aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/rcupreempt.c
diff options
context:
space:
mode:
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>2009-08-22 16:56:53 -0400
committerIngo Molnar <mingo@elte.hu>2009-08-23 04:32:40 -0400
commit6b3ef48adf847f7adf11c870e3ffacac150f1564 (patch)
treee1403ce515bf00ade99ec806f6ab6b6db999aa0b /kernel/rcupreempt.c
parentf41d911f8c49a5d65c86504c19e8204bb605c4fd (diff)
rcu: Remove CONFIG_PREEMPT_RCU
Now that CONFIG_TREE_PREEMPT_RCU is in place, there is no further need for CONFIG_PREEMPT_RCU. Remove it, along with whatever subtle bugs it may (or may not) contain. Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: laijs@cn.fujitsu.com Cc: dipankar@in.ibm.com Cc: akpm@linux-foundation.org Cc: mathieu.desnoyers@polymtl.ca Cc: josht@linux.vnet.ibm.com Cc: dvhltc@us.ibm.com Cc: niv@us.ibm.com Cc: peterz@infradead.org Cc: rostedt@goodmis.org LKML-Reference: <125097461396-git-send-email-> Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'kernel/rcupreempt.c')
-rw-r--r--kernel/rcupreempt.c1518
1 files changed, 0 insertions, 1518 deletions
diff --git a/kernel/rcupreempt.c b/kernel/rcupreempt.c
deleted file mode 100644
index 0053ce56e326..000000000000
--- a/kernel/rcupreempt.c
+++ /dev/null
@@ -1,1518 +0,0 @@
1/*
2 * Read-Copy Update mechanism for mutual exclusion, realtime implementation
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, 2006
19 *
20 * Authors: Paul E. McKenney <paulmck@us.ibm.com>
21 * With thanks to Esben Nielsen, Bill Huey, and Ingo Molnar
22 * for pushing me away from locks and towards counters, and
23 * to Suparna Bhattacharya for pushing me completely away
24 * from atomic instructions on the read side.
25 *
26 * - Added handling of Dynamic Ticks
27 * Copyright 2007 - Paul E. Mckenney <paulmck@us.ibm.com>
28 * - Steven Rostedt <srostedt@redhat.com>
29 *
30 * Papers: http://www.rdrop.com/users/paulmck/RCU
31 *
32 * Design Document: http://lwn.net/Articles/253651/
33 *
34 * For detailed explanation of Read-Copy Update mechanism see -
35 * Documentation/RCU/ *.txt
36 *
37 */
38#include <linux/types.h>
39#include <linux/kernel.h>
40#include <linux/init.h>
41#include <linux/spinlock.h>
42#include <linux/smp.h>
43#include <linux/rcupdate.h>
44#include <linux/interrupt.h>
45#include <linux/sched.h>
46#include <asm/atomic.h>
47#include <linux/bitops.h>
48#include <linux/module.h>
49#include <linux/kthread.h>
50#include <linux/completion.h>
51#include <linux/moduleparam.h>
52#include <linux/percpu.h>
53#include <linux/notifier.h>
54#include <linux/cpu.h>
55#include <linux/random.h>
56#include <linux/delay.h>
57#include <linux/cpumask.h>
58#include <linux/rcupreempt_trace.h>
59#include <asm/byteorder.h>
60
61/*
62 * PREEMPT_RCU data structures.
63 */
64
65/*
66 * GP_STAGES specifies the number of times the state machine has
67 * to go through the all the rcu_try_flip_states (see below)
68 * in a single Grace Period.
69 *
70 * GP in GP_STAGES stands for Grace Period ;)
71 */
72#define GP_STAGES 2
73struct rcu_data {
74 spinlock_t lock; /* Protect rcu_data fields. */
75 long completed; /* Number of last completed batch. */
76 int waitlistcount;
77 struct rcu_head *nextlist;
78 struct rcu_head **nexttail;
79 struct rcu_head *waitlist[GP_STAGES];
80 struct rcu_head **waittail[GP_STAGES];
81 struct rcu_head *donelist; /* from waitlist & waitschedlist */
82 struct rcu_head **donetail;
83 long rcu_flipctr[2];
84 struct rcu_head *nextschedlist;
85 struct rcu_head **nextschedtail;
86 struct rcu_head *waitschedlist;
87 struct rcu_head **waitschedtail;
88 int rcu_sched_sleeping;
89#ifdef CONFIG_RCU_TRACE
90 struct rcupreempt_trace trace;
91#endif /* #ifdef CONFIG_RCU_TRACE */
92};
93
94/*
95 * States for rcu_try_flip() and friends.
96 */
97
98enum rcu_try_flip_states {
99
100 /*
101 * Stay here if nothing is happening. Flip the counter if somthing
102 * starts happening. Denoted by "I"
103 */
104 rcu_try_flip_idle_state,
105
106 /*
107 * Wait here for all CPUs to notice that the counter has flipped. This
108 * prevents the old set of counters from ever being incremented once
109 * we leave this state, which in turn is necessary because we cannot
110 * test any individual counter for zero -- we can only check the sum.
111 * Denoted by "A".
112 */
113 rcu_try_flip_waitack_state,
114
115 /*
116 * Wait here for the sum of the old per-CPU counters to reach zero.
117 * Denoted by "Z".
118 */
119 rcu_try_flip_waitzero_state,
120
121 /*
122 * Wait here for each of the other CPUs to execute a memory barrier.
123 * This is necessary to ensure that these other CPUs really have
124 * completed executing their RCU read-side critical sections, despite
125 * their CPUs wildly reordering memory. Denoted by "M".
126 */
127 rcu_try_flip_waitmb_state,
128};
129
130/*
131 * States for rcu_ctrlblk.rcu_sched_sleep.
132 */
133
134enum rcu_sched_sleep_states {
135 rcu_sched_not_sleeping, /* Not sleeping, callbacks need GP. */
136 rcu_sched_sleep_prep, /* Thinking of sleeping, rechecking. */
137 rcu_sched_sleeping, /* Sleeping, awaken if GP needed. */
138};
139
140struct rcu_ctrlblk {
141 spinlock_t fliplock; /* Protect state-machine transitions. */
142 long completed; /* Number of last completed batch. */
143 enum rcu_try_flip_states rcu_try_flip_state; /* The current state of
144 the rcu state machine */
145 spinlock_t schedlock; /* Protect rcu_sched sleep state. */
146 enum rcu_sched_sleep_states sched_sleep; /* rcu_sched state. */
147 wait_queue_head_t sched_wq; /* Place for rcu_sched to sleep. */
148};
149
150struct rcu_dyntick_sched {
151 int dynticks;
152 int dynticks_snap;
153 int sched_qs;
154 int sched_qs_snap;
155 int sched_dynticks_snap;
156};
157
158static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_dyntick_sched, rcu_dyntick_sched) = {
159 .dynticks = 1,
160};
161
162static int rcu_pending(int cpu);
163
164void rcu_sched_qs(int cpu)
165{
166 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
167
168 rdssp->sched_qs++;
169}
170
171#ifdef CONFIG_NO_HZ
172
173void rcu_enter_nohz(void)
174{
175 static DEFINE_RATELIMIT_STATE(rs, 10 * HZ, 1);
176
177 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
178 __get_cpu_var(rcu_dyntick_sched).dynticks++;
179 WARN_ON_RATELIMIT(__get_cpu_var(rcu_dyntick_sched).dynticks & 0x1, &rs);
180}
181
182void rcu_exit_nohz(void)
183{
184 static DEFINE_RATELIMIT_STATE(rs, 10 * HZ, 1);
185
186 __get_cpu_var(rcu_dyntick_sched).dynticks++;
187 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
188 WARN_ON_RATELIMIT(!(__get_cpu_var(rcu_dyntick_sched).dynticks & 0x1),
189 &rs);
190}
191
192#endif /* CONFIG_NO_HZ */
193
194
195static DEFINE_PER_CPU(struct rcu_data, rcu_data);
196
197static struct rcu_ctrlblk rcu_ctrlblk = {
198 .fliplock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.fliplock),
199 .completed = 0,
200 .rcu_try_flip_state = rcu_try_flip_idle_state,
201 .schedlock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.schedlock),
202 .sched_sleep = rcu_sched_not_sleeping,
203 .sched_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rcu_ctrlblk.sched_wq),
204};
205
206static struct task_struct *rcu_sched_grace_period_task;
207
208#ifdef CONFIG_RCU_TRACE
209static char *rcu_try_flip_state_names[] =
210 { "idle", "waitack", "waitzero", "waitmb" };
211#endif /* #ifdef CONFIG_RCU_TRACE */
212
213static DECLARE_BITMAP(rcu_cpu_online_map, NR_CPUS) __read_mostly
214 = CPU_BITS_NONE;
215
216/*
217 * Enum and per-CPU flag to determine when each CPU has seen
218 * the most recent counter flip.
219 */
220
221enum rcu_flip_flag_values {
222 rcu_flip_seen, /* Steady/initial state, last flip seen. */
223 /* Only GP detector can update. */
224 rcu_flipped /* Flip just completed, need confirmation. */
225 /* Only corresponding CPU can update. */
226};
227static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_flip_flag_values, rcu_flip_flag)
228 = rcu_flip_seen;
229
230/*
231 * Enum and per-CPU flag to determine when each CPU has executed the
232 * needed memory barrier to fence in memory references from its last RCU
233 * read-side critical section in the just-completed grace period.
234 */
235
236enum rcu_mb_flag_values {
237 rcu_mb_done, /* Steady/initial state, no mb()s required. */
238 /* Only GP detector can update. */
239 rcu_mb_needed /* Flip just completed, need an mb(). */
240 /* Only corresponding CPU can update. */
241};
242static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_mb_flag_values, rcu_mb_flag)
243 = rcu_mb_done;
244
245/*
246 * RCU_DATA_ME: find the current CPU's rcu_data structure.
247 * RCU_DATA_CPU: find the specified CPU's rcu_data structure.
248 */
249#define RCU_DATA_ME() (&__get_cpu_var(rcu_data))
250#define RCU_DATA_CPU(cpu) (&per_cpu(rcu_data, cpu))
251
252/*
253 * Helper macro for tracing when the appropriate rcu_data is not
254 * cached in a local variable, but where the CPU number is so cached.
255 */
256#define RCU_TRACE_CPU(f, cpu) RCU_TRACE(f, &(RCU_DATA_CPU(cpu)->trace));
257
258/*
259 * Helper macro for tracing when the appropriate rcu_data is not
260 * cached in a local variable.
261 */
262#define RCU_TRACE_ME(f) RCU_TRACE(f, &(RCU_DATA_ME()->trace));
263
264/*
265 * Helper macro for tracing when the appropriate rcu_data is pointed
266 * to by a local variable.
267 */
268#define RCU_TRACE_RDP(f, rdp) RCU_TRACE(f, &((rdp)->trace));
269
270#define RCU_SCHED_BATCH_TIME (HZ / 50)
271
272/*
273 * Return the number of RCU batches processed thus far. Useful
274 * for debug and statistics.
275 */
276long rcu_batches_completed(void)
277{
278 return rcu_ctrlblk.completed;
279}
280EXPORT_SYMBOL_GPL(rcu_batches_completed);
281
282void __rcu_read_lock(void)
283{
284 int idx;
285 struct task_struct *t = current;
286 int nesting;
287
288 nesting = ACCESS_ONCE(t->rcu_read_lock_nesting);
289 if (nesting != 0) {
290
291 /* An earlier rcu_read_lock() covers us, just count it. */
292
293 t->rcu_read_lock_nesting = nesting + 1;
294
295 } else {
296 unsigned long flags;
297
298 /*
299 * We disable interrupts for the following reasons:
300 * - If we get scheduling clock interrupt here, and we
301 * end up acking the counter flip, it's like a promise
302 * that we will never increment the old counter again.
303 * Thus we will break that promise if that
304 * scheduling clock interrupt happens between the time
305 * we pick the .completed field and the time that we
306 * increment our counter.
307 *
308 * - We don't want to be preempted out here.
309 *
310 * NMIs can still occur, of course, and might themselves
311 * contain rcu_read_lock().
312 */
313
314 local_irq_save(flags);
315
316 /*
317 * Outermost nesting of rcu_read_lock(), so increment
318 * the current counter for the current CPU. Use volatile
319 * casts to prevent the compiler from reordering.
320 */
321
322 idx = ACCESS_ONCE(rcu_ctrlblk.completed) & 0x1;
323 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr[idx])++;
324
325 /*
326 * Now that the per-CPU counter has been incremented, we
327 * are protected from races with rcu_read_lock() invoked
328 * from NMI handlers on this CPU. We can therefore safely
329 * increment the nesting counter, relieving further NMIs
330 * of the need to increment the per-CPU counter.
331 */
332
333 ACCESS_ONCE(t->rcu_read_lock_nesting) = nesting + 1;
334
335 /*
336 * Now that we have preventing any NMIs from storing
337 * to the ->rcu_flipctr_idx, we can safely use it to
338 * remember which counter to decrement in the matching
339 * rcu_read_unlock().
340 */
341
342 ACCESS_ONCE(t->rcu_flipctr_idx) = idx;
343 local_irq_restore(flags);
344 }
345}
346EXPORT_SYMBOL_GPL(__rcu_read_lock);
347
348void __rcu_read_unlock(void)
349{
350 int idx;
351 struct task_struct *t = current;
352 int nesting;
353
354 nesting = ACCESS_ONCE(t->rcu_read_lock_nesting);
355 if (nesting > 1) {
356
357 /*
358 * We are still protected by the enclosing rcu_read_lock(),
359 * so simply decrement the counter.
360 */
361
362 t->rcu_read_lock_nesting = nesting - 1;
363
364 } else {
365 unsigned long flags;
366
367 /*
368 * Disable local interrupts to prevent the grace-period
369 * detection state machine from seeing us half-done.
370 * NMIs can still occur, of course, and might themselves
371 * contain rcu_read_lock() and rcu_read_unlock().
372 */
373
374 local_irq_save(flags);
375
376 /*
377 * Outermost nesting of rcu_read_unlock(), so we must
378 * decrement the current counter for the current CPU.
379 * This must be done carefully, because NMIs can
380 * occur at any point in this code, and any rcu_read_lock()
381 * and rcu_read_unlock() pairs in the NMI handlers
382 * must interact non-destructively with this code.
383 * Lots of volatile casts, and -very- careful ordering.
384 *
385 * Changes to this code, including this one, must be
386 * inspected, validated, and tested extremely carefully!!!
387 */
388
389 /*
390 * First, pick up the index.
391 */
392
393 idx = ACCESS_ONCE(t->rcu_flipctr_idx);
394
395 /*
396 * Now that we have fetched the counter index, it is
397 * safe to decrement the per-task RCU nesting counter.
398 * After this, any interrupts or NMIs will increment and
399 * decrement the per-CPU counters.
400 */
401 ACCESS_ONCE(t->rcu_read_lock_nesting) = nesting - 1;
402
403 /*
404 * It is now safe to decrement this task's nesting count.
405 * NMIs that occur after this statement will route their
406 * rcu_read_lock() calls through this "else" clause, and
407 * will thus start incrementing the per-CPU counter on
408 * their own. They will also clobber ->rcu_flipctr_idx,
409 * but that is OK, since we have already fetched it.
410 */
411
412 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr[idx])--;
413 local_irq_restore(flags);
414 }
415}
416EXPORT_SYMBOL_GPL(__rcu_read_unlock);
417
418/*
419 * If a global counter flip has occurred since the last time that we
420 * advanced callbacks, advance them. Hardware interrupts must be
421 * disabled when calling this function.
422 */
423static void __rcu_advance_callbacks(struct rcu_data *rdp)
424{
425 int cpu;
426 int i;
427 int wlc = 0;
428
429 if (rdp->completed != rcu_ctrlblk.completed) {
430 if (rdp->waitlist[GP_STAGES - 1] != NULL) {
431 *rdp->donetail = rdp->waitlist[GP_STAGES - 1];
432 rdp->donetail = rdp->waittail[GP_STAGES - 1];
433 RCU_TRACE_RDP(rcupreempt_trace_move2done, rdp);
434 }
435 for (i = GP_STAGES - 2; i >= 0; i--) {
436 if (rdp->waitlist[i] != NULL) {
437 rdp->waitlist[i + 1] = rdp->waitlist[i];
438 rdp->waittail[i + 1] = rdp->waittail[i];
439 wlc++;
440 } else {
441 rdp->waitlist[i + 1] = NULL;
442 rdp->waittail[i + 1] =
443 &rdp->waitlist[i + 1];
444 }
445 }
446 if (rdp->nextlist != NULL) {
447 rdp->waitlist[0] = rdp->nextlist;
448 rdp->waittail[0] = rdp->nexttail;
449 wlc++;
450 rdp->nextlist = NULL;
451 rdp->nexttail = &rdp->nextlist;
452 RCU_TRACE_RDP(rcupreempt_trace_move2wait, rdp);
453 } else {
454 rdp->waitlist[0] = NULL;
455 rdp->waittail[0] = &rdp->waitlist[0];
456 }
457 rdp->waitlistcount = wlc;
458 rdp->completed = rcu_ctrlblk.completed;
459 }
460
461 /*
462 * Check to see if this CPU needs to report that it has seen
463 * the most recent counter flip, thereby declaring that all
464 * subsequent rcu_read_lock() invocations will respect this flip.
465 */
466
467 cpu = raw_smp_processor_id();
468 if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
469 smp_mb(); /* Subsequent counter accesses must see new value */
470 per_cpu(rcu_flip_flag, cpu) = rcu_flip_seen;
471 smp_mb(); /* Subsequent RCU read-side critical sections */
472 /* seen -after- acknowledgement. */
473 }
474}
475
476#ifdef CONFIG_NO_HZ
477static DEFINE_PER_CPU(int, rcu_update_flag);
478
479/**
480 * rcu_irq_enter - Called from Hard irq handlers and NMI/SMI.
481 *
482 * If the CPU was idle with dynamic ticks active, this updates the
483 * rcu_dyntick_sched.dynticks to let the RCU handling know that the
484 * CPU is active.
485 */
486void rcu_irq_enter(void)
487{
488 int cpu = smp_processor_id();
489 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
490
491 if (per_cpu(rcu_update_flag, cpu))
492 per_cpu(rcu_update_flag, cpu)++;
493
494 /*
495 * Only update if we are coming from a stopped ticks mode
496 * (rcu_dyntick_sched.dynticks is even).
497 */
498 if (!in_interrupt() &&
499 (rdssp->dynticks & 0x1) == 0) {
500 /*
501 * The following might seem like we could have a race
502 * with NMI/SMIs. But this really isn't a problem.
503 * Here we do a read/modify/write, and the race happens
504 * when an NMI/SMI comes in after the read and before
505 * the write. But NMI/SMIs will increment this counter
506 * twice before returning, so the zero bit will not
507 * be corrupted by the NMI/SMI which is the most important
508 * part.
509 *
510 * The only thing is that we would bring back the counter
511 * to a postion that it was in during the NMI/SMI.
512 * But the zero bit would be set, so the rest of the
513 * counter would again be ignored.
514 *
515 * On return from the IRQ, the counter may have the zero
516 * bit be 0 and the counter the same as the return from
517 * the NMI/SMI. If the state machine was so unlucky to
518 * see that, it still doesn't matter, since all
519 * RCU read-side critical sections on this CPU would
520 * have already completed.
521 */
522 rdssp->dynticks++;
523 /*
524 * The following memory barrier ensures that any
525 * rcu_read_lock() primitives in the irq handler
526 * are seen by other CPUs to follow the above
527 * increment to rcu_dyntick_sched.dynticks. This is
528 * required in order for other CPUs to correctly
529 * determine when it is safe to advance the RCU
530 * grace-period state machine.
531 */
532 smp_mb(); /* see above block comment. */
533 /*
534 * Since we can't determine the dynamic tick mode from
535 * the rcu_dyntick_sched.dynticks after this routine,
536 * we use a second flag to acknowledge that we came
537 * from an idle state with ticks stopped.
538 */
539 per_cpu(rcu_update_flag, cpu)++;
540 /*
541 * If we take an NMI/SMI now, they will also increment
542 * the rcu_update_flag, and will not update the
543 * rcu_dyntick_sched.dynticks on exit. That is for
544 * this IRQ to do.
545 */
546 }
547}
548
549/**
550 * rcu_irq_exit - Called from exiting Hard irq context.
551 *
552 * If the CPU was idle with dynamic ticks active, update the
553 * rcu_dyntick_sched.dynticks to let the RCU handling be
554 * aware that the CPU is going back to idle with no ticks.
555 */
556void rcu_irq_exit(void)
557{
558 int cpu = smp_processor_id();
559 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
560
561 /*
562 * rcu_update_flag is set if we interrupted the CPU
563 * when it was idle with ticks stopped.
564 * Once this occurs, we keep track of interrupt nesting
565 * because a NMI/SMI could also come in, and we still
566 * only want the IRQ that started the increment of the
567 * rcu_dyntick_sched.dynticks to be the one that modifies
568 * it on exit.
569 */
570 if (per_cpu(rcu_update_flag, cpu)) {
571 if (--per_cpu(rcu_update_flag, cpu))
572 return;
573
574 /* This must match the interrupt nesting */
575 WARN_ON(in_interrupt());
576
577 /*
578 * If an NMI/SMI happens now we are still
579 * protected by the rcu_dyntick_sched.dynticks being odd.
580 */
581
582 /*
583 * The following memory barrier ensures that any
584 * rcu_read_unlock() primitives in the irq handler
585 * are seen by other CPUs to preceed the following
586 * increment to rcu_dyntick_sched.dynticks. This
587 * is required in order for other CPUs to determine
588 * when it is safe to advance the RCU grace-period
589 * state machine.
590 */
591 smp_mb(); /* see above block comment. */
592 rdssp->dynticks++;
593 WARN_ON(rdssp->dynticks & 0x1);
594 }
595}
596
597void rcu_nmi_enter(void)
598{
599 rcu_irq_enter();
600}
601
602void rcu_nmi_exit(void)
603{
604 rcu_irq_exit();
605}
606
607static void dyntick_save_progress_counter(int cpu)
608{
609 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
610
611 rdssp->dynticks_snap = rdssp->dynticks;
612}
613
614static inline int
615rcu_try_flip_waitack_needed(int cpu)
616{
617 long curr;
618 long snap;
619 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
620
621 curr = rdssp->dynticks;
622 snap = rdssp->dynticks_snap;
623 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
624
625 /*
626 * If the CPU remained in dynticks mode for the entire time
627 * and didn't take any interrupts, NMIs, SMIs, or whatever,
628 * then it cannot be in the middle of an rcu_read_lock(), so
629 * the next rcu_read_lock() it executes must use the new value
630 * of the counter. So we can safely pretend that this CPU
631 * already acknowledged the counter.
632 */
633
634 if ((curr == snap) && ((curr & 0x1) == 0))
635 return 0;
636
637 /*
638 * If the CPU passed through or entered a dynticks idle phase with
639 * no active irq handlers, then, as above, we can safely pretend
640 * that this CPU already acknowledged the counter.
641 */
642
643 if ((curr - snap) > 2 || (curr & 0x1) == 0)
644 return 0;
645
646 /* We need this CPU to explicitly acknowledge the counter flip. */
647
648 return 1;
649}
650
651static inline int
652rcu_try_flip_waitmb_needed(int cpu)
653{
654 long curr;
655 long snap;
656 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
657
658 curr = rdssp->dynticks;
659 snap = rdssp->dynticks_snap;
660 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
661
662 /*
663 * If the CPU remained in dynticks mode for the entire time
664 * and didn't take any interrupts, NMIs, SMIs, or whatever,
665 * then it cannot have executed an RCU read-side critical section
666 * during that time, so there is no need for it to execute a
667 * memory barrier.
668 */
669
670 if ((curr == snap) && ((curr & 0x1) == 0))
671 return 0;
672
673 /*
674 * If the CPU either entered or exited an outermost interrupt,
675 * SMI, NMI, or whatever handler, then we know that it executed
676 * a memory barrier when doing so. So we don't need another one.
677 */
678 if (curr != snap)
679 return 0;
680
681 /* We need the CPU to execute a memory barrier. */
682
683 return 1;
684}
685
686static void dyntick_save_progress_counter_sched(int cpu)
687{
688 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
689
690 rdssp->sched_dynticks_snap = rdssp->dynticks;
691}
692
693static int rcu_qsctr_inc_needed_dyntick(int cpu)
694{
695 long curr;
696 long snap;
697 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
698
699 curr = rdssp->dynticks;
700 snap = rdssp->sched_dynticks_snap;
701 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
702
703 /*
704 * If the CPU remained in dynticks mode for the entire time
705 * and didn't take any interrupts, NMIs, SMIs, or whatever,
706 * then it cannot be in the middle of an rcu_read_lock(), so
707 * the next rcu_read_lock() it executes must use the new value
708 * of the counter. Therefore, this CPU has been in a quiescent
709 * state the entire time, and we don't need to wait for it.
710 */
711
712 if ((curr == snap) && ((curr & 0x1) == 0))
713 return 0;
714
715 /*
716 * If the CPU passed through or entered a dynticks idle phase with
717 * no active irq handlers, then, as above, this CPU has already
718 * passed through a quiescent state.
719 */
720
721 if ((curr - snap) > 2 || (snap & 0x1) == 0)
722 return 0;
723
724 /* We need this CPU to go through a quiescent state. */
725
726 return 1;
727}
728
729#else /* !CONFIG_NO_HZ */
730
731# define dyntick_save_progress_counter(cpu) do { } while (0)
732# define rcu_try_flip_waitack_needed(cpu) (1)
733# define rcu_try_flip_waitmb_needed(cpu) (1)
734
735# define dyntick_save_progress_counter_sched(cpu) do { } while (0)
736# define rcu_qsctr_inc_needed_dyntick(cpu) (1)
737
738#endif /* CONFIG_NO_HZ */
739
740static void save_qsctr_sched(int cpu)
741{
742 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
743
744 rdssp->sched_qs_snap = rdssp->sched_qs;
745}
746
747static inline int rcu_qsctr_inc_needed(int cpu)
748{
749 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
750
751 /*
752 * If there has been a quiescent state, no more need to wait
753 * on this CPU.
754 */
755
756 if (rdssp->sched_qs != rdssp->sched_qs_snap) {
757 smp_mb(); /* force ordering with cpu entering schedule(). */
758 return 0;
759 }
760
761 /* We need this CPU to go through a quiescent state. */
762
763 return 1;
764}
765
766/*
767 * Get here when RCU is idle. Decide whether we need to
768 * move out of idle state, and return non-zero if so.
769 * "Straightforward" approach for the moment, might later
770 * use callback-list lengths, grace-period duration, or
771 * some such to determine when to exit idle state.
772 * Might also need a pre-idle test that does not acquire
773 * the lock, but let's get the simple case working first...
774 */
775
776static int
777rcu_try_flip_idle(void)
778{
779 int cpu;
780
781 RCU_TRACE_ME(rcupreempt_trace_try_flip_i1);
782 if (!rcu_pending(smp_processor_id())) {
783 RCU_TRACE_ME(rcupreempt_trace_try_flip_ie1);
784 return 0;
785 }
786
787 /*
788 * Do the flip.
789 */
790
791 RCU_TRACE_ME(rcupreempt_trace_try_flip_g1);
792 rcu_ctrlblk.completed++; /* stands in for rcu_try_flip_g2 */
793
794 /*
795 * Need a memory barrier so that other CPUs see the new
796 * counter value before they see the subsequent change of all
797 * the rcu_flip_flag instances to rcu_flipped.
798 */
799
800 smp_mb(); /* see above block comment. */
801
802 /* Now ask each CPU for acknowledgement of the flip. */
803
804 for_each_cpu(cpu, to_cpumask(rcu_cpu_online_map)) {
805 per_cpu(rcu_flip_flag, cpu) = rcu_flipped;
806 dyntick_save_progress_counter(cpu);
807 }
808
809 return 1;
810}
811
812/*
813 * Wait for CPUs to acknowledge the flip.
814 */
815
816static int
817rcu_try_flip_waitack(void)
818{
819 int cpu;
820
821 RCU_TRACE_ME(rcupreempt_trace_try_flip_a1);
822 for_each_cpu(cpu, to_cpumask(rcu_cpu_online_map))
823 if (rcu_try_flip_waitack_needed(cpu) &&
824 per_cpu(rcu_flip_flag, cpu) != rcu_flip_seen) {
825 RCU_TRACE_ME(rcupreempt_trace_try_flip_ae1);
826 return 0;
827 }
828
829 /*
830 * Make sure our checks above don't bleed into subsequent
831 * waiting for the sum of the counters to reach zero.
832 */
833
834 smp_mb(); /* see above block comment. */
835 RCU_TRACE_ME(rcupreempt_trace_try_flip_a2);
836 return 1;
837}
838
839/*
840 * Wait for collective ``last'' counter to reach zero,
841 * then tell all CPUs to do an end-of-grace-period memory barrier.
842 */
843
844static int
845rcu_try_flip_waitzero(void)
846{
847 int cpu;
848 int lastidx = !(rcu_ctrlblk.completed & 0x1);
849 int sum = 0;
850
851 /* Check to see if the sum of the "last" counters is zero. */
852
853 RCU_TRACE_ME(rcupreempt_trace_try_flip_z1);
854 for_each_possible_cpu(cpu)
855 sum += RCU_DATA_CPU(cpu)->rcu_flipctr[lastidx];
856 if (sum != 0) {
857 RCU_TRACE_ME(rcupreempt_trace_try_flip_ze1);
858 return 0;
859 }
860
861 /*
862 * This ensures that the other CPUs see the call for
863 * memory barriers -after- the sum to zero has been
864 * detected here
865 */
866 smp_mb(); /* ^^^^^^^^^^^^ */
867
868 /* Call for a memory barrier from each CPU. */
869 for_each_cpu(cpu, to_cpumask(rcu_cpu_online_map)) {
870 per_cpu(rcu_mb_flag, cpu) = rcu_mb_needed;
871 dyntick_save_progress_counter(cpu);
872 }
873
874 RCU_TRACE_ME(rcupreempt_trace_try_flip_z2);
875 return 1;
876}
877
878/*
879 * Wait for all CPUs to do their end-of-grace-period memory barrier.
880 * Return 0 once all CPUs have done so.
881 */
882
883static int
884rcu_try_flip_waitmb(void)
885{
886 int cpu;
887
888 RCU_TRACE_ME(rcupreempt_trace_try_flip_m1);
889 for_each_cpu(cpu, to_cpumask(rcu_cpu_online_map))
890 if (rcu_try_flip_waitmb_needed(cpu) &&
891 per_cpu(rcu_mb_flag, cpu) != rcu_mb_done) {
892 RCU_TRACE_ME(rcupreempt_trace_try_flip_me1);
893 return 0;
894 }
895
896 smp_mb(); /* Ensure that the above checks precede any following flip. */
897 RCU_TRACE_ME(rcupreempt_trace_try_flip_m2);
898 return 1;
899}
900
901/*
902 * Attempt a single flip of the counters. Remember, a single flip does
903 * -not- constitute a grace period. Instead, the interval between
904 * at least GP_STAGES consecutive flips is a grace period.
905 *
906 * If anyone is nuts enough to run this CONFIG_PREEMPT_RCU implementation
907 * on a large SMP, they might want to use a hierarchical organization of
908 * the per-CPU-counter pairs.
909 */
910static void rcu_try_flip(void)
911{
912 unsigned long flags;
913
914 RCU_TRACE_ME(rcupreempt_trace_try_flip_1);
915 if (unlikely(!spin_trylock_irqsave(&rcu_ctrlblk.fliplock, flags))) {
916 RCU_TRACE_ME(rcupreempt_trace_try_flip_e1);
917 return;
918 }
919
920 /*
921 * Take the next transition(s) through the RCU grace-period
922 * flip-counter state machine.
923 */
924
925 switch (rcu_ctrlblk.rcu_try_flip_state) {
926 case rcu_try_flip_idle_state:
927 if (rcu_try_flip_idle())
928 rcu_ctrlblk.rcu_try_flip_state =
929 rcu_try_flip_waitack_state;
930 break;
931 case rcu_try_flip_waitack_state:
932 if (rcu_try_flip_waitack())
933 rcu_ctrlblk.rcu_try_flip_state =
934 rcu_try_flip_waitzero_state;
935 break;
936 case rcu_try_flip_waitzero_state:
937 if (rcu_try_flip_waitzero())
938 rcu_ctrlblk.rcu_try_flip_state =
939 rcu_try_flip_waitmb_state;
940 break;
941 case rcu_try_flip_waitmb_state:
942 if (rcu_try_flip_waitmb())
943 rcu_ctrlblk.rcu_try_flip_state =
944 rcu_try_flip_idle_state;
945 }
946 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
947}
948
949/*
950 * Check to see if this CPU needs to do a memory barrier in order to
951 * ensure that any prior RCU read-side critical sections have committed
952 * their counter manipulations and critical-section memory references
953 * before declaring the grace period to be completed.
954 */
955static void rcu_check_mb(int cpu)
956{
957 if (per_cpu(rcu_mb_flag, cpu) == rcu_mb_needed) {
958 smp_mb(); /* Ensure RCU read-side accesses are visible. */
959 per_cpu(rcu_mb_flag, cpu) = rcu_mb_done;
960 }
961}
962
963void rcu_check_callbacks(int cpu, int user)
964{
965 unsigned long flags;
966 struct rcu_data *rdp;
967
968 if (!rcu_pending(cpu))
969 return; /* if nothing for RCU to do. */
970
971 /*
972 * If this CPU took its interrupt from user mode or from the
973 * idle loop, and this is not a nested interrupt, then
974 * this CPU has to have exited all prior preept-disable
975 * sections of code. So invoke rcu_sched_qs() to note this.
976 *
977 * The memory barrier is needed to handle the case where
978 * writes from a preempt-disable section of code get reordered
979 * into schedule() by this CPU's write buffer. So the memory
980 * barrier makes sure that the rcu_sched_qs() is seen by other
981 * CPUs to happen after any such write.
982 */
983
984 rdp = RCU_DATA_CPU(cpu);
985 if (user ||
986 (idle_cpu(cpu) && !in_softirq() &&
987 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
988 smp_mb(); /* Guard against aggressive schedule(). */
989 rcu_sched_qs(cpu);
990 }
991
992 rcu_check_mb(cpu);
993 if (rcu_ctrlblk.completed == rdp->completed)
994 rcu_try_flip();
995 spin_lock_irqsave(&rdp->lock, flags);
996 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
997 __rcu_advance_callbacks(rdp);
998 if (rdp->donelist == NULL) {
999 spin_unlock_irqrestore(&rdp->lock, flags);
1000 } else {
1001 spin_unlock_irqrestore(&rdp->lock, flags);
1002 raise_softirq(RCU_SOFTIRQ);
1003 }
1004}
1005
1006/*
1007 * Needed by dynticks, to make sure all RCU processing has finished
1008 * when we go idle:
1009 */
1010void rcu_advance_callbacks(int cpu, int user)
1011{
1012 unsigned long flags;
1013 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1014
1015 if (rcu_ctrlblk.completed == rdp->completed) {
1016 rcu_try_flip();
1017 if (rcu_ctrlblk.completed == rdp->completed)
1018 return;
1019 }
1020 spin_lock_irqsave(&rdp->lock, flags);
1021 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
1022 __rcu_advance_callbacks(rdp);
1023 spin_unlock_irqrestore(&rdp->lock, flags);
1024}
1025
1026#ifdef CONFIG_HOTPLUG_CPU
1027#define rcu_offline_cpu_enqueue(srclist, srctail, dstlist, dsttail) do { \
1028 *dsttail = srclist; \
1029 if (srclist != NULL) { \
1030 dsttail = srctail; \
1031 srclist = NULL; \
1032 srctail = &srclist;\
1033 } \
1034 } while (0)
1035
1036void rcu_offline_cpu(int cpu)
1037{
1038 int i;
1039 struct rcu_head *list = NULL;
1040 unsigned long flags;
1041 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1042 struct rcu_head *schedlist = NULL;
1043 struct rcu_head **schedtail = &schedlist;
1044 struct rcu_head **tail = &list;
1045
1046 /*
1047 * Remove all callbacks from the newly dead CPU, retaining order.
1048 * Otherwise rcu_barrier() will fail
1049 */
1050
1051 spin_lock_irqsave(&rdp->lock, flags);
1052 rcu_offline_cpu_enqueue(rdp->donelist, rdp->donetail, list, tail);
1053 for (i = GP_STAGES - 1; i >= 0; i--)
1054 rcu_offline_cpu_enqueue(rdp->waitlist[i], rdp->waittail[i],
1055 list, tail);
1056 rcu_offline_cpu_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
1057 rcu_offline_cpu_enqueue(rdp->waitschedlist, rdp->waitschedtail,
1058 schedlist, schedtail);
1059 rcu_offline_cpu_enqueue(rdp->nextschedlist, rdp->nextschedtail,
1060 schedlist, schedtail);
1061 rdp->rcu_sched_sleeping = 0;
1062 spin_unlock_irqrestore(&rdp->lock, flags);
1063 rdp->waitlistcount = 0;
1064
1065 /* Disengage the newly dead CPU from the grace-period computation. */
1066
1067 spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
1068 rcu_check_mb(cpu);
1069 if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
1070 smp_mb(); /* Subsequent counter accesses must see new value */
1071 per_cpu(rcu_flip_flag, cpu) = rcu_flip_seen;
1072 smp_mb(); /* Subsequent RCU read-side critical sections */
1073 /* seen -after- acknowledgement. */
1074 }
1075
1076 cpumask_clear_cpu(cpu, to_cpumask(rcu_cpu_online_map));
1077
1078 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
1079
1080 /*
1081 * Place the removed callbacks on the current CPU's queue.
1082 * Make them all start a new grace period: simple approach,
1083 * in theory could starve a given set of callbacks, but
1084 * you would need to be doing some serious CPU hotplugging
1085 * to make this happen. If this becomes a problem, adding
1086 * a synchronize_rcu() to the hotplug path would be a simple
1087 * fix.
1088 */
1089
1090 local_irq_save(flags); /* disable preempt till we know what lock. */
1091 rdp = RCU_DATA_ME();
1092 spin_lock(&rdp->lock);
1093 *rdp->nexttail = list;
1094 if (list)
1095 rdp->nexttail = tail;
1096 *rdp->nextschedtail = schedlist;
1097 if (schedlist)
1098 rdp->nextschedtail = schedtail;
1099 spin_unlock_irqrestore(&rdp->lock, flags);
1100}
1101
1102#else /* #ifdef CONFIG_HOTPLUG_CPU */
1103
1104void rcu_offline_cpu(int cpu)
1105{
1106}
1107
1108#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1109
1110void __cpuinit rcu_online_cpu(int cpu)
1111{
1112 unsigned long flags;
1113 struct rcu_data *rdp;
1114
1115 spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
1116 cpumask_set_cpu(cpu, to_cpumask(rcu_cpu_online_map));
1117 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
1118
1119 /*
1120 * The rcu_sched grace-period processing might have bypassed
1121 * this CPU, given that it was not in the rcu_cpu_online_map
1122 * when the grace-period scan started. This means that the
1123 * grace-period task might sleep. So make sure that if this
1124 * should happen, the first callback posted to this CPU will
1125 * wake up the grace-period task if need be.
1126 */
1127
1128 rdp = RCU_DATA_CPU(cpu);
1129 spin_lock_irqsave(&rdp->lock, flags);
1130 rdp->rcu_sched_sleeping = 1;
1131 spin_unlock_irqrestore(&rdp->lock, flags);
1132}
1133
1134static void rcu_process_callbacks(struct softirq_action *unused)
1135{
1136 unsigned long flags;
1137 struct rcu_head *next, *list;
1138 struct rcu_data *rdp;
1139
1140 local_irq_save(flags);
1141 rdp = RCU_DATA_ME();
1142 spin_lock(&rdp->lock);
1143 list = rdp->donelist;
1144 if (list == NULL) {
1145 spin_unlock_irqrestore(&rdp->lock, flags);
1146 return;
1147 }
1148 rdp->donelist = NULL;
1149 rdp->donetail = &rdp->donelist;
1150 RCU_TRACE_RDP(rcupreempt_trace_done_remove, rdp);
1151 spin_unlock_irqrestore(&rdp->lock, flags);
1152 while (list) {
1153 next = list->next;
1154 list->func(list);
1155 list = next;
1156 RCU_TRACE_ME(rcupreempt_trace_invoke);
1157 }
1158}
1159
1160void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1161{
1162 unsigned long flags;
1163 struct rcu_data *rdp;
1164
1165 head->func = func;
1166 head->next = NULL;
1167 local_irq_save(flags);
1168 rdp = RCU_DATA_ME();
1169 spin_lock(&rdp->lock);
1170 __rcu_advance_callbacks(rdp);
1171 *rdp->nexttail = head;
1172 rdp->nexttail = &head->next;
1173 RCU_TRACE_RDP(rcupreempt_trace_next_add, rdp);
1174 spin_unlock_irqrestore(&rdp->lock, flags);
1175}
1176EXPORT_SYMBOL_GPL(call_rcu);
1177
1178void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1179{
1180 unsigned long flags;
1181 struct rcu_data *rdp;
1182 int wake_gp = 0;
1183
1184 head->func = func;
1185 head->next = NULL;
1186 local_irq_save(flags);
1187 rdp = RCU_DATA_ME();
1188 spin_lock(&rdp->lock);
1189 *rdp->nextschedtail = head;
1190 rdp->nextschedtail = &head->next;
1191 if (rdp->rcu_sched_sleeping) {
1192
1193 /* Grace-period processing might be sleeping... */
1194
1195 rdp->rcu_sched_sleeping = 0;
1196 wake_gp = 1;
1197 }
1198 spin_unlock_irqrestore(&rdp->lock, flags);
1199 if (wake_gp) {
1200
1201 /* Wake up grace-period processing, unless someone beat us. */
1202
1203 spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
1204 if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
1205 wake_gp = 0;
1206 rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
1207 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1208 if (wake_gp)
1209 wake_up_interruptible(&rcu_ctrlblk.sched_wq);
1210 }
1211}
1212EXPORT_SYMBOL_GPL(call_rcu_sched);
1213
1214/*
1215 * Wait until all currently running preempt_disable() code segments
1216 * (including hardware-irq-disable segments) complete. Note that
1217 * in -rt this does -not- necessarily result in all currently executing
1218 * interrupt -handlers- having completed.
1219 */
1220void __synchronize_sched(void)
1221{
1222 struct rcu_synchronize rcu;
1223
1224 if (num_online_cpus() == 1)
1225 return; /* blocking is gp if only one CPU! */
1226
1227 init_completion(&rcu.completion);
1228 /* Will wake me after RCU finished. */
1229 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1230 /* Wait for it. */
1231 wait_for_completion(&rcu.completion);
1232}
1233EXPORT_SYMBOL_GPL(__synchronize_sched);
1234
1235/*
1236 * kthread function that manages call_rcu_sched grace periods.
1237 */
1238static int rcu_sched_grace_period(void *arg)
1239{
1240 int couldsleep; /* might sleep after current pass. */
1241 int couldsleepnext = 0; /* might sleep after next pass. */
1242 int cpu;
1243 unsigned long flags;
1244 struct rcu_data *rdp;
1245 int ret;
1246
1247 /*
1248 * Each pass through the following loop handles one
1249 * rcu_sched grace period cycle.
1250 */
1251 do {
1252 /* Save each CPU's current state. */
1253
1254 for_each_online_cpu(cpu) {
1255 dyntick_save_progress_counter_sched(cpu);
1256 save_qsctr_sched(cpu);
1257 }
1258
1259 /*
1260 * Sleep for about an RCU grace-period's worth to
1261 * allow better batching and to consume less CPU.
1262 */
1263 schedule_timeout_interruptible(RCU_SCHED_BATCH_TIME);
1264
1265 /*
1266 * If there was nothing to do last time, prepare to
1267 * sleep at the end of the current grace period cycle.
1268 */
1269 couldsleep = couldsleepnext;
1270 couldsleepnext = 1;
1271 if (couldsleep) {
1272 spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
1273 rcu_ctrlblk.sched_sleep = rcu_sched_sleep_prep;
1274 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1275 }
1276
1277 /*
1278 * Wait on each CPU in turn to have either visited
1279 * a quiescent state or been in dynticks-idle mode.
1280 */
1281 for_each_online_cpu(cpu) {
1282 while (rcu_qsctr_inc_needed(cpu) &&
1283 rcu_qsctr_inc_needed_dyntick(cpu)) {
1284 /* resched_cpu(cpu); @@@ */
1285 schedule_timeout_interruptible(1);
1286 }
1287 }
1288
1289 /* Advance callbacks for each CPU. */
1290
1291 for_each_online_cpu(cpu) {
1292
1293 rdp = RCU_DATA_CPU(cpu);
1294 spin_lock_irqsave(&rdp->lock, flags);
1295
1296 /*
1297 * We are running on this CPU irq-disabled, so no
1298 * CPU can go offline until we re-enable irqs.
1299 * The current CPU might have already gone
1300 * offline (between the for_each_offline_cpu and
1301 * the spin_lock_irqsave), but in that case all its
1302 * callback lists will be empty, so no harm done.
1303 *
1304 * Advance the callbacks! We share normal RCU's
1305 * donelist, since callbacks are invoked the
1306 * same way in either case.
1307 */
1308 if (rdp->waitschedlist != NULL) {
1309 *rdp->donetail = rdp->waitschedlist;
1310 rdp->donetail = rdp->waitschedtail;
1311
1312 /*
1313 * Next rcu_check_callbacks() will
1314 * do the required raise_softirq().
1315 */
1316 }
1317 if (rdp->nextschedlist != NULL) {
1318 rdp->waitschedlist = rdp->nextschedlist;
1319 rdp->waitschedtail = rdp->nextschedtail;
1320 couldsleep = 0;
1321 couldsleepnext = 0;
1322 } else {
1323 rdp->waitschedlist = NULL;
1324 rdp->waitschedtail = &rdp->waitschedlist;
1325 }
1326 rdp->nextschedlist = NULL;
1327 rdp->nextschedtail = &rdp->nextschedlist;
1328
1329 /* Mark sleep intention. */
1330
1331 rdp->rcu_sched_sleeping = couldsleep;
1332
1333 spin_unlock_irqrestore(&rdp->lock, flags);
1334 }
1335
1336 /* If we saw callbacks on the last scan, go deal with them. */
1337
1338 if (!couldsleep)
1339 continue;
1340
1341 /* Attempt to block... */
1342
1343 spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
1344 if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
1345
1346 /*
1347 * Someone posted a callback after we scanned.
1348 * Go take care of it.
1349 */
1350 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1351 couldsleepnext = 0;
1352 continue;
1353 }
1354
1355 /* Block until the next person posts a callback. */
1356
1357 rcu_ctrlblk.sched_sleep = rcu_sched_sleeping;
1358 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1359 ret = 0; /* unused */
1360 __wait_event_interruptible(rcu_ctrlblk.sched_wq,
1361 rcu_ctrlblk.sched_sleep != rcu_sched_sleeping,
1362 ret);
1363
1364 couldsleepnext = 0;
1365
1366 } while (!kthread_should_stop());
1367
1368 return (0);
1369}
1370
1371/*
1372 * Check to see if any future RCU-related work will need to be done
1373 * by the current CPU, even if none need be done immediately, returning
1374 * 1 if so. Assumes that notifiers would take care of handling any
1375 * outstanding requests from the RCU core.
1376 *
1377 * This function is part of the RCU implementation; it is -not-
1378 * an exported member of the RCU API.
1379 */
1380int rcu_needs_cpu(int cpu)
1381{
1382 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1383
1384 return (rdp->donelist != NULL ||
1385 !!rdp->waitlistcount ||
1386 rdp->nextlist != NULL ||
1387 rdp->nextschedlist != NULL ||
1388 rdp->waitschedlist != NULL);
1389}
1390
1391static int rcu_pending(int cpu)
1392{
1393 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1394
1395 /* The CPU has at least one callback queued somewhere. */
1396
1397 if (rdp->donelist != NULL ||
1398 !!rdp->waitlistcount ||
1399 rdp->nextlist != NULL ||
1400 rdp->nextschedlist != NULL ||
1401 rdp->waitschedlist != NULL)
1402 return 1;
1403
1404 /* The RCU core needs an acknowledgement from this CPU. */
1405
1406 if ((per_cpu(rcu_flip_flag, cpu) == rcu_flipped) ||
1407 (per_cpu(rcu_mb_flag, cpu) == rcu_mb_needed))
1408 return 1;
1409
1410 /* This CPU has fallen behind the global grace-period number. */
1411
1412 if (rdp->completed != rcu_ctrlblk.completed)
1413 return 1;
1414
1415 /* Nothing needed from this CPU. */
1416
1417 return 0;
1418}
1419
1420int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1421 unsigned long action, void *hcpu)
1422{
1423 long cpu = (long)hcpu;
1424
1425 switch (action) {
1426 case CPU_UP_PREPARE:
1427 case CPU_UP_PREPARE_FROZEN:
1428 rcu_online_cpu(cpu);
1429 break;
1430 case CPU_UP_CANCELED:
1431 case CPU_UP_CANCELED_FROZEN:
1432 case CPU_DEAD:
1433 case CPU_DEAD_FROZEN:
1434 rcu_offline_cpu(cpu);
1435 break;
1436 default:
1437 break;
1438 }
1439 return NOTIFY_OK;
1440}
1441
1442void __init __rcu_init(void)
1443{
1444 int cpu;
1445 int i;
1446 struct rcu_data *rdp;
1447
1448 printk(KERN_NOTICE "Preemptible RCU implementation.\n");
1449 for_each_possible_cpu(cpu) {
1450 rdp = RCU_DATA_CPU(cpu);
1451 spin_lock_init(&rdp->lock);
1452 rdp->completed = 0;
1453 rdp->waitlistcount = 0;
1454 rdp->nextlist = NULL;
1455 rdp->nexttail = &rdp->nextlist;
1456 for (i = 0; i < GP_STAGES; i++) {
1457 rdp->waitlist[i] = NULL;
1458 rdp->waittail[i] = &rdp->waitlist[i];
1459 }
1460 rdp->donelist = NULL;
1461 rdp->donetail = &rdp->donelist;
1462 rdp->rcu_flipctr[0] = 0;
1463 rdp->rcu_flipctr[1] = 0;
1464 rdp->nextschedlist = NULL;
1465 rdp->nextschedtail = &rdp->nextschedlist;
1466 rdp->waitschedlist = NULL;
1467 rdp->waitschedtail = &rdp->waitschedlist;
1468 rdp->rcu_sched_sleeping = 0;
1469 }
1470 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1471}
1472
1473/*
1474 * Late-boot-time RCU initialization that must wait until after scheduler
1475 * has been initialized.
1476 */
1477void __init rcu_init_sched(void)
1478{
1479 rcu_sched_grace_period_task = kthread_run(rcu_sched_grace_period,
1480 NULL,
1481 "rcu_sched_grace_period");
1482 WARN_ON(IS_ERR(rcu_sched_grace_period_task));
1483}
1484
1485#ifdef CONFIG_RCU_TRACE
1486long *rcupreempt_flipctr(int cpu)
1487{
1488 return &RCU_DATA_CPU(cpu)->rcu_flipctr[0];
1489}
1490EXPORT_SYMBOL_GPL(rcupreempt_flipctr);
1491
1492int rcupreempt_flip_flag(int cpu)
1493{
1494 return per_cpu(rcu_flip_flag, cpu);
1495}
1496EXPORT_SYMBOL_GPL(rcupreempt_flip_flag);
1497
1498int rcupreempt_mb_flag(int cpu)
1499{
1500 return per_cpu(rcu_mb_flag, cpu);
1501}
1502EXPORT_SYMBOL_GPL(rcupreempt_mb_flag);
1503
1504char *rcupreempt_try_flip_state_name(void)
1505{
1506 return rcu_try_flip_state_names[rcu_ctrlblk.rcu_try_flip_state];
1507}
1508EXPORT_SYMBOL_GPL(rcupreempt_try_flip_state_name);
1509
1510struct rcupreempt_trace *rcupreempt_trace_cpu(int cpu)
1511{
1512 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1513
1514 return &rdp->trace;
1515}
1516EXPORT_SYMBOL_GPL(rcupreempt_trace_cpu);
1517
1518#endif /* #ifdef RCU_TRACE */