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 preemptable 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
29/*
30 * Check the RCU kernel configuration parameters and print informative
31 * messages about anything out of the ordinary. If you like #ifdef, you
32 * will love this function.
33 */
34static void __init rcu_bootup_announce_oddness(void)
35{
36#ifdef CONFIG_RCU_TRACE
37    printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n");
38#endif
39#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
40    printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
41           CONFIG_RCU_FANOUT);
42#endif
43#ifdef CONFIG_RCU_FANOUT_EXACT
44    printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n");
45#endif
46#ifdef CONFIG_RCU_FAST_NO_HZ
47    printk(KERN_INFO
48           "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
49#endif
50#ifdef CONFIG_PROVE_RCU
51    printk(KERN_INFO "\tRCU lockdep checking is enabled.\n");
52#endif
53#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
54    printk(KERN_INFO "\tRCU torture testing starts during boot.\n");
55#endif
56#ifndef CONFIG_RCU_CPU_STALL_DETECTOR
57    printk(KERN_INFO
58           "\tRCU-based detection of stalled CPUs is disabled.\n");
59#endif
60#ifndef CONFIG_RCU_CPU_STALL_VERBOSE
61    printk(KERN_INFO "\tVerbose stalled-CPUs detection is disabled.\n");
62#endif
63#if NUM_RCU_LVL_4 != 0
64    printk(KERN_INFO "\tExperimental four-level hierarchy is enabled.\n");
65#endif
66}
67
68#ifdef CONFIG_TREE_PREEMPT_RCU
69
70struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state);
71DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
72
73static int rcu_preempted_readers_exp(struct rcu_node *rnp);
74
75/*
76 * Tell them what RCU they are running.
77 */
78static void __init rcu_bootup_announce(void)
79{
80    printk(KERN_INFO "Preemptable hierarchical RCU implementation.\n");
81    rcu_bootup_announce_oddness();
82}
83
84/*
85 * Return the number of RCU-preempt batches processed thus far
86 * for debug and statistics.
87 */
88long rcu_batches_completed_preempt(void)
89{
90    return rcu_preempt_state.completed;
91}
92EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
93
94/*
95 * Return the number of RCU batches processed thus far for debug & stats.
96 */
97long rcu_batches_completed(void)
98{
99    return rcu_batches_completed_preempt();
100}
101EXPORT_SYMBOL_GPL(rcu_batches_completed);
102
103/*
104 * Force a quiescent state for preemptible RCU.
105 */
106void rcu_force_quiescent_state(void)
107{
108    force_quiescent_state(&rcu_preempt_state, 0);
109}
110EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
111
112/*
113 * Record a preemptable-RCU quiescent state for the specified CPU. Note
114 * that this just means that the task currently running on the CPU is
115 * not in a quiescent state. There might be any number of tasks blocked
116 * while in an RCU read-side critical section.
117 *
118 * Unlike the other rcu_*_qs() functions, callers to this function
119 * must disable irqs in order to protect the assignment to
120 * ->rcu_read_unlock_special.
121 */
122static void rcu_preempt_qs(int cpu)
123{
124    struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
125
126    rdp->passed_quiesc_completed = rdp->gpnum - 1;
127    barrier();
128    rdp->passed_quiesc = 1;
129    current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
130}
131
132/*
133 * We have entered the scheduler, and the current task might soon be
134 * context-switched away from. If this task is in an RCU read-side
135 * critical section, we will no longer be able to rely on the CPU to
136 * record that fact, so we enqueue the task on the appropriate entry
137 * of the blocked_tasks[] array. The task will dequeue itself when
138 * it exits the outermost enclosing RCU read-side critical section.
139 * Therefore, the current grace period cannot be permitted to complete
140 * until the blocked_tasks[] entry indexed by the low-order bit of
141 * rnp->gpnum empties.
142 *
143 * Caller must disable preemption.
144 */
145static void rcu_preempt_note_context_switch(int cpu)
146{
147    struct task_struct *t = current;
148    unsigned long flags;
149    int phase;
150    struct rcu_data *rdp;
151    struct rcu_node *rnp;
152
153    if (t->rcu_read_lock_nesting &&
154        (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
155
156        /* Possibly blocking in an RCU read-side critical section. */
157        rdp = rcu_preempt_state.rda[cpu];
158        rnp = rdp->mynode;
159        raw_spin_lock_irqsave(&rnp->lock, flags);
160        t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
161        t->rcu_blocked_node = rnp;
162
163        /*
164         * If this CPU has already checked in, then this task
165         * will hold up the next grace period rather than the
166         * current grace period. Queue the task accordingly.
167         * If the task is queued for the current grace period
168         * (i.e., this CPU has not yet passed through a quiescent
169         * state for the current grace period), then as long
170         * as that task remains queued, the current grace period
171         * cannot end.
172         *
173         * But first, note that the current CPU must still be
174         * on line!
175         */
176        WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
177        WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
178        phase = (rnp->gpnum + !(rnp->qsmask & rdp->grpmask)) & 0x1;
179        list_add(&t->rcu_node_entry, &rnp->blocked_tasks[phase]);
180        raw_spin_unlock_irqrestore(&rnp->lock, flags);
181    }
182
183    /*
184     * Either we were not in an RCU read-side critical section to
185     * begin with, or we have now recorded that critical section
186     * globally. Either way, we can now note a quiescent state
187     * for this CPU. Again, if we were in an RCU read-side critical
188     * section, and if that critical section was blocking the current
189     * grace period, then the fact that the task has been enqueued
190     * means that we continue to block the current grace period.
191     */
192    local_irq_save(flags);
193    rcu_preempt_qs(cpu);
194    local_irq_restore(flags);
195}
196
197/*
198 * Tree-preemptable RCU implementation for rcu_read_lock().
199 * Just increment ->rcu_read_lock_nesting, shared state will be updated
200 * if we block.
201 */
202void __rcu_read_lock(void)
203{
204    ACCESS_ONCE(current->rcu_read_lock_nesting)++;
205    barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
206}
207EXPORT_SYMBOL_GPL(__rcu_read_lock);
208
209/*
210 * Check for preempted RCU readers blocking the current grace period
211 * for the specified rcu_node structure. If the caller needs a reliable
212 * answer, it must hold the rcu_node's ->lock.
213 */
214static int rcu_preempted_readers(struct rcu_node *rnp)
215{
216    int phase = rnp->gpnum & 0x1;
217
218    return !list_empty(&rnp->blocked_tasks[phase]) ||
219           !list_empty(&rnp->blocked_tasks[phase + 2]);
220}
221
222/*
223 * Record a quiescent state for all tasks that were previously queued
224 * on the specified rcu_node structure and that were blocking the current
225 * RCU grace period. The caller must hold the specified rnp->lock with
226 * irqs disabled, and this lock is released upon return, but irqs remain
227 * disabled.
228 */
229static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
230    __releases(rnp->lock)
231{
232    unsigned long mask;
233    struct rcu_node *rnp_p;
234
235    if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
236        raw_spin_unlock_irqrestore(&rnp->lock, flags);
237        return; /* Still need more quiescent states! */
238    }
239
240    rnp_p = rnp->parent;
241    if (rnp_p == NULL) {
242        /*
243         * Either there is only one rcu_node in the tree,
244         * or tasks were kicked up to root rcu_node due to
245         * CPUs going offline.
246         */
247        rcu_report_qs_rsp(&rcu_preempt_state, flags);
248        return;
249    }
250
251    /* Report up the rest of the hierarchy. */
252    mask = rnp->grpmask;
253    raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
254    raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
255    rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
256}
257
258/*
259 * Handle special cases during rcu_read_unlock(), such as needing to
260 * notify RCU core processing or task having blocked during the RCU
261 * read-side critical section.
262 */
263static void rcu_read_unlock_special(struct task_struct *t)
264{
265    int empty;
266    int empty_exp;
267    unsigned long flags;
268    struct rcu_node *rnp;
269    int special;
270
271    /* NMI handlers cannot block and cannot safely manipulate state. */
272    if (in_nmi())
273        return;
274
275    local_irq_save(flags);
276
277    /*
278     * If RCU core is waiting for this CPU to exit critical section,
279     * let it know that we have done so.
280     */
281    special = t->rcu_read_unlock_special;
282    if (special & RCU_READ_UNLOCK_NEED_QS) {
283        rcu_preempt_qs(smp_processor_id());
284    }
285
286    /* Hardware IRQ handlers cannot block. */
287    if (in_irq()) {
288        local_irq_restore(flags);
289        return;
290    }
291
292    /* Clean up if blocked during RCU read-side critical section. */
293    if (special & RCU_READ_UNLOCK_BLOCKED) {
294        t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
295
296        /*
297         * Remove this task from the list it blocked on. The
298         * task can migrate while we acquire the lock, but at
299         * most one time. So at most two passes through loop.
300         */
301        for (;;) {
302            rnp = t->rcu_blocked_node;
303            raw_spin_lock(&rnp->lock); /* irqs already disabled. */
304            if (rnp == t->rcu_blocked_node)
305                break;
306            raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
307        }
308        empty = !rcu_preempted_readers(rnp);
309        empty_exp = !rcu_preempted_readers_exp(rnp);
310        smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
311        list_del_init(&t->rcu_node_entry);
312        t->rcu_blocked_node = NULL;
313
314        /*
315         * If this was the last task on the current list, and if
316         * we aren't waiting on any CPUs, report the quiescent state.
317         * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
318         */
319        if (empty)
320            raw_spin_unlock_irqrestore(&rnp->lock, flags);
321        else
322            rcu_report_unblock_qs_rnp(rnp, flags);
323
324        /*
325         * If this was the last task on the expedited lists,
326         * then we need to report up the rcu_node hierarchy.
327         */
328        if (!empty_exp && !rcu_preempted_readers_exp(rnp))
329            rcu_report_exp_rnp(&rcu_preempt_state, rnp);
330    } else {
331        local_irq_restore(flags);
332    }
333}
334
335/*
336 * Tree-preemptable RCU implementation for rcu_read_unlock().
337 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
338 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
339 * invoke rcu_read_unlock_special() to clean up after a context switch
340 * in an RCU read-side critical section and other special cases.
341 */
342void __rcu_read_unlock(void)
343{
344    struct task_struct *t = current;
345
346    barrier(); /* needed if we ever invoke rcu_read_unlock in rcutree.c */
347    if (--ACCESS_ONCE(t->rcu_read_lock_nesting) == 0 &&
348        unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
349        rcu_read_unlock_special(t);
350#ifdef CONFIG_PROVE_LOCKING
351    WARN_ON_ONCE(ACCESS_ONCE(t->rcu_read_lock_nesting) < 0);
352#endif /* #ifdef CONFIG_PROVE_LOCKING */
353}
354EXPORT_SYMBOL_GPL(__rcu_read_unlock);
355
356#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
357
358#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
359
360/*
361 * Dump detailed information for all tasks blocking the current RCU
362 * grace period on the specified rcu_node structure.
363 */
364static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
365{
366    unsigned long flags;
367    struct list_head *lp;
368    int phase;
369    struct task_struct *t;
370
371    if (rcu_preempted_readers(rnp)) {
372        raw_spin_lock_irqsave(&rnp->lock, flags);
373        phase = rnp->gpnum & 0x1;
374        lp = &rnp->blocked_tasks[phase];
375        list_for_each_entry(t, lp, rcu_node_entry)
376            sched_show_task(t);
377        raw_spin_unlock_irqrestore(&rnp->lock, flags);
378    }
379}
380
381/*
382 * Dump detailed information for all tasks blocking the current RCU
383 * grace period.
384 */
385static void rcu_print_detail_task_stall(struct rcu_state *rsp)
386{
387    struct rcu_node *rnp = rcu_get_root(rsp);
388
389    rcu_print_detail_task_stall_rnp(rnp);
390    rcu_for_each_leaf_node(rsp, rnp)
391        rcu_print_detail_task_stall_rnp(rnp);
392}
393
394#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
395
396static void rcu_print_detail_task_stall(struct rcu_state *rsp)
397{
398}
399
400#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
401
402/*
403 * Scan the current list of tasks blocked within RCU read-side critical
404 * sections, printing out the tid of each.
405 */
406static void rcu_print_task_stall(struct rcu_node *rnp)
407{
408    struct list_head *lp;
409    int phase;
410    struct task_struct *t;
411
412    if (rcu_preempted_readers(rnp)) {
413        phase = rnp->gpnum & 0x1;
414        lp = &rnp->blocked_tasks[phase];
415        list_for_each_entry(t, lp, rcu_node_entry)
416            printk(" P%d", t->pid);
417    }
418}
419
420#endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
421
422/*
423 * Check that the list of blocked tasks for the newly completed grace
424 * period is in fact empty. It is a serious bug to complete a grace
425 * period that still has RCU readers blocked! This function must be
426 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
427 * must be held by the caller.
428 */
429static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
430{
431    WARN_ON_ONCE(rcu_preempted_readers(rnp));
432    WARN_ON_ONCE(rnp->qsmask);
433}
434
435#ifdef CONFIG_HOTPLUG_CPU
436
437/*
438 * Handle tasklist migration for case in which all CPUs covered by the
439 * specified rcu_node have gone offline. Move them up to the root
440 * rcu_node. The reason for not just moving them to the immediate
441 * parent is to remove the need for rcu_read_unlock_special() to
442 * make more than two attempts to acquire the target rcu_node's lock.
443 * Returns true if there were tasks blocking the current RCU grace
444 * period.
445 *
446 * Returns 1 if there was previously a task blocking the current grace
447 * period on the specified rcu_node structure.
448 *
449 * The caller must hold rnp->lock with irqs disabled.
450 */
451static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
452                     struct rcu_node *rnp,
453                     struct rcu_data *rdp)
454{
455    int i;
456    struct list_head *lp;
457    struct list_head *lp_root;
458    int retval = 0;
459    struct rcu_node *rnp_root = rcu_get_root(rsp);
460    struct task_struct *tp;
461
462    if (rnp == rnp_root) {
463        WARN_ONCE(1, "Last CPU thought to be offlined?");
464        return 0; /* Shouldn't happen: at least one CPU online. */
465    }
466    WARN_ON_ONCE(rnp != rdp->mynode &&
467             (!list_empty(&rnp->blocked_tasks[0]) ||
468              !list_empty(&rnp->blocked_tasks[1]) ||
469              !list_empty(&rnp->blocked_tasks[2]) ||
470              !list_empty(&rnp->blocked_tasks[3])));
471
472    /*
473     * Move tasks up to root rcu_node. Rely on the fact that the
474     * root rcu_node can be at most one ahead of the rest of the
475     * rcu_nodes in terms of gp_num value. This fact allows us to
476     * move the blocked_tasks[] array directly, element by element.
477     */
478    if (rcu_preempted_readers(rnp))
479        retval |= RCU_OFL_TASKS_NORM_GP;
480    if (rcu_preempted_readers_exp(rnp))
481        retval |= RCU_OFL_TASKS_EXP_GP;
482    for (i = 0; i < 4; i++) {
483        lp = &rnp->blocked_tasks[i];
484        lp_root = &rnp_root->blocked_tasks[i];
485        while (!list_empty(lp)) {
486            tp = list_entry(lp->next, typeof(*tp), rcu_node_entry);
487            raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
488            list_del(&tp->rcu_node_entry);
489            tp->rcu_blocked_node = rnp_root;
490            list_add(&tp->rcu_node_entry, lp_root);
491            raw_spin_unlock(&rnp_root->lock); /* irqs remain disabled */
492        }
493    }
494    return retval;
495}
496
497/*
498 * Do CPU-offline processing for preemptable RCU.
499 */
500static void rcu_preempt_offline_cpu(int cpu)
501{
502    __rcu_offline_cpu(cpu, &rcu_preempt_state);
503}
504
505#endif /* #ifdef CONFIG_HOTPLUG_CPU */
506
507/*
508 * Check for a quiescent state from the current CPU. When a task blocks,
509 * the task is recorded in the corresponding CPU's rcu_node structure,
510 * which is checked elsewhere.
511 *
512 * Caller must disable hard irqs.
513 */
514static void rcu_preempt_check_callbacks(int cpu)
515{
516    struct task_struct *t = current;
517
518    if (t->rcu_read_lock_nesting == 0) {
519        rcu_preempt_qs(cpu);
520        return;
521    }
522    if (per_cpu(rcu_preempt_data, cpu).qs_pending)
523        t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
524}
525
526/*
527 * Process callbacks for preemptable RCU.
528 */
529static void rcu_preempt_process_callbacks(void)
530{
531    __rcu_process_callbacks(&rcu_preempt_state,
532                &__get_cpu_var(rcu_preempt_data));
533}
534
535/*
536 * Queue a preemptable-RCU callback for invocation after a grace period.
537 */
538void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
539{
540    __call_rcu(head, func, &rcu_preempt_state);
541}
542EXPORT_SYMBOL_GPL(call_rcu);
543
544/**
545 * synchronize_rcu - wait until a grace period has elapsed.
546 *
547 * Control will return to the caller some time after a full grace
548 * period has elapsed, in other words after all currently executing RCU
549 * read-side critical sections have completed. RCU read-side critical
550 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
551 * and may be nested.
552 */
553void synchronize_rcu(void)
554{
555    struct rcu_synchronize rcu;
556
557    if (!rcu_scheduler_active)
558        return;
559
560    init_rcu_head_on_stack(&rcu.head);
561    init_completion(&rcu.completion);
562    /* Will wake me after RCU finished. */
563    call_rcu(&rcu.head, wakeme_after_rcu);
564    /* Wait for it. */
565    wait_for_completion(&rcu.completion);
566    destroy_rcu_head_on_stack(&rcu.head);
567}
568EXPORT_SYMBOL_GPL(synchronize_rcu);
569
570static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
571static long sync_rcu_preempt_exp_count;
572static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
573
574/*
575 * Return non-zero if there are any tasks in RCU read-side critical
576 * sections blocking the current preemptible-RCU expedited grace period.
577 * If there is no preemptible-RCU expedited grace period currently in
578 * progress, returns zero unconditionally.
579 */
580static int rcu_preempted_readers_exp(struct rcu_node *rnp)
581{
582    return !list_empty(&rnp->blocked_tasks[2]) ||
583           !list_empty(&rnp->blocked_tasks[3]);
584}
585
586/*
587 * return non-zero if there is no RCU expedited grace period in progress
588 * for the specified rcu_node structure, in other words, if all CPUs and
589 * tasks covered by the specified rcu_node structure have done their bit
590 * for the current expedited grace period. Works only for preemptible
591 * RCU -- other RCU implementation use other means.
592 *
593 * Caller must hold sync_rcu_preempt_exp_mutex.
594 */
595static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
596{
597    return !rcu_preempted_readers_exp(rnp) &&
598           ACCESS_ONCE(rnp->expmask) == 0;
599}
600
601/*
602 * Report the exit from RCU read-side critical section for the last task
603 * that queued itself during or before the current expedited preemptible-RCU
604 * grace period. This event is reported either to the rcu_node structure on
605 * which the task was queued or to one of that rcu_node structure's ancestors,
606 * recursively up the tree. (Calm down, calm down, we do the recursion
607 * iteratively!)
608 *
609 * Caller must hold sync_rcu_preempt_exp_mutex.
610 */
611static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
612{
613    unsigned long flags;
614    unsigned long mask;
615
616    raw_spin_lock_irqsave(&rnp->lock, flags);
617    for (;;) {
618        if (!sync_rcu_preempt_exp_done(rnp))
619            break;
620        if (rnp->parent == NULL) {
621            wake_up(&sync_rcu_preempt_exp_wq);
622            break;
623        }
624        mask = rnp->grpmask;
625        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
626        rnp = rnp->parent;
627        raw_spin_lock(&rnp->lock); /* irqs already disabled */
628        rnp->expmask &= ~mask;
629    }
630    raw_spin_unlock_irqrestore(&rnp->lock, flags);
631}
632
633/*
634 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
635 * grace period for the specified rcu_node structure. If there are no such
636 * tasks, report it up the rcu_node hierarchy.
637 *
638 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
639 */
640static void
641sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
642{
643    int must_wait;
644
645    raw_spin_lock(&rnp->lock); /* irqs already disabled */
646    list_splice_init(&rnp->blocked_tasks[0], &rnp->blocked_tasks[2]);
647    list_splice_init(&rnp->blocked_tasks[1], &rnp->blocked_tasks[3]);
648    must_wait = rcu_preempted_readers_exp(rnp);
649    raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
650    if (!must_wait)
651        rcu_report_exp_rnp(rsp, rnp);
652}
653
654/*
655 * Wait for an rcu-preempt grace period, but expedite it. The basic idea
656 * is to invoke synchronize_sched_expedited() to push all the tasks to
657 * the ->blocked_tasks[] lists, move all entries from the first set of
658 * ->blocked_tasks[] lists to the second set, and finally wait for this
659 * second set to drain.
660 */
661void synchronize_rcu_expedited(void)
662{
663    unsigned long flags;
664    struct rcu_node *rnp;
665    struct rcu_state *rsp = &rcu_preempt_state;
666    long snap;
667    int trycount = 0;
668
669    smp_mb(); /* Caller's modifications seen first by other CPUs. */
670    snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
671    smp_mb(); /* Above access cannot bleed into critical section. */
672
673    /*
674     * Acquire lock, falling back to synchronize_rcu() if too many
675     * lock-acquisition failures. Of course, if someone does the
676     * expedited grace period for us, just leave.
677     */
678    while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
679        if (trycount++ < 10)
680            udelay(trycount * num_online_cpus());
681        else {
682            synchronize_rcu();
683            return;
684        }
685        if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
686            goto mb_ret; /* Others did our work for us. */
687    }
688    if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
689        goto unlock_mb_ret; /* Others did our work for us. */
690
691    /* force all RCU readers onto blocked_tasks[]. */
692    synchronize_sched_expedited();
693
694    raw_spin_lock_irqsave(&rsp->onofflock, flags);
695
696    /* Initialize ->expmask for all non-leaf rcu_node structures. */
697    rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
698        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
699        rnp->expmask = rnp->qsmaskinit;
700        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
701    }
702
703    /* Snapshot current state of ->blocked_tasks[] lists. */
704    rcu_for_each_leaf_node(rsp, rnp)
705        sync_rcu_preempt_exp_init(rsp, rnp);
706    if (NUM_RCU_NODES > 1)
707        sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
708
709    raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
710
711    /* Wait for snapshotted ->blocked_tasks[] lists to drain. */
712    rnp = rcu_get_root(rsp);
713    wait_event(sync_rcu_preempt_exp_wq,
714           sync_rcu_preempt_exp_done(rnp));
715
716    /* Clean up and exit. */
717    smp_mb(); /* ensure expedited GP seen before counter increment. */
718    ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
719unlock_mb_ret:
720    mutex_unlock(&sync_rcu_preempt_exp_mutex);
721mb_ret:
722    smp_mb(); /* ensure subsequent action seen after grace period. */
723}
724EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
725
726/*
727 * Check to see if there is any immediate preemptable-RCU-related work
728 * to be done.
729 */
730static int rcu_preempt_pending(int cpu)
731{
732    return __rcu_pending(&rcu_preempt_state,
733                 &per_cpu(rcu_preempt_data, cpu));
734}
735
736/*
737 * Does preemptable RCU need the CPU to stay out of dynticks mode?
738 */
739static int rcu_preempt_needs_cpu(int cpu)
740{
741    return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
742}
743
744/**
745 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
746 */
747void rcu_barrier(void)
748{
749    _rcu_barrier(&rcu_preempt_state, call_rcu);
750}
751EXPORT_SYMBOL_GPL(rcu_barrier);
752
753/*
754 * Initialize preemptable RCU's per-CPU data.
755 */
756static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
757{
758    rcu_init_percpu_data(cpu, &rcu_preempt_state, 1);
759}
760
761/*
762 * Move preemptable RCU's callbacks to ->orphan_cbs_list.
763 */
764static void rcu_preempt_send_cbs_to_orphanage(void)
765{
766    rcu_send_cbs_to_orphanage(&rcu_preempt_state);
767}
768
769/*
770 * Initialize preemptable RCU's state structures.
771 */
772static void __init __rcu_init_preempt(void)
773{
774    RCU_INIT_FLAVOR(&rcu_preempt_state, rcu_preempt_data);
775}
776
777/*
778 * Check for a task exiting while in a preemptable-RCU read-side
779 * critical section, clean up if so. No need to issue warnings,
780 * as debug_check_no_locks_held() already does this if lockdep
781 * is enabled.
782 */
783void exit_rcu(void)
784{
785    struct task_struct *t = current;
786
787    if (t->rcu_read_lock_nesting == 0)
788        return;
789    t->rcu_read_lock_nesting = 1;
790    rcu_read_unlock();
791}
792
793#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
794
795/*
796 * Tell them what RCU they are running.
797 */
798static void __init rcu_bootup_announce(void)
799{
800    printk(KERN_INFO "Hierarchical RCU implementation.\n");
801    rcu_bootup_announce_oddness();
802}
803
804/*
805 * Return the number of RCU batches processed thus far for debug & stats.
806 */
807long rcu_batches_completed(void)
808{
809    return rcu_batches_completed_sched();
810}
811EXPORT_SYMBOL_GPL(rcu_batches_completed);
812
813/*
814 * Force a quiescent state for RCU, which, because there is no preemptible
815 * RCU, becomes the same as rcu-sched.
816 */
817void rcu_force_quiescent_state(void)
818{
819    rcu_sched_force_quiescent_state();
820}
821EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
822
823/*
824 * Because preemptable RCU does not exist, we never have to check for
825 * CPUs being in quiescent states.
826 */
827static void rcu_preempt_note_context_switch(int cpu)
828{
829}
830
831/*
832 * Because preemptable RCU does not exist, there are never any preempted
833 * RCU readers.
834 */
835static int rcu_preempted_readers(struct rcu_node *rnp)
836{
837    return 0;
838}
839
840#ifdef CONFIG_HOTPLUG_CPU
841
842/* Because preemptible RCU does not exist, no quieting of tasks. */
843static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
844{
845    raw_spin_unlock_irqrestore(&rnp->lock, flags);
846}
847
848#endif /* #ifdef CONFIG_HOTPLUG_CPU */
849
850#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
851
852/*
853 * Because preemptable RCU does not exist, we never have to check for
854 * tasks blocked within RCU read-side critical sections.
855 */
856static void rcu_print_detail_task_stall(struct rcu_state *rsp)
857{
858}
859
860/*
861 * Because preemptable RCU does not exist, we never have to check for
862 * tasks blocked within RCU read-side critical sections.
863 */
864static void rcu_print_task_stall(struct rcu_node *rnp)
865{
866}
867
868#endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
869
870/*
871 * Because there is no preemptable RCU, there can be no readers blocked,
872 * so there is no need to check for blocked tasks. So check only for
873 * bogus qsmask values.
874 */
875static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
876{
877    WARN_ON_ONCE(rnp->qsmask);
878}
879
880#ifdef CONFIG_HOTPLUG_CPU
881
882/*
883 * Because preemptable RCU does not exist, it never needs to migrate
884 * tasks that were blocked within RCU read-side critical sections, and
885 * such non-existent tasks cannot possibly have been blocking the current
886 * grace period.
887 */
888static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
889                     struct rcu_node *rnp,
890                     struct rcu_data *rdp)
891{
892    return 0;
893}
894
895/*
896 * Because preemptable RCU does not exist, it never needs CPU-offline
897 * processing.
898 */
899static void rcu_preempt_offline_cpu(int cpu)
900{
901}
902
903#endif /* #ifdef CONFIG_HOTPLUG_CPU */
904
905/*
906 * Because preemptable RCU does not exist, it never has any callbacks
907 * to check.
908 */
909static void rcu_preempt_check_callbacks(int cpu)
910{
911}
912
913/*
914 * Because preemptable RCU does not exist, it never has any callbacks
915 * to process.
916 */
917static void rcu_preempt_process_callbacks(void)
918{
919}
920
921/*
922 * In classic RCU, call_rcu() is just call_rcu_sched().
923 */
924void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
925{
926    call_rcu_sched(head, func);
927}
928EXPORT_SYMBOL_GPL(call_rcu);
929
930/*
931 * Wait for an rcu-preempt grace period, but make it happen quickly.
932 * But because preemptable RCU does not exist, map to rcu-sched.
933 */
934void synchronize_rcu_expedited(void)
935{
936    synchronize_sched_expedited();
937}
938EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
939
940#ifdef CONFIG_HOTPLUG_CPU
941
942/*
943 * Because preemptable RCU does not exist, there is never any need to
944 * report on tasks preempted in RCU read-side critical sections during
945 * expedited RCU grace periods.
946 */
947static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
948{
949    return;
950}
951
952#endif /* #ifdef CONFIG_HOTPLUG_CPU */
953
954/*
955 * Because preemptable RCU does not exist, it never has any work to do.
956 */
957static int rcu_preempt_pending(int cpu)
958{
959    return 0;
960}
961
962/*
963 * Because preemptable RCU does not exist, it never needs any CPU.
964 */
965static int rcu_preempt_needs_cpu(int cpu)
966{
967    return 0;
968}
969
970/*
971 * Because preemptable RCU does not exist, rcu_barrier() is just
972 * another name for rcu_barrier_sched().
973 */
974void rcu_barrier(void)
975{
976    rcu_barrier_sched();
977}
978EXPORT_SYMBOL_GPL(rcu_barrier);
979
980/*
981 * Because preemptable RCU does not exist, there is no per-CPU
982 * data to initialize.
983 */
984static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
985{
986}
987
988/*
989 * Because there is no preemptable RCU, there are no callbacks to move.
990 */
991static void rcu_preempt_send_cbs_to_orphanage(void)
992{
993}
994
995/*
996 * Because preemptable RCU does not exist, it need not be initialized.
997 */
998static void __init __rcu_init_preempt(void)
999{
1000}
1001
1002#endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1003
1004#if !defined(CONFIG_RCU_FAST_NO_HZ)
1005
1006/*
1007 * Check to see if any future RCU-related work will need to be done
1008 * by the current CPU, even if none need be done immediately, returning
1009 * 1 if so. This function is part of the RCU implementation; it is -not-
1010 * an exported member of the RCU API.
1011 *
1012 * Because we have preemptible RCU, just check whether this CPU needs
1013 * any flavor of RCU. Do not chew up lots of CPU cycles with preemption
1014 * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
1015 */
1016int rcu_needs_cpu(int cpu)
1017{
1018    return rcu_needs_cpu_quick_check(cpu);
1019}
1020
1021/*
1022 * Check to see if we need to continue a callback-flush operations to
1023 * allow the last CPU to enter dyntick-idle mode. But fast dyntick-idle
1024 * entry is not configured, so we never do need to.
1025 */
1026static void rcu_needs_cpu_flush(void)
1027{
1028}
1029
1030#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1031
1032#define RCU_NEEDS_CPU_FLUSHES 5
1033static DEFINE_PER_CPU(int, rcu_dyntick_drain);
1034static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff);
1035
1036/*
1037 * Check to see if any future RCU-related work will need to be done
1038 * by the current CPU, even if none need be done immediately, returning
1039 * 1 if so. This function is part of the RCU implementation; it is -not-
1040 * an exported member of the RCU API.
1041 *
1042 * Because we are not supporting preemptible RCU, attempt to accelerate
1043 * any current grace periods so that RCU no longer needs this CPU, but
1044 * only if all other CPUs are already in dynticks-idle mode. This will
1045 * allow the CPU cores to be powered down immediately, as opposed to after
1046 * waiting many milliseconds for grace periods to elapse.
1047 *
1048 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1049 * disabled, we do one pass of force_quiescent_state(), then do a
1050 * raise_softirq() to cause rcu_process_callbacks() to be invoked later.
1051 * The per-cpu rcu_dyntick_drain variable controls the sequencing.
1052 */
1053int rcu_needs_cpu(int cpu)
1054{
1055    int c = 0;
1056    int snap;
1057    int snap_nmi;
1058    int thatcpu;
1059
1060    /* Check for being in the holdoff period. */
1061    if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies)
1062        return rcu_needs_cpu_quick_check(cpu);
1063
1064    /* Don't bother unless we are the last non-dyntick-idle CPU. */
1065    for_each_online_cpu(thatcpu) {
1066        if (thatcpu == cpu)
1067            continue;
1068        snap = per_cpu(rcu_dynticks, thatcpu).dynticks;
1069        snap_nmi = per_cpu(rcu_dynticks, thatcpu).dynticks_nmi;
1070        smp_mb(); /* Order sampling of snap with end of grace period. */
1071        if (((snap & 0x1) != 0) || ((snap_nmi & 0x1) != 0)) {
1072            per_cpu(rcu_dyntick_drain, cpu) = 0;
1073            per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1;
1074            return rcu_needs_cpu_quick_check(cpu);
1075        }
1076    }
1077
1078    /* Check and update the rcu_dyntick_drain sequencing. */
1079    if (per_cpu(rcu_dyntick_drain, cpu) <= 0) {
1080        /* First time through, initialize the counter. */
1081        per_cpu(rcu_dyntick_drain, cpu) = RCU_NEEDS_CPU_FLUSHES;
1082    } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) {
1083        /* We have hit the limit, so time to give up. */
1084        per_cpu(rcu_dyntick_holdoff, cpu) = jiffies;
1085        return rcu_needs_cpu_quick_check(cpu);
1086    }
1087
1088    /* Do one step pushing remaining RCU callbacks through. */
1089    if (per_cpu(rcu_sched_data, cpu).nxtlist) {
1090        rcu_sched_qs(cpu);
1091        force_quiescent_state(&rcu_sched_state, 0);
1092        c = c || per_cpu(rcu_sched_data, cpu).nxtlist;
1093    }
1094    if (per_cpu(rcu_bh_data, cpu).nxtlist) {
1095        rcu_bh_qs(cpu);
1096        force_quiescent_state(&rcu_bh_state, 0);
1097        c = c || per_cpu(rcu_bh_data, cpu).nxtlist;
1098    }
1099
1100    /* If RCU callbacks are still pending, RCU still needs this CPU. */
1101    if (c)
1102        raise_softirq(RCU_SOFTIRQ);
1103    return c;
1104}
1105
1106/*
1107 * Check to see if we need to continue a callback-flush operations to
1108 * allow the last CPU to enter dyntick-idle mode.
1109 */
1110static void rcu_needs_cpu_flush(void)
1111{
1112    int cpu = smp_processor_id();
1113    unsigned long flags;
1114
1115    if (per_cpu(rcu_dyntick_drain, cpu) <= 0)
1116        return;
1117    local_irq_save(flags);
1118    (void)rcu_needs_cpu(cpu);
1119    local_irq_restore(flags);
1120}
1121
1122#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1123

Archive Download this file



interactive