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, 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), \
81 struct rcu_state rcu_sched_state =
82 RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
83 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
85 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
86 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
88 static struct rcu_state *rcu_state;
89 LIST_HEAD(rcu_struct_flavors);
91 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
92 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
93 module_param(rcu_fanout_leaf, int, 0444);
94 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
95 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
102 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
105 * The rcu_scheduler_active variable transitions from zero to one just
106 * before the first task is spawned. So when this variable is zero, RCU
107 * can assume that there is but one task, allowing RCU to (for example)
108 * optimize synchronize_sched() to a simple barrier(). When this variable
109 * is one, RCU must actually do all the hard work required to detect real
110 * grace periods. This variable is also used to suppress boot-time false
111 * positives from lockdep-RCU error checking.
113 int rcu_scheduler_active __read_mostly;
114 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
117 * The rcu_scheduler_fully_active variable transitions from zero to one
118 * during the early_initcall() processing, which is after the scheduler
119 * is capable of creating new tasks. So RCU processing (for example,
120 * creating tasks for RCU priority boosting) must be delayed until after
121 * rcu_scheduler_fully_active transitions from zero to one. We also
122 * currently delay invocation of any RCU callbacks until after this point.
124 * It might later prove better for people registering RCU callbacks during
125 * early boot to take responsibility for these callbacks, but one step at
128 static int rcu_scheduler_fully_active __read_mostly;
130 #ifdef CONFIG_RCU_BOOST
133 * Control variables for per-CPU and per-rcu_node kthreads. These
134 * handle all flavors of RCU.
136 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
139 DEFINE_PER_CPU(char, rcu_cpu_has_work);
141 #endif /* #ifdef CONFIG_RCU_BOOST */
143 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
144 static void invoke_rcu_core(void);
145 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
148 * Track the rcutorture test sequence number and the update version
149 * number within a given test. The rcutorture_testseq is incremented
150 * on every rcutorture module load and unload, so has an odd value
151 * when a test is running. The rcutorture_vernum is set to zero
152 * when rcutorture starts and is incremented on each rcutorture update.
153 * These variables enable correlating rcutorture output with the
154 * RCU tracing information.
156 unsigned long rcutorture_testseq;
157 unsigned long rcutorture_vernum;
160 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
161 * permit this function to be invoked without holding the root rcu_node
162 * structure's ->lock, but of course results can be subject to change.
164 static int rcu_gp_in_progress(struct rcu_state *rsp)
166 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
170 * Note a quiescent state. Because we do not need to know
171 * how many quiescent states passed, just if there was at least
172 * one since the start of the grace period, this just sets a flag.
173 * The caller must have disabled preemption.
175 void rcu_sched_qs(int cpu)
177 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
179 if (rdp->passed_quiesce == 0)
180 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
181 rdp->passed_quiesce = 1;
184 void rcu_bh_qs(int cpu)
186 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
188 if (rdp->passed_quiesce == 0)
189 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
190 rdp->passed_quiesce = 1;
194 * Note a context switch. This is a quiescent state for RCU-sched,
195 * and requires special handling for preemptible RCU.
196 * The caller must have disabled preemption.
198 void rcu_note_context_switch(int cpu)
200 trace_rcu_utilization("Start context switch");
202 rcu_preempt_note_context_switch(cpu);
203 trace_rcu_utilization("End context switch");
205 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
207 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
208 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
209 .dynticks = ATOMIC_INIT(1),
212 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
213 static long qhimark = 10000; /* If this many pending, ignore blimit. */
214 static long qlowmark = 100; /* Once only this many pending, use blimit. */
216 module_param(blimit, long, 0444);
217 module_param(qhimark, long, 0444);
218 module_param(qlowmark, long, 0444);
220 static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
221 static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
223 module_param(jiffies_till_first_fqs, ulong, 0644);
224 module_param(jiffies_till_next_fqs, ulong, 0644);
226 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
227 static void force_quiescent_state(struct rcu_state *rsp);
228 static int rcu_pending(int cpu);
231 * Return the number of RCU-sched batches processed thus far for debug & stats.
233 long rcu_batches_completed_sched(void)
235 return rcu_sched_state.completed;
237 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
240 * Return the number of RCU BH batches processed thus far for debug & stats.
242 long rcu_batches_completed_bh(void)
244 return rcu_bh_state.completed;
246 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
249 * Force a quiescent state for RCU BH.
251 void rcu_bh_force_quiescent_state(void)
253 force_quiescent_state(&rcu_bh_state);
255 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
258 * Record the number of times rcutorture tests have been initiated and
259 * terminated. This information allows the debugfs tracing stats to be
260 * correlated to the rcutorture messages, even when the rcutorture module
261 * is being repeatedly loaded and unloaded. In other words, we cannot
262 * store this state in rcutorture itself.
264 void rcutorture_record_test_transition(void)
266 rcutorture_testseq++;
267 rcutorture_vernum = 0;
269 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
272 * Record the number of writer passes through the current rcutorture test.
273 * This is also used to correlate debugfs tracing stats with the rcutorture
276 void rcutorture_record_progress(unsigned long vernum)
280 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
283 * Force a quiescent state for RCU-sched.
285 void rcu_sched_force_quiescent_state(void)
287 force_quiescent_state(&rcu_sched_state);
289 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
292 * Does the CPU have callbacks ready to be invoked?
295 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
297 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
298 rdp->nxttail[RCU_DONE_TAIL] != NULL;
302 * Does the current CPU require a not-yet-started grace period?
303 * The caller must have disabled interrupts to prevent races with
304 * normal callback registry.
307 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
311 if (rcu_gp_in_progress(rsp))
312 return 0; /* No, a grace period is already in progress. */
313 if (rcu_nocb_needs_gp(rdp))
314 return 1; /* Yes, a no-CBs CPU needs one. */
315 if (!rdp->nxttail[RCU_NEXT_TAIL])
316 return 0; /* No, this is a no-CBs (or offline) CPU. */
317 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
318 return 1; /* Yes, this CPU has newly registered callbacks. */
319 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
320 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
321 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
322 rdp->nxtcompleted[i]))
323 return 1; /* Yes, CBs for future grace period. */
324 return 0; /* No grace period needed. */
328 * Return the root node of the specified rcu_state structure.
330 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
332 return &rsp->node[0];
336 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
338 * If the new value of the ->dynticks_nesting counter now is zero,
339 * we really have entered idle, and must do the appropriate accounting.
340 * The caller must have disabled interrupts.
342 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
345 trace_rcu_dyntick("Start", oldval, rdtp->dynticks_nesting);
346 if (!user && !is_idle_task(current)) {
347 struct task_struct *idle = idle_task(smp_processor_id());
349 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
350 ftrace_dump(DUMP_ORIG);
351 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
352 current->pid, current->comm,
353 idle->pid, idle->comm); /* must be idle task! */
355 rcu_prepare_for_idle(smp_processor_id());
356 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
357 smp_mb__before_atomic_inc(); /* See above. */
358 atomic_inc(&rdtp->dynticks);
359 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
360 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
363 * It is illegal to enter an extended quiescent state while
364 * in an RCU read-side critical section.
366 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
367 "Illegal idle entry in RCU read-side critical section.");
368 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
369 "Illegal idle entry in RCU-bh read-side critical section.");
370 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
371 "Illegal idle entry in RCU-sched read-side critical section.");
375 * Enter an RCU extended quiescent state, which can be either the
376 * idle loop or adaptive-tickless usermode execution.
378 static void rcu_eqs_enter(bool user)
381 struct rcu_dynticks *rdtp;
383 rdtp = &__get_cpu_var(rcu_dynticks);
384 oldval = rdtp->dynticks_nesting;
385 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
386 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
387 rdtp->dynticks_nesting = 0;
389 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
390 rcu_eqs_enter_common(rdtp, oldval, user);
394 * rcu_idle_enter - inform RCU that current CPU is entering idle
396 * Enter idle mode, in other words, -leave- the mode in which RCU
397 * read-side critical sections can occur. (Though RCU read-side
398 * critical sections can occur in irq handlers in idle, a possibility
399 * handled by irq_enter() and irq_exit().)
401 * We crowbar the ->dynticks_nesting field to zero to allow for
402 * the possibility of usermode upcalls having messed up our count
403 * of interrupt nesting level during the prior busy period.
405 void rcu_idle_enter(void)
409 local_irq_save(flags);
410 rcu_eqs_enter(false);
411 local_irq_restore(flags);
413 EXPORT_SYMBOL_GPL(rcu_idle_enter);
415 #ifdef CONFIG_RCU_USER_QS
417 * rcu_user_enter - inform RCU that we are resuming userspace.
419 * Enter RCU idle mode right before resuming userspace. No use of RCU
420 * is permitted between this call and rcu_user_exit(). This way the
421 * CPU doesn't need to maintain the tick for RCU maintenance purposes
422 * when the CPU runs in userspace.
424 void rcu_user_enter(void)
430 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
431 * after the current irq returns.
433 * This is similar to rcu_user_enter() but in the context of a non-nesting
434 * irq. After this call, RCU enters into idle mode when the interrupt
437 void rcu_user_enter_after_irq(void)
440 struct rcu_dynticks *rdtp;
442 local_irq_save(flags);
443 rdtp = &__get_cpu_var(rcu_dynticks);
444 /* Ensure this irq is interrupting a non-idle RCU state. */
445 WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
446 rdtp->dynticks_nesting = 1;
447 local_irq_restore(flags);
449 #endif /* CONFIG_RCU_USER_QS */
452 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
454 * Exit from an interrupt handler, which might possibly result in entering
455 * idle mode, in other words, leaving the mode in which read-side critical
456 * sections can occur.
458 * This code assumes that the idle loop never does anything that might
459 * result in unbalanced calls to irq_enter() and irq_exit(). If your
460 * architecture violates this assumption, RCU will give you what you
461 * deserve, good and hard. But very infrequently and irreproducibly.
463 * Use things like work queues to work around this limitation.
465 * You have been warned.
467 void rcu_irq_exit(void)
471 struct rcu_dynticks *rdtp;
473 local_irq_save(flags);
474 rdtp = &__get_cpu_var(rcu_dynticks);
475 oldval = rdtp->dynticks_nesting;
476 rdtp->dynticks_nesting--;
477 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
478 if (rdtp->dynticks_nesting)
479 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
481 rcu_eqs_enter_common(rdtp, oldval, true);
482 local_irq_restore(flags);
486 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
488 * If the new value of the ->dynticks_nesting counter was previously zero,
489 * we really have exited idle, and must do the appropriate accounting.
490 * The caller must have disabled interrupts.
492 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
495 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
496 atomic_inc(&rdtp->dynticks);
497 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
498 smp_mb__after_atomic_inc(); /* See above. */
499 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
500 rcu_cleanup_after_idle(smp_processor_id());
501 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
502 if (!user && !is_idle_task(current)) {
503 struct task_struct *idle = idle_task(smp_processor_id());
505 trace_rcu_dyntick("Error on exit: not idle task",
506 oldval, rdtp->dynticks_nesting);
507 ftrace_dump(DUMP_ORIG);
508 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
509 current->pid, current->comm,
510 idle->pid, idle->comm); /* must be idle task! */
515 * Exit an RCU extended quiescent state, which can be either the
516 * idle loop or adaptive-tickless usermode execution.
518 static void rcu_eqs_exit(bool user)
520 struct rcu_dynticks *rdtp;
523 rdtp = &__get_cpu_var(rcu_dynticks);
524 oldval = rdtp->dynticks_nesting;
525 WARN_ON_ONCE(oldval < 0);
526 if (oldval & DYNTICK_TASK_NEST_MASK)
527 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
529 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
530 rcu_eqs_exit_common(rdtp, oldval, user);
534 * rcu_idle_exit - inform RCU that current CPU is leaving idle
536 * Exit idle mode, in other words, -enter- the mode in which RCU
537 * read-side critical sections can occur.
539 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
540 * allow for the possibility of usermode upcalls messing up our count
541 * of interrupt nesting level during the busy period that is just
544 void rcu_idle_exit(void)
548 local_irq_save(flags);
550 local_irq_restore(flags);
552 EXPORT_SYMBOL_GPL(rcu_idle_exit);
554 #ifdef CONFIG_RCU_USER_QS
556 * rcu_user_exit - inform RCU that we are exiting userspace.
558 * Exit RCU idle mode while entering the kernel because it can
559 * run a RCU read side critical section anytime.
561 void rcu_user_exit(void)
567 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
568 * idle mode after the current non-nesting irq returns.
570 * This is similar to rcu_user_exit() but in the context of an irq.
571 * This is called when the irq has interrupted a userspace RCU idle mode
572 * context. When the current non-nesting interrupt returns after this call,
573 * the CPU won't restore the RCU idle mode.
575 void rcu_user_exit_after_irq(void)
578 struct rcu_dynticks *rdtp;
580 local_irq_save(flags);
581 rdtp = &__get_cpu_var(rcu_dynticks);
582 /* Ensure we are interrupting an RCU idle mode. */
583 WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
584 rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
585 local_irq_restore(flags);
587 #endif /* CONFIG_RCU_USER_QS */
590 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
592 * Enter an interrupt handler, which might possibly result in exiting
593 * idle mode, in other words, entering the mode in which read-side critical
594 * sections can occur.
596 * Note that the Linux kernel is fully capable of entering an interrupt
597 * handler that it never exits, for example when doing upcalls to
598 * user mode! This code assumes that the idle loop never does upcalls to
599 * user mode. If your architecture does do upcalls from the idle loop (or
600 * does anything else that results in unbalanced calls to the irq_enter()
601 * and irq_exit() functions), RCU will give you what you deserve, good
602 * and hard. But very infrequently and irreproducibly.
604 * Use things like work queues to work around this limitation.
606 * You have been warned.
608 void rcu_irq_enter(void)
611 struct rcu_dynticks *rdtp;
614 local_irq_save(flags);
615 rdtp = &__get_cpu_var(rcu_dynticks);
616 oldval = rdtp->dynticks_nesting;
617 rdtp->dynticks_nesting++;
618 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
620 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
622 rcu_eqs_exit_common(rdtp, oldval, true);
623 local_irq_restore(flags);
627 * rcu_nmi_enter - inform RCU of entry to NMI context
629 * If the CPU was idle with dynamic ticks active, and there is no
630 * irq handler running, this updates rdtp->dynticks_nmi to let the
631 * RCU grace-period handling know that the CPU is active.
633 void rcu_nmi_enter(void)
635 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
637 if (rdtp->dynticks_nmi_nesting == 0 &&
638 (atomic_read(&rdtp->dynticks) & 0x1))
640 rdtp->dynticks_nmi_nesting++;
641 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
642 atomic_inc(&rdtp->dynticks);
643 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
644 smp_mb__after_atomic_inc(); /* See above. */
645 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
649 * rcu_nmi_exit - inform RCU of exit from NMI context
651 * If the CPU was idle with dynamic ticks active, and there is no
652 * irq handler running, this updates rdtp->dynticks_nmi to let the
653 * RCU grace-period handling know that the CPU is no longer active.
655 void rcu_nmi_exit(void)
657 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
659 if (rdtp->dynticks_nmi_nesting == 0 ||
660 --rdtp->dynticks_nmi_nesting != 0)
662 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
663 smp_mb__before_atomic_inc(); /* See above. */
664 atomic_inc(&rdtp->dynticks);
665 smp_mb__after_atomic_inc(); /* Force delay to next write. */
666 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
670 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
672 * If the current CPU is in its idle loop and is neither in an interrupt
673 * or NMI handler, return true.
675 int rcu_is_cpu_idle(void)
680 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
684 EXPORT_SYMBOL(rcu_is_cpu_idle);
686 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
689 * Is the current CPU online? Disable preemption to avoid false positives
690 * that could otherwise happen due to the current CPU number being sampled,
691 * this task being preempted, its old CPU being taken offline, resuming
692 * on some other CPU, then determining that its old CPU is now offline.
693 * It is OK to use RCU on an offline processor during initial boot, hence
694 * the check for rcu_scheduler_fully_active. Note also that it is OK
695 * for a CPU coming online to use RCU for one jiffy prior to marking itself
696 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
697 * offline to continue to use RCU for one jiffy after marking itself
698 * offline in the cpu_online_mask. This leniency is necessary given the
699 * non-atomic nature of the online and offline processing, for example,
700 * the fact that a CPU enters the scheduler after completing the CPU_DYING
703 * This is also why RCU internally marks CPUs online during the
704 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
706 * Disable checking if in an NMI handler because we cannot safely report
707 * errors from NMI handlers anyway.
709 bool rcu_lockdep_current_cpu_online(void)
711 struct rcu_data *rdp;
712 struct rcu_node *rnp;
718 rdp = &__get_cpu_var(rcu_sched_data);
720 ret = (rdp->grpmask & rnp->qsmaskinit) ||
721 !rcu_scheduler_fully_active;
725 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
727 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
730 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
732 * If the current CPU is idle or running at a first-level (not nested)
733 * interrupt from idle, return true. The caller must have at least
734 * disabled preemption.
736 static int rcu_is_cpu_rrupt_from_idle(void)
738 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
742 * Snapshot the specified CPU's dynticks counter so that we can later
743 * credit them with an implicit quiescent state. Return 1 if this CPU
744 * is in dynticks idle mode, which is an extended quiescent state.
746 static int dyntick_save_progress_counter(struct rcu_data *rdp)
748 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
749 return (rdp->dynticks_snap & 0x1) == 0;
753 * Return true if the specified CPU has passed through a quiescent
754 * state by virtue of being in or having passed through an dynticks
755 * idle state since the last call to dyntick_save_progress_counter()
756 * for this same CPU, or by virtue of having been offline.
758 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
763 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
764 snap = (unsigned int)rdp->dynticks_snap;
767 * If the CPU passed through or entered a dynticks idle phase with
768 * no active irq/NMI handlers, then we can safely pretend that the CPU
769 * already acknowledged the request to pass through a quiescent
770 * state. Either way, that CPU cannot possibly be in an RCU
771 * read-side critical section that started before the beginning
772 * of the current RCU grace period.
774 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
775 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
781 * Check for the CPU being offline, but only if the grace period
782 * is old enough. We don't need to worry about the CPU changing
783 * state: If we see it offline even once, it has been through a
786 * The reason for insisting that the grace period be at least
787 * one jiffy old is that CPUs that are not quite online and that
788 * have just gone offline can still execute RCU read-side critical
791 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
792 return 0; /* Grace period is not old enough. */
794 if (cpu_is_offline(rdp->cpu)) {
795 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
802 static void record_gp_stall_check_time(struct rcu_state *rsp)
804 rsp->gp_start = jiffies;
805 rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
809 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
810 * for architectures that do not implement trigger_all_cpu_backtrace().
811 * The NMI-triggered stack traces are more accurate because they are
812 * printed by the target CPU.
814 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
818 struct rcu_node *rnp;
820 rcu_for_each_leaf_node(rsp, rnp) {
821 raw_spin_lock_irqsave(&rnp->lock, flags);
822 if (rnp->qsmask != 0) {
823 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
824 if (rnp->qsmask & (1UL << cpu))
825 dump_cpu_task(rnp->grplo + cpu);
827 raw_spin_unlock_irqrestore(&rnp->lock, flags);
831 static void print_other_cpu_stall(struct rcu_state *rsp)
837 struct rcu_node *rnp = rcu_get_root(rsp);
840 /* Only let one CPU complain about others per time interval. */
842 raw_spin_lock_irqsave(&rnp->lock, flags);
843 delta = jiffies - rsp->jiffies_stall;
844 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
845 raw_spin_unlock_irqrestore(&rnp->lock, flags);
848 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
849 raw_spin_unlock_irqrestore(&rnp->lock, flags);
852 * OK, time to rat on our buddy...
853 * See Documentation/RCU/stallwarn.txt for info on how to debug
854 * RCU CPU stall warnings.
856 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
858 print_cpu_stall_info_begin();
859 rcu_for_each_leaf_node(rsp, rnp) {
860 raw_spin_lock_irqsave(&rnp->lock, flags);
861 ndetected += rcu_print_task_stall(rnp);
862 if (rnp->qsmask != 0) {
863 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
864 if (rnp->qsmask & (1UL << cpu)) {
865 print_cpu_stall_info(rsp,
870 raw_spin_unlock_irqrestore(&rnp->lock, flags);
874 * Now rat on any tasks that got kicked up to the root rcu_node
875 * due to CPU offlining.
877 rnp = rcu_get_root(rsp);
878 raw_spin_lock_irqsave(&rnp->lock, flags);
879 ndetected += rcu_print_task_stall(rnp);
880 raw_spin_unlock_irqrestore(&rnp->lock, flags);
882 print_cpu_stall_info_end();
883 for_each_possible_cpu(cpu)
884 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
885 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
886 smp_processor_id(), (long)(jiffies - rsp->gp_start),
887 rsp->gpnum, rsp->completed, totqlen);
889 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
890 else if (!trigger_all_cpu_backtrace())
891 rcu_dump_cpu_stacks(rsp);
893 /* Complain about tasks blocking the grace period. */
895 rcu_print_detail_task_stall(rsp);
897 force_quiescent_state(rsp); /* Kick them all. */
900 static void print_cpu_stall(struct rcu_state *rsp)
904 struct rcu_node *rnp = rcu_get_root(rsp);
908 * OK, time to rat on ourselves...
909 * See Documentation/RCU/stallwarn.txt for info on how to debug
910 * RCU CPU stall warnings.
912 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
913 print_cpu_stall_info_begin();
914 print_cpu_stall_info(rsp, smp_processor_id());
915 print_cpu_stall_info_end();
916 for_each_possible_cpu(cpu)
917 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
918 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
919 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
920 if (!trigger_all_cpu_backtrace())
923 raw_spin_lock_irqsave(&rnp->lock, flags);
924 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
925 rsp->jiffies_stall = jiffies +
926 3 * rcu_jiffies_till_stall_check() + 3;
927 raw_spin_unlock_irqrestore(&rnp->lock, flags);
929 set_need_resched(); /* kick ourselves to get things going. */
932 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
936 struct rcu_node *rnp;
938 if (rcu_cpu_stall_suppress)
940 j = ACCESS_ONCE(jiffies);
941 js = ACCESS_ONCE(rsp->jiffies_stall);
943 if (rcu_gp_in_progress(rsp) &&
944 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
946 /* We haven't checked in, so go dump stack. */
947 print_cpu_stall(rsp);
949 } else if (rcu_gp_in_progress(rsp) &&
950 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
952 /* They had a few time units to dump stack, so complain. */
953 print_other_cpu_stall(rsp);
958 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
960 * Set the stall-warning timeout way off into the future, thus preventing
961 * any RCU CPU stall-warning messages from appearing in the current set of
964 * The caller must disable hard irqs.
966 void rcu_cpu_stall_reset(void)
968 struct rcu_state *rsp;
970 for_each_rcu_flavor(rsp)
971 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
975 * Update CPU-local rcu_data state to record the newly noticed grace period.
976 * This is used both when we started the grace period and when we notice
977 * that someone else started the grace period. The caller must hold the
978 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
979 * and must have irqs disabled.
981 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
983 if (rdp->gpnum != rnp->gpnum) {
985 * If the current grace period is waiting for this CPU,
986 * set up to detect a quiescent state, otherwise don't
987 * go looking for one.
989 rdp->gpnum = rnp->gpnum;
990 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
991 rdp->passed_quiesce = 0;
992 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
993 zero_cpu_stall_ticks(rdp);
997 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
1000 struct rcu_node *rnp;
1002 local_irq_save(flags);
1004 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1005 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1006 local_irq_restore(flags);
1009 __note_new_gpnum(rsp, rnp, rdp);
1010 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1014 * Did someone else start a new RCU grace period start since we last
1015 * checked? Update local state appropriately if so. Must be called
1016 * on the CPU corresponding to rdp.
1019 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1021 unsigned long flags;
1024 local_irq_save(flags);
1025 if (rdp->gpnum != rsp->gpnum) {
1026 note_new_gpnum(rsp, rdp);
1029 local_irq_restore(flags);
1034 * Initialize the specified rcu_data structure's callback list to empty.
1036 static void init_callback_list(struct rcu_data *rdp)
1040 if (init_nocb_callback_list(rdp))
1042 rdp->nxtlist = NULL;
1043 for (i = 0; i < RCU_NEXT_SIZE; i++)
1044 rdp->nxttail[i] = &rdp->nxtlist;
1048 * Determine the value that ->completed will have at the end of the
1049 * next subsequent grace period. This is used to tag callbacks so that
1050 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1051 * been dyntick-idle for an extended period with callbacks under the
1052 * influence of RCU_FAST_NO_HZ.
1054 * The caller must hold rnp->lock with interrupts disabled.
1056 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1057 struct rcu_node *rnp)
1060 * If RCU is idle, we just wait for the next grace period.
1061 * But we can only be sure that RCU is idle if we are looking
1062 * at the root rcu_node structure -- otherwise, a new grace
1063 * period might have started, but just not yet gotten around
1064 * to initializing the current non-root rcu_node structure.
1066 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1067 return rnp->completed + 1;
1070 * Otherwise, wait for a possible partial grace period and
1071 * then the subsequent full grace period.
1073 return rnp->completed + 2;
1077 * If there is room, assign a ->completed number to any callbacks on
1078 * this CPU that have not already been assigned. Also accelerate any
1079 * callbacks that were previously assigned a ->completed number that has
1080 * since proven to be too conservative, which can happen if callbacks get
1081 * assigned a ->completed number while RCU is idle, but with reference to
1082 * a non-root rcu_node structure. This function is idempotent, so it does
1083 * not hurt to call it repeatedly.
1085 * The caller must hold rnp->lock with interrupts disabled.
1087 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1088 struct rcu_data *rdp)
1093 /* If the CPU has no callbacks, nothing to do. */
1094 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1098 * Starting from the sublist containing the callbacks most
1099 * recently assigned a ->completed number and working down, find the
1100 * first sublist that is not assignable to an upcoming grace period.
1101 * Such a sublist has something in it (first two tests) and has
1102 * a ->completed number assigned that will complete sooner than
1103 * the ->completed number for newly arrived callbacks (last test).
1105 * The key point is that any later sublist can be assigned the
1106 * same ->completed number as the newly arrived callbacks, which
1107 * means that the callbacks in any of these later sublist can be
1108 * grouped into a single sublist, whether or not they have already
1109 * been assigned a ->completed number.
1111 c = rcu_cbs_completed(rsp, rnp);
1112 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1113 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1114 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1118 * If there are no sublist for unassigned callbacks, leave.
1119 * At the same time, advance "i" one sublist, so that "i" will
1120 * index into the sublist where all the remaining callbacks should
1123 if (++i >= RCU_NEXT_TAIL)
1127 * Assign all subsequent callbacks' ->completed number to the next
1128 * full grace period and group them all in the sublist initially
1131 for (; i <= RCU_NEXT_TAIL; i++) {
1132 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1133 rdp->nxtcompleted[i] = c;
1136 /* Trace depending on how much we were able to accelerate. */
1137 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1138 trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccWaitCB");
1140 trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccReadyCB");
1144 * Move any callbacks whose grace period has completed to the
1145 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1146 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1147 * sublist. This function is idempotent, so it does not hurt to
1148 * invoke it repeatedly. As long as it is not invoked -too- often...
1150 * The caller must hold rnp->lock with interrupts disabled.
1152 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1153 struct rcu_data *rdp)
1157 /* If the CPU has no callbacks, nothing to do. */
1158 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1162 * Find all callbacks whose ->completed numbers indicate that they
1163 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1165 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1166 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1168 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1170 /* Clean up any sublist tail pointers that were misordered above. */
1171 for (j = RCU_WAIT_TAIL; j < i; j++)
1172 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1174 /* Copy down callbacks to fill in empty sublists. */
1175 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1176 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1178 rdp->nxttail[j] = rdp->nxttail[i];
1179 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1182 /* Classify any remaining callbacks. */
1183 rcu_accelerate_cbs(rsp, rnp, rdp);
1187 * Advance this CPU's callbacks, but only if the current grace period
1188 * has ended. This may be called only from the CPU to whom the rdp
1189 * belongs. In addition, the corresponding leaf rcu_node structure's
1190 * ->lock must be held by the caller, with irqs disabled.
1193 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1195 /* Did another grace period end? */
1196 if (rdp->completed == rnp->completed) {
1198 /* No, so just accelerate recent callbacks. */
1199 rcu_accelerate_cbs(rsp, rnp, rdp);
1203 /* Advance callbacks. */
1204 rcu_advance_cbs(rsp, rnp, rdp);
1206 /* Remember that we saw this grace-period completion. */
1207 rdp->completed = rnp->completed;
1208 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1211 * If we were in an extended quiescent state, we may have
1212 * missed some grace periods that others CPUs handled on
1213 * our behalf. Catch up with this state to avoid noting
1214 * spurious new grace periods. If another grace period
1215 * has started, then rnp->gpnum will have advanced, so
1216 * we will detect this later on. Of course, any quiescent
1217 * states we found for the old GP are now invalid.
1219 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1220 rdp->gpnum = rdp->completed;
1221 rdp->passed_quiesce = 0;
1225 * If RCU does not need a quiescent state from this CPU,
1226 * then make sure that this CPU doesn't go looking for one.
1228 if ((rnp->qsmask & rdp->grpmask) == 0)
1229 rdp->qs_pending = 0;
1234 * Advance this CPU's callbacks, but only if the current grace period
1235 * has ended. This may be called only from the CPU to whom the rdp
1239 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1241 unsigned long flags;
1242 struct rcu_node *rnp;
1244 local_irq_save(flags);
1246 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1247 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1248 local_irq_restore(flags);
1251 __rcu_process_gp_end(rsp, rnp, rdp);
1252 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1256 * Do per-CPU grace-period initialization for running CPU. The caller
1257 * must hold the lock of the leaf rcu_node structure corresponding to
1261 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1263 /* Prior grace period ended, so advance callbacks for current CPU. */
1264 __rcu_process_gp_end(rsp, rnp, rdp);
1266 /* Set state so that this CPU will detect the next quiescent state. */
1267 __note_new_gpnum(rsp, rnp, rdp);
1271 * Initialize a new grace period.
1273 static int rcu_gp_init(struct rcu_state *rsp)
1275 struct rcu_data *rdp;
1276 struct rcu_node *rnp = rcu_get_root(rsp);
1278 raw_spin_lock_irq(&rnp->lock);
1279 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1281 if (rcu_gp_in_progress(rsp)) {
1282 /* Grace period already in progress, don't start another. */
1283 raw_spin_unlock_irq(&rnp->lock);
1287 /* Advance to a new grace period and initialize state. */
1289 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1290 record_gp_stall_check_time(rsp);
1291 raw_spin_unlock_irq(&rnp->lock);
1293 /* Exclude any concurrent CPU-hotplug operations. */
1294 mutex_lock(&rsp->onoff_mutex);
1297 * Set the quiescent-state-needed bits in all the rcu_node
1298 * structures for all currently online CPUs in breadth-first order,
1299 * starting from the root rcu_node structure, relying on the layout
1300 * of the tree within the rsp->node[] array. Note that other CPUs
1301 * will access only the leaves of the hierarchy, thus seeing that no
1302 * grace period is in progress, at least until the corresponding
1303 * leaf node has been initialized. In addition, we have excluded
1304 * CPU-hotplug operations.
1306 * The grace period cannot complete until the initialization
1307 * process finishes, because this kthread handles both.
1309 rcu_for_each_node_breadth_first(rsp, rnp) {
1310 raw_spin_lock_irq(&rnp->lock);
1311 rdp = this_cpu_ptr(rsp->rda);
1312 rcu_preempt_check_blocked_tasks(rnp);
1313 rnp->qsmask = rnp->qsmaskinit;
1314 rnp->gpnum = rsp->gpnum;
1315 WARN_ON_ONCE(rnp->completed != rsp->completed);
1316 rnp->completed = rsp->completed;
1317 if (rnp == rdp->mynode)
1318 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1319 rcu_preempt_boost_start_gp(rnp);
1320 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1321 rnp->level, rnp->grplo,
1322 rnp->grphi, rnp->qsmask);
1323 raw_spin_unlock_irq(&rnp->lock);
1324 #ifdef CONFIG_PROVE_RCU_DELAY
1325 if ((random32() % (rcu_num_nodes * 8)) == 0)
1326 schedule_timeout_uninterruptible(2);
1327 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1331 mutex_unlock(&rsp->onoff_mutex);
1336 * Do one round of quiescent-state forcing.
1338 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1340 int fqs_state = fqs_state_in;
1341 struct rcu_node *rnp = rcu_get_root(rsp);
1344 if (fqs_state == RCU_SAVE_DYNTICK) {
1345 /* Collect dyntick-idle snapshots. */
1346 force_qs_rnp(rsp, dyntick_save_progress_counter);
1347 fqs_state = RCU_FORCE_QS;
1349 /* Handle dyntick-idle and offline CPUs. */
1350 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1352 /* Clear flag to prevent immediate re-entry. */
1353 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1354 raw_spin_lock_irq(&rnp->lock);
1355 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1356 raw_spin_unlock_irq(&rnp->lock);
1362 * Clean up after the old grace period.
1364 static void rcu_gp_cleanup(struct rcu_state *rsp)
1366 unsigned long gp_duration;
1367 struct rcu_data *rdp;
1368 struct rcu_node *rnp = rcu_get_root(rsp);
1370 raw_spin_lock_irq(&rnp->lock);
1371 gp_duration = jiffies - rsp->gp_start;
1372 if (gp_duration > rsp->gp_max)
1373 rsp->gp_max = gp_duration;
1376 * We know the grace period is complete, but to everyone else
1377 * it appears to still be ongoing. But it is also the case
1378 * that to everyone else it looks like there is nothing that
1379 * they can do to advance the grace period. It is therefore
1380 * safe for us to drop the lock in order to mark the grace
1381 * period as completed in all of the rcu_node structures.
1383 raw_spin_unlock_irq(&rnp->lock);
1386 * Propagate new ->completed value to rcu_node structures so
1387 * that other CPUs don't have to wait until the start of the next
1388 * grace period to process their callbacks. This also avoids
1389 * some nasty RCU grace-period initialization races by forcing
1390 * the end of the current grace period to be completely recorded in
1391 * all of the rcu_node structures before the beginning of the next
1392 * grace period is recorded in any of the rcu_node structures.
1394 rcu_for_each_node_breadth_first(rsp, rnp) {
1395 raw_spin_lock_irq(&rnp->lock);
1396 rnp->completed = rsp->gpnum;
1397 raw_spin_unlock_irq(&rnp->lock);
1400 rnp = rcu_get_root(rsp);
1401 raw_spin_lock_irq(&rnp->lock);
1403 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1404 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1405 rsp->fqs_state = RCU_GP_IDLE;
1406 rdp = this_cpu_ptr(rsp->rda);
1407 if (cpu_needs_another_gp(rsp, rdp))
1409 raw_spin_unlock_irq(&rnp->lock);
1413 * Body of kthread that handles grace periods.
1415 static int __noreturn rcu_gp_kthread(void *arg)
1420 struct rcu_state *rsp = arg;
1421 struct rcu_node *rnp = rcu_get_root(rsp);
1425 /* Handle grace-period start. */
1427 wait_event_interruptible(rsp->gp_wq,
1430 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1434 flush_signals(current);
1437 /* Handle quiescent-state forcing. */
1438 fqs_state = RCU_SAVE_DYNTICK;
1439 j = jiffies_till_first_fqs;
1442 jiffies_till_first_fqs = HZ;
1445 rsp->jiffies_force_qs = jiffies + j;
1446 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1447 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1448 (!ACCESS_ONCE(rnp->qsmask) &&
1449 !rcu_preempt_blocked_readers_cgp(rnp)),
1451 /* If grace period done, leave loop. */
1452 if (!ACCESS_ONCE(rnp->qsmask) &&
1453 !rcu_preempt_blocked_readers_cgp(rnp))
1455 /* If time for quiescent-state forcing, do it. */
1456 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1457 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1460 /* Deal with stray signal. */
1462 flush_signals(current);
1464 j = jiffies_till_next_fqs;
1467 jiffies_till_next_fqs = HZ;
1470 jiffies_till_next_fqs = 1;
1474 /* Handle grace-period end. */
1475 rcu_gp_cleanup(rsp);
1480 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1481 * in preparation for detecting the next grace period. The caller must hold
1482 * the root node's ->lock, which is released before return. Hard irqs must
1485 * Note that it is legal for a dying CPU (which is marked as offline) to
1486 * invoke this function. This can happen when the dying CPU reports its
1490 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1491 __releases(rcu_get_root(rsp)->lock)
1493 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1494 struct rcu_node *rnp = rcu_get_root(rsp);
1496 if (!rsp->gp_kthread ||
1497 !cpu_needs_another_gp(rsp, rdp)) {
1499 * Either we have not yet spawned the grace-period
1500 * task, this CPU does not need another grace period,
1501 * or a grace period is already in progress.
1502 * Either way, don't start a new grace period.
1504 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1509 * Because there is no grace period in progress right now,
1510 * any callbacks we have up to this point will be satisfied
1511 * by the next grace period. So this is a good place to
1512 * assign a grace period number to recently posted callbacks.
1514 rcu_accelerate_cbs(rsp, rnp, rdp);
1516 rsp->gp_flags = RCU_GP_FLAG_INIT;
1517 raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
1519 /* Ensure that CPU is aware of completion of last grace period. */
1520 rcu_process_gp_end(rsp, rdp);
1521 local_irq_restore(flags);
1523 /* Wake up rcu_gp_kthread() to start the grace period. */
1524 wake_up(&rsp->gp_wq);
1528 * Report a full set of quiescent states to the specified rcu_state
1529 * data structure. This involves cleaning up after the prior grace
1530 * period and letting rcu_start_gp() start up the next grace period
1531 * if one is needed. Note that the caller must hold rnp->lock, as
1532 * required by rcu_start_gp(), which will release it.
1534 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1535 __releases(rcu_get_root(rsp)->lock)
1537 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1538 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1539 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1543 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1544 * Allows quiescent states for a group of CPUs to be reported at one go
1545 * to the specified rcu_node structure, though all the CPUs in the group
1546 * must be represented by the same rcu_node structure (which need not be
1547 * a leaf rcu_node structure, though it often will be). That structure's
1548 * lock must be held upon entry, and it is released before return.
1551 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1552 struct rcu_node *rnp, unsigned long flags)
1553 __releases(rnp->lock)
1555 struct rcu_node *rnp_c;
1557 /* Walk up the rcu_node hierarchy. */
1559 if (!(rnp->qsmask & mask)) {
1561 /* Our bit has already been cleared, so done. */
1562 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1565 rnp->qsmask &= ~mask;
1566 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1567 mask, rnp->qsmask, rnp->level,
1568 rnp->grplo, rnp->grphi,
1570 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1572 /* Other bits still set at this level, so done. */
1573 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1576 mask = rnp->grpmask;
1577 if (rnp->parent == NULL) {
1579 /* No more levels. Exit loop holding root lock. */
1583 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1586 raw_spin_lock_irqsave(&rnp->lock, flags);
1587 WARN_ON_ONCE(rnp_c->qsmask);
1591 * Get here if we are the last CPU to pass through a quiescent
1592 * state for this grace period. Invoke rcu_report_qs_rsp()
1593 * to clean up and start the next grace period if one is needed.
1595 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1599 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1600 * structure. This must be either called from the specified CPU, or
1601 * called when the specified CPU is known to be offline (and when it is
1602 * also known that no other CPU is concurrently trying to help the offline
1603 * CPU). The lastcomp argument is used to make sure we are still in the
1604 * grace period of interest. We don't want to end the current grace period
1605 * based on quiescent states detected in an earlier grace period!
1608 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1610 unsigned long flags;
1612 struct rcu_node *rnp;
1615 raw_spin_lock_irqsave(&rnp->lock, flags);
1616 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1617 rnp->completed == rnp->gpnum) {
1620 * The grace period in which this quiescent state was
1621 * recorded has ended, so don't report it upwards.
1622 * We will instead need a new quiescent state that lies
1623 * within the current grace period.
1625 rdp->passed_quiesce = 0; /* need qs for new gp. */
1626 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1629 mask = rdp->grpmask;
1630 if ((rnp->qsmask & mask) == 0) {
1631 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1633 rdp->qs_pending = 0;
1636 * This GP can't end until cpu checks in, so all of our
1637 * callbacks can be processed during the next GP.
1639 rcu_accelerate_cbs(rsp, rnp, rdp);
1641 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1646 * Check to see if there is a new grace period of which this CPU
1647 * is not yet aware, and if so, set up local rcu_data state for it.
1648 * Otherwise, see if this CPU has just passed through its first
1649 * quiescent state for this grace period, and record that fact if so.
1652 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1654 /* If there is now a new grace period, record and return. */
1655 if (check_for_new_grace_period(rsp, rdp))
1659 * Does this CPU still need to do its part for current grace period?
1660 * If no, return and let the other CPUs do their part as well.
1662 if (!rdp->qs_pending)
1666 * Was there a quiescent state since the beginning of the grace
1667 * period? If no, then exit and wait for the next call.
1669 if (!rdp->passed_quiesce)
1673 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1676 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1679 #ifdef CONFIG_HOTPLUG_CPU
1682 * Send the specified CPU's RCU callbacks to the orphanage. The
1683 * specified CPU must be offline, and the caller must hold the
1687 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1688 struct rcu_node *rnp, struct rcu_data *rdp)
1690 /* No-CBs CPUs do not have orphanable callbacks. */
1691 if (is_nocb_cpu(rdp->cpu))
1695 * Orphan the callbacks. First adjust the counts. This is safe
1696 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1697 * cannot be running now. Thus no memory barrier is required.
1699 if (rdp->nxtlist != NULL) {
1700 rsp->qlen_lazy += rdp->qlen_lazy;
1701 rsp->qlen += rdp->qlen;
1702 rdp->n_cbs_orphaned += rdp->qlen;
1704 ACCESS_ONCE(rdp->qlen) = 0;
1708 * Next, move those callbacks still needing a grace period to
1709 * the orphanage, where some other CPU will pick them up.
1710 * Some of the callbacks might have gone partway through a grace
1711 * period, but that is too bad. They get to start over because we
1712 * cannot assume that grace periods are synchronized across CPUs.
1713 * We don't bother updating the ->nxttail[] array yet, instead
1714 * we just reset the whole thing later on.
1716 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1717 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1718 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1719 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1723 * Then move the ready-to-invoke callbacks to the orphanage,
1724 * where some other CPU will pick them up. These will not be
1725 * required to pass though another grace period: They are done.
1727 if (rdp->nxtlist != NULL) {
1728 *rsp->orphan_donetail = rdp->nxtlist;
1729 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1732 /* Finally, initialize the rcu_data structure's list to empty. */
1733 init_callback_list(rdp);
1737 * Adopt the RCU callbacks from the specified rcu_state structure's
1738 * orphanage. The caller must hold the ->orphan_lock.
1740 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1743 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1745 /* No-CBs CPUs are handled specially. */
1746 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1749 /* Do the accounting first. */
1750 rdp->qlen_lazy += rsp->qlen_lazy;
1751 rdp->qlen += rsp->qlen;
1752 rdp->n_cbs_adopted += rsp->qlen;
1753 if (rsp->qlen_lazy != rsp->qlen)
1754 rcu_idle_count_callbacks_posted();
1759 * We do not need a memory barrier here because the only way we
1760 * can get here if there is an rcu_barrier() in flight is if
1761 * we are the task doing the rcu_barrier().
1764 /* First adopt the ready-to-invoke callbacks. */
1765 if (rsp->orphan_donelist != NULL) {
1766 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1767 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1768 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1769 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1770 rdp->nxttail[i] = rsp->orphan_donetail;
1771 rsp->orphan_donelist = NULL;
1772 rsp->orphan_donetail = &rsp->orphan_donelist;
1775 /* And then adopt the callbacks that still need a grace period. */
1776 if (rsp->orphan_nxtlist != NULL) {
1777 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1778 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1779 rsp->orphan_nxtlist = NULL;
1780 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1785 * Trace the fact that this CPU is going offline.
1787 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1789 RCU_TRACE(unsigned long mask);
1790 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1791 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1793 RCU_TRACE(mask = rdp->grpmask);
1794 trace_rcu_grace_period(rsp->name,
1795 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1800 * The CPU has been completely removed, and some other CPU is reporting
1801 * this fact from process context. Do the remainder of the cleanup,
1802 * including orphaning the outgoing CPU's RCU callbacks, and also
1803 * adopting them. There can only be one CPU hotplug operation at a time,
1804 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1806 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1808 unsigned long flags;
1810 int need_report = 0;
1811 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1812 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1814 /* Adjust any no-longer-needed kthreads. */
1815 rcu_boost_kthread_setaffinity(rnp, -1);
1817 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1819 /* Exclude any attempts to start a new grace period. */
1820 mutex_lock(&rsp->onoff_mutex);
1821 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1823 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1824 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1825 rcu_adopt_orphan_cbs(rsp);
1827 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1828 mask = rdp->grpmask; /* rnp->grplo is constant. */
1830 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1831 rnp->qsmaskinit &= ~mask;
1832 if (rnp->qsmaskinit != 0) {
1833 if (rnp != rdp->mynode)
1834 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1837 if (rnp == rdp->mynode)
1838 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1840 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1841 mask = rnp->grpmask;
1843 } while (rnp != NULL);
1846 * We still hold the leaf rcu_node structure lock here, and
1847 * irqs are still disabled. The reason for this subterfuge is
1848 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1849 * held leads to deadlock.
1851 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1853 if (need_report & RCU_OFL_TASKS_NORM_GP)
1854 rcu_report_unblock_qs_rnp(rnp, flags);
1856 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1857 if (need_report & RCU_OFL_TASKS_EXP_GP)
1858 rcu_report_exp_rnp(rsp, rnp, true);
1859 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1860 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1861 cpu, rdp->qlen, rdp->nxtlist);
1862 init_callback_list(rdp);
1863 /* Disallow further callbacks on this CPU. */
1864 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1865 mutex_unlock(&rsp->onoff_mutex);
1868 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1870 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1874 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1878 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1881 * Invoke any RCU callbacks that have made it to the end of their grace
1882 * period. Thottle as specified by rdp->blimit.
1884 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1886 unsigned long flags;
1887 struct rcu_head *next, *list, **tail;
1888 long bl, count, count_lazy;
1891 /* If no callbacks are ready, just return. */
1892 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1893 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1894 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1895 need_resched(), is_idle_task(current),
1896 rcu_is_callbacks_kthread());
1901 * Extract the list of ready callbacks, disabling to prevent
1902 * races with call_rcu() from interrupt handlers.
1904 local_irq_save(flags);
1905 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1907 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1908 list = rdp->nxtlist;
1909 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1910 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1911 tail = rdp->nxttail[RCU_DONE_TAIL];
1912 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1913 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1914 rdp->nxttail[i] = &rdp->nxtlist;
1915 local_irq_restore(flags);
1917 /* Invoke callbacks. */
1918 count = count_lazy = 0;
1922 debug_rcu_head_unqueue(list);
1923 if (__rcu_reclaim(rsp->name, list))
1926 /* Stop only if limit reached and CPU has something to do. */
1927 if (++count >= bl &&
1929 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1933 local_irq_save(flags);
1934 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1935 is_idle_task(current),
1936 rcu_is_callbacks_kthread());
1938 /* Update count, and requeue any remaining callbacks. */
1940 *tail = rdp->nxtlist;
1941 rdp->nxtlist = list;
1942 for (i = 0; i < RCU_NEXT_SIZE; i++)
1943 if (&rdp->nxtlist == rdp->nxttail[i])
1944 rdp->nxttail[i] = tail;
1948 smp_mb(); /* List handling before counting for rcu_barrier(). */
1949 rdp->qlen_lazy -= count_lazy;
1950 ACCESS_ONCE(rdp->qlen) -= count;
1951 rdp->n_cbs_invoked += count;
1953 /* Reinstate batch limit if we have worked down the excess. */
1954 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1955 rdp->blimit = blimit;
1957 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1958 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1959 rdp->qlen_last_fqs_check = 0;
1960 rdp->n_force_qs_snap = rsp->n_force_qs;
1961 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1962 rdp->qlen_last_fqs_check = rdp->qlen;
1963 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1965 local_irq_restore(flags);
1967 /* Re-invoke RCU core processing if there are callbacks remaining. */
1968 if (cpu_has_callbacks_ready_to_invoke(rdp))
1973 * Check to see if this CPU is in a non-context-switch quiescent state
1974 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1975 * Also schedule RCU core processing.
1977 * This function must be called from hardirq context. It is normally
1978 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1979 * false, there is no point in invoking rcu_check_callbacks().
1981 void rcu_check_callbacks(int cpu, int user)
1983 trace_rcu_utilization("Start scheduler-tick");
1984 increment_cpu_stall_ticks();
1985 if (user || rcu_is_cpu_rrupt_from_idle()) {
1988 * Get here if this CPU took its interrupt from user
1989 * mode or from the idle loop, and if this is not a
1990 * nested interrupt. In this case, the CPU is in
1991 * a quiescent state, so note it.
1993 * No memory barrier is required here because both
1994 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1995 * variables that other CPUs neither access nor modify,
1996 * at least not while the corresponding CPU is online.
2002 } else if (!in_softirq()) {
2005 * Get here if this CPU did not take its interrupt from
2006 * softirq, in other words, if it is not interrupting
2007 * a rcu_bh read-side critical section. This is an _bh
2008 * critical section, so note it.
2013 rcu_preempt_check_callbacks(cpu);
2014 if (rcu_pending(cpu))
2016 trace_rcu_utilization("End scheduler-tick");
2020 * Scan the leaf rcu_node structures, processing dyntick state for any that
2021 * have not yet encountered a quiescent state, using the function specified.
2022 * Also initiate boosting for any threads blocked on the root rcu_node.
2024 * The caller must have suppressed start of new grace periods.
2026 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
2030 unsigned long flags;
2032 struct rcu_node *rnp;
2034 rcu_for_each_leaf_node(rsp, rnp) {
2037 raw_spin_lock_irqsave(&rnp->lock, flags);
2038 if (!rcu_gp_in_progress(rsp)) {
2039 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2042 if (rnp->qsmask == 0) {
2043 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2048 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2049 if ((rnp->qsmask & bit) != 0 &&
2050 f(per_cpu_ptr(rsp->rda, cpu)))
2055 /* rcu_report_qs_rnp() releases rnp->lock. */
2056 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2059 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2061 rnp = rcu_get_root(rsp);
2062 if (rnp->qsmask == 0) {
2063 raw_spin_lock_irqsave(&rnp->lock, flags);
2064 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2069 * Force quiescent states on reluctant CPUs, and also detect which
2070 * CPUs are in dyntick-idle mode.
2072 static void force_quiescent_state(struct rcu_state *rsp)
2074 unsigned long flags;
2076 struct rcu_node *rnp;
2077 struct rcu_node *rnp_old = NULL;
2079 /* Funnel through hierarchy to reduce memory contention. */
2080 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2081 for (; rnp != NULL; rnp = rnp->parent) {
2082 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2083 !raw_spin_trylock(&rnp->fqslock);
2084 if (rnp_old != NULL)
2085 raw_spin_unlock(&rnp_old->fqslock);
2087 rsp->n_force_qs_lh++;
2092 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2094 /* Reached the root of the rcu_node tree, acquire lock. */
2095 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2096 raw_spin_unlock(&rnp_old->fqslock);
2097 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2098 rsp->n_force_qs_lh++;
2099 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2100 return; /* Someone beat us to it. */
2102 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2103 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2104 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2108 * This does the RCU core processing work for the specified rcu_state
2109 * and rcu_data structures. This may be called only from the CPU to
2110 * whom the rdp belongs.
2113 __rcu_process_callbacks(struct rcu_state *rsp)
2115 unsigned long flags;
2116 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2118 WARN_ON_ONCE(rdp->beenonline == 0);
2120 /* Handle the end of a grace period that some other CPU ended. */
2121 rcu_process_gp_end(rsp, rdp);
2123 /* Update RCU state based on any recent quiescent states. */
2124 rcu_check_quiescent_state(rsp, rdp);
2126 /* Does this CPU require a not-yet-started grace period? */
2127 local_irq_save(flags);
2128 if (cpu_needs_another_gp(rsp, rdp)) {
2129 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2130 rcu_start_gp(rsp, flags); /* releases above lock */
2132 local_irq_restore(flags);
2135 /* If there are callbacks ready, invoke them. */
2136 if (cpu_has_callbacks_ready_to_invoke(rdp))
2137 invoke_rcu_callbacks(rsp, rdp);
2141 * Do RCU core processing for the current CPU.
2143 static void rcu_process_callbacks(struct softirq_action *unused)
2145 struct rcu_state *rsp;
2147 if (cpu_is_offline(smp_processor_id()))
2149 trace_rcu_utilization("Start RCU core");
2150 for_each_rcu_flavor(rsp)
2151 __rcu_process_callbacks(rsp);
2152 trace_rcu_utilization("End RCU core");
2156 * Schedule RCU callback invocation. If the specified type of RCU
2157 * does not support RCU priority boosting, just do a direct call,
2158 * otherwise wake up the per-CPU kernel kthread. Note that because we
2159 * are running on the current CPU with interrupts disabled, the
2160 * rcu_cpu_kthread_task cannot disappear out from under us.
2162 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2164 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2166 if (likely(!rsp->boost)) {
2167 rcu_do_batch(rsp, rdp);
2170 invoke_rcu_callbacks_kthread();
2173 static void invoke_rcu_core(void)
2175 raise_softirq(RCU_SOFTIRQ);
2179 * Handle any core-RCU processing required by a call_rcu() invocation.
2181 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2182 struct rcu_head *head, unsigned long flags)
2185 * If called from an extended quiescent state, invoke the RCU
2186 * core in order to force a re-evaluation of RCU's idleness.
2188 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2191 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2192 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2196 * Force the grace period if too many callbacks or too long waiting.
2197 * Enforce hysteresis, and don't invoke force_quiescent_state()
2198 * if some other CPU has recently done so. Also, don't bother
2199 * invoking force_quiescent_state() if the newly enqueued callback
2200 * is the only one waiting for a grace period to complete.
2202 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2204 /* Are we ignoring a completed grace period? */
2205 rcu_process_gp_end(rsp, rdp);
2206 check_for_new_grace_period(rsp, rdp);
2208 /* Start a new grace period if one not already started. */
2209 if (!rcu_gp_in_progress(rsp)) {
2210 unsigned long nestflag;
2211 struct rcu_node *rnp_root = rcu_get_root(rsp);
2213 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
2214 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
2216 /* Give the grace period a kick. */
2217 rdp->blimit = LONG_MAX;
2218 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2219 *rdp->nxttail[RCU_DONE_TAIL] != head)
2220 force_quiescent_state(rsp);
2221 rdp->n_force_qs_snap = rsp->n_force_qs;
2222 rdp->qlen_last_fqs_check = rdp->qlen;
2228 * Helper function for call_rcu() and friends. The cpu argument will
2229 * normally be -1, indicating "currently running CPU". It may specify
2230 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2231 * is expected to specify a CPU.
2234 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2235 struct rcu_state *rsp, int cpu, bool lazy)
2237 unsigned long flags;
2238 struct rcu_data *rdp;
2240 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2241 debug_rcu_head_queue(head);
2246 * Opportunistically note grace-period endings and beginnings.
2247 * Note that we might see a beginning right after we see an
2248 * end, but never vice versa, since this CPU has to pass through
2249 * a quiescent state betweentimes.
2251 local_irq_save(flags);
2252 rdp = this_cpu_ptr(rsp->rda);
2254 /* Add the callback to our list. */
2255 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2259 rdp = per_cpu_ptr(rsp->rda, cpu);
2260 offline = !__call_rcu_nocb(rdp, head, lazy);
2261 WARN_ON_ONCE(offline);
2262 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2263 local_irq_restore(flags);
2266 ACCESS_ONCE(rdp->qlen)++;
2270 rcu_idle_count_callbacks_posted();
2271 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2272 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2273 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2275 if (__is_kfree_rcu_offset((unsigned long)func))
2276 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2277 rdp->qlen_lazy, rdp->qlen);
2279 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2281 /* Go handle any RCU core processing required. */
2282 __call_rcu_core(rsp, rdp, head, flags);
2283 local_irq_restore(flags);
2287 * Queue an RCU-sched callback for invocation after a grace period.
2289 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2291 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2293 EXPORT_SYMBOL_GPL(call_rcu_sched);
2296 * Queue an RCU callback for invocation after a quicker grace period.
2298 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2300 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2302 EXPORT_SYMBOL_GPL(call_rcu_bh);
2305 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2306 * any blocking grace-period wait automatically implies a grace period
2307 * if there is only one CPU online at any point time during execution
2308 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2309 * occasionally incorrectly indicate that there are multiple CPUs online
2310 * when there was in fact only one the whole time, as this just adds
2311 * some overhead: RCU still operates correctly.
2313 static inline int rcu_blocking_is_gp(void)
2317 might_sleep(); /* Check for RCU read-side critical section. */
2319 ret = num_online_cpus() <= 1;
2325 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2327 * Control will return to the caller some time after a full rcu-sched
2328 * grace period has elapsed, in other words after all currently executing
2329 * rcu-sched read-side critical sections have completed. These read-side
2330 * critical sections are delimited by rcu_read_lock_sched() and
2331 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2332 * local_irq_disable(), and so on may be used in place of
2333 * rcu_read_lock_sched().
2335 * This means that all preempt_disable code sequences, including NMI and
2336 * non-threaded hardware-interrupt handlers, in progress on entry will
2337 * have completed before this primitive returns. However, this does not
2338 * guarantee that softirq handlers will have completed, since in some
2339 * kernels, these handlers can run in process context, and can block.
2341 * Note that this guarantee implies further memory-ordering guarantees.
2342 * On systems with more than one CPU, when synchronize_sched() returns,
2343 * each CPU is guaranteed to have executed a full memory barrier since the
2344 * end of its last RCU-sched read-side critical section whose beginning
2345 * preceded the call to synchronize_sched(). In addition, each CPU having
2346 * an RCU read-side critical section that extends beyond the return from
2347 * synchronize_sched() is guaranteed to have executed a full memory barrier
2348 * after the beginning of synchronize_sched() and before the beginning of
2349 * that RCU read-side critical section. Note that these guarantees include
2350 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2351 * that are executing in the kernel.
2353 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2354 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2355 * to have executed a full memory barrier during the execution of
2356 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2357 * again only if the system has more than one CPU).
2359 * This primitive provides the guarantees made by the (now removed)
2360 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2361 * guarantees that rcu_read_lock() sections will have completed.
2362 * In "classic RCU", these two guarantees happen to be one and
2363 * the same, but can differ in realtime RCU implementations.
2365 void synchronize_sched(void)
2367 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2368 !lock_is_held(&rcu_lock_map) &&
2369 !lock_is_held(&rcu_sched_lock_map),
2370 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2371 if (rcu_blocking_is_gp())
2374 synchronize_sched_expedited();
2376 wait_rcu_gp(call_rcu_sched);
2378 EXPORT_SYMBOL_GPL(synchronize_sched);
2381 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2383 * Control will return to the caller some time after a full rcu_bh grace
2384 * period has elapsed, in other words after all currently executing rcu_bh
2385 * read-side critical sections have completed. RCU read-side critical
2386 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2387 * and may be nested.
2389 * See the description of synchronize_sched() for more detailed information
2390 * on memory ordering guarantees.
2392 void synchronize_rcu_bh(void)
2394 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2395 !lock_is_held(&rcu_lock_map) &&
2396 !lock_is_held(&rcu_sched_lock_map),
2397 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2398 if (rcu_blocking_is_gp())
2401 synchronize_rcu_bh_expedited();
2403 wait_rcu_gp(call_rcu_bh);
2405 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2407 static int synchronize_sched_expedited_cpu_stop(void *data)
2410 * There must be a full memory barrier on each affected CPU
2411 * between the time that try_stop_cpus() is called and the
2412 * time that it returns.
2414 * In the current initial implementation of cpu_stop, the
2415 * above condition is already met when the control reaches
2416 * this point and the following smp_mb() is not strictly
2417 * necessary. Do smp_mb() anyway for documentation and
2418 * robustness against future implementation changes.
2420 smp_mb(); /* See above comment block. */
2425 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2427 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2428 * approach to force the grace period to end quickly. This consumes
2429 * significant time on all CPUs and is unfriendly to real-time workloads,
2430 * so is thus not recommended for any sort of common-case code. In fact,
2431 * if you are using synchronize_sched_expedited() in a loop, please
2432 * restructure your code to batch your updates, and then use a single
2433 * synchronize_sched() instead.
2435 * Note that it is illegal to call this function while holding any lock
2436 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2437 * to call this function from a CPU-hotplug notifier. Failing to observe
2438 * these restriction will result in deadlock.
2440 * This implementation can be thought of as an application of ticket
2441 * locking to RCU, with sync_sched_expedited_started and
2442 * sync_sched_expedited_done taking on the roles of the halves
2443 * of the ticket-lock word. Each task atomically increments
2444 * sync_sched_expedited_started upon entry, snapshotting the old value,
2445 * then attempts to stop all the CPUs. If this succeeds, then each
2446 * CPU will have executed a context switch, resulting in an RCU-sched
2447 * grace period. We are then done, so we use atomic_cmpxchg() to
2448 * update sync_sched_expedited_done to match our snapshot -- but
2449 * only if someone else has not already advanced past our snapshot.
2451 * On the other hand, if try_stop_cpus() fails, we check the value
2452 * of sync_sched_expedited_done. If it has advanced past our
2453 * initial snapshot, then someone else must have forced a grace period
2454 * some time after we took our snapshot. In this case, our work is
2455 * done for us, and we can simply return. Otherwise, we try again,
2456 * but keep our initial snapshot for purposes of checking for someone
2457 * doing our work for us.
2459 * If we fail too many times in a row, we fall back to synchronize_sched().
2461 void synchronize_sched_expedited(void)
2463 long firstsnap, s, snap;
2465 struct rcu_state *rsp = &rcu_sched_state;
2468 * If we are in danger of counter wrap, just do synchronize_sched().
2469 * By allowing sync_sched_expedited_started to advance no more than
2470 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2471 * that more than 3.5 billion CPUs would be required to force a
2472 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2473 * course be required on a 64-bit system.
2475 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2476 (ulong)atomic_long_read(&rsp->expedited_done) +
2478 synchronize_sched();
2479 atomic_long_inc(&rsp->expedited_wrap);
2484 * Take a ticket. Note that atomic_inc_return() implies a
2485 * full memory barrier.
2487 snap = atomic_long_inc_return(&rsp->expedited_start);
2490 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2493 * Each pass through the following loop attempts to force a
2494 * context switch on each CPU.
2496 while (try_stop_cpus(cpu_online_mask,
2497 synchronize_sched_expedited_cpu_stop,
2500 atomic_long_inc(&rsp->expedited_tryfail);
2502 /* Check to see if someone else did our work for us. */
2503 s = atomic_long_read(&rsp->expedited_done);
2504 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2505 /* ensure test happens before caller kfree */
2506 smp_mb__before_atomic_inc(); /* ^^^ */
2507 atomic_long_inc(&rsp->expedited_workdone1);
2511 /* No joy, try again later. Or just synchronize_sched(). */
2512 if (trycount++ < 10) {
2513 udelay(trycount * num_online_cpus());
2515 wait_rcu_gp(call_rcu_sched);
2516 atomic_long_inc(&rsp->expedited_normal);
2520 /* Recheck to see if someone else did our work for us. */
2521 s = atomic_long_read(&rsp->expedited_done);
2522 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2523 /* ensure test happens before caller kfree */
2524 smp_mb__before_atomic_inc(); /* ^^^ */
2525 atomic_long_inc(&rsp->expedited_workdone2);
2530 * Refetching sync_sched_expedited_started allows later
2531 * callers to piggyback on our grace period. We retry
2532 * after they started, so our grace period works for them,
2533 * and they started after our first try, so their grace
2534 * period works for us.
2537 snap = atomic_long_read(&rsp->expedited_start);
2538 smp_mb(); /* ensure read is before try_stop_cpus(). */
2540 atomic_long_inc(&rsp->expedited_stoppedcpus);
2543 * Everyone up to our most recent fetch is covered by our grace
2544 * period. Update the counter, but only if our work is still
2545 * relevant -- which it won't be if someone who started later
2546 * than we did already did their update.
2549 atomic_long_inc(&rsp->expedited_done_tries);
2550 s = atomic_long_read(&rsp->expedited_done);
2551 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2552 /* ensure test happens before caller kfree */
2553 smp_mb__before_atomic_inc(); /* ^^^ */
2554 atomic_long_inc(&rsp->expedited_done_lost);
2557 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2558 atomic_long_inc(&rsp->expedited_done_exit);
2562 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2565 * Check to see if there is any immediate RCU-related work to be done
2566 * by the current CPU, for the specified type of RCU, returning 1 if so.
2567 * The checks are in order of increasing expense: checks that can be
2568 * carried out against CPU-local state are performed first. However,
2569 * we must check for CPU stalls first, else we might not get a chance.
2571 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2573 struct rcu_node *rnp = rdp->mynode;
2575 rdp->n_rcu_pending++;
2577 /* Check for CPU stalls, if enabled. */
2578 check_cpu_stall(rsp, rdp);
2580 /* Is the RCU core waiting for a quiescent state from this CPU? */
2581 if (rcu_scheduler_fully_active &&
2582 rdp->qs_pending && !rdp->passed_quiesce) {
2583 rdp->n_rp_qs_pending++;
2584 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2585 rdp->n_rp_report_qs++;
2589 /* Does this CPU have callbacks ready to invoke? */
2590 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2591 rdp->n_rp_cb_ready++;
2595 /* Has RCU gone idle with this CPU needing another grace period? */
2596 if (cpu_needs_another_gp(rsp, rdp)) {
2597 rdp->n_rp_cpu_needs_gp++;
2601 /* Has another RCU grace period completed? */
2602 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2603 rdp->n_rp_gp_completed++;
2607 /* Has a new RCU grace period started? */
2608 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2609 rdp->n_rp_gp_started++;
2614 rdp->n_rp_need_nothing++;
2619 * Check to see if there is any immediate RCU-related work to be done
2620 * by the current CPU, returning 1 if so. This function is part of the
2621 * RCU implementation; it is -not- an exported member of the RCU API.
2623 static int rcu_pending(int cpu)
2625 struct rcu_state *rsp;
2627 for_each_rcu_flavor(rsp)
2628 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2634 * Check to see if any future RCU-related work will need to be done
2635 * by the current CPU, even if none need be done immediately, returning
2638 static int rcu_cpu_has_callbacks(int cpu)
2640 struct rcu_state *rsp;
2642 /* RCU callbacks either ready or pending? */
2643 for_each_rcu_flavor(rsp)
2644 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2650 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2651 * the compiler is expected to optimize this away.
2653 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2654 int cpu, unsigned long done)
2656 trace_rcu_barrier(rsp->name, s, cpu,
2657 atomic_read(&rsp->barrier_cpu_count), done);
2661 * RCU callback function for _rcu_barrier(). If we are last, wake
2662 * up the task executing _rcu_barrier().
2664 static void rcu_barrier_callback(struct rcu_head *rhp)
2666 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2667 struct rcu_state *rsp = rdp->rsp;
2669 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2670 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2671 complete(&rsp->barrier_completion);
2673 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2678 * Called with preemption disabled, and from cross-cpu IRQ context.
2680 static void rcu_barrier_func(void *type)
2682 struct rcu_state *rsp = type;
2683 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2685 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2686 atomic_inc(&rsp->barrier_cpu_count);
2687 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2691 * Orchestrate the specified type of RCU barrier, waiting for all
2692 * RCU callbacks of the specified type to complete.
2694 static void _rcu_barrier(struct rcu_state *rsp)
2697 struct rcu_data *rdp;
2698 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2699 unsigned long snap_done;
2701 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2703 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2704 mutex_lock(&rsp->barrier_mutex);
2707 * Ensure that all prior references, including to ->n_barrier_done,
2708 * are ordered before the _rcu_barrier() machinery.
2710 smp_mb(); /* See above block comment. */
2713 * Recheck ->n_barrier_done to see if others did our work for us.
2714 * This means checking ->n_barrier_done for an even-to-odd-to-even
2715 * transition. The "if" expression below therefore rounds the old
2716 * value up to the next even number and adds two before comparing.
2718 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2719 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2720 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2721 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2722 smp_mb(); /* caller's subsequent code after above check. */
2723 mutex_unlock(&rsp->barrier_mutex);
2728 * Increment ->n_barrier_done to avoid duplicate work. Use
2729 * ACCESS_ONCE() to prevent the compiler from speculating
2730 * the increment to precede the early-exit check.
2732 ACCESS_ONCE(rsp->n_barrier_done)++;
2733 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2734 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2735 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2738 * Initialize the count to one rather than to zero in order to
2739 * avoid a too-soon return to zero in case of a short grace period
2740 * (or preemption of this task). Exclude CPU-hotplug operations
2741 * to ensure that no offline CPU has callbacks queued.
2743 init_completion(&rsp->barrier_completion);
2744 atomic_set(&rsp->barrier_cpu_count, 1);
2748 * Force each CPU with callbacks to register a new callback.
2749 * When that callback is invoked, we will know that all of the
2750 * corresponding CPU's preceding callbacks have been invoked.
2752 for_each_possible_cpu(cpu) {
2753 if (!cpu_online(cpu) && !is_nocb_cpu(cpu))
2755 rdp = per_cpu_ptr(rsp->rda, cpu);
2756 if (is_nocb_cpu(cpu)) {
2757 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2758 rsp->n_barrier_done);
2759 atomic_inc(&rsp->barrier_cpu_count);
2760 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2762 } else if (ACCESS_ONCE(rdp->qlen)) {
2763 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2764 rsp->n_barrier_done);
2765 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2767 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2768 rsp->n_barrier_done);
2774 * Now that we have an rcu_barrier_callback() callback on each
2775 * CPU, and thus each counted, remove the initial count.
2777 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2778 complete(&rsp->barrier_completion);
2780 /* Increment ->n_barrier_done to prevent duplicate work. */
2781 smp_mb(); /* Keep increment after above mechanism. */
2782 ACCESS_ONCE(rsp->n_barrier_done)++;
2783 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2784 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2785 smp_mb(); /* Keep increment before caller's subsequent code. */
2787 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2788 wait_for_completion(&rsp->barrier_completion);
2790 /* Other rcu_barrier() invocations can now safely proceed. */
2791 mutex_unlock(&rsp->barrier_mutex);
2795 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2797 void rcu_barrier_bh(void)
2799 _rcu_barrier(&rcu_bh_state);
2801 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2804 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2806 void rcu_barrier_sched(void)
2808 _rcu_barrier(&rcu_sched_state);
2810 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2813 * Do boot-time initialization of a CPU's per-CPU RCU data.
2816 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2818 unsigned long flags;
2819 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2820 struct rcu_node *rnp = rcu_get_root(rsp);
2822 /* Set up local state, ensuring consistent view of global state. */
2823 raw_spin_lock_irqsave(&rnp->lock, flags);
2824 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2825 init_callback_list(rdp);
2827 ACCESS_ONCE(rdp->qlen) = 0;
2828 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2829 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2830 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2833 rcu_boot_init_nocb_percpu_data(rdp);
2834 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2838 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2839 * offline event can be happening at a given time. Note also that we
2840 * can accept some slop in the rsp->completed access due to the fact
2841 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2843 static void __cpuinit
2844 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2846 unsigned long flags;
2848 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2849 struct rcu_node *rnp = rcu_get_root(rsp);
2851 /* Exclude new grace periods. */
2852 mutex_lock(&rsp->onoff_mutex);
2854 /* Set up local state, ensuring consistent view of global state. */
2855 raw_spin_lock_irqsave(&rnp->lock, flags);
2856 rdp->beenonline = 1; /* We have now been online. */
2857 rdp->preemptible = preemptible;
2858 rdp->qlen_last_fqs_check = 0;
2859 rdp->n_force_qs_snap = rsp->n_force_qs;
2860 rdp->blimit = blimit;
2861 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2862 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2863 atomic_set(&rdp->dynticks->dynticks,
2864 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2865 rcu_prepare_for_idle_init(cpu);
2866 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2868 /* Add CPU to rcu_node bitmasks. */
2870 mask = rdp->grpmask;
2872 /* Exclude any attempts to start a new GP on small systems. */
2873 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2874 rnp->qsmaskinit |= mask;
2875 mask = rnp->grpmask;
2876 if (rnp == rdp->mynode) {
2878 * If there is a grace period in progress, we will
2879 * set up to wait for it next time we run the
2882 rdp->gpnum = rnp->completed;
2883 rdp->completed = rnp->completed;
2884 rdp->passed_quiesce = 0;
2885 rdp->qs_pending = 0;
2886 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2888 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2890 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2891 local_irq_restore(flags);
2893 mutex_unlock(&rsp->onoff_mutex);
2896 static void __cpuinit rcu_prepare_cpu(int cpu)
2898 struct rcu_state *rsp;
2900 for_each_rcu_flavor(rsp)
2901 rcu_init_percpu_data(cpu, rsp,
2902 strcmp(rsp->name, "rcu_preempt") == 0);
2906 * Handle CPU online/offline notification events.
2908 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2909 unsigned long action, void *hcpu)
2911 long cpu = (long)hcpu;
2912 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2913 struct rcu_node *rnp = rdp->mynode;
2914 struct rcu_state *rsp;
2916 trace_rcu_utilization("Start CPU hotplug");
2918 case CPU_UP_PREPARE:
2919 case CPU_UP_PREPARE_FROZEN:
2920 rcu_prepare_cpu(cpu);
2921 rcu_prepare_kthreads(cpu);
2924 case CPU_DOWN_FAILED:
2925 rcu_boost_kthread_setaffinity(rnp, -1);
2927 case CPU_DOWN_PREPARE:
2928 rcu_boost_kthread_setaffinity(rnp, cpu);
2931 case CPU_DYING_FROZEN:
2933 * The whole machine is "stopped" except this CPU, so we can
2934 * touch any data without introducing corruption. We send the
2935 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2937 for_each_rcu_flavor(rsp)
2938 rcu_cleanup_dying_cpu(rsp);
2939 rcu_cleanup_after_idle(cpu);
2942 case CPU_DEAD_FROZEN:
2943 case CPU_UP_CANCELED:
2944 case CPU_UP_CANCELED_FROZEN:
2945 for_each_rcu_flavor(rsp)
2946 rcu_cleanup_dead_cpu(cpu, rsp);
2951 trace_rcu_utilization("End CPU hotplug");
2956 * Spawn the kthread that handles this RCU flavor's grace periods.
2958 static int __init rcu_spawn_gp_kthread(void)
2960 unsigned long flags;
2961 struct rcu_node *rnp;
2962 struct rcu_state *rsp;
2963 struct task_struct *t;
2965 for_each_rcu_flavor(rsp) {
2966 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2968 rnp = rcu_get_root(rsp);
2969 raw_spin_lock_irqsave(&rnp->lock, flags);
2970 rsp->gp_kthread = t;
2971 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2972 rcu_spawn_nocb_kthreads(rsp);
2976 early_initcall(rcu_spawn_gp_kthread);
2979 * This function is invoked towards the end of the scheduler's initialization
2980 * process. Before this is called, the idle task might contain
2981 * RCU read-side critical sections (during which time, this idle
2982 * task is booting the system). After this function is called, the
2983 * idle tasks are prohibited from containing RCU read-side critical
2984 * sections. This function also enables RCU lockdep checking.
2986 void rcu_scheduler_starting(void)
2988 WARN_ON(num_online_cpus() != 1);
2989 WARN_ON(nr_context_switches() > 0);
2990 rcu_scheduler_active = 1;
2994 * Compute the per-level fanout, either using the exact fanout specified
2995 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2997 #ifdef CONFIG_RCU_FANOUT_EXACT
2998 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3002 for (i = rcu_num_lvls - 1; i > 0; i--)
3003 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3004 rsp->levelspread[0] = rcu_fanout_leaf;
3006 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3007 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3014 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3015 ccur = rsp->levelcnt[i];
3016 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3020 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3023 * Helper function for rcu_init() that initializes one rcu_state structure.
3025 static void __init rcu_init_one(struct rcu_state *rsp,
3026 struct rcu_data __percpu *rda)
3028 static char *buf[] = { "rcu_node_0",
3031 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3032 static char *fqs[] = { "rcu_node_fqs_0",
3035 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3039 struct rcu_node *rnp;
3041 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3043 /* Silence gcc 4.8 warning about array index out of range. */
3044 if (rcu_num_lvls > RCU_NUM_LVLS)
3045 panic("rcu_init_one: rcu_num_lvls overflow");
3047 /* Initialize the level-tracking arrays. */
3049 for (i = 0; i < rcu_num_lvls; i++)
3050 rsp->levelcnt[i] = num_rcu_lvl[i];
3051 for (i = 1; i < rcu_num_lvls; i++)
3052 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3053 rcu_init_levelspread(rsp);
3055 /* Initialize the elements themselves, starting from the leaves. */
3057 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3058 cpustride *= rsp->levelspread[i];
3059 rnp = rsp->level[i];
3060 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3061 raw_spin_lock_init(&rnp->lock);
3062 lockdep_set_class_and_name(&rnp->lock,
3063 &rcu_node_class[i], buf[i]);
3064 raw_spin_lock_init(&rnp->fqslock);
3065 lockdep_set_class_and_name(&rnp->fqslock,
3066 &rcu_fqs_class[i], fqs[i]);
3067 rnp->gpnum = rsp->gpnum;
3068 rnp->completed = rsp->completed;
3070 rnp->qsmaskinit = 0;
3071 rnp->grplo = j * cpustride;
3072 rnp->grphi = (j + 1) * cpustride - 1;
3073 if (rnp->grphi >= NR_CPUS)
3074 rnp->grphi = NR_CPUS - 1;
3080 rnp->grpnum = j % rsp->levelspread[i - 1];
3081 rnp->grpmask = 1UL << rnp->grpnum;
3082 rnp->parent = rsp->level[i - 1] +
3083 j / rsp->levelspread[i - 1];
3086 INIT_LIST_HEAD(&rnp->blkd_tasks);
3091 init_waitqueue_head(&rsp->gp_wq);
3092 rnp = rsp->level[rcu_num_lvls - 1];
3093 for_each_possible_cpu(i) {
3094 while (i > rnp->grphi)
3096 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3097 rcu_boot_init_percpu_data(i, rsp);
3099 list_add(&rsp->flavors, &rcu_struct_flavors);
3103 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3104 * replace the definitions in rcutree.h because those are needed to size
3105 * the ->node array in the rcu_state structure.
3107 static void __init rcu_init_geometry(void)
3112 int rcu_capacity[MAX_RCU_LVLS + 1];
3114 /* If the compile-time values are accurate, just leave. */
3115 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3116 nr_cpu_ids == NR_CPUS)
3120 * Compute number of nodes that can be handled an rcu_node tree
3121 * with the given number of levels. Setting rcu_capacity[0] makes
3122 * some of the arithmetic easier.
3124 rcu_capacity[0] = 1;
3125 rcu_capacity[1] = rcu_fanout_leaf;
3126 for (i = 2; i <= MAX_RCU_LVLS; i++)
3127 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3130 * The boot-time rcu_fanout_leaf parameter is only permitted
3131 * to increase the leaf-level fanout, not decrease it. Of course,
3132 * the leaf-level fanout cannot exceed the number of bits in
3133 * the rcu_node masks. Finally, the tree must be able to accommodate
3134 * the configured number of CPUs. Complain and fall back to the
3135 * compile-time values if these limits are exceeded.
3137 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3138 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3139 n > rcu_capacity[MAX_RCU_LVLS]) {
3144 /* Calculate the number of rcu_nodes at each level of the tree. */
3145 for (i = 1; i <= MAX_RCU_LVLS; i++)
3146 if (n <= rcu_capacity[i]) {
3147 for (j = 0; j <= i; j++)
3149 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3151 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3156 /* Calculate the total number of rcu_node structures. */
3158 for (i = 0; i <= MAX_RCU_LVLS; i++)
3159 rcu_num_nodes += num_rcu_lvl[i];
3163 void __init rcu_init(void)
3167 rcu_bootup_announce();
3168 rcu_init_geometry();
3169 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3170 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3171 __rcu_init_preempt();
3172 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3175 * We don't need protection against CPU-hotplug here because
3176 * this is called early in boot, before either interrupts
3177 * or the scheduler are operational.
3179 cpu_notifier(rcu_cpu_notify, 0);
3180 for_each_online_cpu(cpu)
3181 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3184 #include "rcutree_plugin.h"