2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/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>
61 #include <trace/events/rcu.h>
65 MODULE_ALIAS("rcutree");
66 #ifdef MODULE_PARAM_PREFIX
67 #undef MODULE_PARAM_PREFIX
69 #define MODULE_PARAM_PREFIX "rcutree."
71 /* Data structures. */
73 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
74 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
77 * In order to export the rcu_state name to the tracing tools, it
78 * needs to be added in the __tracepoint_string section.
79 * This requires defining a separate variable tp_<sname>_varname
80 * that points to the string being used, and this will allow
81 * the tracing userspace tools to be able to decipher the string
82 * address to the matching string.
84 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
85 static char sname##_varname[] = #sname; \
86 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
87 struct rcu_state sname##_state = { \
88 .level = { &sname##_state.node[0] }, \
90 .fqs_state = RCU_GP_IDLE, \
91 .gpnum = 0UL - 300UL, \
92 .completed = 0UL - 300UL, \
93 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
94 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
95 .orphan_donetail = &sname##_state.orphan_donelist, \
96 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
97 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
98 .name = sname##_varname, \
101 DEFINE_PER_CPU(struct rcu_data, sname##_data)
103 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
104 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
106 static struct rcu_state *rcu_state;
107 LIST_HEAD(rcu_struct_flavors);
109 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
110 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
111 module_param(rcu_fanout_leaf, int, 0444);
112 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
113 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
120 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
123 * The rcu_scheduler_active variable transitions from zero to one just
124 * before the first task is spawned. So when this variable is zero, RCU
125 * can assume that there is but one task, allowing RCU to (for example)
126 * optimize synchronize_sched() to a simple barrier(). When this variable
127 * is one, RCU must actually do all the hard work required to detect real
128 * grace periods. This variable is also used to suppress boot-time false
129 * positives from lockdep-RCU error checking.
131 int rcu_scheduler_active __read_mostly;
132 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
135 * The rcu_scheduler_fully_active variable transitions from zero to one
136 * during the early_initcall() processing, which is after the scheduler
137 * is capable of creating new tasks. So RCU processing (for example,
138 * creating tasks for RCU priority boosting) must be delayed until after
139 * rcu_scheduler_fully_active transitions from zero to one. We also
140 * currently delay invocation of any RCU callbacks until after this point.
142 * It might later prove better for people registering RCU callbacks during
143 * early boot to take responsibility for these callbacks, but one step at
146 static int rcu_scheduler_fully_active __read_mostly;
148 #ifdef CONFIG_RCU_BOOST
151 * Control variables for per-CPU and per-rcu_node kthreads. These
152 * handle all flavors of RCU.
154 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
155 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
156 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
157 DEFINE_PER_CPU(char, rcu_cpu_has_work);
159 #endif /* #ifdef CONFIG_RCU_BOOST */
161 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
162 static void invoke_rcu_core(void);
163 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
166 * Track the rcutorture test sequence number and the update version
167 * number within a given test. The rcutorture_testseq is incremented
168 * on every rcutorture module load and unload, so has an odd value
169 * when a test is running. The rcutorture_vernum is set to zero
170 * when rcutorture starts and is incremented on each rcutorture update.
171 * These variables enable correlating rcutorture output with the
172 * RCU tracing information.
174 unsigned long rcutorture_testseq;
175 unsigned long rcutorture_vernum;
178 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
179 * permit this function to be invoked without holding the root rcu_node
180 * structure's ->lock, but of course results can be subject to change.
182 static int rcu_gp_in_progress(struct rcu_state *rsp)
184 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
188 * Note a quiescent state. Because we do not need to know
189 * how many quiescent states passed, just if there was at least
190 * one since the start of the grace period, this just sets a flag.
191 * The caller must have disabled preemption.
193 void rcu_sched_qs(int cpu)
195 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
197 if (rdp->passed_quiesce == 0)
198 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
199 rdp->passed_quiesce = 1;
202 void rcu_bh_qs(int cpu)
204 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
206 if (rdp->passed_quiesce == 0)
207 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
208 rdp->passed_quiesce = 1;
212 * Note a context switch. This is a quiescent state for RCU-sched,
213 * and requires special handling for preemptible RCU.
214 * The caller must have disabled preemption.
216 void rcu_note_context_switch(int cpu)
218 trace_rcu_utilization(TPS("Start context switch"));
220 rcu_preempt_note_context_switch(cpu);
221 trace_rcu_utilization(TPS("End context switch"));
223 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
225 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
226 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
227 .dynticks = ATOMIC_INIT(1),
228 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
229 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
230 .dynticks_idle = ATOMIC_INIT(1),
231 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
234 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
235 static long qhimark = 10000; /* If this many pending, ignore blimit. */
236 static long qlowmark = 100; /* Once only this many pending, use blimit. */
238 module_param(blimit, long, 0444);
239 module_param(qhimark, long, 0444);
240 module_param(qlowmark, long, 0444);
242 static ulong jiffies_till_first_fqs = ULONG_MAX;
243 static ulong jiffies_till_next_fqs = ULONG_MAX;
245 module_param(jiffies_till_first_fqs, ulong, 0644);
246 module_param(jiffies_till_next_fqs, ulong, 0644);
248 static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
249 struct rcu_data *rdp);
250 static void force_qs_rnp(struct rcu_state *rsp,
251 int (*f)(struct rcu_data *rsp, bool *isidle,
252 unsigned long *maxj),
253 bool *isidle, unsigned long *maxj);
254 static void force_quiescent_state(struct rcu_state *rsp);
255 static int rcu_pending(int cpu);
258 * Return the number of RCU-sched batches processed thus far for debug & stats.
260 long rcu_batches_completed_sched(void)
262 return rcu_sched_state.completed;
264 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
267 * Return the number of RCU BH batches processed thus far for debug & stats.
269 long rcu_batches_completed_bh(void)
271 return rcu_bh_state.completed;
273 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
276 * Force a quiescent state for RCU BH.
278 void rcu_bh_force_quiescent_state(void)
280 force_quiescent_state(&rcu_bh_state);
282 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
285 * Record the number of times rcutorture tests have been initiated and
286 * terminated. This information allows the debugfs tracing stats to be
287 * correlated to the rcutorture messages, even when the rcutorture module
288 * is being repeatedly loaded and unloaded. In other words, we cannot
289 * store this state in rcutorture itself.
291 void rcutorture_record_test_transition(void)
293 rcutorture_testseq++;
294 rcutorture_vernum = 0;
296 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
299 * Record the number of writer passes through the current rcutorture test.
300 * This is also used to correlate debugfs tracing stats with the rcutorture
303 void rcutorture_record_progress(unsigned long vernum)
307 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
310 * Force a quiescent state for RCU-sched.
312 void rcu_sched_force_quiescent_state(void)
314 force_quiescent_state(&rcu_sched_state);
316 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
319 * Does the CPU have callbacks ready to be invoked?
322 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
324 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
325 rdp->nxttail[RCU_DONE_TAIL] != NULL;
329 * Does the current CPU require a not-yet-started grace period?
330 * The caller must have disabled interrupts to prevent races with
331 * normal callback registry.
334 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
338 if (rcu_gp_in_progress(rsp))
339 return 0; /* No, a grace period is already in progress. */
340 if (rcu_nocb_needs_gp(rsp))
341 return 1; /* Yes, a no-CBs CPU needs one. */
342 if (!rdp->nxttail[RCU_NEXT_TAIL])
343 return 0; /* No, this is a no-CBs (or offline) CPU. */
344 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
345 return 1; /* Yes, this CPU has newly registered callbacks. */
346 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
347 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
348 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
349 rdp->nxtcompleted[i]))
350 return 1; /* Yes, CBs for future grace period. */
351 return 0; /* No grace period needed. */
355 * Return the root node of the specified rcu_state structure.
357 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
359 return &rsp->node[0];
363 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
365 * If the new value of the ->dynticks_nesting counter now is zero,
366 * we really have entered idle, and must do the appropriate accounting.
367 * The caller must have disabled interrupts.
369 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
372 struct rcu_state *rsp;
373 struct rcu_data *rdp;
375 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
376 if (!user && !is_idle_task(current)) {
377 struct task_struct *idle __maybe_unused =
378 idle_task(smp_processor_id());
380 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
381 ftrace_dump(DUMP_ORIG);
382 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
383 current->pid, current->comm,
384 idle->pid, idle->comm); /* must be idle task! */
386 for_each_rcu_flavor(rsp) {
387 rdp = this_cpu_ptr(rsp->rda);
388 do_nocb_deferred_wakeup(rdp);
390 rcu_prepare_for_idle(smp_processor_id());
391 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
392 smp_mb__before_atomic_inc(); /* See above. */
393 atomic_inc(&rdtp->dynticks);
394 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
395 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
398 * It is illegal to enter an extended quiescent state while
399 * in an RCU read-side critical section.
401 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
402 "Illegal idle entry in RCU read-side critical section.");
403 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
404 "Illegal idle entry in RCU-bh read-side critical section.");
405 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
406 "Illegal idle entry in RCU-sched read-side critical section.");
410 * Enter an RCU extended quiescent state, which can be either the
411 * idle loop or adaptive-tickless usermode execution.
413 static void rcu_eqs_enter(bool user)
416 struct rcu_dynticks *rdtp;
418 rdtp = this_cpu_ptr(&rcu_dynticks);
419 oldval = rdtp->dynticks_nesting;
420 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
421 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
422 rdtp->dynticks_nesting = 0;
424 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
425 rcu_eqs_enter_common(rdtp, oldval, user);
429 * rcu_idle_enter - inform RCU that current CPU is entering idle
431 * Enter idle mode, in other words, -leave- the mode in which RCU
432 * read-side critical sections can occur. (Though RCU read-side
433 * critical sections can occur in irq handlers in idle, a possibility
434 * handled by irq_enter() and irq_exit().)
436 * We crowbar the ->dynticks_nesting field to zero to allow for
437 * the possibility of usermode upcalls having messed up our count
438 * of interrupt nesting level during the prior busy period.
440 void rcu_idle_enter(void)
444 local_irq_save(flags);
445 rcu_eqs_enter(false);
446 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
447 local_irq_restore(flags);
449 EXPORT_SYMBOL_GPL(rcu_idle_enter);
451 #ifdef CONFIG_RCU_USER_QS
453 * rcu_user_enter - inform RCU that we are resuming userspace.
455 * Enter RCU idle mode right before resuming userspace. No use of RCU
456 * is permitted between this call and rcu_user_exit(). This way the
457 * CPU doesn't need to maintain the tick for RCU maintenance purposes
458 * when the CPU runs in userspace.
460 void rcu_user_enter(void)
464 #endif /* CONFIG_RCU_USER_QS */
467 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
469 * Exit from an interrupt handler, which might possibly result in entering
470 * idle mode, in other words, leaving the mode in which read-side critical
471 * sections can occur.
473 * This code assumes that the idle loop never does anything that might
474 * result in unbalanced calls to irq_enter() and irq_exit(). If your
475 * architecture violates this assumption, RCU will give you what you
476 * deserve, good and hard. But very infrequently and irreproducibly.
478 * Use things like work queues to work around this limitation.
480 * You have been warned.
482 void rcu_irq_exit(void)
486 struct rcu_dynticks *rdtp;
488 local_irq_save(flags);
489 rdtp = this_cpu_ptr(&rcu_dynticks);
490 oldval = rdtp->dynticks_nesting;
491 rdtp->dynticks_nesting--;
492 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
493 if (rdtp->dynticks_nesting)
494 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
496 rcu_eqs_enter_common(rdtp, oldval, true);
497 rcu_sysidle_enter(rdtp, 1);
498 local_irq_restore(flags);
502 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
504 * If the new value of the ->dynticks_nesting counter was previously zero,
505 * we really have exited idle, and must do the appropriate accounting.
506 * The caller must have disabled interrupts.
508 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
511 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
512 atomic_inc(&rdtp->dynticks);
513 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
514 smp_mb__after_atomic_inc(); /* See above. */
515 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
516 rcu_cleanup_after_idle(smp_processor_id());
517 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
518 if (!user && !is_idle_task(current)) {
519 struct task_struct *idle __maybe_unused =
520 idle_task(smp_processor_id());
522 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
523 oldval, rdtp->dynticks_nesting);
524 ftrace_dump(DUMP_ORIG);
525 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
526 current->pid, current->comm,
527 idle->pid, idle->comm); /* must be idle task! */
532 * Exit an RCU extended quiescent state, which can be either the
533 * idle loop or adaptive-tickless usermode execution.
535 static void rcu_eqs_exit(bool user)
537 struct rcu_dynticks *rdtp;
540 rdtp = this_cpu_ptr(&rcu_dynticks);
541 oldval = rdtp->dynticks_nesting;
542 WARN_ON_ONCE(oldval < 0);
543 if (oldval & DYNTICK_TASK_NEST_MASK)
544 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
546 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
547 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(ACCESS_ONCE(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 = ACCESS_ONCE(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=%lu, c=%lu, q=%lu)\n",
936 smp_processor_id(), (long)(jiffies - rsp->gp_start),
937 rsp->gpnum, 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=%lu c=%lu q=%lu)\n",
975 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
976 if (!trigger_all_cpu_backtrace())
979 raw_spin_lock_irqsave(&rnp->lock, flags);
980 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
981 rsp->jiffies_stall = jiffies +
982 3 * rcu_jiffies_till_stall_check() + 3;
983 raw_spin_unlock_irqrestore(&rnp->lock, flags);
986 * Attempt to revive the RCU machinery by forcing a context switch.
988 * A context switch would normally allow the RCU state machine to make
989 * progress and it could be we're stuck in kernel space without context
990 * switches for an entirely unreasonable amount of time.
992 resched_cpu(smp_processor_id());
995 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
997 unsigned long completed;
1002 struct rcu_node *rnp;
1004 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1006 j = ACCESS_ONCE(jiffies);
1009 * Lots of memory barriers to reject false positives.
1011 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1012 * then rsp->gp_start, and finally rsp->completed. These values
1013 * are updated in the opposite order with memory barriers (or
1014 * equivalent) during grace-period initialization and cleanup.
1015 * Now, a false positive can occur if we get an new value of
1016 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1017 * the memory barriers, the only way that this can happen is if one
1018 * grace period ends and another starts between these two fetches.
1019 * Detect this by comparing rsp->completed with the previous fetch
1022 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1023 * and rsp->gp_start suffice to forestall false positives.
1025 gpnum = ACCESS_ONCE(rsp->gpnum);
1026 smp_rmb(); /* Pick up ->gpnum first... */
1027 js = ACCESS_ONCE(rsp->jiffies_stall);
1028 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1029 gps = ACCESS_ONCE(rsp->gp_start);
1030 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1031 completed = ACCESS_ONCE(rsp->completed);
1032 if (ULONG_CMP_GE(completed, gpnum) ||
1033 ULONG_CMP_LT(j, js) ||
1034 ULONG_CMP_GE(gps, js))
1035 return; /* No stall or GP completed since entering function. */
1037 if (rcu_gp_in_progress(rsp) &&
1038 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1040 /* We haven't checked in, so go dump stack. */
1041 print_cpu_stall(rsp);
1043 } else if (rcu_gp_in_progress(rsp) &&
1044 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1046 /* They had a few time units to dump stack, so complain. */
1047 print_other_cpu_stall(rsp);
1052 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1054 * Set the stall-warning timeout way off into the future, thus preventing
1055 * any RCU CPU stall-warning messages from appearing in the current set of
1056 * RCU grace periods.
1058 * The caller must disable hard irqs.
1060 void rcu_cpu_stall_reset(void)
1062 struct rcu_state *rsp;
1064 for_each_rcu_flavor(rsp)
1065 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1069 * Initialize the specified rcu_data structure's callback list to empty.
1071 static void init_callback_list(struct rcu_data *rdp)
1075 if (init_nocb_callback_list(rdp))
1077 rdp->nxtlist = NULL;
1078 for (i = 0; i < RCU_NEXT_SIZE; i++)
1079 rdp->nxttail[i] = &rdp->nxtlist;
1083 * Determine the value that ->completed will have at the end of the
1084 * next subsequent grace period. This is used to tag callbacks so that
1085 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1086 * been dyntick-idle for an extended period with callbacks under the
1087 * influence of RCU_FAST_NO_HZ.
1089 * The caller must hold rnp->lock with interrupts disabled.
1091 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1092 struct rcu_node *rnp)
1095 * If RCU is idle, we just wait for the next grace period.
1096 * But we can only be sure that RCU is idle if we are looking
1097 * at the root rcu_node structure -- otherwise, a new grace
1098 * period might have started, but just not yet gotten around
1099 * to initializing the current non-root rcu_node structure.
1101 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1102 return rnp->completed + 1;
1105 * Otherwise, wait for a possible partial grace period and
1106 * then the subsequent full grace period.
1108 return rnp->completed + 2;
1112 * Trace-event helper function for rcu_start_future_gp() and
1113 * rcu_nocb_wait_gp().
1115 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1116 unsigned long c, const char *s)
1118 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1119 rnp->completed, c, rnp->level,
1120 rnp->grplo, rnp->grphi, s);
1124 * Start some future grace period, as needed to handle newly arrived
1125 * callbacks. The required future grace periods are recorded in each
1126 * rcu_node structure's ->need_future_gp field.
1128 * The caller must hold the specified rcu_node structure's ->lock.
1130 static unsigned long __maybe_unused
1131 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1135 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1138 * Pick up grace-period number for new callbacks. If this
1139 * grace period is already marked as needed, return to the caller.
1141 c = rcu_cbs_completed(rdp->rsp, rnp);
1142 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1143 if (rnp->need_future_gp[c & 0x1]) {
1144 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1149 * If either this rcu_node structure or the root rcu_node structure
1150 * believe that a grace period is in progress, then we must wait
1151 * for the one following, which is in "c". Because our request
1152 * will be noticed at the end of the current grace period, we don't
1153 * need to explicitly start one.
1155 if (rnp->gpnum != rnp->completed ||
1156 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1157 rnp->need_future_gp[c & 0x1]++;
1158 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1163 * There might be no grace period in progress. If we don't already
1164 * hold it, acquire the root rcu_node structure's lock in order to
1165 * start one (if needed).
1167 if (rnp != rnp_root)
1168 raw_spin_lock(&rnp_root->lock);
1171 * Get a new grace-period number. If there really is no grace
1172 * period in progress, it will be smaller than the one we obtained
1173 * earlier. Adjust callbacks as needed. Note that even no-CBs
1174 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1176 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1177 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1178 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1179 rdp->nxtcompleted[i] = c;
1182 * If the needed for the required grace period is already
1183 * recorded, trace and leave.
1185 if (rnp_root->need_future_gp[c & 0x1]) {
1186 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1190 /* Record the need for the future grace period. */
1191 rnp_root->need_future_gp[c & 0x1]++;
1193 /* If a grace period is not already in progress, start one. */
1194 if (rnp_root->gpnum != rnp_root->completed) {
1195 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1197 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1198 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1201 if (rnp != rnp_root)
1202 raw_spin_unlock(&rnp_root->lock);
1207 * Clean up any old requests for the just-ended grace period. Also return
1208 * whether any additional grace periods have been requested. Also invoke
1209 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1210 * waiting for this grace period to complete.
1212 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1214 int c = rnp->completed;
1216 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1218 rcu_nocb_gp_cleanup(rsp, rnp);
1219 rnp->need_future_gp[c & 0x1] = 0;
1220 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1221 trace_rcu_future_gp(rnp, rdp, c,
1222 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1227 * If there is room, assign a ->completed number to any callbacks on
1228 * this CPU that have not already been assigned. Also accelerate any
1229 * callbacks that were previously assigned a ->completed number that has
1230 * since proven to be too conservative, which can happen if callbacks get
1231 * assigned a ->completed number while RCU is idle, but with reference to
1232 * a non-root rcu_node structure. This function is idempotent, so it does
1233 * not hurt to call it repeatedly.
1235 * The caller must hold rnp->lock with interrupts disabled.
1237 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1238 struct rcu_data *rdp)
1243 /* If the CPU has no callbacks, nothing to do. */
1244 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1248 * Starting from the sublist containing the callbacks most
1249 * recently assigned a ->completed number and working down, find the
1250 * first sublist that is not assignable to an upcoming grace period.
1251 * Such a sublist has something in it (first two tests) and has
1252 * a ->completed number assigned that will complete sooner than
1253 * the ->completed number for newly arrived callbacks (last test).
1255 * The key point is that any later sublist can be assigned the
1256 * same ->completed number as the newly arrived callbacks, which
1257 * means that the callbacks in any of these later sublist can be
1258 * grouped into a single sublist, whether or not they have already
1259 * been assigned a ->completed number.
1261 c = rcu_cbs_completed(rsp, rnp);
1262 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1263 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1264 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1268 * If there are no sublist for unassigned callbacks, leave.
1269 * At the same time, advance "i" one sublist, so that "i" will
1270 * index into the sublist where all the remaining callbacks should
1273 if (++i >= RCU_NEXT_TAIL)
1277 * Assign all subsequent callbacks' ->completed number to the next
1278 * full grace period and group them all in the sublist initially
1281 for (; i <= RCU_NEXT_TAIL; i++) {
1282 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1283 rdp->nxtcompleted[i] = c;
1285 /* Record any needed additional grace periods. */
1286 rcu_start_future_gp(rnp, rdp);
1288 /* Trace depending on how much we were able to accelerate. */
1289 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1290 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1292 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1296 * Move any callbacks whose grace period has completed to the
1297 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1298 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1299 * sublist. This function is idempotent, so it does not hurt to
1300 * invoke it repeatedly. As long as it is not invoked -too- often...
1302 * The caller must hold rnp->lock with interrupts disabled.
1304 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1305 struct rcu_data *rdp)
1309 /* If the CPU has no callbacks, nothing to do. */
1310 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1314 * Find all callbacks whose ->completed numbers indicate that they
1315 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1317 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1318 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1320 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1322 /* Clean up any sublist tail pointers that were misordered above. */
1323 for (j = RCU_WAIT_TAIL; j < i; j++)
1324 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1326 /* Copy down callbacks to fill in empty sublists. */
1327 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1328 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1330 rdp->nxttail[j] = rdp->nxttail[i];
1331 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1334 /* Classify any remaining callbacks. */
1335 rcu_accelerate_cbs(rsp, rnp, rdp);
1339 * Update CPU-local rcu_data state to record the beginnings and ends of
1340 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1341 * structure corresponding to the current CPU, and must have irqs disabled.
1343 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1345 /* Handle the ends of any preceding grace periods first. */
1346 if (rdp->completed == rnp->completed) {
1348 /* No grace period end, so just accelerate recent callbacks. */
1349 rcu_accelerate_cbs(rsp, rnp, rdp);
1353 /* Advance callbacks. */
1354 rcu_advance_cbs(rsp, rnp, rdp);
1356 /* Remember that we saw this grace-period completion. */
1357 rdp->completed = rnp->completed;
1358 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1361 if (rdp->gpnum != rnp->gpnum) {
1363 * If the current grace period is waiting for this CPU,
1364 * set up to detect a quiescent state, otherwise don't
1365 * go looking for one.
1367 rdp->gpnum = rnp->gpnum;
1368 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1369 rdp->passed_quiesce = 0;
1370 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1371 zero_cpu_stall_ticks(rdp);
1375 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1377 unsigned long flags;
1378 struct rcu_node *rnp;
1380 local_irq_save(flags);
1382 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1383 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1384 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1385 local_irq_restore(flags);
1388 __note_gp_changes(rsp, rnp, rdp);
1389 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1393 * Initialize a new grace period. Return 0 if no grace period required.
1395 static int rcu_gp_init(struct rcu_state *rsp)
1397 struct rcu_data *rdp;
1398 struct rcu_node *rnp = rcu_get_root(rsp);
1400 rcu_bind_gp_kthread();
1401 raw_spin_lock_irq(&rnp->lock);
1402 if (rsp->gp_flags == 0) {
1403 /* Spurious wakeup, tell caller to go back to sleep. */
1404 raw_spin_unlock_irq(&rnp->lock);
1407 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1409 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1411 * Grace period already in progress, don't start another.
1412 * Not supposed to be able to happen.
1414 raw_spin_unlock_irq(&rnp->lock);
1418 /* Advance to a new grace period and initialize state. */
1419 record_gp_stall_check_time(rsp);
1420 smp_wmb(); /* Record GP times before starting GP. */
1422 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1423 raw_spin_unlock_irq(&rnp->lock);
1425 /* Exclude any concurrent CPU-hotplug operations. */
1426 mutex_lock(&rsp->onoff_mutex);
1429 * Set the quiescent-state-needed bits in all the rcu_node
1430 * structures for all currently online CPUs in breadth-first order,
1431 * starting from the root rcu_node structure, relying on the layout
1432 * of the tree within the rsp->node[] array. Note that other CPUs
1433 * will access only the leaves of the hierarchy, thus seeing that no
1434 * grace period is in progress, at least until the corresponding
1435 * leaf node has been initialized. In addition, we have excluded
1436 * CPU-hotplug operations.
1438 * The grace period cannot complete until the initialization
1439 * process finishes, because this kthread handles both.
1441 rcu_for_each_node_breadth_first(rsp, rnp) {
1442 raw_spin_lock_irq(&rnp->lock);
1443 rdp = this_cpu_ptr(rsp->rda);
1444 rcu_preempt_check_blocked_tasks(rnp);
1445 rnp->qsmask = rnp->qsmaskinit;
1446 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1447 WARN_ON_ONCE(rnp->completed != rsp->completed);
1448 ACCESS_ONCE(rnp->completed) = rsp->completed;
1449 if (rnp == rdp->mynode)
1450 __note_gp_changes(rsp, rnp, rdp);
1451 rcu_preempt_boost_start_gp(rnp);
1452 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1453 rnp->level, rnp->grplo,
1454 rnp->grphi, rnp->qsmask);
1455 raw_spin_unlock_irq(&rnp->lock);
1456 #ifdef CONFIG_PROVE_RCU_DELAY
1457 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1458 system_state == SYSTEM_RUNNING)
1460 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1464 mutex_unlock(&rsp->onoff_mutex);
1469 * Do one round of quiescent-state forcing.
1471 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1473 int fqs_state = fqs_state_in;
1474 bool isidle = false;
1476 struct rcu_node *rnp = rcu_get_root(rsp);
1479 if (fqs_state == RCU_SAVE_DYNTICK) {
1480 /* Collect dyntick-idle snapshots. */
1481 if (is_sysidle_rcu_state(rsp)) {
1483 maxj = jiffies - ULONG_MAX / 4;
1485 force_qs_rnp(rsp, dyntick_save_progress_counter,
1487 rcu_sysidle_report_gp(rsp, isidle, maxj);
1488 fqs_state = RCU_FORCE_QS;
1490 /* Handle dyntick-idle and offline CPUs. */
1492 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1494 /* Clear flag to prevent immediate re-entry. */
1495 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1496 raw_spin_lock_irq(&rnp->lock);
1497 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1498 raw_spin_unlock_irq(&rnp->lock);
1504 * Clean up after the old grace period.
1506 static void rcu_gp_cleanup(struct rcu_state *rsp)
1508 unsigned long gp_duration;
1510 struct rcu_data *rdp;
1511 struct rcu_node *rnp = rcu_get_root(rsp);
1513 raw_spin_lock_irq(&rnp->lock);
1514 gp_duration = jiffies - rsp->gp_start;
1515 if (gp_duration > rsp->gp_max)
1516 rsp->gp_max = gp_duration;
1519 * We know the grace period is complete, but to everyone else
1520 * it appears to still be ongoing. But it is also the case
1521 * that to everyone else it looks like there is nothing that
1522 * they can do to advance the grace period. It is therefore
1523 * safe for us to drop the lock in order to mark the grace
1524 * period as completed in all of the rcu_node structures.
1526 raw_spin_unlock_irq(&rnp->lock);
1529 * Propagate new ->completed value to rcu_node structures so
1530 * that other CPUs don't have to wait until the start of the next
1531 * grace period to process their callbacks. This also avoids
1532 * some nasty RCU grace-period initialization races by forcing
1533 * the end of the current grace period to be completely recorded in
1534 * all of the rcu_node structures before the beginning of the next
1535 * grace period is recorded in any of the rcu_node structures.
1537 rcu_for_each_node_breadth_first(rsp, rnp) {
1538 raw_spin_lock_irq(&rnp->lock);
1539 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1540 rdp = this_cpu_ptr(rsp->rda);
1541 if (rnp == rdp->mynode)
1542 __note_gp_changes(rsp, rnp, rdp);
1543 /* smp_mb() provided by prior unlock-lock pair. */
1544 nocb += rcu_future_gp_cleanup(rsp, rnp);
1545 raw_spin_unlock_irq(&rnp->lock);
1548 rnp = rcu_get_root(rsp);
1549 raw_spin_lock_irq(&rnp->lock);
1550 rcu_nocb_gp_set(rnp, nocb);
1552 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1553 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1554 rsp->fqs_state = RCU_GP_IDLE;
1555 rdp = this_cpu_ptr(rsp->rda);
1556 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1557 if (cpu_needs_another_gp(rsp, rdp)) {
1558 rsp->gp_flags = RCU_GP_FLAG_INIT;
1559 trace_rcu_grace_period(rsp->name,
1560 ACCESS_ONCE(rsp->gpnum),
1563 raw_spin_unlock_irq(&rnp->lock);
1567 * Body of kthread that handles grace periods.
1569 static int __noreturn rcu_gp_kthread(void *arg)
1575 struct rcu_state *rsp = arg;
1576 struct rcu_node *rnp = rcu_get_root(rsp);
1580 /* Handle grace-period start. */
1582 trace_rcu_grace_period(rsp->name,
1583 ACCESS_ONCE(rsp->gpnum),
1585 wait_event_interruptible(rsp->gp_wq,
1586 ACCESS_ONCE(rsp->gp_flags) &
1588 /* Locking provides needed memory barrier. */
1589 if (rcu_gp_init(rsp))
1592 flush_signals(current);
1593 trace_rcu_grace_period(rsp->name,
1594 ACCESS_ONCE(rsp->gpnum),
1598 /* Handle quiescent-state forcing. */
1599 fqs_state = RCU_SAVE_DYNTICK;
1600 j = jiffies_till_first_fqs;
1603 jiffies_till_first_fqs = HZ;
1608 rsp->jiffies_force_qs = jiffies + j;
1609 trace_rcu_grace_period(rsp->name,
1610 ACCESS_ONCE(rsp->gpnum),
1612 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1613 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1615 (!ACCESS_ONCE(rnp->qsmask) &&
1616 !rcu_preempt_blocked_readers_cgp(rnp)),
1618 /* Locking provides needed memory barriers. */
1619 /* If grace period done, leave loop. */
1620 if (!ACCESS_ONCE(rnp->qsmask) &&
1621 !rcu_preempt_blocked_readers_cgp(rnp))
1623 /* If time for quiescent-state forcing, do it. */
1624 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1625 (gf & RCU_GP_FLAG_FQS)) {
1626 trace_rcu_grace_period(rsp->name,
1627 ACCESS_ONCE(rsp->gpnum),
1629 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1630 trace_rcu_grace_period(rsp->name,
1631 ACCESS_ONCE(rsp->gpnum),
1635 /* Deal with stray signal. */
1637 flush_signals(current);
1638 trace_rcu_grace_period(rsp->name,
1639 ACCESS_ONCE(rsp->gpnum),
1642 j = jiffies_till_next_fqs;
1645 jiffies_till_next_fqs = HZ;
1648 jiffies_till_next_fqs = 1;
1652 /* Handle grace-period end. */
1653 rcu_gp_cleanup(rsp);
1657 static void rsp_wakeup(struct irq_work *work)
1659 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1661 /* Wake up rcu_gp_kthread() to start the grace period. */
1662 wake_up(&rsp->gp_wq);
1666 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1667 * in preparation for detecting the next grace period. The caller must hold
1668 * the root node's ->lock and hard irqs must be disabled.
1670 * Note that it is legal for a dying CPU (which is marked as offline) to
1671 * invoke this function. This can happen when the dying CPU reports its
1675 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1676 struct rcu_data *rdp)
1678 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1680 * Either we have not yet spawned the grace-period
1681 * task, this CPU does not need another grace period,
1682 * or a grace period is already in progress.
1683 * Either way, don't start a new grace period.
1687 rsp->gp_flags = RCU_GP_FLAG_INIT;
1688 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1692 * We can't do wakeups while holding the rnp->lock, as that
1693 * could cause possible deadlocks with the rq->lock. Defer
1694 * the wakeup to interrupt context. And don't bother waking
1695 * up the running kthread.
1697 if (current != rsp->gp_kthread)
1698 irq_work_queue(&rsp->wakeup_work);
1702 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1703 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1704 * is invoked indirectly from rcu_advance_cbs(), which would result in
1705 * endless recursion -- or would do so if it wasn't for the self-deadlock
1706 * that is encountered beforehand.
1709 rcu_start_gp(struct rcu_state *rsp)
1711 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1712 struct rcu_node *rnp = rcu_get_root(rsp);
1715 * If there is no grace period in progress right now, any
1716 * callbacks we have up to this point will be satisfied by the
1717 * next grace period. Also, advancing the callbacks reduces the
1718 * probability of false positives from cpu_needs_another_gp()
1719 * resulting in pointless grace periods. So, advance callbacks
1720 * then start the grace period!
1722 rcu_advance_cbs(rsp, rnp, rdp);
1723 rcu_start_gp_advanced(rsp, rnp, rdp);
1727 * Report a full set of quiescent states to the specified rcu_state
1728 * data structure. This involves cleaning up after the prior grace
1729 * period and letting rcu_start_gp() start up the next grace period
1730 * if one is needed. Note that the caller must hold rnp->lock, which
1731 * is released before return.
1733 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1734 __releases(rcu_get_root(rsp)->lock)
1736 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1737 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1738 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1742 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1743 * Allows quiescent states for a group of CPUs to be reported at one go
1744 * to the specified rcu_node structure, though all the CPUs in the group
1745 * must be represented by the same rcu_node structure (which need not be
1746 * a leaf rcu_node structure, though it often will be). That structure's
1747 * lock must be held upon entry, and it is released before return.
1750 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1751 struct rcu_node *rnp, unsigned long flags)
1752 __releases(rnp->lock)
1754 struct rcu_node *rnp_c;
1756 /* Walk up the rcu_node hierarchy. */
1758 if (!(rnp->qsmask & mask)) {
1760 /* Our bit has already been cleared, so done. */
1761 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1764 rnp->qsmask &= ~mask;
1765 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1766 mask, rnp->qsmask, rnp->level,
1767 rnp->grplo, rnp->grphi,
1769 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1771 /* Other bits still set at this level, so done. */
1772 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1775 mask = rnp->grpmask;
1776 if (rnp->parent == NULL) {
1778 /* No more levels. Exit loop holding root lock. */
1782 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1785 raw_spin_lock_irqsave(&rnp->lock, flags);
1786 WARN_ON_ONCE(rnp_c->qsmask);
1790 * Get here if we are the last CPU to pass through a quiescent
1791 * state for this grace period. Invoke rcu_report_qs_rsp()
1792 * to clean up and start the next grace period if one is needed.
1794 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1798 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1799 * structure. This must be either called from the specified CPU, or
1800 * called when the specified CPU is known to be offline (and when it is
1801 * also known that no other CPU is concurrently trying to help the offline
1802 * CPU). The lastcomp argument is used to make sure we are still in the
1803 * grace period of interest. We don't want to end the current grace period
1804 * based on quiescent states detected in an earlier grace period!
1807 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1809 unsigned long flags;
1811 struct rcu_node *rnp;
1814 raw_spin_lock_irqsave(&rnp->lock, flags);
1815 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1816 rnp->completed == rnp->gpnum) {
1819 * The grace period in which this quiescent state was
1820 * recorded has ended, so don't report it upwards.
1821 * We will instead need a new quiescent state that lies
1822 * within the current grace period.
1824 rdp->passed_quiesce = 0; /* need qs for new gp. */
1825 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1828 mask = rdp->grpmask;
1829 if ((rnp->qsmask & mask) == 0) {
1830 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1832 rdp->qs_pending = 0;
1835 * This GP can't end until cpu checks in, so all of our
1836 * callbacks can be processed during the next GP.
1838 rcu_accelerate_cbs(rsp, rnp, rdp);
1840 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1845 * Check to see if there is a new grace period of which this CPU
1846 * is not yet aware, and if so, set up local rcu_data state for it.
1847 * Otherwise, see if this CPU has just passed through its first
1848 * quiescent state for this grace period, and record that fact if so.
1851 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1853 /* Check for grace-period ends and beginnings. */
1854 note_gp_changes(rsp, rdp);
1857 * Does this CPU still need to do its part for current grace period?
1858 * If no, return and let the other CPUs do their part as well.
1860 if (!rdp->qs_pending)
1864 * Was there a quiescent state since the beginning of the grace
1865 * period? If no, then exit and wait for the next call.
1867 if (!rdp->passed_quiesce)
1871 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1874 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1877 #ifdef CONFIG_HOTPLUG_CPU
1880 * Send the specified CPU's RCU callbacks to the orphanage. The
1881 * specified CPU must be offline, and the caller must hold the
1885 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1886 struct rcu_node *rnp, struct rcu_data *rdp)
1888 /* No-CBs CPUs do not have orphanable callbacks. */
1889 if (rcu_is_nocb_cpu(rdp->cpu))
1893 * Orphan the callbacks. First adjust the counts. This is safe
1894 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1895 * cannot be running now. Thus no memory barrier is required.
1897 if (rdp->nxtlist != NULL) {
1898 rsp->qlen_lazy += rdp->qlen_lazy;
1899 rsp->qlen += rdp->qlen;
1900 rdp->n_cbs_orphaned += rdp->qlen;
1902 ACCESS_ONCE(rdp->qlen) = 0;
1906 * Next, move those callbacks still needing a grace period to
1907 * the orphanage, where some other CPU will pick them up.
1908 * Some of the callbacks might have gone partway through a grace
1909 * period, but that is too bad. They get to start over because we
1910 * cannot assume that grace periods are synchronized across CPUs.
1911 * We don't bother updating the ->nxttail[] array yet, instead
1912 * we just reset the whole thing later on.
1914 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1915 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1916 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1917 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1921 * Then move the ready-to-invoke callbacks to the orphanage,
1922 * where some other CPU will pick them up. These will not be
1923 * required to pass though another grace period: They are done.
1925 if (rdp->nxtlist != NULL) {
1926 *rsp->orphan_donetail = rdp->nxtlist;
1927 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1930 /* Finally, initialize the rcu_data structure's list to empty. */
1931 init_callback_list(rdp);
1935 * Adopt the RCU callbacks from the specified rcu_state structure's
1936 * orphanage. The caller must hold the ->orphan_lock.
1938 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
1941 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1943 /* No-CBs CPUs are handled specially. */
1944 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
1947 /* Do the accounting first. */
1948 rdp->qlen_lazy += rsp->qlen_lazy;
1949 rdp->qlen += rsp->qlen;
1950 rdp->n_cbs_adopted += rsp->qlen;
1951 if (rsp->qlen_lazy != rsp->qlen)
1952 rcu_idle_count_callbacks_posted();
1957 * We do not need a memory barrier here because the only way we
1958 * can get here if there is an rcu_barrier() in flight is if
1959 * we are the task doing the rcu_barrier().
1962 /* First adopt the ready-to-invoke callbacks. */
1963 if (rsp->orphan_donelist != NULL) {
1964 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1965 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1966 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1967 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1968 rdp->nxttail[i] = rsp->orphan_donetail;
1969 rsp->orphan_donelist = NULL;
1970 rsp->orphan_donetail = &rsp->orphan_donelist;
1973 /* And then adopt the callbacks that still need a grace period. */
1974 if (rsp->orphan_nxtlist != NULL) {
1975 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1976 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1977 rsp->orphan_nxtlist = NULL;
1978 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1983 * Trace the fact that this CPU is going offline.
1985 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1987 RCU_TRACE(unsigned long mask);
1988 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1989 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1991 RCU_TRACE(mask = rdp->grpmask);
1992 trace_rcu_grace_period(rsp->name,
1993 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1998 * The CPU has been completely removed, and some other CPU is reporting
1999 * this fact from process context. Do the remainder of the cleanup,
2000 * including orphaning the outgoing CPU's RCU callbacks, and also
2001 * adopting them. There can only be one CPU hotplug operation at a time,
2002 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2004 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2006 unsigned long flags;
2008 int need_report = 0;
2009 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2010 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2012 /* Adjust any no-longer-needed kthreads. */
2013 rcu_boost_kthread_setaffinity(rnp, -1);
2015 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2017 /* Exclude any attempts to start a new grace period. */
2018 mutex_lock(&rsp->onoff_mutex);
2019 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2021 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2022 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2023 rcu_adopt_orphan_cbs(rsp, flags);
2025 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2026 mask = rdp->grpmask; /* rnp->grplo is constant. */
2028 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2029 rnp->qsmaskinit &= ~mask;
2030 if (rnp->qsmaskinit != 0) {
2031 if (rnp != rdp->mynode)
2032 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2035 if (rnp == rdp->mynode)
2036 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2038 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2039 mask = rnp->grpmask;
2041 } while (rnp != NULL);
2044 * We still hold the leaf rcu_node structure lock here, and
2045 * irqs are still disabled. The reason for this subterfuge is
2046 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2047 * held leads to deadlock.
2049 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2051 if (need_report & RCU_OFL_TASKS_NORM_GP)
2052 rcu_report_unblock_qs_rnp(rnp, flags);
2054 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2055 if (need_report & RCU_OFL_TASKS_EXP_GP)
2056 rcu_report_exp_rnp(rsp, rnp, true);
2057 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2058 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2059 cpu, rdp->qlen, rdp->nxtlist);
2060 init_callback_list(rdp);
2061 /* Disallow further callbacks on this CPU. */
2062 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2063 mutex_unlock(&rsp->onoff_mutex);
2066 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2068 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2072 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2076 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2079 * Invoke any RCU callbacks that have made it to the end of their grace
2080 * period. Thottle as specified by rdp->blimit.
2082 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2084 unsigned long flags;
2085 struct rcu_head *next, *list, **tail;
2086 long bl, count, count_lazy;
2089 /* If no callbacks are ready, just return. */
2090 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2091 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2092 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2093 need_resched(), is_idle_task(current),
2094 rcu_is_callbacks_kthread());
2099 * Extract the list of ready callbacks, disabling to prevent
2100 * races with call_rcu() from interrupt handlers.
2102 local_irq_save(flags);
2103 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2105 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2106 list = rdp->nxtlist;
2107 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2108 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2109 tail = rdp->nxttail[RCU_DONE_TAIL];
2110 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2111 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2112 rdp->nxttail[i] = &rdp->nxtlist;
2113 local_irq_restore(flags);
2115 /* Invoke callbacks. */
2116 count = count_lazy = 0;
2120 debug_rcu_head_unqueue(list);
2121 if (__rcu_reclaim(rsp->name, list))
2124 /* Stop only if limit reached and CPU has something to do. */
2125 if (++count >= bl &&
2127 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2131 local_irq_save(flags);
2132 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2133 is_idle_task(current),
2134 rcu_is_callbacks_kthread());
2136 /* Update count, and requeue any remaining callbacks. */
2138 *tail = rdp->nxtlist;
2139 rdp->nxtlist = list;
2140 for (i = 0; i < RCU_NEXT_SIZE; i++)
2141 if (&rdp->nxtlist == rdp->nxttail[i])
2142 rdp->nxttail[i] = tail;
2146 smp_mb(); /* List handling before counting for rcu_barrier(). */
2147 rdp->qlen_lazy -= count_lazy;
2148 ACCESS_ONCE(rdp->qlen) -= count;
2149 rdp->n_cbs_invoked += count;
2151 /* Reinstate batch limit if we have worked down the excess. */
2152 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2153 rdp->blimit = blimit;
2155 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2156 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2157 rdp->qlen_last_fqs_check = 0;
2158 rdp->n_force_qs_snap = rsp->n_force_qs;
2159 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2160 rdp->qlen_last_fqs_check = rdp->qlen;
2161 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2163 local_irq_restore(flags);
2165 /* Re-invoke RCU core processing if there are callbacks remaining. */
2166 if (cpu_has_callbacks_ready_to_invoke(rdp))
2171 * Check to see if this CPU is in a non-context-switch quiescent state
2172 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2173 * Also schedule RCU core processing.
2175 * This function must be called from hardirq context. It is normally
2176 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2177 * false, there is no point in invoking rcu_check_callbacks().
2179 void rcu_check_callbacks(int cpu, int user)
2181 trace_rcu_utilization(TPS("Start scheduler-tick"));
2182 increment_cpu_stall_ticks();
2183 if (user || rcu_is_cpu_rrupt_from_idle()) {
2186 * Get here if this CPU took its interrupt from user
2187 * mode or from the idle loop, and if this is not a
2188 * nested interrupt. In this case, the CPU is in
2189 * a quiescent state, so note it.
2191 * No memory barrier is required here because both
2192 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2193 * variables that other CPUs neither access nor modify,
2194 * at least not while the corresponding CPU is online.
2200 } else if (!in_softirq()) {
2203 * Get here if this CPU did not take its interrupt from
2204 * softirq, in other words, if it is not interrupting
2205 * a rcu_bh read-side critical section. This is an _bh
2206 * critical section, so note it.
2211 rcu_preempt_check_callbacks(cpu);
2212 if (rcu_pending(cpu))
2214 trace_rcu_utilization(TPS("End scheduler-tick"));
2218 * Scan the leaf rcu_node structures, processing dyntick state for any that
2219 * have not yet encountered a quiescent state, using the function specified.
2220 * Also initiate boosting for any threads blocked on the root rcu_node.
2222 * The caller must have suppressed start of new grace periods.
2224 static void force_qs_rnp(struct rcu_state *rsp,
2225 int (*f)(struct rcu_data *rsp, bool *isidle,
2226 unsigned long *maxj),
2227 bool *isidle, unsigned long *maxj)
2231 unsigned long flags;
2233 struct rcu_node *rnp;
2235 rcu_for_each_leaf_node(rsp, rnp) {
2238 raw_spin_lock_irqsave(&rnp->lock, flags);
2239 if (!rcu_gp_in_progress(rsp)) {
2240 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2243 if (rnp->qsmask == 0) {
2244 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2249 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2250 if ((rnp->qsmask & bit) != 0) {
2251 if ((rnp->qsmaskinit & bit) != 0)
2253 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2259 /* rcu_report_qs_rnp() releases rnp->lock. */
2260 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2263 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2265 rnp = rcu_get_root(rsp);
2266 if (rnp->qsmask == 0) {
2267 raw_spin_lock_irqsave(&rnp->lock, flags);
2268 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2273 * Force quiescent states on reluctant CPUs, and also detect which
2274 * CPUs are in dyntick-idle mode.
2276 static void force_quiescent_state(struct rcu_state *rsp)
2278 unsigned long flags;
2280 struct rcu_node *rnp;
2281 struct rcu_node *rnp_old = NULL;
2283 /* Funnel through hierarchy to reduce memory contention. */
2284 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2285 for (; rnp != NULL; rnp = rnp->parent) {
2286 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2287 !raw_spin_trylock(&rnp->fqslock);
2288 if (rnp_old != NULL)
2289 raw_spin_unlock(&rnp_old->fqslock);
2291 rsp->n_force_qs_lh++;
2296 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2298 /* Reached the root of the rcu_node tree, acquire lock. */
2299 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2300 raw_spin_unlock(&rnp_old->fqslock);
2301 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2302 rsp->n_force_qs_lh++;
2303 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2304 return; /* Someone beat us to it. */
2306 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2307 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2308 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2312 * This does the RCU core processing work for the specified rcu_state
2313 * and rcu_data structures. This may be called only from the CPU to
2314 * whom the rdp belongs.
2317 __rcu_process_callbacks(struct rcu_state *rsp)
2319 unsigned long flags;
2320 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2322 WARN_ON_ONCE(rdp->beenonline == 0);
2324 /* Update RCU state based on any recent quiescent states. */
2325 rcu_check_quiescent_state(rsp, rdp);
2327 /* Does this CPU require a not-yet-started grace period? */
2328 local_irq_save(flags);
2329 if (cpu_needs_another_gp(rsp, rdp)) {
2330 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2332 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2334 local_irq_restore(flags);
2337 /* If there are callbacks ready, invoke them. */
2338 if (cpu_has_callbacks_ready_to_invoke(rdp))
2339 invoke_rcu_callbacks(rsp, rdp);
2341 /* Do any needed deferred wakeups of rcuo kthreads. */
2342 do_nocb_deferred_wakeup(rdp);
2346 * Do RCU core processing for the current CPU.
2348 static void rcu_process_callbacks(struct softirq_action *unused)
2350 struct rcu_state *rsp;
2352 if (cpu_is_offline(smp_processor_id()))
2354 trace_rcu_utilization(TPS("Start RCU core"));
2355 for_each_rcu_flavor(rsp)
2356 __rcu_process_callbacks(rsp);
2357 trace_rcu_utilization(TPS("End RCU core"));
2361 * Schedule RCU callback invocation. If the specified type of RCU
2362 * does not support RCU priority boosting, just do a direct call,
2363 * otherwise wake up the per-CPU kernel kthread. Note that because we
2364 * are running on the current CPU with interrupts disabled, the
2365 * rcu_cpu_kthread_task cannot disappear out from under us.
2367 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2369 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2371 if (likely(!rsp->boost)) {
2372 rcu_do_batch(rsp, rdp);
2375 invoke_rcu_callbacks_kthread();
2378 static void invoke_rcu_core(void)
2380 if (cpu_online(smp_processor_id()))
2381 raise_softirq(RCU_SOFTIRQ);
2385 * Handle any core-RCU processing required by a call_rcu() invocation.
2387 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2388 struct rcu_head *head, unsigned long flags)
2391 * If called from an extended quiescent state, invoke the RCU
2392 * core in order to force a re-evaluation of RCU's idleness.
2394 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2397 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2398 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2402 * Force the grace period if too many callbacks or too long waiting.
2403 * Enforce hysteresis, and don't invoke force_quiescent_state()
2404 * if some other CPU has recently done so. Also, don't bother
2405 * invoking force_quiescent_state() if the newly enqueued callback
2406 * is the only one waiting for a grace period to complete.
2408 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2410 /* Are we ignoring a completed grace period? */
2411 note_gp_changes(rsp, rdp);
2413 /* Start a new grace period if one not already started. */
2414 if (!rcu_gp_in_progress(rsp)) {
2415 struct rcu_node *rnp_root = rcu_get_root(rsp);
2417 raw_spin_lock(&rnp_root->lock);
2419 raw_spin_unlock(&rnp_root->lock);
2421 /* Give the grace period a kick. */
2422 rdp->blimit = LONG_MAX;
2423 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2424 *rdp->nxttail[RCU_DONE_TAIL] != head)
2425 force_quiescent_state(rsp);
2426 rdp->n_force_qs_snap = rsp->n_force_qs;
2427 rdp->qlen_last_fqs_check = rdp->qlen;
2433 * RCU callback function to leak a callback.
2435 static void rcu_leak_callback(struct rcu_head *rhp)
2440 * Helper function for call_rcu() and friends. The cpu argument will
2441 * normally be -1, indicating "currently running CPU". It may specify
2442 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2443 * is expected to specify a CPU.
2446 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2447 struct rcu_state *rsp, int cpu, bool lazy)
2449 unsigned long flags;
2450 struct rcu_data *rdp;
2452 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2453 if (debug_rcu_head_queue(head)) {
2454 /* Probable double call_rcu(), so leak the callback. */
2455 ACCESS_ONCE(head->func) = rcu_leak_callback;
2456 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2463 * Opportunistically note grace-period endings and beginnings.
2464 * Note that we might see a beginning right after we see an
2465 * end, but never vice versa, since this CPU has to pass through
2466 * a quiescent state betweentimes.
2468 local_irq_save(flags);
2469 rdp = this_cpu_ptr(rsp->rda);
2471 /* Add the callback to our list. */
2472 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2476 rdp = per_cpu_ptr(rsp->rda, cpu);
2477 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2478 WARN_ON_ONCE(offline);
2479 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2480 local_irq_restore(flags);
2483 ACCESS_ONCE(rdp->qlen)++;
2487 rcu_idle_count_callbacks_posted();
2488 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2489 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2490 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2492 if (__is_kfree_rcu_offset((unsigned long)func))
2493 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2494 rdp->qlen_lazy, rdp->qlen);
2496 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2498 /* Go handle any RCU core processing required. */
2499 __call_rcu_core(rsp, rdp, head, flags);
2500 local_irq_restore(flags);
2504 * Queue an RCU-sched callback for invocation after a grace period.
2506 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2508 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2510 EXPORT_SYMBOL_GPL(call_rcu_sched);
2513 * Queue an RCU callback for invocation after a quicker grace period.
2515 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2517 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2519 EXPORT_SYMBOL_GPL(call_rcu_bh);
2522 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2523 * any blocking grace-period wait automatically implies a grace period
2524 * if there is only one CPU online at any point time during execution
2525 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2526 * occasionally incorrectly indicate that there are multiple CPUs online
2527 * when there was in fact only one the whole time, as this just adds
2528 * some overhead: RCU still operates correctly.
2530 static inline int rcu_blocking_is_gp(void)
2534 might_sleep(); /* Check for RCU read-side critical section. */
2536 ret = num_online_cpus() <= 1;
2542 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2544 * Control will return to the caller some time after a full rcu-sched
2545 * grace period has elapsed, in other words after all currently executing
2546 * rcu-sched read-side critical sections have completed. These read-side
2547 * critical sections are delimited by rcu_read_lock_sched() and
2548 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2549 * local_irq_disable(), and so on may be used in place of
2550 * rcu_read_lock_sched().
2552 * This means that all preempt_disable code sequences, including NMI and
2553 * non-threaded hardware-interrupt handlers, in progress on entry will
2554 * have completed before this primitive returns. However, this does not
2555 * guarantee that softirq handlers will have completed, since in some
2556 * kernels, these handlers can run in process context, and can block.
2558 * Note that this guarantee implies further memory-ordering guarantees.
2559 * On systems with more than one CPU, when synchronize_sched() returns,
2560 * each CPU is guaranteed to have executed a full memory barrier since the
2561 * end of its last RCU-sched read-side critical section whose beginning
2562 * preceded the call to synchronize_sched(). In addition, each CPU having
2563 * an RCU read-side critical section that extends beyond the return from
2564 * synchronize_sched() is guaranteed to have executed a full memory barrier
2565 * after the beginning of synchronize_sched() and before the beginning of
2566 * that RCU read-side critical section. Note that these guarantees include
2567 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2568 * that are executing in the kernel.
2570 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2571 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2572 * to have executed a full memory barrier during the execution of
2573 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2574 * again only if the system has more than one CPU).
2576 * This primitive provides the guarantees made by the (now removed)
2577 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2578 * guarantees that rcu_read_lock() sections will have completed.
2579 * In "classic RCU", these two guarantees happen to be one and
2580 * the same, but can differ in realtime RCU implementations.
2582 void synchronize_sched(void)
2584 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2585 !lock_is_held(&rcu_lock_map) &&
2586 !lock_is_held(&rcu_sched_lock_map),
2587 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2588 if (rcu_blocking_is_gp())
2591 synchronize_sched_expedited();
2593 wait_rcu_gp(call_rcu_sched);
2595 EXPORT_SYMBOL_GPL(synchronize_sched);
2598 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2600 * Control will return to the caller some time after a full rcu_bh grace
2601 * period has elapsed, in other words after all currently executing rcu_bh
2602 * read-side critical sections have completed. RCU read-side critical
2603 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2604 * and may be nested.
2606 * See the description of synchronize_sched() for more detailed information
2607 * on memory ordering guarantees.
2609 void synchronize_rcu_bh(void)
2611 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2612 !lock_is_held(&rcu_lock_map) &&
2613 !lock_is_held(&rcu_sched_lock_map),
2614 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2615 if (rcu_blocking_is_gp())
2618 synchronize_rcu_bh_expedited();
2620 wait_rcu_gp(call_rcu_bh);
2622 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2624 static int synchronize_sched_expedited_cpu_stop(void *data)
2627 * There must be a full memory barrier on each affected CPU
2628 * between the time that try_stop_cpus() is called and the
2629 * time that it returns.
2631 * In the current initial implementation of cpu_stop, the
2632 * above condition is already met when the control reaches
2633 * this point and the following smp_mb() is not strictly
2634 * necessary. Do smp_mb() anyway for documentation and
2635 * robustness against future implementation changes.
2637 smp_mb(); /* See above comment block. */
2642 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2644 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2645 * approach to force the grace period to end quickly. This consumes
2646 * significant time on all CPUs and is unfriendly to real-time workloads,
2647 * so is thus not recommended for any sort of common-case code. In fact,
2648 * if you are using synchronize_sched_expedited() in a loop, please
2649 * restructure your code to batch your updates, and then use a single
2650 * synchronize_sched() instead.
2652 * Note that it is illegal to call this function while holding any lock
2653 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2654 * to call this function from a CPU-hotplug notifier. Failing to observe
2655 * these restriction will result in deadlock.
2657 * This implementation can be thought of as an application of ticket
2658 * locking to RCU, with sync_sched_expedited_started and
2659 * sync_sched_expedited_done taking on the roles of the halves
2660 * of the ticket-lock word. Each task atomically increments
2661 * sync_sched_expedited_started upon entry, snapshotting the old value,
2662 * then attempts to stop all the CPUs. If this succeeds, then each
2663 * CPU will have executed a context switch, resulting in an RCU-sched
2664 * grace period. We are then done, so we use atomic_cmpxchg() to
2665 * update sync_sched_expedited_done to match our snapshot -- but
2666 * only if someone else has not already advanced past our snapshot.
2668 * On the other hand, if try_stop_cpus() fails, we check the value
2669 * of sync_sched_expedited_done. If it has advanced past our
2670 * initial snapshot, then someone else must have forced a grace period
2671 * some time after we took our snapshot. In this case, our work is
2672 * done for us, and we can simply return. Otherwise, we try again,
2673 * but keep our initial snapshot for purposes of checking for someone
2674 * doing our work for us.
2676 * If we fail too many times in a row, we fall back to synchronize_sched().
2678 void synchronize_sched_expedited(void)
2680 long firstsnap, s, snap;
2682 struct rcu_state *rsp = &rcu_sched_state;
2685 * If we are in danger of counter wrap, just do synchronize_sched().
2686 * By allowing sync_sched_expedited_started to advance no more than
2687 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2688 * that more than 3.5 billion CPUs would be required to force a
2689 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2690 * course be required on a 64-bit system.
2692 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2693 (ulong)atomic_long_read(&rsp->expedited_done) +
2695 synchronize_sched();
2696 atomic_long_inc(&rsp->expedited_wrap);
2701 * Take a ticket. Note that atomic_inc_return() implies a
2702 * full memory barrier.
2704 snap = atomic_long_inc_return(&rsp->expedited_start);
2707 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2710 * Each pass through the following loop attempts to force a
2711 * context switch on each CPU.
2713 while (try_stop_cpus(cpu_online_mask,
2714 synchronize_sched_expedited_cpu_stop,
2717 atomic_long_inc(&rsp->expedited_tryfail);
2719 /* Check to see if someone else did our work for us. */
2720 s = atomic_long_read(&rsp->expedited_done);
2721 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2722 /* ensure test happens before caller kfree */
2723 smp_mb__before_atomic_inc(); /* ^^^ */
2724 atomic_long_inc(&rsp->expedited_workdone1);
2728 /* No joy, try again later. Or just synchronize_sched(). */
2729 if (trycount++ < 10) {
2730 udelay(trycount * num_online_cpus());
2732 wait_rcu_gp(call_rcu_sched);
2733 atomic_long_inc(&rsp->expedited_normal);
2737 /* Recheck to see if someone else did our work for us. */
2738 s = atomic_long_read(&rsp->expedited_done);
2739 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2740 /* ensure test happens before caller kfree */
2741 smp_mb__before_atomic_inc(); /* ^^^ */
2742 atomic_long_inc(&rsp->expedited_workdone2);
2747 * Refetching sync_sched_expedited_started allows later
2748 * callers to piggyback on our grace period. We retry
2749 * after they started, so our grace period works for them,
2750 * and they started after our first try, so their grace
2751 * period works for us.
2754 snap = atomic_long_read(&rsp->expedited_start);
2755 smp_mb(); /* ensure read is before try_stop_cpus(). */
2757 atomic_long_inc(&rsp->expedited_stoppedcpus);
2760 * Everyone up to our most recent fetch is covered by our grace
2761 * period. Update the counter, but only if our work is still
2762 * relevant -- which it won't be if someone who started later
2763 * than we did already did their update.
2766 atomic_long_inc(&rsp->expedited_done_tries);
2767 s = atomic_long_read(&rsp->expedited_done);
2768 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2769 /* ensure test happens before caller kfree */
2770 smp_mb__before_atomic_inc(); /* ^^^ */
2771 atomic_long_inc(&rsp->expedited_done_lost);
2774 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2775 atomic_long_inc(&rsp->expedited_done_exit);
2779 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2782 * Check to see if there is any immediate RCU-related work to be done
2783 * by the current CPU, for the specified type of RCU, returning 1 if so.
2784 * The checks are in order of increasing expense: checks that can be
2785 * carried out against CPU-local state are performed first. However,
2786 * we must check for CPU stalls first, else we might not get a chance.
2788 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2790 struct rcu_node *rnp = rdp->mynode;
2792 rdp->n_rcu_pending++;
2794 /* Check for CPU stalls, if enabled. */
2795 check_cpu_stall(rsp, rdp);
2797 /* Is the RCU core waiting for a quiescent state from this CPU? */
2798 if (rcu_scheduler_fully_active &&
2799 rdp->qs_pending && !rdp->passed_quiesce) {
2800 rdp->n_rp_qs_pending++;
2801 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2802 rdp->n_rp_report_qs++;
2806 /* Does this CPU have callbacks ready to invoke? */
2807 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2808 rdp->n_rp_cb_ready++;
2812 /* Has RCU gone idle with this CPU needing another grace period? */
2813 if (cpu_needs_another_gp(rsp, rdp)) {
2814 rdp->n_rp_cpu_needs_gp++;
2818 /* Has another RCU grace period completed? */
2819 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2820 rdp->n_rp_gp_completed++;
2824 /* Has a new RCU grace period started? */
2825 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2826 rdp->n_rp_gp_started++;
2830 /* Does this CPU need a deferred NOCB wakeup? */
2831 if (rcu_nocb_need_deferred_wakeup(rdp)) {
2832 rdp->n_rp_nocb_defer_wakeup++;
2837 rdp->n_rp_need_nothing++;
2842 * Check to see if there is any immediate RCU-related work to be done
2843 * by the current CPU, returning 1 if so. This function is part of the
2844 * RCU implementation; it is -not- an exported member of the RCU API.
2846 static int rcu_pending(int cpu)
2848 struct rcu_state *rsp;
2850 for_each_rcu_flavor(rsp)
2851 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2857 * Return true if the specified CPU has any callback. If all_lazy is
2858 * non-NULL, store an indication of whether all callbacks are lazy.
2859 * (If there are no callbacks, all of them are deemed to be lazy.)
2861 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2865 struct rcu_data *rdp;
2866 struct rcu_state *rsp;
2868 for_each_rcu_flavor(rsp) {
2869 rdp = per_cpu_ptr(rsp->rda, cpu);
2873 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2884 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2885 * the compiler is expected to optimize this away.
2887 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2888 int cpu, unsigned long done)
2890 trace_rcu_barrier(rsp->name, s, cpu,
2891 atomic_read(&rsp->barrier_cpu_count), done);
2895 * RCU callback function for _rcu_barrier(). If we are last, wake
2896 * up the task executing _rcu_barrier().
2898 static void rcu_barrier_callback(struct rcu_head *rhp)
2900 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2901 struct rcu_state *rsp = rdp->rsp;
2903 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2904 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2905 complete(&rsp->barrier_completion);
2907 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2912 * Called with preemption disabled, and from cross-cpu IRQ context.
2914 static void rcu_barrier_func(void *type)
2916 struct rcu_state *rsp = type;
2917 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2919 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2920 atomic_inc(&rsp->barrier_cpu_count);
2921 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2925 * Orchestrate the specified type of RCU barrier, waiting for all
2926 * RCU callbacks of the specified type to complete.
2928 static void _rcu_barrier(struct rcu_state *rsp)
2931 struct rcu_data *rdp;
2932 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2933 unsigned long snap_done;
2935 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2937 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2938 mutex_lock(&rsp->barrier_mutex);
2941 * Ensure that all prior references, including to ->n_barrier_done,
2942 * are ordered before the _rcu_barrier() machinery.
2944 smp_mb(); /* See above block comment. */
2947 * Recheck ->n_barrier_done to see if others did our work for us.
2948 * This means checking ->n_barrier_done for an even-to-odd-to-even
2949 * transition. The "if" expression below therefore rounds the old
2950 * value up to the next even number and adds two before comparing.
2952 snap_done = rsp->n_barrier_done;
2953 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2956 * If the value in snap is odd, we needed to wait for the current
2957 * rcu_barrier() to complete, then wait for the next one, in other
2958 * words, we need the value of snap_done to be three larger than
2959 * the value of snap. On the other hand, if the value in snap is
2960 * even, we only had to wait for the next rcu_barrier() to complete,
2961 * in other words, we need the value of snap_done to be only two
2962 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
2963 * this for us (thank you, Linus!).
2965 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
2966 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2967 smp_mb(); /* caller's subsequent code after above check. */
2968 mutex_unlock(&rsp->barrier_mutex);
2973 * Increment ->n_barrier_done to avoid duplicate work. Use
2974 * ACCESS_ONCE() to prevent the compiler from speculating
2975 * the increment to precede the early-exit check.
2977 ACCESS_ONCE(rsp->n_barrier_done)++;
2978 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2979 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2980 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2983 * Initialize the count to one rather than to zero in order to
2984 * avoid a too-soon return to zero in case of a short grace period
2985 * (or preemption of this task). Exclude CPU-hotplug operations
2986 * to ensure that no offline CPU has callbacks queued.
2988 init_completion(&rsp->barrier_completion);
2989 atomic_set(&rsp->barrier_cpu_count, 1);
2993 * Force each CPU with callbacks to register a new callback.
2994 * When that callback is invoked, we will know that all of the
2995 * corresponding CPU's preceding callbacks have been invoked.
2997 for_each_possible_cpu(cpu) {
2998 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3000 rdp = per_cpu_ptr(rsp->rda, cpu);
3001 if (rcu_is_nocb_cpu(cpu)) {
3002 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3003 rsp->n_barrier_done);
3004 atomic_inc(&rsp->barrier_cpu_count);
3005 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3007 } else if (ACCESS_ONCE(rdp->qlen)) {
3008 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3009 rsp->n_barrier_done);
3010 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3012 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3013 rsp->n_barrier_done);
3019 * Now that we have an rcu_barrier_callback() callback on each
3020 * CPU, and thus each counted, remove the initial count.
3022 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3023 complete(&rsp->barrier_completion);
3025 /* Increment ->n_barrier_done to prevent duplicate work. */
3026 smp_mb(); /* Keep increment after above mechanism. */
3027 ACCESS_ONCE(rsp->n_barrier_done)++;
3028 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3029 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3030 smp_mb(); /* Keep increment before caller's subsequent code. */
3032 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3033 wait_for_completion(&rsp->barrier_completion);
3035 /* Other rcu_barrier() invocations can now safely proceed. */
3036 mutex_unlock(&rsp->barrier_mutex);
3040 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3042 void rcu_barrier_bh(void)
3044 _rcu_barrier(&rcu_bh_state);
3046 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3049 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3051 void rcu_barrier_sched(void)
3053 _rcu_barrier(&rcu_sched_state);
3055 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3058 * Do boot-time initialization of a CPU's per-CPU RCU data.
3061 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3063 unsigned long flags;
3064 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3065 struct rcu_node *rnp = rcu_get_root(rsp);
3067 /* Set up local state, ensuring consistent view of global state. */
3068 raw_spin_lock_irqsave(&rnp->lock, flags);
3069 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3070 init_callback_list(rdp);
3072 ACCESS_ONCE(rdp->qlen) = 0;
3073 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3074 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3075 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3078 rcu_boot_init_nocb_percpu_data(rdp);
3079 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3083 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3084 * offline event can be happening at a given time. Note also that we
3085 * can accept some slop in the rsp->completed access due to the fact
3086 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3089 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3091 unsigned long flags;
3093 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3094 struct rcu_node *rnp = rcu_get_root(rsp);
3096 /* Exclude new grace periods. */
3097 mutex_lock(&rsp->onoff_mutex);
3099 /* Set up local state, ensuring consistent view of global state. */
3100 raw_spin_lock_irqsave(&rnp->lock, flags);
3101 rdp->beenonline = 1; /* We have now been online. */
3102 rdp->preemptible = preemptible;
3103 rdp->qlen_last_fqs_check = 0;
3104 rdp->n_force_qs_snap = rsp->n_force_qs;
3105 rdp->blimit = blimit;
3106 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3107 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3108 rcu_sysidle_init_percpu_data(rdp->dynticks);
3109 atomic_set(&rdp->dynticks->dynticks,
3110 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3111 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3113 /* Add CPU to rcu_node bitmasks. */
3115 mask = rdp->grpmask;
3117 /* Exclude any attempts to start a new GP on small systems. */
3118 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3119 rnp->qsmaskinit |= mask;
3120 mask = rnp->grpmask;
3121 if (rnp == rdp->mynode) {
3123 * If there is a grace period in progress, we will
3124 * set up to wait for it next time we run the
3127 rdp->gpnum = rnp->completed;
3128 rdp->completed = rnp->completed;
3129 rdp->passed_quiesce = 0;
3130 rdp->qs_pending = 0;
3131 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3133 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3135 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3136 local_irq_restore(flags);
3138 mutex_unlock(&rsp->onoff_mutex);
3141 static void rcu_prepare_cpu(int cpu)
3143 struct rcu_state *rsp;
3145 for_each_rcu_flavor(rsp)
3146 rcu_init_percpu_data(cpu, rsp,
3147 strcmp(rsp->name, "rcu_preempt") == 0);
3151 * Handle CPU online/offline notification events.
3153 static int rcu_cpu_notify(struct notifier_block *self,
3154 unsigned long action, void *hcpu)
3156 long cpu = (long)hcpu;
3157 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3158 struct rcu_node *rnp = rdp->mynode;
3159 struct rcu_state *rsp;
3161 trace_rcu_utilization(TPS("Start CPU hotplug"));
3163 case CPU_UP_PREPARE:
3164 case CPU_UP_PREPARE_FROZEN:
3165 rcu_prepare_cpu(cpu);
3166 rcu_prepare_kthreads(cpu);
3169 case CPU_DOWN_FAILED:
3170 rcu_boost_kthread_setaffinity(rnp, -1);
3172 case CPU_DOWN_PREPARE:
3173 rcu_boost_kthread_setaffinity(rnp, cpu);
3176 case CPU_DYING_FROZEN:
3177 for_each_rcu_flavor(rsp)
3178 rcu_cleanup_dying_cpu(rsp);
3181 case CPU_DEAD_FROZEN:
3182 case CPU_UP_CANCELED:
3183 case CPU_UP_CANCELED_FROZEN:
3184 for_each_rcu_flavor(rsp)
3185 rcu_cleanup_dead_cpu(cpu, rsp);
3190 trace_rcu_utilization(TPS("End CPU hotplug"));
3194 static int rcu_pm_notify(struct notifier_block *self,
3195 unsigned long action, void *hcpu)
3198 case PM_HIBERNATION_PREPARE:
3199 case PM_SUSPEND_PREPARE:
3200 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3203 case PM_POST_HIBERNATION:
3204 case PM_POST_SUSPEND:
3214 * Spawn the kthread that handles this RCU flavor's grace periods.
3216 static int __init rcu_spawn_gp_kthread(void)
3218 unsigned long flags;
3219 struct rcu_node *rnp;
3220 struct rcu_state *rsp;
3221 struct task_struct *t;
3223 for_each_rcu_flavor(rsp) {
3224 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3226 rnp = rcu_get_root(rsp);
3227 raw_spin_lock_irqsave(&rnp->lock, flags);
3228 rsp->gp_kthread = t;
3229 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3230 rcu_spawn_nocb_kthreads(rsp);
3234 early_initcall(rcu_spawn_gp_kthread);
3237 * This function is invoked towards the end of the scheduler's initialization
3238 * process. Before this is called, the idle task might contain
3239 * RCU read-side critical sections (during which time, this idle
3240 * task is booting the system). After this function is called, the
3241 * idle tasks are prohibited from containing RCU read-side critical
3242 * sections. This function also enables RCU lockdep checking.
3244 void rcu_scheduler_starting(void)
3246 WARN_ON(num_online_cpus() != 1);
3247 WARN_ON(nr_context_switches() > 0);
3248 rcu_scheduler_active = 1;
3252 * Compute the per-level fanout, either using the exact fanout specified
3253 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3255 #ifdef CONFIG_RCU_FANOUT_EXACT
3256 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3260 for (i = rcu_num_lvls - 1; i > 0; i--)
3261 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3262 rsp->levelspread[0] = rcu_fanout_leaf;
3264 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3265 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3272 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3273 ccur = rsp->levelcnt[i];
3274 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3278 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3281 * Helper function for rcu_init() that initializes one rcu_state structure.
3283 static void __init rcu_init_one(struct rcu_state *rsp,
3284 struct rcu_data __percpu *rda)
3286 static char *buf[] = { "rcu_node_0",
3289 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3290 static char *fqs[] = { "rcu_node_fqs_0",
3293 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3297 struct rcu_node *rnp;
3299 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3301 /* Silence gcc 4.8 warning about array index out of range. */
3302 if (rcu_num_lvls > RCU_NUM_LVLS)
3303 panic("rcu_init_one: rcu_num_lvls overflow");
3305 /* Initialize the level-tracking arrays. */
3307 for (i = 0; i < rcu_num_lvls; i++)
3308 rsp->levelcnt[i] = num_rcu_lvl[i];
3309 for (i = 1; i < rcu_num_lvls; i++)
3310 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3311 rcu_init_levelspread(rsp);
3313 /* Initialize the elements themselves, starting from the leaves. */
3315 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3316 cpustride *= rsp->levelspread[i];
3317 rnp = rsp->level[i];
3318 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3319 raw_spin_lock_init(&rnp->lock);
3320 lockdep_set_class_and_name(&rnp->lock,
3321 &rcu_node_class[i], buf[i]);
3322 raw_spin_lock_init(&rnp->fqslock);
3323 lockdep_set_class_and_name(&rnp->fqslock,
3324 &rcu_fqs_class[i], fqs[i]);
3325 rnp->gpnum = rsp->gpnum;
3326 rnp->completed = rsp->completed;
3328 rnp->qsmaskinit = 0;
3329 rnp->grplo = j * cpustride;
3330 rnp->grphi = (j + 1) * cpustride - 1;
3331 if (rnp->grphi >= NR_CPUS)
3332 rnp->grphi = NR_CPUS - 1;
3338 rnp->grpnum = j % rsp->levelspread[i - 1];
3339 rnp->grpmask = 1UL << rnp->grpnum;
3340 rnp->parent = rsp->level[i - 1] +
3341 j / rsp->levelspread[i - 1];
3344 INIT_LIST_HEAD(&rnp->blkd_tasks);
3345 rcu_init_one_nocb(rnp);
3350 init_waitqueue_head(&rsp->gp_wq);
3351 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3352 rnp = rsp->level[rcu_num_lvls - 1];
3353 for_each_possible_cpu(i) {
3354 while (i > rnp->grphi)
3356 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3357 rcu_boot_init_percpu_data(i, rsp);
3359 list_add(&rsp->flavors, &rcu_struct_flavors);
3363 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3364 * replace the definitions in tree.h because those are needed to size
3365 * the ->node array in the rcu_state structure.
3367 static void __init rcu_init_geometry(void)
3373 int rcu_capacity[MAX_RCU_LVLS + 1];
3376 * Initialize any unspecified boot parameters.
3377 * The default values of jiffies_till_first_fqs and
3378 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3379 * value, which is a function of HZ, then adding one for each
3380 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3382 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3383 if (jiffies_till_first_fqs == ULONG_MAX)
3384 jiffies_till_first_fqs = d;
3385 if (jiffies_till_next_fqs == ULONG_MAX)
3386 jiffies_till_next_fqs = d;
3388 /* If the compile-time values are accurate, just leave. */
3389 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3390 nr_cpu_ids == NR_CPUS)
3394 * Compute number of nodes that can be handled an rcu_node tree
3395 * with the given number of levels. Setting rcu_capacity[0] makes
3396 * some of the arithmetic easier.
3398 rcu_capacity[0] = 1;
3399 rcu_capacity[1] = rcu_fanout_leaf;
3400 for (i = 2; i <= MAX_RCU_LVLS; i++)
3401 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3404 * The boot-time rcu_fanout_leaf parameter is only permitted
3405 * to increase the leaf-level fanout, not decrease it. Of course,
3406 * the leaf-level fanout cannot exceed the number of bits in
3407 * the rcu_node masks. Finally, the tree must be able to accommodate
3408 * the configured number of CPUs. Complain and fall back to the
3409 * compile-time values if these limits are exceeded.
3411 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3412 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3413 n > rcu_capacity[MAX_RCU_LVLS]) {
3418 /* Calculate the number of rcu_nodes at each level of the tree. */
3419 for (i = 1; i <= MAX_RCU_LVLS; i++)
3420 if (n <= rcu_capacity[i]) {
3421 for (j = 0; j <= i; j++)
3423 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3425 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3430 /* Calculate the total number of rcu_node structures. */
3432 for (i = 0; i <= MAX_RCU_LVLS; i++)
3433 rcu_num_nodes += num_rcu_lvl[i];
3437 void __init rcu_init(void)
3441 rcu_bootup_announce();
3442 rcu_init_geometry();
3443 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3444 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3445 __rcu_init_preempt();
3446 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3449 * We don't need protection against CPU-hotplug here because
3450 * this is called early in boot, before either interrupts
3451 * or the scheduler are operational.
3453 cpu_notifier(rcu_cpu_notify, 0);
3454 pm_notifier(rcu_pm_notify, 0);
3455 for_each_online_cpu(cpu)
3456 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3459 #include "tree_plugin.h"