rcu: Fix grace-period-stall bug on large systems with CPU hotplug
[deliverable/linux.git] / 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
28 #ifdef CONFIG_TREE_PREEMPT_RCU
29
30 struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state);
31 DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
32
33 /*
34 * Tell them what RCU they are running.
35 */
36 static void __init rcu_bootup_announce(void)
37 {
38 printk(KERN_INFO
39 "Experimental preemptable hierarchical RCU implementation.\n");
40 }
41
42 /*
43 * Return the number of RCU-preempt batches processed thus far
44 * for debug and statistics.
45 */
46 long rcu_batches_completed_preempt(void)
47 {
48 return rcu_preempt_state.completed;
49 }
50 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
51
52 /*
53 * Return the number of RCU batches processed thus far for debug & stats.
54 */
55 long rcu_batches_completed(void)
56 {
57 return rcu_batches_completed_preempt();
58 }
59 EXPORT_SYMBOL_GPL(rcu_batches_completed);
60
61 /*
62 * Record a preemptable-RCU quiescent state for the specified CPU. Note
63 * that this just means that the task currently running on the CPU is
64 * not in a quiescent state. There might be any number of tasks blocked
65 * while in an RCU read-side critical section.
66 */
67 static void rcu_preempt_qs(int cpu)
68 {
69 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
70 rdp->passed_quiesc_completed = rdp->gpnum - 1;
71 barrier();
72 rdp->passed_quiesc = 1;
73 }
74
75 /*
76 * We have entered the scheduler, and the current task might soon be
77 * context-switched away from. If this task is in an RCU read-side
78 * critical section, we will no longer be able to rely on the CPU to
79 * record that fact, so we enqueue the task on the appropriate entry
80 * of the blocked_tasks[] array. The task will dequeue itself when
81 * it exits the outermost enclosing RCU read-side critical section.
82 * Therefore, the current grace period cannot be permitted to complete
83 * until the blocked_tasks[] entry indexed by the low-order bit of
84 * rnp->gpnum empties.
85 *
86 * Caller must disable preemption.
87 */
88 static void rcu_preempt_note_context_switch(int cpu)
89 {
90 struct task_struct *t = current;
91 unsigned long flags;
92 int phase;
93 struct rcu_data *rdp;
94 struct rcu_node *rnp;
95
96 if (t->rcu_read_lock_nesting &&
97 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
98
99 /* Possibly blocking in an RCU read-side critical section. */
100 rdp = rcu_preempt_state.rda[cpu];
101 rnp = rdp->mynode;
102 spin_lock_irqsave(&rnp->lock, flags);
103 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
104 t->rcu_blocked_node = rnp;
105
106 /*
107 * If this CPU has already checked in, then this task
108 * will hold up the next grace period rather than the
109 * current grace period. Queue the task accordingly.
110 * If the task is queued for the current grace period
111 * (i.e., this CPU has not yet passed through a quiescent
112 * state for the current grace period), then as long
113 * as that task remains queued, the current grace period
114 * cannot end.
115 *
116 * But first, note that the current CPU must still be
117 * on line!
118 */
119 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
120 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
121 phase = (rnp->gpnum + !(rnp->qsmask & rdp->grpmask)) & 0x1;
122 list_add(&t->rcu_node_entry, &rnp->blocked_tasks[phase]);
123 spin_unlock_irqrestore(&rnp->lock, flags);
124 }
125
126 /*
127 * Either we were not in an RCU read-side critical section to
128 * begin with, or we have now recorded that critical section
129 * globally. Either way, we can now note a quiescent state
130 * for this CPU. Again, if we were in an RCU read-side critical
131 * section, and if that critical section was blocking the current
132 * grace period, then the fact that the task has been enqueued
133 * means that we continue to block the current grace period.
134 */
135 rcu_preempt_qs(cpu);
136 local_irq_save(flags);
137 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
138 local_irq_restore(flags);
139 }
140
141 /*
142 * Tree-preemptable RCU implementation for rcu_read_lock().
143 * Just increment ->rcu_read_lock_nesting, shared state will be updated
144 * if we block.
145 */
146 void __rcu_read_lock(void)
147 {
148 ACCESS_ONCE(current->rcu_read_lock_nesting)++;
149 barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */
150 }
151 EXPORT_SYMBOL_GPL(__rcu_read_lock);
152
153 /*
154 * Check for preempted RCU readers blocking the current grace period
155 * for the specified rcu_node structure. If the caller needs a reliable
156 * answer, it must hold the rcu_node's ->lock.
157 */
158 static int rcu_preempted_readers(struct rcu_node *rnp)
159 {
160 return !list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1]);
161 }
162
163 /*
164 * Record a quiescent state for all tasks that were previously queued
165 * on the specified rcu_node structure and that were blocking the current
166 * RCU grace period. The caller must hold the specified rnp->lock with
167 * irqs disabled, and this lock is released upon return, but irqs remain
168 * disabled.
169 */
170 static void task_quiet(struct rcu_node *rnp, unsigned long flags)
171 __releases(rnp->lock)
172 {
173 unsigned long mask;
174 struct rcu_node *rnp_p;
175
176 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
177 spin_unlock_irqrestore(&rnp->lock, flags);
178 return; /* Still need more quiescent states! */
179 }
180
181 rnp_p = rnp->parent;
182 if (rnp_p == NULL) {
183 /*
184 * Either there is only one rcu_node in the tree,
185 * or tasks were kicked up to root rcu_node due to
186 * CPUs going offline.
187 */
188 cpu_quiet_msk_finish(&rcu_preempt_state, flags);
189 return;
190 }
191
192 /* Report up the rest of the hierarchy. */
193 mask = rnp->grpmask;
194 spin_unlock(&rnp->lock); /* irqs remain disabled. */
195 spin_lock(&rnp_p->lock); /* irqs already disabled. */
196 cpu_quiet_msk(mask, &rcu_preempt_state, rnp_p, flags);
197 }
198
199 /*
200 * Handle special cases during rcu_read_unlock(), such as needing to
201 * notify RCU core processing or task having blocked during the RCU
202 * read-side critical section.
203 */
204 static void rcu_read_unlock_special(struct task_struct *t)
205 {
206 int empty;
207 unsigned long flags;
208 struct rcu_node *rnp;
209 int special;
210
211 /* NMI handlers cannot block and cannot safely manipulate state. */
212 if (in_nmi())
213 return;
214
215 local_irq_save(flags);
216
217 /*
218 * If RCU core is waiting for this CPU to exit critical section,
219 * let it know that we have done so.
220 */
221 special = t->rcu_read_unlock_special;
222 if (special & RCU_READ_UNLOCK_NEED_QS) {
223 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
224 rcu_preempt_qs(smp_processor_id());
225 }
226
227 /* Hardware IRQ handlers cannot block. */
228 if (in_irq()) {
229 local_irq_restore(flags);
230 return;
231 }
232
233 /* Clean up if blocked during RCU read-side critical section. */
234 if (special & RCU_READ_UNLOCK_BLOCKED) {
235 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
236
237 /*
238 * Remove this task from the list it blocked on. The
239 * task can migrate while we acquire the lock, but at
240 * most one time. So at most two passes through loop.
241 */
242 for (;;) {
243 rnp = t->rcu_blocked_node;
244 spin_lock(&rnp->lock); /* irqs already disabled. */
245 if (rnp == t->rcu_blocked_node)
246 break;
247 spin_unlock(&rnp->lock); /* irqs remain disabled. */
248 }
249 empty = !rcu_preempted_readers(rnp);
250 list_del_init(&t->rcu_node_entry);
251 t->rcu_blocked_node = NULL;
252
253 /*
254 * If this was the last task on the current list, and if
255 * we aren't waiting on any CPUs, report the quiescent state.
256 * Note that task_quiet() releases rnp->lock.
257 */
258 if (empty)
259 spin_unlock_irqrestore(&rnp->lock, flags);
260 else
261 task_quiet(rnp, flags);
262 } else {
263 local_irq_restore(flags);
264 }
265 }
266
267 /*
268 * Tree-preemptable RCU implementation for rcu_read_unlock().
269 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
270 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
271 * invoke rcu_read_unlock_special() to clean up after a context switch
272 * in an RCU read-side critical section and other special cases.
273 */
274 void __rcu_read_unlock(void)
275 {
276 struct task_struct *t = current;
277
278 barrier(); /* needed if we ever invoke rcu_read_unlock in rcutree.c */
279 if (--ACCESS_ONCE(t->rcu_read_lock_nesting) == 0 &&
280 unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
281 rcu_read_unlock_special(t);
282 }
283 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
284
285 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
286
287 /*
288 * Scan the current list of tasks blocked within RCU read-side critical
289 * sections, printing out the tid of each.
290 */
291 static void rcu_print_task_stall(struct rcu_node *rnp)
292 {
293 unsigned long flags;
294 struct list_head *lp;
295 int phase;
296 struct task_struct *t;
297
298 if (rcu_preempted_readers(rnp)) {
299 spin_lock_irqsave(&rnp->lock, flags);
300 phase = rnp->gpnum & 0x1;
301 lp = &rnp->blocked_tasks[phase];
302 list_for_each_entry(t, lp, rcu_node_entry)
303 printk(" P%d", t->pid);
304 spin_unlock_irqrestore(&rnp->lock, flags);
305 }
306 }
307
308 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
309
310 /*
311 * Check that the list of blocked tasks for the newly completed grace
312 * period is in fact empty. It is a serious bug to complete a grace
313 * period that still has RCU readers blocked! This function must be
314 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
315 * must be held by the caller.
316 */
317 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
318 {
319 WARN_ON_ONCE(rcu_preempted_readers(rnp));
320 WARN_ON_ONCE(rnp->qsmask);
321 }
322
323 #ifdef CONFIG_HOTPLUG_CPU
324
325 /*
326 * Handle tasklist migration for case in which all CPUs covered by the
327 * specified rcu_node have gone offline. Move them up to the root
328 * rcu_node. The reason for not just moving them to the immediate
329 * parent is to remove the need for rcu_read_unlock_special() to
330 * make more than two attempts to acquire the target rcu_node's lock.
331 * Returns true if there were tasks blocking the current RCU grace
332 * period.
333 *
334 * Returns 1 if there was previously a task blocking the current grace
335 * period on the specified rcu_node structure.
336 *
337 * The caller must hold rnp->lock with irqs disabled.
338 */
339 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
340 struct rcu_node *rnp,
341 struct rcu_data *rdp)
342 {
343 int i;
344 struct list_head *lp;
345 struct list_head *lp_root;
346 int retval;
347 struct rcu_node *rnp_root = rcu_get_root(rsp);
348 struct task_struct *tp;
349
350 if (rnp == rnp_root) {
351 WARN_ONCE(1, "Last CPU thought to be offlined?");
352 return 0; /* Shouldn't happen: at least one CPU online. */
353 }
354 WARN_ON_ONCE(rnp != rdp->mynode &&
355 (!list_empty(&rnp->blocked_tasks[0]) ||
356 !list_empty(&rnp->blocked_tasks[1])));
357
358 /*
359 * Move tasks up to root rcu_node. Rely on the fact that the
360 * root rcu_node can be at most one ahead of the rest of the
361 * rcu_nodes in terms of gp_num value. This fact allows us to
362 * move the blocked_tasks[] array directly, element by element.
363 */
364 retval = rcu_preempted_readers(rnp);
365 for (i = 0; i < 2; i++) {
366 lp = &rnp->blocked_tasks[i];
367 lp_root = &rnp_root->blocked_tasks[i];
368 while (!list_empty(lp)) {
369 tp = list_entry(lp->next, typeof(*tp), rcu_node_entry);
370 spin_lock(&rnp_root->lock); /* irqs already disabled */
371 list_del(&tp->rcu_node_entry);
372 tp->rcu_blocked_node = rnp_root;
373 list_add(&tp->rcu_node_entry, lp_root);
374 spin_unlock(&rnp_root->lock); /* irqs remain disabled */
375 }
376 }
377 return retval;
378 }
379
380 /*
381 * Do CPU-offline processing for preemptable RCU.
382 */
383 static void rcu_preempt_offline_cpu(int cpu)
384 {
385 __rcu_offline_cpu(cpu, &rcu_preempt_state);
386 }
387
388 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
389
390 /*
391 * Check for a quiescent state from the current CPU. When a task blocks,
392 * the task is recorded in the corresponding CPU's rcu_node structure,
393 * which is checked elsewhere.
394 *
395 * Caller must disable hard irqs.
396 */
397 static void rcu_preempt_check_callbacks(int cpu)
398 {
399 struct task_struct *t = current;
400
401 if (t->rcu_read_lock_nesting == 0) {
402 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
403 rcu_preempt_qs(cpu);
404 return;
405 }
406 if (per_cpu(rcu_preempt_data, cpu).qs_pending)
407 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
408 }
409
410 /*
411 * Process callbacks for preemptable RCU.
412 */
413 static void rcu_preempt_process_callbacks(void)
414 {
415 __rcu_process_callbacks(&rcu_preempt_state,
416 &__get_cpu_var(rcu_preempt_data));
417 }
418
419 /*
420 * Queue a preemptable-RCU callback for invocation after a grace period.
421 */
422 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
423 {
424 __call_rcu(head, func, &rcu_preempt_state);
425 }
426 EXPORT_SYMBOL_GPL(call_rcu);
427
428 /*
429 * Wait for an rcu-preempt grace period. We are supposed to expedite the
430 * grace period, but this is the crude slow compatability hack, so just
431 * invoke synchronize_rcu().
432 */
433 void synchronize_rcu_expedited(void)
434 {
435 synchronize_rcu();
436 }
437 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
438
439 /*
440 * Check to see if there is any immediate preemptable-RCU-related work
441 * to be done.
442 */
443 static int rcu_preempt_pending(int cpu)
444 {
445 return __rcu_pending(&rcu_preempt_state,
446 &per_cpu(rcu_preempt_data, cpu));
447 }
448
449 /*
450 * Does preemptable RCU need the CPU to stay out of dynticks mode?
451 */
452 static int rcu_preempt_needs_cpu(int cpu)
453 {
454 return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
455 }
456
457 /**
458 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
459 */
460 void rcu_barrier(void)
461 {
462 _rcu_barrier(&rcu_preempt_state, call_rcu);
463 }
464 EXPORT_SYMBOL_GPL(rcu_barrier);
465
466 /*
467 * Initialize preemptable RCU's per-CPU data.
468 */
469 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
470 {
471 rcu_init_percpu_data(cpu, &rcu_preempt_state, 1);
472 }
473
474 /*
475 * Move preemptable RCU's callbacks to ->orphan_cbs_list.
476 */
477 static void rcu_preempt_send_cbs_to_orphanage(void)
478 {
479 rcu_send_cbs_to_orphanage(&rcu_preempt_state);
480 }
481
482 /*
483 * Initialize preemptable RCU's state structures.
484 */
485 static void __init __rcu_init_preempt(void)
486 {
487 RCU_INIT_FLAVOR(&rcu_preempt_state, rcu_preempt_data);
488 }
489
490 /*
491 * Check for a task exiting while in a preemptable-RCU read-side
492 * critical section, clean up if so. No need to issue warnings,
493 * as debug_check_no_locks_held() already does this if lockdep
494 * is enabled.
495 */
496 void exit_rcu(void)
497 {
498 struct task_struct *t = current;
499
500 if (t->rcu_read_lock_nesting == 0)
501 return;
502 t->rcu_read_lock_nesting = 1;
503 rcu_read_unlock();
504 }
505
506 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
507
508 /*
509 * Tell them what RCU they are running.
510 */
511 static void __init rcu_bootup_announce(void)
512 {
513 printk(KERN_INFO "Hierarchical RCU implementation.\n");
514 }
515
516 /*
517 * Return the number of RCU batches processed thus far for debug & stats.
518 */
519 long rcu_batches_completed(void)
520 {
521 return rcu_batches_completed_sched();
522 }
523 EXPORT_SYMBOL_GPL(rcu_batches_completed);
524
525 /*
526 * Because preemptable RCU does not exist, we never have to check for
527 * CPUs being in quiescent states.
528 */
529 static void rcu_preempt_note_context_switch(int cpu)
530 {
531 }
532
533 /*
534 * Because preemptable RCU does not exist, there are never any preempted
535 * RCU readers.
536 */
537 static int rcu_preempted_readers(struct rcu_node *rnp)
538 {
539 return 0;
540 }
541
542 #ifdef CONFIG_HOTPLUG_CPU
543
544 /* Because preemptible RCU does not exist, no quieting of tasks. */
545 static void task_quiet(struct rcu_node *rnp, unsigned long flags)
546 {
547 spin_unlock_irqrestore(&rnp->lock, flags);
548 }
549
550 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
551
552 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
553
554 /*
555 * Because preemptable RCU does not exist, we never have to check for
556 * tasks blocked within RCU read-side critical sections.
557 */
558 static void rcu_print_task_stall(struct rcu_node *rnp)
559 {
560 }
561
562 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
563
564 /*
565 * Because there is no preemptable RCU, there can be no readers blocked,
566 * so there is no need to check for blocked tasks. So check only for
567 * bogus qsmask values.
568 */
569 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
570 {
571 WARN_ON_ONCE(rnp->qsmask);
572 }
573
574 #ifdef CONFIG_HOTPLUG_CPU
575
576 /*
577 * Because preemptable RCU does not exist, it never needs to migrate
578 * tasks that were blocked within RCU read-side critical sections, and
579 * such non-existent tasks cannot possibly have been blocking the current
580 * grace period.
581 */
582 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
583 struct rcu_node *rnp,
584 struct rcu_data *rdp)
585 {
586 return 0;
587 }
588
589 /*
590 * Because preemptable RCU does not exist, it never needs CPU-offline
591 * processing.
592 */
593 static void rcu_preempt_offline_cpu(int cpu)
594 {
595 }
596
597 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
598
599 /*
600 * Because preemptable RCU does not exist, it never has any callbacks
601 * to check.
602 */
603 static void rcu_preempt_check_callbacks(int cpu)
604 {
605 }
606
607 /*
608 * Because preemptable RCU does not exist, it never has any callbacks
609 * to process.
610 */
611 static void rcu_preempt_process_callbacks(void)
612 {
613 }
614
615 /*
616 * In classic RCU, call_rcu() is just call_rcu_sched().
617 */
618 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
619 {
620 call_rcu_sched(head, func);
621 }
622 EXPORT_SYMBOL_GPL(call_rcu);
623
624 /*
625 * Wait for an rcu-preempt grace period, but make it happen quickly.
626 * But because preemptable RCU does not exist, map to rcu-sched.
627 */
628 void synchronize_rcu_expedited(void)
629 {
630 synchronize_sched_expedited();
631 }
632 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
633
634 /*
635 * Because preemptable RCU does not exist, it never has any work to do.
636 */
637 static int rcu_preempt_pending(int cpu)
638 {
639 return 0;
640 }
641
642 /*
643 * Because preemptable RCU does not exist, it never needs any CPU.
644 */
645 static int rcu_preempt_needs_cpu(int cpu)
646 {
647 return 0;
648 }
649
650 /*
651 * Because preemptable RCU does not exist, rcu_barrier() is just
652 * another name for rcu_barrier_sched().
653 */
654 void rcu_barrier(void)
655 {
656 rcu_barrier_sched();
657 }
658 EXPORT_SYMBOL_GPL(rcu_barrier);
659
660 /*
661 * Because preemptable RCU does not exist, there is no per-CPU
662 * data to initialize.
663 */
664 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
665 {
666 }
667
668 /*
669 * Because there is no preemptable RCU, there are no callbacks to move.
670 */
671 static void rcu_preempt_send_cbs_to_orphanage(void)
672 {
673 }
674
675 /*
676 * Because preemptable RCU does not exist, it need not be initialized.
677 */
678 static void __init __rcu_init_preempt(void)
679 {
680 }
681
682 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
This page took 0.085365 seconds and 6 git commands to generate.