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 IBM Corporation, 2001
19 *
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 *
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)
28 *
29 * For detailed explanation of Read-Copy Update mechanism see -
30 * Documentation/RCU
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/cpu.h>
49 #include <linux/mutex.h>
50 #include <linux/time.h>
51
52 #ifdef CONFIG_DEBUG_LOCK_ALLOC
53 static struct lock_class_key rcu_lock_key;
54 struct lockdep_map rcu_lock_map =
55 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
56 EXPORT_SYMBOL_GPL(rcu_lock_map);
57 #endif
58
59
60 /* Definition for rcupdate control block. */
61 static struct rcu_ctrlblk rcu_ctrlblk = {
62 .cur = -300,
63 .completed = -300,
64 .pending = -300,
65 .lock = __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk.lock),
66 .cpumask = CPU_BITS_NONE,
67 };
68 static struct rcu_ctrlblk rcu_bh_ctrlblk = {
69 .cur = -300,
70 .completed = -300,
71 .pending = -300,
72 .lock = __SPIN_LOCK_UNLOCKED(&rcu_bh_ctrlblk.lock),
73 .cpumask = CPU_BITS_NONE,
74 };
75
76 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
77 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
78
79 static int blimit = 10;
80 static int qhimark = 10000;
81 static int qlowmark = 100;
82
83 #ifdef CONFIG_SMP
force_quiescent_state(struct rcu_data * rdp,struct rcu_ctrlblk * rcp)84 static void force_quiescent_state(struct rcu_data *rdp,
85 struct rcu_ctrlblk *rcp)
86 {
87 int cpu;
88 unsigned long flags;
89
90 set_need_resched();
91 spin_lock_irqsave(&rcp->lock, flags);
92 if (unlikely(!rcp->signaled)) {
93 rcp->signaled = 1;
94 /*
95 * Don't send IPI to itself. With irqs disabled,
96 * rdp->cpu is the current cpu.
97 *
98 * cpu_online_mask is updated by the _cpu_down()
99 * using __stop_machine(). Since we're in irqs disabled
100 * section, __stop_machine() is not exectuting, hence
101 * the cpu_online_mask is stable.
102 *
103 * However, a cpu might have been offlined _just_ before
104 * we disabled irqs while entering here.
105 * And rcu subsystem might not yet have handled the CPU_DEAD
106 * notification, leading to the offlined cpu's bit
107 * being set in the rcp->cpumask.
108 *
109 * Hence cpumask = (rcp->cpumask & cpu_online_mask) to prevent
110 * sending smp_reschedule() to an offlined CPU.
111 */
112 for_each_cpu_and(cpu,
113 to_cpumask(rcp->cpumask), cpu_online_mask) {
114 if (cpu != rdp->cpu)
115 smp_send_reschedule(cpu);
116 }
117 }
118 spin_unlock_irqrestore(&rcp->lock, flags);
119 }
120 #else
force_quiescent_state(struct rcu_data * rdp,struct rcu_ctrlblk * rcp)121 static inline void force_quiescent_state(struct rcu_data *rdp,
122 struct rcu_ctrlblk *rcp)
123 {
124 set_need_resched();
125 }
126 #endif
127
__call_rcu(struct rcu_head * head,struct rcu_ctrlblk * rcp,struct rcu_data * rdp)128 static void __call_rcu(struct rcu_head *head, struct rcu_ctrlblk *rcp,
129 struct rcu_data *rdp)
130 {
131 long batch;
132
133 head->next = NULL;
134 smp_mb(); /* Read of rcu->cur must happen after any change by caller. */
135
136 /*
137 * Determine the batch number of this callback.
138 *
139 * Using ACCESS_ONCE to avoid the following error when gcc eliminates
140 * local variable "batch" and emits codes like this:
141 * 1) rdp->batch = rcp->cur + 1 # gets old value
142 * ......
143 * 2)rcu_batch_after(rcp->cur + 1, rdp->batch) # gets new value
144 * then [*nxttail[0], *nxttail[1]) may contain callbacks
145 * that batch# = rdp->batch, see the comment of struct rcu_data.
146 */
147 batch = ACCESS_ONCE(rcp->cur) + 1;
148
149 if (rdp->nxtlist && rcu_batch_after(batch, rdp->batch)) {
150 /* process callbacks */
151 rdp->nxttail[0] = rdp->nxttail[1];
152 rdp->nxttail[1] = rdp->nxttail[2];
153 if (rcu_batch_after(batch - 1, rdp->batch))
154 rdp->nxttail[0] = rdp->nxttail[2];
155 }
156
157 rdp->batch = batch;
158 *rdp->nxttail[2] = head;
159 rdp->nxttail[2] = &head->next;
160
161 if (unlikely(++rdp->qlen > qhimark)) {
162 rdp->blimit = INT_MAX;
163 force_quiescent_state(rdp, &rcu_ctrlblk);
164 }
165 }
166
167 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
168
record_gp_stall_check_time(struct rcu_ctrlblk * rcp)169 static void record_gp_stall_check_time(struct rcu_ctrlblk *rcp)
170 {
171 rcp->gp_start = jiffies;
172 rcp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
173 }
174
print_other_cpu_stall(struct rcu_ctrlblk * rcp)175 static void print_other_cpu_stall(struct rcu_ctrlblk *rcp)
176 {
177 int cpu;
178 long delta;
179 unsigned long flags;
180
181 /* Only let one CPU complain about others per time interval. */
182
183 spin_lock_irqsave(&rcp->lock, flags);
184 delta = jiffies - rcp->jiffies_stall;
185 if (delta < 2 || rcp->cur != rcp->completed) {
186 spin_unlock_irqrestore(&rcp->lock, flags);
187 return;
188 }
189 rcp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
190 spin_unlock_irqrestore(&rcp->lock, flags);
191
192 /* OK, time to rat on our buddy... */
193
194 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
195 for_each_possible_cpu(cpu) {
196 if (cpumask_test_cpu(cpu, to_cpumask(rcp->cpumask)))
197 printk(" %d", cpu);
198 }
199 printk(" (detected by %d, t=%ld jiffies)\n",
200 smp_processor_id(), (long)(jiffies - rcp->gp_start));
201 }
202
print_cpu_stall(struct rcu_ctrlblk * rcp)203 static void print_cpu_stall(struct rcu_ctrlblk *rcp)
204 {
205 unsigned long flags;
206
207 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu/%lu jiffies)\n",
208 smp_processor_id(), jiffies,
209 jiffies - rcp->gp_start);
210 dump_stack();
211 spin_lock_irqsave(&rcp->lock, flags);
212 if ((long)(jiffies - rcp->jiffies_stall) >= 0)
213 rcp->jiffies_stall =
214 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
215 spin_unlock_irqrestore(&rcp->lock, flags);
216 set_need_resched(); /* kick ourselves to get things going. */
217 }
218
check_cpu_stall(struct rcu_ctrlblk * rcp)219 static void check_cpu_stall(struct rcu_ctrlblk *rcp)
220 {
221 long delta;
222
223 delta = jiffies - rcp->jiffies_stall;
224 if (cpumask_test_cpu(smp_processor_id(), to_cpumask(rcp->cpumask)) &&
225 delta >= 0) {
226
227 /* We haven't checked in, so go dump stack. */
228 print_cpu_stall(rcp);
229
230 } else if (rcp->cur != rcp->completed && delta >= 2) {
231
232 /* They had two seconds to dump stack, so complain. */
233 print_other_cpu_stall(rcp);
234 }
235 }
236
237 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
238
record_gp_stall_check_time(struct rcu_ctrlblk * rcp)239 static void record_gp_stall_check_time(struct rcu_ctrlblk *rcp)
240 {
241 }
242
check_cpu_stall(struct rcu_ctrlblk * rcp)243 static inline void check_cpu_stall(struct rcu_ctrlblk *rcp)
244 {
245 }
246
247 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
248
249 /**
250 * call_rcu - Queue an RCU callback for invocation after a grace period.
251 * @head: structure to be used for queueing the RCU updates.
252 * @func: actual update function to be invoked after the grace period
253 *
254 * The update function will be invoked some time after a full grace
255 * period elapses, in other words after all currently executing RCU
256 * read-side critical sections have completed. RCU read-side critical
257 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
258 * and may be nested.
259 */
call_rcu(struct rcu_head * head,void (* func)(struct rcu_head * rcu))260 void call_rcu(struct rcu_head *head,
261 void (*func)(struct rcu_head *rcu))
262 {
263 unsigned long flags;
264
265 head->func = func;
266 local_irq_save(flags);
267 __call_rcu(head, &rcu_ctrlblk, &__get_cpu_var(rcu_data));
268 local_irq_restore(flags);
269 }
270 EXPORT_SYMBOL_GPL(call_rcu);
271
272 /**
273 * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
274 * @head: structure to be used for queueing the RCU updates.
275 * @func: actual update function to be invoked after the grace period
276 *
277 * The update function will be invoked some time after a full grace
278 * period elapses, in other words after all currently executing RCU
279 * read-side critical sections have completed. call_rcu_bh() assumes
280 * that the read-side critical sections end on completion of a softirq
281 * handler. This means that read-side critical sections in process
282 * context must not be interrupted by softirqs. This interface is to be
283 * used when most of the read-side critical sections are in softirq context.
284 * RCU read-side critical sections are delimited by rcu_read_lock() and
285 * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
286 * and rcu_read_unlock_bh(), if in process context. These may be nested.
287 */
call_rcu_bh(struct rcu_head * head,void (* func)(struct rcu_head * rcu))288 void call_rcu_bh(struct rcu_head *head,
289 void (*func)(struct rcu_head *rcu))
290 {
291 unsigned long flags;
292
293 head->func = func;
294 local_irq_save(flags);
295 __call_rcu(head, &rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
296 local_irq_restore(flags);
297 }
298 EXPORT_SYMBOL_GPL(call_rcu_bh);
299
300 /*
301 * Return the number of RCU batches processed thus far. Useful
302 * for debug and statistics.
303 */
rcu_batches_completed(void)304 long rcu_batches_completed(void)
305 {
306 return rcu_ctrlblk.completed;
307 }
308 EXPORT_SYMBOL_GPL(rcu_batches_completed);
309
310 /*
311 * Return the number of RCU batches processed thus far. Useful
312 * for debug and statistics.
313 */
rcu_batches_completed_bh(void)314 long rcu_batches_completed_bh(void)
315 {
316 return rcu_bh_ctrlblk.completed;
317 }
318 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
319
320 /* Raises the softirq for processing rcu_callbacks. */
raise_rcu_softirq(void)321 static inline void raise_rcu_softirq(void)
322 {
323 raise_softirq(RCU_SOFTIRQ);
324 }
325
326 /*
327 * Invoke the completed RCU callbacks. They are expected to be in
328 * a per-cpu list.
329 */
rcu_do_batch(struct rcu_data * rdp)330 static void rcu_do_batch(struct rcu_data *rdp)
331 {
332 unsigned long flags;
333 struct rcu_head *next, *list;
334 int count = 0;
335
336 list = rdp->donelist;
337 while (list) {
338 next = list->next;
339 prefetch(next);
340 list->func(list);
341 list = next;
342 if (++count >= rdp->blimit)
343 break;
344 }
345 rdp->donelist = list;
346
347 local_irq_save(flags);
348 rdp->qlen -= count;
349 local_irq_restore(flags);
350 if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark)
351 rdp->blimit = blimit;
352
353 if (!rdp->donelist)
354 rdp->donetail = &rdp->donelist;
355 else
356 raise_rcu_softirq();
357 }
358
359 /*
360 * Grace period handling:
361 * The grace period handling consists out of two steps:
362 * - A new grace period is started.
363 * This is done by rcu_start_batch. The start is not broadcasted to
364 * all cpus, they must pick this up by comparing rcp->cur with
365 * rdp->quiescbatch. All cpus are recorded in the
366 * rcu_ctrlblk.cpumask bitmap.
367 * - All cpus must go through a quiescent state.
368 * Since the start of the grace period is not broadcasted, at least two
369 * calls to rcu_check_quiescent_state are required:
370 * The first call just notices that a new grace period is running. The
371 * following calls check if there was a quiescent state since the beginning
372 * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
373 * the bitmap is empty, then the grace period is completed.
374 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
375 * period (if necessary).
376 */
377
378 /*
379 * Register a new batch of callbacks, and start it up if there is currently no
380 * active batch and the batch to be registered has not already occurred.
381 * Caller must hold rcu_ctrlblk.lock.
382 */
rcu_start_batch(struct rcu_ctrlblk * rcp)383 static void rcu_start_batch(struct rcu_ctrlblk *rcp)
384 {
385 if (rcp->cur != rcp->pending &&
386 rcp->completed == rcp->cur) {
387 rcp->cur++;
388 record_gp_stall_check_time(rcp);
389
390 /*
391 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
392 * Barrier Otherwise it can cause tickless idle CPUs to be
393 * included in rcp->cpumask, which will extend graceperiods
394 * unnecessarily.
395 */
396 smp_mb();
397 cpumask_andnot(to_cpumask(rcp->cpumask),
398 cpu_online_mask, nohz_cpu_mask);
399
400 rcp->signaled = 0;
401 }
402 }
403
404 /*
405 * cpu went through a quiescent state since the beginning of the grace period.
406 * Clear it from the cpu mask and complete the grace period if it was the last
407 * cpu. Start another grace period if someone has further entries pending
408 */
cpu_quiet(int cpu,struct rcu_ctrlblk * rcp)409 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
410 {
411 cpumask_clear_cpu(cpu, to_cpumask(rcp->cpumask));
412 if (cpumask_empty(to_cpumask(rcp->cpumask))) {
413 /* batch completed ! */
414 rcp->completed = rcp->cur;
415 rcu_start_batch(rcp);
416 }
417 }
418
419 /*
420 * Check if the cpu has gone through a quiescent state (say context
421 * switch). If so and if it already hasn't done so in this RCU
422 * quiescent cycle, then indicate that it has done so.
423 */
rcu_check_quiescent_state(struct rcu_ctrlblk * rcp,struct rcu_data * rdp)424 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
425 struct rcu_data *rdp)
426 {
427 unsigned long flags;
428
429 if (rdp->quiescbatch != rcp->cur) {
430 /* start new grace period: */
431 rdp->qs_pending = 1;
432 rdp->passed_quiesc = 0;
433 rdp->quiescbatch = rcp->cur;
434 return;
435 }
436
437 /* Grace period already completed for this cpu?
438 * qs_pending is checked instead of the actual bitmap to avoid
439 * cacheline trashing.
440 */
441 if (!rdp->qs_pending)
442 return;
443
444 /*
445 * Was there a quiescent state since the beginning of the grace
446 * period? If no, then exit and wait for the next call.
447 */
448 if (!rdp->passed_quiesc)
449 return;
450 rdp->qs_pending = 0;
451
452 spin_lock_irqsave(&rcp->lock, flags);
453 /*
454 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
455 * during cpu startup. Ignore the quiescent state.
456 */
457 if (likely(rdp->quiescbatch == rcp->cur))
458 cpu_quiet(rdp->cpu, rcp);
459
460 spin_unlock_irqrestore(&rcp->lock, flags);
461 }
462
463
464 #ifdef CONFIG_HOTPLUG_CPU
465
466 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
467 * locking requirements, the list it's pulling from has to belong to a cpu
468 * which is dead and hence not processing interrupts.
469 */
rcu_move_batch(struct rcu_data * this_rdp,struct rcu_head * list,struct rcu_head ** tail,long batch)470 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
471 struct rcu_head **tail, long batch)
472 {
473 unsigned long flags;
474
475 if (list) {
476 local_irq_save(flags);
477 this_rdp->batch = batch;
478 *this_rdp->nxttail[2] = list;
479 this_rdp->nxttail[2] = tail;
480 local_irq_restore(flags);
481 }
482 }
483
__rcu_offline_cpu(struct rcu_data * this_rdp,struct rcu_ctrlblk * rcp,struct rcu_data * rdp)484 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
485 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
486 {
487 unsigned long flags;
488
489 /*
490 * if the cpu going offline owns the grace period
491 * we can block indefinitely waiting for it, so flush
492 * it here
493 */
494 spin_lock_irqsave(&rcp->lock, flags);
495 if (rcp->cur != rcp->completed)
496 cpu_quiet(rdp->cpu, rcp);
497 rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail, rcp->cur + 1);
498 rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail[2], rcp->cur + 1);
499 spin_unlock(&rcp->lock);
500
501 this_rdp->qlen += rdp->qlen;
502 local_irq_restore(flags);
503 }
504
rcu_offline_cpu(int cpu)505 static void rcu_offline_cpu(int cpu)
506 {
507 struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
508 struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
509
510 __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
511 &per_cpu(rcu_data, cpu));
512 __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
513 &per_cpu(rcu_bh_data, cpu));
514 put_cpu_var(rcu_data);
515 put_cpu_var(rcu_bh_data);
516 }
517
518 #else
519
rcu_offline_cpu(int cpu)520 static void rcu_offline_cpu(int cpu)
521 {
522 }
523
524 #endif
525
526 /*
527 * This does the RCU processing work from softirq context.
528 */
__rcu_process_callbacks(struct rcu_ctrlblk * rcp,struct rcu_data * rdp)529 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
530 struct rcu_data *rdp)
531 {
532 unsigned long flags;
533 long completed_snap;
534
535 if (rdp->nxtlist) {
536 local_irq_save(flags);
537 completed_snap = ACCESS_ONCE(rcp->completed);
538
539 /*
540 * move the other grace-period-completed entries to
541 * [rdp->nxtlist, *rdp->nxttail[0]) temporarily
542 */
543 if (!rcu_batch_before(completed_snap, rdp->batch))
544 rdp->nxttail[0] = rdp->nxttail[1] = rdp->nxttail[2];
545 else if (!rcu_batch_before(completed_snap, rdp->batch - 1))
546 rdp->nxttail[0] = rdp->nxttail[1];
547
548 /*
549 * the grace period for entries in
550 * [rdp->nxtlist, *rdp->nxttail[0]) has completed and
551 * move these entries to donelist
552 */
553 if (rdp->nxttail[0] != &rdp->nxtlist) {
554 *rdp->donetail = rdp->nxtlist;
555 rdp->donetail = rdp->nxttail[0];
556 rdp->nxtlist = *rdp->nxttail[0];
557 *rdp->donetail = NULL;
558
559 if (rdp->nxttail[1] == rdp->nxttail[0])
560 rdp->nxttail[1] = &rdp->nxtlist;
561 if (rdp->nxttail[2] == rdp->nxttail[0])
562 rdp->nxttail[2] = &rdp->nxtlist;
563 rdp->nxttail[0] = &rdp->nxtlist;
564 }
565
566 local_irq_restore(flags);
567
568 if (rcu_batch_after(rdp->batch, rcp->pending)) {
569 unsigned long flags2;
570
571 /* and start it/schedule start if it's a new batch */
572 spin_lock_irqsave(&rcp->lock, flags2);
573 if (rcu_batch_after(rdp->batch, rcp->pending)) {
574 rcp->pending = rdp->batch;
575 rcu_start_batch(rcp);
576 }
577 spin_unlock_irqrestore(&rcp->lock, flags2);
578 }
579 }
580
581 rcu_check_quiescent_state(rcp, rdp);
582 if (rdp->donelist)
583 rcu_do_batch(rdp);
584 }
585
rcu_process_callbacks(struct softirq_action * unused)586 static void rcu_process_callbacks(struct softirq_action *unused)
587 {
588 /*
589 * Memory references from any prior RCU read-side critical sections
590 * executed by the interrupted code must be see before any RCU
591 * grace-period manupulations below.
592 */
593
594 smp_mb(); /* See above block comment. */
595
596 __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
597 __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
598
599 /*
600 * Memory references from any later RCU read-side critical sections
601 * executed by the interrupted code must be see after any RCU
602 * grace-period manupulations above.
603 */
604
605 smp_mb(); /* See above block comment. */
606 }
607
__rcu_pending(struct rcu_ctrlblk * rcp,struct rcu_data * rdp)608 static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
609 {
610 /* Check for CPU stalls, if enabled. */
611 check_cpu_stall(rcp);
612
613 if (rdp->nxtlist) {
614 long completed_snap = ACCESS_ONCE(rcp->completed);
615
616 /*
617 * This cpu has pending rcu entries and the grace period
618 * for them has completed.
619 */
620 if (!rcu_batch_before(completed_snap, rdp->batch))
621 return 1;
622 if (!rcu_batch_before(completed_snap, rdp->batch - 1) &&
623 rdp->nxttail[0] != rdp->nxttail[1])
624 return 1;
625 if (rdp->nxttail[0] != &rdp->nxtlist)
626 return 1;
627
628 /*
629 * This cpu has pending rcu entries and the new batch
630 * for then hasn't been started nor scheduled start
631 */
632 if (rcu_batch_after(rdp->batch, rcp->pending))
633 return 1;
634 }
635
636 /* This cpu has finished callbacks to invoke */
637 if (rdp->donelist)
638 return 1;
639
640 /* The rcu core waits for a quiescent state from the cpu */
641 if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
642 return 1;
643
644 /* nothing to do */
645 return 0;
646 }
647
648 /*
649 * Check to see if there is any immediate RCU-related work to be done
650 * by the current CPU, returning 1 if so. This function is part of the
651 * RCU implementation; it is -not- an exported member of the RCU API.
652 */
rcu_pending(int cpu)653 int rcu_pending(int cpu)
654 {
655 return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
656 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
657 }
658
659 /*
660 * Check to see if any future RCU-related work will need to be done
661 * by the current CPU, even if none need be done immediately, returning
662 * 1 if so. This function is part of the RCU implementation; it is -not-
663 * an exported member of the RCU API.
664 */
rcu_needs_cpu(int cpu)665 int rcu_needs_cpu(int cpu)
666 {
667 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
668 struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
669
670 return !!rdp->nxtlist || !!rdp_bh->nxtlist || rcu_pending(cpu);
671 }
672
673 /*
674 * Top-level function driving RCU grace-period detection, normally
675 * invoked from the scheduler-clock interrupt. This function simply
676 * increments counters that are read only from softirq by this same
677 * CPU, so there are no memory barriers required.
678 */
rcu_check_callbacks(int cpu,int user)679 void rcu_check_callbacks(int cpu, int user)
680 {
681 if (user ||
682 (idle_cpu(cpu) && rcu_scheduler_active &&
683 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
684
685 /*
686 * Get here if this CPU took its interrupt from user
687 * mode or from the idle loop, and if this is not a
688 * nested interrupt. In this case, the CPU is in
689 * a quiescent state, so count it.
690 *
691 * Also do a memory barrier. This is needed to handle
692 * the case where writes from a preempt-disable section
693 * of code get reordered into schedule() by this CPU's
694 * write buffer. The memory barrier makes sure that
695 * the rcu_qsctr_inc() and rcu_bh_qsctr_inc() are see
696 * by other CPUs to happen after any such write.
697 */
698
699 smp_mb(); /* See above block comment. */
700 rcu_qsctr_inc(cpu);
701 rcu_bh_qsctr_inc(cpu);
702
703 } else if (!in_softirq()) {
704
705 /*
706 * Get here if this CPU did not take its interrupt from
707 * softirq, in other words, if it is not interrupting
708 * a rcu_bh read-side critical section. This is an _bh
709 * critical section, so count it. The memory barrier
710 * is needed for the same reason as is the above one.
711 */
712
713 smp_mb(); /* See above block comment. */
714 rcu_bh_qsctr_inc(cpu);
715 }
716 raise_rcu_softirq();
717 }
718
rcu_init_percpu_data(int cpu,struct rcu_ctrlblk * rcp,struct rcu_data * rdp)719 static void __cpuinit rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
720 struct rcu_data *rdp)
721 {
722 unsigned long flags;
723
724 spin_lock_irqsave(&rcp->lock, flags);
725 memset(rdp, 0, sizeof(*rdp));
726 rdp->nxttail[0] = rdp->nxttail[1] = rdp->nxttail[2] = &rdp->nxtlist;
727 rdp->donetail = &rdp->donelist;
728 rdp->quiescbatch = rcp->completed;
729 rdp->qs_pending = 0;
730 rdp->cpu = cpu;
731 rdp->blimit = blimit;
732 spin_unlock_irqrestore(&rcp->lock, flags);
733 }
734
rcu_online_cpu(int cpu)735 static void __cpuinit rcu_online_cpu(int cpu)
736 {
737 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
738 struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
739
740 rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
741 rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
742 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
743 }
744
rcu_cpu_notify(struct notifier_block * self,unsigned long action,void * hcpu)745 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
746 unsigned long action, void *hcpu)
747 {
748 long cpu = (long)hcpu;
749
750 switch (action) {
751 case CPU_UP_PREPARE:
752 case CPU_UP_PREPARE_FROZEN:
753 rcu_online_cpu(cpu);
754 break;
755 case CPU_DEAD:
756 case CPU_DEAD_FROZEN:
757 rcu_offline_cpu(cpu);
758 break;
759 default:
760 break;
761 }
762 return NOTIFY_OK;
763 }
764
765 static struct notifier_block __cpuinitdata rcu_nb = {
766 .notifier_call = rcu_cpu_notify,
767 };
768
769 /*
770 * Initializes rcu mechanism. Assumed to be called early.
771 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
772 * Note that rcu_qsctr and friends are implicitly
773 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
774 */
__rcu_init(void)775 void __init __rcu_init(void)
776 {
777 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
778 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
779 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
780 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
781 (void *)(long)smp_processor_id());
782 /* Register notifier for non-boot CPUs */
783 register_cpu_notifier(&rcu_nb);
784 }
785
786 module_param(blimit, int, 0);
787 module_param(qhimark, int, 0);
788 module_param(qlowmark, int, 0);
789