2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
58 #include <trace/events/rcu.h>
62 /* Data structures. */
64 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
65 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
67 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) { \
68 .level = { &sname##_state.node[0] }, \
70 .fqs_state = RCU_GP_IDLE, \
71 .gpnum = 0UL - 300UL, \
72 .completed = 0UL - 300UL, \
73 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
77 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
82 struct rcu_state rcu_sched_state
=
83 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
84 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
86 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
87 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
89 static struct rcu_state
*rcu_state
;
90 LIST_HEAD(rcu_struct_flavors
);
92 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
93 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
94 module_param(rcu_fanout_leaf
, int, 0444);
95 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
96 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
103 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
106 * The rcu_scheduler_active variable transitions from zero to one just
107 * before the first task is spawned. So when this variable is zero, RCU
108 * can assume that there is but one task, allowing RCU to (for example)
109 * optimize synchronize_sched() to a simple barrier(). When this variable
110 * is one, RCU must actually do all the hard work required to detect real
111 * grace periods. This variable is also used to suppress boot-time false
112 * positives from lockdep-RCU error checking.
114 int rcu_scheduler_active __read_mostly
;
115 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
118 * The rcu_scheduler_fully_active variable transitions from zero to one
119 * during the early_initcall() processing, which is after the scheduler
120 * is capable of creating new tasks. So RCU processing (for example,
121 * creating tasks for RCU priority boosting) must be delayed until after
122 * rcu_scheduler_fully_active transitions from zero to one. We also
123 * currently delay invocation of any RCU callbacks until after this point.
125 * It might later prove better for people registering RCU callbacks during
126 * early boot to take responsibility for these callbacks, but one step at
129 static int rcu_scheduler_fully_active __read_mostly
;
131 #ifdef CONFIG_RCU_BOOST
134 * Control variables for per-CPU and per-rcu_node kthreads. These
135 * handle all flavors of RCU.
137 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
139 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
140 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
142 #endif /* #ifdef CONFIG_RCU_BOOST */
144 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
145 static void invoke_rcu_core(void);
146 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
149 * Track the rcutorture test sequence number and the update version
150 * number within a given test. The rcutorture_testseq is incremented
151 * on every rcutorture module load and unload, so has an odd value
152 * when a test is running. The rcutorture_vernum is set to zero
153 * when rcutorture starts and is incremented on each rcutorture update.
154 * These variables enable correlating rcutorture output with the
155 * RCU tracing information.
157 unsigned long rcutorture_testseq
;
158 unsigned long rcutorture_vernum
;
161 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
162 * permit this function to be invoked without holding the root rcu_node
163 * structure's ->lock, but of course results can be subject to change.
165 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
167 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
171 * Note a quiescent state. Because we do not need to know
172 * how many quiescent states passed, just if there was at least
173 * one since the start of the grace period, this just sets a flag.
174 * The caller must have disabled preemption.
176 void rcu_sched_qs(int cpu
)
178 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
180 if (rdp
->passed_quiesce
== 0)
181 trace_rcu_grace_period("rcu_sched", rdp
->gpnum
, "cpuqs");
182 rdp
->passed_quiesce
= 1;
185 void rcu_bh_qs(int cpu
)
187 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
189 if (rdp
->passed_quiesce
== 0)
190 trace_rcu_grace_period("rcu_bh", rdp
->gpnum
, "cpuqs");
191 rdp
->passed_quiesce
= 1;
195 * Note a context switch. This is a quiescent state for RCU-sched,
196 * and requires special handling for preemptible RCU.
197 * The caller must have disabled preemption.
199 void rcu_note_context_switch(int cpu
)
201 trace_rcu_utilization("Start context switch");
203 rcu_preempt_note_context_switch(cpu
);
204 trace_rcu_utilization("End context switch");
206 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
208 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
209 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
210 .dynticks
= ATOMIC_INIT(1),
213 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
214 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
215 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
217 module_param(blimit
, long, 0444);
218 module_param(qhimark
, long, 0444);
219 module_param(qlowmark
, long, 0444);
221 static ulong jiffies_till_first_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
222 static ulong jiffies_till_next_fqs
= RCU_JIFFIES_TILL_FORCE_QS
;
224 module_param(jiffies_till_first_fqs
, ulong
, 0644);
225 module_param(jiffies_till_next_fqs
, ulong
, 0644);
227 static void rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
228 struct rcu_data
*rdp
);
229 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*));
230 static void force_quiescent_state(struct rcu_state
*rsp
);
231 static int rcu_pending(int cpu
);
234 * Return the number of RCU-sched batches processed thus far for debug & stats.
236 long rcu_batches_completed_sched(void)
238 return rcu_sched_state
.completed
;
240 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
243 * Return the number of RCU BH batches processed thus far for debug & stats.
245 long rcu_batches_completed_bh(void)
247 return rcu_bh_state
.completed
;
249 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
252 * Force a quiescent state for RCU BH.
254 void rcu_bh_force_quiescent_state(void)
256 force_quiescent_state(&rcu_bh_state
);
258 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
261 * Record the number of times rcutorture tests have been initiated and
262 * terminated. This information allows the debugfs tracing stats to be
263 * correlated to the rcutorture messages, even when the rcutorture module
264 * is being repeatedly loaded and unloaded. In other words, we cannot
265 * store this state in rcutorture itself.
267 void rcutorture_record_test_transition(void)
269 rcutorture_testseq
++;
270 rcutorture_vernum
= 0;
272 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
275 * Record the number of writer passes through the current rcutorture test.
276 * This is also used to correlate debugfs tracing stats with the rcutorture
279 void rcutorture_record_progress(unsigned long vernum
)
283 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
286 * Force a quiescent state for RCU-sched.
288 void rcu_sched_force_quiescent_state(void)
290 force_quiescent_state(&rcu_sched_state
);
292 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
295 * Does the CPU have callbacks ready to be invoked?
298 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
300 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
301 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
305 * Does the current CPU require a not-yet-started grace period?
306 * The caller must have disabled interrupts to prevent races with
307 * normal callback registry.
310 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
314 if (rcu_gp_in_progress(rsp
))
315 return 0; /* No, a grace period is already in progress. */
316 if (rcu_nocb_needs_gp(rsp
))
317 return 1; /* Yes, a no-CBs CPU needs one. */
318 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
319 return 0; /* No, this is a no-CBs (or offline) CPU. */
320 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
321 return 1; /* Yes, this CPU has newly registered callbacks. */
322 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
323 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
324 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
325 rdp
->nxtcompleted
[i
]))
326 return 1; /* Yes, CBs for future grace period. */
327 return 0; /* No grace period needed. */
331 * Return the root node of the specified rcu_state structure.
333 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
335 return &rsp
->node
[0];
339 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
341 * If the new value of the ->dynticks_nesting counter now is zero,
342 * we really have entered idle, and must do the appropriate accounting.
343 * The caller must have disabled interrupts.
345 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
348 trace_rcu_dyntick("Start", oldval
, rdtp
->dynticks_nesting
);
349 if (!user
&& !is_idle_task(current
)) {
350 struct task_struct
*idle
= idle_task(smp_processor_id());
352 trace_rcu_dyntick("Error on entry: not idle task", oldval
, 0);
353 ftrace_dump(DUMP_ORIG
);
354 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
355 current
->pid
, current
->comm
,
356 idle
->pid
, idle
->comm
); /* must be idle task! */
358 rcu_prepare_for_idle(smp_processor_id());
359 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
360 smp_mb__before_atomic_inc(); /* See above. */
361 atomic_inc(&rdtp
->dynticks
);
362 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
363 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
366 * It is illegal to enter an extended quiescent state while
367 * in an RCU read-side critical section.
369 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
370 "Illegal idle entry in RCU read-side critical section.");
371 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
372 "Illegal idle entry in RCU-bh read-side critical section.");
373 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
374 "Illegal idle entry in RCU-sched read-side critical section.");
378 * Enter an RCU extended quiescent state, which can be either the
379 * idle loop or adaptive-tickless usermode execution.
381 static void rcu_eqs_enter(bool user
)
384 struct rcu_dynticks
*rdtp
;
386 rdtp
= &__get_cpu_var(rcu_dynticks
);
387 oldval
= rdtp
->dynticks_nesting
;
388 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
389 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
)
390 rdtp
->dynticks_nesting
= 0;
392 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
393 rcu_eqs_enter_common(rdtp
, oldval
, user
);
397 * rcu_idle_enter - inform RCU that current CPU is entering idle
399 * Enter idle mode, in other words, -leave- the mode in which RCU
400 * read-side critical sections can occur. (Though RCU read-side
401 * critical sections can occur in irq handlers in idle, a possibility
402 * handled by irq_enter() and irq_exit().)
404 * We crowbar the ->dynticks_nesting field to zero to allow for
405 * the possibility of usermode upcalls having messed up our count
406 * of interrupt nesting level during the prior busy period.
408 void rcu_idle_enter(void)
412 local_irq_save(flags
);
413 rcu_eqs_enter(false);
414 local_irq_restore(flags
);
416 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
418 #ifdef CONFIG_RCU_USER_QS
420 * rcu_user_enter - inform RCU that we are resuming userspace.
422 * Enter RCU idle mode right before resuming userspace. No use of RCU
423 * is permitted between this call and rcu_user_exit(). This way the
424 * CPU doesn't need to maintain the tick for RCU maintenance purposes
425 * when the CPU runs in userspace.
427 void rcu_user_enter(void)
433 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
434 * after the current irq returns.
436 * This is similar to rcu_user_enter() but in the context of a non-nesting
437 * irq. After this call, RCU enters into idle mode when the interrupt
440 void rcu_user_enter_after_irq(void)
443 struct rcu_dynticks
*rdtp
;
445 local_irq_save(flags
);
446 rdtp
= &__get_cpu_var(rcu_dynticks
);
447 /* Ensure this irq is interrupting a non-idle RCU state. */
448 WARN_ON_ONCE(!(rdtp
->dynticks_nesting
& DYNTICK_TASK_MASK
));
449 rdtp
->dynticks_nesting
= 1;
450 local_irq_restore(flags
);
452 #endif /* CONFIG_RCU_USER_QS */
455 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
457 * Exit from an interrupt handler, which might possibly result in entering
458 * idle mode, in other words, leaving the mode in which read-side critical
459 * sections can occur.
461 * This code assumes that the idle loop never does anything that might
462 * result in unbalanced calls to irq_enter() and irq_exit(). If your
463 * architecture violates this assumption, RCU will give you what you
464 * deserve, good and hard. But very infrequently and irreproducibly.
466 * Use things like work queues to work around this limitation.
468 * You have been warned.
470 void rcu_irq_exit(void)
474 struct rcu_dynticks
*rdtp
;
476 local_irq_save(flags
);
477 rdtp
= &__get_cpu_var(rcu_dynticks
);
478 oldval
= rdtp
->dynticks_nesting
;
479 rdtp
->dynticks_nesting
--;
480 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
481 if (rdtp
->dynticks_nesting
)
482 trace_rcu_dyntick("--=", oldval
, rdtp
->dynticks_nesting
);
484 rcu_eqs_enter_common(rdtp
, oldval
, true);
485 local_irq_restore(flags
);
489 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
491 * If the new value of the ->dynticks_nesting counter was previously zero,
492 * we really have exited idle, and must do the appropriate accounting.
493 * The caller must have disabled interrupts.
495 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
498 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
499 atomic_inc(&rdtp
->dynticks
);
500 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
501 smp_mb__after_atomic_inc(); /* See above. */
502 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
503 rcu_cleanup_after_idle(smp_processor_id());
504 trace_rcu_dyntick("End", oldval
, rdtp
->dynticks_nesting
);
505 if (!user
&& !is_idle_task(current
)) {
506 struct task_struct
*idle
= idle_task(smp_processor_id());
508 trace_rcu_dyntick("Error on exit: not idle task",
509 oldval
, rdtp
->dynticks_nesting
);
510 ftrace_dump(DUMP_ORIG
);
511 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
512 current
->pid
, current
->comm
,
513 idle
->pid
, idle
->comm
); /* must be idle task! */
518 * Exit an RCU extended quiescent state, which can be either the
519 * idle loop or adaptive-tickless usermode execution.
521 static void rcu_eqs_exit(bool user
)
523 struct rcu_dynticks
*rdtp
;
526 rdtp
= &__get_cpu_var(rcu_dynticks
);
527 oldval
= rdtp
->dynticks_nesting
;
528 WARN_ON_ONCE(oldval
< 0);
529 if (oldval
& DYNTICK_TASK_NEST_MASK
)
530 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
532 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
533 rcu_eqs_exit_common(rdtp
, oldval
, user
);
537 * rcu_idle_exit - inform RCU that current CPU is leaving idle
539 * Exit idle mode, in other words, -enter- the mode in which RCU
540 * read-side critical sections can occur.
542 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
543 * allow for the possibility of usermode upcalls messing up our count
544 * of interrupt nesting level during the busy period that is just
547 void rcu_idle_exit(void)
551 local_irq_save(flags
);
553 local_irq_restore(flags
);
555 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
557 #ifdef CONFIG_RCU_USER_QS
559 * rcu_user_exit - inform RCU that we are exiting userspace.
561 * Exit RCU idle mode while entering the kernel because it can
562 * run a RCU read side critical section anytime.
564 void rcu_user_exit(void)
570 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
571 * idle mode after the current non-nesting irq returns.
573 * This is similar to rcu_user_exit() but in the context of an irq.
574 * This is called when the irq has interrupted a userspace RCU idle mode
575 * context. When the current non-nesting interrupt returns after this call,
576 * the CPU won't restore the RCU idle mode.
578 void rcu_user_exit_after_irq(void)
581 struct rcu_dynticks
*rdtp
;
583 local_irq_save(flags
);
584 rdtp
= &__get_cpu_var(rcu_dynticks
);
585 /* Ensure we are interrupting an RCU idle mode. */
586 WARN_ON_ONCE(rdtp
->dynticks_nesting
& DYNTICK_TASK_NEST_MASK
);
587 rdtp
->dynticks_nesting
+= DYNTICK_TASK_EXIT_IDLE
;
588 local_irq_restore(flags
);
590 #endif /* CONFIG_RCU_USER_QS */
593 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
595 * Enter an interrupt handler, which might possibly result in exiting
596 * idle mode, in other words, entering the mode in which read-side critical
597 * sections can occur.
599 * Note that the Linux kernel is fully capable of entering an interrupt
600 * handler that it never exits, for example when doing upcalls to
601 * user mode! This code assumes that the idle loop never does upcalls to
602 * user mode. If your architecture does do upcalls from the idle loop (or
603 * does anything else that results in unbalanced calls to the irq_enter()
604 * and irq_exit() functions), RCU will give you what you deserve, good
605 * and hard. But very infrequently and irreproducibly.
607 * Use things like work queues to work around this limitation.
609 * You have been warned.
611 void rcu_irq_enter(void)
614 struct rcu_dynticks
*rdtp
;
617 local_irq_save(flags
);
618 rdtp
= &__get_cpu_var(rcu_dynticks
);
619 oldval
= rdtp
->dynticks_nesting
;
620 rdtp
->dynticks_nesting
++;
621 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
623 trace_rcu_dyntick("++=", oldval
, rdtp
->dynticks_nesting
);
625 rcu_eqs_exit_common(rdtp
, oldval
, true);
626 local_irq_restore(flags
);
630 * rcu_nmi_enter - inform RCU of entry to NMI context
632 * If the CPU was idle with dynamic ticks active, and there is no
633 * irq handler running, this updates rdtp->dynticks_nmi to let the
634 * RCU grace-period handling know that the CPU is active.
636 void rcu_nmi_enter(void)
638 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
640 if (rdtp
->dynticks_nmi_nesting
== 0 &&
641 (atomic_read(&rdtp
->dynticks
) & 0x1))
643 rdtp
->dynticks_nmi_nesting
++;
644 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
645 atomic_inc(&rdtp
->dynticks
);
646 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
647 smp_mb__after_atomic_inc(); /* See above. */
648 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
652 * rcu_nmi_exit - inform RCU of exit from NMI context
654 * If the CPU was idle with dynamic ticks active, and there is no
655 * irq handler running, this updates rdtp->dynticks_nmi to let the
656 * RCU grace-period handling know that the CPU is no longer active.
658 void rcu_nmi_exit(void)
660 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
662 if (rdtp
->dynticks_nmi_nesting
== 0 ||
663 --rdtp
->dynticks_nmi_nesting
!= 0)
665 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
666 smp_mb__before_atomic_inc(); /* See above. */
667 atomic_inc(&rdtp
->dynticks
);
668 smp_mb__after_atomic_inc(); /* Force delay to next write. */
669 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
673 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
675 * If the current CPU is in its idle loop and is neither in an interrupt
676 * or NMI handler, return true.
678 int rcu_is_cpu_idle(void)
683 ret
= (atomic_read(&__get_cpu_var(rcu_dynticks
).dynticks
) & 0x1) == 0;
687 EXPORT_SYMBOL(rcu_is_cpu_idle
);
689 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
692 * Is the current CPU online? Disable preemption to avoid false positives
693 * that could otherwise happen due to the current CPU number being sampled,
694 * this task being preempted, its old CPU being taken offline, resuming
695 * on some other CPU, then determining that its old CPU is now offline.
696 * It is OK to use RCU on an offline processor during initial boot, hence
697 * the check for rcu_scheduler_fully_active. Note also that it is OK
698 * for a CPU coming online to use RCU for one jiffy prior to marking itself
699 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
700 * offline to continue to use RCU for one jiffy after marking itself
701 * offline in the cpu_online_mask. This leniency is necessary given the
702 * non-atomic nature of the online and offline processing, for example,
703 * the fact that a CPU enters the scheduler after completing the CPU_DYING
706 * This is also why RCU internally marks CPUs online during the
707 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
709 * Disable checking if in an NMI handler because we cannot safely report
710 * errors from NMI handlers anyway.
712 bool rcu_lockdep_current_cpu_online(void)
714 struct rcu_data
*rdp
;
715 struct rcu_node
*rnp
;
721 rdp
= &__get_cpu_var(rcu_sched_data
);
723 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
724 !rcu_scheduler_fully_active
;
728 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
730 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
733 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
735 * If the current CPU is idle or running at a first-level (not nested)
736 * interrupt from idle, return true. The caller must have at least
737 * disabled preemption.
739 static int rcu_is_cpu_rrupt_from_idle(void)
741 return __get_cpu_var(rcu_dynticks
).dynticks_nesting
<= 1;
745 * Snapshot the specified CPU's dynticks counter so that we can later
746 * credit them with an implicit quiescent state. Return 1 if this CPU
747 * is in dynticks idle mode, which is an extended quiescent state.
749 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
751 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
752 return (rdp
->dynticks_snap
& 0x1) == 0;
756 * Return true if the specified CPU has passed through a quiescent
757 * state by virtue of being in or having passed through an dynticks
758 * idle state since the last call to dyntick_save_progress_counter()
759 * for this same CPU, or by virtue of having been offline.
761 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
766 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
767 snap
= (unsigned int)rdp
->dynticks_snap
;
770 * If the CPU passed through or entered a dynticks idle phase with
771 * no active irq/NMI handlers, then we can safely pretend that the CPU
772 * already acknowledged the request to pass through a quiescent
773 * state. Either way, that CPU cannot possibly be in an RCU
774 * read-side critical section that started before the beginning
775 * of the current RCU grace period.
777 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
778 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "dti");
784 * Check for the CPU being offline, but only if the grace period
785 * is old enough. We don't need to worry about the CPU changing
786 * state: If we see it offline even once, it has been through a
789 * The reason for insisting that the grace period be at least
790 * one jiffy old is that CPUs that are not quite online and that
791 * have just gone offline can still execute RCU read-side critical
794 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
795 return 0; /* Grace period is not old enough. */
797 if (cpu_is_offline(rdp
->cpu
)) {
798 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, "ofl");
805 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
807 rsp
->gp_start
= jiffies
;
808 rsp
->jiffies_stall
= jiffies
+ rcu_jiffies_till_stall_check();
812 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
813 * for architectures that do not implement trigger_all_cpu_backtrace().
814 * The NMI-triggered stack traces are more accurate because they are
815 * printed by the target CPU.
817 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
821 struct rcu_node
*rnp
;
823 rcu_for_each_leaf_node(rsp
, rnp
) {
824 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
825 if (rnp
->qsmask
!= 0) {
826 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
827 if (rnp
->qsmask
& (1UL << cpu
))
828 dump_cpu_task(rnp
->grplo
+ cpu
);
830 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
834 static void print_other_cpu_stall(struct rcu_state
*rsp
)
840 struct rcu_node
*rnp
= rcu_get_root(rsp
);
843 /* Only let one CPU complain about others per time interval. */
845 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
846 delta
= jiffies
- rsp
->jiffies_stall
;
847 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
848 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
851 rsp
->jiffies_stall
= jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
852 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
855 * OK, time to rat on our buddy...
856 * See Documentation/RCU/stallwarn.txt for info on how to debug
857 * RCU CPU stall warnings.
859 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks:",
861 print_cpu_stall_info_begin();
862 rcu_for_each_leaf_node(rsp
, rnp
) {
863 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
864 ndetected
+= rcu_print_task_stall(rnp
);
865 if (rnp
->qsmask
!= 0) {
866 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
867 if (rnp
->qsmask
& (1UL << cpu
)) {
868 print_cpu_stall_info(rsp
,
873 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
877 * Now rat on any tasks that got kicked up to the root rcu_node
878 * due to CPU offlining.
880 rnp
= rcu_get_root(rsp
);
881 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
882 ndetected
+= rcu_print_task_stall(rnp
);
883 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
885 print_cpu_stall_info_end();
886 for_each_possible_cpu(cpu
)
887 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
888 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
889 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
890 rsp
->gpnum
, rsp
->completed
, totqlen
);
892 printk(KERN_ERR
"INFO: Stall ended before state dump start\n");
893 else if (!trigger_all_cpu_backtrace())
894 rcu_dump_cpu_stacks(rsp
);
896 /* Complain about tasks blocking the grace period. */
898 rcu_print_detail_task_stall(rsp
);
900 force_quiescent_state(rsp
); /* Kick them all. */
903 static void print_cpu_stall(struct rcu_state
*rsp
)
907 struct rcu_node
*rnp
= rcu_get_root(rsp
);
911 * OK, time to rat on ourselves...
912 * See Documentation/RCU/stallwarn.txt for info on how to debug
913 * RCU CPU stall warnings.
915 printk(KERN_ERR
"INFO: %s self-detected stall on CPU", rsp
->name
);
916 print_cpu_stall_info_begin();
917 print_cpu_stall_info(rsp
, smp_processor_id());
918 print_cpu_stall_info_end();
919 for_each_possible_cpu(cpu
)
920 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
921 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
922 jiffies
- rsp
->gp_start
, rsp
->gpnum
, rsp
->completed
, totqlen
);
923 if (!trigger_all_cpu_backtrace())
926 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
927 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
928 rsp
->jiffies_stall
= jiffies
+
929 3 * rcu_jiffies_till_stall_check() + 3;
930 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
932 set_need_resched(); /* kick ourselves to get things going. */
935 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
939 struct rcu_node
*rnp
;
941 if (rcu_cpu_stall_suppress
)
943 j
= ACCESS_ONCE(jiffies
);
944 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
946 if (rcu_gp_in_progress(rsp
) &&
947 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
949 /* We haven't checked in, so go dump stack. */
950 print_cpu_stall(rsp
);
952 } else if (rcu_gp_in_progress(rsp
) &&
953 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
955 /* They had a few time units to dump stack, so complain. */
956 print_other_cpu_stall(rsp
);
961 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
963 * Set the stall-warning timeout way off into the future, thus preventing
964 * any RCU CPU stall-warning messages from appearing in the current set of
967 * The caller must disable hard irqs.
969 void rcu_cpu_stall_reset(void)
971 struct rcu_state
*rsp
;
973 for_each_rcu_flavor(rsp
)
974 rsp
->jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
978 * Update CPU-local rcu_data state to record the newly noticed grace period.
979 * This is used both when we started the grace period and when we notice
980 * that someone else started the grace period. The caller must hold the
981 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
982 * and must have irqs disabled.
984 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
986 if (rdp
->gpnum
!= rnp
->gpnum
) {
988 * If the current grace period is waiting for this CPU,
989 * set up to detect a quiescent state, otherwise don't
990 * go looking for one.
992 rdp
->gpnum
= rnp
->gpnum
;
993 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpustart");
994 rdp
->passed_quiesce
= 0;
995 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
996 zero_cpu_stall_ticks(rdp
);
1000 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1002 unsigned long flags
;
1003 struct rcu_node
*rnp
;
1005 local_irq_save(flags
);
1007 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
1008 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1009 local_irq_restore(flags
);
1012 __note_new_gpnum(rsp
, rnp
, rdp
);
1013 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1017 * Did someone else start a new RCU grace period start since we last
1018 * checked? Update local state appropriately if so. Must be called
1019 * on the CPU corresponding to rdp.
1022 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1024 unsigned long flags
;
1027 local_irq_save(flags
);
1028 if (rdp
->gpnum
!= rsp
->gpnum
) {
1029 note_new_gpnum(rsp
, rdp
);
1032 local_irq_restore(flags
);
1037 * Initialize the specified rcu_data structure's callback list to empty.
1039 static void init_callback_list(struct rcu_data
*rdp
)
1043 if (init_nocb_callback_list(rdp
))
1045 rdp
->nxtlist
= NULL
;
1046 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1047 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1051 * Determine the value that ->completed will have at the end of the
1052 * next subsequent grace period. This is used to tag callbacks so that
1053 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1054 * been dyntick-idle for an extended period with callbacks under the
1055 * influence of RCU_FAST_NO_HZ.
1057 * The caller must hold rnp->lock with interrupts disabled.
1059 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1060 struct rcu_node
*rnp
)
1063 * If RCU is idle, we just wait for the next grace period.
1064 * But we can only be sure that RCU is idle if we are looking
1065 * at the root rcu_node structure -- otherwise, a new grace
1066 * period might have started, but just not yet gotten around
1067 * to initializing the current non-root rcu_node structure.
1069 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1070 return rnp
->completed
+ 1;
1073 * Otherwise, wait for a possible partial grace period and
1074 * then the subsequent full grace period.
1076 return rnp
->completed
+ 2;
1080 * Trace-event helper function for rcu_start_future_gp() and
1081 * rcu_nocb_wait_gp().
1083 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1084 unsigned long c
, char *s
)
1086 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1087 rnp
->completed
, c
, rnp
->level
,
1088 rnp
->grplo
, rnp
->grphi
, s
);
1092 * Start some future grace period, as needed to handle newly arrived
1093 * callbacks. The required future grace periods are recorded in each
1094 * rcu_node structure's ->need_future_gp field.
1096 * The caller must hold the specified rcu_node structure's ->lock.
1098 static unsigned long __maybe_unused
1099 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1103 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1106 * Pick up grace-period number for new callbacks. If this
1107 * grace period is already marked as needed, return to the caller.
1109 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1110 trace_rcu_future_gp(rnp
, rdp
, c
, "Startleaf");
1111 if (rnp
->need_future_gp
[c
& 0x1]) {
1112 trace_rcu_future_gp(rnp
, rdp
, c
, "Prestartleaf");
1117 * If either this rcu_node structure or the root rcu_node structure
1118 * believe that a grace period is in progress, then we must wait
1119 * for the one following, which is in "c". Because our request
1120 * will be noticed at the end of the current grace period, we don't
1121 * need to explicitly start one.
1123 if (rnp
->gpnum
!= rnp
->completed
||
1124 ACCESS_ONCE(rnp
->gpnum
) != ACCESS_ONCE(rnp
->completed
)) {
1125 rnp
->need_future_gp
[c
& 0x1]++;
1126 trace_rcu_future_gp(rnp
, rdp
, c
, "Startedleaf");
1131 * There might be no grace period in progress. If we don't already
1132 * hold it, acquire the root rcu_node structure's lock in order to
1133 * start one (if needed).
1135 if (rnp
!= rnp_root
)
1136 raw_spin_lock(&rnp_root
->lock
);
1139 * Get a new grace-period number. If there really is no grace
1140 * period in progress, it will be smaller than the one we obtained
1141 * earlier. Adjust callbacks as needed. Note that even no-CBs
1142 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1144 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1145 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1146 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1147 rdp
->nxtcompleted
[i
] = c
;
1150 * If the needed for the required grace period is already
1151 * recorded, trace and leave.
1153 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1154 trace_rcu_future_gp(rnp
, rdp
, c
, "Prestartedroot");
1158 /* Record the need for the future grace period. */
1159 rnp_root
->need_future_gp
[c
& 0x1]++;
1161 /* If a grace period is not already in progress, start one. */
1162 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1163 trace_rcu_future_gp(rnp
, rdp
, c
, "Startedleafroot");
1165 trace_rcu_future_gp(rnp
, rdp
, c
, "Startedroot");
1166 rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1169 if (rnp
!= rnp_root
)
1170 raw_spin_unlock(&rnp_root
->lock
);
1175 * Clean up any old requests for the just-ended grace period. Also return
1176 * whether any additional grace periods have been requested. Also invoke
1177 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1178 * waiting for this grace period to complete.
1180 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1182 int c
= rnp
->completed
;
1184 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1186 rcu_nocb_gp_cleanup(rsp
, rnp
);
1187 rnp
->need_future_gp
[c
& 0x1] = 0;
1188 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1189 trace_rcu_future_gp(rnp
, rdp
, c
, needmore
? "CleanupMore" : "Cleanup");
1194 * If there is room, assign a ->completed number to any callbacks on
1195 * this CPU that have not already been assigned. Also accelerate any
1196 * callbacks that were previously assigned a ->completed number that has
1197 * since proven to be too conservative, which can happen if callbacks get
1198 * assigned a ->completed number while RCU is idle, but with reference to
1199 * a non-root rcu_node structure. This function is idempotent, so it does
1200 * not hurt to call it repeatedly.
1202 * The caller must hold rnp->lock with interrupts disabled.
1204 static void rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1205 struct rcu_data
*rdp
)
1210 /* If the CPU has no callbacks, nothing to do. */
1211 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1215 * Starting from the sublist containing the callbacks most
1216 * recently assigned a ->completed number and working down, find the
1217 * first sublist that is not assignable to an upcoming grace period.
1218 * Such a sublist has something in it (first two tests) and has
1219 * a ->completed number assigned that will complete sooner than
1220 * the ->completed number for newly arrived callbacks (last test).
1222 * The key point is that any later sublist can be assigned the
1223 * same ->completed number as the newly arrived callbacks, which
1224 * means that the callbacks in any of these later sublist can be
1225 * grouped into a single sublist, whether or not they have already
1226 * been assigned a ->completed number.
1228 c
= rcu_cbs_completed(rsp
, rnp
);
1229 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1230 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1231 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1235 * If there are no sublist for unassigned callbacks, leave.
1236 * At the same time, advance "i" one sublist, so that "i" will
1237 * index into the sublist where all the remaining callbacks should
1240 if (++i
>= RCU_NEXT_TAIL
)
1244 * Assign all subsequent callbacks' ->completed number to the next
1245 * full grace period and group them all in the sublist initially
1248 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1249 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1250 rdp
->nxtcompleted
[i
] = c
;
1252 /* Record any needed additional grace periods. */
1253 rcu_start_future_gp(rnp
, rdp
);
1255 /* Trace depending on how much we were able to accelerate. */
1256 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1257 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "AccWaitCB");
1259 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "AccReadyCB");
1263 * Move any callbacks whose grace period has completed to the
1264 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1265 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1266 * sublist. This function is idempotent, so it does not hurt to
1267 * invoke it repeatedly. As long as it is not invoked -too- often...
1269 * The caller must hold rnp->lock with interrupts disabled.
1271 static void rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1272 struct rcu_data
*rdp
)
1276 /* If the CPU has no callbacks, nothing to do. */
1277 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1281 * Find all callbacks whose ->completed numbers indicate that they
1282 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1284 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1285 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1287 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1289 /* Clean up any sublist tail pointers that were misordered above. */
1290 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1291 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1293 /* Copy down callbacks to fill in empty sublists. */
1294 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1295 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1297 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1298 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1301 /* Classify any remaining callbacks. */
1302 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1306 * Advance this CPU's callbacks, but only if the current grace period
1307 * has ended. This may be called only from the CPU to whom the rdp
1308 * belongs. In addition, the corresponding leaf rcu_node structure's
1309 * ->lock must be held by the caller, with irqs disabled.
1312 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1314 /* Did another grace period end? */
1315 if (rdp
->completed
== rnp
->completed
) {
1317 /* No, so just accelerate recent callbacks. */
1318 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1322 /* Advance callbacks. */
1323 rcu_advance_cbs(rsp
, rnp
, rdp
);
1325 /* Remember that we saw this grace-period completion. */
1326 rdp
->completed
= rnp
->completed
;
1327 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuend");
1330 * If we were in an extended quiescent state, we may have
1331 * missed some grace periods that others CPUs handled on
1332 * our behalf. Catch up with this state to avoid noting
1333 * spurious new grace periods. If another grace period
1334 * has started, then rnp->gpnum will have advanced, so
1335 * we will detect this later on. Of course, any quiescent
1336 * states we found for the old GP are now invalid.
1338 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
)) {
1339 rdp
->gpnum
= rdp
->completed
;
1340 rdp
->passed_quiesce
= 0;
1344 * If RCU does not need a quiescent state from this CPU,
1345 * then make sure that this CPU doesn't go looking for one.
1347 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
1348 rdp
->qs_pending
= 0;
1353 * Advance this CPU's callbacks, but only if the current grace period
1354 * has ended. This may be called only from the CPU to whom the rdp
1358 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1360 unsigned long flags
;
1361 struct rcu_node
*rnp
;
1363 local_irq_save(flags
);
1365 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
1366 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1367 local_irq_restore(flags
);
1370 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1371 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1375 * Do per-CPU grace-period initialization for running CPU. The caller
1376 * must hold the lock of the leaf rcu_node structure corresponding to
1380 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1382 /* Prior grace period ended, so advance callbacks for current CPU. */
1383 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1385 /* Set state so that this CPU will detect the next quiescent state. */
1386 __note_new_gpnum(rsp
, rnp
, rdp
);
1390 * Initialize a new grace period.
1392 static int rcu_gp_init(struct rcu_state
*rsp
)
1394 struct rcu_data
*rdp
;
1395 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1397 raw_spin_lock_irq(&rnp
->lock
);
1398 rsp
->gp_flags
= 0; /* Clear all flags: New grace period. */
1400 if (rcu_gp_in_progress(rsp
)) {
1401 /* Grace period already in progress, don't start another. */
1402 raw_spin_unlock_irq(&rnp
->lock
);
1406 /* Advance to a new grace period and initialize state. */
1408 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, "start");
1409 record_gp_stall_check_time(rsp
);
1410 raw_spin_unlock_irq(&rnp
->lock
);
1412 /* Exclude any concurrent CPU-hotplug operations. */
1413 mutex_lock(&rsp
->onoff_mutex
);
1416 * Set the quiescent-state-needed bits in all the rcu_node
1417 * structures for all currently online CPUs in breadth-first order,
1418 * starting from the root rcu_node structure, relying on the layout
1419 * of the tree within the rsp->node[] array. Note that other CPUs
1420 * will access only the leaves of the hierarchy, thus seeing that no
1421 * grace period is in progress, at least until the corresponding
1422 * leaf node has been initialized. In addition, we have excluded
1423 * CPU-hotplug operations.
1425 * The grace period cannot complete until the initialization
1426 * process finishes, because this kthread handles both.
1428 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1429 raw_spin_lock_irq(&rnp
->lock
);
1430 rdp
= this_cpu_ptr(rsp
->rda
);
1431 rcu_preempt_check_blocked_tasks(rnp
);
1432 rnp
->qsmask
= rnp
->qsmaskinit
;
1433 ACCESS_ONCE(rnp
->gpnum
) = rsp
->gpnum
;
1434 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1435 ACCESS_ONCE(rnp
->completed
) = rsp
->completed
;
1436 if (rnp
== rdp
->mynode
)
1437 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
1438 rcu_preempt_boost_start_gp(rnp
);
1439 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1440 rnp
->level
, rnp
->grplo
,
1441 rnp
->grphi
, rnp
->qsmask
);
1442 raw_spin_unlock_irq(&rnp
->lock
);
1443 #ifdef CONFIG_PROVE_RCU_DELAY
1444 if ((prandom_u32() % (rcu_num_nodes
* 8)) == 0 &&
1445 system_state
== SYSTEM_RUNNING
)
1446 schedule_timeout_uninterruptible(2);
1447 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1451 mutex_unlock(&rsp
->onoff_mutex
);
1456 * Do one round of quiescent-state forcing.
1458 int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1460 int fqs_state
= fqs_state_in
;
1461 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1464 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1465 /* Collect dyntick-idle snapshots. */
1466 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1467 fqs_state
= RCU_FORCE_QS
;
1469 /* Handle dyntick-idle and offline CPUs. */
1470 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1472 /* Clear flag to prevent immediate re-entry. */
1473 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1474 raw_spin_lock_irq(&rnp
->lock
);
1475 rsp
->gp_flags
&= ~RCU_GP_FLAG_FQS
;
1476 raw_spin_unlock_irq(&rnp
->lock
);
1482 * Clean up after the old grace period.
1484 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1486 unsigned long gp_duration
;
1488 struct rcu_data
*rdp
;
1489 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1491 raw_spin_lock_irq(&rnp
->lock
);
1492 gp_duration
= jiffies
- rsp
->gp_start
;
1493 if (gp_duration
> rsp
->gp_max
)
1494 rsp
->gp_max
= gp_duration
;
1497 * We know the grace period is complete, but to everyone else
1498 * it appears to still be ongoing. But it is also the case
1499 * that to everyone else it looks like there is nothing that
1500 * they can do to advance the grace period. It is therefore
1501 * safe for us to drop the lock in order to mark the grace
1502 * period as completed in all of the rcu_node structures.
1504 raw_spin_unlock_irq(&rnp
->lock
);
1507 * Propagate new ->completed value to rcu_node structures so
1508 * that other CPUs don't have to wait until the start of the next
1509 * grace period to process their callbacks. This also avoids
1510 * some nasty RCU grace-period initialization races by forcing
1511 * the end of the current grace period to be completely recorded in
1512 * all of the rcu_node structures before the beginning of the next
1513 * grace period is recorded in any of the rcu_node structures.
1515 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1516 raw_spin_lock_irq(&rnp
->lock
);
1517 ACCESS_ONCE(rnp
->completed
) = rsp
->gpnum
;
1518 rdp
= this_cpu_ptr(rsp
->rda
);
1519 if (rnp
== rdp
->mynode
)
1520 __rcu_process_gp_end(rsp
, rnp
, rdp
);
1521 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1522 raw_spin_unlock_irq(&rnp
->lock
);
1525 rnp
= rcu_get_root(rsp
);
1526 raw_spin_lock_irq(&rnp
->lock
);
1527 rcu_nocb_gp_set(rnp
, nocb
);
1529 rsp
->completed
= rsp
->gpnum
; /* Declare grace period done. */
1530 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, "end");
1531 rsp
->fqs_state
= RCU_GP_IDLE
;
1532 rdp
= this_cpu_ptr(rsp
->rda
);
1533 rcu_advance_cbs(rsp
, rnp
, rdp
); /* Reduce false positives below. */
1534 if (cpu_needs_another_gp(rsp
, rdp
))
1536 raw_spin_unlock_irq(&rnp
->lock
);
1540 * Body of kthread that handles grace periods.
1542 static int __noreturn
rcu_gp_kthread(void *arg
)
1547 struct rcu_state
*rsp
= arg
;
1548 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1552 /* Handle grace-period start. */
1554 wait_event_interruptible(rsp
->gp_wq
,
1557 if ((rsp
->gp_flags
& RCU_GP_FLAG_INIT
) &&
1561 flush_signals(current
);
1564 /* Handle quiescent-state forcing. */
1565 fqs_state
= RCU_SAVE_DYNTICK
;
1566 j
= jiffies_till_first_fqs
;
1569 jiffies_till_first_fqs
= HZ
;
1572 rsp
->jiffies_force_qs
= jiffies
+ j
;
1573 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1574 (rsp
->gp_flags
& RCU_GP_FLAG_FQS
) ||
1575 (!ACCESS_ONCE(rnp
->qsmask
) &&
1576 !rcu_preempt_blocked_readers_cgp(rnp
)),
1578 /* If grace period done, leave loop. */
1579 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1580 !rcu_preempt_blocked_readers_cgp(rnp
))
1582 /* If time for quiescent-state forcing, do it. */
1583 if (ret
== 0 || (rsp
->gp_flags
& RCU_GP_FLAG_FQS
)) {
1584 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1587 /* Deal with stray signal. */
1589 flush_signals(current
);
1591 j
= jiffies_till_next_fqs
;
1594 jiffies_till_next_fqs
= HZ
;
1597 jiffies_till_next_fqs
= 1;
1601 /* Handle grace-period end. */
1602 rcu_gp_cleanup(rsp
);
1607 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1608 * in preparation for detecting the next grace period. The caller must hold
1609 * the root node's ->lock and hard irqs must be disabled.
1611 * Note that it is legal for a dying CPU (which is marked as offline) to
1612 * invoke this function. This can happen when the dying CPU reports its
1616 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1617 struct rcu_data
*rdp
)
1619 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
1621 * Either we have not yet spawned the grace-period
1622 * task, this CPU does not need another grace period,
1623 * or a grace period is already in progress.
1624 * Either way, don't start a new grace period.
1628 rsp
->gp_flags
= RCU_GP_FLAG_INIT
;
1630 /* Wake up rcu_gp_kthread() to start the grace period. */
1631 wake_up(&rsp
->gp_wq
);
1635 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1636 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1637 * is invoked indirectly from rcu_advance_cbs(), which would result in
1638 * endless recursion -- or would do so if it wasn't for the self-deadlock
1639 * that is encountered beforehand.
1642 rcu_start_gp(struct rcu_state
*rsp
)
1644 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1645 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1648 * If there is no grace period in progress right now, any
1649 * callbacks we have up to this point will be satisfied by the
1650 * next grace period. Also, advancing the callbacks reduces the
1651 * probability of false positives from cpu_needs_another_gp()
1652 * resulting in pointless grace periods. So, advance callbacks
1653 * then start the grace period!
1655 rcu_advance_cbs(rsp
, rnp
, rdp
);
1656 rcu_start_gp_advanced(rsp
, rnp
, rdp
);
1660 * Report a full set of quiescent states to the specified rcu_state
1661 * data structure. This involves cleaning up after the prior grace
1662 * period and letting rcu_start_gp() start up the next grace period
1663 * if one is needed. Note that the caller must hold rnp->lock, which
1664 * is released before return.
1666 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1667 __releases(rcu_get_root(rsp
)->lock
)
1669 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1670 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1671 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1675 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1676 * Allows quiescent states for a group of CPUs to be reported at one go
1677 * to the specified rcu_node structure, though all the CPUs in the group
1678 * must be represented by the same rcu_node structure (which need not be
1679 * a leaf rcu_node structure, though it often will be). That structure's
1680 * lock must be held upon entry, and it is released before return.
1683 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1684 struct rcu_node
*rnp
, unsigned long flags
)
1685 __releases(rnp
->lock
)
1687 struct rcu_node
*rnp_c
;
1689 /* Walk up the rcu_node hierarchy. */
1691 if (!(rnp
->qsmask
& mask
)) {
1693 /* Our bit has already been cleared, so done. */
1694 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1697 rnp
->qsmask
&= ~mask
;
1698 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1699 mask
, rnp
->qsmask
, rnp
->level
,
1700 rnp
->grplo
, rnp
->grphi
,
1702 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1704 /* Other bits still set at this level, so done. */
1705 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1708 mask
= rnp
->grpmask
;
1709 if (rnp
->parent
== NULL
) {
1711 /* No more levels. Exit loop holding root lock. */
1715 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1718 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1719 WARN_ON_ONCE(rnp_c
->qsmask
);
1723 * Get here if we are the last CPU to pass through a quiescent
1724 * state for this grace period. Invoke rcu_report_qs_rsp()
1725 * to clean up and start the next grace period if one is needed.
1727 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1731 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1732 * structure. This must be either called from the specified CPU, or
1733 * called when the specified CPU is known to be offline (and when it is
1734 * also known that no other CPU is concurrently trying to help the offline
1735 * CPU). The lastcomp argument is used to make sure we are still in the
1736 * grace period of interest. We don't want to end the current grace period
1737 * based on quiescent states detected in an earlier grace period!
1740 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1742 unsigned long flags
;
1744 struct rcu_node
*rnp
;
1747 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1748 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
1749 rnp
->completed
== rnp
->gpnum
) {
1752 * The grace period in which this quiescent state was
1753 * recorded has ended, so don't report it upwards.
1754 * We will instead need a new quiescent state that lies
1755 * within the current grace period.
1757 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
1758 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1761 mask
= rdp
->grpmask
;
1762 if ((rnp
->qsmask
& mask
) == 0) {
1763 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1765 rdp
->qs_pending
= 0;
1768 * This GP can't end until cpu checks in, so all of our
1769 * callbacks can be processed during the next GP.
1771 rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1773 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1778 * Check to see if there is a new grace period of which this CPU
1779 * is not yet aware, and if so, set up local rcu_data state for it.
1780 * Otherwise, see if this CPU has just passed through its first
1781 * quiescent state for this grace period, and record that fact if so.
1784 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1786 /* If there is now a new grace period, record and return. */
1787 if (check_for_new_grace_period(rsp
, rdp
))
1791 * Does this CPU still need to do its part for current grace period?
1792 * If no, return and let the other CPUs do their part as well.
1794 if (!rdp
->qs_pending
)
1798 * Was there a quiescent state since the beginning of the grace
1799 * period? If no, then exit and wait for the next call.
1801 if (!rdp
->passed_quiesce
)
1805 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1808 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
1811 #ifdef CONFIG_HOTPLUG_CPU
1814 * Send the specified CPU's RCU callbacks to the orphanage. The
1815 * specified CPU must be offline, and the caller must hold the
1819 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
1820 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1822 /* No-CBs CPUs do not have orphanable callbacks. */
1823 if (is_nocb_cpu(rdp
->cpu
))
1827 * Orphan the callbacks. First adjust the counts. This is safe
1828 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1829 * cannot be running now. Thus no memory barrier is required.
1831 if (rdp
->nxtlist
!= NULL
) {
1832 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
1833 rsp
->qlen
+= rdp
->qlen
;
1834 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1836 ACCESS_ONCE(rdp
->qlen
) = 0;
1840 * Next, move those callbacks still needing a grace period to
1841 * the orphanage, where some other CPU will pick them up.
1842 * Some of the callbacks might have gone partway through a grace
1843 * period, but that is too bad. They get to start over because we
1844 * cannot assume that grace periods are synchronized across CPUs.
1845 * We don't bother updating the ->nxttail[] array yet, instead
1846 * we just reset the whole thing later on.
1848 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
1849 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1850 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
1851 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1855 * Then move the ready-to-invoke callbacks to the orphanage,
1856 * where some other CPU will pick them up. These will not be
1857 * required to pass though another grace period: They are done.
1859 if (rdp
->nxtlist
!= NULL
) {
1860 *rsp
->orphan_donetail
= rdp
->nxtlist
;
1861 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1864 /* Finally, initialize the rcu_data structure's list to empty. */
1865 init_callback_list(rdp
);
1869 * Adopt the RCU callbacks from the specified rcu_state structure's
1870 * orphanage. The caller must hold the ->orphan_lock.
1872 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1875 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
1877 /* No-CBs CPUs are handled specially. */
1878 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
))
1881 /* Do the accounting first. */
1882 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
1883 rdp
->qlen
+= rsp
->qlen
;
1884 rdp
->n_cbs_adopted
+= rsp
->qlen
;
1885 if (rsp
->qlen_lazy
!= rsp
->qlen
)
1886 rcu_idle_count_callbacks_posted();
1891 * We do not need a memory barrier here because the only way we
1892 * can get here if there is an rcu_barrier() in flight is if
1893 * we are the task doing the rcu_barrier().
1896 /* First adopt the ready-to-invoke callbacks. */
1897 if (rsp
->orphan_donelist
!= NULL
) {
1898 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1899 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
1900 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
1901 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1902 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
1903 rsp
->orphan_donelist
= NULL
;
1904 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
1907 /* And then adopt the callbacks that still need a grace period. */
1908 if (rsp
->orphan_nxtlist
!= NULL
) {
1909 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
1910 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
1911 rsp
->orphan_nxtlist
= NULL
;
1912 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
1917 * Trace the fact that this CPU is going offline.
1919 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
1921 RCU_TRACE(unsigned long mask
);
1922 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
1923 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
1925 RCU_TRACE(mask
= rdp
->grpmask
);
1926 trace_rcu_grace_period(rsp
->name
,
1927 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
1932 * The CPU has been completely removed, and some other CPU is reporting
1933 * this fact from process context. Do the remainder of the cleanup,
1934 * including orphaning the outgoing CPU's RCU callbacks, and also
1935 * adopting them. There can only be one CPU hotplug operation at a time,
1936 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1938 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
1940 unsigned long flags
;
1942 int need_report
= 0;
1943 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1944 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
1946 /* Adjust any no-longer-needed kthreads. */
1947 rcu_boost_kthread_setaffinity(rnp
, -1);
1949 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1951 /* Exclude any attempts to start a new grace period. */
1952 mutex_lock(&rsp
->onoff_mutex
);
1953 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
1955 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1956 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
1957 rcu_adopt_orphan_cbs(rsp
);
1959 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1960 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1962 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1963 rnp
->qsmaskinit
&= ~mask
;
1964 if (rnp
->qsmaskinit
!= 0) {
1965 if (rnp
!= rdp
->mynode
)
1966 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1969 if (rnp
== rdp
->mynode
)
1970 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1972 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1973 mask
= rnp
->grpmask
;
1975 } while (rnp
!= NULL
);
1978 * We still hold the leaf rcu_node structure lock here, and
1979 * irqs are still disabled. The reason for this subterfuge is
1980 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1981 * held leads to deadlock.
1983 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
1985 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1986 rcu_report_unblock_qs_rnp(rnp
, flags
);
1988 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1989 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1990 rcu_report_exp_rnp(rsp
, rnp
, true);
1991 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
1992 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1993 cpu
, rdp
->qlen
, rdp
->nxtlist
);
1994 init_callback_list(rdp
);
1995 /* Disallow further callbacks on this CPU. */
1996 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
1997 mutex_unlock(&rsp
->onoff_mutex
);
2000 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2002 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2006 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2010 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2013 * Invoke any RCU callbacks that have made it to the end of their grace
2014 * period. Thottle as specified by rdp->blimit.
2016 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2018 unsigned long flags
;
2019 struct rcu_head
*next
, *list
, **tail
;
2020 long bl
, count
, count_lazy
;
2023 /* If no callbacks are ready, just return. */
2024 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2025 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2026 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
2027 need_resched(), is_idle_task(current
),
2028 rcu_is_callbacks_kthread());
2033 * Extract the list of ready callbacks, disabling to prevent
2034 * races with call_rcu() from interrupt handlers.
2036 local_irq_save(flags
);
2037 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2039 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2040 list
= rdp
->nxtlist
;
2041 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2042 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2043 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2044 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2045 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2046 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2047 local_irq_restore(flags
);
2049 /* Invoke callbacks. */
2050 count
= count_lazy
= 0;
2054 debug_rcu_head_unqueue(list
);
2055 if (__rcu_reclaim(rsp
->name
, list
))
2058 /* Stop only if limit reached and CPU has something to do. */
2059 if (++count
>= bl
&&
2061 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2065 local_irq_save(flags
);
2066 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2067 is_idle_task(current
),
2068 rcu_is_callbacks_kthread());
2070 /* Update count, and requeue any remaining callbacks. */
2072 *tail
= rdp
->nxtlist
;
2073 rdp
->nxtlist
= list
;
2074 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2075 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2076 rdp
->nxttail
[i
] = tail
;
2080 smp_mb(); /* List handling before counting for rcu_barrier(). */
2081 rdp
->qlen_lazy
-= count_lazy
;
2082 ACCESS_ONCE(rdp
->qlen
) -= count
;
2083 rdp
->n_cbs_invoked
+= count
;
2085 /* Reinstate batch limit if we have worked down the excess. */
2086 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2087 rdp
->blimit
= blimit
;
2089 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2090 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2091 rdp
->qlen_last_fqs_check
= 0;
2092 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2093 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2094 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2095 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2097 local_irq_restore(flags
);
2099 /* Re-invoke RCU core processing if there are callbacks remaining. */
2100 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2105 * Check to see if this CPU is in a non-context-switch quiescent state
2106 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2107 * Also schedule RCU core processing.
2109 * This function must be called from hardirq context. It is normally
2110 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2111 * false, there is no point in invoking rcu_check_callbacks().
2113 void rcu_check_callbacks(int cpu
, int user
)
2115 trace_rcu_utilization("Start scheduler-tick");
2116 increment_cpu_stall_ticks();
2117 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2120 * Get here if this CPU took its interrupt from user
2121 * mode or from the idle loop, and if this is not a
2122 * nested interrupt. In this case, the CPU is in
2123 * a quiescent state, so note it.
2125 * No memory barrier is required here because both
2126 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2127 * variables that other CPUs neither access nor modify,
2128 * at least not while the corresponding CPU is online.
2134 } else if (!in_softirq()) {
2137 * Get here if this CPU did not take its interrupt from
2138 * softirq, in other words, if it is not interrupting
2139 * a rcu_bh read-side critical section. This is an _bh
2140 * critical section, so note it.
2145 rcu_preempt_check_callbacks(cpu
);
2146 if (rcu_pending(cpu
))
2148 trace_rcu_utilization("End scheduler-tick");
2152 * Scan the leaf rcu_node structures, processing dyntick state for any that
2153 * have not yet encountered a quiescent state, using the function specified.
2154 * Also initiate boosting for any threads blocked on the root rcu_node.
2156 * The caller must have suppressed start of new grace periods.
2158 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
2162 unsigned long flags
;
2164 struct rcu_node
*rnp
;
2166 rcu_for_each_leaf_node(rsp
, rnp
) {
2169 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2170 if (!rcu_gp_in_progress(rsp
)) {
2171 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2174 if (rnp
->qsmask
== 0) {
2175 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
2180 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2181 if ((rnp
->qsmask
& bit
) != 0 &&
2182 f(per_cpu_ptr(rsp
->rda
, cpu
)))
2187 /* rcu_report_qs_rnp() releases rnp->lock. */
2188 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
2191 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2193 rnp
= rcu_get_root(rsp
);
2194 if (rnp
->qsmask
== 0) {
2195 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2196 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
2201 * Force quiescent states on reluctant CPUs, and also detect which
2202 * CPUs are in dyntick-idle mode.
2204 static void force_quiescent_state(struct rcu_state
*rsp
)
2206 unsigned long flags
;
2208 struct rcu_node
*rnp
;
2209 struct rcu_node
*rnp_old
= NULL
;
2211 /* Funnel through hierarchy to reduce memory contention. */
2212 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
2213 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2214 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2215 !raw_spin_trylock(&rnp
->fqslock
);
2216 if (rnp_old
!= NULL
)
2217 raw_spin_unlock(&rnp_old
->fqslock
);
2219 rsp
->n_force_qs_lh
++;
2224 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2226 /* Reached the root of the rcu_node tree, acquire lock. */
2227 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2228 raw_spin_unlock(&rnp_old
->fqslock
);
2229 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2230 rsp
->n_force_qs_lh
++;
2231 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2232 return; /* Someone beat us to it. */
2234 rsp
->gp_flags
|= RCU_GP_FLAG_FQS
;
2235 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2236 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
2240 * This does the RCU core processing work for the specified rcu_state
2241 * and rcu_data structures. This may be called only from the CPU to
2242 * whom the rdp belongs.
2245 __rcu_process_callbacks(struct rcu_state
*rsp
)
2247 unsigned long flags
;
2248 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2250 WARN_ON_ONCE(rdp
->beenonline
== 0);
2252 /* Handle the end of a grace period that some other CPU ended. */
2253 rcu_process_gp_end(rsp
, rdp
);
2255 /* Update RCU state based on any recent quiescent states. */
2256 rcu_check_quiescent_state(rsp
, rdp
);
2258 /* Does this CPU require a not-yet-started grace period? */
2259 local_irq_save(flags
);
2260 if (cpu_needs_another_gp(rsp
, rdp
)) {
2261 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2263 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2265 local_irq_restore(flags
);
2268 /* If there are callbacks ready, invoke them. */
2269 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2270 invoke_rcu_callbacks(rsp
, rdp
);
2274 * Do RCU core processing for the current CPU.
2276 static void rcu_process_callbacks(struct softirq_action
*unused
)
2278 struct rcu_state
*rsp
;
2280 if (cpu_is_offline(smp_processor_id()))
2282 trace_rcu_utilization("Start RCU core");
2283 for_each_rcu_flavor(rsp
)
2284 __rcu_process_callbacks(rsp
);
2285 trace_rcu_utilization("End RCU core");
2289 * Schedule RCU callback invocation. If the specified type of RCU
2290 * does not support RCU priority boosting, just do a direct call,
2291 * otherwise wake up the per-CPU kernel kthread. Note that because we
2292 * are running on the current CPU with interrupts disabled, the
2293 * rcu_cpu_kthread_task cannot disappear out from under us.
2295 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2297 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2299 if (likely(!rsp
->boost
)) {
2300 rcu_do_batch(rsp
, rdp
);
2303 invoke_rcu_callbacks_kthread();
2306 static void invoke_rcu_core(void)
2308 if (cpu_online(smp_processor_id()))
2309 raise_softirq(RCU_SOFTIRQ
);
2313 * Handle any core-RCU processing required by a call_rcu() invocation.
2315 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2316 struct rcu_head
*head
, unsigned long flags
)
2319 * If called from an extended quiescent state, invoke the RCU
2320 * core in order to force a re-evaluation of RCU's idleness.
2322 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2325 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2326 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2330 * Force the grace period if too many callbacks or too long waiting.
2331 * Enforce hysteresis, and don't invoke force_quiescent_state()
2332 * if some other CPU has recently done so. Also, don't bother
2333 * invoking force_quiescent_state() if the newly enqueued callback
2334 * is the only one waiting for a grace period to complete.
2336 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2338 /* Are we ignoring a completed grace period? */
2339 rcu_process_gp_end(rsp
, rdp
);
2340 check_for_new_grace_period(rsp
, rdp
);
2342 /* Start a new grace period if one not already started. */
2343 if (!rcu_gp_in_progress(rsp
)) {
2344 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2346 raw_spin_lock(&rnp_root
->lock
);
2348 raw_spin_unlock(&rnp_root
->lock
);
2350 /* Give the grace period a kick. */
2351 rdp
->blimit
= LONG_MAX
;
2352 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2353 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2354 force_quiescent_state(rsp
);
2355 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2356 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2362 * Helper function for call_rcu() and friends. The cpu argument will
2363 * normally be -1, indicating "currently running CPU". It may specify
2364 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2365 * is expected to specify a CPU.
2368 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2369 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2371 unsigned long flags
;
2372 struct rcu_data
*rdp
;
2374 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2375 debug_rcu_head_queue(head
);
2380 * Opportunistically note grace-period endings and beginnings.
2381 * Note that we might see a beginning right after we see an
2382 * end, but never vice versa, since this CPU has to pass through
2383 * a quiescent state betweentimes.
2385 local_irq_save(flags
);
2386 rdp
= this_cpu_ptr(rsp
->rda
);
2388 /* Add the callback to our list. */
2389 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2393 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2394 offline
= !__call_rcu_nocb(rdp
, head
, lazy
);
2395 WARN_ON_ONCE(offline
);
2396 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2397 local_irq_restore(flags
);
2400 ACCESS_ONCE(rdp
->qlen
)++;
2404 rcu_idle_count_callbacks_posted();
2405 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2406 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2407 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2409 if (__is_kfree_rcu_offset((unsigned long)func
))
2410 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2411 rdp
->qlen_lazy
, rdp
->qlen
);
2413 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2415 /* Go handle any RCU core processing required. */
2416 __call_rcu_core(rsp
, rdp
, head
, flags
);
2417 local_irq_restore(flags
);
2421 * Queue an RCU-sched callback for invocation after a grace period.
2423 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2425 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2427 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2430 * Queue an RCU callback for invocation after a quicker grace period.
2432 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2434 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2436 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2439 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2440 * any blocking grace-period wait automatically implies a grace period
2441 * if there is only one CPU online at any point time during execution
2442 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2443 * occasionally incorrectly indicate that there are multiple CPUs online
2444 * when there was in fact only one the whole time, as this just adds
2445 * some overhead: RCU still operates correctly.
2447 static inline int rcu_blocking_is_gp(void)
2451 might_sleep(); /* Check for RCU read-side critical section. */
2453 ret
= num_online_cpus() <= 1;
2459 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2461 * Control will return to the caller some time after a full rcu-sched
2462 * grace period has elapsed, in other words after all currently executing
2463 * rcu-sched read-side critical sections have completed. These read-side
2464 * critical sections are delimited by rcu_read_lock_sched() and
2465 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2466 * local_irq_disable(), and so on may be used in place of
2467 * rcu_read_lock_sched().
2469 * This means that all preempt_disable code sequences, including NMI and
2470 * non-threaded hardware-interrupt handlers, in progress on entry will
2471 * have completed before this primitive returns. However, this does not
2472 * guarantee that softirq handlers will have completed, since in some
2473 * kernels, these handlers can run in process context, and can block.
2475 * Note that this guarantee implies further memory-ordering guarantees.
2476 * On systems with more than one CPU, when synchronize_sched() returns,
2477 * each CPU is guaranteed to have executed a full memory barrier since the
2478 * end of its last RCU-sched read-side critical section whose beginning
2479 * preceded the call to synchronize_sched(). In addition, each CPU having
2480 * an RCU read-side critical section that extends beyond the return from
2481 * synchronize_sched() is guaranteed to have executed a full memory barrier
2482 * after the beginning of synchronize_sched() and before the beginning of
2483 * that RCU read-side critical section. Note that these guarantees include
2484 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2485 * that are executing in the kernel.
2487 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2488 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2489 * to have executed a full memory barrier during the execution of
2490 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2491 * again only if the system has more than one CPU).
2493 * This primitive provides the guarantees made by the (now removed)
2494 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2495 * guarantees that rcu_read_lock() sections will have completed.
2496 * In "classic RCU", these two guarantees happen to be one and
2497 * the same, but can differ in realtime RCU implementations.
2499 void synchronize_sched(void)
2501 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2502 !lock_is_held(&rcu_lock_map
) &&
2503 !lock_is_held(&rcu_sched_lock_map
),
2504 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2505 if (rcu_blocking_is_gp())
2508 synchronize_sched_expedited();
2510 wait_rcu_gp(call_rcu_sched
);
2512 EXPORT_SYMBOL_GPL(synchronize_sched
);
2515 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2517 * Control will return to the caller some time after a full rcu_bh grace
2518 * period has elapsed, in other words after all currently executing rcu_bh
2519 * read-side critical sections have completed. RCU read-side critical
2520 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2521 * and may be nested.
2523 * See the description of synchronize_sched() for more detailed information
2524 * on memory ordering guarantees.
2526 void synchronize_rcu_bh(void)
2528 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2529 !lock_is_held(&rcu_lock_map
) &&
2530 !lock_is_held(&rcu_sched_lock_map
),
2531 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2532 if (rcu_blocking_is_gp())
2535 synchronize_rcu_bh_expedited();
2537 wait_rcu_gp(call_rcu_bh
);
2539 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2541 static int synchronize_sched_expedited_cpu_stop(void *data
)
2544 * There must be a full memory barrier on each affected CPU
2545 * between the time that try_stop_cpus() is called and the
2546 * time that it returns.
2548 * In the current initial implementation of cpu_stop, the
2549 * above condition is already met when the control reaches
2550 * this point and the following smp_mb() is not strictly
2551 * necessary. Do smp_mb() anyway for documentation and
2552 * robustness against future implementation changes.
2554 smp_mb(); /* See above comment block. */
2559 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2561 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2562 * approach to force the grace period to end quickly. This consumes
2563 * significant time on all CPUs and is unfriendly to real-time workloads,
2564 * so is thus not recommended for any sort of common-case code. In fact,
2565 * if you are using synchronize_sched_expedited() in a loop, please
2566 * restructure your code to batch your updates, and then use a single
2567 * synchronize_sched() instead.
2569 * Note that it is illegal to call this function while holding any lock
2570 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2571 * to call this function from a CPU-hotplug notifier. Failing to observe
2572 * these restriction will result in deadlock.
2574 * This implementation can be thought of as an application of ticket
2575 * locking to RCU, with sync_sched_expedited_started and
2576 * sync_sched_expedited_done taking on the roles of the halves
2577 * of the ticket-lock word. Each task atomically increments
2578 * sync_sched_expedited_started upon entry, snapshotting the old value,
2579 * then attempts to stop all the CPUs. If this succeeds, then each
2580 * CPU will have executed a context switch, resulting in an RCU-sched
2581 * grace period. We are then done, so we use atomic_cmpxchg() to
2582 * update sync_sched_expedited_done to match our snapshot -- but
2583 * only if someone else has not already advanced past our snapshot.
2585 * On the other hand, if try_stop_cpus() fails, we check the value
2586 * of sync_sched_expedited_done. If it has advanced past our
2587 * initial snapshot, then someone else must have forced a grace period
2588 * some time after we took our snapshot. In this case, our work is
2589 * done for us, and we can simply return. Otherwise, we try again,
2590 * but keep our initial snapshot for purposes of checking for someone
2591 * doing our work for us.
2593 * If we fail too many times in a row, we fall back to synchronize_sched().
2595 void synchronize_sched_expedited(void)
2597 long firstsnap
, s
, snap
;
2599 struct rcu_state
*rsp
= &rcu_sched_state
;
2602 * If we are in danger of counter wrap, just do synchronize_sched().
2603 * By allowing sync_sched_expedited_started to advance no more than
2604 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2605 * that more than 3.5 billion CPUs would be required to force a
2606 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2607 * course be required on a 64-bit system.
2609 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
2610 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
2612 synchronize_sched();
2613 atomic_long_inc(&rsp
->expedited_wrap
);
2618 * Take a ticket. Note that atomic_inc_return() implies a
2619 * full memory barrier.
2621 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
2624 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2627 * Each pass through the following loop attempts to force a
2628 * context switch on each CPU.
2630 while (try_stop_cpus(cpu_online_mask
,
2631 synchronize_sched_expedited_cpu_stop
,
2634 atomic_long_inc(&rsp
->expedited_tryfail
);
2636 /* Check to see if someone else did our work for us. */
2637 s
= atomic_long_read(&rsp
->expedited_done
);
2638 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2639 /* ensure test happens before caller kfree */
2640 smp_mb__before_atomic_inc(); /* ^^^ */
2641 atomic_long_inc(&rsp
->expedited_workdone1
);
2645 /* No joy, try again later. Or just synchronize_sched(). */
2646 if (trycount
++ < 10) {
2647 udelay(trycount
* num_online_cpus());
2649 wait_rcu_gp(call_rcu_sched
);
2650 atomic_long_inc(&rsp
->expedited_normal
);
2654 /* Recheck to see if someone else did our work for us. */
2655 s
= atomic_long_read(&rsp
->expedited_done
);
2656 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
2657 /* ensure test happens before caller kfree */
2658 smp_mb__before_atomic_inc(); /* ^^^ */
2659 atomic_long_inc(&rsp
->expedited_workdone2
);
2664 * Refetching sync_sched_expedited_started allows later
2665 * callers to piggyback on our grace period. We retry
2666 * after they started, so our grace period works for them,
2667 * and they started after our first try, so their grace
2668 * period works for us.
2671 snap
= atomic_long_read(&rsp
->expedited_start
);
2672 smp_mb(); /* ensure read is before try_stop_cpus(). */
2674 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
2677 * Everyone up to our most recent fetch is covered by our grace
2678 * period. Update the counter, but only if our work is still
2679 * relevant -- which it won't be if someone who started later
2680 * than we did already did their update.
2683 atomic_long_inc(&rsp
->expedited_done_tries
);
2684 s
= atomic_long_read(&rsp
->expedited_done
);
2685 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
2686 /* ensure test happens before caller kfree */
2687 smp_mb__before_atomic_inc(); /* ^^^ */
2688 atomic_long_inc(&rsp
->expedited_done_lost
);
2691 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
2692 atomic_long_inc(&rsp
->expedited_done_exit
);
2696 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
2699 * Check to see if there is any immediate RCU-related work to be done
2700 * by the current CPU, for the specified type of RCU, returning 1 if so.
2701 * The checks are in order of increasing expense: checks that can be
2702 * carried out against CPU-local state are performed first. However,
2703 * we must check for CPU stalls first, else we might not get a chance.
2705 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2707 struct rcu_node
*rnp
= rdp
->mynode
;
2709 rdp
->n_rcu_pending
++;
2711 /* Check for CPU stalls, if enabled. */
2712 check_cpu_stall(rsp
, rdp
);
2714 /* Is the RCU core waiting for a quiescent state from this CPU? */
2715 if (rcu_scheduler_fully_active
&&
2716 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
2717 rdp
->n_rp_qs_pending
++;
2718 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
2719 rdp
->n_rp_report_qs
++;
2723 /* Does this CPU have callbacks ready to invoke? */
2724 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
2725 rdp
->n_rp_cb_ready
++;
2729 /* Has RCU gone idle with this CPU needing another grace period? */
2730 if (cpu_needs_another_gp(rsp
, rdp
)) {
2731 rdp
->n_rp_cpu_needs_gp
++;
2735 /* Has another RCU grace period completed? */
2736 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
2737 rdp
->n_rp_gp_completed
++;
2741 /* Has a new RCU grace period started? */
2742 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
2743 rdp
->n_rp_gp_started
++;
2748 rdp
->n_rp_need_nothing
++;
2753 * Check to see if there is any immediate RCU-related work to be done
2754 * by the current CPU, returning 1 if so. This function is part of the
2755 * RCU implementation; it is -not- an exported member of the RCU API.
2757 static int rcu_pending(int cpu
)
2759 struct rcu_state
*rsp
;
2761 for_each_rcu_flavor(rsp
)
2762 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
2768 * Return true if the specified CPU has any callback. If all_lazy is
2769 * non-NULL, store an indication of whether all callbacks are lazy.
2770 * (If there are no callbacks, all of them are deemed to be lazy.)
2772 static int rcu_cpu_has_callbacks(int cpu
, bool *all_lazy
)
2776 struct rcu_data
*rdp
;
2777 struct rcu_state
*rsp
;
2779 for_each_rcu_flavor(rsp
) {
2780 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2781 if (rdp
->qlen
!= rdp
->qlen_lazy
)
2792 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2793 * the compiler is expected to optimize this away.
2795 static void _rcu_barrier_trace(struct rcu_state
*rsp
, char *s
,
2796 int cpu
, unsigned long done
)
2798 trace_rcu_barrier(rsp
->name
, s
, cpu
,
2799 atomic_read(&rsp
->barrier_cpu_count
), done
);
2803 * RCU callback function for _rcu_barrier(). If we are last, wake
2804 * up the task executing _rcu_barrier().
2806 static void rcu_barrier_callback(struct rcu_head
*rhp
)
2808 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
2809 struct rcu_state
*rsp
= rdp
->rsp
;
2811 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
2812 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
2813 complete(&rsp
->barrier_completion
);
2815 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
2820 * Called with preemption disabled, and from cross-cpu IRQ context.
2822 static void rcu_barrier_func(void *type
)
2824 struct rcu_state
*rsp
= type
;
2825 struct rcu_data
*rdp
= __this_cpu_ptr(rsp
->rda
);
2827 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
2828 atomic_inc(&rsp
->barrier_cpu_count
);
2829 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
2833 * Orchestrate the specified type of RCU barrier, waiting for all
2834 * RCU callbacks of the specified type to complete.
2836 static void _rcu_barrier(struct rcu_state
*rsp
)
2839 struct rcu_data
*rdp
;
2840 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
2841 unsigned long snap_done
;
2843 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
2845 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2846 mutex_lock(&rsp
->barrier_mutex
);
2849 * Ensure that all prior references, including to ->n_barrier_done,
2850 * are ordered before the _rcu_barrier() machinery.
2852 smp_mb(); /* See above block comment. */
2855 * Recheck ->n_barrier_done to see if others did our work for us.
2856 * This means checking ->n_barrier_done for an even-to-odd-to-even
2857 * transition. The "if" expression below therefore rounds the old
2858 * value up to the next even number and adds two before comparing.
2860 snap_done
= ACCESS_ONCE(rsp
->n_barrier_done
);
2861 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
2862 if (ULONG_CMP_GE(snap_done
, ((snap
+ 1) & ~0x1) + 2)) {
2863 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
2864 smp_mb(); /* caller's subsequent code after above check. */
2865 mutex_unlock(&rsp
->barrier_mutex
);
2870 * Increment ->n_barrier_done to avoid duplicate work. Use
2871 * ACCESS_ONCE() to prevent the compiler from speculating
2872 * the increment to precede the early-exit check.
2874 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2875 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
2876 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
2877 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2880 * Initialize the count to one rather than to zero in order to
2881 * avoid a too-soon return to zero in case of a short grace period
2882 * (or preemption of this task). Exclude CPU-hotplug operations
2883 * to ensure that no offline CPU has callbacks queued.
2885 init_completion(&rsp
->barrier_completion
);
2886 atomic_set(&rsp
->barrier_cpu_count
, 1);
2890 * Force each CPU with callbacks to register a new callback.
2891 * When that callback is invoked, we will know that all of the
2892 * corresponding CPU's preceding callbacks have been invoked.
2894 for_each_possible_cpu(cpu
) {
2895 if (!cpu_online(cpu
) && !is_nocb_cpu(cpu
))
2897 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2898 if (is_nocb_cpu(cpu
)) {
2899 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
2900 rsp
->n_barrier_done
);
2901 atomic_inc(&rsp
->barrier_cpu_count
);
2902 __call_rcu(&rdp
->barrier_head
, rcu_barrier_callback
,
2904 } else if (ACCESS_ONCE(rdp
->qlen
)) {
2905 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
2906 rsp
->n_barrier_done
);
2907 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
2909 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
2910 rsp
->n_barrier_done
);
2916 * Now that we have an rcu_barrier_callback() callback on each
2917 * CPU, and thus each counted, remove the initial count.
2919 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
2920 complete(&rsp
->barrier_completion
);
2922 /* Increment ->n_barrier_done to prevent duplicate work. */
2923 smp_mb(); /* Keep increment after above mechanism. */
2924 ACCESS_ONCE(rsp
->n_barrier_done
)++;
2925 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
2926 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
2927 smp_mb(); /* Keep increment before caller's subsequent code. */
2929 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2930 wait_for_completion(&rsp
->barrier_completion
);
2932 /* Other rcu_barrier() invocations can now safely proceed. */
2933 mutex_unlock(&rsp
->barrier_mutex
);
2937 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2939 void rcu_barrier_bh(void)
2941 _rcu_barrier(&rcu_bh_state
);
2943 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
2946 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2948 void rcu_barrier_sched(void)
2950 _rcu_barrier(&rcu_sched_state
);
2952 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
2955 * Do boot-time initialization of a CPU's per-CPU RCU data.
2958 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
2960 unsigned long flags
;
2961 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2962 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2964 /* Set up local state, ensuring consistent view of global state. */
2965 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2966 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
2967 init_callback_list(rdp
);
2969 ACCESS_ONCE(rdp
->qlen
) = 0;
2970 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
2971 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
2972 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
2975 rcu_boot_init_nocb_percpu_data(rdp
);
2976 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2980 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2981 * offline event can be happening at a given time. Note also that we
2982 * can accept some slop in the rsp->completed access due to the fact
2983 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2985 static void __cpuinit
2986 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
2988 unsigned long flags
;
2990 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2991 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2993 /* Exclude new grace periods. */
2994 mutex_lock(&rsp
->onoff_mutex
);
2996 /* Set up local state, ensuring consistent view of global state. */
2997 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2998 rdp
->beenonline
= 1; /* We have now been online. */
2999 rdp
->preemptible
= preemptible
;
3000 rdp
->qlen_last_fqs_check
= 0;
3001 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3002 rdp
->blimit
= blimit
;
3003 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3004 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3005 atomic_set(&rdp
->dynticks
->dynticks
,
3006 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3007 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3009 /* Add CPU to rcu_node bitmasks. */
3011 mask
= rdp
->grpmask
;
3013 /* Exclude any attempts to start a new GP on small systems. */
3014 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3015 rnp
->qsmaskinit
|= mask
;
3016 mask
= rnp
->grpmask
;
3017 if (rnp
== rdp
->mynode
) {
3019 * If there is a grace period in progress, we will
3020 * set up to wait for it next time we run the
3023 rdp
->gpnum
= rnp
->completed
;
3024 rdp
->completed
= rnp
->completed
;
3025 rdp
->passed_quiesce
= 0;
3026 rdp
->qs_pending
= 0;
3027 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, "cpuonl");
3029 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
3031 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
3032 local_irq_restore(flags
);
3034 mutex_unlock(&rsp
->onoff_mutex
);
3037 static void __cpuinit
rcu_prepare_cpu(int cpu
)
3039 struct rcu_state
*rsp
;
3041 for_each_rcu_flavor(rsp
)
3042 rcu_init_percpu_data(cpu
, rsp
,
3043 strcmp(rsp
->name
, "rcu_preempt") == 0);
3047 * Handle CPU online/offline notification events.
3049 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
3050 unsigned long action
, void *hcpu
)
3052 long cpu
= (long)hcpu
;
3053 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
3054 struct rcu_node
*rnp
= rdp
->mynode
;
3055 struct rcu_state
*rsp
;
3057 trace_rcu_utilization("Start CPU hotplug");
3059 case CPU_UP_PREPARE
:
3060 case CPU_UP_PREPARE_FROZEN
:
3061 rcu_prepare_cpu(cpu
);
3062 rcu_prepare_kthreads(cpu
);
3065 case CPU_DOWN_FAILED
:
3066 rcu_boost_kthread_setaffinity(rnp
, -1);
3068 case CPU_DOWN_PREPARE
:
3069 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3072 case CPU_DYING_FROZEN
:
3073 for_each_rcu_flavor(rsp
)
3074 rcu_cleanup_dying_cpu(rsp
);
3077 case CPU_DEAD_FROZEN
:
3078 case CPU_UP_CANCELED
:
3079 case CPU_UP_CANCELED_FROZEN
:
3080 for_each_rcu_flavor(rsp
)
3081 rcu_cleanup_dead_cpu(cpu
, rsp
);
3086 trace_rcu_utilization("End CPU hotplug");
3091 * Spawn the kthread that handles this RCU flavor's grace periods.
3093 static int __init
rcu_spawn_gp_kthread(void)
3095 unsigned long flags
;
3096 struct rcu_node
*rnp
;
3097 struct rcu_state
*rsp
;
3098 struct task_struct
*t
;
3100 for_each_rcu_flavor(rsp
) {
3101 t
= kthread_run(rcu_gp_kthread
, rsp
, rsp
->name
);
3103 rnp
= rcu_get_root(rsp
);
3104 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3105 rsp
->gp_kthread
= t
;
3106 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3107 rcu_spawn_nocb_kthreads(rsp
);
3111 early_initcall(rcu_spawn_gp_kthread
);
3114 * This function is invoked towards the end of the scheduler's initialization
3115 * process. Before this is called, the idle task might contain
3116 * RCU read-side critical sections (during which time, this idle
3117 * task is booting the system). After this function is called, the
3118 * idle tasks are prohibited from containing RCU read-side critical
3119 * sections. This function also enables RCU lockdep checking.
3121 void rcu_scheduler_starting(void)
3123 WARN_ON(num_online_cpus() != 1);
3124 WARN_ON(nr_context_switches() > 0);
3125 rcu_scheduler_active
= 1;
3129 * Compute the per-level fanout, either using the exact fanout specified
3130 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3132 #ifdef CONFIG_RCU_FANOUT_EXACT
3133 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3137 for (i
= rcu_num_lvls
- 1; i
> 0; i
--)
3138 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3139 rsp
->levelspread
[0] = rcu_fanout_leaf
;
3141 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3142 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3149 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3150 ccur
= rsp
->levelcnt
[i
];
3151 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3155 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3158 * Helper function for rcu_init() that initializes one rcu_state structure.
3160 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3161 struct rcu_data __percpu
*rda
)
3163 static char *buf
[] = { "rcu_node_0",
3166 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3167 static char *fqs
[] = { "rcu_node_fqs_0",
3170 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3174 struct rcu_node
*rnp
;
3176 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3178 /* Silence gcc 4.8 warning about array index out of range. */
3179 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3180 panic("rcu_init_one: rcu_num_lvls overflow");
3182 /* Initialize the level-tracking arrays. */
3184 for (i
= 0; i
< rcu_num_lvls
; i
++)
3185 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3186 for (i
= 1; i
< rcu_num_lvls
; i
++)
3187 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3188 rcu_init_levelspread(rsp
);
3190 /* Initialize the elements themselves, starting from the leaves. */
3192 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3193 cpustride
*= rsp
->levelspread
[i
];
3194 rnp
= rsp
->level
[i
];
3195 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3196 raw_spin_lock_init(&rnp
->lock
);
3197 lockdep_set_class_and_name(&rnp
->lock
,
3198 &rcu_node_class
[i
], buf
[i
]);
3199 raw_spin_lock_init(&rnp
->fqslock
);
3200 lockdep_set_class_and_name(&rnp
->fqslock
,
3201 &rcu_fqs_class
[i
], fqs
[i
]);
3202 rnp
->gpnum
= rsp
->gpnum
;
3203 rnp
->completed
= rsp
->completed
;
3205 rnp
->qsmaskinit
= 0;
3206 rnp
->grplo
= j
* cpustride
;
3207 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
3208 if (rnp
->grphi
>= NR_CPUS
)
3209 rnp
->grphi
= NR_CPUS
- 1;
3215 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
3216 rnp
->grpmask
= 1UL << rnp
->grpnum
;
3217 rnp
->parent
= rsp
->level
[i
- 1] +
3218 j
/ rsp
->levelspread
[i
- 1];
3221 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3222 rcu_init_one_nocb(rnp
);
3227 init_waitqueue_head(&rsp
->gp_wq
);
3228 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3229 for_each_possible_cpu(i
) {
3230 while (i
> rnp
->grphi
)
3232 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3233 rcu_boot_init_percpu_data(i
, rsp
);
3235 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3239 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3240 * replace the definitions in rcutree.h because those are needed to size
3241 * the ->node array in the rcu_state structure.
3243 static void __init
rcu_init_geometry(void)
3248 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3250 /* If the compile-time values are accurate, just leave. */
3251 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3252 nr_cpu_ids
== NR_CPUS
)
3256 * Compute number of nodes that can be handled an rcu_node tree
3257 * with the given number of levels. Setting rcu_capacity[0] makes
3258 * some of the arithmetic easier.
3260 rcu_capacity
[0] = 1;
3261 rcu_capacity
[1] = rcu_fanout_leaf
;
3262 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3263 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3266 * The boot-time rcu_fanout_leaf parameter is only permitted
3267 * to increase the leaf-level fanout, not decrease it. Of course,
3268 * the leaf-level fanout cannot exceed the number of bits in
3269 * the rcu_node masks. Finally, the tree must be able to accommodate
3270 * the configured number of CPUs. Complain and fall back to the
3271 * compile-time values if these limits are exceeded.
3273 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3274 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3275 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3280 /* Calculate the number of rcu_nodes at each level of the tree. */
3281 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3282 if (n
<= rcu_capacity
[i
]) {
3283 for (j
= 0; j
<= i
; j
++)
3285 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3287 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3292 /* Calculate the total number of rcu_node structures. */
3294 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3295 rcu_num_nodes
+= num_rcu_lvl
[i
];
3299 void __init
rcu_init(void)
3303 rcu_bootup_announce();
3304 rcu_init_geometry();
3305 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3306 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3307 __rcu_init_preempt();
3308 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3311 * We don't need protection against CPU-hotplug here because
3312 * this is called early in boot, before either interrupts
3313 * or the scheduler are operational.
3315 cpu_notifier(rcu_cpu_notify
, 0);
3316 for_each_online_cpu(cpu
)
3317 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3320 #include "rcutree_plugin.h"