2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
56 #include <linux/ftrace_event.h>
59 #include <trace/events/rcu.h>
64 * Strings used in tracepoints need to be exported via the
65 * tracing system such that tools like perf and trace-cmd can
66 * translate the string address pointers to actual text.
68 #define TPS(x) tracepoint_string(x)
70 /* Data structures. */
72 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
73 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
76 * In order to export the rcu_state name to the tracing tools, it
77 * needs to be added in the __tracepoint_string section.
78 * This requires defining a separate variable tp_<sname>_varname
79 * that points to the string being used, and this will allow
80 * the tracing userspace tools to be able to decipher the string
81 * address to the matching string.
83 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
86 struct rcu_state sname##_state = { \
87 .level = { &sname##_state.node[0] }, \
89 .fqs_state = RCU_GP_IDLE, \
90 .gpnum = 0UL - 300UL, \
91 .completed = 0UL - 300UL, \
92 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
93 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
94 .orphan_donetail = &sname##_state.orphan_donelist, \
95 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
96 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
97 .name = sname##_varname, \
100 DEFINE_PER_CPU(struct rcu_data, sname##_data)
102 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
103 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
105 static struct rcu_state
*rcu_state
;
106 LIST_HEAD(rcu_struct_flavors
);
108 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
109 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
110 module_param(rcu_fanout_leaf
, int, 0444);
111 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
112 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
119 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
122 * The rcu_scheduler_active variable transitions from zero to one just
123 * before the first task is spawned. So when this variable is zero, RCU
124 * can assume that there is but one task, allowing RCU to (for example)
125 * optimize synchronize_sched() to a simple barrier(). When this variable
126 * is one, RCU must actually do all the hard work required to detect real
127 * grace periods. This variable is also used to suppress boot-time false
128 * positives from lockdep-RCU error checking.
130 int rcu_scheduler_active __read_mostly
;
131 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
134 * The rcu_scheduler_fully_active variable transitions from zero to one
135 * during the early_initcall() processing, which is after the scheduler
136 * is capable of creating new tasks. So RCU processing (for example,
137 * creating tasks for RCU priority boosting) must be delayed until after
138 * rcu_scheduler_fully_active transitions from zero to one. We also
139 * currently delay invocation of any RCU callbacks until after this point.
141 * It might later prove better for people registering RCU callbacks during
142 * early boot to take responsibility for these callbacks, but one step at
145 static int rcu_scheduler_fully_active __read_mostly
;
147 #ifdef CONFIG_RCU_BOOST
150 * Control variables for per-CPU and per-rcu_node kthreads. These
151 * handle all flavors of RCU.
153 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
154 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
155 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
156 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
158 #endif /* #ifdef CONFIG_RCU_BOOST */
160 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
161 static void invoke_rcu_core(void);
162 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
165 * Track the rcutorture test sequence number and the update version
166 * number within a given test. The rcutorture_testseq is incremented
167 * on every rcutorture module load and unload, so has an odd value
168 * when a test is running. The rcutorture_vernum is set to zero
169 * when rcutorture starts and is incremented on each rcutorture update.
170 * These variables enable correlating rcutorture output with the
171 * RCU tracing information.
173 unsigned long rcutorture_testseq
;
174 unsigned long rcutorture_vernum
;
177 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
178 * permit this function to be invoked without holding the root rcu_node
179 * structure's ->lock, but of course results can be subject to change.
181 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
183 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
187 * Note a quiescent state. Because we do not need to know
188 * how many quiescent states passed, just if there was at least
189 * one since the start of the grace period, this just sets a flag.
190 * The caller must have disabled preemption.
192 void rcu_sched_qs(int cpu
)
194 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
196 if (rdp
->passed_quiesce
== 0)
197 trace_rcu_grace_period(TPS("rcu_sched"), rdp
->gpnum
, TPS("cpuqs"));
198 rdp
->passed_quiesce
= 1;
201 void rcu_bh_qs(int cpu
)
203 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
205 if (rdp
->passed_quiesce
== 0)
206 trace_rcu_grace_period(TPS("rcu_bh"), rdp
->gpnum
, TPS("cpuqs"));
207 rdp
->passed_quiesce
= 1;
211 * Note a context switch. This is a quiescent state for RCU-sched,
212 * and requires special handling for preemptible RCU.
213 * The caller must have disabled preemption.
215 void rcu_note_context_switch(int cpu
)
217 trace_rcu_utilization(TPS("Start context switch"));
219 rcu_preempt_note_context_switch(cpu
);
220 trace_rcu_utilization(TPS("End context switch"));
222 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
224 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
225 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
226 .dynticks
= ATOMIC_INIT(1),
229 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
230 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
231 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
233 module_param(blimit
, long, 0444);
234 module_param(qhimark
, long, 0444);
235 module_param(qlowmark
, long, 0444);
237 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
238 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
240 module_param(jiffies_till_first_fqs
, ulong
, 0644);
241 module_param(jiffies_till_next_fqs
, ulong
, 0644);
243 static void rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
244 struct rcu_data
*rdp
);
245 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*));
246 static void force_quiescent_state(struct rcu_state
*rsp
);
247 static int rcu_pending(int cpu
);
250 * Return the number of RCU-sched batches processed thus far for debug & stats.
252 long rcu_batches_completed_sched(void)
254 return rcu_sched_state
.completed
;
256 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
259 * Return the number of RCU BH batches processed thus far for debug & stats.
261 long rcu_batches_completed_bh(void)
263 return rcu_bh_state
.completed
;
265 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
268 * Force a quiescent state for RCU BH.
270 void rcu_bh_force_quiescent_state(void)
272 force_quiescent_state(&rcu_bh_state
);
274 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
277 * Record the number of times rcutorture tests have been initiated and
278 * terminated. This information allows the debugfs tracing stats to be
279 * correlated to the rcutorture messages, even when the rcutorture module
280 * is being repeatedly loaded and unloaded. In other words, we cannot
281 * store this state in rcutorture itself.
283 void rcutorture_record_test_transition(void)
285 rcutorture_testseq
++;
286 rcutorture_vernum
= 0;
288 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
291 * Record the number of writer passes through the current rcutorture test.
292 * This is also used to correlate debugfs tracing stats with the rcutorture
295 void rcutorture_record_progress(unsigned long vernum
)
299 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
302 * Force a quiescent state for RCU-sched.
304 void rcu_sched_force_quiescent_state(void)
306 force_quiescent_state(&rcu_sched_state
);
308 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
311 * Does the CPU have callbacks ready to be invoked?
314 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
316 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
317 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
321 * Does the current CPU require a not-yet-started grace period?
322 * The caller must have disabled interrupts to prevent races with
323 * normal callback registry.
326 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
330 if (rcu_gp_in_progress(rsp
))
331 return 0; /* No, a grace period is already in progress. */
332 if (rcu_nocb_needs_gp(rsp
))
333 return 1; /* Yes, a no-CBs CPU needs one. */
334 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
335 return 0; /* No, this is a no-CBs (or offline) CPU. */
336 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
337 return 1; /* Yes, this CPU has newly registered callbacks. */
338 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
339 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
340 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
341 rdp
->nxtcompleted
[i
]))
342 return 1; /* Yes, CBs for future grace period. */
343 return 0; /* No grace period needed. */
347 * Return the root node of the specified rcu_state structure.
349 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
351 return &rsp
->node
[0];
355 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
357 * If the new value of the ->dynticks_nesting counter now is zero,
358 * we really have entered idle, and must do the appropriate accounting.
359 * The caller must have disabled interrupts.
361 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
364 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
365 if (!user
&& !is_idle_task(current
)) {
366 struct task_struct
*idle
= idle_task(smp_processor_id());
368 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
369 ftrace_dump(DUMP_ORIG
);
370 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
371 current
->pid
, current
->comm
,
372 idle
->pid
, idle
->comm
); /* must be idle task! */
374 rcu_prepare_for_idle(smp_processor_id());
375 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
376 smp_mb__before_atomic_inc(); /* See above. */
377 atomic_inc(&rdtp
->dynticks
);
378 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
379 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
382 * It is illegal to enter an extended quiescent state while
383 * in an RCU read-side critical section.
385 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
386 "Illegal idle entry in RCU read-side critical section.");
387 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
388 "Illegal idle entry in RCU-bh read-side critical section.");
389 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
390 "Illegal idle entry in RCU-sched read-side critical section.");
394 * Enter an RCU extended quiescent state, which can be either the
395 * idle loop or adaptive-tickless usermode execution.
397 static void rcu_eqs_enter(bool user
)
400 struct rcu_dynticks
*rdtp
;
402 rdtp
= &__get_cpu_var(rcu_dynticks
);
403 oldval
= rdtp
->dynticks_nesting
;
404 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
405 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
406 rdtp
->dynticks_nesting
= 0;
408 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
409 rcu_eqs_enter_common(rdtp
, oldval
, user
);
413 * rcu_idle_enter - inform RCU that current CPU is entering idle
415 * Enter idle mode, in other words, -leave- the mode in which RCU
416 * read-side critical sections can occur. (Though RCU read-side
417 * critical sections can occur in irq handlers in idle, a possibility
418 * handled by irq_enter() and irq_exit().)
420 * We crowbar the ->dynticks_nesting field to zero to allow for
421 * the possibility of usermode upcalls having messed up our count
422 * of interrupt nesting level during the prior busy period.
424 void rcu_idle_enter(void)
428 local_irq_save(flags
);
429 rcu_eqs_enter(false);
430 local_irq_restore(flags
);
432 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
434 #ifdef CONFIG_RCU_USER_QS
436 * rcu_user_enter - inform RCU that we are resuming userspace.
438 * Enter RCU idle mode right before resuming userspace. No use of RCU
439 * is permitted between this call and rcu_user_exit(). This way the
440 * CPU doesn't need to maintain the tick for RCU maintenance purposes
441 * when the CPU runs in userspace.
443 void rcu_user_enter(void)
449 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
450 * after the current irq returns.
452 * This is similar to rcu_user_enter() but in the context of a non-nesting
453 * irq. After this call, RCU enters into idle mode when the interrupt
456 void rcu_user_enter_after_irq(void)
459 struct rcu_dynticks
*rdtp
;
461 local_irq_save(flags
);
462 rdtp
= &__get_cpu_var(rcu_dynticks
);
463 /* Ensure this irq is interrupting a non-idle RCU state. */
464 WARN_ON_ONCE(!(rdtp
->dynticks_nesting
& DYNTICK_TASK_MASK
));
465 rdtp
->dynticks_nesting
= 1;
466 local_irq_restore(flags
);
468 #endif /* CONFIG_RCU_USER_QS */
471 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
473 * Exit from an interrupt handler, which might possibly result in entering
474 * idle mode, in other words, leaving the mode in which read-side critical
475 * sections can occur.
477 * This code assumes that the idle loop never does anything that might
478 * result in unbalanced calls to irq_enter() and irq_exit(). If your
479 * architecture violates this assumption, RCU will give you what you
480 * deserve, good and hard. But very infrequently and irreproducibly.
482 * Use things like work queues to work around this limitation.
484 * You have been warned.
486 void rcu_irq_exit(void)
490 struct rcu_dynticks
*rdtp
;
492 local_irq_save(flags
);
493 rdtp
= &__get_cpu_var(rcu_dynticks
);
494 oldval
= rdtp
->dynticks_nesting
;
495 rdtp
->dynticks_nesting
--;
496 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
497 if (rdtp
->dynticks_nesting
)
498 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
500 rcu_eqs_enter_common(rdtp
, oldval
, true);
501 local_irq_restore(flags
);
505 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
507 * If the new value of the ->dynticks_nesting counter was previously zero,
508 * we really have exited idle, and must do the appropriate accounting.
509 * The caller must have disabled interrupts.
511 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
514 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
515 atomic_inc(&rdtp
->dynticks
);
516 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
517 smp_mb__after_atomic_inc(); /* See above. */
518 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
519 rcu_cleanup_after_idle(smp_processor_id());
520 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
521 if (!user
&& !is_idle_task(current
)) {
522 struct task_struct
*idle
= idle_task(smp_processor_id());
524 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
525 oldval
, rdtp
->dynticks_nesting
);
526 ftrace_dump(DUMP_ORIG
);
527 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
528 current
->pid
, current
->comm
,
529 idle
->pid
, idle
->comm
); /* must be idle task! */
534 * Exit an RCU extended quiescent state, which can be either the
535 * idle loop or adaptive-tickless usermode execution.
537 static void rcu_eqs_exit(bool user
)
539 struct rcu_dynticks
*rdtp
;
542 rdtp
= &__get_cpu_var(rcu_dynticks
);
543 oldval
= rdtp
->dynticks_nesting
;
544 WARN_ON_ONCE(oldval
< 0);
545 if (oldval
& DYNTICK_TASK_NEST_MASK
)
546 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
548 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
549 rcu_eqs_exit_common(rdtp
, oldval
, user
);
553 * rcu_idle_exit - inform RCU that current CPU is leaving idle
555 * Exit idle mode, in other words, -enter- the mode in which RCU
556 * read-side critical sections can occur.
558 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
559 * allow for the possibility of usermode upcalls messing up our count
560 * of interrupt nesting level during the busy period that is just
563 void rcu_idle_exit(void)
567 local_irq_save(flags
);
569 local_irq_restore(flags
);
571 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
573 #ifdef CONFIG_RCU_USER_QS
575 * rcu_user_exit - inform RCU that we are exiting userspace.
577 * Exit RCU idle mode while entering the kernel because it can
578 * run a RCU read side critical section anytime.
580 void rcu_user_exit(void)
586 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
587 * idle mode after the current non-nesting irq returns.
589 * This is similar to rcu_user_exit() but in the context of an irq.
590 * This is called when the irq has interrupted a userspace RCU idle mode
591 * context. When the current non-nesting interrupt returns after this call,
592 * the CPU won't restore the RCU idle mode.
594 void rcu_user_exit_after_irq(void)
597 struct rcu_dynticks
*rdtp
;
599 local_irq_save(flags
);
600 rdtp
= &__get_cpu_var(rcu_dynticks
);
601 /* Ensure we are interrupting an RCU idle mode. */
602 WARN_ON_ONCE(rdtp
->dynticks_nesting
& DYNTICK_TASK_NEST_MASK
);
603 rdtp
->dynticks_nesting
+= DYNTICK_TASK_EXIT_IDLE
;
604 local_irq_restore(flags
);
606 #endif /* CONFIG_RCU_USER_QS */
609 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
611 * Enter an interrupt handler, which might possibly result in exiting
612 * idle mode, in other words, entering the mode in which read-side critical
613 * sections can occur.
615 * Note that the Linux kernel is fully capable of entering an interrupt
616 * handler that it never exits, for example when doing upcalls to
617 * user mode! This code assumes that the idle loop never does upcalls to
618 * user mode. If your architecture does do upcalls from the idle loop (or
619 * does anything else that results in unbalanced calls to the irq_enter()
620 * and irq_exit() functions), RCU will give you what you deserve, good
621 * and hard. But very infrequently and irreproducibly.
623 * Use things like work queues to work around this limitation.
625 * You have been warned.
627 void rcu_irq_enter(void)
630 struct rcu_dynticks
*rdtp
;
633 local_irq_save(flags
);
634 rdtp
= &__get_cpu_var(rcu_dynticks
);
635 oldval
= rdtp
->dynticks_nesting
;
636 rdtp
->dynticks_nesting
++;
637 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
639 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
641 rcu_eqs_exit_common(rdtp
, oldval
, true);
642 local_irq_restore(flags
);
646 * rcu_nmi_enter - inform RCU of entry to 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 active.
652 void rcu_nmi_enter(void)
654 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
656 if (rdtp
->dynticks_nmi_nesting
== 0 &&
657 (atomic_read(&rdtp
->dynticks
) & 0x1))
659 rdtp
->dynticks_nmi_nesting
++;
660 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
661 atomic_inc(&rdtp
->dynticks
);
662 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
663 smp_mb__after_atomic_inc(); /* See above. */
664 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
668 * rcu_nmi_exit - inform RCU of exit from NMI context
670 * If the CPU was idle with dynamic ticks active, and there is no
671 * irq handler running, this updates rdtp->dynticks_nmi to let the
672 * RCU grace-period handling know that the CPU is no longer active.
674 void rcu_nmi_exit(void)
676 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
678 if (rdtp
->dynticks_nmi_nesting
== 0 ||
679 --rdtp
->dynticks_nmi_nesting
!= 0)
681 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
682 smp_mb__before_atomic_inc(); /* See above. */
683 atomic_inc(&rdtp
->dynticks
);
684 smp_mb__after_atomic_inc(); /* Force delay to next write. */
685 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
689 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
691 * If the current CPU is in its idle loop and is neither in an interrupt
692 * or NMI handler, return true.
694 int rcu_is_cpu_idle(void)
699 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
703 EXPORT_SYMBOL(rcu_is_cpu_idle
);
705 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
708 * Is the current CPU online? Disable preemption to avoid false positives
709 * that could otherwise happen due to the current CPU number being sampled,
710 * this task being preempted, its old CPU being taken offline, resuming
711 * on some other CPU, then determining that its old CPU is now offline.
712 * It is OK to use RCU on an offline processor during initial boot, hence
713 * the check for rcu_scheduler_fully_active. Note also that it is OK
714 * for a CPU coming online to use RCU for one jiffy prior to marking itself
715 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
716 * offline to continue to use RCU for one jiffy after marking itself
717 * offline in the cpu_online_mask. This leniency is necessary given the
718 * non-atomic nature of the online and offline processing, for example,
719 * the fact that a CPU enters the scheduler after completing the CPU_DYING
722 * This is also why RCU internally marks CPUs online during the
723 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
725 * Disable checking if in an NMI handler because we cannot safely report
726 * errors from NMI handlers anyway.
728 bool rcu_lockdep_current_cpu_online(void)
730 struct rcu_data
*rdp
;
731 struct rcu_node
*rnp
;
737 rdp
= &__get_cpu_var(rcu_sched_data
);
739 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
740 !rcu_scheduler_fully_active
;
744 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
746 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
749 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
751 * If the current CPU is idle or running at a first-level (not nested)
752 * interrupt from idle, return true. The caller must have at least
753 * disabled preemption.
755 static int rcu_is_cpu_rrupt_from_idle(void)
757 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
761 * Snapshot the specified CPU's dynticks counter so that we can later
762 * credit them with an implicit quiescent state. Return 1 if this CPU
763 * is in dynticks idle mode, which is an extended quiescent state.
765 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
767 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
768 return (rdp
->dynticks_snap
& 0x1) == 0;
772 * Return true if the specified CPU has passed through a quiescent
773 * state by virtue of being in or having passed through an dynticks
774 * idle state since the last call to dyntick_save_progress_counter()
775 * for this same CPU, or by virtue of having been offline.
777 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
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
);
831 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
833 rsp
->gp_start
= jiffies
;
834 rsp
->jiffies_stall
= jiffies
+ rcu_jiffies_till_stall_check();
838 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
839 * for architectures that do not implement trigger_all_cpu_backtrace().
840 * The NMI-triggered stack traces are more accurate because they are
841 * printed by the target CPU.
843 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
847 struct rcu_node
*rnp
;
849 rcu_for_each_leaf_node(rsp
, rnp
) {
850 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
851 if (rnp
->qsmask
!= 0) {
852 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
853 if (rnp
->qsmask
& (1UL << cpu
))
854 dump_cpu_task(rnp
->grplo
+ cpu
);
856 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
860 static void print_other_cpu_stall(struct rcu_state
*rsp
)
866 struct rcu_node
*rnp
= rcu_get_root(rsp
);
869 /* Only let one CPU complain about others per time interval. */
871 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
872 delta
= jiffies
- rsp
->jiffies_stall
;
873 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
874 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
877 rsp
->jiffies_stall
= jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
878 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
881 * OK, time to rat on our buddy...
882 * See Documentation/RCU/stallwarn.txt for info on how to debug
883 * RCU CPU stall warnings.
885 pr_err("INFO: %s detected stalls on CPUs/tasks:",
887 print_cpu_stall_info_begin();
888 rcu_for_each_leaf_node(rsp
, rnp
) {
889 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
890 ndetected
+= rcu_print_task_stall(rnp
);
891 if (rnp
->qsmask
!= 0) {
892 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
893 if (rnp
->qsmask
& (1UL << cpu
)) {
894 print_cpu_stall_info(rsp
,
899 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
903 * Now rat on any tasks that got kicked up to the root rcu_node
904 * due to CPU offlining.
906 rnp
= rcu_get_root(rsp
);
907 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
908 ndetected
+= rcu_print_task_stall(rnp
);
909 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
911 print_cpu_stall_info_end();
912 for_each_possible_cpu(cpu
)
913 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
914 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
915 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
916 rsp
->gpnum
, rsp
->completed
, totqlen
);
918 pr_err("INFO: Stall ended before state dump start\n");
919 else if (!trigger_all_cpu_backtrace())
920 rcu_dump_cpu_stacks(rsp
);
922 /* Complain about tasks blocking the grace period. */
924 rcu_print_detail_task_stall(rsp
);
926 force_quiescent_state(rsp
); /* Kick them all. */
929 static void print_cpu_stall(struct rcu_state
*rsp
)
933 struct rcu_node
*rnp
= rcu_get_root(rsp
);
937 * OK, time to rat on ourselves...
938 * See Documentation/RCU/stallwarn.txt for info on how to debug
939 * RCU CPU stall warnings.
941 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
942 print_cpu_stall_info_begin();
943 print_cpu_stall_info(rsp
, smp_processor_id());
944 print_cpu_stall_info_end();
945 for_each_possible_cpu(cpu
)
946 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
947 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
948 jiffies
- rsp
->gp_start
, rsp
->gpnum
, rsp
->completed
, totqlen
);
949 if (!trigger_all_cpu_backtrace())
952 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
953 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
954 rsp
->jiffies_stall
= jiffies
+
955 3 * rcu_jiffies_till_stall_check() + 3;
956 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
958 set_need_resched(); /* kick ourselves to get things going. */
961 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
965 struct rcu_node
*rnp
;
967 if (rcu_cpu_stall_suppress
)
969 j
= ACCESS_ONCE(jiffies
);
970 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
972 if (rcu_gp_in_progress(rsp
) &&
973 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
975 /* We haven't checked in, so go dump stack. */
976 print_cpu_stall(rsp
);
978 } else if (rcu_gp_in_progress(rsp
) &&
979 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
981 /* They had a few time units to dump stack, so complain. */
982 print_other_cpu_stall(rsp
);
987 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
989 * Set the stall-warning timeout way off into the future, thus preventing
990 * any RCU CPU stall-warning messages from appearing in the current set of
993 * The caller must disable hard irqs.
995 void rcu_cpu_stall_reset(void)
997 struct rcu_state
*rsp
;
999 for_each_rcu_flavor(rsp
)
1000 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
1004 * Initialize the specified rcu_data structure's callback list to empty.
1006 static void init_callback_list(struct rcu_data
*rdp
)
1010 if (init_nocb_callback_list(rdp
))
1012 rdp
->nxtlist
= NULL
;
1013 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1014 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1018 * Determine the value that ->completed will have at the end of the
1019 * next subsequent grace period. This is used to tag callbacks so that
1020 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1021 * been dyntick-idle for an extended period with callbacks under the
1022 * influence of RCU_FAST_NO_HZ.
1024 * The caller must hold rnp->lock with interrupts disabled.
1026 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1027 struct rcu_node
*rnp
)
1030 * If RCU is idle, we just wait for the next grace period.
1031 * But we can only be sure that RCU is idle if we are looking
1032 * at the root rcu_node structure -- otherwise, a new grace
1033 * period might have started, but just not yet gotten around
1034 * to initializing the current non-root rcu_node structure.
1036 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1037 return rnp
->completed
+ 1;
1040 * Otherwise, wait for a possible partial grace period and
1041 * then the subsequent full grace period.
1043 return rnp
->completed
+ 2;
1047 * Trace-event helper function for rcu_start_future_gp() and
1048 * rcu_nocb_wait_gp().
1050 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1051 unsigned long c
, const char *s
)
1053 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1054 rnp
->completed
, c
, rnp
->level
,
1055 rnp
->grplo
, rnp
->grphi
, s
);
1059 * Start some future grace period, as needed to handle newly arrived
1060 * callbacks. The required future grace periods are recorded in each
1061 * rcu_node structure's ->need_future_gp field.
1063 * The caller must hold the specified rcu_node structure's ->lock.
1065 static unsigned long __maybe_unused
1066 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1070 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1073 * Pick up grace-period number for new callbacks. If this
1074 * grace period is already marked as needed, return to the caller.
1076 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1077 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1078 if (rnp
->need_future_gp
[c
& 0x1]) {
1079 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1084 * If either this rcu_node structure or the root rcu_node structure
1085 * believe that a grace period is in progress, then we must wait
1086 * for the one following, which is in "c". Because our request
1087 * will be noticed at the end of the current grace period, we don't
1088 * need to explicitly start one.
1090 if (rnp
->gpnum
!= rnp
->completed
||
1091 ACCESS_ONCE(rnp
->gpnum
) != ACCESS_ONCE(rnp
->completed
)) {
1092 rnp
->need_future_gp
[c
& 0x1]++;
1093 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1098 * There might be no grace period in progress. If we don't already
1099 * hold it, acquire the root rcu_node structure's lock in order to
1100 * start one (if needed).
1102 if (rnp
!= rnp_root
)
1103 raw_spin_lock(&rnp_root
->lock
);
1106 * Get a new grace-period number. If there really is no grace
1107 * period in progress, it will be smaller than the one we obtained
1108 * earlier. Adjust callbacks as needed. Note that even no-CBs
1109 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1111 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1112 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1113 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1114 rdp
->nxtcompleted
[i
] = c
;
1117 * If the needed for the required grace period is already
1118 * recorded, trace and leave.
1120 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1121 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1125 /* Record the need for the future grace period. */
1126 rnp_root
->need_future_gp
[c
& 0x1]++;
1128 /* If a grace period is not already in progress, start one. */
1129 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1130 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1132 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1133 rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1136 if (rnp
!= rnp_root
)
1137 raw_spin_unlock(&rnp_root
->lock
);
1142 * Clean up any old requests for the just-ended grace period. Also return
1143 * whether any additional grace periods have been requested. Also invoke
1144 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1145 * waiting for this grace period to complete.
1147 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1149 int c
= rnp
->completed
;
1151 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1153 rcu_nocb_gp_cleanup(rsp
, rnp
);
1154 rnp
->need_future_gp
[c
& 0x1] = 0;
1155 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1156 trace_rcu_future_gp(rnp
, rdp
, c
,
1157 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1162 * If there is room, assign a ->completed number to any callbacks on
1163 * this CPU that have not already been assigned. Also accelerate any
1164 * callbacks that were previously assigned a ->completed number that has
1165 * since proven to be too conservative, which can happen if callbacks get
1166 * assigned a ->completed number while RCU is idle, but with reference to
1167 * a non-root rcu_node structure. This function is idempotent, so it does
1168 * not hurt to call it repeatedly.
1170 * The caller must hold rnp->lock with interrupts disabled.
1172 static void rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1173 struct rcu_data
*rdp
)
1178 /* If the CPU has no callbacks, nothing to do. */
1179 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1183 * Starting from the sublist containing the callbacks most
1184 * recently assigned a ->completed number and working down, find the
1185 * first sublist that is not assignable to an upcoming grace period.
1186 * Such a sublist has something in it (first two tests) and has
1187 * a ->completed number assigned that will complete sooner than
1188 * the ->completed number for newly arrived callbacks (last test).
1190 * The key point is that any later sublist can be assigned the
1191 * same ->completed number as the newly arrived callbacks, which
1192 * means that the callbacks in any of these later sublist can be
1193 * grouped into a single sublist, whether or not they have already
1194 * been assigned a ->completed number.
1196 c
= rcu_cbs_completed(rsp
, rnp
);
1197 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1198 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1199 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1203 * If there are no sublist for unassigned callbacks, leave.
1204 * At the same time, advance "i" one sublist, so that "i" will
1205 * index into the sublist where all the remaining callbacks should
1208 if (++i
>= RCU_NEXT_TAIL
)
1212 * Assign all subsequent callbacks' ->completed number to the next
1213 * full grace period and group them all in the sublist initially
1216 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1217 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1218 rdp
->nxtcompleted
[i
] = c
;
1220 /* Record any needed additional grace periods. */
1221 rcu_start_future_gp(rnp
, rdp
);
1223 /* Trace depending on how much we were able to accelerate. */
1224 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1225 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1227 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1231 * Move any callbacks whose grace period has completed to the
1232 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1233 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1234 * sublist. This function is idempotent, so it does not hurt to
1235 * invoke it repeatedly. As long as it is not invoked -too- often...
1237 * The caller must hold rnp->lock with interrupts disabled.
1239 static void rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1240 struct rcu_data
*rdp
)
1244 /* If the CPU has no callbacks, nothing to do. */
1245 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1249 * Find all callbacks whose ->completed numbers indicate that they
1250 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1252 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1253 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1255 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1257 /* Clean up any sublist tail pointers that were misordered above. */
1258 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1259 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1261 /* Copy down callbacks to fill in empty sublists. */
1262 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1263 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1265 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1266 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1269 /* Classify any remaining callbacks. */
1270 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1274 * Update CPU-local rcu_data state to record the beginnings and ends of
1275 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1276 * structure corresponding to the current CPU, and must have irqs disabled.
1278 static void __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1280 /* Handle the ends of any preceding grace periods first. */
1281 if (rdp
->completed
== rnp
->completed
) {
1283 /* No grace period end, so just accelerate recent callbacks. */
1284 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1288 /* Advance callbacks. */
1289 rcu_advance_cbs(rsp
, rnp
, rdp
);
1291 /* Remember that we saw this grace-period completion. */
1292 rdp
->completed
= rnp
->completed
;
1293 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1296 if (rdp
->gpnum
!= rnp
->gpnum
) {
1298 * If the current grace period is waiting for this CPU,
1299 * set up to detect a quiescent state, otherwise don't
1300 * go looking for one.
1302 rdp
->gpnum
= rnp
->gpnum
;
1303 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1304 rdp
->passed_quiesce
= 0;
1305 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1306 zero_cpu_stall_ticks(rdp
);
1310 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1312 unsigned long flags
;
1313 struct rcu_node
*rnp
;
1315 local_irq_save(flags
);
1317 if ((rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) &&
1318 rdp
->completed
== ACCESS_ONCE(rnp
->completed
)) || /* w/out lock. */
1319 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1320 local_irq_restore(flags
);
1323 __note_gp_changes(rsp
, rnp
, rdp
);
1324 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1328 * Initialize a new grace period.
1330 static int rcu_gp_init(struct rcu_state
*rsp
)
1332 struct rcu_data
*rdp
;
1333 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1335 raw_spin_lock_irq(&rnp
->lock
);
1336 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1338 if (rcu_gp_in_progress(rsp
)) {
1339 /* Grace period already in progress, don't start another. */
1340 raw_spin_unlock_irq(&rnp
->lock
);
1344 /* Advance to a new grace period and initialize state. */
1346 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1347 record_gp_stall_check_time(rsp
);
1348 raw_spin_unlock_irq(&rnp
->lock
);
1350 /* Exclude any concurrent CPU-hotplug operations. */
1351 mutex_lock(&rsp
->onoff_mutex
);
1354 * Set the quiescent-state-needed bits in all the rcu_node
1355 * structures for all currently online CPUs in breadth-first order,
1356 * starting from the root rcu_node structure, relying on the layout
1357 * of the tree within the rsp->node[] array. Note that other CPUs
1358 * will access only the leaves of the hierarchy, thus seeing that no
1359 * grace period is in progress, at least until the corresponding
1360 * leaf node has been initialized. In addition, we have excluded
1361 * CPU-hotplug operations.
1363 * The grace period cannot complete until the initialization
1364 * process finishes, because this kthread handles both.
1366 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1367 raw_spin_lock_irq(&rnp
->lock
);
1368 rdp
= this_cpu_ptr(rsp
->rda
);
1369 rcu_preempt_check_blocked_tasks(rnp
);
1370 rnp
->qsmask
= rnp
->qsmaskinit
;
1371 ACCESS_ONCE(rnp
->gpnum
) = rsp
->gpnum
;
1372 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1373 ACCESS_ONCE(rnp
->completed
) = rsp
->completed
;
1374 if (rnp
== rdp
->mynode
)
1375 __note_gp_changes(rsp
, rnp
, rdp
);
1376 rcu_preempt_boost_start_gp(rnp
);
1377 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1378 rnp
->level
, rnp
->grplo
,
1379 rnp
->grphi
, rnp
->qsmask
);
1380 raw_spin_unlock_irq(&rnp
->lock
);
1381 #ifdef CONFIG_PROVE_RCU_DELAY
1382 if ((prandom_u32() % (rcu_num_nodes
+ 1)) == 0 &&
1383 system_state
== SYSTEM_RUNNING
)
1385 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1389 mutex_unlock(&rsp
->onoff_mutex
);
1394 * Do one round of quiescent-state forcing.
1396 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1398 int fqs_state
= fqs_state_in
;
1399 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1402 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1403 /* Collect dyntick-idle snapshots. */
1404 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1405 fqs_state
= RCU_FORCE_QS
;
1407 /* Handle dyntick-idle and offline CPUs. */
1408 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1410 /* Clear flag to prevent immediate re-entry. */
1411 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1412 raw_spin_lock_irq(&rnp
->lock
);
1413 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1414 raw_spin_unlock_irq(&rnp
->lock
);
1420 * Clean up after the old grace period.
1422 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1424 unsigned long gp_duration
;
1426 struct rcu_data
*rdp
;
1427 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1429 raw_spin_lock_irq(&rnp
->lock
);
1430 gp_duration
= jiffies
- rsp
->gp_start
;
1431 if (gp_duration
> rsp
->gp_max
)
1432 rsp
->gp_max
= gp_duration
;
1435 * We know the grace period is complete, but to everyone else
1436 * it appears to still be ongoing. But it is also the case
1437 * that to everyone else it looks like there is nothing that
1438 * they can do to advance the grace period. It is therefore
1439 * safe for us to drop the lock in order to mark the grace
1440 * period as completed in all of the rcu_node structures.
1442 raw_spin_unlock_irq(&rnp
->lock
);
1445 * Propagate new ->completed value to rcu_node structures so
1446 * that other CPUs don't have to wait until the start of the next
1447 * grace period to process their callbacks. This also avoids
1448 * some nasty RCU grace-period initialization races by forcing
1449 * the end of the current grace period to be completely recorded in
1450 * all of the rcu_node structures before the beginning of the next
1451 * grace period is recorded in any of the rcu_node structures.
1453 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1454 raw_spin_lock_irq(&rnp
->lock
);
1455 ACCESS_ONCE(rnp
->completed
) = rsp
->gpnum
;
1456 rdp
= this_cpu_ptr(rsp
->rda
);
1457 if (rnp
== rdp
->mynode
)
1458 __note_gp_changes(rsp
, rnp
, rdp
);
1459 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1460 raw_spin_unlock_irq(&rnp
->lock
);
1463 rnp
= rcu_get_root(rsp
);
1464 raw_spin_lock_irq(&rnp
->lock
);
1465 rcu_nocb_gp_set(rnp
, nocb
);
1467 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1468 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1469 rsp
->fqs_state
= RCU_GP_IDLE
;
1470 rdp
= this_cpu_ptr(rsp
->rda
);
1471 rcu_advance_cbs(rsp
, rnp
, rdp
); /* Reduce false positives below. */
1472 if (cpu_needs_another_gp(rsp
, rdp
))
1474 raw_spin_unlock_irq(&rnp
->lock
);
1478 * Body of kthread that handles grace periods.
1480 static int __noreturn
rcu_gp_kthread(void *arg
)
1485 struct rcu_state
*rsp
= arg
;
1486 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1490 /* Handle grace-period start. */
1492 wait_event_interruptible(rsp
->gp_wq
,
1495 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1499 flush_signals(current
);
1502 /* Handle quiescent-state forcing. */
1503 fqs_state
= RCU_SAVE_DYNTICK
;
1504 j
= jiffies_till_first_fqs
;
1507 jiffies_till_first_fqs
= HZ
;
1510 rsp
->jiffies_force_qs
= jiffies
+ j
;
1511 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1512 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1513 (!ACCESS_ONCE(rnp
->qsmask
) &&
1514 !rcu_preempt_blocked_readers_cgp(rnp
)),
1516 /* If grace period done, leave loop. */
1517 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1518 !rcu_preempt_blocked_readers_cgp(rnp
))
1520 /* If time for quiescent-state forcing, do it. */
1521 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1522 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1525 /* Deal with stray signal. */
1527 flush_signals(current
);
1529 j
= jiffies_till_next_fqs
;
1532 jiffies_till_next_fqs
= HZ
;
1535 jiffies_till_next_fqs
= 1;
1539 /* Handle grace-period end. */
1540 rcu_gp_cleanup(rsp
);
1544 static void rsp_wakeup(struct irq_work
*work
)
1546 struct rcu_state
*rsp
= container_of(work
, struct rcu_state
, wakeup_work
);
1548 /* Wake up rcu_gp_kthread() to start the grace period. */
1549 wake_up(&rsp
->gp_wq
);
1553 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1554 * in preparation for detecting the next grace period. The caller must hold
1555 * the root node's ->lock and hard irqs must be disabled.
1557 * Note that it is legal for a dying CPU (which is marked as offline) to
1558 * invoke this function. This can happen when the dying CPU reports its
1562 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1563 struct rcu_data
*rdp
)
1565 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
1567 * Either we have not yet spawned the grace-period
1568 * task, this CPU does not need another grace period,
1569 * or a grace period is already in progress.
1570 * Either way, don't start a new grace period.
1574 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1577 * We can't do wakeups while holding the rnp->lock, as that
1578 * could cause possible deadlocks with the rq->lock. Deter
1579 * the wakeup to interrupt context.
1581 irq_work_queue(&rsp
->wakeup_work
);
1585 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1586 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1587 * is invoked indirectly from rcu_advance_cbs(), which would result in
1588 * endless recursion -- or would do so if it wasn't for the self-deadlock
1589 * that is encountered beforehand.
1592 rcu_start_gp(struct rcu_state
*rsp
)
1594 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1595 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1598 * If there is no grace period in progress right now, any
1599 * callbacks we have up to this point will be satisfied by the
1600 * next grace period. Also, advancing the callbacks reduces the
1601 * probability of false positives from cpu_needs_another_gp()
1602 * resulting in pointless grace periods. So, advance callbacks
1603 * then start the grace period!
1605 rcu_advance_cbs(rsp
, rnp
, rdp
);
1606 rcu_start_gp_advanced(rsp
, rnp
, rdp
);
1610 * Report a full set of quiescent states to the specified rcu_state
1611 * data structure. This involves cleaning up after the prior grace
1612 * period and letting rcu_start_gp() start up the next grace period
1613 * if one is needed. Note that the caller must hold rnp->lock, which
1614 * is released before return.
1616 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1617 __releases(rcu_get_root(rsp
)->lock
)
1619 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1620 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1621 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1625 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1626 * Allows quiescent states for a group of CPUs to be reported at one go
1627 * to the specified rcu_node structure, though all the CPUs in the group
1628 * must be represented by the same rcu_node structure (which need not be
1629 * a leaf rcu_node structure, though it often will be). That structure's
1630 * lock must be held upon entry, and it is released before return.
1633 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1634 struct rcu_node
*rnp
, unsigned long flags
)
1635 __releases(rnp
->lock
)
1637 struct rcu_node
*rnp_c
;
1639 /* Walk up the rcu_node hierarchy. */
1641 if (!(rnp
->qsmask
& mask
)) {
1643 /* Our bit has already been cleared, so done. */
1644 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1647 rnp
->qsmask
&= ~mask
;
1648 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1649 mask
, rnp
->qsmask
, rnp
->level
,
1650 rnp
->grplo
, rnp
->grphi
,
1652 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1654 /* Other bits still set at this level, so done. */
1655 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1658 mask
= rnp
->grpmask
;
1659 if (rnp
->parent
== NULL
) {
1661 /* No more levels. Exit loop holding root lock. */
1665 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1668 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1669 WARN_ON_ONCE(rnp_c
->qsmask
);
1673 * Get here if we are the last CPU to pass through a quiescent
1674 * state for this grace period. Invoke rcu_report_qs_rsp()
1675 * to clean up and start the next grace period if one is needed.
1677 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1681 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1682 * structure. This must be either called from the specified CPU, or
1683 * called when the specified CPU is known to be offline (and when it is
1684 * also known that no other CPU is concurrently trying to help the offline
1685 * CPU). The lastcomp argument is used to make sure we are still in the
1686 * grace period of interest. We don't want to end the current grace period
1687 * based on quiescent states detected in an earlier grace period!
1690 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1692 unsigned long flags
;
1694 struct rcu_node
*rnp
;
1697 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1698 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1699 rnp
->completed
== rnp
->gpnum
) {
1702 * The grace period in which this quiescent state was
1703 * recorded has ended, so don't report it upwards.
1704 * We will instead need a new quiescent state that lies
1705 * within the current grace period.
1707 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1708 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1711 mask
= rdp
->grpmask
;
1712 if ((rnp
->qsmask
& mask
) == 0) {
1713 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1715 rdp
->qs_pending
= 0;
1718 * This GP can't end until cpu checks in, so all of our
1719 * callbacks can be processed during the next GP.
1721 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1723 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1728 * Check to see if there is a new grace period of which this CPU
1729 * is not yet aware, and if so, set up local rcu_data state for it.
1730 * Otherwise, see if this CPU has just passed through its first
1731 * quiescent state for this grace period, and record that fact if so.
1734 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1736 /* Check for grace-period ends and beginnings. */
1737 note_gp_changes(rsp
, rdp
);
1740 * Does this CPU still need to do its part for current grace period?
1741 * If no, return and let the other CPUs do their part as well.
1743 if (!rdp
->qs_pending
)
1747 * Was there a quiescent state since the beginning of the grace
1748 * period? If no, then exit and wait for the next call.
1750 if (!rdp
->passed_quiesce
)
1754 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1757 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1760 #ifdef CONFIG_HOTPLUG_CPU
1763 * Send the specified CPU's RCU callbacks to the orphanage. The
1764 * specified CPU must be offline, and the caller must hold the
1768 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1769 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1771 /* No-CBs CPUs do not have orphanable callbacks. */
1772 if (rcu_is_nocb_cpu(rdp
->cpu
))
1776 * Orphan the callbacks. First adjust the counts. This is safe
1777 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1778 * cannot be running now. Thus no memory barrier is required.
1780 if (rdp
->nxtlist
!= NULL
) {
1781 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1782 rsp
->qlen
+= rdp
->qlen
;
1783 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1785 ACCESS_ONCE(rdp
->qlen
) = 0;
1789 * Next, move those callbacks still needing a grace period to
1790 * the orphanage, where some other CPU will pick them up.
1791 * Some of the callbacks might have gone partway through a grace
1792 * period, but that is too bad. They get to start over because we
1793 * cannot assume that grace periods are synchronized across CPUs.
1794 * We don't bother updating the ->nxttail[] array yet, instead
1795 * we just reset the whole thing later on.
1797 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1798 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1799 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1800 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1804 * Then move the ready-to-invoke callbacks to the orphanage,
1805 * where some other CPU will pick them up. These will not be
1806 * required to pass though another grace period: They are done.
1808 if (rdp
->nxtlist
!= NULL
) {
1809 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1810 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1813 /* Finally, initialize the rcu_data structure's list to empty. */
1814 init_callback_list(rdp
);
1818 * Adopt the RCU callbacks from the specified rcu_state structure's
1819 * orphanage. The caller must hold the ->orphan_lock.
1821 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1824 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1826 /* No-CBs CPUs are handled specially. */
1827 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
))
1830 /* Do the accounting first. */
1831 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1832 rdp
->qlen
+= rsp
->qlen
;
1833 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1834 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1835 rcu_idle_count_callbacks_posted();
1840 * We do not need a memory barrier here because the only way we
1841 * can get here if there is an rcu_barrier() in flight is if
1842 * we are the task doing the rcu_barrier().
1845 /* First adopt the ready-to-invoke callbacks. */
1846 if (rsp
->orphan_donelist
!= NULL
) {
1847 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1848 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1849 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1850 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1851 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1852 rsp
->orphan_donelist
= NULL
;
1853 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1856 /* And then adopt the callbacks that still need a grace period. */
1857 if (rsp
->orphan_nxtlist
!= NULL
) {
1858 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1859 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1860 rsp
->orphan_nxtlist
= NULL
;
1861 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1866 * Trace the fact that this CPU is going offline.
1868 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1870 RCU_TRACE(unsigned long mask
);
1871 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1872 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1874 RCU_TRACE(mask
= rdp
->grpmask
);
1875 trace_rcu_grace_period(rsp
->name
,
1876 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1881 * The CPU has been completely removed, and some other CPU is reporting
1882 * this fact from process context. Do the remainder of the cleanup,
1883 * including orphaning the outgoing CPU's RCU callbacks, and also
1884 * adopting them. There can only be one CPU hotplug operation at a time,
1885 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1887 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1889 unsigned long flags
;
1891 int need_report
= 0;
1892 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1893 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1895 /* Adjust any no-longer-needed kthreads. */
1896 rcu_boost_kthread_setaffinity(rnp
, -1);
1898 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1900 /* Exclude any attempts to start a new grace period. */
1901 mutex_lock(&rsp
->onoff_mutex
);
1902 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
1904 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1905 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1906 rcu_adopt_orphan_cbs(rsp
);
1908 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1909 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1911 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1912 rnp
->qsmaskinit
&= ~mask
;
1913 if (rnp
->qsmaskinit
!= 0) {
1914 if (rnp
!= rdp
->mynode
)
1915 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1918 if (rnp
== rdp
->mynode
)
1919 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1921 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1922 mask
= rnp
->grpmask
;
1924 } while (rnp
!= NULL
);
1927 * We still hold the leaf rcu_node structure lock here, and
1928 * irqs are still disabled. The reason for this subterfuge is
1929 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1930 * held leads to deadlock.
1932 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
1934 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1935 rcu_report_unblock_qs_rnp(rnp
, flags
);
1937 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1938 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1939 rcu_report_exp_rnp(rsp
, rnp
, true);
1940 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1941 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1942 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1943 init_callback_list(rdp
);
1944 /* Disallow further callbacks on this CPU. */
1945 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1946 mutex_unlock(&rsp
->onoff_mutex
);
1949 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1951 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1955 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1959 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1962 * Invoke any RCU callbacks that have made it to the end of their grace
1963 * period. Thottle as specified by rdp->blimit.
1965 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1967 unsigned long flags
;
1968 struct rcu_head
*next
, *list
, **tail
;
1969 long bl
, count
, count_lazy
;
1972 /* If no callbacks are ready, just return. */
1973 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1974 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
1975 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
1976 need_resched(), is_idle_task(current
),
1977 rcu_is_callbacks_kthread());
1982 * Extract the list of ready callbacks, disabling to prevent
1983 * races with call_rcu() from interrupt handlers.
1985 local_irq_save(flags
);
1986 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1988 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
1989 list
= rdp
->nxtlist
;
1990 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1991 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1992 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1993 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
1994 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1995 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1996 local_irq_restore(flags
);
1998 /* Invoke callbacks. */
1999 count
= count_lazy
= 0;
2003 debug_rcu_head_unqueue(list
);
2004 if (__rcu_reclaim(rsp
->name
, list
))
2007 /* Stop only if limit reached and CPU has something to do. */
2008 if (++count
>= bl
&&
2010 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2014 local_irq_save(flags
);
2015 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2016 is_idle_task(current
),
2017 rcu_is_callbacks_kthread());
2019 /* Update count, and requeue any remaining callbacks. */
2021 *tail
= rdp
->nxtlist
;
2022 rdp
->nxtlist
= list
;
2023 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2024 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2025 rdp
->nxttail
[i
] = tail
;
2029 smp_mb(); /* List handling before counting for rcu_barrier(). */
2030 rdp
->qlen_lazy
-= count_lazy
;
2031 ACCESS_ONCE(rdp
->qlen
) -= count
;
2032 rdp
->n_cbs_invoked
+= count
;
2034 /* Reinstate batch limit if we have worked down the excess. */
2035 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2036 rdp
->blimit
= blimit
;
2038 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2039 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2040 rdp
->qlen_last_fqs_check
= 0;
2041 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2042 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2043 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2044 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2046 local_irq_restore(flags
);
2048 /* Re-invoke RCU core processing if there are callbacks remaining. */
2049 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2054 * Check to see if this CPU is in a non-context-switch quiescent state
2055 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2056 * Also schedule RCU core processing.
2058 * This function must be called from hardirq context. It is normally
2059 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2060 * false, there is no point in invoking rcu_check_callbacks().
2062 void rcu_check_callbacks(int cpu
, int user
)
2064 trace_rcu_utilization(TPS("Start scheduler-tick"));
2065 increment_cpu_stall_ticks();
2066 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2069 * Get here if this CPU took its interrupt from user
2070 * mode or from the idle loop, and if this is not a
2071 * nested interrupt. In this case, the CPU is in
2072 * a quiescent state, so note it.
2074 * No memory barrier is required here because both
2075 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2076 * variables that other CPUs neither access nor modify,
2077 * at least not while the corresponding CPU is online.
2083 } else if (!in_softirq()) {
2086 * Get here if this CPU did not take its interrupt from
2087 * softirq, in other words, if it is not interrupting
2088 * a rcu_bh read-side critical section. This is an _bh
2089 * critical section, so note it.
2094 rcu_preempt_check_callbacks(cpu
);
2095 if (rcu_pending(cpu
))
2097 trace_rcu_utilization(TPS("End scheduler-tick"));
2101 * Scan the leaf rcu_node structures, processing dyntick state for any that
2102 * have not yet encountered a quiescent state, using the function specified.
2103 * Also initiate boosting for any threads blocked on the root rcu_node.
2105 * The caller must have suppressed start of new grace periods.
2107 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
2111 unsigned long flags
;
2113 struct rcu_node
*rnp
;
2115 rcu_for_each_leaf_node(rsp
, rnp
) {
2118 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2119 if (!rcu_gp_in_progress(rsp
)) {
2120 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2123 if (rnp
->qsmask
== 0) {
2124 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
2129 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2130 if ((rnp
->qsmask
& bit
) != 0 &&
2131 f(per_cpu_ptr(rsp
->rda
, cpu
)))
2136 /* rcu_report_qs_rnp() releases rnp->lock. */
2137 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
2140 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2142 rnp
= rcu_get_root(rsp
);
2143 if (rnp
->qsmask
== 0) {
2144 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2145 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
2150 * Force quiescent states on reluctant CPUs, and also detect which
2151 * CPUs are in dyntick-idle mode.
2153 static void force_quiescent_state(struct rcu_state
*rsp
)
2155 unsigned long flags
;
2157 struct rcu_node
*rnp
;
2158 struct rcu_node
*rnp_old
= NULL
;
2160 /* Funnel through hierarchy to reduce memory contention. */
2161 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
2162 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2163 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2164 !raw_spin_trylock(&rnp
->fqslock
);
2165 if (rnp_old
!= NULL
)
2166 raw_spin_unlock(&rnp_old
->fqslock
);
2168 rsp
->n_force_qs_lh
++;
2173 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2175 /* Reached the root of the rcu_node tree, acquire lock. */
2176 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2177 raw_spin_unlock(&rnp_old
->fqslock
);
2178 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2179 rsp
->n_force_qs_lh
++;
2180 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2181 return; /* Someone beat us to it. */
2183 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
2184 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2185 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
2189 * This does the RCU core processing work for the specified rcu_state
2190 * and rcu_data structures. This may be called only from the CPU to
2191 * whom the rdp belongs.
2194 __rcu_process_callbacks(struct rcu_state
*rsp
)
2196 unsigned long flags
;
2197 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2199 WARN_ON_ONCE(rdp
->beenonline
== 0);
2201 /* Update RCU state based on any recent quiescent states. */
2202 rcu_check_quiescent_state(rsp
, rdp
);
2204 /* Does this CPU require a not-yet-started grace period? */
2205 local_irq_save(flags
);
2206 if (cpu_needs_another_gp(rsp
, rdp
)) {
2207 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2209 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2211 local_irq_restore(flags
);
2214 /* If there are callbacks ready, invoke them. */
2215 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2216 invoke_rcu_callbacks(rsp
, rdp
);
2220 * Do RCU core processing for the current CPU.
2222 static void rcu_process_callbacks(struct softirq_action
*unused
)
2224 struct rcu_state
*rsp
;
2226 if (cpu_is_offline(smp_processor_id()))
2228 trace_rcu_utilization(TPS("Start RCU core"));
2229 for_each_rcu_flavor(rsp
)
2230 __rcu_process_callbacks(rsp
);
2231 trace_rcu_utilization(TPS("End RCU core"));
2235 * Schedule RCU callback invocation. If the specified type of RCU
2236 * does not support RCU priority boosting, just do a direct call,
2237 * otherwise wake up the per-CPU kernel kthread. Note that because we
2238 * are running on the current CPU with interrupts disabled, the
2239 * rcu_cpu_kthread_task cannot disappear out from under us.
2241 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2243 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2245 if (likely(!rsp
->boost
)) {
2246 rcu_do_batch(rsp
, rdp
);
2249 invoke_rcu_callbacks_kthread();
2252 static void invoke_rcu_core(void)
2254 if (cpu_online(smp_processor_id()))
2255 raise_softirq(RCU_SOFTIRQ
);
2259 * Handle any core-RCU processing required by a call_rcu() invocation.
2261 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2262 struct rcu_head
*head
, unsigned long flags
)
2265 * If called from an extended quiescent state, invoke the RCU
2266 * core in order to force a re-evaluation of RCU's idleness.
2268 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2271 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2272 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2276 * Force the grace period if too many callbacks or too long waiting.
2277 * Enforce hysteresis, and don't invoke force_quiescent_state()
2278 * if some other CPU has recently done so. Also, don't bother
2279 * invoking force_quiescent_state() if the newly enqueued callback
2280 * is the only one waiting for a grace period to complete.
2282 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2284 /* Are we ignoring a completed grace period? */
2285 note_gp_changes(rsp
, rdp
);
2287 /* Start a new grace period if one not already started. */
2288 if (!rcu_gp_in_progress(rsp
)) {
2289 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2291 raw_spin_lock(&rnp_root
->lock
);
2293 raw_spin_unlock(&rnp_root
->lock
);
2295 /* Give the grace period a kick. */
2296 rdp
->blimit
= LONG_MAX
;
2297 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2298 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2299 force_quiescent_state(rsp
);
2300 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2301 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2307 * Helper function for call_rcu() and friends. The cpu argument will
2308 * normally be -1, indicating "currently running CPU". It may specify
2309 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2310 * is expected to specify a CPU.
2313 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2314 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2316 unsigned long flags
;
2317 struct rcu_data
*rdp
;
2319 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2320 debug_rcu_head_queue(head
);
2325 * Opportunistically note grace-period endings and beginnings.
2326 * Note that we might see a beginning right after we see an
2327 * end, but never vice versa, since this CPU has to pass through
2328 * a quiescent state betweentimes.
2330 local_irq_save(flags
);
2331 rdp
= this_cpu_ptr(rsp
->rda
);
2333 /* Add the callback to our list. */
2334 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2338 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2339 offline
= !__call_rcu_nocb(rdp
, head
, lazy
);
2340 WARN_ON_ONCE(offline
);
2341 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2342 local_irq_restore(flags
);
2345 ACCESS_ONCE(rdp
->qlen
)++;
2349 rcu_idle_count_callbacks_posted();
2350 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2351 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2352 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2354 if (__is_kfree_rcu_offset((unsigned long)func
))
2355 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2356 rdp
->qlen_lazy
, rdp
->qlen
);
2358 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2360 /* Go handle any RCU core processing required. */
2361 __call_rcu_core(rsp
, rdp
, head
, flags
);
2362 local_irq_restore(flags
);
2366 * Queue an RCU-sched callback for invocation after a grace period.
2368 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2370 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2372 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2375 * Queue an RCU callback for invocation after a quicker grace period.
2377 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2379 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2381 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2384 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2385 * any blocking grace-period wait automatically implies a grace period
2386 * if there is only one CPU online at any point time during execution
2387 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2388 * occasionally incorrectly indicate that there are multiple CPUs online
2389 * when there was in fact only one the whole time, as this just adds
2390 * some overhead: RCU still operates correctly.
2392 static inline int rcu_blocking_is_gp(void)
2396 might_sleep(); /* Check for RCU read-side critical section. */
2398 ret
= num_online_cpus() <= 1;
2404 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2406 * Control will return to the caller some time after a full rcu-sched
2407 * grace period has elapsed, in other words after all currently executing
2408 * rcu-sched read-side critical sections have completed. These read-side
2409 * critical sections are delimited by rcu_read_lock_sched() and
2410 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2411 * local_irq_disable(), and so on may be used in place of
2412 * rcu_read_lock_sched().
2414 * This means that all preempt_disable code sequences, including NMI and
2415 * non-threaded hardware-interrupt handlers, in progress on entry will
2416 * have completed before this primitive returns. However, this does not
2417 * guarantee that softirq handlers will have completed, since in some
2418 * kernels, these handlers can run in process context, and can block.
2420 * Note that this guarantee implies further memory-ordering guarantees.
2421 * On systems with more than one CPU, when synchronize_sched() returns,
2422 * each CPU is guaranteed to have executed a full memory barrier since the
2423 * end of its last RCU-sched read-side critical section whose beginning
2424 * preceded the call to synchronize_sched(). In addition, each CPU having
2425 * an RCU read-side critical section that extends beyond the return from
2426 * synchronize_sched() is guaranteed to have executed a full memory barrier
2427 * after the beginning of synchronize_sched() and before the beginning of
2428 * that RCU read-side critical section. Note that these guarantees include
2429 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2430 * that are executing in the kernel.
2432 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2433 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2434 * to have executed a full memory barrier during the execution of
2435 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2436 * again only if the system has more than one CPU).
2438 * This primitive provides the guarantees made by the (now removed)
2439 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2440 * guarantees that rcu_read_lock() sections will have completed.
2441 * In "classic RCU", these two guarantees happen to be one and
2442 * the same, but can differ in realtime RCU implementations.
2444 void synchronize_sched(void)
2446 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2447 !lock_is_held(&rcu_lock_map
) &&
2448 !lock_is_held(&rcu_sched_lock_map
),
2449 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2450 if (rcu_blocking_is_gp())
2453 synchronize_sched_expedited();
2455 wait_rcu_gp(call_rcu_sched
);
2457 EXPORT_SYMBOL_GPL(synchronize_sched
);
2460 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2462 * Control will return to the caller some time after a full rcu_bh grace
2463 * period has elapsed, in other words after all currently executing rcu_bh
2464 * read-side critical sections have completed. RCU read-side critical
2465 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2466 * and may be nested.
2468 * See the description of synchronize_sched() for more detailed information
2469 * on memory ordering guarantees.
2471 void synchronize_rcu_bh(void)
2473 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2474 !lock_is_held(&rcu_lock_map
) &&
2475 !lock_is_held(&rcu_sched_lock_map
),
2476 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2477 if (rcu_blocking_is_gp())
2480 synchronize_rcu_bh_expedited();
2482 wait_rcu_gp(call_rcu_bh
);
2484 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2486 static int synchronize_sched_expedited_cpu_stop(void *data
)
2489 * There must be a full memory barrier on each affected CPU
2490 * between the time that try_stop_cpus() is called and the
2491 * time that it returns.
2493 * In the current initial implementation of cpu_stop, the
2494 * above condition is already met when the control reaches
2495 * this point and the following smp_mb() is not strictly
2496 * necessary. Do smp_mb() anyway for documentation and
2497 * robustness against future implementation changes.
2499 smp_mb(); /* See above comment block. */
2504 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2506 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2507 * approach to force the grace period to end quickly. This consumes
2508 * significant time on all CPUs and is unfriendly to real-time workloads,
2509 * so is thus not recommended for any sort of common-case code. In fact,
2510 * if you are using synchronize_sched_expedited() in a loop, please
2511 * restructure your code to batch your updates, and then use a single
2512 * synchronize_sched() instead.
2514 * Note that it is illegal to call this function while holding any lock
2515 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2516 * to call this function from a CPU-hotplug notifier. Failing to observe
2517 * these restriction will result in deadlock.
2519 * This implementation can be thought of as an application of ticket
2520 * locking to RCU, with sync_sched_expedited_started and
2521 * sync_sched_expedited_done taking on the roles of the halves
2522 * of the ticket-lock word. Each task atomically increments
2523 * sync_sched_expedited_started upon entry, snapshotting the old value,
2524 * then attempts to stop all the CPUs. If this succeeds, then each
2525 * CPU will have executed a context switch, resulting in an RCU-sched
2526 * grace period. We are then done, so we use atomic_cmpxchg() to
2527 * update sync_sched_expedited_done to match our snapshot -- but
2528 * only if someone else has not already advanced past our snapshot.
2530 * On the other hand, if try_stop_cpus() fails, we check the value
2531 * of sync_sched_expedited_done. If it has advanced past our
2532 * initial snapshot, then someone else must have forced a grace period
2533 * some time after we took our snapshot. In this case, our work is
2534 * done for us, and we can simply return. Otherwise, we try again,
2535 * but keep our initial snapshot for purposes of checking for someone
2536 * doing our work for us.
2538 * If we fail too many times in a row, we fall back to synchronize_sched().
2540 void synchronize_sched_expedited(void)
2542 long firstsnap
, s
, snap
;
2544 struct rcu_state
*rsp
= &rcu_sched_state
;
2547 * If we are in danger of counter wrap, just do synchronize_sched().
2548 * By allowing sync_sched_expedited_started to advance no more than
2549 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2550 * that more than 3.5 billion CPUs would be required to force a
2551 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2552 * course be required on a 64-bit system.
2554 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
2555 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
2557 synchronize_sched();
2558 atomic_long_inc(&rsp
->expedited_wrap
);
2563 * Take a ticket. Note that atomic_inc_return() implies a
2564 * full memory barrier.
2566 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
2569 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2572 * Each pass through the following loop attempts to force a
2573 * context switch on each CPU.
2575 while (try_stop_cpus(cpu_online_mask
,
2576 synchronize_sched_expedited_cpu_stop
,
2579 atomic_long_inc(&rsp
->expedited_tryfail
);
2581 /* Check to see if someone else did our work for us. */
2582 s
= atomic_long_read(&rsp
->expedited_done
);
2583 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2584 /* ensure test happens before caller kfree */
2585 smp_mb__before_atomic_inc(); /* ^^^ */
2586 atomic_long_inc(&rsp
->expedited_workdone1
);
2590 /* No joy, try again later. Or just synchronize_sched(). */
2591 if (trycount
++ < 10) {
2592 udelay(trycount
* num_online_cpus());
2594 wait_rcu_gp(call_rcu_sched
);
2595 atomic_long_inc(&rsp
->expedited_normal
);
2599 /* Recheck to see if someone else did our work for us. */
2600 s
= atomic_long_read(&rsp
->expedited_done
);
2601 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2602 /* ensure test happens before caller kfree */
2603 smp_mb__before_atomic_inc(); /* ^^^ */
2604 atomic_long_inc(&rsp
->expedited_workdone2
);
2609 * Refetching sync_sched_expedited_started allows later
2610 * callers to piggyback on our grace period. We retry
2611 * after they started, so our grace period works for them,
2612 * and they started after our first try, so their grace
2613 * period works for us.
2616 snap
= atomic_long_read(&rsp
->expedited_start
);
2617 smp_mb(); /* ensure read is before try_stop_cpus(). */
2619 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
2622 * Everyone up to our most recent fetch is covered by our grace
2623 * period. Update the counter, but only if our work is still
2624 * relevant -- which it won't be if someone who started later
2625 * than we did already did their update.
2628 atomic_long_inc(&rsp
->expedited_done_tries
);
2629 s
= atomic_long_read(&rsp
->expedited_done
);
2630 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
2631 /* ensure test happens before caller kfree */
2632 smp_mb__before_atomic_inc(); /* ^^^ */
2633 atomic_long_inc(&rsp
->expedited_done_lost
);
2636 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
2637 atomic_long_inc(&rsp
->expedited_done_exit
);
2641 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2644 * Check to see if there is any immediate RCU-related work to be done
2645 * by the current CPU, for the specified type of RCU, returning 1 if so.
2646 * The checks are in order of increasing expense: checks that can be
2647 * carried out against CPU-local state are performed first. However,
2648 * we must check for CPU stalls first, else we might not get a chance.
2650 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2652 struct rcu_node
*rnp
= rdp
->mynode
;
2654 rdp
->n_rcu_pending
++;
2656 /* Check for CPU stalls, if enabled. */
2657 check_cpu_stall(rsp
, rdp
);
2659 /* Is the RCU core waiting for a quiescent state from this CPU? */
2660 if (rcu_scheduler_fully_active
&&
2661 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2662 rdp
->n_rp_qs_pending
++;
2663 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2664 rdp
->n_rp_report_qs
++;
2668 /* Does this CPU have callbacks ready to invoke? */
2669 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2670 rdp
->n_rp_cb_ready
++;
2674 /* Has RCU gone idle with this CPU needing another grace period? */
2675 if (cpu_needs_another_gp(rsp
, rdp
)) {
2676 rdp
->n_rp_cpu_needs_gp
++;
2680 /* Has another RCU grace period completed? */
2681 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2682 rdp
->n_rp_gp_completed
++;
2686 /* Has a new RCU grace period started? */
2687 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2688 rdp
->n_rp_gp_started
++;
2693 rdp
->n_rp_need_nothing
++;
2698 * Check to see if there is any immediate RCU-related work to be done
2699 * by the current CPU, returning 1 if so. This function is part of the
2700 * RCU implementation; it is -not- an exported member of the RCU API.
2702 static int rcu_pending(int cpu
)
2704 struct rcu_state
*rsp
;
2706 for_each_rcu_flavor(rsp
)
2707 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2713 * Return true if the specified CPU has any callback. If all_lazy is
2714 * non-NULL, store an indication of whether all callbacks are lazy.
2715 * (If there are no callbacks, all of them are deemed to be lazy.)
2717 static int rcu_cpu_has_callbacks(int cpu
, bool *all_lazy
)
2721 struct rcu_data
*rdp
;
2722 struct rcu_state
*rsp
;
2724 for_each_rcu_flavor(rsp
) {
2725 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2726 if (rdp
->qlen
!= rdp
->qlen_lazy
)
2737 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2738 * the compiler is expected to optimize this away.
2740 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
2741 int cpu
, unsigned long done
)
2743 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2744 atomic_read(&rsp
->barrier_cpu_count
), done
);
2748 * RCU callback function for _rcu_barrier(). If we are last, wake
2749 * up the task executing _rcu_barrier().
2751 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2753 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2754 struct rcu_state
*rsp
= rdp
->rsp
;
2756 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2757 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2758 complete(&rsp
->barrier_completion
);
2760 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2765 * Called with preemption disabled, and from cross-cpu IRQ context.
2767 static void rcu_barrier_func(void *type
)
2769 struct rcu_state
*rsp
= type
;
2770 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2772 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2773 atomic_inc(&rsp
->barrier_cpu_count
);
2774 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2778 * Orchestrate the specified type of RCU barrier, waiting for all
2779 * RCU callbacks of the specified type to complete.
2781 static void _rcu_barrier(struct rcu_state
*rsp
)
2784 struct rcu_data
*rdp
;
2785 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2786 unsigned long snap_done
;
2788 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2790 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2791 mutex_lock(&rsp
->barrier_mutex
);
2794 * Ensure that all prior references, including to ->n_barrier_done,
2795 * are ordered before the _rcu_barrier() machinery.
2797 smp_mb(); /* See above block comment. */
2800 * Recheck ->n_barrier_done to see if others did our work for us.
2801 * This means checking ->n_barrier_done for an even-to-odd-to-even
2802 * transition. The "if" expression below therefore rounds the old
2803 * value up to the next even number and adds two before comparing.
2805 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2806 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2807 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2808 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2809 smp_mb(); /* caller's subsequent code after above check. */
2810 mutex_unlock(&rsp
->barrier_mutex
);
2815 * Increment ->n_barrier_done to avoid duplicate work. Use
2816 * ACCESS_ONCE() to prevent the compiler from speculating
2817 * the increment to precede the early-exit check.
2819 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2820 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2821 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2822 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2825 * Initialize the count to one rather than to zero in order to
2826 * avoid a too-soon return to zero in case of a short grace period
2827 * (or preemption of this task). Exclude CPU-hotplug operations
2828 * to ensure that no offline CPU has callbacks queued.
2830 init_completion(&rsp
->barrier_completion
);
2831 atomic_set(&rsp
->barrier_cpu_count
, 1);
2835 * Force each CPU with callbacks to register a new callback.
2836 * When that callback is invoked, we will know that all of the
2837 * corresponding CPU's preceding callbacks have been invoked.
2839 for_each_possible_cpu(cpu
) {
2840 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
2842 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2843 if (rcu_is_nocb_cpu(cpu
)) {
2844 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
2845 rsp
->n_barrier_done
);
2846 atomic_inc(&rsp
->barrier_cpu_count
);
2847 __call_rcu(&rdp
->barrier_head
, rcu_barrier_callback
,
2849 } else if (ACCESS_ONCE(rdp
->qlen
)) {
2850 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2851 rsp
->n_barrier_done
);
2852 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2854 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2855 rsp
->n_barrier_done
);
2861 * Now that we have an rcu_barrier_callback() callback on each
2862 * CPU, and thus each counted, remove the initial count.
2864 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2865 complete(&rsp
->barrier_completion
);
2867 /* Increment ->n_barrier_done to prevent duplicate work. */
2868 smp_mb(); /* Keep increment after above mechanism. */
2869 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2870 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2871 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2872 smp_mb(); /* Keep increment before caller's subsequent code. */
2874 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2875 wait_for_completion(&rsp
->barrier_completion
);
2877 /* Other rcu_barrier() invocations can now safely proceed. */
2878 mutex_unlock(&rsp
->barrier_mutex
);
2882 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2884 void rcu_barrier_bh(void)
2886 _rcu_barrier(&rcu_bh_state
);
2888 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2891 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2893 void rcu_barrier_sched(void)
2895 _rcu_barrier(&rcu_sched_state
);
2897 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2900 * Do boot-time initialization of a CPU's per-CPU RCU data.
2903 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2905 unsigned long flags
;
2906 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2907 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2909 /* Set up local state, ensuring consistent view of global state. */
2910 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2911 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2912 init_callback_list(rdp
);
2914 ACCESS_ONCE(rdp
->qlen
) = 0;
2915 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2916 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2917 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2920 rcu_boot_init_nocb_percpu_data(rdp
);
2921 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2925 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2926 * offline event can be happening at a given time. Note also that we
2927 * can accept some slop in the rsp->completed access due to the fact
2928 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2931 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2933 unsigned long flags
;
2935 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2936 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2938 /* Exclude new grace periods. */
2939 mutex_lock(&rsp
->onoff_mutex
);
2941 /* Set up local state, ensuring consistent view of global state. */
2942 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2943 rdp
->beenonline
= 1; /* We have now been online. */
2944 rdp
->preemptible
= preemptible
;
2945 rdp
->qlen_last_fqs_check
= 0;
2946 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2947 rdp
->blimit
= blimit
;
2948 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
2949 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
2950 atomic_set(&rdp
->dynticks
->dynticks
,
2951 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
2952 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2954 /* Add CPU to rcu_node bitmasks. */
2956 mask
= rdp
->grpmask
;
2958 /* Exclude any attempts to start a new GP on small systems. */
2959 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2960 rnp
->qsmaskinit
|= mask
;
2961 mask
= rnp
->grpmask
;
2962 if (rnp
== rdp
->mynode
) {
2964 * If there is a grace period in progress, we will
2965 * set up to wait for it next time we run the
2968 rdp
->gpnum
= rnp
->completed
;
2969 rdp
->completed
= rnp
->completed
;
2970 rdp
->passed_quiesce
= 0;
2971 rdp
->qs_pending
= 0;
2972 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
2974 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
2976 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
2977 local_irq_restore(flags
);
2979 mutex_unlock(&rsp
->onoff_mutex
);
2982 static void rcu_prepare_cpu(int cpu
)
2984 struct rcu_state
*rsp
;
2986 for_each_rcu_flavor(rsp
)
2987 rcu_init_percpu_data(cpu
, rsp
,
2988 strcmp(rsp
->name
, "rcu_preempt") == 0);
2992 * Handle CPU online/offline notification events.
2994 static int rcu_cpu_notify(struct notifier_block
*self
,
2995 unsigned long action
, void *hcpu
)
2997 long cpu
= (long)hcpu
;
2998 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
2999 struct rcu_node
*rnp
= rdp
->mynode
;
3000 struct rcu_state
*rsp
;
3002 trace_rcu_utilization(TPS("Start CPU hotplug"));
3004 case CPU_UP_PREPARE
:
3005 case CPU_UP_PREPARE_FROZEN
:
3006 rcu_prepare_cpu(cpu
);
3007 rcu_prepare_kthreads(cpu
);
3010 case CPU_DOWN_FAILED
:
3011 rcu_boost_kthread_setaffinity(rnp
, -1);
3013 case CPU_DOWN_PREPARE
:
3014 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3017 case CPU_DYING_FROZEN
:
3018 for_each_rcu_flavor(rsp
)
3019 rcu_cleanup_dying_cpu(rsp
);
3022 case CPU_DEAD_FROZEN
:
3023 case CPU_UP_CANCELED
:
3024 case CPU_UP_CANCELED_FROZEN
:
3025 for_each_rcu_flavor(rsp
)
3026 rcu_cleanup_dead_cpu(cpu
, rsp
);
3031 trace_rcu_utilization(TPS("End CPU hotplug"));
3036 * Spawn the kthread that handles this RCU flavor's grace periods.
3038 static int __init
rcu_spawn_gp_kthread(void)
3040 unsigned long flags
;
3041 struct rcu_node
*rnp
;
3042 struct rcu_state
*rsp
;
3043 struct task_struct
*t
;
3045 for_each_rcu_flavor(rsp
) {
3046 t
= kthread_run(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3048 rnp
= rcu_get_root(rsp
);
3049 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3050 rsp
->gp_kthread
= t
;
3051 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3052 rcu_spawn_nocb_kthreads(rsp
);
3056 early_initcall(rcu_spawn_gp_kthread
);
3059 * This function is invoked towards the end of the scheduler's initialization
3060 * process. Before this is called, the idle task might contain
3061 * RCU read-side critical sections (during which time, this idle
3062 * task is booting the system). After this function is called, the
3063 * idle tasks are prohibited from containing RCU read-side critical
3064 * sections. This function also enables RCU lockdep checking.
3066 void rcu_scheduler_starting(void)
3068 WARN_ON(num_online_cpus() != 1);
3069 WARN_ON(nr_context_switches() > 0);
3070 rcu_scheduler_active
= 1;
3074 * Compute the per-level fanout, either using the exact fanout specified
3075 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3077 #ifdef CONFIG_RCU_FANOUT_EXACT
3078 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3082 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
3083 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3084 rsp
->levelspread
[0] = rcu_fanout_leaf
;
3086 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3087 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3094 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3095 ccur
= rsp
->levelcnt
[i
];
3096 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3100 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3103 * Helper function for rcu_init() that initializes one rcu_state structure.
3105 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3106 struct rcu_data __percpu
*rda
)
3108 static char *buf
[] = { "rcu_node_0",
3111 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3112 static char *fqs
[] = { "rcu_node_fqs_0",
3115 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3119 struct rcu_node
*rnp
;
3121 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3123 /* Silence gcc 4.8 warning about array index out of range. */
3124 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3125 panic("rcu_init_one: rcu_num_lvls overflow");
3127 /* Initialize the level-tracking arrays. */
3129 for (i
= 0; i
< rcu_num_lvls
; i
++)
3130 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3131 for (i
= 1; i
< rcu_num_lvls
; i
++)
3132 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3133 rcu_init_levelspread(rsp
);
3135 /* Initialize the elements themselves, starting from the leaves. */
3137 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3138 cpustride
*= rsp
->levelspread
[i
];
3139 rnp
= rsp
->level
[i
];
3140 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3141 raw_spin_lock_init(&rnp
->lock
);
3142 lockdep_set_class_and_name(&rnp
->lock
,
3143 &rcu_node_class
[i
], buf
[i
]);
3144 raw_spin_lock_init(&rnp
->fqslock
);
3145 lockdep_set_class_and_name(&rnp
->fqslock
,
3146 &rcu_fqs_class
[i
], fqs
[i
]);
3147 rnp
->gpnum
= rsp
->gpnum
;
3148 rnp
->completed
= rsp
->completed
;
3150 rnp
->qsmaskinit
= 0;
3151 rnp
->grplo
= j
* cpustride
;
3152 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
3153 if (rnp
->grphi
>= NR_CPUS
)
3154 rnp
->grphi
= NR_CPUS
- 1;
3160 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
3161 rnp
->grpmask
= 1UL << rnp
->grpnum
;
3162 rnp
->parent
= rsp
->level
[i
- 1] +
3163 j
/ rsp
->levelspread
[i
- 1];
3166 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3167 rcu_init_one_nocb(rnp
);
3172 init_waitqueue_head(&rsp
->gp_wq
);
3173 init_irq_work(&rsp
->wakeup_work
, rsp_wakeup
);
3174 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3175 for_each_possible_cpu(i
) {
3176 while (i
> rnp
->grphi
)
3178 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3179 rcu_boot_init_percpu_data(i
, rsp
);
3181 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3185 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3186 * replace the definitions in rcutree.h because those are needed to size
3187 * the ->node array in the rcu_state structure.
3189 static void __init
rcu_init_geometry(void)
3195 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3198 * Initialize any unspecified boot parameters.
3199 * The default values of jiffies_till_first_fqs and
3200 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3201 * value, which is a function of HZ, then adding one for each
3202 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3204 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
3205 if (jiffies_till_first_fqs
== ULONG_MAX
)
3206 jiffies_till_first_fqs
= d
;
3207 if (jiffies_till_next_fqs
== ULONG_MAX
)
3208 jiffies_till_next_fqs
= d
;
3210 /* If the compile-time values are accurate, just leave. */
3211 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3212 nr_cpu_ids
== NR_CPUS
)
3216 * Compute number of nodes that can be handled an rcu_node tree
3217 * with the given number of levels. Setting rcu_capacity[0] makes
3218 * some of the arithmetic easier.
3220 rcu_capacity
[0] = 1;
3221 rcu_capacity
[1] = rcu_fanout_leaf
;
3222 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3223 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3226 * The boot-time rcu_fanout_leaf parameter is only permitted
3227 * to increase the leaf-level fanout, not decrease it. Of course,
3228 * the leaf-level fanout cannot exceed the number of bits in
3229 * the rcu_node masks. Finally, the tree must be able to accommodate
3230 * the configured number of CPUs. Complain and fall back to the
3231 * compile-time values if these limits are exceeded.
3233 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3234 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3235 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3240 /* Calculate the number of rcu_nodes at each level of the tree. */
3241 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3242 if (n
<= rcu_capacity
[i
]) {
3243 for (j
= 0; j
<= i
; j
++)
3245 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3247 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3252 /* Calculate the total number of rcu_node structures. */
3254 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3255 rcu_num_nodes
+= num_rcu_lvl
[i
];
3259 void __init
rcu_init(void)
3263 rcu_bootup_announce();
3264 rcu_init_geometry();
3265 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3266 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3267 __rcu_init_preempt();
3268 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3271 * We don't need protection against CPU-hotplug here because
3272 * this is called early in boot, before either interrupts
3273 * or the scheduler are operational.
3275 cpu_notifier(rcu_cpu_notify
, 0);
3276 for_each_online_cpu(cpu
)
3277 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3280 #include "rcutree_plugin.h"