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/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/byteorder/swabb.h>
58 #include <linux/cpumask.h>
59 #include <linux/rcupreempt_trace.h>
62 * PREEMPT_RCU data structures.
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.
70 * GP in GP_STAGES stands for Grace Period ;)
74 spinlock_t lock; /* Protect rcu_data fields. */
75 long completed; /* Number of last completed batch. */
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;
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 */
95 * States for rcu_try_flip() and friends.
98 enum rcu_try_flip_states {
101 * Stay here if nothing is happening. Flip the counter if somthing
102 * starts happening. Denoted by "I"
104 rcu_try_flip_idle_state,
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.
113 rcu_try_flip_waitack_state,
116 * Wait here for the sum of the old per-CPU counters to reach zero.
119 rcu_try_flip_waitzero_state,
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".
127 rcu_try_flip_waitmb_state,
131 * States for rcu_ctrlblk.rcu_sched_sleep.
134 enum 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. */
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. */
150 static DEFINE_PER_CPU(struct rcu_data, rcu_data);
151 static struct rcu_ctrlblk rcu_ctrlblk = {
152 .fliplock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.fliplock),
154 .rcu_try_flip_state = rcu_try_flip_idle_state,
155 .schedlock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.schedlock),
156 .sched_sleep = rcu_sched_not_sleeping,
157 .sched_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rcu_ctrlblk.sched_wq),
160 static struct task_struct *rcu_sched_grace_period_task;
162 #ifdef CONFIG_RCU_TRACE
163 static char *rcu_try_flip_state_names[] =
164 { "idle", "waitack", "waitzero", "waitmb" };
165 #endif /* #ifdef CONFIG_RCU_TRACE */
167 static cpumask_t rcu_cpu_online_map __read_mostly = CPU_MASK_NONE;
170 * Enum and per-CPU flag to determine when each CPU has seen
171 * the most recent counter flip.
174 enum rcu_flip_flag_values {
175 rcu_flip_seen, /* Steady/initial state, last flip seen. */
176 /* Only GP detector can update. */
177 rcu_flipped /* Flip just completed, need confirmation. */
178 /* Only corresponding CPU can update. */
180 static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_flip_flag_values, rcu_flip_flag)
184 * Enum and per-CPU flag to determine when each CPU has executed the
185 * needed memory barrier to fence in memory references from its last RCU
186 * read-side critical section in the just-completed grace period.
189 enum rcu_mb_flag_values {
190 rcu_mb_done, /* Steady/initial state, no mb()s required. */
191 /* Only GP detector can update. */
192 rcu_mb_needed /* Flip just completed, need an mb(). */
193 /* Only corresponding CPU can update. */
195 static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_mb_flag_values, rcu_mb_flag)
199 * RCU_DATA_ME: find the current CPU's rcu_data structure.
200 * RCU_DATA_CPU: find the specified CPU's rcu_data structure.
202 #define RCU_DATA_ME() (&__get_cpu_var(rcu_data))
203 #define RCU_DATA_CPU(cpu) (&per_cpu(rcu_data, cpu))
206 * Helper macro for tracing when the appropriate rcu_data is not
207 * cached in a local variable, but where the CPU number is so cached.
209 #define RCU_TRACE_CPU(f, cpu) RCU_TRACE(f, &(RCU_DATA_CPU(cpu)->trace));
212 * Helper macro for tracing when the appropriate rcu_data is not
213 * cached in a local variable.
215 #define RCU_TRACE_ME(f) RCU_TRACE(f, &(RCU_DATA_ME()->trace));
218 * Helper macro for tracing when the appropriate rcu_data is pointed
219 * to by a local variable.
221 #define RCU_TRACE_RDP(f, rdp) RCU_TRACE(f, &((rdp)->trace));
223 #define RCU_SCHED_BATCH_TIME (HZ / 50)
226 * Return the number of RCU batches processed thus far. Useful
227 * for debug and statistics.
229 long rcu_batches_completed(void)
231 return rcu_ctrlblk.completed;
233 EXPORT_SYMBOL_GPL(rcu_batches_completed);
235 void __rcu_read_lock(void)
238 struct task_struct *t = current;
241 nesting = ACCESS_ONCE(t->rcu_read_lock_nesting);
244 /* An earlier rcu_read_lock() covers us, just count it. */
246 t->rcu_read_lock_nesting = nesting + 1;
252 * We disable interrupts for the following reasons:
253 * - If we get scheduling clock interrupt here, and we
254 * end up acking the counter flip, it's like a promise
255 * that we will never increment the old counter again.
256 * Thus we will break that promise if that
257 * scheduling clock interrupt happens between the time
258 * we pick the .completed field and the time that we
259 * increment our counter.
261 * - We don't want to be preempted out here.
263 * NMIs can still occur, of course, and might themselves
264 * contain rcu_read_lock().
267 local_irq_save(flags);
270 * Outermost nesting of rcu_read_lock(), so increment
271 * the current counter for the current CPU. Use volatile
272 * casts to prevent the compiler from reordering.
275 idx = ACCESS_ONCE(rcu_ctrlblk.completed) & 0x1;
276 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr[idx])++;
279 * Now that the per-CPU counter has been incremented, we
280 * are protected from races with rcu_read_lock() invoked
281 * from NMI handlers on this CPU. We can therefore safely
282 * increment the nesting counter, relieving further NMIs
283 * of the need to increment the per-CPU counter.
286 ACCESS_ONCE(t->rcu_read_lock_nesting) = nesting + 1;
289 * Now that we have preventing any NMIs from storing
290 * to the ->rcu_flipctr_idx, we can safely use it to
291 * remember which counter to decrement in the matching
295 ACCESS_ONCE(t->rcu_flipctr_idx) = idx;
296 local_irq_restore(flags);
299 EXPORT_SYMBOL_GPL(__rcu_read_lock);
301 void __rcu_read_unlock(void)
304 struct task_struct *t = current;
307 nesting = ACCESS_ONCE(t->rcu_read_lock_nesting);
311 * We are still protected by the enclosing rcu_read_lock(),
312 * so simply decrement the counter.
315 t->rcu_read_lock_nesting = nesting - 1;
321 * Disable local interrupts to prevent the grace-period
322 * detection state machine from seeing us half-done.
323 * NMIs can still occur, of course, and might themselves
324 * contain rcu_read_lock() and rcu_read_unlock().
327 local_irq_save(flags);
330 * Outermost nesting of rcu_read_unlock(), so we must
331 * decrement the current counter for the current CPU.
332 * This must be done carefully, because NMIs can
333 * occur at any point in this code, and any rcu_read_lock()
334 * and rcu_read_unlock() pairs in the NMI handlers
335 * must interact non-destructively with this code.
336 * Lots of volatile casts, and -very- careful ordering.
338 * Changes to this code, including this one, must be
339 * inspected, validated, and tested extremely carefully!!!
343 * First, pick up the index.
346 idx = ACCESS_ONCE(t->rcu_flipctr_idx);
349 * Now that we have fetched the counter index, it is
350 * safe to decrement the per-task RCU nesting counter.
351 * After this, any interrupts or NMIs will increment and
352 * decrement the per-CPU counters.
354 ACCESS_ONCE(t->rcu_read_lock_nesting) = nesting - 1;
357 * It is now safe to decrement this task's nesting count.
358 * NMIs that occur after this statement will route their
359 * rcu_read_lock() calls through this "else" clause, and
360 * will thus start incrementing the per-CPU counter on
361 * their own. They will also clobber ->rcu_flipctr_idx,
362 * but that is OK, since we have already fetched it.
365 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr[idx])--;
366 local_irq_restore(flags);
369 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
372 * If a global counter flip has occurred since the last time that we
373 * advanced callbacks, advance them. Hardware interrupts must be
374 * disabled when calling this function.
376 static void __rcu_advance_callbacks(struct rcu_data *rdp)
382 if (rdp->completed != rcu_ctrlblk.completed) {
383 if (rdp->waitlist[GP_STAGES - 1] != NULL) {
384 *rdp->donetail = rdp->waitlist[GP_STAGES - 1];
385 rdp->donetail = rdp->waittail[GP_STAGES - 1];
386 RCU_TRACE_RDP(rcupreempt_trace_move2done, rdp);
388 for (i = GP_STAGES - 2; i >= 0; i--) {
389 if (rdp->waitlist[i] != NULL) {
390 rdp->waitlist[i + 1] = rdp->waitlist[i];
391 rdp->waittail[i + 1] = rdp->waittail[i];
394 rdp->waitlist[i + 1] = NULL;
395 rdp->waittail[i + 1] =
396 &rdp->waitlist[i + 1];
399 if (rdp->nextlist != NULL) {
400 rdp->waitlist[0] = rdp->nextlist;
401 rdp->waittail[0] = rdp->nexttail;
403 rdp->nextlist = NULL;
404 rdp->nexttail = &rdp->nextlist;
405 RCU_TRACE_RDP(rcupreempt_trace_move2wait, rdp);
407 rdp->waitlist[0] = NULL;
408 rdp->waittail[0] = &rdp->waitlist[0];
410 rdp->waitlistcount = wlc;
411 rdp->completed = rcu_ctrlblk.completed;
415 * Check to see if this CPU needs to report that it has seen
416 * the most recent counter flip, thereby declaring that all
417 * subsequent rcu_read_lock() invocations will respect this flip.
420 cpu = raw_smp_processor_id();
421 if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
422 smp_mb(); /* Subsequent counter accesses must see new value */
423 per_cpu(rcu_flip_flag, cpu) = rcu_flip_seen;
424 smp_mb(); /* Subsequent RCU read-side critical sections */
425 /* seen -after- acknowledgement. */
429 DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_dyntick_sched, rcu_dyntick_sched) = {
434 static DEFINE_PER_CPU(int, rcu_update_flag);
437 * rcu_irq_enter - Called from Hard irq handlers and NMI/SMI.
439 * If the CPU was idle with dynamic ticks active, this updates the
440 * rcu_dyntick_sched.dynticks to let the RCU handling know that the
443 void rcu_irq_enter(void)
445 int cpu = smp_processor_id();
446 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
448 if (per_cpu(rcu_update_flag, cpu))
449 per_cpu(rcu_update_flag, cpu)++;
452 * Only update if we are coming from a stopped ticks mode
453 * (rcu_dyntick_sched.dynticks is even).
455 if (!in_interrupt() &&
456 (rdssp->dynticks & 0x1) == 0) {
458 * The following might seem like we could have a race
459 * with NMI/SMIs. But this really isn't a problem.
460 * Here we do a read/modify/write, and the race happens
461 * when an NMI/SMI comes in after the read and before
462 * the write. But NMI/SMIs will increment this counter
463 * twice before returning, so the zero bit will not
464 * be corrupted by the NMI/SMI which is the most important
467 * The only thing is that we would bring back the counter
468 * to a postion that it was in during the NMI/SMI.
469 * But the zero bit would be set, so the rest of the
470 * counter would again be ignored.
472 * On return from the IRQ, the counter may have the zero
473 * bit be 0 and the counter the same as the return from
474 * the NMI/SMI. If the state machine was so unlucky to
475 * see that, it still doesn't matter, since all
476 * RCU read-side critical sections on this CPU would
477 * have already completed.
481 * The following memory barrier ensures that any
482 * rcu_read_lock() primitives in the irq handler
483 * are seen by other CPUs to follow the above
484 * increment to rcu_dyntick_sched.dynticks. This is
485 * required in order for other CPUs to correctly
486 * determine when it is safe to advance the RCU
487 * grace-period state machine.
489 smp_mb(); /* see above block comment. */
491 * Since we can't determine the dynamic tick mode from
492 * the rcu_dyntick_sched.dynticks after this routine,
493 * we use a second flag to acknowledge that we came
494 * from an idle state with ticks stopped.
496 per_cpu(rcu_update_flag, cpu)++;
498 * If we take an NMI/SMI now, they will also increment
499 * the rcu_update_flag, and will not update the
500 * rcu_dyntick_sched.dynticks on exit. That is for
507 * rcu_irq_exit - Called from exiting Hard irq context.
509 * If the CPU was idle with dynamic ticks active, update the
510 * rcu_dyntick_sched.dynticks to put let the RCU handling be
511 * aware that the CPU is going back to idle with no ticks.
513 void rcu_irq_exit(void)
515 int cpu = smp_processor_id();
516 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
519 * rcu_update_flag is set if we interrupted the CPU
520 * when it was idle with ticks stopped.
521 * Once this occurs, we keep track of interrupt nesting
522 * because a NMI/SMI could also come in, and we still
523 * only want the IRQ that started the increment of the
524 * rcu_dyntick_sched.dynticks to be the one that modifies
527 if (per_cpu(rcu_update_flag, cpu)) {
528 if (--per_cpu(rcu_update_flag, cpu))
531 /* This must match the interrupt nesting */
532 WARN_ON(in_interrupt());
535 * If an NMI/SMI happens now we are still
536 * protected by the rcu_dyntick_sched.dynticks being odd.
540 * The following memory barrier ensures that any
541 * rcu_read_unlock() primitives in the irq handler
542 * are seen by other CPUs to preceed the following
543 * increment to rcu_dyntick_sched.dynticks. This
544 * is required in order for other CPUs to determine
545 * when it is safe to advance the RCU grace-period
548 smp_mb(); /* see above block comment. */
550 WARN_ON(rdssp->dynticks & 0x1);
554 static void dyntick_save_progress_counter(int cpu)
556 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
558 rdssp->dynticks_snap = rdssp->dynticks;
562 rcu_try_flip_waitack_needed(int cpu)
566 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
568 curr = rdssp->dynticks;
569 snap = rdssp->dynticks_snap;
570 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
573 * If the CPU remained in dynticks mode for the entire time
574 * and didn't take any interrupts, NMIs, SMIs, or whatever,
575 * then it cannot be in the middle of an rcu_read_lock(), so
576 * the next rcu_read_lock() it executes must use the new value
577 * of the counter. So we can safely pretend that this CPU
578 * already acknowledged the counter.
581 if ((curr == snap) && ((curr & 0x1) == 0))
585 * If the CPU passed through or entered a dynticks idle phase with
586 * no active irq handlers, then, as above, we can safely pretend
587 * that this CPU already acknowledged the counter.
590 if ((curr - snap) > 2 || (curr & 0x1) == 0)
593 /* We need this CPU to explicitly acknowledge the counter flip. */
599 rcu_try_flip_waitmb_needed(int cpu)
603 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
605 curr = rdssp->dynticks;
606 snap = rdssp->dynticks_snap;
607 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
610 * If the CPU remained in dynticks mode for the entire time
611 * and didn't take any interrupts, NMIs, SMIs, or whatever,
612 * then it cannot have executed an RCU read-side critical section
613 * during that time, so there is no need for it to execute a
617 if ((curr == snap) && ((curr & 0x1) == 0))
621 * If the CPU either entered or exited an outermost interrupt,
622 * SMI, NMI, or whatever handler, then we know that it executed
623 * a memory barrier when doing so. So we don't need another one.
628 /* We need the CPU to execute a memory barrier. */
633 static void dyntick_save_progress_counter_sched(int cpu)
635 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
637 rdssp->sched_dynticks_snap = rdssp->dynticks;
640 static int rcu_qsctr_inc_needed_dyntick(int cpu)
644 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
646 curr = rdssp->dynticks;
647 snap = rdssp->sched_dynticks_snap;
648 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
651 * If the CPU remained in dynticks mode for the entire time
652 * and didn't take any interrupts, NMIs, SMIs, or whatever,
653 * then it cannot be in the middle of an rcu_read_lock(), so
654 * the next rcu_read_lock() it executes must use the new value
655 * of the counter. Therefore, this CPU has been in a quiescent
656 * state the entire time, and we don't need to wait for it.
659 if ((curr == snap) && ((curr & 0x1) == 0))
663 * If the CPU passed through or entered a dynticks idle phase with
664 * no active irq handlers, then, as above, this CPU has already
665 * passed through a quiescent state.
668 if ((curr - snap) > 2 || (snap & 0x1) == 0)
671 /* We need this CPU to go through a quiescent state. */
676 #else /* !CONFIG_NO_HZ */
678 # define dyntick_save_progress_counter(cpu) do { } while (0)
679 # define rcu_try_flip_waitack_needed(cpu) (1)
680 # define rcu_try_flip_waitmb_needed(cpu) (1)
682 # define dyntick_save_progress_counter_sched(cpu) do { } while (0)
683 # define rcu_qsctr_inc_needed_dyntick(cpu) (1)
685 #endif /* CONFIG_NO_HZ */
687 static void save_qsctr_sched(int cpu)
689 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
691 rdssp->sched_qs_snap = rdssp->sched_qs;
694 static inline int rcu_qsctr_inc_needed(int cpu)
696 struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
699 * If there has been a quiescent state, no more need to wait
703 if (rdssp->sched_qs != rdssp->sched_qs_snap) {
704 smp_mb(); /* force ordering with cpu entering schedule(). */
708 /* We need this CPU to go through a quiescent state. */
714 * Get here when RCU is idle. Decide whether we need to
715 * move out of idle state, and return non-zero if so.
716 * "Straightforward" approach for the moment, might later
717 * use callback-list lengths, grace-period duration, or
718 * some such to determine when to exit idle state.
719 * Might also need a pre-idle test that does not acquire
720 * the lock, but let's get the simple case working first...
724 rcu_try_flip_idle(void)
728 RCU_TRACE_ME(rcupreempt_trace_try_flip_i1);
729 if (!rcu_pending(smp_processor_id())) {
730 RCU_TRACE_ME(rcupreempt_trace_try_flip_ie1);
738 RCU_TRACE_ME(rcupreempt_trace_try_flip_g1);
739 rcu_ctrlblk.completed++; /* stands in for rcu_try_flip_g2 */
742 * Need a memory barrier so that other CPUs see the new
743 * counter value before they see the subsequent change of all
744 * the rcu_flip_flag instances to rcu_flipped.
747 smp_mb(); /* see above block comment. */
749 /* Now ask each CPU for acknowledgement of the flip. */
751 for_each_cpu_mask_nr(cpu, rcu_cpu_online_map) {
752 per_cpu(rcu_flip_flag, cpu) = rcu_flipped;
753 dyntick_save_progress_counter(cpu);
760 * Wait for CPUs to acknowledge the flip.
764 rcu_try_flip_waitack(void)
768 RCU_TRACE_ME(rcupreempt_trace_try_flip_a1);
769 for_each_cpu_mask_nr(cpu, rcu_cpu_online_map)
770 if (rcu_try_flip_waitack_needed(cpu) &&
771 per_cpu(rcu_flip_flag, cpu) != rcu_flip_seen) {
772 RCU_TRACE_ME(rcupreempt_trace_try_flip_ae1);
777 * Make sure our checks above don't bleed into subsequent
778 * waiting for the sum of the counters to reach zero.
781 smp_mb(); /* see above block comment. */
782 RCU_TRACE_ME(rcupreempt_trace_try_flip_a2);
787 * Wait for collective ``last'' counter to reach zero,
788 * then tell all CPUs to do an end-of-grace-period memory barrier.
792 rcu_try_flip_waitzero(void)
795 int lastidx = !(rcu_ctrlblk.completed & 0x1);
798 /* Check to see if the sum of the "last" counters is zero. */
800 RCU_TRACE_ME(rcupreempt_trace_try_flip_z1);
801 for_each_cpu_mask_nr(cpu, rcu_cpu_online_map)
802 sum += RCU_DATA_CPU(cpu)->rcu_flipctr[lastidx];
804 RCU_TRACE_ME(rcupreempt_trace_try_flip_ze1);
809 * This ensures that the other CPUs see the call for
810 * memory barriers -after- the sum to zero has been
813 smp_mb(); /* ^^^^^^^^^^^^ */
815 /* Call for a memory barrier from each CPU. */
816 for_each_cpu_mask_nr(cpu, rcu_cpu_online_map) {
817 per_cpu(rcu_mb_flag, cpu) = rcu_mb_needed;
818 dyntick_save_progress_counter(cpu);
821 RCU_TRACE_ME(rcupreempt_trace_try_flip_z2);
826 * Wait for all CPUs to do their end-of-grace-period memory barrier.
827 * Return 0 once all CPUs have done so.
831 rcu_try_flip_waitmb(void)
835 RCU_TRACE_ME(rcupreempt_trace_try_flip_m1);
836 for_each_cpu_mask_nr(cpu, rcu_cpu_online_map)
837 if (rcu_try_flip_waitmb_needed(cpu) &&
838 per_cpu(rcu_mb_flag, cpu) != rcu_mb_done) {
839 RCU_TRACE_ME(rcupreempt_trace_try_flip_me1);
843 smp_mb(); /* Ensure that the above checks precede any following flip. */
844 RCU_TRACE_ME(rcupreempt_trace_try_flip_m2);
849 * Attempt a single flip of the counters. Remember, a single flip does
850 * -not- constitute a grace period. Instead, the interval between
851 * at least GP_STAGES consecutive flips is a grace period.
853 * If anyone is nuts enough to run this CONFIG_PREEMPT_RCU implementation
854 * on a large SMP, they might want to use a hierarchical organization of
855 * the per-CPU-counter pairs.
857 static void rcu_try_flip(void)
861 RCU_TRACE_ME(rcupreempt_trace_try_flip_1);
862 if (unlikely(!spin_trylock_irqsave(&rcu_ctrlblk.fliplock, flags))) {
863 RCU_TRACE_ME(rcupreempt_trace_try_flip_e1);
868 * Take the next transition(s) through the RCU grace-period
869 * flip-counter state machine.
872 switch (rcu_ctrlblk.rcu_try_flip_state) {
873 case rcu_try_flip_idle_state:
874 if (rcu_try_flip_idle())
875 rcu_ctrlblk.rcu_try_flip_state =
876 rcu_try_flip_waitack_state;
878 case rcu_try_flip_waitack_state:
879 if (rcu_try_flip_waitack())
880 rcu_ctrlblk.rcu_try_flip_state =
881 rcu_try_flip_waitzero_state;
883 case rcu_try_flip_waitzero_state:
884 if (rcu_try_flip_waitzero())
885 rcu_ctrlblk.rcu_try_flip_state =
886 rcu_try_flip_waitmb_state;
888 case rcu_try_flip_waitmb_state:
889 if (rcu_try_flip_waitmb())
890 rcu_ctrlblk.rcu_try_flip_state =
891 rcu_try_flip_idle_state;
893 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
897 * Check to see if this CPU needs to do a memory barrier in order to
898 * ensure that any prior RCU read-side critical sections have committed
899 * their counter manipulations and critical-section memory references
900 * before declaring the grace period to be completed.
902 static void rcu_check_mb(int cpu)
904 if (per_cpu(rcu_mb_flag, cpu) == rcu_mb_needed) {
905 smp_mb(); /* Ensure RCU read-side accesses are visible. */
906 per_cpu(rcu_mb_flag, cpu) = rcu_mb_done;
910 void rcu_check_callbacks(int cpu, int user)
913 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
916 * If this CPU took its interrupt from user mode or from the
917 * idle loop, and this is not a nested interrupt, then
918 * this CPU has to have exited all prior preept-disable
919 * sections of code. So increment the counter to note this.
921 * The memory barrier is needed to handle the case where
922 * writes from a preempt-disable section of code get reordered
923 * into schedule() by this CPU's write buffer. So the memory
924 * barrier makes sure that the rcu_qsctr_inc() is seen by other
925 * CPUs to happen after any such write.
929 (idle_cpu(cpu) && !in_softirq() &&
930 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
931 smp_mb(); /* Guard against aggressive schedule(). */
936 if (rcu_ctrlblk.completed == rdp->completed)
938 spin_lock_irqsave(&rdp->lock, flags);
939 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
940 __rcu_advance_callbacks(rdp);
941 if (rdp->donelist == NULL) {
942 spin_unlock_irqrestore(&rdp->lock, flags);
944 spin_unlock_irqrestore(&rdp->lock, flags);
945 raise_softirq(RCU_SOFTIRQ);
950 * Needed by dynticks, to make sure all RCU processing has finished
953 void rcu_advance_callbacks(int cpu, int user)
956 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
958 if (rcu_ctrlblk.completed == rdp->completed) {
960 if (rcu_ctrlblk.completed == rdp->completed)
963 spin_lock_irqsave(&rdp->lock, flags);
964 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
965 __rcu_advance_callbacks(rdp);
966 spin_unlock_irqrestore(&rdp->lock, flags);
969 #ifdef CONFIG_HOTPLUG_CPU
970 #define rcu_offline_cpu_enqueue(srclist, srctail, dstlist, dsttail) do { \
971 *dsttail = srclist; \
972 if (srclist != NULL) { \
979 void rcu_offline_cpu(int cpu)
982 struct rcu_head *list = NULL;
984 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
985 struct rcu_head *schedlist = NULL;
986 struct rcu_head **schedtail = &schedlist;
987 struct rcu_head **tail = &list;
990 * Remove all callbacks from the newly dead CPU, retaining order.
991 * Otherwise rcu_barrier() will fail
994 spin_lock_irqsave(&rdp->lock, flags);
995 rcu_offline_cpu_enqueue(rdp->donelist, rdp->donetail, list, tail);
996 for (i = GP_STAGES - 1; i >= 0; i--)
997 rcu_offline_cpu_enqueue(rdp->waitlist[i], rdp->waittail[i],
999 rcu_offline_cpu_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
1000 rcu_offline_cpu_enqueue(rdp->waitschedlist, rdp->waitschedtail,
1001 schedlist, schedtail);
1002 rcu_offline_cpu_enqueue(rdp->nextschedlist, rdp->nextschedtail,
1003 schedlist, schedtail);
1004 rdp->rcu_sched_sleeping = 0;
1005 spin_unlock_irqrestore(&rdp->lock, flags);
1006 rdp->waitlistcount = 0;
1008 /* Disengage the newly dead CPU from the grace-period computation. */
1010 spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
1012 if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
1013 smp_mb(); /* Subsequent counter accesses must see new value */
1014 per_cpu(rcu_flip_flag, cpu) = rcu_flip_seen;
1015 smp_mb(); /* Subsequent RCU read-side critical sections */
1016 /* seen -after- acknowledgement. */
1019 RCU_DATA_ME()->rcu_flipctr[0] += RCU_DATA_CPU(cpu)->rcu_flipctr[0];
1020 RCU_DATA_ME()->rcu_flipctr[1] += RCU_DATA_CPU(cpu)->rcu_flipctr[1];
1022 RCU_DATA_CPU(cpu)->rcu_flipctr[0] = 0;
1023 RCU_DATA_CPU(cpu)->rcu_flipctr[1] = 0;
1025 cpu_clear(cpu, rcu_cpu_online_map);
1027 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
1030 * Place the removed callbacks on the current CPU's queue.
1031 * Make them all start a new grace period: simple approach,
1032 * in theory could starve a given set of callbacks, but
1033 * you would need to be doing some serious CPU hotplugging
1034 * to make this happen. If this becomes a problem, adding
1035 * a synchronize_rcu() to the hotplug path would be a simple
1039 local_irq_save(flags); /* disable preempt till we know what lock. */
1040 rdp = RCU_DATA_ME();
1041 spin_lock(&rdp->lock);
1042 *rdp->nexttail = list;
1044 rdp->nexttail = tail;
1045 *rdp->nextschedtail = schedlist;
1047 rdp->nextschedtail = schedtail;
1048 spin_unlock_irqrestore(&rdp->lock, flags);
1051 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1053 void rcu_offline_cpu(int cpu)
1057 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1059 void __cpuinit rcu_online_cpu(int cpu)
1061 unsigned long flags;
1062 struct rcu_data *rdp;
1064 spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
1065 cpu_set(cpu, rcu_cpu_online_map);
1066 spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
1069 * The rcu_sched grace-period processing might have bypassed
1070 * this CPU, given that it was not in the rcu_cpu_online_map
1071 * when the grace-period scan started. This means that the
1072 * grace-period task might sleep. So make sure that if this
1073 * should happen, the first callback posted to this CPU will
1074 * wake up the grace-period task if need be.
1077 rdp = RCU_DATA_CPU(cpu);
1078 spin_lock_irqsave(&rdp->lock, flags);
1079 rdp->rcu_sched_sleeping = 1;
1080 spin_unlock_irqrestore(&rdp->lock, flags);
1083 static void rcu_process_callbacks(struct softirq_action *unused)
1085 unsigned long flags;
1086 struct rcu_head *next, *list;
1087 struct rcu_data *rdp;
1089 local_irq_save(flags);
1090 rdp = RCU_DATA_ME();
1091 spin_lock(&rdp->lock);
1092 list = rdp->donelist;
1094 spin_unlock_irqrestore(&rdp->lock, flags);
1097 rdp->donelist = NULL;
1098 rdp->donetail = &rdp->donelist;
1099 RCU_TRACE_RDP(rcupreempt_trace_done_remove, rdp);
1100 spin_unlock_irqrestore(&rdp->lock, flags);
1105 RCU_TRACE_ME(rcupreempt_trace_invoke);
1109 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1111 unsigned long flags;
1112 struct rcu_data *rdp;
1116 local_irq_save(flags);
1117 rdp = RCU_DATA_ME();
1118 spin_lock(&rdp->lock);
1119 __rcu_advance_callbacks(rdp);
1120 *rdp->nexttail = head;
1121 rdp->nexttail = &head->next;
1122 RCU_TRACE_RDP(rcupreempt_trace_next_add, rdp);
1123 spin_unlock_irqrestore(&rdp->lock, flags);
1125 EXPORT_SYMBOL_GPL(call_rcu);
1127 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1129 unsigned long flags;
1130 struct rcu_data *rdp;
1135 local_irq_save(flags);
1136 rdp = RCU_DATA_ME();
1137 spin_lock(&rdp->lock);
1138 *rdp->nextschedtail = head;
1139 rdp->nextschedtail = &head->next;
1140 if (rdp->rcu_sched_sleeping) {
1142 /* Grace-period processing might be sleeping... */
1144 rdp->rcu_sched_sleeping = 0;
1147 spin_unlock_irqrestore(&rdp->lock, flags);
1150 /* Wake up grace-period processing, unless someone beat us. */
1152 spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
1153 if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
1155 rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
1156 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1158 wake_up_interruptible(&rcu_ctrlblk.sched_wq);
1161 EXPORT_SYMBOL_GPL(call_rcu_sched);
1164 * Wait until all currently running preempt_disable() code segments
1165 * (including hardware-irq-disable segments) complete. Note that
1166 * in -rt this does -not- necessarily result in all currently executing
1167 * interrupt -handlers- having completed.
1169 synchronize_rcu_xxx(__synchronize_sched, call_rcu_sched)
1170 EXPORT_SYMBOL_GPL(__synchronize_sched);
1173 * kthread function that manages call_rcu_sched grace periods.
1175 static int rcu_sched_grace_period(void *arg)
1177 int couldsleep; /* might sleep after current pass. */
1178 int couldsleepnext = 0; /* might sleep after next pass. */
1180 unsigned long flags;
1181 struct rcu_data *rdp;
1185 * Each pass through the following loop handles one
1186 * rcu_sched grace period cycle.
1189 /* Save each CPU's current state. */
1191 for_each_online_cpu(cpu) {
1192 dyntick_save_progress_counter_sched(cpu);
1193 save_qsctr_sched(cpu);
1197 * Sleep for about an RCU grace-period's worth to
1198 * allow better batching and to consume less CPU.
1200 schedule_timeout_interruptible(RCU_SCHED_BATCH_TIME);
1203 * If there was nothing to do last time, prepare to
1204 * sleep at the end of the current grace period cycle.
1206 couldsleep = couldsleepnext;
1209 spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
1210 rcu_ctrlblk.sched_sleep = rcu_sched_sleep_prep;
1211 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1215 * Wait on each CPU in turn to have either visited
1216 * a quiescent state or been in dynticks-idle mode.
1218 for_each_online_cpu(cpu) {
1219 while (rcu_qsctr_inc_needed(cpu) &&
1220 rcu_qsctr_inc_needed_dyntick(cpu)) {
1221 /* resched_cpu(cpu); @@@ */
1222 schedule_timeout_interruptible(1);
1226 /* Advance callbacks for each CPU. */
1228 for_each_online_cpu(cpu) {
1230 rdp = RCU_DATA_CPU(cpu);
1231 spin_lock_irqsave(&rdp->lock, flags);
1234 * We are running on this CPU irq-disabled, so no
1235 * CPU can go offline until we re-enable irqs.
1236 * The current CPU might have already gone
1237 * offline (between the for_each_offline_cpu and
1238 * the spin_lock_irqsave), but in that case all its
1239 * callback lists will be empty, so no harm done.
1241 * Advance the callbacks! We share normal RCU's
1242 * donelist, since callbacks are invoked the
1243 * same way in either case.
1245 if (rdp->waitschedlist != NULL) {
1246 *rdp->donetail = rdp->waitschedlist;
1247 rdp->donetail = rdp->waitschedtail;
1250 * Next rcu_check_callbacks() will
1251 * do the required raise_softirq().
1254 if (rdp->nextschedlist != NULL) {
1255 rdp->waitschedlist = rdp->nextschedlist;
1256 rdp->waitschedtail = rdp->nextschedtail;
1260 rdp->waitschedlist = NULL;
1261 rdp->waitschedtail = &rdp->waitschedlist;
1263 rdp->nextschedlist = NULL;
1264 rdp->nextschedtail = &rdp->nextschedlist;
1266 /* Mark sleep intention. */
1268 rdp->rcu_sched_sleeping = couldsleep;
1270 spin_unlock_irqrestore(&rdp->lock, flags);
1273 /* If we saw callbacks on the last scan, go deal with them. */
1278 /* Attempt to block... */
1280 spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
1281 if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
1284 * Someone posted a callback after we scanned.
1285 * Go take care of it.
1287 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1292 /* Block until the next person posts a callback. */
1294 rcu_ctrlblk.sched_sleep = rcu_sched_sleeping;
1295 spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
1297 __wait_event_interruptible(rcu_ctrlblk.sched_wq,
1298 rcu_ctrlblk.sched_sleep != rcu_sched_sleeping,
1302 * Signals would prevent us from sleeping, and we cannot
1303 * do much with them in any case. So flush them.
1306 flush_signals(current);
1309 } while (!kthread_should_stop());
1315 * Check to see if any future RCU-related work will need to be done
1316 * by the current CPU, even if none need be done immediately, returning
1317 * 1 if so. Assumes that notifiers would take care of handling any
1318 * outstanding requests from the RCU core.
1320 * This function is part of the RCU implementation; it is -not-
1321 * an exported member of the RCU API.
1323 int rcu_needs_cpu(int cpu)
1325 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1327 return (rdp->donelist != NULL ||
1328 !!rdp->waitlistcount ||
1329 rdp->nextlist != NULL ||
1330 rdp->nextschedlist != NULL ||
1331 rdp->waitschedlist != NULL);
1334 int rcu_pending(int cpu)
1336 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1338 /* The CPU has at least one callback queued somewhere. */
1340 if (rdp->donelist != NULL ||
1341 !!rdp->waitlistcount ||
1342 rdp->nextlist != NULL ||
1343 rdp->nextschedlist != NULL ||
1344 rdp->waitschedlist != NULL)
1347 /* The RCU core needs an acknowledgement from this CPU. */
1349 if ((per_cpu(rcu_flip_flag, cpu) == rcu_flipped) ||
1350 (per_cpu(rcu_mb_flag, cpu) == rcu_mb_needed))
1353 /* This CPU has fallen behind the global grace-period number. */
1355 if (rdp->completed != rcu_ctrlblk.completed)
1358 /* Nothing needed from this CPU. */
1363 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1364 unsigned long action, void *hcpu)
1366 long cpu = (long)hcpu;
1369 case CPU_UP_PREPARE:
1370 case CPU_UP_PREPARE_FROZEN:
1371 rcu_online_cpu(cpu);
1373 case CPU_UP_CANCELED:
1374 case CPU_UP_CANCELED_FROZEN:
1376 case CPU_DEAD_FROZEN:
1377 rcu_offline_cpu(cpu);
1385 static struct notifier_block __cpuinitdata rcu_nb = {
1386 .notifier_call = rcu_cpu_notify,
1389 void __init __rcu_init(void)
1393 struct rcu_data *rdp;
1395 printk(KERN_NOTICE "Preemptible RCU implementation.\n");
1396 for_each_possible_cpu(cpu) {
1397 rdp = RCU_DATA_CPU(cpu);
1398 spin_lock_init(&rdp->lock);
1400 rdp->waitlistcount = 0;
1401 rdp->nextlist = NULL;
1402 rdp->nexttail = &rdp->nextlist;
1403 for (i = 0; i < GP_STAGES; i++) {
1404 rdp->waitlist[i] = NULL;
1405 rdp->waittail[i] = &rdp->waitlist[i];
1407 rdp->donelist = NULL;
1408 rdp->donetail = &rdp->donelist;
1409 rdp->rcu_flipctr[0] = 0;
1410 rdp->rcu_flipctr[1] = 0;
1411 rdp->nextschedlist = NULL;
1412 rdp->nextschedtail = &rdp->nextschedlist;
1413 rdp->waitschedlist = NULL;
1414 rdp->waitschedtail = &rdp->waitschedlist;
1415 rdp->rcu_sched_sleeping = 0;
1417 register_cpu_notifier(&rcu_nb);
1420 * We don't need protection against CPU-Hotplug here
1422 * a) If a CPU comes online while we are iterating over the
1423 * cpu_online_map below, we would only end up making a
1424 * duplicate call to rcu_online_cpu() which sets the corresponding
1425 * CPU's mask in the rcu_cpu_online_map.
1427 * b) A CPU cannot go offline at this point in time since the user
1428 * does not have access to the sysfs interface, nor do we
1429 * suspend the system.
1431 for_each_online_cpu(cpu)
1432 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE, (void *)(long) cpu);
1434 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1438 * Late-boot-time RCU initialization that must wait until after scheduler
1439 * has been initialized.
1441 void __init rcu_init_sched(void)
1443 rcu_sched_grace_period_task = kthread_run(rcu_sched_grace_period,
1445 "rcu_sched_grace_period");
1446 WARN_ON(IS_ERR(rcu_sched_grace_period_task));
1449 #ifdef CONFIG_RCU_TRACE
1450 long *rcupreempt_flipctr(int cpu)
1452 return &RCU_DATA_CPU(cpu)->rcu_flipctr[0];
1454 EXPORT_SYMBOL_GPL(rcupreempt_flipctr);
1456 int rcupreempt_flip_flag(int cpu)
1458 return per_cpu(rcu_flip_flag, cpu);
1460 EXPORT_SYMBOL_GPL(rcupreempt_flip_flag);
1462 int rcupreempt_mb_flag(int cpu)
1464 return per_cpu(rcu_mb_flag, cpu);
1466 EXPORT_SYMBOL_GPL(rcupreempt_mb_flag);
1468 char *rcupreempt_try_flip_state_name(void)
1470 return rcu_try_flip_state_names[rcu_ctrlblk.rcu_try_flip_state];
1472 EXPORT_SYMBOL_GPL(rcupreempt_try_flip_state_name);
1474 struct rcupreempt_trace *rcupreempt_trace_cpu(int cpu)
1476 struct rcu_data *rdp = RCU_DATA_CPU(cpu);
1480 EXPORT_SYMBOL_GPL(rcupreempt_trace_cpu);
1482 #endif /* #ifdef RCU_TRACE */