2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
58 #include <trace/events/rcu.h>
62 /* Data structures. */
64 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
65 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
67 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) { \
68 .level = { &sname##_state.node[0] }, \
70 .fqs_state = RCU_GP_IDLE, \
71 .gpnum = 0UL - 300UL, \
72 .completed = 0UL - 300UL, \
73 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
77 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
82 struct rcu_state rcu_sched_state =
83 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
84 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
86 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
87 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
89 static struct rcu_state *rcu_state;
90 LIST_HEAD(rcu_struct_flavors);
92 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
93 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
94 module_param(rcu_fanout_leaf, int, 0444);
95 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
96 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
103 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
106 * The rcu_scheduler_active variable transitions from zero to one just
107 * before the first task is spawned. So when this variable is zero, RCU
108 * can assume that there is but one task, allowing RCU to (for example)
109 * optimize synchronize_sched() to a simple barrier(). When this variable
110 * is one, RCU must actually do all the hard work required to detect real
111 * grace periods. This variable is also used to suppress boot-time false
112 * positives from lockdep-RCU error checking.
114 int rcu_scheduler_active __read_mostly;
115 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
118 * The rcu_scheduler_fully_active variable transitions from zero to one
119 * during the early_initcall() processing, which is after the scheduler
120 * is capable of creating new tasks. So RCU processing (for example,
121 * creating tasks for RCU priority boosting) must be delayed until after
122 * rcu_scheduler_fully_active transitions from zero to one. We also
123 * currently delay invocation of any RCU callbacks until after this point.
125 * It might later prove better for people registering RCU callbacks during
126 * early boot to take responsibility for these callbacks, but one step at
129 static int rcu_scheduler_fully_active __read_mostly;
131 #ifdef CONFIG_RCU_BOOST
134 * Control variables for per-CPU and per-rcu_node kthreads. These
135 * handle all flavors of RCU.
137 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
139 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
140 DEFINE_PER_CPU(char, rcu_cpu_has_work);
142 #endif /* #ifdef CONFIG_RCU_BOOST */
144 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
145 static void invoke_rcu_core(void);
146 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
149 * Track the rcutorture test sequence number and the update version
150 * number within a given test. The rcutorture_testseq is incremented
151 * on every rcutorture module load and unload, so has an odd value
152 * when a test is running. The rcutorture_vernum is set to zero
153 * when rcutorture starts and is incremented on each rcutorture update.
154 * These variables enable correlating rcutorture output with the
155 * RCU tracing information.
157 unsigned long rcutorture_testseq;
158 unsigned long rcutorture_vernum;
161 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
162 * permit this function to be invoked without holding the root rcu_node
163 * structure's ->lock, but of course results can be subject to change.
165 static int rcu_gp_in_progress(struct rcu_state *rsp)
167 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
171 * Note a quiescent state. Because we do not need to know
172 * how many quiescent states passed, just if there was at least
173 * one since the start of the grace period, this just sets a flag.
174 * The caller must have disabled preemption.
176 void rcu_sched_qs(int cpu)
178 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
180 if (rdp->passed_quiesce == 0)
181 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
182 rdp->passed_quiesce = 1;
185 void rcu_bh_qs(int cpu)
187 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
189 if (rdp->passed_quiesce == 0)
190 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
191 rdp->passed_quiesce = 1;
195 * Note a context switch. This is a quiescent state for RCU-sched,
196 * and requires special handling for preemptible RCU.
197 * The caller must have disabled preemption.
199 void rcu_note_context_switch(int cpu)
201 trace_rcu_utilization("Start context switch");
203 rcu_preempt_note_context_switch(cpu);
204 trace_rcu_utilization("End context switch");
206 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
208 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
209 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
210 .dynticks = ATOMIC_INIT(1),
213 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
214 static long qhimark = 10000; /* If this many pending, ignore blimit. */
215 static long qlowmark = 100; /* Once only this many pending, use blimit. */
217 module_param(blimit, long, 0444);
218 module_param(qhimark, long, 0444);
219 module_param(qlowmark, long, 0444);
221 static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
222 static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
224 module_param(jiffies_till_first_fqs, ulong, 0644);
225 module_param(jiffies_till_next_fqs, ulong, 0644);
227 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
228 static void force_quiescent_state(struct rcu_state *rsp);
229 static int rcu_pending(int cpu);
232 * Return the number of RCU-sched batches processed thus far for debug & stats.
234 long rcu_batches_completed_sched(void)
236 return rcu_sched_state.completed;
238 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
241 * Return the number of RCU BH batches processed thus far for debug & stats.
243 long rcu_batches_completed_bh(void)
245 return rcu_bh_state.completed;
247 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
250 * Force a quiescent state for RCU BH.
252 void rcu_bh_force_quiescent_state(void)
254 force_quiescent_state(&rcu_bh_state);
256 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
259 * Record the number of times rcutorture tests have been initiated and
260 * terminated. This information allows the debugfs tracing stats to be
261 * correlated to the rcutorture messages, even when the rcutorture module
262 * is being repeatedly loaded and unloaded. In other words, we cannot
263 * store this state in rcutorture itself.
265 void rcutorture_record_test_transition(void)
267 rcutorture_testseq++;
268 rcutorture_vernum = 0;
270 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
273 * Record the number of writer passes through the current rcutorture test.
274 * This is also used to correlate debugfs tracing stats with the rcutorture
277 void rcutorture_record_progress(unsigned long vernum)
281 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
284 * Force a quiescent state for RCU-sched.
286 void rcu_sched_force_quiescent_state(void)
288 force_quiescent_state(&rcu_sched_state);
290 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
293 * Does the CPU have callbacks ready to be invoked?
296 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
298 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
299 rdp->nxttail[RCU_DONE_TAIL] != NULL;
303 * Does the current CPU require a not-yet-started grace period?
304 * The caller must have disabled interrupts to prevent races with
305 * normal callback registry.
308 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
312 if (rcu_gp_in_progress(rsp))
313 return 0; /* No, a grace period is already in progress. */
314 if (rcu_nocb_needs_gp(rsp))
315 return 1; /* Yes, a no-CBs CPU needs one. */
316 if (!rdp->nxttail[RCU_NEXT_TAIL])
317 return 0; /* No, this is a no-CBs (or offline) CPU. */
318 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
319 return 1; /* Yes, this CPU has newly registered callbacks. */
320 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
321 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
322 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
323 rdp->nxtcompleted[i]))
324 return 1; /* Yes, CBs for future grace period. */
325 return 0; /* No grace period needed. */
329 * Return the root node of the specified rcu_state structure.
331 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
333 return &rsp->node[0];
337 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
339 * If the new value of the ->dynticks_nesting counter now is zero,
340 * we really have entered idle, and must do the appropriate accounting.
341 * The caller must have disabled interrupts.
343 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
346 trace_rcu_dyntick("Start", oldval, rdtp->dynticks_nesting);
347 if (!user && !is_idle_task(current)) {
348 struct task_struct *idle = idle_task(smp_processor_id());
350 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
351 ftrace_dump(DUMP_ORIG);
352 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
353 current->pid, current->comm,
354 idle->pid, idle->comm); /* must be idle task! */
356 rcu_prepare_for_idle(smp_processor_id());
357 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
358 smp_mb__before_atomic_inc(); /* See above. */
359 atomic_inc(&rdtp->dynticks);
360 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
361 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
364 * It is illegal to enter an extended quiescent state while
365 * in an RCU read-side critical section.
367 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
368 "Illegal idle entry in RCU read-side critical section.");
369 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
370 "Illegal idle entry in RCU-bh read-side critical section.");
371 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
372 "Illegal idle entry in RCU-sched read-side critical section.");
376 * Enter an RCU extended quiescent state, which can be either the
377 * idle loop or adaptive-tickless usermode execution.
379 static void rcu_eqs_enter(bool user)
382 struct rcu_dynticks *rdtp;
384 rdtp = &__get_cpu_var(rcu_dynticks);
385 oldval = rdtp->dynticks_nesting;
386 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
387 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
388 rdtp->dynticks_nesting = 0;
390 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
391 rcu_eqs_enter_common(rdtp, oldval, user);
395 * rcu_idle_enter - inform RCU that current CPU is entering idle
397 * Enter idle mode, in other words, -leave- the mode in which RCU
398 * read-side critical sections can occur. (Though RCU read-side
399 * critical sections can occur in irq handlers in idle, a possibility
400 * handled by irq_enter() and irq_exit().)
402 * We crowbar the ->dynticks_nesting field to zero to allow for
403 * the possibility of usermode upcalls having messed up our count
404 * of interrupt nesting level during the prior busy period.
406 void rcu_idle_enter(void)
410 local_irq_save(flags);
411 rcu_eqs_enter(false);
412 local_irq_restore(flags);
414 EXPORT_SYMBOL_GPL(rcu_idle_enter);
416 #ifdef CONFIG_RCU_USER_QS
418 * rcu_user_enter - inform RCU that we are resuming userspace.
420 * Enter RCU idle mode right before resuming userspace. No use of RCU
421 * is permitted between this call and rcu_user_exit(). This way the
422 * CPU doesn't need to maintain the tick for RCU maintenance purposes
423 * when the CPU runs in userspace.
425 void rcu_user_enter(void)
431 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
432 * after the current irq returns.
434 * This is similar to rcu_user_enter() but in the context of a non-nesting
435 * irq. After this call, RCU enters into idle mode when the interrupt
438 void rcu_user_enter_after_irq(void)
441 struct rcu_dynticks *rdtp;
443 local_irq_save(flags);
444 rdtp = &__get_cpu_var(rcu_dynticks);
445 /* Ensure this irq is interrupting a non-idle RCU state. */
446 WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
447 rdtp->dynticks_nesting = 1;
448 local_irq_restore(flags);
450 #endif /* CONFIG_RCU_USER_QS */
453 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
455 * Exit from an interrupt handler, which might possibly result in entering
456 * idle mode, in other words, leaving the mode in which read-side critical
457 * sections can occur.
459 * This code assumes that the idle loop never does anything that might
460 * result in unbalanced calls to irq_enter() and irq_exit(). If your
461 * architecture violates this assumption, RCU will give you what you
462 * deserve, good and hard. But very infrequently and irreproducibly.
464 * Use things like work queues to work around this limitation.
466 * You have been warned.
468 void rcu_irq_exit(void)
472 struct rcu_dynticks *rdtp;
474 local_irq_save(flags);
475 rdtp = &__get_cpu_var(rcu_dynticks);
476 oldval = rdtp->dynticks_nesting;
477 rdtp->dynticks_nesting--;
478 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
479 if (rdtp->dynticks_nesting)
480 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
482 rcu_eqs_enter_common(rdtp, oldval, true);
483 local_irq_restore(flags);
487 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
489 * If the new value of the ->dynticks_nesting counter was previously zero,
490 * we really have exited idle, and must do the appropriate accounting.
491 * The caller must have disabled interrupts.
493 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
496 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
497 atomic_inc(&rdtp->dynticks);
498 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
499 smp_mb__after_atomic_inc(); /* See above. */
500 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
501 rcu_cleanup_after_idle(smp_processor_id());
502 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
503 if (!user && !is_idle_task(current)) {
504 struct task_struct *idle = idle_task(smp_processor_id());
506 trace_rcu_dyntick("Error on exit: not idle task",
507 oldval, rdtp->dynticks_nesting);
508 ftrace_dump(DUMP_ORIG);
509 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
510 current->pid, current->comm,
511 idle->pid, idle->comm); /* must be idle task! */
516 * Exit an RCU extended quiescent state, which can be either the
517 * idle loop or adaptive-tickless usermode execution.
519 static void rcu_eqs_exit(bool user)
521 struct rcu_dynticks *rdtp;
524 rdtp = &__get_cpu_var(rcu_dynticks);
525 oldval = rdtp->dynticks_nesting;
526 WARN_ON_ONCE(oldval < 0);
527 if (oldval & DYNTICK_TASK_NEST_MASK)
528 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
530 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
531 rcu_eqs_exit_common(rdtp, oldval, user);
535 * rcu_idle_exit - inform RCU that current CPU is leaving idle
537 * Exit idle mode, in other words, -enter- the mode in which RCU
538 * read-side critical sections can occur.
540 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
541 * allow for the possibility of usermode upcalls messing up our count
542 * of interrupt nesting level during the busy period that is just
545 void rcu_idle_exit(void)
549 local_irq_save(flags);
551 local_irq_restore(flags);
553 EXPORT_SYMBOL_GPL(rcu_idle_exit);
555 #ifdef CONFIG_RCU_USER_QS
557 * rcu_user_exit - inform RCU that we are exiting userspace.
559 * Exit RCU idle mode while entering the kernel because it can
560 * run a RCU read side critical section anytime.
562 void rcu_user_exit(void)
568 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
569 * idle mode after the current non-nesting irq returns.
571 * This is similar to rcu_user_exit() but in the context of an irq.
572 * This is called when the irq has interrupted a userspace RCU idle mode
573 * context. When the current non-nesting interrupt returns after this call,
574 * the CPU won't restore the RCU idle mode.
576 void rcu_user_exit_after_irq(void)
579 struct rcu_dynticks *rdtp;
581 local_irq_save(flags);
582 rdtp = &__get_cpu_var(rcu_dynticks);
583 /* Ensure we are interrupting an RCU idle mode. */
584 WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
585 rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
586 local_irq_restore(flags);
588 #endif /* CONFIG_RCU_USER_QS */
591 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
593 * Enter an interrupt handler, which might possibly result in exiting
594 * idle mode, in other words, entering the mode in which read-side critical
595 * sections can occur.
597 * Note that the Linux kernel is fully capable of entering an interrupt
598 * handler that it never exits, for example when doing upcalls to
599 * user mode! This code assumes that the idle loop never does upcalls to
600 * user mode. If your architecture does do upcalls from the idle loop (or
601 * does anything else that results in unbalanced calls to the irq_enter()
602 * and irq_exit() functions), RCU will give you what you deserve, good
603 * and hard. But very infrequently and irreproducibly.
605 * Use things like work queues to work around this limitation.
607 * You have been warned.
609 void rcu_irq_enter(void)
612 struct rcu_dynticks *rdtp;
615 local_irq_save(flags);
616 rdtp = &__get_cpu_var(rcu_dynticks);
617 oldval = rdtp->dynticks_nesting;
618 rdtp->dynticks_nesting++;
619 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
621 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
623 rcu_eqs_exit_common(rdtp, oldval, true);
624 local_irq_restore(flags);
628 * rcu_nmi_enter - inform RCU of entry to NMI context
630 * If the CPU was idle with dynamic ticks active, and there is no
631 * irq handler running, this updates rdtp->dynticks_nmi to let the
632 * RCU grace-period handling know that the CPU is active.
634 void rcu_nmi_enter(void)
636 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
638 if (rdtp->dynticks_nmi_nesting == 0 &&
639 (atomic_read(&rdtp->dynticks) & 0x1))
641 rdtp->dynticks_nmi_nesting++;
642 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
643 atomic_inc(&rdtp->dynticks);
644 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
645 smp_mb__after_atomic_inc(); /* See above. */
646 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
650 * rcu_nmi_exit - inform RCU of exit from NMI context
652 * If the CPU was idle with dynamic ticks active, and there is no
653 * irq handler running, this updates rdtp->dynticks_nmi to let the
654 * RCU grace-period handling know that the CPU is no longer active.
656 void rcu_nmi_exit(void)
658 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
660 if (rdtp->dynticks_nmi_nesting == 0 ||
661 --rdtp->dynticks_nmi_nesting != 0)
663 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
664 smp_mb__before_atomic_inc(); /* See above. */
665 atomic_inc(&rdtp->dynticks);
666 smp_mb__after_atomic_inc(); /* Force delay to next write. */
667 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
671 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
673 * If the current CPU is in its idle loop and is neither in an interrupt
674 * or NMI handler, return true.
676 int rcu_is_cpu_idle(void)
681 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
685 EXPORT_SYMBOL(rcu_is_cpu_idle);
687 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
690 * Is the current CPU online? Disable preemption to avoid false positives
691 * that could otherwise happen due to the current CPU number being sampled,
692 * this task being preempted, its old CPU being taken offline, resuming
693 * on some other CPU, then determining that its old CPU is now offline.
694 * It is OK to use RCU on an offline processor during initial boot, hence
695 * the check for rcu_scheduler_fully_active. Note also that it is OK
696 * for a CPU coming online to use RCU for one jiffy prior to marking itself
697 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
698 * offline to continue to use RCU for one jiffy after marking itself
699 * offline in the cpu_online_mask. This leniency is necessary given the
700 * non-atomic nature of the online and offline processing, for example,
701 * the fact that a CPU enters the scheduler after completing the CPU_DYING
704 * This is also why RCU internally marks CPUs online during the
705 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
707 * Disable checking if in an NMI handler because we cannot safely report
708 * errors from NMI handlers anyway.
710 bool rcu_lockdep_current_cpu_online(void)
712 struct rcu_data *rdp;
713 struct rcu_node *rnp;
719 rdp = &__get_cpu_var(rcu_sched_data);
721 ret = (rdp->grpmask & rnp->qsmaskinit) ||
722 !rcu_scheduler_fully_active;
726 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
728 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
731 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
733 * If the current CPU is idle or running at a first-level (not nested)
734 * interrupt from idle, return true. The caller must have at least
735 * disabled preemption.
737 static int rcu_is_cpu_rrupt_from_idle(void)
739 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
743 * Snapshot the specified CPU's dynticks counter so that we can later
744 * credit them with an implicit quiescent state. Return 1 if this CPU
745 * is in dynticks idle mode, which is an extended quiescent state.
747 static int dyntick_save_progress_counter(struct rcu_data *rdp)
749 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
750 return (rdp->dynticks_snap & 0x1) == 0;
754 * Return true if the specified CPU has passed through a quiescent
755 * state by virtue of being in or having passed through an dynticks
756 * idle state since the last call to dyntick_save_progress_counter()
757 * for this same CPU, or by virtue of having been offline.
759 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
764 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
765 snap = (unsigned int)rdp->dynticks_snap;
768 * If the CPU passed through or entered a dynticks idle phase with
769 * no active irq/NMI handlers, then we can safely pretend that the CPU
770 * already acknowledged the request to pass through a quiescent
771 * state. Either way, that CPU cannot possibly be in an RCU
772 * read-side critical section that started before the beginning
773 * of the current RCU grace period.
775 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
776 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
782 * Check for the CPU being offline, but only if the grace period
783 * is old enough. We don't need to worry about the CPU changing
784 * state: If we see it offline even once, it has been through a
787 * The reason for insisting that the grace period be at least
788 * one jiffy old is that CPUs that are not quite online and that
789 * have just gone offline can still execute RCU read-side critical
792 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
793 return 0; /* Grace period is not old enough. */
795 if (cpu_is_offline(rdp->cpu)) {
796 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
803 static void record_gp_stall_check_time(struct rcu_state *rsp)
805 rsp->gp_start = jiffies;
806 rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
810 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
811 * for architectures that do not implement trigger_all_cpu_backtrace().
812 * The NMI-triggered stack traces are more accurate because they are
813 * printed by the target CPU.
815 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
819 struct rcu_node *rnp;
821 rcu_for_each_leaf_node(rsp, rnp) {
822 raw_spin_lock_irqsave(&rnp->lock, flags);
823 if (rnp->qsmask != 0) {
824 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
825 if (rnp->qsmask & (1UL << cpu))
826 dump_cpu_task(rnp->grplo + cpu);
828 raw_spin_unlock_irqrestore(&rnp->lock, flags);
832 static void print_other_cpu_stall(struct rcu_state *rsp)
838 struct rcu_node *rnp = rcu_get_root(rsp);
841 /* Only let one CPU complain about others per time interval. */
843 raw_spin_lock_irqsave(&rnp->lock, flags);
844 delta = jiffies - rsp->jiffies_stall;
845 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
846 raw_spin_unlock_irqrestore(&rnp->lock, flags);
849 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
850 raw_spin_unlock_irqrestore(&rnp->lock, flags);
853 * OK, time to rat on our buddy...
854 * See Documentation/RCU/stallwarn.txt for info on how to debug
855 * RCU CPU stall warnings.
857 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
859 print_cpu_stall_info_begin();
860 rcu_for_each_leaf_node(rsp, rnp) {
861 raw_spin_lock_irqsave(&rnp->lock, flags);
862 ndetected += rcu_print_task_stall(rnp);
863 if (rnp->qsmask != 0) {
864 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
865 if (rnp->qsmask & (1UL << cpu)) {
866 print_cpu_stall_info(rsp,
871 raw_spin_unlock_irqrestore(&rnp->lock, flags);
875 * Now rat on any tasks that got kicked up to the root rcu_node
876 * due to CPU offlining.
878 rnp = rcu_get_root(rsp);
879 raw_spin_lock_irqsave(&rnp->lock, flags);
880 ndetected += rcu_print_task_stall(rnp);
881 raw_spin_unlock_irqrestore(&rnp->lock, flags);
883 print_cpu_stall_info_end();
884 for_each_possible_cpu(cpu)
885 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
886 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
887 smp_processor_id(), (long)(jiffies - rsp->gp_start),
888 rsp->gpnum, rsp->completed, totqlen);
890 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
891 else if (!trigger_all_cpu_backtrace())
892 rcu_dump_cpu_stacks(rsp);
894 /* Complain about tasks blocking the grace period. */
896 rcu_print_detail_task_stall(rsp);
898 force_quiescent_state(rsp); /* Kick them all. */
901 static void print_cpu_stall(struct rcu_state *rsp)
905 struct rcu_node *rnp = rcu_get_root(rsp);
909 * OK, time to rat on ourselves...
910 * See Documentation/RCU/stallwarn.txt for info on how to debug
911 * RCU CPU stall warnings.
913 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
914 print_cpu_stall_info_begin();
915 print_cpu_stall_info(rsp, smp_processor_id());
916 print_cpu_stall_info_end();
917 for_each_possible_cpu(cpu)
918 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
919 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
920 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
921 if (!trigger_all_cpu_backtrace())
924 raw_spin_lock_irqsave(&rnp->lock, flags);
925 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
926 rsp->jiffies_stall = jiffies +
927 3 * rcu_jiffies_till_stall_check() + 3;
928 raw_spin_unlock_irqrestore(&rnp->lock, flags);
930 set_need_resched(); /* kick ourselves to get things going. */
933 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
937 struct rcu_node *rnp;
939 if (rcu_cpu_stall_suppress)
941 j = ACCESS_ONCE(jiffies);
942 js = ACCESS_ONCE(rsp->jiffies_stall);
944 if (rcu_gp_in_progress(rsp) &&
945 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
947 /* We haven't checked in, so go dump stack. */
948 print_cpu_stall(rsp);
950 } else if (rcu_gp_in_progress(rsp) &&
951 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
953 /* They had a few time units to dump stack, so complain. */
954 print_other_cpu_stall(rsp);
959 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
961 * Set the stall-warning timeout way off into the future, thus preventing
962 * any RCU CPU stall-warning messages from appearing in the current set of
965 * The caller must disable hard irqs.
967 void rcu_cpu_stall_reset(void)
969 struct rcu_state *rsp;
971 for_each_rcu_flavor(rsp)
972 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
976 * Update CPU-local rcu_data state to record the newly noticed grace period.
977 * This is used both when we started the grace period and when we notice
978 * that someone else started the grace period. The caller must hold the
979 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
980 * and must have irqs disabled.
982 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
984 if (rdp->gpnum != rnp->gpnum) {
986 * If the current grace period is waiting for this CPU,
987 * set up to detect a quiescent state, otherwise don't
988 * go looking for one.
990 rdp->gpnum = rnp->gpnum;
991 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
992 rdp->passed_quiesce = 0;
993 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
994 zero_cpu_stall_ticks(rdp);
998 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
1000 unsigned long flags;
1001 struct rcu_node *rnp;
1003 local_irq_save(flags);
1005 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1006 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1007 local_irq_restore(flags);
1010 __note_new_gpnum(rsp, rnp, rdp);
1011 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1015 * Did someone else start a new RCU grace period start since we last
1016 * checked? Update local state appropriately if so. Must be called
1017 * on the CPU corresponding to rdp.
1020 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1022 unsigned long flags;
1025 local_irq_save(flags);
1026 if (rdp->gpnum != rsp->gpnum) {
1027 note_new_gpnum(rsp, rdp);
1030 local_irq_restore(flags);
1035 * Initialize the specified rcu_data structure's callback list to empty.
1037 static void init_callback_list(struct rcu_data *rdp)
1041 if (init_nocb_callback_list(rdp))
1043 rdp->nxtlist = NULL;
1044 for (i = 0; i < RCU_NEXT_SIZE; i++)
1045 rdp->nxttail[i] = &rdp->nxtlist;
1049 * Determine the value that ->completed will have at the end of the
1050 * next subsequent grace period. This is used to tag callbacks so that
1051 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1052 * been dyntick-idle for an extended period with callbacks under the
1053 * influence of RCU_FAST_NO_HZ.
1055 * The caller must hold rnp->lock with interrupts disabled.
1057 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1058 struct rcu_node *rnp)
1061 * If RCU is idle, we just wait for the next grace period.
1062 * But we can only be sure that RCU is idle if we are looking
1063 * at the root rcu_node structure -- otherwise, a new grace
1064 * period might have started, but just not yet gotten around
1065 * to initializing the current non-root rcu_node structure.
1067 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1068 return rnp->completed + 1;
1071 * Otherwise, wait for a possible partial grace period and
1072 * then the subsequent full grace period.
1074 return rnp->completed + 2;
1078 * If there is room, assign a ->completed number to any callbacks on
1079 * this CPU that have not already been assigned. Also accelerate any
1080 * callbacks that were previously assigned a ->completed number that has
1081 * since proven to be too conservative, which can happen if callbacks get
1082 * assigned a ->completed number while RCU is idle, but with reference to
1083 * a non-root rcu_node structure. This function is idempotent, so it does
1084 * not hurt to call it repeatedly.
1086 * The caller must hold rnp->lock with interrupts disabled.
1088 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1089 struct rcu_data *rdp)
1094 /* If the CPU has no callbacks, nothing to do. */
1095 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1099 * Starting from the sublist containing the callbacks most
1100 * recently assigned a ->completed number and working down, find the
1101 * first sublist that is not assignable to an upcoming grace period.
1102 * Such a sublist has something in it (first two tests) and has
1103 * a ->completed number assigned that will complete sooner than
1104 * the ->completed number for newly arrived callbacks (last test).
1106 * The key point is that any later sublist can be assigned the
1107 * same ->completed number as the newly arrived callbacks, which
1108 * means that the callbacks in any of these later sublist can be
1109 * grouped into a single sublist, whether or not they have already
1110 * been assigned a ->completed number.
1112 c = rcu_cbs_completed(rsp, rnp);
1113 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1114 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1115 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1119 * If there are no sublist for unassigned callbacks, leave.
1120 * At the same time, advance "i" one sublist, so that "i" will
1121 * index into the sublist where all the remaining callbacks should
1124 if (++i >= RCU_NEXT_TAIL)
1128 * Assign all subsequent callbacks' ->completed number to the next
1129 * full grace period and group them all in the sublist initially
1132 for (; i <= RCU_NEXT_TAIL; i++) {
1133 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1134 rdp->nxtcompleted[i] = c;
1137 /* Trace depending on how much we were able to accelerate. */
1138 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1139 trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccWaitCB");
1141 trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccReadyCB");
1145 * Move any callbacks whose grace period has completed to the
1146 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1147 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1148 * sublist. This function is idempotent, so it does not hurt to
1149 * invoke it repeatedly. As long as it is not invoked -too- often...
1151 * The caller must hold rnp->lock with interrupts disabled.
1153 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1154 struct rcu_data *rdp)
1158 /* If the CPU has no callbacks, nothing to do. */
1159 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1163 * Find all callbacks whose ->completed numbers indicate that they
1164 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1166 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1167 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1169 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1171 /* Clean up any sublist tail pointers that were misordered above. */
1172 for (j = RCU_WAIT_TAIL; j < i; j++)
1173 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1175 /* Copy down callbacks to fill in empty sublists. */
1176 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1177 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1179 rdp->nxttail[j] = rdp->nxttail[i];
1180 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1183 /* Classify any remaining callbacks. */
1184 rcu_accelerate_cbs(rsp, rnp, rdp);
1188 * Advance this CPU's callbacks, but only if the current grace period
1189 * has ended. This may be called only from the CPU to whom the rdp
1190 * belongs. In addition, the corresponding leaf rcu_node structure's
1191 * ->lock must be held by the caller, with irqs disabled.
1194 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1196 /* Did another grace period end? */
1197 if (rdp->completed == rnp->completed) {
1199 /* No, so just accelerate recent callbacks. */
1200 rcu_accelerate_cbs(rsp, rnp, rdp);
1204 /* Advance callbacks. */
1205 rcu_advance_cbs(rsp, rnp, rdp);
1207 /* Remember that we saw this grace-period completion. */
1208 rdp->completed = rnp->completed;
1209 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1212 * If we were in an extended quiescent state, we may have
1213 * missed some grace periods that others CPUs handled on
1214 * our behalf. Catch up with this state to avoid noting
1215 * spurious new grace periods. If another grace period
1216 * has started, then rnp->gpnum will have advanced, so
1217 * we will detect this later on. Of course, any quiescent
1218 * states we found for the old GP are now invalid.
1220 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1221 rdp->gpnum = rdp->completed;
1222 rdp->passed_quiesce = 0;
1226 * If RCU does not need a quiescent state from this CPU,
1227 * then make sure that this CPU doesn't go looking for one.
1229 if ((rnp->qsmask & rdp->grpmask) == 0)
1230 rdp->qs_pending = 0;
1235 * Advance this CPU's callbacks, but only if the current grace period
1236 * has ended. This may be called only from the CPU to whom the rdp
1240 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1242 unsigned long flags;
1243 struct rcu_node *rnp;
1245 local_irq_save(flags);
1247 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1248 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1249 local_irq_restore(flags);
1252 __rcu_process_gp_end(rsp, rnp, rdp);
1253 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1257 * Do per-CPU grace-period initialization for running CPU. The caller
1258 * must hold the lock of the leaf rcu_node structure corresponding to
1262 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1264 /* Prior grace period ended, so advance callbacks for current CPU. */
1265 __rcu_process_gp_end(rsp, rnp, rdp);
1267 /* Set state so that this CPU will detect the next quiescent state. */
1268 __note_new_gpnum(rsp, rnp, rdp);
1272 * Initialize a new grace period.
1274 static int rcu_gp_init(struct rcu_state *rsp)
1276 struct rcu_data *rdp;
1277 struct rcu_node *rnp = rcu_get_root(rsp);
1279 raw_spin_lock_irq(&rnp->lock);
1280 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1282 if (rcu_gp_in_progress(rsp)) {
1283 /* Grace period already in progress, don't start another. */
1284 raw_spin_unlock_irq(&rnp->lock);
1288 /* Advance to a new grace period and initialize state. */
1290 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1291 record_gp_stall_check_time(rsp);
1292 raw_spin_unlock_irq(&rnp->lock);
1294 /* Exclude any concurrent CPU-hotplug operations. */
1295 mutex_lock(&rsp->onoff_mutex);
1298 * Set the quiescent-state-needed bits in all the rcu_node
1299 * structures for all currently online CPUs in breadth-first order,
1300 * starting from the root rcu_node structure, relying on the layout
1301 * of the tree within the rsp->node[] array. Note that other CPUs
1302 * will access only the leaves of the hierarchy, thus seeing that no
1303 * grace period is in progress, at least until the corresponding
1304 * leaf node has been initialized. In addition, we have excluded
1305 * CPU-hotplug operations.
1307 * The grace period cannot complete until the initialization
1308 * process finishes, because this kthread handles both.
1310 rcu_for_each_node_breadth_first(rsp, rnp) {
1311 raw_spin_lock_irq(&rnp->lock);
1312 rdp = this_cpu_ptr(rsp->rda);
1313 rcu_preempt_check_blocked_tasks(rnp);
1314 rnp->qsmask = rnp->qsmaskinit;
1315 rnp->gpnum = rsp->gpnum;
1316 WARN_ON_ONCE(rnp->completed != rsp->completed);
1317 rnp->completed = rsp->completed;
1318 if (rnp == rdp->mynode)
1319 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1320 rcu_preempt_boost_start_gp(rnp);
1321 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1322 rnp->level, rnp->grplo,
1323 rnp->grphi, rnp->qsmask);
1324 raw_spin_unlock_irq(&rnp->lock);
1325 #ifdef CONFIG_PROVE_RCU_DELAY
1326 if ((random32() % (rcu_num_nodes * 8)) == 0)
1327 schedule_timeout_uninterruptible(2);
1328 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1332 mutex_unlock(&rsp->onoff_mutex);
1337 * Do one round of quiescent-state forcing.
1339 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1341 int fqs_state = fqs_state_in;
1342 struct rcu_node *rnp = rcu_get_root(rsp);
1345 if (fqs_state == RCU_SAVE_DYNTICK) {
1346 /* Collect dyntick-idle snapshots. */
1347 force_qs_rnp(rsp, dyntick_save_progress_counter);
1348 fqs_state = RCU_FORCE_QS;
1350 /* Handle dyntick-idle and offline CPUs. */
1351 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1353 /* Clear flag to prevent immediate re-entry. */
1354 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1355 raw_spin_lock_irq(&rnp->lock);
1356 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1357 raw_spin_unlock_irq(&rnp->lock);
1363 * Clean up after the old grace period.
1365 static void rcu_gp_cleanup(struct rcu_state *rsp)
1367 unsigned long gp_duration;
1369 struct rcu_data *rdp;
1370 struct rcu_node *rnp = rcu_get_root(rsp);
1372 raw_spin_lock_irq(&rnp->lock);
1373 gp_duration = jiffies - rsp->gp_start;
1374 if (gp_duration > rsp->gp_max)
1375 rsp->gp_max = gp_duration;
1378 * We know the grace period is complete, but to everyone else
1379 * it appears to still be ongoing. But it is also the case
1380 * that to everyone else it looks like there is nothing that
1381 * they can do to advance the grace period. It is therefore
1382 * safe for us to drop the lock in order to mark the grace
1383 * period as completed in all of the rcu_node structures.
1385 raw_spin_unlock_irq(&rnp->lock);
1388 * Propagate new ->completed value to rcu_node structures so
1389 * that other CPUs don't have to wait until the start of the next
1390 * grace period to process their callbacks. This also avoids
1391 * some nasty RCU grace-period initialization races by forcing
1392 * the end of the current grace period to be completely recorded in
1393 * all of the rcu_node structures before the beginning of the next
1394 * grace period is recorded in any of the rcu_node structures.
1396 rcu_for_each_node_breadth_first(rsp, rnp) {
1397 raw_spin_lock_irq(&rnp->lock);
1398 rnp->completed = rsp->gpnum;
1399 rdp = this_cpu_ptr(rsp->rda);
1400 if (rnp == rdp->mynode)
1401 __rcu_process_gp_end(rsp, rnp, rdp);
1402 nocb += rcu_nocb_gp_cleanup(rsp, rnp);
1403 raw_spin_unlock_irq(&rnp->lock);
1406 rnp = rcu_get_root(rsp);
1407 raw_spin_lock_irq(&rnp->lock);
1408 rcu_nocb_gp_set(rnp, nocb);
1410 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1411 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1412 rsp->fqs_state = RCU_GP_IDLE;
1413 rdp = this_cpu_ptr(rsp->rda);
1414 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1415 if (cpu_needs_another_gp(rsp, rdp))
1417 raw_spin_unlock_irq(&rnp->lock);
1421 * Body of kthread that handles grace periods.
1423 static int __noreturn rcu_gp_kthread(void *arg)
1428 struct rcu_state *rsp = arg;
1429 struct rcu_node *rnp = rcu_get_root(rsp);
1433 /* Handle grace-period start. */
1435 wait_event_interruptible(rsp->gp_wq,
1438 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1442 flush_signals(current);
1445 /* Handle quiescent-state forcing. */
1446 fqs_state = RCU_SAVE_DYNTICK;
1447 j = jiffies_till_first_fqs;
1450 jiffies_till_first_fqs = HZ;
1453 rsp->jiffies_force_qs = jiffies + j;
1454 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1455 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1456 (!ACCESS_ONCE(rnp->qsmask) &&
1457 !rcu_preempt_blocked_readers_cgp(rnp)),
1459 /* If grace period done, leave loop. */
1460 if (!ACCESS_ONCE(rnp->qsmask) &&
1461 !rcu_preempt_blocked_readers_cgp(rnp))
1463 /* If time for quiescent-state forcing, do it. */
1464 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1465 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1468 /* Deal with stray signal. */
1470 flush_signals(current);
1472 j = jiffies_till_next_fqs;
1475 jiffies_till_next_fqs = HZ;
1478 jiffies_till_next_fqs = 1;
1482 /* Handle grace-period end. */
1483 rcu_gp_cleanup(rsp);
1488 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1489 * in preparation for detecting the next grace period. The caller must hold
1490 * the root node's ->lock, which is released before return. Hard irqs must
1493 * Note that it is legal for a dying CPU (which is marked as offline) to
1494 * invoke this function. This can happen when the dying CPU reports its
1498 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1499 __releases(rcu_get_root(rsp)->lock)
1501 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1502 struct rcu_node *rnp = rcu_get_root(rsp);
1505 * If there is no grace period in progress right now, any
1506 * callbacks we have up to this point will be satisfied by the
1507 * next grace period. Also, advancing the callbacks reduces the
1508 * probability of false positives from cpu_needs_another_gp()
1509 * resulting in pointless grace periods. So, advance callbacks!
1511 rcu_advance_cbs(rsp, rnp, rdp);
1513 if (!rsp->gp_kthread ||
1514 !cpu_needs_another_gp(rsp, rdp)) {
1516 * Either we have not yet spawned the grace-period
1517 * task, this CPU does not need another grace period,
1518 * or a grace period is already in progress.
1519 * Either way, don't start a new grace period.
1521 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1525 rsp->gp_flags = RCU_GP_FLAG_INIT;
1526 raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
1528 /* Ensure that CPU is aware of completion of last grace period. */
1529 rcu_process_gp_end(rsp, rdp);
1530 local_irq_restore(flags);
1532 /* Wake up rcu_gp_kthread() to start the grace period. */
1533 wake_up(&rsp->gp_wq);
1537 * Report a full set of quiescent states to the specified rcu_state
1538 * data structure. This involves cleaning up after the prior grace
1539 * period and letting rcu_start_gp() start up the next grace period
1540 * if one is needed. Note that the caller must hold rnp->lock, as
1541 * required by rcu_start_gp(), which will release it.
1543 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1544 __releases(rcu_get_root(rsp)->lock)
1546 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1547 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1548 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1552 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1553 * Allows quiescent states for a group of CPUs to be reported at one go
1554 * to the specified rcu_node structure, though all the CPUs in the group
1555 * must be represented by the same rcu_node structure (which need not be
1556 * a leaf rcu_node structure, though it often will be). That structure's
1557 * lock must be held upon entry, and it is released before return.
1560 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1561 struct rcu_node *rnp, unsigned long flags)
1562 __releases(rnp->lock)
1564 struct rcu_node *rnp_c;
1566 /* Walk up the rcu_node hierarchy. */
1568 if (!(rnp->qsmask & mask)) {
1570 /* Our bit has already been cleared, so done. */
1571 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1574 rnp->qsmask &= ~mask;
1575 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1576 mask, rnp->qsmask, rnp->level,
1577 rnp->grplo, rnp->grphi,
1579 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1581 /* Other bits still set at this level, so done. */
1582 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1585 mask = rnp->grpmask;
1586 if (rnp->parent == NULL) {
1588 /* No more levels. Exit loop holding root lock. */
1592 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1595 raw_spin_lock_irqsave(&rnp->lock, flags);
1596 WARN_ON_ONCE(rnp_c->qsmask);
1600 * Get here if we are the last CPU to pass through a quiescent
1601 * state for this grace period. Invoke rcu_report_qs_rsp()
1602 * to clean up and start the next grace period if one is needed.
1604 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1608 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1609 * structure. This must be either called from the specified CPU, or
1610 * called when the specified CPU is known to be offline (and when it is
1611 * also known that no other CPU is concurrently trying to help the offline
1612 * CPU). The lastcomp argument is used to make sure we are still in the
1613 * grace period of interest. We don't want to end the current grace period
1614 * based on quiescent states detected in an earlier grace period!
1617 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1619 unsigned long flags;
1621 struct rcu_node *rnp;
1624 raw_spin_lock_irqsave(&rnp->lock, flags);
1625 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1626 rnp->completed == rnp->gpnum) {
1629 * The grace period in which this quiescent state was
1630 * recorded has ended, so don't report it upwards.
1631 * We will instead need a new quiescent state that lies
1632 * within the current grace period.
1634 rdp->passed_quiesce = 0; /* need qs for new gp. */
1635 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1638 mask = rdp->grpmask;
1639 if ((rnp->qsmask & mask) == 0) {
1640 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1642 rdp->qs_pending = 0;
1645 * This GP can't end until cpu checks in, so all of our
1646 * callbacks can be processed during the next GP.
1648 rcu_accelerate_cbs(rsp, rnp, rdp);
1650 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1655 * Check to see if there is a new grace period of which this CPU
1656 * is not yet aware, and if so, set up local rcu_data state for it.
1657 * Otherwise, see if this CPU has just passed through its first
1658 * quiescent state for this grace period, and record that fact if so.
1661 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1663 /* If there is now a new grace period, record and return. */
1664 if (check_for_new_grace_period(rsp, rdp))
1668 * Does this CPU still need to do its part for current grace period?
1669 * If no, return and let the other CPUs do their part as well.
1671 if (!rdp->qs_pending)
1675 * Was there a quiescent state since the beginning of the grace
1676 * period? If no, then exit and wait for the next call.
1678 if (!rdp->passed_quiesce)
1682 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1685 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1688 #ifdef CONFIG_HOTPLUG_CPU
1691 * Send the specified CPU's RCU callbacks to the orphanage. The
1692 * specified CPU must be offline, and the caller must hold the
1696 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1697 struct rcu_node *rnp, struct rcu_data *rdp)
1699 /* No-CBs CPUs do not have orphanable callbacks. */
1700 if (is_nocb_cpu(rdp->cpu))
1704 * Orphan the callbacks. First adjust the counts. This is safe
1705 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1706 * cannot be running now. Thus no memory barrier is required.
1708 if (rdp->nxtlist != NULL) {
1709 rsp->qlen_lazy += rdp->qlen_lazy;
1710 rsp->qlen += rdp->qlen;
1711 rdp->n_cbs_orphaned += rdp->qlen;
1713 ACCESS_ONCE(rdp->qlen) = 0;
1717 * Next, move those callbacks still needing a grace period to
1718 * the orphanage, where some other CPU will pick them up.
1719 * Some of the callbacks might have gone partway through a grace
1720 * period, but that is too bad. They get to start over because we
1721 * cannot assume that grace periods are synchronized across CPUs.
1722 * We don't bother updating the ->nxttail[] array yet, instead
1723 * we just reset the whole thing later on.
1725 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1726 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1727 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1728 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1732 * Then move the ready-to-invoke callbacks to the orphanage,
1733 * where some other CPU will pick them up. These will not be
1734 * required to pass though another grace period: They are done.
1736 if (rdp->nxtlist != NULL) {
1737 *rsp->orphan_donetail = rdp->nxtlist;
1738 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1741 /* Finally, initialize the rcu_data structure's list to empty. */
1742 init_callback_list(rdp);
1746 * Adopt the RCU callbacks from the specified rcu_state structure's
1747 * orphanage. The caller must hold the ->orphan_lock.
1749 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1752 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1754 /* No-CBs CPUs are handled specially. */
1755 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1758 /* Do the accounting first. */
1759 rdp->qlen_lazy += rsp->qlen_lazy;
1760 rdp->qlen += rsp->qlen;
1761 rdp->n_cbs_adopted += rsp->qlen;
1762 if (rsp->qlen_lazy != rsp->qlen)
1763 rcu_idle_count_callbacks_posted();
1768 * We do not need a memory barrier here because the only way we
1769 * can get here if there is an rcu_barrier() in flight is if
1770 * we are the task doing the rcu_barrier().
1773 /* First adopt the ready-to-invoke callbacks. */
1774 if (rsp->orphan_donelist != NULL) {
1775 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1776 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1777 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1778 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1779 rdp->nxttail[i] = rsp->orphan_donetail;
1780 rsp->orphan_donelist = NULL;
1781 rsp->orphan_donetail = &rsp->orphan_donelist;
1784 /* And then adopt the callbacks that still need a grace period. */
1785 if (rsp->orphan_nxtlist != NULL) {
1786 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1787 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1788 rsp->orphan_nxtlist = NULL;
1789 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1794 * Trace the fact that this CPU is going offline.
1796 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1798 RCU_TRACE(unsigned long mask);
1799 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1800 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1802 RCU_TRACE(mask = rdp->grpmask);
1803 trace_rcu_grace_period(rsp->name,
1804 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1809 * The CPU has been completely removed, and some other CPU is reporting
1810 * this fact from process context. Do the remainder of the cleanup,
1811 * including orphaning the outgoing CPU's RCU callbacks, and also
1812 * adopting them. There can only be one CPU hotplug operation at a time,
1813 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1815 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1817 unsigned long flags;
1819 int need_report = 0;
1820 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1821 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1823 /* Adjust any no-longer-needed kthreads. */
1824 rcu_boost_kthread_setaffinity(rnp, -1);
1826 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1828 /* Exclude any attempts to start a new grace period. */
1829 mutex_lock(&rsp->onoff_mutex);
1830 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1832 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1833 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1834 rcu_adopt_orphan_cbs(rsp);
1836 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1837 mask = rdp->grpmask; /* rnp->grplo is constant. */
1839 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1840 rnp->qsmaskinit &= ~mask;
1841 if (rnp->qsmaskinit != 0) {
1842 if (rnp != rdp->mynode)
1843 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1846 if (rnp == rdp->mynode)
1847 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1849 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1850 mask = rnp->grpmask;
1852 } while (rnp != NULL);
1855 * We still hold the leaf rcu_node structure lock here, and
1856 * irqs are still disabled. The reason for this subterfuge is
1857 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1858 * held leads to deadlock.
1860 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1862 if (need_report & RCU_OFL_TASKS_NORM_GP)
1863 rcu_report_unblock_qs_rnp(rnp, flags);
1865 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1866 if (need_report & RCU_OFL_TASKS_EXP_GP)
1867 rcu_report_exp_rnp(rsp, rnp, true);
1868 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1869 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1870 cpu, rdp->qlen, rdp->nxtlist);
1871 init_callback_list(rdp);
1872 /* Disallow further callbacks on this CPU. */
1873 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1874 mutex_unlock(&rsp->onoff_mutex);
1877 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1879 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1883 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1887 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1890 * Invoke any RCU callbacks that have made it to the end of their grace
1891 * period. Thottle as specified by rdp->blimit.
1893 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1895 unsigned long flags;
1896 struct rcu_head *next, *list, **tail;
1897 long bl, count, count_lazy;
1900 /* If no callbacks are ready, just return. */
1901 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1902 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1903 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1904 need_resched(), is_idle_task(current),
1905 rcu_is_callbacks_kthread());
1910 * Extract the list of ready callbacks, disabling to prevent
1911 * races with call_rcu() from interrupt handlers.
1913 local_irq_save(flags);
1914 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1916 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1917 list = rdp->nxtlist;
1918 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1919 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1920 tail = rdp->nxttail[RCU_DONE_TAIL];
1921 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1922 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1923 rdp->nxttail[i] = &rdp->nxtlist;
1924 local_irq_restore(flags);
1926 /* Invoke callbacks. */
1927 count = count_lazy = 0;
1931 debug_rcu_head_unqueue(list);
1932 if (__rcu_reclaim(rsp->name, list))
1935 /* Stop only if limit reached and CPU has something to do. */
1936 if (++count >= bl &&
1938 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1942 local_irq_save(flags);
1943 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1944 is_idle_task(current),
1945 rcu_is_callbacks_kthread());
1947 /* Update count, and requeue any remaining callbacks. */
1949 *tail = rdp->nxtlist;
1950 rdp->nxtlist = list;
1951 for (i = 0; i < RCU_NEXT_SIZE; i++)
1952 if (&rdp->nxtlist == rdp->nxttail[i])
1953 rdp->nxttail[i] = tail;
1957 smp_mb(); /* List handling before counting for rcu_barrier(). */
1958 rdp->qlen_lazy -= count_lazy;
1959 ACCESS_ONCE(rdp->qlen) -= count;
1960 rdp->n_cbs_invoked += count;
1962 /* Reinstate batch limit if we have worked down the excess. */
1963 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1964 rdp->blimit = blimit;
1966 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1967 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1968 rdp->qlen_last_fqs_check = 0;
1969 rdp->n_force_qs_snap = rsp->n_force_qs;
1970 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1971 rdp->qlen_last_fqs_check = rdp->qlen;
1972 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1974 local_irq_restore(flags);
1976 /* Re-invoke RCU core processing if there are callbacks remaining. */
1977 if (cpu_has_callbacks_ready_to_invoke(rdp))
1982 * Check to see if this CPU is in a non-context-switch quiescent state
1983 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1984 * Also schedule RCU core processing.
1986 * This function must be called from hardirq context. It is normally
1987 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1988 * false, there is no point in invoking rcu_check_callbacks().
1990 void rcu_check_callbacks(int cpu, int user)
1992 trace_rcu_utilization("Start scheduler-tick");
1993 increment_cpu_stall_ticks();
1994 if (user || rcu_is_cpu_rrupt_from_idle()) {
1997 * Get here if this CPU took its interrupt from user
1998 * mode or from the idle loop, and if this is not a
1999 * nested interrupt. In this case, the CPU is in
2000 * a quiescent state, so note it.
2002 * No memory barrier is required here because both
2003 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2004 * variables that other CPUs neither access nor modify,
2005 * at least not while the corresponding CPU is online.
2011 } else if (!in_softirq()) {
2014 * Get here if this CPU did not take its interrupt from
2015 * softirq, in other words, if it is not interrupting
2016 * a rcu_bh read-side critical section. This is an _bh
2017 * critical section, so note it.
2022 rcu_preempt_check_callbacks(cpu);
2023 if (rcu_pending(cpu))
2025 trace_rcu_utilization("End scheduler-tick");
2029 * Scan the leaf rcu_node structures, processing dyntick state for any that
2030 * have not yet encountered a quiescent state, using the function specified.
2031 * Also initiate boosting for any threads blocked on the root rcu_node.
2033 * The caller must have suppressed start of new grace periods.
2035 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
2039 unsigned long flags;
2041 struct rcu_node *rnp;
2043 rcu_for_each_leaf_node(rsp, rnp) {
2046 raw_spin_lock_irqsave(&rnp->lock, flags);
2047 if (!rcu_gp_in_progress(rsp)) {
2048 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2051 if (rnp->qsmask == 0) {
2052 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2057 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2058 if ((rnp->qsmask & bit) != 0 &&
2059 f(per_cpu_ptr(rsp->rda, cpu)))
2064 /* rcu_report_qs_rnp() releases rnp->lock. */
2065 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2068 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2070 rnp = rcu_get_root(rsp);
2071 if (rnp->qsmask == 0) {
2072 raw_spin_lock_irqsave(&rnp->lock, flags);
2073 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2078 * Force quiescent states on reluctant CPUs, and also detect which
2079 * CPUs are in dyntick-idle mode.
2081 static void force_quiescent_state(struct rcu_state *rsp)
2083 unsigned long flags;
2085 struct rcu_node *rnp;
2086 struct rcu_node *rnp_old = NULL;
2088 /* Funnel through hierarchy to reduce memory contention. */
2089 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2090 for (; rnp != NULL; rnp = rnp->parent) {
2091 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2092 !raw_spin_trylock(&rnp->fqslock);
2093 if (rnp_old != NULL)
2094 raw_spin_unlock(&rnp_old->fqslock);
2096 rsp->n_force_qs_lh++;
2101 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2103 /* Reached the root of the rcu_node tree, acquire lock. */
2104 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2105 raw_spin_unlock(&rnp_old->fqslock);
2106 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2107 rsp->n_force_qs_lh++;
2108 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2109 return; /* Someone beat us to it. */
2111 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2112 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2113 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2117 * This does the RCU core processing work for the specified rcu_state
2118 * and rcu_data structures. This may be called only from the CPU to
2119 * whom the rdp belongs.
2122 __rcu_process_callbacks(struct rcu_state *rsp)
2124 unsigned long flags;
2125 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2127 WARN_ON_ONCE(rdp->beenonline == 0);
2129 /* Handle the end of a grace period that some other CPU ended. */
2130 rcu_process_gp_end(rsp, rdp);
2132 /* Update RCU state based on any recent quiescent states. */
2133 rcu_check_quiescent_state(rsp, rdp);
2135 /* Does this CPU require a not-yet-started grace period? */
2136 local_irq_save(flags);
2137 if (cpu_needs_another_gp(rsp, rdp)) {
2138 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2139 rcu_start_gp(rsp, flags); /* releases above lock */
2141 local_irq_restore(flags);
2144 /* If there are callbacks ready, invoke them. */
2145 if (cpu_has_callbacks_ready_to_invoke(rdp))
2146 invoke_rcu_callbacks(rsp, rdp);
2150 * Do RCU core processing for the current CPU.
2152 static void rcu_process_callbacks(struct softirq_action *unused)
2154 struct rcu_state *rsp;
2156 if (cpu_is_offline(smp_processor_id()))
2158 trace_rcu_utilization("Start RCU core");
2159 for_each_rcu_flavor(rsp)
2160 __rcu_process_callbacks(rsp);
2161 trace_rcu_utilization("End RCU core");
2165 * Schedule RCU callback invocation. If the specified type of RCU
2166 * does not support RCU priority boosting, just do a direct call,
2167 * otherwise wake up the per-CPU kernel kthread. Note that because we
2168 * are running on the current CPU with interrupts disabled, the
2169 * rcu_cpu_kthread_task cannot disappear out from under us.
2171 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2173 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2175 if (likely(!rsp->boost)) {
2176 rcu_do_batch(rsp, rdp);
2179 invoke_rcu_callbacks_kthread();
2182 static void invoke_rcu_core(void)
2184 raise_softirq(RCU_SOFTIRQ);
2188 * Handle any core-RCU processing required by a call_rcu() invocation.
2190 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2191 struct rcu_head *head, unsigned long flags)
2194 * If called from an extended quiescent state, invoke the RCU
2195 * core in order to force a re-evaluation of RCU's idleness.
2197 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2200 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2201 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2205 * Force the grace period if too many callbacks or too long waiting.
2206 * Enforce hysteresis, and don't invoke force_quiescent_state()
2207 * if some other CPU has recently done so. Also, don't bother
2208 * invoking force_quiescent_state() if the newly enqueued callback
2209 * is the only one waiting for a grace period to complete.
2211 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2213 /* Are we ignoring a completed grace period? */
2214 rcu_process_gp_end(rsp, rdp);
2215 check_for_new_grace_period(rsp, rdp);
2217 /* Start a new grace period if one not already started. */
2218 if (!rcu_gp_in_progress(rsp)) {
2219 unsigned long nestflag;
2220 struct rcu_node *rnp_root = rcu_get_root(rsp);
2222 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
2223 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
2225 /* Give the grace period a kick. */
2226 rdp->blimit = LONG_MAX;
2227 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2228 *rdp->nxttail[RCU_DONE_TAIL] != head)
2229 force_quiescent_state(rsp);
2230 rdp->n_force_qs_snap = rsp->n_force_qs;
2231 rdp->qlen_last_fqs_check = rdp->qlen;
2237 * Helper function for call_rcu() and friends. The cpu argument will
2238 * normally be -1, indicating "currently running CPU". It may specify
2239 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2240 * is expected to specify a CPU.
2243 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2244 struct rcu_state *rsp, int cpu, bool lazy)
2246 unsigned long flags;
2247 struct rcu_data *rdp;
2249 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2250 debug_rcu_head_queue(head);
2255 * Opportunistically note grace-period endings and beginnings.
2256 * Note that we might see a beginning right after we see an
2257 * end, but never vice versa, since this CPU has to pass through
2258 * a quiescent state betweentimes.
2260 local_irq_save(flags);
2261 rdp = this_cpu_ptr(rsp->rda);
2263 /* Add the callback to our list. */
2264 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2268 rdp = per_cpu_ptr(rsp->rda, cpu);
2269 offline = !__call_rcu_nocb(rdp, head, lazy);
2270 WARN_ON_ONCE(offline);
2271 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2272 local_irq_restore(flags);
2275 ACCESS_ONCE(rdp->qlen)++;
2279 rcu_idle_count_callbacks_posted();
2280 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2281 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2282 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2284 if (__is_kfree_rcu_offset((unsigned long)func))
2285 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2286 rdp->qlen_lazy, rdp->qlen);
2288 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2290 /* Go handle any RCU core processing required. */
2291 __call_rcu_core(rsp, rdp, head, flags);
2292 local_irq_restore(flags);
2296 * Queue an RCU-sched callback for invocation after a grace period.
2298 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2300 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2302 EXPORT_SYMBOL_GPL(call_rcu_sched);
2305 * Queue an RCU callback for invocation after a quicker grace period.
2307 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2309 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2311 EXPORT_SYMBOL_GPL(call_rcu_bh);
2314 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2315 * any blocking grace-period wait automatically implies a grace period
2316 * if there is only one CPU online at any point time during execution
2317 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2318 * occasionally incorrectly indicate that there are multiple CPUs online
2319 * when there was in fact only one the whole time, as this just adds
2320 * some overhead: RCU still operates correctly.
2322 static inline int rcu_blocking_is_gp(void)
2326 might_sleep(); /* Check for RCU read-side critical section. */
2328 ret = num_online_cpus() <= 1;
2334 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2336 * Control will return to the caller some time after a full rcu-sched
2337 * grace period has elapsed, in other words after all currently executing
2338 * rcu-sched read-side critical sections have completed. These read-side
2339 * critical sections are delimited by rcu_read_lock_sched() and
2340 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2341 * local_irq_disable(), and so on may be used in place of
2342 * rcu_read_lock_sched().
2344 * This means that all preempt_disable code sequences, including NMI and
2345 * non-threaded hardware-interrupt handlers, in progress on entry will
2346 * have completed before this primitive returns. However, this does not
2347 * guarantee that softirq handlers will have completed, since in some
2348 * kernels, these handlers can run in process context, and can block.
2350 * Note that this guarantee implies further memory-ordering guarantees.
2351 * On systems with more than one CPU, when synchronize_sched() returns,
2352 * each CPU is guaranteed to have executed a full memory barrier since the
2353 * end of its last RCU-sched read-side critical section whose beginning
2354 * preceded the call to synchronize_sched(). In addition, each CPU having
2355 * an RCU read-side critical section that extends beyond the return from
2356 * synchronize_sched() is guaranteed to have executed a full memory barrier
2357 * after the beginning of synchronize_sched() and before the beginning of
2358 * that RCU read-side critical section. Note that these guarantees include
2359 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2360 * that are executing in the kernel.
2362 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2363 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2364 * to have executed a full memory barrier during the execution of
2365 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2366 * again only if the system has more than one CPU).
2368 * This primitive provides the guarantees made by the (now removed)
2369 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2370 * guarantees that rcu_read_lock() sections will have completed.
2371 * In "classic RCU", these two guarantees happen to be one and
2372 * the same, but can differ in realtime RCU implementations.
2374 void synchronize_sched(void)
2376 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2377 !lock_is_held(&rcu_lock_map) &&
2378 !lock_is_held(&rcu_sched_lock_map),
2379 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2380 if (rcu_blocking_is_gp())
2383 synchronize_sched_expedited();
2385 wait_rcu_gp(call_rcu_sched);
2387 EXPORT_SYMBOL_GPL(synchronize_sched);
2390 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2392 * Control will return to the caller some time after a full rcu_bh grace
2393 * period has elapsed, in other words after all currently executing rcu_bh
2394 * read-side critical sections have completed. RCU read-side critical
2395 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2396 * and may be nested.
2398 * See the description of synchronize_sched() for more detailed information
2399 * on memory ordering guarantees.
2401 void synchronize_rcu_bh(void)
2403 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2404 !lock_is_held(&rcu_lock_map) &&
2405 !lock_is_held(&rcu_sched_lock_map),
2406 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2407 if (rcu_blocking_is_gp())
2410 synchronize_rcu_bh_expedited();
2412 wait_rcu_gp(call_rcu_bh);
2414 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2416 static int synchronize_sched_expedited_cpu_stop(void *data)
2419 * There must be a full memory barrier on each affected CPU
2420 * between the time that try_stop_cpus() is called and the
2421 * time that it returns.
2423 * In the current initial implementation of cpu_stop, the
2424 * above condition is already met when the control reaches
2425 * this point and the following smp_mb() is not strictly
2426 * necessary. Do smp_mb() anyway for documentation and
2427 * robustness against future implementation changes.
2429 smp_mb(); /* See above comment block. */
2434 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2436 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2437 * approach to force the grace period to end quickly. This consumes
2438 * significant time on all CPUs and is unfriendly to real-time workloads,
2439 * so is thus not recommended for any sort of common-case code. In fact,
2440 * if you are using synchronize_sched_expedited() in a loop, please
2441 * restructure your code to batch your updates, and then use a single
2442 * synchronize_sched() instead.
2444 * Note that it is illegal to call this function while holding any lock
2445 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2446 * to call this function from a CPU-hotplug notifier. Failing to observe
2447 * these restriction will result in deadlock.
2449 * This implementation can be thought of as an application of ticket
2450 * locking to RCU, with sync_sched_expedited_started and
2451 * sync_sched_expedited_done taking on the roles of the halves
2452 * of the ticket-lock word. Each task atomically increments
2453 * sync_sched_expedited_started upon entry, snapshotting the old value,
2454 * then attempts to stop all the CPUs. If this succeeds, then each
2455 * CPU will have executed a context switch, resulting in an RCU-sched
2456 * grace period. We are then done, so we use atomic_cmpxchg() to
2457 * update sync_sched_expedited_done to match our snapshot -- but
2458 * only if someone else has not already advanced past our snapshot.
2460 * On the other hand, if try_stop_cpus() fails, we check the value
2461 * of sync_sched_expedited_done. If it has advanced past our
2462 * initial snapshot, then someone else must have forced a grace period
2463 * some time after we took our snapshot. In this case, our work is
2464 * done for us, and we can simply return. Otherwise, we try again,
2465 * but keep our initial snapshot for purposes of checking for someone
2466 * doing our work for us.
2468 * If we fail too many times in a row, we fall back to synchronize_sched().
2470 void synchronize_sched_expedited(void)
2472 long firstsnap, s, snap;
2474 struct rcu_state *rsp = &rcu_sched_state;
2477 * If we are in danger of counter wrap, just do synchronize_sched().
2478 * By allowing sync_sched_expedited_started to advance no more than
2479 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2480 * that more than 3.5 billion CPUs would be required to force a
2481 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2482 * course be required on a 64-bit system.
2484 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2485 (ulong)atomic_long_read(&rsp->expedited_done) +
2487 synchronize_sched();
2488 atomic_long_inc(&rsp->expedited_wrap);
2493 * Take a ticket. Note that atomic_inc_return() implies a
2494 * full memory barrier.
2496 snap = atomic_long_inc_return(&rsp->expedited_start);
2499 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2502 * Each pass through the following loop attempts to force a
2503 * context switch on each CPU.
2505 while (try_stop_cpus(cpu_online_mask,
2506 synchronize_sched_expedited_cpu_stop,
2509 atomic_long_inc(&rsp->expedited_tryfail);
2511 /* Check to see if someone else did our work for us. */
2512 s = atomic_long_read(&rsp->expedited_done);
2513 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2514 /* ensure test happens before caller kfree */
2515 smp_mb__before_atomic_inc(); /* ^^^ */
2516 atomic_long_inc(&rsp->expedited_workdone1);
2520 /* No joy, try again later. Or just synchronize_sched(). */
2521 if (trycount++ < 10) {
2522 udelay(trycount * num_online_cpus());
2524 wait_rcu_gp(call_rcu_sched);
2525 atomic_long_inc(&rsp->expedited_normal);
2529 /* Recheck to see if someone else did our work for us. */
2530 s = atomic_long_read(&rsp->expedited_done);
2531 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2532 /* ensure test happens before caller kfree */
2533 smp_mb__before_atomic_inc(); /* ^^^ */
2534 atomic_long_inc(&rsp->expedited_workdone2);
2539 * Refetching sync_sched_expedited_started allows later
2540 * callers to piggyback on our grace period. We retry
2541 * after they started, so our grace period works for them,
2542 * and they started after our first try, so their grace
2543 * period works for us.
2546 snap = atomic_long_read(&rsp->expedited_start);
2547 smp_mb(); /* ensure read is before try_stop_cpus(). */
2549 atomic_long_inc(&rsp->expedited_stoppedcpus);
2552 * Everyone up to our most recent fetch is covered by our grace
2553 * period. Update the counter, but only if our work is still
2554 * relevant -- which it won't be if someone who started later
2555 * than we did already did their update.
2558 atomic_long_inc(&rsp->expedited_done_tries);
2559 s = atomic_long_read(&rsp->expedited_done);
2560 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2561 /* ensure test happens before caller kfree */
2562 smp_mb__before_atomic_inc(); /* ^^^ */
2563 atomic_long_inc(&rsp->expedited_done_lost);
2566 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2567 atomic_long_inc(&rsp->expedited_done_exit);
2571 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2574 * Check to see if there is any immediate RCU-related work to be done
2575 * by the current CPU, for the specified type of RCU, returning 1 if so.
2576 * The checks are in order of increasing expense: checks that can be
2577 * carried out against CPU-local state are performed first. However,
2578 * we must check for CPU stalls first, else we might not get a chance.
2580 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2582 struct rcu_node *rnp = rdp->mynode;
2584 rdp->n_rcu_pending++;
2586 /* Check for CPU stalls, if enabled. */
2587 check_cpu_stall(rsp, rdp);
2589 /* Is the RCU core waiting for a quiescent state from this CPU? */
2590 if (rcu_scheduler_fully_active &&
2591 rdp->qs_pending && !rdp->passed_quiesce) {
2592 rdp->n_rp_qs_pending++;
2593 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2594 rdp->n_rp_report_qs++;
2598 /* Does this CPU have callbacks ready to invoke? */
2599 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2600 rdp->n_rp_cb_ready++;
2604 /* Has RCU gone idle with this CPU needing another grace period? */
2605 if (cpu_needs_another_gp(rsp, rdp)) {
2606 rdp->n_rp_cpu_needs_gp++;
2610 /* Has another RCU grace period completed? */
2611 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2612 rdp->n_rp_gp_completed++;
2616 /* Has a new RCU grace period started? */
2617 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2618 rdp->n_rp_gp_started++;
2623 rdp->n_rp_need_nothing++;
2628 * Check to see if there is any immediate RCU-related work to be done
2629 * by the current CPU, returning 1 if so. This function is part of the
2630 * RCU implementation; it is -not- an exported member of the RCU API.
2632 static int rcu_pending(int cpu)
2634 struct rcu_state *rsp;
2636 for_each_rcu_flavor(rsp)
2637 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2643 * Return true if the specified CPU has any callback. If all_lazy is
2644 * non-NULL, store an indication of whether all callbacks are lazy.
2645 * (If there are no callbacks, all of them are deemed to be lazy.)
2647 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2651 struct rcu_data *rdp;
2652 struct rcu_state *rsp;
2654 for_each_rcu_flavor(rsp) {
2655 rdp = per_cpu_ptr(rsp->rda, cpu);
2656 if (rdp->qlen != rdp->qlen_lazy)
2667 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2668 * the compiler is expected to optimize this away.
2670 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2671 int cpu, unsigned long done)
2673 trace_rcu_barrier(rsp->name, s, cpu,
2674 atomic_read(&rsp->barrier_cpu_count), done);
2678 * RCU callback function for _rcu_barrier(). If we are last, wake
2679 * up the task executing _rcu_barrier().
2681 static void rcu_barrier_callback(struct rcu_head *rhp)
2683 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2684 struct rcu_state *rsp = rdp->rsp;
2686 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2687 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2688 complete(&rsp->barrier_completion);
2690 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2695 * Called with preemption disabled, and from cross-cpu IRQ context.
2697 static void rcu_barrier_func(void *type)
2699 struct rcu_state *rsp = type;
2700 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2702 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2703 atomic_inc(&rsp->barrier_cpu_count);
2704 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2708 * Orchestrate the specified type of RCU barrier, waiting for all
2709 * RCU callbacks of the specified type to complete.
2711 static void _rcu_barrier(struct rcu_state *rsp)
2714 struct rcu_data *rdp;
2715 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2716 unsigned long snap_done;
2718 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2720 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2721 mutex_lock(&rsp->barrier_mutex);
2724 * Ensure that all prior references, including to ->n_barrier_done,
2725 * are ordered before the _rcu_barrier() machinery.
2727 smp_mb(); /* See above block comment. */
2730 * Recheck ->n_barrier_done to see if others did our work for us.
2731 * This means checking ->n_barrier_done for an even-to-odd-to-even
2732 * transition. The "if" expression below therefore rounds the old
2733 * value up to the next even number and adds two before comparing.
2735 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2736 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2737 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2738 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2739 smp_mb(); /* caller's subsequent code after above check. */
2740 mutex_unlock(&rsp->barrier_mutex);
2745 * Increment ->n_barrier_done to avoid duplicate work. Use
2746 * ACCESS_ONCE() to prevent the compiler from speculating
2747 * the increment to precede the early-exit check.
2749 ACCESS_ONCE(rsp->n_barrier_done)++;
2750 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2751 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2752 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2755 * Initialize the count to one rather than to zero in order to
2756 * avoid a too-soon return to zero in case of a short grace period
2757 * (or preemption of this task). Exclude CPU-hotplug operations
2758 * to ensure that no offline CPU has callbacks queued.
2760 init_completion(&rsp->barrier_completion);
2761 atomic_set(&rsp->barrier_cpu_count, 1);
2765 * Force each CPU with callbacks to register a new callback.
2766 * When that callback is invoked, we will know that all of the
2767 * corresponding CPU's preceding callbacks have been invoked.
2769 for_each_possible_cpu(cpu) {
2770 if (!cpu_online(cpu) && !is_nocb_cpu(cpu))
2772 rdp = per_cpu_ptr(rsp->rda, cpu);
2773 if (is_nocb_cpu(cpu)) {
2774 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2775 rsp->n_barrier_done);
2776 atomic_inc(&rsp->barrier_cpu_count);
2777 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2779 } else if (ACCESS_ONCE(rdp->qlen)) {
2780 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2781 rsp->n_barrier_done);
2782 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2784 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2785 rsp->n_barrier_done);
2791 * Now that we have an rcu_barrier_callback() callback on each
2792 * CPU, and thus each counted, remove the initial count.
2794 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2795 complete(&rsp->barrier_completion);
2797 /* Increment ->n_barrier_done to prevent duplicate work. */
2798 smp_mb(); /* Keep increment after above mechanism. */
2799 ACCESS_ONCE(rsp->n_barrier_done)++;
2800 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2801 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2802 smp_mb(); /* Keep increment before caller's subsequent code. */
2804 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2805 wait_for_completion(&rsp->barrier_completion);
2807 /* Other rcu_barrier() invocations can now safely proceed. */
2808 mutex_unlock(&rsp->barrier_mutex);
2812 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2814 void rcu_barrier_bh(void)
2816 _rcu_barrier(&rcu_bh_state);
2818 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2821 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2823 void rcu_barrier_sched(void)
2825 _rcu_barrier(&rcu_sched_state);
2827 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2830 * Do boot-time initialization of a CPU's per-CPU RCU data.
2833 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2835 unsigned long flags;
2836 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2837 struct rcu_node *rnp = rcu_get_root(rsp);
2839 /* Set up local state, ensuring consistent view of global state. */
2840 raw_spin_lock_irqsave(&rnp->lock, flags);
2841 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2842 init_callback_list(rdp);
2844 ACCESS_ONCE(rdp->qlen) = 0;
2845 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2846 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2847 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2850 rcu_boot_init_nocb_percpu_data(rdp);
2851 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2855 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2856 * offline event can be happening at a given time. Note also that we
2857 * can accept some slop in the rsp->completed access due to the fact
2858 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2860 static void __cpuinit
2861 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2863 unsigned long flags;
2865 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2866 struct rcu_node *rnp = rcu_get_root(rsp);
2868 /* Exclude new grace periods. */
2869 mutex_lock(&rsp->onoff_mutex);
2871 /* Set up local state, ensuring consistent view of global state. */
2872 raw_spin_lock_irqsave(&rnp->lock, flags);
2873 rdp->beenonline = 1; /* We have now been online. */
2874 rdp->preemptible = preemptible;
2875 rdp->qlen_last_fqs_check = 0;
2876 rdp->n_force_qs_snap = rsp->n_force_qs;
2877 rdp->blimit = blimit;
2878 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2879 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2880 atomic_set(&rdp->dynticks->dynticks,
2881 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2882 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2884 /* Add CPU to rcu_node bitmasks. */
2886 mask = rdp->grpmask;
2888 /* Exclude any attempts to start a new GP on small systems. */
2889 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2890 rnp->qsmaskinit |= mask;
2891 mask = rnp->grpmask;
2892 if (rnp == rdp->mynode) {
2894 * If there is a grace period in progress, we will
2895 * set up to wait for it next time we run the
2898 rdp->gpnum = rnp->completed;
2899 rdp->completed = rnp->completed;
2900 rdp->passed_quiesce = 0;
2901 rdp->qs_pending = 0;
2902 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2904 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2906 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2907 local_irq_restore(flags);
2909 mutex_unlock(&rsp->onoff_mutex);
2912 static void __cpuinit rcu_prepare_cpu(int cpu)
2914 struct rcu_state *rsp;
2916 for_each_rcu_flavor(rsp)
2917 rcu_init_percpu_data(cpu, rsp,
2918 strcmp(rsp->name, "rcu_preempt") == 0);
2922 * Handle CPU online/offline notification events.
2924 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2925 unsigned long action, void *hcpu)
2927 long cpu = (long)hcpu;
2928 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2929 struct rcu_node *rnp = rdp->mynode;
2930 struct rcu_state *rsp;
2932 trace_rcu_utilization("Start CPU hotplug");
2934 case CPU_UP_PREPARE:
2935 case CPU_UP_PREPARE_FROZEN:
2936 rcu_prepare_cpu(cpu);
2937 rcu_prepare_kthreads(cpu);
2940 case CPU_DOWN_FAILED:
2941 rcu_boost_kthread_setaffinity(rnp, -1);
2943 case CPU_DOWN_PREPARE:
2944 rcu_boost_kthread_setaffinity(rnp, cpu);
2947 case CPU_DYING_FROZEN:
2949 * The whole machine is "stopped" except this CPU, so we can
2950 * touch any data without introducing corruption. We send the
2951 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2953 for_each_rcu_flavor(rsp)
2954 rcu_cleanup_dying_cpu(rsp);
2957 case CPU_DEAD_FROZEN:
2958 case CPU_UP_CANCELED:
2959 case CPU_UP_CANCELED_FROZEN:
2960 for_each_rcu_flavor(rsp)
2961 rcu_cleanup_dead_cpu(cpu, rsp);
2966 trace_rcu_utilization("End CPU hotplug");
2971 * Spawn the kthread that handles this RCU flavor's grace periods.
2973 static int __init rcu_spawn_gp_kthread(void)
2975 unsigned long flags;
2976 struct rcu_node *rnp;
2977 struct rcu_state *rsp;
2978 struct task_struct *t;
2980 for_each_rcu_flavor(rsp) {
2981 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2983 rnp = rcu_get_root(rsp);
2984 raw_spin_lock_irqsave(&rnp->lock, flags);
2985 rsp->gp_kthread = t;
2986 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2987 rcu_spawn_nocb_kthreads(rsp);
2991 early_initcall(rcu_spawn_gp_kthread);
2994 * This function is invoked towards the end of the scheduler's initialization
2995 * process. Before this is called, the idle task might contain
2996 * RCU read-side critical sections (during which time, this idle
2997 * task is booting the system). After this function is called, the
2998 * idle tasks are prohibited from containing RCU read-side critical
2999 * sections. This function also enables RCU lockdep checking.
3001 void rcu_scheduler_starting(void)
3003 WARN_ON(num_online_cpus() != 1);
3004 WARN_ON(nr_context_switches() > 0);
3005 rcu_scheduler_active = 1;
3009 * Compute the per-level fanout, either using the exact fanout specified
3010 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3012 #ifdef CONFIG_RCU_FANOUT_EXACT
3013 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3017 for (i = rcu_num_lvls - 1; i > 0; i--)
3018 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3019 rsp->levelspread[0] = rcu_fanout_leaf;
3021 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3022 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3029 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3030 ccur = rsp->levelcnt[i];
3031 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3035 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3038 * Helper function for rcu_init() that initializes one rcu_state structure.
3040 static void __init rcu_init_one(struct rcu_state *rsp,
3041 struct rcu_data __percpu *rda)
3043 static char *buf[] = { "rcu_node_0",
3046 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3047 static char *fqs[] = { "rcu_node_fqs_0",
3050 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3054 struct rcu_node *rnp;
3056 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3058 /* Silence gcc 4.8 warning about array index out of range. */
3059 if (rcu_num_lvls > RCU_NUM_LVLS)
3060 panic("rcu_init_one: rcu_num_lvls overflow");
3062 /* Initialize the level-tracking arrays. */
3064 for (i = 0; i < rcu_num_lvls; i++)
3065 rsp->levelcnt[i] = num_rcu_lvl[i];
3066 for (i = 1; i < rcu_num_lvls; i++)
3067 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3068 rcu_init_levelspread(rsp);
3070 /* Initialize the elements themselves, starting from the leaves. */
3072 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3073 cpustride *= rsp->levelspread[i];
3074 rnp = rsp->level[i];
3075 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3076 raw_spin_lock_init(&rnp->lock);
3077 lockdep_set_class_and_name(&rnp->lock,
3078 &rcu_node_class[i], buf[i]);
3079 raw_spin_lock_init(&rnp->fqslock);
3080 lockdep_set_class_and_name(&rnp->fqslock,
3081 &rcu_fqs_class[i], fqs[i]);
3082 rnp->gpnum = rsp->gpnum;
3083 rnp->completed = rsp->completed;
3085 rnp->qsmaskinit = 0;
3086 rnp->grplo = j * cpustride;
3087 rnp->grphi = (j + 1) * cpustride - 1;
3088 if (rnp->grphi >= NR_CPUS)
3089 rnp->grphi = NR_CPUS - 1;
3095 rnp->grpnum = j % rsp->levelspread[i - 1];
3096 rnp->grpmask = 1UL << rnp->grpnum;
3097 rnp->parent = rsp->level[i - 1] +
3098 j / rsp->levelspread[i - 1];
3101 INIT_LIST_HEAD(&rnp->blkd_tasks);
3102 rcu_init_one_nocb(rnp);
3107 init_waitqueue_head(&rsp->gp_wq);
3108 rnp = rsp->level[rcu_num_lvls - 1];
3109 for_each_possible_cpu(i) {
3110 while (i > rnp->grphi)
3112 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3113 rcu_boot_init_percpu_data(i, rsp);
3115 list_add(&rsp->flavors, &rcu_struct_flavors);
3119 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3120 * replace the definitions in rcutree.h because those are needed to size
3121 * the ->node array in the rcu_state structure.
3123 static void __init rcu_init_geometry(void)
3128 int rcu_capacity[MAX_RCU_LVLS + 1];
3130 /* If the compile-time values are accurate, just leave. */
3131 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3132 nr_cpu_ids == NR_CPUS)
3136 * Compute number of nodes that can be handled an rcu_node tree
3137 * with the given number of levels. Setting rcu_capacity[0] makes
3138 * some of the arithmetic easier.
3140 rcu_capacity[0] = 1;
3141 rcu_capacity[1] = rcu_fanout_leaf;
3142 for (i = 2; i <= MAX_RCU_LVLS; i++)
3143 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3146 * The boot-time rcu_fanout_leaf parameter is only permitted
3147 * to increase the leaf-level fanout, not decrease it. Of course,
3148 * the leaf-level fanout cannot exceed the number of bits in
3149 * the rcu_node masks. Finally, the tree must be able to accommodate
3150 * the configured number of CPUs. Complain and fall back to the
3151 * compile-time values if these limits are exceeded.
3153 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3154 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3155 n > rcu_capacity[MAX_RCU_LVLS]) {
3160 /* Calculate the number of rcu_nodes at each level of the tree. */
3161 for (i = 1; i <= MAX_RCU_LVLS; i++)
3162 if (n <= rcu_capacity[i]) {
3163 for (j = 0; j <= i; j++)
3165 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3167 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3172 /* Calculate the total number of rcu_node structures. */
3174 for (i = 0; i <= MAX_RCU_LVLS; i++)
3175 rcu_num_nodes += num_rcu_lvl[i];
3179 void __init rcu_init(void)
3183 rcu_bootup_announce();
3184 rcu_init_geometry();
3185 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3186 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3187 __rcu_init_preempt();
3188 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3191 * We don't need protection against CPU-hotplug here because
3192 * this is called early in boot, before either interrupts
3193 * or the scheduler are operational.
3195 cpu_notifier(rcu_cpu_notify, 0);
3196 for_each_online_cpu(cpu)
3197 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3200 #include "rcutree_plugin.h"