projects/linux·blob v3.2 · refs

2026 lines · 58.8 KiB

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 <[email protected]>
24 * Paul E. McKenney <[email protected]>
25 */
26
27#include <linux/delay.h>
28#include <linux/stop_machine.h>
29
30#define RCU_KTHREAD_PRIO 1
31
32#ifdef CONFIG_RCU_BOOST
33#define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
34#else
35#define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
36#endif
37
38/*
39 * Check the RCU kernel configuration parameters and print informative
40 * messages about anything out of the ordinary. If you like #ifdef, you
41 * will love this function.
42 */
43static void __init rcu_bootup_announce_oddness(void)
44{
45#ifdef CONFIG_RCU_TRACE
46 printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n");
47#endif
48#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
49 printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
50 CONFIG_RCU_FANOUT);
51#endif
52#ifdef CONFIG_RCU_FANOUT_EXACT
53 printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n");
54#endif
55#ifdef CONFIG_RCU_FAST_NO_HZ
56 printk(KERN_INFO
57 "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
58#endif
59#ifdef CONFIG_PROVE_RCU
60 printk(KERN_INFO "\tRCU lockdep checking is enabled.\n");
61#endif
62#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
63 printk(KERN_INFO "\tRCU torture testing starts during boot.\n");
64#endif
65#if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
66 printk(KERN_INFO "\tVerbose stalled-CPUs detection is disabled.\n");
67#endif
68#if NUM_RCU_LVL_4 != 0
69 printk(KERN_INFO "\tExperimental four-level hierarchy is enabled.\n");
70#endif
71}
72
73#ifdef CONFIG_TREE_PREEMPT_RCU
74
75struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt);
76DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
77static struct rcu_state *rcu_state = &rcu_preempt_state;
78
79static void rcu_read_unlock_special(struct task_struct *t);
80static int rcu_preempted_readers_exp(struct rcu_node *rnp);
81
82/*
83 * Tell them what RCU they are running.
84 */
85static void __init rcu_bootup_announce(void)
86{
87 printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n");
88 rcu_bootup_announce_oddness();
89}
90
91/*
92 * Return the number of RCU-preempt batches processed thus far
93 * for debug and statistics.
94 */
95long rcu_batches_completed_preempt(void)
96{
97 return rcu_preempt_state.completed;
98}
99EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
100
101/*
102 * Return the number of RCU batches processed thus far for debug & stats.
103 */
104long rcu_batches_completed(void)
105{
106 return rcu_batches_completed_preempt();
107}
108EXPORT_SYMBOL_GPL(rcu_batches_completed);
109
110/*
111 * Force a quiescent state for preemptible RCU.
112 */
113void rcu_force_quiescent_state(void)
114{
115 force_quiescent_state(&rcu_preempt_state, 0);
116}
117EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
118
119/*
120 * Record a preemptible-RCU quiescent state for the specified CPU. Note
121 * that this just means that the task currently running on the CPU is
122 * not in a quiescent state. There might be any number of tasks blocked
123 * while in an RCU read-side critical section.
124 *
125 * Unlike the other rcu_*_qs() functions, callers to this function
126 * must disable irqs in order to protect the assignment to
127 * ->rcu_read_unlock_special.
128 */
129static void rcu_preempt_qs(int cpu)
130{
131 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
132
133 rdp->passed_quiesce_gpnum = rdp->gpnum;
134 barrier();
135 if (rdp->passed_quiesce == 0)
136 trace_rcu_grace_period("rcu_preempt", rdp->gpnum, "cpuqs");
137 rdp->passed_quiesce = 1;
138 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
139}
140
141/*
142 * We have entered the scheduler, and the current task might soon be
143 * context-switched away from. If this task is in an RCU read-side
144 * critical section, we will no longer be able to rely on the CPU to
145 * record that fact, so we enqueue the task on the blkd_tasks list.
146 * The task will dequeue itself when it exits the outermost enclosing
147 * RCU read-side critical section. Therefore, the current grace period
148 * cannot be permitted to complete until the blkd_tasks list entries
149 * predating the current grace period drain, in other words, until
150 * rnp->gp_tasks becomes NULL.
151 *
152 * Caller must disable preemption.
153 */
154static void rcu_preempt_note_context_switch(int cpu)
155{
156 struct task_struct *t = current;
157 unsigned long flags;
158 struct rcu_data *rdp;
159 struct rcu_node *rnp;
160
161 if (t->rcu_read_lock_nesting > 0 &&
162 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
163
164 /* Possibly blocking in an RCU read-side critical section. */
165 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
166 rnp = rdp->mynode;
167 raw_spin_lock_irqsave(&rnp->lock, flags);
168 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
169 t->rcu_blocked_node = rnp;
170
171 /*
172 * If this CPU has already checked in, then this task
173 * will hold up the next grace period rather than the
174 * current grace period. Queue the task accordingly.
175 * If the task is queued for the current grace period
176 * (i.e., this CPU has not yet passed through a quiescent
177 * state for the current grace period), then as long
178 * as that task remains queued, the current grace period
179 * cannot end. Note that there is some uncertainty as
180 * to exactly when the current grace period started.
181 * We take a conservative approach, which can result
182 * in unnecessarily waiting on tasks that started very
183 * slightly after the current grace period began. C'est
184 * la vie!!!
185 *
186 * But first, note that the current CPU must still be
187 * on line!
188 */
189 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
190 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
191 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
192 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
193 rnp->gp_tasks = &t->rcu_node_entry;
194#ifdef CONFIG_RCU_BOOST
195 if (rnp->boost_tasks != NULL)
196 rnp->boost_tasks = rnp->gp_tasks;
197#endif /* #ifdef CONFIG_RCU_BOOST */
198 } else {
199 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
200 if (rnp->qsmask & rdp->grpmask)
201 rnp->gp_tasks = &t->rcu_node_entry;
202 }
203 trace_rcu_preempt_task(rdp->rsp->name,
204 t->pid,
205 (rnp->qsmask & rdp->grpmask)
206 ? rnp->gpnum
207 : rnp->gpnum + 1);
208 raw_spin_unlock_irqrestore(&rnp->lock, flags);
209 } else if (t->rcu_read_lock_nesting < 0 &&
210 t->rcu_read_unlock_special) {
211
212 /*
213 * Complete exit from RCU read-side critical section on
214 * behalf of preempted instance of __rcu_read_unlock().
215 */
216 rcu_read_unlock_special(t);
217 }
218
219 /*
220 * Either we were not in an RCU read-side critical section to
221 * begin with, or we have now recorded that critical section
222 * globally. Either way, we can now note a quiescent state
223 * for this CPU. Again, if we were in an RCU read-side critical
224 * section, and if that critical section was blocking the current
225 * grace period, then the fact that the task has been enqueued
226 * means that we continue to block the current grace period.
227 */
228 local_irq_save(flags);
229 rcu_preempt_qs(cpu);
230 local_irq_restore(flags);
231}
232
233/*
234 * Tree-preemptible RCU implementation for rcu_read_lock().
235 * Just increment ->rcu_read_lock_nesting, shared state will be updated
236 * if we block.
237 */
238void __rcu_read_lock(void)
239{
240 current->rcu_read_lock_nesting++;
241 barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
242}
243EXPORT_SYMBOL_GPL(__rcu_read_lock);
244
245/*
246 * Check for preempted RCU readers blocking the current grace period
247 * for the specified rcu_node structure. If the caller needs a reliable
248 * answer, it must hold the rcu_node's ->lock.
249 */
250static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
251{
252 return rnp->gp_tasks != NULL;
253}
254
255/*
256 * Record a quiescent state for all tasks that were previously queued
257 * on the specified rcu_node structure and that were blocking the current
258 * RCU grace period. The caller must hold the specified rnp->lock with
259 * irqs disabled, and this lock is released upon return, but irqs remain
260 * disabled.
261 */
262static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
263 __releases(rnp->lock)
264{
265 unsigned long mask;
266 struct rcu_node *rnp_p;
267
268 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
269 raw_spin_unlock_irqrestore(&rnp->lock, flags);
270 return; /* Still need more quiescent states! */
271 }
272
273 rnp_p = rnp->parent;
274 if (rnp_p == NULL) {
275 /*
276 * Either there is only one rcu_node in the tree,
277 * or tasks were kicked up to root rcu_node due to
278 * CPUs going offline.
279 */
280 rcu_report_qs_rsp(&rcu_preempt_state, flags);
281 return;
282 }
283
284 /* Report up the rest of the hierarchy. */
285 mask = rnp->grpmask;
286 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
287 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
288 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
289}
290
291/*
292 * Advance a ->blkd_tasks-list pointer to the next entry, instead
293 * returning NULL if at the end of the list.
294 */
295static struct list_head *rcu_next_node_entry(struct task_struct *t,
296 struct rcu_node *rnp)
297{
298 struct list_head *np;
299
300 np = t->rcu_node_entry.next;
301 if (np == &rnp->blkd_tasks)
302 np = NULL;
303 return np;
304}
305
306/*
307 * Handle special cases during rcu_read_unlock(), such as needing to
308 * notify RCU core processing or task having blocked during the RCU
309 * read-side critical section.
310 */
311static noinline void rcu_read_unlock_special(struct task_struct *t)
312{
313 int empty;
314 int empty_exp;
315 unsigned long flags;
316 struct list_head *np;
317#ifdef CONFIG_RCU_BOOST
318 struct rt_mutex *rbmp = NULL;
319#endif /* #ifdef CONFIG_RCU_BOOST */
320 struct rcu_node *rnp;
321 int special;
322
323 /* NMI handlers cannot block and cannot safely manipulate state. */
324 if (in_nmi())
325 return;
326
327 local_irq_save(flags);
328
329 /*
330 * If RCU core is waiting for this CPU to exit critical section,
331 * let it know that we have done so.
332 */
333 special = t->rcu_read_unlock_special;
334 if (special & RCU_READ_UNLOCK_NEED_QS) {
335 rcu_preempt_qs(smp_processor_id());
336 }
337
338 /* Hardware IRQ handlers cannot block. */
339 if (in_irq() || in_serving_softirq()) {
340 local_irq_restore(flags);
341 return;
342 }
343
344 /* Clean up if blocked during RCU read-side critical section. */
345 if (special & RCU_READ_UNLOCK_BLOCKED) {
346 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
347
348 /*
349 * Remove this task from the list it blocked on. The
350 * task can migrate while we acquire the lock, but at
351 * most one time. So at most two passes through loop.
352 */
353 for (;;) {
354 rnp = t->rcu_blocked_node;
355 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
356 if (rnp == t->rcu_blocked_node)
357 break;
358 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
359 }
360 empty = !rcu_preempt_blocked_readers_cgp(rnp);
361 empty_exp = !rcu_preempted_readers_exp(rnp);
362 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
363 np = rcu_next_node_entry(t, rnp);
364 list_del_init(&t->rcu_node_entry);
365 t->rcu_blocked_node = NULL;
366 trace_rcu_unlock_preempted_task("rcu_preempt",
367 rnp->gpnum, t->pid);
368 if (&t->rcu_node_entry == rnp->gp_tasks)
369 rnp->gp_tasks = np;
370 if (&t->rcu_node_entry == rnp->exp_tasks)
371 rnp->exp_tasks = np;
372#ifdef CONFIG_RCU_BOOST
373 if (&t->rcu_node_entry == rnp->boost_tasks)
374 rnp->boost_tasks = np;
375 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
376 if (t->rcu_boost_mutex) {
377 rbmp = t->rcu_boost_mutex;
378 t->rcu_boost_mutex = NULL;
379 }
380#endif /* #ifdef CONFIG_RCU_BOOST */
381
382 /*
383 * If this was the last task on the current list, and if
384 * we aren't waiting on any CPUs, report the quiescent state.
385 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
386 */
387 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
388 trace_rcu_quiescent_state_report("preempt_rcu",
389 rnp->gpnum,
390 0, rnp->qsmask,
391 rnp->level,
392 rnp->grplo,
393 rnp->grphi,
394 !!rnp->gp_tasks);
395 rcu_report_unblock_qs_rnp(rnp, flags);
396 } else
397 raw_spin_unlock_irqrestore(&rnp->lock, flags);
398
399#ifdef CONFIG_RCU_BOOST
400 /* Unboost if we were boosted. */
401 if (rbmp)
402 rt_mutex_unlock(rbmp);
403#endif /* #ifdef CONFIG_RCU_BOOST */
404
405 /*
406 * If this was the last task on the expedited lists,
407 * then we need to report up the rcu_node hierarchy.
408 */
409 if (!empty_exp && !rcu_preempted_readers_exp(rnp))
410 rcu_report_exp_rnp(&rcu_preempt_state, rnp);
411 } else {
412 local_irq_restore(flags);
413 }
414}
415
416/*
417 * Tree-preemptible RCU implementation for rcu_read_unlock().
418 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
419 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
420 * invoke rcu_read_unlock_special() to clean up after a context switch
421 * in an RCU read-side critical section and other special cases.
422 */
423void __rcu_read_unlock(void)
424{
425 struct task_struct *t = current;
426
427 if (t->rcu_read_lock_nesting != 1)
428 --t->rcu_read_lock_nesting;
429 else {
430 barrier(); /* critical section before exit code. */
431 t->rcu_read_lock_nesting = INT_MIN;
432 barrier(); /* assign before ->rcu_read_unlock_special load */
433 if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
434 rcu_read_unlock_special(t);
435 barrier(); /* ->rcu_read_unlock_special load before assign */
436 t->rcu_read_lock_nesting = 0;
437 }
438#ifdef CONFIG_PROVE_LOCKING
439 {
440 int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting);
441
442 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
443 }
444#endif /* #ifdef CONFIG_PROVE_LOCKING */
445}
446EXPORT_SYMBOL_GPL(__rcu_read_unlock);
447
448#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
449
450/*
451 * Dump detailed information for all tasks blocking the current RCU
452 * grace period on the specified rcu_node structure.
453 */
454static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
455{
456 unsigned long flags;
457 struct task_struct *t;
458
459 if (!rcu_preempt_blocked_readers_cgp(rnp))
460 return;
461 raw_spin_lock_irqsave(&rnp->lock, flags);
462 t = list_entry(rnp->gp_tasks,
463 struct task_struct, rcu_node_entry);
464 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
465 sched_show_task(t);
466 raw_spin_unlock_irqrestore(&rnp->lock, flags);
467}
468
469/*
470 * Dump detailed information for all tasks blocking the current RCU
471 * grace period.
472 */
473static void rcu_print_detail_task_stall(struct rcu_state *rsp)
474{
475 struct rcu_node *rnp = rcu_get_root(rsp);
476
477 rcu_print_detail_task_stall_rnp(rnp);
478 rcu_for_each_leaf_node(rsp, rnp)
479 rcu_print_detail_task_stall_rnp(rnp);
480}
481
482#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
483
484static void rcu_print_detail_task_stall(struct rcu_state *rsp)
485{
486}
487
488#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
489
490/*
491 * Scan the current list of tasks blocked within RCU read-side critical
492 * sections, printing out the tid of each.
493 */
494static int rcu_print_task_stall(struct rcu_node *rnp)
495{
496 struct task_struct *t;
497 int ndetected = 0;
498
499 if (!rcu_preempt_blocked_readers_cgp(rnp))
500 return 0;
501 t = list_entry(rnp->gp_tasks,
502 struct task_struct, rcu_node_entry);
503 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
504 printk(" P%d", t->pid);
505 ndetected++;
506 }
507 return ndetected;
508}
509
510/*
511 * Suppress preemptible RCU's CPU stall warnings by pushing the
512 * time of the next stall-warning message comfortably far into the
513 * future.
514 */
515static void rcu_preempt_stall_reset(void)
516{
517 rcu_preempt_state.jiffies_stall = jiffies + ULONG_MAX / 2;
518}
519
520/*
521 * Check that the list of blocked tasks for the newly completed grace
522 * period is in fact empty. It is a serious bug to complete a grace
523 * period that still has RCU readers blocked! This function must be
524 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
525 * must be held by the caller.
526 *
527 * Also, if there are blocked tasks on the list, they automatically
528 * block the newly created grace period, so set up ->gp_tasks accordingly.
529 */
530static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
531{
532 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
533 if (!list_empty(&rnp->blkd_tasks))
534 rnp->gp_tasks = rnp->blkd_tasks.next;
535 WARN_ON_ONCE(rnp->qsmask);
536}
537
538#ifdef CONFIG_HOTPLUG_CPU
539
540/*
541 * Handle tasklist migration for case in which all CPUs covered by the
542 * specified rcu_node have gone offline. Move them up to the root
543 * rcu_node. The reason for not just moving them to the immediate
544 * parent is to remove the need for rcu_read_unlock_special() to
545 * make more than two attempts to acquire the target rcu_node's lock.
546 * Returns true if there were tasks blocking the current RCU grace
547 * period.
548 *
549 * Returns 1 if there was previously a task blocking the current grace
550 * period on the specified rcu_node structure.
551 *
552 * The caller must hold rnp->lock with irqs disabled.
553 */
554static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
555 struct rcu_node *rnp,
556 struct rcu_data *rdp)
557{
558 struct list_head *lp;
559 struct list_head *lp_root;
560 int retval = 0;
561 struct rcu_node *rnp_root = rcu_get_root(rsp);
562 struct task_struct *t;
563
564 if (rnp == rnp_root) {
565 WARN_ONCE(1, "Last CPU thought to be offlined?");
566 return 0; /* Shouldn't happen: at least one CPU online. */
567 }
568
569 /* If we are on an internal node, complain bitterly. */
570 WARN_ON_ONCE(rnp != rdp->mynode);
571
572 /*
573 * Move tasks up to root rcu_node. Don't try to get fancy for
574 * this corner-case operation -- just put this node's tasks
575 * at the head of the root node's list, and update the root node's
576 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
577 * if non-NULL. This might result in waiting for more tasks than
578 * absolutely necessary, but this is a good performance/complexity
579 * tradeoff.
580 */
581 if (rcu_preempt_blocked_readers_cgp(rnp))
582 retval |= RCU_OFL_TASKS_NORM_GP;
583 if (rcu_preempted_readers_exp(rnp))
584 retval |= RCU_OFL_TASKS_EXP_GP;
585 lp = &rnp->blkd_tasks;
586 lp_root = &rnp_root->blkd_tasks;
587 while (!list_empty(lp)) {
588 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
589 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
590 list_del(&t->rcu_node_entry);
591 t->rcu_blocked_node = rnp_root;
592 list_add(&t->rcu_node_entry, lp_root);
593 if (&t->rcu_node_entry == rnp->gp_tasks)
594 rnp_root->gp_tasks = rnp->gp_tasks;
595 if (&t->rcu_node_entry == rnp->exp_tasks)
596 rnp_root->exp_tasks = rnp->exp_tasks;
597#ifdef CONFIG_RCU_BOOST
598 if (&t->rcu_node_entry == rnp->boost_tasks)
599 rnp_root->boost_tasks = rnp->boost_tasks;
600#endif /* #ifdef CONFIG_RCU_BOOST */
601 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
602 }
603
604#ifdef CONFIG_RCU_BOOST
605 /* In case root is being boosted and leaf is not. */
606 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
607 if (rnp_root->boost_tasks != NULL &&
608 rnp_root->boost_tasks != rnp_root->gp_tasks)
609 rnp_root->boost_tasks = rnp_root->gp_tasks;
610 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
611#endif /* #ifdef CONFIG_RCU_BOOST */
612
613 rnp->gp_tasks = NULL;
614 rnp->exp_tasks = NULL;
615 return retval;
616}
617
618/*
619 * Do CPU-offline processing for preemptible RCU.
620 */
621static void rcu_preempt_offline_cpu(int cpu)
622{
623 __rcu_offline_cpu(cpu, &rcu_preempt_state);
624}
625
626#endif /* #ifdef CONFIG_HOTPLUG_CPU */
627
628/*
629 * Check for a quiescent state from the current CPU. When a task blocks,
630 * the task is recorded in the corresponding CPU's rcu_node structure,
631 * which is checked elsewhere.
632 *
633 * Caller must disable hard irqs.
634 */
635static void rcu_preempt_check_callbacks(int cpu)
636{
637 struct task_struct *t = current;
638
639 if (t->rcu_read_lock_nesting == 0) {
640 rcu_preempt_qs(cpu);
641 return;
642 }
643 if (t->rcu_read_lock_nesting > 0 &&
644 per_cpu(rcu_preempt_data, cpu).qs_pending)
645 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
646}
647
648/*
649 * Process callbacks for preemptible RCU.
650 */
651static void rcu_preempt_process_callbacks(void)
652{
653 __rcu_process_callbacks(&rcu_preempt_state,
654 &__get_cpu_var(rcu_preempt_data));
655}
656
657#ifdef CONFIG_RCU_BOOST
658
659static void rcu_preempt_do_callbacks(void)
660{
661 rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
662}
663
664#endif /* #ifdef CONFIG_RCU_BOOST */
665
666/*
667 * Queue a preemptible-RCU callback for invocation after a grace period.
668 */
669void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
670{
671 __call_rcu(head, func, &rcu_preempt_state);
672}
673EXPORT_SYMBOL_GPL(call_rcu);
674
675/**
676 * synchronize_rcu - wait until a grace period has elapsed.
677 *
678 * Control will return to the caller some time after a full grace
679 * period has elapsed, in other words after all currently executing RCU
680 * read-side critical sections have completed. Note, however, that
681 * upon return from synchronize_rcu(), the caller might well be executing
682 * concurrently with new RCU read-side critical sections that began while
683 * synchronize_rcu() was waiting. RCU read-side critical sections are
684 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
685 */
686void synchronize_rcu(void)
687{
688 if (!rcu_scheduler_active)
689 return;
690 wait_rcu_gp(call_rcu);
691}
692EXPORT_SYMBOL_GPL(synchronize_rcu);
693
694static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
695static long sync_rcu_preempt_exp_count;
696static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
697
698/*
699 * Return non-zero if there are any tasks in RCU read-side critical
700 * sections blocking the current preemptible-RCU expedited grace period.
701 * If there is no preemptible-RCU expedited grace period currently in
702 * progress, returns zero unconditionally.
703 */
704static int rcu_preempted_readers_exp(struct rcu_node *rnp)
705{
706 return rnp->exp_tasks != NULL;
707}
708
709/*
710 * return non-zero if there is no RCU expedited grace period in progress
711 * for the specified rcu_node structure, in other words, if all CPUs and
712 * tasks covered by the specified rcu_node structure have done their bit
713 * for the current expedited grace period. Works only for preemptible
714 * RCU -- other RCU implementation use other means.
715 *
716 * Caller must hold sync_rcu_preempt_exp_mutex.
717 */
718static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
719{
720 return !rcu_preempted_readers_exp(rnp) &&
721 ACCESS_ONCE(rnp->expmask) == 0;
722}
723
724/*
725 * Report the exit from RCU read-side critical section for the last task
726 * that queued itself during or before the current expedited preemptible-RCU
727 * grace period. This event is reported either to the rcu_node structure on
728 * which the task was queued or to one of that rcu_node structure's ancestors,
729 * recursively up the tree. (Calm down, calm down, we do the recursion
730 * iteratively!)
731 *
732 * Caller must hold sync_rcu_preempt_exp_mutex.
733 */
734static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
735{
736 unsigned long flags;
737 unsigned long mask;
738
739 raw_spin_lock_irqsave(&rnp->lock, flags);
740 for (;;) {
741 if (!sync_rcu_preempt_exp_done(rnp)) {
742 raw_spin_unlock_irqrestore(&rnp->lock, flags);
743 break;
744 }
745 if (rnp->parent == NULL) {
746 raw_spin_unlock_irqrestore(&rnp->lock, flags);
747 wake_up(&sync_rcu_preempt_exp_wq);
748 break;
749 }
750 mask = rnp->grpmask;
751 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
752 rnp = rnp->parent;
753 raw_spin_lock(&rnp->lock); /* irqs already disabled */
754 rnp->expmask &= ~mask;
755 }
756}
757
758/*
759 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
760 * grace period for the specified rcu_node structure. If there are no such
761 * tasks, report it up the rcu_node hierarchy.
762 *
763 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
764 */
765static void
766sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
767{
768 unsigned long flags;
769 int must_wait = 0;
770
771 raw_spin_lock_irqsave(&rnp->lock, flags);
772 if (list_empty(&rnp->blkd_tasks))
773 raw_spin_unlock_irqrestore(&rnp->lock, flags);
774 else {
775 rnp->exp_tasks = rnp->blkd_tasks.next;
776 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
777 must_wait = 1;
778 }
779 if (!must_wait)
780 rcu_report_exp_rnp(rsp, rnp);
781}
782
783/*
784 * Wait for an rcu-preempt grace period, but expedite it. The basic idea
785 * is to invoke synchronize_sched_expedited() to push all the tasks to
786 * the ->blkd_tasks lists and wait for this list to drain.
787 */
788void synchronize_rcu_expedited(void)
789{
790 unsigned long flags;
791 struct rcu_node *rnp;
792 struct rcu_state *rsp = &rcu_preempt_state;
793 long snap;
794 int trycount = 0;
795
796 smp_mb(); /* Caller's modifications seen first by other CPUs. */
797 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
798 smp_mb(); /* Above access cannot bleed into critical section. */
799
800 /*
801 * Acquire lock, falling back to synchronize_rcu() if too many
802 * lock-acquisition failures. Of course, if someone does the
803 * expedited grace period for us, just leave.
804 */
805 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
806 if (trycount++ < 10)
807 udelay(trycount * num_online_cpus());
808 else {
809 synchronize_rcu();
810 return;
811 }
812 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
813 goto mb_ret; /* Others did our work for us. */
814 }
815 if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
816 goto unlock_mb_ret; /* Others did our work for us. */
817
818 /* force all RCU readers onto ->blkd_tasks lists. */
819 synchronize_sched_expedited();
820
821 raw_spin_lock_irqsave(&rsp->onofflock, flags);
822
823 /* Initialize ->expmask for all non-leaf rcu_node structures. */
824 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
825 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
826 rnp->expmask = rnp->qsmaskinit;
827 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
828 }
829
830 /* Snapshot current state of ->blkd_tasks lists. */
831 rcu_for_each_leaf_node(rsp, rnp)
832 sync_rcu_preempt_exp_init(rsp, rnp);
833 if (NUM_RCU_NODES > 1)
834 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
835
836 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
837
838 /* Wait for snapshotted ->blkd_tasks lists to drain. */
839 rnp = rcu_get_root(rsp);
840 wait_event(sync_rcu_preempt_exp_wq,
841 sync_rcu_preempt_exp_done(rnp));
842
843 /* Clean up and exit. */
844 smp_mb(); /* ensure expedited GP seen before counter increment. */
845 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
846unlock_mb_ret:
847 mutex_unlock(&sync_rcu_preempt_exp_mutex);
848mb_ret:
849 smp_mb(); /* ensure subsequent action seen after grace period. */
850}
851EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
852
853/*
854 * Check to see if there is any immediate preemptible-RCU-related work
855 * to be done.
856 */
857static int rcu_preempt_pending(int cpu)
858{
859 return __rcu_pending(&rcu_preempt_state,
860 &per_cpu(rcu_preempt_data, cpu));
861}
862
863/*
864 * Does preemptible RCU need the CPU to stay out of dynticks mode?
865 */
866static int rcu_preempt_needs_cpu(int cpu)
867{
868 return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
869}
870
871/**
872 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
873 */
874void rcu_barrier(void)
875{
876 _rcu_barrier(&rcu_preempt_state, call_rcu);
877}
878EXPORT_SYMBOL_GPL(rcu_barrier);
879
880/*
881 * Initialize preemptible RCU's per-CPU data.
882 */
883static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
884{
885 rcu_init_percpu_data(cpu, &rcu_preempt_state, 1);
886}
887
888/*
889 * Move preemptible RCU's callbacks from dying CPU to other online CPU.
890 */
891static void rcu_preempt_send_cbs_to_online(void)
892{
893 rcu_send_cbs_to_online(&rcu_preempt_state);
894}
895
896/*
897 * Initialize preemptible RCU's state structures.
898 */
899static void __init __rcu_init_preempt(void)
900{
901 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
902}
903
904/*
905 * Check for a task exiting while in a preemptible-RCU read-side
906 * critical section, clean up if so. No need to issue warnings,
907 * as debug_check_no_locks_held() already does this if lockdep
908 * is enabled.
909 */
910void exit_rcu(void)
911{
912 struct task_struct *t = current;
913
914 if (t->rcu_read_lock_nesting == 0)
915 return;
916 t->rcu_read_lock_nesting = 1;
917 __rcu_read_unlock();
918}
919
920#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
921
922static struct rcu_state *rcu_state = &rcu_sched_state;
923
924/*
925 * Tell them what RCU they are running.
926 */
927static void __init rcu_bootup_announce(void)
928{
929 printk(KERN_INFO "Hierarchical RCU implementation.\n");
930 rcu_bootup_announce_oddness();
931}
932
933/*
934 * Return the number of RCU batches processed thus far for debug & stats.
935 */
936long rcu_batches_completed(void)
937{
938 return rcu_batches_completed_sched();
939}
940EXPORT_SYMBOL_GPL(rcu_batches_completed);
941
942/*
943 * Force a quiescent state for RCU, which, because there is no preemptible
944 * RCU, becomes the same as rcu-sched.
945 */
946void rcu_force_quiescent_state(void)
947{
948 rcu_sched_force_quiescent_state();
949}
950EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
951
952/*
953 * Because preemptible RCU does not exist, we never have to check for
954 * CPUs being in quiescent states.
955 */
956static void rcu_preempt_note_context_switch(int cpu)
957{
958}
959
960/*
961 * Because preemptible RCU does not exist, there are never any preempted
962 * RCU readers.
963 */
964static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
965{
966 return 0;
967}
968
969#ifdef CONFIG_HOTPLUG_CPU
970
971/* Because preemptible RCU does not exist, no quieting of tasks. */
972static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
973{
974 raw_spin_unlock_irqrestore(&rnp->lock, flags);
975}
976
977#endif /* #ifdef CONFIG_HOTPLUG_CPU */
978
979/*
980 * Because preemptible RCU does not exist, we never have to check for
981 * tasks blocked within RCU read-side critical sections.
982 */
983static void rcu_print_detail_task_stall(struct rcu_state *rsp)
984{
985}
986
987/*
988 * Because preemptible RCU does not exist, we never have to check for
989 * tasks blocked within RCU read-side critical sections.
990 */
991static int rcu_print_task_stall(struct rcu_node *rnp)
992{
993 return 0;
994}
995
996/*
997 * Because preemptible RCU does not exist, there is no need to suppress
998 * its CPU stall warnings.
999 */
1000static void rcu_preempt_stall_reset(void)
1001{
1002}
1003
1004/*
1005 * Because there is no preemptible RCU, there can be no readers blocked,
1006 * so there is no need to check for blocked tasks. So check only for
1007 * bogus qsmask values.
1008 */
1009static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1010{
1011 WARN_ON_ONCE(rnp->qsmask);
1012}
1013
1014#ifdef CONFIG_HOTPLUG_CPU
1015
1016/*
1017 * Because preemptible RCU does not exist, it never needs to migrate
1018 * tasks that were blocked within RCU read-side critical sections, and
1019 * such non-existent tasks cannot possibly have been blocking the current
1020 * grace period.
1021 */
1022static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1023 struct rcu_node *rnp,
1024 struct rcu_data *rdp)
1025{
1026 return 0;
1027}
1028
1029/*
1030 * Because preemptible RCU does not exist, it never needs CPU-offline
1031 * processing.
1032 */
1033static void rcu_preempt_offline_cpu(int cpu)
1034{
1035}
1036
1037#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1038
1039/*
1040 * Because preemptible RCU does not exist, it never has any callbacks
1041 * to check.
1042 */
1043static void rcu_preempt_check_callbacks(int cpu)
1044{
1045}
1046
1047/*
1048 * Because preemptible RCU does not exist, it never has any callbacks
1049 * to process.
1050 */
1051static void rcu_preempt_process_callbacks(void)
1052{
1053}
1054
1055/*
1056 * Wait for an rcu-preempt grace period, but make it happen quickly.
1057 * But because preemptible RCU does not exist, map to rcu-sched.
1058 */
1059void synchronize_rcu_expedited(void)
1060{
1061 synchronize_sched_expedited();
1062}
1063EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1064
1065#ifdef CONFIG_HOTPLUG_CPU
1066
1067/*
1068 * Because preemptible RCU does not exist, there is never any need to
1069 * report on tasks preempted in RCU read-side critical sections during
1070 * expedited RCU grace periods.
1071 */
1072static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
1073{
1074 return;
1075}
1076
1077#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1078
1079/*
1080 * Because preemptible RCU does not exist, it never has any work to do.
1081 */
1082static int rcu_preempt_pending(int cpu)
1083{
1084 return 0;
1085}
1086
1087/*
1088 * Because preemptible RCU does not exist, it never needs any CPU.
1089 */
1090static int rcu_preempt_needs_cpu(int cpu)
1091{
1092 return 0;
1093}
1094
1095/*
1096 * Because preemptible RCU does not exist, rcu_barrier() is just
1097 * another name for rcu_barrier_sched().
1098 */
1099void rcu_barrier(void)
1100{
1101 rcu_barrier_sched();
1102}
1103EXPORT_SYMBOL_GPL(rcu_barrier);
1104
1105/*
1106 * Because preemptible RCU does not exist, there is no per-CPU
1107 * data to initialize.
1108 */
1109static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
1110{
1111}
1112
1113/*
1114 * Because there is no preemptible RCU, there are no callbacks to move.
1115 */
1116static void rcu_preempt_send_cbs_to_online(void)
1117{
1118}
1119
1120/*
1121 * Because preemptible RCU does not exist, it need not be initialized.
1122 */
1123static void __init __rcu_init_preempt(void)
1124{
1125}
1126
1127#endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1128
1129#ifdef CONFIG_RCU_BOOST
1130
1131#include "rtmutex_common.h"
1132
1133#ifdef CONFIG_RCU_TRACE
1134
1135static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1136{
1137 if (list_empty(&rnp->blkd_tasks))
1138 rnp->n_balk_blkd_tasks++;
1139 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1140 rnp->n_balk_exp_gp_tasks++;
1141 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1142 rnp->n_balk_boost_tasks++;
1143 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1144 rnp->n_balk_notblocked++;
1145 else if (rnp->gp_tasks != NULL &&
1146 ULONG_CMP_LT(jiffies, rnp->boost_time))
1147 rnp->n_balk_notyet++;
1148 else
1149 rnp->n_balk_nos++;
1150}
1151
1152#else /* #ifdef CONFIG_RCU_TRACE */
1153
1154static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1155{
1156}
1157
1158#endif /* #else #ifdef CONFIG_RCU_TRACE */
1159
1160static struct lock_class_key rcu_boost_class;
1161
1162/*
1163 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1164 * or ->boost_tasks, advancing the pointer to the next task in the
1165 * ->blkd_tasks list.
1166 *
1167 * Note that irqs must be enabled: boosting the task can block.
1168 * Returns 1 if there are more tasks needing to be boosted.
1169 */
1170static int rcu_boost(struct rcu_node *rnp)
1171{
1172 unsigned long flags;
1173 struct rt_mutex mtx;
1174 struct task_struct *t;
1175 struct list_head *tb;
1176
1177 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1178 return 0; /* Nothing left to boost. */
1179
1180 raw_spin_lock_irqsave(&rnp->lock, flags);
1181
1182 /*
1183 * Recheck under the lock: all tasks in need of boosting
1184 * might exit their RCU read-side critical sections on their own.
1185 */
1186 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1187 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1188 return 0;
1189 }
1190
1191 /*
1192 * Preferentially boost tasks blocking expedited grace periods.
1193 * This cannot starve the normal grace periods because a second
1194 * expedited grace period must boost all blocked tasks, including
1195 * those blocking the pre-existing normal grace period.
1196 */
1197 if (rnp->exp_tasks != NULL) {
1198 tb = rnp->exp_tasks;
1199 rnp->n_exp_boosts++;
1200 } else {
1201 tb = rnp->boost_tasks;
1202 rnp->n_normal_boosts++;
1203 }
1204 rnp->n_tasks_boosted++;
1205
1206 /*
1207 * We boost task t by manufacturing an rt_mutex that appears to
1208 * be held by task t. We leave a pointer to that rt_mutex where
1209 * task t can find it, and task t will release the mutex when it
1210 * exits its outermost RCU read-side critical section. Then
1211 * simply acquiring this artificial rt_mutex will boost task
1212 * t's priority. (Thanks to tglx for suggesting this approach!)
1213 *
1214 * Note that task t must acquire rnp->lock to remove itself from
1215 * the ->blkd_tasks list, which it will do from exit() if from
1216 * nowhere else. We therefore are guaranteed that task t will
1217 * stay around at least until we drop rnp->lock. Note that
1218 * rnp->lock also resolves races between our priority boosting
1219 * and task t's exiting its outermost RCU read-side critical
1220 * section.
1221 */
1222 t = container_of(tb, struct task_struct, rcu_node_entry);
1223 rt_mutex_init_proxy_locked(&mtx, t);
1224 /* Avoid lockdep false positives. This rt_mutex is its own thing. */
1225 lockdep_set_class_and_name(&mtx.wait_lock, &rcu_boost_class,
1226 "rcu_boost_mutex");
1227 t->rcu_boost_mutex = &mtx;
1228 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1229 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1230 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1231
1232 return rnp->exp_tasks != NULL || rnp->boost_tasks != NULL;
1233}
1234
1235/*
1236 * Timer handler to initiate waking up of boost kthreads that
1237 * have yielded the CPU due to excessive numbers of tasks to
1238 * boost. We wake up the per-rcu_node kthread, which in turn
1239 * will wake up the booster kthread.
1240 */
1241static void rcu_boost_kthread_timer(unsigned long arg)
1242{
1243 invoke_rcu_node_kthread((struct rcu_node *)arg);
1244}
1245
1246/*
1247 * Priority-boosting kthread. One per leaf rcu_node and one for the
1248 * root rcu_node.
1249 */
1250static int rcu_boost_kthread(void *arg)
1251{
1252 struct rcu_node *rnp = (struct rcu_node *)arg;
1253 int spincnt = 0;
1254 int more2boost;
1255
1256 trace_rcu_utilization("Start boost kthread@init");
1257 for (;;) {
1258 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1259 trace_rcu_utilization("End boost kthread@rcu_wait");
1260 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1261 trace_rcu_utilization("Start boost kthread@rcu_wait");
1262 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1263 more2boost = rcu_boost(rnp);
1264 if (more2boost)
1265 spincnt++;
1266 else
1267 spincnt = 0;
1268 if (spincnt > 10) {
1269 trace_rcu_utilization("End boost kthread@rcu_yield");
1270 rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp);
1271 trace_rcu_utilization("Start boost kthread@rcu_yield");
1272 spincnt = 0;
1273 }
1274 }
1275 /* NOTREACHED */
1276 trace_rcu_utilization("End boost kthread@notreached");
1277 return 0;
1278}
1279
1280/*
1281 * Check to see if it is time to start boosting RCU readers that are
1282 * blocking the current grace period, and, if so, tell the per-rcu_node
1283 * kthread to start boosting them. If there is an expedited grace
1284 * period in progress, it is always time to boost.
1285 *
1286 * The caller must hold rnp->lock, which this function releases,
1287 * but irqs remain disabled. The ->boost_kthread_task is immortal,
1288 * so we don't need to worry about it going away.
1289 */
1290static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1291{
1292 struct task_struct *t;
1293
1294 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1295 rnp->n_balk_exp_gp_tasks++;
1296 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1297 return;
1298 }
1299 if (rnp->exp_tasks != NULL ||
1300 (rnp->gp_tasks != NULL &&
1301 rnp->boost_tasks == NULL &&
1302 rnp->qsmask == 0 &&
1303 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1304 if (rnp->exp_tasks == NULL)
1305 rnp->boost_tasks = rnp->gp_tasks;
1306 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1307 t = rnp->boost_kthread_task;
1308 if (t != NULL)
1309 wake_up_process(t);
1310 } else {
1311 rcu_initiate_boost_trace(rnp);
1312 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1313 }
1314}
1315
1316/*
1317 * Wake up the per-CPU kthread to invoke RCU callbacks.
1318 */
1319static void invoke_rcu_callbacks_kthread(void)
1320{
1321 unsigned long flags;
1322
1323 local_irq_save(flags);
1324 __this_cpu_write(rcu_cpu_has_work, 1);
1325 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1326 current != __this_cpu_read(rcu_cpu_kthread_task))
1327 wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
1328 local_irq_restore(flags);
1329}
1330
1331/*
1332 * Set the affinity of the boost kthread. The CPU-hotplug locks are
1333 * held, so no one should be messing with the existence of the boost
1334 * kthread.
1335 */
1336static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
1337 cpumask_var_t cm)
1338{
1339 struct task_struct *t;
1340
1341 t = rnp->boost_kthread_task;
1342 if (t != NULL)
1343 set_cpus_allowed_ptr(rnp->boost_kthread_task, cm);
1344}
1345
1346#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1347
1348/*
1349 * Do priority-boost accounting for the start of a new grace period.
1350 */
1351static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1352{
1353 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1354}
1355
1356/*
1357 * Create an RCU-boost kthread for the specified node if one does not
1358 * already exist. We only create this kthread for preemptible RCU.
1359 * Returns zero if all is well, a negated errno otherwise.
1360 */
1361static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1362 struct rcu_node *rnp,
1363 int rnp_index)
1364{
1365 unsigned long flags;
1366 struct sched_param sp;
1367 struct task_struct *t;
1368
1369 if (&rcu_preempt_state != rsp)
1370 return 0;
1371 rsp->boost = 1;
1372 if (rnp->boost_kthread_task != NULL)
1373 return 0;
1374 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1375 "rcub/%d", rnp_index);
1376 if (IS_ERR(t))
1377 return PTR_ERR(t);
1378 raw_spin_lock_irqsave(&rnp->lock, flags);
1379 rnp->boost_kthread_task = t;
1380 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1381 sp.sched_priority = RCU_BOOST_PRIO;
1382 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1383 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1384 return 0;
1385}
1386
1387#ifdef CONFIG_HOTPLUG_CPU
1388
1389/*
1390 * Stop the RCU's per-CPU kthread when its CPU goes offline,.
1391 */
1392static void rcu_stop_cpu_kthread(int cpu)
1393{
1394 struct task_struct *t;
1395
1396 /* Stop the CPU's kthread. */
1397 t = per_cpu(rcu_cpu_kthread_task, cpu);
1398 if (t != NULL) {
1399 per_cpu(rcu_cpu_kthread_task, cpu) = NULL;
1400 kthread_stop(t);
1401 }
1402}
1403
1404#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1405
1406static void rcu_kthread_do_work(void)
1407{
1408 rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1409 rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1410 rcu_preempt_do_callbacks();
1411}
1412
1413/*
1414 * Wake up the specified per-rcu_node-structure kthread.
1415 * Because the per-rcu_node kthreads are immortal, we don't need
1416 * to do anything to keep them alive.
1417 */
1418static void invoke_rcu_node_kthread(struct rcu_node *rnp)
1419{
1420 struct task_struct *t;
1421
1422 t = rnp->node_kthread_task;
1423 if (t != NULL)
1424 wake_up_process(t);
1425}
1426
1427/*
1428 * Set the specified CPU's kthread to run RT or not, as specified by
1429 * the to_rt argument. The CPU-hotplug locks are held, so the task
1430 * is not going away.
1431 */
1432static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1433{
1434 int policy;
1435 struct sched_param sp;
1436 struct task_struct *t;
1437
1438 t = per_cpu(rcu_cpu_kthread_task, cpu);
1439 if (t == NULL)
1440 return;
1441 if (to_rt) {
1442 policy = SCHED_FIFO;
1443 sp.sched_priority = RCU_KTHREAD_PRIO;
1444 } else {
1445 policy = SCHED_NORMAL;
1446 sp.sched_priority = 0;
1447 }
1448 sched_setscheduler_nocheck(t, policy, &sp);
1449}
1450
1451/*
1452 * Timer handler to initiate the waking up of per-CPU kthreads that
1453 * have yielded the CPU due to excess numbers of RCU callbacks.
1454 * We wake up the per-rcu_node kthread, which in turn will wake up
1455 * the booster kthread.
1456 */
1457static void rcu_cpu_kthread_timer(unsigned long arg)
1458{
1459 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
1460 struct rcu_node *rnp = rdp->mynode;
1461
1462 atomic_or(rdp->grpmask, &rnp->wakemask);
1463 invoke_rcu_node_kthread(rnp);
1464}
1465
1466/*
1467 * Drop to non-real-time priority and yield, but only after posting a
1468 * timer that will cause us to regain our real-time priority if we
1469 * remain preempted. Either way, we restore our real-time priority
1470 * before returning.
1471 */
1472static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
1473{
1474 struct sched_param sp;
1475 struct timer_list yield_timer;
1476 int prio = current->rt_priority;
1477
1478 setup_timer_on_stack(&yield_timer, f, arg);
1479 mod_timer(&yield_timer, jiffies + 2);
1480 sp.sched_priority = 0;
1481 sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
1482 set_user_nice(current, 19);
1483 schedule();
1484 set_user_nice(current, 0);
1485 sp.sched_priority = prio;
1486 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1487 del_timer(&yield_timer);
1488}
1489
1490/*
1491 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1492 * This can happen while the corresponding CPU is either coming online
1493 * or going offline. We cannot wait until the CPU is fully online
1494 * before starting the kthread, because the various notifier functions
1495 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1496 * the corresponding CPU is online.
1497 *
1498 * Return 1 if the kthread needs to stop, 0 otherwise.
1499 *
1500 * Caller must disable bh. This function can momentarily enable it.
1501 */
1502static int rcu_cpu_kthread_should_stop(int cpu)
1503{
1504 while (cpu_is_offline(cpu) ||
1505 !cpumask_equal(&current->cpus_allowed, cpumask_of(cpu)) ||
1506 smp_processor_id() != cpu) {
1507 if (kthread_should_stop())
1508 return 1;
1509 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1510 per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id();
1511 local_bh_enable();
1512 schedule_timeout_uninterruptible(1);
1513 if (!cpumask_equal(&current->cpus_allowed, cpumask_of(cpu)))
1514 set_cpus_allowed_ptr(current, cpumask_of(cpu));
1515 local_bh_disable();
1516 }
1517 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1518 return 0;
1519}
1520
1521/*
1522 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1523 * RCU softirq used in flavors and configurations of RCU that do not
1524 * support RCU priority boosting.
1525 */
1526static int rcu_cpu_kthread(void *arg)
1527{
1528 int cpu = (int)(long)arg;
1529 unsigned long flags;
1530 int spincnt = 0;
1531 unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu);
1532 char work;
1533 char *workp = &per_cpu(rcu_cpu_has_work, cpu);
1534
1535 trace_rcu_utilization("Start CPU kthread@init");
1536 for (;;) {
1537 *statusp = RCU_KTHREAD_WAITING;
1538 trace_rcu_utilization("End CPU kthread@rcu_wait");
1539 rcu_wait(*workp != 0 || kthread_should_stop());
1540 trace_rcu_utilization("Start CPU kthread@rcu_wait");
1541 local_bh_disable();
1542 if (rcu_cpu_kthread_should_stop(cpu)) {
1543 local_bh_enable();
1544 break;
1545 }
1546 *statusp = RCU_KTHREAD_RUNNING;
1547 per_cpu(rcu_cpu_kthread_loops, cpu)++;
1548 local_irq_save(flags);
1549 work = *workp;
1550 *workp = 0;
1551 local_irq_restore(flags);
1552 if (work)
1553 rcu_kthread_do_work();
1554 local_bh_enable();
1555 if (*workp != 0)
1556 spincnt++;
1557 else
1558 spincnt = 0;
1559 if (spincnt > 10) {
1560 *statusp = RCU_KTHREAD_YIELDING;
1561 trace_rcu_utilization("End CPU kthread@rcu_yield");
1562 rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
1563 trace_rcu_utilization("Start CPU kthread@rcu_yield");
1564 spincnt = 0;
1565 }
1566 }
1567 *statusp = RCU_KTHREAD_STOPPED;
1568 trace_rcu_utilization("End CPU kthread@term");
1569 return 0;
1570}
1571
1572/*
1573 * Spawn a per-CPU kthread, setting up affinity and priority.
1574 * Because the CPU hotplug lock is held, no other CPU will be attempting
1575 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1576 * attempting to access it during boot, but the locking in kthread_bind()
1577 * will enforce sufficient ordering.
1578 *
1579 * Please note that we cannot simply refuse to wake up the per-CPU
1580 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1581 * which can result in softlockup complaints if the task ends up being
1582 * idle for more than a couple of minutes.
1583 *
1584 * However, please note also that we cannot bind the per-CPU kthread to its
1585 * CPU until that CPU is fully online. We also cannot wait until the
1586 * CPU is fully online before we create its per-CPU kthread, as this would
1587 * deadlock the system when CPU notifiers tried waiting for grace
1588 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1589 * is online. If its CPU is not yet fully online, then the code in
1590 * rcu_cpu_kthread() will wait until it is fully online, and then do
1591 * the binding.
1592 */
1593static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu)
1594{
1595 struct sched_param sp;
1596 struct task_struct *t;
1597
1598 if (!rcu_scheduler_fully_active ||
1599 per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
1600 return 0;
1601 t = kthread_create_on_node(rcu_cpu_kthread,
1602 (void *)(long)cpu,
1603 cpu_to_node(cpu),
1604 "rcuc/%d", cpu);
1605 if (IS_ERR(t))
1606 return PTR_ERR(t);
1607 if (cpu_online(cpu))
1608 kthread_bind(t, cpu);
1609 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1610 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL);
1611 sp.sched_priority = RCU_KTHREAD_PRIO;
1612 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1613 per_cpu(rcu_cpu_kthread_task, cpu) = t;
1614 wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */
1615 return 0;
1616}
1617
1618/*
1619 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1620 * kthreads when needed. We ignore requests to wake up kthreads
1621 * for offline CPUs, which is OK because force_quiescent_state()
1622 * takes care of this case.
1623 */
1624static int rcu_node_kthread(void *arg)
1625{
1626 int cpu;
1627 unsigned long flags;
1628 unsigned long mask;
1629 struct rcu_node *rnp = (struct rcu_node *)arg;
1630 struct sched_param sp;
1631 struct task_struct *t;
1632
1633 for (;;) {
1634 rnp->node_kthread_status = RCU_KTHREAD_WAITING;
1635 rcu_wait(atomic_read(&rnp->wakemask) != 0);
1636 rnp->node_kthread_status = RCU_KTHREAD_RUNNING;
1637 raw_spin_lock_irqsave(&rnp->lock, flags);
1638 mask = atomic_xchg(&rnp->wakemask, 0);
1639 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1640 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
1641 if ((mask & 0x1) == 0)
1642 continue;
1643 preempt_disable();
1644 t = per_cpu(rcu_cpu_kthread_task, cpu);
1645 if (!cpu_online(cpu) || t == NULL) {
1646 preempt_enable();
1647 continue;
1648 }
1649 per_cpu(rcu_cpu_has_work, cpu) = 1;
1650 sp.sched_priority = RCU_KTHREAD_PRIO;
1651 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1652 preempt_enable();
1653 }
1654 }
1655 /* NOTREACHED */
1656 rnp->node_kthread_status = RCU_KTHREAD_STOPPED;
1657 return 0;
1658}
1659
1660/*
1661 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1662 * served by the rcu_node in question. The CPU hotplug lock is still
1663 * held, so the value of rnp->qsmaskinit will be stable.
1664 *
1665 * We don't include outgoingcpu in the affinity set, use -1 if there is
1666 * no outgoing CPU. If there are no CPUs left in the affinity set,
1667 * this function allows the kthread to execute on any CPU.
1668 */
1669static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1670{
1671 cpumask_var_t cm;
1672 int cpu;
1673 unsigned long mask = rnp->qsmaskinit;
1674
1675 if (rnp->node_kthread_task == NULL)
1676 return;
1677 if (!alloc_cpumask_var(&cm, GFP_KERNEL))
1678 return;
1679 cpumask_clear(cm);
1680 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1681 if ((mask & 0x1) && cpu != outgoingcpu)
1682 cpumask_set_cpu(cpu, cm);
1683 if (cpumask_weight(cm) == 0) {
1684 cpumask_setall(cm);
1685 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1686 cpumask_clear_cpu(cpu, cm);
1687 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1688 }
1689 set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
1690 rcu_boost_kthread_setaffinity(rnp, cm);
1691 free_cpumask_var(cm);
1692}
1693
1694/*
1695 * Spawn a per-rcu_node kthread, setting priority and affinity.
1696 * Called during boot before online/offline can happen, or, if
1697 * during runtime, with the main CPU-hotplug locks held. So only
1698 * one of these can be executing at a time.
1699 */
1700static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
1701 struct rcu_node *rnp)
1702{
1703 unsigned long flags;
1704 int rnp_index = rnp - &rsp->node[0];
1705 struct sched_param sp;
1706 struct task_struct *t;
1707
1708 if (!rcu_scheduler_fully_active ||
1709 rnp->qsmaskinit == 0)
1710 return 0;
1711 if (rnp->node_kthread_task == NULL) {
1712 t = kthread_create(rcu_node_kthread, (void *)rnp,
1713 "rcun/%d", rnp_index);
1714 if (IS_ERR(t))
1715 return PTR_ERR(t);
1716 raw_spin_lock_irqsave(&rnp->lock, flags);
1717 rnp->node_kthread_task = t;
1718 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1719 sp.sched_priority = 99;
1720 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1721 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1722 }
1723 return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
1724}
1725
1726/*
1727 * Spawn all kthreads -- called as soon as the scheduler is running.
1728 */
1729static int __init rcu_spawn_kthreads(void)
1730{
1731 int cpu;
1732 struct rcu_node *rnp;
1733
1734 rcu_scheduler_fully_active = 1;
1735 for_each_possible_cpu(cpu) {
1736 per_cpu(rcu_cpu_has_work, cpu) = 0;
1737 if (cpu_online(cpu))
1738 (void)rcu_spawn_one_cpu_kthread(cpu);
1739 }
1740 rnp = rcu_get_root(rcu_state);
1741 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1742 if (NUM_RCU_NODES > 1) {
1743 rcu_for_each_leaf_node(rcu_state, rnp)
1744 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1745 }
1746 return 0;
1747}
1748early_initcall(rcu_spawn_kthreads);
1749
1750static void __cpuinit rcu_prepare_kthreads(int cpu)
1751{
1752 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1753 struct rcu_node *rnp = rdp->mynode;
1754
1755 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1756 if (rcu_scheduler_fully_active) {
1757 (void)rcu_spawn_one_cpu_kthread(cpu);
1758 if (rnp->node_kthread_task == NULL)
1759 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1760 }
1761}
1762
1763#else /* #ifdef CONFIG_RCU_BOOST */
1764
1765static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1766{
1767 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1768}
1769
1770static void invoke_rcu_callbacks_kthread(void)
1771{
1772 WARN_ON_ONCE(1);
1773}
1774
1775static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1776{
1777}
1778
1779#ifdef CONFIG_HOTPLUG_CPU
1780
1781static void rcu_stop_cpu_kthread(int cpu)
1782{
1783}
1784
1785#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1786
1787static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1788{
1789}
1790
1791static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1792{
1793}
1794
1795static int __init rcu_scheduler_really_started(void)
1796{
1797 rcu_scheduler_fully_active = 1;
1798 return 0;
1799}
1800early_initcall(rcu_scheduler_really_started);
1801
1802static void __cpuinit rcu_prepare_kthreads(int cpu)
1803{
1804}
1805
1806#endif /* #else #ifdef CONFIG_RCU_BOOST */
1807
1808#ifndef CONFIG_SMP
1809
1810void synchronize_sched_expedited(void)
1811{
1812 cond_resched();
1813}
1814EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
1815
1816#else /* #ifndef CONFIG_SMP */
1817
1818static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
1819static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
1820
1821static int synchronize_sched_expedited_cpu_stop(void *data)
1822{
1823 /*
1824 * There must be a full memory barrier on each affected CPU
1825 * between the time that try_stop_cpus() is called and the
1826 * time that it returns.
1827 *
1828 * In the current initial implementation of cpu_stop, the
1829 * above condition is already met when the control reaches
1830 * this point and the following smp_mb() is not strictly
1831 * necessary. Do smp_mb() anyway for documentation and
1832 * robustness against future implementation changes.
1833 */
1834 smp_mb(); /* See above comment block. */
1835 return 0;
1836}
1837
1838/*
1839 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
1840 * approach to force grace period to end quickly. This consumes
1841 * significant time on all CPUs, and is thus not recommended for
1842 * any sort of common-case code.
1843 *
1844 * Note that it is illegal to call this function while holding any
1845 * lock that is acquired by a CPU-hotplug notifier. Failing to
1846 * observe this restriction will result in deadlock.
1847 *
1848 * This implementation can be thought of as an application of ticket
1849 * locking to RCU, with sync_sched_expedited_started and
1850 * sync_sched_expedited_done taking on the roles of the halves
1851 * of the ticket-lock word. Each task atomically increments
1852 * sync_sched_expedited_started upon entry, snapshotting the old value,
1853 * then attempts to stop all the CPUs. If this succeeds, then each
1854 * CPU will have executed a context switch, resulting in an RCU-sched
1855 * grace period. We are then done, so we use atomic_cmpxchg() to
1856 * update sync_sched_expedited_done to match our snapshot -- but
1857 * only if someone else has not already advanced past our snapshot.
1858 *
1859 * On the other hand, if try_stop_cpus() fails, we check the value
1860 * of sync_sched_expedited_done. If it has advanced past our
1861 * initial snapshot, then someone else must have forced a grace period
1862 * some time after we took our snapshot. In this case, our work is
1863 * done for us, and we can simply return. Otherwise, we try again,
1864 * but keep our initial snapshot for purposes of checking for someone
1865 * doing our work for us.
1866 *
1867 * If we fail too many times in a row, we fall back to synchronize_sched().
1868 */
1869void synchronize_sched_expedited(void)
1870{
1871 int firstsnap, s, snap, trycount = 0;
1872
1873 /* Note that atomic_inc_return() implies full memory barrier. */
1874 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
1875 get_online_cpus();
1876
1877 /*
1878 * Each pass through the following loop attempts to force a
1879 * context switch on each CPU.
1880 */
1881 while (try_stop_cpus(cpu_online_mask,
1882 synchronize_sched_expedited_cpu_stop,
1883 NULL) == -EAGAIN) {
1884 put_online_cpus();
1885
1886 /* No joy, try again later. Or just synchronize_sched(). */
1887 if (trycount++ < 10)
1888 udelay(trycount * num_online_cpus());
1889 else {
1890 synchronize_sched();
1891 return;
1892 }
1893
1894 /* Check to see if someone else did our work for us. */
1895 s = atomic_read(&sync_sched_expedited_done);
1896 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
1897 smp_mb(); /* ensure test happens before caller kfree */
1898 return;
1899 }
1900
1901 /*
1902 * Refetching sync_sched_expedited_started allows later
1903 * callers to piggyback on our grace period. We subtract
1904 * 1 to get the same token that the last incrementer got.
1905 * We retry after they started, so our grace period works
1906 * for them, and they started after our first try, so their
1907 * grace period works for us.
1908 */
1909 get_online_cpus();
1910 snap = atomic_read(&sync_sched_expedited_started) - 1;
1911 smp_mb(); /* ensure read is before try_stop_cpus(). */
1912 }
1913
1914 /*
1915 * Everyone up to our most recent fetch is covered by our grace
1916 * period. Update the counter, but only if our work is still
1917 * relevant -- which it won't be if someone who started later
1918 * than we did beat us to the punch.
1919 */
1920 do {
1921 s = atomic_read(&sync_sched_expedited_done);
1922 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
1923 smp_mb(); /* ensure test happens before caller kfree */
1924 break;
1925 }
1926 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
1927
1928 put_online_cpus();
1929}
1930EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
1931
1932#endif /* #else #ifndef CONFIG_SMP */
1933
1934#if !defined(CONFIG_RCU_FAST_NO_HZ)
1935
1936/*
1937 * Check to see if any future RCU-related work will need to be done
1938 * by the current CPU, even if none need be done immediately, returning
1939 * 1 if so. This function is part of the RCU implementation; it is -not-
1940 * an exported member of the RCU API.
1941 *
1942 * Because we have preemptible RCU, just check whether this CPU needs
1943 * any flavor of RCU. Do not chew up lots of CPU cycles with preemption
1944 * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
1945 */
1946int rcu_needs_cpu(int cpu)
1947{
1948 return rcu_needs_cpu_quick_check(cpu);
1949}
1950
1951#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1952
1953#define RCU_NEEDS_CPU_FLUSHES 5
1954static DEFINE_PER_CPU(int, rcu_dyntick_drain);
1955static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff);
1956
1957/*
1958 * Check to see if any future RCU-related work will need to be done
1959 * by the current CPU, even if none need be done immediately, returning
1960 * 1 if so. This function is part of the RCU implementation; it is -not-
1961 * an exported member of the RCU API.
1962 *
1963 * Because we are not supporting preemptible RCU, attempt to accelerate
1964 * any current grace periods so that RCU no longer needs this CPU, but
1965 * only if all other CPUs are already in dynticks-idle mode. This will
1966 * allow the CPU cores to be powered down immediately, as opposed to after
1967 * waiting many milliseconds for grace periods to elapse.
1968 *
1969 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1970 * disabled, we do one pass of force_quiescent_state(), then do a
1971 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
1972 * later. The per-cpu rcu_dyntick_drain variable controls the sequencing.
1973 */
1974int rcu_needs_cpu(int cpu)
1975{
1976 int c = 0;
1977 int snap;
1978 int thatcpu;
1979
1980 /* Check for being in the holdoff period. */
1981 if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies)
1982 return rcu_needs_cpu_quick_check(cpu);
1983
1984 /* Don't bother unless we are the last non-dyntick-idle CPU. */
1985 for_each_online_cpu(thatcpu) {
1986 if (thatcpu == cpu)
1987 continue;
1988 snap = atomic_add_return(0, &per_cpu(rcu_dynticks,
1989 thatcpu).dynticks);
1990 smp_mb(); /* Order sampling of snap with end of grace period. */
1991 if ((snap & 0x1) != 0) {
1992 per_cpu(rcu_dyntick_drain, cpu) = 0;
1993 per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1;
1994 return rcu_needs_cpu_quick_check(cpu);
1995 }
1996 }
1997
1998 /* Check and update the rcu_dyntick_drain sequencing. */
1999 if (per_cpu(rcu_dyntick_drain, cpu) <= 0) {
2000 /* First time through, initialize the counter. */
2001 per_cpu(rcu_dyntick_drain, cpu) = RCU_NEEDS_CPU_FLUSHES;
2002 } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) {
2003 /* We have hit the limit, so time to give up. */
2004 per_cpu(rcu_dyntick_holdoff, cpu) = jiffies;
2005 return rcu_needs_cpu_quick_check(cpu);
2006 }
2007
2008 /* Do one step pushing remaining RCU callbacks through. */
2009 if (per_cpu(rcu_sched_data, cpu).nxtlist) {
2010 rcu_sched_qs(cpu);
2011 force_quiescent_state(&rcu_sched_state, 0);
2012 c = c || per_cpu(rcu_sched_data, cpu).nxtlist;
2013 }
2014 if (per_cpu(rcu_bh_data, cpu).nxtlist) {
2015 rcu_bh_qs(cpu);
2016 force_quiescent_state(&rcu_bh_state, 0);
2017 c = c || per_cpu(rcu_bh_data, cpu).nxtlist;
2018 }
2019
2020 /* If RCU callbacks are still pending, RCU still needs this CPU. */
2021 if (c)
2022 invoke_rcu_core();
2023 return c;
2024}
2025
2026#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */