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 DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
95 struct rcu_state sname##_state = { \
96 .level = { &sname##_state.node[0] }, \
97 .rda = &sname##_data, \
99 .fqs_state = RCU_GP_IDLE, \
100 .gpnum = 0UL - 300UL, \
101 .completed = 0UL - 300UL, \
102 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
103 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
104 .orphan_donetail = &sname##_state.orphan_donelist, \
105 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
106 .name = RCU_STATE_NAME(sname), \
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_init_new_rnp(struct rcu_node
*rnp_leaf
);
156 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
157 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
158 static void invoke_rcu_core(void);
159 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
161 /* rcuc/rcub kthread realtime priority */
162 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
163 module_param(kthread_prio
, int, 0644);
165 /* Delay in jiffies for grace-period initialization delays. */
166 static int gp_init_delay
= IS_ENABLED(CONFIG_RCU_TORTURE_TEST_SLOW_INIT
)
167 ? CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
169 module_param(gp_init_delay
, int, 0644);
172 * Track the rcutorture test sequence number and the update version
173 * number within a given test. The rcutorture_testseq is incremented
174 * on every rcutorture module load and unload, so has an odd value
175 * when a test is running. The rcutorture_vernum is set to zero
176 * when rcutorture starts and is incremented on each rcutorture update.
177 * These variables enable correlating rcutorture output with the
178 * RCU tracing information.
180 unsigned long rcutorture_testseq
;
181 unsigned long rcutorture_vernum
;
184 * Compute the mask of online CPUs for the specified rcu_node structure.
185 * This will not be stable unless the rcu_node structure's ->lock is
186 * held, but the bit corresponding to the current CPU will be stable
189 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
191 return ACCESS_ONCE(rnp
->qsmaskinitnext
);
195 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
196 * permit this function to be invoked without holding the root rcu_node
197 * structure's ->lock, but of course results can be subject to change.
199 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
201 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
205 * Note a quiescent state. Because we do not need to know
206 * how many quiescent states passed, just if there was at least
207 * one since the start of the grace period, this just sets a flag.
208 * The caller must have disabled preemption.
210 void rcu_sched_qs(void)
212 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
213 trace_rcu_grace_period(TPS("rcu_sched"),
214 __this_cpu_read(rcu_sched_data
.gpnum
),
216 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
222 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
223 trace_rcu_grace_period(TPS("rcu_bh"),
224 __this_cpu_read(rcu_bh_data
.gpnum
),
226 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
230 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
232 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
233 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
234 .dynticks
= ATOMIC_INIT(1),
235 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
236 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
237 .dynticks_idle
= ATOMIC_INIT(1),
238 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
241 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
242 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
245 * Let the RCU core know that this CPU has gone through the scheduler,
246 * which is a quiescent state. This is called when the need for a
247 * quiescent state is urgent, so we burn an atomic operation and full
248 * memory barriers to let the RCU core know about it, regardless of what
249 * this CPU might (or might not) do in the near future.
251 * We inform the RCU core by emulating a zero-duration dyntick-idle
252 * period, which we in turn do by incrementing the ->dynticks counter
255 static void rcu_momentary_dyntick_idle(void)
258 struct rcu_data
*rdp
;
259 struct rcu_dynticks
*rdtp
;
261 struct rcu_state
*rsp
;
263 local_irq_save(flags
);
266 * Yes, we can lose flag-setting operations. This is OK, because
267 * the flag will be set again after some delay.
269 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
270 raw_cpu_write(rcu_sched_qs_mask
, 0);
272 /* Find the flavor that needs a quiescent state. */
273 for_each_rcu_flavor(rsp
) {
274 rdp
= raw_cpu_ptr(rsp
->rda
);
275 if (!(resched_mask
& rsp
->flavor_mask
))
277 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
278 if (ACCESS_ONCE(rdp
->mynode
->completed
) !=
279 ACCESS_ONCE(rdp
->cond_resched_completed
))
283 * Pretend to be momentarily idle for the quiescent state.
284 * This allows the grace-period kthread to record the
285 * quiescent state, with no need for this CPU to do anything
288 rdtp
= this_cpu_ptr(&rcu_dynticks
);
289 smp_mb__before_atomic(); /* Earlier stuff before QS. */
290 atomic_add(2, &rdtp
->dynticks
); /* QS. */
291 smp_mb__after_atomic(); /* Later stuff after QS. */
294 local_irq_restore(flags
);
298 * Note a context switch. This is a quiescent state for RCU-sched,
299 * and requires special handling for preemptible RCU.
300 * The caller must have disabled preemption.
302 void rcu_note_context_switch(void)
304 trace_rcu_utilization(TPS("Start context switch"));
306 rcu_preempt_note_context_switch();
307 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
308 rcu_momentary_dyntick_idle();
309 trace_rcu_utilization(TPS("End context switch"));
311 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
314 * Register a quiescent state for all RCU flavors. If there is an
315 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
316 * dyntick-idle quiescent state visible to other CPUs (but only for those
317 * RCU flavors in desperate need of a quiescent state, which will normally
318 * be none of them). Either way, do a lightweight quiescent state for
321 void rcu_all_qs(void)
323 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
324 rcu_momentary_dyntick_idle();
325 this_cpu_inc(rcu_qs_ctr
);
327 EXPORT_SYMBOL_GPL(rcu_all_qs
);
329 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
330 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
331 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
333 module_param(blimit
, long, 0444);
334 module_param(qhimark
, long, 0444);
335 module_param(qlowmark
, long, 0444);
337 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
338 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
340 module_param(jiffies_till_first_fqs
, ulong
, 0644);
341 module_param(jiffies_till_next_fqs
, ulong
, 0644);
344 * How long the grace period must be before we start recruiting
345 * quiescent-state help from rcu_note_context_switch().
347 static ulong jiffies_till_sched_qs
= HZ
/ 20;
348 module_param(jiffies_till_sched_qs
, ulong
, 0644);
350 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
351 struct rcu_data
*rdp
);
352 static void force_qs_rnp(struct rcu_state
*rsp
,
353 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
354 unsigned long *maxj
),
355 bool *isidle
, unsigned long *maxj
);
356 static void force_quiescent_state(struct rcu_state
*rsp
);
357 static int rcu_pending(void);
360 * Return the number of RCU batches started thus far for debug & stats.
362 unsigned long rcu_batches_started(void)
364 return rcu_state_p
->gpnum
;
366 EXPORT_SYMBOL_GPL(rcu_batches_started
);
369 * Return the number of RCU-sched batches started thus far for debug & stats.
371 unsigned long rcu_batches_started_sched(void)
373 return rcu_sched_state
.gpnum
;
375 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
378 * Return the number of RCU BH batches started thus far for debug & stats.
380 unsigned long rcu_batches_started_bh(void)
382 return rcu_bh_state
.gpnum
;
384 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
387 * Return the number of RCU batches completed thus far for debug & stats.
389 unsigned long rcu_batches_completed(void)
391 return rcu_state_p
->completed
;
393 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
396 * Return the number of RCU-sched batches completed thus far for debug & stats.
398 unsigned long rcu_batches_completed_sched(void)
400 return rcu_sched_state
.completed
;
402 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
405 * Return the number of RCU BH batches completed thus far for debug & stats.
407 unsigned long rcu_batches_completed_bh(void)
409 return rcu_bh_state
.completed
;
411 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
414 * Force a quiescent state.
416 void rcu_force_quiescent_state(void)
418 force_quiescent_state(rcu_state_p
);
420 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
423 * Force a quiescent state for RCU BH.
425 void rcu_bh_force_quiescent_state(void)
427 force_quiescent_state(&rcu_bh_state
);
429 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
432 * Force a quiescent state for RCU-sched.
434 void rcu_sched_force_quiescent_state(void)
436 force_quiescent_state(&rcu_sched_state
);
438 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
441 * Show the state of the grace-period kthreads.
443 void show_rcu_gp_kthreads(void)
445 struct rcu_state
*rsp
;
447 for_each_rcu_flavor(rsp
) {
448 pr_info("%s: wait state: %d ->state: %#lx\n",
449 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
450 /* sched_show_task(rsp->gp_kthread); */
453 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
456 * Record the number of times rcutorture tests have been initiated and
457 * terminated. This information allows the debugfs tracing stats to be
458 * correlated to the rcutorture messages, even when the rcutorture module
459 * is being repeatedly loaded and unloaded. In other words, we cannot
460 * store this state in rcutorture itself.
462 void rcutorture_record_test_transition(void)
464 rcutorture_testseq
++;
465 rcutorture_vernum
= 0;
467 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
470 * Send along grace-period-related data for rcutorture diagnostics.
472 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
473 unsigned long *gpnum
, unsigned long *completed
)
475 struct rcu_state
*rsp
= NULL
;
484 case RCU_SCHED_FLAVOR
:
485 rsp
= &rcu_sched_state
;
491 *flags
= ACCESS_ONCE(rsp
->gp_flags
);
492 *gpnum
= ACCESS_ONCE(rsp
->gpnum
);
493 *completed
= ACCESS_ONCE(rsp
->completed
);
500 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
503 * Record the number of writer passes through the current rcutorture test.
504 * This is also used to correlate debugfs tracing stats with the rcutorture
507 void rcutorture_record_progress(unsigned long vernum
)
511 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
514 * Does the CPU have callbacks ready to be invoked?
517 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
519 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
520 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
524 * Return the root node of the specified rcu_state structure.
526 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
528 return &rsp
->node
[0];
532 * Is there any need for future grace periods?
533 * Interrupts must be disabled. If the caller does not hold the root
534 * rnp_node structure's ->lock, the results are advisory only.
536 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
538 struct rcu_node
*rnp
= rcu_get_root(rsp
);
539 int idx
= (ACCESS_ONCE(rnp
->completed
) + 1) & 0x1;
540 int *fp
= &rnp
->need_future_gp
[idx
];
542 return ACCESS_ONCE(*fp
);
546 * Does the current CPU require a not-yet-started grace period?
547 * The caller must have disabled interrupts to prevent races with
548 * normal callback registry.
551 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
555 if (rcu_gp_in_progress(rsp
))
556 return 0; /* No, a grace period is already in progress. */
557 if (rcu_future_needs_gp(rsp
))
558 return 1; /* Yes, a no-CBs CPU needs one. */
559 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
560 return 0; /* No, this is a no-CBs (or offline) CPU. */
561 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
562 return 1; /* Yes, this CPU has newly registered callbacks. */
563 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
564 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
565 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
566 rdp
->nxtcompleted
[i
]))
567 return 1; /* Yes, CBs for future grace period. */
568 return 0; /* No grace period needed. */
572 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
574 * If the new value of the ->dynticks_nesting counter now is zero,
575 * we really have entered idle, and must do the appropriate accounting.
576 * The caller must have disabled interrupts.
578 static void rcu_eqs_enter_common(long long oldval
, bool user
)
580 struct rcu_state
*rsp
;
581 struct rcu_data
*rdp
;
582 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
584 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
585 if (!user
&& !is_idle_task(current
)) {
586 struct task_struct
*idle __maybe_unused
=
587 idle_task(smp_processor_id());
589 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
590 ftrace_dump(DUMP_ORIG
);
591 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
592 current
->pid
, current
->comm
,
593 idle
->pid
, idle
->comm
); /* must be idle task! */
595 for_each_rcu_flavor(rsp
) {
596 rdp
= this_cpu_ptr(rsp
->rda
);
597 do_nocb_deferred_wakeup(rdp
);
599 rcu_prepare_for_idle();
600 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
601 smp_mb__before_atomic(); /* See above. */
602 atomic_inc(&rdtp
->dynticks
);
603 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
604 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
605 rcu_dynticks_task_enter();
608 * It is illegal to enter an extended quiescent state while
609 * in an RCU read-side critical section.
611 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
612 "Illegal idle entry in RCU read-side critical section.");
613 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
614 "Illegal idle entry in RCU-bh read-side critical section.");
615 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
616 "Illegal idle entry in RCU-sched read-side critical section.");
620 * Enter an RCU extended quiescent state, which can be either the
621 * idle loop or adaptive-tickless usermode execution.
623 static void rcu_eqs_enter(bool user
)
626 struct rcu_dynticks
*rdtp
;
628 rdtp
= this_cpu_ptr(&rcu_dynticks
);
629 oldval
= rdtp
->dynticks_nesting
;
630 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
631 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
632 rdtp
->dynticks_nesting
= 0;
633 rcu_eqs_enter_common(oldval
, user
);
635 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
640 * rcu_idle_enter - inform RCU that current CPU is entering idle
642 * Enter idle mode, in other words, -leave- the mode in which RCU
643 * read-side critical sections can occur. (Though RCU read-side
644 * critical sections can occur in irq handlers in idle, a possibility
645 * handled by irq_enter() and irq_exit().)
647 * We crowbar the ->dynticks_nesting field to zero to allow for
648 * the possibility of usermode upcalls having messed up our count
649 * of interrupt nesting level during the prior busy period.
651 void rcu_idle_enter(void)
655 local_irq_save(flags
);
656 rcu_eqs_enter(false);
657 rcu_sysidle_enter(0);
658 local_irq_restore(flags
);
660 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
662 #ifdef CONFIG_RCU_USER_QS
664 * rcu_user_enter - inform RCU that we are resuming userspace.
666 * Enter RCU idle mode right before resuming userspace. No use of RCU
667 * is permitted between this call and rcu_user_exit(). This way the
668 * CPU doesn't need to maintain the tick for RCU maintenance purposes
669 * when the CPU runs in userspace.
671 void rcu_user_enter(void)
675 #endif /* CONFIG_RCU_USER_QS */
678 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
680 * Exit from an interrupt handler, which might possibly result in entering
681 * idle mode, in other words, leaving the mode in which read-side critical
682 * sections can occur.
684 * This code assumes that the idle loop never does anything that might
685 * result in unbalanced calls to irq_enter() and irq_exit(). If your
686 * architecture violates this assumption, RCU will give you what you
687 * deserve, good and hard. But very infrequently and irreproducibly.
689 * Use things like work queues to work around this limitation.
691 * You have been warned.
693 void rcu_irq_exit(void)
697 struct rcu_dynticks
*rdtp
;
699 local_irq_save(flags
);
700 rdtp
= this_cpu_ptr(&rcu_dynticks
);
701 oldval
= rdtp
->dynticks_nesting
;
702 rdtp
->dynticks_nesting
--;
703 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
704 if (rdtp
->dynticks_nesting
)
705 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
707 rcu_eqs_enter_common(oldval
, true);
708 rcu_sysidle_enter(1);
709 local_irq_restore(flags
);
713 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
715 * If the new value of the ->dynticks_nesting counter was previously zero,
716 * we really have exited idle, and must do the appropriate accounting.
717 * The caller must have disabled interrupts.
719 static void rcu_eqs_exit_common(long long oldval
, int user
)
721 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
723 rcu_dynticks_task_exit();
724 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
725 atomic_inc(&rdtp
->dynticks
);
726 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
727 smp_mb__after_atomic(); /* See above. */
728 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
729 rcu_cleanup_after_idle();
730 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
731 if (!user
&& !is_idle_task(current
)) {
732 struct task_struct
*idle __maybe_unused
=
733 idle_task(smp_processor_id());
735 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
736 oldval
, rdtp
->dynticks_nesting
);
737 ftrace_dump(DUMP_ORIG
);
738 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
739 current
->pid
, current
->comm
,
740 idle
->pid
, idle
->comm
); /* must be idle task! */
745 * Exit an RCU extended quiescent state, which can be either the
746 * idle loop or adaptive-tickless usermode execution.
748 static void rcu_eqs_exit(bool user
)
750 struct rcu_dynticks
*rdtp
;
753 rdtp
= this_cpu_ptr(&rcu_dynticks
);
754 oldval
= rdtp
->dynticks_nesting
;
755 WARN_ON_ONCE(oldval
< 0);
756 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
757 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
759 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
760 rcu_eqs_exit_common(oldval
, user
);
765 * rcu_idle_exit - inform RCU that current CPU is leaving idle
767 * Exit idle mode, in other words, -enter- the mode in which RCU
768 * read-side critical sections can occur.
770 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
771 * allow for the possibility of usermode upcalls messing up our count
772 * of interrupt nesting level during the busy period that is just
775 void rcu_idle_exit(void)
779 local_irq_save(flags
);
782 local_irq_restore(flags
);
784 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
786 #ifdef CONFIG_RCU_USER_QS
788 * rcu_user_exit - inform RCU that we are exiting userspace.
790 * Exit RCU idle mode while entering the kernel because it can
791 * run a RCU read side critical section anytime.
793 void rcu_user_exit(void)
797 #endif /* CONFIG_RCU_USER_QS */
800 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
802 * Enter an interrupt handler, which might possibly result in exiting
803 * idle mode, in other words, entering the mode in which read-side critical
804 * sections can occur.
806 * Note that the Linux kernel is fully capable of entering an interrupt
807 * handler that it never exits, for example when doing upcalls to
808 * user mode! This code assumes that the idle loop never does upcalls to
809 * user mode. If your architecture does do upcalls from the idle loop (or
810 * does anything else that results in unbalanced calls to the irq_enter()
811 * and irq_exit() functions), RCU will give you what you deserve, good
812 * and hard. But very infrequently and irreproducibly.
814 * Use things like work queues to work around this limitation.
816 * You have been warned.
818 void rcu_irq_enter(void)
821 struct rcu_dynticks
*rdtp
;
824 local_irq_save(flags
);
825 rdtp
= this_cpu_ptr(&rcu_dynticks
);
826 oldval
= rdtp
->dynticks_nesting
;
827 rdtp
->dynticks_nesting
++;
828 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
830 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
832 rcu_eqs_exit_common(oldval
, true);
834 local_irq_restore(flags
);
838 * rcu_nmi_enter - inform RCU of entry to NMI context
840 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
841 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
842 * that the CPU is active. This implementation permits nested NMIs, as
843 * long as the nesting level does not overflow an int. (You will probably
844 * run out of stack space first.)
846 void rcu_nmi_enter(void)
848 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
851 /* Complain about underflow. */
852 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
855 * If idle from RCU viewpoint, atomically increment ->dynticks
856 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
857 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
858 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
859 * to be in the outermost NMI handler that interrupted an RCU-idle
860 * period (observation due to Andy Lutomirski).
862 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
863 smp_mb__before_atomic(); /* Force delay from prior write. */
864 atomic_inc(&rdtp
->dynticks
);
865 /* atomic_inc() before later RCU read-side crit sects */
866 smp_mb__after_atomic(); /* See above. */
867 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
870 rdtp
->dynticks_nmi_nesting
+= incby
;
875 * rcu_nmi_exit - inform RCU of exit from NMI context
877 * If we are returning from the outermost NMI handler that interrupted an
878 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
879 * to let the RCU grace-period handling know that the CPU is back to
882 void rcu_nmi_exit(void)
884 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
887 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
888 * (We are exiting an NMI handler, so RCU better be paying attention
891 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
892 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
895 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
896 * leave it in non-RCU-idle state.
898 if (rdtp
->dynticks_nmi_nesting
!= 1) {
899 rdtp
->dynticks_nmi_nesting
-= 2;
903 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
904 rdtp
->dynticks_nmi_nesting
= 0;
905 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
906 smp_mb__before_atomic(); /* See above. */
907 atomic_inc(&rdtp
->dynticks
);
908 smp_mb__after_atomic(); /* Force delay to next write. */
909 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
913 * __rcu_is_watching - are RCU read-side critical sections safe?
915 * Return true if RCU is watching the running CPU, which means that
916 * this CPU can safely enter RCU read-side critical sections. Unlike
917 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
918 * least disabled preemption.
920 bool notrace
__rcu_is_watching(void)
922 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
926 * rcu_is_watching - see if RCU thinks that the current CPU is idle
928 * If the current CPU is in its idle loop and is neither in an interrupt
929 * or NMI handler, return true.
931 bool notrace
rcu_is_watching(void)
936 ret
= __rcu_is_watching();
940 EXPORT_SYMBOL_GPL(rcu_is_watching
);
942 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
945 * Is the current CPU online? Disable preemption to avoid false positives
946 * that could otherwise happen due to the current CPU number being sampled,
947 * this task being preempted, its old CPU being taken offline, resuming
948 * on some other CPU, then determining that its old CPU is now offline.
949 * It is OK to use RCU on an offline processor during initial boot, hence
950 * the check for rcu_scheduler_fully_active. Note also that it is OK
951 * for a CPU coming online to use RCU for one jiffy prior to marking itself
952 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
953 * offline to continue to use RCU for one jiffy after marking itself
954 * offline in the cpu_online_mask. This leniency is necessary given the
955 * non-atomic nature of the online and offline processing, for example,
956 * the fact that a CPU enters the scheduler after completing the CPU_DYING
959 * This is also why RCU internally marks CPUs online during the
960 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
962 * Disable checking if in an NMI handler because we cannot safely report
963 * errors from NMI handlers anyway.
965 bool rcu_lockdep_current_cpu_online(void)
967 struct rcu_data
*rdp
;
968 struct rcu_node
*rnp
;
974 rdp
= this_cpu_ptr(&rcu_sched_data
);
976 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
977 !rcu_scheduler_fully_active
;
981 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
983 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
986 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
988 * If the current CPU is idle or running at a first-level (not nested)
989 * interrupt from idle, return true. The caller must have at least
990 * disabled preemption.
992 static int rcu_is_cpu_rrupt_from_idle(void)
994 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
998 * Snapshot the specified CPU's dynticks counter so that we can later
999 * credit them with an implicit quiescent state. Return 1 if this CPU
1000 * is in dynticks idle mode, which is an extended quiescent state.
1002 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1003 bool *isidle
, unsigned long *maxj
)
1005 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1006 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1007 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1008 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1011 if (ULONG_CMP_LT(ACCESS_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1012 rdp
->mynode
->gpnum
))
1013 ACCESS_ONCE(rdp
->gpwrap
) = true;
1019 * Return true if the specified CPU has passed through a quiescent
1020 * state by virtue of being in or having passed through an dynticks
1021 * idle state since the last call to dyntick_save_progress_counter()
1022 * for this same CPU, or by virtue of having been offline.
1024 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1025 bool *isidle
, unsigned long *maxj
)
1031 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1032 snap
= (unsigned int)rdp
->dynticks_snap
;
1035 * If the CPU passed through or entered a dynticks idle phase with
1036 * no active irq/NMI handlers, then we can safely pretend that the CPU
1037 * already acknowledged the request to pass through a quiescent
1038 * state. Either way, that CPU cannot possibly be in an RCU
1039 * read-side critical section that started before the beginning
1040 * of the current RCU grace period.
1042 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1043 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1044 rdp
->dynticks_fqs
++;
1049 * Check for the CPU being offline, but only if the grace period
1050 * is old enough. We don't need to worry about the CPU changing
1051 * state: If we see it offline even once, it has been through a
1054 * The reason for insisting that the grace period be at least
1055 * one jiffy old is that CPUs that are not quite online and that
1056 * have just gone offline can still execute RCU read-side critical
1059 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1060 return 0; /* Grace period is not old enough. */
1062 if (cpu_is_offline(rdp
->cpu
)) {
1063 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1069 * A CPU running for an extended time within the kernel can
1070 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1071 * even context-switching back and forth between a pair of
1072 * in-kernel CPU-bound tasks cannot advance grace periods.
1073 * So if the grace period is old enough, make the CPU pay attention.
1074 * Note that the unsynchronized assignments to the per-CPU
1075 * rcu_sched_qs_mask variable are safe. Yes, setting of
1076 * bits can be lost, but they will be set again on the next
1077 * force-quiescent-state pass. So lost bit sets do not result
1078 * in incorrect behavior, merely in a grace period lasting
1079 * a few jiffies longer than it might otherwise. Because
1080 * there are at most four threads involved, and because the
1081 * updates are only once every few jiffies, the probability of
1082 * lossage (and thus of slight grace-period extension) is
1085 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1086 * is set too high, we override with half of the RCU CPU stall
1089 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1090 if (ULONG_CMP_GE(jiffies
,
1091 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1092 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1093 if (!(ACCESS_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1094 ACCESS_ONCE(rdp
->cond_resched_completed
) =
1095 ACCESS_ONCE(rdp
->mynode
->completed
);
1096 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1097 ACCESS_ONCE(*rcrmp
) =
1098 ACCESS_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
;
1099 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1100 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1101 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1102 /* Time to beat on that CPU again! */
1103 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1104 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1111 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1113 unsigned long j
= jiffies
;
1117 smp_wmb(); /* Record start time before stall time. */
1118 j1
= rcu_jiffies_till_stall_check();
1119 ACCESS_ONCE(rsp
->jiffies_stall
) = j
+ j1
;
1120 rsp
->jiffies_resched
= j
+ j1
/ 2;
1121 rsp
->n_force_qs_gpstart
= ACCESS_ONCE(rsp
->n_force_qs
);
1125 * Complain about starvation of grace-period kthread.
1127 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1133 gpa
= ACCESS_ONCE(rsp
->gp_activity
);
1134 if (j
- gpa
> 2 * HZ
)
1135 pr_err("%s kthread starved for %ld jiffies!\n",
1136 rsp
->name
, j
- gpa
);
1140 * Dump stacks of all tasks running on stalled CPUs.
1142 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1145 unsigned long flags
;
1146 struct rcu_node
*rnp
;
1148 rcu_for_each_leaf_node(rsp
, rnp
) {
1149 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1150 if (rnp
->qsmask
!= 0) {
1151 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1152 if (rnp
->qsmask
& (1UL << cpu
))
1153 dump_cpu_task(rnp
->grplo
+ cpu
);
1155 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1159 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1163 unsigned long flags
;
1167 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1170 /* Only let one CPU complain about others per time interval. */
1172 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1173 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
1174 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1175 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1178 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
1179 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1182 * OK, time to rat on our buddy...
1183 * See Documentation/RCU/stallwarn.txt for info on how to debug
1184 * RCU CPU stall warnings.
1186 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1188 print_cpu_stall_info_begin();
1189 rcu_for_each_leaf_node(rsp
, rnp
) {
1190 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1191 ndetected
+= rcu_print_task_stall(rnp
);
1192 if (rnp
->qsmask
!= 0) {
1193 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1194 if (rnp
->qsmask
& (1UL << cpu
)) {
1195 print_cpu_stall_info(rsp
,
1200 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1203 print_cpu_stall_info_end();
1204 for_each_possible_cpu(cpu
)
1205 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1206 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1207 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1208 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1210 rcu_dump_cpu_stacks(rsp
);
1212 if (ACCESS_ONCE(rsp
->gpnum
) != gpnum
||
1213 ACCESS_ONCE(rsp
->completed
) == gpnum
) {
1214 pr_err("INFO: Stall ended before state dump start\n");
1217 gpa
= ACCESS_ONCE(rsp
->gp_activity
);
1218 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1219 rsp
->name
, j
- gpa
, j
, gpa
,
1220 jiffies_till_next_fqs
,
1221 rcu_get_root(rsp
)->qsmask
);
1222 /* In this case, the current CPU might be at fault. */
1223 sched_show_task(current
);
1227 /* Complain about tasks blocking the grace period. */
1228 rcu_print_detail_task_stall(rsp
);
1230 rcu_check_gp_kthread_starvation(rsp
);
1232 force_quiescent_state(rsp
); /* Kick them all. */
1235 static void print_cpu_stall(struct rcu_state
*rsp
)
1238 unsigned long flags
;
1239 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1243 * OK, time to rat on ourselves...
1244 * See Documentation/RCU/stallwarn.txt for info on how to debug
1245 * RCU CPU stall warnings.
1247 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1248 print_cpu_stall_info_begin();
1249 print_cpu_stall_info(rsp
, smp_processor_id());
1250 print_cpu_stall_info_end();
1251 for_each_possible_cpu(cpu
)
1252 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1253 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1254 jiffies
- rsp
->gp_start
,
1255 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1257 rcu_check_gp_kthread_starvation(rsp
);
1259 rcu_dump_cpu_stacks(rsp
);
1261 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1262 if (ULONG_CMP_GE(jiffies
, ACCESS_ONCE(rsp
->jiffies_stall
)))
1263 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+
1264 3 * rcu_jiffies_till_stall_check() + 3;
1265 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1268 * Attempt to revive the RCU machinery by forcing a context switch.
1270 * A context switch would normally allow the RCU state machine to make
1271 * progress and it could be we're stuck in kernel space without context
1272 * switches for an entirely unreasonable amount of time.
1274 resched_cpu(smp_processor_id());
1277 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1279 unsigned long completed
;
1280 unsigned long gpnum
;
1284 struct rcu_node
*rnp
;
1286 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1291 * Lots of memory barriers to reject false positives.
1293 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1294 * then rsp->gp_start, and finally rsp->completed. These values
1295 * are updated in the opposite order with memory barriers (or
1296 * equivalent) during grace-period initialization and cleanup.
1297 * Now, a false positive can occur if we get an new value of
1298 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1299 * the memory barriers, the only way that this can happen is if one
1300 * grace period ends and another starts between these two fetches.
1301 * Detect this by comparing rsp->completed with the previous fetch
1304 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1305 * and rsp->gp_start suffice to forestall false positives.
1307 gpnum
= ACCESS_ONCE(rsp
->gpnum
);
1308 smp_rmb(); /* Pick up ->gpnum first... */
1309 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
1310 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1311 gps
= ACCESS_ONCE(rsp
->gp_start
);
1312 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1313 completed
= ACCESS_ONCE(rsp
->completed
);
1314 if (ULONG_CMP_GE(completed
, gpnum
) ||
1315 ULONG_CMP_LT(j
, js
) ||
1316 ULONG_CMP_GE(gps
, js
))
1317 return; /* No stall or GP completed since entering function. */
1319 if (rcu_gp_in_progress(rsp
) &&
1320 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1322 /* We haven't checked in, so go dump stack. */
1323 print_cpu_stall(rsp
);
1325 } else if (rcu_gp_in_progress(rsp
) &&
1326 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1328 /* They had a few time units to dump stack, so complain. */
1329 print_other_cpu_stall(rsp
, gpnum
);
1334 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1336 * Set the stall-warning timeout way off into the future, thus preventing
1337 * any RCU CPU stall-warning messages from appearing in the current set of
1338 * RCU grace periods.
1340 * The caller must disable hard irqs.
1342 void rcu_cpu_stall_reset(void)
1344 struct rcu_state
*rsp
;
1346 for_each_rcu_flavor(rsp
)
1347 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ ULONG_MAX
/ 2;
1351 * Initialize the specified rcu_data structure's default callback list
1352 * to empty. The default callback list is the one that is not used by
1353 * no-callbacks CPUs.
1355 static void init_default_callback_list(struct rcu_data
*rdp
)
1359 rdp
->nxtlist
= NULL
;
1360 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1361 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1365 * Initialize the specified rcu_data structure's callback list to empty.
1367 static void init_callback_list(struct rcu_data
*rdp
)
1369 if (init_nocb_callback_list(rdp
))
1371 init_default_callback_list(rdp
);
1375 * Determine the value that ->completed will have at the end of the
1376 * next subsequent grace period. This is used to tag callbacks so that
1377 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1378 * been dyntick-idle for an extended period with callbacks under the
1379 * influence of RCU_FAST_NO_HZ.
1381 * The caller must hold rnp->lock with interrupts disabled.
1383 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1384 struct rcu_node
*rnp
)
1387 * If RCU is idle, we just wait for the next grace period.
1388 * But we can only be sure that RCU is idle if we are looking
1389 * at the root rcu_node structure -- otherwise, a new grace
1390 * period might have started, but just not yet gotten around
1391 * to initializing the current non-root rcu_node structure.
1393 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1394 return rnp
->completed
+ 1;
1397 * Otherwise, wait for a possible partial grace period and
1398 * then the subsequent full grace period.
1400 return rnp
->completed
+ 2;
1404 * Trace-event helper function for rcu_start_future_gp() and
1405 * rcu_nocb_wait_gp().
1407 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1408 unsigned long c
, const char *s
)
1410 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1411 rnp
->completed
, c
, rnp
->level
,
1412 rnp
->grplo
, rnp
->grphi
, s
);
1416 * Start some future grace period, as needed to handle newly arrived
1417 * callbacks. The required future grace periods are recorded in each
1418 * rcu_node structure's ->need_future_gp field. Returns true if there
1419 * is reason to awaken the grace-period kthread.
1421 * The caller must hold the specified rcu_node structure's ->lock.
1423 static bool __maybe_unused
1424 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1425 unsigned long *c_out
)
1430 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1433 * Pick up grace-period number for new callbacks. If this
1434 * grace period is already marked as needed, return to the caller.
1436 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1437 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1438 if (rnp
->need_future_gp
[c
& 0x1]) {
1439 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1444 * If either this rcu_node structure or the root rcu_node structure
1445 * believe that a grace period is in progress, then we must wait
1446 * for the one following, which is in "c". Because our request
1447 * will be noticed at the end of the current grace period, we don't
1448 * need to explicitly start one. We only do the lockless check
1449 * of rnp_root's fields if the current rcu_node structure thinks
1450 * there is no grace period in flight, and because we hold rnp->lock,
1451 * the only possible change is when rnp_root's two fields are
1452 * equal, in which case rnp_root->gpnum might be concurrently
1453 * incremented. But that is OK, as it will just result in our
1454 * doing some extra useless work.
1456 if (rnp
->gpnum
!= rnp
->completed
||
1457 ACCESS_ONCE(rnp_root
->gpnum
) != ACCESS_ONCE(rnp_root
->completed
)) {
1458 rnp
->need_future_gp
[c
& 0x1]++;
1459 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1464 * There might be no grace period in progress. If we don't already
1465 * hold it, acquire the root rcu_node structure's lock in order to
1466 * start one (if needed).
1468 if (rnp
!= rnp_root
) {
1469 raw_spin_lock(&rnp_root
->lock
);
1470 smp_mb__after_unlock_lock();
1474 * Get a new grace-period number. If there really is no grace
1475 * period in progress, it will be smaller than the one we obtained
1476 * earlier. Adjust callbacks as needed. Note that even no-CBs
1477 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1479 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1480 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1481 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1482 rdp
->nxtcompleted
[i
] = c
;
1485 * If the needed for the required grace period is already
1486 * recorded, trace and leave.
1488 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1489 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1493 /* Record the need for the future grace period. */
1494 rnp_root
->need_future_gp
[c
& 0x1]++;
1496 /* If a grace period is not already in progress, start one. */
1497 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1498 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1500 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1501 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1504 if (rnp
!= rnp_root
)
1505 raw_spin_unlock(&rnp_root
->lock
);
1513 * Clean up any old requests for the just-ended grace period. Also return
1514 * whether any additional grace periods have been requested. Also invoke
1515 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1516 * waiting for this grace period to complete.
1518 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1520 int c
= rnp
->completed
;
1522 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1524 rcu_nocb_gp_cleanup(rsp
, rnp
);
1525 rnp
->need_future_gp
[c
& 0x1] = 0;
1526 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1527 trace_rcu_future_gp(rnp
, rdp
, c
,
1528 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1533 * Awaken the grace-period kthread for the specified flavor of RCU.
1534 * Don't do a self-awaken, and don't bother awakening when there is
1535 * nothing for the grace-period kthread to do (as in several CPUs
1536 * raced to awaken, and we lost), and finally don't try to awaken
1537 * a kthread that has not yet been created.
1539 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1541 if (current
== rsp
->gp_kthread
||
1542 !ACCESS_ONCE(rsp
->gp_flags
) ||
1545 wake_up(&rsp
->gp_wq
);
1549 * If there is room, assign a ->completed number to any callbacks on
1550 * this CPU that have not already been assigned. Also accelerate any
1551 * callbacks that were previously assigned a ->completed number that has
1552 * since proven to be too conservative, which can happen if callbacks get
1553 * assigned a ->completed number while RCU is idle, but with reference to
1554 * a non-root rcu_node structure. This function is idempotent, so it does
1555 * not hurt to call it repeatedly. Returns an flag saying that we should
1556 * awaken the RCU grace-period kthread.
1558 * The caller must hold rnp->lock with interrupts disabled.
1560 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1561 struct rcu_data
*rdp
)
1567 /* If the CPU has no callbacks, nothing to do. */
1568 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1572 * Starting from the sublist containing the callbacks most
1573 * recently assigned a ->completed number and working down, find the
1574 * first sublist that is not assignable to an upcoming grace period.
1575 * Such a sublist has something in it (first two tests) and has
1576 * a ->completed number assigned that will complete sooner than
1577 * the ->completed number for newly arrived callbacks (last test).
1579 * The key point is that any later sublist can be assigned the
1580 * same ->completed number as the newly arrived callbacks, which
1581 * means that the callbacks in any of these later sublist can be
1582 * grouped into a single sublist, whether or not they have already
1583 * been assigned a ->completed number.
1585 c
= rcu_cbs_completed(rsp
, rnp
);
1586 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1587 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1588 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1592 * If there are no sublist for unassigned callbacks, leave.
1593 * At the same time, advance "i" one sublist, so that "i" will
1594 * index into the sublist where all the remaining callbacks should
1597 if (++i
>= RCU_NEXT_TAIL
)
1601 * Assign all subsequent callbacks' ->completed number to the next
1602 * full grace period and group them all in the sublist initially
1605 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1606 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1607 rdp
->nxtcompleted
[i
] = c
;
1609 /* Record any needed additional grace periods. */
1610 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1612 /* Trace depending on how much we were able to accelerate. */
1613 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1614 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1616 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1621 * Move any callbacks whose grace period has completed to the
1622 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1623 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1624 * sublist. This function is idempotent, so it does not hurt to
1625 * invoke it repeatedly. As long as it is not invoked -too- often...
1626 * Returns true if the RCU grace-period kthread needs to be awakened.
1628 * The caller must hold rnp->lock with interrupts disabled.
1630 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1631 struct rcu_data
*rdp
)
1635 /* If the CPU has no callbacks, nothing to do. */
1636 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1640 * Find all callbacks whose ->completed numbers indicate that they
1641 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1643 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1644 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1646 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1648 /* Clean up any sublist tail pointers that were misordered above. */
1649 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1650 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1652 /* Copy down callbacks to fill in empty sublists. */
1653 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1654 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1656 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1657 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1660 /* Classify any remaining callbacks. */
1661 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1665 * Update CPU-local rcu_data state to record the beginnings and ends of
1666 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1667 * structure corresponding to the current CPU, and must have irqs disabled.
1668 * Returns true if the grace-period kthread needs to be awakened.
1670 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1671 struct rcu_data
*rdp
)
1675 /* Handle the ends of any preceding grace periods first. */
1676 if (rdp
->completed
== rnp
->completed
&&
1677 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1679 /* No grace period end, so just accelerate recent callbacks. */
1680 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1684 /* Advance callbacks. */
1685 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1687 /* Remember that we saw this grace-period completion. */
1688 rdp
->completed
= rnp
->completed
;
1689 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1692 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1694 * If the current grace period is waiting for this CPU,
1695 * set up to detect a quiescent state, otherwise don't
1696 * go looking for one.
1698 rdp
->gpnum
= rnp
->gpnum
;
1699 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1700 rdp
->passed_quiesce
= 0;
1701 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1702 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1703 zero_cpu_stall_ticks(rdp
);
1704 ACCESS_ONCE(rdp
->gpwrap
) = false;
1709 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1711 unsigned long flags
;
1713 struct rcu_node
*rnp
;
1715 local_irq_save(flags
);
1717 if ((rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) &&
1718 rdp
->completed
== ACCESS_ONCE(rnp
->completed
) &&
1719 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1720 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1721 local_irq_restore(flags
);
1724 smp_mb__after_unlock_lock();
1725 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1726 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1728 rcu_gp_kthread_wake(rsp
);
1732 * Initialize a new grace period. Return 0 if no grace period required.
1734 static int rcu_gp_init(struct rcu_state
*rsp
)
1736 unsigned long oldmask
;
1737 struct rcu_data
*rdp
;
1738 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1740 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1741 raw_spin_lock_irq(&rnp
->lock
);
1742 smp_mb__after_unlock_lock();
1743 if (!ACCESS_ONCE(rsp
->gp_flags
)) {
1744 /* Spurious wakeup, tell caller to go back to sleep. */
1745 raw_spin_unlock_irq(&rnp
->lock
);
1748 ACCESS_ONCE(rsp
->gp_flags
) = 0; /* Clear all flags: New grace period. */
1750 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1752 * Grace period already in progress, don't start another.
1753 * Not supposed to be able to happen.
1755 raw_spin_unlock_irq(&rnp
->lock
);
1759 /* Advance to a new grace period and initialize state. */
1760 record_gp_stall_check_time(rsp
);
1761 /* Record GP times before starting GP, hence smp_store_release(). */
1762 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1763 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1764 raw_spin_unlock_irq(&rnp
->lock
);
1767 * Apply per-leaf buffered online and offline operations to the
1768 * rcu_node tree. Note that this new grace period need not wait
1769 * for subsequent online CPUs, and that quiescent-state forcing
1770 * will handle subsequent offline CPUs.
1772 rcu_for_each_leaf_node(rsp
, rnp
) {
1773 raw_spin_lock_irq(&rnp
->lock
);
1774 smp_mb__after_unlock_lock();
1775 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1776 !rnp
->wait_blkd_tasks
) {
1777 /* Nothing to do on this leaf rcu_node structure. */
1778 raw_spin_unlock_irq(&rnp
->lock
);
1782 /* Record old state, apply changes to ->qsmaskinit field. */
1783 oldmask
= rnp
->qsmaskinit
;
1784 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1786 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1787 if (!oldmask
!= !rnp
->qsmaskinit
) {
1788 if (!oldmask
) /* First online CPU for this rcu_node. */
1789 rcu_init_new_rnp(rnp
);
1790 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1791 rnp
->wait_blkd_tasks
= true;
1792 else /* Last offline CPU and can propagate. */
1793 rcu_cleanup_dead_rnp(rnp
);
1797 * If all waited-on tasks from prior grace period are
1798 * done, and if all this rcu_node structure's CPUs are
1799 * still offline, propagate up the rcu_node tree and
1800 * clear ->wait_blkd_tasks. Otherwise, if one of this
1801 * rcu_node structure's CPUs has since come back online,
1802 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1803 * checks for this, so just call it unconditionally).
1805 if (rnp
->wait_blkd_tasks
&&
1806 (!rcu_preempt_has_tasks(rnp
) ||
1808 rnp
->wait_blkd_tasks
= false;
1809 rcu_cleanup_dead_rnp(rnp
);
1812 raw_spin_unlock_irq(&rnp
->lock
);
1816 * Set the quiescent-state-needed bits in all the rcu_node
1817 * structures for all currently online CPUs in breadth-first order,
1818 * starting from the root rcu_node structure, relying on the layout
1819 * of the tree within the rsp->node[] array. Note that other CPUs
1820 * will access only the leaves of the hierarchy, thus seeing that no
1821 * grace period is in progress, at least until the corresponding
1822 * leaf node has been initialized. In addition, we have excluded
1823 * CPU-hotplug operations.
1825 * The grace period cannot complete until the initialization
1826 * process finishes, because this kthread handles both.
1828 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1829 raw_spin_lock_irq(&rnp
->lock
);
1830 smp_mb__after_unlock_lock();
1831 rdp
= this_cpu_ptr(rsp
->rda
);
1832 rcu_preempt_check_blocked_tasks(rnp
);
1833 rnp
->qsmask
= rnp
->qsmaskinit
;
1834 ACCESS_ONCE(rnp
->gpnum
) = rsp
->gpnum
;
1835 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1836 ACCESS_ONCE(rnp
->completed
) = rsp
->completed
;
1837 if (rnp
== rdp
->mynode
)
1838 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1839 rcu_preempt_boost_start_gp(rnp
);
1840 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1841 rnp
->level
, rnp
->grplo
,
1842 rnp
->grphi
, rnp
->qsmask
);
1843 raw_spin_unlock_irq(&rnp
->lock
);
1844 cond_resched_rcu_qs();
1845 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1846 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_SLOW_INIT
) &&
1847 gp_init_delay
> 0 &&
1848 !(rsp
->gpnum
% (rcu_num_nodes
* 10)))
1849 schedule_timeout_uninterruptible(gp_init_delay
);
1856 * Do one round of quiescent-state forcing.
1858 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1860 int fqs_state
= fqs_state_in
;
1861 bool isidle
= false;
1863 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1865 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1867 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1868 /* Collect dyntick-idle snapshots. */
1869 if (is_sysidle_rcu_state(rsp
)) {
1871 maxj
= jiffies
- ULONG_MAX
/ 4;
1873 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1875 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1876 fqs_state
= RCU_FORCE_QS
;
1878 /* Handle dyntick-idle and offline CPUs. */
1880 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1882 /* Clear flag to prevent immediate re-entry. */
1883 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1884 raw_spin_lock_irq(&rnp
->lock
);
1885 smp_mb__after_unlock_lock();
1886 ACCESS_ONCE(rsp
->gp_flags
) =
1887 ACCESS_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
;
1888 raw_spin_unlock_irq(&rnp
->lock
);
1894 * Clean up after the old grace period.
1896 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1898 unsigned long gp_duration
;
1899 bool needgp
= false;
1901 struct rcu_data
*rdp
;
1902 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1904 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1905 raw_spin_lock_irq(&rnp
->lock
);
1906 smp_mb__after_unlock_lock();
1907 gp_duration
= jiffies
- rsp
->gp_start
;
1908 if (gp_duration
> rsp
->gp_max
)
1909 rsp
->gp_max
= gp_duration
;
1912 * We know the grace period is complete, but to everyone else
1913 * it appears to still be ongoing. But it is also the case
1914 * that to everyone else it looks like there is nothing that
1915 * they can do to advance the grace period. It is therefore
1916 * safe for us to drop the lock in order to mark the grace
1917 * period as completed in all of the rcu_node structures.
1919 raw_spin_unlock_irq(&rnp
->lock
);
1922 * Propagate new ->completed value to rcu_node structures so
1923 * that other CPUs don't have to wait until the start of the next
1924 * grace period to process their callbacks. This also avoids
1925 * some nasty RCU grace-period initialization races by forcing
1926 * the end of the current grace period to be completely recorded in
1927 * all of the rcu_node structures before the beginning of the next
1928 * grace period is recorded in any of the rcu_node structures.
1930 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1931 raw_spin_lock_irq(&rnp
->lock
);
1932 smp_mb__after_unlock_lock();
1933 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
1934 WARN_ON_ONCE(rnp
->qsmask
);
1935 ACCESS_ONCE(rnp
->completed
) = rsp
->gpnum
;
1936 rdp
= this_cpu_ptr(rsp
->rda
);
1937 if (rnp
== rdp
->mynode
)
1938 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1939 /* smp_mb() provided by prior unlock-lock pair. */
1940 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1941 raw_spin_unlock_irq(&rnp
->lock
);
1942 cond_resched_rcu_qs();
1943 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1945 rnp
= rcu_get_root(rsp
);
1946 raw_spin_lock_irq(&rnp
->lock
);
1947 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1948 rcu_nocb_gp_set(rnp
, nocb
);
1950 /* Declare grace period done. */
1951 ACCESS_ONCE(rsp
->completed
) = rsp
->gpnum
;
1952 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1953 rsp
->fqs_state
= RCU_GP_IDLE
;
1954 rdp
= this_cpu_ptr(rsp
->rda
);
1955 /* Advance CBs to reduce false positives below. */
1956 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1957 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1958 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
1959 trace_rcu_grace_period(rsp
->name
,
1960 ACCESS_ONCE(rsp
->gpnum
),
1963 raw_spin_unlock_irq(&rnp
->lock
);
1967 * Body of kthread that handles grace periods.
1969 static int __noreturn
rcu_gp_kthread(void *arg
)
1975 struct rcu_state
*rsp
= arg
;
1976 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1978 rcu_bind_gp_kthread();
1981 /* Handle grace-period start. */
1983 trace_rcu_grace_period(rsp
->name
,
1984 ACCESS_ONCE(rsp
->gpnum
),
1986 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
1987 wait_event_interruptible(rsp
->gp_wq
,
1988 ACCESS_ONCE(rsp
->gp_flags
) &
1990 /* Locking provides needed memory barrier. */
1991 if (rcu_gp_init(rsp
))
1993 cond_resched_rcu_qs();
1994 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1995 WARN_ON(signal_pending(current
));
1996 trace_rcu_grace_period(rsp
->name
,
1997 ACCESS_ONCE(rsp
->gpnum
),
2001 /* Handle quiescent-state forcing. */
2002 fqs_state
= RCU_SAVE_DYNTICK
;
2003 j
= jiffies_till_first_fqs
;
2006 jiffies_till_first_fqs
= HZ
;
2011 rsp
->jiffies_force_qs
= jiffies
+ j
;
2012 trace_rcu_grace_period(rsp
->name
,
2013 ACCESS_ONCE(rsp
->gpnum
),
2015 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2016 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
2017 ((gf
= ACCESS_ONCE(rsp
->gp_flags
)) &
2019 (!ACCESS_ONCE(rnp
->qsmask
) &&
2020 !rcu_preempt_blocked_readers_cgp(rnp
)),
2022 /* Locking provides needed memory barriers. */
2023 /* If grace period done, leave loop. */
2024 if (!ACCESS_ONCE(rnp
->qsmask
) &&
2025 !rcu_preempt_blocked_readers_cgp(rnp
))
2027 /* If time for quiescent-state forcing, do it. */
2028 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2029 (gf
& RCU_GP_FLAG_FQS
)) {
2030 trace_rcu_grace_period(rsp
->name
,
2031 ACCESS_ONCE(rsp
->gpnum
),
2033 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
2034 trace_rcu_grace_period(rsp
->name
,
2035 ACCESS_ONCE(rsp
->gpnum
),
2037 cond_resched_rcu_qs();
2038 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
2040 /* Deal with stray signal. */
2041 cond_resched_rcu_qs();
2042 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
2043 WARN_ON(signal_pending(current
));
2044 trace_rcu_grace_period(rsp
->name
,
2045 ACCESS_ONCE(rsp
->gpnum
),
2048 j
= jiffies_till_next_fqs
;
2051 jiffies_till_next_fqs
= HZ
;
2054 jiffies_till_next_fqs
= 1;
2058 /* Handle grace-period end. */
2059 rcu_gp_cleanup(rsp
);
2064 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2065 * in preparation for detecting the next grace period. The caller must hold
2066 * the root node's ->lock and hard irqs must be disabled.
2068 * Note that it is legal for a dying CPU (which is marked as offline) to
2069 * invoke this function. This can happen when the dying CPU reports its
2072 * Returns true if the grace-period kthread must be awakened.
2075 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2076 struct rcu_data
*rdp
)
2078 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2080 * Either we have not yet spawned the grace-period
2081 * task, this CPU does not need another grace period,
2082 * or a grace period is already in progress.
2083 * Either way, don't start a new grace period.
2087 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
2088 trace_rcu_grace_period(rsp
->name
, ACCESS_ONCE(rsp
->gpnum
),
2092 * We can't do wakeups while holding the rnp->lock, as that
2093 * could cause possible deadlocks with the rq->lock. Defer
2094 * the wakeup to our caller.
2100 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2101 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2102 * is invoked indirectly from rcu_advance_cbs(), which would result in
2103 * endless recursion -- or would do so if it wasn't for the self-deadlock
2104 * that is encountered beforehand.
2106 * Returns true if the grace-period kthread needs to be awakened.
2108 static bool rcu_start_gp(struct rcu_state
*rsp
)
2110 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2111 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2115 * If there is no grace period in progress right now, any
2116 * callbacks we have up to this point will be satisfied by the
2117 * next grace period. Also, advancing the callbacks reduces the
2118 * probability of false positives from cpu_needs_another_gp()
2119 * resulting in pointless grace periods. So, advance callbacks
2120 * then start the grace period!
2122 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2123 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2128 * Report a full set of quiescent states to the specified rcu_state
2129 * data structure. This involves cleaning up after the prior grace
2130 * period and letting rcu_start_gp() start up the next grace period
2131 * if one is needed. Note that the caller must hold rnp->lock, which
2132 * is released before return.
2134 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2135 __releases(rcu_get_root(rsp
)->lock
)
2137 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2138 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2139 rcu_gp_kthread_wake(rsp
);
2143 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2144 * Allows quiescent states for a group of CPUs to be reported at one go
2145 * to the specified rcu_node structure, though all the CPUs in the group
2146 * must be represented by the same rcu_node structure (which need not be a
2147 * leaf rcu_node structure, though it often will be). The gps parameter
2148 * is the grace-period snapshot, which means that the quiescent states
2149 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2150 * must be held upon entry, and it is released before return.
2153 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2154 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2155 __releases(rnp
->lock
)
2157 unsigned long oldmask
= 0;
2158 struct rcu_node
*rnp_c
;
2160 /* Walk up the rcu_node hierarchy. */
2162 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2165 * Our bit has already been cleared, or the
2166 * relevant grace period is already over, so done.
2168 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2171 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2172 rnp
->qsmask
&= ~mask
;
2173 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2174 mask
, rnp
->qsmask
, rnp
->level
,
2175 rnp
->grplo
, rnp
->grphi
,
2177 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2179 /* Other bits still set at this level, so done. */
2180 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2183 mask
= rnp
->grpmask
;
2184 if (rnp
->parent
== NULL
) {
2186 /* No more levels. Exit loop holding root lock. */
2190 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2193 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2194 smp_mb__after_unlock_lock();
2195 oldmask
= rnp_c
->qsmask
;
2199 * Get here if we are the last CPU to pass through a quiescent
2200 * state for this grace period. Invoke rcu_report_qs_rsp()
2201 * to clean up and start the next grace period if one is needed.
2203 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2207 * Record a quiescent state for all tasks that were previously queued
2208 * on the specified rcu_node structure and that were blocking the current
2209 * RCU grace period. The caller must hold the specified rnp->lock with
2210 * irqs disabled, and this lock is released upon return, but irqs remain
2213 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2214 struct rcu_node
*rnp
, unsigned long flags
)
2215 __releases(rnp
->lock
)
2219 struct rcu_node
*rnp_p
;
2221 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2222 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2223 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2224 return; /* Still need more quiescent states! */
2227 rnp_p
= rnp
->parent
;
2228 if (rnp_p
== NULL
) {
2230 * Only one rcu_node structure in the tree, so don't
2231 * try to report up to its nonexistent parent!
2233 rcu_report_qs_rsp(rsp
, flags
);
2237 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2239 mask
= rnp
->grpmask
;
2240 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2241 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
2242 smp_mb__after_unlock_lock();
2243 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2247 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2248 * structure. This must be either called from the specified CPU, or
2249 * called when the specified CPU is known to be offline (and when it is
2250 * also known that no other CPU is concurrently trying to help the offline
2251 * CPU). The lastcomp argument is used to make sure we are still in the
2252 * grace period of interest. We don't want to end the current grace period
2253 * based on quiescent states detected in an earlier grace period!
2256 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2258 unsigned long flags
;
2261 struct rcu_node
*rnp
;
2264 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2265 smp_mb__after_unlock_lock();
2266 if ((rdp
->passed_quiesce
== 0 &&
2267 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2268 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2272 * The grace period in which this quiescent state was
2273 * recorded has ended, so don't report it upwards.
2274 * We will instead need a new quiescent state that lies
2275 * within the current grace period.
2277 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2278 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2279 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2282 mask
= rdp
->grpmask
;
2283 if ((rnp
->qsmask
& mask
) == 0) {
2284 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2286 rdp
->qs_pending
= 0;
2289 * This GP can't end until cpu checks in, so all of our
2290 * callbacks can be processed during the next GP.
2292 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2294 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2295 /* ^^^ Released rnp->lock */
2297 rcu_gp_kthread_wake(rsp
);
2302 * Check to see if there is a new grace period of which this CPU
2303 * is not yet aware, and if so, set up local rcu_data state for it.
2304 * Otherwise, see if this CPU has just passed through its first
2305 * quiescent state for this grace period, and record that fact if so.
2308 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2310 /* Check for grace-period ends and beginnings. */
2311 note_gp_changes(rsp
, rdp
);
2314 * Does this CPU still need to do its part for current grace period?
2315 * If no, return and let the other CPUs do their part as well.
2317 if (!rdp
->qs_pending
)
2321 * Was there a quiescent state since the beginning of the grace
2322 * period? If no, then exit and wait for the next call.
2324 if (!rdp
->passed_quiesce
&&
2325 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2329 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2332 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2335 #ifdef CONFIG_HOTPLUG_CPU
2338 * Send the specified CPU's RCU callbacks to the orphanage. The
2339 * specified CPU must be offline, and the caller must hold the
2343 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2344 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2346 /* No-CBs CPUs do not have orphanable callbacks. */
2347 if (rcu_is_nocb_cpu(rdp
->cpu
))
2351 * Orphan the callbacks. First adjust the counts. This is safe
2352 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2353 * cannot be running now. Thus no memory barrier is required.
2355 if (rdp
->nxtlist
!= NULL
) {
2356 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2357 rsp
->qlen
+= rdp
->qlen
;
2358 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2360 ACCESS_ONCE(rdp
->qlen
) = 0;
2364 * Next, move those callbacks still needing a grace period to
2365 * the orphanage, where some other CPU will pick them up.
2366 * Some of the callbacks might have gone partway through a grace
2367 * period, but that is too bad. They get to start over because we
2368 * cannot assume that grace periods are synchronized across CPUs.
2369 * We don't bother updating the ->nxttail[] array yet, instead
2370 * we just reset the whole thing later on.
2372 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2373 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2374 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2375 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2379 * Then move the ready-to-invoke callbacks to the orphanage,
2380 * where some other CPU will pick them up. These will not be
2381 * required to pass though another grace period: They are done.
2383 if (rdp
->nxtlist
!= NULL
) {
2384 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2385 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2389 * Finally, initialize the rcu_data structure's list to empty and
2390 * disallow further callbacks on this CPU.
2392 init_callback_list(rdp
);
2393 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2397 * Adopt the RCU callbacks from the specified rcu_state structure's
2398 * orphanage. The caller must hold the ->orphan_lock.
2400 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2403 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2405 /* No-CBs CPUs are handled specially. */
2406 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2409 /* Do the accounting first. */
2410 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2411 rdp
->qlen
+= rsp
->qlen
;
2412 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2413 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2414 rcu_idle_count_callbacks_posted();
2419 * We do not need a memory barrier here because the only way we
2420 * can get here if there is an rcu_barrier() in flight is if
2421 * we are the task doing the rcu_barrier().
2424 /* First adopt the ready-to-invoke callbacks. */
2425 if (rsp
->orphan_donelist
!= NULL
) {
2426 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2427 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2428 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2429 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2430 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2431 rsp
->orphan_donelist
= NULL
;
2432 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2435 /* And then adopt the callbacks that still need a grace period. */
2436 if (rsp
->orphan_nxtlist
!= NULL
) {
2437 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2438 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2439 rsp
->orphan_nxtlist
= NULL
;
2440 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2445 * Trace the fact that this CPU is going offline.
2447 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2449 RCU_TRACE(unsigned long mask
);
2450 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2451 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2453 RCU_TRACE(mask
= rdp
->grpmask
);
2454 trace_rcu_grace_period(rsp
->name
,
2455 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2460 * All CPUs for the specified rcu_node structure have gone offline,
2461 * and all tasks that were preempted within an RCU read-side critical
2462 * section while running on one of those CPUs have since exited their RCU
2463 * read-side critical section. Some other CPU is reporting this fact with
2464 * the specified rcu_node structure's ->lock held and interrupts disabled.
2465 * This function therefore goes up the tree of rcu_node structures,
2466 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2467 * the leaf rcu_node structure's ->qsmaskinit field has already been
2470 * This function does check that the specified rcu_node structure has
2471 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2472 * prematurely. That said, invoking it after the fact will cost you
2473 * a needless lock acquisition. So once it has done its work, don't
2476 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2479 struct rcu_node
*rnp
= rnp_leaf
;
2481 if (rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2484 mask
= rnp
->grpmask
;
2488 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2489 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2490 rnp
->qsmaskinit
&= ~mask
;
2491 rnp
->qsmask
&= ~mask
;
2492 if (rnp
->qsmaskinit
) {
2493 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2496 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2501 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2502 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2505 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2507 unsigned long flags
;
2509 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2510 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2512 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2513 mask
= rdp
->grpmask
;
2514 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2515 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2516 rnp
->qsmaskinitnext
&= ~mask
;
2517 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2521 * The CPU has been completely removed, and some other CPU is reporting
2522 * this fact from process context. Do the remainder of the cleanup,
2523 * including orphaning the outgoing CPU's RCU callbacks, and also
2524 * adopting them. There can only be one CPU hotplug operation at a time,
2525 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2527 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2529 unsigned long flags
;
2530 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2531 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2533 /* Adjust any no-longer-needed kthreads. */
2534 rcu_boost_kthread_setaffinity(rnp
, -1);
2536 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2537 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2538 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2539 rcu_adopt_orphan_cbs(rsp
, flags
);
2540 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2542 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2543 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2544 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2547 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2549 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2553 static void __maybe_unused
rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2557 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2561 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2565 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2568 * Invoke any RCU callbacks that have made it to the end of their grace
2569 * period. Thottle as specified by rdp->blimit.
2571 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2573 unsigned long flags
;
2574 struct rcu_head
*next
, *list
, **tail
;
2575 long bl
, count
, count_lazy
;
2578 /* If no callbacks are ready, just return. */
2579 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2580 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2581 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
2582 need_resched(), is_idle_task(current
),
2583 rcu_is_callbacks_kthread());
2588 * Extract the list of ready callbacks, disabling to prevent
2589 * races with call_rcu() from interrupt handlers.
2591 local_irq_save(flags
);
2592 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2594 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2595 list
= rdp
->nxtlist
;
2596 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2597 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2598 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2599 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2600 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2601 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2602 local_irq_restore(flags
);
2604 /* Invoke callbacks. */
2605 count
= count_lazy
= 0;
2609 debug_rcu_head_unqueue(list
);
2610 if (__rcu_reclaim(rsp
->name
, list
))
2613 /* Stop only if limit reached and CPU has something to do. */
2614 if (++count
>= bl
&&
2616 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2620 local_irq_save(flags
);
2621 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2622 is_idle_task(current
),
2623 rcu_is_callbacks_kthread());
2625 /* Update count, and requeue any remaining callbacks. */
2627 *tail
= rdp
->nxtlist
;
2628 rdp
->nxtlist
= list
;
2629 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2630 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2631 rdp
->nxttail
[i
] = tail
;
2635 smp_mb(); /* List handling before counting for rcu_barrier(). */
2636 rdp
->qlen_lazy
-= count_lazy
;
2637 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
- count
;
2638 rdp
->n_cbs_invoked
+= count
;
2640 /* Reinstate batch limit if we have worked down the excess. */
2641 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2642 rdp
->blimit
= blimit
;
2644 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2645 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2646 rdp
->qlen_last_fqs_check
= 0;
2647 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2648 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2649 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2650 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2652 local_irq_restore(flags
);
2654 /* Re-invoke RCU core processing if there are callbacks remaining. */
2655 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2660 * Check to see if this CPU is in a non-context-switch quiescent state
2661 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2662 * Also schedule RCU core processing.
2664 * This function must be called from hardirq context. It is normally
2665 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2666 * false, there is no point in invoking rcu_check_callbacks().
2668 void rcu_check_callbacks(int user
)
2670 trace_rcu_utilization(TPS("Start scheduler-tick"));
2671 increment_cpu_stall_ticks();
2672 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2675 * Get here if this CPU took its interrupt from user
2676 * mode or from the idle loop, and if this is not a
2677 * nested interrupt. In this case, the CPU is in
2678 * a quiescent state, so note it.
2680 * No memory barrier is required here because both
2681 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2682 * variables that other CPUs neither access nor modify,
2683 * at least not while the corresponding CPU is online.
2689 } else if (!in_softirq()) {
2692 * Get here if this CPU did not take its interrupt from
2693 * softirq, in other words, if it is not interrupting
2694 * a rcu_bh read-side critical section. This is an _bh
2695 * critical section, so note it.
2700 rcu_preempt_check_callbacks();
2704 rcu_note_voluntary_context_switch(current
);
2705 trace_rcu_utilization(TPS("End scheduler-tick"));
2709 * Scan the leaf rcu_node structures, processing dyntick state for any that
2710 * have not yet encountered a quiescent state, using the function specified.
2711 * Also initiate boosting for any threads blocked on the root rcu_node.
2713 * The caller must have suppressed start of new grace periods.
2715 static void force_qs_rnp(struct rcu_state
*rsp
,
2716 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2717 unsigned long *maxj
),
2718 bool *isidle
, unsigned long *maxj
)
2722 unsigned long flags
;
2724 struct rcu_node
*rnp
;
2726 rcu_for_each_leaf_node(rsp
, rnp
) {
2727 cond_resched_rcu_qs();
2729 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2730 smp_mb__after_unlock_lock();
2731 if (!rcu_gp_in_progress(rsp
)) {
2732 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2735 if (rnp
->qsmask
== 0) {
2736 if (rcu_state_p
== &rcu_sched_state
||
2737 rsp
!= rcu_state_p
||
2738 rcu_preempt_blocked_readers_cgp(rnp
)) {
2740 * No point in scanning bits because they
2741 * are all zero. But we might need to
2742 * priority-boost blocked readers.
2744 rcu_initiate_boost(rnp
, flags
);
2745 /* rcu_initiate_boost() releases rnp->lock */
2749 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2751 * Race between grace-period
2752 * initialization and task exiting RCU
2753 * read-side critical section: Report.
2755 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2756 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2762 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2763 if ((rnp
->qsmask
& bit
) != 0) {
2764 if ((rnp
->qsmaskinit
& bit
) == 0)
2765 *isidle
= false; /* Pending hotplug. */
2766 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2771 /* Idle/offline CPUs, report (releases rnp->lock. */
2772 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2774 /* Nothing to do here, so just drop the lock. */
2775 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2781 * Force quiescent states on reluctant CPUs, and also detect which
2782 * CPUs are in dyntick-idle mode.
2784 static void force_quiescent_state(struct rcu_state
*rsp
)
2786 unsigned long flags
;
2788 struct rcu_node
*rnp
;
2789 struct rcu_node
*rnp_old
= NULL
;
2791 /* Funnel through hierarchy to reduce memory contention. */
2792 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2793 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2794 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2795 !raw_spin_trylock(&rnp
->fqslock
);
2796 if (rnp_old
!= NULL
)
2797 raw_spin_unlock(&rnp_old
->fqslock
);
2799 rsp
->n_force_qs_lh
++;
2804 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2806 /* Reached the root of the rcu_node tree, acquire lock. */
2807 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2808 smp_mb__after_unlock_lock();
2809 raw_spin_unlock(&rnp_old
->fqslock
);
2810 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2811 rsp
->n_force_qs_lh
++;
2812 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2813 return; /* Someone beat us to it. */
2815 ACCESS_ONCE(rsp
->gp_flags
) =
2816 ACCESS_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
;
2817 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2818 rcu_gp_kthread_wake(rsp
);
2822 * This does the RCU core processing work for the specified rcu_state
2823 * and rcu_data structures. This may be called only from the CPU to
2824 * whom the rdp belongs.
2827 __rcu_process_callbacks(struct rcu_state
*rsp
)
2829 unsigned long flags
;
2831 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2833 WARN_ON_ONCE(rdp
->beenonline
== 0);
2835 /* Update RCU state based on any recent quiescent states. */
2836 rcu_check_quiescent_state(rsp
, rdp
);
2838 /* Does this CPU require a not-yet-started grace period? */
2839 local_irq_save(flags
);
2840 if (cpu_needs_another_gp(rsp
, rdp
)) {
2841 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2842 needwake
= rcu_start_gp(rsp
);
2843 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2845 rcu_gp_kthread_wake(rsp
);
2847 local_irq_restore(flags
);
2850 /* If there are callbacks ready, invoke them. */
2851 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2852 invoke_rcu_callbacks(rsp
, rdp
);
2854 /* Do any needed deferred wakeups of rcuo kthreads. */
2855 do_nocb_deferred_wakeup(rdp
);
2859 * Do RCU core processing for the current CPU.
2861 static void rcu_process_callbacks(struct softirq_action
*unused
)
2863 struct rcu_state
*rsp
;
2865 if (cpu_is_offline(smp_processor_id()))
2867 trace_rcu_utilization(TPS("Start RCU core"));
2868 for_each_rcu_flavor(rsp
)
2869 __rcu_process_callbacks(rsp
);
2870 trace_rcu_utilization(TPS("End RCU core"));
2874 * Schedule RCU callback invocation. If the specified type of RCU
2875 * does not support RCU priority boosting, just do a direct call,
2876 * otherwise wake up the per-CPU kernel kthread. Note that because we
2877 * are running on the current CPU with softirqs disabled, the
2878 * rcu_cpu_kthread_task cannot disappear out from under us.
2880 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2882 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2884 if (likely(!rsp
->boost
)) {
2885 rcu_do_batch(rsp
, rdp
);
2888 invoke_rcu_callbacks_kthread();
2891 static void invoke_rcu_core(void)
2893 if (cpu_online(smp_processor_id()))
2894 raise_softirq(RCU_SOFTIRQ
);
2898 * Handle any core-RCU processing required by a call_rcu() invocation.
2900 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2901 struct rcu_head
*head
, unsigned long flags
)
2906 * If called from an extended quiescent state, invoke the RCU
2907 * core in order to force a re-evaluation of RCU's idleness.
2909 if (!rcu_is_watching())
2912 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2913 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2917 * Force the grace period if too many callbacks or too long waiting.
2918 * Enforce hysteresis, and don't invoke force_quiescent_state()
2919 * if some other CPU has recently done so. Also, don't bother
2920 * invoking force_quiescent_state() if the newly enqueued callback
2921 * is the only one waiting for a grace period to complete.
2923 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2925 /* Are we ignoring a completed grace period? */
2926 note_gp_changes(rsp
, rdp
);
2928 /* Start a new grace period if one not already started. */
2929 if (!rcu_gp_in_progress(rsp
)) {
2930 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2932 raw_spin_lock(&rnp_root
->lock
);
2933 smp_mb__after_unlock_lock();
2934 needwake
= rcu_start_gp(rsp
);
2935 raw_spin_unlock(&rnp_root
->lock
);
2937 rcu_gp_kthread_wake(rsp
);
2939 /* Give the grace period a kick. */
2940 rdp
->blimit
= LONG_MAX
;
2941 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2942 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2943 force_quiescent_state(rsp
);
2944 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2945 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2951 * RCU callback function to leak a callback.
2953 static void rcu_leak_callback(struct rcu_head
*rhp
)
2958 * Helper function for call_rcu() and friends. The cpu argument will
2959 * normally be -1, indicating "currently running CPU". It may specify
2960 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2961 * is expected to specify a CPU.
2964 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2965 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2967 unsigned long flags
;
2968 struct rcu_data
*rdp
;
2970 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
2971 if (debug_rcu_head_queue(head
)) {
2972 /* Probable double call_rcu(), so leak the callback. */
2973 ACCESS_ONCE(head
->func
) = rcu_leak_callback
;
2974 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2981 * Opportunistically note grace-period endings and beginnings.
2982 * Note that we might see a beginning right after we see an
2983 * end, but never vice versa, since this CPU has to pass through
2984 * a quiescent state betweentimes.
2986 local_irq_save(flags
);
2987 rdp
= this_cpu_ptr(rsp
->rda
);
2989 /* Add the callback to our list. */
2990 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2994 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2995 if (likely(rdp
->mynode
)) {
2996 /* Post-boot, so this should be for a no-CBs CPU. */
2997 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
2998 WARN_ON_ONCE(offline
);
2999 /* Offline CPU, _call_rcu() illegal, leak callback. */
3000 local_irq_restore(flags
);
3004 * Very early boot, before rcu_init(). Initialize if needed
3005 * and then drop through to queue the callback.
3008 WARN_ON_ONCE(!rcu_is_watching());
3009 if (!likely(rdp
->nxtlist
))
3010 init_default_callback_list(rdp
);
3012 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
+ 1;
3016 rcu_idle_count_callbacks_posted();
3017 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3018 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3019 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3021 if (__is_kfree_rcu_offset((unsigned long)func
))
3022 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3023 rdp
->qlen_lazy
, rdp
->qlen
);
3025 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3027 /* Go handle any RCU core processing required. */
3028 __call_rcu_core(rsp
, rdp
, head
, flags
);
3029 local_irq_restore(flags
);
3033 * Queue an RCU-sched callback for invocation after a grace period.
3035 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3037 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3039 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3042 * Queue an RCU callback for invocation after a quicker grace period.
3044 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3046 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3048 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3051 * Queue an RCU callback for lazy invocation after a grace period.
3052 * This will likely be later named something like "call_rcu_lazy()",
3053 * but this change will require some way of tagging the lazy RCU
3054 * callbacks in the list of pending callbacks. Until then, this
3055 * function may only be called from __kfree_rcu().
3057 void kfree_call_rcu(struct rcu_head
*head
,
3058 void (*func
)(struct rcu_head
*rcu
))
3060 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3062 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3065 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3066 * any blocking grace-period wait automatically implies a grace period
3067 * if there is only one CPU online at any point time during execution
3068 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3069 * occasionally incorrectly indicate that there are multiple CPUs online
3070 * when there was in fact only one the whole time, as this just adds
3071 * some overhead: RCU still operates correctly.
3073 static inline int rcu_blocking_is_gp(void)
3077 might_sleep(); /* Check for RCU read-side critical section. */
3079 ret
= num_online_cpus() <= 1;
3085 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3087 * Control will return to the caller some time after a full rcu-sched
3088 * grace period has elapsed, in other words after all currently executing
3089 * rcu-sched read-side critical sections have completed. These read-side
3090 * critical sections are delimited by rcu_read_lock_sched() and
3091 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3092 * local_irq_disable(), and so on may be used in place of
3093 * rcu_read_lock_sched().
3095 * This means that all preempt_disable code sequences, including NMI and
3096 * non-threaded hardware-interrupt handlers, in progress on entry will
3097 * have completed before this primitive returns. However, this does not
3098 * guarantee that softirq handlers will have completed, since in some
3099 * kernels, these handlers can run in process context, and can block.
3101 * Note that this guarantee implies further memory-ordering guarantees.
3102 * On systems with more than one CPU, when synchronize_sched() returns,
3103 * each CPU is guaranteed to have executed a full memory barrier since the
3104 * end of its last RCU-sched read-side critical section whose beginning
3105 * preceded the call to synchronize_sched(). In addition, each CPU having
3106 * an RCU read-side critical section that extends beyond the return from
3107 * synchronize_sched() is guaranteed to have executed a full memory barrier
3108 * after the beginning of synchronize_sched() and before the beginning of
3109 * that RCU read-side critical section. Note that these guarantees include
3110 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3111 * that are executing in the kernel.
3113 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3114 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3115 * to have executed a full memory barrier during the execution of
3116 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3117 * again only if the system has more than one CPU).
3119 * This primitive provides the guarantees made by the (now removed)
3120 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3121 * guarantees that rcu_read_lock() sections will have completed.
3122 * In "classic RCU", these two guarantees happen to be one and
3123 * the same, but can differ in realtime RCU implementations.
3125 void synchronize_sched(void)
3127 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3128 !lock_is_held(&rcu_lock_map
) &&
3129 !lock_is_held(&rcu_sched_lock_map
),
3130 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3131 if (rcu_blocking_is_gp())
3133 if (rcu_gp_is_expedited())
3134 synchronize_sched_expedited();
3136 wait_rcu_gp(call_rcu_sched
);
3138 EXPORT_SYMBOL_GPL(synchronize_sched
);
3141 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3143 * Control will return to the caller some time after a full rcu_bh grace
3144 * period has elapsed, in other words after all currently executing rcu_bh
3145 * read-side critical sections have completed. RCU read-side critical
3146 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3147 * and may be nested.
3149 * See the description of synchronize_sched() for more detailed information
3150 * on memory ordering guarantees.
3152 void synchronize_rcu_bh(void)
3154 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3155 !lock_is_held(&rcu_lock_map
) &&
3156 !lock_is_held(&rcu_sched_lock_map
),
3157 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3158 if (rcu_blocking_is_gp())
3160 if (rcu_gp_is_expedited())
3161 synchronize_rcu_bh_expedited();
3163 wait_rcu_gp(call_rcu_bh
);
3165 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3168 * get_state_synchronize_rcu - Snapshot current RCU state
3170 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3171 * to determine whether or not a full grace period has elapsed in the
3174 unsigned long get_state_synchronize_rcu(void)
3177 * Any prior manipulation of RCU-protected data must happen
3178 * before the load from ->gpnum.
3183 * Make sure this load happens before the purportedly
3184 * time-consuming work between get_state_synchronize_rcu()
3185 * and cond_synchronize_rcu().
3187 return smp_load_acquire(&rcu_state_p
->gpnum
);
3189 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3192 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3194 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3196 * If a full RCU grace period has elapsed since the earlier call to
3197 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3198 * synchronize_rcu() to wait for a full grace period.
3200 * Yes, this function does not take counter wrap into account. But
3201 * counter wrap is harmless. If the counter wraps, we have waited for
3202 * more than 2 billion grace periods (and way more on a 64-bit system!),
3203 * so waiting for one additional grace period should be just fine.
3205 void cond_synchronize_rcu(unsigned long oldstate
)
3207 unsigned long newstate
;
3210 * Ensure that this load happens before any RCU-destructive
3211 * actions the caller might carry out after we return.
3213 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3214 if (ULONG_CMP_GE(oldstate
, newstate
))
3217 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3219 static int synchronize_sched_expedited_cpu_stop(void *data
)
3222 * There must be a full memory barrier on each affected CPU
3223 * between the time that try_stop_cpus() is called and the
3224 * time that it returns.
3226 * In the current initial implementation of cpu_stop, the
3227 * above condition is already met when the control reaches
3228 * this point and the following smp_mb() is not strictly
3229 * necessary. Do smp_mb() anyway for documentation and
3230 * robustness against future implementation changes.
3232 smp_mb(); /* See above comment block. */
3237 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3239 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3240 * approach to force the grace period to end quickly. This consumes
3241 * significant time on all CPUs and is unfriendly to real-time workloads,
3242 * so is thus not recommended for any sort of common-case code. In fact,
3243 * if you are using synchronize_sched_expedited() in a loop, please
3244 * restructure your code to batch your updates, and then use a single
3245 * synchronize_sched() instead.
3247 * This implementation can be thought of as an application of ticket
3248 * locking to RCU, with sync_sched_expedited_started and
3249 * sync_sched_expedited_done taking on the roles of the halves
3250 * of the ticket-lock word. Each task atomically increments
3251 * sync_sched_expedited_started upon entry, snapshotting the old value,
3252 * then attempts to stop all the CPUs. If this succeeds, then each
3253 * CPU will have executed a context switch, resulting in an RCU-sched
3254 * grace period. We are then done, so we use atomic_cmpxchg() to
3255 * update sync_sched_expedited_done to match our snapshot -- but
3256 * only if someone else has not already advanced past our snapshot.
3258 * On the other hand, if try_stop_cpus() fails, we check the value
3259 * of sync_sched_expedited_done. If it has advanced past our
3260 * initial snapshot, then someone else must have forced a grace period
3261 * some time after we took our snapshot. In this case, our work is
3262 * done for us, and we can simply return. Otherwise, we try again,
3263 * but keep our initial snapshot for purposes of checking for someone
3264 * doing our work for us.
3266 * If we fail too many times in a row, we fall back to synchronize_sched().
3268 void synchronize_sched_expedited(void)
3273 long firstsnap
, s
, snap
;
3275 struct rcu_state
*rsp
= &rcu_sched_state
;
3278 * If we are in danger of counter wrap, just do synchronize_sched().
3279 * By allowing sync_sched_expedited_started to advance no more than
3280 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3281 * that more than 3.5 billion CPUs would be required to force a
3282 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3283 * course be required on a 64-bit system.
3285 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3286 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3288 synchronize_sched();
3289 atomic_long_inc(&rsp
->expedited_wrap
);
3294 * Take a ticket. Note that atomic_inc_return() implies a
3295 * full memory barrier.
3297 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3299 if (!try_get_online_cpus()) {
3300 /* CPU hotplug operation in flight, fall back to normal GP. */
3301 wait_rcu_gp(call_rcu_sched
);
3302 atomic_long_inc(&rsp
->expedited_normal
);
3305 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3307 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3308 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3310 cpumask_copy(cm
, cpu_online_mask
);
3311 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3312 for_each_cpu(cpu
, cm
) {
3313 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3315 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3316 cpumask_clear_cpu(cpu
, cm
);
3318 if (cpumask_weight(cm
) == 0)
3323 * Each pass through the following loop attempts to force a
3324 * context switch on each CPU.
3326 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3327 synchronize_sched_expedited_cpu_stop
,
3330 atomic_long_inc(&rsp
->expedited_tryfail
);
3332 /* Check to see if someone else did our work for us. */
3333 s
= atomic_long_read(&rsp
->expedited_done
);
3334 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3335 /* ensure test happens before caller kfree */
3336 smp_mb__before_atomic(); /* ^^^ */
3337 atomic_long_inc(&rsp
->expedited_workdone1
);
3338 free_cpumask_var(cm
);
3342 /* No joy, try again later. Or just synchronize_sched(). */
3343 if (trycount
++ < 10) {
3344 udelay(trycount
* num_online_cpus());
3346 wait_rcu_gp(call_rcu_sched
);
3347 atomic_long_inc(&rsp
->expedited_normal
);
3348 free_cpumask_var(cm
);
3352 /* Recheck to see if someone else did our work for us. */
3353 s
= atomic_long_read(&rsp
->expedited_done
);
3354 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3355 /* ensure test happens before caller kfree */
3356 smp_mb__before_atomic(); /* ^^^ */
3357 atomic_long_inc(&rsp
->expedited_workdone2
);
3358 free_cpumask_var(cm
);
3363 * Refetching sync_sched_expedited_started allows later
3364 * callers to piggyback on our grace period. We retry
3365 * after they started, so our grace period works for them,
3366 * and they started after our first try, so their grace
3367 * period works for us.
3369 if (!try_get_online_cpus()) {
3370 /* CPU hotplug operation in flight, use normal GP. */
3371 wait_rcu_gp(call_rcu_sched
);
3372 atomic_long_inc(&rsp
->expedited_normal
);
3373 free_cpumask_var(cm
);
3376 snap
= atomic_long_read(&rsp
->expedited_start
);
3377 smp_mb(); /* ensure read is before try_stop_cpus(). */
3379 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3382 free_cpumask_var(cm
);
3385 * Everyone up to our most recent fetch is covered by our grace
3386 * period. Update the counter, but only if our work is still
3387 * relevant -- which it won't be if someone who started later
3388 * than we did already did their update.
3391 atomic_long_inc(&rsp
->expedited_done_tries
);
3392 s
= atomic_long_read(&rsp
->expedited_done
);
3393 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3394 /* ensure test happens before caller kfree */
3395 smp_mb__before_atomic(); /* ^^^ */
3396 atomic_long_inc(&rsp
->expedited_done_lost
);
3399 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3400 atomic_long_inc(&rsp
->expedited_done_exit
);
3404 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3407 * Check to see if there is any immediate RCU-related work to be done
3408 * by the current CPU, for the specified type of RCU, returning 1 if so.
3409 * The checks are in order of increasing expense: checks that can be
3410 * carried out against CPU-local state are performed first. However,
3411 * we must check for CPU stalls first, else we might not get a chance.
3413 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3415 struct rcu_node
*rnp
= rdp
->mynode
;
3417 rdp
->n_rcu_pending
++;
3419 /* Check for CPU stalls, if enabled. */
3420 check_cpu_stall(rsp
, rdp
);
3422 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3423 if (rcu_nohz_full_cpu(rsp
))
3426 /* Is the RCU core waiting for a quiescent state from this CPU? */
3427 if (rcu_scheduler_fully_active
&&
3428 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3429 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3430 rdp
->n_rp_qs_pending
++;
3431 } else if (rdp
->qs_pending
&&
3432 (rdp
->passed_quiesce
||
3433 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3434 rdp
->n_rp_report_qs
++;
3438 /* Does this CPU have callbacks ready to invoke? */
3439 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3440 rdp
->n_rp_cb_ready
++;
3444 /* Has RCU gone idle with this CPU needing another grace period? */
3445 if (cpu_needs_another_gp(rsp
, rdp
)) {
3446 rdp
->n_rp_cpu_needs_gp
++;
3450 /* Has another RCU grace period completed? */
3451 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3452 rdp
->n_rp_gp_completed
++;
3456 /* Has a new RCU grace period started? */
3457 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3458 unlikely(ACCESS_ONCE(rdp
->gpwrap
))) { /* outside lock */
3459 rdp
->n_rp_gp_started
++;
3463 /* Does this CPU need a deferred NOCB wakeup? */
3464 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3465 rdp
->n_rp_nocb_defer_wakeup
++;
3470 rdp
->n_rp_need_nothing
++;
3475 * Check to see if there is any immediate RCU-related work to be done
3476 * by the current CPU, returning 1 if so. This function is part of the
3477 * RCU implementation; it is -not- an exported member of the RCU API.
3479 static int rcu_pending(void)
3481 struct rcu_state
*rsp
;
3483 for_each_rcu_flavor(rsp
)
3484 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3490 * Return true if the specified CPU has any callback. If all_lazy is
3491 * non-NULL, store an indication of whether all callbacks are lazy.
3492 * (If there are no callbacks, all of them are deemed to be lazy.)
3494 static int __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3498 struct rcu_data
*rdp
;
3499 struct rcu_state
*rsp
;
3501 for_each_rcu_flavor(rsp
) {
3502 rdp
= this_cpu_ptr(rsp
->rda
);
3506 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3517 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3518 * the compiler is expected to optimize this away.
3520 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3521 int cpu
, unsigned long done
)
3523 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3524 atomic_read(&rsp
->barrier_cpu_count
), done
);
3528 * RCU callback function for _rcu_barrier(). If we are last, wake
3529 * up the task executing _rcu_barrier().
3531 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3533 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3534 struct rcu_state
*rsp
= rdp
->rsp
;
3536 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3537 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3538 complete(&rsp
->barrier_completion
);
3540 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3545 * Called with preemption disabled, and from cross-cpu IRQ context.
3547 static void rcu_barrier_func(void *type
)
3549 struct rcu_state
*rsp
= type
;
3550 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3552 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3553 atomic_inc(&rsp
->barrier_cpu_count
);
3554 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3558 * Orchestrate the specified type of RCU barrier, waiting for all
3559 * RCU callbacks of the specified type to complete.
3561 static void _rcu_barrier(struct rcu_state
*rsp
)
3564 struct rcu_data
*rdp
;
3565 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
3566 unsigned long snap_done
;
3568 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3570 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3571 mutex_lock(&rsp
->barrier_mutex
);
3574 * Ensure that all prior references, including to ->n_barrier_done,
3575 * are ordered before the _rcu_barrier() machinery.
3577 smp_mb(); /* See above block comment. */
3580 * Recheck ->n_barrier_done to see if others did our work for us.
3581 * This means checking ->n_barrier_done for an even-to-odd-to-even
3582 * transition. The "if" expression below therefore rounds the old
3583 * value up to the next even number and adds two before comparing.
3585 snap_done
= rsp
->n_barrier_done
;
3586 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3589 * If the value in snap is odd, we needed to wait for the current
3590 * rcu_barrier() to complete, then wait for the next one, in other
3591 * words, we need the value of snap_done to be three larger than
3592 * the value of snap. On the other hand, if the value in snap is
3593 * even, we only had to wait for the next rcu_barrier() to complete,
3594 * in other words, we need the value of snap_done to be only two
3595 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3596 * this for us (thank you, Linus!).
3598 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3599 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3600 smp_mb(); /* caller's subsequent code after above check. */
3601 mutex_unlock(&rsp
->barrier_mutex
);
3606 * Increment ->n_barrier_done to avoid duplicate work. Use
3607 * ACCESS_ONCE() to prevent the compiler from speculating
3608 * the increment to precede the early-exit check.
3610 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3611 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3612 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3613 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3616 * Initialize the count to one rather than to zero in order to
3617 * avoid a too-soon return to zero in case of a short grace period
3618 * (or preemption of this task). Exclude CPU-hotplug operations
3619 * to ensure that no offline CPU has callbacks queued.
3621 init_completion(&rsp
->barrier_completion
);
3622 atomic_set(&rsp
->barrier_cpu_count
, 1);
3626 * Force each CPU with callbacks to register a new callback.
3627 * When that callback is invoked, we will know that all of the
3628 * corresponding CPU's preceding callbacks have been invoked.
3630 for_each_possible_cpu(cpu
) {
3631 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3633 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3634 if (rcu_is_nocb_cpu(cpu
)) {
3635 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3636 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3637 rsp
->n_barrier_done
);
3639 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3640 rsp
->n_barrier_done
);
3641 smp_mb__before_atomic();
3642 atomic_inc(&rsp
->barrier_cpu_count
);
3643 __call_rcu(&rdp
->barrier_head
,
3644 rcu_barrier_callback
, rsp
, cpu
, 0);
3646 } else if (ACCESS_ONCE(rdp
->qlen
)) {
3647 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3648 rsp
->n_barrier_done
);
3649 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3651 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3652 rsp
->n_barrier_done
);
3658 * Now that we have an rcu_barrier_callback() callback on each
3659 * CPU, and thus each counted, remove the initial count.
3661 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3662 complete(&rsp
->barrier_completion
);
3664 /* Increment ->n_barrier_done to prevent duplicate work. */
3665 smp_mb(); /* Keep increment after above mechanism. */
3666 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3667 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3668 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3669 smp_mb(); /* Keep increment before caller's subsequent code. */
3671 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3672 wait_for_completion(&rsp
->barrier_completion
);
3674 /* Other rcu_barrier() invocations can now safely proceed. */
3675 mutex_unlock(&rsp
->barrier_mutex
);
3679 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3681 void rcu_barrier_bh(void)
3683 _rcu_barrier(&rcu_bh_state
);
3685 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3688 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3690 void rcu_barrier_sched(void)
3692 _rcu_barrier(&rcu_sched_state
);
3694 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3697 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3698 * first CPU in a given leaf rcu_node structure coming online. The caller
3699 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3702 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
3705 struct rcu_node
*rnp
= rnp_leaf
;
3708 mask
= rnp
->grpmask
;
3712 raw_spin_lock(&rnp
->lock
); /* Interrupts already disabled. */
3713 rnp
->qsmaskinit
|= mask
;
3714 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
3719 * Do boot-time initialization of a CPU's per-CPU RCU data.
3722 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3724 unsigned long flags
;
3725 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3726 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3728 /* Set up local state, ensuring consistent view of global state. */
3729 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3730 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3731 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3732 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3733 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3736 rcu_boot_init_nocb_percpu_data(rdp
);
3737 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3741 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3742 * offline event can be happening at a given time. Note also that we
3743 * can accept some slop in the rsp->completed access due to the fact
3744 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3747 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3749 unsigned long flags
;
3751 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3752 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3754 /* Set up local state, ensuring consistent view of global state. */
3755 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3756 rdp
->beenonline
= 1; /* We have now been online. */
3757 rdp
->qlen_last_fqs_check
= 0;
3758 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3759 rdp
->blimit
= blimit
;
3761 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3762 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3763 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3764 atomic_set(&rdp
->dynticks
->dynticks
,
3765 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3766 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3769 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3770 * propagation up the rcu_node tree will happen at the beginning
3771 * of the next grace period.
3774 mask
= rdp
->grpmask
;
3775 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3776 smp_mb__after_unlock_lock();
3777 rnp
->qsmaskinitnext
|= mask
;
3778 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
3779 rdp
->completed
= rnp
->completed
;
3780 rdp
->passed_quiesce
= false;
3781 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
3782 rdp
->qs_pending
= false;
3783 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3784 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3787 static void rcu_prepare_cpu(int cpu
)
3789 struct rcu_state
*rsp
;
3791 for_each_rcu_flavor(rsp
)
3792 rcu_init_percpu_data(cpu
, rsp
);
3796 * Handle CPU online/offline notification events.
3798 int rcu_cpu_notify(struct notifier_block
*self
,
3799 unsigned long action
, void *hcpu
)
3801 long cpu
= (long)hcpu
;
3802 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3803 struct rcu_node
*rnp
= rdp
->mynode
;
3804 struct rcu_state
*rsp
;
3807 case CPU_UP_PREPARE
:
3808 case CPU_UP_PREPARE_FROZEN
:
3809 rcu_prepare_cpu(cpu
);
3810 rcu_prepare_kthreads(cpu
);
3811 rcu_spawn_all_nocb_kthreads(cpu
);
3814 case CPU_DOWN_FAILED
:
3815 rcu_boost_kthread_setaffinity(rnp
, -1);
3817 case CPU_DOWN_PREPARE
:
3818 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3821 case CPU_DYING_FROZEN
:
3822 for_each_rcu_flavor(rsp
)
3823 rcu_cleanup_dying_cpu(rsp
);
3825 case CPU_DYING_IDLE
:
3826 for_each_rcu_flavor(rsp
) {
3827 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
3831 case CPU_DEAD_FROZEN
:
3832 case CPU_UP_CANCELED
:
3833 case CPU_UP_CANCELED_FROZEN
:
3834 for_each_rcu_flavor(rsp
) {
3835 rcu_cleanup_dead_cpu(cpu
, rsp
);
3836 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3845 static int rcu_pm_notify(struct notifier_block
*self
,
3846 unsigned long action
, void *hcpu
)
3849 case PM_HIBERNATION_PREPARE
:
3850 case PM_SUSPEND_PREPARE
:
3851 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3854 case PM_POST_HIBERNATION
:
3855 case PM_POST_SUSPEND
:
3856 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3857 rcu_unexpedite_gp();
3866 * Spawn the kthreads that handle each RCU flavor's grace periods.
3868 static int __init
rcu_spawn_gp_kthread(void)
3870 unsigned long flags
;
3871 int kthread_prio_in
= kthread_prio
;
3872 struct rcu_node
*rnp
;
3873 struct rcu_state
*rsp
;
3874 struct sched_param sp
;
3875 struct task_struct
*t
;
3877 /* Force priority into range. */
3878 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3880 else if (kthread_prio
< 0)
3882 else if (kthread_prio
> 99)
3884 if (kthread_prio
!= kthread_prio_in
)
3885 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3886 kthread_prio
, kthread_prio_in
);
3888 rcu_scheduler_fully_active
= 1;
3889 for_each_rcu_flavor(rsp
) {
3890 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3892 rnp
= rcu_get_root(rsp
);
3893 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3894 rsp
->gp_kthread
= t
;
3896 sp
.sched_priority
= kthread_prio
;
3897 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3900 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3902 rcu_spawn_nocb_kthreads();
3903 rcu_spawn_boost_kthreads();
3906 early_initcall(rcu_spawn_gp_kthread
);
3909 * This function is invoked towards the end of the scheduler's initialization
3910 * process. Before this is called, the idle task might contain
3911 * RCU read-side critical sections (during which time, this idle
3912 * task is booting the system). After this function is called, the
3913 * idle tasks are prohibited from containing RCU read-side critical
3914 * sections. This function also enables RCU lockdep checking.
3916 void rcu_scheduler_starting(void)
3918 WARN_ON(num_online_cpus() != 1);
3919 WARN_ON(nr_context_switches() > 0);
3920 rcu_scheduler_active
= 1;
3924 * Compute the per-level fanout, either using the exact fanout specified
3925 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3927 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3931 if (IS_ENABLED(CONFIG_RCU_FANOUT_EXACT
)) {
3932 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3933 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3934 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3940 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3941 ccur
= rsp
->levelcnt
[i
];
3942 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3949 * Helper function for rcu_init() that initializes one rcu_state structure.
3951 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3952 struct rcu_data __percpu
*rda
)
3954 static const char * const buf
[] = {
3958 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3959 static const char * const fqs
[] = {
3963 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3964 static u8 fl_mask
= 0x1;
3968 struct rcu_node
*rnp
;
3970 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3972 /* Silence gcc 4.8 warning about array index out of range. */
3973 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3974 panic("rcu_init_one: rcu_num_lvls overflow");
3976 /* Initialize the level-tracking arrays. */
3978 for (i
= 0; i
< rcu_num_lvls
; i
++)
3979 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3980 for (i
= 1; i
< rcu_num_lvls
; i
++)
3981 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3982 rcu_init_levelspread(rsp
);
3983 rsp
->flavor_mask
= fl_mask
;
3986 /* Initialize the elements themselves, starting from the leaves. */
3988 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3989 cpustride
*= rsp
->levelspread
[i
];
3990 rnp
= rsp
->level
[i
];
3991 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3992 raw_spin_lock_init(&rnp
->lock
);
3993 lockdep_set_class_and_name(&rnp
->lock
,
3994 &rcu_node_class
[i
], buf
[i
]);
3995 raw_spin_lock_init(&rnp
->fqslock
);
3996 lockdep_set_class_and_name(&rnp
->fqslock
,
3997 &rcu_fqs_class
[i
], fqs
[i
]);
3998 rnp
->gpnum
= rsp
->gpnum
;
3999 rnp
->completed
= rsp
->completed
;
4001 rnp
->qsmaskinit
= 0;
4002 rnp
->grplo
= j
* cpustride
;
4003 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4004 if (rnp
->grphi
>= nr_cpu_ids
)
4005 rnp
->grphi
= nr_cpu_ids
- 1;
4011 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
4012 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4013 rnp
->parent
= rsp
->level
[i
- 1] +
4014 j
/ rsp
->levelspread
[i
- 1];
4017 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4018 rcu_init_one_nocb(rnp
);
4022 init_waitqueue_head(&rsp
->gp_wq
);
4023 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4024 for_each_possible_cpu(i
) {
4025 while (i
> rnp
->grphi
)
4027 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4028 rcu_boot_init_percpu_data(i
, rsp
);
4030 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4034 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4035 * replace the definitions in tree.h because those are needed to size
4036 * the ->node array in the rcu_state structure.
4038 static void __init
rcu_init_geometry(void)
4044 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
4047 * Initialize any unspecified boot parameters.
4048 * The default values of jiffies_till_first_fqs and
4049 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4050 * value, which is a function of HZ, then adding one for each
4051 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4053 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4054 if (jiffies_till_first_fqs
== ULONG_MAX
)
4055 jiffies_till_first_fqs
= d
;
4056 if (jiffies_till_next_fqs
== ULONG_MAX
)
4057 jiffies_till_next_fqs
= d
;
4059 /* If the compile-time values are accurate, just leave. */
4060 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
4061 nr_cpu_ids
== NR_CPUS
)
4063 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4064 rcu_fanout_leaf
, nr_cpu_ids
);
4067 * Compute number of nodes that can be handled an rcu_node tree
4068 * with the given number of levels. Setting rcu_capacity[0] makes
4069 * some of the arithmetic easier.
4071 rcu_capacity
[0] = 1;
4072 rcu_capacity
[1] = rcu_fanout_leaf
;
4073 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
4074 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
4077 * The boot-time rcu_fanout_leaf parameter is only permitted
4078 * to increase the leaf-level fanout, not decrease it. Of course,
4079 * the leaf-level fanout cannot exceed the number of bits in
4080 * the rcu_node masks. Finally, the tree must be able to accommodate
4081 * the configured number of CPUs. Complain and fall back to the
4082 * compile-time values if these limits are exceeded.
4084 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
4085 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
4086 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
4091 /* Calculate the number of rcu_nodes at each level of the tree. */
4092 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
4093 if (n
<= rcu_capacity
[i
]) {
4094 for (j
= 0; j
<= i
; j
++)
4096 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
4098 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
4103 /* Calculate the total number of rcu_node structures. */
4105 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
4106 rcu_num_nodes
+= num_rcu_lvl
[i
];
4110 void __init
rcu_init(void)
4114 rcu_early_boot_tests();
4116 rcu_bootup_announce();
4117 rcu_init_geometry();
4118 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
4119 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
4120 __rcu_init_preempt();
4121 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4124 * We don't need protection against CPU-hotplug here because
4125 * this is called early in boot, before either interrupts
4126 * or the scheduler are operational.
4128 cpu_notifier(rcu_cpu_notify
, 0);
4129 pm_notifier(rcu_pm_notify
, 0);
4130 for_each_online_cpu(cpu
)
4131 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4134 #include "tree_plugin.h"