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 */
174 * Number of grace periods between delays, normalized by the duration of
175 * the delay. The longer the the delay, the more the grace periods between
176 * each delay. The reason for this normalization is that it means that,
177 * for non-zero delays, the overall slowdown of grace periods is constant
178 * regardless of the duration of the delay. This arrangement balances
179 * the need for long delays to increase some race probabilities with the
180 * need for fast grace periods to increase other race probabilities.
182 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
185 * Track the rcutorture test sequence number and the update version
186 * number within a given test. The rcutorture_testseq is incremented
187 * on every rcutorture module load and unload, so has an odd value
188 * when a test is running. The rcutorture_vernum is set to zero
189 * when rcutorture starts and is incremented on each rcutorture update.
190 * These variables enable correlating rcutorture output with the
191 * RCU tracing information.
193 unsigned long rcutorture_testseq
;
194 unsigned long rcutorture_vernum
;
197 * Compute the mask of online CPUs for the specified rcu_node structure.
198 * This will not be stable unless the rcu_node structure's ->lock is
199 * held, but the bit corresponding to the current CPU will be stable
202 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
204 return READ_ONCE(rnp
->qsmaskinitnext
);
208 * Return true if an RCU grace period is in progress. The READ_ONCE()s
209 * permit this function to be invoked without holding the root rcu_node
210 * structure's ->lock, but of course results can be subject to change.
212 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
214 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
218 * Note a quiescent state. Because we do not need to know
219 * how many quiescent states passed, just if there was at least
220 * one since the start of the grace period, this just sets a flag.
221 * The caller must have disabled preemption.
223 void rcu_sched_qs(void)
225 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
226 trace_rcu_grace_period(TPS("rcu_sched"),
227 __this_cpu_read(rcu_sched_data
.gpnum
),
229 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
235 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
236 trace_rcu_grace_period(TPS("rcu_bh"),
237 __this_cpu_read(rcu_bh_data
.gpnum
),
239 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
243 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
245 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
246 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
247 .dynticks
= ATOMIC_INIT(1),
248 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
249 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
250 .dynticks_idle
= ATOMIC_INIT(1),
251 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
254 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
255 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
258 * Let the RCU core know that this CPU has gone through the scheduler,
259 * which is a quiescent state. This is called when the need for a
260 * quiescent state is urgent, so we burn an atomic operation and full
261 * memory barriers to let the RCU core know about it, regardless of what
262 * this CPU might (or might not) do in the near future.
264 * We inform the RCU core by emulating a zero-duration dyntick-idle
265 * period, which we in turn do by incrementing the ->dynticks counter
268 static void rcu_momentary_dyntick_idle(void)
271 struct rcu_data
*rdp
;
272 struct rcu_dynticks
*rdtp
;
274 struct rcu_state
*rsp
;
276 local_irq_save(flags
);
279 * Yes, we can lose flag-setting operations. This is OK, because
280 * the flag will be set again after some delay.
282 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
283 raw_cpu_write(rcu_sched_qs_mask
, 0);
285 /* Find the flavor that needs a quiescent state. */
286 for_each_rcu_flavor(rsp
) {
287 rdp
= raw_cpu_ptr(rsp
->rda
);
288 if (!(resched_mask
& rsp
->flavor_mask
))
290 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
291 if (READ_ONCE(rdp
->mynode
->completed
) !=
292 READ_ONCE(rdp
->cond_resched_completed
))
296 * Pretend to be momentarily idle for the quiescent state.
297 * This allows the grace-period kthread to record the
298 * quiescent state, with no need for this CPU to do anything
301 rdtp
= this_cpu_ptr(&rcu_dynticks
);
302 smp_mb__before_atomic(); /* Earlier stuff before QS. */
303 atomic_add(2, &rdtp
->dynticks
); /* QS. */
304 smp_mb__after_atomic(); /* Later stuff after QS. */
307 local_irq_restore(flags
);
311 * Note a context switch. This is a quiescent state for RCU-sched,
312 * and requires special handling for preemptible RCU.
313 * The caller must have disabled preemption.
315 void rcu_note_context_switch(void)
317 trace_rcu_utilization(TPS("Start context switch"));
319 rcu_preempt_note_context_switch();
320 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
321 rcu_momentary_dyntick_idle();
322 trace_rcu_utilization(TPS("End context switch"));
324 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
327 * Register a quiescent state for all RCU flavors. If there is an
328 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
329 * dyntick-idle quiescent state visible to other CPUs (but only for those
330 * RCU flavors in desperate need of a quiescent state, which will normally
331 * be none of them). Either way, do a lightweight quiescent state for
334 void rcu_all_qs(void)
336 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
337 rcu_momentary_dyntick_idle();
338 this_cpu_inc(rcu_qs_ctr
);
340 EXPORT_SYMBOL_GPL(rcu_all_qs
);
342 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
343 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
344 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
346 module_param(blimit
, long, 0444);
347 module_param(qhimark
, long, 0444);
348 module_param(qlowmark
, long, 0444);
350 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
351 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
353 module_param(jiffies_till_first_fqs
, ulong
, 0644);
354 module_param(jiffies_till_next_fqs
, ulong
, 0644);
357 * How long the grace period must be before we start recruiting
358 * quiescent-state help from rcu_note_context_switch().
360 static ulong jiffies_till_sched_qs
= HZ
/ 20;
361 module_param(jiffies_till_sched_qs
, ulong
, 0644);
363 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
364 struct rcu_data
*rdp
);
365 static void force_qs_rnp(struct rcu_state
*rsp
,
366 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
367 unsigned long *maxj
),
368 bool *isidle
, unsigned long *maxj
);
369 static void force_quiescent_state(struct rcu_state
*rsp
);
370 static int rcu_pending(void);
373 * Return the number of RCU batches started thus far for debug & stats.
375 unsigned long rcu_batches_started(void)
377 return rcu_state_p
->gpnum
;
379 EXPORT_SYMBOL_GPL(rcu_batches_started
);
382 * Return the number of RCU-sched batches started thus far for debug & stats.
384 unsigned long rcu_batches_started_sched(void)
386 return rcu_sched_state
.gpnum
;
388 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
391 * Return the number of RCU BH batches started thus far for debug & stats.
393 unsigned long rcu_batches_started_bh(void)
395 return rcu_bh_state
.gpnum
;
397 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
400 * Return the number of RCU batches completed thus far for debug & stats.
402 unsigned long rcu_batches_completed(void)
404 return rcu_state_p
->completed
;
406 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
409 * Return the number of RCU-sched batches completed thus far for debug & stats.
411 unsigned long rcu_batches_completed_sched(void)
413 return rcu_sched_state
.completed
;
415 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
418 * Return the number of RCU BH batches completed thus far for debug & stats.
420 unsigned long rcu_batches_completed_bh(void)
422 return rcu_bh_state
.completed
;
424 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
427 * Force a quiescent state.
429 void rcu_force_quiescent_state(void)
431 force_quiescent_state(rcu_state_p
);
433 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
436 * Force a quiescent state for RCU BH.
438 void rcu_bh_force_quiescent_state(void)
440 force_quiescent_state(&rcu_bh_state
);
442 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
445 * Force a quiescent state for RCU-sched.
447 void rcu_sched_force_quiescent_state(void)
449 force_quiescent_state(&rcu_sched_state
);
451 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
454 * Show the state of the grace-period kthreads.
456 void show_rcu_gp_kthreads(void)
458 struct rcu_state
*rsp
;
460 for_each_rcu_flavor(rsp
) {
461 pr_info("%s: wait state: %d ->state: %#lx\n",
462 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
463 /* sched_show_task(rsp->gp_kthread); */
466 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
469 * Record the number of times rcutorture tests have been initiated and
470 * terminated. This information allows the debugfs tracing stats to be
471 * correlated to the rcutorture messages, even when the rcutorture module
472 * is being repeatedly loaded and unloaded. In other words, we cannot
473 * store this state in rcutorture itself.
475 void rcutorture_record_test_transition(void)
477 rcutorture_testseq
++;
478 rcutorture_vernum
= 0;
480 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
483 * Send along grace-period-related data for rcutorture diagnostics.
485 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
486 unsigned long *gpnum
, unsigned long *completed
)
488 struct rcu_state
*rsp
= NULL
;
497 case RCU_SCHED_FLAVOR
:
498 rsp
= &rcu_sched_state
;
504 *flags
= READ_ONCE(rsp
->gp_flags
);
505 *gpnum
= READ_ONCE(rsp
->gpnum
);
506 *completed
= READ_ONCE(rsp
->completed
);
513 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
516 * Record the number of writer passes through the current rcutorture test.
517 * This is also used to correlate debugfs tracing stats with the rcutorture
520 void rcutorture_record_progress(unsigned long vernum
)
524 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
527 * Does the CPU have callbacks ready to be invoked?
530 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
532 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
533 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
537 * Return the root node of the specified rcu_state structure.
539 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
541 return &rsp
->node
[0];
545 * Is there any need for future grace periods?
546 * Interrupts must be disabled. If the caller does not hold the root
547 * rnp_node structure's ->lock, the results are advisory only.
549 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
551 struct rcu_node
*rnp
= rcu_get_root(rsp
);
552 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
553 int *fp
= &rnp
->need_future_gp
[idx
];
555 return READ_ONCE(*fp
);
559 * Does the current CPU require a not-yet-started grace period?
560 * The caller must have disabled interrupts to prevent races with
561 * normal callback registry.
564 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
568 if (rcu_gp_in_progress(rsp
))
569 return 0; /* No, a grace period is already in progress. */
570 if (rcu_future_needs_gp(rsp
))
571 return 1; /* Yes, a no-CBs CPU needs one. */
572 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
573 return 0; /* No, this is a no-CBs (or offline) CPU. */
574 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
575 return 1; /* Yes, this CPU has newly registered callbacks. */
576 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
577 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
578 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
579 rdp
->nxtcompleted
[i
]))
580 return 1; /* Yes, CBs for future grace period. */
581 return 0; /* No grace period needed. */
585 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
587 * If the new value of the ->dynticks_nesting counter now is zero,
588 * we really have entered idle, and must do the appropriate accounting.
589 * The caller must have disabled interrupts.
591 static void rcu_eqs_enter_common(long long oldval
, bool user
)
593 struct rcu_state
*rsp
;
594 struct rcu_data
*rdp
;
595 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
597 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
598 if (!user
&& !is_idle_task(current
)) {
599 struct task_struct
*idle __maybe_unused
=
600 idle_task(smp_processor_id());
602 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
603 ftrace_dump(DUMP_ORIG
);
604 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
605 current
->pid
, current
->comm
,
606 idle
->pid
, idle
->comm
); /* must be idle task! */
608 for_each_rcu_flavor(rsp
) {
609 rdp
= this_cpu_ptr(rsp
->rda
);
610 do_nocb_deferred_wakeup(rdp
);
612 rcu_prepare_for_idle();
613 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
614 smp_mb__before_atomic(); /* See above. */
615 atomic_inc(&rdtp
->dynticks
);
616 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
617 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
618 rcu_dynticks_task_enter();
621 * It is illegal to enter an extended quiescent state while
622 * in an RCU read-side critical section.
624 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
625 "Illegal idle entry in RCU read-side critical section.");
626 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
627 "Illegal idle entry in RCU-bh read-side critical section.");
628 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
629 "Illegal idle entry in RCU-sched read-side critical section.");
633 * Enter an RCU extended quiescent state, which can be either the
634 * idle loop or adaptive-tickless usermode execution.
636 static void rcu_eqs_enter(bool user
)
639 struct rcu_dynticks
*rdtp
;
641 rdtp
= this_cpu_ptr(&rcu_dynticks
);
642 oldval
= rdtp
->dynticks_nesting
;
643 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
644 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
645 rdtp
->dynticks_nesting
= 0;
646 rcu_eqs_enter_common(oldval
, user
);
648 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
653 * rcu_idle_enter - inform RCU that current CPU is entering idle
655 * Enter idle mode, in other words, -leave- the mode in which RCU
656 * read-side critical sections can occur. (Though RCU read-side
657 * critical sections can occur in irq handlers in idle, a possibility
658 * handled by irq_enter() and irq_exit().)
660 * We crowbar the ->dynticks_nesting field to zero to allow for
661 * the possibility of usermode upcalls having messed up our count
662 * of interrupt nesting level during the prior busy period.
664 void rcu_idle_enter(void)
668 local_irq_save(flags
);
669 rcu_eqs_enter(false);
670 rcu_sysidle_enter(0);
671 local_irq_restore(flags
);
673 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
675 #ifdef CONFIG_RCU_USER_QS
677 * rcu_user_enter - inform RCU that we are resuming userspace.
679 * Enter RCU idle mode right before resuming userspace. No use of RCU
680 * is permitted between this call and rcu_user_exit(). This way the
681 * CPU doesn't need to maintain the tick for RCU maintenance purposes
682 * when the CPU runs in userspace.
684 void rcu_user_enter(void)
688 #endif /* CONFIG_RCU_USER_QS */
691 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
693 * Exit from an interrupt handler, which might possibly result in entering
694 * idle mode, in other words, leaving the mode in which read-side critical
695 * sections can occur.
697 * This code assumes that the idle loop never does anything that might
698 * result in unbalanced calls to irq_enter() and irq_exit(). If your
699 * architecture violates this assumption, RCU will give you what you
700 * deserve, good and hard. But very infrequently and irreproducibly.
702 * Use things like work queues to work around this limitation.
704 * You have been warned.
706 void rcu_irq_exit(void)
710 struct rcu_dynticks
*rdtp
;
712 local_irq_save(flags
);
713 rdtp
= this_cpu_ptr(&rcu_dynticks
);
714 oldval
= rdtp
->dynticks_nesting
;
715 rdtp
->dynticks_nesting
--;
716 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
717 if (rdtp
->dynticks_nesting
)
718 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
720 rcu_eqs_enter_common(oldval
, true);
721 rcu_sysidle_enter(1);
722 local_irq_restore(flags
);
726 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
728 * If the new value of the ->dynticks_nesting counter was previously zero,
729 * we really have exited idle, and must do the appropriate accounting.
730 * The caller must have disabled interrupts.
732 static void rcu_eqs_exit_common(long long oldval
, int user
)
734 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
736 rcu_dynticks_task_exit();
737 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
738 atomic_inc(&rdtp
->dynticks
);
739 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
740 smp_mb__after_atomic(); /* See above. */
741 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
742 rcu_cleanup_after_idle();
743 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
744 if (!user
&& !is_idle_task(current
)) {
745 struct task_struct
*idle __maybe_unused
=
746 idle_task(smp_processor_id());
748 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
749 oldval
, rdtp
->dynticks_nesting
);
750 ftrace_dump(DUMP_ORIG
);
751 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
752 current
->pid
, current
->comm
,
753 idle
->pid
, idle
->comm
); /* must be idle task! */
758 * Exit an RCU extended quiescent state, which can be either the
759 * idle loop or adaptive-tickless usermode execution.
761 static void rcu_eqs_exit(bool user
)
763 struct rcu_dynticks
*rdtp
;
766 rdtp
= this_cpu_ptr(&rcu_dynticks
);
767 oldval
= rdtp
->dynticks_nesting
;
768 WARN_ON_ONCE(oldval
< 0);
769 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
770 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
772 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
773 rcu_eqs_exit_common(oldval
, user
);
778 * rcu_idle_exit - inform RCU that current CPU is leaving idle
780 * Exit idle mode, in other words, -enter- the mode in which RCU
781 * read-side critical sections can occur.
783 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
784 * allow for the possibility of usermode upcalls messing up our count
785 * of interrupt nesting level during the busy period that is just
788 void rcu_idle_exit(void)
792 local_irq_save(flags
);
795 local_irq_restore(flags
);
797 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
799 #ifdef CONFIG_RCU_USER_QS
801 * rcu_user_exit - inform RCU that we are exiting userspace.
803 * Exit RCU idle mode while entering the kernel because it can
804 * run a RCU read side critical section anytime.
806 void rcu_user_exit(void)
810 #endif /* CONFIG_RCU_USER_QS */
813 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
815 * Enter an interrupt handler, which might possibly result in exiting
816 * idle mode, in other words, entering the mode in which read-side critical
817 * sections can occur.
819 * Note that the Linux kernel is fully capable of entering an interrupt
820 * handler that it never exits, for example when doing upcalls to
821 * user mode! This code assumes that the idle loop never does upcalls to
822 * user mode. If your architecture does do upcalls from the idle loop (or
823 * does anything else that results in unbalanced calls to the irq_enter()
824 * and irq_exit() functions), RCU will give you what you deserve, good
825 * and hard. But very infrequently and irreproducibly.
827 * Use things like work queues to work around this limitation.
829 * You have been warned.
831 void rcu_irq_enter(void)
834 struct rcu_dynticks
*rdtp
;
837 local_irq_save(flags
);
838 rdtp
= this_cpu_ptr(&rcu_dynticks
);
839 oldval
= rdtp
->dynticks_nesting
;
840 rdtp
->dynticks_nesting
++;
841 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
843 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
845 rcu_eqs_exit_common(oldval
, true);
847 local_irq_restore(flags
);
851 * rcu_nmi_enter - inform RCU of entry to NMI context
853 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
854 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
855 * that the CPU is active. This implementation permits nested NMIs, as
856 * long as the nesting level does not overflow an int. (You will probably
857 * run out of stack space first.)
859 void rcu_nmi_enter(void)
861 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
864 /* Complain about underflow. */
865 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
868 * If idle from RCU viewpoint, atomically increment ->dynticks
869 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
870 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
871 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
872 * to be in the outermost NMI handler that interrupted an RCU-idle
873 * period (observation due to Andy Lutomirski).
875 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
876 smp_mb__before_atomic(); /* Force delay from prior write. */
877 atomic_inc(&rdtp
->dynticks
);
878 /* atomic_inc() before later RCU read-side crit sects */
879 smp_mb__after_atomic(); /* See above. */
880 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
883 rdtp
->dynticks_nmi_nesting
+= incby
;
888 * rcu_nmi_exit - inform RCU of exit from NMI context
890 * If we are returning from the outermost NMI handler that interrupted an
891 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
892 * to let the RCU grace-period handling know that the CPU is back to
895 void rcu_nmi_exit(void)
897 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
900 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
901 * (We are exiting an NMI handler, so RCU better be paying attention
904 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
905 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
908 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
909 * leave it in non-RCU-idle state.
911 if (rdtp
->dynticks_nmi_nesting
!= 1) {
912 rdtp
->dynticks_nmi_nesting
-= 2;
916 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
917 rdtp
->dynticks_nmi_nesting
= 0;
918 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
919 smp_mb__before_atomic(); /* See above. */
920 atomic_inc(&rdtp
->dynticks
);
921 smp_mb__after_atomic(); /* Force delay to next write. */
922 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
926 * __rcu_is_watching - are RCU read-side critical sections safe?
928 * Return true if RCU is watching the running CPU, which means that
929 * this CPU can safely enter RCU read-side critical sections. Unlike
930 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
931 * least disabled preemption.
933 bool notrace
__rcu_is_watching(void)
935 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
939 * rcu_is_watching - see if RCU thinks that the current CPU is idle
941 * If the current CPU is in its idle loop and is neither in an interrupt
942 * or NMI handler, return true.
944 bool notrace
rcu_is_watching(void)
949 ret
= __rcu_is_watching();
953 EXPORT_SYMBOL_GPL(rcu_is_watching
);
955 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
958 * Is the current CPU online? Disable preemption to avoid false positives
959 * that could otherwise happen due to the current CPU number being sampled,
960 * this task being preempted, its old CPU being taken offline, resuming
961 * on some other CPU, then determining that its old CPU is now offline.
962 * It is OK to use RCU on an offline processor during initial boot, hence
963 * the check for rcu_scheduler_fully_active. Note also that it is OK
964 * for a CPU coming online to use RCU for one jiffy prior to marking itself
965 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
966 * offline to continue to use RCU for one jiffy after marking itself
967 * offline in the cpu_online_mask. This leniency is necessary given the
968 * non-atomic nature of the online and offline processing, for example,
969 * the fact that a CPU enters the scheduler after completing the CPU_DYING
972 * This is also why RCU internally marks CPUs online during the
973 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
975 * Disable checking if in an NMI handler because we cannot safely report
976 * errors from NMI handlers anyway.
978 bool rcu_lockdep_current_cpu_online(void)
980 struct rcu_data
*rdp
;
981 struct rcu_node
*rnp
;
987 rdp
= this_cpu_ptr(&rcu_sched_data
);
989 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
990 !rcu_scheduler_fully_active
;
994 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
996 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
999 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1001 * If the current CPU is idle or running at a first-level (not nested)
1002 * interrupt from idle, return true. The caller must have at least
1003 * disabled preemption.
1005 static int rcu_is_cpu_rrupt_from_idle(void)
1007 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1011 * Snapshot the specified CPU's dynticks counter so that we can later
1012 * credit them with an implicit quiescent state. Return 1 if this CPU
1013 * is in dynticks idle mode, which is an extended quiescent state.
1015 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1016 bool *isidle
, unsigned long *maxj
)
1018 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1019 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1020 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1021 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1024 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1025 rdp
->mynode
->gpnum
))
1026 WRITE_ONCE(rdp
->gpwrap
, true);
1032 * Return true if the specified CPU has passed through a quiescent
1033 * state by virtue of being in or having passed through an dynticks
1034 * idle state since the last call to dyntick_save_progress_counter()
1035 * for this same CPU, or by virtue of having been offline.
1037 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1038 bool *isidle
, unsigned long *maxj
)
1044 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1045 snap
= (unsigned int)rdp
->dynticks_snap
;
1048 * If the CPU passed through or entered a dynticks idle phase with
1049 * no active irq/NMI handlers, then we can safely pretend that the CPU
1050 * already acknowledged the request to pass through a quiescent
1051 * state. Either way, that CPU cannot possibly be in an RCU
1052 * read-side critical section that started before the beginning
1053 * of the current RCU grace period.
1055 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1056 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1057 rdp
->dynticks_fqs
++;
1062 * Check for the CPU being offline, but only if the grace period
1063 * is old enough. We don't need to worry about the CPU changing
1064 * state: If we see it offline even once, it has been through a
1067 * The reason for insisting that the grace period be at least
1068 * one jiffy old is that CPUs that are not quite online and that
1069 * have just gone offline can still execute RCU read-side critical
1072 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1073 return 0; /* Grace period is not old enough. */
1075 if (cpu_is_offline(rdp
->cpu
)) {
1076 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1082 * A CPU running for an extended time within the kernel can
1083 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1084 * even context-switching back and forth between a pair of
1085 * in-kernel CPU-bound tasks cannot advance grace periods.
1086 * So if the grace period is old enough, make the CPU pay attention.
1087 * Note that the unsynchronized assignments to the per-CPU
1088 * rcu_sched_qs_mask variable are safe. Yes, setting of
1089 * bits can be lost, but they will be set again on the next
1090 * force-quiescent-state pass. So lost bit sets do not result
1091 * in incorrect behavior, merely in a grace period lasting
1092 * a few jiffies longer than it might otherwise. Because
1093 * there are at most four threads involved, and because the
1094 * updates are only once every few jiffies, the probability of
1095 * lossage (and thus of slight grace-period extension) is
1098 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1099 * is set too high, we override with half of the RCU CPU stall
1102 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1103 if (ULONG_CMP_GE(jiffies
,
1104 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1105 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1106 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1107 WRITE_ONCE(rdp
->cond_resched_completed
,
1108 READ_ONCE(rdp
->mynode
->completed
));
1109 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1111 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1112 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1113 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1114 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1115 /* Time to beat on that CPU again! */
1116 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1117 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1124 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1126 unsigned long j
= jiffies
;
1130 smp_wmb(); /* Record start time before stall time. */
1131 j1
= rcu_jiffies_till_stall_check();
1132 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1133 rsp
->jiffies_resched
= j
+ j1
/ 2;
1134 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1138 * Complain about starvation of grace-period kthread.
1140 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1146 gpa
= READ_ONCE(rsp
->gp_activity
);
1147 if (j
- gpa
> 2 * HZ
)
1148 pr_err("%s kthread starved for %ld jiffies!\n",
1149 rsp
->name
, j
- gpa
);
1153 * Dump stacks of all tasks running on stalled CPUs.
1155 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1158 unsigned long flags
;
1159 struct rcu_node
*rnp
;
1161 rcu_for_each_leaf_node(rsp
, rnp
) {
1162 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1163 if (rnp
->qsmask
!= 0) {
1164 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1165 if (rnp
->qsmask
& (1UL << cpu
))
1166 dump_cpu_task(rnp
->grplo
+ cpu
);
1168 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1172 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1176 unsigned long flags
;
1180 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1183 /* Only let one CPU complain about others per time interval. */
1185 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1186 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1187 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1188 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1191 WRITE_ONCE(rsp
->jiffies_stall
,
1192 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1193 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1196 * OK, time to rat on our buddy...
1197 * See Documentation/RCU/stallwarn.txt for info on how to debug
1198 * RCU CPU stall warnings.
1200 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1202 print_cpu_stall_info_begin();
1203 rcu_for_each_leaf_node(rsp
, rnp
) {
1204 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1205 ndetected
+= rcu_print_task_stall(rnp
);
1206 if (rnp
->qsmask
!= 0) {
1207 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1208 if (rnp
->qsmask
& (1UL << cpu
)) {
1209 print_cpu_stall_info(rsp
,
1214 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1217 print_cpu_stall_info_end();
1218 for_each_possible_cpu(cpu
)
1219 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1220 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1221 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1222 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1224 rcu_dump_cpu_stacks(rsp
);
1226 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1227 READ_ONCE(rsp
->completed
) == gpnum
) {
1228 pr_err("INFO: Stall ended before state dump start\n");
1231 gpa
= READ_ONCE(rsp
->gp_activity
);
1232 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1233 rsp
->name
, j
- gpa
, j
, gpa
,
1234 jiffies_till_next_fqs
,
1235 rcu_get_root(rsp
)->qsmask
);
1236 /* In this case, the current CPU might be at fault. */
1237 sched_show_task(current
);
1241 /* Complain about tasks blocking the grace period. */
1242 rcu_print_detail_task_stall(rsp
);
1244 rcu_check_gp_kthread_starvation(rsp
);
1246 force_quiescent_state(rsp
); /* Kick them all. */
1249 static void print_cpu_stall(struct rcu_state
*rsp
)
1252 unsigned long flags
;
1253 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1257 * OK, time to rat on ourselves...
1258 * See Documentation/RCU/stallwarn.txt for info on how to debug
1259 * RCU CPU stall warnings.
1261 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1262 print_cpu_stall_info_begin();
1263 print_cpu_stall_info(rsp
, smp_processor_id());
1264 print_cpu_stall_info_end();
1265 for_each_possible_cpu(cpu
)
1266 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1267 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1268 jiffies
- rsp
->gp_start
,
1269 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1271 rcu_check_gp_kthread_starvation(rsp
);
1273 rcu_dump_cpu_stacks(rsp
);
1275 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1276 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1277 WRITE_ONCE(rsp
->jiffies_stall
,
1278 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1279 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1282 * Attempt to revive the RCU machinery by forcing a context switch.
1284 * A context switch would normally allow the RCU state machine to make
1285 * progress and it could be we're stuck in kernel space without context
1286 * switches for an entirely unreasonable amount of time.
1288 resched_cpu(smp_processor_id());
1291 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1293 unsigned long completed
;
1294 unsigned long gpnum
;
1298 struct rcu_node
*rnp
;
1300 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1305 * Lots of memory barriers to reject false positives.
1307 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1308 * then rsp->gp_start, and finally rsp->completed. These values
1309 * are updated in the opposite order with memory barriers (or
1310 * equivalent) during grace-period initialization and cleanup.
1311 * Now, a false positive can occur if we get an new value of
1312 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1313 * the memory barriers, the only way that this can happen is if one
1314 * grace period ends and another starts between these two fetches.
1315 * Detect this by comparing rsp->completed with the previous fetch
1318 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1319 * and rsp->gp_start suffice to forestall false positives.
1321 gpnum
= READ_ONCE(rsp
->gpnum
);
1322 smp_rmb(); /* Pick up ->gpnum first... */
1323 js
= READ_ONCE(rsp
->jiffies_stall
);
1324 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1325 gps
= READ_ONCE(rsp
->gp_start
);
1326 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1327 completed
= READ_ONCE(rsp
->completed
);
1328 if (ULONG_CMP_GE(completed
, gpnum
) ||
1329 ULONG_CMP_LT(j
, js
) ||
1330 ULONG_CMP_GE(gps
, js
))
1331 return; /* No stall or GP completed since entering function. */
1333 if (rcu_gp_in_progress(rsp
) &&
1334 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1336 /* We haven't checked in, so go dump stack. */
1337 print_cpu_stall(rsp
);
1339 } else if (rcu_gp_in_progress(rsp
) &&
1340 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1342 /* They had a few time units to dump stack, so complain. */
1343 print_other_cpu_stall(rsp
, gpnum
);
1348 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1350 * Set the stall-warning timeout way off into the future, thus preventing
1351 * any RCU CPU stall-warning messages from appearing in the current set of
1352 * RCU grace periods.
1354 * The caller must disable hard irqs.
1356 void rcu_cpu_stall_reset(void)
1358 struct rcu_state
*rsp
;
1360 for_each_rcu_flavor(rsp
)
1361 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1365 * Initialize the specified rcu_data structure's default callback list
1366 * to empty. The default callback list is the one that is not used by
1367 * no-callbacks CPUs.
1369 static void init_default_callback_list(struct rcu_data
*rdp
)
1373 rdp
->nxtlist
= NULL
;
1374 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1375 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1379 * Initialize the specified rcu_data structure's callback list to empty.
1381 static void init_callback_list(struct rcu_data
*rdp
)
1383 if (init_nocb_callback_list(rdp
))
1385 init_default_callback_list(rdp
);
1389 * Determine the value that ->completed will have at the end of the
1390 * next subsequent grace period. This is used to tag callbacks so that
1391 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1392 * been dyntick-idle for an extended period with callbacks under the
1393 * influence of RCU_FAST_NO_HZ.
1395 * The caller must hold rnp->lock with interrupts disabled.
1397 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1398 struct rcu_node
*rnp
)
1401 * If RCU is idle, we just wait for the next grace period.
1402 * But we can only be sure that RCU is idle if we are looking
1403 * at the root rcu_node structure -- otherwise, a new grace
1404 * period might have started, but just not yet gotten around
1405 * to initializing the current non-root rcu_node structure.
1407 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1408 return rnp
->completed
+ 1;
1411 * Otherwise, wait for a possible partial grace period and
1412 * then the subsequent full grace period.
1414 return rnp
->completed
+ 2;
1418 * Trace-event helper function for rcu_start_future_gp() and
1419 * rcu_nocb_wait_gp().
1421 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1422 unsigned long c
, const char *s
)
1424 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1425 rnp
->completed
, c
, rnp
->level
,
1426 rnp
->grplo
, rnp
->grphi
, s
);
1430 * Start some future grace period, as needed to handle newly arrived
1431 * callbacks. The required future grace periods are recorded in each
1432 * rcu_node structure's ->need_future_gp field. Returns true if there
1433 * is reason to awaken the grace-period kthread.
1435 * The caller must hold the specified rcu_node structure's ->lock.
1437 static bool __maybe_unused
1438 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1439 unsigned long *c_out
)
1444 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1447 * Pick up grace-period number for new callbacks. If this
1448 * grace period is already marked as needed, return to the caller.
1450 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1451 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1452 if (rnp
->need_future_gp
[c
& 0x1]) {
1453 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1458 * If either this rcu_node structure or the root rcu_node structure
1459 * believe that a grace period is in progress, then we must wait
1460 * for the one following, which is in "c". Because our request
1461 * will be noticed at the end of the current grace period, we don't
1462 * need to explicitly start one. We only do the lockless check
1463 * of rnp_root's fields if the current rcu_node structure thinks
1464 * there is no grace period in flight, and because we hold rnp->lock,
1465 * the only possible change is when rnp_root's two fields are
1466 * equal, in which case rnp_root->gpnum might be concurrently
1467 * incremented. But that is OK, as it will just result in our
1468 * doing some extra useless work.
1470 if (rnp
->gpnum
!= rnp
->completed
||
1471 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1472 rnp
->need_future_gp
[c
& 0x1]++;
1473 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1478 * There might be no grace period in progress. If we don't already
1479 * hold it, acquire the root rcu_node structure's lock in order to
1480 * start one (if needed).
1482 if (rnp
!= rnp_root
) {
1483 raw_spin_lock(&rnp_root
->lock
);
1484 smp_mb__after_unlock_lock();
1488 * Get a new grace-period number. If there really is no grace
1489 * period in progress, it will be smaller than the one we obtained
1490 * earlier. Adjust callbacks as needed. Note that even no-CBs
1491 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1493 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1494 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1495 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1496 rdp
->nxtcompleted
[i
] = c
;
1499 * If the needed for the required grace period is already
1500 * recorded, trace and leave.
1502 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1503 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1507 /* Record the need for the future grace period. */
1508 rnp_root
->need_future_gp
[c
& 0x1]++;
1510 /* If a grace period is not already in progress, start one. */
1511 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1512 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1514 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1515 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1518 if (rnp
!= rnp_root
)
1519 raw_spin_unlock(&rnp_root
->lock
);
1527 * Clean up any old requests for the just-ended grace period. Also return
1528 * whether any additional grace periods have been requested. Also invoke
1529 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1530 * waiting for this grace period to complete.
1532 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1534 int c
= rnp
->completed
;
1536 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1538 rcu_nocb_gp_cleanup(rsp
, rnp
);
1539 rnp
->need_future_gp
[c
& 0x1] = 0;
1540 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1541 trace_rcu_future_gp(rnp
, rdp
, c
,
1542 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1547 * Awaken the grace-period kthread for the specified flavor of RCU.
1548 * Don't do a self-awaken, and don't bother awakening when there is
1549 * nothing for the grace-period kthread to do (as in several CPUs
1550 * raced to awaken, and we lost), and finally don't try to awaken
1551 * a kthread that has not yet been created.
1553 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1555 if (current
== rsp
->gp_kthread
||
1556 !READ_ONCE(rsp
->gp_flags
) ||
1559 wake_up(&rsp
->gp_wq
);
1563 * If there is room, assign a ->completed number to any callbacks on
1564 * this CPU that have not already been assigned. Also accelerate any
1565 * callbacks that were previously assigned a ->completed number that has
1566 * since proven to be too conservative, which can happen if callbacks get
1567 * assigned a ->completed number while RCU is idle, but with reference to
1568 * a non-root rcu_node structure. This function is idempotent, so it does
1569 * not hurt to call it repeatedly. Returns an flag saying that we should
1570 * awaken the RCU grace-period kthread.
1572 * The caller must hold rnp->lock with interrupts disabled.
1574 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1575 struct rcu_data
*rdp
)
1581 /* If the CPU has no callbacks, nothing to do. */
1582 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1586 * Starting from the sublist containing the callbacks most
1587 * recently assigned a ->completed number and working down, find the
1588 * first sublist that is not assignable to an upcoming grace period.
1589 * Such a sublist has something in it (first two tests) and has
1590 * a ->completed number assigned that will complete sooner than
1591 * the ->completed number for newly arrived callbacks (last test).
1593 * The key point is that any later sublist can be assigned the
1594 * same ->completed number as the newly arrived callbacks, which
1595 * means that the callbacks in any of these later sublist can be
1596 * grouped into a single sublist, whether or not they have already
1597 * been assigned a ->completed number.
1599 c
= rcu_cbs_completed(rsp
, rnp
);
1600 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1601 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1602 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1606 * If there are no sublist for unassigned callbacks, leave.
1607 * At the same time, advance "i" one sublist, so that "i" will
1608 * index into the sublist where all the remaining callbacks should
1611 if (++i
>= RCU_NEXT_TAIL
)
1615 * Assign all subsequent callbacks' ->completed number to the next
1616 * full grace period and group them all in the sublist initially
1619 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1620 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1621 rdp
->nxtcompleted
[i
] = c
;
1623 /* Record any needed additional grace periods. */
1624 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1626 /* Trace depending on how much we were able to accelerate. */
1627 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1628 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1630 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1635 * Move any callbacks whose grace period has completed to the
1636 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1637 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1638 * sublist. This function is idempotent, so it does not hurt to
1639 * invoke it repeatedly. As long as it is not invoked -too- often...
1640 * Returns true if the RCU grace-period kthread needs to be awakened.
1642 * The caller must hold rnp->lock with interrupts disabled.
1644 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1645 struct rcu_data
*rdp
)
1649 /* If the CPU has no callbacks, nothing to do. */
1650 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1654 * Find all callbacks whose ->completed numbers indicate that they
1655 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1657 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1658 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1660 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1662 /* Clean up any sublist tail pointers that were misordered above. */
1663 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1664 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1666 /* Copy down callbacks to fill in empty sublists. */
1667 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1668 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1670 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1671 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1674 /* Classify any remaining callbacks. */
1675 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1679 * Update CPU-local rcu_data state to record the beginnings and ends of
1680 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1681 * structure corresponding to the current CPU, and must have irqs disabled.
1682 * Returns true if the grace-period kthread needs to be awakened.
1684 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1685 struct rcu_data
*rdp
)
1689 /* Handle the ends of any preceding grace periods first. */
1690 if (rdp
->completed
== rnp
->completed
&&
1691 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1693 /* No grace period end, so just accelerate recent callbacks. */
1694 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1698 /* Advance callbacks. */
1699 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1701 /* Remember that we saw this grace-period completion. */
1702 rdp
->completed
= rnp
->completed
;
1703 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1706 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1708 * If the current grace period is waiting for this CPU,
1709 * set up to detect a quiescent state, otherwise don't
1710 * go looking for one.
1712 rdp
->gpnum
= rnp
->gpnum
;
1713 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1714 rdp
->passed_quiesce
= 0;
1715 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1716 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1717 zero_cpu_stall_ticks(rdp
);
1718 WRITE_ONCE(rdp
->gpwrap
, false);
1723 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1725 unsigned long flags
;
1727 struct rcu_node
*rnp
;
1729 local_irq_save(flags
);
1731 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1732 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1733 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1734 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1735 local_irq_restore(flags
);
1738 smp_mb__after_unlock_lock();
1739 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1740 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1742 rcu_gp_kthread_wake(rsp
);
1746 * Initialize a new grace period. Return 0 if no grace period required.
1748 static int rcu_gp_init(struct rcu_state
*rsp
)
1750 unsigned long oldmask
;
1751 struct rcu_data
*rdp
;
1752 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1754 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1755 raw_spin_lock_irq(&rnp
->lock
);
1756 smp_mb__after_unlock_lock();
1757 if (!READ_ONCE(rsp
->gp_flags
)) {
1758 /* Spurious wakeup, tell caller to go back to sleep. */
1759 raw_spin_unlock_irq(&rnp
->lock
);
1762 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1764 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1766 * Grace period already in progress, don't start another.
1767 * Not supposed to be able to happen.
1769 raw_spin_unlock_irq(&rnp
->lock
);
1773 /* Advance to a new grace period and initialize state. */
1774 record_gp_stall_check_time(rsp
);
1775 /* Record GP times before starting GP, hence smp_store_release(). */
1776 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1777 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1778 raw_spin_unlock_irq(&rnp
->lock
);
1781 * Apply per-leaf buffered online and offline operations to the
1782 * rcu_node tree. Note that this new grace period need not wait
1783 * for subsequent online CPUs, and that quiescent-state forcing
1784 * will handle subsequent offline CPUs.
1786 rcu_for_each_leaf_node(rsp
, rnp
) {
1787 raw_spin_lock_irq(&rnp
->lock
);
1788 smp_mb__after_unlock_lock();
1789 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1790 !rnp
->wait_blkd_tasks
) {
1791 /* Nothing to do on this leaf rcu_node structure. */
1792 raw_spin_unlock_irq(&rnp
->lock
);
1796 /* Record old state, apply changes to ->qsmaskinit field. */
1797 oldmask
= rnp
->qsmaskinit
;
1798 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1800 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1801 if (!oldmask
!= !rnp
->qsmaskinit
) {
1802 if (!oldmask
) /* First online CPU for this rcu_node. */
1803 rcu_init_new_rnp(rnp
);
1804 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1805 rnp
->wait_blkd_tasks
= true;
1806 else /* Last offline CPU and can propagate. */
1807 rcu_cleanup_dead_rnp(rnp
);
1811 * If all waited-on tasks from prior grace period are
1812 * done, and if all this rcu_node structure's CPUs are
1813 * still offline, propagate up the rcu_node tree and
1814 * clear ->wait_blkd_tasks. Otherwise, if one of this
1815 * rcu_node structure's CPUs has since come back online,
1816 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1817 * checks for this, so just call it unconditionally).
1819 if (rnp
->wait_blkd_tasks
&&
1820 (!rcu_preempt_has_tasks(rnp
) ||
1822 rnp
->wait_blkd_tasks
= false;
1823 rcu_cleanup_dead_rnp(rnp
);
1826 raw_spin_unlock_irq(&rnp
->lock
);
1830 * Set the quiescent-state-needed bits in all the rcu_node
1831 * structures for all currently online CPUs in breadth-first order,
1832 * starting from the root rcu_node structure, relying on the layout
1833 * of the tree within the rsp->node[] array. Note that other CPUs
1834 * will access only the leaves of the hierarchy, thus seeing that no
1835 * grace period is in progress, at least until the corresponding
1836 * leaf node has been initialized. In addition, we have excluded
1837 * CPU-hotplug operations.
1839 * The grace period cannot complete until the initialization
1840 * process finishes, because this kthread handles both.
1842 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1843 raw_spin_lock_irq(&rnp
->lock
);
1844 smp_mb__after_unlock_lock();
1845 rdp
= this_cpu_ptr(rsp
->rda
);
1846 rcu_preempt_check_blocked_tasks(rnp
);
1847 rnp
->qsmask
= rnp
->qsmaskinit
;
1848 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
1849 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1850 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
1851 if (rnp
== rdp
->mynode
)
1852 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1853 rcu_preempt_boost_start_gp(rnp
);
1854 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1855 rnp
->level
, rnp
->grplo
,
1856 rnp
->grphi
, rnp
->qsmask
);
1857 raw_spin_unlock_irq(&rnp
->lock
);
1858 cond_resched_rcu_qs();
1859 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1860 if (gp_init_delay
> 0 &&
1862 (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* gp_init_delay
)))
1863 schedule_timeout_uninterruptible(gp_init_delay
);
1870 * Do one round of quiescent-state forcing.
1872 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1874 int fqs_state
= fqs_state_in
;
1875 bool isidle
= false;
1877 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1879 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1881 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1882 /* Collect dyntick-idle snapshots. */
1883 if (is_sysidle_rcu_state(rsp
)) {
1885 maxj
= jiffies
- ULONG_MAX
/ 4;
1887 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1889 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1890 fqs_state
= RCU_FORCE_QS
;
1892 /* Handle dyntick-idle and offline CPUs. */
1894 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1896 /* Clear flag to prevent immediate re-entry. */
1897 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1898 raw_spin_lock_irq(&rnp
->lock
);
1899 smp_mb__after_unlock_lock();
1900 WRITE_ONCE(rsp
->gp_flags
,
1901 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
1902 raw_spin_unlock_irq(&rnp
->lock
);
1908 * Clean up after the old grace period.
1910 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1912 unsigned long gp_duration
;
1913 bool needgp
= false;
1915 struct rcu_data
*rdp
;
1916 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1918 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1919 raw_spin_lock_irq(&rnp
->lock
);
1920 smp_mb__after_unlock_lock();
1921 gp_duration
= jiffies
- rsp
->gp_start
;
1922 if (gp_duration
> rsp
->gp_max
)
1923 rsp
->gp_max
= gp_duration
;
1926 * We know the grace period is complete, but to everyone else
1927 * it appears to still be ongoing. But it is also the case
1928 * that to everyone else it looks like there is nothing that
1929 * they can do to advance the grace period. It is therefore
1930 * safe for us to drop the lock in order to mark the grace
1931 * period as completed in all of the rcu_node structures.
1933 raw_spin_unlock_irq(&rnp
->lock
);
1936 * Propagate new ->completed value to rcu_node structures so
1937 * that other CPUs don't have to wait until the start of the next
1938 * grace period to process their callbacks. This also avoids
1939 * some nasty RCU grace-period initialization races by forcing
1940 * the end of the current grace period to be completely recorded in
1941 * all of the rcu_node structures before the beginning of the next
1942 * grace period is recorded in any of the rcu_node structures.
1944 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1945 raw_spin_lock_irq(&rnp
->lock
);
1946 smp_mb__after_unlock_lock();
1947 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
1948 WARN_ON_ONCE(rnp
->qsmask
);
1949 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
1950 rdp
= this_cpu_ptr(rsp
->rda
);
1951 if (rnp
== rdp
->mynode
)
1952 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1953 /* smp_mb() provided by prior unlock-lock pair. */
1954 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1955 raw_spin_unlock_irq(&rnp
->lock
);
1956 cond_resched_rcu_qs();
1957 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1959 rnp
= rcu_get_root(rsp
);
1960 raw_spin_lock_irq(&rnp
->lock
);
1961 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1962 rcu_nocb_gp_set(rnp
, nocb
);
1964 /* Declare grace period done. */
1965 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
1966 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1967 rsp
->fqs_state
= RCU_GP_IDLE
;
1968 rdp
= this_cpu_ptr(rsp
->rda
);
1969 /* Advance CBs to reduce false positives below. */
1970 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1971 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1972 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
1973 trace_rcu_grace_period(rsp
->name
,
1974 READ_ONCE(rsp
->gpnum
),
1977 raw_spin_unlock_irq(&rnp
->lock
);
1981 * Body of kthread that handles grace periods.
1983 static int __noreturn
rcu_gp_kthread(void *arg
)
1989 struct rcu_state
*rsp
= arg
;
1990 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1992 rcu_bind_gp_kthread();
1995 /* Handle grace-period start. */
1997 trace_rcu_grace_period(rsp
->name
,
1998 READ_ONCE(rsp
->gpnum
),
2000 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
2001 wait_event_interruptible(rsp
->gp_wq
,
2002 READ_ONCE(rsp
->gp_flags
) &
2004 /* Locking provides needed memory barrier. */
2005 if (rcu_gp_init(rsp
))
2007 cond_resched_rcu_qs();
2008 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2009 WARN_ON(signal_pending(current
));
2010 trace_rcu_grace_period(rsp
->name
,
2011 READ_ONCE(rsp
->gpnum
),
2015 /* Handle quiescent-state forcing. */
2016 fqs_state
= RCU_SAVE_DYNTICK
;
2017 j
= jiffies_till_first_fqs
;
2020 jiffies_till_first_fqs
= HZ
;
2025 rsp
->jiffies_force_qs
= jiffies
+ j
;
2026 trace_rcu_grace_period(rsp
->name
,
2027 READ_ONCE(rsp
->gpnum
),
2029 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2030 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
2031 ((gf
= READ_ONCE(rsp
->gp_flags
)) &
2033 (!READ_ONCE(rnp
->qsmask
) &&
2034 !rcu_preempt_blocked_readers_cgp(rnp
)),
2036 /* Locking provides needed memory barriers. */
2037 /* If grace period done, leave loop. */
2038 if (!READ_ONCE(rnp
->qsmask
) &&
2039 !rcu_preempt_blocked_readers_cgp(rnp
))
2041 /* If time for quiescent-state forcing, do it. */
2042 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2043 (gf
& RCU_GP_FLAG_FQS
)) {
2044 trace_rcu_grace_period(rsp
->name
,
2045 READ_ONCE(rsp
->gpnum
),
2047 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
2048 trace_rcu_grace_period(rsp
->name
,
2049 READ_ONCE(rsp
->gpnum
),
2051 cond_resched_rcu_qs();
2052 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2054 /* Deal with stray signal. */
2055 cond_resched_rcu_qs();
2056 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2057 WARN_ON(signal_pending(current
));
2058 trace_rcu_grace_period(rsp
->name
,
2059 READ_ONCE(rsp
->gpnum
),
2062 j
= jiffies_till_next_fqs
;
2065 jiffies_till_next_fqs
= HZ
;
2068 jiffies_till_next_fqs
= 1;
2072 /* Handle grace-period end. */
2073 rcu_gp_cleanup(rsp
);
2078 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2079 * in preparation for detecting the next grace period. The caller must hold
2080 * the root node's ->lock and hard irqs must be disabled.
2082 * Note that it is legal for a dying CPU (which is marked as offline) to
2083 * invoke this function. This can happen when the dying CPU reports its
2086 * Returns true if the grace-period kthread must be awakened.
2089 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2090 struct rcu_data
*rdp
)
2092 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2094 * Either we have not yet spawned the grace-period
2095 * task, this CPU does not need another grace period,
2096 * or a grace period is already in progress.
2097 * Either way, don't start a new grace period.
2101 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2102 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2106 * We can't do wakeups while holding the rnp->lock, as that
2107 * could cause possible deadlocks with the rq->lock. Defer
2108 * the wakeup to our caller.
2114 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2115 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2116 * is invoked indirectly from rcu_advance_cbs(), which would result in
2117 * endless recursion -- or would do so if it wasn't for the self-deadlock
2118 * that is encountered beforehand.
2120 * Returns true if the grace-period kthread needs to be awakened.
2122 static bool rcu_start_gp(struct rcu_state
*rsp
)
2124 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2125 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2129 * If there is no grace period in progress right now, any
2130 * callbacks we have up to this point will be satisfied by the
2131 * next grace period. Also, advancing the callbacks reduces the
2132 * probability of false positives from cpu_needs_another_gp()
2133 * resulting in pointless grace periods. So, advance callbacks
2134 * then start the grace period!
2136 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2137 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2142 * Report a full set of quiescent states to the specified rcu_state
2143 * data structure. This involves cleaning up after the prior grace
2144 * period and letting rcu_start_gp() start up the next grace period
2145 * if one is needed. Note that the caller must hold rnp->lock, which
2146 * is released before return.
2148 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2149 __releases(rcu_get_root(rsp
)->lock
)
2151 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2152 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2153 rcu_gp_kthread_wake(rsp
);
2157 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2158 * Allows quiescent states for a group of CPUs to be reported at one go
2159 * to the specified rcu_node structure, though all the CPUs in the group
2160 * must be represented by the same rcu_node structure (which need not be a
2161 * leaf rcu_node structure, though it often will be). The gps parameter
2162 * is the grace-period snapshot, which means that the quiescent states
2163 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2164 * must be held upon entry, and it is released before return.
2167 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2168 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2169 __releases(rnp
->lock
)
2171 unsigned long oldmask
= 0;
2172 struct rcu_node
*rnp_c
;
2174 /* Walk up the rcu_node hierarchy. */
2176 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2179 * Our bit has already been cleared, or the
2180 * relevant grace period is already over, so done.
2182 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2185 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2186 rnp
->qsmask
&= ~mask
;
2187 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2188 mask
, rnp
->qsmask
, rnp
->level
,
2189 rnp
->grplo
, rnp
->grphi
,
2191 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2193 /* Other bits still set at this level, so done. */
2194 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2197 mask
= rnp
->grpmask
;
2198 if (rnp
->parent
== NULL
) {
2200 /* No more levels. Exit loop holding root lock. */
2204 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2207 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2208 smp_mb__after_unlock_lock();
2209 oldmask
= rnp_c
->qsmask
;
2213 * Get here if we are the last CPU to pass through a quiescent
2214 * state for this grace period. Invoke rcu_report_qs_rsp()
2215 * to clean up and start the next grace period if one is needed.
2217 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2221 * Record a quiescent state for all tasks that were previously queued
2222 * on the specified rcu_node structure and that were blocking the current
2223 * RCU grace period. The caller must hold the specified rnp->lock with
2224 * irqs disabled, and this lock is released upon return, but irqs remain
2227 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2228 struct rcu_node
*rnp
, unsigned long flags
)
2229 __releases(rnp
->lock
)
2233 struct rcu_node
*rnp_p
;
2235 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2236 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2237 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2238 return; /* Still need more quiescent states! */
2241 rnp_p
= rnp
->parent
;
2242 if (rnp_p
== NULL
) {
2244 * Only one rcu_node structure in the tree, so don't
2245 * try to report up to its nonexistent parent!
2247 rcu_report_qs_rsp(rsp
, flags
);
2251 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2253 mask
= rnp
->grpmask
;
2254 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2255 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
2256 smp_mb__after_unlock_lock();
2257 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2261 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2262 * structure. This must be either called from the specified CPU, or
2263 * called when the specified CPU is known to be offline (and when it is
2264 * also known that no other CPU is concurrently trying to help the offline
2265 * CPU). The lastcomp argument is used to make sure we are still in the
2266 * grace period of interest. We don't want to end the current grace period
2267 * based on quiescent states detected in an earlier grace period!
2270 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2272 unsigned long flags
;
2275 struct rcu_node
*rnp
;
2278 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2279 smp_mb__after_unlock_lock();
2280 if ((rdp
->passed_quiesce
== 0 &&
2281 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2282 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2286 * The grace period in which this quiescent state was
2287 * recorded has ended, so don't report it upwards.
2288 * We will instead need a new quiescent state that lies
2289 * within the current grace period.
2291 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2292 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2293 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2296 mask
= rdp
->grpmask
;
2297 if ((rnp
->qsmask
& mask
) == 0) {
2298 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2300 rdp
->qs_pending
= 0;
2303 * This GP can't end until cpu checks in, so all of our
2304 * callbacks can be processed during the next GP.
2306 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2308 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2309 /* ^^^ Released rnp->lock */
2311 rcu_gp_kthread_wake(rsp
);
2316 * Check to see if there is a new grace period of which this CPU
2317 * is not yet aware, and if so, set up local rcu_data state for it.
2318 * Otherwise, see if this CPU has just passed through its first
2319 * quiescent state for this grace period, and record that fact if so.
2322 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2324 /* Check for grace-period ends and beginnings. */
2325 note_gp_changes(rsp
, rdp
);
2328 * Does this CPU still need to do its part for current grace period?
2329 * If no, return and let the other CPUs do their part as well.
2331 if (!rdp
->qs_pending
)
2335 * Was there a quiescent state since the beginning of the grace
2336 * period? If no, then exit and wait for the next call.
2338 if (!rdp
->passed_quiesce
&&
2339 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2343 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2346 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2349 #ifdef CONFIG_HOTPLUG_CPU
2352 * Send the specified CPU's RCU callbacks to the orphanage. The
2353 * specified CPU must be offline, and the caller must hold the
2357 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2358 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2360 /* No-CBs CPUs do not have orphanable callbacks. */
2361 if (rcu_is_nocb_cpu(rdp
->cpu
))
2365 * Orphan the callbacks. First adjust the counts. This is safe
2366 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2367 * cannot be running now. Thus no memory barrier is required.
2369 if (rdp
->nxtlist
!= NULL
) {
2370 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2371 rsp
->qlen
+= rdp
->qlen
;
2372 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2374 WRITE_ONCE(rdp
->qlen
, 0);
2378 * Next, move those callbacks still needing a grace period to
2379 * the orphanage, where some other CPU will pick them up.
2380 * Some of the callbacks might have gone partway through a grace
2381 * period, but that is too bad. They get to start over because we
2382 * cannot assume that grace periods are synchronized across CPUs.
2383 * We don't bother updating the ->nxttail[] array yet, instead
2384 * we just reset the whole thing later on.
2386 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2387 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2388 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2389 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2393 * Then move the ready-to-invoke callbacks to the orphanage,
2394 * where some other CPU will pick them up. These will not be
2395 * required to pass though another grace period: They are done.
2397 if (rdp
->nxtlist
!= NULL
) {
2398 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2399 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2403 * Finally, initialize the rcu_data structure's list to empty and
2404 * disallow further callbacks on this CPU.
2406 init_callback_list(rdp
);
2407 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2411 * Adopt the RCU callbacks from the specified rcu_state structure's
2412 * orphanage. The caller must hold the ->orphan_lock.
2414 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2417 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2419 /* No-CBs CPUs are handled specially. */
2420 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2423 /* Do the accounting first. */
2424 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2425 rdp
->qlen
+= rsp
->qlen
;
2426 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2427 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2428 rcu_idle_count_callbacks_posted();
2433 * We do not need a memory barrier here because the only way we
2434 * can get here if there is an rcu_barrier() in flight is if
2435 * we are the task doing the rcu_barrier().
2438 /* First adopt the ready-to-invoke callbacks. */
2439 if (rsp
->orphan_donelist
!= NULL
) {
2440 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2441 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2442 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2443 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2444 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2445 rsp
->orphan_donelist
= NULL
;
2446 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2449 /* And then adopt the callbacks that still need a grace period. */
2450 if (rsp
->orphan_nxtlist
!= NULL
) {
2451 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2452 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2453 rsp
->orphan_nxtlist
= NULL
;
2454 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2459 * Trace the fact that this CPU is going offline.
2461 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2463 RCU_TRACE(unsigned long mask
);
2464 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2465 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2467 RCU_TRACE(mask
= rdp
->grpmask
);
2468 trace_rcu_grace_period(rsp
->name
,
2469 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2474 * All CPUs for the specified rcu_node structure have gone offline,
2475 * and all tasks that were preempted within an RCU read-side critical
2476 * section while running on one of those CPUs have since exited their RCU
2477 * read-side critical section. Some other CPU is reporting this fact with
2478 * the specified rcu_node structure's ->lock held and interrupts disabled.
2479 * This function therefore goes up the tree of rcu_node structures,
2480 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2481 * the leaf rcu_node structure's ->qsmaskinit field has already been
2484 * This function does check that the specified rcu_node structure has
2485 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2486 * prematurely. That said, invoking it after the fact will cost you
2487 * a needless lock acquisition. So once it has done its work, don't
2490 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2493 struct rcu_node
*rnp
= rnp_leaf
;
2495 if (rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2498 mask
= rnp
->grpmask
;
2502 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2503 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2504 rnp
->qsmaskinit
&= ~mask
;
2505 rnp
->qsmask
&= ~mask
;
2506 if (rnp
->qsmaskinit
) {
2507 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2510 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2515 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2516 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2519 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2521 unsigned long flags
;
2523 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2524 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2526 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2527 mask
= rdp
->grpmask
;
2528 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2529 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2530 rnp
->qsmaskinitnext
&= ~mask
;
2531 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2535 * The CPU has been completely removed, and some other CPU is reporting
2536 * this fact from process context. Do the remainder of the cleanup,
2537 * including orphaning the outgoing CPU's RCU callbacks, and also
2538 * adopting them. There can only be one CPU hotplug operation at a time,
2539 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2541 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2543 unsigned long flags
;
2544 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2545 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2547 /* Adjust any no-longer-needed kthreads. */
2548 rcu_boost_kthread_setaffinity(rnp
, -1);
2550 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2551 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2552 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2553 rcu_adopt_orphan_cbs(rsp
, flags
);
2554 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2556 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2557 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2558 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2561 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2563 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2567 static void __maybe_unused
rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2571 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2575 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2579 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2582 * Invoke any RCU callbacks that have made it to the end of their grace
2583 * period. Thottle as specified by rdp->blimit.
2585 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2587 unsigned long flags
;
2588 struct rcu_head
*next
, *list
, **tail
;
2589 long bl
, count
, count_lazy
;
2592 /* If no callbacks are ready, just return. */
2593 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2594 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2595 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2596 need_resched(), is_idle_task(current
),
2597 rcu_is_callbacks_kthread());
2602 * Extract the list of ready callbacks, disabling to prevent
2603 * races with call_rcu() from interrupt handlers.
2605 local_irq_save(flags
);
2606 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2608 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2609 list
= rdp
->nxtlist
;
2610 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2611 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2612 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2613 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2614 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2615 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2616 local_irq_restore(flags
);
2618 /* Invoke callbacks. */
2619 count
= count_lazy
= 0;
2623 debug_rcu_head_unqueue(list
);
2624 if (__rcu_reclaim(rsp
->name
, list
))
2627 /* Stop only if limit reached and CPU has something to do. */
2628 if (++count
>= bl
&&
2630 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2634 local_irq_save(flags
);
2635 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2636 is_idle_task(current
),
2637 rcu_is_callbacks_kthread());
2639 /* Update count, and requeue any remaining callbacks. */
2641 *tail
= rdp
->nxtlist
;
2642 rdp
->nxtlist
= list
;
2643 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2644 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2645 rdp
->nxttail
[i
] = tail
;
2649 smp_mb(); /* List handling before counting for rcu_barrier(). */
2650 rdp
->qlen_lazy
-= count_lazy
;
2651 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2652 rdp
->n_cbs_invoked
+= count
;
2654 /* Reinstate batch limit if we have worked down the excess. */
2655 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2656 rdp
->blimit
= blimit
;
2658 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2659 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2660 rdp
->qlen_last_fqs_check
= 0;
2661 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2662 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2663 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2664 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2666 local_irq_restore(flags
);
2668 /* Re-invoke RCU core processing if there are callbacks remaining. */
2669 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2674 * Check to see if this CPU is in a non-context-switch quiescent state
2675 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2676 * Also schedule RCU core processing.
2678 * This function must be called from hardirq context. It is normally
2679 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2680 * false, there is no point in invoking rcu_check_callbacks().
2682 void rcu_check_callbacks(int user
)
2684 trace_rcu_utilization(TPS("Start scheduler-tick"));
2685 increment_cpu_stall_ticks();
2686 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2689 * Get here if this CPU took its interrupt from user
2690 * mode or from the idle loop, and if this is not a
2691 * nested interrupt. In this case, the CPU is in
2692 * a quiescent state, so note it.
2694 * No memory barrier is required here because both
2695 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2696 * variables that other CPUs neither access nor modify,
2697 * at least not while the corresponding CPU is online.
2703 } else if (!in_softirq()) {
2706 * Get here if this CPU did not take its interrupt from
2707 * softirq, in other words, if it is not interrupting
2708 * a rcu_bh read-side critical section. This is an _bh
2709 * critical section, so note it.
2714 rcu_preempt_check_callbacks();
2718 rcu_note_voluntary_context_switch(current
);
2719 trace_rcu_utilization(TPS("End scheduler-tick"));
2723 * Scan the leaf rcu_node structures, processing dyntick state for any that
2724 * have not yet encountered a quiescent state, using the function specified.
2725 * Also initiate boosting for any threads blocked on the root rcu_node.
2727 * The caller must have suppressed start of new grace periods.
2729 static void force_qs_rnp(struct rcu_state
*rsp
,
2730 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2731 unsigned long *maxj
),
2732 bool *isidle
, unsigned long *maxj
)
2736 unsigned long flags
;
2738 struct rcu_node
*rnp
;
2740 rcu_for_each_leaf_node(rsp
, rnp
) {
2741 cond_resched_rcu_qs();
2743 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2744 smp_mb__after_unlock_lock();
2745 if (!rcu_gp_in_progress(rsp
)) {
2746 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2749 if (rnp
->qsmask
== 0) {
2750 if (rcu_state_p
== &rcu_sched_state
||
2751 rsp
!= rcu_state_p
||
2752 rcu_preempt_blocked_readers_cgp(rnp
)) {
2754 * No point in scanning bits because they
2755 * are all zero. But we might need to
2756 * priority-boost blocked readers.
2758 rcu_initiate_boost(rnp
, flags
);
2759 /* rcu_initiate_boost() releases rnp->lock */
2763 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2765 * Race between grace-period
2766 * initialization and task exiting RCU
2767 * read-side critical section: Report.
2769 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2770 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2776 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2777 if ((rnp
->qsmask
& bit
) != 0) {
2778 if ((rnp
->qsmaskinit
& bit
) == 0)
2779 *isidle
= false; /* Pending hotplug. */
2780 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2785 /* Idle/offline CPUs, report (releases rnp->lock. */
2786 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2788 /* Nothing to do here, so just drop the lock. */
2789 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2795 * Force quiescent states on reluctant CPUs, and also detect which
2796 * CPUs are in dyntick-idle mode.
2798 static void force_quiescent_state(struct rcu_state
*rsp
)
2800 unsigned long flags
;
2802 struct rcu_node
*rnp
;
2803 struct rcu_node
*rnp_old
= NULL
;
2805 /* Funnel through hierarchy to reduce memory contention. */
2806 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2807 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2808 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2809 !raw_spin_trylock(&rnp
->fqslock
);
2810 if (rnp_old
!= NULL
)
2811 raw_spin_unlock(&rnp_old
->fqslock
);
2813 rsp
->n_force_qs_lh
++;
2818 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2820 /* Reached the root of the rcu_node tree, acquire lock. */
2821 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2822 smp_mb__after_unlock_lock();
2823 raw_spin_unlock(&rnp_old
->fqslock
);
2824 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2825 rsp
->n_force_qs_lh
++;
2826 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2827 return; /* Someone beat us to it. */
2829 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2830 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2831 rcu_gp_kthread_wake(rsp
);
2835 * This does the RCU core processing work for the specified rcu_state
2836 * and rcu_data structures. This may be called only from the CPU to
2837 * whom the rdp belongs.
2840 __rcu_process_callbacks(struct rcu_state
*rsp
)
2842 unsigned long flags
;
2844 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2846 WARN_ON_ONCE(rdp
->beenonline
== 0);
2848 /* Update RCU state based on any recent quiescent states. */
2849 rcu_check_quiescent_state(rsp
, rdp
);
2851 /* Does this CPU require a not-yet-started grace period? */
2852 local_irq_save(flags
);
2853 if (cpu_needs_another_gp(rsp
, rdp
)) {
2854 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2855 needwake
= rcu_start_gp(rsp
);
2856 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2858 rcu_gp_kthread_wake(rsp
);
2860 local_irq_restore(flags
);
2863 /* If there are callbacks ready, invoke them. */
2864 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2865 invoke_rcu_callbacks(rsp
, rdp
);
2867 /* Do any needed deferred wakeups of rcuo kthreads. */
2868 do_nocb_deferred_wakeup(rdp
);
2872 * Do RCU core processing for the current CPU.
2874 static void rcu_process_callbacks(struct softirq_action
*unused
)
2876 struct rcu_state
*rsp
;
2878 if (cpu_is_offline(smp_processor_id()))
2880 trace_rcu_utilization(TPS("Start RCU core"));
2881 for_each_rcu_flavor(rsp
)
2882 __rcu_process_callbacks(rsp
);
2883 trace_rcu_utilization(TPS("End RCU core"));
2887 * Schedule RCU callback invocation. If the specified type of RCU
2888 * does not support RCU priority boosting, just do a direct call,
2889 * otherwise wake up the per-CPU kernel kthread. Note that because we
2890 * are running on the current CPU with softirqs disabled, the
2891 * rcu_cpu_kthread_task cannot disappear out from under us.
2893 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2895 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
2897 if (likely(!rsp
->boost
)) {
2898 rcu_do_batch(rsp
, rdp
);
2901 invoke_rcu_callbacks_kthread();
2904 static void invoke_rcu_core(void)
2906 if (cpu_online(smp_processor_id()))
2907 raise_softirq(RCU_SOFTIRQ
);
2911 * Handle any core-RCU processing required by a call_rcu() invocation.
2913 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2914 struct rcu_head
*head
, unsigned long flags
)
2919 * If called from an extended quiescent state, invoke the RCU
2920 * core in order to force a re-evaluation of RCU's idleness.
2922 if (!rcu_is_watching())
2925 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2926 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2930 * Force the grace period if too many callbacks or too long waiting.
2931 * Enforce hysteresis, and don't invoke force_quiescent_state()
2932 * if some other CPU has recently done so. Also, don't bother
2933 * invoking force_quiescent_state() if the newly enqueued callback
2934 * is the only one waiting for a grace period to complete.
2936 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2938 /* Are we ignoring a completed grace period? */
2939 note_gp_changes(rsp
, rdp
);
2941 /* Start a new grace period if one not already started. */
2942 if (!rcu_gp_in_progress(rsp
)) {
2943 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2945 raw_spin_lock(&rnp_root
->lock
);
2946 smp_mb__after_unlock_lock();
2947 needwake
= rcu_start_gp(rsp
);
2948 raw_spin_unlock(&rnp_root
->lock
);
2950 rcu_gp_kthread_wake(rsp
);
2952 /* Give the grace period a kick. */
2953 rdp
->blimit
= LONG_MAX
;
2954 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2955 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2956 force_quiescent_state(rsp
);
2957 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2958 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2964 * RCU callback function to leak a callback.
2966 static void rcu_leak_callback(struct rcu_head
*rhp
)
2971 * Helper function for call_rcu() and friends. The cpu argument will
2972 * normally be -1, indicating "currently running CPU". It may specify
2973 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2974 * is expected to specify a CPU.
2977 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2978 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2980 unsigned long flags
;
2981 struct rcu_data
*rdp
;
2983 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
2984 if (debug_rcu_head_queue(head
)) {
2985 /* Probable double call_rcu(), so leak the callback. */
2986 WRITE_ONCE(head
->func
, rcu_leak_callback
);
2987 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2994 * Opportunistically note grace-period endings and beginnings.
2995 * Note that we might see a beginning right after we see an
2996 * end, but never vice versa, since this CPU has to pass through
2997 * a quiescent state betweentimes.
2999 local_irq_save(flags
);
3000 rdp
= this_cpu_ptr(rsp
->rda
);
3002 /* Add the callback to our list. */
3003 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
3007 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3008 if (likely(rdp
->mynode
)) {
3009 /* Post-boot, so this should be for a no-CBs CPU. */
3010 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3011 WARN_ON_ONCE(offline
);
3012 /* Offline CPU, _call_rcu() illegal, leak callback. */
3013 local_irq_restore(flags
);
3017 * Very early boot, before rcu_init(). Initialize if needed
3018 * and then drop through to queue the callback.
3021 WARN_ON_ONCE(!rcu_is_watching());
3022 if (!likely(rdp
->nxtlist
))
3023 init_default_callback_list(rdp
);
3025 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3029 rcu_idle_count_callbacks_posted();
3030 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3031 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3032 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3034 if (__is_kfree_rcu_offset((unsigned long)func
))
3035 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3036 rdp
->qlen_lazy
, rdp
->qlen
);
3038 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3040 /* Go handle any RCU core processing required. */
3041 __call_rcu_core(rsp
, rdp
, head
, flags
);
3042 local_irq_restore(flags
);
3046 * Queue an RCU-sched callback for invocation after a grace period.
3048 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3050 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3052 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3055 * Queue an RCU callback for invocation after a quicker grace period.
3057 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3059 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3061 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3064 * Queue an RCU callback for lazy invocation after a grace period.
3065 * This will likely be later named something like "call_rcu_lazy()",
3066 * but this change will require some way of tagging the lazy RCU
3067 * callbacks in the list of pending callbacks. Until then, this
3068 * function may only be called from __kfree_rcu().
3070 void kfree_call_rcu(struct rcu_head
*head
,
3071 void (*func
)(struct rcu_head
*rcu
))
3073 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3075 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3078 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3079 * any blocking grace-period wait automatically implies a grace period
3080 * if there is only one CPU online at any point time during execution
3081 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3082 * occasionally incorrectly indicate that there are multiple CPUs online
3083 * when there was in fact only one the whole time, as this just adds
3084 * some overhead: RCU still operates correctly.
3086 static inline int rcu_blocking_is_gp(void)
3090 might_sleep(); /* Check for RCU read-side critical section. */
3092 ret
= num_online_cpus() <= 1;
3098 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3100 * Control will return to the caller some time after a full rcu-sched
3101 * grace period has elapsed, in other words after all currently executing
3102 * rcu-sched read-side critical sections have completed. These read-side
3103 * critical sections are delimited by rcu_read_lock_sched() and
3104 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3105 * local_irq_disable(), and so on may be used in place of
3106 * rcu_read_lock_sched().
3108 * This means that all preempt_disable code sequences, including NMI and
3109 * non-threaded hardware-interrupt handlers, in progress on entry will
3110 * have completed before this primitive returns. However, this does not
3111 * guarantee that softirq handlers will have completed, since in some
3112 * kernels, these handlers can run in process context, and can block.
3114 * Note that this guarantee implies further memory-ordering guarantees.
3115 * On systems with more than one CPU, when synchronize_sched() returns,
3116 * each CPU is guaranteed to have executed a full memory barrier since the
3117 * end of its last RCU-sched read-side critical section whose beginning
3118 * preceded the call to synchronize_sched(). In addition, each CPU having
3119 * an RCU read-side critical section that extends beyond the return from
3120 * synchronize_sched() is guaranteed to have executed a full memory barrier
3121 * after the beginning of synchronize_sched() and before the beginning of
3122 * that RCU read-side critical section. Note that these guarantees include
3123 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3124 * that are executing in the kernel.
3126 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3127 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3128 * to have executed a full memory barrier during the execution of
3129 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3130 * again only if the system has more than one CPU).
3132 * This primitive provides the guarantees made by the (now removed)
3133 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3134 * guarantees that rcu_read_lock() sections will have completed.
3135 * In "classic RCU", these two guarantees happen to be one and
3136 * the same, but can differ in realtime RCU implementations.
3138 void synchronize_sched(void)
3140 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3141 !lock_is_held(&rcu_lock_map
) &&
3142 !lock_is_held(&rcu_sched_lock_map
),
3143 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3144 if (rcu_blocking_is_gp())
3146 if (rcu_gp_is_expedited())
3147 synchronize_sched_expedited();
3149 wait_rcu_gp(call_rcu_sched
);
3151 EXPORT_SYMBOL_GPL(synchronize_sched
);
3154 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3156 * Control will return to the caller some time after a full rcu_bh grace
3157 * period has elapsed, in other words after all currently executing rcu_bh
3158 * read-side critical sections have completed. RCU read-side critical
3159 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3160 * and may be nested.
3162 * See the description of synchronize_sched() for more detailed information
3163 * on memory ordering guarantees.
3165 void synchronize_rcu_bh(void)
3167 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3168 !lock_is_held(&rcu_lock_map
) &&
3169 !lock_is_held(&rcu_sched_lock_map
),
3170 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3171 if (rcu_blocking_is_gp())
3173 if (rcu_gp_is_expedited())
3174 synchronize_rcu_bh_expedited();
3176 wait_rcu_gp(call_rcu_bh
);
3178 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3181 * get_state_synchronize_rcu - Snapshot current RCU state
3183 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3184 * to determine whether or not a full grace period has elapsed in the
3187 unsigned long get_state_synchronize_rcu(void)
3190 * Any prior manipulation of RCU-protected data must happen
3191 * before the load from ->gpnum.
3196 * Make sure this load happens before the purportedly
3197 * time-consuming work between get_state_synchronize_rcu()
3198 * and cond_synchronize_rcu().
3200 return smp_load_acquire(&rcu_state_p
->gpnum
);
3202 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3205 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3207 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3209 * If a full RCU grace period has elapsed since the earlier call to
3210 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3211 * synchronize_rcu() to wait for a full grace period.
3213 * Yes, this function does not take counter wrap into account. But
3214 * counter wrap is harmless. If the counter wraps, we have waited for
3215 * more than 2 billion grace periods (and way more on a 64-bit system!),
3216 * so waiting for one additional grace period should be just fine.
3218 void cond_synchronize_rcu(unsigned long oldstate
)
3220 unsigned long newstate
;
3223 * Ensure that this load happens before any RCU-destructive
3224 * actions the caller might carry out after we return.
3226 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3227 if (ULONG_CMP_GE(oldstate
, newstate
))
3230 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3232 static int synchronize_sched_expedited_cpu_stop(void *data
)
3235 * There must be a full memory barrier on each affected CPU
3236 * between the time that try_stop_cpus() is called and the
3237 * time that it returns.
3239 * In the current initial implementation of cpu_stop, the
3240 * above condition is already met when the control reaches
3241 * this point and the following smp_mb() is not strictly
3242 * necessary. Do smp_mb() anyway for documentation and
3243 * robustness against future implementation changes.
3245 smp_mb(); /* See above comment block. */
3250 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3252 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3253 * approach to force the grace period to end quickly. This consumes
3254 * significant time on all CPUs and is unfriendly to real-time workloads,
3255 * so is thus not recommended for any sort of common-case code. In fact,
3256 * if you are using synchronize_sched_expedited() in a loop, please
3257 * restructure your code to batch your updates, and then use a single
3258 * synchronize_sched() instead.
3260 * This implementation can be thought of as an application of ticket
3261 * locking to RCU, with sync_sched_expedited_started and
3262 * sync_sched_expedited_done taking on the roles of the halves
3263 * of the ticket-lock word. Each task atomically increments
3264 * sync_sched_expedited_started upon entry, snapshotting the old value,
3265 * then attempts to stop all the CPUs. If this succeeds, then each
3266 * CPU will have executed a context switch, resulting in an RCU-sched
3267 * grace period. We are then done, so we use atomic_cmpxchg() to
3268 * update sync_sched_expedited_done to match our snapshot -- but
3269 * only if someone else has not already advanced past our snapshot.
3271 * On the other hand, if try_stop_cpus() fails, we check the value
3272 * of sync_sched_expedited_done. If it has advanced past our
3273 * initial snapshot, then someone else must have forced a grace period
3274 * some time after we took our snapshot. In this case, our work is
3275 * done for us, and we can simply return. Otherwise, we try again,
3276 * but keep our initial snapshot for purposes of checking for someone
3277 * doing our work for us.
3279 * If we fail too many times in a row, we fall back to synchronize_sched().
3281 void synchronize_sched_expedited(void)
3286 long firstsnap
, s
, snap
;
3288 struct rcu_state
*rsp
= &rcu_sched_state
;
3291 * If we are in danger of counter wrap, just do synchronize_sched().
3292 * By allowing sync_sched_expedited_started to advance no more than
3293 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3294 * that more than 3.5 billion CPUs would be required to force a
3295 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3296 * course be required on a 64-bit system.
3298 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3299 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3301 synchronize_sched();
3302 atomic_long_inc(&rsp
->expedited_wrap
);
3307 * Take a ticket. Note that atomic_inc_return() implies a
3308 * full memory barrier.
3310 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3312 if (!try_get_online_cpus()) {
3313 /* CPU hotplug operation in flight, fall back to normal GP. */
3314 wait_rcu_gp(call_rcu_sched
);
3315 atomic_long_inc(&rsp
->expedited_normal
);
3318 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3320 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3321 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3323 cpumask_copy(cm
, cpu_online_mask
);
3324 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3325 for_each_cpu(cpu
, cm
) {
3326 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3328 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3329 cpumask_clear_cpu(cpu
, cm
);
3331 if (cpumask_weight(cm
) == 0)
3336 * Each pass through the following loop attempts to force a
3337 * context switch on each CPU.
3339 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3340 synchronize_sched_expedited_cpu_stop
,
3343 atomic_long_inc(&rsp
->expedited_tryfail
);
3345 /* Check to see if someone else did our work for us. */
3346 s
= atomic_long_read(&rsp
->expedited_done
);
3347 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3348 /* ensure test happens before caller kfree */
3349 smp_mb__before_atomic(); /* ^^^ */
3350 atomic_long_inc(&rsp
->expedited_workdone1
);
3351 free_cpumask_var(cm
);
3355 /* No joy, try again later. Or just synchronize_sched(). */
3356 if (trycount
++ < 10) {
3357 udelay(trycount
* num_online_cpus());
3359 wait_rcu_gp(call_rcu_sched
);
3360 atomic_long_inc(&rsp
->expedited_normal
);
3361 free_cpumask_var(cm
);
3365 /* Recheck to see if someone else did our work for us. */
3366 s
= atomic_long_read(&rsp
->expedited_done
);
3367 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3368 /* ensure test happens before caller kfree */
3369 smp_mb__before_atomic(); /* ^^^ */
3370 atomic_long_inc(&rsp
->expedited_workdone2
);
3371 free_cpumask_var(cm
);
3376 * Refetching sync_sched_expedited_started allows later
3377 * callers to piggyback on our grace period. We retry
3378 * after they started, so our grace period works for them,
3379 * and they started after our first try, so their grace
3380 * period works for us.
3382 if (!try_get_online_cpus()) {
3383 /* CPU hotplug operation in flight, use normal GP. */
3384 wait_rcu_gp(call_rcu_sched
);
3385 atomic_long_inc(&rsp
->expedited_normal
);
3386 free_cpumask_var(cm
);
3389 snap
= atomic_long_read(&rsp
->expedited_start
);
3390 smp_mb(); /* ensure read is before try_stop_cpus(). */
3392 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3395 free_cpumask_var(cm
);
3398 * Everyone up to our most recent fetch is covered by our grace
3399 * period. Update the counter, but only if our work is still
3400 * relevant -- which it won't be if someone who started later
3401 * than we did already did their update.
3404 atomic_long_inc(&rsp
->expedited_done_tries
);
3405 s
= atomic_long_read(&rsp
->expedited_done
);
3406 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3407 /* ensure test happens before caller kfree */
3408 smp_mb__before_atomic(); /* ^^^ */
3409 atomic_long_inc(&rsp
->expedited_done_lost
);
3412 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3413 atomic_long_inc(&rsp
->expedited_done_exit
);
3417 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3420 * Check to see if there is any immediate RCU-related work to be done
3421 * by the current CPU, for the specified type of RCU, returning 1 if so.
3422 * The checks are in order of increasing expense: checks that can be
3423 * carried out against CPU-local state are performed first. However,
3424 * we must check for CPU stalls first, else we might not get a chance.
3426 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3428 struct rcu_node
*rnp
= rdp
->mynode
;
3430 rdp
->n_rcu_pending
++;
3432 /* Check for CPU stalls, if enabled. */
3433 check_cpu_stall(rsp
, rdp
);
3435 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3436 if (rcu_nohz_full_cpu(rsp
))
3439 /* Is the RCU core waiting for a quiescent state from this CPU? */
3440 if (rcu_scheduler_fully_active
&&
3441 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3442 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3443 rdp
->n_rp_qs_pending
++;
3444 } else if (rdp
->qs_pending
&&
3445 (rdp
->passed_quiesce
||
3446 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3447 rdp
->n_rp_report_qs
++;
3451 /* Does this CPU have callbacks ready to invoke? */
3452 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3453 rdp
->n_rp_cb_ready
++;
3457 /* Has RCU gone idle with this CPU needing another grace period? */
3458 if (cpu_needs_another_gp(rsp
, rdp
)) {
3459 rdp
->n_rp_cpu_needs_gp
++;
3463 /* Has another RCU grace period completed? */
3464 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3465 rdp
->n_rp_gp_completed
++;
3469 /* Has a new RCU grace period started? */
3470 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3471 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3472 rdp
->n_rp_gp_started
++;
3476 /* Does this CPU need a deferred NOCB wakeup? */
3477 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3478 rdp
->n_rp_nocb_defer_wakeup
++;
3483 rdp
->n_rp_need_nothing
++;
3488 * Check to see if there is any immediate RCU-related work to be done
3489 * by the current CPU, returning 1 if so. This function is part of the
3490 * RCU implementation; it is -not- an exported member of the RCU API.
3492 static int rcu_pending(void)
3494 struct rcu_state
*rsp
;
3496 for_each_rcu_flavor(rsp
)
3497 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3503 * Return true if the specified CPU has any callback. If all_lazy is
3504 * non-NULL, store an indication of whether all callbacks are lazy.
3505 * (If there are no callbacks, all of them are deemed to be lazy.)
3507 static int __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3511 struct rcu_data
*rdp
;
3512 struct rcu_state
*rsp
;
3514 for_each_rcu_flavor(rsp
) {
3515 rdp
= this_cpu_ptr(rsp
->rda
);
3519 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3530 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3531 * the compiler is expected to optimize this away.
3533 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3534 int cpu
, unsigned long done
)
3536 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3537 atomic_read(&rsp
->barrier_cpu_count
), done
);
3541 * RCU callback function for _rcu_barrier(). If we are last, wake
3542 * up the task executing _rcu_barrier().
3544 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3546 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3547 struct rcu_state
*rsp
= rdp
->rsp
;
3549 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3550 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3551 complete(&rsp
->barrier_completion
);
3553 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3558 * Called with preemption disabled, and from cross-cpu IRQ context.
3560 static void rcu_barrier_func(void *type
)
3562 struct rcu_state
*rsp
= type
;
3563 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3565 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3566 atomic_inc(&rsp
->barrier_cpu_count
);
3567 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3571 * Orchestrate the specified type of RCU barrier, waiting for all
3572 * RCU callbacks of the specified type to complete.
3574 static void _rcu_barrier(struct rcu_state
*rsp
)
3577 struct rcu_data
*rdp
;
3578 unsigned long snap
= READ_ONCE(rsp
->n_barrier_done
);
3579 unsigned long snap_done
;
3581 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3583 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3584 mutex_lock(&rsp
->barrier_mutex
);
3587 * Ensure that all prior references, including to ->n_barrier_done,
3588 * are ordered before the _rcu_barrier() machinery.
3590 smp_mb(); /* See above block comment. */
3593 * Recheck ->n_barrier_done to see if others did our work for us.
3594 * This means checking ->n_barrier_done for an even-to-odd-to-even
3595 * transition. The "if" expression below therefore rounds the old
3596 * value up to the next even number and adds two before comparing.
3598 snap_done
= rsp
->n_barrier_done
;
3599 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3602 * If the value in snap is odd, we needed to wait for the current
3603 * rcu_barrier() to complete, then wait for the next one, in other
3604 * words, we need the value of snap_done to be three larger than
3605 * the value of snap. On the other hand, if the value in snap is
3606 * even, we only had to wait for the next rcu_barrier() to complete,
3607 * in other words, we need the value of snap_done to be only two
3608 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3609 * this for us (thank you, Linus!).
3611 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3612 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3613 smp_mb(); /* caller's subsequent code after above check. */
3614 mutex_unlock(&rsp
->barrier_mutex
);
3619 * Increment ->n_barrier_done to avoid duplicate work. Use
3620 * WRITE_ONCE() to prevent the compiler from speculating
3621 * the increment to precede the early-exit check.
3623 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3624 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3625 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3626 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3629 * Initialize the count to one rather than to zero in order to
3630 * avoid a too-soon return to zero in case of a short grace period
3631 * (or preemption of this task). Exclude CPU-hotplug operations
3632 * to ensure that no offline CPU has callbacks queued.
3634 init_completion(&rsp
->barrier_completion
);
3635 atomic_set(&rsp
->barrier_cpu_count
, 1);
3639 * Force each CPU with callbacks to register a new callback.
3640 * When that callback is invoked, we will know that all of the
3641 * corresponding CPU's preceding callbacks have been invoked.
3643 for_each_possible_cpu(cpu
) {
3644 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3646 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3647 if (rcu_is_nocb_cpu(cpu
)) {
3648 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3649 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3650 rsp
->n_barrier_done
);
3652 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3653 rsp
->n_barrier_done
);
3654 smp_mb__before_atomic();
3655 atomic_inc(&rsp
->barrier_cpu_count
);
3656 __call_rcu(&rdp
->barrier_head
,
3657 rcu_barrier_callback
, rsp
, cpu
, 0);
3659 } else if (READ_ONCE(rdp
->qlen
)) {
3660 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3661 rsp
->n_barrier_done
);
3662 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3664 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3665 rsp
->n_barrier_done
);
3671 * Now that we have an rcu_barrier_callback() callback on each
3672 * CPU, and thus each counted, remove the initial count.
3674 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3675 complete(&rsp
->barrier_completion
);
3677 /* Increment ->n_barrier_done to prevent duplicate work. */
3678 smp_mb(); /* Keep increment after above mechanism. */
3679 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3680 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3681 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3682 smp_mb(); /* Keep increment before caller's subsequent code. */
3684 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3685 wait_for_completion(&rsp
->barrier_completion
);
3687 /* Other rcu_barrier() invocations can now safely proceed. */
3688 mutex_unlock(&rsp
->barrier_mutex
);
3692 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3694 void rcu_barrier_bh(void)
3696 _rcu_barrier(&rcu_bh_state
);
3698 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3701 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3703 void rcu_barrier_sched(void)
3705 _rcu_barrier(&rcu_sched_state
);
3707 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3710 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3711 * first CPU in a given leaf rcu_node structure coming online. The caller
3712 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3715 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
3718 struct rcu_node
*rnp
= rnp_leaf
;
3721 mask
= rnp
->grpmask
;
3725 raw_spin_lock(&rnp
->lock
); /* Interrupts already disabled. */
3726 rnp
->qsmaskinit
|= mask
;
3727 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
3732 * Do boot-time initialization of a CPU's per-CPU RCU data.
3735 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3737 unsigned long flags
;
3738 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3739 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3741 /* Set up local state, ensuring consistent view of global state. */
3742 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3743 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3744 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3745 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3746 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3749 rcu_boot_init_nocb_percpu_data(rdp
);
3750 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3754 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3755 * offline event can be happening at a given time. Note also that we
3756 * can accept some slop in the rsp->completed access due to the fact
3757 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3760 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3762 unsigned long flags
;
3764 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3765 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3767 /* Set up local state, ensuring consistent view of global state. */
3768 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3769 rdp
->beenonline
= 1; /* We have now been online. */
3770 rdp
->qlen_last_fqs_check
= 0;
3771 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3772 rdp
->blimit
= blimit
;
3774 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3775 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3776 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3777 atomic_set(&rdp
->dynticks
->dynticks
,
3778 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3779 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3782 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3783 * propagation up the rcu_node tree will happen at the beginning
3784 * of the next grace period.
3787 mask
= rdp
->grpmask
;
3788 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3789 smp_mb__after_unlock_lock();
3790 rnp
->qsmaskinitnext
|= mask
;
3791 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
3792 rdp
->completed
= rnp
->completed
;
3793 rdp
->passed_quiesce
= false;
3794 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
3795 rdp
->qs_pending
= false;
3796 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3797 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3800 static void rcu_prepare_cpu(int cpu
)
3802 struct rcu_state
*rsp
;
3804 for_each_rcu_flavor(rsp
)
3805 rcu_init_percpu_data(cpu
, rsp
);
3809 * Handle CPU online/offline notification events.
3811 int rcu_cpu_notify(struct notifier_block
*self
,
3812 unsigned long action
, void *hcpu
)
3814 long cpu
= (long)hcpu
;
3815 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3816 struct rcu_node
*rnp
= rdp
->mynode
;
3817 struct rcu_state
*rsp
;
3820 case CPU_UP_PREPARE
:
3821 case CPU_UP_PREPARE_FROZEN
:
3822 rcu_prepare_cpu(cpu
);
3823 rcu_prepare_kthreads(cpu
);
3824 rcu_spawn_all_nocb_kthreads(cpu
);
3827 case CPU_DOWN_FAILED
:
3828 rcu_boost_kthread_setaffinity(rnp
, -1);
3830 case CPU_DOWN_PREPARE
:
3831 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3834 case CPU_DYING_FROZEN
:
3835 for_each_rcu_flavor(rsp
)
3836 rcu_cleanup_dying_cpu(rsp
);
3838 case CPU_DYING_IDLE
:
3839 for_each_rcu_flavor(rsp
) {
3840 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
3844 case CPU_DEAD_FROZEN
:
3845 case CPU_UP_CANCELED
:
3846 case CPU_UP_CANCELED_FROZEN
:
3847 for_each_rcu_flavor(rsp
) {
3848 rcu_cleanup_dead_cpu(cpu
, rsp
);
3849 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3858 static int rcu_pm_notify(struct notifier_block
*self
,
3859 unsigned long action
, void *hcpu
)
3862 case PM_HIBERNATION_PREPARE
:
3863 case PM_SUSPEND_PREPARE
:
3864 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3867 case PM_POST_HIBERNATION
:
3868 case PM_POST_SUSPEND
:
3869 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3870 rcu_unexpedite_gp();
3879 * Spawn the kthreads that handle each RCU flavor's grace periods.
3881 static int __init
rcu_spawn_gp_kthread(void)
3883 unsigned long flags
;
3884 int kthread_prio_in
= kthread_prio
;
3885 struct rcu_node
*rnp
;
3886 struct rcu_state
*rsp
;
3887 struct sched_param sp
;
3888 struct task_struct
*t
;
3890 /* Force priority into range. */
3891 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3893 else if (kthread_prio
< 0)
3895 else if (kthread_prio
> 99)
3897 if (kthread_prio
!= kthread_prio_in
)
3898 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3899 kthread_prio
, kthread_prio_in
);
3901 rcu_scheduler_fully_active
= 1;
3902 for_each_rcu_flavor(rsp
) {
3903 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3905 rnp
= rcu_get_root(rsp
);
3906 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3907 rsp
->gp_kthread
= t
;
3909 sp
.sched_priority
= kthread_prio
;
3910 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3913 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3915 rcu_spawn_nocb_kthreads();
3916 rcu_spawn_boost_kthreads();
3919 early_initcall(rcu_spawn_gp_kthread
);
3922 * This function is invoked towards the end of the scheduler's initialization
3923 * process. Before this is called, the idle task might contain
3924 * RCU read-side critical sections (during which time, this idle
3925 * task is booting the system). After this function is called, the
3926 * idle tasks are prohibited from containing RCU read-side critical
3927 * sections. This function also enables RCU lockdep checking.
3929 void rcu_scheduler_starting(void)
3931 WARN_ON(num_online_cpus() != 1);
3932 WARN_ON(nr_context_switches() > 0);
3933 rcu_scheduler_active
= 1;
3937 * Compute the per-level fanout, either using the exact fanout specified
3938 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3940 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3944 if (IS_ENABLED(CONFIG_RCU_FANOUT_EXACT
)) {
3945 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3946 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3947 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3953 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3954 ccur
= rsp
->levelcnt
[i
];
3955 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3962 * Helper function for rcu_init() that initializes one rcu_state structure.
3964 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3965 struct rcu_data __percpu
*rda
)
3967 static const char * const buf
[] = {
3971 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3972 static const char * const fqs
[] = {
3976 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3977 static u8 fl_mask
= 0x1;
3981 struct rcu_node
*rnp
;
3983 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3985 /* Silence gcc 4.8 false positive about array index out of range. */
3986 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
3987 panic("rcu_init_one: rcu_num_lvls out of range");
3989 /* Initialize the level-tracking arrays. */
3991 for (i
= 0; i
< rcu_num_lvls
; i
++)
3992 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3993 for (i
= 1; i
< rcu_num_lvls
; i
++)
3994 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3995 rcu_init_levelspread(rsp
);
3996 rsp
->flavor_mask
= fl_mask
;
3999 /* Initialize the elements themselves, starting from the leaves. */
4001 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4002 cpustride
*= rsp
->levelspread
[i
];
4003 rnp
= rsp
->level
[i
];
4004 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
4005 raw_spin_lock_init(&rnp
->lock
);
4006 lockdep_set_class_and_name(&rnp
->lock
,
4007 &rcu_node_class
[i
], buf
[i
]);
4008 raw_spin_lock_init(&rnp
->fqslock
);
4009 lockdep_set_class_and_name(&rnp
->fqslock
,
4010 &rcu_fqs_class
[i
], fqs
[i
]);
4011 rnp
->gpnum
= rsp
->gpnum
;
4012 rnp
->completed
= rsp
->completed
;
4014 rnp
->qsmaskinit
= 0;
4015 rnp
->grplo
= j
* cpustride
;
4016 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4017 if (rnp
->grphi
>= nr_cpu_ids
)
4018 rnp
->grphi
= nr_cpu_ids
- 1;
4024 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
4025 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4026 rnp
->parent
= rsp
->level
[i
- 1] +
4027 j
/ rsp
->levelspread
[i
- 1];
4030 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4031 rcu_init_one_nocb(rnp
);
4035 init_waitqueue_head(&rsp
->gp_wq
);
4036 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4037 for_each_possible_cpu(i
) {
4038 while (i
> rnp
->grphi
)
4040 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4041 rcu_boot_init_percpu_data(i
, rsp
);
4043 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4047 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4048 * replace the definitions in tree.h because those are needed to size
4049 * the ->node array in the rcu_state structure.
4051 static void __init
rcu_init_geometry(void)
4057 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
4060 * Initialize any unspecified boot parameters.
4061 * The default values of jiffies_till_first_fqs and
4062 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4063 * value, which is a function of HZ, then adding one for each
4064 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4066 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4067 if (jiffies_till_first_fqs
== ULONG_MAX
)
4068 jiffies_till_first_fqs
= d
;
4069 if (jiffies_till_next_fqs
== ULONG_MAX
)
4070 jiffies_till_next_fqs
= d
;
4072 /* If the compile-time values are accurate, just leave. */
4073 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
4074 nr_cpu_ids
== NR_CPUS
)
4076 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4077 rcu_fanout_leaf
, nr_cpu_ids
);
4080 * Compute number of nodes that can be handled an rcu_node tree
4081 * with the given number of levels. Setting rcu_capacity[0] makes
4082 * some of the arithmetic easier.
4084 rcu_capacity
[0] = 1;
4085 rcu_capacity
[1] = rcu_fanout_leaf
;
4086 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
4087 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
4090 * The boot-time rcu_fanout_leaf parameter is only permitted
4091 * to increase the leaf-level fanout, not decrease it. Of course,
4092 * the leaf-level fanout cannot exceed the number of bits in
4093 * the rcu_node masks. Finally, the tree must be able to accommodate
4094 * the configured number of CPUs. Complain and fall back to the
4095 * compile-time values if these limits are exceeded.
4097 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
4098 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
4099 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
4104 /* Calculate the number of rcu_nodes at each level of the tree. */
4105 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
4106 if (n
<= rcu_capacity
[i
]) {
4107 for (j
= 0; j
<= i
; j
++)
4109 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
4111 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
4116 /* Calculate the total number of rcu_node structures. */
4118 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
4119 rcu_num_nodes
+= num_rcu_lvl
[i
];
4123 void __init
rcu_init(void)
4127 rcu_early_boot_tests();
4129 rcu_bootup_announce();
4130 rcu_init_geometry();
4131 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
4132 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
4133 __rcu_init_preempt();
4134 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4137 * We don't need protection against CPU-hotplug here because
4138 * this is called early in boot, before either interrupts
4139 * or the scheduler are operational.
4141 cpu_notifier(rcu_cpu_notify
, 0);
4142 pm_notifier(rcu_pm_notify
, 0);
4143 for_each_online_cpu(cpu
)
4144 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4147 #include "tree_plugin.h"