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
)
572 bool couldrecurse
= rcu_preempt_depth() >= 0;
574 if (!rcu_preempt_need_deferred_qs(t
))
577 rcu_preempt_depth_set(rcu_preempt_depth() - RCU_NEST_BIAS
);
578 local_irq_save(flags
);
579 rcu_preempt_deferred_qs_irqrestore(t
, flags
);
581 rcu_preempt_depth_set(rcu_preempt_depth() + RCU_NEST_BIAS
);
585 * Minimal handler to give the scheduler a chance to re-evaluate.
587 static void rcu_preempt_deferred_qs_handler(struct irq_work
*iwp
)
589 struct rcu_data
*rdp
;
591 rdp
= container_of(iwp
, struct rcu_data
, defer_qs_iw
);
592 rdp
->defer_qs_iw_pending
= false;
596 * Handle special cases during rcu_read_unlock(), such as needing to
597 * notify RCU core processing or task having blocked during the RCU
598 * read-side critical section.
600 static void rcu_read_unlock_special(struct task_struct
*t
)
603 bool preempt_bh_were_disabled
=
604 !!(preempt_count() & (PREEMPT_MASK
| SOFTIRQ_MASK
));
605 bool irqs_were_disabled
;
607 /* NMI handlers cannot block and cannot safely manipulate state. */
611 local_irq_save(flags
);
612 irqs_were_disabled
= irqs_disabled_flags(flags
);
613 if (preempt_bh_were_disabled
|| irqs_were_disabled
) {
615 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
616 struct rcu_node
*rnp
= rdp
->mynode
;
618 exp
= (t
->rcu_blocked_node
&& t
->rcu_blocked_node
->exp_tasks
) ||
619 (rdp
->grpmask
& READ_ONCE(rnp
->expmask
)) ||
620 tick_nohz_full_cpu(rdp
->cpu
);
621 // Need to defer quiescent state until everything is enabled.
622 if (irqs_were_disabled
&& use_softirq
&&
624 (exp
&& !t
->rcu_read_unlock_special
.b
.deferred_qs
))) {
625 // Using softirq, safe to awaken, and we get
626 // no help from enabling irqs, unlike bh/preempt.
627 raise_softirq_irqoff(RCU_SOFTIRQ
);
629 // Enabling BH or preempt does reschedule, so...
630 // Also if no expediting or NO_HZ_FULL, slow is OK.
631 set_tsk_need_resched(current
);
632 set_preempt_need_resched();
633 if (IS_ENABLED(CONFIG_IRQ_WORK
) && irqs_were_disabled
&&
634 !rdp
->defer_qs_iw_pending
&& exp
) {
635 // Get scheduler to re-evaluate and call hooks.
636 // If !IRQ_WORK, FQS scan will eventually IPI.
637 init_irq_work(&rdp
->defer_qs_iw
,
638 rcu_preempt_deferred_qs_handler
);
639 rdp
->defer_qs_iw_pending
= true;
640 irq_work_queue_on(&rdp
->defer_qs_iw
, rdp
->cpu
);
643 t
->rcu_read_unlock_special
.b
.deferred_qs
= true;
644 local_irq_restore(flags
);
647 rcu_preempt_deferred_qs_irqrestore(t
, flags
);
651 * Check that the list of blocked tasks for the newly completed grace
652 * period is in fact empty. It is a serious bug to complete a grace
653 * period that still has RCU readers blocked! This function must be
654 * invoked -before- updating this rnp's ->gp_seq.
656 * Also, if there are blocked tasks on the list, they automatically
657 * block the newly created grace period, so set up ->gp_tasks accordingly.
659 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
661 struct task_struct
*t
;
663 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
664 raw_lockdep_assert_held_rcu_node(rnp
);
665 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
)))
666 dump_blkd_tasks(rnp
, 10);
667 if (rcu_preempt_has_tasks(rnp
) &&
668 (rnp
->qsmaskinit
|| rnp
->wait_blkd_tasks
)) {
669 WRITE_ONCE(rnp
->gp_tasks
, rnp
->blkd_tasks
.next
);
670 t
= container_of(rnp
->gp_tasks
, struct task_struct
,
672 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
673 rnp
->gp_seq
, t
->pid
);
675 WARN_ON_ONCE(rnp
->qsmask
);
679 * Check for a quiescent state from the current CPU, including voluntary
680 * context switches for Tasks RCU. When a task blocks, the task is
681 * recorded in the corresponding CPU's rcu_node structure, which is checked
682 * elsewhere, hence this function need only check for quiescent states
683 * related to the current CPU, not to those related to tasks.
685 static void rcu_flavor_sched_clock_irq(int user
)
687 struct task_struct
*t
= current
;
689 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
690 rcu_note_voluntary_context_switch(current
);
692 if (rcu_preempt_depth() > 0 ||
693 (preempt_count() & (PREEMPT_MASK
| SOFTIRQ_MASK
))) {
694 /* No QS, force context switch if deferred. */
695 if (rcu_preempt_need_deferred_qs(t
)) {
696 set_tsk_need_resched(t
);
697 set_preempt_need_resched();
699 } else if (rcu_preempt_need_deferred_qs(t
)) {
700 rcu_preempt_deferred_qs(t
); /* Report deferred QS. */
702 } else if (!rcu_preempt_depth()) {
703 rcu_qs(); /* Report immediate QS. */
707 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
708 if (rcu_preempt_depth() > 0 &&
709 __this_cpu_read(rcu_data
.core_needs_qs
) &&
710 __this_cpu_read(rcu_data
.cpu_no_qs
.b
.norm
) &&
711 !t
->rcu_read_unlock_special
.b
.need_qs
&&
712 time_after(jiffies
, rcu_state
.gp_start
+ HZ
))
713 t
->rcu_read_unlock_special
.b
.need_qs
= true;
717 * Check for a task exiting while in a preemptible-RCU read-side
718 * critical section, clean up if so. No need to issue warnings, as
719 * debug_check_no_locks_held() already does this if lockdep is enabled.
720 * Besides, if this function does anything other than just immediately
721 * return, there was a bug of some sort. Spewing warnings from this
722 * function is like as not to simply obscure important prior warnings.
726 struct task_struct
*t
= current
;
728 if (unlikely(!list_empty(¤t
->rcu_node_entry
))) {
729 rcu_preempt_depth_set(1);
731 WRITE_ONCE(t
->rcu_read_unlock_special
.b
.blocked
, true);
732 } else if (unlikely(rcu_preempt_depth())) {
733 rcu_preempt_depth_set(1);
738 rcu_preempt_deferred_qs(current
);
742 * Dump the blocked-tasks state, but limit the list dump to the
743 * specified number of elements.
746 dump_blkd_tasks(struct rcu_node
*rnp
, int ncheck
)
750 struct list_head
*lhp
;
752 struct rcu_data
*rdp
;
753 struct rcu_node
*rnp1
;
755 raw_lockdep_assert_held_rcu_node(rnp
);
756 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
757 __func__
, rnp
->grplo
, rnp
->grphi
, rnp
->level
,
758 (long)READ_ONCE(rnp
->gp_seq
), (long)rnp
->completedqs
);
759 for (rnp1
= rnp
; rnp1
; rnp1
= rnp1
->parent
)
760 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
761 __func__
, rnp1
->grplo
, rnp1
->grphi
, rnp1
->qsmask
, rnp1
->qsmaskinit
, rnp1
->qsmaskinitnext
);
762 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
763 __func__
, READ_ONCE(rnp
->gp_tasks
), rnp
->boost_tasks
,
765 pr_info("%s: ->blkd_tasks", __func__
);
767 list_for_each(lhp
, &rnp
->blkd_tasks
) {
773 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++) {
774 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
775 onl
= !!(rdp
->grpmask
& rcu_rnp_online_cpus(rnp
));
776 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
778 (long)rdp
->rcu_onl_gp_seq
, rdp
->rcu_onl_gp_flags
,
779 (long)rdp
->rcu_ofl_gp_seq
, rdp
->rcu_ofl_gp_flags
);
783 #else /* #ifdef CONFIG_PREEMPT_RCU */
786 * Tell them what RCU they are running.
788 static void __init
rcu_bootup_announce(void)
790 pr_info("Hierarchical RCU implementation.\n");
791 rcu_bootup_announce_oddness();
795 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
796 * how many quiescent states passed, just if there was at least one since
797 * the start of the grace period, this just sets a flag. The caller must
798 * have disabled preemption.
800 static void rcu_qs(void)
802 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
803 if (!__this_cpu_read(rcu_data
.cpu_no_qs
.s
))
805 trace_rcu_grace_period(TPS("rcu_sched"),
806 __this_cpu_read(rcu_data
.gp_seq
), TPS("cpuqs"));
807 __this_cpu_write(rcu_data
.cpu_no_qs
.b
.norm
, false);
808 if (!__this_cpu_read(rcu_data
.cpu_no_qs
.b
.exp
))
810 __this_cpu_write(rcu_data
.cpu_no_qs
.b
.exp
, false);
811 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data
));
815 * Register an urgently needed quiescent state. If there is an
816 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
817 * dyntick-idle quiescent state visible to other CPUs, which will in
818 * some cases serve for expedited as well as normal grace periods.
819 * Either way, register a lightweight quiescent state.
821 void rcu_all_qs(void)
825 if (!raw_cpu_read(rcu_data
.rcu_urgent_qs
))
828 /* Load rcu_urgent_qs before other flags. */
829 if (!smp_load_acquire(this_cpu_ptr(&rcu_data
.rcu_urgent_qs
))) {
833 this_cpu_write(rcu_data
.rcu_urgent_qs
, false);
834 if (unlikely(raw_cpu_read(rcu_data
.rcu_need_heavy_qs
))) {
835 local_irq_save(flags
);
836 rcu_momentary_dyntick_idle();
837 local_irq_restore(flags
);
842 EXPORT_SYMBOL_GPL(rcu_all_qs
);
845 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
847 void rcu_note_context_switch(bool preempt
)
849 trace_rcu_utilization(TPS("Start context switch"));
851 /* Load rcu_urgent_qs before other flags. */
852 if (!smp_load_acquire(this_cpu_ptr(&rcu_data
.rcu_urgent_qs
)))
854 this_cpu_write(rcu_data
.rcu_urgent_qs
, false);
855 if (unlikely(raw_cpu_read(rcu_data
.rcu_need_heavy_qs
)))
856 rcu_momentary_dyntick_idle();
858 rcu_tasks_qs(current
);
860 trace_rcu_utilization(TPS("End context switch"));
862 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
865 * Because preemptible RCU does not exist, there are never any preempted
868 static int rcu_preempt_blocked_readers_cgp(struct rcu_node
*rnp
)
874 * Because there is no preemptible RCU, there can be no readers blocked.
876 static bool rcu_preempt_has_tasks(struct rcu_node
*rnp
)
882 * Because there is no preemptible RCU, there can be no deferred quiescent
885 static bool rcu_preempt_need_deferred_qs(struct task_struct
*t
)
889 static void rcu_preempt_deferred_qs(struct task_struct
*t
) { }
892 * Because there is no preemptible RCU, there can be no readers blocked,
893 * so there is no need to check for blocked tasks. So check only for
894 * bogus qsmask values.
896 static void rcu_preempt_check_blocked_tasks(struct rcu_node
*rnp
)
898 WARN_ON_ONCE(rnp
->qsmask
);
902 * Check to see if this CPU is in a non-context-switch quiescent state,
903 * namely user mode and idle loop.
905 static void rcu_flavor_sched_clock_irq(int user
)
907 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
910 * Get here if this CPU took its interrupt from user
911 * mode or from the idle loop, and if this is not a
912 * nested interrupt. In this case, the CPU is in
913 * a quiescent state, so note it.
915 * No memory barrier is required here because rcu_qs()
916 * references only CPU-local variables that other CPUs
917 * neither access nor modify, at least not while the
918 * corresponding CPU is online.
926 * Because preemptible RCU does not exist, tasks cannot possibly exit
927 * while in preemptible RCU read-side critical sections.
934 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
937 dump_blkd_tasks(struct rcu_node
*rnp
, int ncheck
)
939 WARN_ON_ONCE(!list_empty(&rnp
->blkd_tasks
));
942 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
945 * If boosting, set rcuc kthreads to realtime priority.
947 static void rcu_cpu_kthread_setup(unsigned int cpu
)
949 #ifdef CONFIG_RCU_BOOST
950 struct sched_param sp
;
952 sp
.sched_priority
= kthread_prio
;
953 sched_setscheduler_nocheck(current
, SCHED_FIFO
, &sp
);
954 #endif /* #ifdef CONFIG_RCU_BOOST */
957 #ifdef CONFIG_RCU_BOOST
960 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
961 * or ->boost_tasks, advancing the pointer to the next task in the
964 * Note that irqs must be enabled: boosting the task can block.
965 * Returns 1 if there are more tasks needing to be boosted.
967 static int rcu_boost(struct rcu_node
*rnp
)
970 struct task_struct
*t
;
971 struct list_head
*tb
;
973 if (READ_ONCE(rnp
->exp_tasks
) == NULL
&&
974 READ_ONCE(rnp
->boost_tasks
) == NULL
)
975 return 0; /* Nothing left to boost. */
977 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
980 * Recheck under the lock: all tasks in need of boosting
981 * might exit their RCU read-side critical sections on their own.
983 if (rnp
->exp_tasks
== NULL
&& rnp
->boost_tasks
== NULL
) {
984 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
989 * Preferentially boost tasks blocking expedited grace periods.
990 * This cannot starve the normal grace periods because a second
991 * expedited grace period must boost all blocked tasks, including
992 * those blocking the pre-existing normal grace period.
994 if (rnp
->exp_tasks
!= NULL
)
997 tb
= rnp
->boost_tasks
;
1000 * We boost task t by manufacturing an rt_mutex that appears to
1001 * be held by task t. We leave a pointer to that rt_mutex where
1002 * task t can find it, and task t will release the mutex when it
1003 * exits its outermost RCU read-side critical section. Then
1004 * simply acquiring this artificial rt_mutex will boost task
1005 * t's priority. (Thanks to tglx for suggesting this approach!)
1007 * Note that task t must acquire rnp->lock to remove itself from
1008 * the ->blkd_tasks list, which it will do from exit() if from
1009 * nowhere else. We therefore are guaranteed that task t will
1010 * stay around at least until we drop rnp->lock. Note that
1011 * rnp->lock also resolves races between our priority boosting
1012 * and task t's exiting its outermost RCU read-side critical
1015 t
= container_of(tb
, struct task_struct
, rcu_node_entry
);
1016 rt_mutex_init_proxy_locked(&rnp
->boost_mtx
, t
);
1017 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1018 /* Lock only for side effect: boosts task t's priority. */
1019 rt_mutex_lock(&rnp
->boost_mtx
);
1020 rt_mutex_unlock(&rnp
->boost_mtx
); /* Then keep lockdep happy. */
1022 return READ_ONCE(rnp
->exp_tasks
) != NULL
||
1023 READ_ONCE(rnp
->boost_tasks
) != NULL
;
1027 * Priority-boosting kthread, one per leaf rcu_node.
1029 static int rcu_boost_kthread(void *arg
)
1031 struct rcu_node
*rnp
= (struct rcu_node
*)arg
;
1035 trace_rcu_utilization(TPS("Start boost kthread@init"));
1037 WRITE_ONCE(rnp
->boost_kthread_status
, RCU_KTHREAD_WAITING
);
1038 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1039 rcu_wait(rnp
->boost_tasks
|| rnp
->exp_tasks
);
1040 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1041 WRITE_ONCE(rnp
->boost_kthread_status
, RCU_KTHREAD_RUNNING
);
1042 more2boost
= rcu_boost(rnp
);
1048 WRITE_ONCE(rnp
->boost_kthread_status
, RCU_KTHREAD_YIELDING
);
1049 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1050 schedule_timeout_interruptible(2);
1051 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1056 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1061 * Check to see if it is time to start boosting RCU readers that are
1062 * blocking the current grace period, and, if so, tell the per-rcu_node
1063 * kthread to start boosting them. If there is an expedited grace
1064 * period in progress, it is always time to boost.
1066 * The caller must hold rnp->lock, which this function releases.
1067 * The ->boost_kthread_task is immortal, so we don't need to worry
1068 * about it going away.
1070 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1071 __releases(rnp
->lock
)
1073 raw_lockdep_assert_held_rcu_node(rnp
);
1074 if (!rcu_preempt_blocked_readers_cgp(rnp
) && rnp
->exp_tasks
== NULL
) {
1075 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1078 if (rnp
->exp_tasks
!= NULL
||
1079 (rnp
->gp_tasks
!= NULL
&&
1080 rnp
->boost_tasks
== NULL
&&
1082 (ULONG_CMP_GE(jiffies
, rnp
->boost_time
) || rcu_state
.cbovld
))) {
1083 if (rnp
->exp_tasks
== NULL
)
1084 rnp
->boost_tasks
= rnp
->gp_tasks
;
1085 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1086 rcu_wake_cond(rnp
->boost_kthread_task
,
1087 READ_ONCE(rnp
->boost_kthread_status
));
1089 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1094 * Is the current CPU running the RCU-callbacks kthread?
1095 * Caller must have preemption disabled.
1097 static bool rcu_is_callbacks_kthread(void)
1099 return __this_cpu_read(rcu_data
.rcu_cpu_kthread_task
) == current
;
1102 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1105 * Do priority-boost accounting for the start of a new grace period.
1107 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1109 rnp
->boost_time
= jiffies
+ RCU_BOOST_DELAY_JIFFIES
;
1113 * Create an RCU-boost kthread for the specified node if one does not
1114 * already exist. We only create this kthread for preemptible RCU.
1115 * Returns zero if all is well, a negated errno otherwise.
1117 static void rcu_spawn_one_boost_kthread(struct rcu_node
*rnp
)
1119 int rnp_index
= rnp
- rcu_get_root();
1120 unsigned long flags
;
1121 struct sched_param sp
;
1122 struct task_struct
*t
;
1124 if (!IS_ENABLED(CONFIG_PREEMPT_RCU
))
1127 if (!rcu_scheduler_fully_active
|| rcu_rnp_online_cpus(rnp
) == 0)
1130 rcu_state
.boost
= 1;
1132 if (rnp
->boost_kthread_task
!= NULL
)
1135 t
= kthread_create(rcu_boost_kthread
, (void *)rnp
,
1136 "rcub/%d", rnp_index
);
1137 if (WARN_ON_ONCE(IS_ERR(t
)))
1140 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1141 rnp
->boost_kthread_task
= t
;
1142 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1143 sp
.sched_priority
= kthread_prio
;
1144 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
1145 wake_up_process(t
); /* get to TASK_INTERRUPTIBLE quickly. */
1149 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1150 * served by the rcu_node in question. The CPU hotplug lock is still
1151 * held, so the value of rnp->qsmaskinit will be stable.
1153 * We don't include outgoingcpu in the affinity set, use -1 if there is
1154 * no outgoing CPU. If there are no CPUs left in the affinity set,
1155 * this function allows the kthread to execute on any CPU.
1157 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1159 struct task_struct
*t
= rnp
->boost_kthread_task
;
1160 unsigned long mask
= rcu_rnp_online_cpus(rnp
);
1166 if (!zalloc_cpumask_var(&cm
, GFP_KERNEL
))
1168 for_each_leaf_node_possible_cpu(rnp
, cpu
)
1169 if ((mask
& leaf_node_cpu_bit(rnp
, cpu
)) &&
1171 cpumask_set_cpu(cpu
, cm
);
1172 if (cpumask_weight(cm
) == 0)
1174 set_cpus_allowed_ptr(t
, cm
);
1175 free_cpumask_var(cm
);
1179 * Spawn boost kthreads -- called as soon as the scheduler is running.
1181 static void __init
rcu_spawn_boost_kthreads(void)
1183 struct rcu_node
*rnp
;
1185 rcu_for_each_leaf_node(rnp
)
1186 rcu_spawn_one_boost_kthread(rnp
);
1189 static void rcu_prepare_kthreads(int cpu
)
1191 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
1192 struct rcu_node
*rnp
= rdp
->mynode
;
1194 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1195 if (rcu_scheduler_fully_active
)
1196 rcu_spawn_one_boost_kthread(rnp
);
1199 #else /* #ifdef CONFIG_RCU_BOOST */
1201 static void rcu_initiate_boost(struct rcu_node
*rnp
, unsigned long flags
)
1202 __releases(rnp
->lock
)
1204 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1207 static bool rcu_is_callbacks_kthread(void)
1212 static void rcu_preempt_boost_start_gp(struct rcu_node
*rnp
)
1216 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
)
1220 static void __init
rcu_spawn_boost_kthreads(void)
1224 static void rcu_prepare_kthreads(int cpu
)
1228 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1230 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1233 * Check to see if any future non-offloaded RCU-related work will need
1234 * to be done by the current CPU, even if none need be done immediately,
1235 * returning 1 if so. This function is part of the RCU implementation;
1236 * it is -not- an exported member of the RCU API.
1238 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1239 * CPU has RCU callbacks queued.
1241 int rcu_needs_cpu(u64 basemono
, u64
*nextevt
)
1243 *nextevt
= KTIME_MAX
;
1244 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data
)->cblist
) &&
1245 !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data
)->cblist
);
1249 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1252 static void rcu_cleanup_after_idle(void)
1257 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1260 static void rcu_prepare_for_idle(void)
1264 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1267 * This code is invoked when a CPU goes idle, at which point we want
1268 * to have the CPU do everything required for RCU so that it can enter
1269 * the energy-efficient dyntick-idle mode.
1271 * The following preprocessor symbol controls this:
1273 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1274 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1275 * is sized to be roughly one RCU grace period. Those energy-efficiency
1276 * benchmarkers who might otherwise be tempted to set this to a large
1277 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1278 * system. And if you are -that- concerned about energy efficiency,
1279 * just power the system down and be done with it!
1281 * The value below works well in practice. If future workloads require
1282 * adjustment, they can be converted into kernel config parameters, though
1283 * making the state machine smarter might be a better option.
1285 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1287 static int rcu_idle_gp_delay
= RCU_IDLE_GP_DELAY
;
1288 module_param(rcu_idle_gp_delay
, int, 0644);
1291 * Try to advance callbacks on the current CPU, but only if it has been
1292 * awhile since the last time we did so. Afterwards, if there are any
1293 * callbacks ready for immediate invocation, return true.
1295 static bool __maybe_unused
rcu_try_advance_all_cbs(void)
1297 bool cbs_ready
= false;
1298 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1299 struct rcu_node
*rnp
;
1301 /* Exit early if we advanced recently. */
1302 if (jiffies
== rdp
->last_advance_all
)
1304 rdp
->last_advance_all
= jiffies
;
1309 * Don't bother checking unless a grace period has
1310 * completed since we last checked and there are
1311 * callbacks not yet ready to invoke.
1313 if ((rcu_seq_completed_gp(rdp
->gp_seq
,
1314 rcu_seq_current(&rnp
->gp_seq
)) ||
1315 unlikely(READ_ONCE(rdp
->gpwrap
))) &&
1316 rcu_segcblist_pend_cbs(&rdp
->cblist
))
1317 note_gp_changes(rdp
);
1319 if (rcu_segcblist_ready_cbs(&rdp
->cblist
))
1325 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1326 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1327 * caller about what to set the timeout.
1329 * The caller must have disabled interrupts.
1331 int rcu_needs_cpu(u64 basemono
, u64
*nextevt
)
1333 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1336 lockdep_assert_irqs_disabled();
1338 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1339 if (rcu_segcblist_empty(&rdp
->cblist
) ||
1340 rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data
)->cblist
)) {
1341 *nextevt
= KTIME_MAX
;
1345 /* Attempt to advance callbacks. */
1346 if (rcu_try_advance_all_cbs()) {
1347 /* Some ready to invoke, so initiate later invocation. */
1351 rdp
->last_accelerate
= jiffies
;
1353 /* Request timer and round. */
1354 dj
= round_up(rcu_idle_gp_delay
+ jiffies
, rcu_idle_gp_delay
) - jiffies
;
1356 *nextevt
= basemono
+ dj
* TICK_NSEC
;
1361 * Prepare a CPU for idle from an RCU perspective. The first major task is to
1362 * sense whether nohz mode has been enabled or disabled via sysfs. The second
1363 * major task is to accelerate (that is, assign grace-period numbers to) any
1364 * recently arrived callbacks.
1366 * The caller must have disabled interrupts.
1368 static void rcu_prepare_for_idle(void)
1371 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1372 struct rcu_node
*rnp
;
1375 lockdep_assert_irqs_disabled();
1376 if (rcu_segcblist_is_offloaded(&rdp
->cblist
))
1379 /* Handle nohz enablement switches conservatively. */
1380 tne
= READ_ONCE(tick_nohz_active
);
1381 if (tne
!= rdp
->tick_nohz_enabled_snap
) {
1382 if (!rcu_segcblist_empty(&rdp
->cblist
))
1383 invoke_rcu_core(); /* force nohz to see update. */
1384 rdp
->tick_nohz_enabled_snap
= tne
;
1391 * If we have not yet accelerated this jiffy, accelerate all
1392 * callbacks on this CPU.
1394 if (rdp
->last_accelerate
== jiffies
)
1396 rdp
->last_accelerate
= jiffies
;
1397 if (rcu_segcblist_pend_cbs(&rdp
->cblist
)) {
1399 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
1400 needwake
= rcu_accelerate_cbs(rnp
, rdp
);
1401 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
1403 rcu_gp_kthread_wake();
1408 * Clean up for exit from idle. Attempt to advance callbacks based on
1409 * any grace periods that elapsed while the CPU was idle, and if any
1410 * callbacks are now ready to invoke, initiate invocation.
1412 static void rcu_cleanup_after_idle(void)
1414 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1416 lockdep_assert_irqs_disabled();
1417 if (rcu_segcblist_is_offloaded(&rdp
->cblist
))
1419 if (rcu_try_advance_all_cbs())
1423 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1425 #ifdef CONFIG_RCU_NOCB_CPU
1428 * Offload callback processing from the boot-time-specified set of CPUs
1429 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1430 * created that pull the callbacks from the corresponding CPU, wait for
1431 * a grace period to elapse, and invoke the callbacks. These kthreads
1432 * are organized into GP kthreads, which manage incoming callbacks, wait for
1433 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1434 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1435 * do a wake_up() on their GP kthread when they insert a callback into any
1436 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1437 * in which case each kthread actively polls its CPU. (Which isn't so great
1438 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1440 * This is intended to be used in conjunction with Frederic Weisbecker's
1441 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1442 * running CPU-bound user-mode computations.
1444 * Offloading of callbacks can also be used as an energy-efficiency
1445 * measure because CPUs with no RCU callbacks queued are more aggressive
1446 * about entering dyntick-idle mode.
1451 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1452 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1453 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1454 * given, a warning is emitted and all CPUs are offloaded.
1456 static int __init
rcu_nocb_setup(char *str
)
1458 alloc_bootmem_cpumask_var(&rcu_nocb_mask
);
1459 if (!strcasecmp(str
, "all"))
1460 cpumask_setall(rcu_nocb_mask
);
1462 if (cpulist_parse(str
, rcu_nocb_mask
)) {
1463 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1464 cpumask_setall(rcu_nocb_mask
);
1468 __setup("rcu_nocbs=", rcu_nocb_setup
);
1470 static int __init
parse_rcu_nocb_poll(char *arg
)
1472 rcu_nocb_poll
= true;
1475 early_param("rcu_nocb_poll", parse_rcu_nocb_poll
);
1478 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1479 * After all, the main point of bypassing is to avoid lock contention
1480 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1482 int nocb_nobypass_lim_per_jiffy
= 16 * 1000 / HZ
;
1483 module_param(nocb_nobypass_lim_per_jiffy
, int, 0);
1486 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1487 * lock isn't immediately available, increment ->nocb_lock_contended to
1488 * flag the contention.
1490 static void rcu_nocb_bypass_lock(struct rcu_data
*rdp
)
1491 __acquires(&rdp
->nocb_bypass_lock
)
1493 lockdep_assert_irqs_disabled();
1494 if (raw_spin_trylock(&rdp
->nocb_bypass_lock
))
1496 atomic_inc(&rdp
->nocb_lock_contended
);
1497 WARN_ON_ONCE(smp_processor_id() != rdp
->cpu
);
1498 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1499 raw_spin_lock(&rdp
->nocb_bypass_lock
);
1500 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1501 atomic_dec(&rdp
->nocb_lock_contended
);
1505 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1506 * not contended. Please note that this is extremely special-purpose,
1507 * relying on the fact that at most two kthreads and one CPU contend for
1508 * this lock, and also that the two kthreads are guaranteed to have frequent
1509 * grace-period-duration time intervals between successive acquisitions
1510 * of the lock. This allows us to use an extremely simple throttling
1511 * mechanism, and further to apply it only to the CPU doing floods of
1512 * call_rcu() invocations. Don't try this at home!
1514 static void rcu_nocb_wait_contended(struct rcu_data
*rdp
)
1516 WARN_ON_ONCE(smp_processor_id() != rdp
->cpu
);
1517 while (WARN_ON_ONCE(atomic_read(&rdp
->nocb_lock_contended
)))
1522 * Conditionally acquire the specified rcu_data structure's
1523 * ->nocb_bypass_lock.
1525 static bool rcu_nocb_bypass_trylock(struct rcu_data
*rdp
)
1527 lockdep_assert_irqs_disabled();
1528 return raw_spin_trylock(&rdp
->nocb_bypass_lock
);
1532 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1534 static void rcu_nocb_bypass_unlock(struct rcu_data
*rdp
)
1535 __releases(&rdp
->nocb_bypass_lock
)
1537 lockdep_assert_irqs_disabled();
1538 raw_spin_unlock(&rdp
->nocb_bypass_lock
);
1542 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1543 * if it corresponds to a no-CBs CPU.
1545 static void rcu_nocb_lock(struct rcu_data
*rdp
)
1547 lockdep_assert_irqs_disabled();
1548 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
))
1550 raw_spin_lock(&rdp
->nocb_lock
);
1554 * Release the specified rcu_data structure's ->nocb_lock, but only
1555 * if it corresponds to a no-CBs CPU.
1557 static void rcu_nocb_unlock(struct rcu_data
*rdp
)
1559 if (rcu_segcblist_is_offloaded(&rdp
->cblist
)) {
1560 lockdep_assert_irqs_disabled();
1561 raw_spin_unlock(&rdp
->nocb_lock
);
1566 * Release the specified rcu_data structure's ->nocb_lock and restore
1567 * interrupts, but only if it corresponds to a no-CBs CPU.
1569 static void rcu_nocb_unlock_irqrestore(struct rcu_data
*rdp
,
1570 unsigned long flags
)
1572 if (rcu_segcblist_is_offloaded(&rdp
->cblist
)) {
1573 lockdep_assert_irqs_disabled();
1574 raw_spin_unlock_irqrestore(&rdp
->nocb_lock
, flags
);
1576 local_irq_restore(flags
);
1580 /* Lockdep check that ->cblist may be safely accessed. */
1581 static void rcu_lockdep_assert_cblist_protected(struct rcu_data
*rdp
)
1583 lockdep_assert_irqs_disabled();
1584 if (rcu_segcblist_is_offloaded(&rdp
->cblist
))
1585 lockdep_assert_held(&rdp
->nocb_lock
);
1589 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1592 static void rcu_nocb_gp_cleanup(struct swait_queue_head
*sq
)
1597 static struct swait_queue_head
*rcu_nocb_gp_get(struct rcu_node
*rnp
)
1599 return &rnp
->nocb_gp_wq
[rcu_seq_ctr(rnp
->gp_seq
) & 0x1];
1602 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
1604 init_swait_queue_head(&rnp
->nocb_gp_wq
[0]);
1605 init_swait_queue_head(&rnp
->nocb_gp_wq
[1]);
1608 /* Is the specified CPU a no-CBs CPU? */
1609 bool rcu_is_nocb_cpu(int cpu
)
1611 if (cpumask_available(rcu_nocb_mask
))
1612 return cpumask_test_cpu(cpu
, rcu_nocb_mask
);
1617 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1618 * and this function releases it.
1620 static void wake_nocb_gp(struct rcu_data
*rdp
, bool force
,
1621 unsigned long flags
)
1622 __releases(rdp
->nocb_lock
)
1624 bool needwake
= false;
1625 struct rcu_data
*rdp_gp
= rdp
->nocb_gp_rdp
;
1627 lockdep_assert_held(&rdp
->nocb_lock
);
1628 if (!READ_ONCE(rdp_gp
->nocb_gp_kthread
)) {
1629 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1630 TPS("AlreadyAwake"));
1631 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1634 del_timer(&rdp
->nocb_timer
);
1635 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1636 raw_spin_lock_irqsave(&rdp_gp
->nocb_gp_lock
, flags
);
1637 if (force
|| READ_ONCE(rdp_gp
->nocb_gp_sleep
)) {
1638 WRITE_ONCE(rdp_gp
->nocb_gp_sleep
, false);
1640 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("DoWake"));
1642 raw_spin_unlock_irqrestore(&rdp_gp
->nocb_gp_lock
, flags
);
1644 wake_up_process(rdp_gp
->nocb_gp_kthread
);
1648 * Arrange to wake the GP kthread for this NOCB group at some future
1649 * time when it is safe to do so.
1651 static void wake_nocb_gp_defer(struct rcu_data
*rdp
, int waketype
,
1654 if (rdp
->nocb_defer_wakeup
== RCU_NOCB_WAKE_NOT
)
1655 mod_timer(&rdp
->nocb_timer
, jiffies
+ 1);
1656 if (rdp
->nocb_defer_wakeup
< waketype
)
1657 WRITE_ONCE(rdp
->nocb_defer_wakeup
, waketype
);
1658 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, reason
);
1662 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1663 * However, if there is a callback to be enqueued and if ->nocb_bypass
1664 * proves to be initially empty, just return false because the no-CB GP
1665 * kthread may need to be awakened in this case.
1667 * Note that this function always returns true if rhp is NULL.
1669 static bool rcu_nocb_do_flush_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
1672 struct rcu_cblist rcl
;
1674 WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp
->cblist
));
1675 rcu_lockdep_assert_cblist_protected(rdp
);
1676 lockdep_assert_held(&rdp
->nocb_bypass_lock
);
1677 if (rhp
&& !rcu_cblist_n_cbs(&rdp
->nocb_bypass
)) {
1678 raw_spin_unlock(&rdp
->nocb_bypass_lock
);
1681 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1683 rcu_segcblist_inc_len(&rdp
->cblist
); /* Must precede enqueue. */
1684 rcu_cblist_flush_enqueue(&rcl
, &rdp
->nocb_bypass
, rhp
);
1685 rcu_segcblist_insert_pend_cbs(&rdp
->cblist
, &rcl
);
1686 WRITE_ONCE(rdp
->nocb_bypass_first
, j
);
1687 rcu_nocb_bypass_unlock(rdp
);
1692 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1693 * However, if there is a callback to be enqueued and if ->nocb_bypass
1694 * proves to be initially empty, just return false because the no-CB GP
1695 * kthread may need to be awakened in this case.
1697 * Note that this function always returns true if rhp is NULL.
1699 static bool rcu_nocb_flush_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
1702 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
))
1704 rcu_lockdep_assert_cblist_protected(rdp
);
1705 rcu_nocb_bypass_lock(rdp
);
1706 return rcu_nocb_do_flush_bypass(rdp
, rhp
, j
);
1710 * If the ->nocb_bypass_lock is immediately available, flush the
1711 * ->nocb_bypass queue into ->cblist.
1713 static void rcu_nocb_try_flush_bypass(struct rcu_data
*rdp
, unsigned long j
)
1715 rcu_lockdep_assert_cblist_protected(rdp
);
1716 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
) ||
1717 !rcu_nocb_bypass_trylock(rdp
))
1719 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp
, NULL
, j
));
1723 * See whether it is appropriate to use the ->nocb_bypass list in order
1724 * to control contention on ->nocb_lock. A limited number of direct
1725 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1726 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1727 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1728 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1729 * used if ->cblist is empty, because otherwise callbacks can be stranded
1730 * on ->nocb_bypass because we cannot count on the current CPU ever again
1731 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1732 * non-empty, the corresponding no-CBs grace-period kthread must not be
1733 * in an indefinite sleep state.
1735 * Finally, it is not permitted to use the bypass during early boot,
1736 * as doing so would confuse the auto-initialization code. Besides
1737 * which, there is no point in worrying about lock contention while
1738 * there is only one CPU in operation.
1740 static bool rcu_nocb_try_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
1741 bool *was_alldone
, unsigned long flags
)
1744 unsigned long cur_gp_seq
;
1745 unsigned long j
= jiffies
;
1746 long ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1748 if (!rcu_segcblist_is_offloaded(&rdp
->cblist
)) {
1749 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1750 return false; /* Not offloaded, no bypassing. */
1752 lockdep_assert_irqs_disabled();
1754 // Don't use ->nocb_bypass during early boot.
1755 if (rcu_scheduler_active
!= RCU_SCHEDULER_RUNNING
) {
1757 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp
->nocb_bypass
));
1758 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1762 // If we have advanced to a new jiffy, reset counts to allow
1763 // moving back from ->nocb_bypass to ->cblist.
1764 if (j
== rdp
->nocb_nobypass_last
) {
1765 c
= rdp
->nocb_nobypass_count
+ 1;
1767 WRITE_ONCE(rdp
->nocb_nobypass_last
, j
);
1768 c
= rdp
->nocb_nobypass_count
- nocb_nobypass_lim_per_jiffy
;
1769 if (ULONG_CMP_LT(rdp
->nocb_nobypass_count
,
1770 nocb_nobypass_lim_per_jiffy
))
1772 else if (c
> nocb_nobypass_lim_per_jiffy
)
1773 c
= nocb_nobypass_lim_per_jiffy
;
1775 WRITE_ONCE(rdp
->nocb_nobypass_count
, c
);
1777 // If there hasn't yet been all that many ->cblist enqueues
1778 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1779 // ->nocb_bypass first.
1780 if (rdp
->nocb_nobypass_count
< nocb_nobypass_lim_per_jiffy
) {
1782 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1784 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1786 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp
, NULL
, j
));
1787 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp
->nocb_bypass
));
1788 return false; // Caller must enqueue the callback.
1791 // If ->nocb_bypass has been used too long or is too full,
1792 // flush ->nocb_bypass to ->cblist.
1793 if ((ncbs
&& j
!= READ_ONCE(rdp
->nocb_bypass_first
)) ||
1796 if (!rcu_nocb_flush_bypass(rdp
, rhp
, j
)) {
1797 *was_alldone
= !rcu_segcblist_pend_cbs(&rdp
->cblist
);
1799 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1801 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp
->nocb_bypass
));
1802 return false; // Caller must enqueue the callback.
1804 if (j
!= rdp
->nocb_gp_adv_time
&&
1805 rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
1806 rcu_seq_done(&rdp
->mynode
->gp_seq
, cur_gp_seq
)) {
1807 rcu_advance_cbs_nowake(rdp
->mynode
, rdp
);
1808 rdp
->nocb_gp_adv_time
= j
;
1810 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1811 return true; // Callback already enqueued.
1814 // We need to use the bypass.
1815 rcu_nocb_wait_contended(rdp
);
1816 rcu_nocb_bypass_lock(rdp
);
1817 ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1818 rcu_segcblist_inc_len(&rdp
->cblist
); /* Must precede enqueue. */
1819 rcu_cblist_enqueue(&rdp
->nocb_bypass
, rhp
);
1821 WRITE_ONCE(rdp
->nocb_bypass_first
, j
);
1822 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("FirstBQ"));
1824 rcu_nocb_bypass_unlock(rdp
);
1825 smp_mb(); /* Order enqueue before wake. */
1827 local_irq_restore(flags
);
1829 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1830 rcu_nocb_lock(rdp
); // Rare during call_rcu() flood.
1831 if (!rcu_segcblist_pend_cbs(&rdp
->cblist
)) {
1832 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1833 TPS("FirstBQwake"));
1834 __call_rcu_nocb_wake(rdp
, true, flags
);
1836 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1837 TPS("FirstBQnoWake"));
1838 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1841 return true; // Callback already enqueued.
1845 * Awaken the no-CBs grace-period kthead if needed, either due to it
1846 * legitimately being asleep or due to overload conditions.
1848 * If warranted, also wake up the kthread servicing this CPUs queues.
1850 static void __call_rcu_nocb_wake(struct rcu_data
*rdp
, bool was_alldone
,
1851 unsigned long flags
)
1852 __releases(rdp
->nocb_lock
)
1854 unsigned long cur_gp_seq
;
1857 struct task_struct
*t
;
1859 // If we are being polled or there is no kthread, just leave.
1860 t
= READ_ONCE(rdp
->nocb_gp_kthread
);
1861 if (rcu_nocb_poll
|| !t
) {
1862 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1863 TPS("WakeNotPoll"));
1864 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1867 // Need to actually to a wakeup.
1868 len
= rcu_segcblist_n_cbs(&rdp
->cblist
);
1870 rdp
->qlen_last_fqs_check
= len
;
1871 if (!irqs_disabled_flags(flags
)) {
1872 /* ... if queue was empty ... */
1873 wake_nocb_gp(rdp
, false, flags
);
1874 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1877 wake_nocb_gp_defer(rdp
, RCU_NOCB_WAKE
,
1878 TPS("WakeEmptyIsDeferred"));
1879 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1881 } else if (len
> rdp
->qlen_last_fqs_check
+ qhimark
) {
1882 /* ... or if many callbacks queued. */
1883 rdp
->qlen_last_fqs_check
= len
;
1885 if (j
!= rdp
->nocb_gp_adv_time
&&
1886 rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
1887 rcu_seq_done(&rdp
->mynode
->gp_seq
, cur_gp_seq
)) {
1888 rcu_advance_cbs_nowake(rdp
->mynode
, rdp
);
1889 rdp
->nocb_gp_adv_time
= j
;
1891 smp_mb(); /* Enqueue before timer_pending(). */
1892 if ((rdp
->nocb_cb_sleep
||
1893 !rcu_segcblist_ready_cbs(&rdp
->cblist
)) &&
1894 !timer_pending(&rdp
->nocb_bypass_timer
))
1895 wake_nocb_gp_defer(rdp
, RCU_NOCB_WAKE_FORCE
,
1896 TPS("WakeOvfIsDeferred"));
1897 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1899 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("WakeNot"));
1900 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1905 /* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1906 static void do_nocb_bypass_wakeup_timer(struct timer_list
*t
)
1908 unsigned long flags
;
1909 struct rcu_data
*rdp
= from_timer(rdp
, t
, nocb_bypass_timer
);
1911 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("Timer"));
1912 rcu_nocb_lock_irqsave(rdp
, flags
);
1913 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1914 __call_rcu_nocb_wake(rdp
, true, flags
);
1918 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1919 * or for grace periods to end.
1921 static void nocb_gp_wait(struct rcu_data
*my_rdp
)
1923 bool bypass
= false;
1925 int __maybe_unused cpu
= my_rdp
->cpu
;
1926 unsigned long cur_gp_seq
;
1927 unsigned long flags
;
1928 bool gotcbs
= false;
1929 unsigned long j
= jiffies
;
1930 bool needwait_gp
= false; // This prevents actual uninitialized use.
1933 struct rcu_data
*rdp
;
1934 struct rcu_node
*rnp
;
1935 unsigned long wait_gp_seq
= 0; // Suppress "use uninitialized" warning.
1936 bool wasempty
= false;
1939 * Each pass through the following loop checks for CBs and for the
1940 * nearest grace period (if any) to wait for next. The CB kthreads
1941 * and the global grace-period kthread are awakened if needed.
1943 for (rdp
= my_rdp
; rdp
; rdp
= rdp
->nocb_next_cb_rdp
) {
1944 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("Check"));
1945 rcu_nocb_lock_irqsave(rdp
, flags
);
1946 bypass_ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1948 (time_after(j
, READ_ONCE(rdp
->nocb_bypass_first
) + 1) ||
1949 bypass_ncbs
> 2 * qhimark
)) {
1950 // Bypass full or old, so flush it.
1951 (void)rcu_nocb_try_flush_bypass(rdp
, j
);
1952 bypass_ncbs
= rcu_cblist_n_cbs(&rdp
->nocb_bypass
);
1953 } else if (!bypass_ncbs
&& rcu_segcblist_empty(&rdp
->cblist
)) {
1954 rcu_nocb_unlock_irqrestore(rdp
, flags
);
1955 continue; /* No callbacks here, try next. */
1958 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1963 if (bypass
) { // Avoid race with first bypass CB.
1964 WRITE_ONCE(my_rdp
->nocb_defer_wakeup
,
1966 del_timer(&my_rdp
->nocb_timer
);
1968 // Advance callbacks if helpful and low contention.
1969 needwake_gp
= false;
1970 if (!rcu_segcblist_restempty(&rdp
->cblist
,
1971 RCU_NEXT_READY_TAIL
) ||
1972 (rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
1973 rcu_seq_done(&rnp
->gp_seq
, cur_gp_seq
))) {
1974 raw_spin_lock_rcu_node(rnp
); /* irqs disabled. */
1975 needwake_gp
= rcu_advance_cbs(rnp
, rdp
);
1976 wasempty
= rcu_segcblist_restempty(&rdp
->cblist
,
1977 RCU_NEXT_READY_TAIL
);
1978 raw_spin_unlock_rcu_node(rnp
); /* irqs disabled. */
1980 // Need to wait on some grace period?
1981 WARN_ON_ONCE(wasempty
&&
1982 !rcu_segcblist_restempty(&rdp
->cblist
,
1983 RCU_NEXT_READY_TAIL
));
1984 if (rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
)) {
1986 ULONG_CMP_LT(cur_gp_seq
, wait_gp_seq
))
1987 wait_gp_seq
= cur_gp_seq
;
1989 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
,
1992 if (rcu_segcblist_ready_cbs(&rdp
->cblist
)) {
1993 needwake
= rdp
->nocb_cb_sleep
;
1994 WRITE_ONCE(rdp
->nocb_cb_sleep
, false);
1995 smp_mb(); /* CB invocation -after- GP end. */
1999 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2001 swake_up_one(&rdp
->nocb_cb_wq
);
2005 rcu_gp_kthread_wake();
2008 my_rdp
->nocb_gp_bypass
= bypass
;
2009 my_rdp
->nocb_gp_gp
= needwait_gp
;
2010 my_rdp
->nocb_gp_seq
= needwait_gp
? wait_gp_seq
: 0;
2011 if (bypass
&& !rcu_nocb_poll
) {
2012 // At least one child with non-empty ->nocb_bypass, so set
2013 // timer in order to avoid stranding its callbacks.
2014 raw_spin_lock_irqsave(&my_rdp
->nocb_gp_lock
, flags
);
2015 mod_timer(&my_rdp
->nocb_bypass_timer
, j
+ 2);
2016 raw_spin_unlock_irqrestore(&my_rdp
->nocb_gp_lock
, flags
);
2018 if (rcu_nocb_poll
) {
2019 /* Polling, so trace if first poll in the series. */
2021 trace_rcu_nocb_wake(rcu_state
.name
, cpu
, TPS("Poll"));
2022 schedule_timeout_interruptible(1);
2023 } else if (!needwait_gp
) {
2024 /* Wait for callbacks to appear. */
2025 trace_rcu_nocb_wake(rcu_state
.name
, cpu
, TPS("Sleep"));
2026 swait_event_interruptible_exclusive(my_rdp
->nocb_gp_wq
,
2027 !READ_ONCE(my_rdp
->nocb_gp_sleep
));
2028 trace_rcu_nocb_wake(rcu_state
.name
, cpu
, TPS("EndSleep"));
2030 rnp
= my_rdp
->mynode
;
2031 trace_rcu_this_gp(rnp
, my_rdp
, wait_gp_seq
, TPS("StartWait"));
2032 swait_event_interruptible_exclusive(
2033 rnp
->nocb_gp_wq
[rcu_seq_ctr(wait_gp_seq
) & 0x1],
2034 rcu_seq_done(&rnp
->gp_seq
, wait_gp_seq
) ||
2035 !READ_ONCE(my_rdp
->nocb_gp_sleep
));
2036 trace_rcu_this_gp(rnp
, my_rdp
, wait_gp_seq
, TPS("EndWait"));
2038 if (!rcu_nocb_poll
) {
2039 raw_spin_lock_irqsave(&my_rdp
->nocb_gp_lock
, flags
);
2041 del_timer(&my_rdp
->nocb_bypass_timer
);
2042 WRITE_ONCE(my_rdp
->nocb_gp_sleep
, true);
2043 raw_spin_unlock_irqrestore(&my_rdp
->nocb_gp_lock
, flags
);
2045 my_rdp
->nocb_gp_seq
= -1;
2046 WARN_ON(signal_pending(current
));
2050 * No-CBs grace-period-wait kthread. There is one of these per group
2051 * of CPUs, but only once at least one CPU in that group has come online
2052 * at least once since boot. This kthread checks for newly posted
2053 * callbacks from any of the CPUs it is responsible for, waits for a
2054 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2055 * that then have callback-invocation work to do.
2057 static int rcu_nocb_gp_kthread(void *arg
)
2059 struct rcu_data
*rdp
= arg
;
2062 WRITE_ONCE(rdp
->nocb_gp_loops
, rdp
->nocb_gp_loops
+ 1);
2064 cond_resched_tasks_rcu_qs();
2070 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2071 * then, if there are no more, wait for more to appear.
2073 static void nocb_cb_wait(struct rcu_data
*rdp
)
2075 unsigned long cur_gp_seq
;
2076 unsigned long flags
;
2077 bool needwake_gp
= false;
2078 struct rcu_node
*rnp
= rdp
->mynode
;
2080 local_irq_save(flags
);
2081 rcu_momentary_dyntick_idle();
2082 local_irq_restore(flags
);
2086 lockdep_assert_irqs_enabled();
2087 rcu_nocb_lock_irqsave(rdp
, flags
);
2088 if (rcu_segcblist_nextgp(&rdp
->cblist
, &cur_gp_seq
) &&
2089 rcu_seq_done(&rnp
->gp_seq
, cur_gp_seq
) &&
2090 raw_spin_trylock_rcu_node(rnp
)) { /* irqs already disabled. */
2091 needwake_gp
= rcu_advance_cbs(rdp
->mynode
, rdp
);
2092 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2094 if (rcu_segcblist_ready_cbs(&rdp
->cblist
)) {
2095 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2097 rcu_gp_kthread_wake();
2101 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("CBSleep"));
2102 WRITE_ONCE(rdp
->nocb_cb_sleep
, true);
2103 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2105 rcu_gp_kthread_wake();
2106 swait_event_interruptible_exclusive(rdp
->nocb_cb_wq
,
2107 !READ_ONCE(rdp
->nocb_cb_sleep
));
2108 if (!smp_load_acquire(&rdp
->nocb_cb_sleep
)) { /* VVV */
2109 /* ^^^ Ensure CB invocation follows _sleep test. */
2112 WARN_ON(signal_pending(current
));
2113 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("WokeEmpty"));
2117 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2118 * nocb_cb_wait() to do the dirty work.
2120 static int rcu_nocb_cb_kthread(void *arg
)
2122 struct rcu_data
*rdp
= arg
;
2124 // Each pass through this loop does one callback batch, and,
2125 // if there are no more ready callbacks, waits for them.
2128 cond_resched_tasks_rcu_qs();
2133 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2134 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2136 return READ_ONCE(rdp
->nocb_defer_wakeup
);
2139 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2140 static void do_nocb_deferred_wakeup_common(struct rcu_data
*rdp
)
2142 unsigned long flags
;
2145 rcu_nocb_lock_irqsave(rdp
, flags
);
2146 if (!rcu_nocb_need_deferred_wakeup(rdp
)) {
2147 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2150 ndw
= READ_ONCE(rdp
->nocb_defer_wakeup
);
2151 WRITE_ONCE(rdp
->nocb_defer_wakeup
, RCU_NOCB_WAKE_NOT
);
2152 wake_nocb_gp(rdp
, ndw
== RCU_NOCB_WAKE_FORCE
, flags
);
2153 trace_rcu_nocb_wake(rcu_state
.name
, rdp
->cpu
, TPS("DeferredWake"));
2156 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2157 static void do_nocb_deferred_wakeup_timer(struct timer_list
*t
)
2159 struct rcu_data
*rdp
= from_timer(rdp
, t
, nocb_timer
);
2161 do_nocb_deferred_wakeup_common(rdp
);
2165 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2166 * This means we do an inexact common-case check. Note that if
2167 * we miss, ->nocb_timer will eventually clean things up.
2169 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2171 if (rcu_nocb_need_deferred_wakeup(rdp
))
2172 do_nocb_deferred_wakeup_common(rdp
);
2175 void __init
rcu_init_nohz(void)
2178 bool need_rcu_nocb_mask
= false;
2179 struct rcu_data
*rdp
;
2181 #if defined(CONFIG_NO_HZ_FULL)
2182 if (tick_nohz_full_running
&& cpumask_weight(tick_nohz_full_mask
))
2183 need_rcu_nocb_mask
= true;
2184 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2186 if (!cpumask_available(rcu_nocb_mask
) && need_rcu_nocb_mask
) {
2187 if (!zalloc_cpumask_var(&rcu_nocb_mask
, GFP_KERNEL
)) {
2188 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2192 if (!cpumask_available(rcu_nocb_mask
))
2195 #if defined(CONFIG_NO_HZ_FULL)
2196 if (tick_nohz_full_running
)
2197 cpumask_or(rcu_nocb_mask
, rcu_nocb_mask
, tick_nohz_full_mask
);
2198 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2200 if (!cpumask_subset(rcu_nocb_mask
, cpu_possible_mask
)) {
2201 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2202 cpumask_and(rcu_nocb_mask
, cpu_possible_mask
,
2205 if (cpumask_empty(rcu_nocb_mask
))
2206 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2208 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2209 cpumask_pr_args(rcu_nocb_mask
));
2211 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2213 for_each_cpu(cpu
, rcu_nocb_mask
) {
2214 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2215 if (rcu_segcblist_empty(&rdp
->cblist
))
2216 rcu_segcblist_init(&rdp
->cblist
);
2217 rcu_segcblist_offload(&rdp
->cblist
);
2219 rcu_organize_nocb_kthreads();
2222 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2223 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2225 init_swait_queue_head(&rdp
->nocb_cb_wq
);
2226 init_swait_queue_head(&rdp
->nocb_gp_wq
);
2227 raw_spin_lock_init(&rdp
->nocb_lock
);
2228 raw_spin_lock_init(&rdp
->nocb_bypass_lock
);
2229 raw_spin_lock_init(&rdp
->nocb_gp_lock
);
2230 timer_setup(&rdp
->nocb_timer
, do_nocb_deferred_wakeup_timer
, 0);
2231 timer_setup(&rdp
->nocb_bypass_timer
, do_nocb_bypass_wakeup_timer
, 0);
2232 rcu_cblist_init(&rdp
->nocb_bypass
);
2236 * If the specified CPU is a no-CBs CPU that does not already have its
2237 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2238 * for this CPU's group has not yet been created, spawn it as well.
2240 static void rcu_spawn_one_nocb_kthread(int cpu
)
2242 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2243 struct rcu_data
*rdp_gp
;
2244 struct task_struct
*t
;
2247 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2248 * then nothing to do.
2250 if (!rcu_is_nocb_cpu(cpu
) || rdp
->nocb_cb_kthread
)
2253 /* If we didn't spawn the GP kthread first, reorganize! */
2254 rdp_gp
= rdp
->nocb_gp_rdp
;
2255 if (!rdp_gp
->nocb_gp_kthread
) {
2256 t
= kthread_run(rcu_nocb_gp_kthread
, rdp_gp
,
2257 "rcuog/%d", rdp_gp
->cpu
);
2258 if (WARN_ONCE(IS_ERR(t
), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__
))
2260 WRITE_ONCE(rdp_gp
->nocb_gp_kthread
, t
);
2263 /* Spawn the kthread for this CPU. */
2264 t
= kthread_run(rcu_nocb_cb_kthread
, rdp
,
2265 "rcuo%c/%d", rcu_state
.abbr
, cpu
);
2266 if (WARN_ONCE(IS_ERR(t
), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__
))
2268 WRITE_ONCE(rdp
->nocb_cb_kthread
, t
);
2269 WRITE_ONCE(rdp
->nocb_gp_kthread
, rdp_gp
->nocb_gp_kthread
);
2273 * If the specified CPU is a no-CBs CPU that does not already have its
2274 * rcuo kthread, spawn it.
2276 static void rcu_spawn_cpu_nocb_kthread(int cpu
)
2278 if (rcu_scheduler_fully_active
)
2279 rcu_spawn_one_nocb_kthread(cpu
);
2283 * Once the scheduler is running, spawn rcuo kthreads for all online
2284 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2285 * non-boot CPUs come online -- if this changes, we will need to add
2286 * some mutual exclusion.
2288 static void __init
rcu_spawn_nocb_kthreads(void)
2292 for_each_online_cpu(cpu
)
2293 rcu_spawn_cpu_nocb_kthread(cpu
);
2296 /* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2297 static int rcu_nocb_gp_stride
= -1;
2298 module_param(rcu_nocb_gp_stride
, int, 0444);
2301 * Initialize GP-CB relationships for all no-CBs CPU.
2303 static void __init
rcu_organize_nocb_kthreads(void)
2306 bool firsttime
= true;
2307 bool gotnocbs
= false;
2308 bool gotnocbscbs
= true;
2309 int ls
= rcu_nocb_gp_stride
;
2310 int nl
= 0; /* Next GP kthread. */
2311 struct rcu_data
*rdp
;
2312 struct rcu_data
*rdp_gp
= NULL
; /* Suppress misguided gcc warn. */
2313 struct rcu_data
*rdp_prev
= NULL
;
2315 if (!cpumask_available(rcu_nocb_mask
))
2318 ls
= nr_cpu_ids
/ int_sqrt(nr_cpu_ids
);
2319 rcu_nocb_gp_stride
= ls
;
2323 * Each pass through this loop sets up one rcu_data structure.
2324 * Should the corresponding CPU come online in the future, then
2325 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2327 for_each_cpu(cpu
, rcu_nocb_mask
) {
2328 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2329 if (rdp
->cpu
>= nl
) {
2330 /* New GP kthread, set up for CBs & next GP. */
2332 nl
= DIV_ROUND_UP(rdp
->cpu
+ 1, ls
) * ls
;
2333 rdp
->nocb_gp_rdp
= rdp
;
2337 pr_cont("%s\n", gotnocbscbs
2338 ? "" : " (self only)");
2339 gotnocbscbs
= false;
2341 pr_alert("%s: No-CB GP kthread CPU %d:",
2345 /* Another CB kthread, link to previous GP kthread. */
2347 rdp
->nocb_gp_rdp
= rdp_gp
;
2348 rdp_prev
->nocb_next_cb_rdp
= rdp
;
2350 pr_cont(" %d", cpu
);
2354 if (gotnocbs
&& dump_tree
)
2355 pr_cont("%s\n", gotnocbscbs
? "" : " (self only)");
2359 * Bind the current task to the offloaded CPUs. If there are no offloaded
2360 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2362 void rcu_bind_current_to_nocb(void)
2364 if (cpumask_available(rcu_nocb_mask
) && cpumask_weight(rcu_nocb_mask
))
2365 WARN_ON(sched_setaffinity(current
->pid
, rcu_nocb_mask
));
2367 EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb
);
2370 * Dump out nocb grace-period kthread state for the specified rcu_data
2373 static void show_rcu_nocb_gp_state(struct rcu_data
*rdp
)
2375 struct rcu_node
*rnp
= rdp
->mynode
;
2377 pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2379 "kK"[!!rdp
->nocb_gp_kthread
],
2380 "lL"[raw_spin_is_locked(&rdp
->nocb_gp_lock
)],
2381 "dD"[!!rdp
->nocb_defer_wakeup
],
2382 "tT"[timer_pending(&rdp
->nocb_timer
)],
2383 "bB"[timer_pending(&rdp
->nocb_bypass_timer
)],
2384 "sS"[!!rdp
->nocb_gp_sleep
],
2385 ".W"[swait_active(&rdp
->nocb_gp_wq
)],
2386 ".W"[swait_active(&rnp
->nocb_gp_wq
[0])],
2387 ".W"[swait_active(&rnp
->nocb_gp_wq
[1])],
2388 ".B"[!!rdp
->nocb_gp_bypass
],
2389 ".G"[!!rdp
->nocb_gp_gp
],
2390 (long)rdp
->nocb_gp_seq
,
2391 rnp
->grplo
, rnp
->grphi
, READ_ONCE(rdp
->nocb_gp_loops
));
2394 /* Dump out nocb kthread state for the specified rcu_data structure. */
2395 static void show_rcu_nocb_state(struct rcu_data
*rdp
)
2397 struct rcu_segcblist
*rsclp
= &rdp
->cblist
;
2402 if (rdp
->nocb_gp_rdp
== rdp
)
2403 show_rcu_nocb_gp_state(rdp
);
2405 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",
2406 rdp
->cpu
, rdp
->nocb_gp_rdp
->cpu
,
2407 "kK"[!!rdp
->nocb_cb_kthread
],
2408 "bB"[raw_spin_is_locked(&rdp
->nocb_bypass_lock
)],
2409 "cC"[!!atomic_read(&rdp
->nocb_lock_contended
)],
2410 "lL"[raw_spin_is_locked(&rdp
->nocb_lock
)],
2411 "sS"[!!rdp
->nocb_cb_sleep
],
2412 ".W"[swait_active(&rdp
->nocb_cb_wq
)],
2413 jiffies
- rdp
->nocb_bypass_first
,
2414 jiffies
- rdp
->nocb_nobypass_last
,
2415 rdp
->nocb_nobypass_count
,
2416 ".D"[rcu_segcblist_ready_cbs(rsclp
)],
2417 ".W"[!rcu_segcblist_restempty(rsclp
, RCU_DONE_TAIL
)],
2418 ".R"[!rcu_segcblist_restempty(rsclp
, RCU_WAIT_TAIL
)],
2419 ".N"[!rcu_segcblist_restempty(rsclp
, RCU_NEXT_READY_TAIL
)],
2420 ".B"[!!rcu_cblist_n_cbs(&rdp
->nocb_bypass
)],
2421 rcu_segcblist_n_cbs(&rdp
->cblist
));
2423 /* It is OK for GP kthreads to have GP state. */
2424 if (rdp
->nocb_gp_rdp
== rdp
)
2427 waslocked
= raw_spin_is_locked(&rdp
->nocb_gp_lock
);
2428 wastimer
= timer_pending(&rdp
->nocb_timer
);
2429 wassleep
= swait_active(&rdp
->nocb_gp_wq
);
2430 if (!rdp
->nocb_defer_wakeup
&& !rdp
->nocb_gp_sleep
&&
2431 !waslocked
&& !wastimer
&& !wassleep
)
2432 return; /* Nothing untowards. */
2434 pr_info(" !!! %c%c%c%c %c\n",
2436 "dD"[!!rdp
->nocb_defer_wakeup
],
2438 "sS"[!!rdp
->nocb_gp_sleep
],
2442 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2444 /* No ->nocb_lock to acquire. */
2445 static void rcu_nocb_lock(struct rcu_data
*rdp
)
2449 /* No ->nocb_lock to release. */
2450 static void rcu_nocb_unlock(struct rcu_data
*rdp
)
2454 /* No ->nocb_lock to release. */
2455 static void rcu_nocb_unlock_irqrestore(struct rcu_data
*rdp
,
2456 unsigned long flags
)
2458 local_irq_restore(flags
);
2461 /* Lockdep check that ->cblist may be safely accessed. */
2462 static void rcu_lockdep_assert_cblist_protected(struct rcu_data
*rdp
)
2464 lockdep_assert_irqs_disabled();
2467 static void rcu_nocb_gp_cleanup(struct swait_queue_head
*sq
)
2471 static struct swait_queue_head
*rcu_nocb_gp_get(struct rcu_node
*rnp
)
2476 static void rcu_init_one_nocb(struct rcu_node
*rnp
)
2480 static bool rcu_nocb_flush_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2486 static bool rcu_nocb_try_bypass(struct rcu_data
*rdp
, struct rcu_head
*rhp
,
2487 bool *was_alldone
, unsigned long flags
)
2492 static void __call_rcu_nocb_wake(struct rcu_data
*rdp
, bool was_empty
,
2493 unsigned long flags
)
2495 WARN_ON_ONCE(1); /* Should be dead code! */
2498 static void __init
rcu_boot_init_nocb_percpu_data(struct rcu_data
*rdp
)
2502 static int rcu_nocb_need_deferred_wakeup(struct rcu_data
*rdp
)
2507 static void do_nocb_deferred_wakeup(struct rcu_data
*rdp
)
2511 static void rcu_spawn_cpu_nocb_kthread(int cpu
)
2515 static void __init
rcu_spawn_nocb_kthreads(void)
2519 static void show_rcu_nocb_state(struct rcu_data
*rdp
)
2523 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2526 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2527 * grace-period kthread will do force_quiescent_state() processing?
2528 * The idea is to avoid waking up RCU core processing on such a
2529 * CPU unless the grace period has extended for too long.
2531 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2532 * CONFIG_RCU_NOCB_CPU CPUs.
2534 static bool rcu_nohz_full_cpu(void)
2536 #ifdef CONFIG_NO_HZ_FULL
2537 if (tick_nohz_full_cpu(smp_processor_id()) &&
2538 (!rcu_gp_in_progress() ||
2539 ULONG_CMP_LT(jiffies
, READ_ONCE(rcu_state
.gp_start
) + HZ
)))
2541 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2546 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2548 static void rcu_bind_gp_kthread(void)
2550 if (!tick_nohz_full_enabled())
2552 housekeeping_affine(current
, HK_FLAG_RCU
);
2555 /* Record the current task on dyntick-idle entry. */
2556 static void rcu_dynticks_task_enter(void)
2558 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2559 WRITE_ONCE(current
->rcu_tasks_idle_cpu
, smp_processor_id());
2560 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2563 /* Record no current task on dyntick-idle exit. */
2564 static void rcu_dynticks_task_exit(void)
2566 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2567 WRITE_ONCE(current
->rcu_tasks_idle_cpu
, -1);
2568 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */