[PATCH] sched: fix cond_resched_softirq() offset
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / rcupdate.c
blob26bb5ffe1ef14ef61824180dd6163acdc104006d
1 /*
2 * Read-Copy Update mechanism for mutual exclusion
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.
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.
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.
18 * Copyright (C) IBM Corporation, 2001
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
23 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
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)
29 * For detailed explanation of Read-Copy Update mechanism see -
30 * http://lse.sourceforge.net/locking/rcupdate.html
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>
52 /* Definition for rcupdate control block. */
53 static struct rcu_ctrlblk rcu_ctrlblk = {
54 .cur = -300,
55 .completed = -300,
56 .lock = __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk.lock),
57 .cpumask = CPU_MASK_NONE,
59 static 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,
66 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
67 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
69 /* Fake initialization required by compiler */
70 static DEFINE_PER_CPU(struct tasklet_struct, rcu_tasklet) = {NULL};
71 static int blimit = 10;
72 static int qhimark = 10000;
73 static int qlowmark = 100;
75 static atomic_t rcu_barrier_cpu_count;
76 static DEFINE_MUTEX(rcu_barrier_mutex);
77 static struct completion rcu_barrier_completion;
79 #ifdef CONFIG_SMP
80 static void force_quiescent_state(struct rcu_data *rdp,
81 struct rcu_ctrlblk *rcp)
83 int cpu;
84 cpumask_t cpumask;
85 set_need_resched();
86 if (unlikely(!rcp->signaled)) {
87 rcp->signaled = 1;
89 * Don't send IPI to itself. With irqs disabled,
90 * rdp->cpu is the current cpu.
92 cpumask = rcp->cpumask;
93 cpu_clear(rdp->cpu, cpumask);
94 for_each_cpu_mask(cpu, cpumask)
95 smp_send_reschedule(cpu);
98 #else
99 static inline void force_quiescent_state(struct rcu_data *rdp,
100 struct rcu_ctrlblk *rcp)
102 set_need_resched();
104 #endif
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
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.
117 void fastcall call_rcu(struct rcu_head *head,
118 void (*func)(struct rcu_head *rcu))
120 unsigned long flags;
121 struct rcu_data *rdp;
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);
133 local_irq_restore(flags);
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
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.
152 void fastcall call_rcu_bh(struct rcu_head *head,
153 void (*func)(struct rcu_head *rcu))
155 unsigned long flags;
156 struct rcu_data *rdp;
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;
165 if (unlikely(++rdp->qlen > qhimark)) {
166 rdp->blimit = INT_MAX;
167 force_quiescent_state(rdp, &rcu_bh_ctrlblk);
170 local_irq_restore(flags);
174 * Return the number of RCU batches processed thus far. Useful
175 * for debug and statistics.
177 long rcu_batches_completed(void)
179 return rcu_ctrlblk.completed;
183 * Return the number of RCU batches processed thus far. Useful
184 * for debug and statistics.
186 long rcu_batches_completed_bh(void)
188 return rcu_bh_ctrlblk.completed;
191 static void rcu_barrier_callback(struct rcu_head *notused)
193 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
194 complete(&rcu_barrier_completion);
198 * Called with preemption disabled, and from cross-cpu IRQ context.
200 static void rcu_barrier_func(void *notused)
202 int cpu = smp_processor_id();
203 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
204 struct rcu_head *head;
206 head = &rdp->barrier;
207 atomic_inc(&rcu_barrier_cpu_count);
208 call_rcu(head, rcu_barrier_callback);
212 * rcu_barrier - Wait until all the in-flight RCUs are complete.
214 void rcu_barrier(void)
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);
225 EXPORT_SYMBOL_GPL(rcu_barrier);
228 * Invoke the completed RCU callbacks. They are expected to be in
229 * a per-cpu list.
231 static void rcu_do_batch(struct rcu_data *rdp)
233 struct rcu_head *next, *list;
234 int count = 0;
236 list = rdp->donelist;
237 while (list) {
238 next = rdp->donelist = list->next;
239 list->func(list);
240 list = next;
241 if (++count >= rdp->blimit)
242 break;
245 local_irq_disable();
246 rdp->qlen -= count;
247 local_irq_enable();
248 if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark)
249 rdp->blimit = blimit;
251 if (!rdp->donelist)
252 rdp->donetail = &rdp->donelist;
253 else
254 tasklet_schedule(&per_cpu(rcu_tasklet, rdp->cpu));
258 * Grace period handling:
259 * The grace period handling consists out of two steps:
260 * - A new grace period is started.
261 * This is done by rcu_start_batch. The start is not broadcasted to
262 * all cpus, they must pick this up by comparing rcp->cur with
263 * rdp->quiescbatch. All cpus are recorded in the
264 * rcu_ctrlblk.cpumask bitmap.
265 * - All cpus must go through a quiescent state.
266 * Since the start of the grace period is not broadcasted, at least two
267 * calls to rcu_check_quiescent_state are required:
268 * The first call just notices that a new grace period is running. The
269 * following calls check if there was a quiescent state since the beginning
270 * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
271 * the bitmap is empty, then the grace period is completed.
272 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
273 * period (if necessary).
276 * Register a new batch of callbacks, and start it up if there is currently no
277 * active batch and the batch to be registered has not already occurred.
278 * Caller must hold rcu_ctrlblk.lock.
280 static void rcu_start_batch(struct rcu_ctrlblk *rcp)
282 if (rcp->next_pending &&
283 rcp->completed == rcp->cur) {
284 rcp->next_pending = 0;
286 * next_pending == 0 must be visible in
287 * __rcu_process_callbacks() before it can see new value of cur.
289 smp_wmb();
290 rcp->cur++;
293 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
294 * Barrier Otherwise it can cause tickless idle CPUs to be
295 * included in rcp->cpumask, which will extend graceperiods
296 * unnecessarily.
298 smp_mb();
299 cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
301 rcp->signaled = 0;
306 * cpu went through a quiescent state since the beginning of the grace period.
307 * Clear it from the cpu mask and complete the grace period if it was the last
308 * cpu. Start another grace period if someone has further entries pending
310 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
312 cpu_clear(cpu, rcp->cpumask);
313 if (cpus_empty(rcp->cpumask)) {
314 /* batch completed ! */
315 rcp->completed = rcp->cur;
316 rcu_start_batch(rcp);
321 * Check if the cpu has gone through a quiescent state (say context
322 * switch). If so and if it already hasn't done so in this RCU
323 * quiescent cycle, then indicate that it has done so.
325 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
326 struct rcu_data *rdp)
328 if (rdp->quiescbatch != rcp->cur) {
329 /* start new grace period: */
330 rdp->qs_pending = 1;
331 rdp->passed_quiesc = 0;
332 rdp->quiescbatch = rcp->cur;
333 return;
336 /* Grace period already completed for this cpu?
337 * qs_pending is checked instead of the actual bitmap to avoid
338 * cacheline trashing.
340 if (!rdp->qs_pending)
341 return;
344 * Was there a quiescent state since the beginning of the grace
345 * period? If no, then exit and wait for the next call.
347 if (!rdp->passed_quiesc)
348 return;
349 rdp->qs_pending = 0;
351 spin_lock(&rcp->lock);
353 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
354 * during cpu startup. Ignore the quiescent state.
356 if (likely(rdp->quiescbatch == rcp->cur))
357 cpu_quiet(rdp->cpu, rcp);
359 spin_unlock(&rcp->lock);
363 #ifdef CONFIG_HOTPLUG_CPU
365 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
366 * locking requirements, the list it's pulling from has to belong to a cpu
367 * which is dead and hence not processing interrupts.
369 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
370 struct rcu_head **tail)
372 local_irq_disable();
373 *this_rdp->nxttail = list;
374 if (list)
375 this_rdp->nxttail = tail;
376 local_irq_enable();
379 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
380 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
382 /* if the cpu going offline owns the grace period
383 * we can block indefinitely waiting for it, so flush
384 * it here
386 spin_lock_bh(&rcp->lock);
387 if (rcp->cur != rcp->completed)
388 cpu_quiet(rdp->cpu, rcp);
389 spin_unlock_bh(&rcp->lock);
390 rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail);
391 rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail);
392 rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail);
395 static void rcu_offline_cpu(int cpu)
397 struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
398 struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
400 __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
401 &per_cpu(rcu_data, cpu));
402 __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
403 &per_cpu(rcu_bh_data, cpu));
404 put_cpu_var(rcu_data);
405 put_cpu_var(rcu_bh_data);
406 tasklet_kill_immediate(&per_cpu(rcu_tasklet, cpu), cpu);
409 #else
411 static void rcu_offline_cpu(int cpu)
415 #endif
418 * This does the RCU processing work from tasklet context.
420 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
421 struct rcu_data *rdp)
423 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) {
424 *rdp->donetail = rdp->curlist;
425 rdp->donetail = rdp->curtail;
426 rdp->curlist = NULL;
427 rdp->curtail = &rdp->curlist;
430 if (rdp->nxtlist && !rdp->curlist) {
431 local_irq_disable();
432 rdp->curlist = rdp->nxtlist;
433 rdp->curtail = rdp->nxttail;
434 rdp->nxtlist = NULL;
435 rdp->nxttail = &rdp->nxtlist;
436 local_irq_enable();
439 * start the next batch of callbacks
442 /* determine batch number */
443 rdp->batch = rcp->cur + 1;
444 /* see the comment and corresponding wmb() in
445 * the rcu_start_batch()
447 smp_rmb();
449 if (!rcp->next_pending) {
450 /* and start it/schedule start if it's a new batch */
451 spin_lock(&rcp->lock);
452 rcp->next_pending = 1;
453 rcu_start_batch(rcp);
454 spin_unlock(&rcp->lock);
458 rcu_check_quiescent_state(rcp, rdp);
459 if (rdp->donelist)
460 rcu_do_batch(rdp);
463 static void rcu_process_callbacks(unsigned long unused)
465 __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
466 __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
469 static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
471 /* This cpu has pending rcu entries and the grace period
472 * for them has completed.
474 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch))
475 return 1;
477 /* This cpu has no pending entries, but there are new entries */
478 if (!rdp->curlist && rdp->nxtlist)
479 return 1;
481 /* This cpu has finished callbacks to invoke */
482 if (rdp->donelist)
483 return 1;
485 /* The rcu core waits for a quiescent state from the cpu */
486 if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
487 return 1;
489 /* nothing to do */
490 return 0;
494 * Check to see if there is any immediate RCU-related work to be done
495 * by the current CPU, returning 1 if so. This function is part of the
496 * RCU implementation; it is -not- an exported member of the RCU API.
498 int rcu_pending(int cpu)
500 return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
501 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
505 * Check to see if any future RCU-related work will need to be done
506 * by the current CPU, even if none need be done immediately, returning
507 * 1 if so. This function is part of the RCU implementation; it is -not-
508 * an exported member of the RCU API.
510 int rcu_needs_cpu(int cpu)
512 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
513 struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
515 return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
518 void rcu_check_callbacks(int cpu, int user)
520 if (user ||
521 (idle_cpu(cpu) && !in_softirq() &&
522 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
523 rcu_qsctr_inc(cpu);
524 rcu_bh_qsctr_inc(cpu);
525 } else if (!in_softirq())
526 rcu_bh_qsctr_inc(cpu);
527 tasklet_schedule(&per_cpu(rcu_tasklet, cpu));
530 static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
531 struct rcu_data *rdp)
533 memset(rdp, 0, sizeof(*rdp));
534 rdp->curtail = &rdp->curlist;
535 rdp->nxttail = &rdp->nxtlist;
536 rdp->donetail = &rdp->donelist;
537 rdp->quiescbatch = rcp->completed;
538 rdp->qs_pending = 0;
539 rdp->cpu = cpu;
540 rdp->blimit = blimit;
543 static void __devinit rcu_online_cpu(int cpu)
545 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
546 struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
548 rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
549 rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
550 tasklet_init(&per_cpu(rcu_tasklet, cpu), rcu_process_callbacks, 0UL);
553 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
554 unsigned long action, void *hcpu)
556 long cpu = (long)hcpu;
557 switch (action) {
558 case CPU_UP_PREPARE:
559 rcu_online_cpu(cpu);
560 break;
561 case CPU_DEAD:
562 rcu_offline_cpu(cpu);
563 break;
564 default:
565 break;
567 return NOTIFY_OK;
570 static struct notifier_block __cpuinitdata rcu_nb = {
571 .notifier_call = rcu_cpu_notify,
575 * Initializes rcu mechanism. Assumed to be called early.
576 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
577 * Note that rcu_qsctr and friends are implicitly
578 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
580 void __init rcu_init(void)
582 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
583 (void *)(long)smp_processor_id());
584 /* Register notifier for non-boot CPUs */
585 register_cpu_notifier(&rcu_nb);
588 struct rcu_synchronize {
589 struct rcu_head head;
590 struct completion completion;
593 /* Because of FASTCALL declaration of complete, we use this wrapper */
594 static void wakeme_after_rcu(struct rcu_head *head)
596 struct rcu_synchronize *rcu;
598 rcu = container_of(head, struct rcu_synchronize, head);
599 complete(&rcu->completion);
603 * synchronize_rcu - wait until a grace period has elapsed.
605 * Control will return to the caller some time after a full grace
606 * period has elapsed, in other words after all currently executing RCU
607 * read-side critical sections have completed. RCU read-side critical
608 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
609 * and may be nested.
611 * If your read-side code is not protected by rcu_read_lock(), do -not-
612 * use synchronize_rcu().
614 void synchronize_rcu(void)
616 struct rcu_synchronize rcu;
618 init_completion(&rcu.completion);
619 /* Will wake me after RCU finished */
620 call_rcu(&rcu.head, wakeme_after_rcu);
622 /* Wait for it */
623 wait_for_completion(&rcu.completion);
626 module_param(blimit, int, 0);
627 module_param(qhimark, int, 0);
628 module_param(qlowmark, int, 0);
629 EXPORT_SYMBOL_GPL(rcu_batches_completed);
630 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
631 EXPORT_SYMBOL_GPL(call_rcu);
632 EXPORT_SYMBOL_GPL(call_rcu_bh);
633 EXPORT_SYMBOL_GPL(synchronize_rcu);