
1970    } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) {
1971        /* We have hit the limit, so time to give up. */
1972        per_cpu(rcu_dyntick_holdoff, cpu) = jiffies;
1973        return rcu_needs_cpu_quick_check(cpu);
1974    }
1975
1976    /* Do one step pushing remaining RCU callbacks through. */
1977    if (per_cpu(rcu_sched_data, cpu).nxtlist) {
1978        rcu_sched_qs(cpu);
1979        force_quiescent_state(&rcu_sched_state, 0);
1980        c = c || per_cpu(rcu_sched_data, cpu).nxtlist;
1981    }
1982    if (per_cpu(rcu_bh_data, cpu).nxtlist) {
1983        rcu_bh_qs(cpu);
1984        force_quiescent_state(&rcu_bh_state, 0);
1985        c = c || per_cpu(rcu_bh_data, cpu).nxtlist;
1986    }
1987
1988    /* If RCU callbacks are still pending, RCU still needs this CPU. */
1989    if (c)
1990        invoke_rcu_core();
1991    return c;
1992}
1993
1994/*
1995 * Check to see if we need to continue a callback-flush operations to
1996 * allow the last CPU to enter dyntick-idle mode.
1997 */
1998static void rcu_needs_cpu_flush(void)
1999{
2000    int cpu = smp_processor_id();
2001    unsigned long flags;
2002
2003    if (per_cpu(rcu_dyntick_drain, cpu) <= 0)
2004        return;
2005    local_irq_save(flags);
2006    (void)rcu_needs_cpu(cpu);
2007    local_irq_restore(flags);
2008}
2009
2010#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
2011

Archive Download this file



interactive