2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
72 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
83 # define DEFINE_RCU_TPS(sname) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
86 # define RCU_STATE_NAME(sname) sname##_varname
88 # define DEFINE_RCU_TPS(sname)
89 # define RCU_STATE_NAME(sname) __stringify(sname)
92 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
93 DEFINE_RCU_TPS(sname) \
94 struct rcu_state sname##_state = { \
95 .level = { &sname##_state.node[0] }, \
97 .fqs_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
101 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
102 .orphan_donetail = &sname##_state.orphan_donelist, \
103 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
104 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
105 .name = RCU_STATE_NAME(sname), \
108 DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data)
110 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
111 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
113 static struct rcu_state
*rcu_state_p
;
114 LIST_HEAD(rcu_struct_flavors
);
116 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
117 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
118 module_param(rcu_fanout_leaf
, int, 0444);
119 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
120 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
127 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
130 * The rcu_scheduler_active variable transitions from zero to one just
131 * before the first task is spawned. So when this variable is zero, RCU
132 * can assume that there is but one task, allowing RCU to (for example)
133 * optimize synchronize_sched() to a simple barrier(). When this variable
134 * is one, RCU must actually do all the hard work required to detect real
135 * grace periods. This variable is also used to suppress boot-time false
136 * positives from lockdep-RCU error checking.
138 int rcu_scheduler_active __read_mostly
;
139 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
142 * The rcu_scheduler_fully_active variable transitions from zero to one
143 * during the early_initcall() processing, which is after the scheduler
144 * is capable of creating new tasks. So RCU processing (for example,
145 * creating tasks for RCU priority boosting) must be delayed until after
146 * rcu_scheduler_fully_active transitions from zero to one. We also
147 * currently delay invocation of any RCU callbacks until after this point.
149 * It might later prove better for people registering RCU callbacks during
150 * early boot to take responsibility for these callbacks, but one step at
153 static int rcu_scheduler_fully_active __read_mostly
;
155 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
156 static void invoke_rcu_core(void);
157 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
159 /* rcuc/rcub kthread realtime priority */
160 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
161 module_param(kthread_prio
, int, 0644);
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
172 unsigned long rcutorture_testseq
;
173 unsigned long rcutorture_vernum
;
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
180 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
182 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
191 void rcu_sched_qs(void)
193 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
194 trace_rcu_grace_period(TPS("rcu_sched"),
195 __this_cpu_read(rcu_sched_data
.gpnum
),
197 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
203 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
204 trace_rcu_grace_period(TPS("rcu_bh"),
205 __this_cpu_read(rcu_bh_data
.gpnum
),
207 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
211 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
213 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
214 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
215 .dynticks
= ATOMIC_INIT(1),
216 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
217 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
218 .dynticks_idle
= ATOMIC_INIT(1),
219 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
222 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
223 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
226 * Let the RCU core know that this CPU has gone through the scheduler,
227 * which is a quiescent state. This is called when the need for a
228 * quiescent state is urgent, so we burn an atomic operation and full
229 * memory barriers to let the RCU core know about it, regardless of what
230 * this CPU might (or might not) do in the near future.
232 * We inform the RCU core by emulating a zero-duration dyntick-idle
233 * period, which we in turn do by incrementing the ->dynticks counter
236 static void rcu_momentary_dyntick_idle(void)
239 struct rcu_data
*rdp
;
240 struct rcu_dynticks
*rdtp
;
242 struct rcu_state
*rsp
;
244 local_irq_save(flags
);
247 * Yes, we can lose flag-setting operations. This is OK, because
248 * the flag will be set again after some delay.
250 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
251 raw_cpu_write(rcu_sched_qs_mask
, 0);
253 /* Find the flavor that needs a quiescent state. */
254 for_each_rcu_flavor(rsp
) {
255 rdp
= raw_cpu_ptr(rsp
->rda
);
256 if (!(resched_mask
& rsp
->flavor_mask
))
258 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
259 if (ACCESS_ONCE(rdp
->mynode
->completed
) !=
260 ACCESS_ONCE(rdp
->cond_resched_completed
))
264 * Pretend to be momentarily idle for the quiescent state.
265 * This allows the grace-period kthread to record the
266 * quiescent state, with no need for this CPU to do anything
269 rdtp
= this_cpu_ptr(&rcu_dynticks
);
270 smp_mb__before_atomic(); /* Earlier stuff before QS. */
271 atomic_add(2, &rdtp
->dynticks
); /* QS. */
272 smp_mb__after_atomic(); /* Later stuff after QS. */
275 local_irq_restore(flags
);
279 * Note a context switch. This is a quiescent state for RCU-sched,
280 * and requires special handling for preemptible RCU.
281 * The caller must have disabled preemption.
283 void rcu_note_context_switch(void)
285 trace_rcu_utilization(TPS("Start context switch"));
287 rcu_preempt_note_context_switch();
288 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
289 rcu_momentary_dyntick_idle();
290 trace_rcu_utilization(TPS("End context switch"));
292 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
295 * Register a quiesecent state for all RCU flavors. If there is an
296 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
297 * dyntick-idle quiescent state visible to other CPUs (but only for those
298 * RCU flavors in desparate need of a quiescent state, which will normally
299 * be none of them). Either way, do a lightweight quiescent state for
302 void rcu_all_qs(void)
304 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
305 rcu_momentary_dyntick_idle();
306 this_cpu_inc(rcu_qs_ctr
);
308 EXPORT_SYMBOL_GPL(rcu_all_qs
);
310 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
311 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
312 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
314 module_param(blimit
, long, 0444);
315 module_param(qhimark
, long, 0444);
316 module_param(qlowmark
, long, 0444);
318 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
319 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
321 module_param(jiffies_till_first_fqs
, ulong
, 0644);
322 module_param(jiffies_till_next_fqs
, ulong
, 0644);
325 * How long the grace period must be before we start recruiting
326 * quiescent-state help from rcu_note_context_switch().
328 static ulong jiffies_till_sched_qs
= HZ
/ 20;
329 module_param(jiffies_till_sched_qs
, ulong
, 0644);
331 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
332 struct rcu_data
*rdp
);
333 static void force_qs_rnp(struct rcu_state
*rsp
,
334 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
335 unsigned long *maxj
),
336 bool *isidle
, unsigned long *maxj
);
337 static void force_quiescent_state(struct rcu_state
*rsp
);
338 static int rcu_pending(void);
341 * Return the number of RCU-sched batches processed thus far for debug & stats.
343 long rcu_batches_completed_sched(void)
345 return rcu_sched_state
.completed
;
347 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
350 * Return the number of RCU BH batches processed thus far for debug & stats.
352 long rcu_batches_completed_bh(void)
354 return rcu_bh_state
.completed
;
356 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
359 * Force a quiescent state.
361 void rcu_force_quiescent_state(void)
363 force_quiescent_state(rcu_state_p
);
365 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
368 * Force a quiescent state for RCU BH.
370 void rcu_bh_force_quiescent_state(void)
372 force_quiescent_state(&rcu_bh_state
);
374 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
377 * Show the state of the grace-period kthreads.
379 void show_rcu_gp_kthreads(void)
381 struct rcu_state
*rsp
;
383 for_each_rcu_flavor(rsp
) {
384 pr_info("%s: wait state: %d ->state: %#lx\n",
385 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
386 /* sched_show_task(rsp->gp_kthread); */
389 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
392 * Record the number of times rcutorture tests have been initiated and
393 * terminated. This information allows the debugfs tracing stats to be
394 * correlated to the rcutorture messages, even when the rcutorture module
395 * is being repeatedly loaded and unloaded. In other words, we cannot
396 * store this state in rcutorture itself.
398 void rcutorture_record_test_transition(void)
400 rcutorture_testseq
++;
401 rcutorture_vernum
= 0;
403 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
406 * Send along grace-period-related data for rcutorture diagnostics.
408 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
409 unsigned long *gpnum
, unsigned long *completed
)
411 struct rcu_state
*rsp
= NULL
;
420 case RCU_SCHED_FLAVOR
:
421 rsp
= &rcu_sched_state
;
427 *flags
= ACCESS_ONCE(rsp
->gp_flags
);
428 *gpnum
= ACCESS_ONCE(rsp
->gpnum
);
429 *completed
= ACCESS_ONCE(rsp
->completed
);
436 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
439 * Record the number of writer passes through the current rcutorture test.
440 * This is also used to correlate debugfs tracing stats with the rcutorture
443 void rcutorture_record_progress(unsigned long vernum
)
447 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
450 * Force a quiescent state for RCU-sched.
452 void rcu_sched_force_quiescent_state(void)
454 force_quiescent_state(&rcu_sched_state
);
456 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
459 * Does the CPU have callbacks ready to be invoked?
462 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
464 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
465 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
469 * Return the root node of the specified rcu_state structure.
471 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
473 return &rsp
->node
[0];
477 * Is there any need for future grace periods?
478 * Interrupts must be disabled. If the caller does not hold the root
479 * rnp_node structure's ->lock, the results are advisory only.
481 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
483 struct rcu_node
*rnp
= rcu_get_root(rsp
);
484 int idx
= (ACCESS_ONCE(rnp
->completed
) + 1) & 0x1;
485 int *fp
= &rnp
->need_future_gp
[idx
];
487 return ACCESS_ONCE(*fp
);
491 * Does the current CPU require a not-yet-started grace period?
492 * The caller must have disabled interrupts to prevent races with
493 * normal callback registry.
496 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
500 if (rcu_gp_in_progress(rsp
))
501 return 0; /* No, a grace period is already in progress. */
502 if (rcu_future_needs_gp(rsp
))
503 return 1; /* Yes, a no-CBs CPU needs one. */
504 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
505 return 0; /* No, this is a no-CBs (or offline) CPU. */
506 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
507 return 1; /* Yes, this CPU has newly registered callbacks. */
508 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
509 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
510 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
511 rdp
->nxtcompleted
[i
]))
512 return 1; /* Yes, CBs for future grace period. */
513 return 0; /* No grace period needed. */
517 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
519 * If the new value of the ->dynticks_nesting counter now is zero,
520 * we really have entered idle, and must do the appropriate accounting.
521 * The caller must have disabled interrupts.
523 static void rcu_eqs_enter_common(long long oldval
, bool user
)
525 struct rcu_state
*rsp
;
526 struct rcu_data
*rdp
;
527 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
529 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
530 if (!user
&& !is_idle_task(current
)) {
531 struct task_struct
*idle __maybe_unused
=
532 idle_task(smp_processor_id());
534 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
535 ftrace_dump(DUMP_ORIG
);
536 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
537 current
->pid
, current
->comm
,
538 idle
->pid
, idle
->comm
); /* must be idle task! */
540 for_each_rcu_flavor(rsp
) {
541 rdp
= this_cpu_ptr(rsp
->rda
);
542 do_nocb_deferred_wakeup(rdp
);
544 rcu_prepare_for_idle();
545 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
546 smp_mb__before_atomic(); /* See above. */
547 atomic_inc(&rdtp
->dynticks
);
548 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
549 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
550 rcu_dynticks_task_enter();
553 * It is illegal to enter an extended quiescent state while
554 * in an RCU read-side critical section.
556 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
557 "Illegal idle entry in RCU read-side critical section.");
558 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
559 "Illegal idle entry in RCU-bh read-side critical section.");
560 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
561 "Illegal idle entry in RCU-sched read-side critical section.");
565 * Enter an RCU extended quiescent state, which can be either the
566 * idle loop or adaptive-tickless usermode execution.
568 static void rcu_eqs_enter(bool user
)
571 struct rcu_dynticks
*rdtp
;
573 rdtp
= this_cpu_ptr(&rcu_dynticks
);
574 oldval
= rdtp
->dynticks_nesting
;
575 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
576 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
577 rdtp
->dynticks_nesting
= 0;
578 rcu_eqs_enter_common(oldval
, user
);
580 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
585 * rcu_idle_enter - inform RCU that current CPU is entering idle
587 * Enter idle mode, in other words, -leave- the mode in which RCU
588 * read-side critical sections can occur. (Though RCU read-side
589 * critical sections can occur in irq handlers in idle, a possibility
590 * handled by irq_enter() and irq_exit().)
592 * We crowbar the ->dynticks_nesting field to zero to allow for
593 * the possibility of usermode upcalls having messed up our count
594 * of interrupt nesting level during the prior busy period.
596 void rcu_idle_enter(void)
600 local_irq_save(flags
);
601 rcu_eqs_enter(false);
602 rcu_sysidle_enter(0);
603 local_irq_restore(flags
);
605 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
607 #ifdef CONFIG_RCU_USER_QS
609 * rcu_user_enter - inform RCU that we are resuming userspace.
611 * Enter RCU idle mode right before resuming userspace. No use of RCU
612 * is permitted between this call and rcu_user_exit(). This way the
613 * CPU doesn't need to maintain the tick for RCU maintenance purposes
614 * when the CPU runs in userspace.
616 void rcu_user_enter(void)
620 #endif /* CONFIG_RCU_USER_QS */
623 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
625 * Exit from an interrupt handler, which might possibly result in entering
626 * idle mode, in other words, leaving the mode in which read-side critical
627 * sections can occur.
629 * This code assumes that the idle loop never does anything that might
630 * result in unbalanced calls to irq_enter() and irq_exit(). If your
631 * architecture violates this assumption, RCU will give you what you
632 * deserve, good and hard. But very infrequently and irreproducibly.
634 * Use things like work queues to work around this limitation.
636 * You have been warned.
638 void rcu_irq_exit(void)
642 struct rcu_dynticks
*rdtp
;
644 local_irq_save(flags
);
645 rdtp
= this_cpu_ptr(&rcu_dynticks
);
646 oldval
= rdtp
->dynticks_nesting
;
647 rdtp
->dynticks_nesting
--;
648 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
649 if (rdtp
->dynticks_nesting
)
650 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
652 rcu_eqs_enter_common(oldval
, true);
653 rcu_sysidle_enter(1);
654 local_irq_restore(flags
);
658 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
660 * If the new value of the ->dynticks_nesting counter was previously zero,
661 * we really have exited idle, and must do the appropriate accounting.
662 * The caller must have disabled interrupts.
664 static void rcu_eqs_exit_common(long long oldval
, int user
)
666 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
668 rcu_dynticks_task_exit();
669 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
670 atomic_inc(&rdtp
->dynticks
);
671 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
672 smp_mb__after_atomic(); /* See above. */
673 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
674 rcu_cleanup_after_idle();
675 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
676 if (!user
&& !is_idle_task(current
)) {
677 struct task_struct
*idle __maybe_unused
=
678 idle_task(smp_processor_id());
680 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
681 oldval
, rdtp
->dynticks_nesting
);
682 ftrace_dump(DUMP_ORIG
);
683 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
684 current
->pid
, current
->comm
,
685 idle
->pid
, idle
->comm
); /* must be idle task! */
690 * Exit an RCU extended quiescent state, which can be either the
691 * idle loop or adaptive-tickless usermode execution.
693 static void rcu_eqs_exit(bool user
)
695 struct rcu_dynticks
*rdtp
;
698 rdtp
= this_cpu_ptr(&rcu_dynticks
);
699 oldval
= rdtp
->dynticks_nesting
;
700 WARN_ON_ONCE(oldval
< 0);
701 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
702 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
704 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
705 rcu_eqs_exit_common(oldval
, user
);
710 * rcu_idle_exit - inform RCU that current CPU is leaving idle
712 * Exit idle mode, in other words, -enter- the mode in which RCU
713 * read-side critical sections can occur.
715 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
716 * allow for the possibility of usermode upcalls messing up our count
717 * of interrupt nesting level during the busy period that is just
720 void rcu_idle_exit(void)
724 local_irq_save(flags
);
727 local_irq_restore(flags
);
729 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
731 #ifdef CONFIG_RCU_USER_QS
733 * rcu_user_exit - inform RCU that we are exiting userspace.
735 * Exit RCU idle mode while entering the kernel because it can
736 * run a RCU read side critical section anytime.
738 void rcu_user_exit(void)
742 #endif /* CONFIG_RCU_USER_QS */
745 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
747 * Enter an interrupt handler, which might possibly result in exiting
748 * idle mode, in other words, entering the mode in which read-side critical
749 * sections can occur.
751 * Note that the Linux kernel is fully capable of entering an interrupt
752 * handler that it never exits, for example when doing upcalls to
753 * user mode! This code assumes that the idle loop never does upcalls to
754 * user mode. If your architecture does do upcalls from the idle loop (or
755 * does anything else that results in unbalanced calls to the irq_enter()
756 * and irq_exit() functions), RCU will give you what you deserve, good
757 * and hard. But very infrequently and irreproducibly.
759 * Use things like work queues to work around this limitation.
761 * You have been warned.
763 void rcu_irq_enter(void)
766 struct rcu_dynticks
*rdtp
;
769 local_irq_save(flags
);
770 rdtp
= this_cpu_ptr(&rcu_dynticks
);
771 oldval
= rdtp
->dynticks_nesting
;
772 rdtp
->dynticks_nesting
++;
773 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
775 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
777 rcu_eqs_exit_common(oldval
, true);
779 local_irq_restore(flags
);
783 * rcu_nmi_enter - inform RCU of entry to NMI context
785 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
786 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
787 * that the CPU is active. This implementation permits nested NMIs, as
788 * long as the nesting level does not overflow an int. (You will probably
789 * run out of stack space first.)
791 void rcu_nmi_enter(void)
793 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
796 /* Complain about underflow. */
797 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
800 * If idle from RCU viewpoint, atomically increment ->dynticks
801 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
802 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
803 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
804 * to be in the outermost NMI handler that interrupted an RCU-idle
805 * period (observation due to Andy Lutomirski).
807 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
808 smp_mb__before_atomic(); /* Force delay from prior write. */
809 atomic_inc(&rdtp
->dynticks
);
810 /* atomic_inc() before later RCU read-side crit sects */
811 smp_mb__after_atomic(); /* See above. */
812 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
815 rdtp
->dynticks_nmi_nesting
+= incby
;
820 * rcu_nmi_exit - inform RCU of exit from NMI context
822 * If we are returning from the outermost NMI handler that interrupted an
823 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
824 * to let the RCU grace-period handling know that the CPU is back to
827 void rcu_nmi_exit(void)
829 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
832 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
833 * (We are exiting an NMI handler, so RCU better be paying attention
836 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
837 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
840 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
841 * leave it in non-RCU-idle state.
843 if (rdtp
->dynticks_nmi_nesting
!= 1) {
844 rdtp
->dynticks_nmi_nesting
-= 2;
848 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
849 rdtp
->dynticks_nmi_nesting
= 0;
850 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
851 smp_mb__before_atomic(); /* See above. */
852 atomic_inc(&rdtp
->dynticks
);
853 smp_mb__after_atomic(); /* Force delay to next write. */
854 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
858 * __rcu_is_watching - are RCU read-side critical sections safe?
860 * Return true if RCU is watching the running CPU, which means that
861 * this CPU can safely enter RCU read-side critical sections. Unlike
862 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
863 * least disabled preemption.
865 bool notrace
__rcu_is_watching(void)
867 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
871 * rcu_is_watching - see if RCU thinks that the current CPU is idle
873 * If the current CPU is in its idle loop and is neither in an interrupt
874 * or NMI handler, return true.
876 bool notrace
rcu_is_watching(void)
881 ret
= __rcu_is_watching();
885 EXPORT_SYMBOL_GPL(rcu_is_watching
);
887 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
890 * Is the current CPU online? Disable preemption to avoid false positives
891 * that could otherwise happen due to the current CPU number being sampled,
892 * this task being preempted, its old CPU being taken offline, resuming
893 * on some other CPU, then determining that its old CPU is now offline.
894 * It is OK to use RCU on an offline processor during initial boot, hence
895 * the check for rcu_scheduler_fully_active. Note also that it is OK
896 * for a CPU coming online to use RCU for one jiffy prior to marking itself
897 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
898 * offline to continue to use RCU for one jiffy after marking itself
899 * offline in the cpu_online_mask. This leniency is necessary given the
900 * non-atomic nature of the online and offline processing, for example,
901 * the fact that a CPU enters the scheduler after completing the CPU_DYING
904 * This is also why RCU internally marks CPUs online during the
905 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
907 * Disable checking if in an NMI handler because we cannot safely report
908 * errors from NMI handlers anyway.
910 bool rcu_lockdep_current_cpu_online(void)
912 struct rcu_data
*rdp
;
913 struct rcu_node
*rnp
;
919 rdp
= this_cpu_ptr(&rcu_sched_data
);
921 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
922 !rcu_scheduler_fully_active
;
926 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
928 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
931 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
933 * If the current CPU is idle or running at a first-level (not nested)
934 * interrupt from idle, return true. The caller must have at least
935 * disabled preemption.
937 static int rcu_is_cpu_rrupt_from_idle(void)
939 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
943 * Snapshot the specified CPU's dynticks counter so that we can later
944 * credit them with an implicit quiescent state. Return 1 if this CPU
945 * is in dynticks idle mode, which is an extended quiescent state.
947 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
948 bool *isidle
, unsigned long *maxj
)
950 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
951 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
952 if ((rdp
->dynticks_snap
& 0x1) == 0) {
953 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
956 if (ULONG_CMP_LT(ACCESS_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
958 ACCESS_ONCE(rdp
->gpwrap
) = true;
964 * This function really isn't for public consumption, but RCU is special in
965 * that context switches can allow the state machine to make progress.
967 extern void resched_cpu(int cpu
);
970 * Return true if the specified CPU has passed through a quiescent
971 * state by virtue of being in or having passed through an dynticks
972 * idle state since the last call to dyntick_save_progress_counter()
973 * for this same CPU, or by virtue of having been offline.
975 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
976 bool *isidle
, unsigned long *maxj
)
982 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
983 snap
= (unsigned int)rdp
->dynticks_snap
;
986 * If the CPU passed through or entered a dynticks idle phase with
987 * no active irq/NMI handlers, then we can safely pretend that the CPU
988 * already acknowledged the request to pass through a quiescent
989 * state. Either way, that CPU cannot possibly be in an RCU
990 * read-side critical section that started before the beginning
991 * of the current RCU grace period.
993 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
994 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1000 * Check for the CPU being offline, but only if the grace period
1001 * is old enough. We don't need to worry about the CPU changing
1002 * state: If we see it offline even once, it has been through a
1005 * The reason for insisting that the grace period be at least
1006 * one jiffy old is that CPUs that are not quite online and that
1007 * have just gone offline can still execute RCU read-side critical
1010 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1011 return 0; /* Grace period is not old enough. */
1013 if (cpu_is_offline(rdp
->cpu
)) {
1014 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1020 * A CPU running for an extended time within the kernel can
1021 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1022 * even context-switching back and forth between a pair of
1023 * in-kernel CPU-bound tasks cannot advance grace periods.
1024 * So if the grace period is old enough, make the CPU pay attention.
1025 * Note that the unsynchronized assignments to the per-CPU
1026 * rcu_sched_qs_mask variable are safe. Yes, setting of
1027 * bits can be lost, but they will be set again on the next
1028 * force-quiescent-state pass. So lost bit sets do not result
1029 * in incorrect behavior, merely in a grace period lasting
1030 * a few jiffies longer than it might otherwise. Because
1031 * there are at most four threads involved, and because the
1032 * updates are only once every few jiffies, the probability of
1033 * lossage (and thus of slight grace-period extension) is
1036 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1037 * is set too high, we override with half of the RCU CPU stall
1040 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1041 if (ULONG_CMP_GE(jiffies
,
1042 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1043 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1044 if (!(ACCESS_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1045 ACCESS_ONCE(rdp
->cond_resched_completed
) =
1046 ACCESS_ONCE(rdp
->mynode
->completed
);
1047 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1048 ACCESS_ONCE(*rcrmp
) =
1049 ACCESS_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
;
1050 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1051 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1052 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1053 /* Time to beat on that CPU again! */
1054 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1055 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1062 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1064 unsigned long j
= jiffies
;
1068 smp_wmb(); /* Record start time before stall time. */
1069 j1
= rcu_jiffies_till_stall_check();
1070 ACCESS_ONCE(rsp
->jiffies_stall
) = j
+ j1
;
1071 rsp
->jiffies_resched
= j
+ j1
/ 2;
1072 rsp
->n_force_qs_gpstart
= ACCESS_ONCE(rsp
->n_force_qs
);
1076 * Dump stacks of all tasks running on stalled CPUs.
1078 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1081 unsigned long flags
;
1082 struct rcu_node
*rnp
;
1084 rcu_for_each_leaf_node(rsp
, rnp
) {
1085 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1086 if (rnp
->qsmask
!= 0) {
1087 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1088 if (rnp
->qsmask
& (1UL << cpu
))
1089 dump_cpu_task(rnp
->grplo
+ cpu
);
1091 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1095 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1099 unsigned long flags
;
1103 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1106 /* Only let one CPU complain about others per time interval. */
1108 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1109 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
1110 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1111 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1114 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
1115 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1118 * OK, time to rat on our buddy...
1119 * See Documentation/RCU/stallwarn.txt for info on how to debug
1120 * RCU CPU stall warnings.
1122 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1124 print_cpu_stall_info_begin();
1125 rcu_for_each_leaf_node(rsp
, rnp
) {
1126 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1127 ndetected
+= rcu_print_task_stall(rnp
);
1128 if (rnp
->qsmask
!= 0) {
1129 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1130 if (rnp
->qsmask
& (1UL << cpu
)) {
1131 print_cpu_stall_info(rsp
,
1136 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1140 * Now rat on any tasks that got kicked up to the root rcu_node
1141 * due to CPU offlining.
1143 rnp
= rcu_get_root(rsp
);
1144 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1145 ndetected
+= rcu_print_task_stall(rnp
);
1146 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1148 print_cpu_stall_info_end();
1149 for_each_possible_cpu(cpu
)
1150 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1151 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1152 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1153 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1155 rcu_dump_cpu_stacks(rsp
);
1157 if (ACCESS_ONCE(rsp
->gpnum
) != gpnum
||
1158 ACCESS_ONCE(rsp
->completed
) == gpnum
) {
1159 pr_err("INFO: Stall ended before state dump start\n");
1162 gpa
= ACCESS_ONCE(rsp
->gp_activity
);
1163 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld\n",
1164 rsp
->name
, j
- gpa
, j
, gpa
,
1165 jiffies_till_next_fqs
);
1166 /* In this case, the current CPU might be at fault. */
1167 sched_show_task(current
);
1171 /* Complain about tasks blocking the grace period. */
1173 rcu_print_detail_task_stall(rsp
);
1175 force_quiescent_state(rsp
); /* Kick them all. */
1178 static void print_cpu_stall(struct rcu_state
*rsp
)
1181 unsigned long flags
;
1182 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1186 * OK, time to rat on ourselves...
1187 * See Documentation/RCU/stallwarn.txt for info on how to debug
1188 * RCU CPU stall warnings.
1190 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1191 print_cpu_stall_info_begin();
1192 print_cpu_stall_info(rsp
, smp_processor_id());
1193 print_cpu_stall_info_end();
1194 for_each_possible_cpu(cpu
)
1195 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1196 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1197 jiffies
- rsp
->gp_start
,
1198 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1199 rcu_dump_cpu_stacks(rsp
);
1201 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1202 if (ULONG_CMP_GE(jiffies
, ACCESS_ONCE(rsp
->jiffies_stall
)))
1203 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+
1204 3 * rcu_jiffies_till_stall_check() + 3;
1205 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1208 * Attempt to revive the RCU machinery by forcing a context switch.
1210 * A context switch would normally allow the RCU state machine to make
1211 * progress and it could be we're stuck in kernel space without context
1212 * switches for an entirely unreasonable amount of time.
1214 resched_cpu(smp_processor_id());
1217 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1219 unsigned long completed
;
1220 unsigned long gpnum
;
1224 struct rcu_node
*rnp
;
1226 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1231 * Lots of memory barriers to reject false positives.
1233 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1234 * then rsp->gp_start, and finally rsp->completed. These values
1235 * are updated in the opposite order with memory barriers (or
1236 * equivalent) during grace-period initialization and cleanup.
1237 * Now, a false positive can occur if we get an new value of
1238 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1239 * the memory barriers, the only way that this can happen is if one
1240 * grace period ends and another starts between these two fetches.
1241 * Detect this by comparing rsp->completed with the previous fetch
1244 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1245 * and rsp->gp_start suffice to forestall false positives.
1247 gpnum
= ACCESS_ONCE(rsp
->gpnum
);
1248 smp_rmb(); /* Pick up ->gpnum first... */
1249 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
1250 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1251 gps
= ACCESS_ONCE(rsp
->gp_start
);
1252 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1253 completed
= ACCESS_ONCE(rsp
->completed
);
1254 if (ULONG_CMP_GE(completed
, gpnum
) ||
1255 ULONG_CMP_LT(j
, js
) ||
1256 ULONG_CMP_GE(gps
, js
))
1257 return; /* No stall or GP completed since entering function. */
1259 if (rcu_gp_in_progress(rsp
) &&
1260 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1262 /* We haven't checked in, so go dump stack. */
1263 print_cpu_stall(rsp
);
1265 } else if (rcu_gp_in_progress(rsp
) &&
1266 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1268 /* They had a few time units to dump stack, so complain. */
1269 print_other_cpu_stall(rsp
, gpnum
);
1274 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1276 * Set the stall-warning timeout way off into the future, thus preventing
1277 * any RCU CPU stall-warning messages from appearing in the current set of
1278 * RCU grace periods.
1280 * The caller must disable hard irqs.
1282 void rcu_cpu_stall_reset(void)
1284 struct rcu_state
*rsp
;
1286 for_each_rcu_flavor(rsp
)
1287 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ ULONG_MAX
/ 2;
1291 * Initialize the specified rcu_data structure's callback list to empty.
1293 static void init_callback_list(struct rcu_data
*rdp
)
1297 if (init_nocb_callback_list(rdp
))
1299 rdp
->nxtlist
= NULL
;
1300 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1301 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1305 * Determine the value that ->completed will have at the end of the
1306 * next subsequent grace period. This is used to tag callbacks so that
1307 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1308 * been dyntick-idle for an extended period with callbacks under the
1309 * influence of RCU_FAST_NO_HZ.
1311 * The caller must hold rnp->lock with interrupts disabled.
1313 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1314 struct rcu_node
*rnp
)
1317 * If RCU is idle, we just wait for the next grace period.
1318 * But we can only be sure that RCU is idle if we are looking
1319 * at the root rcu_node structure -- otherwise, a new grace
1320 * period might have started, but just not yet gotten around
1321 * to initializing the current non-root rcu_node structure.
1323 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1324 return rnp
->completed
+ 1;
1327 * Otherwise, wait for a possible partial grace period and
1328 * then the subsequent full grace period.
1330 return rnp
->completed
+ 2;
1334 * Trace-event helper function for rcu_start_future_gp() and
1335 * rcu_nocb_wait_gp().
1337 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1338 unsigned long c
, const char *s
)
1340 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1341 rnp
->completed
, c
, rnp
->level
,
1342 rnp
->grplo
, rnp
->grphi
, s
);
1346 * Start some future grace period, as needed to handle newly arrived
1347 * callbacks. The required future grace periods are recorded in each
1348 * rcu_node structure's ->need_future_gp field. Returns true if there
1349 * is reason to awaken the grace-period kthread.
1351 * The caller must hold the specified rcu_node structure's ->lock.
1353 static bool __maybe_unused
1354 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1355 unsigned long *c_out
)
1360 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1363 * Pick up grace-period number for new callbacks. If this
1364 * grace period is already marked as needed, return to the caller.
1366 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1367 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1368 if (rnp
->need_future_gp
[c
& 0x1]) {
1369 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1374 * If either this rcu_node structure or the root rcu_node structure
1375 * believe that a grace period is in progress, then we must wait
1376 * for the one following, which is in "c". Because our request
1377 * will be noticed at the end of the current grace period, we don't
1378 * need to explicitly start one. We only do the lockless check
1379 * of rnp_root's fields if the current rcu_node structure thinks
1380 * there is no grace period in flight, and because we hold rnp->lock,
1381 * the only possible change is when rnp_root's two fields are
1382 * equal, in which case rnp_root->gpnum might be concurrently
1383 * incremented. But that is OK, as it will just result in our
1384 * doing some extra useless work.
1386 if (rnp
->gpnum
!= rnp
->completed
||
1387 ACCESS_ONCE(rnp_root
->gpnum
) != ACCESS_ONCE(rnp_root
->completed
)) {
1388 rnp
->need_future_gp
[c
& 0x1]++;
1389 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1394 * There might be no grace period in progress. If we don't already
1395 * hold it, acquire the root rcu_node structure's lock in order to
1396 * start one (if needed).
1398 if (rnp
!= rnp_root
) {
1399 raw_spin_lock(&rnp_root
->lock
);
1400 smp_mb__after_unlock_lock();
1404 * Get a new grace-period number. If there really is no grace
1405 * period in progress, it will be smaller than the one we obtained
1406 * earlier. Adjust callbacks as needed. Note that even no-CBs
1407 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1409 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1410 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1411 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1412 rdp
->nxtcompleted
[i
] = c
;
1415 * If the needed for the required grace period is already
1416 * recorded, trace and leave.
1418 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1419 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1423 /* Record the need for the future grace period. */
1424 rnp_root
->need_future_gp
[c
& 0x1]++;
1426 /* If a grace period is not already in progress, start one. */
1427 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1428 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1430 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1431 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1434 if (rnp
!= rnp_root
)
1435 raw_spin_unlock(&rnp_root
->lock
);
1443 * Clean up any old requests for the just-ended grace period. Also return
1444 * whether any additional grace periods have been requested. Also invoke
1445 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1446 * waiting for this grace period to complete.
1448 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1450 int c
= rnp
->completed
;
1452 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1454 rcu_nocb_gp_cleanup(rsp
, rnp
);
1455 rnp
->need_future_gp
[c
& 0x1] = 0;
1456 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1457 trace_rcu_future_gp(rnp
, rdp
, c
,
1458 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1463 * Awaken the grace-period kthread for the specified flavor of RCU.
1464 * Don't do a self-awaken, and don't bother awakening when there is
1465 * nothing for the grace-period kthread to do (as in several CPUs
1466 * raced to awaken, and we lost), and finally don't try to awaken
1467 * a kthread that has not yet been created.
1469 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1471 if (current
== rsp
->gp_kthread
||
1472 !ACCESS_ONCE(rsp
->gp_flags
) ||
1475 wake_up(&rsp
->gp_wq
);
1479 * If there is room, assign a ->completed number to any callbacks on
1480 * this CPU that have not already been assigned. Also accelerate any
1481 * callbacks that were previously assigned a ->completed number that has
1482 * since proven to be too conservative, which can happen if callbacks get
1483 * assigned a ->completed number while RCU is idle, but with reference to
1484 * a non-root rcu_node structure. This function is idempotent, so it does
1485 * not hurt to call it repeatedly. Returns an flag saying that we should
1486 * awaken the RCU grace-period kthread.
1488 * The caller must hold rnp->lock with interrupts disabled.
1490 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1491 struct rcu_data
*rdp
)
1497 /* If the CPU has no callbacks, nothing to do. */
1498 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1502 * Starting from the sublist containing the callbacks most
1503 * recently assigned a ->completed number and working down, find the
1504 * first sublist that is not assignable to an upcoming grace period.
1505 * Such a sublist has something in it (first two tests) and has
1506 * a ->completed number assigned that will complete sooner than
1507 * the ->completed number for newly arrived callbacks (last test).
1509 * The key point is that any later sublist can be assigned the
1510 * same ->completed number as the newly arrived callbacks, which
1511 * means that the callbacks in any of these later sublist can be
1512 * grouped into a single sublist, whether or not they have already
1513 * been assigned a ->completed number.
1515 c
= rcu_cbs_completed(rsp
, rnp
);
1516 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1517 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1518 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1522 * If there are no sublist for unassigned callbacks, leave.
1523 * At the same time, advance "i" one sublist, so that "i" will
1524 * index into the sublist where all the remaining callbacks should
1527 if (++i
>= RCU_NEXT_TAIL
)
1531 * Assign all subsequent callbacks' ->completed number to the next
1532 * full grace period and group them all in the sublist initially
1535 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1536 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1537 rdp
->nxtcompleted
[i
] = c
;
1539 /* Record any needed additional grace periods. */
1540 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1542 /* Trace depending on how much we were able to accelerate. */
1543 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1544 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1546 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1551 * Move any callbacks whose grace period has completed to the
1552 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1553 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1554 * sublist. This function is idempotent, so it does not hurt to
1555 * invoke it repeatedly. As long as it is not invoked -too- often...
1556 * Returns true if the RCU grace-period kthread needs to be awakened.
1558 * The caller must hold rnp->lock with interrupts disabled.
1560 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1561 struct rcu_data
*rdp
)
1565 /* If the CPU has no callbacks, nothing to do. */
1566 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1570 * Find all callbacks whose ->completed numbers indicate that they
1571 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1573 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1574 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1576 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1578 /* Clean up any sublist tail pointers that were misordered above. */
1579 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1580 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1582 /* Copy down callbacks to fill in empty sublists. */
1583 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1584 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1586 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1587 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1590 /* Classify any remaining callbacks. */
1591 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1595 * Update CPU-local rcu_data state to record the beginnings and ends of
1596 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1597 * structure corresponding to the current CPU, and must have irqs disabled.
1598 * Returns true if the grace-period kthread needs to be awakened.
1600 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1601 struct rcu_data
*rdp
)
1605 /* Handle the ends of any preceding grace periods first. */
1606 if (rdp
->completed
== rnp
->completed
&&
1607 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1609 /* No grace period end, so just accelerate recent callbacks. */
1610 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1614 /* Advance callbacks. */
1615 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1617 /* Remember that we saw this grace-period completion. */
1618 rdp
->completed
= rnp
->completed
;
1619 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1622 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1624 * If the current grace period is waiting for this CPU,
1625 * set up to detect a quiescent state, otherwise don't
1626 * go looking for one.
1628 rdp
->gpnum
= rnp
->gpnum
;
1629 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1630 rdp
->passed_quiesce
= 0;
1631 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1632 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1633 zero_cpu_stall_ticks(rdp
);
1634 ACCESS_ONCE(rdp
->gpwrap
) = false;
1639 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1641 unsigned long flags
;
1643 struct rcu_node
*rnp
;
1645 local_irq_save(flags
);
1647 if ((rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) &&
1648 rdp
->completed
== ACCESS_ONCE(rnp
->completed
) &&
1649 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1650 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1651 local_irq_restore(flags
);
1654 smp_mb__after_unlock_lock();
1655 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1656 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1658 rcu_gp_kthread_wake(rsp
);
1662 * Initialize a new grace period. Return 0 if no grace period required.
1664 static int rcu_gp_init(struct rcu_state
*rsp
)
1666 struct rcu_data
*rdp
;
1667 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1669 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1670 rcu_bind_gp_kthread();
1671 raw_spin_lock_irq(&rnp
->lock
);
1672 smp_mb__after_unlock_lock();
1673 if (!ACCESS_ONCE(rsp
->gp_flags
)) {
1674 /* Spurious wakeup, tell caller to go back to sleep. */
1675 raw_spin_unlock_irq(&rnp
->lock
);
1678 ACCESS_ONCE(rsp
->gp_flags
) = 0; /* Clear all flags: New grace period. */
1680 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1682 * Grace period already in progress, don't start another.
1683 * Not supposed to be able to happen.
1685 raw_spin_unlock_irq(&rnp
->lock
);
1689 /* Advance to a new grace period and initialize state. */
1690 record_gp_stall_check_time(rsp
);
1691 /* Record GP times before starting GP, hence smp_store_release(). */
1692 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1693 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1694 raw_spin_unlock_irq(&rnp
->lock
);
1696 /* Exclude any concurrent CPU-hotplug operations. */
1697 mutex_lock(&rsp
->onoff_mutex
);
1698 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1701 * Set the quiescent-state-needed bits in all the rcu_node
1702 * structures for all currently online CPUs in breadth-first order,
1703 * starting from the root rcu_node structure, relying on the layout
1704 * of the tree within the rsp->node[] array. Note that other CPUs
1705 * will access only the leaves of the hierarchy, thus seeing that no
1706 * grace period is in progress, at least until the corresponding
1707 * leaf node has been initialized. In addition, we have excluded
1708 * CPU-hotplug operations.
1710 * The grace period cannot complete until the initialization
1711 * process finishes, because this kthread handles both.
1713 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1714 raw_spin_lock_irq(&rnp
->lock
);
1715 smp_mb__after_unlock_lock();
1716 rdp
= this_cpu_ptr(rsp
->rda
);
1717 rcu_preempt_check_blocked_tasks(rnp
);
1718 rnp
->qsmask
= rnp
->qsmaskinit
;
1719 ACCESS_ONCE(rnp
->gpnum
) = rsp
->gpnum
;
1720 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1721 ACCESS_ONCE(rnp
->completed
) = rsp
->completed
;
1722 if (rnp
== rdp
->mynode
)
1723 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1724 rcu_preempt_boost_start_gp(rnp
);
1725 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1726 rnp
->level
, rnp
->grplo
,
1727 rnp
->grphi
, rnp
->qsmask
);
1728 raw_spin_unlock_irq(&rnp
->lock
);
1729 cond_resched_rcu_qs();
1730 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1733 mutex_unlock(&rsp
->onoff_mutex
);
1738 * Do one round of quiescent-state forcing.
1740 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1742 int fqs_state
= fqs_state_in
;
1743 bool isidle
= false;
1745 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1747 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1749 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1750 /* Collect dyntick-idle snapshots. */
1751 if (is_sysidle_rcu_state(rsp
)) {
1753 maxj
= jiffies
- ULONG_MAX
/ 4;
1755 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1757 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1758 fqs_state
= RCU_FORCE_QS
;
1760 /* Handle dyntick-idle and offline CPUs. */
1762 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1764 /* Clear flag to prevent immediate re-entry. */
1765 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1766 raw_spin_lock_irq(&rnp
->lock
);
1767 smp_mb__after_unlock_lock();
1768 ACCESS_ONCE(rsp
->gp_flags
) =
1769 ACCESS_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
;
1770 raw_spin_unlock_irq(&rnp
->lock
);
1776 * Clean up after the old grace period.
1778 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1780 unsigned long gp_duration
;
1781 bool needgp
= false;
1783 struct rcu_data
*rdp
;
1784 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1786 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1787 raw_spin_lock_irq(&rnp
->lock
);
1788 smp_mb__after_unlock_lock();
1789 gp_duration
= jiffies
- rsp
->gp_start
;
1790 if (gp_duration
> rsp
->gp_max
)
1791 rsp
->gp_max
= gp_duration
;
1794 * We know the grace period is complete, but to everyone else
1795 * it appears to still be ongoing. But it is also the case
1796 * that to everyone else it looks like there is nothing that
1797 * they can do to advance the grace period. It is therefore
1798 * safe for us to drop the lock in order to mark the grace
1799 * period as completed in all of the rcu_node structures.
1801 raw_spin_unlock_irq(&rnp
->lock
);
1804 * Propagate new ->completed value to rcu_node structures so
1805 * that other CPUs don't have to wait until the start of the next
1806 * grace period to process their callbacks. This also avoids
1807 * some nasty RCU grace-period initialization races by forcing
1808 * the end of the current grace period to be completely recorded in
1809 * all of the rcu_node structures before the beginning of the next
1810 * grace period is recorded in any of the rcu_node structures.
1812 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1813 raw_spin_lock_irq(&rnp
->lock
);
1814 smp_mb__after_unlock_lock();
1815 ACCESS_ONCE(rnp
->completed
) = rsp
->gpnum
;
1816 rdp
= this_cpu_ptr(rsp
->rda
);
1817 if (rnp
== rdp
->mynode
)
1818 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1819 /* smp_mb() provided by prior unlock-lock pair. */
1820 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1821 raw_spin_unlock_irq(&rnp
->lock
);
1822 cond_resched_rcu_qs();
1823 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1825 rnp
= rcu_get_root(rsp
);
1826 raw_spin_lock_irq(&rnp
->lock
);
1827 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1828 rcu_nocb_gp_set(rnp
, nocb
);
1830 /* Declare grace period done. */
1831 ACCESS_ONCE(rsp
->completed
) = rsp
->gpnum
;
1832 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1833 rsp
->fqs_state
= RCU_GP_IDLE
;
1834 rdp
= this_cpu_ptr(rsp
->rda
);
1835 /* Advance CBs to reduce false positives below. */
1836 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1837 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1838 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
1839 trace_rcu_grace_period(rsp
->name
,
1840 ACCESS_ONCE(rsp
->gpnum
),
1843 raw_spin_unlock_irq(&rnp
->lock
);
1847 * Body of kthread that handles grace periods.
1849 static int __noreturn
rcu_gp_kthread(void *arg
)
1855 struct rcu_state
*rsp
= arg
;
1856 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1860 /* Handle grace-period start. */
1862 trace_rcu_grace_period(rsp
->name
,
1863 ACCESS_ONCE(rsp
->gpnum
),
1865 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
1866 wait_event_interruptible(rsp
->gp_wq
,
1867 ACCESS_ONCE(rsp
->gp_flags
) &
1869 /* Locking provides needed memory barrier. */
1870 if (rcu_gp_init(rsp
))
1872 cond_resched_rcu_qs();
1873 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1874 WARN_ON(signal_pending(current
));
1875 trace_rcu_grace_period(rsp
->name
,
1876 ACCESS_ONCE(rsp
->gpnum
),
1880 /* Handle quiescent-state forcing. */
1881 fqs_state
= RCU_SAVE_DYNTICK
;
1882 j
= jiffies_till_first_fqs
;
1885 jiffies_till_first_fqs
= HZ
;
1890 rsp
->jiffies_force_qs
= jiffies
+ j
;
1891 trace_rcu_grace_period(rsp
->name
,
1892 ACCESS_ONCE(rsp
->gpnum
),
1894 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
1895 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1896 ((gf
= ACCESS_ONCE(rsp
->gp_flags
)) &
1898 (!ACCESS_ONCE(rnp
->qsmask
) &&
1899 !rcu_preempt_blocked_readers_cgp(rnp
)),
1901 /* Locking provides needed memory barriers. */
1902 /* If grace period done, leave loop. */
1903 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1904 !rcu_preempt_blocked_readers_cgp(rnp
))
1906 /* If time for quiescent-state forcing, do it. */
1907 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
1908 (gf
& RCU_GP_FLAG_FQS
)) {
1909 trace_rcu_grace_period(rsp
->name
,
1910 ACCESS_ONCE(rsp
->gpnum
),
1912 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1913 trace_rcu_grace_period(rsp
->name
,
1914 ACCESS_ONCE(rsp
->gpnum
),
1916 cond_resched_rcu_qs();
1917 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1919 /* Deal with stray signal. */
1920 cond_resched_rcu_qs();
1921 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1922 WARN_ON(signal_pending(current
));
1923 trace_rcu_grace_period(rsp
->name
,
1924 ACCESS_ONCE(rsp
->gpnum
),
1927 j
= jiffies_till_next_fqs
;
1930 jiffies_till_next_fqs
= HZ
;
1933 jiffies_till_next_fqs
= 1;
1937 /* Handle grace-period end. */
1938 rcu_gp_cleanup(rsp
);
1943 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1944 * in preparation for detecting the next grace period. The caller must hold
1945 * the root node's ->lock and hard irqs must be disabled.
1947 * Note that it is legal for a dying CPU (which is marked as offline) to
1948 * invoke this function. This can happen when the dying CPU reports its
1951 * Returns true if the grace-period kthread must be awakened.
1954 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1955 struct rcu_data
*rdp
)
1957 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
1959 * Either we have not yet spawned the grace-period
1960 * task, this CPU does not need another grace period,
1961 * or a grace period is already in progress.
1962 * Either way, don't start a new grace period.
1966 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
1967 trace_rcu_grace_period(rsp
->name
, ACCESS_ONCE(rsp
->gpnum
),
1971 * We can't do wakeups while holding the rnp->lock, as that
1972 * could cause possible deadlocks with the rq->lock. Defer
1973 * the wakeup to our caller.
1979 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1980 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1981 * is invoked indirectly from rcu_advance_cbs(), which would result in
1982 * endless recursion -- or would do so if it wasn't for the self-deadlock
1983 * that is encountered beforehand.
1985 * Returns true if the grace-period kthread needs to be awakened.
1987 static bool rcu_start_gp(struct rcu_state
*rsp
)
1989 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1990 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1994 * If there is no grace period in progress right now, any
1995 * callbacks we have up to this point will be satisfied by the
1996 * next grace period. Also, advancing the callbacks reduces the
1997 * probability of false positives from cpu_needs_another_gp()
1998 * resulting in pointless grace periods. So, advance callbacks
1999 * then start the grace period!
2001 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2002 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2007 * Report a full set of quiescent states to the specified rcu_state
2008 * data structure. This involves cleaning up after the prior grace
2009 * period and letting rcu_start_gp() start up the next grace period
2010 * if one is needed. Note that the caller must hold rnp->lock, which
2011 * is released before return.
2013 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2014 __releases(rcu_get_root(rsp
)->lock
)
2016 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2017 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2018 rcu_gp_kthread_wake(rsp
);
2022 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2023 * Allows quiescent states for a group of CPUs to be reported at one go
2024 * to the specified rcu_node structure, though all the CPUs in the group
2025 * must be represented by the same rcu_node structure (which need not be
2026 * a leaf rcu_node structure, though it often will be). That structure's
2027 * lock must be held upon entry, and it is released before return.
2030 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2031 struct rcu_node
*rnp
, unsigned long flags
)
2032 __releases(rnp
->lock
)
2034 struct rcu_node
*rnp_c
;
2036 /* Walk up the rcu_node hierarchy. */
2038 if (!(rnp
->qsmask
& mask
)) {
2040 /* Our bit has already been cleared, so done. */
2041 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2044 rnp
->qsmask
&= ~mask
;
2045 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2046 mask
, rnp
->qsmask
, rnp
->level
,
2047 rnp
->grplo
, rnp
->grphi
,
2049 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2051 /* Other bits still set at this level, so done. */
2052 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2055 mask
= rnp
->grpmask
;
2056 if (rnp
->parent
== NULL
) {
2058 /* No more levels. Exit loop holding root lock. */
2062 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2065 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2066 smp_mb__after_unlock_lock();
2067 WARN_ON_ONCE(rnp_c
->qsmask
);
2071 * Get here if we are the last CPU to pass through a quiescent
2072 * state for this grace period. Invoke rcu_report_qs_rsp()
2073 * to clean up and start the next grace period if one is needed.
2075 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2079 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2080 * structure. This must be either called from the specified CPU, or
2081 * called when the specified CPU is known to be offline (and when it is
2082 * also known that no other CPU is concurrently trying to help the offline
2083 * CPU). The lastcomp argument is used to make sure we are still in the
2084 * grace period of interest. We don't want to end the current grace period
2085 * based on quiescent states detected in an earlier grace period!
2088 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2090 unsigned long flags
;
2093 struct rcu_node
*rnp
;
2096 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2097 smp_mb__after_unlock_lock();
2098 if ((rdp
->passed_quiesce
== 0 &&
2099 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2100 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2104 * The grace period in which this quiescent state was
2105 * recorded has ended, so don't report it upwards.
2106 * We will instead need a new quiescent state that lies
2107 * within the current grace period.
2109 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2110 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2111 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2114 mask
= rdp
->grpmask
;
2115 if ((rnp
->qsmask
& mask
) == 0) {
2116 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2118 rdp
->qs_pending
= 0;
2121 * This GP can't end until cpu checks in, so all of our
2122 * callbacks can be processed during the next GP.
2124 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2126 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
2128 rcu_gp_kthread_wake(rsp
);
2133 * Check to see if there is a new grace period of which this CPU
2134 * is not yet aware, and if so, set up local rcu_data state for it.
2135 * Otherwise, see if this CPU has just passed through its first
2136 * quiescent state for this grace period, and record that fact if so.
2139 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2141 /* Check for grace-period ends and beginnings. */
2142 note_gp_changes(rsp
, rdp
);
2145 * Does this CPU still need to do its part for current grace period?
2146 * If no, return and let the other CPUs do their part as well.
2148 if (!rdp
->qs_pending
)
2152 * Was there a quiescent state since the beginning of the grace
2153 * period? If no, then exit and wait for the next call.
2155 if (!rdp
->passed_quiesce
&&
2156 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2160 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2163 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2166 #ifdef CONFIG_HOTPLUG_CPU
2169 * Send the specified CPU's RCU callbacks to the orphanage. The
2170 * specified CPU must be offline, and the caller must hold the
2174 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2175 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2177 /* No-CBs CPUs do not have orphanable callbacks. */
2178 if (rcu_is_nocb_cpu(rdp
->cpu
))
2182 * Orphan the callbacks. First adjust the counts. This is safe
2183 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2184 * cannot be running now. Thus no memory barrier is required.
2186 if (rdp
->nxtlist
!= NULL
) {
2187 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2188 rsp
->qlen
+= rdp
->qlen
;
2189 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2191 ACCESS_ONCE(rdp
->qlen
) = 0;
2195 * Next, move those callbacks still needing a grace period to
2196 * the orphanage, where some other CPU will pick them up.
2197 * Some of the callbacks might have gone partway through a grace
2198 * period, but that is too bad. They get to start over because we
2199 * cannot assume that grace periods are synchronized across CPUs.
2200 * We don't bother updating the ->nxttail[] array yet, instead
2201 * we just reset the whole thing later on.
2203 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2204 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2205 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2206 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2210 * Then move the ready-to-invoke callbacks to the orphanage,
2211 * where some other CPU will pick them up. These will not be
2212 * required to pass though another grace period: They are done.
2214 if (rdp
->nxtlist
!= NULL
) {
2215 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2216 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2219 /* Finally, initialize the rcu_data structure's list to empty. */
2220 init_callback_list(rdp
);
2224 * Adopt the RCU callbacks from the specified rcu_state structure's
2225 * orphanage. The caller must hold the ->orphan_lock.
2227 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2230 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2232 /* No-CBs CPUs are handled specially. */
2233 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2236 /* Do the accounting first. */
2237 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2238 rdp
->qlen
+= rsp
->qlen
;
2239 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2240 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2241 rcu_idle_count_callbacks_posted();
2246 * We do not need a memory barrier here because the only way we
2247 * can get here if there is an rcu_barrier() in flight is if
2248 * we are the task doing the rcu_barrier().
2251 /* First adopt the ready-to-invoke callbacks. */
2252 if (rsp
->orphan_donelist
!= NULL
) {
2253 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2254 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2255 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2256 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2257 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2258 rsp
->orphan_donelist
= NULL
;
2259 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2262 /* And then adopt the callbacks that still need a grace period. */
2263 if (rsp
->orphan_nxtlist
!= NULL
) {
2264 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2265 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2266 rsp
->orphan_nxtlist
= NULL
;
2267 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2272 * Trace the fact that this CPU is going offline.
2274 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2276 RCU_TRACE(unsigned long mask
);
2277 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2278 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2280 RCU_TRACE(mask
= rdp
->grpmask
);
2281 trace_rcu_grace_period(rsp
->name
,
2282 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2287 * The CPU has been completely removed, and some other CPU is reporting
2288 * this fact from process context. Do the remainder of the cleanup,
2289 * including orphaning the outgoing CPU's RCU callbacks, and also
2290 * adopting them. There can only be one CPU hotplug operation at a time,
2291 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2293 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2295 unsigned long flags
;
2297 int need_report
= 0;
2298 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2299 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2301 /* Adjust any no-longer-needed kthreads. */
2302 rcu_boost_kthread_setaffinity(rnp
, -1);
2304 /* Exclude any attempts to start a new grace period. */
2305 mutex_lock(&rsp
->onoff_mutex
);
2306 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2308 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2309 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2310 rcu_adopt_orphan_cbs(rsp
, flags
);
2312 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2313 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
2315 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2316 smp_mb__after_unlock_lock();
2317 rnp
->qsmaskinit
&= ~mask
;
2318 if (rnp
->qsmaskinit
!= 0) {
2319 if (rnp
!= rdp
->mynode
)
2320 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2323 if (rnp
== rdp
->mynode
)
2324 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
2326 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2327 mask
= rnp
->grpmask
;
2329 } while (rnp
!= NULL
);
2332 * We still hold the leaf rcu_node structure lock here, and
2333 * irqs are still disabled. The reason for this subterfuge is
2334 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2335 * held leads to deadlock.
2337 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
2339 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
2340 rcu_report_unblock_qs_rnp(rnp
, flags
);
2342 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2343 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
2344 rcu_report_exp_rnp(rsp
, rnp
, true);
2345 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2346 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2347 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2348 init_callback_list(rdp
);
2349 /* Disallow further callbacks on this CPU. */
2350 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2351 mutex_unlock(&rsp
->onoff_mutex
);
2354 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2356 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2360 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2364 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2367 * Invoke any RCU callbacks that have made it to the end of their grace
2368 * period. Thottle as specified by rdp->blimit.
2370 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2372 unsigned long flags
;
2373 struct rcu_head
*next
, *list
, **tail
;
2374 long bl
, count
, count_lazy
;
2377 /* If no callbacks are ready, just return. */
2378 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2379 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2380 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
2381 need_resched(), is_idle_task(current
),
2382 rcu_is_callbacks_kthread());
2387 * Extract the list of ready callbacks, disabling to prevent
2388 * races with call_rcu() from interrupt handlers.
2390 local_irq_save(flags
);
2391 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2393 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2394 list
= rdp
->nxtlist
;
2395 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2396 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2397 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2398 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2399 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2400 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2401 local_irq_restore(flags
);
2403 /* Invoke callbacks. */
2404 count
= count_lazy
= 0;
2408 debug_rcu_head_unqueue(list
);
2409 if (__rcu_reclaim(rsp
->name
, list
))
2412 /* Stop only if limit reached and CPU has something to do. */
2413 if (++count
>= bl
&&
2415 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2419 local_irq_save(flags
);
2420 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2421 is_idle_task(current
),
2422 rcu_is_callbacks_kthread());
2424 /* Update count, and requeue any remaining callbacks. */
2426 *tail
= rdp
->nxtlist
;
2427 rdp
->nxtlist
= list
;
2428 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2429 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2430 rdp
->nxttail
[i
] = tail
;
2434 smp_mb(); /* List handling before counting for rcu_barrier(). */
2435 rdp
->qlen_lazy
-= count_lazy
;
2436 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
- count
;
2437 rdp
->n_cbs_invoked
+= count
;
2439 /* Reinstate batch limit if we have worked down the excess. */
2440 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2441 rdp
->blimit
= blimit
;
2443 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2444 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2445 rdp
->qlen_last_fqs_check
= 0;
2446 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2447 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2448 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2449 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2451 local_irq_restore(flags
);
2453 /* Re-invoke RCU core processing if there are callbacks remaining. */
2454 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2459 * Check to see if this CPU is in a non-context-switch quiescent state
2460 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2461 * Also schedule RCU core processing.
2463 * This function must be called from hardirq context. It is normally
2464 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2465 * false, there is no point in invoking rcu_check_callbacks().
2467 void rcu_check_callbacks(int user
)
2469 trace_rcu_utilization(TPS("Start scheduler-tick"));
2470 increment_cpu_stall_ticks();
2471 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2474 * Get here if this CPU took its interrupt from user
2475 * mode or from the idle loop, and if this is not a
2476 * nested interrupt. In this case, the CPU is in
2477 * a quiescent state, so note it.
2479 * No memory barrier is required here because both
2480 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2481 * variables that other CPUs neither access nor modify,
2482 * at least not while the corresponding CPU is online.
2488 } else if (!in_softirq()) {
2491 * Get here if this CPU did not take its interrupt from
2492 * softirq, in other words, if it is not interrupting
2493 * a rcu_bh read-side critical section. This is an _bh
2494 * critical section, so note it.
2499 rcu_preempt_check_callbacks();
2503 rcu_note_voluntary_context_switch(current
);
2504 trace_rcu_utilization(TPS("End scheduler-tick"));
2508 * Scan the leaf rcu_node structures, processing dyntick state for any that
2509 * have not yet encountered a quiescent state, using the function specified.
2510 * Also initiate boosting for any threads blocked on the root rcu_node.
2512 * The caller must have suppressed start of new grace periods.
2514 static void force_qs_rnp(struct rcu_state
*rsp
,
2515 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2516 unsigned long *maxj
),
2517 bool *isidle
, unsigned long *maxj
)
2521 unsigned long flags
;
2523 struct rcu_node
*rnp
;
2525 rcu_for_each_leaf_node(rsp
, rnp
) {
2526 cond_resched_rcu_qs();
2528 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2529 smp_mb__after_unlock_lock();
2530 if (!rcu_gp_in_progress(rsp
)) {
2531 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2534 if (rnp
->qsmask
== 0) {
2535 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
2540 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2541 if ((rnp
->qsmask
& bit
) != 0) {
2542 if ((rnp
->qsmaskinit
& bit
) != 0)
2544 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2550 /* rcu_report_qs_rnp() releases rnp->lock. */
2551 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
2554 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2556 rnp
= rcu_get_root(rsp
);
2557 if (rnp
->qsmask
== 0) {
2558 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2559 smp_mb__after_unlock_lock();
2560 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
2565 * Force quiescent states on reluctant CPUs, and also detect which
2566 * CPUs are in dyntick-idle mode.
2568 static void force_quiescent_state(struct rcu_state
*rsp
)
2570 unsigned long flags
;
2572 struct rcu_node
*rnp
;
2573 struct rcu_node
*rnp_old
= NULL
;
2575 /* Funnel through hierarchy to reduce memory contention. */
2576 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2577 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2578 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2579 !raw_spin_trylock(&rnp
->fqslock
);
2580 if (rnp_old
!= NULL
)
2581 raw_spin_unlock(&rnp_old
->fqslock
);
2583 rsp
->n_force_qs_lh
++;
2588 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2590 /* Reached the root of the rcu_node tree, acquire lock. */
2591 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2592 smp_mb__after_unlock_lock();
2593 raw_spin_unlock(&rnp_old
->fqslock
);
2594 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2595 rsp
->n_force_qs_lh
++;
2596 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2597 return; /* Someone beat us to it. */
2599 ACCESS_ONCE(rsp
->gp_flags
) =
2600 ACCESS_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
;
2601 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2602 rcu_gp_kthread_wake(rsp
);
2606 * This does the RCU core processing work for the specified rcu_state
2607 * and rcu_data structures. This may be called only from the CPU to
2608 * whom the rdp belongs.
2611 __rcu_process_callbacks(struct rcu_state
*rsp
)
2613 unsigned long flags
;
2615 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2617 WARN_ON_ONCE(rdp
->beenonline
== 0);
2619 /* Update RCU state based on any recent quiescent states. */
2620 rcu_check_quiescent_state(rsp
, rdp
);
2622 /* Does this CPU require a not-yet-started grace period? */
2623 local_irq_save(flags
);
2624 if (cpu_needs_another_gp(rsp
, rdp
)) {
2625 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2626 needwake
= rcu_start_gp(rsp
);
2627 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2629 rcu_gp_kthread_wake(rsp
);
2631 local_irq_restore(flags
);
2634 /* If there are callbacks ready, invoke them. */
2635 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2636 invoke_rcu_callbacks(rsp
, rdp
);
2638 /* Do any needed deferred wakeups of rcuo kthreads. */
2639 do_nocb_deferred_wakeup(rdp
);
2643 * Do RCU core processing for the current CPU.
2645 static void rcu_process_callbacks(struct softirq_action
*unused
)
2647 struct rcu_state
*rsp
;
2649 if (cpu_is_offline(smp_processor_id()))
2651 trace_rcu_utilization(TPS("Start RCU core"));
2652 for_each_rcu_flavor(rsp
)
2653 __rcu_process_callbacks(rsp
);
2654 trace_rcu_utilization(TPS("End RCU core"));
2658 * Schedule RCU callback invocation. If the specified type of RCU
2659 * does not support RCU priority boosting, just do a direct call,
2660 * otherwise wake up the per-CPU kernel kthread. Note that because we
2661 * are running on the current CPU with interrupts disabled, the
2662 * rcu_cpu_kthread_task cannot disappear out from under us.
2664 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2666 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2668 if (likely(!rsp
->boost
)) {
2669 rcu_do_batch(rsp
, rdp
);
2672 invoke_rcu_callbacks_kthread();
2675 static void invoke_rcu_core(void)
2677 if (cpu_online(smp_processor_id()))
2678 raise_softirq(RCU_SOFTIRQ
);
2682 * Handle any core-RCU processing required by a call_rcu() invocation.
2684 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2685 struct rcu_head
*head
, unsigned long flags
)
2690 * If called from an extended quiescent state, invoke the RCU
2691 * core in order to force a re-evaluation of RCU's idleness.
2693 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2696 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2697 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2701 * Force the grace period if too many callbacks or too long waiting.
2702 * Enforce hysteresis, and don't invoke force_quiescent_state()
2703 * if some other CPU has recently done so. Also, don't bother
2704 * invoking force_quiescent_state() if the newly enqueued callback
2705 * is the only one waiting for a grace period to complete.
2707 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2709 /* Are we ignoring a completed grace period? */
2710 note_gp_changes(rsp
, rdp
);
2712 /* Start a new grace period if one not already started. */
2713 if (!rcu_gp_in_progress(rsp
)) {
2714 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2716 raw_spin_lock(&rnp_root
->lock
);
2717 smp_mb__after_unlock_lock();
2718 needwake
= rcu_start_gp(rsp
);
2719 raw_spin_unlock(&rnp_root
->lock
);
2721 rcu_gp_kthread_wake(rsp
);
2723 /* Give the grace period a kick. */
2724 rdp
->blimit
= LONG_MAX
;
2725 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2726 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2727 force_quiescent_state(rsp
);
2728 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2729 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2735 * RCU callback function to leak a callback.
2737 static void rcu_leak_callback(struct rcu_head
*rhp
)
2742 * Helper function for call_rcu() and friends. The cpu argument will
2743 * normally be -1, indicating "currently running CPU". It may specify
2744 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2745 * is expected to specify a CPU.
2748 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2749 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2751 unsigned long flags
;
2752 struct rcu_data
*rdp
;
2754 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
2755 if (debug_rcu_head_queue(head
)) {
2756 /* Probable double call_rcu(), so leak the callback. */
2757 ACCESS_ONCE(head
->func
) = rcu_leak_callback
;
2758 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2765 * Opportunistically note grace-period endings and beginnings.
2766 * Note that we might see a beginning right after we see an
2767 * end, but never vice versa, since this CPU has to pass through
2768 * a quiescent state betweentimes.
2770 local_irq_save(flags
);
2771 rdp
= this_cpu_ptr(rsp
->rda
);
2773 /* Add the callback to our list. */
2774 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2778 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2779 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
2780 WARN_ON_ONCE(offline
);
2781 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2782 local_irq_restore(flags
);
2785 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
+ 1;
2789 rcu_idle_count_callbacks_posted();
2790 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2791 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2792 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2794 if (__is_kfree_rcu_offset((unsigned long)func
))
2795 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2796 rdp
->qlen_lazy
, rdp
->qlen
);
2798 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2800 /* Go handle any RCU core processing required. */
2801 __call_rcu_core(rsp
, rdp
, head
, flags
);
2802 local_irq_restore(flags
);
2806 * Queue an RCU-sched callback for invocation after a grace period.
2808 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2810 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2812 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2815 * Queue an RCU callback for invocation after a quicker grace period.
2817 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2819 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2821 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2824 * Queue an RCU callback for lazy invocation after a grace period.
2825 * This will likely be later named something like "call_rcu_lazy()",
2826 * but this change will require some way of tagging the lazy RCU
2827 * callbacks in the list of pending callbacks. Until then, this
2828 * function may only be called from __kfree_rcu().
2830 void kfree_call_rcu(struct rcu_head
*head
,
2831 void (*func
)(struct rcu_head
*rcu
))
2833 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
2835 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
2838 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2839 * any blocking grace-period wait automatically implies a grace period
2840 * if there is only one CPU online at any point time during execution
2841 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2842 * occasionally incorrectly indicate that there are multiple CPUs online
2843 * when there was in fact only one the whole time, as this just adds
2844 * some overhead: RCU still operates correctly.
2846 static inline int rcu_blocking_is_gp(void)
2850 might_sleep(); /* Check for RCU read-side critical section. */
2852 ret
= num_online_cpus() <= 1;
2858 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2860 * Control will return to the caller some time after a full rcu-sched
2861 * grace period has elapsed, in other words after all currently executing
2862 * rcu-sched read-side critical sections have completed. These read-side
2863 * critical sections are delimited by rcu_read_lock_sched() and
2864 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2865 * local_irq_disable(), and so on may be used in place of
2866 * rcu_read_lock_sched().
2868 * This means that all preempt_disable code sequences, including NMI and
2869 * non-threaded hardware-interrupt handlers, in progress on entry will
2870 * have completed before this primitive returns. However, this does not
2871 * guarantee that softirq handlers will have completed, since in some
2872 * kernels, these handlers can run in process context, and can block.
2874 * Note that this guarantee implies further memory-ordering guarantees.
2875 * On systems with more than one CPU, when synchronize_sched() returns,
2876 * each CPU is guaranteed to have executed a full memory barrier since the
2877 * end of its last RCU-sched read-side critical section whose beginning
2878 * preceded the call to synchronize_sched(). In addition, each CPU having
2879 * an RCU read-side critical section that extends beyond the return from
2880 * synchronize_sched() is guaranteed to have executed a full memory barrier
2881 * after the beginning of synchronize_sched() and before the beginning of
2882 * that RCU read-side critical section. Note that these guarantees include
2883 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2884 * that are executing in the kernel.
2886 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2887 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2888 * to have executed a full memory barrier during the execution of
2889 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2890 * again only if the system has more than one CPU).
2892 * This primitive provides the guarantees made by the (now removed)
2893 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2894 * guarantees that rcu_read_lock() sections will have completed.
2895 * In "classic RCU", these two guarantees happen to be one and
2896 * the same, but can differ in realtime RCU implementations.
2898 void synchronize_sched(void)
2900 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2901 !lock_is_held(&rcu_lock_map
) &&
2902 !lock_is_held(&rcu_sched_lock_map
),
2903 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2904 if (rcu_blocking_is_gp())
2907 synchronize_sched_expedited();
2909 wait_rcu_gp(call_rcu_sched
);
2911 EXPORT_SYMBOL_GPL(synchronize_sched
);
2914 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2916 * Control will return to the caller some time after a full rcu_bh grace
2917 * period has elapsed, in other words after all currently executing rcu_bh
2918 * read-side critical sections have completed. RCU read-side critical
2919 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2920 * and may be nested.
2922 * See the description of synchronize_sched() for more detailed information
2923 * on memory ordering guarantees.
2925 void synchronize_rcu_bh(void)
2927 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2928 !lock_is_held(&rcu_lock_map
) &&
2929 !lock_is_held(&rcu_sched_lock_map
),
2930 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2931 if (rcu_blocking_is_gp())
2934 synchronize_rcu_bh_expedited();
2936 wait_rcu_gp(call_rcu_bh
);
2938 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2941 * get_state_synchronize_rcu - Snapshot current RCU state
2943 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2944 * to determine whether or not a full grace period has elapsed in the
2947 unsigned long get_state_synchronize_rcu(void)
2950 * Any prior manipulation of RCU-protected data must happen
2951 * before the load from ->gpnum.
2956 * Make sure this load happens before the purportedly
2957 * time-consuming work between get_state_synchronize_rcu()
2958 * and cond_synchronize_rcu().
2960 return smp_load_acquire(&rcu_state_p
->gpnum
);
2962 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
2965 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2967 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2969 * If a full RCU grace period has elapsed since the earlier call to
2970 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2971 * synchronize_rcu() to wait for a full grace period.
2973 * Yes, this function does not take counter wrap into account. But
2974 * counter wrap is harmless. If the counter wraps, we have waited for
2975 * more than 2 billion grace periods (and way more on a 64-bit system!),
2976 * so waiting for one additional grace period should be just fine.
2978 void cond_synchronize_rcu(unsigned long oldstate
)
2980 unsigned long newstate
;
2983 * Ensure that this load happens before any RCU-destructive
2984 * actions the caller might carry out after we return.
2986 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
2987 if (ULONG_CMP_GE(oldstate
, newstate
))
2990 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
2992 static int synchronize_sched_expedited_cpu_stop(void *data
)
2995 * There must be a full memory barrier on each affected CPU
2996 * between the time that try_stop_cpus() is called and the
2997 * time that it returns.
2999 * In the current initial implementation of cpu_stop, the
3000 * above condition is already met when the control reaches
3001 * this point and the following smp_mb() is not strictly
3002 * necessary. Do smp_mb() anyway for documentation and
3003 * robustness against future implementation changes.
3005 smp_mb(); /* See above comment block. */
3010 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3012 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3013 * approach to force the grace period to end quickly. This consumes
3014 * significant time on all CPUs and is unfriendly to real-time workloads,
3015 * so is thus not recommended for any sort of common-case code. In fact,
3016 * if you are using synchronize_sched_expedited() in a loop, please
3017 * restructure your code to batch your updates, and then use a single
3018 * synchronize_sched() instead.
3020 * This implementation can be thought of as an application of ticket
3021 * locking to RCU, with sync_sched_expedited_started and
3022 * sync_sched_expedited_done taking on the roles of the halves
3023 * of the ticket-lock word. Each task atomically increments
3024 * sync_sched_expedited_started upon entry, snapshotting the old value,
3025 * then attempts to stop all the CPUs. If this succeeds, then each
3026 * CPU will have executed a context switch, resulting in an RCU-sched
3027 * grace period. We are then done, so we use atomic_cmpxchg() to
3028 * update sync_sched_expedited_done to match our snapshot -- but
3029 * only if someone else has not already advanced past our snapshot.
3031 * On the other hand, if try_stop_cpus() fails, we check the value
3032 * of sync_sched_expedited_done. If it has advanced past our
3033 * initial snapshot, then someone else must have forced a grace period
3034 * some time after we took our snapshot. In this case, our work is
3035 * done for us, and we can simply return. Otherwise, we try again,
3036 * but keep our initial snapshot for purposes of checking for someone
3037 * doing our work for us.
3039 * If we fail too many times in a row, we fall back to synchronize_sched().
3041 void synchronize_sched_expedited(void)
3046 long firstsnap
, s
, snap
;
3048 struct rcu_state
*rsp
= &rcu_sched_state
;
3051 * If we are in danger of counter wrap, just do synchronize_sched().
3052 * By allowing sync_sched_expedited_started to advance no more than
3053 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3054 * that more than 3.5 billion CPUs would be required to force a
3055 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3056 * course be required on a 64-bit system.
3058 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3059 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3061 synchronize_sched();
3062 atomic_long_inc(&rsp
->expedited_wrap
);
3067 * Take a ticket. Note that atomic_inc_return() implies a
3068 * full memory barrier.
3070 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3072 if (!try_get_online_cpus()) {
3073 /* CPU hotplug operation in flight, fall back to normal GP. */
3074 wait_rcu_gp(call_rcu_sched
);
3075 atomic_long_inc(&rsp
->expedited_normal
);
3078 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3080 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3081 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3083 cpumask_copy(cm
, cpu_online_mask
);
3084 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3085 for_each_cpu(cpu
, cm
) {
3086 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3088 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3089 cpumask_clear_cpu(cpu
, cm
);
3091 if (cpumask_weight(cm
) == 0)
3096 * Each pass through the following loop attempts to force a
3097 * context switch on each CPU.
3099 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3100 synchronize_sched_expedited_cpu_stop
,
3103 atomic_long_inc(&rsp
->expedited_tryfail
);
3105 /* Check to see if someone else did our work for us. */
3106 s
= atomic_long_read(&rsp
->expedited_done
);
3107 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3108 /* ensure test happens before caller kfree */
3109 smp_mb__before_atomic(); /* ^^^ */
3110 atomic_long_inc(&rsp
->expedited_workdone1
);
3111 free_cpumask_var(cm
);
3115 /* No joy, try again later. Or just synchronize_sched(). */
3116 if (trycount
++ < 10) {
3117 udelay(trycount
* num_online_cpus());
3119 wait_rcu_gp(call_rcu_sched
);
3120 atomic_long_inc(&rsp
->expedited_normal
);
3121 free_cpumask_var(cm
);
3125 /* Recheck to see if someone else did our work for us. */
3126 s
= atomic_long_read(&rsp
->expedited_done
);
3127 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3128 /* ensure test happens before caller kfree */
3129 smp_mb__before_atomic(); /* ^^^ */
3130 atomic_long_inc(&rsp
->expedited_workdone2
);
3131 free_cpumask_var(cm
);
3136 * Refetching sync_sched_expedited_started allows later
3137 * callers to piggyback on our grace period. We retry
3138 * after they started, so our grace period works for them,
3139 * and they started after our first try, so their grace
3140 * period works for us.
3142 if (!try_get_online_cpus()) {
3143 /* CPU hotplug operation in flight, use normal GP. */
3144 wait_rcu_gp(call_rcu_sched
);
3145 atomic_long_inc(&rsp
->expedited_normal
);
3146 free_cpumask_var(cm
);
3149 snap
= atomic_long_read(&rsp
->expedited_start
);
3150 smp_mb(); /* ensure read is before try_stop_cpus(). */
3152 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3155 free_cpumask_var(cm
);
3158 * Everyone up to our most recent fetch is covered by our grace
3159 * period. Update the counter, but only if our work is still
3160 * relevant -- which it won't be if someone who started later
3161 * than we did already did their update.
3164 atomic_long_inc(&rsp
->expedited_done_tries
);
3165 s
= atomic_long_read(&rsp
->expedited_done
);
3166 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3167 /* ensure test happens before caller kfree */
3168 smp_mb__before_atomic(); /* ^^^ */
3169 atomic_long_inc(&rsp
->expedited_done_lost
);
3172 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3173 atomic_long_inc(&rsp
->expedited_done_exit
);
3177 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3180 * Check to see if there is any immediate RCU-related work to be done
3181 * by the current CPU, for the specified type of RCU, returning 1 if so.
3182 * The checks are in order of increasing expense: checks that can be
3183 * carried out against CPU-local state are performed first. However,
3184 * we must check for CPU stalls first, else we might not get a chance.
3186 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3188 struct rcu_node
*rnp
= rdp
->mynode
;
3190 rdp
->n_rcu_pending
++;
3192 /* Check for CPU stalls, if enabled. */
3193 check_cpu_stall(rsp
, rdp
);
3195 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3196 if (rcu_nohz_full_cpu(rsp
))
3199 /* Is the RCU core waiting for a quiescent state from this CPU? */
3200 if (rcu_scheduler_fully_active
&&
3201 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3202 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3203 rdp
->n_rp_qs_pending
++;
3204 } else if (rdp
->qs_pending
&&
3205 (rdp
->passed_quiesce
||
3206 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3207 rdp
->n_rp_report_qs
++;
3211 /* Does this CPU have callbacks ready to invoke? */
3212 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3213 rdp
->n_rp_cb_ready
++;
3217 /* Has RCU gone idle with this CPU needing another grace period? */
3218 if (cpu_needs_another_gp(rsp
, rdp
)) {
3219 rdp
->n_rp_cpu_needs_gp
++;
3223 /* Has another RCU grace period completed? */
3224 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3225 rdp
->n_rp_gp_completed
++;
3229 /* Has a new RCU grace period started? */
3230 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3231 unlikely(ACCESS_ONCE(rdp
->gpwrap
))) { /* outside lock */
3232 rdp
->n_rp_gp_started
++;
3236 /* Does this CPU need a deferred NOCB wakeup? */
3237 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3238 rdp
->n_rp_nocb_defer_wakeup
++;
3243 rdp
->n_rp_need_nothing
++;
3248 * Check to see if there is any immediate RCU-related work to be done
3249 * by the current CPU, returning 1 if so. This function is part of the
3250 * RCU implementation; it is -not- an exported member of the RCU API.
3252 static int rcu_pending(void)
3254 struct rcu_state
*rsp
;
3256 for_each_rcu_flavor(rsp
)
3257 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3263 * Return true if the specified CPU has any callback. If all_lazy is
3264 * non-NULL, store an indication of whether all callbacks are lazy.
3265 * (If there are no callbacks, all of them are deemed to be lazy.)
3267 static int __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3271 struct rcu_data
*rdp
;
3272 struct rcu_state
*rsp
;
3274 for_each_rcu_flavor(rsp
) {
3275 rdp
= this_cpu_ptr(rsp
->rda
);
3279 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3290 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3291 * the compiler is expected to optimize this away.
3293 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3294 int cpu
, unsigned long done
)
3296 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3297 atomic_read(&rsp
->barrier_cpu_count
), done
);
3301 * RCU callback function for _rcu_barrier(). If we are last, wake
3302 * up the task executing _rcu_barrier().
3304 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3306 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3307 struct rcu_state
*rsp
= rdp
->rsp
;
3309 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3310 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3311 complete(&rsp
->barrier_completion
);
3313 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3318 * Called with preemption disabled, and from cross-cpu IRQ context.
3320 static void rcu_barrier_func(void *type
)
3322 struct rcu_state
*rsp
= type
;
3323 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3325 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3326 atomic_inc(&rsp
->barrier_cpu_count
);
3327 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3331 * Orchestrate the specified type of RCU barrier, waiting for all
3332 * RCU callbacks of the specified type to complete.
3334 static void _rcu_barrier(struct rcu_state
*rsp
)
3337 struct rcu_data
*rdp
;
3338 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
3339 unsigned long snap_done
;
3341 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3343 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3344 mutex_lock(&rsp
->barrier_mutex
);
3347 * Ensure that all prior references, including to ->n_barrier_done,
3348 * are ordered before the _rcu_barrier() machinery.
3350 smp_mb(); /* See above block comment. */
3353 * Recheck ->n_barrier_done to see if others did our work for us.
3354 * This means checking ->n_barrier_done for an even-to-odd-to-even
3355 * transition. The "if" expression below therefore rounds the old
3356 * value up to the next even number and adds two before comparing.
3358 snap_done
= rsp
->n_barrier_done
;
3359 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3362 * If the value in snap is odd, we needed to wait for the current
3363 * rcu_barrier() to complete, then wait for the next one, in other
3364 * words, we need the value of snap_done to be three larger than
3365 * the value of snap. On the other hand, if the value in snap is
3366 * even, we only had to wait for the next rcu_barrier() to complete,
3367 * in other words, we need the value of snap_done to be only two
3368 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3369 * this for us (thank you, Linus!).
3371 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3372 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3373 smp_mb(); /* caller's subsequent code after above check. */
3374 mutex_unlock(&rsp
->barrier_mutex
);
3379 * Increment ->n_barrier_done to avoid duplicate work. Use
3380 * ACCESS_ONCE() to prevent the compiler from speculating
3381 * the increment to precede the early-exit check.
3383 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3384 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3385 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3386 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3389 * Initialize the count to one rather than to zero in order to
3390 * avoid a too-soon return to zero in case of a short grace period
3391 * (or preemption of this task). Exclude CPU-hotplug operations
3392 * to ensure that no offline CPU has callbacks queued.
3394 init_completion(&rsp
->barrier_completion
);
3395 atomic_set(&rsp
->barrier_cpu_count
, 1);
3399 * Force each CPU with callbacks to register a new callback.
3400 * When that callback is invoked, we will know that all of the
3401 * corresponding CPU's preceding callbacks have been invoked.
3403 for_each_possible_cpu(cpu
) {
3404 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3406 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3407 if (rcu_is_nocb_cpu(cpu
)) {
3408 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3409 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3410 rsp
->n_barrier_done
);
3412 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3413 rsp
->n_barrier_done
);
3414 atomic_inc(&rsp
->barrier_cpu_count
);
3415 __call_rcu(&rdp
->barrier_head
,
3416 rcu_barrier_callback
, rsp
, cpu
, 0);
3418 } else if (ACCESS_ONCE(rdp
->qlen
)) {
3419 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3420 rsp
->n_barrier_done
);
3421 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3423 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3424 rsp
->n_barrier_done
);
3430 * Now that we have an rcu_barrier_callback() callback on each
3431 * CPU, and thus each counted, remove the initial count.
3433 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3434 complete(&rsp
->barrier_completion
);
3436 /* Increment ->n_barrier_done to prevent duplicate work. */
3437 smp_mb(); /* Keep increment after above mechanism. */
3438 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3439 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3440 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3441 smp_mb(); /* Keep increment before caller's subsequent code. */
3443 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3444 wait_for_completion(&rsp
->barrier_completion
);
3446 /* Other rcu_barrier() invocations can now safely proceed. */
3447 mutex_unlock(&rsp
->barrier_mutex
);
3451 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3453 void rcu_barrier_bh(void)
3455 _rcu_barrier(&rcu_bh_state
);
3457 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3460 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3462 void rcu_barrier_sched(void)
3464 _rcu_barrier(&rcu_sched_state
);
3466 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3469 * Do boot-time initialization of a CPU's per-CPU RCU data.
3472 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3474 unsigned long flags
;
3475 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3476 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3478 /* Set up local state, ensuring consistent view of global state. */
3479 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3480 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3481 init_callback_list(rdp
);
3483 ACCESS_ONCE(rdp
->qlen
) = 0;
3484 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3485 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3486 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3489 rcu_boot_init_nocb_percpu_data(rdp
);
3490 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3494 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3495 * offline event can be happening at a given time. Note also that we
3496 * can accept some slop in the rsp->completed access due to the fact
3497 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3500 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3502 unsigned long flags
;
3504 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3505 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3507 /* Exclude new grace periods. */
3508 mutex_lock(&rsp
->onoff_mutex
);
3510 /* Set up local state, ensuring consistent view of global state. */
3511 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3512 rdp
->beenonline
= 1; /* We have now been online. */
3513 rdp
->qlen_last_fqs_check
= 0;
3514 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3515 rdp
->blimit
= blimit
;
3516 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3517 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3518 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3519 atomic_set(&rdp
->dynticks
->dynticks
,
3520 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3521 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3523 /* Add CPU to rcu_node bitmasks. */
3525 mask
= rdp
->grpmask
;
3527 /* Exclude any attempts to start a new GP on small systems. */
3528 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3529 rnp
->qsmaskinit
|= mask
;
3530 mask
= rnp
->grpmask
;
3531 if (rnp
== rdp
->mynode
) {
3533 * If there is a grace period in progress, we will
3534 * set up to wait for it next time we run the
3537 rdp
->gpnum
= rnp
->completed
;
3538 rdp
->completed
= rnp
->completed
;
3539 rdp
->passed_quiesce
= 0;
3540 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
3541 rdp
->qs_pending
= 0;
3542 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3544 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
3546 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
3547 local_irq_restore(flags
);
3549 mutex_unlock(&rsp
->onoff_mutex
);
3552 static void rcu_prepare_cpu(int cpu
)
3554 struct rcu_state
*rsp
;
3556 for_each_rcu_flavor(rsp
)
3557 rcu_init_percpu_data(cpu
, rsp
);
3561 * Handle CPU online/offline notification events.
3563 static int rcu_cpu_notify(struct notifier_block
*self
,
3564 unsigned long action
, void *hcpu
)
3566 long cpu
= (long)hcpu
;
3567 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3568 struct rcu_node
*rnp
= rdp
->mynode
;
3569 struct rcu_state
*rsp
;
3571 trace_rcu_utilization(TPS("Start CPU hotplug"));
3573 case CPU_UP_PREPARE
:
3574 case CPU_UP_PREPARE_FROZEN
:
3575 rcu_prepare_cpu(cpu
);
3576 rcu_prepare_kthreads(cpu
);
3577 rcu_spawn_all_nocb_kthreads(cpu
);
3580 case CPU_DOWN_FAILED
:
3581 rcu_boost_kthread_setaffinity(rnp
, -1);
3583 case CPU_DOWN_PREPARE
:
3584 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3587 case CPU_DYING_FROZEN
:
3588 for_each_rcu_flavor(rsp
)
3589 rcu_cleanup_dying_cpu(rsp
);
3592 case CPU_DEAD_FROZEN
:
3593 case CPU_UP_CANCELED
:
3594 case CPU_UP_CANCELED_FROZEN
:
3595 for_each_rcu_flavor(rsp
) {
3596 rcu_cleanup_dead_cpu(cpu
, rsp
);
3597 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3603 trace_rcu_utilization(TPS("End CPU hotplug"));
3607 static int rcu_pm_notify(struct notifier_block
*self
,
3608 unsigned long action
, void *hcpu
)
3611 case PM_HIBERNATION_PREPARE
:
3612 case PM_SUSPEND_PREPARE
:
3613 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3616 case PM_POST_HIBERNATION
:
3617 case PM_POST_SUSPEND
:
3627 * Spawn the kthreads that handle each RCU flavor's grace periods.
3629 static int __init
rcu_spawn_gp_kthread(void)
3631 unsigned long flags
;
3632 int kthread_prio_in
= kthread_prio
;
3633 struct rcu_node
*rnp
;
3634 struct rcu_state
*rsp
;
3635 struct sched_param sp
;
3636 struct task_struct
*t
;
3638 /* Force priority into range. */
3639 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3641 else if (kthread_prio
< 0)
3643 else if (kthread_prio
> 99)
3645 if (kthread_prio
!= kthread_prio_in
)
3646 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3647 kthread_prio
, kthread_prio_in
);
3649 rcu_scheduler_fully_active
= 1;
3650 for_each_rcu_flavor(rsp
) {
3651 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3653 rnp
= rcu_get_root(rsp
);
3654 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3655 rsp
->gp_kthread
= t
;
3657 sp
.sched_priority
= kthread_prio
;
3658 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3661 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3663 rcu_spawn_nocb_kthreads();
3664 rcu_spawn_boost_kthreads();
3667 early_initcall(rcu_spawn_gp_kthread
);
3670 * This function is invoked towards the end of the scheduler's initialization
3671 * process. Before this is called, the idle task might contain
3672 * RCU read-side critical sections (during which time, this idle
3673 * task is booting the system). After this function is called, the
3674 * idle tasks are prohibited from containing RCU read-side critical
3675 * sections. This function also enables RCU lockdep checking.
3677 void rcu_scheduler_starting(void)
3679 WARN_ON(num_online_cpus() != 1);
3680 WARN_ON(nr_context_switches() > 0);
3681 rcu_scheduler_active
= 1;
3685 * Compute the per-level fanout, either using the exact fanout specified
3686 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3688 #ifdef CONFIG_RCU_FANOUT_EXACT
3689 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3693 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3694 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3695 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3697 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3698 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3705 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3706 ccur
= rsp
->levelcnt
[i
];
3707 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3711 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3714 * Helper function for rcu_init() that initializes one rcu_state structure.
3716 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3717 struct rcu_data __percpu
*rda
)
3719 static const char * const buf
[] = {
3723 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3724 static const char * const fqs
[] = {
3728 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3729 static u8 fl_mask
= 0x1;
3733 struct rcu_node
*rnp
;
3735 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3737 /* Silence gcc 4.8 warning about array index out of range. */
3738 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3739 panic("rcu_init_one: rcu_num_lvls overflow");
3741 /* Initialize the level-tracking arrays. */
3743 for (i
= 0; i
< rcu_num_lvls
; i
++)
3744 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3745 for (i
= 1; i
< rcu_num_lvls
; i
++)
3746 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3747 rcu_init_levelspread(rsp
);
3748 rsp
->flavor_mask
= fl_mask
;
3751 /* Initialize the elements themselves, starting from the leaves. */
3753 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3754 cpustride
*= rsp
->levelspread
[i
];
3755 rnp
= rsp
->level
[i
];
3756 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3757 raw_spin_lock_init(&rnp
->lock
);
3758 lockdep_set_class_and_name(&rnp
->lock
,
3759 &rcu_node_class
[i
], buf
[i
]);
3760 raw_spin_lock_init(&rnp
->fqslock
);
3761 lockdep_set_class_and_name(&rnp
->fqslock
,
3762 &rcu_fqs_class
[i
], fqs
[i
]);
3763 rnp
->gpnum
= rsp
->gpnum
;
3764 rnp
->completed
= rsp
->completed
;
3766 rnp
->qsmaskinit
= 0;
3767 rnp
->grplo
= j
* cpustride
;
3768 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
3769 if (rnp
->grphi
>= nr_cpu_ids
)
3770 rnp
->grphi
= nr_cpu_ids
- 1;
3776 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
3777 rnp
->grpmask
= 1UL << rnp
->grpnum
;
3778 rnp
->parent
= rsp
->level
[i
- 1] +
3779 j
/ rsp
->levelspread
[i
- 1];
3782 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3783 rcu_init_one_nocb(rnp
);
3788 init_waitqueue_head(&rsp
->gp_wq
);
3789 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3790 for_each_possible_cpu(i
) {
3791 while (i
> rnp
->grphi
)
3793 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3794 rcu_boot_init_percpu_data(i
, rsp
);
3796 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3800 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3801 * replace the definitions in tree.h because those are needed to size
3802 * the ->node array in the rcu_state structure.
3804 static void __init
rcu_init_geometry(void)
3810 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3813 * Initialize any unspecified boot parameters.
3814 * The default values of jiffies_till_first_fqs and
3815 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3816 * value, which is a function of HZ, then adding one for each
3817 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3819 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
3820 if (jiffies_till_first_fqs
== ULONG_MAX
)
3821 jiffies_till_first_fqs
= d
;
3822 if (jiffies_till_next_fqs
== ULONG_MAX
)
3823 jiffies_till_next_fqs
= d
;
3825 /* If the compile-time values are accurate, just leave. */
3826 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3827 nr_cpu_ids
== NR_CPUS
)
3829 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3830 rcu_fanout_leaf
, nr_cpu_ids
);
3833 * Compute number of nodes that can be handled an rcu_node tree
3834 * with the given number of levels. Setting rcu_capacity[0] makes
3835 * some of the arithmetic easier.
3837 rcu_capacity
[0] = 1;
3838 rcu_capacity
[1] = rcu_fanout_leaf
;
3839 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3840 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3843 * The boot-time rcu_fanout_leaf parameter is only permitted
3844 * to increase the leaf-level fanout, not decrease it. Of course,
3845 * the leaf-level fanout cannot exceed the number of bits in
3846 * the rcu_node masks. Finally, the tree must be able to accommodate
3847 * the configured number of CPUs. Complain and fall back to the
3848 * compile-time values if these limits are exceeded.
3850 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3851 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3852 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3857 /* Calculate the number of rcu_nodes at each level of the tree. */
3858 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3859 if (n
<= rcu_capacity
[i
]) {
3860 for (j
= 0; j
<= i
; j
++)
3862 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3864 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3869 /* Calculate the total number of rcu_node structures. */
3871 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3872 rcu_num_nodes
+= num_rcu_lvl
[i
];
3876 void __init
rcu_init(void)
3880 rcu_bootup_announce();
3881 rcu_init_geometry();
3882 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3883 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3884 __rcu_init_preempt();
3885 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3888 * We don't need protection against CPU-hotplug here because
3889 * this is called early in boot, before either interrupts
3890 * or the scheduler are operational.
3892 cpu_notifier(rcu_cpu_notify
, 0);
3893 pm_notifier(rcu_pm_notify
, 0);
3894 for_each_online_cpu(cpu
)
3895 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3897 rcu_early_boot_tests();
3900 #include "tree_plugin.h"