2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
72 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
83 # define DEFINE_RCU_TPS(sname) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
86 # define RCU_STATE_NAME(sname) sname##_varname
88 # define DEFINE_RCU_TPS(sname)
89 # define RCU_STATE_NAME(sname) __stringify(sname)
92 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
93 DEFINE_RCU_TPS(sname) \
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
95 struct rcu_state sname##_state = { \
96 .level = { &sname##_state.node[0] }, \
97 .rda = &sname##_data, \
99 .fqs_state = RCU_GP_IDLE, \
100 .gpnum = 0UL - 300UL, \
101 .completed = 0UL - 300UL, \
102 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
103 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
104 .orphan_donetail = &sname##_state.orphan_donelist, \
105 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
106 .name = RCU_STATE_NAME(sname), \
110 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
111 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
113 static struct rcu_state
*rcu_state_p
;
114 LIST_HEAD(rcu_struct_flavors
);
116 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
117 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
118 module_param(rcu_fanout_leaf
, int, 0444);
119 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
120 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
127 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
130 * The rcu_scheduler_active variable transitions from zero to one just
131 * before the first task is spawned. So when this variable is zero, RCU
132 * can assume that there is but one task, allowing RCU to (for example)
133 * optimize synchronize_sched() to a simple barrier(). When this variable
134 * is one, RCU must actually do all the hard work required to detect real
135 * grace periods. This variable is also used to suppress boot-time false
136 * positives from lockdep-RCU error checking.
138 int rcu_scheduler_active __read_mostly
;
139 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
142 * The rcu_scheduler_fully_active variable transitions from zero to one
143 * during the early_initcall() processing, which is after the scheduler
144 * is capable of creating new tasks. So RCU processing (for example,
145 * creating tasks for RCU priority boosting) must be delayed until after
146 * rcu_scheduler_fully_active transitions from zero to one. We also
147 * currently delay invocation of any RCU callbacks until after this point.
149 * It might later prove better for people registering RCU callbacks during
150 * early boot to take responsibility for these callbacks, but one step at
153 static int rcu_scheduler_fully_active __read_mostly
;
155 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
156 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
157 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
158 static void invoke_rcu_core(void);
159 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
161 /* rcuc/rcub kthread realtime priority */
162 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
163 module_param(kthread_prio
, int, 0644);
165 /* Delay in jiffies for grace-period initialization delays, debug only. */
166 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
167 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
168 module_param(gp_init_delay
, int, 0644);
169 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
170 static const int gp_init_delay
;
171 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
172 #define PER_RCU_NODE_PERIOD 10 /* Number of grace periods between delays. */
175 * Track the rcutorture test sequence number and the update version
176 * number within a given test. The rcutorture_testseq is incremented
177 * on every rcutorture module load and unload, so has an odd value
178 * when a test is running. The rcutorture_vernum is set to zero
179 * when rcutorture starts and is incremented on each rcutorture update.
180 * These variables enable correlating rcutorture output with the
181 * RCU tracing information.
183 unsigned long rcutorture_testseq
;
184 unsigned long rcutorture_vernum
;
187 * Compute the mask of online CPUs for the specified rcu_node structure.
188 * This will not be stable unless the rcu_node structure's ->lock is
189 * held, but the bit corresponding to the current CPU will be stable
192 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
194 return ACCESS_ONCE(rnp
->qsmaskinitnext
);
198 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
199 * permit this function to be invoked without holding the root rcu_node
200 * structure's ->lock, but of course results can be subject to change.
202 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
204 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
208 * Note a quiescent state. Because we do not need to know
209 * how many quiescent states passed, just if there was at least
210 * one since the start of the grace period, this just sets a flag.
211 * The caller must have disabled preemption.
213 void rcu_sched_qs(void)
215 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
216 trace_rcu_grace_period(TPS("rcu_sched"),
217 __this_cpu_read(rcu_sched_data
.gpnum
),
219 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
225 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
226 trace_rcu_grace_period(TPS("rcu_bh"),
227 __this_cpu_read(rcu_bh_data
.gpnum
),
229 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
233 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
235 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
236 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
237 .dynticks
= ATOMIC_INIT(1),
238 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
239 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
240 .dynticks_idle
= ATOMIC_INIT(1),
241 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
244 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
245 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
248 * Let the RCU core know that this CPU has gone through the scheduler,
249 * which is a quiescent state. This is called when the need for a
250 * quiescent state is urgent, so we burn an atomic operation and full
251 * memory barriers to let the RCU core know about it, regardless of what
252 * this CPU might (or might not) do in the near future.
254 * We inform the RCU core by emulating a zero-duration dyntick-idle
255 * period, which we in turn do by incrementing the ->dynticks counter
258 static void rcu_momentary_dyntick_idle(void)
261 struct rcu_data
*rdp
;
262 struct rcu_dynticks
*rdtp
;
264 struct rcu_state
*rsp
;
266 local_irq_save(flags
);
269 * Yes, we can lose flag-setting operations. This is OK, because
270 * the flag will be set again after some delay.
272 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
273 raw_cpu_write(rcu_sched_qs_mask
, 0);
275 /* Find the flavor that needs a quiescent state. */
276 for_each_rcu_flavor(rsp
) {
277 rdp
= raw_cpu_ptr(rsp
->rda
);
278 if (!(resched_mask
& rsp
->flavor_mask
))
280 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
281 if (ACCESS_ONCE(rdp
->mynode
->completed
) !=
282 ACCESS_ONCE(rdp
->cond_resched_completed
))
286 * Pretend to be momentarily idle for the quiescent state.
287 * This allows the grace-period kthread to record the
288 * quiescent state, with no need for this CPU to do anything
291 rdtp
= this_cpu_ptr(&rcu_dynticks
);
292 smp_mb__before_atomic(); /* Earlier stuff before QS. */
293 atomic_add(2, &rdtp
->dynticks
); /* QS. */
294 smp_mb__after_atomic(); /* Later stuff after QS. */
297 local_irq_restore(flags
);
301 * Note a context switch. This is a quiescent state for RCU-sched,
302 * and requires special handling for preemptible RCU.
303 * The caller must have disabled preemption.
305 void rcu_note_context_switch(void)
307 trace_rcu_utilization(TPS("Start context switch"));
309 rcu_preempt_note_context_switch();
310 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
311 rcu_momentary_dyntick_idle();
312 trace_rcu_utilization(TPS("End context switch"));
314 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
317 * Register a quiescent state for all RCU flavors. If there is an
318 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
319 * dyntick-idle quiescent state visible to other CPUs (but only for those
320 * RCU flavors in desperate need of a quiescent state, which will normally
321 * be none of them). Either way, do a lightweight quiescent state for
324 void rcu_all_qs(void)
326 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
327 rcu_momentary_dyntick_idle();
328 this_cpu_inc(rcu_qs_ctr
);
330 EXPORT_SYMBOL_GPL(rcu_all_qs
);
332 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
333 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
334 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
336 module_param(blimit
, long, 0444);
337 module_param(qhimark
, long, 0444);
338 module_param(qlowmark
, long, 0444);
340 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
341 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
343 module_param(jiffies_till_first_fqs
, ulong
, 0644);
344 module_param(jiffies_till_next_fqs
, ulong
, 0644);
347 * How long the grace period must be before we start recruiting
348 * quiescent-state help from rcu_note_context_switch().
350 static ulong jiffies_till_sched_qs
= HZ
/ 20;
351 module_param(jiffies_till_sched_qs
, ulong
, 0644);
353 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
354 struct rcu_data
*rdp
);
355 static void force_qs_rnp(struct rcu_state
*rsp
,
356 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
357 unsigned long *maxj
),
358 bool *isidle
, unsigned long *maxj
);
359 static void force_quiescent_state(struct rcu_state
*rsp
);
360 static int rcu_pending(void);
363 * Return the number of RCU batches started thus far for debug & stats.
365 unsigned long rcu_batches_started(void)
367 return rcu_state_p
->gpnum
;
369 EXPORT_SYMBOL_GPL(rcu_batches_started
);
372 * Return the number of RCU-sched batches started thus far for debug & stats.
374 unsigned long rcu_batches_started_sched(void)
376 return rcu_sched_state
.gpnum
;
378 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
381 * Return the number of RCU BH batches started thus far for debug & stats.
383 unsigned long rcu_batches_started_bh(void)
385 return rcu_bh_state
.gpnum
;
387 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
390 * Return the number of RCU batches completed thus far for debug & stats.
392 unsigned long rcu_batches_completed(void)
394 return rcu_state_p
->completed
;
396 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
399 * Return the number of RCU-sched batches completed thus far for debug & stats.
401 unsigned long rcu_batches_completed_sched(void)
403 return rcu_sched_state
.completed
;
405 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
408 * Return the number of RCU BH batches completed thus far for debug & stats.
410 unsigned long rcu_batches_completed_bh(void)
412 return rcu_bh_state
.completed
;
414 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
417 * Force a quiescent state.
419 void rcu_force_quiescent_state(void)
421 force_quiescent_state(rcu_state_p
);
423 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
426 * Force a quiescent state for RCU BH.
428 void rcu_bh_force_quiescent_state(void)
430 force_quiescent_state(&rcu_bh_state
);
432 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
435 * Force a quiescent state for RCU-sched.
437 void rcu_sched_force_quiescent_state(void)
439 force_quiescent_state(&rcu_sched_state
);
441 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
444 * Show the state of the grace-period kthreads.
446 void show_rcu_gp_kthreads(void)
448 struct rcu_state
*rsp
;
450 for_each_rcu_flavor(rsp
) {
451 pr_info("%s: wait state: %d ->state: %#lx\n",
452 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
453 /* sched_show_task(rsp->gp_kthread); */
456 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
459 * Record the number of times rcutorture tests have been initiated and
460 * terminated. This information allows the debugfs tracing stats to be
461 * correlated to the rcutorture messages, even when the rcutorture module
462 * is being repeatedly loaded and unloaded. In other words, we cannot
463 * store this state in rcutorture itself.
465 void rcutorture_record_test_transition(void)
467 rcutorture_testseq
++;
468 rcutorture_vernum
= 0;
470 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
473 * Send along grace-period-related data for rcutorture diagnostics.
475 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
476 unsigned long *gpnum
, unsigned long *completed
)
478 struct rcu_state
*rsp
= NULL
;
487 case RCU_SCHED_FLAVOR
:
488 rsp
= &rcu_sched_state
;
494 *flags
= ACCESS_ONCE(rsp
->gp_flags
);
495 *gpnum
= ACCESS_ONCE(rsp
->gpnum
);
496 *completed
= ACCESS_ONCE(rsp
->completed
);
503 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
506 * Record the number of writer passes through the current rcutorture test.
507 * This is also used to correlate debugfs tracing stats with the rcutorture
510 void rcutorture_record_progress(unsigned long vernum
)
514 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
517 * Does the CPU have callbacks ready to be invoked?
520 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
522 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
523 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
527 * Return the root node of the specified rcu_state structure.
529 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
531 return &rsp
->node
[0];
535 * Is there any need for future grace periods?
536 * Interrupts must be disabled. If the caller does not hold the root
537 * rnp_node structure's ->lock, the results are advisory only.
539 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
541 struct rcu_node
*rnp
= rcu_get_root(rsp
);
542 int idx
= (ACCESS_ONCE(rnp
->completed
) + 1) & 0x1;
543 int *fp
= &rnp
->need_future_gp
[idx
];
545 return ACCESS_ONCE(*fp
);
549 * Does the current CPU require a not-yet-started grace period?
550 * The caller must have disabled interrupts to prevent races with
551 * normal callback registry.
554 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
558 if (rcu_gp_in_progress(rsp
))
559 return 0; /* No, a grace period is already in progress. */
560 if (rcu_future_needs_gp(rsp
))
561 return 1; /* Yes, a no-CBs CPU needs one. */
562 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
563 return 0; /* No, this is a no-CBs (or offline) CPU. */
564 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
565 return 1; /* Yes, this CPU has newly registered callbacks. */
566 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
567 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
568 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
569 rdp
->nxtcompleted
[i
]))
570 return 1; /* Yes, CBs for future grace period. */
571 return 0; /* No grace period needed. */
575 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
577 * If the new value of the ->dynticks_nesting counter now is zero,
578 * we really have entered idle, and must do the appropriate accounting.
579 * The caller must have disabled interrupts.
581 static void rcu_eqs_enter_common(long long oldval
, bool user
)
583 struct rcu_state
*rsp
;
584 struct rcu_data
*rdp
;
585 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
587 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
588 if (!user
&& !is_idle_task(current
)) {
589 struct task_struct
*idle __maybe_unused
=
590 idle_task(smp_processor_id());
592 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
593 ftrace_dump(DUMP_ORIG
);
594 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
595 current
->pid
, current
->comm
,
596 idle
->pid
, idle
->comm
); /* must be idle task! */
598 for_each_rcu_flavor(rsp
) {
599 rdp
= this_cpu_ptr(rsp
->rda
);
600 do_nocb_deferred_wakeup(rdp
);
602 rcu_prepare_for_idle();
603 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
604 smp_mb__before_atomic(); /* See above. */
605 atomic_inc(&rdtp
->dynticks
);
606 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
607 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
608 rcu_dynticks_task_enter();
611 * It is illegal to enter an extended quiescent state while
612 * in an RCU read-side critical section.
614 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
615 "Illegal idle entry in RCU read-side critical section.");
616 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
617 "Illegal idle entry in RCU-bh read-side critical section.");
618 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
619 "Illegal idle entry in RCU-sched read-side critical section.");
623 * Enter an RCU extended quiescent state, which can be either the
624 * idle loop or adaptive-tickless usermode execution.
626 static void rcu_eqs_enter(bool user
)
629 struct rcu_dynticks
*rdtp
;
631 rdtp
= this_cpu_ptr(&rcu_dynticks
);
632 oldval
= rdtp
->dynticks_nesting
;
633 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
634 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
635 rdtp
->dynticks_nesting
= 0;
636 rcu_eqs_enter_common(oldval
, user
);
638 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
643 * rcu_idle_enter - inform RCU that current CPU is entering idle
645 * Enter idle mode, in other words, -leave- the mode in which RCU
646 * read-side critical sections can occur. (Though RCU read-side
647 * critical sections can occur in irq handlers in idle, a possibility
648 * handled by irq_enter() and irq_exit().)
650 * We crowbar the ->dynticks_nesting field to zero to allow for
651 * the possibility of usermode upcalls having messed up our count
652 * of interrupt nesting level during the prior busy period.
654 void rcu_idle_enter(void)
658 local_irq_save(flags
);
659 rcu_eqs_enter(false);
660 rcu_sysidle_enter(0);
661 local_irq_restore(flags
);
663 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
665 #ifdef CONFIG_RCU_USER_QS
667 * rcu_user_enter - inform RCU that we are resuming userspace.
669 * Enter RCU idle mode right before resuming userspace. No use of RCU
670 * is permitted between this call and rcu_user_exit(). This way the
671 * CPU doesn't need to maintain the tick for RCU maintenance purposes
672 * when the CPU runs in userspace.
674 void rcu_user_enter(void)
678 #endif /* CONFIG_RCU_USER_QS */
681 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
683 * Exit from an interrupt handler, which might possibly result in entering
684 * idle mode, in other words, leaving the mode in which read-side critical
685 * sections can occur.
687 * This code assumes that the idle loop never does anything that might
688 * result in unbalanced calls to irq_enter() and irq_exit(). If your
689 * architecture violates this assumption, RCU will give you what you
690 * deserve, good and hard. But very infrequently and irreproducibly.
692 * Use things like work queues to work around this limitation.
694 * You have been warned.
696 void rcu_irq_exit(void)
700 struct rcu_dynticks
*rdtp
;
702 local_irq_save(flags
);
703 rdtp
= this_cpu_ptr(&rcu_dynticks
);
704 oldval
= rdtp
->dynticks_nesting
;
705 rdtp
->dynticks_nesting
--;
706 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
707 if (rdtp
->dynticks_nesting
)
708 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
710 rcu_eqs_enter_common(oldval
, true);
711 rcu_sysidle_enter(1);
712 local_irq_restore(flags
);
716 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
718 * If the new value of the ->dynticks_nesting counter was previously zero,
719 * we really have exited idle, and must do the appropriate accounting.
720 * The caller must have disabled interrupts.
722 static void rcu_eqs_exit_common(long long oldval
, int user
)
724 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
726 rcu_dynticks_task_exit();
727 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
728 atomic_inc(&rdtp
->dynticks
);
729 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
730 smp_mb__after_atomic(); /* See above. */
731 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
732 rcu_cleanup_after_idle();
733 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
734 if (!user
&& !is_idle_task(current
)) {
735 struct task_struct
*idle __maybe_unused
=
736 idle_task(smp_processor_id());
738 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
739 oldval
, rdtp
->dynticks_nesting
);
740 ftrace_dump(DUMP_ORIG
);
741 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
742 current
->pid
, current
->comm
,
743 idle
->pid
, idle
->comm
); /* must be idle task! */
748 * Exit an RCU extended quiescent state, which can be either the
749 * idle loop or adaptive-tickless usermode execution.
751 static void rcu_eqs_exit(bool user
)
753 struct rcu_dynticks
*rdtp
;
756 rdtp
= this_cpu_ptr(&rcu_dynticks
);
757 oldval
= rdtp
->dynticks_nesting
;
758 WARN_ON_ONCE(oldval
< 0);
759 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
760 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
762 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
763 rcu_eqs_exit_common(oldval
, user
);
768 * rcu_idle_exit - inform RCU that current CPU is leaving idle
770 * Exit idle mode, in other words, -enter- the mode in which RCU
771 * read-side critical sections can occur.
773 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
774 * allow for the possibility of usermode upcalls messing up our count
775 * of interrupt nesting level during the busy period that is just
778 void rcu_idle_exit(void)
782 local_irq_save(flags
);
785 local_irq_restore(flags
);
787 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
789 #ifdef CONFIG_RCU_USER_QS
791 * rcu_user_exit - inform RCU that we are exiting userspace.
793 * Exit RCU idle mode while entering the kernel because it can
794 * run a RCU read side critical section anytime.
796 void rcu_user_exit(void)
800 #endif /* CONFIG_RCU_USER_QS */
803 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
805 * Enter an interrupt handler, which might possibly result in exiting
806 * idle mode, in other words, entering the mode in which read-side critical
807 * sections can occur.
809 * Note that the Linux kernel is fully capable of entering an interrupt
810 * handler that it never exits, for example when doing upcalls to
811 * user mode! This code assumes that the idle loop never does upcalls to
812 * user mode. If your architecture does do upcalls from the idle loop (or
813 * does anything else that results in unbalanced calls to the irq_enter()
814 * and irq_exit() functions), RCU will give you what you deserve, good
815 * and hard. But very infrequently and irreproducibly.
817 * Use things like work queues to work around this limitation.
819 * You have been warned.
821 void rcu_irq_enter(void)
824 struct rcu_dynticks
*rdtp
;
827 local_irq_save(flags
);
828 rdtp
= this_cpu_ptr(&rcu_dynticks
);
829 oldval
= rdtp
->dynticks_nesting
;
830 rdtp
->dynticks_nesting
++;
831 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
833 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
835 rcu_eqs_exit_common(oldval
, true);
837 local_irq_restore(flags
);
841 * rcu_nmi_enter - inform RCU of entry to NMI context
843 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
844 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
845 * that the CPU is active. This implementation permits nested NMIs, as
846 * long as the nesting level does not overflow an int. (You will probably
847 * run out of stack space first.)
849 void rcu_nmi_enter(void)
851 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
854 /* Complain about underflow. */
855 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
858 * If idle from RCU viewpoint, atomically increment ->dynticks
859 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
860 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
861 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
862 * to be in the outermost NMI handler that interrupted an RCU-idle
863 * period (observation due to Andy Lutomirski).
865 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
866 smp_mb__before_atomic(); /* Force delay from prior write. */
867 atomic_inc(&rdtp
->dynticks
);
868 /* atomic_inc() before later RCU read-side crit sects */
869 smp_mb__after_atomic(); /* See above. */
870 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
873 rdtp
->dynticks_nmi_nesting
+= incby
;
878 * rcu_nmi_exit - inform RCU of exit from NMI context
880 * If we are returning from the outermost NMI handler that interrupted an
881 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
882 * to let the RCU grace-period handling know that the CPU is back to
885 void rcu_nmi_exit(void)
887 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
890 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
891 * (We are exiting an NMI handler, so RCU better be paying attention
894 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
895 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
898 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
899 * leave it in non-RCU-idle state.
901 if (rdtp
->dynticks_nmi_nesting
!= 1) {
902 rdtp
->dynticks_nmi_nesting
-= 2;
906 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
907 rdtp
->dynticks_nmi_nesting
= 0;
908 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
909 smp_mb__before_atomic(); /* See above. */
910 atomic_inc(&rdtp
->dynticks
);
911 smp_mb__after_atomic(); /* Force delay to next write. */
912 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
916 * __rcu_is_watching - are RCU read-side critical sections safe?
918 * Return true if RCU is watching the running CPU, which means that
919 * this CPU can safely enter RCU read-side critical sections. Unlike
920 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
921 * least disabled preemption.
923 bool notrace
__rcu_is_watching(void)
925 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
929 * rcu_is_watching - see if RCU thinks that the current CPU is idle
931 * If the current CPU is in its idle loop and is neither in an interrupt
932 * or NMI handler, return true.
934 bool notrace
rcu_is_watching(void)
939 ret
= __rcu_is_watching();
943 EXPORT_SYMBOL_GPL(rcu_is_watching
);
945 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
948 * Is the current CPU online? Disable preemption to avoid false positives
949 * that could otherwise happen due to the current CPU number being sampled,
950 * this task being preempted, its old CPU being taken offline, resuming
951 * on some other CPU, then determining that its old CPU is now offline.
952 * It is OK to use RCU on an offline processor during initial boot, hence
953 * the check for rcu_scheduler_fully_active. Note also that it is OK
954 * for a CPU coming online to use RCU for one jiffy prior to marking itself
955 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
956 * offline to continue to use RCU for one jiffy after marking itself
957 * offline in the cpu_online_mask. This leniency is necessary given the
958 * non-atomic nature of the online and offline processing, for example,
959 * the fact that a CPU enters the scheduler after completing the CPU_DYING
962 * This is also why RCU internally marks CPUs online during the
963 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
965 * Disable checking if in an NMI handler because we cannot safely report
966 * errors from NMI handlers anyway.
968 bool rcu_lockdep_current_cpu_online(void)
970 struct rcu_data
*rdp
;
971 struct rcu_node
*rnp
;
977 rdp
= this_cpu_ptr(&rcu_sched_data
);
979 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
980 !rcu_scheduler_fully_active
;
984 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
986 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
989 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
991 * If the current CPU is idle or running at a first-level (not nested)
992 * interrupt from idle, return true. The caller must have at least
993 * disabled preemption.
995 static int rcu_is_cpu_rrupt_from_idle(void)
997 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1001 * Snapshot the specified CPU's dynticks counter so that we can later
1002 * credit them with an implicit quiescent state. Return 1 if this CPU
1003 * is in dynticks idle mode, which is an extended quiescent state.
1005 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1006 bool *isidle
, unsigned long *maxj
)
1008 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1009 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1010 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1011 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1014 if (ULONG_CMP_LT(ACCESS_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1015 rdp
->mynode
->gpnum
))
1016 ACCESS_ONCE(rdp
->gpwrap
) = true;
1022 * Return true if the specified CPU has passed through a quiescent
1023 * state by virtue of being in or having passed through an dynticks
1024 * idle state since the last call to dyntick_save_progress_counter()
1025 * for this same CPU, or by virtue of having been offline.
1027 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1028 bool *isidle
, unsigned long *maxj
)
1034 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1035 snap
= (unsigned int)rdp
->dynticks_snap
;
1038 * If the CPU passed through or entered a dynticks idle phase with
1039 * no active irq/NMI handlers, then we can safely pretend that the CPU
1040 * already acknowledged the request to pass through a quiescent
1041 * state. Either way, that CPU cannot possibly be in an RCU
1042 * read-side critical section that started before the beginning
1043 * of the current RCU grace period.
1045 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1046 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1047 rdp
->dynticks_fqs
++;
1052 * Check for the CPU being offline, but only if the grace period
1053 * is old enough. We don't need to worry about the CPU changing
1054 * state: If we see it offline even once, it has been through a
1057 * The reason for insisting that the grace period be at least
1058 * one jiffy old is that CPUs that are not quite online and that
1059 * have just gone offline can still execute RCU read-side critical
1062 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1063 return 0; /* Grace period is not old enough. */
1065 if (cpu_is_offline(rdp
->cpu
)) {
1066 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1072 * A CPU running for an extended time within the kernel can
1073 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1074 * even context-switching back and forth between a pair of
1075 * in-kernel CPU-bound tasks cannot advance grace periods.
1076 * So if the grace period is old enough, make the CPU pay attention.
1077 * Note that the unsynchronized assignments to the per-CPU
1078 * rcu_sched_qs_mask variable are safe. Yes, setting of
1079 * bits can be lost, but they will be set again on the next
1080 * force-quiescent-state pass. So lost bit sets do not result
1081 * in incorrect behavior, merely in a grace period lasting
1082 * a few jiffies longer than it might otherwise. Because
1083 * there are at most four threads involved, and because the
1084 * updates are only once every few jiffies, the probability of
1085 * lossage (and thus of slight grace-period extension) is
1088 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1089 * is set too high, we override with half of the RCU CPU stall
1092 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1093 if (ULONG_CMP_GE(jiffies
,
1094 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1095 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1096 if (!(ACCESS_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1097 ACCESS_ONCE(rdp
->cond_resched_completed
) =
1098 ACCESS_ONCE(rdp
->mynode
->completed
);
1099 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1100 ACCESS_ONCE(*rcrmp
) =
1101 ACCESS_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
;
1102 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1103 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1104 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1105 /* Time to beat on that CPU again! */
1106 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1107 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1114 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1116 unsigned long j
= jiffies
;
1120 smp_wmb(); /* Record start time before stall time. */
1121 j1
= rcu_jiffies_till_stall_check();
1122 ACCESS_ONCE(rsp
->jiffies_stall
) = j
+ j1
;
1123 rsp
->jiffies_resched
= j
+ j1
/ 2;
1124 rsp
->n_force_qs_gpstart
= ACCESS_ONCE(rsp
->n_force_qs
);
1128 * Complain about starvation of grace-period kthread.
1130 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1136 gpa
= ACCESS_ONCE(rsp
->gp_activity
);
1137 if (j
- gpa
> 2 * HZ
)
1138 pr_err("%s kthread starved for %ld jiffies!\n",
1139 rsp
->name
, j
- gpa
);
1143 * Dump stacks of all tasks running on stalled CPUs.
1145 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1148 unsigned long flags
;
1149 struct rcu_node
*rnp
;
1151 rcu_for_each_leaf_node(rsp
, rnp
) {
1152 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1153 if (rnp
->qsmask
!= 0) {
1154 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1155 if (rnp
->qsmask
& (1UL << cpu
))
1156 dump_cpu_task(rnp
->grplo
+ cpu
);
1158 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1162 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1166 unsigned long flags
;
1170 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1173 /* Only let one CPU complain about others per time interval. */
1175 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1176 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
1177 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1178 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1181 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
1182 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1185 * OK, time to rat on our buddy...
1186 * See Documentation/RCU/stallwarn.txt for info on how to debug
1187 * RCU CPU stall warnings.
1189 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1191 print_cpu_stall_info_begin();
1192 rcu_for_each_leaf_node(rsp
, rnp
) {
1193 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1194 ndetected
+= rcu_print_task_stall(rnp
);
1195 if (rnp
->qsmask
!= 0) {
1196 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1197 if (rnp
->qsmask
& (1UL << cpu
)) {
1198 print_cpu_stall_info(rsp
,
1203 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1206 print_cpu_stall_info_end();
1207 for_each_possible_cpu(cpu
)
1208 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1209 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1210 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1211 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1213 rcu_dump_cpu_stacks(rsp
);
1215 if (ACCESS_ONCE(rsp
->gpnum
) != gpnum
||
1216 ACCESS_ONCE(rsp
->completed
) == gpnum
) {
1217 pr_err("INFO: Stall ended before state dump start\n");
1220 gpa
= ACCESS_ONCE(rsp
->gp_activity
);
1221 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1222 rsp
->name
, j
- gpa
, j
, gpa
,
1223 jiffies_till_next_fqs
,
1224 rcu_get_root(rsp
)->qsmask
);
1225 /* In this case, the current CPU might be at fault. */
1226 sched_show_task(current
);
1230 /* Complain about tasks blocking the grace period. */
1231 rcu_print_detail_task_stall(rsp
);
1233 rcu_check_gp_kthread_starvation(rsp
);
1235 force_quiescent_state(rsp
); /* Kick them all. */
1238 static void print_cpu_stall(struct rcu_state
*rsp
)
1241 unsigned long flags
;
1242 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1246 * OK, time to rat on ourselves...
1247 * See Documentation/RCU/stallwarn.txt for info on how to debug
1248 * RCU CPU stall warnings.
1250 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1251 print_cpu_stall_info_begin();
1252 print_cpu_stall_info(rsp
, smp_processor_id());
1253 print_cpu_stall_info_end();
1254 for_each_possible_cpu(cpu
)
1255 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1256 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1257 jiffies
- rsp
->gp_start
,
1258 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1260 rcu_check_gp_kthread_starvation(rsp
);
1262 rcu_dump_cpu_stacks(rsp
);
1264 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1265 if (ULONG_CMP_GE(jiffies
, ACCESS_ONCE(rsp
->jiffies_stall
)))
1266 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+
1267 3 * rcu_jiffies_till_stall_check() + 3;
1268 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1271 * Attempt to revive the RCU machinery by forcing a context switch.
1273 * A context switch would normally allow the RCU state machine to make
1274 * progress and it could be we're stuck in kernel space without context
1275 * switches for an entirely unreasonable amount of time.
1277 resched_cpu(smp_processor_id());
1280 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1282 unsigned long completed
;
1283 unsigned long gpnum
;
1287 struct rcu_node
*rnp
;
1289 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1294 * Lots of memory barriers to reject false positives.
1296 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1297 * then rsp->gp_start, and finally rsp->completed. These values
1298 * are updated in the opposite order with memory barriers (or
1299 * equivalent) during grace-period initialization and cleanup.
1300 * Now, a false positive can occur if we get an new value of
1301 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1302 * the memory barriers, the only way that this can happen is if one
1303 * grace period ends and another starts between these two fetches.
1304 * Detect this by comparing rsp->completed with the previous fetch
1307 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1308 * and rsp->gp_start suffice to forestall false positives.
1310 gpnum
= ACCESS_ONCE(rsp
->gpnum
);
1311 smp_rmb(); /* Pick up ->gpnum first... */
1312 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
1313 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1314 gps
= ACCESS_ONCE(rsp
->gp_start
);
1315 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1316 completed
= ACCESS_ONCE(rsp
->completed
);
1317 if (ULONG_CMP_GE(completed
, gpnum
) ||
1318 ULONG_CMP_LT(j
, js
) ||
1319 ULONG_CMP_GE(gps
, js
))
1320 return; /* No stall or GP completed since entering function. */
1322 if (rcu_gp_in_progress(rsp
) &&
1323 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1325 /* We haven't checked in, so go dump stack. */
1326 print_cpu_stall(rsp
);
1328 } else if (rcu_gp_in_progress(rsp
) &&
1329 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1331 /* They had a few time units to dump stack, so complain. */
1332 print_other_cpu_stall(rsp
, gpnum
);
1337 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1339 * Set the stall-warning timeout way off into the future, thus preventing
1340 * any RCU CPU stall-warning messages from appearing in the current set of
1341 * RCU grace periods.
1343 * The caller must disable hard irqs.
1345 void rcu_cpu_stall_reset(void)
1347 struct rcu_state
*rsp
;
1349 for_each_rcu_flavor(rsp
)
1350 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ ULONG_MAX
/ 2;
1354 * Initialize the specified rcu_data structure's default callback list
1355 * to empty. The default callback list is the one that is not used by
1356 * no-callbacks CPUs.
1358 static void init_default_callback_list(struct rcu_data
*rdp
)
1362 rdp
->nxtlist
= NULL
;
1363 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1364 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1368 * Initialize the specified rcu_data structure's callback list to empty.
1370 static void init_callback_list(struct rcu_data
*rdp
)
1372 if (init_nocb_callback_list(rdp
))
1374 init_default_callback_list(rdp
);
1378 * Determine the value that ->completed will have at the end of the
1379 * next subsequent grace period. This is used to tag callbacks so that
1380 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1381 * been dyntick-idle for an extended period with callbacks under the
1382 * influence of RCU_FAST_NO_HZ.
1384 * The caller must hold rnp->lock with interrupts disabled.
1386 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1387 struct rcu_node
*rnp
)
1390 * If RCU is idle, we just wait for the next grace period.
1391 * But we can only be sure that RCU is idle if we are looking
1392 * at the root rcu_node structure -- otherwise, a new grace
1393 * period might have started, but just not yet gotten around
1394 * to initializing the current non-root rcu_node structure.
1396 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1397 return rnp
->completed
+ 1;
1400 * Otherwise, wait for a possible partial grace period and
1401 * then the subsequent full grace period.
1403 return rnp
->completed
+ 2;
1407 * Trace-event helper function for rcu_start_future_gp() and
1408 * rcu_nocb_wait_gp().
1410 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1411 unsigned long c
, const char *s
)
1413 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1414 rnp
->completed
, c
, rnp
->level
,
1415 rnp
->grplo
, rnp
->grphi
, s
);
1419 * Start some future grace period, as needed to handle newly arrived
1420 * callbacks. The required future grace periods are recorded in each
1421 * rcu_node structure's ->need_future_gp field. Returns true if there
1422 * is reason to awaken the grace-period kthread.
1424 * The caller must hold the specified rcu_node structure's ->lock.
1426 static bool __maybe_unused
1427 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1428 unsigned long *c_out
)
1433 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1436 * Pick up grace-period number for new callbacks. If this
1437 * grace period is already marked as needed, return to the caller.
1439 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1440 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1441 if (rnp
->need_future_gp
[c
& 0x1]) {
1442 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1447 * If either this rcu_node structure or the root rcu_node structure
1448 * believe that a grace period is in progress, then we must wait
1449 * for the one following, which is in "c". Because our request
1450 * will be noticed at the end of the current grace period, we don't
1451 * need to explicitly start one. We only do the lockless check
1452 * of rnp_root's fields if the current rcu_node structure thinks
1453 * there is no grace period in flight, and because we hold rnp->lock,
1454 * the only possible change is when rnp_root's two fields are
1455 * equal, in which case rnp_root->gpnum might be concurrently
1456 * incremented. But that is OK, as it will just result in our
1457 * doing some extra useless work.
1459 if (rnp
->gpnum
!= rnp
->completed
||
1460 ACCESS_ONCE(rnp_root
->gpnum
) != ACCESS_ONCE(rnp_root
->completed
)) {
1461 rnp
->need_future_gp
[c
& 0x1]++;
1462 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1467 * There might be no grace period in progress. If we don't already
1468 * hold it, acquire the root rcu_node structure's lock in order to
1469 * start one (if needed).
1471 if (rnp
!= rnp_root
) {
1472 raw_spin_lock(&rnp_root
->lock
);
1473 smp_mb__after_unlock_lock();
1477 * Get a new grace-period number. If there really is no grace
1478 * period in progress, it will be smaller than the one we obtained
1479 * earlier. Adjust callbacks as needed. Note that even no-CBs
1480 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1482 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1483 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1484 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1485 rdp
->nxtcompleted
[i
] = c
;
1488 * If the needed for the required grace period is already
1489 * recorded, trace and leave.
1491 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1492 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1496 /* Record the need for the future grace period. */
1497 rnp_root
->need_future_gp
[c
& 0x1]++;
1499 /* If a grace period is not already in progress, start one. */
1500 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1501 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1503 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1504 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1507 if (rnp
!= rnp_root
)
1508 raw_spin_unlock(&rnp_root
->lock
);
1516 * Clean up any old requests for the just-ended grace period. Also return
1517 * whether any additional grace periods have been requested. Also invoke
1518 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1519 * waiting for this grace period to complete.
1521 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1523 int c
= rnp
->completed
;
1525 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1527 rcu_nocb_gp_cleanup(rsp
, rnp
);
1528 rnp
->need_future_gp
[c
& 0x1] = 0;
1529 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1530 trace_rcu_future_gp(rnp
, rdp
, c
,
1531 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1536 * Awaken the grace-period kthread for the specified flavor of RCU.
1537 * Don't do a self-awaken, and don't bother awakening when there is
1538 * nothing for the grace-period kthread to do (as in several CPUs
1539 * raced to awaken, and we lost), and finally don't try to awaken
1540 * a kthread that has not yet been created.
1542 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1544 if (current
== rsp
->gp_kthread
||
1545 !ACCESS_ONCE(rsp
->gp_flags
) ||
1548 wake_up(&rsp
->gp_wq
);
1552 * If there is room, assign a ->completed number to any callbacks on
1553 * this CPU that have not already been assigned. Also accelerate any
1554 * callbacks that were previously assigned a ->completed number that has
1555 * since proven to be too conservative, which can happen if callbacks get
1556 * assigned a ->completed number while RCU is idle, but with reference to
1557 * a non-root rcu_node structure. This function is idempotent, so it does
1558 * not hurt to call it repeatedly. Returns an flag saying that we should
1559 * awaken the RCU grace-period kthread.
1561 * The caller must hold rnp->lock with interrupts disabled.
1563 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1564 struct rcu_data
*rdp
)
1570 /* If the CPU has no callbacks, nothing to do. */
1571 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1575 * Starting from the sublist containing the callbacks most
1576 * recently assigned a ->completed number and working down, find the
1577 * first sublist that is not assignable to an upcoming grace period.
1578 * Such a sublist has something in it (first two tests) and has
1579 * a ->completed number assigned that will complete sooner than
1580 * the ->completed number for newly arrived callbacks (last test).
1582 * The key point is that any later sublist can be assigned the
1583 * same ->completed number as the newly arrived callbacks, which
1584 * means that the callbacks in any of these later sublist can be
1585 * grouped into a single sublist, whether or not they have already
1586 * been assigned a ->completed number.
1588 c
= rcu_cbs_completed(rsp
, rnp
);
1589 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1590 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1591 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1595 * If there are no sublist for unassigned callbacks, leave.
1596 * At the same time, advance "i" one sublist, so that "i" will
1597 * index into the sublist where all the remaining callbacks should
1600 if (++i
>= RCU_NEXT_TAIL
)
1604 * Assign all subsequent callbacks' ->completed number to the next
1605 * full grace period and group them all in the sublist initially
1608 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1609 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1610 rdp
->nxtcompleted
[i
] = c
;
1612 /* Record any needed additional grace periods. */
1613 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1615 /* Trace depending on how much we were able to accelerate. */
1616 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1617 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1619 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1624 * Move any callbacks whose grace period has completed to the
1625 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1626 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1627 * sublist. This function is idempotent, so it does not hurt to
1628 * invoke it repeatedly. As long as it is not invoked -too- often...
1629 * Returns true if the RCU grace-period kthread needs to be awakened.
1631 * The caller must hold rnp->lock with interrupts disabled.
1633 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1634 struct rcu_data
*rdp
)
1638 /* If the CPU has no callbacks, nothing to do. */
1639 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1643 * Find all callbacks whose ->completed numbers indicate that they
1644 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1646 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1647 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1649 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1651 /* Clean up any sublist tail pointers that were misordered above. */
1652 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1653 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1655 /* Copy down callbacks to fill in empty sublists. */
1656 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1657 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1659 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1660 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1663 /* Classify any remaining callbacks. */
1664 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1668 * Update CPU-local rcu_data state to record the beginnings and ends of
1669 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1670 * structure corresponding to the current CPU, and must have irqs disabled.
1671 * Returns true if the grace-period kthread needs to be awakened.
1673 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1674 struct rcu_data
*rdp
)
1678 /* Handle the ends of any preceding grace periods first. */
1679 if (rdp
->completed
== rnp
->completed
&&
1680 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1682 /* No grace period end, so just accelerate recent callbacks. */
1683 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1687 /* Advance callbacks. */
1688 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1690 /* Remember that we saw this grace-period completion. */
1691 rdp
->completed
= rnp
->completed
;
1692 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1695 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(ACCESS_ONCE(rdp
->gpwrap
))) {
1697 * If the current grace period is waiting for this CPU,
1698 * set up to detect a quiescent state, otherwise don't
1699 * go looking for one.
1701 rdp
->gpnum
= rnp
->gpnum
;
1702 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1703 rdp
->passed_quiesce
= 0;
1704 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1705 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1706 zero_cpu_stall_ticks(rdp
);
1707 ACCESS_ONCE(rdp
->gpwrap
) = false;
1712 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1714 unsigned long flags
;
1716 struct rcu_node
*rnp
;
1718 local_irq_save(flags
);
1720 if ((rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) &&
1721 rdp
->completed
== ACCESS_ONCE(rnp
->completed
) &&
1722 !unlikely(ACCESS_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1723 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1724 local_irq_restore(flags
);
1727 smp_mb__after_unlock_lock();
1728 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1729 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1731 rcu_gp_kthread_wake(rsp
);
1735 * Initialize a new grace period. Return 0 if no grace period required.
1737 static int rcu_gp_init(struct rcu_state
*rsp
)
1739 unsigned long oldmask
;
1740 struct rcu_data
*rdp
;
1741 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1743 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1744 raw_spin_lock_irq(&rnp
->lock
);
1745 smp_mb__after_unlock_lock();
1746 if (!ACCESS_ONCE(rsp
->gp_flags
)) {
1747 /* Spurious wakeup, tell caller to go back to sleep. */
1748 raw_spin_unlock_irq(&rnp
->lock
);
1751 ACCESS_ONCE(rsp
->gp_flags
) = 0; /* Clear all flags: New grace period. */
1753 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1755 * Grace period already in progress, don't start another.
1756 * Not supposed to be able to happen.
1758 raw_spin_unlock_irq(&rnp
->lock
);
1762 /* Advance to a new grace period and initialize state. */
1763 record_gp_stall_check_time(rsp
);
1764 /* Record GP times before starting GP, hence smp_store_release(). */
1765 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1766 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1767 raw_spin_unlock_irq(&rnp
->lock
);
1770 * Apply per-leaf buffered online and offline operations to the
1771 * rcu_node tree. Note that this new grace period need not wait
1772 * for subsequent online CPUs, and that quiescent-state forcing
1773 * will handle subsequent offline CPUs.
1775 rcu_for_each_leaf_node(rsp
, rnp
) {
1776 raw_spin_lock_irq(&rnp
->lock
);
1777 smp_mb__after_unlock_lock();
1778 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1779 !rnp
->wait_blkd_tasks
) {
1780 /* Nothing to do on this leaf rcu_node structure. */
1781 raw_spin_unlock_irq(&rnp
->lock
);
1785 /* Record old state, apply changes to ->qsmaskinit field. */
1786 oldmask
= rnp
->qsmaskinit
;
1787 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1789 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1790 if (!oldmask
!= !rnp
->qsmaskinit
) {
1791 if (!oldmask
) /* First online CPU for this rcu_node. */
1792 rcu_init_new_rnp(rnp
);
1793 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1794 rnp
->wait_blkd_tasks
= true;
1795 else /* Last offline CPU and can propagate. */
1796 rcu_cleanup_dead_rnp(rnp
);
1800 * If all waited-on tasks from prior grace period are
1801 * done, and if all this rcu_node structure's CPUs are
1802 * still offline, propagate up the rcu_node tree and
1803 * clear ->wait_blkd_tasks. Otherwise, if one of this
1804 * rcu_node structure's CPUs has since come back online,
1805 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1806 * checks for this, so just call it unconditionally).
1808 if (rnp
->wait_blkd_tasks
&&
1809 (!rcu_preempt_has_tasks(rnp
) ||
1811 rnp
->wait_blkd_tasks
= false;
1812 rcu_cleanup_dead_rnp(rnp
);
1815 raw_spin_unlock_irq(&rnp
->lock
);
1819 * Set the quiescent-state-needed bits in all the rcu_node
1820 * structures for all currently online CPUs in breadth-first order,
1821 * starting from the root rcu_node structure, relying on the layout
1822 * of the tree within the rsp->node[] array. Note that other CPUs
1823 * will access only the leaves of the hierarchy, thus seeing that no
1824 * grace period is in progress, at least until the corresponding
1825 * leaf node has been initialized. In addition, we have excluded
1826 * CPU-hotplug operations.
1828 * The grace period cannot complete until the initialization
1829 * process finishes, because this kthread handles both.
1831 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1832 raw_spin_lock_irq(&rnp
->lock
);
1833 smp_mb__after_unlock_lock();
1834 rdp
= this_cpu_ptr(rsp
->rda
);
1835 rcu_preempt_check_blocked_tasks(rnp
);
1836 rnp
->qsmask
= rnp
->qsmaskinit
;
1837 ACCESS_ONCE(rnp
->gpnum
) = rsp
->gpnum
;
1838 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1839 ACCESS_ONCE(rnp
->completed
) = rsp
->completed
;
1840 if (rnp
== rdp
->mynode
)
1841 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1842 rcu_preempt_boost_start_gp(rnp
);
1843 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1844 rnp
->level
, rnp
->grplo
,
1845 rnp
->grphi
, rnp
->qsmask
);
1846 raw_spin_unlock_irq(&rnp
->lock
);
1847 cond_resched_rcu_qs();
1848 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1849 if (gp_init_delay
> 0 &&
1850 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
)))
1851 schedule_timeout_uninterruptible(gp_init_delay
);
1858 * Do one round of quiescent-state forcing.
1860 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1862 int fqs_state
= fqs_state_in
;
1863 bool isidle
= false;
1865 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1867 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1869 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1870 /* Collect dyntick-idle snapshots. */
1871 if (is_sysidle_rcu_state(rsp
)) {
1873 maxj
= jiffies
- ULONG_MAX
/ 4;
1875 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1877 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1878 fqs_state
= RCU_FORCE_QS
;
1880 /* Handle dyntick-idle and offline CPUs. */
1882 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1884 /* Clear flag to prevent immediate re-entry. */
1885 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1886 raw_spin_lock_irq(&rnp
->lock
);
1887 smp_mb__after_unlock_lock();
1888 ACCESS_ONCE(rsp
->gp_flags
) =
1889 ACCESS_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
;
1890 raw_spin_unlock_irq(&rnp
->lock
);
1896 * Clean up after the old grace period.
1898 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1900 unsigned long gp_duration
;
1901 bool needgp
= false;
1903 struct rcu_data
*rdp
;
1904 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1906 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1907 raw_spin_lock_irq(&rnp
->lock
);
1908 smp_mb__after_unlock_lock();
1909 gp_duration
= jiffies
- rsp
->gp_start
;
1910 if (gp_duration
> rsp
->gp_max
)
1911 rsp
->gp_max
= gp_duration
;
1914 * We know the grace period is complete, but to everyone else
1915 * it appears to still be ongoing. But it is also the case
1916 * that to everyone else it looks like there is nothing that
1917 * they can do to advance the grace period. It is therefore
1918 * safe for us to drop the lock in order to mark the grace
1919 * period as completed in all of the rcu_node structures.
1921 raw_spin_unlock_irq(&rnp
->lock
);
1924 * Propagate new ->completed value to rcu_node structures so
1925 * that other CPUs don't have to wait until the start of the next
1926 * grace period to process their callbacks. This also avoids
1927 * some nasty RCU grace-period initialization races by forcing
1928 * the end of the current grace period to be completely recorded in
1929 * all of the rcu_node structures before the beginning of the next
1930 * grace period is recorded in any of the rcu_node structures.
1932 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1933 raw_spin_lock_irq(&rnp
->lock
);
1934 smp_mb__after_unlock_lock();
1935 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
1936 WARN_ON_ONCE(rnp
->qsmask
);
1937 ACCESS_ONCE(rnp
->completed
) = rsp
->gpnum
;
1938 rdp
= this_cpu_ptr(rsp
->rda
);
1939 if (rnp
== rdp
->mynode
)
1940 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1941 /* smp_mb() provided by prior unlock-lock pair. */
1942 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1943 raw_spin_unlock_irq(&rnp
->lock
);
1944 cond_resched_rcu_qs();
1945 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1947 rnp
= rcu_get_root(rsp
);
1948 raw_spin_lock_irq(&rnp
->lock
);
1949 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1950 rcu_nocb_gp_set(rnp
, nocb
);
1952 /* Declare grace period done. */
1953 ACCESS_ONCE(rsp
->completed
) = rsp
->gpnum
;
1954 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1955 rsp
->fqs_state
= RCU_GP_IDLE
;
1956 rdp
= this_cpu_ptr(rsp
->rda
);
1957 /* Advance CBs to reduce false positives below. */
1958 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1959 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1960 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
1961 trace_rcu_grace_period(rsp
->name
,
1962 ACCESS_ONCE(rsp
->gpnum
),
1965 raw_spin_unlock_irq(&rnp
->lock
);
1969 * Body of kthread that handles grace periods.
1971 static int __noreturn
rcu_gp_kthread(void *arg
)
1977 struct rcu_state
*rsp
= arg
;
1978 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1980 rcu_bind_gp_kthread();
1983 /* Handle grace-period start. */
1985 trace_rcu_grace_period(rsp
->name
,
1986 ACCESS_ONCE(rsp
->gpnum
),
1988 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
1989 wait_event_interruptible(rsp
->gp_wq
,
1990 ACCESS_ONCE(rsp
->gp_flags
) &
1992 /* Locking provides needed memory barrier. */
1993 if (rcu_gp_init(rsp
))
1995 cond_resched_rcu_qs();
1996 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
1997 WARN_ON(signal_pending(current
));
1998 trace_rcu_grace_period(rsp
->name
,
1999 ACCESS_ONCE(rsp
->gpnum
),
2003 /* Handle quiescent-state forcing. */
2004 fqs_state
= RCU_SAVE_DYNTICK
;
2005 j
= jiffies_till_first_fqs
;
2008 jiffies_till_first_fqs
= HZ
;
2013 rsp
->jiffies_force_qs
= jiffies
+ j
;
2014 trace_rcu_grace_period(rsp
->name
,
2015 ACCESS_ONCE(rsp
->gpnum
),
2017 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2018 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
2019 ((gf
= ACCESS_ONCE(rsp
->gp_flags
)) &
2021 (!ACCESS_ONCE(rnp
->qsmask
) &&
2022 !rcu_preempt_blocked_readers_cgp(rnp
)),
2024 /* Locking provides needed memory barriers. */
2025 /* If grace period done, leave loop. */
2026 if (!ACCESS_ONCE(rnp
->qsmask
) &&
2027 !rcu_preempt_blocked_readers_cgp(rnp
))
2029 /* If time for quiescent-state forcing, do it. */
2030 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2031 (gf
& RCU_GP_FLAG_FQS
)) {
2032 trace_rcu_grace_period(rsp
->name
,
2033 ACCESS_ONCE(rsp
->gpnum
),
2035 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
2036 trace_rcu_grace_period(rsp
->name
,
2037 ACCESS_ONCE(rsp
->gpnum
),
2039 cond_resched_rcu_qs();
2040 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
2042 /* Deal with stray signal. */
2043 cond_resched_rcu_qs();
2044 ACCESS_ONCE(rsp
->gp_activity
) = jiffies
;
2045 WARN_ON(signal_pending(current
));
2046 trace_rcu_grace_period(rsp
->name
,
2047 ACCESS_ONCE(rsp
->gpnum
),
2050 j
= jiffies_till_next_fqs
;
2053 jiffies_till_next_fqs
= HZ
;
2056 jiffies_till_next_fqs
= 1;
2060 /* Handle grace-period end. */
2061 rcu_gp_cleanup(rsp
);
2066 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2067 * in preparation for detecting the next grace period. The caller must hold
2068 * the root node's ->lock and hard irqs must be disabled.
2070 * Note that it is legal for a dying CPU (which is marked as offline) to
2071 * invoke this function. This can happen when the dying CPU reports its
2074 * Returns true if the grace-period kthread must be awakened.
2077 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2078 struct rcu_data
*rdp
)
2080 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2082 * Either we have not yet spawned the grace-period
2083 * task, this CPU does not need another grace period,
2084 * or a grace period is already in progress.
2085 * Either way, don't start a new grace period.
2089 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
2090 trace_rcu_grace_period(rsp
->name
, ACCESS_ONCE(rsp
->gpnum
),
2094 * We can't do wakeups while holding the rnp->lock, as that
2095 * could cause possible deadlocks with the rq->lock. Defer
2096 * the wakeup to our caller.
2102 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2103 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2104 * is invoked indirectly from rcu_advance_cbs(), which would result in
2105 * endless recursion -- or would do so if it wasn't for the self-deadlock
2106 * that is encountered beforehand.
2108 * Returns true if the grace-period kthread needs to be awakened.
2110 static bool rcu_start_gp(struct rcu_state
*rsp
)
2112 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2113 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2117 * If there is no grace period in progress right now, any
2118 * callbacks we have up to this point will be satisfied by the
2119 * next grace period. Also, advancing the callbacks reduces the
2120 * probability of false positives from cpu_needs_another_gp()
2121 * resulting in pointless grace periods. So, advance callbacks
2122 * then start the grace period!
2124 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2125 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2130 * Report a full set of quiescent states to the specified rcu_state
2131 * data structure. This involves cleaning up after the prior grace
2132 * period and letting rcu_start_gp() start up the next grace period
2133 * if one is needed. Note that the caller must hold rnp->lock, which
2134 * is released before return.
2136 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2137 __releases(rcu_get_root(rsp
)->lock
)
2139 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2140 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2141 rcu_gp_kthread_wake(rsp
);
2145 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2146 * Allows quiescent states for a group of CPUs to be reported at one go
2147 * to the specified rcu_node structure, though all the CPUs in the group
2148 * must be represented by the same rcu_node structure (which need not be a
2149 * leaf rcu_node structure, though it often will be). The gps parameter
2150 * is the grace-period snapshot, which means that the quiescent states
2151 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2152 * must be held upon entry, and it is released before return.
2155 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2156 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2157 __releases(rnp
->lock
)
2159 unsigned long oldmask
= 0;
2160 struct rcu_node
*rnp_c
;
2162 /* Walk up the rcu_node hierarchy. */
2164 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2167 * Our bit has already been cleared, or the
2168 * relevant grace period is already over, so done.
2170 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2173 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2174 rnp
->qsmask
&= ~mask
;
2175 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2176 mask
, rnp
->qsmask
, rnp
->level
,
2177 rnp
->grplo
, rnp
->grphi
,
2179 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2181 /* Other bits still set at this level, so done. */
2182 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2185 mask
= rnp
->grpmask
;
2186 if (rnp
->parent
== NULL
) {
2188 /* No more levels. Exit loop holding root lock. */
2192 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2195 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2196 smp_mb__after_unlock_lock();
2197 oldmask
= rnp_c
->qsmask
;
2201 * Get here if we are the last CPU to pass through a quiescent
2202 * state for this grace period. Invoke rcu_report_qs_rsp()
2203 * to clean up and start the next grace period if one is needed.
2205 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2209 * Record a quiescent state for all tasks that were previously queued
2210 * on the specified rcu_node structure and that were blocking the current
2211 * RCU grace period. The caller must hold the specified rnp->lock with
2212 * irqs disabled, and this lock is released upon return, but irqs remain
2215 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2216 struct rcu_node
*rnp
, unsigned long flags
)
2217 __releases(rnp
->lock
)
2221 struct rcu_node
*rnp_p
;
2223 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2224 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2225 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2226 return; /* Still need more quiescent states! */
2229 rnp_p
= rnp
->parent
;
2230 if (rnp_p
== NULL
) {
2232 * Only one rcu_node structure in the tree, so don't
2233 * try to report up to its nonexistent parent!
2235 rcu_report_qs_rsp(rsp
, flags
);
2239 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2241 mask
= rnp
->grpmask
;
2242 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2243 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
2244 smp_mb__after_unlock_lock();
2245 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2249 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2250 * structure. This must be either called from the specified CPU, or
2251 * called when the specified CPU is known to be offline (and when it is
2252 * also known that no other CPU is concurrently trying to help the offline
2253 * CPU). The lastcomp argument is used to make sure we are still in the
2254 * grace period of interest. We don't want to end the current grace period
2255 * based on quiescent states detected in an earlier grace period!
2258 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2260 unsigned long flags
;
2263 struct rcu_node
*rnp
;
2266 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2267 smp_mb__after_unlock_lock();
2268 if ((rdp
->passed_quiesce
== 0 &&
2269 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2270 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2274 * The grace period in which this quiescent state was
2275 * recorded has ended, so don't report it upwards.
2276 * We will instead need a new quiescent state that lies
2277 * within the current grace period.
2279 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2280 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2281 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2284 mask
= rdp
->grpmask
;
2285 if ((rnp
->qsmask
& mask
) == 0) {
2286 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2288 rdp
->qs_pending
= 0;
2291 * This GP can't end until cpu checks in, so all of our
2292 * callbacks can be processed during the next GP.
2294 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2296 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2297 /* ^^^ Released rnp->lock */
2299 rcu_gp_kthread_wake(rsp
);
2304 * Check to see if there is a new grace period of which this CPU
2305 * is not yet aware, and if so, set up local rcu_data state for it.
2306 * Otherwise, see if this CPU has just passed through its first
2307 * quiescent state for this grace period, and record that fact if so.
2310 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2312 /* Check for grace-period ends and beginnings. */
2313 note_gp_changes(rsp
, rdp
);
2316 * Does this CPU still need to do its part for current grace period?
2317 * If no, return and let the other CPUs do their part as well.
2319 if (!rdp
->qs_pending
)
2323 * Was there a quiescent state since the beginning of the grace
2324 * period? If no, then exit and wait for the next call.
2326 if (!rdp
->passed_quiesce
&&
2327 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2331 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2334 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2337 #ifdef CONFIG_HOTPLUG_CPU
2340 * Send the specified CPU's RCU callbacks to the orphanage. The
2341 * specified CPU must be offline, and the caller must hold the
2345 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2346 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2348 /* No-CBs CPUs do not have orphanable callbacks. */
2349 if (rcu_is_nocb_cpu(rdp
->cpu
))
2353 * Orphan the callbacks. First adjust the counts. This is safe
2354 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2355 * cannot be running now. Thus no memory barrier is required.
2357 if (rdp
->nxtlist
!= NULL
) {
2358 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2359 rsp
->qlen
+= rdp
->qlen
;
2360 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2362 ACCESS_ONCE(rdp
->qlen
) = 0;
2366 * Next, move those callbacks still needing a grace period to
2367 * the orphanage, where some other CPU will pick them up.
2368 * Some of the callbacks might have gone partway through a grace
2369 * period, but that is too bad. They get to start over because we
2370 * cannot assume that grace periods are synchronized across CPUs.
2371 * We don't bother updating the ->nxttail[] array yet, instead
2372 * we just reset the whole thing later on.
2374 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2375 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2376 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2377 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2381 * Then move the ready-to-invoke callbacks to the orphanage,
2382 * where some other CPU will pick them up. These will not be
2383 * required to pass though another grace period: They are done.
2385 if (rdp
->nxtlist
!= NULL
) {
2386 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2387 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2391 * Finally, initialize the rcu_data structure's list to empty and
2392 * disallow further callbacks on this CPU.
2394 init_callback_list(rdp
);
2395 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2399 * Adopt the RCU callbacks from the specified rcu_state structure's
2400 * orphanage. The caller must hold the ->orphan_lock.
2402 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2405 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2407 /* No-CBs CPUs are handled specially. */
2408 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2411 /* Do the accounting first. */
2412 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2413 rdp
->qlen
+= rsp
->qlen
;
2414 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2415 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2416 rcu_idle_count_callbacks_posted();
2421 * We do not need a memory barrier here because the only way we
2422 * can get here if there is an rcu_barrier() in flight is if
2423 * we are the task doing the rcu_barrier().
2426 /* First adopt the ready-to-invoke callbacks. */
2427 if (rsp
->orphan_donelist
!= NULL
) {
2428 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2429 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2430 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2431 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2432 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2433 rsp
->orphan_donelist
= NULL
;
2434 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2437 /* And then adopt the callbacks that still need a grace period. */
2438 if (rsp
->orphan_nxtlist
!= NULL
) {
2439 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2440 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2441 rsp
->orphan_nxtlist
= NULL
;
2442 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2447 * Trace the fact that this CPU is going offline.
2449 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2451 RCU_TRACE(unsigned long mask
);
2452 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2453 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2455 RCU_TRACE(mask
= rdp
->grpmask
);
2456 trace_rcu_grace_period(rsp
->name
,
2457 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2462 * All CPUs for the specified rcu_node structure have gone offline,
2463 * and all tasks that were preempted within an RCU read-side critical
2464 * section while running on one of those CPUs have since exited their RCU
2465 * read-side critical section. Some other CPU is reporting this fact with
2466 * the specified rcu_node structure's ->lock held and interrupts disabled.
2467 * This function therefore goes up the tree of rcu_node structures,
2468 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2469 * the leaf rcu_node structure's ->qsmaskinit field has already been
2472 * This function does check that the specified rcu_node structure has
2473 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2474 * prematurely. That said, invoking it after the fact will cost you
2475 * a needless lock acquisition. So once it has done its work, don't
2478 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2481 struct rcu_node
*rnp
= rnp_leaf
;
2483 if (rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2486 mask
= rnp
->grpmask
;
2490 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2491 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2492 rnp
->qsmaskinit
&= ~mask
;
2493 rnp
->qsmask
&= ~mask
;
2494 if (rnp
->qsmaskinit
) {
2495 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2498 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2503 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2504 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2507 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2509 unsigned long flags
;
2511 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2512 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2514 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2515 mask
= rdp
->grpmask
;
2516 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2517 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2518 rnp
->qsmaskinitnext
&= ~mask
;
2519 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2523 * The CPU has been completely removed, and some other CPU is reporting
2524 * this fact from process context. Do the remainder of the cleanup,
2525 * including orphaning the outgoing CPU's RCU callbacks, and also
2526 * adopting them. There can only be one CPU hotplug operation at a time,
2527 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2529 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2531 unsigned long flags
;
2532 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2533 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2535 /* Adjust any no-longer-needed kthreads. */
2536 rcu_boost_kthread_setaffinity(rnp
, -1);
2538 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2539 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2540 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2541 rcu_adopt_orphan_cbs(rsp
, flags
);
2542 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2544 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2545 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2546 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2549 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2551 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2555 static void __maybe_unused
rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2559 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2563 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2567 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2570 * Invoke any RCU callbacks that have made it to the end of their grace
2571 * period. Thottle as specified by rdp->blimit.
2573 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2575 unsigned long flags
;
2576 struct rcu_head
*next
, *list
, **tail
;
2577 long bl
, count
, count_lazy
;
2580 /* If no callbacks are ready, just return. */
2581 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2582 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2583 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
2584 need_resched(), is_idle_task(current
),
2585 rcu_is_callbacks_kthread());
2590 * Extract the list of ready callbacks, disabling to prevent
2591 * races with call_rcu() from interrupt handlers.
2593 local_irq_save(flags
);
2594 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2596 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2597 list
= rdp
->nxtlist
;
2598 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2599 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2600 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2601 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2602 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2603 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2604 local_irq_restore(flags
);
2606 /* Invoke callbacks. */
2607 count
= count_lazy
= 0;
2611 debug_rcu_head_unqueue(list
);
2612 if (__rcu_reclaim(rsp
->name
, list
))
2615 /* Stop only if limit reached and CPU has something to do. */
2616 if (++count
>= bl
&&
2618 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2622 local_irq_save(flags
);
2623 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2624 is_idle_task(current
),
2625 rcu_is_callbacks_kthread());
2627 /* Update count, and requeue any remaining callbacks. */
2629 *tail
= rdp
->nxtlist
;
2630 rdp
->nxtlist
= list
;
2631 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2632 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2633 rdp
->nxttail
[i
] = tail
;
2637 smp_mb(); /* List handling before counting for rcu_barrier(). */
2638 rdp
->qlen_lazy
-= count_lazy
;
2639 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
- count
;
2640 rdp
->n_cbs_invoked
+= count
;
2642 /* Reinstate batch limit if we have worked down the excess. */
2643 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2644 rdp
->blimit
= blimit
;
2646 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2647 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2648 rdp
->qlen_last_fqs_check
= 0;
2649 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2650 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2651 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2652 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2654 local_irq_restore(flags
);
2656 /* Re-invoke RCU core processing if there are callbacks remaining. */
2657 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2662 * Check to see if this CPU is in a non-context-switch quiescent state
2663 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2664 * Also schedule RCU core processing.
2666 * This function must be called from hardirq context. It is normally
2667 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2668 * false, there is no point in invoking rcu_check_callbacks().
2670 void rcu_check_callbacks(int user
)
2672 trace_rcu_utilization(TPS("Start scheduler-tick"));
2673 increment_cpu_stall_ticks();
2674 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2677 * Get here if this CPU took its interrupt from user
2678 * mode or from the idle loop, and if this is not a
2679 * nested interrupt. In this case, the CPU is in
2680 * a quiescent state, so note it.
2682 * No memory barrier is required here because both
2683 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2684 * variables that other CPUs neither access nor modify,
2685 * at least not while the corresponding CPU is online.
2691 } else if (!in_softirq()) {
2694 * Get here if this CPU did not take its interrupt from
2695 * softirq, in other words, if it is not interrupting
2696 * a rcu_bh read-side critical section. This is an _bh
2697 * critical section, so note it.
2702 rcu_preempt_check_callbacks();
2706 rcu_note_voluntary_context_switch(current
);
2707 trace_rcu_utilization(TPS("End scheduler-tick"));
2711 * Scan the leaf rcu_node structures, processing dyntick state for any that
2712 * have not yet encountered a quiescent state, using the function specified.
2713 * Also initiate boosting for any threads blocked on the root rcu_node.
2715 * The caller must have suppressed start of new grace periods.
2717 static void force_qs_rnp(struct rcu_state
*rsp
,
2718 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2719 unsigned long *maxj
),
2720 bool *isidle
, unsigned long *maxj
)
2724 unsigned long flags
;
2726 struct rcu_node
*rnp
;
2728 rcu_for_each_leaf_node(rsp
, rnp
) {
2729 cond_resched_rcu_qs();
2731 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2732 smp_mb__after_unlock_lock();
2733 if (!rcu_gp_in_progress(rsp
)) {
2734 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2737 if (rnp
->qsmask
== 0) {
2738 if (rcu_state_p
== &rcu_sched_state
||
2739 rsp
!= rcu_state_p
||
2740 rcu_preempt_blocked_readers_cgp(rnp
)) {
2742 * No point in scanning bits because they
2743 * are all zero. But we might need to
2744 * priority-boost blocked readers.
2746 rcu_initiate_boost(rnp
, flags
);
2747 /* rcu_initiate_boost() releases rnp->lock */
2751 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2753 * Race between grace-period
2754 * initialization and task exiting RCU
2755 * read-side critical section: Report.
2757 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2758 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2764 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2765 if ((rnp
->qsmask
& bit
) != 0) {
2766 if ((rnp
->qsmaskinit
& bit
) == 0)
2767 *isidle
= false; /* Pending hotplug. */
2768 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2773 /* Idle/offline CPUs, report (releases rnp->lock. */
2774 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2776 /* Nothing to do here, so just drop the lock. */
2777 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2783 * Force quiescent states on reluctant CPUs, and also detect which
2784 * CPUs are in dyntick-idle mode.
2786 static void force_quiescent_state(struct rcu_state
*rsp
)
2788 unsigned long flags
;
2790 struct rcu_node
*rnp
;
2791 struct rcu_node
*rnp_old
= NULL
;
2793 /* Funnel through hierarchy to reduce memory contention. */
2794 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2795 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2796 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2797 !raw_spin_trylock(&rnp
->fqslock
);
2798 if (rnp_old
!= NULL
)
2799 raw_spin_unlock(&rnp_old
->fqslock
);
2801 rsp
->n_force_qs_lh
++;
2806 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2808 /* Reached the root of the rcu_node tree, acquire lock. */
2809 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2810 smp_mb__after_unlock_lock();
2811 raw_spin_unlock(&rnp_old
->fqslock
);
2812 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2813 rsp
->n_force_qs_lh
++;
2814 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2815 return; /* Someone beat us to it. */
2817 ACCESS_ONCE(rsp
->gp_flags
) =
2818 ACCESS_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
;
2819 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2820 rcu_gp_kthread_wake(rsp
);
2824 * This does the RCU core processing work for the specified rcu_state
2825 * and rcu_data structures. This may be called only from the CPU to
2826 * whom the rdp belongs.
2829 __rcu_process_callbacks(struct rcu_state
*rsp
)
2831 unsigned long flags
;
2833 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2835 WARN_ON_ONCE(rdp
->beenonline
== 0);
2837 /* Update RCU state based on any recent quiescent states. */
2838 rcu_check_quiescent_state(rsp
, rdp
);
2840 /* Does this CPU require a not-yet-started grace period? */
2841 local_irq_save(flags
);
2842 if (cpu_needs_another_gp(rsp
, rdp
)) {
2843 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2844 needwake
= rcu_start_gp(rsp
);
2845 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2847 rcu_gp_kthread_wake(rsp
);
2849 local_irq_restore(flags
);
2852 /* If there are callbacks ready, invoke them. */
2853 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2854 invoke_rcu_callbacks(rsp
, rdp
);
2856 /* Do any needed deferred wakeups of rcuo kthreads. */
2857 do_nocb_deferred_wakeup(rdp
);
2861 * Do RCU core processing for the current CPU.
2863 static void rcu_process_callbacks(struct softirq_action
*unused
)
2865 struct rcu_state
*rsp
;
2867 if (cpu_is_offline(smp_processor_id()))
2869 trace_rcu_utilization(TPS("Start RCU core"));
2870 for_each_rcu_flavor(rsp
)
2871 __rcu_process_callbacks(rsp
);
2872 trace_rcu_utilization(TPS("End RCU core"));
2876 * Schedule RCU callback invocation. If the specified type of RCU
2877 * does not support RCU priority boosting, just do a direct call,
2878 * otherwise wake up the per-CPU kernel kthread. Note that because we
2879 * are running on the current CPU with softirqs disabled, the
2880 * rcu_cpu_kthread_task cannot disappear out from under us.
2882 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2884 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2886 if (likely(!rsp
->boost
)) {
2887 rcu_do_batch(rsp
, rdp
);
2890 invoke_rcu_callbacks_kthread();
2893 static void invoke_rcu_core(void)
2895 if (cpu_online(smp_processor_id()))
2896 raise_softirq(RCU_SOFTIRQ
);
2900 * Handle any core-RCU processing required by a call_rcu() invocation.
2902 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2903 struct rcu_head
*head
, unsigned long flags
)
2908 * If called from an extended quiescent state, invoke the RCU
2909 * core in order to force a re-evaluation of RCU's idleness.
2911 if (!rcu_is_watching())
2914 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2915 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2919 * Force the grace period if too many callbacks or too long waiting.
2920 * Enforce hysteresis, and don't invoke force_quiescent_state()
2921 * if some other CPU has recently done so. Also, don't bother
2922 * invoking force_quiescent_state() if the newly enqueued callback
2923 * is the only one waiting for a grace period to complete.
2925 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2927 /* Are we ignoring a completed grace period? */
2928 note_gp_changes(rsp
, rdp
);
2930 /* Start a new grace period if one not already started. */
2931 if (!rcu_gp_in_progress(rsp
)) {
2932 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2934 raw_spin_lock(&rnp_root
->lock
);
2935 smp_mb__after_unlock_lock();
2936 needwake
= rcu_start_gp(rsp
);
2937 raw_spin_unlock(&rnp_root
->lock
);
2939 rcu_gp_kthread_wake(rsp
);
2941 /* Give the grace period a kick. */
2942 rdp
->blimit
= LONG_MAX
;
2943 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2944 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2945 force_quiescent_state(rsp
);
2946 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2947 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2953 * RCU callback function to leak a callback.
2955 static void rcu_leak_callback(struct rcu_head
*rhp
)
2960 * Helper function for call_rcu() and friends. The cpu argument will
2961 * normally be -1, indicating "currently running CPU". It may specify
2962 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2963 * is expected to specify a CPU.
2966 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2967 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2969 unsigned long flags
;
2970 struct rcu_data
*rdp
;
2972 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
2973 if (debug_rcu_head_queue(head
)) {
2974 /* Probable double call_rcu(), so leak the callback. */
2975 ACCESS_ONCE(head
->func
) = rcu_leak_callback
;
2976 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2983 * Opportunistically note grace-period endings and beginnings.
2984 * Note that we might see a beginning right after we see an
2985 * end, but never vice versa, since this CPU has to pass through
2986 * a quiescent state betweentimes.
2988 local_irq_save(flags
);
2989 rdp
= this_cpu_ptr(rsp
->rda
);
2991 /* Add the callback to our list. */
2992 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2996 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2997 if (likely(rdp
->mynode
)) {
2998 /* Post-boot, so this should be for a no-CBs CPU. */
2999 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3000 WARN_ON_ONCE(offline
);
3001 /* Offline CPU, _call_rcu() illegal, leak callback. */
3002 local_irq_restore(flags
);
3006 * Very early boot, before rcu_init(). Initialize if needed
3007 * and then drop through to queue the callback.
3010 WARN_ON_ONCE(!rcu_is_watching());
3011 if (!likely(rdp
->nxtlist
))
3012 init_default_callback_list(rdp
);
3014 ACCESS_ONCE(rdp
->qlen
) = rdp
->qlen
+ 1;
3018 rcu_idle_count_callbacks_posted();
3019 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3020 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3021 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3023 if (__is_kfree_rcu_offset((unsigned long)func
))
3024 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3025 rdp
->qlen_lazy
, rdp
->qlen
);
3027 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3029 /* Go handle any RCU core processing required. */
3030 __call_rcu_core(rsp
, rdp
, head
, flags
);
3031 local_irq_restore(flags
);
3035 * Queue an RCU-sched callback for invocation after a grace period.
3037 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3039 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3041 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3044 * Queue an RCU callback for invocation after a quicker grace period.
3046 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3048 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3050 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3053 * Queue an RCU callback for lazy invocation after a grace period.
3054 * This will likely be later named something like "call_rcu_lazy()",
3055 * but this change will require some way of tagging the lazy RCU
3056 * callbacks in the list of pending callbacks. Until then, this
3057 * function may only be called from __kfree_rcu().
3059 void kfree_call_rcu(struct rcu_head
*head
,
3060 void (*func
)(struct rcu_head
*rcu
))
3062 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3064 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3067 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3068 * any blocking grace-period wait automatically implies a grace period
3069 * if there is only one CPU online at any point time during execution
3070 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3071 * occasionally incorrectly indicate that there are multiple CPUs online
3072 * when there was in fact only one the whole time, as this just adds
3073 * some overhead: RCU still operates correctly.
3075 static inline int rcu_blocking_is_gp(void)
3079 might_sleep(); /* Check for RCU read-side critical section. */
3081 ret
= num_online_cpus() <= 1;
3087 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3089 * Control will return to the caller some time after a full rcu-sched
3090 * grace period has elapsed, in other words after all currently executing
3091 * rcu-sched read-side critical sections have completed. These read-side
3092 * critical sections are delimited by rcu_read_lock_sched() and
3093 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3094 * local_irq_disable(), and so on may be used in place of
3095 * rcu_read_lock_sched().
3097 * This means that all preempt_disable code sequences, including NMI and
3098 * non-threaded hardware-interrupt handlers, in progress on entry will
3099 * have completed before this primitive returns. However, this does not
3100 * guarantee that softirq handlers will have completed, since in some
3101 * kernels, these handlers can run in process context, and can block.
3103 * Note that this guarantee implies further memory-ordering guarantees.
3104 * On systems with more than one CPU, when synchronize_sched() returns,
3105 * each CPU is guaranteed to have executed a full memory barrier since the
3106 * end of its last RCU-sched read-side critical section whose beginning
3107 * preceded the call to synchronize_sched(). In addition, each CPU having
3108 * an RCU read-side critical section that extends beyond the return from
3109 * synchronize_sched() is guaranteed to have executed a full memory barrier
3110 * after the beginning of synchronize_sched() and before the beginning of
3111 * that RCU read-side critical section. Note that these guarantees include
3112 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3113 * that are executing in the kernel.
3115 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3116 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3117 * to have executed a full memory barrier during the execution of
3118 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3119 * again only if the system has more than one CPU).
3121 * This primitive provides the guarantees made by the (now removed)
3122 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3123 * guarantees that rcu_read_lock() sections will have completed.
3124 * In "classic RCU", these two guarantees happen to be one and
3125 * the same, but can differ in realtime RCU implementations.
3127 void synchronize_sched(void)
3129 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3130 !lock_is_held(&rcu_lock_map
) &&
3131 !lock_is_held(&rcu_sched_lock_map
),
3132 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3133 if (rcu_blocking_is_gp())
3135 if (rcu_gp_is_expedited())
3136 synchronize_sched_expedited();
3138 wait_rcu_gp(call_rcu_sched
);
3140 EXPORT_SYMBOL_GPL(synchronize_sched
);
3143 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3145 * Control will return to the caller some time after a full rcu_bh grace
3146 * period has elapsed, in other words after all currently executing rcu_bh
3147 * read-side critical sections have completed. RCU read-side critical
3148 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3149 * and may be nested.
3151 * See the description of synchronize_sched() for more detailed information
3152 * on memory ordering guarantees.
3154 void synchronize_rcu_bh(void)
3156 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3157 !lock_is_held(&rcu_lock_map
) &&
3158 !lock_is_held(&rcu_sched_lock_map
),
3159 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3160 if (rcu_blocking_is_gp())
3162 if (rcu_gp_is_expedited())
3163 synchronize_rcu_bh_expedited();
3165 wait_rcu_gp(call_rcu_bh
);
3167 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3170 * get_state_synchronize_rcu - Snapshot current RCU state
3172 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3173 * to determine whether or not a full grace period has elapsed in the
3176 unsigned long get_state_synchronize_rcu(void)
3179 * Any prior manipulation of RCU-protected data must happen
3180 * before the load from ->gpnum.
3185 * Make sure this load happens before the purportedly
3186 * time-consuming work between get_state_synchronize_rcu()
3187 * and cond_synchronize_rcu().
3189 return smp_load_acquire(&rcu_state_p
->gpnum
);
3191 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3194 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3196 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3198 * If a full RCU grace period has elapsed since the earlier call to
3199 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3200 * synchronize_rcu() to wait for a full grace period.
3202 * Yes, this function does not take counter wrap into account. But
3203 * counter wrap is harmless. If the counter wraps, we have waited for
3204 * more than 2 billion grace periods (and way more on a 64-bit system!),
3205 * so waiting for one additional grace period should be just fine.
3207 void cond_synchronize_rcu(unsigned long oldstate
)
3209 unsigned long newstate
;
3212 * Ensure that this load happens before any RCU-destructive
3213 * actions the caller might carry out after we return.
3215 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3216 if (ULONG_CMP_GE(oldstate
, newstate
))
3219 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3221 static int synchronize_sched_expedited_cpu_stop(void *data
)
3224 * There must be a full memory barrier on each affected CPU
3225 * between the time that try_stop_cpus() is called and the
3226 * time that it returns.
3228 * In the current initial implementation of cpu_stop, the
3229 * above condition is already met when the control reaches
3230 * this point and the following smp_mb() is not strictly
3231 * necessary. Do smp_mb() anyway for documentation and
3232 * robustness against future implementation changes.
3234 smp_mb(); /* See above comment block. */
3239 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3241 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3242 * approach to force the grace period to end quickly. This consumes
3243 * significant time on all CPUs and is unfriendly to real-time workloads,
3244 * so is thus not recommended for any sort of common-case code. In fact,
3245 * if you are using synchronize_sched_expedited() in a loop, please
3246 * restructure your code to batch your updates, and then use a single
3247 * synchronize_sched() instead.
3249 * This implementation can be thought of as an application of ticket
3250 * locking to RCU, with sync_sched_expedited_started and
3251 * sync_sched_expedited_done taking on the roles of the halves
3252 * of the ticket-lock word. Each task atomically increments
3253 * sync_sched_expedited_started upon entry, snapshotting the old value,
3254 * then attempts to stop all the CPUs. If this succeeds, then each
3255 * CPU will have executed a context switch, resulting in an RCU-sched
3256 * grace period. We are then done, so we use atomic_cmpxchg() to
3257 * update sync_sched_expedited_done to match our snapshot -- but
3258 * only if someone else has not already advanced past our snapshot.
3260 * On the other hand, if try_stop_cpus() fails, we check the value
3261 * of sync_sched_expedited_done. If it has advanced past our
3262 * initial snapshot, then someone else must have forced a grace period
3263 * some time after we took our snapshot. In this case, our work is
3264 * done for us, and we can simply return. Otherwise, we try again,
3265 * but keep our initial snapshot for purposes of checking for someone
3266 * doing our work for us.
3268 * If we fail too many times in a row, we fall back to synchronize_sched().
3270 void synchronize_sched_expedited(void)
3275 long firstsnap
, s
, snap
;
3277 struct rcu_state
*rsp
= &rcu_sched_state
;
3280 * If we are in danger of counter wrap, just do synchronize_sched().
3281 * By allowing sync_sched_expedited_started to advance no more than
3282 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3283 * that more than 3.5 billion CPUs would be required to force a
3284 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3285 * course be required on a 64-bit system.
3287 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3288 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3290 synchronize_sched();
3291 atomic_long_inc(&rsp
->expedited_wrap
);
3296 * Take a ticket. Note that atomic_inc_return() implies a
3297 * full memory barrier.
3299 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3301 if (!try_get_online_cpus()) {
3302 /* CPU hotplug operation in flight, fall back to normal GP. */
3303 wait_rcu_gp(call_rcu_sched
);
3304 atomic_long_inc(&rsp
->expedited_normal
);
3307 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3309 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3310 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3312 cpumask_copy(cm
, cpu_online_mask
);
3313 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3314 for_each_cpu(cpu
, cm
) {
3315 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3317 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3318 cpumask_clear_cpu(cpu
, cm
);
3320 if (cpumask_weight(cm
) == 0)
3325 * Each pass through the following loop attempts to force a
3326 * context switch on each CPU.
3328 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3329 synchronize_sched_expedited_cpu_stop
,
3332 atomic_long_inc(&rsp
->expedited_tryfail
);
3334 /* Check to see if someone else did our work for us. */
3335 s
= atomic_long_read(&rsp
->expedited_done
);
3336 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3337 /* ensure test happens before caller kfree */
3338 smp_mb__before_atomic(); /* ^^^ */
3339 atomic_long_inc(&rsp
->expedited_workdone1
);
3340 free_cpumask_var(cm
);
3344 /* No joy, try again later. Or just synchronize_sched(). */
3345 if (trycount
++ < 10) {
3346 udelay(trycount
* num_online_cpus());
3348 wait_rcu_gp(call_rcu_sched
);
3349 atomic_long_inc(&rsp
->expedited_normal
);
3350 free_cpumask_var(cm
);
3354 /* Recheck to see if someone else did our work for us. */
3355 s
= atomic_long_read(&rsp
->expedited_done
);
3356 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3357 /* ensure test happens before caller kfree */
3358 smp_mb__before_atomic(); /* ^^^ */
3359 atomic_long_inc(&rsp
->expedited_workdone2
);
3360 free_cpumask_var(cm
);
3365 * Refetching sync_sched_expedited_started allows later
3366 * callers to piggyback on our grace period. We retry
3367 * after they started, so our grace period works for them,
3368 * and they started after our first try, so their grace
3369 * period works for us.
3371 if (!try_get_online_cpus()) {
3372 /* CPU hotplug operation in flight, use normal GP. */
3373 wait_rcu_gp(call_rcu_sched
);
3374 atomic_long_inc(&rsp
->expedited_normal
);
3375 free_cpumask_var(cm
);
3378 snap
= atomic_long_read(&rsp
->expedited_start
);
3379 smp_mb(); /* ensure read is before try_stop_cpus(). */
3381 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3384 free_cpumask_var(cm
);
3387 * Everyone up to our most recent fetch is covered by our grace
3388 * period. Update the counter, but only if our work is still
3389 * relevant -- which it won't be if someone who started later
3390 * than we did already did their update.
3393 atomic_long_inc(&rsp
->expedited_done_tries
);
3394 s
= atomic_long_read(&rsp
->expedited_done
);
3395 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3396 /* ensure test happens before caller kfree */
3397 smp_mb__before_atomic(); /* ^^^ */
3398 atomic_long_inc(&rsp
->expedited_done_lost
);
3401 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3402 atomic_long_inc(&rsp
->expedited_done_exit
);
3406 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3409 * Check to see if there is any immediate RCU-related work to be done
3410 * by the current CPU, for the specified type of RCU, returning 1 if so.
3411 * The checks are in order of increasing expense: checks that can be
3412 * carried out against CPU-local state are performed first. However,
3413 * we must check for CPU stalls first, else we might not get a chance.
3415 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3417 struct rcu_node
*rnp
= rdp
->mynode
;
3419 rdp
->n_rcu_pending
++;
3421 /* Check for CPU stalls, if enabled. */
3422 check_cpu_stall(rsp
, rdp
);
3424 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3425 if (rcu_nohz_full_cpu(rsp
))
3428 /* Is the RCU core waiting for a quiescent state from this CPU? */
3429 if (rcu_scheduler_fully_active
&&
3430 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3431 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3432 rdp
->n_rp_qs_pending
++;
3433 } else if (rdp
->qs_pending
&&
3434 (rdp
->passed_quiesce
||
3435 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3436 rdp
->n_rp_report_qs
++;
3440 /* Does this CPU have callbacks ready to invoke? */
3441 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3442 rdp
->n_rp_cb_ready
++;
3446 /* Has RCU gone idle with this CPU needing another grace period? */
3447 if (cpu_needs_another_gp(rsp
, rdp
)) {
3448 rdp
->n_rp_cpu_needs_gp
++;
3452 /* Has another RCU grace period completed? */
3453 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3454 rdp
->n_rp_gp_completed
++;
3458 /* Has a new RCU grace period started? */
3459 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3460 unlikely(ACCESS_ONCE(rdp
->gpwrap
))) { /* outside lock */
3461 rdp
->n_rp_gp_started
++;
3465 /* Does this CPU need a deferred NOCB wakeup? */
3466 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3467 rdp
->n_rp_nocb_defer_wakeup
++;
3472 rdp
->n_rp_need_nothing
++;
3477 * Check to see if there is any immediate RCU-related work to be done
3478 * by the current CPU, returning 1 if so. This function is part of the
3479 * RCU implementation; it is -not- an exported member of the RCU API.
3481 static int rcu_pending(void)
3483 struct rcu_state
*rsp
;
3485 for_each_rcu_flavor(rsp
)
3486 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3492 * Return true if the specified CPU has any callback. If all_lazy is
3493 * non-NULL, store an indication of whether all callbacks are lazy.
3494 * (If there are no callbacks, all of them are deemed to be lazy.)
3496 static int __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3500 struct rcu_data
*rdp
;
3501 struct rcu_state
*rsp
;
3503 for_each_rcu_flavor(rsp
) {
3504 rdp
= this_cpu_ptr(rsp
->rda
);
3508 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3519 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3520 * the compiler is expected to optimize this away.
3522 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3523 int cpu
, unsigned long done
)
3525 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3526 atomic_read(&rsp
->barrier_cpu_count
), done
);
3530 * RCU callback function for _rcu_barrier(). If we are last, wake
3531 * up the task executing _rcu_barrier().
3533 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3535 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3536 struct rcu_state
*rsp
= rdp
->rsp
;
3538 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3539 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3540 complete(&rsp
->barrier_completion
);
3542 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3547 * Called with preemption disabled, and from cross-cpu IRQ context.
3549 static void rcu_barrier_func(void *type
)
3551 struct rcu_state
*rsp
= type
;
3552 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3554 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3555 atomic_inc(&rsp
->barrier_cpu_count
);
3556 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3560 * Orchestrate the specified type of RCU barrier, waiting for all
3561 * RCU callbacks of the specified type to complete.
3563 static void _rcu_barrier(struct rcu_state
*rsp
)
3566 struct rcu_data
*rdp
;
3567 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
3568 unsigned long snap_done
;
3570 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3572 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3573 mutex_lock(&rsp
->barrier_mutex
);
3576 * Ensure that all prior references, including to ->n_barrier_done,
3577 * are ordered before the _rcu_barrier() machinery.
3579 smp_mb(); /* See above block comment. */
3582 * Recheck ->n_barrier_done to see if others did our work for us.
3583 * This means checking ->n_barrier_done for an even-to-odd-to-even
3584 * transition. The "if" expression below therefore rounds the old
3585 * value up to the next even number and adds two before comparing.
3587 snap_done
= rsp
->n_barrier_done
;
3588 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3591 * If the value in snap is odd, we needed to wait for the current
3592 * rcu_barrier() to complete, then wait for the next one, in other
3593 * words, we need the value of snap_done to be three larger than
3594 * the value of snap. On the other hand, if the value in snap is
3595 * even, we only had to wait for the next rcu_barrier() to complete,
3596 * in other words, we need the value of snap_done to be only two
3597 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3598 * this for us (thank you, Linus!).
3600 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3601 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3602 smp_mb(); /* caller's subsequent code after above check. */
3603 mutex_unlock(&rsp
->barrier_mutex
);
3608 * Increment ->n_barrier_done to avoid duplicate work. Use
3609 * ACCESS_ONCE() to prevent the compiler from speculating
3610 * the increment to precede the early-exit check.
3612 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3613 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3614 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3615 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3618 * Initialize the count to one rather than to zero in order to
3619 * avoid a too-soon return to zero in case of a short grace period
3620 * (or preemption of this task). Exclude CPU-hotplug operations
3621 * to ensure that no offline CPU has callbacks queued.
3623 init_completion(&rsp
->barrier_completion
);
3624 atomic_set(&rsp
->barrier_cpu_count
, 1);
3628 * Force each CPU with callbacks to register a new callback.
3629 * When that callback is invoked, we will know that all of the
3630 * corresponding CPU's preceding callbacks have been invoked.
3632 for_each_possible_cpu(cpu
) {
3633 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3635 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3636 if (rcu_is_nocb_cpu(cpu
)) {
3637 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3638 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3639 rsp
->n_barrier_done
);
3641 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3642 rsp
->n_barrier_done
);
3643 smp_mb__before_atomic();
3644 atomic_inc(&rsp
->barrier_cpu_count
);
3645 __call_rcu(&rdp
->barrier_head
,
3646 rcu_barrier_callback
, rsp
, cpu
, 0);
3648 } else if (ACCESS_ONCE(rdp
->qlen
)) {
3649 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3650 rsp
->n_barrier_done
);
3651 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3653 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3654 rsp
->n_barrier_done
);
3660 * Now that we have an rcu_barrier_callback() callback on each
3661 * CPU, and thus each counted, remove the initial count.
3663 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3664 complete(&rsp
->barrier_completion
);
3666 /* Increment ->n_barrier_done to prevent duplicate work. */
3667 smp_mb(); /* Keep increment after above mechanism. */
3668 ACCESS_ONCE(rsp
->n_barrier_done
) = rsp
->n_barrier_done
+ 1;
3669 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3670 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3671 smp_mb(); /* Keep increment before caller's subsequent code. */
3673 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3674 wait_for_completion(&rsp
->barrier_completion
);
3676 /* Other rcu_barrier() invocations can now safely proceed. */
3677 mutex_unlock(&rsp
->barrier_mutex
);
3681 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3683 void rcu_barrier_bh(void)
3685 _rcu_barrier(&rcu_bh_state
);
3687 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3690 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3692 void rcu_barrier_sched(void)
3694 _rcu_barrier(&rcu_sched_state
);
3696 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3699 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3700 * first CPU in a given leaf rcu_node structure coming online. The caller
3701 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3704 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
3707 struct rcu_node
*rnp
= rnp_leaf
;
3710 mask
= rnp
->grpmask
;
3714 raw_spin_lock(&rnp
->lock
); /* Interrupts already disabled. */
3715 rnp
->qsmaskinit
|= mask
;
3716 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
3721 * Do boot-time initialization of a CPU's per-CPU RCU data.
3724 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3726 unsigned long flags
;
3727 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3728 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3730 /* Set up local state, ensuring consistent view of global state. */
3731 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3732 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3733 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3734 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3735 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3738 rcu_boot_init_nocb_percpu_data(rdp
);
3739 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3743 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3744 * offline event can be happening at a given time. Note also that we
3745 * can accept some slop in the rsp->completed access due to the fact
3746 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3749 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3751 unsigned long flags
;
3753 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3754 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3756 /* Set up local state, ensuring consistent view of global state. */
3757 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3758 rdp
->beenonline
= 1; /* We have now been online. */
3759 rdp
->qlen_last_fqs_check
= 0;
3760 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3761 rdp
->blimit
= blimit
;
3763 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3764 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3765 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3766 atomic_set(&rdp
->dynticks
->dynticks
,
3767 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3768 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3771 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3772 * propagation up the rcu_node tree will happen at the beginning
3773 * of the next grace period.
3776 mask
= rdp
->grpmask
;
3777 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3778 smp_mb__after_unlock_lock();
3779 rnp
->qsmaskinitnext
|= mask
;
3780 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
3781 rdp
->completed
= rnp
->completed
;
3782 rdp
->passed_quiesce
= false;
3783 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
3784 rdp
->qs_pending
= false;
3785 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3786 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3789 static void rcu_prepare_cpu(int cpu
)
3791 struct rcu_state
*rsp
;
3793 for_each_rcu_flavor(rsp
)
3794 rcu_init_percpu_data(cpu
, rsp
);
3798 * Handle CPU online/offline notification events.
3800 int rcu_cpu_notify(struct notifier_block
*self
,
3801 unsigned long action
, void *hcpu
)
3803 long cpu
= (long)hcpu
;
3804 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3805 struct rcu_node
*rnp
= rdp
->mynode
;
3806 struct rcu_state
*rsp
;
3809 case CPU_UP_PREPARE
:
3810 case CPU_UP_PREPARE_FROZEN
:
3811 rcu_prepare_cpu(cpu
);
3812 rcu_prepare_kthreads(cpu
);
3813 rcu_spawn_all_nocb_kthreads(cpu
);
3816 case CPU_DOWN_FAILED
:
3817 rcu_boost_kthread_setaffinity(rnp
, -1);
3819 case CPU_DOWN_PREPARE
:
3820 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3823 case CPU_DYING_FROZEN
:
3824 for_each_rcu_flavor(rsp
)
3825 rcu_cleanup_dying_cpu(rsp
);
3827 case CPU_DYING_IDLE
:
3828 for_each_rcu_flavor(rsp
) {
3829 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
3833 case CPU_DEAD_FROZEN
:
3834 case CPU_UP_CANCELED
:
3835 case CPU_UP_CANCELED_FROZEN
:
3836 for_each_rcu_flavor(rsp
) {
3837 rcu_cleanup_dead_cpu(cpu
, rsp
);
3838 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3847 static int rcu_pm_notify(struct notifier_block
*self
,
3848 unsigned long action
, void *hcpu
)
3851 case PM_HIBERNATION_PREPARE
:
3852 case PM_SUSPEND_PREPARE
:
3853 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3856 case PM_POST_HIBERNATION
:
3857 case PM_POST_SUSPEND
:
3858 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3859 rcu_unexpedite_gp();
3868 * Spawn the kthreads that handle each RCU flavor's grace periods.
3870 static int __init
rcu_spawn_gp_kthread(void)
3872 unsigned long flags
;
3873 int kthread_prio_in
= kthread_prio
;
3874 struct rcu_node
*rnp
;
3875 struct rcu_state
*rsp
;
3876 struct sched_param sp
;
3877 struct task_struct
*t
;
3879 /* Force priority into range. */
3880 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3882 else if (kthread_prio
< 0)
3884 else if (kthread_prio
> 99)
3886 if (kthread_prio
!= kthread_prio_in
)
3887 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3888 kthread_prio
, kthread_prio_in
);
3890 rcu_scheduler_fully_active
= 1;
3891 for_each_rcu_flavor(rsp
) {
3892 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3894 rnp
= rcu_get_root(rsp
);
3895 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3896 rsp
->gp_kthread
= t
;
3898 sp
.sched_priority
= kthread_prio
;
3899 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3902 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3904 rcu_spawn_nocb_kthreads();
3905 rcu_spawn_boost_kthreads();
3908 early_initcall(rcu_spawn_gp_kthread
);
3911 * This function is invoked towards the end of the scheduler's initialization
3912 * process. Before this is called, the idle task might contain
3913 * RCU read-side critical sections (during which time, this idle
3914 * task is booting the system). After this function is called, the
3915 * idle tasks are prohibited from containing RCU read-side critical
3916 * sections. This function also enables RCU lockdep checking.
3918 void rcu_scheduler_starting(void)
3920 WARN_ON(num_online_cpus() != 1);
3921 WARN_ON(nr_context_switches() > 0);
3922 rcu_scheduler_active
= 1;
3926 * Compute the per-level fanout, either using the exact fanout specified
3927 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3929 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3933 if (IS_ENABLED(CONFIG_RCU_FANOUT_EXACT
)) {
3934 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3935 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3936 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3942 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3943 ccur
= rsp
->levelcnt
[i
];
3944 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3951 * Helper function for rcu_init() that initializes one rcu_state structure.
3953 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3954 struct rcu_data __percpu
*rda
)
3956 static const char * const buf
[] = {
3960 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3961 static const char * const fqs
[] = {
3965 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3966 static u8 fl_mask
= 0x1;
3970 struct rcu_node
*rnp
;
3972 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3974 /* Silence gcc 4.8 warning about array index out of range. */
3975 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3976 panic("rcu_init_one: rcu_num_lvls overflow");
3978 /* Initialize the level-tracking arrays. */
3980 for (i
= 0; i
< rcu_num_lvls
; i
++)
3981 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3982 for (i
= 1; i
< rcu_num_lvls
; i
++)
3983 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3984 rcu_init_levelspread(rsp
);
3985 rsp
->flavor_mask
= fl_mask
;
3988 /* Initialize the elements themselves, starting from the leaves. */
3990 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3991 cpustride
*= rsp
->levelspread
[i
];
3992 rnp
= rsp
->level
[i
];
3993 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3994 raw_spin_lock_init(&rnp
->lock
);
3995 lockdep_set_class_and_name(&rnp
->lock
,
3996 &rcu_node_class
[i
], buf
[i
]);
3997 raw_spin_lock_init(&rnp
->fqslock
);
3998 lockdep_set_class_and_name(&rnp
->fqslock
,
3999 &rcu_fqs_class
[i
], fqs
[i
]);
4000 rnp
->gpnum
= rsp
->gpnum
;
4001 rnp
->completed
= rsp
->completed
;
4003 rnp
->qsmaskinit
= 0;
4004 rnp
->grplo
= j
* cpustride
;
4005 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4006 if (rnp
->grphi
>= nr_cpu_ids
)
4007 rnp
->grphi
= nr_cpu_ids
- 1;
4013 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
4014 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4015 rnp
->parent
= rsp
->level
[i
- 1] +
4016 j
/ rsp
->levelspread
[i
- 1];
4019 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4020 rcu_init_one_nocb(rnp
);
4024 init_waitqueue_head(&rsp
->gp_wq
);
4025 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4026 for_each_possible_cpu(i
) {
4027 while (i
> rnp
->grphi
)
4029 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4030 rcu_boot_init_percpu_data(i
, rsp
);
4032 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4036 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4037 * replace the definitions in tree.h because those are needed to size
4038 * the ->node array in the rcu_state structure.
4040 static void __init
rcu_init_geometry(void)
4046 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
4049 * Initialize any unspecified boot parameters.
4050 * The default values of jiffies_till_first_fqs and
4051 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4052 * value, which is a function of HZ, then adding one for each
4053 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4055 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4056 if (jiffies_till_first_fqs
== ULONG_MAX
)
4057 jiffies_till_first_fqs
= d
;
4058 if (jiffies_till_next_fqs
== ULONG_MAX
)
4059 jiffies_till_next_fqs
= d
;
4061 /* If the compile-time values are accurate, just leave. */
4062 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
4063 nr_cpu_ids
== NR_CPUS
)
4065 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4066 rcu_fanout_leaf
, nr_cpu_ids
);
4069 * Compute number of nodes that can be handled an rcu_node tree
4070 * with the given number of levels. Setting rcu_capacity[0] makes
4071 * some of the arithmetic easier.
4073 rcu_capacity
[0] = 1;
4074 rcu_capacity
[1] = rcu_fanout_leaf
;
4075 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
4076 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
4079 * The boot-time rcu_fanout_leaf parameter is only permitted
4080 * to increase the leaf-level fanout, not decrease it. Of course,
4081 * the leaf-level fanout cannot exceed the number of bits in
4082 * the rcu_node masks. Finally, the tree must be able to accommodate
4083 * the configured number of CPUs. Complain and fall back to the
4084 * compile-time values if these limits are exceeded.
4086 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
4087 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
4088 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
4093 /* Calculate the number of rcu_nodes at each level of the tree. */
4094 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
4095 if (n
<= rcu_capacity
[i
]) {
4096 for (j
= 0; j
<= i
; j
++)
4098 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
4100 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
4105 /* Calculate the total number of rcu_node structures. */
4107 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
4108 rcu_num_nodes
+= num_rcu_lvl
[i
];
4112 void __init
rcu_init(void)
4116 rcu_early_boot_tests();
4118 rcu_bootup_announce();
4119 rcu_init_geometry();
4120 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
4121 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
4122 __rcu_init_preempt();
4123 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4126 * We don't need protection against CPU-hotplug here because
4127 * this is called early in boot, before either interrupts
4128 * or the scheduler are operational.
4130 cpu_notifier(rcu_cpu_notify
, 0);
4131 pm_notifier(rcu_pm_notify
, 0);
4132 for_each_online_cpu(cpu
)
4133 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4136 #include "tree_plugin.h"