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, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
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/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
82 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
83 static char sname##_varname[] = #sname; \
84 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
85 struct rcu_state sname##_state = { \
86 .level = { &sname##_state.node[0] }, \
88 .fqs_state = RCU_GP_IDLE, \
89 .gpnum = 0UL - 300UL, \
90 .completed = 0UL - 300UL, \
91 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
92 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
93 .orphan_donetail = &sname##_state.orphan_donelist, \
94 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
95 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
96 .name = sname##_varname, \
99 DEFINE_PER_CPU(struct rcu_data, sname##_data)
101 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
102 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
104 static struct rcu_state *rcu_state;
105 LIST_HEAD(rcu_struct_flavors);
107 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
108 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
109 module_param(rcu_fanout_leaf, int, 0444);
110 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
111 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
118 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
121 * The rcu_scheduler_active variable transitions from zero to one just
122 * before the first task is spawned. So when this variable is zero, RCU
123 * can assume that there is but one task, allowing RCU to (for example)
124 * optimize synchronize_sched() to a simple barrier(). When this variable
125 * is one, RCU must actually do all the hard work required to detect real
126 * grace periods. This variable is also used to suppress boot-time false
127 * positives from lockdep-RCU error checking.
129 int rcu_scheduler_active __read_mostly;
130 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
144 static int rcu_scheduler_fully_active __read_mostly;
146 #ifdef CONFIG_RCU_BOOST
149 * Control variables for per-CPU and per-rcu_node kthreads. These
150 * handle all flavors of RCU.
152 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
153 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
154 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
155 DEFINE_PER_CPU(char, rcu_cpu_has_work);
157 #endif /* #ifdef CONFIG_RCU_BOOST */
159 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
160 static void invoke_rcu_core(void);
161 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
172 unsigned long rcutorture_testseq;
173 unsigned long rcutorture_vernum;
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
180 static int rcu_gp_in_progress(struct rcu_state *rsp)
182 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
191 void rcu_sched_qs(int cpu)
193 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
195 if (rdp->passed_quiesce == 0)
196 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
197 rdp->passed_quiesce = 1;
200 void rcu_bh_qs(int cpu)
202 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
204 if (rdp->passed_quiesce == 0)
205 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
206 rdp->passed_quiesce = 1;
210 * Note a context switch. This is a quiescent state for RCU-sched,
211 * and requires special handling for preemptible RCU.
212 * The caller must have disabled preemption.
214 void rcu_note_context_switch(int cpu)
216 trace_rcu_utilization(TPS("Start context switch"));
218 rcu_preempt_note_context_switch(cpu);
219 trace_rcu_utilization(TPS("End context switch"));
221 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
223 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
224 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
225 .dynticks = ATOMIC_INIT(1),
226 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
227 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
228 .dynticks_idle = ATOMIC_INIT(1),
229 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
232 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
233 static long qhimark = 10000; /* If this many pending, ignore blimit. */
234 static long qlowmark = 100; /* Once only this many pending, use blimit. */
236 module_param(blimit, long, 0444);
237 module_param(qhimark, long, 0444);
238 module_param(qlowmark, long, 0444);
240 static ulong jiffies_till_first_fqs = ULONG_MAX;
241 static ulong jiffies_till_next_fqs = ULONG_MAX;
243 module_param(jiffies_till_first_fqs, ulong, 0644);
244 module_param(jiffies_till_next_fqs, ulong, 0644);
246 static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
247 struct rcu_data *rdp);
248 static void force_qs_rnp(struct rcu_state *rsp,
249 int (*f)(struct rcu_data *rsp, bool *isidle,
250 unsigned long *maxj),
251 bool *isidle, unsigned long *maxj);
252 static void force_quiescent_state(struct rcu_state *rsp);
253 static int rcu_pending(int cpu);
256 * Return the number of RCU-sched batches processed thus far for debug & stats.
258 long rcu_batches_completed_sched(void)
260 return rcu_sched_state.completed;
262 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
265 * Return the number of RCU BH batches processed thus far for debug & stats.
267 long rcu_batches_completed_bh(void)
269 return rcu_bh_state.completed;
271 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
274 * Force a quiescent state for RCU BH.
276 void rcu_bh_force_quiescent_state(void)
278 force_quiescent_state(&rcu_bh_state);
280 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
283 * Record the number of times rcutorture tests have been initiated and
284 * terminated. This information allows the debugfs tracing stats to be
285 * correlated to the rcutorture messages, even when the rcutorture module
286 * is being repeatedly loaded and unloaded. In other words, we cannot
287 * store this state in rcutorture itself.
289 void rcutorture_record_test_transition(void)
291 rcutorture_testseq++;
292 rcutorture_vernum = 0;
294 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
297 * Record the number of writer passes through the current rcutorture test.
298 * This is also used to correlate debugfs tracing stats with the rcutorture
301 void rcutorture_record_progress(unsigned long vernum)
305 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
308 * Force a quiescent state for RCU-sched.
310 void rcu_sched_force_quiescent_state(void)
312 force_quiescent_state(&rcu_sched_state);
314 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
317 * Does the CPU have callbacks ready to be invoked?
320 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
322 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
323 rdp->nxttail[RCU_DONE_TAIL] != NULL;
327 * Does the current CPU require a not-yet-started grace period?
328 * The caller must have disabled interrupts to prevent races with
329 * normal callback registry.
332 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
336 if (rcu_gp_in_progress(rsp))
337 return 0; /* No, a grace period is already in progress. */
338 if (rcu_nocb_needs_gp(rsp))
339 return 1; /* Yes, a no-CBs CPU needs one. */
340 if (!rdp->nxttail[RCU_NEXT_TAIL])
341 return 0; /* No, this is a no-CBs (or offline) CPU. */
342 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
343 return 1; /* Yes, this CPU has newly registered callbacks. */
344 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
345 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
346 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
347 rdp->nxtcompleted[i]))
348 return 1; /* Yes, CBs for future grace period. */
349 return 0; /* No grace period needed. */
353 * Return the root node of the specified rcu_state structure.
355 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
357 return &rsp->node[0];
361 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
363 * If the new value of the ->dynticks_nesting counter now is zero,
364 * we really have entered idle, and must do the appropriate accounting.
365 * The caller must have disabled interrupts.
367 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
370 struct rcu_state *rsp;
371 struct rcu_data *rdp;
373 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
374 if (!user && !is_idle_task(current)) {
375 struct task_struct *idle __maybe_unused =
376 idle_task(smp_processor_id());
378 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
379 ftrace_dump(DUMP_ORIG);
380 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
381 current->pid, current->comm,
382 idle->pid, idle->comm); /* must be idle task! */
384 for_each_rcu_flavor(rsp) {
385 rdp = this_cpu_ptr(rsp->rda);
386 do_nocb_deferred_wakeup(rdp);
388 rcu_prepare_for_idle(smp_processor_id());
389 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
390 smp_mb__before_atomic_inc(); /* See above. */
391 atomic_inc(&rdtp->dynticks);
392 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
393 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
396 * It is illegal to enter an extended quiescent state while
397 * in an RCU read-side critical section.
399 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
400 "Illegal idle entry in RCU read-side critical section.");
401 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
402 "Illegal idle entry in RCU-bh read-side critical section.");
403 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
404 "Illegal idle entry in RCU-sched read-side critical section.");
408 * Enter an RCU extended quiescent state, which can be either the
409 * idle loop or adaptive-tickless usermode execution.
411 static void rcu_eqs_enter(bool user)
414 struct rcu_dynticks *rdtp;
416 rdtp = this_cpu_ptr(&rcu_dynticks);
417 oldval = rdtp->dynticks_nesting;
418 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
419 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
420 rdtp->dynticks_nesting = 0;
421 rcu_eqs_enter_common(rdtp, oldval, user);
423 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
428 * rcu_idle_enter - inform RCU that current CPU is entering idle
430 * Enter idle mode, in other words, -leave- the mode in which RCU
431 * read-side critical sections can occur. (Though RCU read-side
432 * critical sections can occur in irq handlers in idle, a possibility
433 * handled by irq_enter() and irq_exit().)
435 * We crowbar the ->dynticks_nesting field to zero to allow for
436 * the possibility of usermode upcalls having messed up our count
437 * of interrupt nesting level during the prior busy period.
439 void rcu_idle_enter(void)
443 local_irq_save(flags);
444 rcu_eqs_enter(false);
445 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
446 local_irq_restore(flags);
448 EXPORT_SYMBOL_GPL(rcu_idle_enter);
450 #ifdef CONFIG_RCU_USER_QS
452 * rcu_user_enter - inform RCU that we are resuming userspace.
454 * Enter RCU idle mode right before resuming userspace. No use of RCU
455 * is permitted between this call and rcu_user_exit(). This way the
456 * CPU doesn't need to maintain the tick for RCU maintenance purposes
457 * when the CPU runs in userspace.
459 void rcu_user_enter(void)
463 #endif /* CONFIG_RCU_USER_QS */
466 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
468 * Exit from an interrupt handler, which might possibly result in entering
469 * idle mode, in other words, leaving the mode in which read-side critical
470 * sections can occur.
472 * This code assumes that the idle loop never does anything that might
473 * result in unbalanced calls to irq_enter() and irq_exit(). If your
474 * architecture violates this assumption, RCU will give you what you
475 * deserve, good and hard. But very infrequently and irreproducibly.
477 * Use things like work queues to work around this limitation.
479 * You have been warned.
481 void rcu_irq_exit(void)
485 struct rcu_dynticks *rdtp;
487 local_irq_save(flags);
488 rdtp = this_cpu_ptr(&rcu_dynticks);
489 oldval = rdtp->dynticks_nesting;
490 rdtp->dynticks_nesting--;
491 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
492 if (rdtp->dynticks_nesting)
493 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
495 rcu_eqs_enter_common(rdtp, oldval, true);
496 rcu_sysidle_enter(rdtp, 1);
497 local_irq_restore(flags);
501 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
503 * If the new value of the ->dynticks_nesting counter was previously zero,
504 * we really have exited idle, and must do the appropriate accounting.
505 * The caller must have disabled interrupts.
507 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
510 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
511 atomic_inc(&rdtp->dynticks);
512 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
513 smp_mb__after_atomic_inc(); /* See above. */
514 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
515 rcu_cleanup_after_idle(smp_processor_id());
516 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
517 if (!user && !is_idle_task(current)) {
518 struct task_struct *idle __maybe_unused =
519 idle_task(smp_processor_id());
521 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
522 oldval, rdtp->dynticks_nesting);
523 ftrace_dump(DUMP_ORIG);
524 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
525 current->pid, current->comm,
526 idle->pid, idle->comm); /* must be idle task! */
531 * Exit an RCU extended quiescent state, which can be either the
532 * idle loop or adaptive-tickless usermode execution.
534 static void rcu_eqs_exit(bool user)
536 struct rcu_dynticks *rdtp;
539 rdtp = this_cpu_ptr(&rcu_dynticks);
540 oldval = rdtp->dynticks_nesting;
541 WARN_ON_ONCE(oldval < 0);
542 if (oldval & DYNTICK_TASK_NEST_MASK) {
543 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
545 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
546 rcu_eqs_exit_common(rdtp, oldval, user);
551 * rcu_idle_exit - inform RCU that current CPU is leaving idle
553 * Exit idle mode, in other words, -enter- the mode in which RCU
554 * read-side critical sections can occur.
556 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
557 * allow for the possibility of usermode upcalls messing up our count
558 * of interrupt nesting level during the busy period that is just
561 void rcu_idle_exit(void)
565 local_irq_save(flags);
567 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
568 local_irq_restore(flags);
570 EXPORT_SYMBOL_GPL(rcu_idle_exit);
572 #ifdef CONFIG_RCU_USER_QS
574 * rcu_user_exit - inform RCU that we are exiting userspace.
576 * Exit RCU idle mode while entering the kernel because it can
577 * run a RCU read side critical section anytime.
579 void rcu_user_exit(void)
583 #endif /* CONFIG_RCU_USER_QS */
586 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
588 * Enter an interrupt handler, which might possibly result in exiting
589 * idle mode, in other words, entering the mode in which read-side critical
590 * sections can occur.
592 * Note that the Linux kernel is fully capable of entering an interrupt
593 * handler that it never exits, for example when doing upcalls to
594 * user mode! This code assumes that the idle loop never does upcalls to
595 * user mode. If your architecture does do upcalls from the idle loop (or
596 * does anything else that results in unbalanced calls to the irq_enter()
597 * and irq_exit() functions), RCU will give you what you deserve, good
598 * and hard. But very infrequently and irreproducibly.
600 * Use things like work queues to work around this limitation.
602 * You have been warned.
604 void rcu_irq_enter(void)
607 struct rcu_dynticks *rdtp;
610 local_irq_save(flags);
611 rdtp = this_cpu_ptr(&rcu_dynticks);
612 oldval = rdtp->dynticks_nesting;
613 rdtp->dynticks_nesting++;
614 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
616 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
618 rcu_eqs_exit_common(rdtp, oldval, true);
619 rcu_sysidle_exit(rdtp, 1);
620 local_irq_restore(flags);
624 * rcu_nmi_enter - inform RCU of entry to NMI context
626 * If the CPU was idle with dynamic ticks active, and there is no
627 * irq handler running, this updates rdtp->dynticks_nmi to let the
628 * RCU grace-period handling know that the CPU is active.
630 void rcu_nmi_enter(void)
632 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
634 if (rdtp->dynticks_nmi_nesting == 0 &&
635 (atomic_read(&rdtp->dynticks) & 0x1))
637 rdtp->dynticks_nmi_nesting++;
638 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
639 atomic_inc(&rdtp->dynticks);
640 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
641 smp_mb__after_atomic_inc(); /* See above. */
642 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
646 * rcu_nmi_exit - inform RCU of exit from NMI context
648 * If the CPU was idle with dynamic ticks active, and there is no
649 * irq handler running, this updates rdtp->dynticks_nmi to let the
650 * RCU grace-period handling know that the CPU is no longer active.
652 void rcu_nmi_exit(void)
654 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
656 if (rdtp->dynticks_nmi_nesting == 0 ||
657 --rdtp->dynticks_nmi_nesting != 0)
659 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
660 smp_mb__before_atomic_inc(); /* See above. */
661 atomic_inc(&rdtp->dynticks);
662 smp_mb__after_atomic_inc(); /* Force delay to next write. */
663 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
667 * __rcu_is_watching - are RCU read-side critical sections safe?
669 * Return true if RCU is watching the running CPU, which means that
670 * this CPU can safely enter RCU read-side critical sections. Unlike
671 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
672 * least disabled preemption.
674 bool notrace __rcu_is_watching(void)
676 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
680 * rcu_is_watching - see if RCU thinks that the current CPU is idle
682 * If the current CPU is in its idle loop and is neither in an interrupt
683 * or NMI handler, return true.
685 bool notrace rcu_is_watching(void)
690 ret = __rcu_is_watching();
694 EXPORT_SYMBOL_GPL(rcu_is_watching);
696 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
699 * Is the current CPU online? Disable preemption to avoid false positives
700 * that could otherwise happen due to the current CPU number being sampled,
701 * this task being preempted, its old CPU being taken offline, resuming
702 * on some other CPU, then determining that its old CPU is now offline.
703 * It is OK to use RCU on an offline processor during initial boot, hence
704 * the check for rcu_scheduler_fully_active. Note also that it is OK
705 * for a CPU coming online to use RCU for one jiffy prior to marking itself
706 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
707 * offline to continue to use RCU for one jiffy after marking itself
708 * offline in the cpu_online_mask. This leniency is necessary given the
709 * non-atomic nature of the online and offline processing, for example,
710 * the fact that a CPU enters the scheduler after completing the CPU_DYING
713 * This is also why RCU internally marks CPUs online during the
714 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
716 * Disable checking if in an NMI handler because we cannot safely report
717 * errors from NMI handlers anyway.
719 bool rcu_lockdep_current_cpu_online(void)
721 struct rcu_data *rdp;
722 struct rcu_node *rnp;
728 rdp = this_cpu_ptr(&rcu_sched_data);
730 ret = (rdp->grpmask & rnp->qsmaskinit) ||
731 !rcu_scheduler_fully_active;
735 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
737 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
740 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
742 * If the current CPU is idle or running at a first-level (not nested)
743 * interrupt from idle, return true. The caller must have at least
744 * disabled preemption.
746 static int rcu_is_cpu_rrupt_from_idle(void)
748 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
752 * Snapshot the specified CPU's dynticks counter so that we can later
753 * credit them with an implicit quiescent state. Return 1 if this CPU
754 * is in dynticks idle mode, which is an extended quiescent state.
756 static int dyntick_save_progress_counter(struct rcu_data *rdp,
757 bool *isidle, unsigned long *maxj)
759 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
760 rcu_sysidle_check_cpu(rdp, isidle, maxj);
761 return (rdp->dynticks_snap & 0x1) == 0;
765 * This function really isn't for public consumption, but RCU is special in
766 * that context switches can allow the state machine to make progress.
768 extern void resched_cpu(int cpu);
771 * Return true if the specified CPU has passed through a quiescent
772 * state by virtue of being in or having passed through an dynticks
773 * idle state since the last call to dyntick_save_progress_counter()
774 * for this same CPU, or by virtue of having been offline.
776 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
777 bool *isidle, unsigned long *maxj)
782 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
783 snap = (unsigned int)rdp->dynticks_snap;
786 * If the CPU passed through or entered a dynticks idle phase with
787 * no active irq/NMI handlers, then we can safely pretend that the CPU
788 * already acknowledged the request to pass through a quiescent
789 * state. Either way, that CPU cannot possibly be in an RCU
790 * read-side critical section that started before the beginning
791 * of the current RCU grace period.
793 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
794 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
800 * Check for the CPU being offline, but only if the grace period
801 * is old enough. We don't need to worry about the CPU changing
802 * state: If we see it offline even once, it has been through a
805 * The reason for insisting that the grace period be at least
806 * one jiffy old is that CPUs that are not quite online and that
807 * have just gone offline can still execute RCU read-side critical
810 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
811 return 0; /* Grace period is not old enough. */
813 if (cpu_is_offline(rdp->cpu)) {
814 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
820 * There is a possibility that a CPU in adaptive-ticks state
821 * might run in the kernel with the scheduling-clock tick disabled
822 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
823 * force the CPU to restart the scheduling-clock tick in this
824 * CPU is in this state.
826 rcu_kick_nohz_cpu(rdp->cpu);
829 * Alternatively, the CPU might be running in the kernel
830 * for an extended period of time without a quiescent state.
831 * Attempt to force the CPU through the scheduler to gain the
832 * needed quiescent state, but only if the grace period has gone
833 * on for an uncommonly long time. If there are many stuck CPUs,
834 * we will beat on the first one until it gets unstuck, then move
835 * to the next. Only do this for the primary flavor of RCU.
837 if (rdp->rsp == rcu_state &&
838 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
839 rdp->rsp->jiffies_resched += 5;
840 resched_cpu(rdp->cpu);
846 static void record_gp_stall_check_time(struct rcu_state *rsp)
848 unsigned long j = jiffies;
852 smp_wmb(); /* Record start time before stall time. */
853 j1 = rcu_jiffies_till_stall_check();
854 rsp->jiffies_stall = j + j1;
855 rsp->jiffies_resched = j + j1 / 2;
859 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
860 * for architectures that do not implement trigger_all_cpu_backtrace().
861 * The NMI-triggered stack traces are more accurate because they are
862 * printed by the target CPU.
864 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
868 struct rcu_node *rnp;
870 rcu_for_each_leaf_node(rsp, rnp) {
871 raw_spin_lock_irqsave(&rnp->lock, flags);
872 if (rnp->qsmask != 0) {
873 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
874 if (rnp->qsmask & (1UL << cpu))
875 dump_cpu_task(rnp->grplo + cpu);
877 raw_spin_unlock_irqrestore(&rnp->lock, flags);
881 static void print_other_cpu_stall(struct rcu_state *rsp)
887 struct rcu_node *rnp = rcu_get_root(rsp);
890 /* Only let one CPU complain about others per time interval. */
892 raw_spin_lock_irqsave(&rnp->lock, flags);
893 delta = jiffies - rsp->jiffies_stall;
894 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
895 raw_spin_unlock_irqrestore(&rnp->lock, flags);
898 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
899 raw_spin_unlock_irqrestore(&rnp->lock, flags);
902 * OK, time to rat on our buddy...
903 * See Documentation/RCU/stallwarn.txt for info on how to debug
904 * RCU CPU stall warnings.
906 pr_err("INFO: %s detected stalls on CPUs/tasks:",
908 print_cpu_stall_info_begin();
909 rcu_for_each_leaf_node(rsp, rnp) {
910 raw_spin_lock_irqsave(&rnp->lock, flags);
911 ndetected += rcu_print_task_stall(rnp);
912 if (rnp->qsmask != 0) {
913 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
914 if (rnp->qsmask & (1UL << cpu)) {
915 print_cpu_stall_info(rsp,
920 raw_spin_unlock_irqrestore(&rnp->lock, flags);
924 * Now rat on any tasks that got kicked up to the root rcu_node
925 * due to CPU offlining.
927 rnp = rcu_get_root(rsp);
928 raw_spin_lock_irqsave(&rnp->lock, flags);
929 ndetected += rcu_print_task_stall(rnp);
930 raw_spin_unlock_irqrestore(&rnp->lock, flags);
932 print_cpu_stall_info_end();
933 for_each_possible_cpu(cpu)
934 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
935 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
936 smp_processor_id(), (long)(jiffies - rsp->gp_start),
937 (long)rsp->gpnum, (long)rsp->completed, totqlen);
939 pr_err("INFO: Stall ended before state dump start\n");
940 else if (!trigger_all_cpu_backtrace())
941 rcu_dump_cpu_stacks(rsp);
943 /* Complain about tasks blocking the grace period. */
945 rcu_print_detail_task_stall(rsp);
947 force_quiescent_state(rsp); /* Kick them all. */
951 * This function really isn't for public consumption, but RCU is special in
952 * that context switches can allow the state machine to make progress.
954 extern void resched_cpu(int cpu);
956 static void print_cpu_stall(struct rcu_state *rsp)
960 struct rcu_node *rnp = rcu_get_root(rsp);
964 * OK, time to rat on ourselves...
965 * See Documentation/RCU/stallwarn.txt for info on how to debug
966 * RCU CPU stall warnings.
968 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
969 print_cpu_stall_info_begin();
970 print_cpu_stall_info(rsp, smp_processor_id());
971 print_cpu_stall_info_end();
972 for_each_possible_cpu(cpu)
973 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
974 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
975 jiffies - rsp->gp_start,
976 (long)rsp->gpnum, (long)rsp->completed, totqlen);
977 if (!trigger_all_cpu_backtrace())
980 raw_spin_lock_irqsave(&rnp->lock, flags);
981 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
982 rsp->jiffies_stall = jiffies +
983 3 * rcu_jiffies_till_stall_check() + 3;
984 raw_spin_unlock_irqrestore(&rnp->lock, flags);
987 * Attempt to revive the RCU machinery by forcing a context switch.
989 * A context switch would normally allow the RCU state machine to make
990 * progress and it could be we're stuck in kernel space without context
991 * switches for an entirely unreasonable amount of time.
993 resched_cpu(smp_processor_id());
996 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
998 unsigned long completed;
1003 struct rcu_node *rnp;
1005 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1010 * Lots of memory barriers to reject false positives.
1012 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1013 * then rsp->gp_start, and finally rsp->completed. These values
1014 * are updated in the opposite order with memory barriers (or
1015 * equivalent) during grace-period initialization and cleanup.
1016 * Now, a false positive can occur if we get an new value of
1017 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1018 * the memory barriers, the only way that this can happen is if one
1019 * grace period ends and another starts between these two fetches.
1020 * Detect this by comparing rsp->completed with the previous fetch
1023 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1024 * and rsp->gp_start suffice to forestall false positives.
1026 gpnum = ACCESS_ONCE(rsp->gpnum);
1027 smp_rmb(); /* Pick up ->gpnum first... */
1028 js = ACCESS_ONCE(rsp->jiffies_stall);
1029 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1030 gps = ACCESS_ONCE(rsp->gp_start);
1031 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1032 completed = ACCESS_ONCE(rsp->completed);
1033 if (ULONG_CMP_GE(completed, gpnum) ||
1034 ULONG_CMP_LT(j, js) ||
1035 ULONG_CMP_GE(gps, js))
1036 return; /* No stall or GP completed since entering function. */
1038 if (rcu_gp_in_progress(rsp) &&
1039 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1041 /* We haven't checked in, so go dump stack. */
1042 print_cpu_stall(rsp);
1044 } else if (rcu_gp_in_progress(rsp) &&
1045 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1047 /* They had a few time units to dump stack, so complain. */
1048 print_other_cpu_stall(rsp);
1053 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1055 * Set the stall-warning timeout way off into the future, thus preventing
1056 * any RCU CPU stall-warning messages from appearing in the current set of
1057 * RCU grace periods.
1059 * The caller must disable hard irqs.
1061 void rcu_cpu_stall_reset(void)
1063 struct rcu_state *rsp;
1065 for_each_rcu_flavor(rsp)
1066 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1070 * Initialize the specified rcu_data structure's callback list to empty.
1072 static void init_callback_list(struct rcu_data *rdp)
1076 if (init_nocb_callback_list(rdp))
1078 rdp->nxtlist = NULL;
1079 for (i = 0; i < RCU_NEXT_SIZE; i++)
1080 rdp->nxttail[i] = &rdp->nxtlist;
1084 * Determine the value that ->completed will have at the end of the
1085 * next subsequent grace period. This is used to tag callbacks so that
1086 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1087 * been dyntick-idle for an extended period with callbacks under the
1088 * influence of RCU_FAST_NO_HZ.
1090 * The caller must hold rnp->lock with interrupts disabled.
1092 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1093 struct rcu_node *rnp)
1096 * If RCU is idle, we just wait for the next grace period.
1097 * But we can only be sure that RCU is idle if we are looking
1098 * at the root rcu_node structure -- otherwise, a new grace
1099 * period might have started, but just not yet gotten around
1100 * to initializing the current non-root rcu_node structure.
1102 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1103 return rnp->completed + 1;
1106 * Otherwise, wait for a possible partial grace period and
1107 * then the subsequent full grace period.
1109 return rnp->completed + 2;
1113 * Trace-event helper function for rcu_start_future_gp() and
1114 * rcu_nocb_wait_gp().
1116 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1117 unsigned long c, const char *s)
1119 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1120 rnp->completed, c, rnp->level,
1121 rnp->grplo, rnp->grphi, s);
1125 * Start some future grace period, as needed to handle newly arrived
1126 * callbacks. The required future grace periods are recorded in each
1127 * rcu_node structure's ->need_future_gp field.
1129 * The caller must hold the specified rcu_node structure's ->lock.
1131 static unsigned long __maybe_unused
1132 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1136 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1139 * Pick up grace-period number for new callbacks. If this
1140 * grace period is already marked as needed, return to the caller.
1142 c = rcu_cbs_completed(rdp->rsp, rnp);
1143 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1144 if (rnp->need_future_gp[c & 0x1]) {
1145 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1150 * If either this rcu_node structure or the root rcu_node structure
1151 * believe that a grace period is in progress, then we must wait
1152 * for the one following, which is in "c". Because our request
1153 * will be noticed at the end of the current grace period, we don't
1154 * need to explicitly start one.
1156 if (rnp->gpnum != rnp->completed ||
1157 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1158 rnp->need_future_gp[c & 0x1]++;
1159 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1164 * There might be no grace period in progress. If we don't already
1165 * hold it, acquire the root rcu_node structure's lock in order to
1166 * start one (if needed).
1168 if (rnp != rnp_root) {
1169 raw_spin_lock(&rnp_root->lock);
1170 smp_mb__after_unlock_lock();
1174 * Get a new grace-period number. If there really is no grace
1175 * period in progress, it will be smaller than the one we obtained
1176 * earlier. Adjust callbacks as needed. Note that even no-CBs
1177 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1179 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1180 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1181 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1182 rdp->nxtcompleted[i] = c;
1185 * If the needed for the required grace period is already
1186 * recorded, trace and leave.
1188 if (rnp_root->need_future_gp[c & 0x1]) {
1189 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1193 /* Record the need for the future grace period. */
1194 rnp_root->need_future_gp[c & 0x1]++;
1196 /* If a grace period is not already in progress, start one. */
1197 if (rnp_root->gpnum != rnp_root->completed) {
1198 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1200 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1201 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1204 if (rnp != rnp_root)
1205 raw_spin_unlock(&rnp_root->lock);
1210 * Clean up any old requests for the just-ended grace period. Also return
1211 * whether any additional grace periods have been requested. Also invoke
1212 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1213 * waiting for this grace period to complete.
1215 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1217 int c = rnp->completed;
1219 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1221 rcu_nocb_gp_cleanup(rsp, rnp);
1222 rnp->need_future_gp[c & 0x1] = 0;
1223 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1224 trace_rcu_future_gp(rnp, rdp, c,
1225 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1230 * If there is room, assign a ->completed number to any callbacks on
1231 * this CPU that have not already been assigned. Also accelerate any
1232 * callbacks that were previously assigned a ->completed number that has
1233 * since proven to be too conservative, which can happen if callbacks get
1234 * assigned a ->completed number while RCU is idle, but with reference to
1235 * a non-root rcu_node structure. This function is idempotent, so it does
1236 * not hurt to call it repeatedly.
1238 * The caller must hold rnp->lock with interrupts disabled.
1240 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1241 struct rcu_data *rdp)
1246 /* If the CPU has no callbacks, nothing to do. */
1247 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1251 * Starting from the sublist containing the callbacks most
1252 * recently assigned a ->completed number and working down, find the
1253 * first sublist that is not assignable to an upcoming grace period.
1254 * Such a sublist has something in it (first two tests) and has
1255 * a ->completed number assigned that will complete sooner than
1256 * the ->completed number for newly arrived callbacks (last test).
1258 * The key point is that any later sublist can be assigned the
1259 * same ->completed number as the newly arrived callbacks, which
1260 * means that the callbacks in any of these later sublist can be
1261 * grouped into a single sublist, whether or not they have already
1262 * been assigned a ->completed number.
1264 c = rcu_cbs_completed(rsp, rnp);
1265 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1266 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1267 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1271 * If there are no sublist for unassigned callbacks, leave.
1272 * At the same time, advance "i" one sublist, so that "i" will
1273 * index into the sublist where all the remaining callbacks should
1276 if (++i >= RCU_NEXT_TAIL)
1280 * Assign all subsequent callbacks' ->completed number to the next
1281 * full grace period and group them all in the sublist initially
1284 for (; i <= RCU_NEXT_TAIL; i++) {
1285 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1286 rdp->nxtcompleted[i] = c;
1288 /* Record any needed additional grace periods. */
1289 rcu_start_future_gp(rnp, rdp);
1291 /* Trace depending on how much we were able to accelerate. */
1292 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1293 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1295 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1299 * Move any callbacks whose grace period has completed to the
1300 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1301 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1302 * sublist. This function is idempotent, so it does not hurt to
1303 * invoke it repeatedly. As long as it is not invoked -too- often...
1305 * The caller must hold rnp->lock with interrupts disabled.
1307 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1308 struct rcu_data *rdp)
1312 /* If the CPU has no callbacks, nothing to do. */
1313 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1317 * Find all callbacks whose ->completed numbers indicate that they
1318 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1320 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1321 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1323 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1325 /* Clean up any sublist tail pointers that were misordered above. */
1326 for (j = RCU_WAIT_TAIL; j < i; j++)
1327 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1329 /* Copy down callbacks to fill in empty sublists. */
1330 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1331 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1333 rdp->nxttail[j] = rdp->nxttail[i];
1334 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1337 /* Classify any remaining callbacks. */
1338 rcu_accelerate_cbs(rsp, rnp, rdp);
1342 * Update CPU-local rcu_data state to record the beginnings and ends of
1343 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1344 * structure corresponding to the current CPU, and must have irqs disabled.
1346 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1348 /* Handle the ends of any preceding grace periods first. */
1349 if (rdp->completed == rnp->completed) {
1351 /* No grace period end, so just accelerate recent callbacks. */
1352 rcu_accelerate_cbs(rsp, rnp, rdp);
1356 /* Advance callbacks. */
1357 rcu_advance_cbs(rsp, rnp, rdp);
1359 /* Remember that we saw this grace-period completion. */
1360 rdp->completed = rnp->completed;
1361 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1364 if (rdp->gpnum != rnp->gpnum) {
1366 * If the current grace period is waiting for this CPU,
1367 * set up to detect a quiescent state, otherwise don't
1368 * go looking for one.
1370 rdp->gpnum = rnp->gpnum;
1371 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1372 rdp->passed_quiesce = 0;
1373 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1374 zero_cpu_stall_ticks(rdp);
1378 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1380 unsigned long flags;
1381 struct rcu_node *rnp;
1383 local_irq_save(flags);
1385 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1386 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1387 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1388 local_irq_restore(flags);
1391 smp_mb__after_unlock_lock();
1392 __note_gp_changes(rsp, rnp, rdp);
1393 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1397 * Initialize a new grace period. Return 0 if no grace period required.
1399 static int rcu_gp_init(struct rcu_state *rsp)
1401 struct rcu_data *rdp;
1402 struct rcu_node *rnp = rcu_get_root(rsp);
1404 rcu_bind_gp_kthread();
1405 raw_spin_lock_irq(&rnp->lock);
1406 smp_mb__after_unlock_lock();
1407 if (!ACCESS_ONCE(rsp->gp_flags)) {
1408 /* Spurious wakeup, tell caller to go back to sleep. */
1409 raw_spin_unlock_irq(&rnp->lock);
1412 ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */
1414 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1416 * Grace period already in progress, don't start another.
1417 * Not supposed to be able to happen.
1419 raw_spin_unlock_irq(&rnp->lock);
1423 /* Advance to a new grace period and initialize state. */
1424 record_gp_stall_check_time(rsp);
1425 /* Record GP times before starting GP, hence smp_store_release(). */
1426 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1427 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1428 raw_spin_unlock_irq(&rnp->lock);
1430 /* Exclude any concurrent CPU-hotplug operations. */
1431 mutex_lock(&rsp->onoff_mutex);
1432 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1435 * Set the quiescent-state-needed bits in all the rcu_node
1436 * structures for all currently online CPUs in breadth-first order,
1437 * starting from the root rcu_node structure, relying on the layout
1438 * of the tree within the rsp->node[] array. Note that other CPUs
1439 * will access only the leaves of the hierarchy, thus seeing that no
1440 * grace period is in progress, at least until the corresponding
1441 * leaf node has been initialized. In addition, we have excluded
1442 * CPU-hotplug operations.
1444 * The grace period cannot complete until the initialization
1445 * process finishes, because this kthread handles both.
1447 rcu_for_each_node_breadth_first(rsp, rnp) {
1448 raw_spin_lock_irq(&rnp->lock);
1449 smp_mb__after_unlock_lock();
1450 rdp = this_cpu_ptr(rsp->rda);
1451 rcu_preempt_check_blocked_tasks(rnp);
1452 rnp->qsmask = rnp->qsmaskinit;
1453 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1454 WARN_ON_ONCE(rnp->completed != rsp->completed);
1455 ACCESS_ONCE(rnp->completed) = rsp->completed;
1456 if (rnp == rdp->mynode)
1457 __note_gp_changes(rsp, rnp, rdp);
1458 rcu_preempt_boost_start_gp(rnp);
1459 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1460 rnp->level, rnp->grplo,
1461 rnp->grphi, rnp->qsmask);
1462 raw_spin_unlock_irq(&rnp->lock);
1463 #ifdef CONFIG_PROVE_RCU_DELAY
1464 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1465 system_state == SYSTEM_RUNNING)
1467 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1471 mutex_unlock(&rsp->onoff_mutex);
1476 * Do one round of quiescent-state forcing.
1478 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1480 int fqs_state = fqs_state_in;
1481 bool isidle = false;
1483 struct rcu_node *rnp = rcu_get_root(rsp);
1486 if (fqs_state == RCU_SAVE_DYNTICK) {
1487 /* Collect dyntick-idle snapshots. */
1488 if (is_sysidle_rcu_state(rsp)) {
1490 maxj = jiffies - ULONG_MAX / 4;
1492 force_qs_rnp(rsp, dyntick_save_progress_counter,
1494 rcu_sysidle_report_gp(rsp, isidle, maxj);
1495 fqs_state = RCU_FORCE_QS;
1497 /* Handle dyntick-idle and offline CPUs. */
1499 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1501 /* Clear flag to prevent immediate re-entry. */
1502 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1503 raw_spin_lock_irq(&rnp->lock);
1504 smp_mb__after_unlock_lock();
1505 ACCESS_ONCE(rsp->gp_flags) &= ~RCU_GP_FLAG_FQS;
1506 raw_spin_unlock_irq(&rnp->lock);
1512 * Clean up after the old grace period.
1514 static void rcu_gp_cleanup(struct rcu_state *rsp)
1516 unsigned long gp_duration;
1518 struct rcu_data *rdp;
1519 struct rcu_node *rnp = rcu_get_root(rsp);
1521 raw_spin_lock_irq(&rnp->lock);
1522 smp_mb__after_unlock_lock();
1523 gp_duration = jiffies - rsp->gp_start;
1524 if (gp_duration > rsp->gp_max)
1525 rsp->gp_max = gp_duration;
1528 * We know the grace period is complete, but to everyone else
1529 * it appears to still be ongoing. But it is also the case
1530 * that to everyone else it looks like there is nothing that
1531 * they can do to advance the grace period. It is therefore
1532 * safe for us to drop the lock in order to mark the grace
1533 * period as completed in all of the rcu_node structures.
1535 raw_spin_unlock_irq(&rnp->lock);
1538 * Propagate new ->completed value to rcu_node structures so
1539 * that other CPUs don't have to wait until the start of the next
1540 * grace period to process their callbacks. This also avoids
1541 * some nasty RCU grace-period initialization races by forcing
1542 * the end of the current grace period to be completely recorded in
1543 * all of the rcu_node structures before the beginning of the next
1544 * grace period is recorded in any of the rcu_node structures.
1546 rcu_for_each_node_breadth_first(rsp, rnp) {
1547 raw_spin_lock_irq(&rnp->lock);
1548 smp_mb__after_unlock_lock();
1549 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1550 rdp = this_cpu_ptr(rsp->rda);
1551 if (rnp == rdp->mynode)
1552 __note_gp_changes(rsp, rnp, rdp);
1553 /* smp_mb() provided by prior unlock-lock pair. */
1554 nocb += rcu_future_gp_cleanup(rsp, rnp);
1555 raw_spin_unlock_irq(&rnp->lock);
1558 rnp = rcu_get_root(rsp);
1559 raw_spin_lock_irq(&rnp->lock);
1560 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1561 rcu_nocb_gp_set(rnp, nocb);
1563 /* Declare grace period done. */
1564 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1565 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1566 rsp->fqs_state = RCU_GP_IDLE;
1567 rdp = this_cpu_ptr(rsp->rda);
1568 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1569 if (cpu_needs_another_gp(rsp, rdp)) {
1570 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1571 trace_rcu_grace_period(rsp->name,
1572 ACCESS_ONCE(rsp->gpnum),
1575 raw_spin_unlock_irq(&rnp->lock);
1579 * Body of kthread that handles grace periods.
1581 static int __noreturn rcu_gp_kthread(void *arg)
1587 struct rcu_state *rsp = arg;
1588 struct rcu_node *rnp = rcu_get_root(rsp);
1592 /* Handle grace-period start. */
1594 trace_rcu_grace_period(rsp->name,
1595 ACCESS_ONCE(rsp->gpnum),
1597 wait_event_interruptible(rsp->gp_wq,
1598 ACCESS_ONCE(rsp->gp_flags) &
1600 /* Locking provides needed memory barrier. */
1601 if (rcu_gp_init(rsp))
1604 flush_signals(current);
1605 trace_rcu_grace_period(rsp->name,
1606 ACCESS_ONCE(rsp->gpnum),
1610 /* Handle quiescent-state forcing. */
1611 fqs_state = RCU_SAVE_DYNTICK;
1612 j = jiffies_till_first_fqs;
1615 jiffies_till_first_fqs = HZ;
1620 rsp->jiffies_force_qs = jiffies + j;
1621 trace_rcu_grace_period(rsp->name,
1622 ACCESS_ONCE(rsp->gpnum),
1624 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1625 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1627 (!ACCESS_ONCE(rnp->qsmask) &&
1628 !rcu_preempt_blocked_readers_cgp(rnp)),
1630 /* Locking provides needed memory barriers. */
1631 /* If grace period done, leave loop. */
1632 if (!ACCESS_ONCE(rnp->qsmask) &&
1633 !rcu_preempt_blocked_readers_cgp(rnp))
1635 /* If time for quiescent-state forcing, do it. */
1636 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1637 (gf & RCU_GP_FLAG_FQS)) {
1638 trace_rcu_grace_period(rsp->name,
1639 ACCESS_ONCE(rsp->gpnum),
1641 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1642 trace_rcu_grace_period(rsp->name,
1643 ACCESS_ONCE(rsp->gpnum),
1647 /* Deal with stray signal. */
1649 flush_signals(current);
1650 trace_rcu_grace_period(rsp->name,
1651 ACCESS_ONCE(rsp->gpnum),
1654 j = jiffies_till_next_fqs;
1657 jiffies_till_next_fqs = HZ;
1660 jiffies_till_next_fqs = 1;
1664 /* Handle grace-period end. */
1665 rcu_gp_cleanup(rsp);
1669 static void rsp_wakeup(struct irq_work *work)
1671 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1673 /* Wake up rcu_gp_kthread() to start the grace period. */
1674 wake_up(&rsp->gp_wq);
1678 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1679 * in preparation for detecting the next grace period. The caller must hold
1680 * the root node's ->lock and hard irqs must be disabled.
1682 * Note that it is legal for a dying CPU (which is marked as offline) to
1683 * invoke this function. This can happen when the dying CPU reports its
1687 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1688 struct rcu_data *rdp)
1690 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1692 * Either we have not yet spawned the grace-period
1693 * task, this CPU does not need another grace period,
1694 * or a grace period is already in progress.
1695 * Either way, don't start a new grace period.
1699 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1700 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1704 * We can't do wakeups while holding the rnp->lock, as that
1705 * could cause possible deadlocks with the rq->lock. Defer
1706 * the wakeup to interrupt context. And don't bother waking
1707 * up the running kthread.
1709 if (current != rsp->gp_kthread)
1710 irq_work_queue(&rsp->wakeup_work);
1714 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1715 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1716 * is invoked indirectly from rcu_advance_cbs(), which would result in
1717 * endless recursion -- or would do so if it wasn't for the self-deadlock
1718 * that is encountered beforehand.
1721 rcu_start_gp(struct rcu_state *rsp)
1723 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1724 struct rcu_node *rnp = rcu_get_root(rsp);
1727 * If there is no grace period in progress right now, any
1728 * callbacks we have up to this point will be satisfied by the
1729 * next grace period. Also, advancing the callbacks reduces the
1730 * probability of false positives from cpu_needs_another_gp()
1731 * resulting in pointless grace periods. So, advance callbacks
1732 * then start the grace period!
1734 rcu_advance_cbs(rsp, rnp, rdp);
1735 rcu_start_gp_advanced(rsp, rnp, rdp);
1739 * Report a full set of quiescent states to the specified rcu_state
1740 * data structure. This involves cleaning up after the prior grace
1741 * period and letting rcu_start_gp() start up the next grace period
1742 * if one is needed. Note that the caller must hold rnp->lock, which
1743 * is released before return.
1745 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1746 __releases(rcu_get_root(rsp)->lock)
1748 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1749 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1750 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1754 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1755 * Allows quiescent states for a group of CPUs to be reported at one go
1756 * to the specified rcu_node structure, though all the CPUs in the group
1757 * must be represented by the same rcu_node structure (which need not be
1758 * a leaf rcu_node structure, though it often will be). That structure's
1759 * lock must be held upon entry, and it is released before return.
1762 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1763 struct rcu_node *rnp, unsigned long flags)
1764 __releases(rnp->lock)
1766 struct rcu_node *rnp_c;
1768 /* Walk up the rcu_node hierarchy. */
1770 if (!(rnp->qsmask & mask)) {
1772 /* Our bit has already been cleared, so done. */
1773 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1776 rnp->qsmask &= ~mask;
1777 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1778 mask, rnp->qsmask, rnp->level,
1779 rnp->grplo, rnp->grphi,
1781 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1783 /* Other bits still set at this level, so done. */
1784 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1787 mask = rnp->grpmask;
1788 if (rnp->parent == NULL) {
1790 /* No more levels. Exit loop holding root lock. */
1794 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1797 raw_spin_lock_irqsave(&rnp->lock, flags);
1798 smp_mb__after_unlock_lock();
1799 WARN_ON_ONCE(rnp_c->qsmask);
1803 * Get here if we are the last CPU to pass through a quiescent
1804 * state for this grace period. Invoke rcu_report_qs_rsp()
1805 * to clean up and start the next grace period if one is needed.
1807 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1811 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1812 * structure. This must be either called from the specified CPU, or
1813 * called when the specified CPU is known to be offline (and when it is
1814 * also known that no other CPU is concurrently trying to help the offline
1815 * CPU). The lastcomp argument is used to make sure we are still in the
1816 * grace period of interest. We don't want to end the current grace period
1817 * based on quiescent states detected in an earlier grace period!
1820 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1822 unsigned long flags;
1824 struct rcu_node *rnp;
1827 raw_spin_lock_irqsave(&rnp->lock, flags);
1828 smp_mb__after_unlock_lock();
1829 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1830 rnp->completed == rnp->gpnum) {
1833 * The grace period in which this quiescent state was
1834 * recorded has ended, so don't report it upwards.
1835 * We will instead need a new quiescent state that lies
1836 * within the current grace period.
1838 rdp->passed_quiesce = 0; /* need qs for new gp. */
1839 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1842 mask = rdp->grpmask;
1843 if ((rnp->qsmask & mask) == 0) {
1844 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1846 rdp->qs_pending = 0;
1849 * This GP can't end until cpu checks in, so all of our
1850 * callbacks can be processed during the next GP.
1852 rcu_accelerate_cbs(rsp, rnp, rdp);
1854 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1859 * Check to see if there is a new grace period of which this CPU
1860 * is not yet aware, and if so, set up local rcu_data state for it.
1861 * Otherwise, see if this CPU has just passed through its first
1862 * quiescent state for this grace period, and record that fact if so.
1865 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1867 /* Check for grace-period ends and beginnings. */
1868 note_gp_changes(rsp, rdp);
1871 * Does this CPU still need to do its part for current grace period?
1872 * If no, return and let the other CPUs do their part as well.
1874 if (!rdp->qs_pending)
1878 * Was there a quiescent state since the beginning of the grace
1879 * period? If no, then exit and wait for the next call.
1881 if (!rdp->passed_quiesce)
1885 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1888 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1891 #ifdef CONFIG_HOTPLUG_CPU
1894 * Send the specified CPU's RCU callbacks to the orphanage. The
1895 * specified CPU must be offline, and the caller must hold the
1899 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1900 struct rcu_node *rnp, struct rcu_data *rdp)
1902 /* No-CBs CPUs do not have orphanable callbacks. */
1903 if (rcu_is_nocb_cpu(rdp->cpu))
1907 * Orphan the callbacks. First adjust the counts. This is safe
1908 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1909 * cannot be running now. Thus no memory barrier is required.
1911 if (rdp->nxtlist != NULL) {
1912 rsp->qlen_lazy += rdp->qlen_lazy;
1913 rsp->qlen += rdp->qlen;
1914 rdp->n_cbs_orphaned += rdp->qlen;
1916 ACCESS_ONCE(rdp->qlen) = 0;
1920 * Next, move those callbacks still needing a grace period to
1921 * the orphanage, where some other CPU will pick them up.
1922 * Some of the callbacks might have gone partway through a grace
1923 * period, but that is too bad. They get to start over because we
1924 * cannot assume that grace periods are synchronized across CPUs.
1925 * We don't bother updating the ->nxttail[] array yet, instead
1926 * we just reset the whole thing later on.
1928 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1929 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1930 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1931 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1935 * Then move the ready-to-invoke callbacks to the orphanage,
1936 * where some other CPU will pick them up. These will not be
1937 * required to pass though another grace period: They are done.
1939 if (rdp->nxtlist != NULL) {
1940 *rsp->orphan_donetail = rdp->nxtlist;
1941 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1944 /* Finally, initialize the rcu_data structure's list to empty. */
1945 init_callback_list(rdp);
1949 * Adopt the RCU callbacks from the specified rcu_state structure's
1950 * orphanage. The caller must hold the ->orphan_lock.
1952 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
1955 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1957 /* No-CBs CPUs are handled specially. */
1958 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
1961 /* Do the accounting first. */
1962 rdp->qlen_lazy += rsp->qlen_lazy;
1963 rdp->qlen += rsp->qlen;
1964 rdp->n_cbs_adopted += rsp->qlen;
1965 if (rsp->qlen_lazy != rsp->qlen)
1966 rcu_idle_count_callbacks_posted();
1971 * We do not need a memory barrier here because the only way we
1972 * can get here if there is an rcu_barrier() in flight is if
1973 * we are the task doing the rcu_barrier().
1976 /* First adopt the ready-to-invoke callbacks. */
1977 if (rsp->orphan_donelist != NULL) {
1978 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1979 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1980 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1981 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1982 rdp->nxttail[i] = rsp->orphan_donetail;
1983 rsp->orphan_donelist = NULL;
1984 rsp->orphan_donetail = &rsp->orphan_donelist;
1987 /* And then adopt the callbacks that still need a grace period. */
1988 if (rsp->orphan_nxtlist != NULL) {
1989 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1990 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1991 rsp->orphan_nxtlist = NULL;
1992 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1997 * Trace the fact that this CPU is going offline.
1999 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2001 RCU_TRACE(unsigned long mask);
2002 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2003 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2005 RCU_TRACE(mask = rdp->grpmask);
2006 trace_rcu_grace_period(rsp->name,
2007 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2012 * The CPU has been completely removed, and some other CPU is reporting
2013 * this fact from process context. Do the remainder of the cleanup,
2014 * including orphaning the outgoing CPU's RCU callbacks, and also
2015 * adopting them. There can only be one CPU hotplug operation at a time,
2016 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2018 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2020 unsigned long flags;
2022 int need_report = 0;
2023 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2024 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2026 /* Adjust any no-longer-needed kthreads. */
2027 rcu_boost_kthread_setaffinity(rnp, -1);
2029 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2031 /* Exclude any attempts to start a new grace period. */
2032 mutex_lock(&rsp->onoff_mutex);
2033 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2035 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2036 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2037 rcu_adopt_orphan_cbs(rsp, flags);
2039 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2040 mask = rdp->grpmask; /* rnp->grplo is constant. */
2042 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2043 smp_mb__after_unlock_lock();
2044 rnp->qsmaskinit &= ~mask;
2045 if (rnp->qsmaskinit != 0) {
2046 if (rnp != rdp->mynode)
2047 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2050 if (rnp == rdp->mynode)
2051 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2053 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2054 mask = rnp->grpmask;
2056 } while (rnp != NULL);
2059 * We still hold the leaf rcu_node structure lock here, and
2060 * irqs are still disabled. The reason for this subterfuge is
2061 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2062 * held leads to deadlock.
2064 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2066 if (need_report & RCU_OFL_TASKS_NORM_GP)
2067 rcu_report_unblock_qs_rnp(rnp, flags);
2069 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2070 if (need_report & RCU_OFL_TASKS_EXP_GP)
2071 rcu_report_exp_rnp(rsp, rnp, true);
2072 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2073 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2074 cpu, rdp->qlen, rdp->nxtlist);
2075 init_callback_list(rdp);
2076 /* Disallow further callbacks on this CPU. */
2077 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2078 mutex_unlock(&rsp->onoff_mutex);
2081 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2083 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2087 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2091 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2094 * Invoke any RCU callbacks that have made it to the end of their grace
2095 * period. Thottle as specified by rdp->blimit.
2097 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2099 unsigned long flags;
2100 struct rcu_head *next, *list, **tail;
2101 long bl, count, count_lazy;
2104 /* If no callbacks are ready, just return. */
2105 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2106 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2107 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2108 need_resched(), is_idle_task(current),
2109 rcu_is_callbacks_kthread());
2114 * Extract the list of ready callbacks, disabling to prevent
2115 * races with call_rcu() from interrupt handlers.
2117 local_irq_save(flags);
2118 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2120 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2121 list = rdp->nxtlist;
2122 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2123 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2124 tail = rdp->nxttail[RCU_DONE_TAIL];
2125 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2126 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2127 rdp->nxttail[i] = &rdp->nxtlist;
2128 local_irq_restore(flags);
2130 /* Invoke callbacks. */
2131 count = count_lazy = 0;
2135 debug_rcu_head_unqueue(list);
2136 if (__rcu_reclaim(rsp->name, list))
2139 /* Stop only if limit reached and CPU has something to do. */
2140 if (++count >= bl &&
2142 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2146 local_irq_save(flags);
2147 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2148 is_idle_task(current),
2149 rcu_is_callbacks_kthread());
2151 /* Update count, and requeue any remaining callbacks. */
2153 *tail = rdp->nxtlist;
2154 rdp->nxtlist = list;
2155 for (i = 0; i < RCU_NEXT_SIZE; i++)
2156 if (&rdp->nxtlist == rdp->nxttail[i])
2157 rdp->nxttail[i] = tail;
2161 smp_mb(); /* List handling before counting for rcu_barrier(). */
2162 rdp->qlen_lazy -= count_lazy;
2163 ACCESS_ONCE(rdp->qlen) -= count;
2164 rdp->n_cbs_invoked += count;
2166 /* Reinstate batch limit if we have worked down the excess. */
2167 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2168 rdp->blimit = blimit;
2170 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2171 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2172 rdp->qlen_last_fqs_check = 0;
2173 rdp->n_force_qs_snap = rsp->n_force_qs;
2174 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2175 rdp->qlen_last_fqs_check = rdp->qlen;
2176 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2178 local_irq_restore(flags);
2180 /* Re-invoke RCU core processing if there are callbacks remaining. */
2181 if (cpu_has_callbacks_ready_to_invoke(rdp))
2186 * Check to see if this CPU is in a non-context-switch quiescent state
2187 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2188 * Also schedule RCU core processing.
2190 * This function must be called from hardirq context. It is normally
2191 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2192 * false, there is no point in invoking rcu_check_callbacks().
2194 void rcu_check_callbacks(int cpu, int user)
2196 trace_rcu_utilization(TPS("Start scheduler-tick"));
2197 increment_cpu_stall_ticks();
2198 if (user || rcu_is_cpu_rrupt_from_idle()) {
2201 * Get here if this CPU took its interrupt from user
2202 * mode or from the idle loop, and if this is not a
2203 * nested interrupt. In this case, the CPU is in
2204 * a quiescent state, so note it.
2206 * No memory barrier is required here because both
2207 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2208 * variables that other CPUs neither access nor modify,
2209 * at least not while the corresponding CPU is online.
2215 } else if (!in_softirq()) {
2218 * Get here if this CPU did not take its interrupt from
2219 * softirq, in other words, if it is not interrupting
2220 * a rcu_bh read-side critical section. This is an _bh
2221 * critical section, so note it.
2226 rcu_preempt_check_callbacks(cpu);
2227 if (rcu_pending(cpu))
2229 trace_rcu_utilization(TPS("End scheduler-tick"));
2233 * Scan the leaf rcu_node structures, processing dyntick state for any that
2234 * have not yet encountered a quiescent state, using the function specified.
2235 * Also initiate boosting for any threads blocked on the root rcu_node.
2237 * The caller must have suppressed start of new grace periods.
2239 static void force_qs_rnp(struct rcu_state *rsp,
2240 int (*f)(struct rcu_data *rsp, bool *isidle,
2241 unsigned long *maxj),
2242 bool *isidle, unsigned long *maxj)
2246 unsigned long flags;
2248 struct rcu_node *rnp;
2250 rcu_for_each_leaf_node(rsp, rnp) {
2253 raw_spin_lock_irqsave(&rnp->lock, flags);
2254 smp_mb__after_unlock_lock();
2255 if (!rcu_gp_in_progress(rsp)) {
2256 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2259 if (rnp->qsmask == 0) {
2260 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2265 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2266 if ((rnp->qsmask & bit) != 0) {
2267 if ((rnp->qsmaskinit & bit) != 0)
2269 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2275 /* rcu_report_qs_rnp() releases rnp->lock. */
2276 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2279 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2281 rnp = rcu_get_root(rsp);
2282 if (rnp->qsmask == 0) {
2283 raw_spin_lock_irqsave(&rnp->lock, flags);
2284 smp_mb__after_unlock_lock();
2285 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2290 * Force quiescent states on reluctant CPUs, and also detect which
2291 * CPUs are in dyntick-idle mode.
2293 static void force_quiescent_state(struct rcu_state *rsp)
2295 unsigned long flags;
2297 struct rcu_node *rnp;
2298 struct rcu_node *rnp_old = NULL;
2300 /* Funnel through hierarchy to reduce memory contention. */
2301 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2302 for (; rnp != NULL; rnp = rnp->parent) {
2303 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2304 !raw_spin_trylock(&rnp->fqslock);
2305 if (rnp_old != NULL)
2306 raw_spin_unlock(&rnp_old->fqslock);
2308 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2313 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2315 /* Reached the root of the rcu_node tree, acquire lock. */
2316 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2317 smp_mb__after_unlock_lock();
2318 raw_spin_unlock(&rnp_old->fqslock);
2319 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2320 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2321 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2322 return; /* Someone beat us to it. */
2324 ACCESS_ONCE(rsp->gp_flags) |= RCU_GP_FLAG_FQS;
2325 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2326 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2330 * This does the RCU core processing work for the specified rcu_state
2331 * and rcu_data structures. This may be called only from the CPU to
2332 * whom the rdp belongs.
2335 __rcu_process_callbacks(struct rcu_state *rsp)
2337 unsigned long flags;
2338 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2340 WARN_ON_ONCE(rdp->beenonline == 0);
2342 /* Update RCU state based on any recent quiescent states. */
2343 rcu_check_quiescent_state(rsp, rdp);
2345 /* Does this CPU require a not-yet-started grace period? */
2346 local_irq_save(flags);
2347 if (cpu_needs_another_gp(rsp, rdp)) {
2348 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2350 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2352 local_irq_restore(flags);
2355 /* If there are callbacks ready, invoke them. */
2356 if (cpu_has_callbacks_ready_to_invoke(rdp))
2357 invoke_rcu_callbacks(rsp, rdp);
2359 /* Do any needed deferred wakeups of rcuo kthreads. */
2360 do_nocb_deferred_wakeup(rdp);
2364 * Do RCU core processing for the current CPU.
2366 static void rcu_process_callbacks(struct softirq_action *unused)
2368 struct rcu_state *rsp;
2370 if (cpu_is_offline(smp_processor_id()))
2372 trace_rcu_utilization(TPS("Start RCU core"));
2373 for_each_rcu_flavor(rsp)
2374 __rcu_process_callbacks(rsp);
2375 trace_rcu_utilization(TPS("End RCU core"));
2379 * Schedule RCU callback invocation. If the specified type of RCU
2380 * does not support RCU priority boosting, just do a direct call,
2381 * otherwise wake up the per-CPU kernel kthread. Note that because we
2382 * are running on the current CPU with interrupts disabled, the
2383 * rcu_cpu_kthread_task cannot disappear out from under us.
2385 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2387 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2389 if (likely(!rsp->boost)) {
2390 rcu_do_batch(rsp, rdp);
2393 invoke_rcu_callbacks_kthread();
2396 static void invoke_rcu_core(void)
2398 if (cpu_online(smp_processor_id()))
2399 raise_softirq(RCU_SOFTIRQ);
2403 * Handle any core-RCU processing required by a call_rcu() invocation.
2405 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2406 struct rcu_head *head, unsigned long flags)
2409 * If called from an extended quiescent state, invoke the RCU
2410 * core in order to force a re-evaluation of RCU's idleness.
2412 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2415 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2416 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2420 * Force the grace period if too many callbacks or too long waiting.
2421 * Enforce hysteresis, and don't invoke force_quiescent_state()
2422 * if some other CPU has recently done so. Also, don't bother
2423 * invoking force_quiescent_state() if the newly enqueued callback
2424 * is the only one waiting for a grace period to complete.
2426 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2428 /* Are we ignoring a completed grace period? */
2429 note_gp_changes(rsp, rdp);
2431 /* Start a new grace period if one not already started. */
2432 if (!rcu_gp_in_progress(rsp)) {
2433 struct rcu_node *rnp_root = rcu_get_root(rsp);
2435 raw_spin_lock(&rnp_root->lock);
2436 smp_mb__after_unlock_lock();
2438 raw_spin_unlock(&rnp_root->lock);
2440 /* Give the grace period a kick. */
2441 rdp->blimit = LONG_MAX;
2442 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2443 *rdp->nxttail[RCU_DONE_TAIL] != head)
2444 force_quiescent_state(rsp);
2445 rdp->n_force_qs_snap = rsp->n_force_qs;
2446 rdp->qlen_last_fqs_check = rdp->qlen;
2452 * RCU callback function to leak a callback.
2454 static void rcu_leak_callback(struct rcu_head *rhp)
2459 * Helper function for call_rcu() and friends. The cpu argument will
2460 * normally be -1, indicating "currently running CPU". It may specify
2461 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2462 * is expected to specify a CPU.
2465 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2466 struct rcu_state *rsp, int cpu, bool lazy)
2468 unsigned long flags;
2469 struct rcu_data *rdp;
2471 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2472 if (debug_rcu_head_queue(head)) {
2473 /* Probable double call_rcu(), so leak the callback. */
2474 ACCESS_ONCE(head->func) = rcu_leak_callback;
2475 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2482 * Opportunistically note grace-period endings and beginnings.
2483 * Note that we might see a beginning right after we see an
2484 * end, but never vice versa, since this CPU has to pass through
2485 * a quiescent state betweentimes.
2487 local_irq_save(flags);
2488 rdp = this_cpu_ptr(rsp->rda);
2490 /* Add the callback to our list. */
2491 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2495 rdp = per_cpu_ptr(rsp->rda, cpu);
2496 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2497 WARN_ON_ONCE(offline);
2498 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2499 local_irq_restore(flags);
2502 ACCESS_ONCE(rdp->qlen)++;
2506 rcu_idle_count_callbacks_posted();
2507 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2508 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2509 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2511 if (__is_kfree_rcu_offset((unsigned long)func))
2512 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2513 rdp->qlen_lazy, rdp->qlen);
2515 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2517 /* Go handle any RCU core processing required. */
2518 __call_rcu_core(rsp, rdp, head, flags);
2519 local_irq_restore(flags);
2523 * Queue an RCU-sched callback for invocation after a grace period.
2525 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2527 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2529 EXPORT_SYMBOL_GPL(call_rcu_sched);
2532 * Queue an RCU callback for invocation after a quicker grace period.
2534 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2536 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2538 EXPORT_SYMBOL_GPL(call_rcu_bh);
2541 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2542 * any blocking grace-period wait automatically implies a grace period
2543 * if there is only one CPU online at any point time during execution
2544 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2545 * occasionally incorrectly indicate that there are multiple CPUs online
2546 * when there was in fact only one the whole time, as this just adds
2547 * some overhead: RCU still operates correctly.
2549 static inline int rcu_blocking_is_gp(void)
2553 might_sleep(); /* Check for RCU read-side critical section. */
2555 ret = num_online_cpus() <= 1;
2561 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2563 * Control will return to the caller some time after a full rcu-sched
2564 * grace period has elapsed, in other words after all currently executing
2565 * rcu-sched read-side critical sections have completed. These read-side
2566 * critical sections are delimited by rcu_read_lock_sched() and
2567 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2568 * local_irq_disable(), and so on may be used in place of
2569 * rcu_read_lock_sched().
2571 * This means that all preempt_disable code sequences, including NMI and
2572 * non-threaded hardware-interrupt handlers, in progress on entry will
2573 * have completed before this primitive returns. However, this does not
2574 * guarantee that softirq handlers will have completed, since in some
2575 * kernels, these handlers can run in process context, and can block.
2577 * Note that this guarantee implies further memory-ordering guarantees.
2578 * On systems with more than one CPU, when synchronize_sched() returns,
2579 * each CPU is guaranteed to have executed a full memory barrier since the
2580 * end of its last RCU-sched read-side critical section whose beginning
2581 * preceded the call to synchronize_sched(). In addition, each CPU having
2582 * an RCU read-side critical section that extends beyond the return from
2583 * synchronize_sched() is guaranteed to have executed a full memory barrier
2584 * after the beginning of synchronize_sched() and before the beginning of
2585 * that RCU read-side critical section. Note that these guarantees include
2586 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2587 * that are executing in the kernel.
2589 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2590 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2591 * to have executed a full memory barrier during the execution of
2592 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2593 * again only if the system has more than one CPU).
2595 * This primitive provides the guarantees made by the (now removed)
2596 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2597 * guarantees that rcu_read_lock() sections will have completed.
2598 * In "classic RCU", these two guarantees happen to be one and
2599 * the same, but can differ in realtime RCU implementations.
2601 void synchronize_sched(void)
2603 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2604 !lock_is_held(&rcu_lock_map) &&
2605 !lock_is_held(&rcu_sched_lock_map),
2606 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2607 if (rcu_blocking_is_gp())
2610 synchronize_sched_expedited();
2612 wait_rcu_gp(call_rcu_sched);
2614 EXPORT_SYMBOL_GPL(synchronize_sched);
2617 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2619 * Control will return to the caller some time after a full rcu_bh grace
2620 * period has elapsed, in other words after all currently executing rcu_bh
2621 * read-side critical sections have completed. RCU read-side critical
2622 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2623 * and may be nested.
2625 * See the description of synchronize_sched() for more detailed information
2626 * on memory ordering guarantees.
2628 void synchronize_rcu_bh(void)
2630 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2631 !lock_is_held(&rcu_lock_map) &&
2632 !lock_is_held(&rcu_sched_lock_map),
2633 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2634 if (rcu_blocking_is_gp())
2637 synchronize_rcu_bh_expedited();
2639 wait_rcu_gp(call_rcu_bh);
2641 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2644 * get_state_synchronize_rcu - Snapshot current RCU state
2646 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2647 * to determine whether or not a full grace period has elapsed in the
2650 unsigned long get_state_synchronize_rcu(void)
2653 * Any prior manipulation of RCU-protected data must happen
2654 * before the load from ->gpnum.
2659 * Make sure this load happens before the purportedly
2660 * time-consuming work between get_state_synchronize_rcu()
2661 * and cond_synchronize_rcu().
2663 return smp_load_acquire(&rcu_state->gpnum);
2665 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2668 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2670 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2672 * If a full RCU grace period has elapsed since the earlier call to
2673 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2674 * synchronize_rcu() to wait for a full grace period.
2676 * Yes, this function does not take counter wrap into account. But
2677 * counter wrap is harmless. If the counter wraps, we have waited for
2678 * more than 2 billion grace periods (and way more on a 64-bit system!),
2679 * so waiting for one additional grace period should be just fine.
2681 void cond_synchronize_rcu(unsigned long oldstate)
2683 unsigned long newstate;
2686 * Ensure that this load happens before any RCU-destructive
2687 * actions the caller might carry out after we return.
2689 newstate = smp_load_acquire(&rcu_state->completed);
2690 if (ULONG_CMP_GE(oldstate, newstate))
2693 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2695 static int synchronize_sched_expedited_cpu_stop(void *data)
2698 * There must be a full memory barrier on each affected CPU
2699 * between the time that try_stop_cpus() is called and the
2700 * time that it returns.
2702 * In the current initial implementation of cpu_stop, the
2703 * above condition is already met when the control reaches
2704 * this point and the following smp_mb() is not strictly
2705 * necessary. Do smp_mb() anyway for documentation and
2706 * robustness against future implementation changes.
2708 smp_mb(); /* See above comment block. */
2713 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2715 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2716 * approach to force the grace period to end quickly. This consumes
2717 * significant time on all CPUs and is unfriendly to real-time workloads,
2718 * so is thus not recommended for any sort of common-case code. In fact,
2719 * if you are using synchronize_sched_expedited() in a loop, please
2720 * restructure your code to batch your updates, and then use a single
2721 * synchronize_sched() instead.
2723 * Note that it is illegal to call this function while holding any lock
2724 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2725 * to call this function from a CPU-hotplug notifier. Failing to observe
2726 * these restriction will result in deadlock.
2728 * This implementation can be thought of as an application of ticket
2729 * locking to RCU, with sync_sched_expedited_started and
2730 * sync_sched_expedited_done taking on the roles of the halves
2731 * of the ticket-lock word. Each task atomically increments
2732 * sync_sched_expedited_started upon entry, snapshotting the old value,
2733 * then attempts to stop all the CPUs. If this succeeds, then each
2734 * CPU will have executed a context switch, resulting in an RCU-sched
2735 * grace period. We are then done, so we use atomic_cmpxchg() to
2736 * update sync_sched_expedited_done to match our snapshot -- but
2737 * only if someone else has not already advanced past our snapshot.
2739 * On the other hand, if try_stop_cpus() fails, we check the value
2740 * of sync_sched_expedited_done. If it has advanced past our
2741 * initial snapshot, then someone else must have forced a grace period
2742 * some time after we took our snapshot. In this case, our work is
2743 * done for us, and we can simply return. Otherwise, we try again,
2744 * but keep our initial snapshot for purposes of checking for someone
2745 * doing our work for us.
2747 * If we fail too many times in a row, we fall back to synchronize_sched().
2749 void synchronize_sched_expedited(void)
2751 long firstsnap, s, snap;
2753 struct rcu_state *rsp = &rcu_sched_state;
2756 * If we are in danger of counter wrap, just do synchronize_sched().
2757 * By allowing sync_sched_expedited_started to advance no more than
2758 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2759 * that more than 3.5 billion CPUs would be required to force a
2760 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2761 * course be required on a 64-bit system.
2763 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2764 (ulong)atomic_long_read(&rsp->expedited_done) +
2766 synchronize_sched();
2767 atomic_long_inc(&rsp->expedited_wrap);
2772 * Take a ticket. Note that atomic_inc_return() implies a
2773 * full memory barrier.
2775 snap = atomic_long_inc_return(&rsp->expedited_start);
2778 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2781 * Each pass through the following loop attempts to force a
2782 * context switch on each CPU.
2784 while (try_stop_cpus(cpu_online_mask,
2785 synchronize_sched_expedited_cpu_stop,
2788 atomic_long_inc(&rsp->expedited_tryfail);
2790 /* Check to see if someone else did our work for us. */
2791 s = atomic_long_read(&rsp->expedited_done);
2792 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2793 /* ensure test happens before caller kfree */
2794 smp_mb__before_atomic_inc(); /* ^^^ */
2795 atomic_long_inc(&rsp->expedited_workdone1);
2799 /* No joy, try again later. Or just synchronize_sched(). */
2800 if (trycount++ < 10) {
2801 udelay(trycount * num_online_cpus());
2803 wait_rcu_gp(call_rcu_sched);
2804 atomic_long_inc(&rsp->expedited_normal);
2808 /* Recheck to see if someone else did our work for us. */
2809 s = atomic_long_read(&rsp->expedited_done);
2810 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2811 /* ensure test happens before caller kfree */
2812 smp_mb__before_atomic_inc(); /* ^^^ */
2813 atomic_long_inc(&rsp->expedited_workdone2);
2818 * Refetching sync_sched_expedited_started allows later
2819 * callers to piggyback on our grace period. We retry
2820 * after they started, so our grace period works for them,
2821 * and they started after our first try, so their grace
2822 * period works for us.
2825 snap = atomic_long_read(&rsp->expedited_start);
2826 smp_mb(); /* ensure read is before try_stop_cpus(). */
2828 atomic_long_inc(&rsp->expedited_stoppedcpus);
2831 * Everyone up to our most recent fetch is covered by our grace
2832 * period. Update the counter, but only if our work is still
2833 * relevant -- which it won't be if someone who started later
2834 * than we did already did their update.
2837 atomic_long_inc(&rsp->expedited_done_tries);
2838 s = atomic_long_read(&rsp->expedited_done);
2839 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2840 /* ensure test happens before caller kfree */
2841 smp_mb__before_atomic_inc(); /* ^^^ */
2842 atomic_long_inc(&rsp->expedited_done_lost);
2845 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2846 atomic_long_inc(&rsp->expedited_done_exit);
2850 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2853 * Check to see if there is any immediate RCU-related work to be done
2854 * by the current CPU, for the specified type of RCU, returning 1 if so.
2855 * The checks are in order of increasing expense: checks that can be
2856 * carried out against CPU-local state are performed first. However,
2857 * we must check for CPU stalls first, else we might not get a chance.
2859 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2861 struct rcu_node *rnp = rdp->mynode;
2863 rdp->n_rcu_pending++;
2865 /* Check for CPU stalls, if enabled. */
2866 check_cpu_stall(rsp, rdp);
2868 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2869 if (rcu_nohz_full_cpu(rsp))
2872 /* Is the RCU core waiting for a quiescent state from this CPU? */
2873 if (rcu_scheduler_fully_active &&
2874 rdp->qs_pending && !rdp->passed_quiesce) {
2875 rdp->n_rp_qs_pending++;
2876 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2877 rdp->n_rp_report_qs++;
2881 /* Does this CPU have callbacks ready to invoke? */
2882 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2883 rdp->n_rp_cb_ready++;
2887 /* Has RCU gone idle with this CPU needing another grace period? */
2888 if (cpu_needs_another_gp(rsp, rdp)) {
2889 rdp->n_rp_cpu_needs_gp++;
2893 /* Has another RCU grace period completed? */
2894 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2895 rdp->n_rp_gp_completed++;
2899 /* Has a new RCU grace period started? */
2900 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2901 rdp->n_rp_gp_started++;
2905 /* Does this CPU need a deferred NOCB wakeup? */
2906 if (rcu_nocb_need_deferred_wakeup(rdp)) {
2907 rdp->n_rp_nocb_defer_wakeup++;
2912 rdp->n_rp_need_nothing++;
2917 * Check to see if there is any immediate RCU-related work to be done
2918 * by the current CPU, returning 1 if so. This function is part of the
2919 * RCU implementation; it is -not- an exported member of the RCU API.
2921 static int rcu_pending(int cpu)
2923 struct rcu_state *rsp;
2925 for_each_rcu_flavor(rsp)
2926 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2932 * Return true if the specified CPU has any callback. If all_lazy is
2933 * non-NULL, store an indication of whether all callbacks are lazy.
2934 * (If there are no callbacks, all of them are deemed to be lazy.)
2936 static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2940 struct rcu_data *rdp;
2941 struct rcu_state *rsp;
2943 for_each_rcu_flavor(rsp) {
2944 rdp = per_cpu_ptr(rsp->rda, cpu);
2948 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2959 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2960 * the compiler is expected to optimize this away.
2962 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2963 int cpu, unsigned long done)
2965 trace_rcu_barrier(rsp->name, s, cpu,
2966 atomic_read(&rsp->barrier_cpu_count), done);
2970 * RCU callback function for _rcu_barrier(). If we are last, wake
2971 * up the task executing _rcu_barrier().
2973 static void rcu_barrier_callback(struct rcu_head *rhp)
2975 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2976 struct rcu_state *rsp = rdp->rsp;
2978 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2979 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2980 complete(&rsp->barrier_completion);
2982 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2987 * Called with preemption disabled, and from cross-cpu IRQ context.
2989 static void rcu_barrier_func(void *type)
2991 struct rcu_state *rsp = type;
2992 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2994 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2995 atomic_inc(&rsp->barrier_cpu_count);
2996 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3000 * Orchestrate the specified type of RCU barrier, waiting for all
3001 * RCU callbacks of the specified type to complete.
3003 static void _rcu_barrier(struct rcu_state *rsp)
3006 struct rcu_data *rdp;
3007 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3008 unsigned long snap_done;
3010 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3012 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3013 mutex_lock(&rsp->barrier_mutex);
3016 * Ensure that all prior references, including to ->n_barrier_done,
3017 * are ordered before the _rcu_barrier() machinery.
3019 smp_mb(); /* See above block comment. */
3022 * Recheck ->n_barrier_done to see if others did our work for us.
3023 * This means checking ->n_barrier_done for an even-to-odd-to-even
3024 * transition. The "if" expression below therefore rounds the old
3025 * value up to the next even number and adds two before comparing.
3027 snap_done = rsp->n_barrier_done;
3028 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3031 * If the value in snap is odd, we needed to wait for the current
3032 * rcu_barrier() to complete, then wait for the next one, in other
3033 * words, we need the value of snap_done to be three larger than
3034 * the value of snap. On the other hand, if the value in snap is
3035 * even, we only had to wait for the next rcu_barrier() to complete,
3036 * in other words, we need the value of snap_done to be only two
3037 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3038 * this for us (thank you, Linus!).
3040 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3041 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3042 smp_mb(); /* caller's subsequent code after above check. */
3043 mutex_unlock(&rsp->barrier_mutex);
3048 * Increment ->n_barrier_done to avoid duplicate work. Use
3049 * ACCESS_ONCE() to prevent the compiler from speculating
3050 * the increment to precede the early-exit check.
3052 ACCESS_ONCE(rsp->n_barrier_done)++;
3053 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3054 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3055 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3058 * Initialize the count to one rather than to zero in order to
3059 * avoid a too-soon return to zero in case of a short grace period
3060 * (or preemption of this task). Exclude CPU-hotplug operations
3061 * to ensure that no offline CPU has callbacks queued.
3063 init_completion(&rsp->barrier_completion);
3064 atomic_set(&rsp->barrier_cpu_count, 1);
3068 * Force each CPU with callbacks to register a new callback.
3069 * When that callback is invoked, we will know that all of the
3070 * corresponding CPU's preceding callbacks have been invoked.
3072 for_each_possible_cpu(cpu) {
3073 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3075 rdp = per_cpu_ptr(rsp->rda, cpu);
3076 if (rcu_is_nocb_cpu(cpu)) {
3077 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3078 rsp->n_barrier_done);
3079 atomic_inc(&rsp->barrier_cpu_count);
3080 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3082 } else if (ACCESS_ONCE(rdp->qlen)) {
3083 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3084 rsp->n_barrier_done);
3085 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3087 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3088 rsp->n_barrier_done);
3094 * Now that we have an rcu_barrier_callback() callback on each
3095 * CPU, and thus each counted, remove the initial count.
3097 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3098 complete(&rsp->barrier_completion);
3100 /* Increment ->n_barrier_done to prevent duplicate work. */
3101 smp_mb(); /* Keep increment after above mechanism. */
3102 ACCESS_ONCE(rsp->n_barrier_done)++;
3103 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3104 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3105 smp_mb(); /* Keep increment before caller's subsequent code. */
3107 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3108 wait_for_completion(&rsp->barrier_completion);
3110 /* Other rcu_barrier() invocations can now safely proceed. */
3111 mutex_unlock(&rsp->barrier_mutex);
3115 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3117 void rcu_barrier_bh(void)
3119 _rcu_barrier(&rcu_bh_state);
3121 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3124 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3126 void rcu_barrier_sched(void)
3128 _rcu_barrier(&rcu_sched_state);
3130 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3133 * Do boot-time initialization of a CPU's per-CPU RCU data.
3136 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3138 unsigned long flags;
3139 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3140 struct rcu_node *rnp = rcu_get_root(rsp);
3142 /* Set up local state, ensuring consistent view of global state. */
3143 raw_spin_lock_irqsave(&rnp->lock, flags);
3144 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3145 init_callback_list(rdp);
3147 ACCESS_ONCE(rdp->qlen) = 0;
3148 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3149 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3150 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3153 rcu_boot_init_nocb_percpu_data(rdp);
3154 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3158 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3159 * offline event can be happening at a given time. Note also that we
3160 * can accept some slop in the rsp->completed access due to the fact
3161 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3164 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3166 unsigned long flags;
3168 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3169 struct rcu_node *rnp = rcu_get_root(rsp);
3171 /* Exclude new grace periods. */
3172 mutex_lock(&rsp->onoff_mutex);
3174 /* Set up local state, ensuring consistent view of global state. */
3175 raw_spin_lock_irqsave(&rnp->lock, flags);
3176 rdp->beenonline = 1; /* We have now been online. */
3177 rdp->preemptible = preemptible;
3178 rdp->qlen_last_fqs_check = 0;
3179 rdp->n_force_qs_snap = rsp->n_force_qs;
3180 rdp->blimit = blimit;
3181 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3182 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3183 rcu_sysidle_init_percpu_data(rdp->dynticks);
3184 atomic_set(&rdp->dynticks->dynticks,
3185 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3186 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3188 /* Add CPU to rcu_node bitmasks. */
3190 mask = rdp->grpmask;
3192 /* Exclude any attempts to start a new GP on small systems. */
3193 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3194 rnp->qsmaskinit |= mask;
3195 mask = rnp->grpmask;
3196 if (rnp == rdp->mynode) {
3198 * If there is a grace period in progress, we will
3199 * set up to wait for it next time we run the
3202 rdp->gpnum = rnp->completed;
3203 rdp->completed = rnp->completed;
3204 rdp->passed_quiesce = 0;
3205 rdp->qs_pending = 0;
3206 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3208 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3210 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3211 local_irq_restore(flags);
3213 mutex_unlock(&rsp->onoff_mutex);
3216 static void rcu_prepare_cpu(int cpu)
3218 struct rcu_state *rsp;
3220 for_each_rcu_flavor(rsp)
3221 rcu_init_percpu_data(cpu, rsp,
3222 strcmp(rsp->name, "rcu_preempt") == 0);
3226 * Handle CPU online/offline notification events.
3228 static int rcu_cpu_notify(struct notifier_block *self,
3229 unsigned long action, void *hcpu)
3231 long cpu = (long)hcpu;
3232 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3233 struct rcu_node *rnp = rdp->mynode;
3234 struct rcu_state *rsp;
3236 trace_rcu_utilization(TPS("Start CPU hotplug"));
3238 case CPU_UP_PREPARE:
3239 case CPU_UP_PREPARE_FROZEN:
3240 rcu_prepare_cpu(cpu);
3241 rcu_prepare_kthreads(cpu);
3244 case CPU_DOWN_FAILED:
3245 rcu_boost_kthread_setaffinity(rnp, -1);
3247 case CPU_DOWN_PREPARE:
3248 rcu_boost_kthread_setaffinity(rnp, cpu);
3251 case CPU_DYING_FROZEN:
3252 for_each_rcu_flavor(rsp)
3253 rcu_cleanup_dying_cpu(rsp);
3256 case CPU_DEAD_FROZEN:
3257 case CPU_UP_CANCELED:
3258 case CPU_UP_CANCELED_FROZEN:
3259 for_each_rcu_flavor(rsp)
3260 rcu_cleanup_dead_cpu(cpu, rsp);
3265 trace_rcu_utilization(TPS("End CPU hotplug"));
3269 static int rcu_pm_notify(struct notifier_block *self,
3270 unsigned long action, void *hcpu)
3273 case PM_HIBERNATION_PREPARE:
3274 case PM_SUSPEND_PREPARE:
3275 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3278 case PM_POST_HIBERNATION:
3279 case PM_POST_SUSPEND:
3289 * Spawn the kthread that handles this RCU flavor's grace periods.
3291 static int __init rcu_spawn_gp_kthread(void)
3293 unsigned long flags;
3294 struct rcu_node *rnp;
3295 struct rcu_state *rsp;
3296 struct task_struct *t;
3298 for_each_rcu_flavor(rsp) {
3299 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3301 rnp = rcu_get_root(rsp);
3302 raw_spin_lock_irqsave(&rnp->lock, flags);
3303 rsp->gp_kthread = t;
3304 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3305 rcu_spawn_nocb_kthreads(rsp);
3309 early_initcall(rcu_spawn_gp_kthread);
3312 * This function is invoked towards the end of the scheduler's initialization
3313 * process. Before this is called, the idle task might contain
3314 * RCU read-side critical sections (during which time, this idle
3315 * task is booting the system). After this function is called, the
3316 * idle tasks are prohibited from containing RCU read-side critical
3317 * sections. This function also enables RCU lockdep checking.
3319 void rcu_scheduler_starting(void)
3321 WARN_ON(num_online_cpus() != 1);
3322 WARN_ON(nr_context_switches() > 0);
3323 rcu_scheduler_active = 1;
3327 * Compute the per-level fanout, either using the exact fanout specified
3328 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3330 #ifdef CONFIG_RCU_FANOUT_EXACT
3331 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3335 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3336 for (i = rcu_num_lvls - 2; i >= 0; i--)
3337 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3339 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3340 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3347 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3348 ccur = rsp->levelcnt[i];
3349 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3353 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3356 * Helper function for rcu_init() that initializes one rcu_state structure.
3358 static void __init rcu_init_one(struct rcu_state *rsp,
3359 struct rcu_data __percpu *rda)
3361 static char *buf[] = { "rcu_node_0",
3364 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3365 static char *fqs[] = { "rcu_node_fqs_0",
3368 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3372 struct rcu_node *rnp;
3374 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3376 /* Silence gcc 4.8 warning about array index out of range. */
3377 if (rcu_num_lvls > RCU_NUM_LVLS)
3378 panic("rcu_init_one: rcu_num_lvls overflow");
3380 /* Initialize the level-tracking arrays. */
3382 for (i = 0; i < rcu_num_lvls; i++)
3383 rsp->levelcnt[i] = num_rcu_lvl[i];
3384 for (i = 1; i < rcu_num_lvls; i++)
3385 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3386 rcu_init_levelspread(rsp);
3388 /* Initialize the elements themselves, starting from the leaves. */
3390 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3391 cpustride *= rsp->levelspread[i];
3392 rnp = rsp->level[i];
3393 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3394 raw_spin_lock_init(&rnp->lock);
3395 lockdep_set_class_and_name(&rnp->lock,
3396 &rcu_node_class[i], buf[i]);
3397 raw_spin_lock_init(&rnp->fqslock);
3398 lockdep_set_class_and_name(&rnp->fqslock,
3399 &rcu_fqs_class[i], fqs[i]);
3400 rnp->gpnum = rsp->gpnum;
3401 rnp->completed = rsp->completed;
3403 rnp->qsmaskinit = 0;
3404 rnp->grplo = j * cpustride;
3405 rnp->grphi = (j + 1) * cpustride - 1;
3406 if (rnp->grphi >= NR_CPUS)
3407 rnp->grphi = NR_CPUS - 1;
3413 rnp->grpnum = j % rsp->levelspread[i - 1];
3414 rnp->grpmask = 1UL << rnp->grpnum;
3415 rnp->parent = rsp->level[i - 1] +
3416 j / rsp->levelspread[i - 1];
3419 INIT_LIST_HEAD(&rnp->blkd_tasks);
3420 rcu_init_one_nocb(rnp);
3425 init_waitqueue_head(&rsp->gp_wq);
3426 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3427 rnp = rsp->level[rcu_num_lvls - 1];
3428 for_each_possible_cpu(i) {
3429 while (i > rnp->grphi)
3431 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3432 rcu_boot_init_percpu_data(i, rsp);
3434 list_add(&rsp->flavors, &rcu_struct_flavors);
3438 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3439 * replace the definitions in tree.h because those are needed to size
3440 * the ->node array in the rcu_state structure.
3442 static void __init rcu_init_geometry(void)
3448 int rcu_capacity[MAX_RCU_LVLS + 1];
3451 * Initialize any unspecified boot parameters.
3452 * The default values of jiffies_till_first_fqs and
3453 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3454 * value, which is a function of HZ, then adding one for each
3455 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3457 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3458 if (jiffies_till_first_fqs == ULONG_MAX)
3459 jiffies_till_first_fqs = d;
3460 if (jiffies_till_next_fqs == ULONG_MAX)
3461 jiffies_till_next_fqs = d;
3463 /* If the compile-time values are accurate, just leave. */
3464 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3465 nr_cpu_ids == NR_CPUS)
3467 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3468 rcu_fanout_leaf, nr_cpu_ids);
3471 * Compute number of nodes that can be handled an rcu_node tree
3472 * with the given number of levels. Setting rcu_capacity[0] makes
3473 * some of the arithmetic easier.
3475 rcu_capacity[0] = 1;
3476 rcu_capacity[1] = rcu_fanout_leaf;
3477 for (i = 2; i <= MAX_RCU_LVLS; i++)
3478 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3481 * The boot-time rcu_fanout_leaf parameter is only permitted
3482 * to increase the leaf-level fanout, not decrease it. Of course,
3483 * the leaf-level fanout cannot exceed the number of bits in
3484 * the rcu_node masks. Finally, the tree must be able to accommodate
3485 * the configured number of CPUs. Complain and fall back to the
3486 * compile-time values if these limits are exceeded.
3488 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3489 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3490 n > rcu_capacity[MAX_RCU_LVLS]) {
3495 /* Calculate the number of rcu_nodes at each level of the tree. */
3496 for (i = 1; i <= MAX_RCU_LVLS; i++)
3497 if (n <= rcu_capacity[i]) {
3498 for (j = 0; j <= i; j++)
3500 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3502 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3507 /* Calculate the total number of rcu_node structures. */
3509 for (i = 0; i <= MAX_RCU_LVLS; i++)
3510 rcu_num_nodes += num_rcu_lvl[i];
3514 void __init rcu_init(void)
3518 rcu_bootup_announce();
3519 rcu_init_geometry();
3520 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3521 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3522 __rcu_init_preempt();
3523 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3526 * We don't need protection against CPU-hotplug here because
3527 * this is called early in boot, before either interrupts
3528 * or the scheduler are operational.
3530 cpu_notifier(rcu_cpu_notify, 0);
3531 pm_notifier(rcu_pm_notify, 0);
3532 for_each_online_cpu(cpu)
3533 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3536 #include "tree_plugin.h"