2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
72 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
83 # define DEFINE_RCU_TPS(sname) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
86 # define RCU_STATE_NAME(sname) sname##_varname
88 # define DEFINE_RCU_TPS(sname)
89 # define RCU_STATE_NAME(sname) __stringify(sname)
92 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
93 DEFINE_RCU_TPS(sname) \
94 struct rcu_state sname##_state = { \
95 .level = { &sname##_state.node[0] }, \
97 .fqs_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
101 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
102 .orphan_donetail = &sname##_state.orphan_donelist, \
103 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
104 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
105 .name = RCU_STATE_NAME(sname), \
108 DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data)
110 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
111 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
113 static struct rcu_state
*rcu_state_p
;
114 LIST_HEAD(rcu_struct_flavors
);
116 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
117 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
118 module_param(rcu_fanout_leaf
, int, 0444);
119 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
120 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
127 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
130 * The rcu_scheduler_active variable transitions from zero to one just
131 * before the first task is spawned. So when this variable is zero, RCU
132 * can assume that there is but one task, allowing RCU to (for example)
133 * optimize synchronize_sched() to a simple barrier(). When this variable
134 * is one, RCU must actually do all the hard work required to detect real
135 * grace periods. This variable is also used to suppress boot-time false
136 * positives from lockdep-RCU error checking.
138 int rcu_scheduler_active __read_mostly
;
139 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
142 * The rcu_scheduler_fully_active variable transitions from zero to one
143 * during the early_initcall() processing, which is after the scheduler
144 * is capable of creating new tasks. So RCU processing (for example,
145 * creating tasks for RCU priority boosting) must be delayed until after
146 * rcu_scheduler_fully_active transitions from zero to one. We also
147 * currently delay invocation of any RCU callbacks until after this point.
149 * It might later prove better for people registering RCU callbacks during
150 * early boot to take responsibility for these callbacks, but one step at
153 static int rcu_scheduler_fully_active __read_mostly
;
155 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
156 static void invoke_rcu_core(void);
157 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
159 /* rcuc/rcub kthread realtime priority */
160 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
161 module_param(kthread_prio
, int, 0644);
163 /* Delay in jiffies for grace-period initialization delays. */
164 static int gp_init_delay
= IS_ENABLED(CONFIG_RCU_TORTURE_TEST_SLOW_INIT
)
165 ? CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
167 module_param(gp_init_delay
, int, 0644);
170 * Track the rcutorture test sequence number and the update version
171 * number within a given test. The rcutorture_testseq is incremented
172 * on every rcutorture module load and unload, so has an odd value
173 * when a test is running. The rcutorture_vernum is set to zero
174 * when rcutorture starts and is incremented on each rcutorture update.
175 * These variables enable correlating rcutorture output with the
176 * RCU tracing information.
178 unsigned long rcutorture_testseq
;
179 unsigned long rcutorture_vernum
;
182 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
183 * permit this function to be invoked without holding the root rcu_node
184 * structure's ->lock, but of course results can be subject to change.
186 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
188 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
192 * Note a quiescent state. Because we do not need to know
193 * how many quiescent states passed, just if there was at least
194 * one since the start of the grace period, this just sets a flag.
195 * The caller must have disabled preemption.
197 void rcu_sched_qs(void)
199 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
200 trace_rcu_grace_period(TPS("rcu_sched"),
201 __this_cpu_read(rcu_sched_data
.gpnum
),
203 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
209 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
210 trace_rcu_grace_period(TPS("rcu_bh"),
211 __this_cpu_read(rcu_bh_data
.gpnum
),
213 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
217 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
219 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
220 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
221 .dynticks
= ATOMIC_INIT(1),
222 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
223 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
224 .dynticks_idle
= ATOMIC_INIT(1),
225 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
228 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
229 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
232 * Let the RCU core know that this CPU has gone through the scheduler,
233 * which is a quiescent state. This is called when the need for a
234 * quiescent state is urgent, so we burn an atomic operation and full
235 * memory barriers to let the RCU core know about it, regardless of what
236 * this CPU might (or might not) do in the near future.
238 * We inform the RCU core by emulating a zero-duration dyntick-idle
239 * period, which we in turn do by incrementing the ->dynticks counter
242 static void rcu_momentary_dyntick_idle(void)
245 struct rcu_data
*rdp
;
246 struct rcu_dynticks
*rdtp
;
248 struct rcu_state
*rsp
;
250 local_irq_save(flags
);
253 * Yes, we can lose flag-setting operations. This is OK, because
254 * the flag will be set again after some delay.
256 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
257 raw_cpu_write(rcu_sched_qs_mask
, 0);
259 /* Find the flavor that needs a quiescent state. */
260 for_each_rcu_flavor(rsp
) {
261 rdp
= raw_cpu_ptr(rsp
->rda
);
262 if (!(resched_mask
& rsp
->flavor_mask
))
264 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
265 if (ACCESS_ONCE(rdp
->mynode
->completed
) !=
266 ACCESS_ONCE(rdp
->cond_resched_completed
))
270 * Pretend to be momentarily idle for the quiescent state.
271 * This allows the grace-period kthread to record the
272 * quiescent state, with no need for this CPU to do anything
275 rdtp
= this_cpu_ptr(&rcu_dynticks
);
276 smp_mb__before_atomic(); /* Earlier stuff before QS. */
277 atomic_add(2, &rdtp
->dynticks
); /* QS. */
278 smp_mb__after_atomic(); /* Later stuff after QS. */
281 local_irq_restore(flags
);
285 * Note a context switch. This is a quiescent state for RCU-sched,
286 * and requires special handling for preemptible RCU.
287 * The caller must have disabled preemption.
289 void rcu_note_context_switch(void)
291 trace_rcu_utilization(TPS("Start context switch"));
293 rcu_preempt_note_context_switch();
294 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
295 rcu_momentary_dyntick_idle();
296 trace_rcu_utilization(TPS("End context switch"));
298 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
301 * Register a quiesecent state for all RCU flavors. If there is an
302 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
303 * dyntick-idle quiescent state visible to other CPUs (but only for those
304 * RCU flavors in desparate need of a quiescent state, which will normally
305 * be none of them). Either way, do a lightweight quiescent state for
308 void rcu_all_qs(void)
310 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
311 rcu_momentary_dyntick_idle();
312 this_cpu_inc(rcu_qs_ctr
);
314 EXPORT_SYMBOL_GPL(rcu_all_qs
);
316 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
317 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
318 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
320 module_param(blimit
, long, 0444);
321 module_param(qhimark
, long, 0444);
322 module_param(qlowmark
, long, 0444);
324 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
325 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
327 module_param(jiffies_till_first_fqs
, ulong
, 0644);
328 module_param(jiffies_till_next_fqs
, ulong
, 0644);
331 * How long the grace period must be before we start recruiting
332 * quiescent-state help from rcu_note_context_switch().
334 static ulong jiffies_till_sched_qs
= HZ
/ 20;
335 module_param(jiffies_till_sched_qs
, ulong
, 0644);
337 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
338 struct rcu_data
*rdp
);
339 static void force_qs_rnp(struct rcu_state
*rsp
,
340 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
341 unsigned long *maxj
),
342 bool *isidle
, unsigned long *maxj
);
343 static void force_quiescent_state(struct rcu_state
*rsp
);
344 static int rcu_pending(void);
347 * Return the number of RCU batches started thus far for debug & stats.
349 unsigned long rcu_batches_started(void)
351 return rcu_state_p
->gpnum
;
353 EXPORT_SYMBOL_GPL(rcu_batches_started
);
356 * Return the number of RCU-sched batches started thus far for debug & stats.
358 unsigned long rcu_batches_started_sched(void)
360 return rcu_sched_state
.gpnum
;
362 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
365 * Return the number of RCU BH batches started thus far for debug & stats.
367 unsigned long rcu_batches_started_bh(void)
369 return rcu_bh_state
.gpnum
;
371 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
374 * Return the number of RCU batches completed thus far for debug & stats.
376 unsigned long rcu_batches_completed(void)
378 return rcu_state_p
->completed
;
380 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
383 * Return the number of RCU-sched batches completed thus far for debug & stats.
385 unsigned long rcu_batches_completed_sched(void)
387 return rcu_sched_state
.completed
;
389 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
392 * Return the number of RCU BH batches completed thus far for debug & stats.
394 unsigned long rcu_batches_completed_bh(void)
396 return rcu_bh_state
.completed
;
398 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
401 * Force a quiescent state.
403 void rcu_force_quiescent_state(void)
405 force_quiescent_state(rcu_state_p
);
407 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
410 * Force a quiescent state for RCU BH.
412 void rcu_bh_force_quiescent_state(void)
414 force_quiescent_state(&rcu_bh_state
);
416 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
419 * Show the state of the grace-period kthreads.
421 void show_rcu_gp_kthreads(void)
423 struct rcu_state
*rsp
;
425 for_each_rcu_flavor(rsp
) {
426 pr_info("%s: wait state: %d ->state: %#lx\n",
427 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
428 /* sched_show_task(rsp->gp_kthread); */
431 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
434 * Record the number of times rcutorture tests have been initiated and
435 * terminated. This information allows the debugfs tracing stats to be
436 * correlated to the rcutorture messages, even when the rcutorture module
437 * is being repeatedly loaded and unloaded. In other words, we cannot
438 * store this state in rcutorture itself.
440 void rcutorture_record_test_transition(void)
442 rcutorture_testseq
++;
443 rcutorture_vernum
= 0;
445 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
448 * Send along grace-period-related data for rcutorture diagnostics.
450 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
451 unsigned long *gpnum
, unsigned long *completed
)
453 struct rcu_state
*rsp
= NULL
;
462 case RCU_SCHED_FLAVOR
:
463 rsp
= &rcu_sched_state
;
469 *flags
= ACCESS_ONCE(rsp
->gp_flags
);
470 *gpnum
= ACCESS_ONCE(rsp
->gpnum
);
471 *completed
= ACCESS_ONCE(rsp
->completed
);
478 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
481 * Record the number of writer passes through the current rcutorture test.
482 * This is also used to correlate debugfs tracing stats with the rcutorture
485 void rcutorture_record_progress(unsigned long vernum
)
489 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
492 * Force a quiescent state for RCU-sched.
494 void rcu_sched_force_quiescent_state(void)
496 force_quiescent_state(&rcu_sched_state
);
498 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
501 * Does the CPU have callbacks ready to be invoked?
504 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
506 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
507 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
511 * Return the root node of the specified rcu_state structure.
513 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
515 return &rsp
->node
[0];
519 * Is there any need for future grace periods?
520 * Interrupts must be disabled. If the caller does not hold the root
521 * rnp_node structure's ->lock, the results are advisory only.
523 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
525 struct rcu_node
*rnp
= rcu_get_root(rsp
);
526 int idx
= (ACCESS_ONCE(rnp
->completed
) + 1) & 0x1;
527 int *fp
= &rnp
->need_future_gp
[idx
];
529 return ACCESS_ONCE(*fp
);
533 * Does the current CPU require a not-yet-started grace period?
534 * The caller must have disabled interrupts to prevent races with
535 * normal callback registry.
538 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
542 if (rcu_gp_in_progress(rsp
))
543 return 0; /* No, a grace period is already in progress. */
544 if (rcu_future_needs_gp(rsp
))
545 return 1; /* Yes, a no-CBs CPU needs one. */
546 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
547 return 0; /* No, this is a no-CBs (or offline) CPU. */
548 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
549 return 1; /* Yes, this CPU has newly registered callbacks. */
550 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
551 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
552 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
553 rdp
->nxtcompleted
[i
]))
554 return 1; /* Yes, CBs for future grace period. */
555 return 0; /* No grace period needed. */
559 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
561 * If the new value of the ->dynticks_nesting counter now is zero,
562 * we really have entered idle, and must do the appropriate accounting.
563 * The caller must have disabled interrupts.
565 static void rcu_eqs_enter_common(long long oldval
, bool user
)
567 struct rcu_state
*rsp
;
568 struct rcu_data
*rdp
;
569 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
571 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
572 if (!user
&& !is_idle_task(current
)) {
573 struct task_struct
*idle __maybe_unused
=
574 idle_task(smp_processor_id());
576 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
577 ftrace_dump(DUMP_ORIG
);
578 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
579 current
->pid
, current
->comm
,
580 idle
->pid
, idle
->comm
); /* must be idle task! */
582 for_each_rcu_flavor(rsp
) {
583 rdp
= this_cpu_ptr(rsp
->rda
);
584 do_nocb_deferred_wakeup(rdp
);
586 rcu_prepare_for_idle();
587 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
588 smp_mb__before_atomic(); /* See above. */
589 atomic_inc(&rdtp
->dynticks
);
590 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
591 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
592 rcu_dynticks_task_enter();
595 * It is illegal to enter an extended quiescent state while
596 * in an RCU read-side critical section.
598 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
599 "Illegal idle entry in RCU read-side critical section.");
600 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
601 "Illegal idle entry in RCU-bh read-side critical section.");
602 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
603 "Illegal idle entry in RCU-sched read-side critical section.");
607 * Enter an RCU extended quiescent state, which can be either the
608 * idle loop or adaptive-tickless usermode execution.
610 static void rcu_eqs_enter(bool user
)
613 struct rcu_dynticks
*rdtp
;
615 rdtp
= this_cpu_ptr(&rcu_dynticks
);
616 oldval
= rdtp
->dynticks_nesting
;
617 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
618 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
619 rdtp
->dynticks_nesting
= 0;
620 rcu_eqs_enter_common(oldval
, user
);
622 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
627 * rcu_idle_enter - inform RCU that current CPU is entering idle
629 * Enter idle mode, in other words, -leave- the mode in which RCU
630 * read-side critical sections can occur. (Though RCU read-side
631 * critical sections can occur in irq handlers in idle, a possibility
632 * handled by irq_enter() and irq_exit().)
634 * We crowbar the ->dynticks_nesting field to zero to allow for
635 * the possibility of usermode upcalls having messed up our count
636 * of interrupt nesting level during the prior busy period.
638 void rcu_idle_enter(void)
642 local_irq_save(flags
);
643 rcu_eqs_enter(false);
644 rcu_sysidle_enter(0);
645 local_irq_restore(flags
);
647 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
649 #ifdef CONFIG_RCU_USER_QS
651 * rcu_user_enter - inform RCU that we are resuming userspace.
653 * Enter RCU idle mode right before resuming userspace. No use of RCU
654 * is permitted between this call and rcu_user_exit(). This way the
655 * CPU doesn't need to maintain the tick for RCU maintenance purposes
656 * when the CPU runs in userspace.
658 void rcu_user_enter(void)
662 #endif /* CONFIG_RCU_USER_QS */
665 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
667 * Exit from an interrupt handler, which might possibly result in entering
668 * idle mode, in other words, leaving the mode in which read-side critical
669 * sections can occur.
671 * This code assumes that the idle loop never does anything that might
672 * result in unbalanced calls to irq_enter() and irq_exit(). If your
673 * architecture violates this assumption, RCU will give you what you
674 * deserve, good and hard. But very infrequently and irreproducibly.
676 * Use things like work queues to work around this limitation.
678 * You have been warned.
680 void rcu_irq_exit(void)
684 struct rcu_dynticks
*rdtp
;
686 local_irq_save(flags
);
687 rdtp
= this_cpu_ptr(&rcu_dynticks
);
688 oldval
= rdtp
->dynticks_nesting
;
689 rdtp
->dynticks_nesting
--;
690 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
691 if (rdtp
->dynticks_nesting
)
692 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
694 rcu_eqs_enter_common(oldval
, true);
695 rcu_sysidle_enter(1);
696 local_irq_restore(flags
);
700 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
702 * If the new value of the ->dynticks_nesting counter was previously zero,
703 * we really have exited idle, and must do the appropriate accounting.
704 * The caller must have disabled interrupts.
706 static void rcu_eqs_exit_common(long long oldval
, int user
)
708 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
710 rcu_dynticks_task_exit();
711 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
712 atomic_inc(&rdtp
->dynticks
);
713 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
714 smp_mb__after_atomic(); /* See above. */
715 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
716 rcu_cleanup_after_idle();
717 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
718 if (!user
&& !is_idle_task(current
)) {
719 struct task_struct
*idle __maybe_unused
=
720 idle_task(smp_processor_id());
722 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
723 oldval
, rdtp
->dynticks_nesting
);
724 ftrace_dump(DUMP_ORIG
);
725 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
726 current
->pid
, current
->comm
,
727 idle
->pid
, idle
->comm
); /* must be idle task! */
732 * Exit an RCU extended quiescent state, which can be either the
733 * idle loop or adaptive-tickless usermode execution.
735 static void rcu_eqs_exit(bool user
)
737 struct rcu_dynticks
*rdtp
;
740 rdtp
= this_cpu_ptr(&rcu_dynticks
);
741 oldval
= rdtp
->dynticks_nesting
;
742 WARN_ON_ONCE(oldval
< 0);
743 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
744 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
746 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
747 rcu_eqs_exit_common(oldval
, user
);
752 * rcu_idle_exit - inform RCU that current CPU is leaving idle
754 * Exit idle mode, in other words, -enter- the mode in which RCU
755 * read-side critical sections can occur.
757 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
758 * allow for the possibility of usermode upcalls messing up our count
759 * of interrupt nesting level during the busy period that is just
762 void rcu_idle_exit(void)
766 local_irq_save(flags
);
769 local_irq_restore(flags
);
771 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
773 #ifdef CONFIG_RCU_USER_QS
775 * rcu_user_exit - inform RCU that we are exiting userspace.
777 * Exit RCU idle mode while entering the kernel because it can
778 * run a RCU read side critical section anytime.
780 void rcu_user_exit(void)
784 #endif /* CONFIG_RCU_USER_QS */
787 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
789 * Enter an interrupt handler, which might possibly result in exiting
790 * idle mode, in other words, entering the mode in which read-side critical
791 * sections can occur.
793 * Note that the Linux kernel is fully capable of entering an interrupt
794 * handler that it never exits, for example when doing upcalls to
795 * user mode! This code assumes that the idle loop never does upcalls to
796 * user mode. If your architecture does do upcalls from the idle loop (or
797 * does anything else that results in unbalanced calls to the irq_enter()
798 * and irq_exit() functions), RCU will give you what you deserve, good
799 * and hard. But very infrequently and irreproducibly.
801 * Use things like work queues to work around this limitation.
803 * You have been warned.
805 void rcu_irq_enter(void)
808 struct rcu_dynticks
*rdtp
;
811 local_irq_save(flags
);
812 rdtp
= this_cpu_ptr(&rcu_dynticks
);
813 oldval
= rdtp
->dynticks_nesting
;
814 rdtp
->dynticks_nesting
++;
815 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
817 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
819 rcu_eqs_exit_common(oldval
, true);
821 local_irq_restore(flags
);
825 * rcu_nmi_enter - inform RCU of entry to NMI context
827 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
828 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
829 * that the CPU is active. This implementation permits nested NMIs, as
830 * long as the nesting level does not overflow an int. (You will probably
831 * run out of stack space first.)
833 void rcu_nmi_enter(void)
835 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
838 /* Complain about underflow. */
839 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
842 * If idle from RCU viewpoint, atomically increment ->dynticks
843 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
844 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
845 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
846 * to be in the outermost NMI handler that interrupted an RCU-idle
847 * period (observation due to Andy Lutomirski).
849 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
850 smp_mb__before_atomic(); /* Force delay from prior write. */
851 atomic_inc(&rdtp
->dynticks
);
852 /* atomic_inc() before later RCU read-side crit sects */
853 smp_mb__after_atomic(); /* See above. */
854 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
857 rdtp
->dynticks_nmi_nesting
+= incby
;
862 * rcu_nmi_exit - inform RCU of exit from NMI context
864 * If we are returning from the outermost NMI handler that interrupted an
865 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
866 * to let the RCU grace-period handling know that the CPU is back to
869 void rcu_nmi_exit(void)
871 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
874 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
875 * (We are exiting an NMI handler, so RCU better be paying attention
878 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
879 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
882 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
883 * leave it in non-RCU-idle state.
885 if (rdtp
->dynticks_nmi_nesting
!= 1) {
886 rdtp
->dynticks_nmi_nesting
-= 2;
890 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
891 rdtp
->dynticks_nmi_nesting
= 0;
892 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
893 smp_mb__before_atomic(); /* See above. */
894 atomic_inc(&rdtp
->dynticks
);
895 smp_mb__after_atomic(); /* Force delay to next write. */
896 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
900 * __rcu_is_watching - are RCU read-side critical sections safe?
902 * Return true if RCU is watching the running CPU, which means that
903 * this CPU can safely enter RCU read-side critical sections. Unlike
904 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
905 * least disabled preemption.
907 bool notrace
__rcu_is_watching(void)
909 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
913 * rcu_is_watching - see if RCU thinks that the current CPU is idle
915 * If the current CPU is in its idle loop and is neither in an interrupt
916 * or NMI handler, return true.
918 bool notrace
rcu_is_watching(void)
923 ret
= __rcu_is_watching();
927 EXPORT_SYMBOL_GPL(rcu_is_watching
);
929 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
932 * Is the current CPU online? Disable preemption to avoid false positives
933 * that could otherwise happen due to the current CPU number being sampled,
934 * this task being preempted, its old CPU being taken offline, resuming
935 * on some other CPU, then determining that its old CPU is now offline.
936 * It is OK to use RCU on an offline processor during initial boot, hence
937 * the check for rcu_scheduler_fully_active. Note also that it is OK
938 * for a CPU coming online to use RCU for one jiffy prior to marking itself
939 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
940 * offline to continue to use RCU for one jiffy after marking itself
941 * offline in the cpu_online_mask. This leniency is necessary given the
942 * non-atomic nature of the online and offline processing, for example,
943 * the fact that a CPU enters the scheduler after completing the CPU_DYING
946 * This is also why RCU internally marks CPUs online during the
947 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
949 * Disable checking if in an NMI handler because we cannot safely report
950 * errors from NMI handlers anyway.
952 bool rcu_lockdep_current_cpu_online(void)
954 struct rcu_data
*rdp
;
955 struct rcu_node
*rnp
;
961 rdp
= this_cpu_ptr(&rcu_sched_data
);
963 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
964 !rcu_scheduler_fully_active
;
968 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
970 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
973 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
975 * If the current CPU is idle or running at a first-level (not nested)
976 * interrupt from idle, return true. The caller must have at least
977 * disabled preemption.
979 static int rcu_is_cpu_rrupt_from_idle(void)
981 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
985 * Snapshot the specified CPU's dynticks counter so that we can later
986 * credit them with an implicit quiescent state. Return 1 if this CPU
987 * is in dynticks idle mode, which is an extended quiescent state.
989 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
990 bool *isidle
, unsigned long *maxj
)
992 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
993 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
994 if ((rdp
->dynticks_snap
& 0x1) == 0) {
995 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
998 if (ULONG_CMP_LT(ACCESS_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1000 ACCESS_ONCE(rdp
->gpwrap
) = true;
1006 * Return true if the specified CPU has passed through a quiescent
1007 * state by virtue of being in or having passed through an dynticks
1008 * idle state since the last call to dyntick_save_progress_counter()
1009 * for this same CPU, or by virtue of having been offline.
1011 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1012 bool *isidle
, unsigned long *maxj
)
1018 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1019 snap
= (unsigned int)rdp
->dynticks_snap
;
1022 * If the CPU passed through or entered a dynticks idle phase with
1023 * no active irq/NMI handlers, then we can safely pretend that the CPU
1024 * already acknowledged the request to pass through a quiescent
1025 * state. Either way, that CPU cannot possibly be in an RCU
1026 * read-side critical section that started before the beginning
1027 * of the current RCU grace period.
1029 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1030 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1031 rdp
->dynticks_fqs
++;
1036 * Check for the CPU being offline, but only if the grace period
1037 * is old enough. We don't need to worry about the CPU changing
1038 * state: If we see it offline even once, it has been through a
1041 * The reason for insisting that the grace period be at least
1042 * one jiffy old is that CPUs that are not quite online and that
1043 * have just gone offline can still execute RCU read-side critical
1046 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1047 return 0; /* Grace period is not old enough. */
1049 if (cpu_is_offline(rdp
->cpu
)) {
1050 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1056 * A CPU running for an extended time within the kernel can
1057 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1058 * even context-switching back and forth between a pair of
1059 * in-kernel CPU-bound tasks cannot advance grace periods.
1060 * So if the grace period is old enough, make the CPU pay attention.
1061 * Note that the unsynchronized assignments to the per-CPU
1062 * rcu_sched_qs_mask variable are safe. Yes, setting of
1063 * bits can be lost, but they will be set again on the next
1064 * force-quiescent-state pass. So lost bit sets do not result
1065 * in incorrect behavior, merely in a grace period lasting
1066 * a few jiffies longer than it might otherwise. Because
1067 * there are at most four threads involved, and because the
1068 * updates are only once every few jiffies, the probability of
1069 * lossage (and thus of slight grace-period extension) is
1072 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1073 * is set too high, we override with half of the RCU CPU stall
1076 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1077 if (ULONG_CMP_GE(jiffies
,
1078 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1079 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1080 if (!(ACCESS_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1081 ACCESS_ONCE(rdp
->cond_resched_completed
) =
1082 ACCESS_ONCE(rdp
->mynode
->completed
);
1083 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1084 ACCESS_ONCE(*rcrmp
) =
1085 ACCESS_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
;
1086 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1087 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1088 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1089 /* Time to beat on that CPU again! */
1090 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1091 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1098 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1100 unsigned long j
= jiffies
;
1104 smp_wmb(); /* Record start time before stall time. */
1105 j1
= rcu_jiffies_till_stall_check();
1106 ACCESS_ONCE(rsp
->jiffies_stall
) = j
+ j1
;
1107 rsp
->jiffies_resched
= j
+ j1
/ 2;
1108 rsp
->n_force_qs_gpstart
= ACCESS_ONCE(rsp
->n_force_qs
);
1112 * Complain about starvation of grace-period kthread.
1114 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1120 gpa
= ACCESS_ONCE(rsp
->gp_activity
);
1121 if (j
- gpa
> 2 * HZ
)
1122 pr_err("%s kthread starved for %ld jiffies!\n",
1123 rsp
->name
, j
- gpa
);
1127 * Dump stacks of all tasks running on stalled CPUs.
1129 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1132 unsigned long flags
;
1133 struct rcu_node
*rnp
;
1135 rcu_for_each_leaf_node(rsp
, rnp
) {
1136 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1137 if (rnp
->qsmask
!= 0) {
1138 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1139 if (rnp
->qsmask
& (1UL << cpu
))
1140 dump_cpu_task(rnp
->grplo
+ cpu
);
1142 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1146 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1150 unsigned long flags
;
1154 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1157 /* Only let one CPU complain about others per time interval. */
1159 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1160 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
1161 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1162 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1165 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
1166 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1169 * OK, time to rat on our buddy...
1170 * See Documentation/RCU/stallwarn.txt for info on how to debug
1171 * RCU CPU stall warnings.
1173 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1175 print_cpu_stall_info_begin();
1176 rcu_for_each_leaf_node(rsp
, rnp
) {
1177 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1178 ndetected
+= rcu_print_task_stall(rnp
);
1179 if (rnp
->qsmask
!= 0) {
1180 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1181 if (rnp
->qsmask
& (1UL << cpu
)) {
1182 print_cpu_stall_info(rsp
,
1187 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1190 print_cpu_stall_info_end();
1191 for_each_possible_cpu(cpu
)
1192 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1193 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1194 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1195 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1197 rcu_dump_cpu_stacks(rsp
);
1199 if (ACCESS_ONCE(rsp
->gpnum
) != gpnum
||
1200 ACCESS_ONCE(rsp
->completed
) == gpnum
) {
1201 pr_err("INFO: Stall ended before state dump start\n");
1204 gpa
= ACCESS_ONCE(rsp
->gp_activity
);
1205 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1206 rsp
->name
, j
- gpa
, j
, gpa
,
1207 jiffies_till_next_fqs
,
1208 rcu_get_root(rsp
)->qsmask
);
1209 /* In this case, the current CPU might be at fault. */
1210 sched_show_task(current
);
1214 /* Complain about tasks blocking the grace period. */
1215 rcu_print_detail_task_stall(rsp
);
1217 rcu_check_gp_kthread_starvation(rsp
);
1219 force_quiescent_state(rsp
); /* Kick them all. */
1222 static void print_cpu_stall(struct rcu_state
*rsp
)
1225 unsigned long flags
;
1226 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1230 * OK, time to rat on ourselves...
1231 * See Documentation/RCU/stallwarn.txt for info on how to debug
1232 * RCU CPU stall warnings.
1234 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1235 print_cpu_stall_info_begin();
1236 print_cpu_stall_info(rsp
, smp_processor_id());
1237 print_cpu_stall_info_end();
1238 for_each_possible_cpu(cpu
)
1239 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1240 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1241 jiffies
- rsp
->gp_start
,
1242 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1244 rcu_check_gp_kthread_starvation(rsp
);
1246 rcu_dump_cpu_stacks(rsp
);
1248 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1249 if (ULONG_CMP_GE(jiffies
, ACCESS_ONCE(rsp
->jiffies_stall
)))
1250 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+
1251 3 * rcu_jiffies_till_stall_check() + 3;
1252 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1255 * Attempt to revive the RCU machinery by forcing a context switch.
1257 * A context switch would normally allow the RCU state machine to make
1258 * progress and it could be we're stuck in kernel space without context
1259 * switches for an entirely unreasonable amount of time.
1261 resched_cpu(smp_processor_id());
1264 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1266 unsigned long completed
;
1267 unsigned long gpnum
;
1271 struct rcu_node
*rnp
;
1273 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1278 * Lots of memory barriers to reject false positives.
1280 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1281 * then rsp->gp_start, and finally rsp->completed. These values
1282 * are updated in the opposite order with memory barriers (or
1283 * equivalent) during grace-period initialization and cleanup.
1284 * Now, a false positive can occur if we get an new value of
1285 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1286 * the memory barriers, the only way that this can happen is if one
1287 * grace period ends and another starts between these two fetches.
1288 * Detect this by comparing rsp->completed with the previous fetch
1291 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1292 * and rsp->gp_start suffice to forestall false positives.
1294 gpnum
= ACCESS_ONCE(rsp
->gpnum
);
1295 smp_rmb(); /* Pick up ->gpnum first... */
1296 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
1297 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1298 gps
= ACCESS_ONCE(rsp
->gp_start
);
1299 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1300 completed
= ACCESS_ONCE(rsp
->completed
);
1301 if (ULONG_CMP_GE(completed
, gpnum
) ||
1302 ULONG_CMP_LT(j
, js
) ||
1303 ULONG_CMP_GE(gps
, js
))
1304 return; /* No stall or GP completed since entering function. */
1306 if (rcu_gp_in_progress(rsp
) &&
1307 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1309 /* We haven't checked in, so go dump stack. */
1310 print_cpu_stall(rsp
);
1312 } else if (rcu_gp_in_progress(rsp
) &&
1313 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1315 /* They had a few time units to dump stack, so complain. */
1316 print_other_cpu_stall(rsp
, gpnum
);
1321 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1323 * Set the stall-warning timeout way off into the future, thus preventing
1324 * any RCU CPU stall-warning messages from appearing in the current set of
1325 * RCU grace periods.
1327 * The caller must disable hard irqs.
1329 void rcu_cpu_stall_reset(void)
1331 struct rcu_state
*rsp
;
1333 for_each_rcu_flavor(rsp
)
1334 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ ULONG_MAX
/ 2;
1338 * Initialize the specified rcu_data structure's callback list to empty.
1340 static void init_callback_list(struct rcu_data
*rdp
)
1344 if (init_nocb_callback_list(rdp
))
1346 rdp
->nxtlist
= NULL
;
1347 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1348 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1352 * Determine the value that ->completed will have at the end of the
1353 * next subsequent grace period. This is used to tag callbacks so that
1354 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1355 * been dyntick-idle for an extended period with callbacks under the
1356 * influence of RCU_FAST_NO_HZ.
1358 * The caller must hold rnp->lock with interrupts disabled.
1360 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1361 struct rcu_node
*rnp
)
1364 * If RCU is idle, we just wait for the next grace period.
1365 * But we can only be sure that RCU is idle if we are looking
1366 * at the root rcu_node structure -- otherwise, a new grace
1367 * period might have started, but just not yet gotten around
1368 * to initializing the current non-root rcu_node structure.
1370 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1371 return rnp
->completed
+ 1;
1374 * Otherwise, wait for a possible partial grace period and
1375 * then the subsequent full grace period.
1377 return rnp
->completed
+ 2;
1381 * Trace-event helper function for rcu_start_future_gp() and
1382 * rcu_nocb_wait_gp().
1384 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1385 unsigned long c
, const char *s
)
1387 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1388 rnp
->completed
, c
, rnp
->level
,
1389 rnp
->grplo
, rnp
->grphi
, s
);
1393 * Start some future grace period, as needed to handle newly arrived
1394 * callbacks. The required future grace periods are recorded in each
1395 * rcu_node structure's ->need_future_gp field. Returns true if there
1396 * is reason to awaken the grace-period kthread.
1398 * The caller must hold the specified rcu_node structure's ->lock.
1400 static bool __maybe_unused
1401 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1402 unsigned long *c_out
)
1407 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1410 * Pick up grace-period number for new callbacks. If this
1411 * grace period is already marked as needed, return to the caller.
1413 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1414 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1415 if (rnp
->need_future_gp
[c
& 0x1]) {
1416 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1421 * If either this rcu_node structure or the root rcu_node structure
1422 * believe that a grace period is in progress, then we must wait
1423 * for the one following, which is in "c". Because our request
1424 * will be noticed at the end of the current grace period, we don't
1425 * need to explicitly start one. We only do the lockless check
1426 * of rnp_root's fields if the current rcu_node structure thinks
1427 * there is no grace period in flight, and because we hold rnp->lock,
1428 * the only possible change is when rnp_root's two fields are
1429 * equal, in which case rnp_root->gpnum might be concurrently
1430 * incremented. But that is OK, as it will just result in our
1431 * doing some extra useless work.
1433 if (rnp
->gpnum
!= rnp
->completed
||
1434 ACCESS_ONCE(rnp_root
->gpnum
) != ACCESS_ONCE(rnp_root
->completed
)) {
1435 rnp
->need_future_gp
[c
& 0x1]++;
1436 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1441 * There might be no grace period in progress. If we don't already
1442 * hold it, acquire the root rcu_node structure's lock in order to
1443 * start one (if needed).
1445 if (rnp
!= rnp_root
) {
1446 raw_spin_lock(&rnp_root
->lock
);
1447 smp_mb__after_unlock_lock();
1451 * Get a new grace-period number. If there really is no grace
1452 * period in progress, it will be smaller than the one we obtained
1453 * earlier. Adjust callbacks as needed. Note that even no-CBs
1454 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1456 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1457 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1458 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1459 rdp
->nxtcompleted
[i
] = c
;
1462 * If the needed for the required grace period is already
1463 * recorded, trace and leave.
1465 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1466 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1470 /* Record the need for the future grace period. */
1471 rnp_root
->need_future_gp
[c
& 0x1]++;
1473 /* If a grace period is not already in progress, start one. */
1474 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1475 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1477 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1478 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1481 if (rnp
!= rnp_root
)
1482 raw_spin_unlock(&rnp_root
->lock
);
1490 * Clean up any old requests for the just-ended grace period. Also return
1491 * whether any additional grace periods have been requested. Also invoke
1492 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1493 * waiting for this grace period to complete.
1495 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1497 int c
= rnp
->completed
;
1499 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1501 rcu_nocb_gp_cleanup(rsp
, rnp
);
1502 rnp
->need_future_gp
[c
& 0x1] = 0;
1503 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1504 trace_rcu_future_gp(rnp
, rdp
, c
,
1505 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1510 * Awaken the grace-period kthread for the specified flavor of RCU.
1511 * Don't do a self-awaken, and don't bother awakening when there is
1512 * nothing for the grace-period kthread to do (as in several CPUs
1513 * raced to awaken, and we lost), and finally don't try to awaken
1514 * a kthread that has not yet been created.
1516 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1518 if (current
== rsp
->gp_kthread
||
1519 !ACCESS_ONCE(rsp
->gp_flags
) ||
1522 wake_up(&rsp
->gp_wq
);
1526 * If there is room, assign a ->completed number to any callbacks on
1527 * this CPU that have not already been assigned. Also accelerate any
1528 * callbacks that were previously assigned a ->completed number that has
1529 * since proven to be too conservative, which can happen if callbacks get
1530 * assigned a ->completed number while RCU is idle, but with reference to
1531 * a non-root rcu_node structure. This function is idempotent, so it does
1532 * not hurt to call it repeatedly. Returns an flag saying that we should
1533 * awaken the RCU grace-period kthread.
1535 * The caller must hold rnp->lock with interrupts disabled.
1537 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1538 struct rcu_data
*rdp
)
1544 /* If the CPU has no callbacks, nothing to do. */
1545 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1549 * Starting from the sublist containing the callbacks most
1550 * recently assigned a ->completed number and working down, find the
1551 * first sublist that is not assignable to an upcoming grace period.
1552 * Such a sublist has something in it (first two tests) and has
1553 * a ->completed number assigned that will complete sooner than
1554 * the ->completed number for newly arrived callbacks (last test).
1556 * The key point is that any later sublist can be assigned the
1557 * same ->completed number as the newly arrived callbacks, which
1558 * means that the callbacks in any of these later sublist can be
1559 * grouped into a single sublist, whether or not they have already
1560 * been assigned a ->completed number.
1562 c
= rcu_cbs_completed(rsp
, rnp
);
1563 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1564 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1565 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1569 * If there are no sublist for unassigned callbacks, leave.
1570 * At the same time, advance "i" one sublist, so that "i" will
1571 * index into the sublist where all the remaining callbacks should
1574 if (++i
>= RCU_NEXT_TAIL
)
1578 * Assign all subsequent callbacks' ->completed number to the next
1579 * full grace period and group them all in the sublist initially
1582 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1583 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1584 rdp
->nxtcompleted
[i
] = c
;
1586 /* Record any needed additional grace periods. */
1587 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1589 /* Trace depending on how much we were able to accelerate. */
1590 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1591 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1593 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1598 * Move any callbacks whose grace period has completed to the
1599 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1600 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1601 * sublist. This function is idempotent, so it does not hurt to
1602 * invoke it repeatedly. As long as it is not invoked -too- often...
1603 * Returns true if the RCU grace-period kthread needs to be awakened.
1605 * The caller must hold rnp->lock with interrupts disabled.
1607 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1608 struct rcu_data
*rdp
)
1612 /* If the CPU has no callbacks, nothing to do. */
1613 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1617 * Find all callbacks whose ->completed numbers indicate that they
1618 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1620 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1621 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1623 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1625 /* Clean up any sublist tail pointers that were misordered above. */
1626 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1627 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1629 /* Copy down callbacks to fill in empty sublists. */
1630 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1631 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1633 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1634 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1637 /* Classify any remaining callbacks. */
1638 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1642 * Update CPU-local rcu_data state to record the beginnings and ends of
1643 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1644 * structure corresponding to the current CPU, and must have irqs disabled.
1645 * Returns true if the grace-period kthread needs to be awakened.
1647 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1648 struct rcu_data
*rdp
)
1652 /* Handle the ends of any preceding grace periods first. */
1653 if (rdp
->completed
== rnp
->completed
&&
1654 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1656 /* No grace period end, so just accelerate recent callbacks. */
1657 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1661 /* Advance callbacks. */
1662 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1664 /* Remember that we saw this grace-period completion. */
1665 rdp
->completed
= rnp
->completed
;
1666 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1669 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1671 * If the current grace period is waiting for this CPU,
1672 * set up to detect a quiescent state, otherwise don't
1673 * go looking for one.
1675 rdp
->gpnum
= rnp
->gpnum
;
1676 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1677 rdp
->passed_quiesce
= 0;
1678 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1679 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1680 zero_cpu_stall_ticks(rdp
);
1681 ACCESS_ONCE(rdp
->gpwrap
) = false;
1686 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1688 unsigned long flags
;
1690 struct rcu_node
*rnp
;
1692 local_irq_save(flags
);
1694 if ((rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) &&
1695 rdp
->completed
== ACCESS_ONCE(rnp
->completed
) &&
1696 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1697 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1698 local_irq_restore(flags
);
1701 smp_mb__after_unlock_lock();
1702 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1703 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1705 rcu_gp_kthread_wake(rsp
);
1709 * Initialize a new grace period. Return 0 if no grace period required.
1711 static int rcu_gp_init(struct rcu_state
*rsp
)
1713 struct rcu_data
*rdp
;
1714 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1716 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1717 rcu_bind_gp_kthread();
1718 raw_spin_lock_irq(&rnp
->lock
);
1719 smp_mb__after_unlock_lock();
1720 if (!ACCESS_ONCE(rsp
->gp_flags
)) {
1721 /* Spurious wakeup, tell caller to go back to sleep. */
1722 raw_spin_unlock_irq(&rnp
->lock
);
1725 ACCESS_ONCE(rsp
->gp_flags
) = 0; /* Clear all flags: New grace period. */
1727 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1729 * Grace period already in progress, don't start another.
1730 * Not supposed to be able to happen.
1732 raw_spin_unlock_irq(&rnp
->lock
);
1736 /* Advance to a new grace period and initialize state. */
1737 record_gp_stall_check_time(rsp
);
1738 /* Record GP times before starting GP, hence smp_store_release(). */
1739 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1740 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1741 raw_spin_unlock_irq(&rnp
->lock
);
1743 /* Exclude any concurrent CPU-hotplug operations. */
1744 mutex_lock(&rsp
->onoff_mutex
);
1745 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1748 * Set the quiescent-state-needed bits in all the rcu_node
1749 * structures for all currently online CPUs in breadth-first order,
1750 * starting from the root rcu_node structure, relying on the layout
1751 * of the tree within the rsp->node[] array. Note that other CPUs
1752 * will access only the leaves of the hierarchy, thus seeing that no
1753 * grace period is in progress, at least until the corresponding
1754 * leaf node has been initialized. In addition, we have excluded
1755 * CPU-hotplug operations.
1757 * The grace period cannot complete until the initialization
1758 * process finishes, because this kthread handles both.
1760 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1761 raw_spin_lock_irq(&rnp
->lock
);
1762 smp_mb__after_unlock_lock();
1763 rdp
= this_cpu_ptr(rsp
->rda
);
1764 rcu_preempt_check_blocked_tasks(rnp
);
1765 rnp
->qsmask
= rnp
->qsmaskinit
;
1766 ACCESS_ONCE(rnp
->gpnum
) = rsp
->gpnum
;
1767 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1768 ACCESS_ONCE(rnp
->completed
) = rsp
->completed
;
1769 if (rnp
== rdp
->mynode
)
1770 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1771 rcu_preempt_boost_start_gp(rnp
);
1772 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1773 rnp
->level
, rnp
->grplo
,
1774 rnp
->grphi
, rnp
->qsmask
);
1775 raw_spin_unlock_irq(&rnp
->lock
);
1776 cond_resched_rcu_qs();
1777 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1778 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_SLOW_INIT
) &&
1779 gp_init_delay
> 0 &&
1780 !(rsp
->gpnum
% (rcu_num_nodes
* 10)))
1781 schedule_timeout_uninterruptible(gp_init_delay
);
1784 mutex_unlock(&rsp
->onoff_mutex
);
1789 * Do one round of quiescent-state forcing.
1791 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1793 int fqs_state
= fqs_state_in
;
1794 bool isidle
= false;
1796 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1798 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1800 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1801 /* Collect dyntick-idle snapshots. */
1802 if (is_sysidle_rcu_state(rsp
)) {
1804 maxj
= jiffies
- ULONG_MAX
/ 4;
1806 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1808 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1809 fqs_state
= RCU_FORCE_QS
;
1811 /* Handle dyntick-idle and offline CPUs. */
1813 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1815 /* Clear flag to prevent immediate re-entry. */
1816 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1817 raw_spin_lock_irq(&rnp
->lock
);
1818 smp_mb__after_unlock_lock();
1819 ACCESS_ONCE(rsp
->gp_flags
) =
1820 ACCESS_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
;
1821 raw_spin_unlock_irq(&rnp
->lock
);
1827 * Clean up after the old grace period.
1829 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1831 unsigned long gp_duration
;
1832 bool needgp
= false;
1834 struct rcu_data
*rdp
;
1835 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1837 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1838 raw_spin_lock_irq(&rnp
->lock
);
1839 smp_mb__after_unlock_lock();
1840 gp_duration
= jiffies
- rsp
->gp_start
;
1841 if (gp_duration
> rsp
->gp_max
)
1842 rsp
->gp_max
= gp_duration
;
1845 * We know the grace period is complete, but to everyone else
1846 * it appears to still be ongoing. But it is also the case
1847 * that to everyone else it looks like there is nothing that
1848 * they can do to advance the grace period. It is therefore
1849 * safe for us to drop the lock in order to mark the grace
1850 * period as completed in all of the rcu_node structures.
1852 raw_spin_unlock_irq(&rnp
->lock
);
1855 * Propagate new ->completed value to rcu_node structures so
1856 * that other CPUs don't have to wait until the start of the next
1857 * grace period to process their callbacks. This also avoids
1858 * some nasty RCU grace-period initialization races by forcing
1859 * the end of the current grace period to be completely recorded in
1860 * all of the rcu_node structures before the beginning of the next
1861 * grace period is recorded in any of the rcu_node structures.
1863 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1864 raw_spin_lock_irq(&rnp
->lock
);
1865 smp_mb__after_unlock_lock();
1866 ACCESS_ONCE(rnp
->completed
) = rsp
->gpnum
;
1867 rdp
= this_cpu_ptr(rsp
->rda
);
1868 if (rnp
== rdp
->mynode
)
1869 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1870 /* smp_mb() provided by prior unlock-lock pair. */
1871 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1872 raw_spin_unlock_irq(&rnp
->lock
);
1873 cond_resched_rcu_qs();
1874 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1876 rnp
= rcu_get_root(rsp
);
1877 raw_spin_lock_irq(&rnp
->lock
);
1878 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1879 rcu_nocb_gp_set(rnp
, nocb
);
1881 /* Declare grace period done. */
1882 ACCESS_ONCE(rsp
->completed
) = rsp
->gpnum
;
1883 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1884 rsp
->fqs_state
= RCU_GP_IDLE
;
1885 rdp
= this_cpu_ptr(rsp
->rda
);
1886 /* Advance CBs to reduce false positives below. */
1887 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1888 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1889 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
1890 trace_rcu_grace_period(rsp
->name
,
1891 ACCESS_ONCE(rsp
->gpnum
),
1894 raw_spin_unlock_irq(&rnp
->lock
);
1898 * Body of kthread that handles grace periods.
1900 static int __noreturn
rcu_gp_kthread(void *arg
)
1906 struct rcu_state
*rsp
= arg
;
1907 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1911 /* Handle grace-period start. */
1913 trace_rcu_grace_period(rsp
->name
,
1914 ACCESS_ONCE(rsp
->gpnum
),
1916 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
1917 wait_event_interruptible(rsp
->gp_wq
,
1918 ACCESS_ONCE(rsp
->gp_flags
) &
1920 /* Locking provides needed memory barrier. */
1921 if (rcu_gp_init(rsp
))
1923 cond_resched_rcu_qs();
1924 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1925 WARN_ON(signal_pending(current
));
1926 trace_rcu_grace_period(rsp
->name
,
1927 ACCESS_ONCE(rsp
->gpnum
),
1931 /* Handle quiescent-state forcing. */
1932 fqs_state
= RCU_SAVE_DYNTICK
;
1933 j
= jiffies_till_first_fqs
;
1936 jiffies_till_first_fqs
= HZ
;
1941 rsp
->jiffies_force_qs
= jiffies
+ j
;
1942 trace_rcu_grace_period(rsp
->name
,
1943 ACCESS_ONCE(rsp
->gpnum
),
1945 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
1946 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1947 ((gf
= ACCESS_ONCE(rsp
->gp_flags
)) &
1949 (!ACCESS_ONCE(rnp
->qsmask
) &&
1950 !rcu_preempt_blocked_readers_cgp(rnp
)),
1952 /* Locking provides needed memory barriers. */
1953 /* If grace period done, leave loop. */
1954 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1955 !rcu_preempt_blocked_readers_cgp(rnp
))
1957 /* If time for quiescent-state forcing, do it. */
1958 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
1959 (gf
& RCU_GP_FLAG_FQS
)) {
1960 trace_rcu_grace_period(rsp
->name
,
1961 ACCESS_ONCE(rsp
->gpnum
),
1963 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1964 trace_rcu_grace_period(rsp
->name
,
1965 ACCESS_ONCE(rsp
->gpnum
),
1967 cond_resched_rcu_qs();
1968 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1970 /* Deal with stray signal. */
1971 cond_resched_rcu_qs();
1972 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1973 WARN_ON(signal_pending(current
));
1974 trace_rcu_grace_period(rsp
->name
,
1975 ACCESS_ONCE(rsp
->gpnum
),
1978 j
= jiffies_till_next_fqs
;
1981 jiffies_till_next_fqs
= HZ
;
1984 jiffies_till_next_fqs
= 1;
1988 /* Handle grace-period end. */
1989 rcu_gp_cleanup(rsp
);
1994 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1995 * in preparation for detecting the next grace period. The caller must hold
1996 * the root node's ->lock and hard irqs must be disabled.
1998 * Note that it is legal for a dying CPU (which is marked as offline) to
1999 * invoke this function. This can happen when the dying CPU reports its
2002 * Returns true if the grace-period kthread must be awakened.
2005 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2006 struct rcu_data
*rdp
)
2008 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2010 * Either we have not yet spawned the grace-period
2011 * task, this CPU does not need another grace period,
2012 * or a grace period is already in progress.
2013 * Either way, don't start a new grace period.
2017 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
2018 trace_rcu_grace_period(rsp
->name
, ACCESS_ONCE(rsp
->gpnum
),
2022 * We can't do wakeups while holding the rnp->lock, as that
2023 * could cause possible deadlocks with the rq->lock. Defer
2024 * the wakeup to our caller.
2030 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2031 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2032 * is invoked indirectly from rcu_advance_cbs(), which would result in
2033 * endless recursion -- or would do so if it wasn't for the self-deadlock
2034 * that is encountered beforehand.
2036 * Returns true if the grace-period kthread needs to be awakened.
2038 static bool rcu_start_gp(struct rcu_state
*rsp
)
2040 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2041 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2045 * If there is no grace period in progress right now, any
2046 * callbacks we have up to this point will be satisfied by the
2047 * next grace period. Also, advancing the callbacks reduces the
2048 * probability of false positives from cpu_needs_another_gp()
2049 * resulting in pointless grace periods. So, advance callbacks
2050 * then start the grace period!
2052 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2053 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2058 * Report a full set of quiescent states to the specified rcu_state
2059 * data structure. This involves cleaning up after the prior grace
2060 * period and letting rcu_start_gp() start up the next grace period
2061 * if one is needed. Note that the caller must hold rnp->lock, which
2062 * is released before return.
2064 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2065 __releases(rcu_get_root(rsp
)->lock
)
2067 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2068 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2069 rcu_gp_kthread_wake(rsp
);
2073 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2074 * Allows quiescent states for a group of CPUs to be reported at one go
2075 * to the specified rcu_node structure, though all the CPUs in the group
2076 * must be represented by the same rcu_node structure (which need not be
2077 * a leaf rcu_node structure, though it often will be). That structure's
2078 * lock must be held upon entry, and it is released before return.
2081 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2082 struct rcu_node
*rnp
, unsigned long flags
)
2083 __releases(rnp
->lock
)
2085 struct rcu_node
*rnp_c
;
2087 /* Walk up the rcu_node hierarchy. */
2089 if (!(rnp
->qsmask
& mask
)) {
2091 /* Our bit has already been cleared, so done. */
2092 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2095 rnp
->qsmask
&= ~mask
;
2096 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2097 mask
, rnp
->qsmask
, rnp
->level
,
2098 rnp
->grplo
, rnp
->grphi
,
2100 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2102 /* Other bits still set at this level, so done. */
2103 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2106 mask
= rnp
->grpmask
;
2107 if (rnp
->parent
== NULL
) {
2109 /* No more levels. Exit loop holding root lock. */
2113 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2116 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2117 smp_mb__after_unlock_lock();
2118 WARN_ON_ONCE(rnp_c
->qsmask
);
2122 * Get here if we are the last CPU to pass through a quiescent
2123 * state for this grace period. Invoke rcu_report_qs_rsp()
2124 * to clean up and start the next grace period if one is needed.
2126 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2130 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2131 * structure. This must be either called from the specified CPU, or
2132 * called when the specified CPU is known to be offline (and when it is
2133 * also known that no other CPU is concurrently trying to help the offline
2134 * CPU). The lastcomp argument is used to make sure we are still in the
2135 * grace period of interest. We don't want to end the current grace period
2136 * based on quiescent states detected in an earlier grace period!
2139 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2141 unsigned long flags
;
2144 struct rcu_node
*rnp
;
2147 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2148 smp_mb__after_unlock_lock();
2149 if ((rdp
->passed_quiesce
== 0 &&
2150 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2151 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2155 * The grace period in which this quiescent state was
2156 * recorded has ended, so don't report it upwards.
2157 * We will instead need a new quiescent state that lies
2158 * within the current grace period.
2160 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2161 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2162 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2165 mask
= rdp
->grpmask
;
2166 if ((rnp
->qsmask
& mask
) == 0) {
2167 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2169 rdp
->qs_pending
= 0;
2172 * This GP can't end until cpu checks in, so all of our
2173 * callbacks can be processed during the next GP.
2175 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2177 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
2179 rcu_gp_kthread_wake(rsp
);
2184 * Check to see if there is a new grace period of which this CPU
2185 * is not yet aware, and if so, set up local rcu_data state for it.
2186 * Otherwise, see if this CPU has just passed through its first
2187 * quiescent state for this grace period, and record that fact if so.
2190 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2192 /* Check for grace-period ends and beginnings. */
2193 note_gp_changes(rsp
, rdp
);
2196 * Does this CPU still need to do its part for current grace period?
2197 * If no, return and let the other CPUs do their part as well.
2199 if (!rdp
->qs_pending
)
2203 * Was there a quiescent state since the beginning of the grace
2204 * period? If no, then exit and wait for the next call.
2206 if (!rdp
->passed_quiesce
&&
2207 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2211 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2214 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2217 #ifdef CONFIG_HOTPLUG_CPU
2220 * Send the specified CPU's RCU callbacks to the orphanage. The
2221 * specified CPU must be offline, and the caller must hold the
2225 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2226 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2228 /* No-CBs CPUs do not have orphanable callbacks. */
2229 if (rcu_is_nocb_cpu(rdp
->cpu
))
2233 * Orphan the callbacks. First adjust the counts. This is safe
2234 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2235 * cannot be running now. Thus no memory barrier is required.
2237 if (rdp
->nxtlist
!= NULL
) {
2238 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2239 rsp
->qlen
+= rdp
->qlen
;
2240 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2242 ACCESS_ONCE(rdp
->qlen
) = 0;
2246 * Next, move those callbacks still needing a grace period to
2247 * the orphanage, where some other CPU will pick them up.
2248 * Some of the callbacks might have gone partway through a grace
2249 * period, but that is too bad. They get to start over because we
2250 * cannot assume that grace periods are synchronized across CPUs.
2251 * We don't bother updating the ->nxttail[] array yet, instead
2252 * we just reset the whole thing later on.
2254 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2255 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2256 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2257 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2261 * Then move the ready-to-invoke callbacks to the orphanage,
2262 * where some other CPU will pick them up. These will not be
2263 * required to pass though another grace period: They are done.
2265 if (rdp
->nxtlist
!= NULL
) {
2266 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2267 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2271 * Finally, initialize the rcu_data structure's list to empty and
2272 * disallow further callbacks on this CPU.
2274 init_callback_list(rdp
);
2275 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2279 * Adopt the RCU callbacks from the specified rcu_state structure's
2280 * orphanage. The caller must hold the ->orphan_lock.
2282 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2285 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2287 /* No-CBs CPUs are handled specially. */
2288 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2291 /* Do the accounting first. */
2292 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2293 rdp
->qlen
+= rsp
->qlen
;
2294 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2295 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2296 rcu_idle_count_callbacks_posted();
2301 * We do not need a memory barrier here because the only way we
2302 * can get here if there is an rcu_barrier() in flight is if
2303 * we are the task doing the rcu_barrier().
2306 /* First adopt the ready-to-invoke callbacks. */
2307 if (rsp
->orphan_donelist
!= NULL
) {
2308 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2309 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2310 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2311 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2312 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2313 rsp
->orphan_donelist
= NULL
;
2314 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2317 /* And then adopt the callbacks that still need a grace period. */
2318 if (rsp
->orphan_nxtlist
!= NULL
) {
2319 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2320 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2321 rsp
->orphan_nxtlist
= NULL
;
2322 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2327 * Trace the fact that this CPU is going offline.
2329 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2331 RCU_TRACE(unsigned long mask
);
2332 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2333 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2335 RCU_TRACE(mask
= rdp
->grpmask
);
2336 trace_rcu_grace_period(rsp
->name
,
2337 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2342 * All CPUs for the specified rcu_node structure have gone offline,
2343 * and all tasks that were preempted within an RCU read-side critical
2344 * section while running on one of those CPUs have since exited their RCU
2345 * read-side critical section. Some other CPU is reporting this fact with
2346 * the specified rcu_node structure's ->lock held and interrupts disabled.
2347 * This function therefore goes up the tree of rcu_node structures,
2348 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2349 * the leaf rcu_node structure's ->qsmaskinit field has already been
2352 * This function does check that the specified rcu_node structure has
2353 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2354 * prematurely. That said, invoking it after the fact will cost you
2355 * a needless lock acquisition. So once it has done its work, don't
2358 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2361 struct rcu_node
*rnp
= rnp_leaf
;
2363 if (rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2366 mask
= rnp
->grpmask
;
2370 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2371 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2372 rnp
->qsmaskinit
&= ~mask
;
2373 if (rnp
->qsmaskinit
) {
2374 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2377 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2382 * The CPU has been completely removed, and some other CPU is reporting
2383 * this fact from process context. Do the remainder of the cleanup,
2384 * including orphaning the outgoing CPU's RCU callbacks, and also
2385 * adopting them. There can only be one CPU hotplug operation at a time,
2386 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2388 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2390 unsigned long flags
;
2391 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2392 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2394 /* Adjust any no-longer-needed kthreads. */
2395 rcu_boost_kthread_setaffinity(rnp
, -1);
2397 /* Exclude any attempts to start a new grace period. */
2398 mutex_lock(&rsp
->onoff_mutex
);
2400 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2401 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2402 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2403 rcu_adopt_orphan_cbs(rsp
, flags
);
2404 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2406 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2407 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2408 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2409 rnp
->qsmaskinit
&= ~rdp
->grpmask
;
2410 if (rnp
->qsmaskinit
== 0 && !rcu_preempt_has_tasks(rnp
))
2411 rcu_cleanup_dead_rnp(rnp
);
2412 rcu_report_qs_rnp(rdp
->grpmask
, rsp
, rnp
, flags
); /* Rlses rnp->lock. */
2413 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2414 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2415 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2416 mutex_unlock(&rsp
->onoff_mutex
);
2419 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2421 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2425 static void __maybe_unused
rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2429 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2433 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2436 * Invoke any RCU callbacks that have made it to the end of their grace
2437 * period. Thottle as specified by rdp->blimit.
2439 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2441 unsigned long flags
;
2442 struct rcu_head
*next
, *list
, **tail
;
2443 long bl
, count
, count_lazy
;
2446 /* If no callbacks are ready, just return. */
2447 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2448 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2449 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
2450 need_resched(), is_idle_task(current
),
2451 rcu_is_callbacks_kthread());
2456 * Extract the list of ready callbacks, disabling to prevent
2457 * races with call_rcu() from interrupt handlers.
2459 local_irq_save(flags
);
2460 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2462 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2463 list
= rdp
->nxtlist
;
2464 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2465 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2466 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2467 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2468 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2469 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2470 local_irq_restore(flags
);
2472 /* Invoke callbacks. */
2473 count
= count_lazy
= 0;
2477 debug_rcu_head_unqueue(list
);
2478 if (__rcu_reclaim(rsp
->name
, list
))
2481 /* Stop only if limit reached and CPU has something to do. */
2482 if (++count
>= bl
&&
2484 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2488 local_irq_save(flags
);
2489 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2490 is_idle_task(current
),
2491 rcu_is_callbacks_kthread());
2493 /* Update count, and requeue any remaining callbacks. */
2495 *tail
= rdp
->nxtlist
;
2496 rdp
->nxtlist
= list
;
2497 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2498 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2499 rdp
->nxttail
[i
] = tail
;
2503 smp_mb(); /* List handling before counting for rcu_barrier(). */
2504 rdp
->qlen_lazy
-= count_lazy
;
2505 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
- count
;
2506 rdp
->n_cbs_invoked
+= count
;
2508 /* Reinstate batch limit if we have worked down the excess. */
2509 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2510 rdp
->blimit
= blimit
;
2512 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2513 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2514 rdp
->qlen_last_fqs_check
= 0;
2515 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2516 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2517 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2518 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2520 local_irq_restore(flags
);
2522 /* Re-invoke RCU core processing if there are callbacks remaining. */
2523 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2528 * Check to see if this CPU is in a non-context-switch quiescent state
2529 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2530 * Also schedule RCU core processing.
2532 * This function must be called from hardirq context. It is normally
2533 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2534 * false, there is no point in invoking rcu_check_callbacks().
2536 void rcu_check_callbacks(int user
)
2538 trace_rcu_utilization(TPS("Start scheduler-tick"));
2539 increment_cpu_stall_ticks();
2540 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2543 * Get here if this CPU took its interrupt from user
2544 * mode or from the idle loop, and if this is not a
2545 * nested interrupt. In this case, the CPU is in
2546 * a quiescent state, so note it.
2548 * No memory barrier is required here because both
2549 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2550 * variables that other CPUs neither access nor modify,
2551 * at least not while the corresponding CPU is online.
2557 } else if (!in_softirq()) {
2560 * Get here if this CPU did not take its interrupt from
2561 * softirq, in other words, if it is not interrupting
2562 * a rcu_bh read-side critical section. This is an _bh
2563 * critical section, so note it.
2568 rcu_preempt_check_callbacks();
2572 rcu_note_voluntary_context_switch(current
);
2573 trace_rcu_utilization(TPS("End scheduler-tick"));
2577 * Scan the leaf rcu_node structures, processing dyntick state for any that
2578 * have not yet encountered a quiescent state, using the function specified.
2579 * Also initiate boosting for any threads blocked on the root rcu_node.
2581 * The caller must have suppressed start of new grace periods.
2583 static void force_qs_rnp(struct rcu_state
*rsp
,
2584 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2585 unsigned long *maxj
),
2586 bool *isidle
, unsigned long *maxj
)
2590 unsigned long flags
;
2592 struct rcu_node
*rnp
;
2594 rcu_for_each_leaf_node(rsp
, rnp
) {
2595 cond_resched_rcu_qs();
2597 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2598 smp_mb__after_unlock_lock();
2599 if (!rcu_gp_in_progress(rsp
)) {
2600 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2603 if (rnp
->qsmask
== 0) {
2604 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
2609 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2610 if ((rnp
->qsmask
& bit
) != 0) {
2611 if ((rnp
->qsmaskinit
& bit
) != 0)
2613 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2619 /* rcu_report_qs_rnp() releases rnp->lock. */
2620 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
2623 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2628 * Force quiescent states on reluctant CPUs, and also detect which
2629 * CPUs are in dyntick-idle mode.
2631 static void force_quiescent_state(struct rcu_state
*rsp
)
2633 unsigned long flags
;
2635 struct rcu_node
*rnp
;
2636 struct rcu_node
*rnp_old
= NULL
;
2638 /* Funnel through hierarchy to reduce memory contention. */
2639 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2640 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2641 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2642 !raw_spin_trylock(&rnp
->fqslock
);
2643 if (rnp_old
!= NULL
)
2644 raw_spin_unlock(&rnp_old
->fqslock
);
2646 rsp
->n_force_qs_lh
++;
2651 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2653 /* Reached the root of the rcu_node tree, acquire lock. */
2654 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2655 smp_mb__after_unlock_lock();
2656 raw_spin_unlock(&rnp_old
->fqslock
);
2657 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2658 rsp
->n_force_qs_lh
++;
2659 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2660 return; /* Someone beat us to it. */
2662 ACCESS_ONCE(rsp
->gp_flags
) =
2663 ACCESS_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
;
2664 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2665 rcu_gp_kthread_wake(rsp
);
2669 * This does the RCU core processing work for the specified rcu_state
2670 * and rcu_data structures. This may be called only from the CPU to
2671 * whom the rdp belongs.
2674 __rcu_process_callbacks(struct rcu_state
*rsp
)
2676 unsigned long flags
;
2678 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2680 WARN_ON_ONCE(rdp
->beenonline
== 0);
2682 /* Update RCU state based on any recent quiescent states. */
2683 rcu_check_quiescent_state(rsp
, rdp
);
2685 /* Does this CPU require a not-yet-started grace period? */
2686 local_irq_save(flags
);
2687 if (cpu_needs_another_gp(rsp
, rdp
)) {
2688 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2689 needwake
= rcu_start_gp(rsp
);
2690 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2692 rcu_gp_kthread_wake(rsp
);
2694 local_irq_restore(flags
);
2697 /* If there are callbacks ready, invoke them. */
2698 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2699 invoke_rcu_callbacks(rsp
, rdp
);
2701 /* Do any needed deferred wakeups of rcuo kthreads. */
2702 do_nocb_deferred_wakeup(rdp
);
2706 * Do RCU core processing for the current CPU.
2708 static void rcu_process_callbacks(struct softirq_action
*unused
)
2710 struct rcu_state
*rsp
;
2712 if (cpu_is_offline(smp_processor_id()))
2714 trace_rcu_utilization(TPS("Start RCU core"));
2715 for_each_rcu_flavor(rsp
)
2716 __rcu_process_callbacks(rsp
);
2717 trace_rcu_utilization(TPS("End RCU core"));
2721 * Schedule RCU callback invocation. If the specified type of RCU
2722 * does not support RCU priority boosting, just do a direct call,
2723 * otherwise wake up the per-CPU kernel kthread. Note that because we
2724 * are running on the current CPU with softirqs disabled, the
2725 * rcu_cpu_kthread_task cannot disappear out from under us.
2727 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2729 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2731 if (likely(!rsp
->boost
)) {
2732 rcu_do_batch(rsp
, rdp
);
2735 invoke_rcu_callbacks_kthread();
2738 static void invoke_rcu_core(void)
2740 if (cpu_online(smp_processor_id()))
2741 raise_softirq(RCU_SOFTIRQ
);
2745 * Handle any core-RCU processing required by a call_rcu() invocation.
2747 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2748 struct rcu_head
*head
, unsigned long flags
)
2753 * If called from an extended quiescent state, invoke the RCU
2754 * core in order to force a re-evaluation of RCU's idleness.
2756 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2759 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2760 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2764 * Force the grace period if too many callbacks or too long waiting.
2765 * Enforce hysteresis, and don't invoke force_quiescent_state()
2766 * if some other CPU has recently done so. Also, don't bother
2767 * invoking force_quiescent_state() if the newly enqueued callback
2768 * is the only one waiting for a grace period to complete.
2770 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2772 /* Are we ignoring a completed grace period? */
2773 note_gp_changes(rsp
, rdp
);
2775 /* Start a new grace period if one not already started. */
2776 if (!rcu_gp_in_progress(rsp
)) {
2777 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2779 raw_spin_lock(&rnp_root
->lock
);
2780 smp_mb__after_unlock_lock();
2781 needwake
= rcu_start_gp(rsp
);
2782 raw_spin_unlock(&rnp_root
->lock
);
2784 rcu_gp_kthread_wake(rsp
);
2786 /* Give the grace period a kick. */
2787 rdp
->blimit
= LONG_MAX
;
2788 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2789 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2790 force_quiescent_state(rsp
);
2791 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2792 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2798 * RCU callback function to leak a callback.
2800 static void rcu_leak_callback(struct rcu_head
*rhp
)
2805 * Helper function for call_rcu() and friends. The cpu argument will
2806 * normally be -1, indicating "currently running CPU". It may specify
2807 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2808 * is expected to specify a CPU.
2811 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2812 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2814 unsigned long flags
;
2815 struct rcu_data
*rdp
;
2817 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
2818 if (debug_rcu_head_queue(head
)) {
2819 /* Probable double call_rcu(), so leak the callback. */
2820 ACCESS_ONCE(head
->func
) = rcu_leak_callback
;
2821 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2828 * Opportunistically note grace-period endings and beginnings.
2829 * Note that we might see a beginning right after we see an
2830 * end, but never vice versa, since this CPU has to pass through
2831 * a quiescent state betweentimes.
2833 local_irq_save(flags
);
2834 rdp
= this_cpu_ptr(rsp
->rda
);
2836 /* Add the callback to our list. */
2837 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2841 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2842 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
2843 WARN_ON_ONCE(offline
);
2844 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2845 local_irq_restore(flags
);
2848 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
+ 1;
2852 rcu_idle_count_callbacks_posted();
2853 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2854 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2855 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2857 if (__is_kfree_rcu_offset((unsigned long)func
))
2858 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2859 rdp
->qlen_lazy
, rdp
->qlen
);
2861 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2863 /* Go handle any RCU core processing required. */
2864 __call_rcu_core(rsp
, rdp
, head
, flags
);
2865 local_irq_restore(flags
);
2869 * Queue an RCU-sched callback for invocation after a grace period.
2871 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2873 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2875 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2878 * Queue an RCU callback for invocation after a quicker grace period.
2880 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2882 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2884 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2887 * Queue an RCU callback for lazy invocation after a grace period.
2888 * This will likely be later named something like "call_rcu_lazy()",
2889 * but this change will require some way of tagging the lazy RCU
2890 * callbacks in the list of pending callbacks. Until then, this
2891 * function may only be called from __kfree_rcu().
2893 void kfree_call_rcu(struct rcu_head
*head
,
2894 void (*func
)(struct rcu_head
*rcu
))
2896 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
2898 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
2901 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2902 * any blocking grace-period wait automatically implies a grace period
2903 * if there is only one CPU online at any point time during execution
2904 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2905 * occasionally incorrectly indicate that there are multiple CPUs online
2906 * when there was in fact only one the whole time, as this just adds
2907 * some overhead: RCU still operates correctly.
2909 static inline int rcu_blocking_is_gp(void)
2913 might_sleep(); /* Check for RCU read-side critical section. */
2915 ret
= num_online_cpus() <= 1;
2921 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2923 * Control will return to the caller some time after a full rcu-sched
2924 * grace period has elapsed, in other words after all currently executing
2925 * rcu-sched read-side critical sections have completed. These read-side
2926 * critical sections are delimited by rcu_read_lock_sched() and
2927 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2928 * local_irq_disable(), and so on may be used in place of
2929 * rcu_read_lock_sched().
2931 * This means that all preempt_disable code sequences, including NMI and
2932 * non-threaded hardware-interrupt handlers, in progress on entry will
2933 * have completed before this primitive returns. However, this does not
2934 * guarantee that softirq handlers will have completed, since in some
2935 * kernels, these handlers can run in process context, and can block.
2937 * Note that this guarantee implies further memory-ordering guarantees.
2938 * On systems with more than one CPU, when synchronize_sched() returns,
2939 * each CPU is guaranteed to have executed a full memory barrier since the
2940 * end of its last RCU-sched read-side critical section whose beginning
2941 * preceded the call to synchronize_sched(). In addition, each CPU having
2942 * an RCU read-side critical section that extends beyond the return from
2943 * synchronize_sched() is guaranteed to have executed a full memory barrier
2944 * after the beginning of synchronize_sched() and before the beginning of
2945 * that RCU read-side critical section. Note that these guarantees include
2946 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2947 * that are executing in the kernel.
2949 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2950 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2951 * to have executed a full memory barrier during the execution of
2952 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2953 * again only if the system has more than one CPU).
2955 * This primitive provides the guarantees made by the (now removed)
2956 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2957 * guarantees that rcu_read_lock() sections will have completed.
2958 * In "classic RCU", these two guarantees happen to be one and
2959 * the same, but can differ in realtime RCU implementations.
2961 void synchronize_sched(void)
2963 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2964 !lock_is_held(&rcu_lock_map
) &&
2965 !lock_is_held(&rcu_sched_lock_map
),
2966 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2967 if (rcu_blocking_is_gp())
2970 synchronize_sched_expedited();
2972 wait_rcu_gp(call_rcu_sched
);
2974 EXPORT_SYMBOL_GPL(synchronize_sched
);
2977 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2979 * Control will return to the caller some time after a full rcu_bh grace
2980 * period has elapsed, in other words after all currently executing rcu_bh
2981 * read-side critical sections have completed. RCU read-side critical
2982 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2983 * and may be nested.
2985 * See the description of synchronize_sched() for more detailed information
2986 * on memory ordering guarantees.
2988 void synchronize_rcu_bh(void)
2990 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2991 !lock_is_held(&rcu_lock_map
) &&
2992 !lock_is_held(&rcu_sched_lock_map
),
2993 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2994 if (rcu_blocking_is_gp())
2997 synchronize_rcu_bh_expedited();
2999 wait_rcu_gp(call_rcu_bh
);
3001 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3004 * get_state_synchronize_rcu - Snapshot current RCU state
3006 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3007 * to determine whether or not a full grace period has elapsed in the
3010 unsigned long get_state_synchronize_rcu(void)
3013 * Any prior manipulation of RCU-protected data must happen
3014 * before the load from ->gpnum.
3019 * Make sure this load happens before the purportedly
3020 * time-consuming work between get_state_synchronize_rcu()
3021 * and cond_synchronize_rcu().
3023 return smp_load_acquire(&rcu_state_p
->gpnum
);
3025 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3028 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3030 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3032 * If a full RCU grace period has elapsed since the earlier call to
3033 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3034 * synchronize_rcu() to wait for a full grace period.
3036 * Yes, this function does not take counter wrap into account. But
3037 * counter wrap is harmless. If the counter wraps, we have waited for
3038 * more than 2 billion grace periods (and way more on a 64-bit system!),
3039 * so waiting for one additional grace period should be just fine.
3041 void cond_synchronize_rcu(unsigned long oldstate
)
3043 unsigned long newstate
;
3046 * Ensure that this load happens before any RCU-destructive
3047 * actions the caller might carry out after we return.
3049 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3050 if (ULONG_CMP_GE(oldstate
, newstate
))
3053 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3055 static int synchronize_sched_expedited_cpu_stop(void *data
)
3058 * There must be a full memory barrier on each affected CPU
3059 * between the time that try_stop_cpus() is called and the
3060 * time that it returns.
3062 * In the current initial implementation of cpu_stop, the
3063 * above condition is already met when the control reaches
3064 * this point and the following smp_mb() is not strictly
3065 * necessary. Do smp_mb() anyway for documentation and
3066 * robustness against future implementation changes.
3068 smp_mb(); /* See above comment block. */
3073 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3075 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3076 * approach to force the grace period to end quickly. This consumes
3077 * significant time on all CPUs and is unfriendly to real-time workloads,
3078 * so is thus not recommended for any sort of common-case code. In fact,
3079 * if you are using synchronize_sched_expedited() in a loop, please
3080 * restructure your code to batch your updates, and then use a single
3081 * synchronize_sched() instead.
3083 * This implementation can be thought of as an application of ticket
3084 * locking to RCU, with sync_sched_expedited_started and
3085 * sync_sched_expedited_done taking on the roles of the halves
3086 * of the ticket-lock word. Each task atomically increments
3087 * sync_sched_expedited_started upon entry, snapshotting the old value,
3088 * then attempts to stop all the CPUs. If this succeeds, then each
3089 * CPU will have executed a context switch, resulting in an RCU-sched
3090 * grace period. We are then done, so we use atomic_cmpxchg() to
3091 * update sync_sched_expedited_done to match our snapshot -- but
3092 * only if someone else has not already advanced past our snapshot.
3094 * On the other hand, if try_stop_cpus() fails, we check the value
3095 * of sync_sched_expedited_done. If it has advanced past our
3096 * initial snapshot, then someone else must have forced a grace period
3097 * some time after we took our snapshot. In this case, our work is
3098 * done for us, and we can simply return. Otherwise, we try again,
3099 * but keep our initial snapshot for purposes of checking for someone
3100 * doing our work for us.
3102 * If we fail too many times in a row, we fall back to synchronize_sched().
3104 void synchronize_sched_expedited(void)
3109 long firstsnap
, s
, snap
;
3111 struct rcu_state
*rsp
= &rcu_sched_state
;
3114 * If we are in danger of counter wrap, just do synchronize_sched().
3115 * By allowing sync_sched_expedited_started to advance no more than
3116 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3117 * that more than 3.5 billion CPUs would be required to force a
3118 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3119 * course be required on a 64-bit system.
3121 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3122 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3124 synchronize_sched();
3125 atomic_long_inc(&rsp
->expedited_wrap
);
3130 * Take a ticket. Note that atomic_inc_return() implies a
3131 * full memory barrier.
3133 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3135 if (!try_get_online_cpus()) {
3136 /* CPU hotplug operation in flight, fall back to normal GP. */
3137 wait_rcu_gp(call_rcu_sched
);
3138 atomic_long_inc(&rsp
->expedited_normal
);
3141 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3143 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3144 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3146 cpumask_copy(cm
, cpu_online_mask
);
3147 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3148 for_each_cpu(cpu
, cm
) {
3149 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3151 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3152 cpumask_clear_cpu(cpu
, cm
);
3154 if (cpumask_weight(cm
) == 0)
3159 * Each pass through the following loop attempts to force a
3160 * context switch on each CPU.
3162 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3163 synchronize_sched_expedited_cpu_stop
,
3166 atomic_long_inc(&rsp
->expedited_tryfail
);
3168 /* Check to see if someone else did our work for us. */
3169 s
= atomic_long_read(&rsp
->expedited_done
);
3170 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3171 /* ensure test happens before caller kfree */
3172 smp_mb__before_atomic(); /* ^^^ */
3173 atomic_long_inc(&rsp
->expedited_workdone1
);
3174 free_cpumask_var(cm
);
3178 /* No joy, try again later. Or just synchronize_sched(). */
3179 if (trycount
++ < 10) {
3180 udelay(trycount
* num_online_cpus());
3182 wait_rcu_gp(call_rcu_sched
);
3183 atomic_long_inc(&rsp
->expedited_normal
);
3184 free_cpumask_var(cm
);
3188 /* Recheck to see if someone else did our work for us. */
3189 s
= atomic_long_read(&rsp
->expedited_done
);
3190 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3191 /* ensure test happens before caller kfree */
3192 smp_mb__before_atomic(); /* ^^^ */
3193 atomic_long_inc(&rsp
->expedited_workdone2
);
3194 free_cpumask_var(cm
);
3199 * Refetching sync_sched_expedited_started allows later
3200 * callers to piggyback on our grace period. We retry
3201 * after they started, so our grace period works for them,
3202 * and they started after our first try, so their grace
3203 * period works for us.
3205 if (!try_get_online_cpus()) {
3206 /* CPU hotplug operation in flight, use normal GP. */
3207 wait_rcu_gp(call_rcu_sched
);
3208 atomic_long_inc(&rsp
->expedited_normal
);
3209 free_cpumask_var(cm
);
3212 snap
= atomic_long_read(&rsp
->expedited_start
);
3213 smp_mb(); /* ensure read is before try_stop_cpus(). */
3215 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3218 free_cpumask_var(cm
);
3221 * Everyone up to our most recent fetch is covered by our grace
3222 * period. Update the counter, but only if our work is still
3223 * relevant -- which it won't be if someone who started later
3224 * than we did already did their update.
3227 atomic_long_inc(&rsp
->expedited_done_tries
);
3228 s
= atomic_long_read(&rsp
->expedited_done
);
3229 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3230 /* ensure test happens before caller kfree */
3231 smp_mb__before_atomic(); /* ^^^ */
3232 atomic_long_inc(&rsp
->expedited_done_lost
);
3235 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3236 atomic_long_inc(&rsp
->expedited_done_exit
);
3240 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3243 * Check to see if there is any immediate RCU-related work to be done
3244 * by the current CPU, for the specified type of RCU, returning 1 if so.
3245 * The checks are in order of increasing expense: checks that can be
3246 * carried out against CPU-local state are performed first. However,
3247 * we must check for CPU stalls first, else we might not get a chance.
3249 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3251 struct rcu_node
*rnp
= rdp
->mynode
;
3253 rdp
->n_rcu_pending
++;
3255 /* Check for CPU stalls, if enabled. */
3256 check_cpu_stall(rsp
, rdp
);
3258 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3259 if (rcu_nohz_full_cpu(rsp
))
3262 /* Is the RCU core waiting for a quiescent state from this CPU? */
3263 if (rcu_scheduler_fully_active
&&
3264 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3265 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3266 rdp
->n_rp_qs_pending
++;
3267 } else if (rdp
->qs_pending
&&
3268 (rdp
->passed_quiesce
||
3269 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3270 rdp
->n_rp_report_qs
++;
3274 /* Does this CPU have callbacks ready to invoke? */
3275 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3276 rdp
->n_rp_cb_ready
++;
3280 /* Has RCU gone idle with this CPU needing another grace period? */
3281 if (cpu_needs_another_gp(rsp
, rdp
)) {
3282 rdp
->n_rp_cpu_needs_gp
++;
3286 /* Has another RCU grace period completed? */
3287 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3288 rdp
->n_rp_gp_completed
++;
3292 /* Has a new RCU grace period started? */
3293 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3294 unlikely(ACCESS_ONCE(rdp
->gpwrap
))) { /* outside lock */
3295 rdp
->n_rp_gp_started
++;
3299 /* Does this CPU need a deferred NOCB wakeup? */
3300 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3301 rdp
->n_rp_nocb_defer_wakeup
++;
3306 rdp
->n_rp_need_nothing
++;
3311 * Check to see if there is any immediate RCU-related work to be done
3312 * by the current CPU, returning 1 if so. This function is part of the
3313 * RCU implementation; it is -not- an exported member of the RCU API.
3315 static int rcu_pending(void)
3317 struct rcu_state
*rsp
;
3319 for_each_rcu_flavor(rsp
)
3320 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3326 * Return true if the specified CPU has any callback. If all_lazy is
3327 * non-NULL, store an indication of whether all callbacks are lazy.
3328 * (If there are no callbacks, all of them are deemed to be lazy.)
3330 static int __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3334 struct rcu_data
*rdp
;
3335 struct rcu_state
*rsp
;
3337 for_each_rcu_flavor(rsp
) {
3338 rdp
= this_cpu_ptr(rsp
->rda
);
3342 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3353 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3354 * the compiler is expected to optimize this away.
3356 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3357 int cpu
, unsigned long done
)
3359 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3360 atomic_read(&rsp
->barrier_cpu_count
), done
);
3364 * RCU callback function for _rcu_barrier(). If we are last, wake
3365 * up the task executing _rcu_barrier().
3367 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3369 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3370 struct rcu_state
*rsp
= rdp
->rsp
;
3372 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3373 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3374 complete(&rsp
->barrier_completion
);
3376 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3381 * Called with preemption disabled, and from cross-cpu IRQ context.
3383 static void rcu_barrier_func(void *type
)
3385 struct rcu_state
*rsp
= type
;
3386 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3388 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3389 atomic_inc(&rsp
->barrier_cpu_count
);
3390 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3394 * Orchestrate the specified type of RCU barrier, waiting for all
3395 * RCU callbacks of the specified type to complete.
3397 static void _rcu_barrier(struct rcu_state
*rsp
)
3400 struct rcu_data
*rdp
;
3401 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
3402 unsigned long snap_done
;
3404 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3406 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3407 mutex_lock(&rsp
->barrier_mutex
);
3410 * Ensure that all prior references, including to ->n_barrier_done,
3411 * are ordered before the _rcu_barrier() machinery.
3413 smp_mb(); /* See above block comment. */
3416 * Recheck ->n_barrier_done to see if others did our work for us.
3417 * This means checking ->n_barrier_done for an even-to-odd-to-even
3418 * transition. The "if" expression below therefore rounds the old
3419 * value up to the next even number and adds two before comparing.
3421 snap_done
= rsp
->n_barrier_done
;
3422 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3425 * If the value in snap is odd, we needed to wait for the current
3426 * rcu_barrier() to complete, then wait for the next one, in other
3427 * words, we need the value of snap_done to be three larger than
3428 * the value of snap. On the other hand, if the value in snap is
3429 * even, we only had to wait for the next rcu_barrier() to complete,
3430 * in other words, we need the value of snap_done to be only two
3431 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3432 * this for us (thank you, Linus!).
3434 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3435 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3436 smp_mb(); /* caller's subsequent code after above check. */
3437 mutex_unlock(&rsp
->barrier_mutex
);
3442 * Increment ->n_barrier_done to avoid duplicate work. Use
3443 * ACCESS_ONCE() to prevent the compiler from speculating
3444 * the increment to precede the early-exit check.
3446 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3447 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3448 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3449 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3452 * Initialize the count to one rather than to zero in order to
3453 * avoid a too-soon return to zero in case of a short grace period
3454 * (or preemption of this task). Exclude CPU-hotplug operations
3455 * to ensure that no offline CPU has callbacks queued.
3457 init_completion(&rsp
->barrier_completion
);
3458 atomic_set(&rsp
->barrier_cpu_count
, 1);
3462 * Force each CPU with callbacks to register a new callback.
3463 * When that callback is invoked, we will know that all of the
3464 * corresponding CPU's preceding callbacks have been invoked.
3466 for_each_possible_cpu(cpu
) {
3467 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3469 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3470 if (rcu_is_nocb_cpu(cpu
)) {
3471 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3472 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3473 rsp
->n_barrier_done
);
3475 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3476 rsp
->n_barrier_done
);
3477 smp_mb__before_atomic();
3478 atomic_inc(&rsp
->barrier_cpu_count
);
3479 __call_rcu(&rdp
->barrier_head
,
3480 rcu_barrier_callback
, rsp
, cpu
, 0);
3482 } else if (ACCESS_ONCE(rdp
->qlen
)) {
3483 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3484 rsp
->n_barrier_done
);
3485 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3487 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3488 rsp
->n_barrier_done
);
3494 * Now that we have an rcu_barrier_callback() callback on each
3495 * CPU, and thus each counted, remove the initial count.
3497 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3498 complete(&rsp
->barrier_completion
);
3500 /* Increment ->n_barrier_done to prevent duplicate work. */
3501 smp_mb(); /* Keep increment after above mechanism. */
3502 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3503 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3504 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3505 smp_mb(); /* Keep increment before caller's subsequent code. */
3507 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3508 wait_for_completion(&rsp
->barrier_completion
);
3510 /* Other rcu_barrier() invocations can now safely proceed. */
3511 mutex_unlock(&rsp
->barrier_mutex
);
3515 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3517 void rcu_barrier_bh(void)
3519 _rcu_barrier(&rcu_bh_state
);
3521 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3524 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3526 void rcu_barrier_sched(void)
3528 _rcu_barrier(&rcu_sched_state
);
3530 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3533 * Do boot-time initialization of a CPU's per-CPU RCU data.
3536 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3538 unsigned long flags
;
3539 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3540 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3542 /* Set up local state, ensuring consistent view of global state. */
3543 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3544 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3545 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3546 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3547 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3550 rcu_boot_init_nocb_percpu_data(rdp
);
3551 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3555 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3556 * offline event can be happening at a given time. Note also that we
3557 * can accept some slop in the rsp->completed access due to the fact
3558 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3561 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3563 unsigned long flags
;
3565 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3566 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3568 /* Exclude new grace periods. */
3569 mutex_lock(&rsp
->onoff_mutex
);
3571 /* Set up local state, ensuring consistent view of global state. */
3572 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3573 rdp
->beenonline
= 1; /* We have now been online. */
3574 rdp
->qlen_last_fqs_check
= 0;
3575 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3576 rdp
->blimit
= blimit
;
3577 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3578 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3579 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3580 atomic_set(&rdp
->dynticks
->dynticks
,
3581 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3582 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3584 /* Add CPU to rcu_node bitmasks. */
3586 mask
= rdp
->grpmask
;
3588 /* Exclude any attempts to start a new GP on small systems. */
3589 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3590 rnp
->qsmaskinit
|= mask
;
3591 mask
= rnp
->grpmask
;
3592 if (rnp
== rdp
->mynode
) {
3594 * If there is a grace period in progress, we will
3595 * set up to wait for it next time we run the
3598 rdp
->gpnum
= rnp
->completed
;
3599 rdp
->completed
= rnp
->completed
;
3600 rdp
->passed_quiesce
= 0;
3601 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
3602 rdp
->qs_pending
= 0;
3603 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3605 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
3607 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
3608 local_irq_restore(flags
);
3610 mutex_unlock(&rsp
->onoff_mutex
);
3613 static void rcu_prepare_cpu(int cpu
)
3615 struct rcu_state
*rsp
;
3617 for_each_rcu_flavor(rsp
)
3618 rcu_init_percpu_data(cpu
, rsp
);
3622 * Handle CPU online/offline notification events.
3624 static int rcu_cpu_notify(struct notifier_block
*self
,
3625 unsigned long action
, void *hcpu
)
3627 long cpu
= (long)hcpu
;
3628 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3629 struct rcu_node
*rnp
= rdp
->mynode
;
3630 struct rcu_state
*rsp
;
3632 trace_rcu_utilization(TPS("Start CPU hotplug"));
3634 case CPU_UP_PREPARE
:
3635 case CPU_UP_PREPARE_FROZEN
:
3636 rcu_prepare_cpu(cpu
);
3637 rcu_prepare_kthreads(cpu
);
3638 rcu_spawn_all_nocb_kthreads(cpu
);
3641 case CPU_DOWN_FAILED
:
3642 rcu_boost_kthread_setaffinity(rnp
, -1);
3644 case CPU_DOWN_PREPARE
:
3645 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3648 case CPU_DYING_FROZEN
:
3649 for_each_rcu_flavor(rsp
)
3650 rcu_cleanup_dying_cpu(rsp
);
3653 case CPU_DEAD_FROZEN
:
3654 case CPU_UP_CANCELED
:
3655 case CPU_UP_CANCELED_FROZEN
:
3656 for_each_rcu_flavor(rsp
) {
3657 rcu_cleanup_dead_cpu(cpu
, rsp
);
3658 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3664 trace_rcu_utilization(TPS("End CPU hotplug"));
3668 static int rcu_pm_notify(struct notifier_block
*self
,
3669 unsigned long action
, void *hcpu
)
3672 case PM_HIBERNATION_PREPARE
:
3673 case PM_SUSPEND_PREPARE
:
3674 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3677 case PM_POST_HIBERNATION
:
3678 case PM_POST_SUSPEND
:
3688 * Spawn the kthreads that handle each RCU flavor's grace periods.
3690 static int __init
rcu_spawn_gp_kthread(void)
3692 unsigned long flags
;
3693 int kthread_prio_in
= kthread_prio
;
3694 struct rcu_node
*rnp
;
3695 struct rcu_state
*rsp
;
3696 struct sched_param sp
;
3697 struct task_struct
*t
;
3699 /* Force priority into range. */
3700 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3702 else if (kthread_prio
< 0)
3704 else if (kthread_prio
> 99)
3706 if (kthread_prio
!= kthread_prio_in
)
3707 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3708 kthread_prio
, kthread_prio_in
);
3710 rcu_scheduler_fully_active
= 1;
3711 for_each_rcu_flavor(rsp
) {
3712 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3714 rnp
= rcu_get_root(rsp
);
3715 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3716 rsp
->gp_kthread
= t
;
3718 sp
.sched_priority
= kthread_prio
;
3719 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3722 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3724 rcu_spawn_nocb_kthreads();
3725 rcu_spawn_boost_kthreads();
3728 early_initcall(rcu_spawn_gp_kthread
);
3731 * This function is invoked towards the end of the scheduler's initialization
3732 * process. Before this is called, the idle task might contain
3733 * RCU read-side critical sections (during which time, this idle
3734 * task is booting the system). After this function is called, the
3735 * idle tasks are prohibited from containing RCU read-side critical
3736 * sections. This function also enables RCU lockdep checking.
3738 void rcu_scheduler_starting(void)
3740 WARN_ON(num_online_cpus() != 1);
3741 WARN_ON(nr_context_switches() > 0);
3742 rcu_scheduler_active
= 1;
3746 * Compute the per-level fanout, either using the exact fanout specified
3747 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3749 #ifdef CONFIG_RCU_FANOUT_EXACT
3750 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3754 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3755 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3756 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3758 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3759 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3766 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3767 ccur
= rsp
->levelcnt
[i
];
3768 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3772 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3775 * Helper function for rcu_init() that initializes one rcu_state structure.
3777 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3778 struct rcu_data __percpu
*rda
)
3780 static const char * const buf
[] = {
3784 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3785 static const char * const fqs
[] = {
3789 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3790 static u8 fl_mask
= 0x1;
3794 struct rcu_node
*rnp
;
3796 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3798 /* Silence gcc 4.8 warning about array index out of range. */
3799 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3800 panic("rcu_init_one: rcu_num_lvls overflow");
3802 /* Initialize the level-tracking arrays. */
3804 for (i
= 0; i
< rcu_num_lvls
; i
++)
3805 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3806 for (i
= 1; i
< rcu_num_lvls
; i
++)
3807 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3808 rcu_init_levelspread(rsp
);
3809 rsp
->flavor_mask
= fl_mask
;
3812 /* Initialize the elements themselves, starting from the leaves. */
3814 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3815 cpustride
*= rsp
->levelspread
[i
];
3816 rnp
= rsp
->level
[i
];
3817 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3818 raw_spin_lock_init(&rnp
->lock
);
3819 lockdep_set_class_and_name(&rnp
->lock
,
3820 &rcu_node_class
[i
], buf
[i
]);
3821 raw_spin_lock_init(&rnp
->fqslock
);
3822 lockdep_set_class_and_name(&rnp
->fqslock
,
3823 &rcu_fqs_class
[i
], fqs
[i
]);
3824 rnp
->gpnum
= rsp
->gpnum
;
3825 rnp
->completed
= rsp
->completed
;
3827 rnp
->qsmaskinit
= 0;
3828 rnp
->grplo
= j
* cpustride
;
3829 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
3830 if (rnp
->grphi
>= nr_cpu_ids
)
3831 rnp
->grphi
= nr_cpu_ids
- 1;
3837 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
3838 rnp
->grpmask
= 1UL << rnp
->grpnum
;
3839 rnp
->parent
= rsp
->level
[i
- 1] +
3840 j
/ rsp
->levelspread
[i
- 1];
3843 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3844 rcu_init_one_nocb(rnp
);
3849 init_waitqueue_head(&rsp
->gp_wq
);
3850 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3851 for_each_possible_cpu(i
) {
3852 while (i
> rnp
->grphi
)
3854 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3855 rcu_boot_init_percpu_data(i
, rsp
);
3857 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3861 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3862 * replace the definitions in tree.h because those are needed to size
3863 * the ->node array in the rcu_state structure.
3865 static void __init
rcu_init_geometry(void)
3871 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3874 * Initialize any unspecified boot parameters.
3875 * The default values of jiffies_till_first_fqs and
3876 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3877 * value, which is a function of HZ, then adding one for each
3878 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3880 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
3881 if (jiffies_till_first_fqs
== ULONG_MAX
)
3882 jiffies_till_first_fqs
= d
;
3883 if (jiffies_till_next_fqs
== ULONG_MAX
)
3884 jiffies_till_next_fqs
= d
;
3886 /* If the compile-time values are accurate, just leave. */
3887 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3888 nr_cpu_ids
== NR_CPUS
)
3890 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3891 rcu_fanout_leaf
, nr_cpu_ids
);
3894 * Compute number of nodes that can be handled an rcu_node tree
3895 * with the given number of levels. Setting rcu_capacity[0] makes
3896 * some of the arithmetic easier.
3898 rcu_capacity
[0] = 1;
3899 rcu_capacity
[1] = rcu_fanout_leaf
;
3900 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3901 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3904 * The boot-time rcu_fanout_leaf parameter is only permitted
3905 * to increase the leaf-level fanout, not decrease it. Of course,
3906 * the leaf-level fanout cannot exceed the number of bits in
3907 * the rcu_node masks. Finally, the tree must be able to accommodate
3908 * the configured number of CPUs. Complain and fall back to the
3909 * compile-time values if these limits are exceeded.
3911 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3912 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3913 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3918 /* Calculate the number of rcu_nodes at each level of the tree. */
3919 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3920 if (n
<= rcu_capacity
[i
]) {
3921 for (j
= 0; j
<= i
; j
++)
3923 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3925 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3930 /* Calculate the total number of rcu_node structures. */
3932 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3933 rcu_num_nodes
+= num_rcu_lvl
[i
];
3937 void __init
rcu_init(void)
3941 rcu_bootup_announce();
3942 rcu_init_geometry();
3943 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3944 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3945 __rcu_init_preempt();
3946 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3949 * We don't need protection against CPU-hotplug here because
3950 * this is called early in boot, before either interrupts
3951 * or the scheduler are operational.
3953 cpu_notifier(rcu_cpu_notify
, 0);
3954 pm_notifier(rcu_pm_notify
, 0);
3955 for_each_online_cpu(cpu
)
3956 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3958 rcu_early_boot_tests();
3961 #include "tree_plugin.h"