Loading...
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions that provide either classic
5 * or preemptible semantics.
6 *
7 * Copyright Red Hat, 2009
8 * Copyright IBM Corporation, 2009
9 *
10 * Author: Ingo Molnar <mingo@elte.hu>
11 * Paul E. McKenney <paulmck@linux.ibm.com>
12 */
13
14#include "../locking/rtmutex_common.h"
15
16#ifdef CONFIG_RCU_NOCB_CPU
17static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
18static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
19#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
20
21/*
22 * Check the RCU kernel configuration parameters and print informative
23 * messages about anything out of the ordinary.
24 */
25static void __init rcu_bootup_announce_oddness(void)
26{
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",
32 RCU_FANOUT);
33 if (rcu_fanout_exact)
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",
43 RCU_FANOUT_LEAF);
44 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
45 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
46 rcu_fanout_leaf);
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);
52#endif
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");
71 if (gp_preinit_delay)
72 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
73 if (gp_init_delay)
74 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
75 if (gp_cleanup_delay)
76 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
77 if (!use_softirq)
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();
82}
83
84#ifdef CONFIG_PREEMPT_RCU
85
86static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
87static void rcu_read_unlock_special(struct task_struct *t);
88
89/*
90 * Tell them what RCU they are running.
91 */
92static void __init rcu_bootup_announce(void)
93{
94 pr_info("Preemptible hierarchical RCU implementation.\n");
95 rcu_bootup_announce_oddness();
96}
97
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
103
104/*
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.
117 *
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.
122 *
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.
131 */
132static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
133 __releases(rnp->lock) /* But leaves rrupts disabled. */
134{
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;
140
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 &
146 rdp->grpmask);
147
148 /*
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.
152 */
153 switch (blkd_state) {
154 case 0:
155 case RCU_EXP_TASKS:
156 case RCU_EXP_TASKS + RCU_GP_BLKD:
157 case RCU_GP_TASKS:
158 case RCU_GP_TASKS + RCU_EXP_TASKS:
159
160 /*
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.
165 */
166 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
167 break;
168
169 case RCU_EXP_BLKD:
170 case RCU_GP_BLKD:
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:
175
176 /*
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.
183 */
184 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
185 break;
186
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:
190
191 /*
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.
196 */
197 list_add(&t->rcu_node_entry, rnp->exp_tasks);
198 break;
199
200 case RCU_GP_TASKS + RCU_GP_BLKD:
201 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
202
203 /*
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.
207 */
208 list_add(&t->rcu_node_entry, rnp->gp_tasks);
209 break;
210
211 default:
212
213 /* Yet another exercise in excessive paranoia. */
214 WARN_ON_ONCE(1);
215 break;
216 }
217
218 /*
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
222 * blocked tasks.
223 */
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);
227 }
228 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
229 WRITE_ONCE(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. */
235
236 /*
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.)
241 */
242 if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
243 rcu_report_exp_rdp(rdp);
244 else
245 WARN_ON_ONCE(rdp->exp_deferred_qs);
246}
247
248/*
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.
258 *
259 * Callers to this function must disable preemption.
260 */
261static void rcu_qs(void)
262{
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),
267 TPS("cpuqs"));
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);
271 }
272}
273
274/*
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.
284 *
285 * Caller must disable interrupts.
286 */
287void rcu_note_context_switch(bool preempt)
288{
289 struct task_struct *t = current;
290 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
291 struct rcu_node *rnp;
292
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) {
298
299 /* Possibly blocking in an RCU read-side critical section. */
300 rnp = rdp->mynode;
301 raw_spin_lock_rcu_node(rnp);
302 t->rcu_read_unlock_special.b.blocked = true;
303 t->rcu_blocked_node = rnp;
304
305 /*
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.
309 */
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,
313 t->pid,
314 (rnp->qsmask & rdp->grpmask)
315 ? rnp->gp_seq
316 : rcu_seq_snap(&rnp->gp_seq));
317 rcu_preempt_ctxt_queue(rnp, rdp);
318 } else {
319 rcu_preempt_deferred_qs(t);
320 }
321
322 /*
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.
330 */
331 rcu_qs();
332 if (rdp->exp_deferred_qs)
333 rcu_report_exp_rdp(rdp);
334 rcu_tasks_qs(current, preempt);
335 trace_rcu_utilization(TPS("End context switch"));
336}
337EXPORT_SYMBOL_GPL(rcu_note_context_switch);
338
339/*
340 * Check for preempted RCU readers blocking the current grace period
341 * for the specified rcu_node structure. If the caller needs a reliable
342 * answer, it must hold the rcu_node's ->lock.
343 */
344static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
345{
346 return READ_ONCE(rnp->gp_tasks) != NULL;
347}
348
349/* limit value for ->rcu_read_lock_nesting. */
350#define RCU_NEST_PMAX (INT_MAX / 2)
351
352static void rcu_preempt_read_enter(void)
353{
354 current->rcu_read_lock_nesting++;
355}
356
357static int rcu_preempt_read_exit(void)
358{
359 return --current->rcu_read_lock_nesting;
360}
361
362static void rcu_preempt_depth_set(int val)
363{
364 current->rcu_read_lock_nesting = val;
365}
366
367/*
368 * Preemptible RCU implementation for rcu_read_lock().
369 * Just increment ->rcu_read_lock_nesting, shared state will be updated
370 * if we block.
371 */
372void __rcu_read_lock(void)
373{
374 rcu_preempt_read_enter();
375 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
376 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
377 barrier(); /* critical section after entry code. */
378}
379EXPORT_SYMBOL_GPL(__rcu_read_lock);
380
381/*
382 * Preemptible RCU implementation for rcu_read_unlock().
383 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
384 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
385 * invoke rcu_read_unlock_special() to clean up after a context switch
386 * in an RCU read-side critical section and other special cases.
387 */
388void __rcu_read_unlock(void)
389{
390 struct task_struct *t = current;
391
392 if (rcu_preempt_read_exit() == 0) {
393 barrier(); /* critical section before exit code. */
394 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
395 rcu_read_unlock_special(t);
396 }
397 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
398 int rrln = rcu_preempt_depth();
399
400 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
401 }
402}
403EXPORT_SYMBOL_GPL(__rcu_read_unlock);
404
405/*
406 * Advance a ->blkd_tasks-list pointer to the next entry, instead
407 * returning NULL if at the end of the list.
408 */
409static struct list_head *rcu_next_node_entry(struct task_struct *t,
410 struct rcu_node *rnp)
411{
412 struct list_head *np;
413
414 np = t->rcu_node_entry.next;
415 if (np == &rnp->blkd_tasks)
416 np = NULL;
417 return np;
418}
419
420/*
421 * Return true if the specified rcu_node structure has tasks that were
422 * preempted within an RCU read-side critical section.
423 */
424static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
425{
426 return !list_empty(&rnp->blkd_tasks);
427}
428
429/*
430 * Report deferred quiescent states. The deferral time can
431 * be quite short, for example, in the case of the call from
432 * rcu_read_unlock_special().
433 */
434static void
435rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
436{
437 bool empty_exp;
438 bool empty_norm;
439 bool empty_exp_now;
440 struct list_head *np;
441 bool drop_boost_mutex = false;
442 struct rcu_data *rdp;
443 struct rcu_node *rnp;
444 union rcu_special special;
445
446 /*
447 * If RCU core is waiting for this CPU to exit its critical section,
448 * report the fact that it has exited. Because irqs are disabled,
449 * t->rcu_read_unlock_special cannot change.
450 */
451 special = t->rcu_read_unlock_special;
452 rdp = this_cpu_ptr(&rcu_data);
453 if (!special.s && !rdp->exp_deferred_qs) {
454 local_irq_restore(flags);
455 return;
456 }
457 t->rcu_read_unlock_special.s = 0;
458 if (special.b.need_qs)
459 rcu_qs();
460
461 /*
462 * Respond to a request by an expedited grace period for a
463 * quiescent state from this CPU. Note that requests from
464 * tasks are handled when removing the task from the
465 * blocked-tasks list below.
466 */
467 if (rdp->exp_deferred_qs)
468 rcu_report_exp_rdp(rdp);
469
470 /* Clean up if blocked during RCU read-side critical section. */
471 if (special.b.blocked) {
472
473 /*
474 * Remove this task from the list it blocked on. The task
475 * now remains queued on the rcu_node corresponding to the
476 * CPU it first blocked on, so there is no longer any need
477 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
478 */
479 rnp = t->rcu_blocked_node;
480 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
481 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
482 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
483 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
484 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
485 (!empty_norm || rnp->qsmask));
486 empty_exp = sync_rcu_exp_done(rnp);
487 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
488 np = rcu_next_node_entry(t, rnp);
489 list_del_init(&t->rcu_node_entry);
490 t->rcu_blocked_node = NULL;
491 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
492 rnp->gp_seq, t->pid);
493 if (&t->rcu_node_entry == rnp->gp_tasks)
494 WRITE_ONCE(rnp->gp_tasks, np);
495 if (&t->rcu_node_entry == rnp->exp_tasks)
496 WRITE_ONCE(rnp->exp_tasks, np);
497 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
498 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
499 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
500 if (&t->rcu_node_entry == rnp->boost_tasks)
501 WRITE_ONCE(rnp->boost_tasks, np);
502 }
503
504 /*
505 * If this was the last task on the current list, and if
506 * we aren't waiting on any CPUs, report the quiescent state.
507 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
508 * so we must take a snapshot of the expedited state.
509 */
510 empty_exp_now = sync_rcu_exp_done(rnp);
511 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
512 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
513 rnp->gp_seq,
514 0, rnp->qsmask,
515 rnp->level,
516 rnp->grplo,
517 rnp->grphi,
518 !!rnp->gp_tasks);
519 rcu_report_unblock_qs_rnp(rnp, flags);
520 } else {
521 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
522 }
523
524 /* Unboost if we were boosted. */
525 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
526 rt_mutex_futex_unlock(&rnp->boost_mtx);
527
528 /*
529 * If this was the last task on the expedited lists,
530 * then we need to report up the rcu_node hierarchy.
531 */
532 if (!empty_exp && empty_exp_now)
533 rcu_report_exp_rnp(rnp, true);
534 } else {
535 local_irq_restore(flags);
536 }
537}
538
539/*
540 * Is a deferred quiescent-state pending, and are we also not in
541 * an RCU read-side critical section? It is the caller's responsibility
542 * to ensure it is otherwise safe to report any deferred quiescent
543 * states. The reason for this is that it is safe to report a
544 * quiescent state during context switch even though preemption
545 * is disabled. This function cannot be expected to understand these
546 * nuances, so the caller must handle them.
547 */
548static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
549{
550 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
551 READ_ONCE(t->rcu_read_unlock_special.s)) &&
552 rcu_preempt_depth() == 0;
553}
554
555/*
556 * Report a deferred quiescent state if needed and safe to do so.
557 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
558 * not being in an RCU read-side critical section. The caller must
559 * evaluate safety in terms of interrupt, softirq, and preemption
560 * disabling.
561 */
562static void rcu_preempt_deferred_qs(struct task_struct *t)
563{
564 unsigned long flags;
565
566 if (!rcu_preempt_need_deferred_qs(t))
567 return;
568 local_irq_save(flags);
569 rcu_preempt_deferred_qs_irqrestore(t, flags);
570}
571
572/*
573 * Minimal handler to give the scheduler a chance to re-evaluate.
574 */
575static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
576{
577 struct rcu_data *rdp;
578
579 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
580 rdp->defer_qs_iw_pending = false;
581}
582
583/*
584 * Handle special cases during rcu_read_unlock(), such as needing to
585 * notify RCU core processing or task having blocked during the RCU
586 * read-side critical section.
587 */
588static void rcu_read_unlock_special(struct task_struct *t)
589{
590 unsigned long flags;
591 bool preempt_bh_were_disabled =
592 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
593 bool irqs_were_disabled;
594
595 /* NMI handlers cannot block and cannot safely manipulate state. */
596 if (in_nmi())
597 return;
598
599 local_irq_save(flags);
600 irqs_were_disabled = irqs_disabled_flags(flags);
601 if (preempt_bh_were_disabled || irqs_were_disabled) {
602 bool exp;
603 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
604 struct rcu_node *rnp = rdp->mynode;
605
606 exp = (t->rcu_blocked_node &&
607 READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
608 (rdp->grpmask & READ_ONCE(rnp->expmask));
609 // Need to defer quiescent state until everything is enabled.
610 if (use_softirq && (in_irq() || (exp && !irqs_were_disabled))) {
611 // Using softirq, safe to awaken, and either the
612 // wakeup is free or there is an expedited GP.
613 raise_softirq_irqoff(RCU_SOFTIRQ);
614 } else {
615 // Enabling BH or preempt does reschedule, so...
616 // Also if no expediting, slow is OK.
617 // Plus nohz_full CPUs eventually get tick enabled.
618 set_tsk_need_resched(current);
619 set_preempt_need_resched();
620 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
621 !rdp->defer_qs_iw_pending && exp) {
622 // Get scheduler to re-evaluate and call hooks.
623 // If !IRQ_WORK, FQS scan will eventually IPI.
624 init_irq_work(&rdp->defer_qs_iw,
625 rcu_preempt_deferred_qs_handler);
626 rdp->defer_qs_iw_pending = true;
627 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
628 }
629 }
630 local_irq_restore(flags);
631 return;
632 }
633 rcu_preempt_deferred_qs_irqrestore(t, flags);
634}
635
636/*
637 * Check that the list of blocked tasks for the newly completed grace
638 * period is in fact empty. It is a serious bug to complete a grace
639 * period that still has RCU readers blocked! This function must be
640 * invoked -before- updating this rnp's ->gp_seq.
641 *
642 * Also, if there are blocked tasks on the list, they automatically
643 * block the newly created grace period, so set up ->gp_tasks accordingly.
644 */
645static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
646{
647 struct task_struct *t;
648
649 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
650 raw_lockdep_assert_held_rcu_node(rnp);
651 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
652 dump_blkd_tasks(rnp, 10);
653 if (rcu_preempt_has_tasks(rnp) &&
654 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
655 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
656 t = container_of(rnp->gp_tasks, struct task_struct,
657 rcu_node_entry);
658 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
659 rnp->gp_seq, t->pid);
660 }
661 WARN_ON_ONCE(rnp->qsmask);
662}
663
664/*
665 * Check for a quiescent state from the current CPU, including voluntary
666 * context switches for Tasks RCU. When a task blocks, the task is
667 * recorded in the corresponding CPU's rcu_node structure, which is checked
668 * elsewhere, hence this function need only check for quiescent states
669 * related to the current CPU, not to those related to tasks.
670 */
671static void rcu_flavor_sched_clock_irq(int user)
672{
673 struct task_struct *t = current;
674
675 if (user || rcu_is_cpu_rrupt_from_idle()) {
676 rcu_note_voluntary_context_switch(current);
677 }
678 if (rcu_preempt_depth() > 0 ||
679 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
680 /* No QS, force context switch if deferred. */
681 if (rcu_preempt_need_deferred_qs(t)) {
682 set_tsk_need_resched(t);
683 set_preempt_need_resched();
684 }
685 } else if (rcu_preempt_need_deferred_qs(t)) {
686 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
687 return;
688 } else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
689 rcu_qs(); /* Report immediate QS. */
690 return;
691 }
692
693 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
694 if (rcu_preempt_depth() > 0 &&
695 __this_cpu_read(rcu_data.core_needs_qs) &&
696 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
697 !t->rcu_read_unlock_special.b.need_qs &&
698 time_after(jiffies, rcu_state.gp_start + HZ))
699 t->rcu_read_unlock_special.b.need_qs = true;
700}
701
702/*
703 * Check for a task exiting while in a preemptible-RCU read-side
704 * critical section, clean up if so. No need to issue warnings, as
705 * debug_check_no_locks_held() already does this if lockdep is enabled.
706 * Besides, if this function does anything other than just immediately
707 * return, there was a bug of some sort. Spewing warnings from this
708 * function is like as not to simply obscure important prior warnings.
709 */
710void exit_rcu(void)
711{
712 struct task_struct *t = current;
713
714 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
715 rcu_preempt_depth_set(1);
716 barrier();
717 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
718 } else if (unlikely(rcu_preempt_depth())) {
719 rcu_preempt_depth_set(1);
720 } else {
721 return;
722 }
723 __rcu_read_unlock();
724 rcu_preempt_deferred_qs(current);
725}
726
727/*
728 * Dump the blocked-tasks state, but limit the list dump to the
729 * specified number of elements.
730 */
731static void
732dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
733{
734 int cpu;
735 int i;
736 struct list_head *lhp;
737 bool onl;
738 struct rcu_data *rdp;
739 struct rcu_node *rnp1;
740
741 raw_lockdep_assert_held_rcu_node(rnp);
742 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
743 __func__, rnp->grplo, rnp->grphi, rnp->level,
744 (long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
745 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
746 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
747 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
748 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
749 __func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
750 READ_ONCE(rnp->exp_tasks));
751 pr_info("%s: ->blkd_tasks", __func__);
752 i = 0;
753 list_for_each(lhp, &rnp->blkd_tasks) {
754 pr_cont(" %p", lhp);
755 if (++i >= ncheck)
756 break;
757 }
758 pr_cont("\n");
759 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
760 rdp = per_cpu_ptr(&rcu_data, cpu);
761 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
762 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
763 cpu, ".o"[onl],
764 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
765 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
766 }
767}
768
769#else /* #ifdef CONFIG_PREEMPT_RCU */
770
771/*
772 * Tell them what RCU they are running.
773 */
774static void __init rcu_bootup_announce(void)
775{
776 pr_info("Hierarchical RCU implementation.\n");
777 rcu_bootup_announce_oddness();
778}
779
780/*
781 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
782 * how many quiescent states passed, just if there was at least one since
783 * the start of the grace period, this just sets a flag. The caller must
784 * have disabled preemption.
785 */
786static void rcu_qs(void)
787{
788 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
789 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
790 return;
791 trace_rcu_grace_period(TPS("rcu_sched"),
792 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
793 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
794 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
795 return;
796 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
797 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
798}
799
800/*
801 * Register an urgently needed quiescent state. If there is an
802 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
803 * dyntick-idle quiescent state visible to other CPUs, which will in
804 * some cases serve for expedited as well as normal grace periods.
805 * Either way, register a lightweight quiescent state.
806 */
807void rcu_all_qs(void)
808{
809 unsigned long flags;
810
811 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
812 return;
813 preempt_disable();
814 /* Load rcu_urgent_qs before other flags. */
815 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
816 preempt_enable();
817 return;
818 }
819 this_cpu_write(rcu_data.rcu_urgent_qs, false);
820 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
821 local_irq_save(flags);
822 rcu_momentary_dyntick_idle();
823 local_irq_restore(flags);
824 }
825 rcu_qs();
826 preempt_enable();
827}
828EXPORT_SYMBOL_GPL(rcu_all_qs);
829
830/*
831 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
832 */
833void rcu_note_context_switch(bool preempt)
834{
835 trace_rcu_utilization(TPS("Start context switch"));
836 rcu_qs();
837 /* Load rcu_urgent_qs before other flags. */
838 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
839 goto out;
840 this_cpu_write(rcu_data.rcu_urgent_qs, false);
841 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
842 rcu_momentary_dyntick_idle();
843 rcu_tasks_qs(current, preempt);
844out:
845 trace_rcu_utilization(TPS("End context switch"));
846}
847EXPORT_SYMBOL_GPL(rcu_note_context_switch);
848
849/*
850 * Because preemptible RCU does not exist, there are never any preempted
851 * RCU readers.
852 */
853static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
854{
855 return 0;
856}
857
858/*
859 * Because there is no preemptible RCU, there can be no readers blocked.
860 */
861static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
862{
863 return false;
864}
865
866/*
867 * Because there is no preemptible RCU, there can be no deferred quiescent
868 * states.
869 */
870static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
871{
872 return false;
873}
874static void rcu_preempt_deferred_qs(struct task_struct *t) { }
875
876/*
877 * Because there is no preemptible RCU, there can be no readers blocked,
878 * so there is no need to check for blocked tasks. So check only for
879 * bogus qsmask values.
880 */
881static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
882{
883 WARN_ON_ONCE(rnp->qsmask);
884}
885
886/*
887 * Check to see if this CPU is in a non-context-switch quiescent state,
888 * namely user mode and idle loop.
889 */
890static void rcu_flavor_sched_clock_irq(int user)
891{
892 if (user || rcu_is_cpu_rrupt_from_idle()) {
893
894 /*
895 * Get here if this CPU took its interrupt from user
896 * mode or from the idle loop, and if this is not a
897 * nested interrupt. In this case, the CPU is in
898 * a quiescent state, so note it.
899 *
900 * No memory barrier is required here because rcu_qs()
901 * references only CPU-local variables that other CPUs
902 * neither access nor modify, at least not while the
903 * corresponding CPU is online.
904 */
905
906 rcu_qs();
907 }
908}
909
910/*
911 * Because preemptible RCU does not exist, tasks cannot possibly exit
912 * while in preemptible RCU read-side critical sections.
913 */
914void exit_rcu(void)
915{
916}
917
918/*
919 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
920 */
921static void
922dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
923{
924 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
925}
926
927#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
928
929/*
930 * If boosting, set rcuc kthreads to realtime priority.
931 */
932static void rcu_cpu_kthread_setup(unsigned int cpu)
933{
934#ifdef CONFIG_RCU_BOOST
935 struct sched_param sp;
936
937 sp.sched_priority = kthread_prio;
938 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
939#endif /* #ifdef CONFIG_RCU_BOOST */
940}
941
942#ifdef CONFIG_RCU_BOOST
943
944/*
945 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
946 * or ->boost_tasks, advancing the pointer to the next task in the
947 * ->blkd_tasks list.
948 *
949 * Note that irqs must be enabled: boosting the task can block.
950 * Returns 1 if there are more tasks needing to be boosted.
951 */
952static int rcu_boost(struct rcu_node *rnp)
953{
954 unsigned long flags;
955 struct task_struct *t;
956 struct list_head *tb;
957
958 if (READ_ONCE(rnp->exp_tasks) == NULL &&
959 READ_ONCE(rnp->boost_tasks) == NULL)
960 return 0; /* Nothing left to boost. */
961
962 raw_spin_lock_irqsave_rcu_node(rnp, flags);
963
964 /*
965 * Recheck under the lock: all tasks in need of boosting
966 * might exit their RCU read-side critical sections on their own.
967 */
968 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
969 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
970 return 0;
971 }
972
973 /*
974 * Preferentially boost tasks blocking expedited grace periods.
975 * This cannot starve the normal grace periods because a second
976 * expedited grace period must boost all blocked tasks, including
977 * those blocking the pre-existing normal grace period.
978 */
979 if (rnp->exp_tasks != NULL)
980 tb = rnp->exp_tasks;
981 else
982 tb = rnp->boost_tasks;
983
984 /*
985 * We boost task t by manufacturing an rt_mutex that appears to
986 * be held by task t. We leave a pointer to that rt_mutex where
987 * task t can find it, and task t will release the mutex when it
988 * exits its outermost RCU read-side critical section. Then
989 * simply acquiring this artificial rt_mutex will boost task
990 * t's priority. (Thanks to tglx for suggesting this approach!)
991 *
992 * Note that task t must acquire rnp->lock to remove itself from
993 * the ->blkd_tasks list, which it will do from exit() if from
994 * nowhere else. We therefore are guaranteed that task t will
995 * stay around at least until we drop rnp->lock. Note that
996 * rnp->lock also resolves races between our priority boosting
997 * and task t's exiting its outermost RCU read-side critical
998 * section.
999 */
1000 t = container_of(tb, struct task_struct, rcu_node_entry);
1001 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1002 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1003 /* Lock only for side effect: boosts task t's priority. */
1004 rt_mutex_lock(&rnp->boost_mtx);
1005 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1006
1007 return READ_ONCE(rnp->exp_tasks) != NULL ||
1008 READ_ONCE(rnp->boost_tasks) != NULL;
1009}
1010
1011/*
1012 * Priority-boosting kthread, one per leaf rcu_node.
1013 */
1014static int rcu_boost_kthread(void *arg)
1015{
1016 struct rcu_node *rnp = (struct rcu_node *)arg;
1017 int spincnt = 0;
1018 int more2boost;
1019
1020 trace_rcu_utilization(TPS("Start boost kthread@init"));
1021 for (;;) {
1022 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1023 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1024 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1025 READ_ONCE(rnp->exp_tasks));
1026 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1027 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1028 more2boost = rcu_boost(rnp);
1029 if (more2boost)
1030 spincnt++;
1031 else
1032 spincnt = 0;
1033 if (spincnt > 10) {
1034 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1035 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1036 schedule_timeout_idle(2);
1037 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1038 spincnt = 0;
1039 }
1040 }
1041 /* NOTREACHED */
1042 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1043 return 0;
1044}
1045
1046/*
1047 * Check to see if it is time to start boosting RCU readers that are
1048 * blocking the current grace period, and, if so, tell the per-rcu_node
1049 * kthread to start boosting them. If there is an expedited grace
1050 * period in progress, it is always time to boost.
1051 *
1052 * The caller must hold rnp->lock, which this function releases.
1053 * The ->boost_kthread_task is immortal, so we don't need to worry
1054 * about it going away.
1055 */
1056static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1057 __releases(rnp->lock)
1058{
1059 raw_lockdep_assert_held_rcu_node(rnp);
1060 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1061 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1062 return;
1063 }
1064 if (rnp->exp_tasks != NULL ||
1065 (rnp->gp_tasks != NULL &&
1066 rnp->boost_tasks == NULL &&
1067 rnp->qsmask == 0 &&
1068 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld))) {
1069 if (rnp->exp_tasks == NULL)
1070 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1071 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1072 rcu_wake_cond(rnp->boost_kthread_task,
1073 READ_ONCE(rnp->boost_kthread_status));
1074 } else {
1075 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1076 }
1077}
1078
1079/*
1080 * Is the current CPU running the RCU-callbacks kthread?
1081 * Caller must have preemption disabled.
1082 */
1083static bool rcu_is_callbacks_kthread(void)
1084{
1085 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1086}
1087
1088#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1089
1090/*
1091 * Do priority-boost accounting for the start of a new grace period.
1092 */
1093static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1094{
1095 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1096}
1097
1098/*
1099 * Create an RCU-boost kthread for the specified node if one does not
1100 * already exist. We only create this kthread for preemptible RCU.
1101 * Returns zero if all is well, a negated errno otherwise.
1102 */
1103static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1104{
1105 int rnp_index = rnp - rcu_get_root();
1106 unsigned long flags;
1107 struct sched_param sp;
1108 struct task_struct *t;
1109
1110 if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1111 return;
1112
1113 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1114 return;
1115
1116 rcu_state.boost = 1;
1117
1118 if (rnp->boost_kthread_task != NULL)
1119 return;
1120
1121 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1122 "rcub/%d", rnp_index);
1123 if (WARN_ON_ONCE(IS_ERR(t)))
1124 return;
1125
1126 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1127 rnp->boost_kthread_task = t;
1128 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1129 sp.sched_priority = kthread_prio;
1130 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1131 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1132}
1133
1134/*
1135 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1136 * served by the rcu_node in question. The CPU hotplug lock is still
1137 * held, so the value of rnp->qsmaskinit will be stable.
1138 *
1139 * We don't include outgoingcpu in the affinity set, use -1 if there is
1140 * no outgoing CPU. If there are no CPUs left in the affinity set,
1141 * this function allows the kthread to execute on any CPU.
1142 */
1143static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1144{
1145 struct task_struct *t = rnp->boost_kthread_task;
1146 unsigned long mask = rcu_rnp_online_cpus(rnp);
1147 cpumask_var_t cm;
1148 int cpu;
1149
1150 if (!t)
1151 return;
1152 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1153 return;
1154 for_each_leaf_node_possible_cpu(rnp, cpu)
1155 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1156 cpu != outgoingcpu)
1157 cpumask_set_cpu(cpu, cm);
1158 if (cpumask_weight(cm) == 0)
1159 cpumask_setall(cm);
1160 set_cpus_allowed_ptr(t, cm);
1161 free_cpumask_var(cm);
1162}
1163
1164/*
1165 * Spawn boost kthreads -- called as soon as the scheduler is running.
1166 */
1167static void __init rcu_spawn_boost_kthreads(void)
1168{
1169 struct rcu_node *rnp;
1170
1171 rcu_for_each_leaf_node(rnp)
1172 rcu_spawn_one_boost_kthread(rnp);
1173}
1174
1175static void rcu_prepare_kthreads(int cpu)
1176{
1177 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1178 struct rcu_node *rnp = rdp->mynode;
1179
1180 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1181 if (rcu_scheduler_fully_active)
1182 rcu_spawn_one_boost_kthread(rnp);
1183}
1184
1185#else /* #ifdef CONFIG_RCU_BOOST */
1186
1187static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1188 __releases(rnp->lock)
1189{
1190 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1191}
1192
1193static bool rcu_is_callbacks_kthread(void)
1194{
1195 return false;
1196}
1197
1198static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1199{
1200}
1201
1202static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1203{
1204}
1205
1206static void __init rcu_spawn_boost_kthreads(void)
1207{
1208}
1209
1210static void rcu_prepare_kthreads(int cpu)
1211{
1212}
1213
1214#endif /* #else #ifdef CONFIG_RCU_BOOST */
1215
1216#if !defined(CONFIG_RCU_FAST_NO_HZ)
1217
1218/*
1219 * Check to see if any future non-offloaded RCU-related work will need
1220 * to be done by the current CPU, even if none need be done immediately,
1221 * returning 1 if so. This function is part of the RCU implementation;
1222 * it is -not- an exported member of the RCU API.
1223 *
1224 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1225 * CPU has RCU callbacks queued.
1226 */
1227int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1228{
1229 *nextevt = KTIME_MAX;
1230 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1231 !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist);
1232}
1233
1234/*
1235 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1236 * after it.
1237 */
1238static void rcu_cleanup_after_idle(void)
1239{
1240}
1241
1242/*
1243 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1244 * is nothing.
1245 */
1246static void rcu_prepare_for_idle(void)
1247{
1248}
1249
1250#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1251
1252/*
1253 * This code is invoked when a CPU goes idle, at which point we want
1254 * to have the CPU do everything required for RCU so that it can enter
1255 * the energy-efficient dyntick-idle mode.
1256 *
1257 * The following preprocessor symbol controls this:
1258 *
1259 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1260 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1261 * is sized to be roughly one RCU grace period. Those energy-efficiency
1262 * benchmarkers who might otherwise be tempted to set this to a large
1263 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1264 * system. And if you are -that- concerned about energy efficiency,
1265 * just power the system down and be done with it!
1266 *
1267 * The value below works well in practice. If future workloads require
1268 * adjustment, they can be converted into kernel config parameters, though
1269 * making the state machine smarter might be a better option.
1270 */
1271#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1272
1273static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1274module_param(rcu_idle_gp_delay, int, 0644);
1275
1276/*
1277 * Try to advance callbacks on the current CPU, but only if it has been
1278 * awhile since the last time we did so. Afterwards, if there are any
1279 * callbacks ready for immediate invocation, return true.
1280 */
1281static bool __maybe_unused rcu_try_advance_all_cbs(void)
1282{
1283 bool cbs_ready = false;
1284 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1285 struct rcu_node *rnp;
1286
1287 /* Exit early if we advanced recently. */
1288 if (jiffies == rdp->last_advance_all)
1289 return false;
1290 rdp->last_advance_all = jiffies;
1291
1292 rnp = rdp->mynode;
1293
1294 /*
1295 * Don't bother checking unless a grace period has
1296 * completed since we last checked and there are
1297 * callbacks not yet ready to invoke.
1298 */
1299 if ((rcu_seq_completed_gp(rdp->gp_seq,
1300 rcu_seq_current(&rnp->gp_seq)) ||
1301 unlikely(READ_ONCE(rdp->gpwrap))) &&
1302 rcu_segcblist_pend_cbs(&rdp->cblist))
1303 note_gp_changes(rdp);
1304
1305 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1306 cbs_ready = true;
1307 return cbs_ready;
1308}
1309
1310/*
1311 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1312 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1313 * caller about what to set the timeout.
1314 *
1315 * The caller must have disabled interrupts.
1316 */
1317int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1318{
1319 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1320 unsigned long dj;
1321
1322 lockdep_assert_irqs_disabled();
1323
1324 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1325 if (rcu_segcblist_empty(&rdp->cblist) ||
1326 rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist)) {
1327 *nextevt = KTIME_MAX;
1328 return 0;
1329 }
1330
1331 /* Attempt to advance callbacks. */
1332 if (rcu_try_advance_all_cbs()) {
1333 /* Some ready to invoke, so initiate later invocation. */
1334 invoke_rcu_core();
1335 return 1;
1336 }
1337 rdp->last_accelerate = jiffies;
1338
1339 /* Request timer and round. */
1340 dj = round_up(rcu_idle_gp_delay + jiffies, rcu_idle_gp_delay) - jiffies;
1341
1342 *nextevt = basemono + dj * TICK_NSEC;
1343 return 0;
1344}
1345
1346/*
1347 * Prepare a CPU for idle from an RCU perspective. The first major task is to
1348 * sense whether nohz mode has been enabled or disabled via sysfs. The second
1349 * major task is to accelerate (that is, assign grace-period numbers to) any
1350 * recently arrived callbacks.
1351 *
1352 * The caller must have disabled interrupts.
1353 */
1354static void rcu_prepare_for_idle(void)
1355{
1356 bool needwake;
1357 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1358 struct rcu_node *rnp;
1359 int tne;
1360
1361 lockdep_assert_irqs_disabled();
1362 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1363 return;
1364
1365 /* Handle nohz enablement switches conservatively. */
1366 tne = READ_ONCE(tick_nohz_active);
1367 if (tne != rdp->tick_nohz_enabled_snap) {
1368 if (!rcu_segcblist_empty(&rdp->cblist))
1369 invoke_rcu_core(); /* force nohz to see update. */
1370 rdp->tick_nohz_enabled_snap = tne;
1371 return;
1372 }
1373 if (!tne)
1374 return;
1375
1376 /*
1377 * If we have not yet accelerated this jiffy, accelerate all
1378 * callbacks on this CPU.
1379 */
1380 if (rdp->last_accelerate == jiffies)
1381 return;
1382 rdp->last_accelerate = jiffies;
1383 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1384 rnp = rdp->mynode;
1385 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1386 needwake = rcu_accelerate_cbs(rnp, rdp);
1387 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1388 if (needwake)
1389 rcu_gp_kthread_wake();
1390 }
1391}
1392
1393/*
1394 * Clean up for exit from idle. Attempt to advance callbacks based on
1395 * any grace periods that elapsed while the CPU was idle, and if any
1396 * callbacks are now ready to invoke, initiate invocation.
1397 */
1398static void rcu_cleanup_after_idle(void)
1399{
1400 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1401
1402 lockdep_assert_irqs_disabled();
1403 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1404 return;
1405 if (rcu_try_advance_all_cbs())
1406 invoke_rcu_core();
1407}
1408
1409#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1410
1411#ifdef CONFIG_RCU_NOCB_CPU
1412
1413/*
1414 * Offload callback processing from the boot-time-specified set of CPUs
1415 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1416 * created that pull the callbacks from the corresponding CPU, wait for
1417 * a grace period to elapse, and invoke the callbacks. These kthreads
1418 * are organized into GP kthreads, which manage incoming callbacks, wait for
1419 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1420 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1421 * do a wake_up() on their GP kthread when they insert a callback into any
1422 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1423 * in which case each kthread actively polls its CPU. (Which isn't so great
1424 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1425 *
1426 * This is intended to be used in conjunction with Frederic Weisbecker's
1427 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1428 * running CPU-bound user-mode computations.
1429 *
1430 * Offloading of callbacks can also be used as an energy-efficiency
1431 * measure because CPUs with no RCU callbacks queued are more aggressive
1432 * about entering dyntick-idle mode.
1433 */
1434
1435
1436/*
1437 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1438 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1439 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1440 * given, a warning is emitted and all CPUs are offloaded.
1441 */
1442static int __init rcu_nocb_setup(char *str)
1443{
1444 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1445 if (!strcasecmp(str, "all"))
1446 cpumask_setall(rcu_nocb_mask);
1447 else
1448 if (cpulist_parse(str, rcu_nocb_mask)) {
1449 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1450 cpumask_setall(rcu_nocb_mask);
1451 }
1452 return 1;
1453}
1454__setup("rcu_nocbs=", rcu_nocb_setup);
1455
1456static int __init parse_rcu_nocb_poll(char *arg)
1457{
1458 rcu_nocb_poll = true;
1459 return 0;
1460}
1461early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1462
1463/*
1464 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1465 * After all, the main point of bypassing is to avoid lock contention
1466 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1467 */
1468int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1469module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1470
1471/*
1472 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1473 * lock isn't immediately available, increment ->nocb_lock_contended to
1474 * flag the contention.
1475 */
1476static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1477 __acquires(&rdp->nocb_bypass_lock)
1478{
1479 lockdep_assert_irqs_disabled();
1480 if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1481 return;
1482 atomic_inc(&rdp->nocb_lock_contended);
1483 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1484 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1485 raw_spin_lock(&rdp->nocb_bypass_lock);
1486 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1487 atomic_dec(&rdp->nocb_lock_contended);
1488}
1489
1490/*
1491 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1492 * not contended. Please note that this is extremely special-purpose,
1493 * relying on the fact that at most two kthreads and one CPU contend for
1494 * this lock, and also that the two kthreads are guaranteed to have frequent
1495 * grace-period-duration time intervals between successive acquisitions
1496 * of the lock. This allows us to use an extremely simple throttling
1497 * mechanism, and further to apply it only to the CPU doing floods of
1498 * call_rcu() invocations. Don't try this at home!
1499 */
1500static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1501{
1502 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1503 while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1504 cpu_relax();
1505}
1506
1507/*
1508 * Conditionally acquire the specified rcu_data structure's
1509 * ->nocb_bypass_lock.
1510 */
1511static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1512{
1513 lockdep_assert_irqs_disabled();
1514 return raw_spin_trylock(&rdp->nocb_bypass_lock);
1515}
1516
1517/*
1518 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1519 */
1520static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1521 __releases(&rdp->nocb_bypass_lock)
1522{
1523 lockdep_assert_irqs_disabled();
1524 raw_spin_unlock(&rdp->nocb_bypass_lock);
1525}
1526
1527/*
1528 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1529 * if it corresponds to a no-CBs CPU.
1530 */
1531static void rcu_nocb_lock(struct rcu_data *rdp)
1532{
1533 lockdep_assert_irqs_disabled();
1534 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1535 return;
1536 raw_spin_lock(&rdp->nocb_lock);
1537}
1538
1539/*
1540 * Release the specified rcu_data structure's ->nocb_lock, but only
1541 * if it corresponds to a no-CBs CPU.
1542 */
1543static void rcu_nocb_unlock(struct rcu_data *rdp)
1544{
1545 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1546 lockdep_assert_irqs_disabled();
1547 raw_spin_unlock(&rdp->nocb_lock);
1548 }
1549}
1550
1551/*
1552 * Release the specified rcu_data structure's ->nocb_lock and restore
1553 * interrupts, but only if it corresponds to a no-CBs CPU.
1554 */
1555static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1556 unsigned long flags)
1557{
1558 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1559 lockdep_assert_irqs_disabled();
1560 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1561 } else {
1562 local_irq_restore(flags);
1563 }
1564}
1565
1566/* Lockdep check that ->cblist may be safely accessed. */
1567static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1568{
1569 lockdep_assert_irqs_disabled();
1570 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1571 lockdep_assert_held(&rdp->nocb_lock);
1572}
1573
1574/*
1575 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1576 * grace period.
1577 */
1578static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1579{
1580 swake_up_all(sq);
1581}
1582
1583static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1584{
1585 return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1586}
1587
1588static void rcu_init_one_nocb(struct rcu_node *rnp)
1589{
1590 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1591 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1592}
1593
1594/* Is the specified CPU a no-CBs CPU? */
1595bool rcu_is_nocb_cpu(int cpu)
1596{
1597 if (cpumask_available(rcu_nocb_mask))
1598 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1599 return false;
1600}
1601
1602/*
1603 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1604 * and this function releases it.
1605 */
1606static void wake_nocb_gp(struct rcu_data *rdp, bool force,
1607 unsigned long flags)
1608 __releases(rdp->nocb_lock)
1609{
1610 bool needwake = false;
1611 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1612
1613 lockdep_assert_held(&rdp->nocb_lock);
1614 if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1615 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1616 TPS("AlreadyAwake"));
1617 rcu_nocb_unlock_irqrestore(rdp, flags);
1618 return;
1619 }
1620 del_timer(&rdp->nocb_timer);
1621 rcu_nocb_unlock_irqrestore(rdp, flags);
1622 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1623 if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1624 WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1625 needwake = true;
1626 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1627 }
1628 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1629 if (needwake)
1630 wake_up_process(rdp_gp->nocb_gp_kthread);
1631}
1632
1633/*
1634 * Arrange to wake the GP kthread for this NOCB group at some future
1635 * time when it is safe to do so.
1636 */
1637static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1638 const char *reason)
1639{
1640 if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1641 mod_timer(&rdp->nocb_timer, jiffies + 1);
1642 if (rdp->nocb_defer_wakeup < waketype)
1643 WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1644 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1645}
1646
1647/*
1648 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1649 * However, if there is a callback to be enqueued and if ->nocb_bypass
1650 * proves to be initially empty, just return false because the no-CB GP
1651 * kthread may need to be awakened in this case.
1652 *
1653 * Note that this function always returns true if rhp is NULL.
1654 */
1655static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1656 unsigned long j)
1657{
1658 struct rcu_cblist rcl;
1659
1660 WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp->cblist));
1661 rcu_lockdep_assert_cblist_protected(rdp);
1662 lockdep_assert_held(&rdp->nocb_bypass_lock);
1663 if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1664 raw_spin_unlock(&rdp->nocb_bypass_lock);
1665 return false;
1666 }
1667 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1668 if (rhp)
1669 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1670 rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1671 rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1672 WRITE_ONCE(rdp->nocb_bypass_first, j);
1673 rcu_nocb_bypass_unlock(rdp);
1674 return true;
1675}
1676
1677/*
1678 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1679 * However, if there is a callback to be enqueued and if ->nocb_bypass
1680 * proves to be initially empty, just return false because the no-CB GP
1681 * kthread may need to be awakened in this case.
1682 *
1683 * Note that this function always returns true if rhp is NULL.
1684 */
1685static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1686 unsigned long j)
1687{
1688 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1689 return true;
1690 rcu_lockdep_assert_cblist_protected(rdp);
1691 rcu_nocb_bypass_lock(rdp);
1692 return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1693}
1694
1695/*
1696 * If the ->nocb_bypass_lock is immediately available, flush the
1697 * ->nocb_bypass queue into ->cblist.
1698 */
1699static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1700{
1701 rcu_lockdep_assert_cblist_protected(rdp);
1702 if (!rcu_segcblist_is_offloaded(&rdp->cblist) ||
1703 !rcu_nocb_bypass_trylock(rdp))
1704 return;
1705 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1706}
1707
1708/*
1709 * See whether it is appropriate to use the ->nocb_bypass list in order
1710 * to control contention on ->nocb_lock. A limited number of direct
1711 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1712 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1713 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1714 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1715 * used if ->cblist is empty, because otherwise callbacks can be stranded
1716 * on ->nocb_bypass because we cannot count on the current CPU ever again
1717 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1718 * non-empty, the corresponding no-CBs grace-period kthread must not be
1719 * in an indefinite sleep state.
1720 *
1721 * Finally, it is not permitted to use the bypass during early boot,
1722 * as doing so would confuse the auto-initialization code. Besides
1723 * which, there is no point in worrying about lock contention while
1724 * there is only one CPU in operation.
1725 */
1726static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1727 bool *was_alldone, unsigned long flags)
1728{
1729 unsigned long c;
1730 unsigned long cur_gp_seq;
1731 unsigned long j = jiffies;
1732 long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1733
1734 if (!rcu_segcblist_is_offloaded(&rdp->cblist)) {
1735 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1736 return false; /* Not offloaded, no bypassing. */
1737 }
1738 lockdep_assert_irqs_disabled();
1739
1740 // Don't use ->nocb_bypass during early boot.
1741 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1742 rcu_nocb_lock(rdp);
1743 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1744 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1745 return false;
1746 }
1747
1748 // If we have advanced to a new jiffy, reset counts to allow
1749 // moving back from ->nocb_bypass to ->cblist.
1750 if (j == rdp->nocb_nobypass_last) {
1751 c = rdp->nocb_nobypass_count + 1;
1752 } else {
1753 WRITE_ONCE(rdp->nocb_nobypass_last, j);
1754 c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1755 if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1756 nocb_nobypass_lim_per_jiffy))
1757 c = 0;
1758 else if (c > nocb_nobypass_lim_per_jiffy)
1759 c = nocb_nobypass_lim_per_jiffy;
1760 }
1761 WRITE_ONCE(rdp->nocb_nobypass_count, c);
1762
1763 // If there hasn't yet been all that many ->cblist enqueues
1764 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1765 // ->nocb_bypass first.
1766 if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1767 rcu_nocb_lock(rdp);
1768 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1769 if (*was_alldone)
1770 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1771 TPS("FirstQ"));
1772 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1773 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1774 return false; // Caller must enqueue the callback.
1775 }
1776
1777 // If ->nocb_bypass has been used too long or is too full,
1778 // flush ->nocb_bypass to ->cblist.
1779 if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1780 ncbs >= qhimark) {
1781 rcu_nocb_lock(rdp);
1782 if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1783 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1784 if (*was_alldone)
1785 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1786 TPS("FirstQ"));
1787 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1788 return false; // Caller must enqueue the callback.
1789 }
1790 if (j != rdp->nocb_gp_adv_time &&
1791 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1792 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1793 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1794 rdp->nocb_gp_adv_time = j;
1795 }
1796 rcu_nocb_unlock_irqrestore(rdp, flags);
1797 return true; // Callback already enqueued.
1798 }
1799
1800 // We need to use the bypass.
1801 rcu_nocb_wait_contended(rdp);
1802 rcu_nocb_bypass_lock(rdp);
1803 ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1804 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1805 rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1806 if (!ncbs) {
1807 WRITE_ONCE(rdp->nocb_bypass_first, j);
1808 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1809 }
1810 rcu_nocb_bypass_unlock(rdp);
1811 smp_mb(); /* Order enqueue before wake. */
1812 if (ncbs) {
1813 local_irq_restore(flags);
1814 } else {
1815 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1816 rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1817 if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1818 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1819 TPS("FirstBQwake"));
1820 __call_rcu_nocb_wake(rdp, true, flags);
1821 } else {
1822 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1823 TPS("FirstBQnoWake"));
1824 rcu_nocb_unlock_irqrestore(rdp, flags);
1825 }
1826 }
1827 return true; // Callback already enqueued.
1828}
1829
1830/*
1831 * Awaken the no-CBs grace-period kthead if needed, either due to it
1832 * legitimately being asleep or due to overload conditions.
1833 *
1834 * If warranted, also wake up the kthread servicing this CPUs queues.
1835 */
1836static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1837 unsigned long flags)
1838 __releases(rdp->nocb_lock)
1839{
1840 unsigned long cur_gp_seq;
1841 unsigned long j;
1842 long len;
1843 struct task_struct *t;
1844
1845 // If we are being polled or there is no kthread, just leave.
1846 t = READ_ONCE(rdp->nocb_gp_kthread);
1847 if (rcu_nocb_poll || !t) {
1848 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1849 TPS("WakeNotPoll"));
1850 rcu_nocb_unlock_irqrestore(rdp, flags);
1851 return;
1852 }
1853 // Need to actually to a wakeup.
1854 len = rcu_segcblist_n_cbs(&rdp->cblist);
1855 if (was_alldone) {
1856 rdp->qlen_last_fqs_check = len;
1857 if (!irqs_disabled_flags(flags)) {
1858 /* ... if queue was empty ... */
1859 wake_nocb_gp(rdp, false, flags);
1860 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1861 TPS("WakeEmpty"));
1862 } else {
1863 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1864 TPS("WakeEmptyIsDeferred"));
1865 rcu_nocb_unlock_irqrestore(rdp, flags);
1866 }
1867 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1868 /* ... or if many callbacks queued. */
1869 rdp->qlen_last_fqs_check = len;
1870 j = jiffies;
1871 if (j != rdp->nocb_gp_adv_time &&
1872 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1873 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1874 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1875 rdp->nocb_gp_adv_time = j;
1876 }
1877 smp_mb(); /* Enqueue before timer_pending(). */
1878 if ((rdp->nocb_cb_sleep ||
1879 !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1880 !timer_pending(&rdp->nocb_bypass_timer))
1881 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1882 TPS("WakeOvfIsDeferred"));
1883 rcu_nocb_unlock_irqrestore(rdp, flags);
1884 } else {
1885 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1886 rcu_nocb_unlock_irqrestore(rdp, flags);
1887 }
1888 return;
1889}
1890
1891/* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1892static void do_nocb_bypass_wakeup_timer(struct timer_list *t)
1893{
1894 unsigned long flags;
1895 struct rcu_data *rdp = from_timer(rdp, t, nocb_bypass_timer);
1896
1897 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1898 rcu_nocb_lock_irqsave(rdp, flags);
1899 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1900 __call_rcu_nocb_wake(rdp, true, flags);
1901}
1902
1903/*
1904 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1905 * or for grace periods to end.
1906 */
1907static void nocb_gp_wait(struct rcu_data *my_rdp)
1908{
1909 bool bypass = false;
1910 long bypass_ncbs;
1911 int __maybe_unused cpu = my_rdp->cpu;
1912 unsigned long cur_gp_seq;
1913 unsigned long flags;
1914 bool gotcbs = false;
1915 unsigned long j = jiffies;
1916 bool needwait_gp = false; // This prevents actual uninitialized use.
1917 bool needwake;
1918 bool needwake_gp;
1919 struct rcu_data *rdp;
1920 struct rcu_node *rnp;
1921 unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
1922 bool wasempty = false;
1923
1924 /*
1925 * Each pass through the following loop checks for CBs and for the
1926 * nearest grace period (if any) to wait for next. The CB kthreads
1927 * and the global grace-period kthread are awakened if needed.
1928 */
1929 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1930 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
1931 rcu_nocb_lock_irqsave(rdp, flags);
1932 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1933 if (bypass_ncbs &&
1934 (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
1935 bypass_ncbs > 2 * qhimark)) {
1936 // Bypass full or old, so flush it.
1937 (void)rcu_nocb_try_flush_bypass(rdp, j);
1938 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1939 } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
1940 rcu_nocb_unlock_irqrestore(rdp, flags);
1941 continue; /* No callbacks here, try next. */
1942 }
1943 if (bypass_ncbs) {
1944 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1945 TPS("Bypass"));
1946 bypass = true;
1947 }
1948 rnp = rdp->mynode;
1949 if (bypass) { // Avoid race with first bypass CB.
1950 WRITE_ONCE(my_rdp->nocb_defer_wakeup,
1951 RCU_NOCB_WAKE_NOT);
1952 del_timer(&my_rdp->nocb_timer);
1953 }
1954 // Advance callbacks if helpful and low contention.
1955 needwake_gp = false;
1956 if (!rcu_segcblist_restempty(&rdp->cblist,
1957 RCU_NEXT_READY_TAIL) ||
1958 (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1959 rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
1960 raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
1961 needwake_gp = rcu_advance_cbs(rnp, rdp);
1962 wasempty = rcu_segcblist_restempty(&rdp->cblist,
1963 RCU_NEXT_READY_TAIL);
1964 raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
1965 }
1966 // Need to wait on some grace period?
1967 WARN_ON_ONCE(wasempty &&
1968 !rcu_segcblist_restempty(&rdp->cblist,
1969 RCU_NEXT_READY_TAIL));
1970 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
1971 if (!needwait_gp ||
1972 ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
1973 wait_gp_seq = cur_gp_seq;
1974 needwait_gp = true;
1975 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1976 TPS("NeedWaitGP"));
1977 }
1978 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
1979 needwake = rdp->nocb_cb_sleep;
1980 WRITE_ONCE(rdp->nocb_cb_sleep, false);
1981 smp_mb(); /* CB invocation -after- GP end. */
1982 } else {
1983 needwake = false;
1984 }
1985 rcu_nocb_unlock_irqrestore(rdp, flags);
1986 if (needwake) {
1987 swake_up_one(&rdp->nocb_cb_wq);
1988 gotcbs = true;
1989 }
1990 if (needwake_gp)
1991 rcu_gp_kthread_wake();
1992 }
1993
1994 my_rdp->nocb_gp_bypass = bypass;
1995 my_rdp->nocb_gp_gp = needwait_gp;
1996 my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
1997 if (bypass && !rcu_nocb_poll) {
1998 // At least one child with non-empty ->nocb_bypass, so set
1999 // timer in order to avoid stranding its callbacks.
2000 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2001 mod_timer(&my_rdp->nocb_bypass_timer, j + 2);
2002 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2003 }
2004 if (rcu_nocb_poll) {
2005 /* Polling, so trace if first poll in the series. */
2006 if (gotcbs)
2007 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2008 schedule_timeout_idle(1);
2009 } else if (!needwait_gp) {
2010 /* Wait for callbacks to appear. */
2011 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2012 swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2013 !READ_ONCE(my_rdp->nocb_gp_sleep));
2014 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2015 } else {
2016 rnp = my_rdp->mynode;
2017 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2018 swait_event_interruptible_exclusive(
2019 rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2020 rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2021 !READ_ONCE(my_rdp->nocb_gp_sleep));
2022 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2023 }
2024 if (!rcu_nocb_poll) {
2025 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2026 if (bypass)
2027 del_timer(&my_rdp->nocb_bypass_timer);
2028 WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2029 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2030 }
2031 my_rdp->nocb_gp_seq = -1;
2032 WARN_ON(signal_pending(current));
2033}
2034
2035/*
2036 * No-CBs grace-period-wait kthread. There is one of these per group
2037 * of CPUs, but only once at least one CPU in that group has come online
2038 * at least once since boot. This kthread checks for newly posted
2039 * callbacks from any of the CPUs it is responsible for, waits for a
2040 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2041 * that then have callback-invocation work to do.
2042 */
2043static int rcu_nocb_gp_kthread(void *arg)
2044{
2045 struct rcu_data *rdp = arg;
2046
2047 for (;;) {
2048 WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2049 nocb_gp_wait(rdp);
2050 cond_resched_tasks_rcu_qs();
2051 }
2052 return 0;
2053}
2054
2055/*
2056 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2057 * then, if there are no more, wait for more to appear.
2058 */
2059static void nocb_cb_wait(struct rcu_data *rdp)
2060{
2061 unsigned long cur_gp_seq;
2062 unsigned long flags;
2063 bool needwake_gp = false;
2064 struct rcu_node *rnp = rdp->mynode;
2065
2066 local_irq_save(flags);
2067 rcu_momentary_dyntick_idle();
2068 local_irq_restore(flags);
2069 local_bh_disable();
2070 rcu_do_batch(rdp);
2071 local_bh_enable();
2072 lockdep_assert_irqs_enabled();
2073 rcu_nocb_lock_irqsave(rdp, flags);
2074 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2075 rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2076 raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2077 needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2078 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2079 }
2080 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2081 rcu_nocb_unlock_irqrestore(rdp, flags);
2082 if (needwake_gp)
2083 rcu_gp_kthread_wake();
2084 return;
2085 }
2086
2087 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2088 WRITE_ONCE(rdp->nocb_cb_sleep, true);
2089 rcu_nocb_unlock_irqrestore(rdp, flags);
2090 if (needwake_gp)
2091 rcu_gp_kthread_wake();
2092 swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2093 !READ_ONCE(rdp->nocb_cb_sleep));
2094 if (!smp_load_acquire(&rdp->nocb_cb_sleep)) { /* VVV */
2095 /* ^^^ Ensure CB invocation follows _sleep test. */
2096 return;
2097 }
2098 WARN_ON(signal_pending(current));
2099 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2100}
2101
2102/*
2103 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2104 * nocb_cb_wait() to do the dirty work.
2105 */
2106static int rcu_nocb_cb_kthread(void *arg)
2107{
2108 struct rcu_data *rdp = arg;
2109
2110 // Each pass through this loop does one callback batch, and,
2111 // if there are no more ready callbacks, waits for them.
2112 for (;;) {
2113 nocb_cb_wait(rdp);
2114 cond_resched_tasks_rcu_qs();
2115 }
2116 return 0;
2117}
2118
2119/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2120static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2121{
2122 return READ_ONCE(rdp->nocb_defer_wakeup);
2123}
2124
2125/* Do a deferred wakeup of rcu_nocb_kthread(). */
2126static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2127{
2128 unsigned long flags;
2129 int ndw;
2130
2131 rcu_nocb_lock_irqsave(rdp, flags);
2132 if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2133 rcu_nocb_unlock_irqrestore(rdp, flags);
2134 return;
2135 }
2136 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2137 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2138 wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2139 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2140}
2141
2142/* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2143static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2144{
2145 struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2146
2147 do_nocb_deferred_wakeup_common(rdp);
2148}
2149
2150/*
2151 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2152 * This means we do an inexact common-case check. Note that if
2153 * we miss, ->nocb_timer will eventually clean things up.
2154 */
2155static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2156{
2157 if (rcu_nocb_need_deferred_wakeup(rdp))
2158 do_nocb_deferred_wakeup_common(rdp);
2159}
2160
2161void __init rcu_init_nohz(void)
2162{
2163 int cpu;
2164 bool need_rcu_nocb_mask = false;
2165 struct rcu_data *rdp;
2166
2167#if defined(CONFIG_NO_HZ_FULL)
2168 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2169 need_rcu_nocb_mask = true;
2170#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2171
2172 if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2173 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2174 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2175 return;
2176 }
2177 }
2178 if (!cpumask_available(rcu_nocb_mask))
2179 return;
2180
2181#if defined(CONFIG_NO_HZ_FULL)
2182 if (tick_nohz_full_running)
2183 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2184#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2185
2186 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2187 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2188 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2189 rcu_nocb_mask);
2190 }
2191 if (cpumask_empty(rcu_nocb_mask))
2192 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2193 else
2194 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2195 cpumask_pr_args(rcu_nocb_mask));
2196 if (rcu_nocb_poll)
2197 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2198
2199 for_each_cpu(cpu, rcu_nocb_mask) {
2200 rdp = per_cpu_ptr(&rcu_data, cpu);
2201 if (rcu_segcblist_empty(&rdp->cblist))
2202 rcu_segcblist_init(&rdp->cblist);
2203 rcu_segcblist_offload(&rdp->cblist);
2204 }
2205 rcu_organize_nocb_kthreads();
2206}
2207
2208/* Initialize per-rcu_data variables for no-CBs CPUs. */
2209static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2210{
2211 init_swait_queue_head(&rdp->nocb_cb_wq);
2212 init_swait_queue_head(&rdp->nocb_gp_wq);
2213 raw_spin_lock_init(&rdp->nocb_lock);
2214 raw_spin_lock_init(&rdp->nocb_bypass_lock);
2215 raw_spin_lock_init(&rdp->nocb_gp_lock);
2216 timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2217 timer_setup(&rdp->nocb_bypass_timer, do_nocb_bypass_wakeup_timer, 0);
2218 rcu_cblist_init(&rdp->nocb_bypass);
2219}
2220
2221/*
2222 * If the specified CPU is a no-CBs CPU that does not already have its
2223 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2224 * for this CPU's group has not yet been created, spawn it as well.
2225 */
2226static void rcu_spawn_one_nocb_kthread(int cpu)
2227{
2228 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2229 struct rcu_data *rdp_gp;
2230 struct task_struct *t;
2231
2232 /*
2233 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2234 * then nothing to do.
2235 */
2236 if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2237 return;
2238
2239 /* If we didn't spawn the GP kthread first, reorganize! */
2240 rdp_gp = rdp->nocb_gp_rdp;
2241 if (!rdp_gp->nocb_gp_kthread) {
2242 t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2243 "rcuog/%d", rdp_gp->cpu);
2244 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2245 return;
2246 WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2247 }
2248
2249 /* Spawn the kthread for this CPU. */
2250 t = kthread_run(rcu_nocb_cb_kthread, rdp,
2251 "rcuo%c/%d", rcu_state.abbr, cpu);
2252 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2253 return;
2254 WRITE_ONCE(rdp->nocb_cb_kthread, t);
2255 WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2256}
2257
2258/*
2259 * If the specified CPU is a no-CBs CPU that does not already have its
2260 * rcuo kthread, spawn it.
2261 */
2262static void rcu_spawn_cpu_nocb_kthread(int cpu)
2263{
2264 if (rcu_scheduler_fully_active)
2265 rcu_spawn_one_nocb_kthread(cpu);
2266}
2267
2268/*
2269 * Once the scheduler is running, spawn rcuo kthreads for all online
2270 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2271 * non-boot CPUs come online -- if this changes, we will need to add
2272 * some mutual exclusion.
2273 */
2274static void __init rcu_spawn_nocb_kthreads(void)
2275{
2276 int cpu;
2277
2278 for_each_online_cpu(cpu)
2279 rcu_spawn_cpu_nocb_kthread(cpu);
2280}
2281
2282/* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2283static int rcu_nocb_gp_stride = -1;
2284module_param(rcu_nocb_gp_stride, int, 0444);
2285
2286/*
2287 * Initialize GP-CB relationships for all no-CBs CPU.
2288 */
2289static void __init rcu_organize_nocb_kthreads(void)
2290{
2291 int cpu;
2292 bool firsttime = true;
2293 bool gotnocbs = false;
2294 bool gotnocbscbs = true;
2295 int ls = rcu_nocb_gp_stride;
2296 int nl = 0; /* Next GP kthread. */
2297 struct rcu_data *rdp;
2298 struct rcu_data *rdp_gp = NULL; /* Suppress misguided gcc warn. */
2299 struct rcu_data *rdp_prev = NULL;
2300
2301 if (!cpumask_available(rcu_nocb_mask))
2302 return;
2303 if (ls == -1) {
2304 ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2305 rcu_nocb_gp_stride = ls;
2306 }
2307
2308 /*
2309 * Each pass through this loop sets up one rcu_data structure.
2310 * Should the corresponding CPU come online in the future, then
2311 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2312 */
2313 for_each_cpu(cpu, rcu_nocb_mask) {
2314 rdp = per_cpu_ptr(&rcu_data, cpu);
2315 if (rdp->cpu >= nl) {
2316 /* New GP kthread, set up for CBs & next GP. */
2317 gotnocbs = true;
2318 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2319 rdp->nocb_gp_rdp = rdp;
2320 rdp_gp = rdp;
2321 if (dump_tree) {
2322 if (!firsttime)
2323 pr_cont("%s\n", gotnocbscbs
2324 ? "" : " (self only)");
2325 gotnocbscbs = false;
2326 firsttime = false;
2327 pr_alert("%s: No-CB GP kthread CPU %d:",
2328 __func__, cpu);
2329 }
2330 } else {
2331 /* Another CB kthread, link to previous GP kthread. */
2332 gotnocbscbs = true;
2333 rdp->nocb_gp_rdp = rdp_gp;
2334 rdp_prev->nocb_next_cb_rdp = rdp;
2335 if (dump_tree)
2336 pr_cont(" %d", cpu);
2337 }
2338 rdp_prev = rdp;
2339 }
2340 if (gotnocbs && dump_tree)
2341 pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
2342}
2343
2344/*
2345 * Bind the current task to the offloaded CPUs. If there are no offloaded
2346 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2347 */
2348void rcu_bind_current_to_nocb(void)
2349{
2350 if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2351 WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2352}
2353EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2354
2355/*
2356 * Dump out nocb grace-period kthread state for the specified rcu_data
2357 * structure.
2358 */
2359static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2360{
2361 struct rcu_node *rnp = rdp->mynode;
2362
2363 pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2364 rdp->cpu,
2365 "kK"[!!rdp->nocb_gp_kthread],
2366 "lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2367 "dD"[!!rdp->nocb_defer_wakeup],
2368 "tT"[timer_pending(&rdp->nocb_timer)],
2369 "bB"[timer_pending(&rdp->nocb_bypass_timer)],
2370 "sS"[!!rdp->nocb_gp_sleep],
2371 ".W"[swait_active(&rdp->nocb_gp_wq)],
2372 ".W"[swait_active(&rnp->nocb_gp_wq[0])],
2373 ".W"[swait_active(&rnp->nocb_gp_wq[1])],
2374 ".B"[!!rdp->nocb_gp_bypass],
2375 ".G"[!!rdp->nocb_gp_gp],
2376 (long)rdp->nocb_gp_seq,
2377 rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops));
2378}
2379
2380/* Dump out nocb kthread state for the specified rcu_data structure. */
2381static void show_rcu_nocb_state(struct rcu_data *rdp)
2382{
2383 struct rcu_segcblist *rsclp = &rdp->cblist;
2384 bool waslocked;
2385 bool wastimer;
2386 bool wassleep;
2387
2388 if (rdp->nocb_gp_rdp == rdp)
2389 show_rcu_nocb_gp_state(rdp);
2390
2391 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",
2392 rdp->cpu, rdp->nocb_gp_rdp->cpu,
2393 "kK"[!!rdp->nocb_cb_kthread],
2394 "bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2395 "cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2396 "lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2397 "sS"[!!rdp->nocb_cb_sleep],
2398 ".W"[swait_active(&rdp->nocb_cb_wq)],
2399 jiffies - rdp->nocb_bypass_first,
2400 jiffies - rdp->nocb_nobypass_last,
2401 rdp->nocb_nobypass_count,
2402 ".D"[rcu_segcblist_ready_cbs(rsclp)],
2403 ".W"[!rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL)],
2404 ".R"[!rcu_segcblist_restempty(rsclp, RCU_WAIT_TAIL)],
2405 ".N"[!rcu_segcblist_restempty(rsclp, RCU_NEXT_READY_TAIL)],
2406 ".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2407 rcu_segcblist_n_cbs(&rdp->cblist));
2408
2409 /* It is OK for GP kthreads to have GP state. */
2410 if (rdp->nocb_gp_rdp == rdp)
2411 return;
2412
2413 waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2414 wastimer = timer_pending(&rdp->nocb_timer);
2415 wassleep = swait_active(&rdp->nocb_gp_wq);
2416 if (!rdp->nocb_defer_wakeup && !rdp->nocb_gp_sleep &&
2417 !waslocked && !wastimer && !wassleep)
2418 return; /* Nothing untowards. */
2419
2420 pr_info(" !!! %c%c%c%c %c\n",
2421 "lL"[waslocked],
2422 "dD"[!!rdp->nocb_defer_wakeup],
2423 "tT"[wastimer],
2424 "sS"[!!rdp->nocb_gp_sleep],
2425 ".W"[wassleep]);
2426}
2427
2428#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2429
2430/* No ->nocb_lock to acquire. */
2431static void rcu_nocb_lock(struct rcu_data *rdp)
2432{
2433}
2434
2435/* No ->nocb_lock to release. */
2436static void rcu_nocb_unlock(struct rcu_data *rdp)
2437{
2438}
2439
2440/* No ->nocb_lock to release. */
2441static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2442 unsigned long flags)
2443{
2444 local_irq_restore(flags);
2445}
2446
2447/* Lockdep check that ->cblist may be safely accessed. */
2448static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2449{
2450 lockdep_assert_irqs_disabled();
2451}
2452
2453static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2454{
2455}
2456
2457static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2458{
2459 return NULL;
2460}
2461
2462static void rcu_init_one_nocb(struct rcu_node *rnp)
2463{
2464}
2465
2466static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2467 unsigned long j)
2468{
2469 return true;
2470}
2471
2472static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2473 bool *was_alldone, unsigned long flags)
2474{
2475 return false;
2476}
2477
2478static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2479 unsigned long flags)
2480{
2481 WARN_ON_ONCE(1); /* Should be dead code! */
2482}
2483
2484static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2485{
2486}
2487
2488static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2489{
2490 return false;
2491}
2492
2493static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2494{
2495}
2496
2497static void rcu_spawn_cpu_nocb_kthread(int cpu)
2498{
2499}
2500
2501static void __init rcu_spawn_nocb_kthreads(void)
2502{
2503}
2504
2505static void show_rcu_nocb_state(struct rcu_data *rdp)
2506{
2507}
2508
2509#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2510
2511/*
2512 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2513 * grace-period kthread will do force_quiescent_state() processing?
2514 * The idea is to avoid waking up RCU core processing on such a
2515 * CPU unless the grace period has extended for too long.
2516 *
2517 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2518 * CONFIG_RCU_NOCB_CPU CPUs.
2519 */
2520static bool rcu_nohz_full_cpu(void)
2521{
2522#ifdef CONFIG_NO_HZ_FULL
2523 if (tick_nohz_full_cpu(smp_processor_id()) &&
2524 (!rcu_gp_in_progress() ||
2525 time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2526 return true;
2527#endif /* #ifdef CONFIG_NO_HZ_FULL */
2528 return false;
2529}
2530
2531/*
2532 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2533 */
2534static void rcu_bind_gp_kthread(void)
2535{
2536 if (!tick_nohz_full_enabled())
2537 return;
2538 housekeeping_affine(current, HK_FLAG_RCU);
2539}
2540
2541/* Record the current task on dyntick-idle entry. */
2542static void noinstr rcu_dynticks_task_enter(void)
2543{
2544#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2545 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2546#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2547}
2548
2549/* Record no current task on dyntick-idle exit. */
2550static void noinstr rcu_dynticks_task_exit(void)
2551{
2552#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2553 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2554#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2555}
2556
2557/* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
2558static void rcu_dynticks_task_trace_enter(void)
2559{
2560#ifdef CONFIG_TASKS_RCU_TRACE
2561 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2562 current->trc_reader_special.b.need_mb = true;
2563#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */
2564}
2565
2566/* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
2567static void rcu_dynticks_task_trace_exit(void)
2568{
2569#ifdef CONFIG_TASKS_RCU_TRACE
2570 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2571 current->trc_reader_special.b.need_mb = false;
2572#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */
2573}
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions that provide either classic
5 * or preemptible semantics.
6 *
7 * Copyright Red Hat, 2009
8 * Copyright IBM Corporation, 2009
9 *
10 * Author: Ingo Molnar <mingo@elte.hu>
11 * Paul E. McKenney <paulmck@linux.ibm.com>
12 */
13
14#include "../locking/rtmutex_common.h"
15
16#ifdef CONFIG_RCU_NOCB_CPU
17static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
18static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
19static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
20{
21 return lockdep_is_held(&rdp->nocb_lock);
22}
23
24static inline bool rcu_current_is_nocb_kthread(struct rcu_data *rdp)
25{
26 /* Race on early boot between thread creation and assignment */
27 if (!rdp->nocb_cb_kthread || !rdp->nocb_gp_kthread)
28 return true;
29
30 if (current == rdp->nocb_cb_kthread || current == rdp->nocb_gp_kthread)
31 if (in_task())
32 return true;
33 return false;
34}
35
36#else
37static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
38{
39 return 0;
40}
41
42static inline bool rcu_current_is_nocb_kthread(struct rcu_data *rdp)
43{
44 return false;
45}
46
47#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
48
49static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
50{
51 /*
52 * In order to read the offloaded state of an rdp is a safe
53 * and stable way and prevent from its value to be changed
54 * under us, we must either hold the barrier mutex, the cpu
55 * hotplug lock (read or write) or the nocb lock. Local
56 * non-preemptible reads are also safe. NOCB kthreads and
57 * timers have their own means of synchronization against the
58 * offloaded state updaters.
59 */
60 RCU_LOCKDEP_WARN(
61 !(lockdep_is_held(&rcu_state.barrier_mutex) ||
62 (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
63 rcu_lockdep_is_held_nocb(rdp) ||
64 (rdp == this_cpu_ptr(&rcu_data) &&
65 !(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible())) ||
66 rcu_current_is_nocb_kthread(rdp)),
67 "Unsafe read of RCU_NOCB offloaded state"
68 );
69
70 return rcu_segcblist_is_offloaded(&rdp->cblist);
71}
72
73/*
74 * Check the RCU kernel configuration parameters and print informative
75 * messages about anything out of the ordinary.
76 */
77static void __init rcu_bootup_announce_oddness(void)
78{
79 if (IS_ENABLED(CONFIG_RCU_TRACE))
80 pr_info("\tRCU event tracing is enabled.\n");
81 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
82 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
83 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
84 RCU_FANOUT);
85 if (rcu_fanout_exact)
86 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
87 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
88 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
89 if (IS_ENABLED(CONFIG_PROVE_RCU))
90 pr_info("\tRCU lockdep checking is enabled.\n");
91 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
92 pr_info("\tRCU strict (and thus non-scalable) grace periods enabled.\n");
93 if (RCU_NUM_LVLS >= 4)
94 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
95 if (RCU_FANOUT_LEAF != 16)
96 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
97 RCU_FANOUT_LEAF);
98 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
99 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
100 rcu_fanout_leaf);
101 if (nr_cpu_ids != NR_CPUS)
102 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
103#ifdef CONFIG_RCU_BOOST
104 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
105 kthread_prio, CONFIG_RCU_BOOST_DELAY);
106#endif
107 if (blimit != DEFAULT_RCU_BLIMIT)
108 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
109 if (qhimark != DEFAULT_RCU_QHIMARK)
110 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
111 if (qlowmark != DEFAULT_RCU_QLOMARK)
112 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
113 if (qovld != DEFAULT_RCU_QOVLD)
114 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
115 if (jiffies_till_first_fqs != ULONG_MAX)
116 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
117 if (jiffies_till_next_fqs != ULONG_MAX)
118 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
119 if (jiffies_till_sched_qs != ULONG_MAX)
120 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
121 if (rcu_kick_kthreads)
122 pr_info("\tKick kthreads if too-long grace period.\n");
123 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
124 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
125 if (gp_preinit_delay)
126 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
127 if (gp_init_delay)
128 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
129 if (gp_cleanup_delay)
130 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
131 if (!use_softirq)
132 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
133 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
134 pr_info("\tRCU debug extended QS entry/exit.\n");
135 rcupdate_announce_bootup_oddness();
136}
137
138#ifdef CONFIG_PREEMPT_RCU
139
140static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
141static void rcu_read_unlock_special(struct task_struct *t);
142
143/*
144 * Tell them what RCU they are running.
145 */
146static void __init rcu_bootup_announce(void)
147{
148 pr_info("Preemptible hierarchical RCU implementation.\n");
149 rcu_bootup_announce_oddness();
150}
151
152/* Flags for rcu_preempt_ctxt_queue() decision table. */
153#define RCU_GP_TASKS 0x8
154#define RCU_EXP_TASKS 0x4
155#define RCU_GP_BLKD 0x2
156#define RCU_EXP_BLKD 0x1
157
158/*
159 * Queues a task preempted within an RCU-preempt read-side critical
160 * section into the appropriate location within the ->blkd_tasks list,
161 * depending on the states of any ongoing normal and expedited grace
162 * periods. The ->gp_tasks pointer indicates which element the normal
163 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
164 * indicates which element the expedited grace period is waiting on (again,
165 * NULL if none). If a grace period is waiting on a given element in the
166 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
167 * adding a task to the tail of the list blocks any grace period that is
168 * already waiting on one of the elements. In contrast, adding a task
169 * to the head of the list won't block any grace period that is already
170 * waiting on one of the elements.
171 *
172 * This queuing is imprecise, and can sometimes make an ongoing grace
173 * period wait for a task that is not strictly speaking blocking it.
174 * Given the choice, we needlessly block a normal grace period rather than
175 * blocking an expedited grace period.
176 *
177 * Note that an endless sequence of expedited grace periods still cannot
178 * indefinitely postpone a normal grace period. Eventually, all of the
179 * fixed number of preempted tasks blocking the normal grace period that are
180 * not also blocking the expedited grace period will resume and complete
181 * their RCU read-side critical sections. At that point, the ->gp_tasks
182 * pointer will equal the ->exp_tasks pointer, at which point the end of
183 * the corresponding expedited grace period will also be the end of the
184 * normal grace period.
185 */
186static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
187 __releases(rnp->lock) /* But leaves rrupts disabled. */
188{
189 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
190 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
191 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
192 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
193 struct task_struct *t = current;
194
195 raw_lockdep_assert_held_rcu_node(rnp);
196 WARN_ON_ONCE(rdp->mynode != rnp);
197 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
198 /* RCU better not be waiting on newly onlined CPUs! */
199 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
200 rdp->grpmask);
201
202 /*
203 * Decide where to queue the newly blocked task. In theory,
204 * this could be an if-statement. In practice, when I tried
205 * that, it was quite messy.
206 */
207 switch (blkd_state) {
208 case 0:
209 case RCU_EXP_TASKS:
210 case RCU_EXP_TASKS + RCU_GP_BLKD:
211 case RCU_GP_TASKS:
212 case RCU_GP_TASKS + RCU_EXP_TASKS:
213
214 /*
215 * Blocking neither GP, or first task blocking the normal
216 * GP but not blocking the already-waiting expedited GP.
217 * Queue at the head of the list to avoid unnecessarily
218 * blocking the already-waiting GPs.
219 */
220 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
221 break;
222
223 case RCU_EXP_BLKD:
224 case RCU_GP_BLKD:
225 case RCU_GP_BLKD + RCU_EXP_BLKD:
226 case RCU_GP_TASKS + RCU_EXP_BLKD:
227 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
228 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
229
230 /*
231 * First task arriving that blocks either GP, or first task
232 * arriving that blocks the expedited GP (with the normal
233 * GP already waiting), or a task arriving that blocks
234 * both GPs with both GPs already waiting. Queue at the
235 * tail of the list to avoid any GP waiting on any of the
236 * already queued tasks that are not blocking it.
237 */
238 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
239 break;
240
241 case RCU_EXP_TASKS + RCU_EXP_BLKD:
242 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
243 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
244
245 /*
246 * Second or subsequent task blocking the expedited GP.
247 * The task either does not block the normal GP, or is the
248 * first task blocking the normal GP. Queue just after
249 * the first task blocking the expedited GP.
250 */
251 list_add(&t->rcu_node_entry, rnp->exp_tasks);
252 break;
253
254 case RCU_GP_TASKS + RCU_GP_BLKD:
255 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
256
257 /*
258 * Second or subsequent task blocking the normal GP.
259 * The task does not block the expedited GP. Queue just
260 * after the first task blocking the normal GP.
261 */
262 list_add(&t->rcu_node_entry, rnp->gp_tasks);
263 break;
264
265 default:
266
267 /* Yet another exercise in excessive paranoia. */
268 WARN_ON_ONCE(1);
269 break;
270 }
271
272 /*
273 * We have now queued the task. If it was the first one to
274 * block either grace period, update the ->gp_tasks and/or
275 * ->exp_tasks pointers, respectively, to reference the newly
276 * blocked tasks.
277 */
278 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
279 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
280 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
281 }
282 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
283 WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
284 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
285 !(rnp->qsmask & rdp->grpmask));
286 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
287 !(rnp->expmask & rdp->grpmask));
288 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
289
290 /*
291 * Report the quiescent state for the expedited GP. This expedited
292 * GP should not be able to end until we report, so there should be
293 * no need to check for a subsequent expedited GP. (Though we are
294 * still in a quiescent state in any case.)
295 */
296 if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
297 rcu_report_exp_rdp(rdp);
298 else
299 WARN_ON_ONCE(rdp->exp_deferred_qs);
300}
301
302/*
303 * Record a preemptible-RCU quiescent state for the specified CPU.
304 * Note that this does not necessarily mean that the task currently running
305 * on the CPU is in a quiescent state: Instead, it means that the current
306 * grace period need not wait on any RCU read-side critical section that
307 * starts later on this CPU. It also means that if the current task is
308 * in an RCU read-side critical section, it has already added itself to
309 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
310 * current task, there might be any number of other tasks blocked while
311 * in an RCU read-side critical section.
312 *
313 * Callers to this function must disable preemption.
314 */
315static void rcu_qs(void)
316{
317 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
318 if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
319 trace_rcu_grace_period(TPS("rcu_preempt"),
320 __this_cpu_read(rcu_data.gp_seq),
321 TPS("cpuqs"));
322 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
323 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
324 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
325 }
326}
327
328/*
329 * We have entered the scheduler, and the current task might soon be
330 * context-switched away from. If this task is in an RCU read-side
331 * critical section, we will no longer be able to rely on the CPU to
332 * record that fact, so we enqueue the task on the blkd_tasks list.
333 * The task will dequeue itself when it exits the outermost enclosing
334 * RCU read-side critical section. Therefore, the current grace period
335 * cannot be permitted to complete until the blkd_tasks list entries
336 * predating the current grace period drain, in other words, until
337 * rnp->gp_tasks becomes NULL.
338 *
339 * Caller must disable interrupts.
340 */
341void rcu_note_context_switch(bool preempt)
342{
343 struct task_struct *t = current;
344 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
345 struct rcu_node *rnp;
346
347 trace_rcu_utilization(TPS("Start context switch"));
348 lockdep_assert_irqs_disabled();
349 WARN_ON_ONCE(!preempt && rcu_preempt_depth() > 0);
350 if (rcu_preempt_depth() > 0 &&
351 !t->rcu_read_unlock_special.b.blocked) {
352
353 /* Possibly blocking in an RCU read-side critical section. */
354 rnp = rdp->mynode;
355 raw_spin_lock_rcu_node(rnp);
356 t->rcu_read_unlock_special.b.blocked = true;
357 t->rcu_blocked_node = rnp;
358
359 /*
360 * Verify the CPU's sanity, trace the preemption, and
361 * then queue the task as required based on the states
362 * of any ongoing and expedited grace periods.
363 */
364 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
365 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
366 trace_rcu_preempt_task(rcu_state.name,
367 t->pid,
368 (rnp->qsmask & rdp->grpmask)
369 ? rnp->gp_seq
370 : rcu_seq_snap(&rnp->gp_seq));
371 rcu_preempt_ctxt_queue(rnp, rdp);
372 } else {
373 rcu_preempt_deferred_qs(t);
374 }
375
376 /*
377 * Either we were not in an RCU read-side critical section to
378 * begin with, or we have now recorded that critical section
379 * globally. Either way, we can now note a quiescent state
380 * for this CPU. Again, if we were in an RCU read-side critical
381 * section, and if that critical section was blocking the current
382 * grace period, then the fact that the task has been enqueued
383 * means that we continue to block the current grace period.
384 */
385 rcu_qs();
386 if (rdp->exp_deferred_qs)
387 rcu_report_exp_rdp(rdp);
388 rcu_tasks_qs(current, preempt);
389 trace_rcu_utilization(TPS("End context switch"));
390}
391EXPORT_SYMBOL_GPL(rcu_note_context_switch);
392
393/*
394 * Check for preempted RCU readers blocking the current grace period
395 * for the specified rcu_node structure. If the caller needs a reliable
396 * answer, it must hold the rcu_node's ->lock.
397 */
398static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
399{
400 return READ_ONCE(rnp->gp_tasks) != NULL;
401}
402
403/* limit value for ->rcu_read_lock_nesting. */
404#define RCU_NEST_PMAX (INT_MAX / 2)
405
406static void rcu_preempt_read_enter(void)
407{
408 current->rcu_read_lock_nesting++;
409}
410
411static int rcu_preempt_read_exit(void)
412{
413 return --current->rcu_read_lock_nesting;
414}
415
416static void rcu_preempt_depth_set(int val)
417{
418 current->rcu_read_lock_nesting = val;
419}
420
421/*
422 * Preemptible RCU implementation for rcu_read_lock().
423 * Just increment ->rcu_read_lock_nesting, shared state will be updated
424 * if we block.
425 */
426void __rcu_read_lock(void)
427{
428 rcu_preempt_read_enter();
429 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
430 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
431 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
432 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
433 barrier(); /* critical section after entry code. */
434}
435EXPORT_SYMBOL_GPL(__rcu_read_lock);
436
437/*
438 * Preemptible RCU implementation for rcu_read_unlock().
439 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
440 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
441 * invoke rcu_read_unlock_special() to clean up after a context switch
442 * in an RCU read-side critical section and other special cases.
443 */
444void __rcu_read_unlock(void)
445{
446 struct task_struct *t = current;
447
448 barrier(); // critical section before exit code.
449 if (rcu_preempt_read_exit() == 0) {
450 barrier(); // critical-section exit before .s check.
451 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
452 rcu_read_unlock_special(t);
453 }
454 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
455 int rrln = rcu_preempt_depth();
456
457 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
458 }
459}
460EXPORT_SYMBOL_GPL(__rcu_read_unlock);
461
462/*
463 * Advance a ->blkd_tasks-list pointer to the next entry, instead
464 * returning NULL if at the end of the list.
465 */
466static struct list_head *rcu_next_node_entry(struct task_struct *t,
467 struct rcu_node *rnp)
468{
469 struct list_head *np;
470
471 np = t->rcu_node_entry.next;
472 if (np == &rnp->blkd_tasks)
473 np = NULL;
474 return np;
475}
476
477/*
478 * Return true if the specified rcu_node structure has tasks that were
479 * preempted within an RCU read-side critical section.
480 */
481static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
482{
483 return !list_empty(&rnp->blkd_tasks);
484}
485
486/*
487 * Report deferred quiescent states. The deferral time can
488 * be quite short, for example, in the case of the call from
489 * rcu_read_unlock_special().
490 */
491static void
492rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
493{
494 bool empty_exp;
495 bool empty_norm;
496 bool empty_exp_now;
497 struct list_head *np;
498 bool drop_boost_mutex = false;
499 struct rcu_data *rdp;
500 struct rcu_node *rnp;
501 union rcu_special special;
502
503 /*
504 * If RCU core is waiting for this CPU to exit its critical section,
505 * report the fact that it has exited. Because irqs are disabled,
506 * t->rcu_read_unlock_special cannot change.
507 */
508 special = t->rcu_read_unlock_special;
509 rdp = this_cpu_ptr(&rcu_data);
510 if (!special.s && !rdp->exp_deferred_qs) {
511 local_irq_restore(flags);
512 return;
513 }
514 t->rcu_read_unlock_special.s = 0;
515 if (special.b.need_qs) {
516 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
517 rcu_report_qs_rdp(rdp);
518 udelay(rcu_unlock_delay);
519 } else {
520 rcu_qs();
521 }
522 }
523
524 /*
525 * Respond to a request by an expedited grace period for a
526 * quiescent state from this CPU. Note that requests from
527 * tasks are handled when removing the task from the
528 * blocked-tasks list below.
529 */
530 if (rdp->exp_deferred_qs)
531 rcu_report_exp_rdp(rdp);
532
533 /* Clean up if blocked during RCU read-side critical section. */
534 if (special.b.blocked) {
535
536 /*
537 * Remove this task from the list it blocked on. The task
538 * now remains queued on the rcu_node corresponding to the
539 * CPU it first blocked on, so there is no longer any need
540 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
541 */
542 rnp = t->rcu_blocked_node;
543 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
544 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
545 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
546 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
547 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
548 (!empty_norm || rnp->qsmask));
549 empty_exp = sync_rcu_exp_done(rnp);
550 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
551 np = rcu_next_node_entry(t, rnp);
552 list_del_init(&t->rcu_node_entry);
553 t->rcu_blocked_node = NULL;
554 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
555 rnp->gp_seq, t->pid);
556 if (&t->rcu_node_entry == rnp->gp_tasks)
557 WRITE_ONCE(rnp->gp_tasks, np);
558 if (&t->rcu_node_entry == rnp->exp_tasks)
559 WRITE_ONCE(rnp->exp_tasks, np);
560 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
561 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
562 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
563 if (&t->rcu_node_entry == rnp->boost_tasks)
564 WRITE_ONCE(rnp->boost_tasks, np);
565 }
566
567 /*
568 * If this was the last task on the current list, and if
569 * we aren't waiting on any CPUs, report the quiescent state.
570 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
571 * so we must take a snapshot of the expedited state.
572 */
573 empty_exp_now = sync_rcu_exp_done(rnp);
574 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
575 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
576 rnp->gp_seq,
577 0, rnp->qsmask,
578 rnp->level,
579 rnp->grplo,
580 rnp->grphi,
581 !!rnp->gp_tasks);
582 rcu_report_unblock_qs_rnp(rnp, flags);
583 } else {
584 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
585 }
586
587 /* Unboost if we were boosted. */
588 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
589 rt_mutex_futex_unlock(&rnp->boost_mtx);
590
591 /*
592 * If this was the last task on the expedited lists,
593 * then we need to report up the rcu_node hierarchy.
594 */
595 if (!empty_exp && empty_exp_now)
596 rcu_report_exp_rnp(rnp, true);
597 } else {
598 local_irq_restore(flags);
599 }
600}
601
602/*
603 * Is a deferred quiescent-state pending, and are we also not in
604 * an RCU read-side critical section? It is the caller's responsibility
605 * to ensure it is otherwise safe to report any deferred quiescent
606 * states. The reason for this is that it is safe to report a
607 * quiescent state during context switch even though preemption
608 * is disabled. This function cannot be expected to understand these
609 * nuances, so the caller must handle them.
610 */
611static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
612{
613 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
614 READ_ONCE(t->rcu_read_unlock_special.s)) &&
615 rcu_preempt_depth() == 0;
616}
617
618/*
619 * Report a deferred quiescent state if needed and safe to do so.
620 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
621 * not being in an RCU read-side critical section. The caller must
622 * evaluate safety in terms of interrupt, softirq, and preemption
623 * disabling.
624 */
625static void rcu_preempt_deferred_qs(struct task_struct *t)
626{
627 unsigned long flags;
628
629 if (!rcu_preempt_need_deferred_qs(t))
630 return;
631 local_irq_save(flags);
632 rcu_preempt_deferred_qs_irqrestore(t, flags);
633}
634
635/*
636 * Minimal handler to give the scheduler a chance to re-evaluate.
637 */
638static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
639{
640 struct rcu_data *rdp;
641
642 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
643 rdp->defer_qs_iw_pending = false;
644}
645
646/*
647 * Handle special cases during rcu_read_unlock(), such as needing to
648 * notify RCU core processing or task having blocked during the RCU
649 * read-side critical section.
650 */
651static void rcu_read_unlock_special(struct task_struct *t)
652{
653 unsigned long flags;
654 bool irqs_were_disabled;
655 bool preempt_bh_were_disabled =
656 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
657
658 /* NMI handlers cannot block and cannot safely manipulate state. */
659 if (in_nmi())
660 return;
661
662 local_irq_save(flags);
663 irqs_were_disabled = irqs_disabled_flags(flags);
664 if (preempt_bh_were_disabled || irqs_were_disabled) {
665 bool expboost; // Expedited GP in flight or possible boosting.
666 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
667 struct rcu_node *rnp = rdp->mynode;
668
669 expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
670 (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
671 IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ||
672 (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
673 t->rcu_blocked_node);
674 // Need to defer quiescent state until everything is enabled.
675 if (use_softirq && (in_irq() || (expboost && !irqs_were_disabled))) {
676 // Using softirq, safe to awaken, and either the
677 // wakeup is free or there is either an expedited
678 // GP in flight or a potential need to deboost.
679 raise_softirq_irqoff(RCU_SOFTIRQ);
680 } else {
681 // Enabling BH or preempt does reschedule, so...
682 // Also if no expediting and no possible deboosting,
683 // slow is OK. Plus nohz_full CPUs eventually get
684 // tick enabled.
685 set_tsk_need_resched(current);
686 set_preempt_need_resched();
687 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
688 expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
689 // Get scheduler to re-evaluate and call hooks.
690 // If !IRQ_WORK, FQS scan will eventually IPI.
691 init_irq_work(&rdp->defer_qs_iw, rcu_preempt_deferred_qs_handler);
692 rdp->defer_qs_iw_pending = true;
693 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
694 }
695 }
696 local_irq_restore(flags);
697 return;
698 }
699 rcu_preempt_deferred_qs_irqrestore(t, flags);
700}
701
702/*
703 * Check that the list of blocked tasks for the newly completed grace
704 * period is in fact empty. It is a serious bug to complete a grace
705 * period that still has RCU readers blocked! This function must be
706 * invoked -before- updating this rnp's ->gp_seq.
707 *
708 * Also, if there are blocked tasks on the list, they automatically
709 * block the newly created grace period, so set up ->gp_tasks accordingly.
710 */
711static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
712{
713 struct task_struct *t;
714
715 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
716 raw_lockdep_assert_held_rcu_node(rnp);
717 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
718 dump_blkd_tasks(rnp, 10);
719 if (rcu_preempt_has_tasks(rnp) &&
720 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
721 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
722 t = container_of(rnp->gp_tasks, struct task_struct,
723 rcu_node_entry);
724 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
725 rnp->gp_seq, t->pid);
726 }
727 WARN_ON_ONCE(rnp->qsmask);
728}
729
730/*
731 * Check for a quiescent state from the current CPU, including voluntary
732 * context switches for Tasks RCU. When a task blocks, the task is
733 * recorded in the corresponding CPU's rcu_node structure, which is checked
734 * elsewhere, hence this function need only check for quiescent states
735 * related to the current CPU, not to those related to tasks.
736 */
737static void rcu_flavor_sched_clock_irq(int user)
738{
739 struct task_struct *t = current;
740
741 lockdep_assert_irqs_disabled();
742 if (user || rcu_is_cpu_rrupt_from_idle()) {
743 rcu_note_voluntary_context_switch(current);
744 }
745 if (rcu_preempt_depth() > 0 ||
746 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
747 /* No QS, force context switch if deferred. */
748 if (rcu_preempt_need_deferred_qs(t)) {
749 set_tsk_need_resched(t);
750 set_preempt_need_resched();
751 }
752 } else if (rcu_preempt_need_deferred_qs(t)) {
753 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
754 return;
755 } else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
756 rcu_qs(); /* Report immediate QS. */
757 return;
758 }
759
760 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
761 if (rcu_preempt_depth() > 0 &&
762 __this_cpu_read(rcu_data.core_needs_qs) &&
763 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
764 !t->rcu_read_unlock_special.b.need_qs &&
765 time_after(jiffies, rcu_state.gp_start + HZ))
766 t->rcu_read_unlock_special.b.need_qs = true;
767}
768
769/*
770 * Check for a task exiting while in a preemptible-RCU read-side
771 * critical section, clean up if so. No need to issue warnings, as
772 * debug_check_no_locks_held() already does this if lockdep is enabled.
773 * Besides, if this function does anything other than just immediately
774 * return, there was a bug of some sort. Spewing warnings from this
775 * function is like as not to simply obscure important prior warnings.
776 */
777void exit_rcu(void)
778{
779 struct task_struct *t = current;
780
781 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
782 rcu_preempt_depth_set(1);
783 barrier();
784 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
785 } else if (unlikely(rcu_preempt_depth())) {
786 rcu_preempt_depth_set(1);
787 } else {
788 return;
789 }
790 __rcu_read_unlock();
791 rcu_preempt_deferred_qs(current);
792}
793
794/*
795 * Dump the blocked-tasks state, but limit the list dump to the
796 * specified number of elements.
797 */
798static void
799dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
800{
801 int cpu;
802 int i;
803 struct list_head *lhp;
804 bool onl;
805 struct rcu_data *rdp;
806 struct rcu_node *rnp1;
807
808 raw_lockdep_assert_held_rcu_node(rnp);
809 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
810 __func__, rnp->grplo, rnp->grphi, rnp->level,
811 (long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
812 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
813 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
814 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
815 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
816 __func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
817 READ_ONCE(rnp->exp_tasks));
818 pr_info("%s: ->blkd_tasks", __func__);
819 i = 0;
820 list_for_each(lhp, &rnp->blkd_tasks) {
821 pr_cont(" %p", lhp);
822 if (++i >= ncheck)
823 break;
824 }
825 pr_cont("\n");
826 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
827 rdp = per_cpu_ptr(&rcu_data, cpu);
828 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
829 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
830 cpu, ".o"[onl],
831 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
832 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
833 }
834}
835
836#else /* #ifdef CONFIG_PREEMPT_RCU */
837
838/*
839 * If strict grace periods are enabled, and if the calling
840 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
841 * report that quiescent state and, if requested, spin for a bit.
842 */
843void rcu_read_unlock_strict(void)
844{
845 struct rcu_data *rdp;
846
847 if (!IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ||
848 irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
849 return;
850 rdp = this_cpu_ptr(&rcu_data);
851 rcu_report_qs_rdp(rdp);
852 udelay(rcu_unlock_delay);
853}
854EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
855
856/*
857 * Tell them what RCU they are running.
858 */
859static void __init rcu_bootup_announce(void)
860{
861 pr_info("Hierarchical RCU implementation.\n");
862 rcu_bootup_announce_oddness();
863}
864
865/*
866 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
867 * how many quiescent states passed, just if there was at least one since
868 * the start of the grace period, this just sets a flag. The caller must
869 * have disabled preemption.
870 */
871static void rcu_qs(void)
872{
873 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
874 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
875 return;
876 trace_rcu_grace_period(TPS("rcu_sched"),
877 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
878 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
879 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
880 return;
881 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
882 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
883}
884
885/*
886 * Register an urgently needed quiescent state. If there is an
887 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
888 * dyntick-idle quiescent state visible to other CPUs, which will in
889 * some cases serve for expedited as well as normal grace periods.
890 * Either way, register a lightweight quiescent state.
891 */
892void rcu_all_qs(void)
893{
894 unsigned long flags;
895
896 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
897 return;
898 preempt_disable();
899 /* Load rcu_urgent_qs before other flags. */
900 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
901 preempt_enable();
902 return;
903 }
904 this_cpu_write(rcu_data.rcu_urgent_qs, false);
905 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
906 local_irq_save(flags);
907 rcu_momentary_dyntick_idle();
908 local_irq_restore(flags);
909 }
910 rcu_qs();
911 preempt_enable();
912}
913EXPORT_SYMBOL_GPL(rcu_all_qs);
914
915/*
916 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
917 */
918void rcu_note_context_switch(bool preempt)
919{
920 trace_rcu_utilization(TPS("Start context switch"));
921 rcu_qs();
922 /* Load rcu_urgent_qs before other flags. */
923 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
924 goto out;
925 this_cpu_write(rcu_data.rcu_urgent_qs, false);
926 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
927 rcu_momentary_dyntick_idle();
928 rcu_tasks_qs(current, preempt);
929out:
930 trace_rcu_utilization(TPS("End context switch"));
931}
932EXPORT_SYMBOL_GPL(rcu_note_context_switch);
933
934/*
935 * Because preemptible RCU does not exist, there are never any preempted
936 * RCU readers.
937 */
938static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
939{
940 return 0;
941}
942
943/*
944 * Because there is no preemptible RCU, there can be no readers blocked.
945 */
946static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
947{
948 return false;
949}
950
951/*
952 * Because there is no preemptible RCU, there can be no deferred quiescent
953 * states.
954 */
955static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
956{
957 return false;
958}
959static void rcu_preempt_deferred_qs(struct task_struct *t) { }
960
961/*
962 * Because there is no preemptible RCU, there can be no readers blocked,
963 * so there is no need to check for blocked tasks. So check only for
964 * bogus qsmask values.
965 */
966static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
967{
968 WARN_ON_ONCE(rnp->qsmask);
969}
970
971/*
972 * Check to see if this CPU is in a non-context-switch quiescent state,
973 * namely user mode and idle loop.
974 */
975static void rcu_flavor_sched_clock_irq(int user)
976{
977 if (user || rcu_is_cpu_rrupt_from_idle()) {
978
979 /*
980 * Get here if this CPU took its interrupt from user
981 * mode or from the idle loop, and if this is not a
982 * nested interrupt. In this case, the CPU is in
983 * a quiescent state, so note it.
984 *
985 * No memory barrier is required here because rcu_qs()
986 * references only CPU-local variables that other CPUs
987 * neither access nor modify, at least not while the
988 * corresponding CPU is online.
989 */
990
991 rcu_qs();
992 }
993}
994
995/*
996 * Because preemptible RCU does not exist, tasks cannot possibly exit
997 * while in preemptible RCU read-side critical sections.
998 */
999void exit_rcu(void)
1000{
1001}
1002
1003/*
1004 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
1005 */
1006static void
1007dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
1008{
1009 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
1010}
1011
1012#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
1013
1014/*
1015 * If boosting, set rcuc kthreads to realtime priority.
1016 */
1017static void rcu_cpu_kthread_setup(unsigned int cpu)
1018{
1019#ifdef CONFIG_RCU_BOOST
1020 struct sched_param sp;
1021
1022 sp.sched_priority = kthread_prio;
1023 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1024#endif /* #ifdef CONFIG_RCU_BOOST */
1025}
1026
1027#ifdef CONFIG_RCU_BOOST
1028
1029/*
1030 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1031 * or ->boost_tasks, advancing the pointer to the next task in the
1032 * ->blkd_tasks list.
1033 *
1034 * Note that irqs must be enabled: boosting the task can block.
1035 * Returns 1 if there are more tasks needing to be boosted.
1036 */
1037static int rcu_boost(struct rcu_node *rnp)
1038{
1039 unsigned long flags;
1040 struct task_struct *t;
1041 struct list_head *tb;
1042
1043 if (READ_ONCE(rnp->exp_tasks) == NULL &&
1044 READ_ONCE(rnp->boost_tasks) == NULL)
1045 return 0; /* Nothing left to boost. */
1046
1047 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1048
1049 /*
1050 * Recheck under the lock: all tasks in need of boosting
1051 * might exit their RCU read-side critical sections on their own.
1052 */
1053 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1054 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1055 return 0;
1056 }
1057
1058 /*
1059 * Preferentially boost tasks blocking expedited grace periods.
1060 * This cannot starve the normal grace periods because a second
1061 * expedited grace period must boost all blocked tasks, including
1062 * those blocking the pre-existing normal grace period.
1063 */
1064 if (rnp->exp_tasks != NULL)
1065 tb = rnp->exp_tasks;
1066 else
1067 tb = rnp->boost_tasks;
1068
1069 /*
1070 * We boost task t by manufacturing an rt_mutex that appears to
1071 * be held by task t. We leave a pointer to that rt_mutex where
1072 * task t can find it, and task t will release the mutex when it
1073 * exits its outermost RCU read-side critical section. Then
1074 * simply acquiring this artificial rt_mutex will boost task
1075 * t's priority. (Thanks to tglx for suggesting this approach!)
1076 *
1077 * Note that task t must acquire rnp->lock to remove itself from
1078 * the ->blkd_tasks list, which it will do from exit() if from
1079 * nowhere else. We therefore are guaranteed that task t will
1080 * stay around at least until we drop rnp->lock. Note that
1081 * rnp->lock also resolves races between our priority boosting
1082 * and task t's exiting its outermost RCU read-side critical
1083 * section.
1084 */
1085 t = container_of(tb, struct task_struct, rcu_node_entry);
1086 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1087 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1088 /* Lock only for side effect: boosts task t's priority. */
1089 rt_mutex_lock(&rnp->boost_mtx);
1090 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1091 rnp->n_boosts++;
1092
1093 return READ_ONCE(rnp->exp_tasks) != NULL ||
1094 READ_ONCE(rnp->boost_tasks) != NULL;
1095}
1096
1097/*
1098 * Priority-boosting kthread, one per leaf rcu_node.
1099 */
1100static int rcu_boost_kthread(void *arg)
1101{
1102 struct rcu_node *rnp = (struct rcu_node *)arg;
1103 int spincnt = 0;
1104 int more2boost;
1105
1106 trace_rcu_utilization(TPS("Start boost kthread@init"));
1107 for (;;) {
1108 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1109 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1110 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1111 READ_ONCE(rnp->exp_tasks));
1112 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1113 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1114 more2boost = rcu_boost(rnp);
1115 if (more2boost)
1116 spincnt++;
1117 else
1118 spincnt = 0;
1119 if (spincnt > 10) {
1120 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1121 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1122 schedule_timeout_idle(2);
1123 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1124 spincnt = 0;
1125 }
1126 }
1127 /* NOTREACHED */
1128 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1129 return 0;
1130}
1131
1132/*
1133 * Check to see if it is time to start boosting RCU readers that are
1134 * blocking the current grace period, and, if so, tell the per-rcu_node
1135 * kthread to start boosting them. If there is an expedited grace
1136 * period in progress, it is always time to boost.
1137 *
1138 * The caller must hold rnp->lock, which this function releases.
1139 * The ->boost_kthread_task is immortal, so we don't need to worry
1140 * about it going away.
1141 */
1142static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1143 __releases(rnp->lock)
1144{
1145 raw_lockdep_assert_held_rcu_node(rnp);
1146 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1147 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1148 return;
1149 }
1150 if (rnp->exp_tasks != NULL ||
1151 (rnp->gp_tasks != NULL &&
1152 rnp->boost_tasks == NULL &&
1153 rnp->qsmask == 0 &&
1154 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld))) {
1155 if (rnp->exp_tasks == NULL)
1156 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1157 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1158 rcu_wake_cond(rnp->boost_kthread_task,
1159 READ_ONCE(rnp->boost_kthread_status));
1160 } else {
1161 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1162 }
1163}
1164
1165/*
1166 * Is the current CPU running the RCU-callbacks kthread?
1167 * Caller must have preemption disabled.
1168 */
1169static bool rcu_is_callbacks_kthread(void)
1170{
1171 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1172}
1173
1174#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1175
1176/*
1177 * Do priority-boost accounting for the start of a new grace period.
1178 */
1179static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1180{
1181 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1182}
1183
1184/*
1185 * Create an RCU-boost kthread for the specified node if one does not
1186 * already exist. We only create this kthread for preemptible RCU.
1187 * Returns zero if all is well, a negated errno otherwise.
1188 */
1189static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1190{
1191 unsigned long flags;
1192 int rnp_index = rnp - rcu_get_root();
1193 struct sched_param sp;
1194 struct task_struct *t;
1195
1196 if (rnp->boost_kthread_task || !rcu_scheduler_fully_active)
1197 return;
1198
1199 rcu_state.boost = 1;
1200
1201 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1202 "rcub/%d", rnp_index);
1203 if (WARN_ON_ONCE(IS_ERR(t)))
1204 return;
1205
1206 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1207 rnp->boost_kthread_task = t;
1208 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1209 sp.sched_priority = kthread_prio;
1210 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1211 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1212}
1213
1214/*
1215 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1216 * served by the rcu_node in question. The CPU hotplug lock is still
1217 * held, so the value of rnp->qsmaskinit will be stable.
1218 *
1219 * We don't include outgoingcpu in the affinity set, use -1 if there is
1220 * no outgoing CPU. If there are no CPUs left in the affinity set,
1221 * this function allows the kthread to execute on any CPU.
1222 */
1223static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1224{
1225 struct task_struct *t = rnp->boost_kthread_task;
1226 unsigned long mask = rcu_rnp_online_cpus(rnp);
1227 cpumask_var_t cm;
1228 int cpu;
1229
1230 if (!t)
1231 return;
1232 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1233 return;
1234 for_each_leaf_node_possible_cpu(rnp, cpu)
1235 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1236 cpu != outgoingcpu)
1237 cpumask_set_cpu(cpu, cm);
1238 if (cpumask_weight(cm) == 0)
1239 cpumask_setall(cm);
1240 set_cpus_allowed_ptr(t, cm);
1241 free_cpumask_var(cm);
1242}
1243
1244/*
1245 * Spawn boost kthreads -- called as soon as the scheduler is running.
1246 */
1247static void __init rcu_spawn_boost_kthreads(void)
1248{
1249 struct rcu_node *rnp;
1250
1251 rcu_for_each_leaf_node(rnp)
1252 if (rcu_rnp_online_cpus(rnp))
1253 rcu_spawn_one_boost_kthread(rnp);
1254}
1255
1256#else /* #ifdef CONFIG_RCU_BOOST */
1257
1258static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1259 __releases(rnp->lock)
1260{
1261 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1262}
1263
1264static bool rcu_is_callbacks_kthread(void)
1265{
1266 return false;
1267}
1268
1269static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1270{
1271}
1272
1273static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1274{
1275}
1276
1277static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1278{
1279}
1280
1281static void __init rcu_spawn_boost_kthreads(void)
1282{
1283}
1284
1285#endif /* #else #ifdef CONFIG_RCU_BOOST */
1286
1287#if !defined(CONFIG_RCU_FAST_NO_HZ)
1288
1289/*
1290 * Check to see if any future non-offloaded RCU-related work will need
1291 * to be done by the current CPU, even if none need be done immediately,
1292 * returning 1 if so. This function is part of the RCU implementation;
1293 * it is -not- an exported member of the RCU API.
1294 *
1295 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1296 * CPU has RCU callbacks queued.
1297 */
1298int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1299{
1300 *nextevt = KTIME_MAX;
1301 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1302 !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
1303}
1304
1305/*
1306 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1307 * after it.
1308 */
1309static void rcu_cleanup_after_idle(void)
1310{
1311}
1312
1313/*
1314 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1315 * is nothing.
1316 */
1317static void rcu_prepare_for_idle(void)
1318{
1319}
1320
1321#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1322
1323/*
1324 * This code is invoked when a CPU goes idle, at which point we want
1325 * to have the CPU do everything required for RCU so that it can enter
1326 * the energy-efficient dyntick-idle mode.
1327 *
1328 * The following preprocessor symbol controls this:
1329 *
1330 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1331 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1332 * is sized to be roughly one RCU grace period. Those energy-efficiency
1333 * benchmarkers who might otherwise be tempted to set this to a large
1334 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1335 * system. And if you are -that- concerned about energy efficiency,
1336 * just power the system down and be done with it!
1337 *
1338 * The value below works well in practice. If future workloads require
1339 * adjustment, they can be converted into kernel config parameters, though
1340 * making the state machine smarter might be a better option.
1341 */
1342#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1343
1344static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1345module_param(rcu_idle_gp_delay, int, 0644);
1346
1347/*
1348 * Try to advance callbacks on the current CPU, but only if it has been
1349 * awhile since the last time we did so. Afterwards, if there are any
1350 * callbacks ready for immediate invocation, return true.
1351 */
1352static bool __maybe_unused rcu_try_advance_all_cbs(void)
1353{
1354 bool cbs_ready = false;
1355 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1356 struct rcu_node *rnp;
1357
1358 /* Exit early if we advanced recently. */
1359 if (jiffies == rdp->last_advance_all)
1360 return false;
1361 rdp->last_advance_all = jiffies;
1362
1363 rnp = rdp->mynode;
1364
1365 /*
1366 * Don't bother checking unless a grace period has
1367 * completed since we last checked and there are
1368 * callbacks not yet ready to invoke.
1369 */
1370 if ((rcu_seq_completed_gp(rdp->gp_seq,
1371 rcu_seq_current(&rnp->gp_seq)) ||
1372 unlikely(READ_ONCE(rdp->gpwrap))) &&
1373 rcu_segcblist_pend_cbs(&rdp->cblist))
1374 note_gp_changes(rdp);
1375
1376 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1377 cbs_ready = true;
1378 return cbs_ready;
1379}
1380
1381/*
1382 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1383 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1384 * caller about what to set the timeout.
1385 *
1386 * The caller must have disabled interrupts.
1387 */
1388int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1389{
1390 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1391 unsigned long dj;
1392
1393 lockdep_assert_irqs_disabled();
1394
1395 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1396 if (rcu_segcblist_empty(&rdp->cblist) ||
1397 rcu_rdp_is_offloaded(rdp)) {
1398 *nextevt = KTIME_MAX;
1399 return 0;
1400 }
1401
1402 /* Attempt to advance callbacks. */
1403 if (rcu_try_advance_all_cbs()) {
1404 /* Some ready to invoke, so initiate later invocation. */
1405 invoke_rcu_core();
1406 return 1;
1407 }
1408 rdp->last_accelerate = jiffies;
1409
1410 /* Request timer and round. */
1411 dj = round_up(rcu_idle_gp_delay + jiffies, rcu_idle_gp_delay) - jiffies;
1412
1413 *nextevt = basemono + dj * TICK_NSEC;
1414 return 0;
1415}
1416
1417/*
1418 * Prepare a CPU for idle from an RCU perspective. The first major task is to
1419 * sense whether nohz mode has been enabled or disabled via sysfs. The second
1420 * major task is to accelerate (that is, assign grace-period numbers to) any
1421 * recently arrived callbacks.
1422 *
1423 * The caller must have disabled interrupts.
1424 */
1425static void rcu_prepare_for_idle(void)
1426{
1427 bool needwake;
1428 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1429 struct rcu_node *rnp;
1430 int tne;
1431
1432 lockdep_assert_irqs_disabled();
1433 if (rcu_rdp_is_offloaded(rdp))
1434 return;
1435
1436 /* Handle nohz enablement switches conservatively. */
1437 tne = READ_ONCE(tick_nohz_active);
1438 if (tne != rdp->tick_nohz_enabled_snap) {
1439 if (!rcu_segcblist_empty(&rdp->cblist))
1440 invoke_rcu_core(); /* force nohz to see update. */
1441 rdp->tick_nohz_enabled_snap = tne;
1442 return;
1443 }
1444 if (!tne)
1445 return;
1446
1447 /*
1448 * If we have not yet accelerated this jiffy, accelerate all
1449 * callbacks on this CPU.
1450 */
1451 if (rdp->last_accelerate == jiffies)
1452 return;
1453 rdp->last_accelerate = jiffies;
1454 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1455 rnp = rdp->mynode;
1456 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1457 needwake = rcu_accelerate_cbs(rnp, rdp);
1458 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1459 if (needwake)
1460 rcu_gp_kthread_wake();
1461 }
1462}
1463
1464/*
1465 * Clean up for exit from idle. Attempt to advance callbacks based on
1466 * any grace periods that elapsed while the CPU was idle, and if any
1467 * callbacks are now ready to invoke, initiate invocation.
1468 */
1469static void rcu_cleanup_after_idle(void)
1470{
1471 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1472
1473 lockdep_assert_irqs_disabled();
1474 if (rcu_rdp_is_offloaded(rdp))
1475 return;
1476 if (rcu_try_advance_all_cbs())
1477 invoke_rcu_core();
1478}
1479
1480#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1481
1482#ifdef CONFIG_RCU_NOCB_CPU
1483
1484/*
1485 * Offload callback processing from the boot-time-specified set of CPUs
1486 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1487 * created that pull the callbacks from the corresponding CPU, wait for
1488 * a grace period to elapse, and invoke the callbacks. These kthreads
1489 * are organized into GP kthreads, which manage incoming callbacks, wait for
1490 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1491 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1492 * do a wake_up() on their GP kthread when they insert a callback into any
1493 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1494 * in which case each kthread actively polls its CPU. (Which isn't so great
1495 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1496 *
1497 * This is intended to be used in conjunction with Frederic Weisbecker's
1498 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1499 * running CPU-bound user-mode computations.
1500 *
1501 * Offloading of callbacks can also be used as an energy-efficiency
1502 * measure because CPUs with no RCU callbacks queued are more aggressive
1503 * about entering dyntick-idle mode.
1504 */
1505
1506
1507/*
1508 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1509 * If the list is invalid, a warning is emitted and all CPUs are offloaded.
1510 */
1511static int __init rcu_nocb_setup(char *str)
1512{
1513 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1514 if (cpulist_parse(str, rcu_nocb_mask)) {
1515 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1516 cpumask_setall(rcu_nocb_mask);
1517 }
1518 return 1;
1519}
1520__setup("rcu_nocbs=", rcu_nocb_setup);
1521
1522static int __init parse_rcu_nocb_poll(char *arg)
1523{
1524 rcu_nocb_poll = true;
1525 return 0;
1526}
1527early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1528
1529/*
1530 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1531 * After all, the main point of bypassing is to avoid lock contention
1532 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1533 */
1534static int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1535module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1536
1537/*
1538 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1539 * lock isn't immediately available, increment ->nocb_lock_contended to
1540 * flag the contention.
1541 */
1542static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1543 __acquires(&rdp->nocb_bypass_lock)
1544{
1545 lockdep_assert_irqs_disabled();
1546 if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1547 return;
1548 atomic_inc(&rdp->nocb_lock_contended);
1549 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1550 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1551 raw_spin_lock(&rdp->nocb_bypass_lock);
1552 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1553 atomic_dec(&rdp->nocb_lock_contended);
1554}
1555
1556/*
1557 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1558 * not contended. Please note that this is extremely special-purpose,
1559 * relying on the fact that at most two kthreads and one CPU contend for
1560 * this lock, and also that the two kthreads are guaranteed to have frequent
1561 * grace-period-duration time intervals between successive acquisitions
1562 * of the lock. This allows us to use an extremely simple throttling
1563 * mechanism, and further to apply it only to the CPU doing floods of
1564 * call_rcu() invocations. Don't try this at home!
1565 */
1566static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1567{
1568 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1569 while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1570 cpu_relax();
1571}
1572
1573/*
1574 * Conditionally acquire the specified rcu_data structure's
1575 * ->nocb_bypass_lock.
1576 */
1577static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1578{
1579 lockdep_assert_irqs_disabled();
1580 return raw_spin_trylock(&rdp->nocb_bypass_lock);
1581}
1582
1583/*
1584 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1585 */
1586static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1587 __releases(&rdp->nocb_bypass_lock)
1588{
1589 lockdep_assert_irqs_disabled();
1590 raw_spin_unlock(&rdp->nocb_bypass_lock);
1591}
1592
1593/*
1594 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1595 * if it corresponds to a no-CBs CPU.
1596 */
1597static void rcu_nocb_lock(struct rcu_data *rdp)
1598{
1599 lockdep_assert_irqs_disabled();
1600 if (!rcu_rdp_is_offloaded(rdp))
1601 return;
1602 raw_spin_lock(&rdp->nocb_lock);
1603}
1604
1605/*
1606 * Release the specified rcu_data structure's ->nocb_lock, but only
1607 * if it corresponds to a no-CBs CPU.
1608 */
1609static void rcu_nocb_unlock(struct rcu_data *rdp)
1610{
1611 if (rcu_rdp_is_offloaded(rdp)) {
1612 lockdep_assert_irqs_disabled();
1613 raw_spin_unlock(&rdp->nocb_lock);
1614 }
1615}
1616
1617/*
1618 * Release the specified rcu_data structure's ->nocb_lock and restore
1619 * interrupts, but only if it corresponds to a no-CBs CPU.
1620 */
1621static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1622 unsigned long flags)
1623{
1624 if (rcu_rdp_is_offloaded(rdp)) {
1625 lockdep_assert_irqs_disabled();
1626 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1627 } else {
1628 local_irq_restore(flags);
1629 }
1630}
1631
1632/* Lockdep check that ->cblist may be safely accessed. */
1633static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1634{
1635 lockdep_assert_irqs_disabled();
1636 if (rcu_rdp_is_offloaded(rdp))
1637 lockdep_assert_held(&rdp->nocb_lock);
1638}
1639
1640/*
1641 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1642 * grace period.
1643 */
1644static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1645{
1646 swake_up_all(sq);
1647}
1648
1649static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1650{
1651 return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1652}
1653
1654static void rcu_init_one_nocb(struct rcu_node *rnp)
1655{
1656 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1657 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1658}
1659
1660/* Is the specified CPU a no-CBs CPU? */
1661bool rcu_is_nocb_cpu(int cpu)
1662{
1663 if (cpumask_available(rcu_nocb_mask))
1664 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1665 return false;
1666}
1667
1668static bool __wake_nocb_gp(struct rcu_data *rdp_gp,
1669 struct rcu_data *rdp,
1670 bool force, unsigned long flags)
1671 __releases(rdp_gp->nocb_gp_lock)
1672{
1673 bool needwake = false;
1674
1675 if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1676 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1677 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1678 TPS("AlreadyAwake"));
1679 return false;
1680 }
1681
1682 if (rdp_gp->nocb_defer_wakeup > RCU_NOCB_WAKE_NOT) {
1683 WRITE_ONCE(rdp_gp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
1684 del_timer(&rdp_gp->nocb_timer);
1685 }
1686
1687 if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1688 WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1689 needwake = true;
1690 }
1691 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1692 if (needwake) {
1693 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1694 wake_up_process(rdp_gp->nocb_gp_kthread);
1695 }
1696
1697 return needwake;
1698}
1699
1700/*
1701 * Kick the GP kthread for this NOCB group.
1702 */
1703static bool wake_nocb_gp(struct rcu_data *rdp, bool force)
1704{
1705 unsigned long flags;
1706 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1707
1708 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1709 return __wake_nocb_gp(rdp_gp, rdp, force, flags);
1710}
1711
1712/*
1713 * Arrange to wake the GP kthread for this NOCB group at some future
1714 * time when it is safe to do so.
1715 */
1716static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1717 const char *reason)
1718{
1719 unsigned long flags;
1720 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1721
1722 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1723
1724 /*
1725 * Bypass wakeup overrides previous deferments. In case
1726 * of callback storm, no need to wake up too early.
1727 */
1728 if (waketype == RCU_NOCB_WAKE_BYPASS) {
1729 mod_timer(&rdp_gp->nocb_timer, jiffies + 2);
1730 WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
1731 } else {
1732 if (rdp_gp->nocb_defer_wakeup < RCU_NOCB_WAKE)
1733 mod_timer(&rdp_gp->nocb_timer, jiffies + 1);
1734 if (rdp_gp->nocb_defer_wakeup < waketype)
1735 WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
1736 }
1737
1738 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1739
1740 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1741}
1742
1743/*
1744 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1745 * However, if there is a callback to be enqueued and if ->nocb_bypass
1746 * proves to be initially empty, just return false because the no-CB GP
1747 * kthread may need to be awakened in this case.
1748 *
1749 * Note that this function always returns true if rhp is NULL.
1750 */
1751static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1752 unsigned long j)
1753{
1754 struct rcu_cblist rcl;
1755
1756 WARN_ON_ONCE(!rcu_rdp_is_offloaded(rdp));
1757 rcu_lockdep_assert_cblist_protected(rdp);
1758 lockdep_assert_held(&rdp->nocb_bypass_lock);
1759 if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1760 raw_spin_unlock(&rdp->nocb_bypass_lock);
1761 return false;
1762 }
1763 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1764 if (rhp)
1765 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1766 rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1767 rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1768 WRITE_ONCE(rdp->nocb_bypass_first, j);
1769 rcu_nocb_bypass_unlock(rdp);
1770 return true;
1771}
1772
1773/*
1774 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1775 * However, if there is a callback to be enqueued and if ->nocb_bypass
1776 * proves to be initially empty, just return false because the no-CB GP
1777 * kthread may need to be awakened in this case.
1778 *
1779 * Note that this function always returns true if rhp is NULL.
1780 */
1781static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1782 unsigned long j)
1783{
1784 if (!rcu_rdp_is_offloaded(rdp))
1785 return true;
1786 rcu_lockdep_assert_cblist_protected(rdp);
1787 rcu_nocb_bypass_lock(rdp);
1788 return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1789}
1790
1791/*
1792 * If the ->nocb_bypass_lock is immediately available, flush the
1793 * ->nocb_bypass queue into ->cblist.
1794 */
1795static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1796{
1797 rcu_lockdep_assert_cblist_protected(rdp);
1798 if (!rcu_rdp_is_offloaded(rdp) ||
1799 !rcu_nocb_bypass_trylock(rdp))
1800 return;
1801 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1802}
1803
1804/*
1805 * See whether it is appropriate to use the ->nocb_bypass list in order
1806 * to control contention on ->nocb_lock. A limited number of direct
1807 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1808 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1809 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1810 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1811 * used if ->cblist is empty, because otherwise callbacks can be stranded
1812 * on ->nocb_bypass because we cannot count on the current CPU ever again
1813 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1814 * non-empty, the corresponding no-CBs grace-period kthread must not be
1815 * in an indefinite sleep state.
1816 *
1817 * Finally, it is not permitted to use the bypass during early boot,
1818 * as doing so would confuse the auto-initialization code. Besides
1819 * which, there is no point in worrying about lock contention while
1820 * there is only one CPU in operation.
1821 */
1822static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1823 bool *was_alldone, unsigned long flags)
1824{
1825 unsigned long c;
1826 unsigned long cur_gp_seq;
1827 unsigned long j = jiffies;
1828 long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1829
1830 lockdep_assert_irqs_disabled();
1831
1832 // Pure softirq/rcuc based processing: no bypassing, no
1833 // locking.
1834 if (!rcu_rdp_is_offloaded(rdp)) {
1835 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1836 return false;
1837 }
1838
1839 // In the process of (de-)offloading: no bypassing, but
1840 // locking.
1841 if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
1842 rcu_nocb_lock(rdp);
1843 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1844 return false; /* Not offloaded, no bypassing. */
1845 }
1846
1847 // Don't use ->nocb_bypass during early boot.
1848 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1849 rcu_nocb_lock(rdp);
1850 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1851 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1852 return false;
1853 }
1854
1855 // If we have advanced to a new jiffy, reset counts to allow
1856 // moving back from ->nocb_bypass to ->cblist.
1857 if (j == rdp->nocb_nobypass_last) {
1858 c = rdp->nocb_nobypass_count + 1;
1859 } else {
1860 WRITE_ONCE(rdp->nocb_nobypass_last, j);
1861 c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1862 if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1863 nocb_nobypass_lim_per_jiffy))
1864 c = 0;
1865 else if (c > nocb_nobypass_lim_per_jiffy)
1866 c = nocb_nobypass_lim_per_jiffy;
1867 }
1868 WRITE_ONCE(rdp->nocb_nobypass_count, c);
1869
1870 // If there hasn't yet been all that many ->cblist enqueues
1871 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1872 // ->nocb_bypass first.
1873 if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1874 rcu_nocb_lock(rdp);
1875 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1876 if (*was_alldone)
1877 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1878 TPS("FirstQ"));
1879 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1880 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1881 return false; // Caller must enqueue the callback.
1882 }
1883
1884 // If ->nocb_bypass has been used too long or is too full,
1885 // flush ->nocb_bypass to ->cblist.
1886 if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1887 ncbs >= qhimark) {
1888 rcu_nocb_lock(rdp);
1889 if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1890 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1891 if (*was_alldone)
1892 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1893 TPS("FirstQ"));
1894 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1895 return false; // Caller must enqueue the callback.
1896 }
1897 if (j != rdp->nocb_gp_adv_time &&
1898 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1899 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1900 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1901 rdp->nocb_gp_adv_time = j;
1902 }
1903 rcu_nocb_unlock_irqrestore(rdp, flags);
1904 return true; // Callback already enqueued.
1905 }
1906
1907 // We need to use the bypass.
1908 rcu_nocb_wait_contended(rdp);
1909 rcu_nocb_bypass_lock(rdp);
1910 ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1911 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1912 rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1913 if (!ncbs) {
1914 WRITE_ONCE(rdp->nocb_bypass_first, j);
1915 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1916 }
1917 rcu_nocb_bypass_unlock(rdp);
1918 smp_mb(); /* Order enqueue before wake. */
1919 if (ncbs) {
1920 local_irq_restore(flags);
1921 } else {
1922 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1923 rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1924 if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1925 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1926 TPS("FirstBQwake"));
1927 __call_rcu_nocb_wake(rdp, true, flags);
1928 } else {
1929 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1930 TPS("FirstBQnoWake"));
1931 rcu_nocb_unlock_irqrestore(rdp, flags);
1932 }
1933 }
1934 return true; // Callback already enqueued.
1935}
1936
1937/*
1938 * Awaken the no-CBs grace-period kthread if needed, either due to it
1939 * legitimately being asleep or due to overload conditions.
1940 *
1941 * If warranted, also wake up the kthread servicing this CPUs queues.
1942 */
1943static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1944 unsigned long flags)
1945 __releases(rdp->nocb_lock)
1946{
1947 unsigned long cur_gp_seq;
1948 unsigned long j;
1949 long len;
1950 struct task_struct *t;
1951
1952 // If we are being polled or there is no kthread, just leave.
1953 t = READ_ONCE(rdp->nocb_gp_kthread);
1954 if (rcu_nocb_poll || !t) {
1955 rcu_nocb_unlock_irqrestore(rdp, flags);
1956 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1957 TPS("WakeNotPoll"));
1958 return;
1959 }
1960 // Need to actually to a wakeup.
1961 len = rcu_segcblist_n_cbs(&rdp->cblist);
1962 if (was_alldone) {
1963 rdp->qlen_last_fqs_check = len;
1964 if (!irqs_disabled_flags(flags)) {
1965 /* ... if queue was empty ... */
1966 rcu_nocb_unlock_irqrestore(rdp, flags);
1967 wake_nocb_gp(rdp, false);
1968 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1969 TPS("WakeEmpty"));
1970 } else {
1971 rcu_nocb_unlock_irqrestore(rdp, flags);
1972 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1973 TPS("WakeEmptyIsDeferred"));
1974 }
1975 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1976 /* ... or if many callbacks queued. */
1977 rdp->qlen_last_fqs_check = len;
1978 j = jiffies;
1979 if (j != rdp->nocb_gp_adv_time &&
1980 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1981 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1982 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1983 rdp->nocb_gp_adv_time = j;
1984 }
1985 smp_mb(); /* Enqueue before timer_pending(). */
1986 if ((rdp->nocb_cb_sleep ||
1987 !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1988 !timer_pending(&rdp->nocb_timer)) {
1989 rcu_nocb_unlock_irqrestore(rdp, flags);
1990 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1991 TPS("WakeOvfIsDeferred"));
1992 } else {
1993 rcu_nocb_unlock_irqrestore(rdp, flags);
1994 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1995 }
1996 } else {
1997 rcu_nocb_unlock_irqrestore(rdp, flags);
1998 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1999 }
2000 return;
2001}
2002
2003/*
2004 * Check if we ignore this rdp.
2005 *
2006 * We check that without holding the nocb lock but
2007 * we make sure not to miss a freshly offloaded rdp
2008 * with the current ordering:
2009 *
2010 * rdp_offload_toggle() nocb_gp_enabled_cb()
2011 * ------------------------- ----------------------------
2012 * WRITE flags LOCK nocb_gp_lock
2013 * LOCK nocb_gp_lock READ/WRITE nocb_gp_sleep
2014 * READ/WRITE nocb_gp_sleep UNLOCK nocb_gp_lock
2015 * UNLOCK nocb_gp_lock READ flags
2016 */
2017static inline bool nocb_gp_enabled_cb(struct rcu_data *rdp)
2018{
2019 u8 flags = SEGCBLIST_OFFLOADED | SEGCBLIST_KTHREAD_GP;
2020
2021 return rcu_segcblist_test_flags(&rdp->cblist, flags);
2022}
2023
2024static inline bool nocb_gp_update_state_deoffloading(struct rcu_data *rdp,
2025 bool *needwake_state)
2026{
2027 struct rcu_segcblist *cblist = &rdp->cblist;
2028
2029 if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED)) {
2030 if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP)) {
2031 rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_GP);
2032 if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
2033 *needwake_state = true;
2034 }
2035 return false;
2036 }
2037
2038 /*
2039 * De-offloading. Clear our flag and notify the de-offload worker.
2040 * We will ignore this rdp until it ever gets re-offloaded.
2041 */
2042 WARN_ON_ONCE(!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP));
2043 rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_GP);
2044 if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
2045 *needwake_state = true;
2046 return true;
2047}
2048
2049
2050/*
2051 * No-CBs GP kthreads come here to wait for additional callbacks to show up
2052 * or for grace periods to end.
2053 */
2054static void nocb_gp_wait(struct rcu_data *my_rdp)
2055{
2056 bool bypass = false;
2057 long bypass_ncbs;
2058 int __maybe_unused cpu = my_rdp->cpu;
2059 unsigned long cur_gp_seq;
2060 unsigned long flags;
2061 bool gotcbs = false;
2062 unsigned long j = jiffies;
2063 bool needwait_gp = false; // This prevents actual uninitialized use.
2064 bool needwake;
2065 bool needwake_gp;
2066 struct rcu_data *rdp;
2067 struct rcu_node *rnp;
2068 unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
2069 bool wasempty = false;
2070
2071 /*
2072 * Each pass through the following loop checks for CBs and for the
2073 * nearest grace period (if any) to wait for next. The CB kthreads
2074 * and the global grace-period kthread are awakened if needed.
2075 */
2076 WARN_ON_ONCE(my_rdp->nocb_gp_rdp != my_rdp);
2077 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
2078 bool needwake_state = false;
2079
2080 if (!nocb_gp_enabled_cb(rdp))
2081 continue;
2082 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
2083 rcu_nocb_lock_irqsave(rdp, flags);
2084 if (nocb_gp_update_state_deoffloading(rdp, &needwake_state)) {
2085 rcu_nocb_unlock_irqrestore(rdp, flags);
2086 if (needwake_state)
2087 swake_up_one(&rdp->nocb_state_wq);
2088 continue;
2089 }
2090 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
2091 if (bypass_ncbs &&
2092 (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
2093 bypass_ncbs > 2 * qhimark)) {
2094 // Bypass full or old, so flush it.
2095 (void)rcu_nocb_try_flush_bypass(rdp, j);
2096 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
2097 } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
2098 rcu_nocb_unlock_irqrestore(rdp, flags);
2099 if (needwake_state)
2100 swake_up_one(&rdp->nocb_state_wq);
2101 continue; /* No callbacks here, try next. */
2102 }
2103 if (bypass_ncbs) {
2104 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2105 TPS("Bypass"));
2106 bypass = true;
2107 }
2108 rnp = rdp->mynode;
2109
2110 // Advance callbacks if helpful and low contention.
2111 needwake_gp = false;
2112 if (!rcu_segcblist_restempty(&rdp->cblist,
2113 RCU_NEXT_READY_TAIL) ||
2114 (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2115 rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
2116 raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
2117 needwake_gp = rcu_advance_cbs(rnp, rdp);
2118 wasempty = rcu_segcblist_restempty(&rdp->cblist,
2119 RCU_NEXT_READY_TAIL);
2120 raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
2121 }
2122 // Need to wait on some grace period?
2123 WARN_ON_ONCE(wasempty &&
2124 !rcu_segcblist_restempty(&rdp->cblist,
2125 RCU_NEXT_READY_TAIL));
2126 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
2127 if (!needwait_gp ||
2128 ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
2129 wait_gp_seq = cur_gp_seq;
2130 needwait_gp = true;
2131 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2132 TPS("NeedWaitGP"));
2133 }
2134 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2135 needwake = rdp->nocb_cb_sleep;
2136 WRITE_ONCE(rdp->nocb_cb_sleep, false);
2137 smp_mb(); /* CB invocation -after- GP end. */
2138 } else {
2139 needwake = false;
2140 }
2141 rcu_nocb_unlock_irqrestore(rdp, flags);
2142 if (needwake) {
2143 swake_up_one(&rdp->nocb_cb_wq);
2144 gotcbs = true;
2145 }
2146 if (needwake_gp)
2147 rcu_gp_kthread_wake();
2148 if (needwake_state)
2149 swake_up_one(&rdp->nocb_state_wq);
2150 }
2151
2152 my_rdp->nocb_gp_bypass = bypass;
2153 my_rdp->nocb_gp_gp = needwait_gp;
2154 my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
2155
2156 if (bypass && !rcu_nocb_poll) {
2157 // At least one child with non-empty ->nocb_bypass, so set
2158 // timer in order to avoid stranding its callbacks.
2159 wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_BYPASS,
2160 TPS("WakeBypassIsDeferred"));
2161 }
2162 if (rcu_nocb_poll) {
2163 /* Polling, so trace if first poll in the series. */
2164 if (gotcbs)
2165 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2166 schedule_timeout_idle(1);
2167 } else if (!needwait_gp) {
2168 /* Wait for callbacks to appear. */
2169 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2170 swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2171 !READ_ONCE(my_rdp->nocb_gp_sleep));
2172 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2173 } else {
2174 rnp = my_rdp->mynode;
2175 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2176 swait_event_interruptible_exclusive(
2177 rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2178 rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2179 !READ_ONCE(my_rdp->nocb_gp_sleep));
2180 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2181 }
2182 if (!rcu_nocb_poll) {
2183 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2184 if (my_rdp->nocb_defer_wakeup > RCU_NOCB_WAKE_NOT) {
2185 WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2186 del_timer(&my_rdp->nocb_timer);
2187 }
2188 WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2189 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2190 }
2191 my_rdp->nocb_gp_seq = -1;
2192 WARN_ON(signal_pending(current));
2193}
2194
2195/*
2196 * No-CBs grace-period-wait kthread. There is one of these per group
2197 * of CPUs, but only once at least one CPU in that group has come online
2198 * at least once since boot. This kthread checks for newly posted
2199 * callbacks from any of the CPUs it is responsible for, waits for a
2200 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2201 * that then have callback-invocation work to do.
2202 */
2203static int rcu_nocb_gp_kthread(void *arg)
2204{
2205 struct rcu_data *rdp = arg;
2206
2207 for (;;) {
2208 WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2209 nocb_gp_wait(rdp);
2210 cond_resched_tasks_rcu_qs();
2211 }
2212 return 0;
2213}
2214
2215static inline bool nocb_cb_can_run(struct rcu_data *rdp)
2216{
2217 u8 flags = SEGCBLIST_OFFLOADED | SEGCBLIST_KTHREAD_CB;
2218 return rcu_segcblist_test_flags(&rdp->cblist, flags);
2219}
2220
2221static inline bool nocb_cb_wait_cond(struct rcu_data *rdp)
2222{
2223 return nocb_cb_can_run(rdp) && !READ_ONCE(rdp->nocb_cb_sleep);
2224}
2225
2226/*
2227 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2228 * then, if there are no more, wait for more to appear.
2229 */
2230static void nocb_cb_wait(struct rcu_data *rdp)
2231{
2232 struct rcu_segcblist *cblist = &rdp->cblist;
2233 unsigned long cur_gp_seq;
2234 unsigned long flags;
2235 bool needwake_state = false;
2236 bool needwake_gp = false;
2237 bool can_sleep = true;
2238 struct rcu_node *rnp = rdp->mynode;
2239
2240 local_irq_save(flags);
2241 rcu_momentary_dyntick_idle();
2242 local_irq_restore(flags);
2243 /*
2244 * Disable BH to provide the expected environment. Also, when
2245 * transitioning to/from NOCB mode, a self-requeuing callback might
2246 * be invoked from softirq. A short grace period could cause both
2247 * instances of this callback would execute concurrently.
2248 */
2249 local_bh_disable();
2250 rcu_do_batch(rdp);
2251 local_bh_enable();
2252 lockdep_assert_irqs_enabled();
2253 rcu_nocb_lock_irqsave(rdp, flags);
2254 if (rcu_segcblist_nextgp(cblist, &cur_gp_seq) &&
2255 rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2256 raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2257 needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2258 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2259 }
2260
2261 if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED)) {
2262 if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB)) {
2263 rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_CB);
2264 if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
2265 needwake_state = true;
2266 }
2267 if (rcu_segcblist_ready_cbs(cblist))
2268 can_sleep = false;
2269 } else {
2270 /*
2271 * De-offloading. Clear our flag and notify the de-offload worker.
2272 * We won't touch the callbacks and keep sleeping until we ever
2273 * get re-offloaded.
2274 */
2275 WARN_ON_ONCE(!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB));
2276 rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_CB);
2277 if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
2278 needwake_state = true;
2279 }
2280
2281 WRITE_ONCE(rdp->nocb_cb_sleep, can_sleep);
2282
2283 if (rdp->nocb_cb_sleep)
2284 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2285
2286 rcu_nocb_unlock_irqrestore(rdp, flags);
2287 if (needwake_gp)
2288 rcu_gp_kthread_wake();
2289
2290 if (needwake_state)
2291 swake_up_one(&rdp->nocb_state_wq);
2292
2293 do {
2294 swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2295 nocb_cb_wait_cond(rdp));
2296
2297 // VVV Ensure CB invocation follows _sleep test.
2298 if (smp_load_acquire(&rdp->nocb_cb_sleep)) { // ^^^
2299 WARN_ON(signal_pending(current));
2300 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2301 }
2302 } while (!nocb_cb_can_run(rdp));
2303}
2304
2305/*
2306 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2307 * nocb_cb_wait() to do the dirty work.
2308 */
2309static int rcu_nocb_cb_kthread(void *arg)
2310{
2311 struct rcu_data *rdp = arg;
2312
2313 // Each pass through this loop does one callback batch, and,
2314 // if there are no more ready callbacks, waits for them.
2315 for (;;) {
2316 nocb_cb_wait(rdp);
2317 cond_resched_tasks_rcu_qs();
2318 }
2319 return 0;
2320}
2321
2322/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2323static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
2324{
2325 return READ_ONCE(rdp->nocb_defer_wakeup) >= level;
2326}
2327
2328/* Do a deferred wakeup of rcu_nocb_kthread(). */
2329static bool do_nocb_deferred_wakeup_common(struct rcu_data *rdp_gp,
2330 struct rcu_data *rdp, int level,
2331 unsigned long flags)
2332 __releases(rdp_gp->nocb_gp_lock)
2333{
2334 int ndw;
2335 int ret;
2336
2337 if (!rcu_nocb_need_deferred_wakeup(rdp_gp, level)) {
2338 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
2339 return false;
2340 }
2341
2342 ndw = rdp_gp->nocb_defer_wakeup;
2343 ret = __wake_nocb_gp(rdp_gp, rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2344 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2345
2346 return ret;
2347}
2348
2349/* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2350static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2351{
2352 unsigned long flags;
2353 struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2354
2355 WARN_ON_ONCE(rdp->nocb_gp_rdp != rdp);
2356 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
2357
2358 raw_spin_lock_irqsave(&rdp->nocb_gp_lock, flags);
2359 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
2360 do_nocb_deferred_wakeup_common(rdp, rdp, RCU_NOCB_WAKE_BYPASS, flags);
2361}
2362
2363/*
2364 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2365 * This means we do an inexact common-case check. Note that if
2366 * we miss, ->nocb_timer will eventually clean things up.
2367 */
2368static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
2369{
2370 unsigned long flags;
2371 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
2372
2373 if (!rdp_gp || !rcu_nocb_need_deferred_wakeup(rdp_gp, RCU_NOCB_WAKE))
2374 return false;
2375
2376 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
2377 return do_nocb_deferred_wakeup_common(rdp_gp, rdp, RCU_NOCB_WAKE, flags);
2378}
2379
2380void rcu_nocb_flush_deferred_wakeup(void)
2381{
2382 do_nocb_deferred_wakeup(this_cpu_ptr(&rcu_data));
2383}
2384EXPORT_SYMBOL_GPL(rcu_nocb_flush_deferred_wakeup);
2385
2386static int rdp_offload_toggle(struct rcu_data *rdp,
2387 bool offload, unsigned long flags)
2388 __releases(rdp->nocb_lock)
2389{
2390 struct rcu_segcblist *cblist = &rdp->cblist;
2391 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
2392 bool wake_gp = false;
2393
2394 rcu_segcblist_offload(cblist, offload);
2395
2396 if (rdp->nocb_cb_sleep)
2397 rdp->nocb_cb_sleep = false;
2398 rcu_nocb_unlock_irqrestore(rdp, flags);
2399
2400 /*
2401 * Ignore former value of nocb_cb_sleep and force wake up as it could
2402 * have been spuriously set to false already.
2403 */
2404 swake_up_one(&rdp->nocb_cb_wq);
2405
2406 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
2407 if (rdp_gp->nocb_gp_sleep) {
2408 rdp_gp->nocb_gp_sleep = false;
2409 wake_gp = true;
2410 }
2411 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
2412
2413 if (wake_gp)
2414 wake_up_process(rdp_gp->nocb_gp_kthread);
2415
2416 return 0;
2417}
2418
2419static long rcu_nocb_rdp_deoffload(void *arg)
2420{
2421 struct rcu_data *rdp = arg;
2422 struct rcu_segcblist *cblist = &rdp->cblist;
2423 unsigned long flags;
2424 int ret;
2425
2426 WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
2427
2428 pr_info("De-offloading %d\n", rdp->cpu);
2429
2430 rcu_nocb_lock_irqsave(rdp, flags);
2431 /*
2432 * Flush once and for all now. This suffices because we are
2433 * running on the target CPU holding ->nocb_lock (thus having
2434 * interrupts disabled), and because rdp_offload_toggle()
2435 * invokes rcu_segcblist_offload(), which clears SEGCBLIST_OFFLOADED.
2436 * Thus future calls to rcu_segcblist_completely_offloaded() will
2437 * return false, which means that future calls to rcu_nocb_try_bypass()
2438 * will refuse to put anything into the bypass.
2439 */
2440 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
2441 ret = rdp_offload_toggle(rdp, false, flags);
2442 swait_event_exclusive(rdp->nocb_state_wq,
2443 !rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB |
2444 SEGCBLIST_KTHREAD_GP));
2445 /*
2446 * Lock one last time to acquire latest callback updates from kthreads
2447 * so we can later handle callbacks locally without locking.
2448 */
2449 rcu_nocb_lock_irqsave(rdp, flags);
2450 /*
2451 * Theoretically we could set SEGCBLIST_SOFTIRQ_ONLY after the nocb
2452 * lock is released but how about being paranoid for once?
2453 */
2454 rcu_segcblist_set_flags(cblist, SEGCBLIST_SOFTIRQ_ONLY);
2455 /*
2456 * With SEGCBLIST_SOFTIRQ_ONLY, we can't use
2457 * rcu_nocb_unlock_irqrestore() anymore.
2458 */
2459 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2460
2461 /* Sanity check */
2462 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
2463
2464
2465 return ret;
2466}
2467
2468int rcu_nocb_cpu_deoffload(int cpu)
2469{
2470 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2471 int ret = 0;
2472
2473 mutex_lock(&rcu_state.barrier_mutex);
2474 cpus_read_lock();
2475 if (rcu_rdp_is_offloaded(rdp)) {
2476 if (cpu_online(cpu)) {
2477 ret = work_on_cpu(cpu, rcu_nocb_rdp_deoffload, rdp);
2478 if (!ret)
2479 cpumask_clear_cpu(cpu, rcu_nocb_mask);
2480 } else {
2481 pr_info("NOCB: Can't CB-deoffload an offline CPU\n");
2482 ret = -EINVAL;
2483 }
2484 }
2485 cpus_read_unlock();
2486 mutex_unlock(&rcu_state.barrier_mutex);
2487
2488 return ret;
2489}
2490EXPORT_SYMBOL_GPL(rcu_nocb_cpu_deoffload);
2491
2492static long rcu_nocb_rdp_offload(void *arg)
2493{
2494 struct rcu_data *rdp = arg;
2495 struct rcu_segcblist *cblist = &rdp->cblist;
2496 unsigned long flags;
2497 int ret;
2498
2499 WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
2500 /*
2501 * For now we only support re-offload, ie: the rdp must have been
2502 * offloaded on boot first.
2503 */
2504 if (!rdp->nocb_gp_rdp)
2505 return -EINVAL;
2506
2507 pr_info("Offloading %d\n", rdp->cpu);
2508 /*
2509 * Can't use rcu_nocb_lock_irqsave() while we are in
2510 * SEGCBLIST_SOFTIRQ_ONLY mode.
2511 */
2512 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2513
2514 /*
2515 * We didn't take the nocb lock while working on the
2516 * rdp->cblist in SEGCBLIST_SOFTIRQ_ONLY mode.
2517 * Every modifications that have been done previously on
2518 * rdp->cblist must be visible remotely by the nocb kthreads
2519 * upon wake up after reading the cblist flags.
2520 *
2521 * The layout against nocb_lock enforces that ordering:
2522 *
2523 * __rcu_nocb_rdp_offload() nocb_cb_wait()/nocb_gp_wait()
2524 * ------------------------- ----------------------------
2525 * WRITE callbacks rcu_nocb_lock()
2526 * rcu_nocb_lock() READ flags
2527 * WRITE flags READ callbacks
2528 * rcu_nocb_unlock() rcu_nocb_unlock()
2529 */
2530 ret = rdp_offload_toggle(rdp, true, flags);
2531 swait_event_exclusive(rdp->nocb_state_wq,
2532 rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB) &&
2533 rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP));
2534
2535 return ret;
2536}
2537
2538int rcu_nocb_cpu_offload(int cpu)
2539{
2540 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2541 int ret = 0;
2542
2543 mutex_lock(&rcu_state.barrier_mutex);
2544 cpus_read_lock();
2545 if (!rcu_rdp_is_offloaded(rdp)) {
2546 if (cpu_online(cpu)) {
2547 ret = work_on_cpu(cpu, rcu_nocb_rdp_offload, rdp);
2548 if (!ret)
2549 cpumask_set_cpu(cpu, rcu_nocb_mask);
2550 } else {
2551 pr_info("NOCB: Can't CB-offload an offline CPU\n");
2552 ret = -EINVAL;
2553 }
2554 }
2555 cpus_read_unlock();
2556 mutex_unlock(&rcu_state.barrier_mutex);
2557
2558 return ret;
2559}
2560EXPORT_SYMBOL_GPL(rcu_nocb_cpu_offload);
2561
2562void __init rcu_init_nohz(void)
2563{
2564 int cpu;
2565 bool need_rcu_nocb_mask = false;
2566 struct rcu_data *rdp;
2567
2568#if defined(CONFIG_NO_HZ_FULL)
2569 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2570 need_rcu_nocb_mask = true;
2571#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2572
2573 if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2574 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2575 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2576 return;
2577 }
2578 }
2579 if (!cpumask_available(rcu_nocb_mask))
2580 return;
2581
2582#if defined(CONFIG_NO_HZ_FULL)
2583 if (tick_nohz_full_running)
2584 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2585#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2586
2587 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2588 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2589 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2590 rcu_nocb_mask);
2591 }
2592 if (cpumask_empty(rcu_nocb_mask))
2593 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2594 else
2595 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2596 cpumask_pr_args(rcu_nocb_mask));
2597 if (rcu_nocb_poll)
2598 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2599
2600 for_each_cpu(cpu, rcu_nocb_mask) {
2601 rdp = per_cpu_ptr(&rcu_data, cpu);
2602 if (rcu_segcblist_empty(&rdp->cblist))
2603 rcu_segcblist_init(&rdp->cblist);
2604 rcu_segcblist_offload(&rdp->cblist, true);
2605 rcu_segcblist_set_flags(&rdp->cblist, SEGCBLIST_KTHREAD_CB);
2606 rcu_segcblist_set_flags(&rdp->cblist, SEGCBLIST_KTHREAD_GP);
2607 }
2608 rcu_organize_nocb_kthreads();
2609}
2610
2611/* Initialize per-rcu_data variables for no-CBs CPUs. */
2612static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2613{
2614 init_swait_queue_head(&rdp->nocb_cb_wq);
2615 init_swait_queue_head(&rdp->nocb_gp_wq);
2616 init_swait_queue_head(&rdp->nocb_state_wq);
2617 raw_spin_lock_init(&rdp->nocb_lock);
2618 raw_spin_lock_init(&rdp->nocb_bypass_lock);
2619 raw_spin_lock_init(&rdp->nocb_gp_lock);
2620 timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2621 rcu_cblist_init(&rdp->nocb_bypass);
2622}
2623
2624/*
2625 * If the specified CPU is a no-CBs CPU that does not already have its
2626 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2627 * for this CPU's group has not yet been created, spawn it as well.
2628 */
2629static void rcu_spawn_one_nocb_kthread(int cpu)
2630{
2631 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2632 struct rcu_data *rdp_gp;
2633 struct task_struct *t;
2634
2635 /*
2636 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2637 * then nothing to do.
2638 */
2639 if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2640 return;
2641
2642 /* If we didn't spawn the GP kthread first, reorganize! */
2643 rdp_gp = rdp->nocb_gp_rdp;
2644 if (!rdp_gp->nocb_gp_kthread) {
2645 t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2646 "rcuog/%d", rdp_gp->cpu);
2647 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2648 return;
2649 WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2650 }
2651
2652 /* Spawn the kthread for this CPU. */
2653 t = kthread_run(rcu_nocb_cb_kthread, rdp,
2654 "rcuo%c/%d", rcu_state.abbr, cpu);
2655 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2656 return;
2657 WRITE_ONCE(rdp->nocb_cb_kthread, t);
2658 WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2659}
2660
2661/*
2662 * If the specified CPU is a no-CBs CPU that does not already have its
2663 * rcuo kthread, spawn it.
2664 */
2665static void rcu_spawn_cpu_nocb_kthread(int cpu)
2666{
2667 if (rcu_scheduler_fully_active)
2668 rcu_spawn_one_nocb_kthread(cpu);
2669}
2670
2671/*
2672 * Once the scheduler is running, spawn rcuo kthreads for all online
2673 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2674 * non-boot CPUs come online -- if this changes, we will need to add
2675 * some mutual exclusion.
2676 */
2677static void __init rcu_spawn_nocb_kthreads(void)
2678{
2679 int cpu;
2680
2681 for_each_online_cpu(cpu)
2682 rcu_spawn_cpu_nocb_kthread(cpu);
2683}
2684
2685/* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2686static int rcu_nocb_gp_stride = -1;
2687module_param(rcu_nocb_gp_stride, int, 0444);
2688
2689/*
2690 * Initialize GP-CB relationships for all no-CBs CPU.
2691 */
2692static void __init rcu_organize_nocb_kthreads(void)
2693{
2694 int cpu;
2695 bool firsttime = true;
2696 bool gotnocbs = false;
2697 bool gotnocbscbs = true;
2698 int ls = rcu_nocb_gp_stride;
2699 int nl = 0; /* Next GP kthread. */
2700 struct rcu_data *rdp;
2701 struct rcu_data *rdp_gp = NULL; /* Suppress misguided gcc warn. */
2702 struct rcu_data *rdp_prev = NULL;
2703
2704 if (!cpumask_available(rcu_nocb_mask))
2705 return;
2706 if (ls == -1) {
2707 ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2708 rcu_nocb_gp_stride = ls;
2709 }
2710
2711 /*
2712 * Each pass through this loop sets up one rcu_data structure.
2713 * Should the corresponding CPU come online in the future, then
2714 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2715 */
2716 for_each_cpu(cpu, rcu_nocb_mask) {
2717 rdp = per_cpu_ptr(&rcu_data, cpu);
2718 if (rdp->cpu >= nl) {
2719 /* New GP kthread, set up for CBs & next GP. */
2720 gotnocbs = true;
2721 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2722 rdp->nocb_gp_rdp = rdp;
2723 rdp_gp = rdp;
2724 if (dump_tree) {
2725 if (!firsttime)
2726 pr_cont("%s\n", gotnocbscbs
2727 ? "" : " (self only)");
2728 gotnocbscbs = false;
2729 firsttime = false;
2730 pr_alert("%s: No-CB GP kthread CPU %d:",
2731 __func__, cpu);
2732 }
2733 } else {
2734 /* Another CB kthread, link to previous GP kthread. */
2735 gotnocbscbs = true;
2736 rdp->nocb_gp_rdp = rdp_gp;
2737 rdp_prev->nocb_next_cb_rdp = rdp;
2738 if (dump_tree)
2739 pr_cont(" %d", cpu);
2740 }
2741 rdp_prev = rdp;
2742 }
2743 if (gotnocbs && dump_tree)
2744 pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
2745}
2746
2747/*
2748 * Bind the current task to the offloaded CPUs. If there are no offloaded
2749 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2750 */
2751void rcu_bind_current_to_nocb(void)
2752{
2753 if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2754 WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2755}
2756EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2757
2758// The ->on_cpu field is available only in CONFIG_SMP=y, so...
2759#ifdef CONFIG_SMP
2760static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
2761{
2762 return tsp && task_is_running(tsp) && !tsp->on_cpu ? "!" : "";
2763}
2764#else // #ifdef CONFIG_SMP
2765static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
2766{
2767 return "";
2768}
2769#endif // #else #ifdef CONFIG_SMP
2770
2771/*
2772 * Dump out nocb grace-period kthread state for the specified rcu_data
2773 * structure.
2774 */
2775static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2776{
2777 struct rcu_node *rnp = rdp->mynode;
2778
2779 pr_info("nocb GP %d %c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu %c CPU %d%s\n",
2780 rdp->cpu,
2781 "kK"[!!rdp->nocb_gp_kthread],
2782 "lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2783 "dD"[!!rdp->nocb_defer_wakeup],
2784 "tT"[timer_pending(&rdp->nocb_timer)],
2785 "sS"[!!rdp->nocb_gp_sleep],
2786 ".W"[swait_active(&rdp->nocb_gp_wq)],
2787 ".W"[swait_active(&rnp->nocb_gp_wq[0])],
2788 ".W"[swait_active(&rnp->nocb_gp_wq[1])],
2789 ".B"[!!rdp->nocb_gp_bypass],
2790 ".G"[!!rdp->nocb_gp_gp],
2791 (long)rdp->nocb_gp_seq,
2792 rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops),
2793 rdp->nocb_gp_kthread ? task_state_to_char(rdp->nocb_gp_kthread) : '.',
2794 rdp->nocb_cb_kthread ? (int)task_cpu(rdp->nocb_gp_kthread) : -1,
2795 show_rcu_should_be_on_cpu(rdp->nocb_cb_kthread));
2796}
2797
2798/* Dump out nocb kthread state for the specified rcu_data structure. */
2799static void show_rcu_nocb_state(struct rcu_data *rdp)
2800{
2801 char bufw[20];
2802 char bufr[20];
2803 struct rcu_segcblist *rsclp = &rdp->cblist;
2804 bool waslocked;
2805 bool wassleep;
2806
2807 if (rdp->nocb_gp_rdp == rdp)
2808 show_rcu_nocb_gp_state(rdp);
2809
2810 sprintf(bufw, "%ld", rsclp->gp_seq[RCU_WAIT_TAIL]);
2811 sprintf(bufr, "%ld", rsclp->gp_seq[RCU_NEXT_READY_TAIL]);
2812 pr_info(" CB %d^%d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%s%c%s%c%c q%ld %c CPU %d%s\n",
2813 rdp->cpu, rdp->nocb_gp_rdp->cpu,
2814 rdp->nocb_next_cb_rdp ? rdp->nocb_next_cb_rdp->cpu : -1,
2815 "kK"[!!rdp->nocb_cb_kthread],
2816 "bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2817 "cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2818 "lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2819 "sS"[!!rdp->nocb_cb_sleep],
2820 ".W"[swait_active(&rdp->nocb_cb_wq)],
2821 jiffies - rdp->nocb_bypass_first,
2822 jiffies - rdp->nocb_nobypass_last,
2823 rdp->nocb_nobypass_count,
2824 ".D"[rcu_segcblist_ready_cbs(rsclp)],
2825 ".W"[!rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL)],
2826 rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL) ? "" : bufw,
2827 ".R"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL)],
2828 rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL) ? "" : bufr,
2829 ".N"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_TAIL)],
2830 ".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2831 rcu_segcblist_n_cbs(&rdp->cblist),
2832 rdp->nocb_cb_kthread ? task_state_to_char(rdp->nocb_cb_kthread) : '.',
2833 rdp->nocb_cb_kthread ? (int)task_cpu(rdp->nocb_gp_kthread) : -1,
2834 show_rcu_should_be_on_cpu(rdp->nocb_cb_kthread));
2835
2836 /* It is OK for GP kthreads to have GP state. */
2837 if (rdp->nocb_gp_rdp == rdp)
2838 return;
2839
2840 waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2841 wassleep = swait_active(&rdp->nocb_gp_wq);
2842 if (!rdp->nocb_gp_sleep && !waslocked && !wassleep)
2843 return; /* Nothing untoward. */
2844
2845 pr_info(" nocb GP activity on CB-only CPU!!! %c%c%c %c\n",
2846 "lL"[waslocked],
2847 "dD"[!!rdp->nocb_defer_wakeup],
2848 "sS"[!!rdp->nocb_gp_sleep],
2849 ".W"[wassleep]);
2850}
2851
2852#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2853
2854/* No ->nocb_lock to acquire. */
2855static void rcu_nocb_lock(struct rcu_data *rdp)
2856{
2857}
2858
2859/* No ->nocb_lock to release. */
2860static void rcu_nocb_unlock(struct rcu_data *rdp)
2861{
2862}
2863
2864/* No ->nocb_lock to release. */
2865static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2866 unsigned long flags)
2867{
2868 local_irq_restore(flags);
2869}
2870
2871/* Lockdep check that ->cblist may be safely accessed. */
2872static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2873{
2874 lockdep_assert_irqs_disabled();
2875}
2876
2877static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2878{
2879}
2880
2881static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2882{
2883 return NULL;
2884}
2885
2886static void rcu_init_one_nocb(struct rcu_node *rnp)
2887{
2888}
2889
2890static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2891 unsigned long j)
2892{
2893 return true;
2894}
2895
2896static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2897 bool *was_alldone, unsigned long flags)
2898{
2899 return false;
2900}
2901
2902static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2903 unsigned long flags)
2904{
2905 WARN_ON_ONCE(1); /* Should be dead code! */
2906}
2907
2908static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2909{
2910}
2911
2912static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
2913{
2914 return false;
2915}
2916
2917static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
2918{
2919 return false;
2920}
2921
2922static void rcu_spawn_cpu_nocb_kthread(int cpu)
2923{
2924}
2925
2926static void __init rcu_spawn_nocb_kthreads(void)
2927{
2928}
2929
2930static void show_rcu_nocb_state(struct rcu_data *rdp)
2931{
2932}
2933
2934#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2935
2936/*
2937 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2938 * grace-period kthread will do force_quiescent_state() processing?
2939 * The idea is to avoid waking up RCU core processing on such a
2940 * CPU unless the grace period has extended for too long.
2941 *
2942 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2943 * CONFIG_RCU_NOCB_CPU CPUs.
2944 */
2945static bool rcu_nohz_full_cpu(void)
2946{
2947#ifdef CONFIG_NO_HZ_FULL
2948 if (tick_nohz_full_cpu(smp_processor_id()) &&
2949 (!rcu_gp_in_progress() ||
2950 time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2951 return true;
2952#endif /* #ifdef CONFIG_NO_HZ_FULL */
2953 return false;
2954}
2955
2956/*
2957 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2958 */
2959static void rcu_bind_gp_kthread(void)
2960{
2961 if (!tick_nohz_full_enabled())
2962 return;
2963 housekeeping_affine(current, HK_FLAG_RCU);
2964}
2965
2966/* Record the current task on dyntick-idle entry. */
2967static void noinstr rcu_dynticks_task_enter(void)
2968{
2969#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2970 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2971#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2972}
2973
2974/* Record no current task on dyntick-idle exit. */
2975static void noinstr rcu_dynticks_task_exit(void)
2976{
2977#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2978 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2979#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2980}
2981
2982/* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
2983static void rcu_dynticks_task_trace_enter(void)
2984{
2985#ifdef CONFIG_TASKS_TRACE_RCU
2986 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2987 current->trc_reader_special.b.need_mb = true;
2988#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
2989}
2990
2991/* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
2992static void rcu_dynticks_task_trace_exit(void)
2993{
2994#ifdef CONFIG_TASKS_TRACE_RCU
2995 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2996 current->trc_reader_special.b.need_mb = false;
2997#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
2998}