projects/linux·blob v2.6.21 · refs

635 lines · 17.4 KiB

1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
17 *
18 * Copyright (C) IBM Corporation, 2001
19 *
20 * Authors: Dipankar Sarma <[email protected]>
21 * Manfred Spraul <[email protected]>
22 *
23 * Based on the original work by Paul McKenney <[email protected]>
24 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
25 * Papers:
26 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
27 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
28 *
29 * For detailed explanation of Read-Copy Update mechanism see -
30 * http://lse.sourceforge.net/locking/rcupdate.html
31 *
32 */
33#include <linux/types.h>
34#include <linux/kernel.h>
35#include <linux/init.h>
36#include <linux/spinlock.h>
37#include <linux/smp.h>
38#include <linux/rcupdate.h>
39#include <linux/interrupt.h>
40#include <linux/sched.h>
41#include <asm/atomic.h>
42#include <linux/bitops.h>
43#include <linux/module.h>
44#include <linux/completion.h>
45#include <linux/moduleparam.h>
46#include <linux/percpu.h>
47#include <linux/notifier.h>
48#include <linux/rcupdate.h>
49#include <linux/cpu.h>
50#include <linux/mutex.h>
51
52/* Definition for rcupdate control block. */
53static struct rcu_ctrlblk rcu_ctrlblk = {
54 .cur = -300,
55 .completed = -300,
56 .lock = __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk.lock),
57 .cpumask = CPU_MASK_NONE,
58};
59static struct rcu_ctrlblk rcu_bh_ctrlblk = {
60 .cur = -300,
61 .completed = -300,
62 .lock = __SPIN_LOCK_UNLOCKED(&rcu_bh_ctrlblk.lock),
63 .cpumask = CPU_MASK_NONE,
64};
65
66DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
67DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
68
69/* Fake initialization required by compiler */
70static DEFINE_PER_CPU(struct tasklet_struct, rcu_tasklet) = {NULL};
71static int blimit = 10;
72static int qhimark = 10000;
73static int qlowmark = 100;
74
75static atomic_t rcu_barrier_cpu_count;
76static DEFINE_MUTEX(rcu_barrier_mutex);
77static struct completion rcu_barrier_completion;
78
79#ifdef CONFIG_SMP
80static void force_quiescent_state(struct rcu_data *rdp,
81 struct rcu_ctrlblk *rcp)
82{
83 int cpu;
84 cpumask_t cpumask;
85 set_need_resched();
86 if (unlikely(!rcp->signaled)) {
87 rcp->signaled = 1;
88 /*
89 * Don't send IPI to itself. With irqs disabled,
90 * rdp->cpu is the current cpu.
91 */
92 cpumask = rcp->cpumask;
93 cpu_clear(rdp->cpu, cpumask);
94 for_each_cpu_mask(cpu, cpumask)
95 smp_send_reschedule(cpu);
96 }
97}
98#else
99static inline void force_quiescent_state(struct rcu_data *rdp,
100 struct rcu_ctrlblk *rcp)
101{
102 set_need_resched();
103}
104#endif
105
106/**
107 * call_rcu - Queue an RCU callback for invocation after a grace period.
108 * @head: structure to be used for queueing the RCU updates.
109 * @func: actual update function to be invoked after the grace period
110 *
111 * The update function will be invoked some time after a full grace
112 * period elapses, in other words after all currently executing RCU
113 * read-side critical sections have completed. RCU read-side critical
114 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
115 * and may be nested.
116 */
117void fastcall call_rcu(struct rcu_head *head,
118 void (*func)(struct rcu_head *rcu))
119{
120 unsigned long flags;
121 struct rcu_data *rdp;
122
123 head->func = func;
124 head->next = NULL;
125 local_irq_save(flags);
126 rdp = &__get_cpu_var(rcu_data);
127 *rdp->nxttail = head;
128 rdp->nxttail = &head->next;
129 if (unlikely(++rdp->qlen > qhimark)) {
130 rdp->blimit = INT_MAX;
131 force_quiescent_state(rdp, &rcu_ctrlblk);
132 }
133 local_irq_restore(flags);
134}
135
136/**
137 * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
138 * @head: structure to be used for queueing the RCU updates.
139 * @func: actual update function to be invoked after the grace period
140 *
141 * The update function will be invoked some time after a full grace
142 * period elapses, in other words after all currently executing RCU
143 * read-side critical sections have completed. call_rcu_bh() assumes
144 * that the read-side critical sections end on completion of a softirq
145 * handler. This means that read-side critical sections in process
146 * context must not be interrupted by softirqs. This interface is to be
147 * used when most of the read-side critical sections are in softirq context.
148 * RCU read-side critical sections are delimited by rcu_read_lock() and
149 * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
150 * and rcu_read_unlock_bh(), if in process context. These may be nested.
151 */
152void fastcall call_rcu_bh(struct rcu_head *head,
153 void (*func)(struct rcu_head *rcu))
154{
155 unsigned long flags;
156 struct rcu_data *rdp;
157
158 head->func = func;
159 head->next = NULL;
160 local_irq_save(flags);
161 rdp = &__get_cpu_var(rcu_bh_data);
162 *rdp->nxttail = head;
163 rdp->nxttail = &head->next;
164
165 if (unlikely(++rdp->qlen > qhimark)) {
166 rdp->blimit = INT_MAX;
167 force_quiescent_state(rdp, &rcu_bh_ctrlblk);
168 }
169
170 local_irq_restore(flags);
171}
172
173/*
174 * Return the number of RCU batches processed thus far. Useful
175 * for debug and statistics.
176 */
177long rcu_batches_completed(void)
178{
179 return rcu_ctrlblk.completed;
180}
181
182/*
183 * Return the number of RCU batches processed thus far. Useful
184 * for debug and statistics.
185 */
186long rcu_batches_completed_bh(void)
187{
188 return rcu_bh_ctrlblk.completed;
189}
190
191static void rcu_barrier_callback(struct rcu_head *notused)
192{
193 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
194 complete(&rcu_barrier_completion);
195}
196
197/*
198 * Called with preemption disabled, and from cross-cpu IRQ context.
199 */
200static void rcu_barrier_func(void *notused)
201{
202 int cpu = smp_processor_id();
203 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
204 struct rcu_head *head;
205
206 head = &rdp->barrier;
207 atomic_inc(&rcu_barrier_cpu_count);
208 call_rcu(head, rcu_barrier_callback);
209}
210
211/**
212 * rcu_barrier - Wait until all the in-flight RCUs are complete.
213 */
214void rcu_barrier(void)
215{
216 BUG_ON(in_interrupt());
217 /* Take cpucontrol mutex to protect against CPU hotplug */
218 mutex_lock(&rcu_barrier_mutex);
219 init_completion(&rcu_barrier_completion);
220 atomic_set(&rcu_barrier_cpu_count, 0);
221 on_each_cpu(rcu_barrier_func, NULL, 0, 1);
222 wait_for_completion(&rcu_barrier_completion);
223 mutex_unlock(&rcu_barrier_mutex);
224}
225EXPORT_SYMBOL_GPL(rcu_barrier);
226
227/*
228 * Invoke the completed RCU callbacks. They are expected to be in
229 * a per-cpu list.
230 */
231static void rcu_do_batch(struct rcu_data *rdp)
232{
233 struct rcu_head *next, *list;
234 int count = 0;
235
236 list = rdp->donelist;
237 while (list) {
238 next = list->next;
239 prefetch(next);
240 list->func(list);
241 list = next;
242 if (++count >= rdp->blimit)
243 break;
244 }
245 rdp->donelist = list;
246
247 local_irq_disable();
248 rdp->qlen -= count;
249 local_irq_enable();
250 if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark)
251 rdp->blimit = blimit;
252
253 if (!rdp->donelist)
254 rdp->donetail = &rdp->donelist;
255 else
256 tasklet_schedule(&per_cpu(rcu_tasklet, rdp->cpu));
257}
258
259/*
260 * Grace period handling:
261 * The grace period handling consists out of two steps:
262 * - A new grace period is started.
263 * This is done by rcu_start_batch. The start is not broadcasted to
264 * all cpus, they must pick this up by comparing rcp->cur with
265 * rdp->quiescbatch. All cpus are recorded in the
266 * rcu_ctrlblk.cpumask bitmap.
267 * - All cpus must go through a quiescent state.
268 * Since the start of the grace period is not broadcasted, at least two
269 * calls to rcu_check_quiescent_state are required:
270 * The first call just notices that a new grace period is running. The
271 * following calls check if there was a quiescent state since the beginning
272 * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
273 * the bitmap is empty, then the grace period is completed.
274 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
275 * period (if necessary).
276 */
277/*
278 * Register a new batch of callbacks, and start it up if there is currently no
279 * active batch and the batch to be registered has not already occurred.
280 * Caller must hold rcu_ctrlblk.lock.
281 */
282static void rcu_start_batch(struct rcu_ctrlblk *rcp)
283{
284 if (rcp->next_pending &&
285 rcp->completed == rcp->cur) {
286 rcp->next_pending = 0;
287 /*
288 * next_pending == 0 must be visible in
289 * __rcu_process_callbacks() before it can see new value of cur.
290 */
291 smp_wmb();
292 rcp->cur++;
293
294 /*
295 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
296 * Barrier Otherwise it can cause tickless idle CPUs to be
297 * included in rcp->cpumask, which will extend graceperiods
298 * unnecessarily.
299 */
300 smp_mb();
301 cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
302
303 rcp->signaled = 0;
304 }
305}
306
307/*
308 * cpu went through a quiescent state since the beginning of the grace period.
309 * Clear it from the cpu mask and complete the grace period if it was the last
310 * cpu. Start another grace period if someone has further entries pending
311 */
312static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
313{
314 cpu_clear(cpu, rcp->cpumask);
315 if (cpus_empty(rcp->cpumask)) {
316 /* batch completed ! */
317 rcp->completed = rcp->cur;
318 rcu_start_batch(rcp);
319 }
320}
321
322/*
323 * Check if the cpu has gone through a quiescent state (say context
324 * switch). If so and if it already hasn't done so in this RCU
325 * quiescent cycle, then indicate that it has done so.
326 */
327static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
328 struct rcu_data *rdp)
329{
330 if (rdp->quiescbatch != rcp->cur) {
331 /* start new grace period: */
332 rdp->qs_pending = 1;
333 rdp->passed_quiesc = 0;
334 rdp->quiescbatch = rcp->cur;
335 return;
336 }
337
338 /* Grace period already completed for this cpu?
339 * qs_pending is checked instead of the actual bitmap to avoid
340 * cacheline trashing.
341 */
342 if (!rdp->qs_pending)
343 return;
344
345 /*
346 * Was there a quiescent state since the beginning of the grace
347 * period? If no, then exit and wait for the next call.
348 */
349 if (!rdp->passed_quiesc)
350 return;
351 rdp->qs_pending = 0;
352
353 spin_lock(&rcp->lock);
354 /*
355 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
356 * during cpu startup. Ignore the quiescent state.
357 */
358 if (likely(rdp->quiescbatch == rcp->cur))
359 cpu_quiet(rdp->cpu, rcp);
360
361 spin_unlock(&rcp->lock);
362}
363
364
365#ifdef CONFIG_HOTPLUG_CPU
366
367/* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
368 * locking requirements, the list it's pulling from has to belong to a cpu
369 * which is dead and hence not processing interrupts.
370 */
371static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
372 struct rcu_head **tail)
373{
374 local_irq_disable();
375 *this_rdp->nxttail = list;
376 if (list)
377 this_rdp->nxttail = tail;
378 local_irq_enable();
379}
380
381static void __rcu_offline_cpu(struct rcu_data *this_rdp,
382 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
383{
384 /* if the cpu going offline owns the grace period
385 * we can block indefinitely waiting for it, so flush
386 * it here
387 */
388 spin_lock_bh(&rcp->lock);
389 if (rcp->cur != rcp->completed)
390 cpu_quiet(rdp->cpu, rcp);
391 spin_unlock_bh(&rcp->lock);
392 rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail);
393 rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail);
394 rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail);
395}
396
397static void rcu_offline_cpu(int cpu)
398{
399 struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
400 struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
401
402 __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
403 &per_cpu(rcu_data, cpu));
404 __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
405 &per_cpu(rcu_bh_data, cpu));
406 put_cpu_var(rcu_data);
407 put_cpu_var(rcu_bh_data);
408 tasklet_kill_immediate(&per_cpu(rcu_tasklet, cpu), cpu);
409}
410
411#else
412
413static void rcu_offline_cpu(int cpu)
414{
415}
416
417#endif
418
419/*
420 * This does the RCU processing work from tasklet context.
421 */
422static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
423 struct rcu_data *rdp)
424{
425 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) {
426 *rdp->donetail = rdp->curlist;
427 rdp->donetail = rdp->curtail;
428 rdp->curlist = NULL;
429 rdp->curtail = &rdp->curlist;
430 }
431
432 if (rdp->nxtlist && !rdp->curlist) {
433 local_irq_disable();
434 rdp->curlist = rdp->nxtlist;
435 rdp->curtail = rdp->nxttail;
436 rdp->nxtlist = NULL;
437 rdp->nxttail = &rdp->nxtlist;
438 local_irq_enable();
439
440 /*
441 * start the next batch of callbacks
442 */
443
444 /* determine batch number */
445 rdp->batch = rcp->cur + 1;
446 /* see the comment and corresponding wmb() in
447 * the rcu_start_batch()
448 */
449 smp_rmb();
450
451 if (!rcp->next_pending) {
452 /* and start it/schedule start if it's a new batch */
453 spin_lock(&rcp->lock);
454 rcp->next_pending = 1;
455 rcu_start_batch(rcp);
456 spin_unlock(&rcp->lock);
457 }
458 }
459
460 rcu_check_quiescent_state(rcp, rdp);
461 if (rdp->donelist)
462 rcu_do_batch(rdp);
463}
464
465static void rcu_process_callbacks(unsigned long unused)
466{
467 __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
468 __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
469}
470
471static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
472{
473 /* This cpu has pending rcu entries and the grace period
474 * for them has completed.
475 */
476 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch))
477 return 1;
478
479 /* This cpu has no pending entries, but there are new entries */
480 if (!rdp->curlist && rdp->nxtlist)
481 return 1;
482
483 /* This cpu has finished callbacks to invoke */
484 if (rdp->donelist)
485 return 1;
486
487 /* The rcu core waits for a quiescent state from the cpu */
488 if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
489 return 1;
490
491 /* nothing to do */
492 return 0;
493}
494
495/*
496 * Check to see if there is any immediate RCU-related work to be done
497 * by the current CPU, returning 1 if so. This function is part of the
498 * RCU implementation; it is -not- an exported member of the RCU API.
499 */
500int rcu_pending(int cpu)
501{
502 return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
503 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
504}
505
506/*
507 * Check to see if any future RCU-related work will need to be done
508 * by the current CPU, even if none need be done immediately, returning
509 * 1 if so. This function is part of the RCU implementation; it is -not-
510 * an exported member of the RCU API.
511 */
512int rcu_needs_cpu(int cpu)
513{
514 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
515 struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
516
517 return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
518}
519
520void rcu_check_callbacks(int cpu, int user)
521{
522 if (user ||
523 (idle_cpu(cpu) && !in_softirq() &&
524 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
525 rcu_qsctr_inc(cpu);
526 rcu_bh_qsctr_inc(cpu);
527 } else if (!in_softirq())
528 rcu_bh_qsctr_inc(cpu);
529 tasklet_schedule(&per_cpu(rcu_tasklet, cpu));
530}
531
532static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
533 struct rcu_data *rdp)
534{
535 memset(rdp, 0, sizeof(*rdp));
536 rdp->curtail = &rdp->curlist;
537 rdp->nxttail = &rdp->nxtlist;
538 rdp->donetail = &rdp->donelist;
539 rdp->quiescbatch = rcp->completed;
540 rdp->qs_pending = 0;
541 rdp->cpu = cpu;
542 rdp->blimit = blimit;
543}
544
545static void __devinit rcu_online_cpu(int cpu)
546{
547 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
548 struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
549
550 rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
551 rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
552 tasklet_init(&per_cpu(rcu_tasklet, cpu), rcu_process_callbacks, 0UL);
553}
554
555static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
556 unsigned long action, void *hcpu)
557{
558 long cpu = (long)hcpu;
559 switch (action) {
560 case CPU_UP_PREPARE:
561 rcu_online_cpu(cpu);
562 break;
563 case CPU_DEAD:
564 rcu_offline_cpu(cpu);
565 break;
566 default:
567 break;
568 }
569 return NOTIFY_OK;
570}
571
572static struct notifier_block __cpuinitdata rcu_nb = {
573 .notifier_call = rcu_cpu_notify,
574};
575
576/*
577 * Initializes rcu mechanism. Assumed to be called early.
578 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
579 * Note that rcu_qsctr and friends are implicitly
580 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
581 */
582void __init rcu_init(void)
583{
584 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
585 (void *)(long)smp_processor_id());
586 /* Register notifier for non-boot CPUs */
587 register_cpu_notifier(&rcu_nb);
588}
589
590struct rcu_synchronize {
591 struct rcu_head head;
592 struct completion completion;
593};
594
595/* Because of FASTCALL declaration of complete, we use this wrapper */
596static void wakeme_after_rcu(struct rcu_head *head)
597{
598 struct rcu_synchronize *rcu;
599
600 rcu = container_of(head, struct rcu_synchronize, head);
601 complete(&rcu->completion);
602}
603
604/**
605 * synchronize_rcu - wait until a grace period has elapsed.
606 *
607 * Control will return to the caller some time after a full grace
608 * period has elapsed, in other words after all currently executing RCU
609 * read-side critical sections have completed. RCU read-side critical
610 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
611 * and may be nested.
612 *
613 * If your read-side code is not protected by rcu_read_lock(), do -not-
614 * use synchronize_rcu().
615 */
616void synchronize_rcu(void)
617{
618 struct rcu_synchronize rcu;
619
620 init_completion(&rcu.completion);
621 /* Will wake me after RCU finished */
622 call_rcu(&rcu.head, wakeme_after_rcu);
623
624 /* Wait for it */
625 wait_for_completion(&rcu.completion);
626}
627
628module_param(blimit, int, 0);
629module_param(qhimark, int, 0);
630module_param(qlowmark, int, 0);
631EXPORT_SYMBOL_GPL(rcu_batches_completed);
632EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
633EXPORT_SYMBOL_GPL(call_rcu);
634EXPORT_SYMBOL_GPL(call_rcu_bh);
635EXPORT_SYMBOL_GPL(synchronize_rcu);