Merge branch 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[linux-2.6/lfs.git] / kernel / rcupreempt.c
blobe1cdf196a51507ae644c8a7369b599c5233f8604
1 /*
2 * Read-Copy Update mechanism for mutual exclusion, realtime implementation
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 IBM Corporation, 2006
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.
26 * - Added handling of Dynamic Ticks
27 * Copyright 2007 - Paul E. Mckenney <paulmck@us.ibm.com>
28 * - Steven Rostedt <srostedt@redhat.com>
30 * Papers: http://www.rdrop.com/users/paulmck/RCU
32 * Design Document: http://lwn.net/Articles/253651/
34 * For detailed explanation of Read-Copy Update mechanism see -
35 * Documentation/RCU/ *.txt
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/completion.h>
50 #include <linux/moduleparam.h>
51 #include <linux/percpu.h>
52 #include <linux/notifier.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/random.h>
56 #include <linux/delay.h>
57 #include <linux/byteorder/swabb.h>
58 #include <linux/cpumask.h>
59 #include <linux/rcupreempt_trace.h>
62 * Macro that prevents the compiler from reordering accesses, but does
63 * absolutely -nothing- to prevent CPUs from reordering. This is used
64 * only to mediate communication between mainline code and hardware
65 * interrupt and NMI handlers.
67 #define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
70 * PREEMPT_RCU data structures.
74 * GP_STAGES specifies the number of times the state machine has
75 * to go through the all the rcu_try_flip_states (see below)
76 * in a single Grace Period.
78 * GP in GP_STAGES stands for Grace Period ;)
80 #define GP_STAGES 2
81 struct rcu_data {
82 spinlock_t lock; /* Protect rcu_data fields. */
83 long completed; /* Number of last completed batch. */
84 int waitlistcount;
85 struct tasklet_struct rcu_tasklet;
86 struct rcu_head *nextlist;
87 struct rcu_head **nexttail;
88 struct rcu_head *waitlist[GP_STAGES];
89 struct rcu_head **waittail[GP_STAGES];
90 struct rcu_head *donelist;
91 struct rcu_head **donetail;
92 long rcu_flipctr[2];
93 #ifdef CONFIG_RCU_TRACE
94 struct rcupreempt_trace trace;
95 #endif /* #ifdef CONFIG_RCU_TRACE */
99 * States for rcu_try_flip() and friends.
102 enum rcu_try_flip_states {
105 * Stay here if nothing is happening. Flip the counter if somthing
106 * starts happening. Denoted by "I"
108 rcu_try_flip_idle_state,
111 * Wait here for all CPUs to notice that the counter has flipped. This
112 * prevents the old set of counters from ever being incremented once
113 * we leave this state, which in turn is necessary because we cannot
114 * test any individual counter for zero -- we can only check the sum.
115 * Denoted by "A".
117 rcu_try_flip_waitack_state,
120 * Wait here for the sum of the old per-CPU counters to reach zero.
121 * Denoted by "Z".
123 rcu_try_flip_waitzero_state,
126 * Wait here for each of the other CPUs to execute a memory barrier.
127 * This is necessary to ensure that these other CPUs really have
128 * completed executing their RCU read-side critical sections, despite
129 * their CPUs wildly reordering memory. Denoted by "M".
131 rcu_try_flip_waitmb_state,
134 struct rcu_ctrlblk {
135 spinlock_t fliplock; /* Protect state-machine transitions. */
136 long completed; /* Number of last completed batch. */
137 enum rcu_try_flip_states rcu_try_flip_state; /* The current state of
138 the rcu state machine */
141 static DEFINE_PER_CPU(struct rcu_data, rcu_data);
142 static struct rcu_ctrlblk rcu_ctrlblk = {
143 .fliplock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.fliplock),
144 .completed = 0,
145 .rcu_try_flip_state = rcu_try_flip_idle_state,
149 #ifdef CONFIG_RCU_TRACE
150 static char *rcu_try_flip_state_names[] =
151 { "idle", "waitack", "waitzero", "waitmb" };
152 #endif /* #ifdef CONFIG_RCU_TRACE */
154 static cpumask_t rcu_cpu_online_map __read_mostly = CPU_MASK_NONE;
157 * Enum and per-CPU flag to determine when each CPU has seen
158 * the most recent counter flip.
161 enum rcu_flip_flag_values {
162 rcu_flip_seen, /* Steady/initial state, last flip seen. */
163 /* Only GP detector can update. */
164 rcu_flipped /* Flip just completed, need confirmation. */
165 /* Only corresponding CPU can update. */
167 static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_flip_flag_values, rcu_flip_flag)
168 = rcu_flip_seen;
171 * Enum and per-CPU flag to determine when each CPU has executed the
172 * needed memory barrier to fence in memory references from its last RCU
173 * read-side critical section in the just-completed grace period.
176 enum rcu_mb_flag_values {
177 rcu_mb_done, /* Steady/initial state, no mb()s required. */
178 /* Only GP detector can update. */
179 rcu_mb_needed /* Flip just completed, need an mb(). */
180 /* Only corresponding CPU can update. */
182 static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_mb_flag_values, rcu_mb_flag)
183 = rcu_mb_done;
186 * RCU_DATA_ME: find the current CPU's rcu_data structure.
187 * RCU_DATA_CPU: find the specified CPU's rcu_data structure.
189 #define RCU_DATA_ME() (&__get_cpu_var(rcu_data))
190 #define RCU_DATA_CPU(cpu) (&per_cpu(rcu_data, cpu))
193 * Helper macro for tracing when the appropriate rcu_data is not
194 * cached in a local variable, but where the CPU number is so cached.
196 #define RCU_TRACE_CPU(f, cpu) RCU_TRACE(f, &(RCU_DATA_CPU(cpu)->trace));
199 * Helper macro for tracing when the appropriate rcu_data is not
200 * cached in a local variable.
202 #define RCU_TRACE_ME(f) RCU_TRACE(f, &(RCU_DATA_ME()->trace));
205 * Helper macro for tracing when the appropriate rcu_data is pointed
206 * to by a local variable.
208 #define RCU_TRACE_RDP(f, rdp) RCU_TRACE(f, &((rdp)->trace));
211 * Return the number of RCU batches processed thus far. Useful
212 * for debug and statistics.
214 long rcu_batches_completed(void)
216 return rcu_ctrlblk.completed;
218 EXPORT_SYMBOL_GPL(rcu_batches_completed);
220 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
222 void __rcu_read_lock(void)
224 int idx;
225 struct task_struct *t = current;
226 int nesting;
228 nesting = ACCESS_ONCE(t->rcu_read_lock_nesting);
229 if (nesting != 0) {
231 /* An earlier rcu_read_lock() covers us, just count it. */
233 t->rcu_read_lock_nesting = nesting + 1;
235 } else {
236 unsigned long flags;
239 * We disable interrupts for the following reasons:
240 * - If we get scheduling clock interrupt here, and we
241 * end up acking the counter flip, it's like a promise
242 * that we will never increment the old counter again.
243 * Thus we will break that promise if that
244 * scheduling clock interrupt happens between the time
245 * we pick the .completed field and the time that we
246 * increment our counter.
248 * - We don't want to be preempted out here.
250 * NMIs can still occur, of course, and might themselves
251 * contain rcu_read_lock().
254 local_irq_save(flags);
257 * Outermost nesting of rcu_read_lock(), so increment
258 * the current counter for the current CPU. Use volatile
259 * casts to prevent the compiler from reordering.
262 idx = ACCESS_ONCE(rcu_ctrlblk.completed) & 0x1;
263 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr[idx])++;
266 * Now that the per-CPU counter has been incremented, we
267 * are protected from races with rcu_read_lock() invoked
268 * from NMI handlers on this CPU. We can therefore safely
269 * increment the nesting counter, relieving further NMIs
270 * of the need to increment the per-CPU counter.
273 ACCESS_ONCE(t->rcu_read_lock_nesting) = nesting + 1;
276 * Now that we have preventing any NMIs from storing
277 * to the ->rcu_flipctr_idx, we can safely use it to
278 * remember which counter to decrement in the matching
279 * rcu_read_unlock().
282 ACCESS_ONCE(t->rcu_flipctr_idx) = idx;
283 local_irq_restore(flags);
286 EXPORT_SYMBOL_GPL(__rcu_read_lock);
288 void __rcu_read_unlock(void)
290 int idx;
291 struct task_struct *t = current;
292 int nesting;
294 nesting = ACCESS_ONCE(t->rcu_read_lock_nesting);
295 if (nesting > 1) {
298 * We are still protected by the enclosing rcu_read_lock(),
299 * so simply decrement the counter.
302 t->rcu_read_lock_nesting = nesting - 1;
304 } else {
305 unsigned long flags;
308 * Disable local interrupts to prevent the grace-period
309 * detection state machine from seeing us half-done.
310 * NMIs can still occur, of course, and might themselves
311 * contain rcu_read_lock() and rcu_read_unlock().
314 local_irq_save(flags);
317 * Outermost nesting of rcu_read_unlock(), so we must
318 * decrement the current counter for the current CPU.
319 * This must be done carefully, because NMIs can
320 * occur at any point in this code, and any rcu_read_lock()
321 * and rcu_read_unlock() pairs in the NMI handlers
322 * must interact non-destructively with this code.
323 * Lots of volatile casts, and -very- careful ordering.
325 * Changes to this code, including this one, must be
326 * inspected, validated, and tested extremely carefully!!!
330 * First, pick up the index.
333 idx = ACCESS_ONCE(t->rcu_flipctr_idx);
336 * Now that we have fetched the counter index, it is
337 * safe to decrement the per-task RCU nesting counter.
338 * After this, any interrupts or NMIs will increment and
339 * decrement the per-CPU counters.
341 ACCESS_ONCE(t->rcu_read_lock_nesting) = nesting - 1;
344 * It is now safe to decrement this task's nesting count.
345 * NMIs that occur after this statement will route their
346 * rcu_read_lock() calls through this "else" clause, and
347 * will thus start incrementing the per-CPU counter on
348 * their own. They will also clobber ->rcu_flipctr_idx,
349 * but that is OK, since we have already fetched it.
352 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr[idx])--;
353 local_irq_restore(flags);
356 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
359 * If a global counter flip has occurred since the last time that we
360 * advanced callbacks, advance them. Hardware interrupts must be
361 * disabled when calling this function.
363 static void __rcu_advance_callbacks(struct rcu_data *rdp)
365 int cpu;
366 int i;
367 int wlc = 0;
369 if (rdp->completed != rcu_ctrlblk.completed) {
370 if (rdp->waitlist[GP_STAGES - 1] != NULL) {
371 *rdp->donetail = rdp->waitlist[GP_STAGES - 1];
372 rdp->donetail = rdp->waittail[GP_STAGES - 1];
373 RCU_TRACE_RDP(rcupreempt_trace_move2done, rdp);
375 for (i = GP_STAGES - 2; i >= 0; i--) {
376 if (rdp->waitlist[i] != NULL) {
377 rdp->waitlist[i + 1] = rdp->waitlist[i];
378 rdp->waittail[i + 1] = rdp->waittail[i];
379 wlc++;
380 } else {
381 rdp->waitlist[i + 1] = NULL;
382 rdp->waittail[i + 1] =
383 &rdp->waitlist[i + 1];
386 if (rdp->nextlist != NULL) {
387 rdp->waitlist[0] = rdp->nextlist;
388 rdp->waittail[0] = rdp->nexttail;
389 wlc++;
390 rdp->nextlist = NULL;
391 rdp->nexttail = &rdp->nextlist;
392 RCU_TRACE_RDP(rcupreempt_trace_move2wait, rdp);
393 } else {
394 rdp->waitlist[0] = NULL;
395 rdp->waittail[0] = &rdp->waitlist[0];
397 rdp->waitlistcount = wlc;
398 rdp->completed = rcu_ctrlblk.completed;
402 * Check to see if this CPU needs to report that it has seen
403 * the most recent counter flip, thereby declaring that all
404 * subsequent rcu_read_lock() invocations will respect this flip.
407 cpu = raw_smp_processor_id();
408 if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
409 smp_mb(); /* Subsequent counter accesses must see new value */
410 per_cpu(rcu_flip_flag, cpu) = rcu_flip_seen;
411 smp_mb(); /* Subsequent RCU read-side critical sections */
412 /* seen -after- acknowledgement. */
416 #ifdef CONFIG_NO_HZ
418 DEFINE_PER_CPU(long, dynticks_progress_counter) = 1;
419 static DEFINE_PER_CPU(long, rcu_dyntick_snapshot);
420 static DEFINE_PER_CPU(int, rcu_update_flag);
423 * rcu_irq_enter - Called from Hard irq handlers and NMI/SMI.
425 * If the CPU was idle with dynamic ticks active, this updates the
426 * dynticks_progress_counter to let the RCU handling know that the
427 * CPU is active.
429 void rcu_irq_enter(void)
431 int cpu = smp_processor_id();
433 if (per_cpu(rcu_update_flag, cpu))
434 per_cpu(rcu_update_flag, cpu)++;
437 * Only update if we are coming from a stopped ticks mode
438 * (dynticks_progress_counter is even).
440 if (!in_interrupt() &&
441 (per_cpu(dynticks_progress_counter, cpu) & 0x1) == 0) {
443 * The following might seem like we could have a race
444 * with NMI/SMIs. But this really isn't a problem.
445 * Here we do a read/modify/write, and the race happens
446 * when an NMI/SMI comes in after the read and before
447 * the write. But NMI/SMIs will increment this counter
448 * twice before returning, so the zero bit will not
449 * be corrupted by the NMI/SMI which is the most important
450 * part.
452 * The only thing is that we would bring back the counter
453 * to a postion that it was in during the NMI/SMI.
454 * But the zero bit would be set, so the rest of the
455 * counter would again be ignored.
457 * On return from the IRQ, the counter may have the zero
458 * bit be 0 and the counter the same as the return from
459 * the NMI/SMI. If the state machine was so unlucky to
460 * see that, it still doesn't matter, since all
461 * RCU read-side critical sections on this CPU would
462 * have already completed.
464 per_cpu(dynticks_progress_counter, cpu)++;
466 * The following memory barrier ensures that any
467 * rcu_read_lock() primitives in the irq handler
468 * are seen by other CPUs to follow the above
469 * increment to dynticks_progress_counter. This is
470 * required in order for other CPUs to correctly
471 * determine when it is safe to advance the RCU
472 * grace-period state machine.
474 smp_mb(); /* see above block comment. */
476 * Since we can't determine the dynamic tick mode from
477 * the dynticks_progress_counter after this routine,
478 * we use a second flag to acknowledge that we came
479 * from an idle state with ticks stopped.
481 per_cpu(rcu_update_flag, cpu)++;
483 * If we take an NMI/SMI now, they will also increment
484 * the rcu_update_flag, and will not update the
485 * dynticks_progress_counter on exit. That is for
486 * this IRQ to do.
492 * rcu_irq_exit - Called from exiting Hard irq context.
494 * If the CPU was idle with dynamic ticks active, update the
495 * dynticks_progress_counter to put let the RCU handling be
496 * aware that the CPU is going back to idle with no ticks.
498 void rcu_irq_exit(void)
500 int cpu = smp_processor_id();
503 * rcu_update_flag is set if we interrupted the CPU
504 * when it was idle with ticks stopped.
505 * Once this occurs, we keep track of interrupt nesting
506 * because a NMI/SMI could also come in, and we still
507 * only want the IRQ that started the increment of the
508 * dynticks_progress_counter to be the one that modifies
509 * it on exit.
511 if (per_cpu(rcu_update_flag, cpu)) {
512 if (--per_cpu(rcu_update_flag, cpu))
513 return;
515 /* This must match the interrupt nesting */
516 WARN_ON(in_interrupt());
519 * If an NMI/SMI happens now we are still
520 * protected by the dynticks_progress_counter being odd.
524 * The following memory barrier ensures that any
525 * rcu_read_unlock() primitives in the irq handler
526 * are seen by other CPUs to preceed the following
527 * increment to dynticks_progress_counter. This
528 * is required in order for other CPUs to determine
529 * when it is safe to advance the RCU grace-period
530 * state machine.
532 smp_mb(); /* see above block comment. */
533 per_cpu(dynticks_progress_counter, cpu)++;
534 WARN_ON(per_cpu(dynticks_progress_counter, cpu) & 0x1);
538 static void dyntick_save_progress_counter(int cpu)
540 per_cpu(rcu_dyntick_snapshot, cpu) =
541 per_cpu(dynticks_progress_counter, cpu);
544 static inline int
545 rcu_try_flip_waitack_needed(int cpu)
547 long curr;
548 long snap;
550 curr = per_cpu(dynticks_progress_counter, cpu);
551 snap = per_cpu(rcu_dyntick_snapshot, cpu);
552 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
555 * If the CPU remained in dynticks mode for the entire time
556 * and didn't take any interrupts, NMIs, SMIs, or whatever,
557 * then it cannot be in the middle of an rcu_read_lock(), so
558 * the next rcu_read_lock() it executes must use the new value
559 * of the counter. So we can safely pretend that this CPU
560 * already acknowledged the counter.
563 if ((curr == snap) && ((curr & 0x1) == 0))
564 return 0;
567 * If the CPU passed through or entered a dynticks idle phase with
568 * no active irq handlers, then, as above, we can safely pretend
569 * that this CPU already acknowledged the counter.
572 if ((curr - snap) > 2 || (snap & 0x1) == 0)
573 return 0;
575 /* We need this CPU to explicitly acknowledge the counter flip. */
577 return 1;
580 static inline int
581 rcu_try_flip_waitmb_needed(int cpu)
583 long curr;
584 long snap;
586 curr = per_cpu(dynticks_progress_counter, cpu);
587 snap = per_cpu(rcu_dyntick_snapshot, cpu);
588 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
591 * If the CPU remained in dynticks mode for the entire time
592 * and didn't take any interrupts, NMIs, SMIs, or whatever,
593 * then it cannot have executed an RCU read-side critical section
594 * during that time, so there is no need for it to execute a
595 * memory barrier.
598 if ((curr == snap) && ((curr & 0x1) == 0))
599 return 0;
602 * If the CPU either entered or exited an outermost interrupt,
603 * SMI, NMI, or whatever handler, then we know that it executed
604 * a memory barrier when doing so. So we don't need another one.
606 if (curr != snap)
607 return 0;
609 /* We need the CPU to execute a memory barrier. */
611 return 1;
614 #else /* !CONFIG_NO_HZ */
616 # define dyntick_save_progress_counter(cpu) do { } while (0)
617 # define rcu_try_flip_waitack_needed(cpu) (1)
618 # define rcu_try_flip_waitmb_needed(cpu) (1)
620 #endif /* CONFIG_NO_HZ */
623 * Get here when RCU is idle. Decide whether we need to
624 * move out of idle state, and return non-zero if so.
625 * "Straightforward" approach for the moment, might later
626 * use callback-list lengths, grace-period duration, or
627 * some such to determine when to exit idle state.
628 * Might also need a pre-idle test that does not acquire
629 * the lock, but let's get the simple case working first...
632 static int
633 rcu_try_flip_idle(void)
635 int cpu;
637 RCU_TRACE_ME(rcupreempt_trace_try_flip_i1);
638 if (!rcu_pending(smp_processor_id())) {
639 RCU_TRACE_ME(rcupreempt_trace_try_flip_ie1);
640 return 0;
644 * Do the flip.
647 RCU_TRACE_ME(rcupreempt_trace_try_flip_g1);
648 rcu_ctrlblk.completed++; /* stands in for rcu_try_flip_g2 */
651 * Need a memory barrier so that other CPUs see the new
652 * counter value before they see the subsequent change of all
653 * the rcu_flip_flag instances to rcu_flipped.
656 smp_mb(); /* see above block comment. */
658 /* Now ask each CPU for acknowledgement of the flip. */
660 for_each_cpu_mask(cpu, rcu_cpu_online_map) {
661 per_cpu(rcu_flip_flag, cpu) = rcu_flipped;
662 dyntick_save_progress_counter(cpu);
665 return 1;
669 * Wait for CPUs to acknowledge the flip.
672 static int
673 rcu_try_flip_waitack(void)
675 int cpu;
677 RCU_TRACE_ME(rcupreempt_trace_try_flip_a1);
678 for_each_cpu_mask(cpu, rcu_cpu_online_map)
679 if (rcu_try_flip_waitack_needed(cpu) &&
680 per_cpu(rcu_flip_flag, cpu) != rcu_flip_seen) {
681 RCU_TRACE_ME(rcupreempt_trace_try_flip_ae1);
682 return 0;
686 * Make sure our checks above don't bleed into subsequent
687 * waiting for the sum of the counters to reach zero.
690 smp_mb(); /* see above block comment. */
691 RCU_TRACE_ME(rcupreempt_trace_try_flip_a2);
692 return 1;
696 * Wait for collective ``last'' counter to reach zero,
697 * then tell all CPUs to do an end-of-grace-period memory barrier.
700 static int
701 rcu_try_flip_waitzero(void)
703 int cpu;
704 int lastidx = !(rcu_ctrlblk.completed & 0x1);
705 int sum = 0;
707 /* Check to see if the sum of the "last" counters is zero. */
709 RCU_TRACE_ME(rcupreempt_trace_try_flip_z1);
710 for_each_cpu_mask(cpu, rcu_cpu_online_map)
711 sum += RCU_DATA_CPU(cpu)->rcu_flipctr[lastidx];
712 if (sum != 0) {
713 RCU_TRACE_ME(rcupreempt_trace_try_flip_ze1);
714 return 0;
718 * This ensures that the other CPUs see the call for
719 * memory barriers -after- the sum to zero has been
720 * detected here
722 smp_mb(); /* ^^^^^^^^^^^^ */
724 /* Call for a memory barrier from each CPU. */
725 for_each_cpu_mask(cpu, rcu_cpu_online_map) {
726 per_cpu(rcu_mb_flag, cpu) = rcu_mb_needed;
727 dyntick_save_progress_counter(cpu);
730 RCU_TRACE_ME(rcupreempt_trace_try_flip_z2);
731 return 1;
735 * Wait for all CPUs to do their end-of-grace-period memory barrier.
736 * Return 0 once all CPUs have done so.
739 static int
740 rcu_try_flip_waitmb(void)
742 int cpu;
744 RCU_TRACE_ME(rcupreempt_trace_try_flip_m1);
745 for_each_cpu_mask(cpu, rcu_cpu_online_map)
746 if (rcu_try_flip_waitmb_needed(cpu) &&
747 per_cpu(rcu_mb_flag, cpu) != rcu_mb_done) {
748 RCU_TRACE_ME(rcupreempt_trace_try_flip_me1);
749 return 0;
752 smp_mb(); /* Ensure that the above checks precede any following flip. */
753 RCU_TRACE_ME(rcupreempt_trace_try_flip_m2);
754 return 1;
758 * Attempt a single flip of the counters. Remember, a single flip does
759 * -not- constitute a grace period. Instead, the interval between
760 * at least GP_STAGES consecutive flips is a grace period.
762 * If anyone is nuts enough to run this CONFIG_PREEMPT_RCU implementation
763 * on a large SMP, they might want to use a hierarchical organization of
764 * the per-CPU-counter pairs.
766 static void rcu_try_flip(void)
768 unsigned long flags;
770 RCU_TRACE_ME(rcupreempt_trace_try_flip_1);
771 if (unlikely(!spin_trylock_irqsave(&rcu_ctrlblk.fliplock, flags))) {
772 RCU_TRACE_ME(rcupreempt_trace_try_flip_e1);
773 return;
777 * Take the next transition(s) through the RCU grace-period
778 * flip-counter state machine.
781 switch (rcu_ctrlblk.rcu_try_flip_state) {
782 case rcu_try_flip_idle_state:
783 if (rcu_try_flip_idle())
784 rcu_ctrlblk.rcu_try_flip_state =
785 rcu_try_flip_waitack_state;
786 break;
787 case rcu_try_flip_waitack_state:
788 if (rcu_try_flip_waitack())
789 rcu_ctrlblk.rcu_try_flip_state =
790 rcu_try_flip_waitzero_state;
791 break;
792 case rcu_try_flip_waitzero_state:
793 if (rcu_try_flip_waitzero())
794 rcu_ctrlblk.rcu_try_flip_state =
795 rcu_try_flip_waitmb_state;
796 break;
797 case rcu_try_flip_waitmb_state:
798 if (rcu_try_flip_waitmb())
799 rcu_ctrlblk.rcu_try_flip_state =
800 rcu_try_flip_idle_state;
802 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
806 * Check to see if this CPU needs to do a memory barrier in order to
807 * ensure that any prior RCU read-side critical sections have committed
808 * their counter manipulations and critical-section memory references
809 * before declaring the grace period to be completed.
811 static void rcu_check_mb(int cpu)
813 if (per_cpu(rcu_mb_flag, cpu) == rcu_mb_needed) {
814 smp_mb(); /* Ensure RCU read-side accesses are visible. */
815 per_cpu(rcu_mb_flag, cpu) = rcu_mb_done;
819 void rcu_check_callbacks(int cpu, int user)
821 unsigned long flags;
822 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
824 rcu_check_mb(cpu);
825 if (rcu_ctrlblk.completed == rdp->completed)
826 rcu_try_flip();
827 spin_lock_irqsave(&rdp->lock, flags);
828 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
829 __rcu_advance_callbacks(rdp);
830 if (rdp->donelist == NULL) {
831 spin_unlock_irqrestore(&rdp->lock, flags);
832 } else {
833 spin_unlock_irqrestore(&rdp->lock, flags);
834 raise_softirq(RCU_SOFTIRQ);
839 * Needed by dynticks, to make sure all RCU processing has finished
840 * when we go idle:
842 void rcu_advance_callbacks(int cpu, int user)
844 unsigned long flags;
845 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
847 if (rcu_ctrlblk.completed == rdp->completed) {
848 rcu_try_flip();
849 if (rcu_ctrlblk.completed == rdp->completed)
850 return;
852 spin_lock_irqsave(&rdp->lock, flags);
853 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
854 __rcu_advance_callbacks(rdp);
855 spin_unlock_irqrestore(&rdp->lock, flags);
858 #ifdef CONFIG_HOTPLUG_CPU
859 #define rcu_offline_cpu_enqueue(srclist, srctail, dstlist, dsttail) do { \
860 *dsttail = srclist; \
861 if (srclist != NULL) { \
862 dsttail = srctail; \
863 srclist = NULL; \
864 srctail = &srclist;\
866 } while (0)
868 void rcu_offline_cpu(int cpu)
870 int i;
871 struct rcu_head *list = NULL;
872 unsigned long flags;
873 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
874 struct rcu_head **tail = &list;
877 * Remove all callbacks from the newly dead CPU, retaining order.
878 * Otherwise rcu_barrier() will fail
881 spin_lock_irqsave(&rdp->lock, flags);
882 rcu_offline_cpu_enqueue(rdp->donelist, rdp->donetail, list, tail);
883 for (i = GP_STAGES - 1; i >= 0; i--)
884 rcu_offline_cpu_enqueue(rdp->waitlist[i], rdp->waittail[i],
885 list, tail);
886 rcu_offline_cpu_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
887 spin_unlock_irqrestore(&rdp->lock, flags);
888 rdp->waitlistcount = 0;
890 /* Disengage the newly dead CPU from the grace-period computation. */
892 spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
893 rcu_check_mb(cpu);
894 if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
895 smp_mb(); /* Subsequent counter accesses must see new value */
896 per_cpu(rcu_flip_flag, cpu) = rcu_flip_seen;
897 smp_mb(); /* Subsequent RCU read-side critical sections */
898 /* seen -after- acknowledgement. */
901 RCU_DATA_ME()->rcu_flipctr[0] += RCU_DATA_CPU(cpu)->rcu_flipctr[0];
902 RCU_DATA_ME()->rcu_flipctr[1] += RCU_DATA_CPU(cpu)->rcu_flipctr[1];
904 RCU_DATA_CPU(cpu)->rcu_flipctr[0] = 0;
905 RCU_DATA_CPU(cpu)->rcu_flipctr[1] = 0;
907 cpu_clear(cpu, rcu_cpu_online_map);
909 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
912 * Place the removed callbacks on the current CPU's queue.
913 * Make them all start a new grace period: simple approach,
914 * in theory could starve a given set of callbacks, but
915 * you would need to be doing some serious CPU hotplugging
916 * to make this happen. If this becomes a problem, adding
917 * a synchronize_rcu() to the hotplug path would be a simple
918 * fix.
921 local_irq_save(flags);
922 rdp = RCU_DATA_ME();
923 spin_lock(&rdp->lock);
924 *rdp->nexttail = list;
925 if (list)
926 rdp->nexttail = tail;
927 spin_unlock_irqrestore(&rdp->lock, flags);
930 void __devinit rcu_online_cpu(int cpu)
932 unsigned long flags;
934 spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
935 cpu_set(cpu, rcu_cpu_online_map);
936 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
939 #else /* #ifdef CONFIG_HOTPLUG_CPU */
941 void rcu_offline_cpu(int cpu)
945 void __devinit rcu_online_cpu(int cpu)
949 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
951 static void rcu_process_callbacks(struct softirq_action *unused)
953 unsigned long flags;
954 struct rcu_head *next, *list;
955 struct rcu_data *rdp;
957 local_irq_save(flags);
958 rdp = RCU_DATA_ME();
959 spin_lock(&rdp->lock);
960 list = rdp->donelist;
961 if (list == NULL) {
962 spin_unlock_irqrestore(&rdp->lock, flags);
963 return;
965 rdp->donelist = NULL;
966 rdp->donetail = &rdp->donelist;
967 RCU_TRACE_RDP(rcupreempt_trace_done_remove, rdp);
968 spin_unlock_irqrestore(&rdp->lock, flags);
969 while (list) {
970 next = list->next;
971 list->func(list);
972 list = next;
973 RCU_TRACE_ME(rcupreempt_trace_invoke);
977 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
979 unsigned long flags;
980 struct rcu_data *rdp;
982 head->func = func;
983 head->next = NULL;
984 local_irq_save(flags);
985 rdp = RCU_DATA_ME();
986 spin_lock(&rdp->lock);
987 __rcu_advance_callbacks(rdp);
988 *rdp->nexttail = head;
989 rdp->nexttail = &head->next;
990 RCU_TRACE_RDP(rcupreempt_trace_next_add, rdp);
991 spin_unlock(&rdp->lock);
992 local_irq_restore(flags);
994 EXPORT_SYMBOL_GPL(call_rcu);
997 * Wait until all currently running preempt_disable() code segments
998 * (including hardware-irq-disable segments) complete. Note that
999 * in -rt this does -not- necessarily result in all currently executing
1000 * interrupt -handlers- having completed.
1002 void __synchronize_sched(void)
1004 cpumask_t oldmask;
1005 int cpu;
1007 if (sched_getaffinity(0, &oldmask) < 0)
1008 oldmask = cpu_possible_map;
1009 for_each_online_cpu(cpu) {
1010 sched_setaffinity(0, &cpumask_of_cpu(cpu));
1011 schedule();
1013 sched_setaffinity(0, &oldmask);
1015 EXPORT_SYMBOL_GPL(__synchronize_sched);
1018 * Check to see if any future RCU-related work will need to be done
1019 * by the current CPU, even if none need be done immediately, returning
1020 * 1 if so. Assumes that notifiers would take care of handling any
1021 * outstanding requests from the RCU core.
1023 * This function is part of the RCU implementation; it is -not-
1024 * an exported member of the RCU API.
1026 int rcu_needs_cpu(int cpu)
1028 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1030 return (rdp->donelist != NULL ||
1031 !!rdp->waitlistcount ||
1032 rdp->nextlist != NULL);
1035 int rcu_pending(int cpu)
1037 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1039 /* The CPU has at least one callback queued somewhere. */
1041 if (rdp->donelist != NULL ||
1042 !!rdp->waitlistcount ||
1043 rdp->nextlist != NULL)
1044 return 1;
1046 /* The RCU core needs an acknowledgement from this CPU. */
1048 if ((per_cpu(rcu_flip_flag, cpu) == rcu_flipped) ||
1049 (per_cpu(rcu_mb_flag, cpu) == rcu_mb_needed))
1050 return 1;
1052 /* This CPU has fallen behind the global grace-period number. */
1054 if (rdp->completed != rcu_ctrlblk.completed)
1055 return 1;
1057 /* Nothing needed from this CPU. */
1059 return 0;
1062 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1063 unsigned long action, void *hcpu)
1065 long cpu = (long)hcpu;
1067 switch (action) {
1068 case CPU_UP_PREPARE:
1069 case CPU_UP_PREPARE_FROZEN:
1070 rcu_online_cpu(cpu);
1071 break;
1072 case CPU_UP_CANCELED:
1073 case CPU_UP_CANCELED_FROZEN:
1074 case CPU_DEAD:
1075 case CPU_DEAD_FROZEN:
1076 rcu_offline_cpu(cpu);
1077 break;
1078 default:
1079 break;
1081 return NOTIFY_OK;
1084 static struct notifier_block __cpuinitdata rcu_nb = {
1085 .notifier_call = rcu_cpu_notify,
1088 void __init __rcu_init(void)
1090 int cpu;
1091 int i;
1092 struct rcu_data *rdp;
1094 printk(KERN_NOTICE "Preemptible RCU implementation.\n");
1095 for_each_possible_cpu(cpu) {
1096 rdp = RCU_DATA_CPU(cpu);
1097 spin_lock_init(&rdp->lock);
1098 rdp->completed = 0;
1099 rdp->waitlistcount = 0;
1100 rdp->nextlist = NULL;
1101 rdp->nexttail = &rdp->nextlist;
1102 for (i = 0; i < GP_STAGES; i++) {
1103 rdp->waitlist[i] = NULL;
1104 rdp->waittail[i] = &rdp->waitlist[i];
1106 rdp->donelist = NULL;
1107 rdp->donetail = &rdp->donelist;
1108 rdp->rcu_flipctr[0] = 0;
1109 rdp->rcu_flipctr[1] = 0;
1111 register_cpu_notifier(&rcu_nb);
1114 * We don't need protection against CPU-Hotplug here
1115 * since
1116 * a) If a CPU comes online while we are iterating over the
1117 * cpu_online_map below, we would only end up making a
1118 * duplicate call to rcu_online_cpu() which sets the corresponding
1119 * CPU's mask in the rcu_cpu_online_map.
1121 * b) A CPU cannot go offline at this point in time since the user
1122 * does not have access to the sysfs interface, nor do we
1123 * suspend the system.
1125 for_each_online_cpu(cpu)
1126 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE, (void *)(long) cpu);
1128 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks, NULL);
1132 * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
1134 void synchronize_kernel(void)
1136 synchronize_rcu();
1139 #ifdef CONFIG_RCU_TRACE
1140 long *rcupreempt_flipctr(int cpu)
1142 return &RCU_DATA_CPU(cpu)->rcu_flipctr[0];
1144 EXPORT_SYMBOL_GPL(rcupreempt_flipctr);
1146 int rcupreempt_flip_flag(int cpu)
1148 return per_cpu(rcu_flip_flag, cpu);
1150 EXPORT_SYMBOL_GPL(rcupreempt_flip_flag);
1152 int rcupreempt_mb_flag(int cpu)
1154 return per_cpu(rcu_mb_flag, cpu);
1156 EXPORT_SYMBOL_GPL(rcupreempt_mb_flag);
1158 char *rcupreempt_try_flip_state_name(void)
1160 return rcu_try_flip_state_names[rcu_ctrlblk.rcu_try_flip_state];
1162 EXPORT_SYMBOL_GPL(rcupreempt_try_flip_state_name);
1164 struct rcupreempt_trace *rcupreempt_trace_cpu(int cpu)
1166 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1168 return &rdp->trace;
1170 EXPORT_SYMBOL_GPL(rcupreempt_trace_cpu);
1172 #endif /* #ifdef RCU_TRACE */