1 /* SPDX-License-Identifier: GPL-2.0+ */
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions that provide either classic
5 * or preemptible semantics.
7 * Copyright Red Hat, 2009
8 * Copyright IBM Corporation, 2009
10 * Author: Ingo Molnar <mingo@elte.hu>
11 * Paul E. McKenney <paulmck@linux.ibm.com>
14 #include "../locking/rtmutex_common.h"
16 #ifdef CONFIG_RCU_NOCB_CPU
17 static cpumask_var_t rcu_nocb_mask
; /* CPUs to have callbacks offloaded. */
18 static bool __read_mostly rcu_nocb_poll
; /* Offload kthread are to poll. */
19 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
22 * Check the RCU kernel configuration parameters and print informative
23 * messages about anything out of the ordinary.
25 static void __init
rcu_bootup_announce_oddness(void)
27 if (IS_ENABLED(CONFIG_RCU_TRACE
))
28 pr_info("\tRCU event tracing is enabled.\n");
29 if ((IS_ENABLED(CONFIG_64BIT
) && RCU_FANOUT
!= 64) ||
30 (!IS_ENABLED(CONFIG_64BIT
) && RCU_FANOUT
!= 32))
31 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
34 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
35 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ
))
36 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
37 if (IS_ENABLED(CONFIG_PROVE_RCU
))
38 pr_info("\tRCU lockdep checking is enabled.\n");
39 if (RCU_NUM_LVLS
>= 4)
40 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
41 if (RCU_FANOUT_LEAF
!= 16)
42 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
44 if (rcu_fanout_leaf
!= RCU_FANOUT_LEAF
)
45 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
47 if (nr_cpu_ids
!= NR_CPUS
)
48 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS
, nr_cpu_ids
);
49 #ifdef CONFIG_RCU_BOOST
50 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
51 kthread_prio
, CONFIG_RCU_BOOST_DELAY
);
53 if (blimit
!= DEFAULT_RCU_BLIMIT
)
54 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit
);
55 if (qhimark
!= DEFAULT_RCU_QHIMARK
)
56 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark
);
57 if (qlowmark
!= DEFAULT_RCU_QLOMARK
)
58 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark
);
59 if (qovld
!= DEFAULT_RCU_QOVLD
)
60 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld
);
61 if (jiffies_till_first_fqs
!= ULONG_MAX
)
62 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs
);
63 if (jiffies_till_next_fqs
!= ULONG_MAX
)
64 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs
);
65 if (jiffies_till_sched_qs
!= ULONG_MAX
)
66 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs
);
67 if (rcu_kick_kthreads
)
68 pr_info("\tKick kthreads if too-long grace period.\n");
69 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD
))
70 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
72 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay
);
74 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay
);
76 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay
);
78 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
79 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
))
80 pr_info("\tRCU debug extended QS entry/exit.\n");
81 rcupdate_announce_bootup_oddness();
84 #ifdef CONFIG_PREEMPT_RCU
86 static void rcu_report_exp_rnp(struct rcu_node
*rnp
, bool wake
);
87 static void rcu_read_unlock_special(struct task_struct
*t
);
90 * Tell them what RCU they are running.
92 static void __init
rcu_bootup_announce(void)
94 pr_info("Preemptible hierarchical RCU implementation.\n");
95 rcu_bootup_announce_oddness();
98 /* Flags for rcu_preempt_ctxt_queue() decision table. */
99 #define RCU_GP_TASKS 0x8
100 #define RCU_EXP_TASKS 0x4
101 #define RCU_GP_BLKD 0x2
102 #define RCU_EXP_BLKD 0x1
105 * Queues a task preempted within an RCU-preempt read-side critical
106 * section into the appropriate location within the ->blkd_tasks list,
107 * depending on the states of any ongoing normal and expedited grace
108 * periods. The ->gp_tasks pointer indicates which element the normal
109 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
110 * indicates which element the expedited grace period is waiting on (again,
111 * NULL if none). If a grace period is waiting on a given element in the
112 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
113 * adding a task to the tail of the list blocks any grace period that is
114 * already waiting on one of the elements. In contrast, adding a task
115 * to the head of the list won't block any grace period that is already
116 * waiting on one of the elements.
118 * This queuing is imprecise, and can sometimes make an ongoing grace
119 * period wait for a task that is not strictly speaking blocking it.
120 * Given the choice, we needlessly block a normal grace period rather than
121 * blocking an expedited grace period.
123 * Note that an endless sequence of expedited grace periods still cannot
124 * indefinitely postpone a normal grace period. Eventually, all of the
125 * fixed number of preempted tasks blocking the normal grace period that are
126 * not also blocking the expedited grace period will resume and complete
127 * their RCU read-side critical sections. At that point, the ->gp_tasks
128 * pointer will equal the ->exp_tasks pointer, at which point the end of
129 * the corresponding expedited grace period will also be the end of the
130 * normal grace period.
132 static void rcu_preempt_ctxt_queue(struct rcu_node
*rnp
, struct rcu_data
*rdp
)
133 __releases(rnp
->lock
) /* But leaves rrupts disabled. */
135 int blkd_state
= (rnp
->gp_tasks
? RCU_GP_TASKS
: 0) +
136 (rnp
->exp_tasks
? RCU_EXP_TASKS
: 0) +
137 (rnp
->qsmask
& rdp
->grpmask
? RCU_GP_BLKD
: 0) +
138 (rnp
->expmask
& rdp
->grpmask
? RCU_EXP_BLKD
: 0);
139 struct task_struct
*t
= current
;
141 raw_lockdep_assert_held_rcu_node(rnp
);
142 WARN_ON_ONCE(rdp
->mynode
!= rnp
);
143 WARN_ON_ONCE(!rcu_is_leaf_node(rnp
));
144 /* RCU better not be waiting on newly onlined CPUs! */
145 WARN_ON_ONCE(rnp
->qsmaskinitnext
& ~rnp
->qsmaskinit
& rnp
->qsmask
&
149 * Decide where to queue the newly blocked task. In theory,
150 * this could be an if-statement. In practice, when I tried
151 * that, it was quite messy.
153 switch (blkd_state
) {
156 case RCU_EXP_TASKS
+ RCU_GP_BLKD
:
158 case RCU_GP_TASKS
+ RCU_EXP_TASKS
:
161 * Blocking neither GP, or first task blocking the normal
162 * GP but not blocking the already-waiting expedited GP.
163 * Queue at the head of the list to avoid unnecessarily
164 * blocking the already-waiting GPs.
166 list_add(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
171 case RCU_GP_BLKD
+ RCU_EXP_BLKD
:
172 case RCU_GP_TASKS
+ RCU_EXP_BLKD
:
173 case RCU_GP_TASKS
+ RCU_GP_BLKD
+ RCU_EXP_BLKD
:
174 case RCU_GP_TASKS
+ RCU_EXP_TASKS
+ RCU_GP_BLKD
+ RCU_EXP_BLKD
:
177 * First task arriving that blocks either GP, or first task
178 * arriving that blocks the expedited GP (with the normal
179 * GP already waiting), or a task arriving that blocks
180 * both GPs with both GPs already waiting. Queue at the
181 * tail of the list to avoid any GP waiting on any of the
182 * already queued tasks that are not blocking it.
184 list_add_tail(&t
->rcu_node_entry
, &rnp
->blkd_tasks
);
187 case RCU_EXP_TASKS
+ RCU_EXP_BLKD
:
188 case RCU_EXP_TASKS
+ RCU_GP_BLKD
+ RCU_EXP_BLKD
:
189 case RCU_GP_TASKS
+ RCU_EXP_TASKS
+ RCU_EXP_BLKD
:
192 * Second or subsequent task blocking the expedited GP.
193 * The task either does not block the normal GP, or is the
194 * first task blocking the normal GP. Queue just after
195 * the first task blocking the expedited GP.
197 list_add(&t
->rcu_node_entry
, rnp
->exp_tasks
);
200 case RCU_GP_TASKS
+ RCU_GP_BLKD
:
201 case RCU_GP_TASKS
+ RCU_EXP_TASKS
+ RCU_GP_BLKD
:
204 * Second or subsequent task blocking the normal GP.
205 * The task does not block the expedited GP. Queue just
206 * after the first task blocking the normal GP.
208 list_add(&t
->rcu_node_entry
, rnp
->gp_tasks
);
213 /* Yet another exercise in excessive paranoia. */
219 * We have now queued the task. If it was the first one to
220 * block either grace period, update the ->gp_tasks and/or
221 * ->exp_tasks pointers, respectively, to reference the newly
224 if (!rnp
->gp_tasks
&& (blkd_state
& RCU_GP_BLKD
)) {
225 WRITE_ONCE(rnp
->gp_tasks
, &t
->rcu_node_entry
);
226 WARN_ON_ONCE(rnp
->completedqs
== rnp
->gp_seq
);
228 if (!rnp
->exp_tasks
&& (blkd_state
& RCU_EXP_BLKD
))
229 rnp
->exp_tasks
= &t
->rcu_node_entry
;
230 WARN_ON_ONCE(!(blkd_state
& RCU_GP_BLKD
) !=
231 !(rnp
->qsmask
& rdp
->grpmask
));
232 WARN_ON_ONCE(!(blkd_state
& RCU_EXP_BLKD
) !=
233 !(rnp
->expmask
& rdp
->grpmask
));
234 raw_spin_unlock_rcu_node(rnp
); /* interrupts remain disabled. */
237 * Report the quiescent state for the expedited GP. This expedited
238 * GP should not be able to end until we report, so there should be
239 * no need to check for a subsequent expedited GP. (Though we are
240 * still in a quiescent state in any case.)
242 if (blkd_state
& RCU_EXP_BLKD
&& rdp
->exp_deferred_qs
)
243 rcu_report_exp_rdp(rdp
);
245 WARN_ON_ONCE(rdp
->exp_deferred_qs
);
249 * Record a preemptible-RCU quiescent state for the specified CPU.
250 * Note that this does not necessarily mean that the task currently running
251 * on the CPU is in a quiescent state: Instead, it means that the current
252 * grace period need not wait on any RCU read-side critical section that
253 * starts later on this CPU. It also means that if the current task is
254 * in an RCU read-side critical section, it has already added itself to
255 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
256 * current task, there might be any number of other tasks blocked while
257 * in an RCU read-side critical section.
259 * Callers to this function must disable preemption.
261 static void rcu_qs(void)
263 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
264 if (__this_cpu_read(rcu_data
.cpu_no_qs
.s
)) {
265 trace_rcu_grace_period(TPS("rcu_preempt"),
266 __this_cpu_read(rcu_data
.gp_seq
),
268 __this_cpu_write(rcu_data
.cpu_no_qs
.b
.norm
, false);
269 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
270 WRITE_ONCE(current
->rcu_read_unlock_special
.b
.need_qs
, false);
275 * We have entered the scheduler, and the current task might soon be
276 * context-switched away from. If this task is in an RCU read-side
277 * critical section, we will no longer be able to rely on the CPU to
278 * record that fact, so we enqueue the task on the blkd_tasks list.
279 * The task will dequeue itself when it exits the outermost enclosing
280 * RCU read-side critical section. Therefore, the current grace period
281 * cannot be permitted to complete until the blkd_tasks list entries
282 * predating the current grace period drain, in other words, until
283 * rnp->gp_tasks becomes NULL.
285 * Caller must disable interrupts.
287 void rcu_note_context_switch(bool preempt
)
289 struct task_struct
*t
= current
;
290 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
291 struct rcu_node
*rnp
;
293 trace_rcu_utilization(TPS("Start context switch"));
294 lockdep_assert_irqs_disabled();
295 WARN_ON_ONCE(!preempt
&& rcu_preempt_depth() > 0);
296 if (rcu_preempt_depth() > 0 &&
297 !t
->rcu_read_unlock_special
.b
.blocked
) {
299 /* Possibly blocking in an RCU read-side critical section. */
301 raw_spin_lock_rcu_node(rnp
);
302 t
->rcu_read_unlock_special
.b
.blocked
= true;
303 t
->rcu_blocked_node
= rnp
;
306 * Verify the CPU's sanity, trace the preemption, and
307 * then queue the task as required based on the states
308 * of any ongoing and expedited grace periods.
310 WARN_ON_ONCE((rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) == 0);
311 WARN_ON_ONCE(!list_empty(&t
->rcu_node_entry
));
312 trace_rcu_preempt_task(rcu_state
.name
,
314 (rnp
->qsmask
& rdp
->grpmask
)
316 : rcu_seq_snap(&rnp
->gp_seq
));
317 rcu_preempt_ctxt_queue(rnp
, rdp
);
319 rcu_preempt_deferred_qs(t
);
323 * Either we were not in an RCU read-side critical section to
324 * begin with, or we have now recorded that critical section
325 * globally. Either way, we can now note a quiescent state
326 * for this CPU. Again, if we were in an RCU read-side critical
327 * section, and if that critical section was blocking the current
328 * grace period, then the fact that the task has been enqueued
329 * means that we continue to block the current grace period.
332 if (rdp
->exp_deferred_qs
)
333 rcu_report_exp_rdp(rdp
);
334 trace_rcu_utilization(TPS("End context switch"));
336 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
339 * Check for preempted RCU readers blocking the current grace period
340 * for the specified rcu_node structure. If the caller needs a reliable
341 * answer, it must hold the rcu_node's ->lock.
343 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
345 return READ_ONCE(rnp
->gp_tasks
) != NULL
;
348 /* Bias and limit values for ->rcu_read_lock_nesting. */
349 #define RCU_NEST_BIAS INT_MAX
350 #define RCU_NEST_NMAX (-INT_MAX / 2)
351 #define RCU_NEST_PMAX (INT_MAX / 2)
353 static void rcu_preempt_read_enter(void)
355 current
->rcu_read_lock_nesting
++;
358 static void rcu_preempt_read_exit(void)
360 current
->rcu_read_lock_nesting
--;
363 static void rcu_preempt_depth_set(int val
)
365 current
->rcu_read_lock_nesting
= val
;
369 * Preemptible RCU implementation for rcu_read_lock().
370 * Just increment ->rcu_read_lock_nesting, shared state will be updated
373 void __rcu_read_lock(void)
375 rcu_preempt_read_enter();
376 if (IS_ENABLED(CONFIG_PROVE_LOCKING
))
377 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX
);
378 barrier(); /* critical section after entry code. */
380 EXPORT_SYMBOL_GPL(__rcu_read_lock
);
383 * Preemptible RCU implementation for rcu_read_unlock().
384 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
385 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
386 * invoke rcu_read_unlock_special() to clean up after a context switch
387 * in an RCU read-side critical section and other special cases.
389 void __rcu_read_unlock(void)
391 struct task_struct
*t
= current
;
393 if (rcu_preempt_depth() != 1) {
394 rcu_preempt_read_exit();
396 barrier(); /* critical section before exit code. */
397 rcu_preempt_depth_set(-RCU_NEST_BIAS
);
398 barrier(); /* assign before ->rcu_read_unlock_special load */
399 if (unlikely(READ_ONCE(t
->rcu_read_unlock_special
.s
)))
400 rcu_read_unlock_special(t
);
401 barrier(); /* ->rcu_read_unlock_special load before assign */
402 rcu_preempt_depth_set(0);
404 if (IS_ENABLED(CONFIG_PROVE_LOCKING
)) {
405 int rrln
= rcu_preempt_depth();
407 WARN_ON_ONCE(rrln
< 0 && rrln
> RCU_NEST_NMAX
);
410 EXPORT_SYMBOL_GPL(__rcu_read_unlock
);
413 * Advance a ->blkd_tasks-list pointer to the next entry, instead
414 * returning NULL if at the end of the list.
416 static struct list_head
*rcu_next_node_entry(struct task_struct
*t
,
417 struct rcu_node
*rnp
)
419 struct list_head
*np
;
421 np
= t
->rcu_node_entry
.next
;
422 if (np
== &rnp
->blkd_tasks
)
428 * Return true if the specified rcu_node structure has tasks that were
429 * preempted within an RCU read-side critical section.
431 static bool rcu_preempt_has_tasks(struct rcu_node
*rnp
)
433 return !list_empty(&rnp
->blkd_tasks
);
437 * Report deferred quiescent states. The deferral time can
438 * be quite short, for example, in the case of the call from
439 * rcu_read_unlock_special().
442 rcu_preempt_deferred_qs_irqrestore(struct task_struct
*t
, unsigned long flags
)
447 struct list_head
*np
;
448 bool drop_boost_mutex
= false;
449 struct rcu_data
*rdp
;
450 struct rcu_node
*rnp
;
451 union rcu_special special
;
454 * If RCU core is waiting for this CPU to exit its critical section,
455 * report the fact that it has exited. Because irqs are disabled,
456 * t->rcu_read_unlock_special cannot change.
458 special
= t
->rcu_read_unlock_special
;
459 rdp
= this_cpu_ptr(&rcu_data
);
460 if (!special
.s
&& !rdp
->exp_deferred_qs
) {
461 local_irq_restore(flags
);
464 t
->rcu_read_unlock_special
.s
= 0;
465 if (special
.b
.need_qs
)
469 * Respond to a request by an expedited grace period for a
470 * quiescent state from this CPU. Note that requests from
471 * tasks are handled when removing the task from the
472 * blocked-tasks list below.
474 if (rdp
->exp_deferred_qs
)
475 rcu_report_exp_rdp(rdp
);
477 /* Clean up if blocked during RCU read-side critical section. */
478 if (special
.b
.blocked
) {
481 * Remove this task from the list it blocked on. The task
482 * now remains queued on the rcu_node corresponding to the
483 * CPU it first blocked on, so there is no longer any need
484 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
486 rnp
= t
->rcu_blocked_node
;
487 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
488 WARN_ON_ONCE(rnp
!= t
->rcu_blocked_node
);
489 WARN_ON_ONCE(!rcu_is_leaf_node(rnp
));
490 empty_norm
= !rcu_preempt_blocked_readers_cgp(rnp
);
491 WARN_ON_ONCE(rnp
->completedqs
== rnp
->gp_seq
&&
492 (!empty_norm
|| rnp
->qsmask
));
493 empty_exp
= sync_rcu_exp_done(rnp
);
494 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
495 np
= rcu_next_node_entry(t
, rnp
);
496 list_del_init(&t
->rcu_node_entry
);
497 t
->rcu_blocked_node
= NULL
;
498 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
499 rnp
->gp_seq
, t
->pid
);
500 if (&t
->rcu_node_entry
== rnp
->gp_tasks
)
501 WRITE_ONCE(rnp
->gp_tasks
, np
);
502 if (&t
->rcu_node_entry
== rnp
->exp_tasks
)
504 if (IS_ENABLED(CONFIG_RCU_BOOST
)) {
505 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
506 drop_boost_mutex
= rt_mutex_owner(&rnp
->boost_mtx
) == t
;
507 if (&t
->rcu_node_entry
== rnp
->boost_tasks
)
508 rnp
->boost_tasks
= np
;
512 * If this was the last task on the current list, and if
513 * we aren't waiting on any CPUs, report the quiescent state.
514 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
515 * so we must take a snapshot of the expedited state.
517 empty_exp_now
= sync_rcu_exp_done(rnp
);
518 if (!empty_norm
&& !rcu_preempt_blocked_readers_cgp(rnp
)) {
519 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
526 rcu_report_unblock_qs_rnp(rnp
, flags
);
528 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
531 /* Unboost if we were boosted. */
532 if (IS_ENABLED(CONFIG_RCU_BOOST
) && drop_boost_mutex
)
533 rt_mutex_futex_unlock(&rnp
->boost_mtx
);
536 * If this was the last task on the expedited lists,
537 * then we need to report up the rcu_node hierarchy.
539 if (!empty_exp
&& empty_exp_now
)
540 rcu_report_exp_rnp(rnp
, true);
542 local_irq_restore(flags
);
547 * Is a deferred quiescent-state pending, and are we also not in
548 * an RCU read-side critical section? It is the caller's responsibility
549 * to ensure it is otherwise safe to report any deferred quiescent
550 * states. The reason for this is that it is safe to report a
551 * quiescent state during context switch even though preemption
552 * is disabled. This function cannot be expected to understand these
553 * nuances, so the caller must handle them.
555 static bool rcu_preempt_need_deferred_qs(struct task_struct
*t
)
557 return (__this_cpu_read(rcu_data
.exp_deferred_qs
) ||
558 READ_ONCE(t
->rcu_read_unlock_special
.s
)) &&
559 rcu_preempt_depth() <= 0;
563 * Report a deferred quiescent state if needed and safe to do so.
564 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
565 * not being in an RCU read-side critical section. The caller must
566 * evaluate safety in terms of interrupt, softirq, and preemption
569 static void rcu_preempt_deferred_qs(struct task_struct
*t
)
573 if (!rcu_preempt_need_deferred_qs(t
))
575 local_irq_save(flags
);
576 rcu_preempt_deferred_qs_irqrestore(t
, flags
);
580 * Minimal handler to give the scheduler a chance to re-evaluate.
582 static void rcu_preempt_deferred_qs_handler(struct irq_work
*iwp
)
584 struct rcu_data
*rdp
;
586 rdp
= container_of(iwp
, struct rcu_data
, defer_qs_iw
);
587 rdp
->defer_qs_iw_pending
= false;
591 * Handle special cases during rcu_read_unlock(), such as needing to
592 * notify RCU core processing or task having blocked during the RCU
593 * read-side critical section.
595 static void rcu_read_unlock_special(struct task_struct
*t
)
598 bool preempt_bh_were_disabled
=
599 !!(preempt_count() & (PREEMPT_MASK
| SOFTIRQ_MASK
));
600 bool irqs_were_disabled
;
602 /* NMI handlers cannot block and cannot safely manipulate state. */
606 local_irq_save(flags
);
607 irqs_were_disabled
= irqs_disabled_flags(flags
);
608 if (preempt_bh_were_disabled
|| irqs_were_disabled
) {
610 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
611 struct rcu_node
*rnp
= rdp
->mynode
;
613 exp
= (t
->rcu_blocked_node
&&
614 READ_ONCE(t
->rcu_blocked_node
->exp_tasks
)) ||
615 (rdp
->grpmask
& READ_ONCE(rnp
->expmask
));
616 // Need to defer quiescent state until everything is enabled.
617 if (use_softirq
&& (in_irq() || (exp
&& !irqs_were_disabled
))) {
618 // Using softirq, safe to awaken, and either the
619 // wakeup is free or there is an expedited GP.
620 raise_softirq_irqoff(RCU_SOFTIRQ
);
622 // Enabling BH or preempt does reschedule, so...
623 // Also if no expediting, slow is OK.
624 // Plus nohz_full CPUs eventually get tick enabled.
625 set_tsk_need_resched(current
);
626 set_preempt_need_resched();
627 if (IS_ENABLED(CONFIG_IRQ_WORK
) && irqs_were_disabled
&&
628 !rdp
->defer_qs_iw_pending
&& exp
) {
629 // Get scheduler to re-evaluate and call hooks.
630 // If !IRQ_WORK, FQS scan will eventually IPI.
631 init_irq_work(&rdp
->defer_qs_iw
,
632 rcu_preempt_deferred_qs_handler
);
633 rdp
->defer_qs_iw_pending
= true;
634 irq_work_queue_on(&rdp
->defer_qs_iw
, rdp
->cpu
);
637 local_irq_restore(flags
);
640 rcu_preempt_deferred_qs_irqrestore(t
, flags
);
644 * Check that the list of blocked tasks for the newly completed grace
645 * period is in fact empty. It is a serious bug to complete a grace
646 * period that still has RCU readers blocked! This function must be
647 * invoked -before- updating this rnp's ->gp_seq.
649 * Also, if there are blocked tasks on the list, they automatically
650 * block the newly created grace period, so set up ->gp_tasks accordingly.
652 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
654 struct task_struct
*t
;
656 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
657 raw_lockdep_assert_held_rcu_node(rnp
);
658 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
)))
659 dump_blkd_tasks(rnp
, 10);
660 if (rcu_preempt_has_tasks(rnp
) &&
661 (rnp
->qsmaskinit
|| rnp
->wait_blkd_tasks
)) {
662 WRITE_ONCE(rnp
->gp_tasks
, rnp
->blkd_tasks
.next
);
663 t
= container_of(rnp
->gp_tasks
, struct task_struct
,
665 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
666 rnp
->gp_seq
, t
->pid
);
668 WARN_ON_ONCE(rnp
->qsmask
);
672 * Check for a quiescent state from the current CPU, including voluntary
673 * context switches for Tasks RCU. When a task blocks, the task is
674 * recorded in the corresponding CPU's rcu_node structure, which is checked
675 * elsewhere, hence this function need only check for quiescent states
676 * related to the current CPU, not to those related to tasks.
678 static void rcu_flavor_sched_clock_irq(int user
)
680 struct task_struct
*t
= current
;
682 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
683 rcu_note_voluntary_context_switch(current
);
685 if (rcu_preempt_depth() > 0 ||
686 (preempt_count() & (PREEMPT_MASK
| SOFTIRQ_MASK
))) {
687 /* No QS, force context switch if deferred. */
688 if (rcu_preempt_need_deferred_qs(t
)) {
689 set_tsk_need_resched(t
);
690 set_preempt_need_resched();
692 } else if (rcu_preempt_need_deferred_qs(t
)) {
693 rcu_preempt_deferred_qs(t
); /* Report deferred QS. */
695 } else if (!rcu_preempt_depth()) {
696 rcu_qs(); /* Report immediate QS. */
700 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
701 if (rcu_preempt_depth() > 0 &&
702 __this_cpu_read(rcu_data
.core_needs_qs
) &&
703 __this_cpu_read(rcu_data
.cpu_no_qs
.b
.norm
) &&
704 !t
->rcu_read_unlock_special
.b
.need_qs
&&
705 time_after(jiffies
, rcu_state
.gp_start
+ HZ
))
706 t
->rcu_read_unlock_special
.b
.need_qs
= true;
710 * Check for a task exiting while in a preemptible-RCU read-side
711 * critical section, clean up if so. No need to issue warnings, as
712 * debug_check_no_locks_held() already does this if lockdep is enabled.
713 * Besides, if this function does anything other than just immediately
714 * return, there was a bug of some sort. Spewing warnings from this
715 * function is like as not to simply obscure important prior warnings.
719 struct task_struct
*t
= current
;
721 if (unlikely(!list_empty(¤t
->rcu_node_entry
))) {
722 rcu_preempt_depth_set(1);
724 WRITE_ONCE(t
->rcu_read_unlock_special
.b
.blocked
, true);
725 } else if (unlikely(rcu_preempt_depth())) {
726 rcu_preempt_depth_set(1);
731 rcu_preempt_deferred_qs(current
);
735 * Dump the blocked-tasks state, but limit the list dump to the
736 * specified number of elements.
739 dump_blkd_tasks(struct rcu_node
*rnp
, int ncheck
)
743 struct list_head
*lhp
;
745 struct rcu_data
*rdp
;
746 struct rcu_node
*rnp1
;
748 raw_lockdep_assert_held_rcu_node(rnp
);
749 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
750 __func__
, rnp
->grplo
, rnp
->grphi
, rnp
->level
,
751 (long)READ_ONCE(rnp
->gp_seq
), (long)rnp
->completedqs
);
752 for (rnp1
= rnp
; rnp1
; rnp1
= rnp1
->parent
)
753 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
754 __func__
, rnp1
->grplo
, rnp1
->grphi
, rnp1
->qsmask
, rnp1
->qsmaskinit
, rnp1
->qsmaskinitnext
);
755 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
756 __func__
, READ_ONCE(rnp
->gp_tasks
), rnp
->boost_tasks
,
758 pr_info("%s: ->blkd_tasks", __func__
);
760 list_for_each(lhp
, &rnp
->blkd_tasks
) {
766 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++) {
767 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
768 onl
= !!(rdp
->grpmask
& rcu_rnp_online_cpus(rnp
));
769 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
771 (long)rdp
->rcu_onl_gp_seq
, rdp
->rcu_onl_gp_flags
,
772 (long)rdp
->rcu_ofl_gp_seq
, rdp
->rcu_ofl_gp_flags
);
776 #else /* #ifdef CONFIG_PREEMPT_RCU */
779 * Tell them what RCU they are running.
781 static void __init
rcu_bootup_announce(void)
783 pr_info("Hierarchical RCU implementation.\n");
784 rcu_bootup_announce_oddness();
788 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
789 * how many quiescent states passed, just if there was at least one since
790 * the start of the grace period, this just sets a flag. The caller must
791 * have disabled preemption.
793 static void rcu_qs(void)
795 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
796 if (!__this_cpu_read(rcu_data
.cpu_no_qs
.s
))
798 trace_rcu_grace_period(TPS("rcu_sched"),
799 __this_cpu_read(rcu_data
.gp_seq
), TPS("cpuqs"));
800 __this_cpu_write(rcu_data
.cpu_no_qs
.b
.norm
, false);
801 if (!__this_cpu_read(rcu_data
.cpu_no_qs
.b
.exp
))
803 __this_cpu_write(rcu_data
.cpu_no_qs
.b
.exp
, false);
804 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data
));
808 * Register an urgently needed quiescent state. If there is an
809 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
810 * dyntick-idle quiescent state visible to other CPUs, which will in
811 * some cases serve for expedited as well as normal grace periods.
812 * Either way, register a lightweight quiescent state.
814 void rcu_all_qs(void)
818 if (!raw_cpu_read(rcu_data
.rcu_urgent_qs
))
821 /* Load rcu_urgent_qs before other flags. */
822 if (!smp_load_acquire(this_cpu_ptr(&rcu_data
.rcu_urgent_qs
))) {
826 this_cpu_write(rcu_data
.rcu_urgent_qs
, false);
827 if (unlikely(raw_cpu_read(rcu_data
.rcu_need_heavy_qs
))) {
828 local_irq_save(flags
);
829 rcu_momentary_dyntick_idle();
830 local_irq_restore(flags
);
835 EXPORT_SYMBOL_GPL(rcu_all_qs
);
838 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
840 void rcu_note_context_switch(bool preempt
)
842 trace_rcu_utilization(TPS("Start context switch"));
844 /* Load rcu_urgent_qs before other flags. */
845 if (!smp_load_acquire(this_cpu_ptr(&rcu_data
.rcu_urgent_qs
)))
847 this_cpu_write(rcu_data
.rcu_urgent_qs
, false);
848 if (unlikely(raw_cpu_read(rcu_data
.rcu_need_heavy_qs
)))
849 rcu_momentary_dyntick_idle();
851 rcu_tasks_qs(current
);
853 trace_rcu_utilization(TPS("End context switch"));
855 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
858 * Because preemptible RCU does not exist, there are never any preempted
861 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
867 * Because there is no preemptible RCU, there can be no readers blocked.
869 static bool rcu_preempt_has_tasks(struct rcu_node
*rnp
)
875 * Because there is no preemptible RCU, there can be no deferred quiescent
878 static bool rcu_preempt_need_deferred_qs(struct task_struct
*t
)
882 static void rcu_preempt_deferred_qs(struct task_struct
*t
) { }
885 * Because there is no preemptible RCU, there can be no readers blocked,
886 * so there is no need to check for blocked tasks. So check only for
887 * bogus qsmask values.
889 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
891 WARN_ON_ONCE(rnp
->qsmask
);
895 * Check to see if this CPU is in a non-context-switch quiescent state,
896 * namely user mode and idle loop.
898 static void rcu_flavor_sched_clock_irq(int user
)
900 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
903 * Get here if this CPU took its interrupt from user
904 * mode or from the idle loop, and if this is not a
905 * nested interrupt. In this case, the CPU is in
906 * a quiescent state, so note it.
908 * No memory barrier is required here because rcu_qs()
909 * references only CPU-local variables that other CPUs
910 * neither access nor modify, at least not while the
911 * corresponding CPU is online.
919 * Because preemptible RCU does not exist, tasks cannot possibly exit
920 * while in preemptible RCU read-side critical sections.
927 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
930 dump_blkd_tasks(struct rcu_node
*rnp
, int ncheck
)
932 WARN_ON_ONCE(!list_empty(&rnp
->blkd_tasks
));
935 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
938 * If boosting, set rcuc kthreads to realtime priority.
940 static void rcu_cpu_kthread_setup(unsigned int cpu
)
942 #ifdef CONFIG_RCU_BOOST
943 struct sched_param sp
;
945 sp
.sched_priority
= kthread_prio
;
946 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
947 #endif /* #ifdef CONFIG_RCU_BOOST */
950 #ifdef CONFIG_RCU_BOOST
953 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
954 * or ->boost_tasks, advancing the pointer to the next task in the
957 * Note that irqs must be enabled: boosting the task can block.
958 * Returns 1 if there are more tasks needing to be boosted.
960 static int rcu_boost(struct rcu_node
*rnp
)
963 struct task_struct
*t
;
964 struct list_head
*tb
;
966 if (READ_ONCE(rnp
->exp_tasks
) == NULL
&&
967 READ_ONCE(rnp
->boost_tasks
) == NULL
)
968 return 0; /* Nothing left to boost. */
970 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
973 * Recheck under the lock: all tasks in need of boosting
974 * might exit their RCU read-side critical sections on their own.
976 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
977 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
982 * Preferentially boost tasks blocking expedited grace periods.
983 * This cannot starve the normal grace periods because a second
984 * expedited grace period must boost all blocked tasks, including
985 * those blocking the pre-existing normal grace period.
987 if (rnp
->exp_tasks
!= NULL
)
990 tb
= rnp
->boost_tasks
;
993 * We boost task t by manufacturing an rt_mutex that appears to
994 * be held by task t. We leave a pointer to that rt_mutex where
995 * task t can find it, and task t will release the mutex when it
996 * exits its outermost RCU read-side critical section. Then
997 * simply acquiring this artificial rt_mutex will boost task
998 * t's priority. (Thanks to tglx for suggesting this approach!)
1000 * Note that task t must acquire rnp->lock to remove itself from
1001 * the ->blkd_tasks list, which it will do from exit() if from
1002 * nowhere else. We therefore are guaranteed that task t will
1003 * stay around at least until we drop rnp->lock. Note that
1004 * rnp->lock also resolves races between our priority boosting
1005 * and task t's exiting its outermost RCU read-side critical
1008 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1009 rt_mutex_init_proxy_locked(&rnp
->boost_mtx
, t
);
1010 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1011 /* Lock only for side effect: boosts task t's priority. */
1012 rt_mutex_lock(&rnp
->boost_mtx
);
1013 rt_mutex_unlock(&rnp
->boost_mtx
); /* Then keep lockdep happy. */
1015 return READ_ONCE(rnp
->exp_tasks
) != NULL
||
1016 READ_ONCE(rnp
->boost_tasks
) != NULL
;
1020 * Priority-boosting kthread, one per leaf rcu_node.
1022 static int rcu_boost_kthread(void *arg
)
1024 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1028 trace_rcu_utilization(TPS("Start boost kthread@init"));
1030 WRITE_ONCE(rnp
->boost_kthread_status
, RCU_KTHREAD_WAITING
);
1031 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1032 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1033 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1034 WRITE_ONCE(rnp
->boost_kthread_status
, RCU_KTHREAD_RUNNING
);
1035 more2boost
= rcu_boost(rnp
);
1041 WRITE_ONCE(rnp
->boost_kthread_status
, RCU_KTHREAD_YIELDING
);
1042 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1043 schedule_timeout_interruptible(2);
1044 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1049 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1054 * Check to see if it is time to start boosting RCU readers that are
1055 * blocking the current grace period, and, if so, tell the per-rcu_node
1056 * kthread to start boosting them. If there is an expedited grace
1057 * period in progress, it is always time to boost.
1059 * The caller must hold rnp->lock, which this function releases.
1060 * The ->boost_kthread_task is immortal, so we don't need to worry
1061 * about it going away.
1063 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1064 __releases(rnp
->lock
)
1066 raw_lockdep_assert_held_rcu_node(rnp
);
1067 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1068 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1071 if (rnp
->exp_tasks
!= NULL
||
1072 (rnp
->gp_tasks
!= NULL
&&
1073 rnp
->boost_tasks
== NULL
&&
1075 (ULONG_CMP_GE(jiffies
, rnp
->boost_time
) || rcu_state
.cbovld
))) {
1076 if (rnp
->exp_tasks
== NULL
)
1077 rnp
->boost_tasks
= rnp
->gp_tasks
;
1078 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1079 rcu_wake_cond(rnp
->boost_kthread_task
,
1080 READ_ONCE(rnp
->boost_kthread_status
));
1082 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1087 * Is the current CPU running the RCU-callbacks kthread?
1088 * Caller must have preemption disabled.
1090 static bool rcu_is_callbacks_kthread(void)
1092 return __this_cpu_read(rcu_data
.rcu_cpu_kthread_task
) == current
;
1095 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1098 * Do priority-boost accounting for the start of a new grace period.
1100 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1102 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1106 * Create an RCU-boost kthread for the specified node if one does not
1107 * already exist. We only create this kthread for preemptible RCU.
1108 * Returns zero if all is well, a negated errno otherwise.
1110 static void rcu_spawn_one_boost_kthread(struct rcu_node
*rnp
)
1112 int rnp_index
= rnp
- rcu_get_root();
1113 unsigned long flags
;
1114 struct sched_param sp
;
1115 struct task_struct
*t
;
1117 if (!IS_ENABLED(CONFIG_PREEMPT_RCU
))
1120 if (!rcu_scheduler_fully_active
|| rcu_rnp_online_cpus(rnp
) == 0)
1123 rcu_state
.boost
= 1;
1125 if (rnp
->boost_kthread_task
!= NULL
)
1128 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1129 "rcub/%d", rnp_index
);
1130 if (WARN_ON_ONCE(IS_ERR(t
)))
1133 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1134 rnp
->boost_kthread_task
= t
;
1135 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1136 sp
.sched_priority
= kthread_prio
;
1137 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1138 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1142 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1143 * served by the rcu_node in question. The CPU hotplug lock is still
1144 * held, so the value of rnp->qsmaskinit will be stable.
1146 * We don't include outgoingcpu in the affinity set, use -1 if there is
1147 * no outgoing CPU. If there are no CPUs left in the affinity set,
1148 * this function allows the kthread to execute on any CPU.
1150 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1152 struct task_struct
*t
= rnp
->boost_kthread_task
;
1153 unsigned long mask
= rcu_rnp_online_cpus(rnp
);
1159 if (!zalloc_cpumask_var(&cm
, GFP_KERNEL
))
1161 for_each_leaf_node_possible_cpu(rnp
, cpu
)
1162 if ((mask
& leaf_node_cpu_bit(rnp
, cpu
)) &&
1164 cpumask_set_cpu(cpu
, cm
);
1165 if (cpumask_weight(cm
) == 0)
1167 set_cpus_allowed_ptr(t
, cm
);
1168 free_cpumask_var(cm
);
1172 * Spawn boost kthreads -- called as soon as the scheduler is running.
1174 static void __init
rcu_spawn_boost_kthreads(void)
1176 struct rcu_node
*rnp
;
1178 rcu_for_each_leaf_node(rnp
)
1179 rcu_spawn_one_boost_kthread(rnp
);
1182 static void rcu_prepare_kthreads(int cpu
)
1184 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
1185 struct rcu_node
*rnp
= rdp
->mynode
;
1187 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1188 if (rcu_scheduler_fully_active
)
1189 rcu_spawn_one_boost_kthread(rnp
);
1192 #else /* #ifdef CONFIG_RCU_BOOST */
1194 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1195 __releases(rnp
->lock
)
1197 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1200 static bool rcu_is_callbacks_kthread(void)
1205 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1209 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1213 static void __init
rcu_spawn_boost_kthreads(void)
1217 static void rcu_prepare_kthreads(int cpu
)
1221 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1223 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1226 * Check to see if any future non-offloaded RCU-related work will need
1227 * to be done by the current CPU, even if none need be done immediately,
1228 * returning 1 if so. This function is part of the RCU implementation;
1229 * it is -not- an exported member of the RCU API.
1231 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1232 * CPU has RCU callbacks queued.
1234 int rcu_needs_cpu(u64 basemono
, u64
*nextevt
)
1236 *nextevt
= KTIME_MAX
;
1237 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data
)->cblist
) &&
1238 !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data
)->cblist
);
1242 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1245 static void rcu_cleanup_after_idle(void)
1250 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1253 static void rcu_prepare_for_idle(void)
1257 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1260 * This code is invoked when a CPU goes idle, at which point we want
1261 * to have the CPU do everything required for RCU so that it can enter
1262 * the energy-efficient dyntick-idle mode.
1264 * The following preprocessor symbol controls this:
1266 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1267 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1268 * is sized to be roughly one RCU grace period. Those energy-efficiency
1269 * benchmarkers who might otherwise be tempted to set this to a large
1270 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1271 * system. And if you are -that- concerned about energy efficiency,
1272 * just power the system down and be done with it!
1274 * The value below works well in practice. If future workloads require
1275 * adjustment, they can be converted into kernel config parameters, though
1276 * making the state machine smarter might be a better option.
1278 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1280 static int rcu_idle_gp_delay
= RCU_IDLE_GP_DELAY
;
1281 module_param(rcu_idle_gp_delay
, int, 0644);
1284 * Try to advance callbacks on the current CPU, but only if it has been
1285 * awhile since the last time we did so. Afterwards, if there are any
1286 * callbacks ready for immediate invocation, return true.
1288 static bool __maybe_unused
rcu_try_advance_all_cbs(void)
1290 bool cbs_ready
= false;
1291 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1292 struct rcu_node
*rnp
;
1294 /* Exit early if we advanced recently. */
1295 if (jiffies
== rdp
->last_advance_all
)
1297 rdp
->last_advance_all
= jiffies
;
1302 * Don't bother checking unless a grace period has
1303 * completed since we last checked and there are
1304 * callbacks not yet ready to invoke.
1306 if ((rcu_seq_completed_gp(rdp
->gp_seq
,
1307 rcu_seq_current(&rnp
->gp_seq
)) ||
1308 unlikely(READ_ONCE(rdp
->gpwrap
))) &&
1309 rcu_segcblist_pend_cbs(&rdp
->cblist
))
1310 note_gp_changes(rdp
);
1312 if (rcu_segcblist_ready_cbs(&rdp
->cblist
))
1318 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1319 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1320 * caller about what to set the timeout.
1322 * The caller must have disabled interrupts.
1324 int rcu_needs_cpu(u64 basemono
, u64
*nextevt
)
1326 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1329 lockdep_assert_irqs_disabled();
1331 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1332 if (rcu_segcblist_empty(&rdp
->cblist
) ||
1333 rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data
)->cblist
)) {
1334 *nextevt
= KTIME_MAX
;
1338 /* Attempt to advance callbacks. */
1339 if (rcu_try_advance_all_cbs()) {
1340 /* Some ready to invoke, so initiate later invocation. */
1344 rdp
->last_accelerate
= jiffies
;
1346 /* Request timer and round. */
1347 dj
= round_up(rcu_idle_gp_delay
+ jiffies
, rcu_idle_gp_delay
) - jiffies
;
1349 *nextevt
= basemono
+ dj
* TICK_NSEC
;
1354 * Prepare a CPU for idle from an RCU perspective. The first major task is to
1355 * sense whether nohz mode has been enabled or disabled via sysfs. The second
1356 * major task is to accelerate (that is, assign grace-period numbers to) any
1357 * recently arrived callbacks.
1359 * The caller must have disabled interrupts.
1361 static void rcu_prepare_for_idle(void)
1364 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1365 struct rcu_node
*rnp
;
1368 lockdep_assert_irqs_disabled();
1369 if (rcu_segcblist_is_offloaded(&rdp
->cblist
))
1372 /* Handle nohz enablement switches conservatively. */
1373 tne
= READ_ONCE(tick_nohz_active
);
1374 if (tne
!= rdp
->tick_nohz_enabled_snap
) {
1375 if (!rcu_segcblist_empty(&rdp
->cblist
))
1376 invoke_rcu_core(); /* force nohz to see update. */
1377 rdp
->tick_nohz_enabled_snap
= tne
;
1384 * If we have not yet accelerated this jiffy, accelerate all
1385 * callbacks on this CPU.
1387 if (rdp
->last_accelerate
== jiffies
)
1389 rdp
->last_accelerate
= jiffies
;
1390 if (rcu_segcblist_pend_cbs(&rdp
->cblist
)) {
1392 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
1393 needwake
= rcu_accelerate_cbs(rnp
, rdp
);
1394 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
1396 rcu_gp_kthread_wake();
1401 * Clean up for exit from idle. Attempt to advance callbacks based on
1402 * any grace periods that elapsed while the CPU was idle, and if any
1403 * callbacks are now ready to invoke, initiate invocation.
1405 static void rcu_cleanup_after_idle(void)
1407 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1409 lockdep_assert_irqs_disabled();
1410 if (rcu_segcblist_is_offloaded(&rdp
->cblist
))
1412 if (rcu_try_advance_all_cbs())
1416 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1418 #ifdef CONFIG_RCU_NOCB_CPU
1421 * Offload callback processing from the boot-time-specified set of CPUs
1422 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1423 * created that pull the callbacks from the corresponding CPU, wait for
1424 * a grace period to elapse, and invoke the callbacks. These kthreads
1425 * are organized into GP kthreads, which manage incoming callbacks, wait for
1426 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1427 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1428 * do a wake_up() on their GP kthread when they insert a callback into any
1429 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1430 * in which case each kthread actively polls its CPU. (Which isn't so great
1431 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1433 * This is intended to be used in conjunction with Frederic Weisbecker's
1434 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1435 * running CPU-bound user-mode computations.
1437 * Offloading of callbacks can also be used as an energy-efficiency
1438 * measure because CPUs with no RCU callbacks queued are more aggressive
1439 * about entering dyntick-idle mode.
1444 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1445 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1446 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1447 * given, a warning is emitted and all CPUs are offloaded.
1449 static int __init
rcu_nocb_setup(char *str
)
1451 alloc_bootmem_cpumask_var(&rcu_nocb_mask
);
1452 if (!strcasecmp(str
, "all"))
1453 cpumask_setall(rcu_nocb_mask
);
1455 if (cpulist_parse(str
, rcu_nocb_mask
)) {
1456 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1457 cpumask_setall(rcu_nocb_mask
);
1461 __setup("rcu_nocbs=", rcu_nocb_setup
);
1463 static int __init
parse_rcu_nocb_poll(char *arg
)
1465 rcu_nocb_poll
= true;
1468 early_param("rcu_nocb_poll", parse_rcu_nocb_poll
);
1471 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1472 * After all, the main point of bypassing is to avoid lock contention
1473 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1475 int nocb_nobypass_lim_per_jiffy
= 16 * 1000 / HZ
;
1476 module_param(nocb_nobypass_lim_per_jiffy
, int, 0);
1479 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1480 * lock isn't immediately available, increment ->nocb_lock_contended to
1481 * flag the contention.
1483 static void rcu_nocb_bypass_lock(struct rcu_data
*rdp
)
1484 __acquires(&rdp
->nocb_bypass_lock
)
1486 lockdep_assert_irqs_disabled();
1487 if (raw_spin_trylock(&rdp
->nocb_bypass_lock
))
1489 atomic_inc(&rdp
->nocb_lock_contended
);
1490 WARN_ON_ONCE(smp_processor_id() != rdp
->cpu
);
1491 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1492 raw_spin_lock(&rdp
->nocb_bypass_lock
);
1493 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1494 atomic_dec(&rdp
->nocb_lock_contended
);
1498 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1499 * not contended. Please note that this is extremely special-purpose,
1500 * relying on the fact that at most two kthreads and one CPU contend for
1501 * this lock, and also that the two kthreads are guaranteed to have frequent
1502 * grace-period-duration time intervals between successive acquisitions
1503 * of the lock. This allows us to use an extremely simple throttling
1504 * mechanism, and further to apply it only to the CPU doing floods of
1505 * call_rcu() invocations. Don't try this at home!
1507 static void rcu_nocb_wait_contended(struct rcu_data
*rdp
)
1509 WARN_ON_ONCE(smp_processor_id() != rdp
->cpu
);
1510 while (WARN_ON_ONCE(atomic_read(&rdp
->nocb_lock_contended
)))
1515 * Conditionally acquire the specified rcu_data structure's
1516 * ->nocb_bypass_lock.
1518 static bool rcu_nocb_bypass_trylock(struct rcu_data
*rdp
)
1520 lockdep_assert_irqs_disabled();
1521 return raw_spin_trylock(&rdp
->nocb_bypass_lock
);
1525 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1527 static void rcu_nocb_bypass_unlock(struct rcu_data
*rdp
)
1528 __releases(&rdp
->nocb_bypass_lock
)
1530 lockdep_assert_irqs_disabled();
1531 raw_spin_unlock(&rdp
->nocb_bypass_lock
);
1535 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1536 * if it corresponds to a no-CBs CPU.
1538 static void rcu_nocb_lock(struct rcu_data
*rdp
)
1540 lockdep_assert_irqs_disabled();
1541 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
))
1543 raw_spin_lock(&rdp
->nocb_lock
);
1547 * Release the specified rcu_data structure's ->nocb_lock, but only
1548 * if it corresponds to a no-CBs CPU.
1550 static void rcu_nocb_unlock(struct rcu_data
*rdp
)
1552 if (rcu_segcblist_is_offloaded(&rdp
->cblist
)) {
1553 lockdep_assert_irqs_disabled();
1554 raw_spin_unlock(&rdp
->nocb_lock
);
1559 * Release the specified rcu_data structure's ->nocb_lock and restore
1560 * interrupts, but only if it corresponds to a no-CBs CPU.
1562 static void rcu_nocb_unlock_irqrestore(struct rcu_data
*rdp
,
1563 unsigned long flags
)
1565 if (rcu_segcblist_is_offloaded(&rdp
->cblist
)) {
1566 lockdep_assert_irqs_disabled();
1567 raw_spin_unlock_irqrestore(&rdp
->nocb_lock
, flags
);
1569 local_irq_restore(flags
);
1573 /* Lockdep check that ->cblist may be safely accessed. */
1574 static void rcu_lockdep_assert_cblist_protected(struct rcu_data
*rdp
)
1576 lockdep_assert_irqs_disabled();
1577 if (rcu_segcblist_is_offloaded(&rdp
->cblist
))
1578 lockdep_assert_held(&rdp
->nocb_lock
);
1582 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1585 static void rcu_nocb_gp_cleanup(struct swait_queue_head
*sq
)
1590 static struct swait_queue_head
*rcu_nocb_gp_get(struct rcu_node
*rnp
)
1592 return &rnp
->nocb_gp_wq
[rcu_seq_ctr(rnp
->gp_seq
) & 0x1];
1595 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
1597 init_swait_queue_head(&rnp
->nocb_gp_wq
[0]);
1598 init_swait_queue_head(&rnp
->nocb_gp_wq
[1]);
1601 /* Is the specified CPU a no-CBs CPU? */
1602 bool rcu_is_nocb_cpu(int cpu
)
1604 if (cpumask_available(rcu_nocb_mask
))
1605 return cpumask_test_cpu(cpu
, rcu_nocb_mask
);
1610 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1611 * and this function releases it.
1613 static void wake_nocb_gp(struct rcu_data
*rdp
, bool force
,
1614 unsigned long flags
)
1615 __releases(rdp
->nocb_lock
)
1617 bool needwake
= false;
1618 struct rcu_data
*rdp_gp
= rdp
->nocb_gp_rdp
;
1620 lockdep_assert_held(&rdp
->nocb_lock
);
1621 if (!READ_ONCE(rdp_gp
->nocb_gp_kthread
)) {
1622 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1623 TPS("AlreadyAwake"));
1624 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1627 del_timer(&rdp
->nocb_timer
);
1628 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1629 raw_spin_lock_irqsave(&rdp_gp
->nocb_gp_lock
, flags
);
1630 if (force
|| READ_ONCE(rdp_gp
->nocb_gp_sleep
)) {
1631 WRITE_ONCE(rdp_gp
->nocb_gp_sleep
, false);
1633 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("DoWake"));
1635 raw_spin_unlock_irqrestore(&rdp_gp
->nocb_gp_lock
, flags
);
1637 wake_up_process(rdp_gp
->nocb_gp_kthread
);
1641 * Arrange to wake the GP kthread for this NOCB group at some future
1642 * time when it is safe to do so.
1644 static void wake_nocb_gp_defer(struct rcu_data
*rdp
, int waketype
,
1647 if (rdp
->nocb_defer_wakeup
== RCU_NOCB_WAKE_NOT
)
1648 mod_timer(&rdp
->nocb_timer
, jiffies
+ 1);
1649 if (rdp
->nocb_defer_wakeup
< waketype
)
1650 WRITE_ONCE(rdp
->nocb_defer_wakeup
, waketype
);
1651 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, reason
);
1655 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1656 * However, if there is a callback to be enqueued and if ->nocb_bypass
1657 * proves to be initially empty, just return false because the no-CB GP
1658 * kthread may need to be awakened in this case.
1660 * Note that this function always returns true if rhp is NULL.
1662 static bool rcu_nocb_do_flush_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
1665 struct rcu_cblist rcl
;
1667 WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp
->cblist
));
1668 rcu_lockdep_assert_cblist_protected(rdp
);
1669 lockdep_assert_held(&rdp
->nocb_bypass_lock
);
1670 if (rhp
&& !rcu_cblist_n_cbs(&rdp
->nocb_bypass
)) {
1671 raw_spin_unlock(&rdp
->nocb_bypass_lock
);
1674 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1676 rcu_segcblist_inc_len(&rdp
->cblist
); /* Must precede enqueue. */
1677 rcu_cblist_flush_enqueue(&rcl
, &rdp
->nocb_bypass
, rhp
);
1678 rcu_segcblist_insert_pend_cbs(&rdp
->cblist
, &rcl
);
1679 WRITE_ONCE(rdp
->nocb_bypass_first
, j
);
1680 rcu_nocb_bypass_unlock(rdp
);
1685 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1686 * However, if there is a callback to be enqueued and if ->nocb_bypass
1687 * proves to be initially empty, just return false because the no-CB GP
1688 * kthread may need to be awakened in this case.
1690 * Note that this function always returns true if rhp is NULL.
1692 static bool rcu_nocb_flush_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
1695 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
))
1697 rcu_lockdep_assert_cblist_protected(rdp
);
1698 rcu_nocb_bypass_lock(rdp
);
1699 return rcu_nocb_do_flush_bypass(rdp
, rhp
, j
);
1703 * If the ->nocb_bypass_lock is immediately available, flush the
1704 * ->nocb_bypass queue into ->cblist.
1706 static void rcu_nocb_try_flush_bypass(struct rcu_data
*rdp
, unsigned long j
)
1708 rcu_lockdep_assert_cblist_protected(rdp
);
1709 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
) ||
1710 !rcu_nocb_bypass_trylock(rdp
))
1712 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp
, NULL
, j
));
1716 * See whether it is appropriate to use the ->nocb_bypass list in order
1717 * to control contention on ->nocb_lock. A limited number of direct
1718 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1719 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1720 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1721 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1722 * used if ->cblist is empty, because otherwise callbacks can be stranded
1723 * on ->nocb_bypass because we cannot count on the current CPU ever again
1724 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1725 * non-empty, the corresponding no-CBs grace-period kthread must not be
1726 * in an indefinite sleep state.
1728 * Finally, it is not permitted to use the bypass during early boot,
1729 * as doing so would confuse the auto-initialization code. Besides
1730 * which, there is no point in worrying about lock contention while
1731 * there is only one CPU in operation.
1733 static bool rcu_nocb_try_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
1734 bool *was_alldone
, unsigned long flags
)
1737 unsigned long cur_gp_seq
;
1738 unsigned long j
= jiffies
;
1739 long ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1741 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
)) {
1742 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1743 return false; /* Not offloaded, no bypassing. */
1745 lockdep_assert_irqs_disabled();
1747 // Don't use ->nocb_bypass during early boot.
1748 if (rcu_scheduler_active
!= RCU_SCHEDULER_RUNNING
) {
1750 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp
->nocb_bypass
));
1751 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1755 // If we have advanced to a new jiffy, reset counts to allow
1756 // moving back from ->nocb_bypass to ->cblist.
1757 if (j
== rdp
->nocb_nobypass_last
) {
1758 c
= rdp
->nocb_nobypass_count
+ 1;
1760 WRITE_ONCE(rdp
->nocb_nobypass_last
, j
);
1761 c
= rdp
->nocb_nobypass_count
- nocb_nobypass_lim_per_jiffy
;
1762 if (ULONG_CMP_LT(rdp
->nocb_nobypass_count
,
1763 nocb_nobypass_lim_per_jiffy
))
1765 else if (c
> nocb_nobypass_lim_per_jiffy
)
1766 c
= nocb_nobypass_lim_per_jiffy
;
1768 WRITE_ONCE(rdp
->nocb_nobypass_count
, c
);
1770 // If there hasn't yet been all that many ->cblist enqueues
1771 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1772 // ->nocb_bypass first.
1773 if (rdp
->nocb_nobypass_count
< nocb_nobypass_lim_per_jiffy
) {
1775 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1777 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1779 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp
, NULL
, j
));
1780 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp
->nocb_bypass
));
1781 return false; // Caller must enqueue the callback.
1784 // If ->nocb_bypass has been used too long or is too full,
1785 // flush ->nocb_bypass to ->cblist.
1786 if ((ncbs
&& j
!= READ_ONCE(rdp
->nocb_bypass_first
)) ||
1789 if (!rcu_nocb_flush_bypass(rdp
, rhp
, j
)) {
1790 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1792 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1794 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp
->nocb_bypass
));
1795 return false; // Caller must enqueue the callback.
1797 if (j
!= rdp
->nocb_gp_adv_time
&&
1798 rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
1799 rcu_seq_done(&rdp
->mynode
->gp_seq
, cur_gp_seq
)) {
1800 rcu_advance_cbs_nowake(rdp
->mynode
, rdp
);
1801 rdp
->nocb_gp_adv_time
= j
;
1803 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1804 return true; // Callback already enqueued.
1807 // We need to use the bypass.
1808 rcu_nocb_wait_contended(rdp
);
1809 rcu_nocb_bypass_lock(rdp
);
1810 ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1811 rcu_segcblist_inc_len(&rdp
->cblist
); /* Must precede enqueue. */
1812 rcu_cblist_enqueue(&rdp
->nocb_bypass
, rhp
);
1814 WRITE_ONCE(rdp
->nocb_bypass_first
, j
);
1815 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("FirstBQ"));
1817 rcu_nocb_bypass_unlock(rdp
);
1818 smp_mb(); /* Order enqueue before wake. */
1820 local_irq_restore(flags
);
1822 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1823 rcu_nocb_lock(rdp
); // Rare during call_rcu() flood.
1824 if (!rcu_segcblist_pend_cbs(&rdp
->cblist
)) {
1825 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1826 TPS("FirstBQwake"));
1827 __call_rcu_nocb_wake(rdp
, true, flags
);
1829 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1830 TPS("FirstBQnoWake"));
1831 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1834 return true; // Callback already enqueued.
1838 * Awaken the no-CBs grace-period kthead if needed, either due to it
1839 * legitimately being asleep or due to overload conditions.
1841 * If warranted, also wake up the kthread servicing this CPUs queues.
1843 static void __call_rcu_nocb_wake(struct rcu_data
*rdp
, bool was_alldone
,
1844 unsigned long flags
)
1845 __releases(rdp
->nocb_lock
)
1847 unsigned long cur_gp_seq
;
1850 struct task_struct
*t
;
1852 // If we are being polled or there is no kthread, just leave.
1853 t
= READ_ONCE(rdp
->nocb_gp_kthread
);
1854 if (rcu_nocb_poll
|| !t
) {
1855 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1856 TPS("WakeNotPoll"));
1857 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1860 // Need to actually to a wakeup.
1861 len
= rcu_segcblist_n_cbs(&rdp
->cblist
);
1863 rdp
->qlen_last_fqs_check
= len
;
1864 if (!irqs_disabled_flags(flags
)) {
1865 /* ... if queue was empty ... */
1866 wake_nocb_gp(rdp
, false, flags
);
1867 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1870 wake_nocb_gp_defer(rdp
, RCU_NOCB_WAKE
,
1871 TPS("WakeEmptyIsDeferred"));
1872 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1874 } else if (len
> rdp
->qlen_last_fqs_check
+ qhimark
) {
1875 /* ... or if many callbacks queued. */
1876 rdp
->qlen_last_fqs_check
= len
;
1878 if (j
!= rdp
->nocb_gp_adv_time
&&
1879 rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
1880 rcu_seq_done(&rdp
->mynode
->gp_seq
, cur_gp_seq
)) {
1881 rcu_advance_cbs_nowake(rdp
->mynode
, rdp
);
1882 rdp
->nocb_gp_adv_time
= j
;
1884 smp_mb(); /* Enqueue before timer_pending(). */
1885 if ((rdp
->nocb_cb_sleep
||
1886 !rcu_segcblist_ready_cbs(&rdp
->cblist
)) &&
1887 !timer_pending(&rdp
->nocb_bypass_timer
))
1888 wake_nocb_gp_defer(rdp
, RCU_NOCB_WAKE_FORCE
,
1889 TPS("WakeOvfIsDeferred"));
1890 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1892 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("WakeNot"));
1893 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1898 /* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1899 static void do_nocb_bypass_wakeup_timer(struct timer_list
*t
)
1901 unsigned long flags
;
1902 struct rcu_data
*rdp
= from_timer(rdp
, t
, nocb_bypass_timer
);
1904 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("Timer"));
1905 rcu_nocb_lock_irqsave(rdp
, flags
);
1906 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1907 __call_rcu_nocb_wake(rdp
, true, flags
);
1911 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1912 * or for grace periods to end.
1914 static void nocb_gp_wait(struct rcu_data
*my_rdp
)
1916 bool bypass
= false;
1918 int __maybe_unused cpu
= my_rdp
->cpu
;
1919 unsigned long cur_gp_seq
;
1920 unsigned long flags
;
1921 bool gotcbs
= false;
1922 unsigned long j
= jiffies
;
1923 bool needwait_gp
= false; // This prevents actual uninitialized use.
1926 struct rcu_data
*rdp
;
1927 struct rcu_node
*rnp
;
1928 unsigned long wait_gp_seq
= 0; // Suppress "use uninitialized" warning.
1929 bool wasempty
= false;
1932 * Each pass through the following loop checks for CBs and for the
1933 * nearest grace period (if any) to wait for next. The CB kthreads
1934 * and the global grace-period kthread are awakened if needed.
1936 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_cb_rdp
) {
1937 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("Check"));
1938 rcu_nocb_lock_irqsave(rdp
, flags
);
1939 bypass_ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1941 (time_after(j
, READ_ONCE(rdp
->nocb_bypass_first
) + 1) ||
1942 bypass_ncbs
> 2 * qhimark
)) {
1943 // Bypass full or old, so flush it.
1944 (void)rcu_nocb_try_flush_bypass(rdp
, j
);
1945 bypass_ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1946 } else if (!bypass_ncbs
&& rcu_segcblist_empty(&rdp
->cblist
)) {
1947 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1948 continue; /* No callbacks here, try next. */
1951 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1956 if (bypass
) { // Avoid race with first bypass CB.
1957 WRITE_ONCE(my_rdp
->nocb_defer_wakeup
,
1959 del_timer(&my_rdp
->nocb_timer
);
1961 // Advance callbacks if helpful and low contention.
1962 needwake_gp
= false;
1963 if (!rcu_segcblist_restempty(&rdp
->cblist
,
1964 RCU_NEXT_READY_TAIL
) ||
1965 (rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
1966 rcu_seq_done(&rnp
->gp_seq
, cur_gp_seq
))) {
1967 raw_spin_lock_rcu_node(rnp
); /* irqs disabled. */
1968 needwake_gp
= rcu_advance_cbs(rnp
, rdp
);
1969 wasempty
= rcu_segcblist_restempty(&rdp
->cblist
,
1970 RCU_NEXT_READY_TAIL
);
1971 raw_spin_unlock_rcu_node(rnp
); /* irqs disabled. */
1973 // Need to wait on some grace period?
1974 WARN_ON_ONCE(wasempty
&&
1975 !rcu_segcblist_restempty(&rdp
->cblist
,
1976 RCU_NEXT_READY_TAIL
));
1977 if (rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
)) {
1979 ULONG_CMP_LT(cur_gp_seq
, wait_gp_seq
))
1980 wait_gp_seq
= cur_gp_seq
;
1982 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1985 if (rcu_segcblist_ready_cbs(&rdp
->cblist
)) {
1986 needwake
= rdp
->nocb_cb_sleep
;
1987 WRITE_ONCE(rdp
->nocb_cb_sleep
, false);
1988 smp_mb(); /* CB invocation -after- GP end. */
1992 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1994 swake_up_one(&rdp
->nocb_cb_wq
);
1998 rcu_gp_kthread_wake();
2001 my_rdp
->nocb_gp_bypass
= bypass
;
2002 my_rdp
->nocb_gp_gp
= needwait_gp
;
2003 my_rdp
->nocb_gp_seq
= needwait_gp
? wait_gp_seq
: 0;
2004 if (bypass
&& !rcu_nocb_poll
) {
2005 // At least one child with non-empty ->nocb_bypass, so set
2006 // timer in order to avoid stranding its callbacks.
2007 raw_spin_lock_irqsave(&my_rdp
->nocb_gp_lock
, flags
);
2008 mod_timer(&my_rdp
->nocb_bypass_timer
, j
+ 2);
2009 raw_spin_unlock_irqrestore(&my_rdp
->nocb_gp_lock
, flags
);
2011 if (rcu_nocb_poll
) {
2012 /* Polling, so trace if first poll in the series. */
2014 trace_rcu_nocb_wake(rcu_state
.name
, cpu
, TPS("Poll"));
2015 schedule_timeout_interruptible(1);
2016 } else if (!needwait_gp
) {
2017 /* Wait for callbacks to appear. */
2018 trace_rcu_nocb_wake(rcu_state
.name
, cpu
, TPS("Sleep"));
2019 swait_event_interruptible_exclusive(my_rdp
->nocb_gp_wq
,
2020 !READ_ONCE(my_rdp
->nocb_gp_sleep
));
2021 trace_rcu_nocb_wake(rcu_state
.name
, cpu
, TPS("EndSleep"));
2023 rnp
= my_rdp
->mynode
;
2024 trace_rcu_this_gp(rnp
, my_rdp
, wait_gp_seq
, TPS("StartWait"));
2025 swait_event_interruptible_exclusive(
2026 rnp
->nocb_gp_wq
[rcu_seq_ctr(wait_gp_seq
) & 0x1],
2027 rcu_seq_done(&rnp
->gp_seq
, wait_gp_seq
) ||
2028 !READ_ONCE(my_rdp
->nocb_gp_sleep
));
2029 trace_rcu_this_gp(rnp
, my_rdp
, wait_gp_seq
, TPS("EndWait"));
2031 if (!rcu_nocb_poll
) {
2032 raw_spin_lock_irqsave(&my_rdp
->nocb_gp_lock
, flags
);
2034 del_timer(&my_rdp
->nocb_bypass_timer
);
2035 WRITE_ONCE(my_rdp
->nocb_gp_sleep
, true);
2036 raw_spin_unlock_irqrestore(&my_rdp
->nocb_gp_lock
, flags
);
2038 my_rdp
->nocb_gp_seq
= -1;
2039 WARN_ON(signal_pending(current
));
2043 * No-CBs grace-period-wait kthread. There is one of these per group
2044 * of CPUs, but only once at least one CPU in that group has come online
2045 * at least once since boot. This kthread checks for newly posted
2046 * callbacks from any of the CPUs it is responsible for, waits for a
2047 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2048 * that then have callback-invocation work to do.
2050 static int rcu_nocb_gp_kthread(void *arg
)
2052 struct rcu_data
*rdp
= arg
;
2055 WRITE_ONCE(rdp
->nocb_gp_loops
, rdp
->nocb_gp_loops
+ 1);
2057 cond_resched_tasks_rcu_qs();
2063 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2064 * then, if there are no more, wait for more to appear.
2066 static void nocb_cb_wait(struct rcu_data
*rdp
)
2068 unsigned long cur_gp_seq
;
2069 unsigned long flags
;
2070 bool needwake_gp
= false;
2071 struct rcu_node
*rnp
= rdp
->mynode
;
2073 local_irq_save(flags
);
2074 rcu_momentary_dyntick_idle();
2075 local_irq_restore(flags
);
2079 lockdep_assert_irqs_enabled();
2080 rcu_nocb_lock_irqsave(rdp
, flags
);
2081 if (rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
2082 rcu_seq_done(&rnp
->gp_seq
, cur_gp_seq
) &&
2083 raw_spin_trylock_rcu_node(rnp
)) { /* irqs already disabled. */
2084 needwake_gp
= rcu_advance_cbs(rdp
->mynode
, rdp
);
2085 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2087 if (rcu_segcblist_ready_cbs(&rdp
->cblist
)) {
2088 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2090 rcu_gp_kthread_wake();
2094 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("CBSleep"));
2095 WRITE_ONCE(rdp
->nocb_cb_sleep
, true);
2096 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2098 rcu_gp_kthread_wake();
2099 swait_event_interruptible_exclusive(rdp
->nocb_cb_wq
,
2100 !READ_ONCE(rdp
->nocb_cb_sleep
));
2101 if (!smp_load_acquire(&rdp
->nocb_cb_sleep
)) { /* VVV */
2102 /* ^^^ Ensure CB invocation follows _sleep test. */
2105 WARN_ON(signal_pending(current
));
2106 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("WokeEmpty"));
2110 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2111 * nocb_cb_wait() to do the dirty work.
2113 static int rcu_nocb_cb_kthread(void *arg
)
2115 struct rcu_data
*rdp
= arg
;
2117 // Each pass through this loop does one callback batch, and,
2118 // if there are no more ready callbacks, waits for them.
2121 cond_resched_tasks_rcu_qs();
2126 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2127 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2129 return READ_ONCE(rdp
->nocb_defer_wakeup
);
2132 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2133 static void do_nocb_deferred_wakeup_common(struct rcu_data
*rdp
)
2135 unsigned long flags
;
2138 rcu_nocb_lock_irqsave(rdp
, flags
);
2139 if (!rcu_nocb_need_deferred_wakeup(rdp
)) {
2140 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2143 ndw
= READ_ONCE(rdp
->nocb_defer_wakeup
);
2144 WRITE_ONCE(rdp
->nocb_defer_wakeup
, RCU_NOCB_WAKE_NOT
);
2145 wake_nocb_gp(rdp
, ndw
== RCU_NOCB_WAKE_FORCE
, flags
);
2146 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("DeferredWake"));
2149 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2150 static void do_nocb_deferred_wakeup_timer(struct timer_list
*t
)
2152 struct rcu_data
*rdp
= from_timer(rdp
, t
, nocb_timer
);
2154 do_nocb_deferred_wakeup_common(rdp
);
2158 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2159 * This means we do an inexact common-case check. Note that if
2160 * we miss, ->nocb_timer will eventually clean things up.
2162 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2164 if (rcu_nocb_need_deferred_wakeup(rdp
))
2165 do_nocb_deferred_wakeup_common(rdp
);
2168 void __init
rcu_init_nohz(void)
2171 bool need_rcu_nocb_mask
= false;
2172 struct rcu_data
*rdp
;
2174 #if defined(CONFIG_NO_HZ_FULL)
2175 if (tick_nohz_full_running
&& cpumask_weight(tick_nohz_full_mask
))
2176 need_rcu_nocb_mask
= true;
2177 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2179 if (!cpumask_available(rcu_nocb_mask
) && need_rcu_nocb_mask
) {
2180 if (!zalloc_cpumask_var(&rcu_nocb_mask
, GFP_KERNEL
)) {
2181 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2185 if (!cpumask_available(rcu_nocb_mask
))
2188 #if defined(CONFIG_NO_HZ_FULL)
2189 if (tick_nohz_full_running
)
2190 cpumask_or(rcu_nocb_mask
, rcu_nocb_mask
, tick_nohz_full_mask
);
2191 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2193 if (!cpumask_subset(rcu_nocb_mask
, cpu_possible_mask
)) {
2194 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2195 cpumask_and(rcu_nocb_mask
, cpu_possible_mask
,
2198 if (cpumask_empty(rcu_nocb_mask
))
2199 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2201 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2202 cpumask_pr_args(rcu_nocb_mask
));
2204 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2206 for_each_cpu(cpu
, rcu_nocb_mask
) {
2207 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2208 if (rcu_segcblist_empty(&rdp
->cblist
))
2209 rcu_segcblist_init(&rdp
->cblist
);
2210 rcu_segcblist_offload(&rdp
->cblist
);
2212 rcu_organize_nocb_kthreads();
2215 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2216 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2218 init_swait_queue_head(&rdp
->nocb_cb_wq
);
2219 init_swait_queue_head(&rdp
->nocb_gp_wq
);
2220 raw_spin_lock_init(&rdp
->nocb_lock
);
2221 raw_spin_lock_init(&rdp
->nocb_bypass_lock
);
2222 raw_spin_lock_init(&rdp
->nocb_gp_lock
);
2223 timer_setup(&rdp
->nocb_timer
, do_nocb_deferred_wakeup_timer
, 0);
2224 timer_setup(&rdp
->nocb_bypass_timer
, do_nocb_bypass_wakeup_timer
, 0);
2225 rcu_cblist_init(&rdp
->nocb_bypass
);
2229 * If the specified CPU is a no-CBs CPU that does not already have its
2230 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2231 * for this CPU's group has not yet been created, spawn it as well.
2233 static void rcu_spawn_one_nocb_kthread(int cpu
)
2235 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2236 struct rcu_data
*rdp_gp
;
2237 struct task_struct
*t
;
2240 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2241 * then nothing to do.
2243 if (!rcu_is_nocb_cpu(cpu
) || rdp
->nocb_cb_kthread
)
2246 /* If we didn't spawn the GP kthread first, reorganize! */
2247 rdp_gp
= rdp
->nocb_gp_rdp
;
2248 if (!rdp_gp
->nocb_gp_kthread
) {
2249 t
= kthread_run(rcu_nocb_gp_kthread
, rdp_gp
,
2250 "rcuog/%d", rdp_gp
->cpu
);
2251 if (WARN_ONCE(IS_ERR(t
), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__
))
2253 WRITE_ONCE(rdp_gp
->nocb_gp_kthread
, t
);
2256 /* Spawn the kthread for this CPU. */
2257 t
= kthread_run(rcu_nocb_cb_kthread
, rdp
,
2258 "rcuo%c/%d", rcu_state
.abbr
, cpu
);
2259 if (WARN_ONCE(IS_ERR(t
), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__
))
2261 WRITE_ONCE(rdp
->nocb_cb_kthread
, t
);
2262 WRITE_ONCE(rdp
->nocb_gp_kthread
, rdp_gp
->nocb_gp_kthread
);
2266 * If the specified CPU is a no-CBs CPU that does not already have its
2267 * rcuo kthread, spawn it.
2269 static void rcu_spawn_cpu_nocb_kthread(int cpu
)
2271 if (rcu_scheduler_fully_active
)
2272 rcu_spawn_one_nocb_kthread(cpu
);
2276 * Once the scheduler is running, spawn rcuo kthreads for all online
2277 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2278 * non-boot CPUs come online -- if this changes, we will need to add
2279 * some mutual exclusion.
2281 static void __init
rcu_spawn_nocb_kthreads(void)
2285 for_each_online_cpu(cpu
)
2286 rcu_spawn_cpu_nocb_kthread(cpu
);
2289 /* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2290 static int rcu_nocb_gp_stride
= -1;
2291 module_param(rcu_nocb_gp_stride
, int, 0444);
2294 * Initialize GP-CB relationships for all no-CBs CPU.
2296 static void __init
rcu_organize_nocb_kthreads(void)
2299 bool firsttime
= true;
2300 bool gotnocbs
= false;
2301 bool gotnocbscbs
= true;
2302 int ls
= rcu_nocb_gp_stride
;
2303 int nl
= 0; /* Next GP kthread. */
2304 struct rcu_data
*rdp
;
2305 struct rcu_data
*rdp_gp
= NULL
; /* Suppress misguided gcc warn. */
2306 struct rcu_data
*rdp_prev
= NULL
;
2308 if (!cpumask_available(rcu_nocb_mask
))
2311 ls
= nr_cpu_ids
/ int_sqrt(nr_cpu_ids
);
2312 rcu_nocb_gp_stride
= ls
;
2316 * Each pass through this loop sets up one rcu_data structure.
2317 * Should the corresponding CPU come online in the future, then
2318 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2320 for_each_cpu(cpu
, rcu_nocb_mask
) {
2321 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2322 if (rdp
->cpu
>= nl
) {
2323 /* New GP kthread, set up for CBs & next GP. */
2325 nl
= DIV_ROUND_UP(rdp
->cpu
+ 1, ls
) * ls
;
2326 rdp
->nocb_gp_rdp
= rdp
;
2330 pr_cont("%s\n", gotnocbscbs
2331 ? "" : " (self only)");
2332 gotnocbscbs
= false;
2334 pr_alert("%s: No-CB GP kthread CPU %d:",
2338 /* Another CB kthread, link to previous GP kthread. */
2340 rdp
->nocb_gp_rdp
= rdp_gp
;
2341 rdp_prev
->nocb_next_cb_rdp
= rdp
;
2343 pr_cont(" %d", cpu
);
2347 if (gotnocbs
&& dump_tree
)
2348 pr_cont("%s\n", gotnocbscbs
? "" : " (self only)");
2352 * Bind the current task to the offloaded CPUs. If there are no offloaded
2353 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2355 void rcu_bind_current_to_nocb(void)
2357 if (cpumask_available(rcu_nocb_mask
) && cpumask_weight(rcu_nocb_mask
))
2358 WARN_ON(sched_setaffinity(current
->pid
, rcu_nocb_mask
));
2360 EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb
);
2363 * Dump out nocb grace-period kthread state for the specified rcu_data
2366 static void show_rcu_nocb_gp_state(struct rcu_data
*rdp
)
2368 struct rcu_node
*rnp
= rdp
->mynode
;
2370 pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2372 "kK"[!!rdp
->nocb_gp_kthread
],
2373 "lL"[raw_spin_is_locked(&rdp
->nocb_gp_lock
)],
2374 "dD"[!!rdp
->nocb_defer_wakeup
],
2375 "tT"[timer_pending(&rdp
->nocb_timer
)],
2376 "bB"[timer_pending(&rdp
->nocb_bypass_timer
)],
2377 "sS"[!!rdp
->nocb_gp_sleep
],
2378 ".W"[swait_active(&rdp
->nocb_gp_wq
)],
2379 ".W"[swait_active(&rnp
->nocb_gp_wq
[0])],
2380 ".W"[swait_active(&rnp
->nocb_gp_wq
[1])],
2381 ".B"[!!rdp
->nocb_gp_bypass
],
2382 ".G"[!!rdp
->nocb_gp_gp
],
2383 (long)rdp
->nocb_gp_seq
,
2384 rnp
->grplo
, rnp
->grphi
, READ_ONCE(rdp
->nocb_gp_loops
));
2387 /* Dump out nocb kthread state for the specified rcu_data structure. */
2388 static void show_rcu_nocb_state(struct rcu_data
*rdp
)
2390 struct rcu_segcblist
*rsclp
= &rdp
->cblist
;
2395 if (rdp
->nocb_gp_rdp
== rdp
)
2396 show_rcu_nocb_gp_state(rdp
);
2398 pr_info(" CB %d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%c%c%c q%ld\n",
2399 rdp
->cpu
, rdp
->nocb_gp_rdp
->cpu
,
2400 "kK"[!!rdp
->nocb_cb_kthread
],
2401 "bB"[raw_spin_is_locked(&rdp
->nocb_bypass_lock
)],
2402 "cC"[!!atomic_read(&rdp
->nocb_lock_contended
)],
2403 "lL"[raw_spin_is_locked(&rdp
->nocb_lock
)],
2404 "sS"[!!rdp
->nocb_cb_sleep
],
2405 ".W"[swait_active(&rdp
->nocb_cb_wq
)],
2406 jiffies
- rdp
->nocb_bypass_first
,
2407 jiffies
- rdp
->nocb_nobypass_last
,
2408 rdp
->nocb_nobypass_count
,
2409 ".D"[rcu_segcblist_ready_cbs(rsclp
)],
2410 ".W"[!rcu_segcblist_restempty(rsclp
, RCU_DONE_TAIL
)],
2411 ".R"[!rcu_segcblist_restempty(rsclp
, RCU_WAIT_TAIL
)],
2412 ".N"[!rcu_segcblist_restempty(rsclp
, RCU_NEXT_READY_TAIL
)],
2413 ".B"[!!rcu_cblist_n_cbs(&rdp
->nocb_bypass
)],
2414 rcu_segcblist_n_cbs(&rdp
->cblist
));
2416 /* It is OK for GP kthreads to have GP state. */
2417 if (rdp
->nocb_gp_rdp
== rdp
)
2420 waslocked
= raw_spin_is_locked(&rdp
->nocb_gp_lock
);
2421 wastimer
= timer_pending(&rdp
->nocb_timer
);
2422 wassleep
= swait_active(&rdp
->nocb_gp_wq
);
2423 if (!rdp
->nocb_defer_wakeup
&& !rdp
->nocb_gp_sleep
&&
2424 !waslocked
&& !wastimer
&& !wassleep
)
2425 return; /* Nothing untowards. */
2427 pr_info(" !!! %c%c%c%c %c\n",
2429 "dD"[!!rdp
->nocb_defer_wakeup
],
2431 "sS"[!!rdp
->nocb_gp_sleep
],
2435 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2437 /* No ->nocb_lock to acquire. */
2438 static void rcu_nocb_lock(struct rcu_data
*rdp
)
2442 /* No ->nocb_lock to release. */
2443 static void rcu_nocb_unlock(struct rcu_data
*rdp
)
2447 /* No ->nocb_lock to release. */
2448 static void rcu_nocb_unlock_irqrestore(struct rcu_data
*rdp
,
2449 unsigned long flags
)
2451 local_irq_restore(flags
);
2454 /* Lockdep check that ->cblist may be safely accessed. */
2455 static void rcu_lockdep_assert_cblist_protected(struct rcu_data
*rdp
)
2457 lockdep_assert_irqs_disabled();
2460 static void rcu_nocb_gp_cleanup(struct swait_queue_head
*sq
)
2464 static struct swait_queue_head
*rcu_nocb_gp_get(struct rcu_node
*rnp
)
2469 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2473 static bool rcu_nocb_flush_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2479 static bool rcu_nocb_try_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2480 bool *was_alldone
, unsigned long flags
)
2485 static void __call_rcu_nocb_wake(struct rcu_data
*rdp
, bool was_empty
,
2486 unsigned long flags
)
2488 WARN_ON_ONCE(1); /* Should be dead code! */
2491 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2495 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2500 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2504 static void rcu_spawn_cpu_nocb_kthread(int cpu
)
2508 static void __init
rcu_spawn_nocb_kthreads(void)
2512 static void show_rcu_nocb_state(struct rcu_data
*rdp
)
2516 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2519 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2520 * grace-period kthread will do force_quiescent_state() processing?
2521 * The idea is to avoid waking up RCU core processing on such a
2522 * CPU unless the grace period has extended for too long.
2524 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2525 * CONFIG_RCU_NOCB_CPU CPUs.
2527 static bool rcu_nohz_full_cpu(void)
2529 #ifdef CONFIG_NO_HZ_FULL
2530 if (tick_nohz_full_cpu(smp_processor_id()) &&
2531 (!rcu_gp_in_progress() ||
2532 ULONG_CMP_LT(jiffies
, READ_ONCE(rcu_state
.gp_start
) + HZ
)))
2534 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2539 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2541 static void rcu_bind_gp_kthread(void)
2543 if (!tick_nohz_full_enabled())
2545 housekeeping_affine(current
, HK_FLAG_RCU
);
2548 /* Record the current task on dyntick-idle entry. */
2549 static void rcu_dynticks_task_enter(void)
2551 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2552 WRITE_ONCE(current
->rcu_tasks_idle_cpu
, smp_processor_id());
2553 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2556 /* Record no current task on dyntick-idle exit. */
2557 static void rcu_dynticks_task_exit(void)
2559 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2560 WRITE_ONCE(current
->rcu_tasks_idle_cpu
, -1);
2561 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */