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 (jiffies_till_first_fqs != ULONG_MAX)
60 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
61 if (jiffies_till_next_fqs != ULONG_MAX)
62 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
63 if (jiffies_till_sched_qs != ULONG_MAX)
64 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
65 if (rcu_kick_kthreads)
66 pr_info("\tKick kthreads if too-long grace period.\n");
67 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
68 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
69 if (gp_preinit_delay)
70 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
71 if (gp_init_delay)
72 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
73 if (gp_cleanup_delay)
74 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
75 if (!use_softirq)
76 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
77 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
78 pr_info("\tRCU debug extended QS entry/exit.\n");
79 rcupdate_announce_bootup_oddness();
80}
81
82#ifdef CONFIG_PREEMPT_RCU
83
84static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
85static void rcu_read_unlock_special(struct task_struct *t);
86
87/*
88 * Tell them what RCU they are running.
89 */
90static void __init rcu_bootup_announce(void)
91{
92 pr_info("Preemptible hierarchical RCU implementation.\n");
93 rcu_bootup_announce_oddness();
94}
95
96/* Flags for rcu_preempt_ctxt_queue() decision table. */
97#define RCU_GP_TASKS 0x8
98#define RCU_EXP_TASKS 0x4
99#define RCU_GP_BLKD 0x2
100#define RCU_EXP_BLKD 0x1
101
102/*
103 * Queues a task preempted within an RCU-preempt read-side critical
104 * section into the appropriate location within the ->blkd_tasks list,
105 * depending on the states of any ongoing normal and expedited grace
106 * periods. The ->gp_tasks pointer indicates which element the normal
107 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
108 * indicates which element the expedited grace period is waiting on (again,
109 * NULL if none). If a grace period is waiting on a given element in the
110 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
111 * adding a task to the tail of the list blocks any grace period that is
112 * already waiting on one of the elements. In contrast, adding a task
113 * to the head of the list won't block any grace period that is already
114 * waiting on one of the elements.
115 *
116 * This queuing is imprecise, and can sometimes make an ongoing grace
117 * period wait for a task that is not strictly speaking blocking it.
118 * Given the choice, we needlessly block a normal grace period rather than
119 * blocking an expedited grace period.
120 *
121 * Note that an endless sequence of expedited grace periods still cannot
122 * indefinitely postpone a normal grace period. Eventually, all of the
123 * fixed number of preempted tasks blocking the normal grace period that are
124 * not also blocking the expedited grace period will resume and complete
125 * their RCU read-side critical sections. At that point, the ->gp_tasks
126 * pointer will equal the ->exp_tasks pointer, at which point the end of
127 * the corresponding expedited grace period will also be the end of the
128 * normal grace period.
129 */
130static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
131 __releases(rnp->lock) /* But leaves rrupts disabled. */
132{
133 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
134 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
135 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
136 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
137 struct task_struct *t = current;
138
139 raw_lockdep_assert_held_rcu_node(rnp);
140 WARN_ON_ONCE(rdp->mynode != rnp);
141 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
142 /* RCU better not be waiting on newly onlined CPUs! */
143 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
144 rdp->grpmask);
145
146 /*
147 * Decide where to queue the newly blocked task. In theory,
148 * this could be an if-statement. In practice, when I tried
149 * that, it was quite messy.
150 */
151 switch (blkd_state) {
152 case 0:
153 case RCU_EXP_TASKS:
154 case RCU_EXP_TASKS + RCU_GP_BLKD:
155 case RCU_GP_TASKS:
156 case RCU_GP_TASKS + RCU_EXP_TASKS:
157
158 /*
159 * Blocking neither GP, or first task blocking the normal
160 * GP but not blocking the already-waiting expedited GP.
161 * Queue at the head of the list to avoid unnecessarily
162 * blocking the already-waiting GPs.
163 */
164 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
165 break;
166
167 case RCU_EXP_BLKD:
168 case RCU_GP_BLKD:
169 case RCU_GP_BLKD + RCU_EXP_BLKD:
170 case RCU_GP_TASKS + RCU_EXP_BLKD:
171 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
172 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
173
174 /*
175 * First task arriving that blocks either GP, or first task
176 * arriving that blocks the expedited GP (with the normal
177 * GP already waiting), or a task arriving that blocks
178 * both GPs with both GPs already waiting. Queue at the
179 * tail of the list to avoid any GP waiting on any of the
180 * already queued tasks that are not blocking it.
181 */
182 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
183 break;
184
185 case RCU_EXP_TASKS + RCU_EXP_BLKD:
186 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
187 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
188
189 /*
190 * Second or subsequent task blocking the expedited GP.
191 * The task either does not block the normal GP, or is the
192 * first task blocking the normal GP. Queue just after
193 * the first task blocking the expedited GP.
194 */
195 list_add(&t->rcu_node_entry, rnp->exp_tasks);
196 break;
197
198 case RCU_GP_TASKS + RCU_GP_BLKD:
199 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
200
201 /*
202 * Second or subsequent task blocking the normal GP.
203 * The task does not block the expedited GP. Queue just
204 * after the first task blocking the normal GP.
205 */
206 list_add(&t->rcu_node_entry, rnp->gp_tasks);
207 break;
208
209 default:
210
211 /* Yet another exercise in excessive paranoia. */
212 WARN_ON_ONCE(1);
213 break;
214 }
215
216 /*
217 * We have now queued the task. If it was the first one to
218 * block either grace period, update the ->gp_tasks and/or
219 * ->exp_tasks pointers, respectively, to reference the newly
220 * blocked tasks.
221 */
222 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
223 rnp->gp_tasks = &t->rcu_node_entry;
224 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
225 }
226 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
227 rnp->exp_tasks = &t->rcu_node_entry;
228 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
229 !(rnp->qsmask & rdp->grpmask));
230 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
231 !(rnp->expmask & rdp->grpmask));
232 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
233
234 /*
235 * Report the quiescent state for the expedited GP. This expedited
236 * GP should not be able to end until we report, so there should be
237 * no need to check for a subsequent expedited GP. (Though we are
238 * still in a quiescent state in any case.)
239 */
240 if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
241 rcu_report_exp_rdp(rdp);
242 else
243 WARN_ON_ONCE(rdp->exp_deferred_qs);
244}
245
246/*
247 * Record a preemptible-RCU quiescent state for the specified CPU.
248 * Note that this does not necessarily mean that the task currently running
249 * on the CPU is in a quiescent state: Instead, it means that the current
250 * grace period need not wait on any RCU read-side critical section that
251 * starts later on this CPU. It also means that if the current task is
252 * in an RCU read-side critical section, it has already added itself to
253 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
254 * current task, there might be any number of other tasks blocked while
255 * in an RCU read-side critical section.
256 *
257 * Callers to this function must disable preemption.
258 */
259static void rcu_qs(void)
260{
261 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
262 if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
263 trace_rcu_grace_period(TPS("rcu_preempt"),
264 __this_cpu_read(rcu_data.gp_seq),
265 TPS("cpuqs"));
266 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
267 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
268 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
269 }
270}
271
272/*
273 * We have entered the scheduler, and the current task might soon be
274 * context-switched away from. If this task is in an RCU read-side
275 * critical section, we will no longer be able to rely on the CPU to
276 * record that fact, so we enqueue the task on the blkd_tasks list.
277 * The task will dequeue itself when it exits the outermost enclosing
278 * RCU read-side critical section. Therefore, the current grace period
279 * cannot be permitted to complete until the blkd_tasks list entries
280 * predating the current grace period drain, in other words, until
281 * rnp->gp_tasks becomes NULL.
282 *
283 * Caller must disable interrupts.
284 */
285void rcu_note_context_switch(bool preempt)
286{
287 struct task_struct *t = current;
288 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
289 struct rcu_node *rnp;
290
291 trace_rcu_utilization(TPS("Start context switch"));
292 lockdep_assert_irqs_disabled();
293 WARN_ON_ONCE(!preempt && t->rcu_read_lock_nesting > 0);
294 if (t->rcu_read_lock_nesting > 0 &&
295 !t->rcu_read_unlock_special.b.blocked) {
296
297 /* Possibly blocking in an RCU read-side critical section. */
298 rnp = rdp->mynode;
299 raw_spin_lock_rcu_node(rnp);
300 t->rcu_read_unlock_special.b.blocked = true;
301 t->rcu_blocked_node = rnp;
302
303 /*
304 * Verify the CPU's sanity, trace the preemption, and
305 * then queue the task as required based on the states
306 * of any ongoing and expedited grace periods.
307 */
308 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
309 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
310 trace_rcu_preempt_task(rcu_state.name,
311 t->pid,
312 (rnp->qsmask & rdp->grpmask)
313 ? rnp->gp_seq
314 : rcu_seq_snap(&rnp->gp_seq));
315 rcu_preempt_ctxt_queue(rnp, rdp);
316 } else {
317 rcu_preempt_deferred_qs(t);
318 }
319
320 /*
321 * Either we were not in an RCU read-side critical section to
322 * begin with, or we have now recorded that critical section
323 * globally. Either way, we can now note a quiescent state
324 * for this CPU. Again, if we were in an RCU read-side critical
325 * section, and if that critical section was blocking the current
326 * grace period, then the fact that the task has been enqueued
327 * means that we continue to block the current grace period.
328 */
329 rcu_qs();
330 if (rdp->exp_deferred_qs)
331 rcu_report_exp_rdp(rdp);
332 trace_rcu_utilization(TPS("End context switch"));
333}
334EXPORT_SYMBOL_GPL(rcu_note_context_switch);
335
336/*
337 * Check for preempted RCU readers blocking the current grace period
338 * for the specified rcu_node structure. If the caller needs a reliable
339 * answer, it must hold the rcu_node's ->lock.
340 */
341static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
342{
343 return rnp->gp_tasks != NULL;
344}
345
346/* Bias and limit values for ->rcu_read_lock_nesting. */
347#define RCU_NEST_BIAS INT_MAX
348#define RCU_NEST_NMAX (-INT_MAX / 2)
349#define RCU_NEST_PMAX (INT_MAX / 2)
350
351/*
352 * Preemptible RCU implementation for rcu_read_lock().
353 * Just increment ->rcu_read_lock_nesting, shared state will be updated
354 * if we block.
355 */
356void __rcu_read_lock(void)
357{
358 current->rcu_read_lock_nesting++;
359 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
360 WARN_ON_ONCE(current->rcu_read_lock_nesting > RCU_NEST_PMAX);
361 barrier(); /* critical section after entry code. */
362}
363EXPORT_SYMBOL_GPL(__rcu_read_lock);
364
365/*
366 * Preemptible RCU implementation for rcu_read_unlock().
367 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
368 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
369 * invoke rcu_read_unlock_special() to clean up after a context switch
370 * in an RCU read-side critical section and other special cases.
371 */
372void __rcu_read_unlock(void)
373{
374 struct task_struct *t = current;
375
376 if (t->rcu_read_lock_nesting != 1) {
377 --t->rcu_read_lock_nesting;
378 } else {
379 barrier(); /* critical section before exit code. */
380 t->rcu_read_lock_nesting = -RCU_NEST_BIAS;
381 barrier(); /* assign before ->rcu_read_unlock_special load */
382 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
383 rcu_read_unlock_special(t);
384 barrier(); /* ->rcu_read_unlock_special load before assign */
385 t->rcu_read_lock_nesting = 0;
386 }
387 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
388 int rrln = t->rcu_read_lock_nesting;
389
390 WARN_ON_ONCE(rrln < 0 && rrln > RCU_NEST_NMAX);
391 }
392}
393EXPORT_SYMBOL_GPL(__rcu_read_unlock);
394
395/*
396 * Advance a ->blkd_tasks-list pointer to the next entry, instead
397 * returning NULL if at the end of the list.
398 */
399static struct list_head *rcu_next_node_entry(struct task_struct *t,
400 struct rcu_node *rnp)
401{
402 struct list_head *np;
403
404 np = t->rcu_node_entry.next;
405 if (np == &rnp->blkd_tasks)
406 np = NULL;
407 return np;
408}
409
410/*
411 * Return true if the specified rcu_node structure has tasks that were
412 * preempted within an RCU read-side critical section.
413 */
414static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
415{
416 return !list_empty(&rnp->blkd_tasks);
417}
418
419/*
420 * Report deferred quiescent states. The deferral time can
421 * be quite short, for example, in the case of the call from
422 * rcu_read_unlock_special().
423 */
424static void
425rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
426{
427 bool empty_exp;
428 bool empty_norm;
429 bool empty_exp_now;
430 struct list_head *np;
431 bool drop_boost_mutex = false;
432 struct rcu_data *rdp;
433 struct rcu_node *rnp;
434 union rcu_special special;
435
436 /*
437 * If RCU core is waiting for this CPU to exit its critical section,
438 * report the fact that it has exited. Because irqs are disabled,
439 * t->rcu_read_unlock_special cannot change.
440 */
441 special = t->rcu_read_unlock_special;
442 rdp = this_cpu_ptr(&rcu_data);
443 if (!special.s && !rdp->exp_deferred_qs) {
444 local_irq_restore(flags);
445 return;
446 }
447 t->rcu_read_unlock_special.b.deferred_qs = false;
448 if (special.b.need_qs) {
449 rcu_qs();
450 t->rcu_read_unlock_special.b.need_qs = false;
451 if (!t->rcu_read_unlock_special.s && !rdp->exp_deferred_qs) {
452 local_irq_restore(flags);
453 return;
454 }
455 }
456
457 /*
458 * Respond to a request by an expedited grace period for a
459 * quiescent state from this CPU. Note that requests from
460 * tasks are handled when removing the task from the
461 * blocked-tasks list below.
462 */
463 if (rdp->exp_deferred_qs) {
464 rcu_report_exp_rdp(rdp);
465 if (!t->rcu_read_unlock_special.s) {
466 local_irq_restore(flags);
467 return;
468 }
469 }
470
471 /* Clean up if blocked during RCU read-side critical section. */
472 if (special.b.blocked) {
473 t->rcu_read_unlock_special.b.blocked = false;
474
475 /*
476 * Remove this task from the list it blocked on. The task
477 * now remains queued on the rcu_node corresponding to the
478 * CPU it first blocked on, so there is no longer any need
479 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
480 */
481 rnp = t->rcu_blocked_node;
482 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
483 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
484 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
485 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
486 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
487 (!empty_norm || rnp->qsmask));
488 empty_exp = sync_rcu_preempt_exp_done(rnp);
489 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
490 np = rcu_next_node_entry(t, rnp);
491 list_del_init(&t->rcu_node_entry);
492 t->rcu_blocked_node = NULL;
493 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
494 rnp->gp_seq, t->pid);
495 if (&t->rcu_node_entry == rnp->gp_tasks)
496 rnp->gp_tasks = np;
497 if (&t->rcu_node_entry == rnp->exp_tasks)
498 rnp->exp_tasks = np;
499 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
500 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
501 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
502 if (&t->rcu_node_entry == rnp->boost_tasks)
503 rnp->boost_tasks = np;
504 }
505
506 /*
507 * If this was the last task on the current list, and if
508 * we aren't waiting on any CPUs, report the quiescent state.
509 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
510 * so we must take a snapshot of the expedited state.
511 */
512 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
513 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
514 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
515 rnp->gp_seq,
516 0, rnp->qsmask,
517 rnp->level,
518 rnp->grplo,
519 rnp->grphi,
520 !!rnp->gp_tasks);
521 rcu_report_unblock_qs_rnp(rnp, flags);
522 } else {
523 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
524 }
525
526 /* Unboost if we were boosted. */
527 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
528 rt_mutex_futex_unlock(&rnp->boost_mtx);
529
530 /*
531 * If this was the last task on the expedited lists,
532 * then we need to report up the rcu_node hierarchy.
533 */
534 if (!empty_exp && empty_exp_now)
535 rcu_report_exp_rnp(rnp, true);
536 } else {
537 local_irq_restore(flags);
538 }
539}
540
541/*
542 * Is a deferred quiescent-state pending, and are we also not in
543 * an RCU read-side critical section? It is the caller's responsibility
544 * to ensure it is otherwise safe to report any deferred quiescent
545 * states. The reason for this is that it is safe to report a
546 * quiescent state during context switch even though preemption
547 * is disabled. This function cannot be expected to understand these
548 * nuances, so the caller must handle them.
549 */
550static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
551{
552 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
553 READ_ONCE(t->rcu_read_unlock_special.s)) &&
554 t->rcu_read_lock_nesting <= 0;
555}
556
557/*
558 * Report a deferred quiescent state if needed and safe to do so.
559 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
560 * not being in an RCU read-side critical section. The caller must
561 * evaluate safety in terms of interrupt, softirq, and preemption
562 * disabling.
563 */
564static void rcu_preempt_deferred_qs(struct task_struct *t)
565{
566 unsigned long flags;
567 bool couldrecurse = t->rcu_read_lock_nesting >= 0;
568
569 if (!rcu_preempt_need_deferred_qs(t))
570 return;
571 if (couldrecurse)
572 t->rcu_read_lock_nesting -= RCU_NEST_BIAS;
573 local_irq_save(flags);
574 rcu_preempt_deferred_qs_irqrestore(t, flags);
575 if (couldrecurse)
576 t->rcu_read_lock_nesting += RCU_NEST_BIAS;
577}
578
579/*
580 * Minimal handler to give the scheduler a chance to re-evaluate.
581 */
582static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
583{
584 struct rcu_data *rdp;
585
586 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
587 rdp->defer_qs_iw_pending = false;
588}
589
590/*
591 * Handle special cases during rcu_read_unlock(), such as needing to
592 * notify RCU core processing or task having blocked during the RCU
593 * read-side critical section.
594 */
595static void rcu_read_unlock_special(struct task_struct *t)
596{
597 unsigned long flags;
598 bool preempt_bh_were_disabled =
599 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
600 bool irqs_were_disabled;
601
602 /* NMI handlers cannot block and cannot safely manipulate state. */
603 if (in_nmi())
604 return;
605
606 local_irq_save(flags);
607 irqs_were_disabled = irqs_disabled_flags(flags);
608 if (preempt_bh_were_disabled || irqs_were_disabled) {
609 bool exp;
610 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
611 struct rcu_node *rnp = rdp->mynode;
612
613 t->rcu_read_unlock_special.b.exp_hint = false;
614 exp = (t->rcu_blocked_node && t->rcu_blocked_node->exp_tasks) ||
615 (rdp->grpmask & rnp->expmask) ||
616 tick_nohz_full_cpu(rdp->cpu);
617 // Need to defer quiescent state until everything is enabled.
618 if (irqs_were_disabled && use_softirq &&
619 (in_interrupt() ||
620 (exp && !t->rcu_read_unlock_special.b.deferred_qs))) {
621 // Using softirq, safe to awaken, and we get
622 // no help from enabling irqs, unlike bh/preempt.
623 raise_softirq_irqoff(RCU_SOFTIRQ);
624 } else {
625 // Enabling BH or preempt does reschedule, so...
626 // Also if no expediting or NO_HZ_FULL, slow is OK.
627 set_tsk_need_resched(current);
628 set_preempt_need_resched();
629 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
630 !rdp->defer_qs_iw_pending && exp) {
631 // Get scheduler to re-evaluate and call hooks.
632 // If !IRQ_WORK, FQS scan will eventually IPI.
633 init_irq_work(&rdp->defer_qs_iw,
634 rcu_preempt_deferred_qs_handler);
635 rdp->defer_qs_iw_pending = true;
636 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
637 }
638 }
639 t->rcu_read_unlock_special.b.deferred_qs = true;
640 local_irq_restore(flags);
641 return;
642 }
643 WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
644 rcu_preempt_deferred_qs_irqrestore(t, flags);
645}
646
647/*
648 * Check that the list of blocked tasks for the newly completed grace
649 * period is in fact empty. It is a serious bug to complete a grace
650 * period that still has RCU readers blocked! This function must be
651 * invoked -before- updating this rnp's ->gp_seq, and the rnp's ->lock
652 * must be held by the caller.
653 *
654 * Also, if there are blocked tasks on the list, they automatically
655 * block the newly created grace period, so set up ->gp_tasks accordingly.
656 */
657static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
658{
659 struct task_struct *t;
660
661 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
662 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
663 dump_blkd_tasks(rnp, 10);
664 if (rcu_preempt_has_tasks(rnp) &&
665 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
666 rnp->gp_tasks = rnp->blkd_tasks.next;
667 t = container_of(rnp->gp_tasks, struct task_struct,
668 rcu_node_entry);
669 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
670 rnp->gp_seq, t->pid);
671 }
672 WARN_ON_ONCE(rnp->qsmask);
673}
674
675/*
676 * Check for a quiescent state from the current CPU, including voluntary
677 * context switches for Tasks RCU. When a task blocks, the task is
678 * recorded in the corresponding CPU's rcu_node structure, which is checked
679 * elsewhere, hence this function need only check for quiescent states
680 * related to the current CPU, not to those related to tasks.
681 */
682static void rcu_flavor_sched_clock_irq(int user)
683{
684 struct task_struct *t = current;
685
686 if (user || rcu_is_cpu_rrupt_from_idle()) {
687 rcu_note_voluntary_context_switch(current);
688 }
689 if (t->rcu_read_lock_nesting > 0 ||
690 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
691 /* No QS, force context switch if deferred. */
692 if (rcu_preempt_need_deferred_qs(t)) {
693 set_tsk_need_resched(t);
694 set_preempt_need_resched();
695 }
696 } else if (rcu_preempt_need_deferred_qs(t)) {
697 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
698 return;
699 } else if (!t->rcu_read_lock_nesting) {
700 rcu_qs(); /* Report immediate QS. */
701 return;
702 }
703
704 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
705 if (t->rcu_read_lock_nesting > 0 &&
706 __this_cpu_read(rcu_data.core_needs_qs) &&
707 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
708 !t->rcu_read_unlock_special.b.need_qs &&
709 time_after(jiffies, rcu_state.gp_start + HZ))
710 t->rcu_read_unlock_special.b.need_qs = true;
711}
712
713/*
714 * Check for a task exiting while in a preemptible-RCU read-side
715 * critical section, clean up if so. No need to issue warnings, as
716 * debug_check_no_locks_held() already does this if lockdep is enabled.
717 * Besides, if this function does anything other than just immediately
718 * return, there was a bug of some sort. Spewing warnings from this
719 * function is like as not to simply obscure important prior warnings.
720 */
721void exit_rcu(void)
722{
723 struct task_struct *t = current;
724
725 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
726 t->rcu_read_lock_nesting = 1;
727 barrier();
728 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
729 } else if (unlikely(t->rcu_read_lock_nesting)) {
730 t->rcu_read_lock_nesting = 1;
731 } else {
732 return;
733 }
734 __rcu_read_unlock();
735 rcu_preempt_deferred_qs(current);
736}
737
738/*
739 * Dump the blocked-tasks state, but limit the list dump to the
740 * specified number of elements.
741 */
742static void
743dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
744{
745 int cpu;
746 int i;
747 struct list_head *lhp;
748 bool onl;
749 struct rcu_data *rdp;
750 struct rcu_node *rnp1;
751
752 raw_lockdep_assert_held_rcu_node(rnp);
753 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
754 __func__, rnp->grplo, rnp->grphi, rnp->level,
755 (long)rnp->gp_seq, (long)rnp->completedqs);
756 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
757 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
758 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
759 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
760 __func__, rnp->gp_tasks, rnp->boost_tasks, rnp->exp_tasks);
761 pr_info("%s: ->blkd_tasks", __func__);
762 i = 0;
763 list_for_each(lhp, &rnp->blkd_tasks) {
764 pr_cont(" %p", lhp);
765 if (++i >= ncheck)
766 break;
767 }
768 pr_cont("\n");
769 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
770 rdp = per_cpu_ptr(&rcu_data, cpu);
771 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
772 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
773 cpu, ".o"[onl],
774 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
775 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
776 }
777}
778
779#else /* #ifdef CONFIG_PREEMPT_RCU */
780
781/*
782 * Tell them what RCU they are running.
783 */
784static void __init rcu_bootup_announce(void)
785{
786 pr_info("Hierarchical RCU implementation.\n");
787 rcu_bootup_announce_oddness();
788}
789
790/*
791 * Note a quiescent state for PREEMPT=n. Because we do not need to know
792 * how many quiescent states passed, just if there was at least one since
793 * the start of the grace period, this just sets a flag. The caller must
794 * have disabled preemption.
795 */
796static void rcu_qs(void)
797{
798 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
799 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
800 return;
801 trace_rcu_grace_period(TPS("rcu_sched"),
802 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
803 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
804 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
805 return;
806 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
807 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
808}
809
810/*
811 * Register an urgently needed quiescent state. If there is an
812 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
813 * dyntick-idle quiescent state visible to other CPUs, which will in
814 * some cases serve for expedited as well as normal grace periods.
815 * Either way, register a lightweight quiescent state.
816 */
817void rcu_all_qs(void)
818{
819 unsigned long flags;
820
821 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
822 return;
823 preempt_disable();
824 /* Load rcu_urgent_qs before other flags. */
825 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
826 preempt_enable();
827 return;
828 }
829 this_cpu_write(rcu_data.rcu_urgent_qs, false);
830 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
831 local_irq_save(flags);
832 rcu_momentary_dyntick_idle();
833 local_irq_restore(flags);
834 }
835 rcu_qs();
836 preempt_enable();
837}
838EXPORT_SYMBOL_GPL(rcu_all_qs);
839
840/*
841 * Note a PREEMPT=n context switch. The caller must have disabled interrupts.
842 */
843void rcu_note_context_switch(bool preempt)
844{
845 trace_rcu_utilization(TPS("Start context switch"));
846 rcu_qs();
847 /* Load rcu_urgent_qs before other flags. */
848 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
849 goto out;
850 this_cpu_write(rcu_data.rcu_urgent_qs, false);
851 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
852 rcu_momentary_dyntick_idle();
853 if (!preempt)
854 rcu_tasks_qs(current);
855out:
856 trace_rcu_utilization(TPS("End context switch"));
857}
858EXPORT_SYMBOL_GPL(rcu_note_context_switch);
859
860/*
861 * Because preemptible RCU does not exist, there are never any preempted
862 * RCU readers.
863 */
864static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
865{
866 return 0;
867}
868
869/*
870 * Because there is no preemptible RCU, there can be no readers blocked.
871 */
872static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
873{
874 return false;
875}
876
877/*
878 * Because there is no preemptible RCU, there can be no deferred quiescent
879 * states.
880 */
881static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
882{
883 return false;
884}
885static void rcu_preempt_deferred_qs(struct task_struct *t) { }
886
887/*
888 * Because there is no preemptible RCU, there can be no readers blocked,
889 * so there is no need to check for blocked tasks. So check only for
890 * bogus qsmask values.
891 */
892static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
893{
894 WARN_ON_ONCE(rnp->qsmask);
895}
896
897/*
898 * Check to see if this CPU is in a non-context-switch quiescent state,
899 * namely user mode and idle loop.
900 */
901static void rcu_flavor_sched_clock_irq(int user)
902{
903 if (user || rcu_is_cpu_rrupt_from_idle()) {
904
905 /*
906 * Get here if this CPU took its interrupt from user
907 * mode or from the idle loop, and if this is not a
908 * nested interrupt. In this case, the CPU is in
909 * a quiescent state, so note it.
910 *
911 * No memory barrier is required here because rcu_qs()
912 * references only CPU-local variables that other CPUs
913 * neither access nor modify, at least not while the
914 * corresponding CPU is online.
915 */
916
917 rcu_qs();
918 }
919}
920
921/*
922 * Because preemptible RCU does not exist, tasks cannot possibly exit
923 * while in preemptible RCU read-side critical sections.
924 */
925void exit_rcu(void)
926{
927}
928
929/*
930 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
931 */
932static void
933dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
934{
935 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
936}
937
938#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
939
940/*
941 * If boosting, set rcuc kthreads to realtime priority.
942 */
943static void rcu_cpu_kthread_setup(unsigned int cpu)
944{
945#ifdef CONFIG_RCU_BOOST
946 struct sched_param sp;
947
948 sp.sched_priority = kthread_prio;
949 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
950#endif /* #ifdef CONFIG_RCU_BOOST */
951}
952
953#ifdef CONFIG_RCU_BOOST
954
955/*
956 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
957 * or ->boost_tasks, advancing the pointer to the next task in the
958 * ->blkd_tasks list.
959 *
960 * Note that irqs must be enabled: boosting the task can block.
961 * Returns 1 if there are more tasks needing to be boosted.
962 */
963static int rcu_boost(struct rcu_node *rnp)
964{
965 unsigned long flags;
966 struct task_struct *t;
967 struct list_head *tb;
968
969 if (READ_ONCE(rnp->exp_tasks) == NULL &&
970 READ_ONCE(rnp->boost_tasks) == NULL)
971 return 0; /* Nothing left to boost. */
972
973 raw_spin_lock_irqsave_rcu_node(rnp, flags);
974
975 /*
976 * Recheck under the lock: all tasks in need of boosting
977 * might exit their RCU read-side critical sections on their own.
978 */
979 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
980 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
981 return 0;
982 }
983
984 /*
985 * Preferentially boost tasks blocking expedited grace periods.
986 * This cannot starve the normal grace periods because a second
987 * expedited grace period must boost all blocked tasks, including
988 * those blocking the pre-existing normal grace period.
989 */
990 if (rnp->exp_tasks != NULL)
991 tb = rnp->exp_tasks;
992 else
993 tb = rnp->boost_tasks;
994
995 /*
996 * We boost task t by manufacturing an rt_mutex that appears to
997 * be held by task t. We leave a pointer to that rt_mutex where
998 * task t can find it, and task t will release the mutex when it
999 * exits its outermost RCU read-side critical section. Then
1000 * simply acquiring this artificial rt_mutex will boost task
1001 * t's priority. (Thanks to tglx for suggesting this approach!)
1002 *
1003 * Note that task t must acquire rnp->lock to remove itself from
1004 * the ->blkd_tasks list, which it will do from exit() if from
1005 * nowhere else. We therefore are guaranteed that task t will
1006 * stay around at least until we drop rnp->lock. Note that
1007 * rnp->lock also resolves races between our priority boosting
1008 * and task t's exiting its outermost RCU read-side critical
1009 * section.
1010 */
1011 t = container_of(tb, struct task_struct, rcu_node_entry);
1012 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1013 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1014 /* Lock only for side effect: boosts task t's priority. */
1015 rt_mutex_lock(&rnp->boost_mtx);
1016 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1017
1018 return READ_ONCE(rnp->exp_tasks) != NULL ||
1019 READ_ONCE(rnp->boost_tasks) != NULL;
1020}
1021
1022/*
1023 * Priority-boosting kthread, one per leaf rcu_node.
1024 */
1025static int rcu_boost_kthread(void *arg)
1026{
1027 struct rcu_node *rnp = (struct rcu_node *)arg;
1028 int spincnt = 0;
1029 int more2boost;
1030
1031 trace_rcu_utilization(TPS("Start boost kthread@init"));
1032 for (;;) {
1033 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1034 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1035 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1036 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1037 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1038 more2boost = rcu_boost(rnp);
1039 if (more2boost)
1040 spincnt++;
1041 else
1042 spincnt = 0;
1043 if (spincnt > 10) {
1044 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1045 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1046 schedule_timeout_interruptible(2);
1047 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1048 spincnt = 0;
1049 }
1050 }
1051 /* NOTREACHED */
1052 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1053 return 0;
1054}
1055
1056/*
1057 * Check to see if it is time to start boosting RCU readers that are
1058 * blocking the current grace period, and, if so, tell the per-rcu_node
1059 * kthread to start boosting them. If there is an expedited grace
1060 * period in progress, it is always time to boost.
1061 *
1062 * The caller must hold rnp->lock, which this function releases.
1063 * The ->boost_kthread_task is immortal, so we don't need to worry
1064 * about it going away.
1065 */
1066static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1067 __releases(rnp->lock)
1068{
1069 raw_lockdep_assert_held_rcu_node(rnp);
1070 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1071 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1072 return;
1073 }
1074 if (rnp->exp_tasks != NULL ||
1075 (rnp->gp_tasks != NULL &&
1076 rnp->boost_tasks == NULL &&
1077 rnp->qsmask == 0 &&
1078 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1079 if (rnp->exp_tasks == NULL)
1080 rnp->boost_tasks = rnp->gp_tasks;
1081 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1082 rcu_wake_cond(rnp->boost_kthread_task,
1083 rnp->boost_kthread_status);
1084 } else {
1085 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1086 }
1087}
1088
1089/*
1090 * Is the current CPU running the RCU-callbacks kthread?
1091 * Caller must have preemption disabled.
1092 */
1093static bool rcu_is_callbacks_kthread(void)
1094{
1095 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1096}
1097
1098#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1099
1100/*
1101 * Do priority-boost accounting for the start of a new grace period.
1102 */
1103static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1104{
1105 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1106}
1107
1108/*
1109 * Create an RCU-boost kthread for the specified node if one does not
1110 * already exist. We only create this kthread for preemptible RCU.
1111 * Returns zero if all is well, a negated errno otherwise.
1112 */
1113static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1114{
1115 int rnp_index = rnp - rcu_get_root();
1116 unsigned long flags;
1117 struct sched_param sp;
1118 struct task_struct *t;
1119
1120 if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1121 return;
1122
1123 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1124 return;
1125
1126 rcu_state.boost = 1;
1127
1128 if (rnp->boost_kthread_task != NULL)
1129 return;
1130
1131 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1132 "rcub/%d", rnp_index);
1133 if (WARN_ON_ONCE(IS_ERR(t)))
1134 return;
1135
1136 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1137 rnp->boost_kthread_task = t;
1138 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1139 sp.sched_priority = kthread_prio;
1140 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1141 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1142}
1143
1144/*
1145 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1146 * served by the rcu_node in question. The CPU hotplug lock is still
1147 * held, so the value of rnp->qsmaskinit will be stable.
1148 *
1149 * We don't include outgoingcpu in the affinity set, use -1 if there is
1150 * no outgoing CPU. If there are no CPUs left in the affinity set,
1151 * this function allows the kthread to execute on any CPU.
1152 */
1153static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1154{
1155 struct task_struct *t = rnp->boost_kthread_task;
1156 unsigned long mask = rcu_rnp_online_cpus(rnp);
1157 cpumask_var_t cm;
1158 int cpu;
1159
1160 if (!t)
1161 return;
1162 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1163 return;
1164 for_each_leaf_node_possible_cpu(rnp, cpu)
1165 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1166 cpu != outgoingcpu)
1167 cpumask_set_cpu(cpu, cm);
1168 if (cpumask_weight(cm) == 0)
1169 cpumask_setall(cm);
1170 set_cpus_allowed_ptr(t, cm);
1171 free_cpumask_var(cm);
1172}
1173
1174/*
1175 * Spawn boost kthreads -- called as soon as the scheduler is running.
1176 */
1177static void __init rcu_spawn_boost_kthreads(void)
1178{
1179 struct rcu_node *rnp;
1180
1181 rcu_for_each_leaf_node(rnp)
1182 rcu_spawn_one_boost_kthread(rnp);
1183}
1184
1185static void rcu_prepare_kthreads(int cpu)
1186{
1187 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1188 struct rcu_node *rnp = rdp->mynode;
1189
1190 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1191 if (rcu_scheduler_fully_active)
1192 rcu_spawn_one_boost_kthread(rnp);
1193}
1194
1195#else /* #ifdef CONFIG_RCU_BOOST */
1196
1197static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1198 __releases(rnp->lock)
1199{
1200 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1201}
1202
1203static bool rcu_is_callbacks_kthread(void)
1204{
1205 return false;
1206}
1207
1208static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1209{
1210}
1211
1212static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1213{
1214}
1215
1216static void __init rcu_spawn_boost_kthreads(void)
1217{
1218}
1219
1220static void rcu_prepare_kthreads(int cpu)
1221{
1222}
1223
1224#endif /* #else #ifdef CONFIG_RCU_BOOST */
1225
1226#if !defined(CONFIG_RCU_FAST_NO_HZ)
1227
1228/*
1229 * Check to see if any future non-offloaded RCU-related work will need
1230 * to be done by the current CPU, even if none need be done immediately,
1231 * returning 1 if so. This function is part of the RCU implementation;
1232 * it is -not- an exported member of the RCU API.
1233 *
1234 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1235 * CPU has RCU callbacks queued.
1236 */
1237int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1238{
1239 *nextevt = KTIME_MAX;
1240 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1241 !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist);
1242}
1243
1244/*
1245 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1246 * after it.
1247 */
1248static void rcu_cleanup_after_idle(void)
1249{
1250}
1251
1252/*
1253 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1254 * is nothing.
1255 */
1256static void rcu_prepare_for_idle(void)
1257{
1258}
1259
1260#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1261
1262/*
1263 * This code is invoked when a CPU goes idle, at which point we want
1264 * to have the CPU do everything required for RCU so that it can enter
1265 * the energy-efficient dyntick-idle mode. This is handled by a
1266 * state machine implemented by rcu_prepare_for_idle() below.
1267 *
1268 * The following three proprocessor symbols control this state machine:
1269 *
1270 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1271 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1272 * is sized to be roughly one RCU grace period. Those energy-efficiency
1273 * benchmarkers who might otherwise be tempted to set this to a large
1274 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1275 * system. And if you are -that- concerned about energy efficiency,
1276 * just power the system down and be done with it!
1277 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1278 * permitted to sleep in dyntick-idle mode with only lazy RCU
1279 * callbacks pending. Setting this too high can OOM your system.
1280 *
1281 * The values below work well in practice. If future workloads require
1282 * adjustment, they can be converted into kernel config parameters, though
1283 * making the state machine smarter might be a better option.
1284 */
1285#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1286#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1287
1288static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1289module_param(rcu_idle_gp_delay, int, 0644);
1290static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1291module_param(rcu_idle_lazy_gp_delay, int, 0644);
1292
1293/*
1294 * Try to advance callbacks on the current CPU, but only if it has been
1295 * awhile since the last time we did so. Afterwards, if there are any
1296 * callbacks ready for immediate invocation, return true.
1297 */
1298static bool __maybe_unused rcu_try_advance_all_cbs(void)
1299{
1300 bool cbs_ready = false;
1301 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1302 struct rcu_node *rnp;
1303
1304 /* Exit early if we advanced recently. */
1305 if (jiffies == rdp->last_advance_all)
1306 return false;
1307 rdp->last_advance_all = jiffies;
1308
1309 rnp = rdp->mynode;
1310
1311 /*
1312 * Don't bother checking unless a grace period has
1313 * completed since we last checked and there are
1314 * callbacks not yet ready to invoke.
1315 */
1316 if ((rcu_seq_completed_gp(rdp->gp_seq,
1317 rcu_seq_current(&rnp->gp_seq)) ||
1318 unlikely(READ_ONCE(rdp->gpwrap))) &&
1319 rcu_segcblist_pend_cbs(&rdp->cblist))
1320 note_gp_changes(rdp);
1321
1322 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1323 cbs_ready = true;
1324 return cbs_ready;
1325}
1326
1327/*
1328 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1329 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1330 * caller to set the timeout based on whether or not there are non-lazy
1331 * callbacks.
1332 *
1333 * The caller must have disabled interrupts.
1334 */
1335int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1336{
1337 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1338 unsigned long dj;
1339
1340 lockdep_assert_irqs_disabled();
1341
1342 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1343 if (rcu_segcblist_empty(&rdp->cblist) ||
1344 rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist)) {
1345 *nextevt = KTIME_MAX;
1346 return 0;
1347 }
1348
1349 /* Attempt to advance callbacks. */
1350 if (rcu_try_advance_all_cbs()) {
1351 /* Some ready to invoke, so initiate later invocation. */
1352 invoke_rcu_core();
1353 return 1;
1354 }
1355 rdp->last_accelerate = jiffies;
1356
1357 /* Request timer delay depending on laziness, and round. */
1358 rdp->all_lazy = !rcu_segcblist_n_nonlazy_cbs(&rdp->cblist);
1359 if (rdp->all_lazy) {
1360 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1361 } else {
1362 dj = round_up(rcu_idle_gp_delay + jiffies,
1363 rcu_idle_gp_delay) - jiffies;
1364 }
1365 *nextevt = basemono + dj * TICK_NSEC;
1366 return 0;
1367}
1368
1369/*
1370 * Prepare a CPU for idle from an RCU perspective. The first major task
1371 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1372 * The second major task is to check to see if a non-lazy callback has
1373 * arrived at a CPU that previously had only lazy callbacks. The third
1374 * major task is to accelerate (that is, assign grace-period numbers to)
1375 * any recently arrived callbacks.
1376 *
1377 * The caller must have disabled interrupts.
1378 */
1379static void rcu_prepare_for_idle(void)
1380{
1381 bool needwake;
1382 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1383 struct rcu_node *rnp;
1384 int tne;
1385
1386 lockdep_assert_irqs_disabled();
1387 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1388 return;
1389
1390 /* Handle nohz enablement switches conservatively. */
1391 tne = READ_ONCE(tick_nohz_active);
1392 if (tne != rdp->tick_nohz_enabled_snap) {
1393 if (!rcu_segcblist_empty(&rdp->cblist))
1394 invoke_rcu_core(); /* force nohz to see update. */
1395 rdp->tick_nohz_enabled_snap = tne;
1396 return;
1397 }
1398 if (!tne)
1399 return;
1400
1401 /*
1402 * If a non-lazy callback arrived at a CPU having only lazy
1403 * callbacks, invoke RCU core for the side-effect of recalculating
1404 * idle duration on re-entry to idle.
1405 */
1406 if (rdp->all_lazy && rcu_segcblist_n_nonlazy_cbs(&rdp->cblist)) {
1407 rdp->all_lazy = false;
1408 invoke_rcu_core();
1409 return;
1410 }
1411
1412 /*
1413 * If we have not yet accelerated this jiffy, accelerate all
1414 * callbacks on this CPU.
1415 */
1416 if (rdp->last_accelerate == jiffies)
1417 return;
1418 rdp->last_accelerate = jiffies;
1419 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1420 rnp = rdp->mynode;
1421 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1422 needwake = rcu_accelerate_cbs(rnp, rdp);
1423 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1424 if (needwake)
1425 rcu_gp_kthread_wake();
1426 }
1427}
1428
1429/*
1430 * Clean up for exit from idle. Attempt to advance callbacks based on
1431 * any grace periods that elapsed while the CPU was idle, and if any
1432 * callbacks are now ready to invoke, initiate invocation.
1433 */
1434static void rcu_cleanup_after_idle(void)
1435{
1436 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1437
1438 lockdep_assert_irqs_disabled();
1439 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1440 return;
1441 if (rcu_try_advance_all_cbs())
1442 invoke_rcu_core();
1443}
1444
1445#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1446
1447#ifdef CONFIG_RCU_NOCB_CPU
1448
1449/*
1450 * Offload callback processing from the boot-time-specified set of CPUs
1451 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1452 * created that pull the callbacks from the corresponding CPU, wait for
1453 * a grace period to elapse, and invoke the callbacks. These kthreads
1454 * are organized into GP kthreads, which manage incoming callbacks, wait for
1455 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1456 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1457 * do a wake_up() on their GP kthread when they insert a callback into any
1458 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1459 * in which case each kthread actively polls its CPU. (Which isn't so great
1460 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1461 *
1462 * This is intended to be used in conjunction with Frederic Weisbecker's
1463 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1464 * running CPU-bound user-mode computations.
1465 *
1466 * Offloading of callbacks can also be used as an energy-efficiency
1467 * measure because CPUs with no RCU callbacks queued are more aggressive
1468 * about entering dyntick-idle mode.
1469 */
1470
1471
1472/*
1473 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1474 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1475 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1476 * given, a warning is emitted and all CPUs are offloaded.
1477 */
1478static int __init rcu_nocb_setup(char *str)
1479{
1480 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1481 if (!strcasecmp(str, "all"))
1482 cpumask_setall(rcu_nocb_mask);
1483 else
1484 if (cpulist_parse(str, rcu_nocb_mask)) {
1485 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1486 cpumask_setall(rcu_nocb_mask);
1487 }
1488 return 1;
1489}
1490__setup("rcu_nocbs=", rcu_nocb_setup);
1491
1492static int __init parse_rcu_nocb_poll(char *arg)
1493{
1494 rcu_nocb_poll = true;
1495 return 0;
1496}
1497early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1498
1499/*
1500 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1501 * After all, the main point of bypassing is to avoid lock contention
1502 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1503 */
1504int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1505module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1506
1507/*
1508 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1509 * lock isn't immediately available, increment ->nocb_lock_contended to
1510 * flag the contention.
1511 */
1512static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1513{
1514 lockdep_assert_irqs_disabled();
1515 if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1516 return;
1517 atomic_inc(&rdp->nocb_lock_contended);
1518 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1519 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1520 raw_spin_lock(&rdp->nocb_bypass_lock);
1521 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1522 atomic_dec(&rdp->nocb_lock_contended);
1523}
1524
1525/*
1526 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1527 * not contended. Please note that this is extremely special-purpose,
1528 * relying on the fact that at most two kthreads and one CPU contend for
1529 * this lock, and also that the two kthreads are guaranteed to have frequent
1530 * grace-period-duration time intervals between successive acquisitions
1531 * of the lock. This allows us to use an extremely simple throttling
1532 * mechanism, and further to apply it only to the CPU doing floods of
1533 * call_rcu() invocations. Don't try this at home!
1534 */
1535static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1536{
1537 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1538 while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1539 cpu_relax();
1540}
1541
1542/*
1543 * Conditionally acquire the specified rcu_data structure's
1544 * ->nocb_bypass_lock.
1545 */
1546static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1547{
1548 lockdep_assert_irqs_disabled();
1549 return raw_spin_trylock(&rdp->nocb_bypass_lock);
1550}
1551
1552/*
1553 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1554 */
1555static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1556{
1557 lockdep_assert_irqs_disabled();
1558 raw_spin_unlock(&rdp->nocb_bypass_lock);
1559}
1560
1561/*
1562 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1563 * if it corresponds to a no-CBs CPU.
1564 */
1565static void rcu_nocb_lock(struct rcu_data *rdp)
1566{
1567 lockdep_assert_irqs_disabled();
1568 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1569 return;
1570 raw_spin_lock(&rdp->nocb_lock);
1571}
1572
1573/*
1574 * Release the specified rcu_data structure's ->nocb_lock, but only
1575 * if it corresponds to a no-CBs CPU.
1576 */
1577static void rcu_nocb_unlock(struct rcu_data *rdp)
1578{
1579 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1580 lockdep_assert_irqs_disabled();
1581 raw_spin_unlock(&rdp->nocb_lock);
1582 }
1583}
1584
1585/*
1586 * Release the specified rcu_data structure's ->nocb_lock and restore
1587 * interrupts, but only if it corresponds to a no-CBs CPU.
1588 */
1589static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1590 unsigned long flags)
1591{
1592 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1593 lockdep_assert_irqs_disabled();
1594 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1595 } else {
1596 local_irq_restore(flags);
1597 }
1598}
1599
1600/* Lockdep check that ->cblist may be safely accessed. */
1601static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1602{
1603 lockdep_assert_irqs_disabled();
1604 if (rcu_segcblist_is_offloaded(&rdp->cblist) &&
1605 cpu_online(rdp->cpu))
1606 lockdep_assert_held(&rdp->nocb_lock);
1607}
1608
1609/*
1610 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1611 * grace period.
1612 */
1613static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1614{
1615 swake_up_all(sq);
1616}
1617
1618static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1619{
1620 return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1621}
1622
1623static void rcu_init_one_nocb(struct rcu_node *rnp)
1624{
1625 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1626 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1627}
1628
1629/* Is the specified CPU a no-CBs CPU? */
1630bool rcu_is_nocb_cpu(int cpu)
1631{
1632 if (cpumask_available(rcu_nocb_mask))
1633 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1634 return false;
1635}
1636
1637/*
1638 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1639 * and this function releases it.
1640 */
1641static void wake_nocb_gp(struct rcu_data *rdp, bool force,
1642 unsigned long flags)
1643 __releases(rdp->nocb_lock)
1644{
1645 bool needwake = false;
1646 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1647
1648 lockdep_assert_held(&rdp->nocb_lock);
1649 if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1650 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1651 TPS("AlreadyAwake"));
1652 rcu_nocb_unlock_irqrestore(rdp, flags);
1653 return;
1654 }
1655 del_timer(&rdp->nocb_timer);
1656 rcu_nocb_unlock_irqrestore(rdp, flags);
1657 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1658 if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1659 WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1660 needwake = true;
1661 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1662 }
1663 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1664 if (needwake)
1665 wake_up_process(rdp_gp->nocb_gp_kthread);
1666}
1667
1668/*
1669 * Arrange to wake the GP kthread for this NOCB group at some future
1670 * time when it is safe to do so.
1671 */
1672static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1673 const char *reason)
1674{
1675 if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1676 mod_timer(&rdp->nocb_timer, jiffies + 1);
1677 if (rdp->nocb_defer_wakeup < waketype)
1678 WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1679 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1680}
1681
1682/*
1683 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1684 * However, if there is a callback to be enqueued and if ->nocb_bypass
1685 * proves to be initially empty, just return false because the no-CB GP
1686 * kthread may need to be awakened in this case.
1687 *
1688 * Note that this function always returns true if rhp is NULL.
1689 */
1690static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1691 unsigned long j)
1692{
1693 struct rcu_cblist rcl;
1694
1695 WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp->cblist));
1696 rcu_lockdep_assert_cblist_protected(rdp);
1697 lockdep_assert_held(&rdp->nocb_bypass_lock);
1698 if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1699 raw_spin_unlock(&rdp->nocb_bypass_lock);
1700 return false;
1701 }
1702 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1703 if (rhp)
1704 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1705 rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1706 rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1707 WRITE_ONCE(rdp->nocb_bypass_first, j);
1708 rcu_nocb_bypass_unlock(rdp);
1709 return true;
1710}
1711
1712/*
1713 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1714 * However, if there is a callback to be enqueued and if ->nocb_bypass
1715 * proves to be initially empty, just return false because the no-CB GP
1716 * kthread may need to be awakened in this case.
1717 *
1718 * Note that this function always returns true if rhp is NULL.
1719 */
1720static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1721 unsigned long j)
1722{
1723 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1724 return true;
1725 rcu_lockdep_assert_cblist_protected(rdp);
1726 rcu_nocb_bypass_lock(rdp);
1727 return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1728}
1729
1730/*
1731 * If the ->nocb_bypass_lock is immediately available, flush the
1732 * ->nocb_bypass queue into ->cblist.
1733 */
1734static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1735{
1736 rcu_lockdep_assert_cblist_protected(rdp);
1737 if (!rcu_segcblist_is_offloaded(&rdp->cblist) ||
1738 !rcu_nocb_bypass_trylock(rdp))
1739 return;
1740 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1741}
1742
1743/*
1744 * See whether it is appropriate to use the ->nocb_bypass list in order
1745 * to control contention on ->nocb_lock. A limited number of direct
1746 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1747 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1748 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1749 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1750 * used if ->cblist is empty, because otherwise callbacks can be stranded
1751 * on ->nocb_bypass because we cannot count on the current CPU ever again
1752 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1753 * non-empty, the corresponding no-CBs grace-period kthread must not be
1754 * in an indefinite sleep state.
1755 *
1756 * Finally, it is not permitted to use the bypass during early boot,
1757 * as doing so would confuse the auto-initialization code. Besides
1758 * which, there is no point in worrying about lock contention while
1759 * there is only one CPU in operation.
1760 */
1761static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1762 bool *was_alldone, unsigned long flags)
1763{
1764 unsigned long c;
1765 unsigned long cur_gp_seq;
1766 unsigned long j = jiffies;
1767 long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1768
1769 if (!rcu_segcblist_is_offloaded(&rdp->cblist)) {
1770 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1771 return false; /* Not offloaded, no bypassing. */
1772 }
1773 lockdep_assert_irqs_disabled();
1774
1775 // Don't use ->nocb_bypass during early boot.
1776 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1777 rcu_nocb_lock(rdp);
1778 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1779 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1780 return false;
1781 }
1782
1783 // If we have advanced to a new jiffy, reset counts to allow
1784 // moving back from ->nocb_bypass to ->cblist.
1785 if (j == rdp->nocb_nobypass_last) {
1786 c = rdp->nocb_nobypass_count + 1;
1787 } else {
1788 WRITE_ONCE(rdp->nocb_nobypass_last, j);
1789 c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1790 if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1791 nocb_nobypass_lim_per_jiffy))
1792 c = 0;
1793 else if (c > nocb_nobypass_lim_per_jiffy)
1794 c = nocb_nobypass_lim_per_jiffy;
1795 }
1796 WRITE_ONCE(rdp->nocb_nobypass_count, c);
1797
1798 // If there hasn't yet been all that many ->cblist enqueues
1799 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1800 // ->nocb_bypass first.
1801 if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1802 rcu_nocb_lock(rdp);
1803 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1804 if (*was_alldone)
1805 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1806 TPS("FirstQ"));
1807 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1808 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1809 return false; // Caller must enqueue the callback.
1810 }
1811
1812 // If ->nocb_bypass has been used too long or is too full,
1813 // flush ->nocb_bypass to ->cblist.
1814 if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1815 ncbs >= qhimark) {
1816 rcu_nocb_lock(rdp);
1817 if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1818 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1819 if (*was_alldone)
1820 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1821 TPS("FirstQ"));
1822 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1823 return false; // Caller must enqueue the callback.
1824 }
1825 if (j != rdp->nocb_gp_adv_time &&
1826 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1827 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1828 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1829 rdp->nocb_gp_adv_time = j;
1830 }
1831 rcu_nocb_unlock_irqrestore(rdp, flags);
1832 return true; // Callback already enqueued.
1833 }
1834
1835 // We need to use the bypass.
1836 rcu_nocb_wait_contended(rdp);
1837 rcu_nocb_bypass_lock(rdp);
1838 ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1839 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1840 rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1841 if (!ncbs) {
1842 WRITE_ONCE(rdp->nocb_bypass_first, j);
1843 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1844 }
1845 rcu_nocb_bypass_unlock(rdp);
1846 smp_mb(); /* Order enqueue before wake. */
1847 if (ncbs) {
1848 local_irq_restore(flags);
1849 } else {
1850 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1851 rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1852 if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1853 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1854 TPS("FirstBQwake"));
1855 __call_rcu_nocb_wake(rdp, true, flags);
1856 } else {
1857 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1858 TPS("FirstBQnoWake"));
1859 rcu_nocb_unlock_irqrestore(rdp, flags);
1860 }
1861 }
1862 return true; // Callback already enqueued.
1863}
1864
1865/*
1866 * Awaken the no-CBs grace-period kthead if needed, either due to it
1867 * legitimately being asleep or due to overload conditions.
1868 *
1869 * If warranted, also wake up the kthread servicing this CPUs queues.
1870 */
1871static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1872 unsigned long flags)
1873 __releases(rdp->nocb_lock)
1874{
1875 unsigned long cur_gp_seq;
1876 unsigned long j;
1877 long len;
1878 struct task_struct *t;
1879
1880 // If we are being polled or there is no kthread, just leave.
1881 t = READ_ONCE(rdp->nocb_gp_kthread);
1882 if (rcu_nocb_poll || !t) {
1883 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1884 TPS("WakeNotPoll"));
1885 rcu_nocb_unlock_irqrestore(rdp, flags);
1886 return;
1887 }
1888 // Need to actually to a wakeup.
1889 len = rcu_segcblist_n_cbs(&rdp->cblist);
1890 if (was_alldone) {
1891 rdp->qlen_last_fqs_check = len;
1892 if (!irqs_disabled_flags(flags)) {
1893 /* ... if queue was empty ... */
1894 wake_nocb_gp(rdp, false, flags);
1895 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1896 TPS("WakeEmpty"));
1897 } else {
1898 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1899 TPS("WakeEmptyIsDeferred"));
1900 rcu_nocb_unlock_irqrestore(rdp, flags);
1901 }
1902 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1903 /* ... or if many callbacks queued. */
1904 rdp->qlen_last_fqs_check = len;
1905 j = jiffies;
1906 if (j != rdp->nocb_gp_adv_time &&
1907 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1908 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1909 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1910 rdp->nocb_gp_adv_time = j;
1911 }
1912 smp_mb(); /* Enqueue before timer_pending(). */
1913 if ((rdp->nocb_cb_sleep ||
1914 !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1915 !timer_pending(&rdp->nocb_bypass_timer))
1916 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1917 TPS("WakeOvfIsDeferred"));
1918 rcu_nocb_unlock_irqrestore(rdp, flags);
1919 } else {
1920 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1921 rcu_nocb_unlock_irqrestore(rdp, flags);
1922 }
1923 return;
1924}
1925
1926/* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1927static void do_nocb_bypass_wakeup_timer(struct timer_list *t)
1928{
1929 unsigned long flags;
1930 struct rcu_data *rdp = from_timer(rdp, t, nocb_bypass_timer);
1931
1932 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1933 rcu_nocb_lock_irqsave(rdp, flags);
1934 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1935 __call_rcu_nocb_wake(rdp, true, flags);
1936}
1937
1938/*
1939 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1940 * or for grace periods to end.
1941 */
1942static void nocb_gp_wait(struct rcu_data *my_rdp)
1943{
1944 bool bypass = false;
1945 long bypass_ncbs;
1946 int __maybe_unused cpu = my_rdp->cpu;
1947 unsigned long cur_gp_seq;
1948 unsigned long flags;
1949 bool gotcbs;
1950 unsigned long j = jiffies;
1951 bool needwait_gp = false; // This prevents actual uninitialized use.
1952 bool needwake;
1953 bool needwake_gp;
1954 struct rcu_data *rdp;
1955 struct rcu_node *rnp;
1956 unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
1957
1958 /*
1959 * Each pass through the following loop checks for CBs and for the
1960 * nearest grace period (if any) to wait for next. The CB kthreads
1961 * and the global grace-period kthread are awakened if needed.
1962 */
1963 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1964 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
1965 rcu_nocb_lock_irqsave(rdp, flags);
1966 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1967 if (bypass_ncbs &&
1968 (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
1969 bypass_ncbs > 2 * qhimark)) {
1970 // Bypass full or old, so flush it.
1971 (void)rcu_nocb_try_flush_bypass(rdp, j);
1972 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1973 } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
1974 rcu_nocb_unlock_irqrestore(rdp, flags);
1975 continue; /* No callbacks here, try next. */
1976 }
1977 if (bypass_ncbs) {
1978 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1979 TPS("Bypass"));
1980 bypass = true;
1981 }
1982 rnp = rdp->mynode;
1983 if (bypass) { // Avoid race with first bypass CB.
1984 WRITE_ONCE(my_rdp->nocb_defer_wakeup,
1985 RCU_NOCB_WAKE_NOT);
1986 del_timer(&my_rdp->nocb_timer);
1987 }
1988 // Advance callbacks if helpful and low contention.
1989 needwake_gp = false;
1990 if (!rcu_segcblist_restempty(&rdp->cblist,
1991 RCU_NEXT_READY_TAIL) ||
1992 (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1993 rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
1994 raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
1995 needwake_gp = rcu_advance_cbs(rnp, rdp);
1996 raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
1997 }
1998 // Need to wait on some grace period?
1999 WARN_ON_ONCE(!rcu_segcblist_restempty(&rdp->cblist,
2000 RCU_NEXT_READY_TAIL));
2001 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
2002 if (!needwait_gp ||
2003 ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
2004 wait_gp_seq = cur_gp_seq;
2005 needwait_gp = true;
2006 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2007 TPS("NeedWaitGP"));
2008 }
2009 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2010 needwake = rdp->nocb_cb_sleep;
2011 WRITE_ONCE(rdp->nocb_cb_sleep, false);
2012 smp_mb(); /* CB invocation -after- GP end. */
2013 } else {
2014 needwake = false;
2015 }
2016 rcu_nocb_unlock_irqrestore(rdp, flags);
2017 if (needwake) {
2018 swake_up_one(&rdp->nocb_cb_wq);
2019 gotcbs = true;
2020 }
2021 if (needwake_gp)
2022 rcu_gp_kthread_wake();
2023 }
2024
2025 my_rdp->nocb_gp_bypass = bypass;
2026 my_rdp->nocb_gp_gp = needwait_gp;
2027 my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
2028 if (bypass && !rcu_nocb_poll) {
2029 // At least one child with non-empty ->nocb_bypass, so set
2030 // timer in order to avoid stranding its callbacks.
2031 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2032 mod_timer(&my_rdp->nocb_bypass_timer, j + 2);
2033 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2034 }
2035 if (rcu_nocb_poll) {
2036 /* Polling, so trace if first poll in the series. */
2037 if (gotcbs)
2038 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2039 schedule_timeout_interruptible(1);
2040 } else if (!needwait_gp) {
2041 /* Wait for callbacks to appear. */
2042 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2043 swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2044 !READ_ONCE(my_rdp->nocb_gp_sleep));
2045 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2046 } else {
2047 rnp = my_rdp->mynode;
2048 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2049 swait_event_interruptible_exclusive(
2050 rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2051 rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2052 !READ_ONCE(my_rdp->nocb_gp_sleep));
2053 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2054 }
2055 if (!rcu_nocb_poll) {
2056 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2057 if (bypass)
2058 del_timer(&my_rdp->nocb_bypass_timer);
2059 WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2060 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2061 }
2062 my_rdp->nocb_gp_seq = -1;
2063 WARN_ON(signal_pending(current));
2064}
2065
2066/*
2067 * No-CBs grace-period-wait kthread. There is one of these per group
2068 * of CPUs, but only once at least one CPU in that group has come online
2069 * at least once since boot. This kthread checks for newly posted
2070 * callbacks from any of the CPUs it is responsible for, waits for a
2071 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2072 * that then have callback-invocation work to do.
2073 */
2074static int rcu_nocb_gp_kthread(void *arg)
2075{
2076 struct rcu_data *rdp = arg;
2077
2078 for (;;) {
2079 WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2080 nocb_gp_wait(rdp);
2081 cond_resched_tasks_rcu_qs();
2082 }
2083 return 0;
2084}
2085
2086/*
2087 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2088 * then, if there are no more, wait for more to appear.
2089 */
2090static void nocb_cb_wait(struct rcu_data *rdp)
2091{
2092 unsigned long cur_gp_seq;
2093 unsigned long flags;
2094 bool needwake_gp = false;
2095 struct rcu_node *rnp = rdp->mynode;
2096
2097 local_irq_save(flags);
2098 rcu_momentary_dyntick_idle();
2099 local_irq_restore(flags);
2100 local_bh_disable();
2101 rcu_do_batch(rdp);
2102 local_bh_enable();
2103 lockdep_assert_irqs_enabled();
2104 rcu_nocb_lock_irqsave(rdp, flags);
2105 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2106 rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2107 raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2108 needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2109 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2110 }
2111 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2112 rcu_nocb_unlock_irqrestore(rdp, flags);
2113 if (needwake_gp)
2114 rcu_gp_kthread_wake();
2115 return;
2116 }
2117
2118 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2119 WRITE_ONCE(rdp->nocb_cb_sleep, true);
2120 rcu_nocb_unlock_irqrestore(rdp, flags);
2121 if (needwake_gp)
2122 rcu_gp_kthread_wake();
2123 swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2124 !READ_ONCE(rdp->nocb_cb_sleep));
2125 if (!smp_load_acquire(&rdp->nocb_cb_sleep)) { /* VVV */
2126 /* ^^^ Ensure CB invocation follows _sleep test. */
2127 return;
2128 }
2129 WARN_ON(signal_pending(current));
2130 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2131}
2132
2133/*
2134 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2135 * nocb_cb_wait() to do the dirty work.
2136 */
2137static int rcu_nocb_cb_kthread(void *arg)
2138{
2139 struct rcu_data *rdp = arg;
2140
2141 // Each pass through this loop does one callback batch, and,
2142 // if there are no more ready callbacks, waits for them.
2143 for (;;) {
2144 nocb_cb_wait(rdp);
2145 cond_resched_tasks_rcu_qs();
2146 }
2147 return 0;
2148}
2149
2150/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2151static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2152{
2153 return READ_ONCE(rdp->nocb_defer_wakeup);
2154}
2155
2156/* Do a deferred wakeup of rcu_nocb_kthread(). */
2157static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2158{
2159 unsigned long flags;
2160 int ndw;
2161
2162 rcu_nocb_lock_irqsave(rdp, flags);
2163 if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2164 rcu_nocb_unlock_irqrestore(rdp, flags);
2165 return;
2166 }
2167 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2168 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2169 wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2170 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2171}
2172
2173/* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2174static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2175{
2176 struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2177
2178 do_nocb_deferred_wakeup_common(rdp);
2179}
2180
2181/*
2182 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2183 * This means we do an inexact common-case check. Note that if
2184 * we miss, ->nocb_timer will eventually clean things up.
2185 */
2186static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2187{
2188 if (rcu_nocb_need_deferred_wakeup(rdp))
2189 do_nocb_deferred_wakeup_common(rdp);
2190}
2191
2192void __init rcu_init_nohz(void)
2193{
2194 int cpu;
2195 bool need_rcu_nocb_mask = false;
2196 struct rcu_data *rdp;
2197
2198#if defined(CONFIG_NO_HZ_FULL)
2199 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2200 need_rcu_nocb_mask = true;
2201#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2202
2203 if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2204 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2205 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2206 return;
2207 }
2208 }
2209 if (!cpumask_available(rcu_nocb_mask))
2210 return;
2211
2212#if defined(CONFIG_NO_HZ_FULL)
2213 if (tick_nohz_full_running)
2214 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2215#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2216
2217 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2218 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2219 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2220 rcu_nocb_mask);
2221 }
2222 if (cpumask_empty(rcu_nocb_mask))
2223 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2224 else
2225 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2226 cpumask_pr_args(rcu_nocb_mask));
2227 if (rcu_nocb_poll)
2228 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2229
2230 for_each_cpu(cpu, rcu_nocb_mask) {
2231 rdp = per_cpu_ptr(&rcu_data, cpu);
2232 if (rcu_segcblist_empty(&rdp->cblist))
2233 rcu_segcblist_init(&rdp->cblist);
2234 rcu_segcblist_offload(&rdp->cblist);
2235 }
2236 rcu_organize_nocb_kthreads();
2237}
2238
2239/* Initialize per-rcu_data variables for no-CBs CPUs. */
2240static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2241{
2242 init_swait_queue_head(&rdp->nocb_cb_wq);
2243 init_swait_queue_head(&rdp->nocb_gp_wq);
2244 raw_spin_lock_init(&rdp->nocb_lock);
2245 raw_spin_lock_init(&rdp->nocb_bypass_lock);
2246 raw_spin_lock_init(&rdp->nocb_gp_lock);
2247 timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2248 timer_setup(&rdp->nocb_bypass_timer, do_nocb_bypass_wakeup_timer, 0);
2249 rcu_cblist_init(&rdp->nocb_bypass);
2250}
2251
2252/*
2253 * If the specified CPU is a no-CBs CPU that does not already have its
2254 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2255 * for this CPU's group has not yet been created, spawn it as well.
2256 */
2257static void rcu_spawn_one_nocb_kthread(int cpu)
2258{
2259 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2260 struct rcu_data *rdp_gp;
2261 struct task_struct *t;
2262
2263 /*
2264 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2265 * then nothing to do.
2266 */
2267 if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2268 return;
2269
2270 /* If we didn't spawn the GP kthread first, reorganize! */
2271 rdp_gp = rdp->nocb_gp_rdp;
2272 if (!rdp_gp->nocb_gp_kthread) {
2273 t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2274 "rcuog/%d", rdp_gp->cpu);
2275 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2276 return;
2277 WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2278 }
2279
2280 /* Spawn the kthread for this CPU. */
2281 t = kthread_run(rcu_nocb_cb_kthread, rdp,
2282 "rcuo%c/%d", rcu_state.abbr, cpu);
2283 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2284 return;
2285 WRITE_ONCE(rdp->nocb_cb_kthread, t);
2286 WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2287}
2288
2289/*
2290 * If the specified CPU is a no-CBs CPU that does not already have its
2291 * rcuo kthread, spawn it.
2292 */
2293static void rcu_spawn_cpu_nocb_kthread(int cpu)
2294{
2295 if (rcu_scheduler_fully_active)
2296 rcu_spawn_one_nocb_kthread(cpu);
2297}
2298
2299/*
2300 * Once the scheduler is running, spawn rcuo kthreads for all online
2301 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2302 * non-boot CPUs come online -- if this changes, we will need to add
2303 * some mutual exclusion.
2304 */
2305static void __init rcu_spawn_nocb_kthreads(void)
2306{
2307 int cpu;
2308
2309 for_each_online_cpu(cpu)
2310 rcu_spawn_cpu_nocb_kthread(cpu);
2311}
2312
2313/* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2314static int rcu_nocb_gp_stride = -1;
2315module_param(rcu_nocb_gp_stride, int, 0444);
2316
2317/*
2318 * Initialize GP-CB relationships for all no-CBs CPU.
2319 */
2320static void __init rcu_organize_nocb_kthreads(void)
2321{
2322 int cpu;
2323 bool firsttime = true;
2324 int ls = rcu_nocb_gp_stride;
2325 int nl = 0; /* Next GP kthread. */
2326 struct rcu_data *rdp;
2327 struct rcu_data *rdp_gp = NULL; /* Suppress misguided gcc warn. */
2328 struct rcu_data *rdp_prev = NULL;
2329
2330 if (!cpumask_available(rcu_nocb_mask))
2331 return;
2332 if (ls == -1) {
2333 ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2334 rcu_nocb_gp_stride = ls;
2335 }
2336
2337 /*
2338 * Each pass through this loop sets up one rcu_data structure.
2339 * Should the corresponding CPU come online in the future, then
2340 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2341 */
2342 for_each_cpu(cpu, rcu_nocb_mask) {
2343 rdp = per_cpu_ptr(&rcu_data, cpu);
2344 if (rdp->cpu >= nl) {
2345 /* New GP kthread, set up for CBs & next GP. */
2346 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2347 rdp->nocb_gp_rdp = rdp;
2348 rdp_gp = rdp;
2349 if (!firsttime && dump_tree)
2350 pr_cont("\n");
2351 firsttime = false;
2352 pr_alert("%s: No-CB GP kthread CPU %d:", __func__, cpu);
2353 } else {
2354 /* Another CB kthread, link to previous GP kthread. */
2355 rdp->nocb_gp_rdp = rdp_gp;
2356 rdp_prev->nocb_next_cb_rdp = rdp;
2357 pr_alert(" %d", cpu);
2358 }
2359 rdp_prev = rdp;
2360 }
2361}
2362
2363/*
2364 * Bind the current task to the offloaded CPUs. If there are no offloaded
2365 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2366 */
2367void rcu_bind_current_to_nocb(void)
2368{
2369 if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2370 WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2371}
2372EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2373
2374/*
2375 * Dump out nocb grace-period kthread state for the specified rcu_data
2376 * structure.
2377 */
2378static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2379{
2380 struct rcu_node *rnp = rdp->mynode;
2381
2382 pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2383 rdp->cpu,
2384 "kK"[!!rdp->nocb_gp_kthread],
2385 "lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2386 "dD"[!!rdp->nocb_defer_wakeup],
2387 "tT"[timer_pending(&rdp->nocb_timer)],
2388 "bB"[timer_pending(&rdp->nocb_bypass_timer)],
2389 "sS"[!!rdp->nocb_gp_sleep],
2390 ".W"[swait_active(&rdp->nocb_gp_wq)],
2391 ".W"[swait_active(&rnp->nocb_gp_wq[0])],
2392 ".W"[swait_active(&rnp->nocb_gp_wq[1])],
2393 ".B"[!!rdp->nocb_gp_bypass],
2394 ".G"[!!rdp->nocb_gp_gp],
2395 (long)rdp->nocb_gp_seq,
2396 rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops));
2397}
2398
2399/* Dump out nocb kthread state for the specified rcu_data structure. */
2400static void show_rcu_nocb_state(struct rcu_data *rdp)
2401{
2402 struct rcu_segcblist *rsclp = &rdp->cblist;
2403 bool waslocked;
2404 bool wastimer;
2405 bool wassleep;
2406
2407 if (rdp->nocb_gp_rdp == rdp)
2408 show_rcu_nocb_gp_state(rdp);
2409
2410 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",
2411 rdp->cpu, rdp->nocb_gp_rdp->cpu,
2412 "kK"[!!rdp->nocb_cb_kthread],
2413 "bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2414 "cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2415 "lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2416 "sS"[!!rdp->nocb_cb_sleep],
2417 ".W"[swait_active(&rdp->nocb_cb_wq)],
2418 jiffies - rdp->nocb_bypass_first,
2419 jiffies - rdp->nocb_nobypass_last,
2420 rdp->nocb_nobypass_count,
2421 ".D"[rcu_segcblist_ready_cbs(rsclp)],
2422 ".W"[!rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL)],
2423 ".R"[!rcu_segcblist_restempty(rsclp, RCU_WAIT_TAIL)],
2424 ".N"[!rcu_segcblist_restempty(rsclp, RCU_NEXT_READY_TAIL)],
2425 ".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2426 rcu_segcblist_n_cbs(&rdp->cblist));
2427
2428 /* It is OK for GP kthreads to have GP state. */
2429 if (rdp->nocb_gp_rdp == rdp)
2430 return;
2431
2432 waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2433 wastimer = timer_pending(&rdp->nocb_timer);
2434 wassleep = swait_active(&rdp->nocb_gp_wq);
2435 if (!rdp->nocb_defer_wakeup && !rdp->nocb_gp_sleep &&
2436 !waslocked && !wastimer && !wassleep)
2437 return; /* Nothing untowards. */
2438
2439 pr_info(" !!! %c%c%c%c %c\n",
2440 "lL"[waslocked],
2441 "dD"[!!rdp->nocb_defer_wakeup],
2442 "tT"[wastimer],
2443 "sS"[!!rdp->nocb_gp_sleep],
2444 ".W"[wassleep]);
2445}
2446
2447#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2448
2449/* No ->nocb_lock to acquire. */
2450static void rcu_nocb_lock(struct rcu_data *rdp)
2451{
2452}
2453
2454/* No ->nocb_lock to release. */
2455static void rcu_nocb_unlock(struct rcu_data *rdp)
2456{
2457}
2458
2459/* No ->nocb_lock to release. */
2460static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2461 unsigned long flags)
2462{
2463 local_irq_restore(flags);
2464}
2465
2466/* Lockdep check that ->cblist may be safely accessed. */
2467static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2468{
2469 lockdep_assert_irqs_disabled();
2470}
2471
2472static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2473{
2474}
2475
2476static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2477{
2478 return NULL;
2479}
2480
2481static void rcu_init_one_nocb(struct rcu_node *rnp)
2482{
2483}
2484
2485static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2486 unsigned long j)
2487{
2488 return true;
2489}
2490
2491static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2492 bool *was_alldone, unsigned long flags)
2493{
2494 return false;
2495}
2496
2497static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2498 unsigned long flags)
2499{
2500 WARN_ON_ONCE(1); /* Should be dead code! */
2501}
2502
2503static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2504{
2505}
2506
2507static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2508{
2509 return false;
2510}
2511
2512static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2513{
2514}
2515
2516static void rcu_spawn_cpu_nocb_kthread(int cpu)
2517{
2518}
2519
2520static void __init rcu_spawn_nocb_kthreads(void)
2521{
2522}
2523
2524static void show_rcu_nocb_state(struct rcu_data *rdp)
2525{
2526}
2527
2528#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2529
2530/*
2531 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2532 * grace-period kthread will do force_quiescent_state() processing?
2533 * The idea is to avoid waking up RCU core processing on such a
2534 * CPU unless the grace period has extended for too long.
2535 *
2536 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2537 * CONFIG_RCU_NOCB_CPU CPUs.
2538 */
2539static bool rcu_nohz_full_cpu(void)
2540{
2541#ifdef CONFIG_NO_HZ_FULL
2542 if (tick_nohz_full_cpu(smp_processor_id()) &&
2543 (!rcu_gp_in_progress() ||
2544 ULONG_CMP_LT(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2545 return true;
2546#endif /* #ifdef CONFIG_NO_HZ_FULL */
2547 return false;
2548}
2549
2550/*
2551 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2552 */
2553static void rcu_bind_gp_kthread(void)
2554{
2555 if (!tick_nohz_full_enabled())
2556 return;
2557 housekeeping_affine(current, HK_FLAG_RCU);
2558}
2559
2560/* Record the current task on dyntick-idle entry. */
2561static void rcu_dynticks_task_enter(void)
2562{
2563#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2564 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2565#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2566}
2567
2568/* Record no current task on dyntick-idle exit. */
2569static void rcu_dynticks_task_exit(void)
2570{
2571#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2572 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2573#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2574}
1/*
2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
19 *
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
22 *
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
25 */
26
27#include <linux/delay.h>
28#include <linux/gfp.h>
29#include <linux/oom.h>
30#include <linux/smpboot.h>
31#include "../time/tick-internal.h"
32
33#ifdef CONFIG_RCU_BOOST
34
35#include "../locking/rtmutex_common.h"
36
37/*
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
40 */
41static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
42DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
43DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
44DEFINE_PER_CPU(char, rcu_cpu_has_work);
45
46#else /* #ifdef CONFIG_RCU_BOOST */
47
48/*
49 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
50 * all uses are in dead code. Provide a definition to keep the compiler
51 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
52 * This probably needs to be excluded from -rt builds.
53 */
54#define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
55
56#endif /* #else #ifdef CONFIG_RCU_BOOST */
57
58#ifdef CONFIG_RCU_NOCB_CPU
59static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
60static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
61static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
62#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
63
64/*
65 * Check the RCU kernel configuration parameters and print informative
66 * messages about anything out of the ordinary.
67 */
68static void __init rcu_bootup_announce_oddness(void)
69{
70 if (IS_ENABLED(CONFIG_RCU_TRACE))
71 pr_info("\tRCU debugfs-based tracing is enabled.\n");
72 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
73 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
74 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
75 RCU_FANOUT);
76 if (rcu_fanout_exact)
77 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
78 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
79 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
80 if (IS_ENABLED(CONFIG_PROVE_RCU))
81 pr_info("\tRCU lockdep checking is enabled.\n");
82 if (RCU_NUM_LVLS >= 4)
83 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
84 if (RCU_FANOUT_LEAF != 16)
85 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
86 RCU_FANOUT_LEAF);
87 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
88 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
89 if (nr_cpu_ids != NR_CPUS)
90 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
91 if (IS_ENABLED(CONFIG_RCU_BOOST))
92 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
93}
94
95#ifdef CONFIG_PREEMPT_RCU
96
97RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
98static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
99static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
100
101static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
102 bool wake);
103
104/*
105 * Tell them what RCU they are running.
106 */
107static void __init rcu_bootup_announce(void)
108{
109 pr_info("Preemptible hierarchical RCU implementation.\n");
110 rcu_bootup_announce_oddness();
111}
112
113/* Flags for rcu_preempt_ctxt_queue() decision table. */
114#define RCU_GP_TASKS 0x8
115#define RCU_EXP_TASKS 0x4
116#define RCU_GP_BLKD 0x2
117#define RCU_EXP_BLKD 0x1
118
119/*
120 * Queues a task preempted within an RCU-preempt read-side critical
121 * section into the appropriate location within the ->blkd_tasks list,
122 * depending on the states of any ongoing normal and expedited grace
123 * periods. The ->gp_tasks pointer indicates which element the normal
124 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
125 * indicates which element the expedited grace period is waiting on (again,
126 * NULL if none). If a grace period is waiting on a given element in the
127 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
128 * adding a task to the tail of the list blocks any grace period that is
129 * already waiting on one of the elements. In contrast, adding a task
130 * to the head of the list won't block any grace period that is already
131 * waiting on one of the elements.
132 *
133 * This queuing is imprecise, and can sometimes make an ongoing grace
134 * period wait for a task that is not strictly speaking blocking it.
135 * Given the choice, we needlessly block a normal grace period rather than
136 * blocking an expedited grace period.
137 *
138 * Note that an endless sequence of expedited grace periods still cannot
139 * indefinitely postpone a normal grace period. Eventually, all of the
140 * fixed number of preempted tasks blocking the normal grace period that are
141 * not also blocking the expedited grace period will resume and complete
142 * their RCU read-side critical sections. At that point, the ->gp_tasks
143 * pointer will equal the ->exp_tasks pointer, at which point the end of
144 * the corresponding expedited grace period will also be the end of the
145 * normal grace period.
146 */
147static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
148 __releases(rnp->lock) /* But leaves rrupts disabled. */
149{
150 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
151 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
152 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
153 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
154 struct task_struct *t = current;
155
156 /*
157 * Decide where to queue the newly blocked task. In theory,
158 * this could be an if-statement. In practice, when I tried
159 * that, it was quite messy.
160 */
161 switch (blkd_state) {
162 case 0:
163 case RCU_EXP_TASKS:
164 case RCU_EXP_TASKS + RCU_GP_BLKD:
165 case RCU_GP_TASKS:
166 case RCU_GP_TASKS + RCU_EXP_TASKS:
167
168 /*
169 * Blocking neither GP, or first task blocking the normal
170 * GP but not blocking the already-waiting expedited GP.
171 * Queue at the head of the list to avoid unnecessarily
172 * blocking the already-waiting GPs.
173 */
174 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
175 break;
176
177 case RCU_EXP_BLKD:
178 case RCU_GP_BLKD:
179 case RCU_GP_BLKD + RCU_EXP_BLKD:
180 case RCU_GP_TASKS + RCU_EXP_BLKD:
181 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
182 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
183
184 /*
185 * First task arriving that blocks either GP, or first task
186 * arriving that blocks the expedited GP (with the normal
187 * GP already waiting), or a task arriving that blocks
188 * both GPs with both GPs already waiting. Queue at the
189 * tail of the list to avoid any GP waiting on any of the
190 * already queued tasks that are not blocking it.
191 */
192 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
193 break;
194
195 case RCU_EXP_TASKS + RCU_EXP_BLKD:
196 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
197 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
198
199 /*
200 * Second or subsequent task blocking the expedited GP.
201 * The task either does not block the normal GP, or is the
202 * first task blocking the normal GP. Queue just after
203 * the first task blocking the expedited GP.
204 */
205 list_add(&t->rcu_node_entry, rnp->exp_tasks);
206 break;
207
208 case RCU_GP_TASKS + RCU_GP_BLKD:
209 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
210
211 /*
212 * Second or subsequent task blocking the normal GP.
213 * The task does not block the expedited GP. Queue just
214 * after the first task blocking the normal GP.
215 */
216 list_add(&t->rcu_node_entry, rnp->gp_tasks);
217 break;
218
219 default:
220
221 /* Yet another exercise in excessive paranoia. */
222 WARN_ON_ONCE(1);
223 break;
224 }
225
226 /*
227 * We have now queued the task. If it was the first one to
228 * block either grace period, update the ->gp_tasks and/or
229 * ->exp_tasks pointers, respectively, to reference the newly
230 * blocked tasks.
231 */
232 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD))
233 rnp->gp_tasks = &t->rcu_node_entry;
234 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
235 rnp->exp_tasks = &t->rcu_node_entry;
236 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
237
238 /*
239 * Report the quiescent state for the expedited GP. This expedited
240 * GP should not be able to end until we report, so there should be
241 * no need to check for a subsequent expedited GP. (Though we are
242 * still in a quiescent state in any case.)
243 */
244 if (blkd_state & RCU_EXP_BLKD &&
245 t->rcu_read_unlock_special.b.exp_need_qs) {
246 t->rcu_read_unlock_special.b.exp_need_qs = false;
247 rcu_report_exp_rdp(rdp->rsp, rdp, true);
248 } else {
249 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
250 }
251}
252
253/*
254 * Record a preemptible-RCU quiescent state for the specified CPU. Note
255 * that this just means that the task currently running on the CPU is
256 * not in a quiescent state. There might be any number of tasks blocked
257 * while in an RCU read-side critical section.
258 *
259 * As with the other rcu_*_qs() functions, callers to this function
260 * must disable preemption.
261 */
262static void rcu_preempt_qs(void)
263{
264 if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
265 trace_rcu_grace_period(TPS("rcu_preempt"),
266 __this_cpu_read(rcu_data_p->gpnum),
267 TPS("cpuqs"));
268 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
269 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
270 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 */
287static void rcu_preempt_note_context_switch(void)
288{
289 struct task_struct *t = current;
290 struct rcu_data *rdp;
291 struct rcu_node *rnp;
292
293 if (t->rcu_read_lock_nesting > 0 &&
294 !t->rcu_read_unlock_special.b.blocked) {
295
296 /* Possibly blocking in an RCU read-side critical section. */
297 rdp = this_cpu_ptr(rcu_state_p->rda);
298 rnp = rdp->mynode;
299 raw_spin_lock_rcu_node(rnp);
300 t->rcu_read_unlock_special.b.blocked = true;
301 t->rcu_blocked_node = rnp;
302
303 /*
304 * Verify the CPU's sanity, trace the preemption, and
305 * then queue the task as required based on the states
306 * of any ongoing and expedited grace periods.
307 */
308 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
309 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
310 trace_rcu_preempt_task(rdp->rsp->name,
311 t->pid,
312 (rnp->qsmask & rdp->grpmask)
313 ? rnp->gpnum
314 : rnp->gpnum + 1);
315 rcu_preempt_ctxt_queue(rnp, rdp);
316 } else if (t->rcu_read_lock_nesting < 0 &&
317 t->rcu_read_unlock_special.s) {
318
319 /*
320 * Complete exit from RCU read-side critical section on
321 * behalf of preempted instance of __rcu_read_unlock().
322 */
323 rcu_read_unlock_special(t);
324 }
325
326 /*
327 * Either we were not in an RCU read-side critical section to
328 * begin with, or we have now recorded that critical section
329 * globally. Either way, we can now note a quiescent state
330 * for this CPU. Again, if we were in an RCU read-side critical
331 * section, and if that critical section was blocking the current
332 * grace period, then the fact that the task has been enqueued
333 * means that we continue to block the current grace period.
334 */
335 rcu_preempt_qs();
336}
337
338/*
339 * Check for preempted RCU readers blocking the current grace period
340 * for the specified rcu_node structure. If the caller needs a reliable
341 * answer, it must hold the rcu_node's ->lock.
342 */
343static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
344{
345 return rnp->gp_tasks != NULL;
346}
347
348/*
349 * Advance a ->blkd_tasks-list pointer to the next entry, instead
350 * returning NULL if at the end of the list.
351 */
352static struct list_head *rcu_next_node_entry(struct task_struct *t,
353 struct rcu_node *rnp)
354{
355 struct list_head *np;
356
357 np = t->rcu_node_entry.next;
358 if (np == &rnp->blkd_tasks)
359 np = NULL;
360 return np;
361}
362
363/*
364 * Return true if the specified rcu_node structure has tasks that were
365 * preempted within an RCU read-side critical section.
366 */
367static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
368{
369 return !list_empty(&rnp->blkd_tasks);
370}
371
372/*
373 * Handle special cases during rcu_read_unlock(), such as needing to
374 * notify RCU core processing or task having blocked during the RCU
375 * read-side critical section.
376 */
377void rcu_read_unlock_special(struct task_struct *t)
378{
379 bool empty_exp;
380 bool empty_norm;
381 bool empty_exp_now;
382 unsigned long flags;
383 struct list_head *np;
384 bool drop_boost_mutex = false;
385 struct rcu_data *rdp;
386 struct rcu_node *rnp;
387 union rcu_special special;
388
389 /* NMI handlers cannot block and cannot safely manipulate state. */
390 if (in_nmi())
391 return;
392
393 local_irq_save(flags);
394
395 /*
396 * If RCU core is waiting for this CPU to exit its critical section,
397 * report the fact that it has exited. Because irqs are disabled,
398 * t->rcu_read_unlock_special cannot change.
399 */
400 special = t->rcu_read_unlock_special;
401 if (special.b.need_qs) {
402 rcu_preempt_qs();
403 t->rcu_read_unlock_special.b.need_qs = false;
404 if (!t->rcu_read_unlock_special.s) {
405 local_irq_restore(flags);
406 return;
407 }
408 }
409
410 /*
411 * Respond to a request for an expedited grace period, but only if
412 * we were not preempted, meaning that we were running on the same
413 * CPU throughout. If we were preempted, the exp_need_qs flag
414 * would have been cleared at the time of the first preemption,
415 * and the quiescent state would be reported when we were dequeued.
416 */
417 if (special.b.exp_need_qs) {
418 WARN_ON_ONCE(special.b.blocked);
419 t->rcu_read_unlock_special.b.exp_need_qs = false;
420 rdp = this_cpu_ptr(rcu_state_p->rda);
421 rcu_report_exp_rdp(rcu_state_p, rdp, true);
422 if (!t->rcu_read_unlock_special.s) {
423 local_irq_restore(flags);
424 return;
425 }
426 }
427
428 /* Hardware IRQ handlers cannot block, complain if they get here. */
429 if (in_irq() || in_serving_softirq()) {
430 lockdep_rcu_suspicious(__FILE__, __LINE__,
431 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
432 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
433 t->rcu_read_unlock_special.s,
434 t->rcu_read_unlock_special.b.blocked,
435 t->rcu_read_unlock_special.b.exp_need_qs,
436 t->rcu_read_unlock_special.b.need_qs);
437 local_irq_restore(flags);
438 return;
439 }
440
441 /* Clean up if blocked during RCU read-side critical section. */
442 if (special.b.blocked) {
443 t->rcu_read_unlock_special.b.blocked = false;
444
445 /*
446 * Remove this task from the list it blocked on. The task
447 * now remains queued on the rcu_node corresponding to the
448 * CPU it first blocked on, so there is no longer any need
449 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
450 */
451 rnp = t->rcu_blocked_node;
452 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
453 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
454 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
455 empty_exp = sync_rcu_preempt_exp_done(rnp);
456 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
457 np = rcu_next_node_entry(t, rnp);
458 list_del_init(&t->rcu_node_entry);
459 t->rcu_blocked_node = NULL;
460 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
461 rnp->gpnum, t->pid);
462 if (&t->rcu_node_entry == rnp->gp_tasks)
463 rnp->gp_tasks = np;
464 if (&t->rcu_node_entry == rnp->exp_tasks)
465 rnp->exp_tasks = np;
466 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
467 if (&t->rcu_node_entry == rnp->boost_tasks)
468 rnp->boost_tasks = np;
469 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
470 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
471 }
472
473 /*
474 * If this was the last task on the current list, and if
475 * we aren't waiting on any CPUs, report the quiescent state.
476 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
477 * so we must take a snapshot of the expedited state.
478 */
479 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
480 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
481 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
482 rnp->gpnum,
483 0, rnp->qsmask,
484 rnp->level,
485 rnp->grplo,
486 rnp->grphi,
487 !!rnp->gp_tasks);
488 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
489 } else {
490 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
491 }
492
493 /* Unboost if we were boosted. */
494 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
495 rt_mutex_unlock(&rnp->boost_mtx);
496
497 /*
498 * If this was the last task on the expedited lists,
499 * then we need to report up the rcu_node hierarchy.
500 */
501 if (!empty_exp && empty_exp_now)
502 rcu_report_exp_rnp(rcu_state_p, rnp, true);
503 } else {
504 local_irq_restore(flags);
505 }
506}
507
508/*
509 * Dump detailed information for all tasks blocking the current RCU
510 * grace period on the specified rcu_node structure.
511 */
512static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
513{
514 unsigned long flags;
515 struct task_struct *t;
516
517 raw_spin_lock_irqsave_rcu_node(rnp, flags);
518 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
519 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
520 return;
521 }
522 t = list_entry(rnp->gp_tasks->prev,
523 struct task_struct, rcu_node_entry);
524 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
525 sched_show_task(t);
526 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
527}
528
529/*
530 * Dump detailed information for all tasks blocking the current RCU
531 * grace period.
532 */
533static void rcu_print_detail_task_stall(struct rcu_state *rsp)
534{
535 struct rcu_node *rnp = rcu_get_root(rsp);
536
537 rcu_print_detail_task_stall_rnp(rnp);
538 rcu_for_each_leaf_node(rsp, rnp)
539 rcu_print_detail_task_stall_rnp(rnp);
540}
541
542static void rcu_print_task_stall_begin(struct rcu_node *rnp)
543{
544 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
545 rnp->level, rnp->grplo, rnp->grphi);
546}
547
548static void rcu_print_task_stall_end(void)
549{
550 pr_cont("\n");
551}
552
553/*
554 * Scan the current list of tasks blocked within RCU read-side critical
555 * sections, printing out the tid of each.
556 */
557static int rcu_print_task_stall(struct rcu_node *rnp)
558{
559 struct task_struct *t;
560 int ndetected = 0;
561
562 if (!rcu_preempt_blocked_readers_cgp(rnp))
563 return 0;
564 rcu_print_task_stall_begin(rnp);
565 t = list_entry(rnp->gp_tasks->prev,
566 struct task_struct, rcu_node_entry);
567 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
568 pr_cont(" P%d", t->pid);
569 ndetected++;
570 }
571 rcu_print_task_stall_end();
572 return ndetected;
573}
574
575/*
576 * Scan the current list of tasks blocked within RCU read-side critical
577 * sections, printing out the tid of each that is blocking the current
578 * expedited grace period.
579 */
580static int rcu_print_task_exp_stall(struct rcu_node *rnp)
581{
582 struct task_struct *t;
583 int ndetected = 0;
584
585 if (!rnp->exp_tasks)
586 return 0;
587 t = list_entry(rnp->exp_tasks->prev,
588 struct task_struct, rcu_node_entry);
589 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
590 pr_cont(" P%d", t->pid);
591 ndetected++;
592 }
593 return ndetected;
594}
595
596/*
597 * Check that the list of blocked tasks for the newly completed grace
598 * period is in fact empty. It is a serious bug to complete a grace
599 * period that still has RCU readers blocked! This function must be
600 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
601 * must be held by the caller.
602 *
603 * Also, if there are blocked tasks on the list, they automatically
604 * block the newly created grace period, so set up ->gp_tasks accordingly.
605 */
606static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
607{
608 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
609 if (rcu_preempt_has_tasks(rnp))
610 rnp->gp_tasks = rnp->blkd_tasks.next;
611 WARN_ON_ONCE(rnp->qsmask);
612}
613
614/*
615 * Check for a quiescent state from the current CPU. When a task blocks,
616 * the task is recorded in the corresponding CPU's rcu_node structure,
617 * which is checked elsewhere.
618 *
619 * Caller must disable hard irqs.
620 */
621static void rcu_preempt_check_callbacks(void)
622{
623 struct task_struct *t = current;
624
625 if (t->rcu_read_lock_nesting == 0) {
626 rcu_preempt_qs();
627 return;
628 }
629 if (t->rcu_read_lock_nesting > 0 &&
630 __this_cpu_read(rcu_data_p->core_needs_qs) &&
631 __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm))
632 t->rcu_read_unlock_special.b.need_qs = true;
633}
634
635#ifdef CONFIG_RCU_BOOST
636
637static void rcu_preempt_do_callbacks(void)
638{
639 rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p));
640}
641
642#endif /* #ifdef CONFIG_RCU_BOOST */
643
644/*
645 * Queue a preemptible-RCU callback for invocation after a grace period.
646 */
647void call_rcu(struct rcu_head *head, rcu_callback_t func)
648{
649 __call_rcu(head, func, rcu_state_p, -1, 0);
650}
651EXPORT_SYMBOL_GPL(call_rcu);
652
653/**
654 * synchronize_rcu - wait until a grace period has elapsed.
655 *
656 * Control will return to the caller some time after a full grace
657 * period has elapsed, in other words after all currently executing RCU
658 * read-side critical sections have completed. Note, however, that
659 * upon return from synchronize_rcu(), the caller might well be executing
660 * concurrently with new RCU read-side critical sections that began while
661 * synchronize_rcu() was waiting. RCU read-side critical sections are
662 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
663 *
664 * See the description of synchronize_sched() for more detailed information
665 * on memory ordering guarantees.
666 */
667void synchronize_rcu(void)
668{
669 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
670 lock_is_held(&rcu_lock_map) ||
671 lock_is_held(&rcu_sched_lock_map),
672 "Illegal synchronize_rcu() in RCU read-side critical section");
673 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
674 return;
675 if (rcu_gp_is_expedited())
676 synchronize_rcu_expedited();
677 else
678 wait_rcu_gp(call_rcu);
679}
680EXPORT_SYMBOL_GPL(synchronize_rcu);
681
682/**
683 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
684 *
685 * Note that this primitive does not necessarily wait for an RCU grace period
686 * to complete. For example, if there are no RCU callbacks queued anywhere
687 * in the system, then rcu_barrier() is within its rights to return
688 * immediately, without waiting for anything, much less an RCU grace period.
689 */
690void rcu_barrier(void)
691{
692 _rcu_barrier(rcu_state_p);
693}
694EXPORT_SYMBOL_GPL(rcu_barrier);
695
696/*
697 * Initialize preemptible RCU's state structures.
698 */
699static void __init __rcu_init_preempt(void)
700{
701 rcu_init_one(rcu_state_p);
702}
703
704/*
705 * Check for a task exiting while in a preemptible-RCU read-side
706 * critical section, clean up if so. No need to issue warnings,
707 * as debug_check_no_locks_held() already does this if lockdep
708 * is enabled.
709 */
710void exit_rcu(void)
711{
712 struct task_struct *t = current;
713
714 if (likely(list_empty(¤t->rcu_node_entry)))
715 return;
716 t->rcu_read_lock_nesting = 1;
717 barrier();
718 t->rcu_read_unlock_special.b.blocked = true;
719 __rcu_read_unlock();
720}
721
722#else /* #ifdef CONFIG_PREEMPT_RCU */
723
724static struct rcu_state *const rcu_state_p = &rcu_sched_state;
725
726/*
727 * Tell them what RCU they are running.
728 */
729static void __init rcu_bootup_announce(void)
730{
731 pr_info("Hierarchical RCU implementation.\n");
732 rcu_bootup_announce_oddness();
733}
734
735/*
736 * Because preemptible RCU does not exist, we never have to check for
737 * CPUs being in quiescent states.
738 */
739static void rcu_preempt_note_context_switch(void)
740{
741}
742
743/*
744 * Because preemptible RCU does not exist, there are never any preempted
745 * RCU readers.
746 */
747static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
748{
749 return 0;
750}
751
752/*
753 * Because there is no preemptible RCU, there can be no readers blocked.
754 */
755static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
756{
757 return false;
758}
759
760/*
761 * Because preemptible RCU does not exist, we never have to check for
762 * tasks blocked within RCU read-side critical sections.
763 */
764static void rcu_print_detail_task_stall(struct rcu_state *rsp)
765{
766}
767
768/*
769 * Because preemptible RCU does not exist, we never have to check for
770 * tasks blocked within RCU read-side critical sections.
771 */
772static int rcu_print_task_stall(struct rcu_node *rnp)
773{
774 return 0;
775}
776
777/*
778 * Because preemptible RCU does not exist, we never have to check for
779 * tasks blocked within RCU read-side critical sections that are
780 * blocking the current expedited grace period.
781 */
782static int rcu_print_task_exp_stall(struct rcu_node *rnp)
783{
784 return 0;
785}
786
787/*
788 * Because there is no preemptible RCU, there can be no readers blocked,
789 * so there is no need to check for blocked tasks. So check only for
790 * bogus qsmask values.
791 */
792static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
793{
794 WARN_ON_ONCE(rnp->qsmask);
795}
796
797/*
798 * Because preemptible RCU does not exist, it never has any callbacks
799 * to check.
800 */
801static void rcu_preempt_check_callbacks(void)
802{
803}
804
805/*
806 * Because preemptible RCU does not exist, rcu_barrier() is just
807 * another name for rcu_barrier_sched().
808 */
809void rcu_barrier(void)
810{
811 rcu_barrier_sched();
812}
813EXPORT_SYMBOL_GPL(rcu_barrier);
814
815/*
816 * Because preemptible RCU does not exist, it need not be initialized.
817 */
818static void __init __rcu_init_preempt(void)
819{
820}
821
822/*
823 * Because preemptible RCU does not exist, tasks cannot possibly exit
824 * while in preemptible RCU read-side critical sections.
825 */
826void exit_rcu(void)
827{
828}
829
830#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
831
832#ifdef CONFIG_RCU_BOOST
833
834#include "../locking/rtmutex_common.h"
835
836#ifdef CONFIG_RCU_TRACE
837
838static void rcu_initiate_boost_trace(struct rcu_node *rnp)
839{
840 if (!rcu_preempt_has_tasks(rnp))
841 rnp->n_balk_blkd_tasks++;
842 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
843 rnp->n_balk_exp_gp_tasks++;
844 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
845 rnp->n_balk_boost_tasks++;
846 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
847 rnp->n_balk_notblocked++;
848 else if (rnp->gp_tasks != NULL &&
849 ULONG_CMP_LT(jiffies, rnp->boost_time))
850 rnp->n_balk_notyet++;
851 else
852 rnp->n_balk_nos++;
853}
854
855#else /* #ifdef CONFIG_RCU_TRACE */
856
857static void rcu_initiate_boost_trace(struct rcu_node *rnp)
858{
859}
860
861#endif /* #else #ifdef CONFIG_RCU_TRACE */
862
863static void rcu_wake_cond(struct task_struct *t, int status)
864{
865 /*
866 * If the thread is yielding, only wake it when this
867 * is invoked from idle
868 */
869 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
870 wake_up_process(t);
871}
872
873/*
874 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
875 * or ->boost_tasks, advancing the pointer to the next task in the
876 * ->blkd_tasks list.
877 *
878 * Note that irqs must be enabled: boosting the task can block.
879 * Returns 1 if there are more tasks needing to be boosted.
880 */
881static int rcu_boost(struct rcu_node *rnp)
882{
883 unsigned long flags;
884 struct task_struct *t;
885 struct list_head *tb;
886
887 if (READ_ONCE(rnp->exp_tasks) == NULL &&
888 READ_ONCE(rnp->boost_tasks) == NULL)
889 return 0; /* Nothing left to boost. */
890
891 raw_spin_lock_irqsave_rcu_node(rnp, flags);
892
893 /*
894 * Recheck under the lock: all tasks in need of boosting
895 * might exit their RCU read-side critical sections on their own.
896 */
897 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
898 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
899 return 0;
900 }
901
902 /*
903 * Preferentially boost tasks blocking expedited grace periods.
904 * This cannot starve the normal grace periods because a second
905 * expedited grace period must boost all blocked tasks, including
906 * those blocking the pre-existing normal grace period.
907 */
908 if (rnp->exp_tasks != NULL) {
909 tb = rnp->exp_tasks;
910 rnp->n_exp_boosts++;
911 } else {
912 tb = rnp->boost_tasks;
913 rnp->n_normal_boosts++;
914 }
915 rnp->n_tasks_boosted++;
916
917 /*
918 * We boost task t by manufacturing an rt_mutex that appears to
919 * be held by task t. We leave a pointer to that rt_mutex where
920 * task t can find it, and task t will release the mutex when it
921 * exits its outermost RCU read-side critical section. Then
922 * simply acquiring this artificial rt_mutex will boost task
923 * t's priority. (Thanks to tglx for suggesting this approach!)
924 *
925 * Note that task t must acquire rnp->lock to remove itself from
926 * the ->blkd_tasks list, which it will do from exit() if from
927 * nowhere else. We therefore are guaranteed that task t will
928 * stay around at least until we drop rnp->lock. Note that
929 * rnp->lock also resolves races between our priority boosting
930 * and task t's exiting its outermost RCU read-side critical
931 * section.
932 */
933 t = container_of(tb, struct task_struct, rcu_node_entry);
934 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
935 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
936 /* Lock only for side effect: boosts task t's priority. */
937 rt_mutex_lock(&rnp->boost_mtx);
938 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
939
940 return READ_ONCE(rnp->exp_tasks) != NULL ||
941 READ_ONCE(rnp->boost_tasks) != NULL;
942}
943
944/*
945 * Priority-boosting kthread, one per leaf rcu_node.
946 */
947static int rcu_boost_kthread(void *arg)
948{
949 struct rcu_node *rnp = (struct rcu_node *)arg;
950 int spincnt = 0;
951 int more2boost;
952
953 trace_rcu_utilization(TPS("Start boost kthread@init"));
954 for (;;) {
955 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
956 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
957 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
958 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
959 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
960 more2boost = rcu_boost(rnp);
961 if (more2boost)
962 spincnt++;
963 else
964 spincnt = 0;
965 if (spincnt > 10) {
966 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
967 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
968 schedule_timeout_interruptible(2);
969 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
970 spincnt = 0;
971 }
972 }
973 /* NOTREACHED */
974 trace_rcu_utilization(TPS("End boost kthread@notreached"));
975 return 0;
976}
977
978/*
979 * Check to see if it is time to start boosting RCU readers that are
980 * blocking the current grace period, and, if so, tell the per-rcu_node
981 * kthread to start boosting them. If there is an expedited grace
982 * period in progress, it is always time to boost.
983 *
984 * The caller must hold rnp->lock, which this function releases.
985 * The ->boost_kthread_task is immortal, so we don't need to worry
986 * about it going away.
987 */
988static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
989 __releases(rnp->lock)
990{
991 struct task_struct *t;
992
993 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
994 rnp->n_balk_exp_gp_tasks++;
995 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
996 return;
997 }
998 if (rnp->exp_tasks != NULL ||
999 (rnp->gp_tasks != NULL &&
1000 rnp->boost_tasks == NULL &&
1001 rnp->qsmask == 0 &&
1002 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1003 if (rnp->exp_tasks == NULL)
1004 rnp->boost_tasks = rnp->gp_tasks;
1005 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1006 t = rnp->boost_kthread_task;
1007 if (t)
1008 rcu_wake_cond(t, rnp->boost_kthread_status);
1009 } else {
1010 rcu_initiate_boost_trace(rnp);
1011 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1012 }
1013}
1014
1015/*
1016 * Wake up the per-CPU kthread to invoke RCU callbacks.
1017 */
1018static void invoke_rcu_callbacks_kthread(void)
1019{
1020 unsigned long flags;
1021
1022 local_irq_save(flags);
1023 __this_cpu_write(rcu_cpu_has_work, 1);
1024 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1025 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1026 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1027 __this_cpu_read(rcu_cpu_kthread_status));
1028 }
1029 local_irq_restore(flags);
1030}
1031
1032/*
1033 * Is the current CPU running the RCU-callbacks kthread?
1034 * Caller must have preemption disabled.
1035 */
1036static bool rcu_is_callbacks_kthread(void)
1037{
1038 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1039}
1040
1041#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1042
1043/*
1044 * Do priority-boost accounting for the start of a new grace period.
1045 */
1046static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1047{
1048 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1049}
1050
1051/*
1052 * Create an RCU-boost kthread for the specified node if one does not
1053 * already exist. We only create this kthread for preemptible RCU.
1054 * Returns zero if all is well, a negated errno otherwise.
1055 */
1056static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1057 struct rcu_node *rnp)
1058{
1059 int rnp_index = rnp - &rsp->node[0];
1060 unsigned long flags;
1061 struct sched_param sp;
1062 struct task_struct *t;
1063
1064 if (rcu_state_p != rsp)
1065 return 0;
1066
1067 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1068 return 0;
1069
1070 rsp->boost = 1;
1071 if (rnp->boost_kthread_task != NULL)
1072 return 0;
1073 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1074 "rcub/%d", rnp_index);
1075 if (IS_ERR(t))
1076 return PTR_ERR(t);
1077 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1078 rnp->boost_kthread_task = t;
1079 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1080 sp.sched_priority = kthread_prio;
1081 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1082 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1083 return 0;
1084}
1085
1086static void rcu_kthread_do_work(void)
1087{
1088 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1089 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1090 rcu_preempt_do_callbacks();
1091}
1092
1093static void rcu_cpu_kthread_setup(unsigned int cpu)
1094{
1095 struct sched_param sp;
1096
1097 sp.sched_priority = kthread_prio;
1098 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1099}
1100
1101static void rcu_cpu_kthread_park(unsigned int cpu)
1102{
1103 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1104}
1105
1106static int rcu_cpu_kthread_should_run(unsigned int cpu)
1107{
1108 return __this_cpu_read(rcu_cpu_has_work);
1109}
1110
1111/*
1112 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1113 * RCU softirq used in flavors and configurations of RCU that do not
1114 * support RCU priority boosting.
1115 */
1116static void rcu_cpu_kthread(unsigned int cpu)
1117{
1118 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1119 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1120 int spincnt;
1121
1122 for (spincnt = 0; spincnt < 10; spincnt++) {
1123 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1124 local_bh_disable();
1125 *statusp = RCU_KTHREAD_RUNNING;
1126 this_cpu_inc(rcu_cpu_kthread_loops);
1127 local_irq_disable();
1128 work = *workp;
1129 *workp = 0;
1130 local_irq_enable();
1131 if (work)
1132 rcu_kthread_do_work();
1133 local_bh_enable();
1134 if (*workp == 0) {
1135 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1136 *statusp = RCU_KTHREAD_WAITING;
1137 return;
1138 }
1139 }
1140 *statusp = RCU_KTHREAD_YIELDING;
1141 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1142 schedule_timeout_interruptible(2);
1143 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1144 *statusp = RCU_KTHREAD_WAITING;
1145}
1146
1147/*
1148 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1149 * served by the rcu_node in question. The CPU hotplug lock is still
1150 * held, so the value of rnp->qsmaskinit will be stable.
1151 *
1152 * We don't include outgoingcpu in the affinity set, use -1 if there is
1153 * no outgoing CPU. If there are no CPUs left in the affinity set,
1154 * this function allows the kthread to execute on any CPU.
1155 */
1156static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1157{
1158 struct task_struct *t = rnp->boost_kthread_task;
1159 unsigned long mask = rcu_rnp_online_cpus(rnp);
1160 cpumask_var_t cm;
1161 int cpu;
1162
1163 if (!t)
1164 return;
1165 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1166 return;
1167 for_each_leaf_node_possible_cpu(rnp, cpu)
1168 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1169 cpu != outgoingcpu)
1170 cpumask_set_cpu(cpu, cm);
1171 if (cpumask_weight(cm) == 0)
1172 cpumask_setall(cm);
1173 set_cpus_allowed_ptr(t, cm);
1174 free_cpumask_var(cm);
1175}
1176
1177static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1178 .store = &rcu_cpu_kthread_task,
1179 .thread_should_run = rcu_cpu_kthread_should_run,
1180 .thread_fn = rcu_cpu_kthread,
1181 .thread_comm = "rcuc/%u",
1182 .setup = rcu_cpu_kthread_setup,
1183 .park = rcu_cpu_kthread_park,
1184};
1185
1186/*
1187 * Spawn boost kthreads -- called as soon as the scheduler is running.
1188 */
1189static void __init rcu_spawn_boost_kthreads(void)
1190{
1191 struct rcu_node *rnp;
1192 int cpu;
1193
1194 for_each_possible_cpu(cpu)
1195 per_cpu(rcu_cpu_has_work, cpu) = 0;
1196 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1197 rcu_for_each_leaf_node(rcu_state_p, rnp)
1198 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1199}
1200
1201static void rcu_prepare_kthreads(int cpu)
1202{
1203 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1204 struct rcu_node *rnp = rdp->mynode;
1205
1206 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1207 if (rcu_scheduler_fully_active)
1208 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1209}
1210
1211#else /* #ifdef CONFIG_RCU_BOOST */
1212
1213static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1214 __releases(rnp->lock)
1215{
1216 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1217}
1218
1219static void invoke_rcu_callbacks_kthread(void)
1220{
1221 WARN_ON_ONCE(1);
1222}
1223
1224static bool rcu_is_callbacks_kthread(void)
1225{
1226 return false;
1227}
1228
1229static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1230{
1231}
1232
1233static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1234{
1235}
1236
1237static void __init rcu_spawn_boost_kthreads(void)
1238{
1239}
1240
1241static void rcu_prepare_kthreads(int cpu)
1242{
1243}
1244
1245#endif /* #else #ifdef CONFIG_RCU_BOOST */
1246
1247#if !defined(CONFIG_RCU_FAST_NO_HZ)
1248
1249/*
1250 * Check to see if any future RCU-related work will need to be done
1251 * by the current CPU, even if none need be done immediately, returning
1252 * 1 if so. This function is part of the RCU implementation; it is -not-
1253 * an exported member of the RCU API.
1254 *
1255 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1256 * any flavor of RCU.
1257 */
1258int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1259{
1260 *nextevt = KTIME_MAX;
1261 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1262 ? 0 : rcu_cpu_has_callbacks(NULL);
1263}
1264
1265/*
1266 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1267 * after it.
1268 */
1269static void rcu_cleanup_after_idle(void)
1270{
1271}
1272
1273/*
1274 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1275 * is nothing.
1276 */
1277static void rcu_prepare_for_idle(void)
1278{
1279}
1280
1281/*
1282 * Don't bother keeping a running count of the number of RCU callbacks
1283 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1284 */
1285static void rcu_idle_count_callbacks_posted(void)
1286{
1287}
1288
1289#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1290
1291/*
1292 * This code is invoked when a CPU goes idle, at which point we want
1293 * to have the CPU do everything required for RCU so that it can enter
1294 * the energy-efficient dyntick-idle mode. This is handled by a
1295 * state machine implemented by rcu_prepare_for_idle() below.
1296 *
1297 * The following three proprocessor symbols control this state machine:
1298 *
1299 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1300 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1301 * is sized to be roughly one RCU grace period. Those energy-efficiency
1302 * benchmarkers who might otherwise be tempted to set this to a large
1303 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1304 * system. And if you are -that- concerned about energy efficiency,
1305 * just power the system down and be done with it!
1306 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1307 * permitted to sleep in dyntick-idle mode with only lazy RCU
1308 * callbacks pending. Setting this too high can OOM your system.
1309 *
1310 * The values below work well in practice. If future workloads require
1311 * adjustment, they can be converted into kernel config parameters, though
1312 * making the state machine smarter might be a better option.
1313 */
1314#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1315#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1316
1317static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1318module_param(rcu_idle_gp_delay, int, 0644);
1319static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1320module_param(rcu_idle_lazy_gp_delay, int, 0644);
1321
1322/*
1323 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1324 * only if it has been awhile since the last time we did so. Afterwards,
1325 * if there are any callbacks ready for immediate invocation, return true.
1326 */
1327static bool __maybe_unused rcu_try_advance_all_cbs(void)
1328{
1329 bool cbs_ready = false;
1330 struct rcu_data *rdp;
1331 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1332 struct rcu_node *rnp;
1333 struct rcu_state *rsp;
1334
1335 /* Exit early if we advanced recently. */
1336 if (jiffies == rdtp->last_advance_all)
1337 return false;
1338 rdtp->last_advance_all = jiffies;
1339
1340 for_each_rcu_flavor(rsp) {
1341 rdp = this_cpu_ptr(rsp->rda);
1342 rnp = rdp->mynode;
1343
1344 /*
1345 * Don't bother checking unless a grace period has
1346 * completed since we last checked and there are
1347 * callbacks not yet ready to invoke.
1348 */
1349 if ((rdp->completed != rnp->completed ||
1350 unlikely(READ_ONCE(rdp->gpwrap))) &&
1351 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1352 note_gp_changes(rsp, rdp);
1353
1354 if (cpu_has_callbacks_ready_to_invoke(rdp))
1355 cbs_ready = true;
1356 }
1357 return cbs_ready;
1358}
1359
1360/*
1361 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1362 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1363 * caller to set the timeout based on whether or not there are non-lazy
1364 * callbacks.
1365 *
1366 * The caller must have disabled interrupts.
1367 */
1368int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1369{
1370 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1371 unsigned long dj;
1372
1373 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1374 *nextevt = KTIME_MAX;
1375 return 0;
1376 }
1377
1378 /* Snapshot to detect later posting of non-lazy callback. */
1379 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1380
1381 /* If no callbacks, RCU doesn't need the CPU. */
1382 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1383 *nextevt = KTIME_MAX;
1384 return 0;
1385 }
1386
1387 /* Attempt to advance callbacks. */
1388 if (rcu_try_advance_all_cbs()) {
1389 /* Some ready to invoke, so initiate later invocation. */
1390 invoke_rcu_core();
1391 return 1;
1392 }
1393 rdtp->last_accelerate = jiffies;
1394
1395 /* Request timer delay depending on laziness, and round. */
1396 if (!rdtp->all_lazy) {
1397 dj = round_up(rcu_idle_gp_delay + jiffies,
1398 rcu_idle_gp_delay) - jiffies;
1399 } else {
1400 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1401 }
1402 *nextevt = basemono + dj * TICK_NSEC;
1403 return 0;
1404}
1405
1406/*
1407 * Prepare a CPU for idle from an RCU perspective. The first major task
1408 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1409 * The second major task is to check to see if a non-lazy callback has
1410 * arrived at a CPU that previously had only lazy callbacks. The third
1411 * major task is to accelerate (that is, assign grace-period numbers to)
1412 * any recently arrived callbacks.
1413 *
1414 * The caller must have disabled interrupts.
1415 */
1416static void rcu_prepare_for_idle(void)
1417{
1418 bool needwake;
1419 struct rcu_data *rdp;
1420 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1421 struct rcu_node *rnp;
1422 struct rcu_state *rsp;
1423 int tne;
1424
1425 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1426 rcu_is_nocb_cpu(smp_processor_id()))
1427 return;
1428
1429 /* Handle nohz enablement switches conservatively. */
1430 tne = READ_ONCE(tick_nohz_active);
1431 if (tne != rdtp->tick_nohz_enabled_snap) {
1432 if (rcu_cpu_has_callbacks(NULL))
1433 invoke_rcu_core(); /* force nohz to see update. */
1434 rdtp->tick_nohz_enabled_snap = tne;
1435 return;
1436 }
1437 if (!tne)
1438 return;
1439
1440 /*
1441 * If a non-lazy callback arrived at a CPU having only lazy
1442 * callbacks, invoke RCU core for the side-effect of recalculating
1443 * idle duration on re-entry to idle.
1444 */
1445 if (rdtp->all_lazy &&
1446 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1447 rdtp->all_lazy = false;
1448 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1449 invoke_rcu_core();
1450 return;
1451 }
1452
1453 /*
1454 * If we have not yet accelerated this jiffy, accelerate all
1455 * callbacks on this CPU.
1456 */
1457 if (rdtp->last_accelerate == jiffies)
1458 return;
1459 rdtp->last_accelerate = jiffies;
1460 for_each_rcu_flavor(rsp) {
1461 rdp = this_cpu_ptr(rsp->rda);
1462 if (!*rdp->nxttail[RCU_DONE_TAIL])
1463 continue;
1464 rnp = rdp->mynode;
1465 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1466 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1467 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1468 if (needwake)
1469 rcu_gp_kthread_wake(rsp);
1470 }
1471}
1472
1473/*
1474 * Clean up for exit from idle. Attempt to advance callbacks based on
1475 * any grace periods that elapsed while the CPU was idle, and if any
1476 * callbacks are now ready to invoke, initiate invocation.
1477 */
1478static void rcu_cleanup_after_idle(void)
1479{
1480 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1481 rcu_is_nocb_cpu(smp_processor_id()))
1482 return;
1483 if (rcu_try_advance_all_cbs())
1484 invoke_rcu_core();
1485}
1486
1487/*
1488 * Keep a running count of the number of non-lazy callbacks posted
1489 * on this CPU. This running counter (which is never decremented) allows
1490 * rcu_prepare_for_idle() to detect when something out of the idle loop
1491 * posts a callback, even if an equal number of callbacks are invoked.
1492 * Of course, callbacks should only be posted from within a trace event
1493 * designed to be called from idle or from within RCU_NONIDLE().
1494 */
1495static void rcu_idle_count_callbacks_posted(void)
1496{
1497 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1498}
1499
1500/*
1501 * Data for flushing lazy RCU callbacks at OOM time.
1502 */
1503static atomic_t oom_callback_count;
1504static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1505
1506/*
1507 * RCU OOM callback -- decrement the outstanding count and deliver the
1508 * wake-up if we are the last one.
1509 */
1510static void rcu_oom_callback(struct rcu_head *rhp)
1511{
1512 if (atomic_dec_and_test(&oom_callback_count))
1513 wake_up(&oom_callback_wq);
1514}
1515
1516/*
1517 * Post an rcu_oom_notify callback on the current CPU if it has at
1518 * least one lazy callback. This will unnecessarily post callbacks
1519 * to CPUs that already have a non-lazy callback at the end of their
1520 * callback list, but this is an infrequent operation, so accept some
1521 * extra overhead to keep things simple.
1522 */
1523static void rcu_oom_notify_cpu(void *unused)
1524{
1525 struct rcu_state *rsp;
1526 struct rcu_data *rdp;
1527
1528 for_each_rcu_flavor(rsp) {
1529 rdp = raw_cpu_ptr(rsp->rda);
1530 if (rdp->qlen_lazy != 0) {
1531 atomic_inc(&oom_callback_count);
1532 rsp->call(&rdp->oom_head, rcu_oom_callback);
1533 }
1534 }
1535}
1536
1537/*
1538 * If low on memory, ensure that each CPU has a non-lazy callback.
1539 * This will wake up CPUs that have only lazy callbacks, in turn
1540 * ensuring that they free up the corresponding memory in a timely manner.
1541 * Because an uncertain amount of memory will be freed in some uncertain
1542 * timeframe, we do not claim to have freed anything.
1543 */
1544static int rcu_oom_notify(struct notifier_block *self,
1545 unsigned long notused, void *nfreed)
1546{
1547 int cpu;
1548
1549 /* Wait for callbacks from earlier instance to complete. */
1550 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1551 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1552
1553 /*
1554 * Prevent premature wakeup: ensure that all increments happen
1555 * before there is a chance of the counter reaching zero.
1556 */
1557 atomic_set(&oom_callback_count, 1);
1558
1559 for_each_online_cpu(cpu) {
1560 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1561 cond_resched_rcu_qs();
1562 }
1563
1564 /* Unconditionally decrement: no need to wake ourselves up. */
1565 atomic_dec(&oom_callback_count);
1566
1567 return NOTIFY_OK;
1568}
1569
1570static struct notifier_block rcu_oom_nb = {
1571 .notifier_call = rcu_oom_notify
1572};
1573
1574static int __init rcu_register_oom_notifier(void)
1575{
1576 register_oom_notifier(&rcu_oom_nb);
1577 return 0;
1578}
1579early_initcall(rcu_register_oom_notifier);
1580
1581#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1582
1583#ifdef CONFIG_RCU_FAST_NO_HZ
1584
1585static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1586{
1587 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1588 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1589
1590 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1591 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1592 ulong2long(nlpd),
1593 rdtp->all_lazy ? 'L' : '.',
1594 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1595}
1596
1597#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1598
1599static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1600{
1601 *cp = '\0';
1602}
1603
1604#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1605
1606/* Initiate the stall-info list. */
1607static void print_cpu_stall_info_begin(void)
1608{
1609 pr_cont("\n");
1610}
1611
1612/*
1613 * Print out diagnostic information for the specified stalled CPU.
1614 *
1615 * If the specified CPU is aware of the current RCU grace period
1616 * (flavor specified by rsp), then print the number of scheduling
1617 * clock interrupts the CPU has taken during the time that it has
1618 * been aware. Otherwise, print the number of RCU grace periods
1619 * that this CPU is ignorant of, for example, "1" if the CPU was
1620 * aware of the previous grace period.
1621 *
1622 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1623 */
1624static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1625{
1626 char fast_no_hz[72];
1627 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1628 struct rcu_dynticks *rdtp = rdp->dynticks;
1629 char *ticks_title;
1630 unsigned long ticks_value;
1631
1632 if (rsp->gpnum == rdp->gpnum) {
1633 ticks_title = "ticks this GP";
1634 ticks_value = rdp->ticks_this_gp;
1635 } else {
1636 ticks_title = "GPs behind";
1637 ticks_value = rsp->gpnum - rdp->gpnum;
1638 }
1639 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1640 pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1641 cpu,
1642 "O."[!!cpu_online(cpu)],
1643 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1644 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1645 ticks_value, ticks_title,
1646 atomic_read(&rdtp->dynticks) & 0xfff,
1647 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1648 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1649 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1650 fast_no_hz);
1651}
1652
1653/* Terminate the stall-info list. */
1654static void print_cpu_stall_info_end(void)
1655{
1656 pr_err("\t");
1657}
1658
1659/* Zero ->ticks_this_gp for all flavors of RCU. */
1660static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1661{
1662 rdp->ticks_this_gp = 0;
1663 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1664}
1665
1666/* Increment ->ticks_this_gp for all flavors of RCU. */
1667static void increment_cpu_stall_ticks(void)
1668{
1669 struct rcu_state *rsp;
1670
1671 for_each_rcu_flavor(rsp)
1672 raw_cpu_inc(rsp->rda->ticks_this_gp);
1673}
1674
1675#ifdef CONFIG_RCU_NOCB_CPU
1676
1677/*
1678 * Offload callback processing from the boot-time-specified set of CPUs
1679 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1680 * kthread created that pulls the callbacks from the corresponding CPU,
1681 * waits for a grace period to elapse, and invokes the callbacks.
1682 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1683 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1684 * has been specified, in which case each kthread actively polls its
1685 * CPU. (Which isn't so great for energy efficiency, but which does
1686 * reduce RCU's overhead on that CPU.)
1687 *
1688 * This is intended to be used in conjunction with Frederic Weisbecker's
1689 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1690 * running CPU-bound user-mode computations.
1691 *
1692 * Offloading of callback processing could also in theory be used as
1693 * an energy-efficiency measure because CPUs with no RCU callbacks
1694 * queued are more aggressive about entering dyntick-idle mode.
1695 */
1696
1697
1698/* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1699static int __init rcu_nocb_setup(char *str)
1700{
1701 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1702 have_rcu_nocb_mask = true;
1703 cpulist_parse(str, rcu_nocb_mask);
1704 return 1;
1705}
1706__setup("rcu_nocbs=", rcu_nocb_setup);
1707
1708static int __init parse_rcu_nocb_poll(char *arg)
1709{
1710 rcu_nocb_poll = 1;
1711 return 0;
1712}
1713early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1714
1715/*
1716 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1717 * grace period.
1718 */
1719static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1720{
1721 swake_up_all(sq);
1722}
1723
1724/*
1725 * Set the root rcu_node structure's ->need_future_gp field
1726 * based on the sum of those of all rcu_node structures. This does
1727 * double-count the root rcu_node structure's requests, but this
1728 * is necessary to handle the possibility of a rcu_nocb_kthread()
1729 * having awakened during the time that the rcu_node structures
1730 * were being updated for the end of the previous grace period.
1731 */
1732static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1733{
1734 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1735}
1736
1737static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1738{
1739 return &rnp->nocb_gp_wq[rnp->completed & 0x1];
1740}
1741
1742static void rcu_init_one_nocb(struct rcu_node *rnp)
1743{
1744 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1745 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1746}
1747
1748#ifndef CONFIG_RCU_NOCB_CPU_ALL
1749/* Is the specified CPU a no-CBs CPU? */
1750bool rcu_is_nocb_cpu(int cpu)
1751{
1752 if (have_rcu_nocb_mask)
1753 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1754 return false;
1755}
1756#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1757
1758/*
1759 * Kick the leader kthread for this NOCB group.
1760 */
1761static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1762{
1763 struct rcu_data *rdp_leader = rdp->nocb_leader;
1764
1765 if (!READ_ONCE(rdp_leader->nocb_kthread))
1766 return;
1767 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1768 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1769 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1770 swake_up(&rdp_leader->nocb_wq);
1771 }
1772}
1773
1774/*
1775 * Does the specified CPU need an RCU callback for the specified flavor
1776 * of rcu_barrier()?
1777 */
1778static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1779{
1780 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1781 unsigned long ret;
1782#ifdef CONFIG_PROVE_RCU
1783 struct rcu_head *rhp;
1784#endif /* #ifdef CONFIG_PROVE_RCU */
1785
1786 /*
1787 * Check count of all no-CBs callbacks awaiting invocation.
1788 * There needs to be a barrier before this function is called,
1789 * but associated with a prior determination that no more
1790 * callbacks would be posted. In the worst case, the first
1791 * barrier in _rcu_barrier() suffices (but the caller cannot
1792 * necessarily rely on this, not a substitute for the caller
1793 * getting the concurrency design right!). There must also be
1794 * a barrier between the following load an posting of a callback
1795 * (if a callback is in fact needed). This is associated with an
1796 * atomic_inc() in the caller.
1797 */
1798 ret = atomic_long_read(&rdp->nocb_q_count);
1799
1800#ifdef CONFIG_PROVE_RCU
1801 rhp = READ_ONCE(rdp->nocb_head);
1802 if (!rhp)
1803 rhp = READ_ONCE(rdp->nocb_gp_head);
1804 if (!rhp)
1805 rhp = READ_ONCE(rdp->nocb_follower_head);
1806
1807 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1808 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1809 rcu_scheduler_fully_active) {
1810 /* RCU callback enqueued before CPU first came online??? */
1811 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1812 cpu, rhp->func);
1813 WARN_ON_ONCE(1);
1814 }
1815#endif /* #ifdef CONFIG_PROVE_RCU */
1816
1817 return !!ret;
1818}
1819
1820/*
1821 * Enqueue the specified string of rcu_head structures onto the specified
1822 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1823 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1824 * counts are supplied by rhcount and rhcount_lazy.
1825 *
1826 * If warranted, also wake up the kthread servicing this CPUs queues.
1827 */
1828static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1829 struct rcu_head *rhp,
1830 struct rcu_head **rhtp,
1831 int rhcount, int rhcount_lazy,
1832 unsigned long flags)
1833{
1834 int len;
1835 struct rcu_head **old_rhpp;
1836 struct task_struct *t;
1837
1838 /* Enqueue the callback on the nocb list and update counts. */
1839 atomic_long_add(rhcount, &rdp->nocb_q_count);
1840 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1841 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1842 WRITE_ONCE(*old_rhpp, rhp);
1843 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1844 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1845
1846 /* If we are not being polled and there is a kthread, awaken it ... */
1847 t = READ_ONCE(rdp->nocb_kthread);
1848 if (rcu_nocb_poll || !t) {
1849 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1850 TPS("WakeNotPoll"));
1851 return;
1852 }
1853 len = atomic_long_read(&rdp->nocb_q_count);
1854 if (old_rhpp == &rdp->nocb_head) {
1855 if (!irqs_disabled_flags(flags)) {
1856 /* ... if queue was empty ... */
1857 wake_nocb_leader(rdp, false);
1858 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1859 TPS("WakeEmpty"));
1860 } else {
1861 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1862 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1863 TPS("WakeEmptyIsDeferred"));
1864 }
1865 rdp->qlen_last_fqs_check = 0;
1866 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1867 /* ... or if many callbacks queued. */
1868 if (!irqs_disabled_flags(flags)) {
1869 wake_nocb_leader(rdp, true);
1870 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1871 TPS("WakeOvf"));
1872 } else {
1873 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1874 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1875 TPS("WakeOvfIsDeferred"));
1876 }
1877 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1878 } else {
1879 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1880 }
1881 return;
1882}
1883
1884/*
1885 * This is a helper for __call_rcu(), which invokes this when the normal
1886 * callback queue is inoperable. If this is not a no-CBs CPU, this
1887 * function returns failure back to __call_rcu(), which can complain
1888 * appropriately.
1889 *
1890 * Otherwise, this function queues the callback where the corresponding
1891 * "rcuo" kthread can find it.
1892 */
1893static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1894 bool lazy, unsigned long flags)
1895{
1896
1897 if (!rcu_is_nocb_cpu(rdp->cpu))
1898 return false;
1899 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1900 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1901 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
1902 (unsigned long)rhp->func,
1903 -atomic_long_read(&rdp->nocb_q_count_lazy),
1904 -atomic_long_read(&rdp->nocb_q_count));
1905 else
1906 trace_rcu_callback(rdp->rsp->name, rhp,
1907 -atomic_long_read(&rdp->nocb_q_count_lazy),
1908 -atomic_long_read(&rdp->nocb_q_count));
1909
1910 /*
1911 * If called from an extended quiescent state with interrupts
1912 * disabled, invoke the RCU core in order to allow the idle-entry
1913 * deferred-wakeup check to function.
1914 */
1915 if (irqs_disabled_flags(flags) &&
1916 !rcu_is_watching() &&
1917 cpu_online(smp_processor_id()))
1918 invoke_rcu_core();
1919
1920 return true;
1921}
1922
1923/*
1924 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
1925 * not a no-CBs CPU.
1926 */
1927static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
1928 struct rcu_data *rdp,
1929 unsigned long flags)
1930{
1931 long ql = rsp->qlen;
1932 long qll = rsp->qlen_lazy;
1933
1934 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
1935 if (!rcu_is_nocb_cpu(smp_processor_id()))
1936 return false;
1937 rsp->qlen = 0;
1938 rsp->qlen_lazy = 0;
1939
1940 /* First, enqueue the donelist, if any. This preserves CB ordering. */
1941 if (rsp->orphan_donelist != NULL) {
1942 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
1943 rsp->orphan_donetail, ql, qll, flags);
1944 ql = qll = 0;
1945 rsp->orphan_donelist = NULL;
1946 rsp->orphan_donetail = &rsp->orphan_donelist;
1947 }
1948 if (rsp->orphan_nxtlist != NULL) {
1949 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
1950 rsp->orphan_nxttail, ql, qll, flags);
1951 ql = qll = 0;
1952 rsp->orphan_nxtlist = NULL;
1953 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1954 }
1955 return true;
1956}
1957
1958/*
1959 * If necessary, kick off a new grace period, and either way wait
1960 * for a subsequent grace period to complete.
1961 */
1962static void rcu_nocb_wait_gp(struct rcu_data *rdp)
1963{
1964 unsigned long c;
1965 bool d;
1966 unsigned long flags;
1967 bool needwake;
1968 struct rcu_node *rnp = rdp->mynode;
1969
1970 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1971 needwake = rcu_start_future_gp(rnp, rdp, &c);
1972 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1973 if (needwake)
1974 rcu_gp_kthread_wake(rdp->rsp);
1975
1976 /*
1977 * Wait for the grace period. Do so interruptibly to avoid messing
1978 * up the load average.
1979 */
1980 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
1981 for (;;) {
1982 swait_event_interruptible(
1983 rnp->nocb_gp_wq[c & 0x1],
1984 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
1985 if (likely(d))
1986 break;
1987 WARN_ON(signal_pending(current));
1988 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
1989 }
1990 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
1991 smp_mb(); /* Ensure that CB invocation happens after GP end. */
1992}
1993
1994/*
1995 * Leaders come here to wait for additional callbacks to show up.
1996 * This function does not return until callbacks appear.
1997 */
1998static void nocb_leader_wait(struct rcu_data *my_rdp)
1999{
2000 bool firsttime = true;
2001 bool gotcbs;
2002 struct rcu_data *rdp;
2003 struct rcu_head **tail;
2004
2005wait_again:
2006
2007 /* Wait for callbacks to appear. */
2008 if (!rcu_nocb_poll) {
2009 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2010 swait_event_interruptible(my_rdp->nocb_wq,
2011 !READ_ONCE(my_rdp->nocb_leader_sleep));
2012 /* Memory barrier handled by smp_mb() calls below and repoll. */
2013 } else if (firsttime) {
2014 firsttime = false; /* Don't drown trace log with "Poll"! */
2015 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2016 }
2017
2018 /*
2019 * Each pass through the following loop checks a follower for CBs.
2020 * We are our own first follower. Any CBs found are moved to
2021 * nocb_gp_head, where they await a grace period.
2022 */
2023 gotcbs = false;
2024 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2025 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2026 if (!rdp->nocb_gp_head)
2027 continue; /* No CBs here, try next follower. */
2028
2029 /* Move callbacks to wait-for-GP list, which is empty. */
2030 WRITE_ONCE(rdp->nocb_head, NULL);
2031 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2032 gotcbs = true;
2033 }
2034
2035 /*
2036 * If there were no callbacks, sleep a bit, rescan after a
2037 * memory barrier, and go retry.
2038 */
2039 if (unlikely(!gotcbs)) {
2040 if (!rcu_nocb_poll)
2041 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2042 "WokeEmpty");
2043 WARN_ON(signal_pending(current));
2044 schedule_timeout_interruptible(1);
2045
2046 /* Rescan in case we were a victim of memory ordering. */
2047 my_rdp->nocb_leader_sleep = true;
2048 smp_mb(); /* Ensure _sleep true before scan. */
2049 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2050 if (READ_ONCE(rdp->nocb_head)) {
2051 /* Found CB, so short-circuit next wait. */
2052 my_rdp->nocb_leader_sleep = false;
2053 break;
2054 }
2055 goto wait_again;
2056 }
2057
2058 /* Wait for one grace period. */
2059 rcu_nocb_wait_gp(my_rdp);
2060
2061 /*
2062 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2063 * We set it now, but recheck for new callbacks while
2064 * traversing our follower list.
2065 */
2066 my_rdp->nocb_leader_sleep = true;
2067 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2068
2069 /* Each pass through the following loop wakes a follower, if needed. */
2070 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2071 if (READ_ONCE(rdp->nocb_head))
2072 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2073 if (!rdp->nocb_gp_head)
2074 continue; /* No CBs, so no need to wake follower. */
2075
2076 /* Append callbacks to follower's "done" list. */
2077 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2078 *tail = rdp->nocb_gp_head;
2079 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2080 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2081 /*
2082 * List was empty, wake up the follower.
2083 * Memory barriers supplied by atomic_long_add().
2084 */
2085 swake_up(&rdp->nocb_wq);
2086 }
2087 }
2088
2089 /* If we (the leader) don't have CBs, go wait some more. */
2090 if (!my_rdp->nocb_follower_head)
2091 goto wait_again;
2092}
2093
2094/*
2095 * Followers come here to wait for additional callbacks to show up.
2096 * This function does not return until callbacks appear.
2097 */
2098static void nocb_follower_wait(struct rcu_data *rdp)
2099{
2100 bool firsttime = true;
2101
2102 for (;;) {
2103 if (!rcu_nocb_poll) {
2104 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2105 "FollowerSleep");
2106 swait_event_interruptible(rdp->nocb_wq,
2107 READ_ONCE(rdp->nocb_follower_head));
2108 } else if (firsttime) {
2109 /* Don't drown trace log with "Poll"! */
2110 firsttime = false;
2111 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2112 }
2113 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2114 /* ^^^ Ensure CB invocation follows _head test. */
2115 return;
2116 }
2117 if (!rcu_nocb_poll)
2118 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2119 "WokeEmpty");
2120 WARN_ON(signal_pending(current));
2121 schedule_timeout_interruptible(1);
2122 }
2123}
2124
2125/*
2126 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2127 * callbacks queued by the corresponding no-CBs CPU, however, there is
2128 * an optional leader-follower relationship so that the grace-period
2129 * kthreads don't have to do quite so many wakeups.
2130 */
2131static int rcu_nocb_kthread(void *arg)
2132{
2133 int c, cl;
2134 struct rcu_head *list;
2135 struct rcu_head *next;
2136 struct rcu_head **tail;
2137 struct rcu_data *rdp = arg;
2138
2139 /* Each pass through this loop invokes one batch of callbacks */
2140 for (;;) {
2141 /* Wait for callbacks. */
2142 if (rdp->nocb_leader == rdp)
2143 nocb_leader_wait(rdp);
2144 else
2145 nocb_follower_wait(rdp);
2146
2147 /* Pull the ready-to-invoke callbacks onto local list. */
2148 list = READ_ONCE(rdp->nocb_follower_head);
2149 BUG_ON(!list);
2150 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2151 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2152 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2153
2154 /* Each pass through the following loop invokes a callback. */
2155 trace_rcu_batch_start(rdp->rsp->name,
2156 atomic_long_read(&rdp->nocb_q_count_lazy),
2157 atomic_long_read(&rdp->nocb_q_count), -1);
2158 c = cl = 0;
2159 while (list) {
2160 next = list->next;
2161 /* Wait for enqueuing to complete, if needed. */
2162 while (next == NULL && &list->next != tail) {
2163 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2164 TPS("WaitQueue"));
2165 schedule_timeout_interruptible(1);
2166 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2167 TPS("WokeQueue"));
2168 next = list->next;
2169 }
2170 debug_rcu_head_unqueue(list);
2171 local_bh_disable();
2172 if (__rcu_reclaim(rdp->rsp->name, list))
2173 cl++;
2174 c++;
2175 local_bh_enable();
2176 cond_resched_rcu_qs();
2177 list = next;
2178 }
2179 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2180 smp_mb__before_atomic(); /* _add after CB invocation. */
2181 atomic_long_add(-c, &rdp->nocb_q_count);
2182 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2183 rdp->n_nocbs_invoked += c;
2184 }
2185 return 0;
2186}
2187
2188/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2189static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2190{
2191 return READ_ONCE(rdp->nocb_defer_wakeup);
2192}
2193
2194/* Do a deferred wakeup of rcu_nocb_kthread(). */
2195static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2196{
2197 int ndw;
2198
2199 if (!rcu_nocb_need_deferred_wakeup(rdp))
2200 return;
2201 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2202 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2203 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2204 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2205}
2206
2207void __init rcu_init_nohz(void)
2208{
2209 int cpu;
2210 bool need_rcu_nocb_mask = true;
2211 struct rcu_state *rsp;
2212
2213#ifdef CONFIG_RCU_NOCB_CPU_NONE
2214 need_rcu_nocb_mask = false;
2215#endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2216
2217#if defined(CONFIG_NO_HZ_FULL)
2218 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2219 need_rcu_nocb_mask = true;
2220#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2221
2222 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2223 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2224 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2225 return;
2226 }
2227 have_rcu_nocb_mask = true;
2228 }
2229 if (!have_rcu_nocb_mask)
2230 return;
2231
2232#ifdef CONFIG_RCU_NOCB_CPU_ZERO
2233 pr_info("\tOffload RCU callbacks from CPU 0\n");
2234 cpumask_set_cpu(0, rcu_nocb_mask);
2235#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2236#ifdef CONFIG_RCU_NOCB_CPU_ALL
2237 pr_info("\tOffload RCU callbacks from all CPUs\n");
2238 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2239#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2240#if defined(CONFIG_NO_HZ_FULL)
2241 if (tick_nohz_full_running)
2242 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2243#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2244
2245 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2246 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2247 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2248 rcu_nocb_mask);
2249 }
2250 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2251 cpumask_pr_args(rcu_nocb_mask));
2252 if (rcu_nocb_poll)
2253 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2254
2255 for_each_rcu_flavor(rsp) {
2256 for_each_cpu(cpu, rcu_nocb_mask)
2257 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2258 rcu_organize_nocb_kthreads(rsp);
2259 }
2260}
2261
2262/* Initialize per-rcu_data variables for no-CBs CPUs. */
2263static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2264{
2265 rdp->nocb_tail = &rdp->nocb_head;
2266 init_swait_queue_head(&rdp->nocb_wq);
2267 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2268}
2269
2270/*
2271 * If the specified CPU is a no-CBs CPU that does not already have its
2272 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2273 * brought online out of order, this can require re-organizing the
2274 * leader-follower relationships.
2275 */
2276static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2277{
2278 struct rcu_data *rdp;
2279 struct rcu_data *rdp_last;
2280 struct rcu_data *rdp_old_leader;
2281 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2282 struct task_struct *t;
2283
2284 /*
2285 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2286 * then nothing to do.
2287 */
2288 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2289 return;
2290
2291 /* If we didn't spawn the leader first, reorganize! */
2292 rdp_old_leader = rdp_spawn->nocb_leader;
2293 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2294 rdp_last = NULL;
2295 rdp = rdp_old_leader;
2296 do {
2297 rdp->nocb_leader = rdp_spawn;
2298 if (rdp_last && rdp != rdp_spawn)
2299 rdp_last->nocb_next_follower = rdp;
2300 if (rdp == rdp_spawn) {
2301 rdp = rdp->nocb_next_follower;
2302 } else {
2303 rdp_last = rdp;
2304 rdp = rdp->nocb_next_follower;
2305 rdp_last->nocb_next_follower = NULL;
2306 }
2307 } while (rdp);
2308 rdp_spawn->nocb_next_follower = rdp_old_leader;
2309 }
2310
2311 /* Spawn the kthread for this CPU and RCU flavor. */
2312 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2313 "rcuo%c/%d", rsp->abbr, cpu);
2314 BUG_ON(IS_ERR(t));
2315 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2316}
2317
2318/*
2319 * If the specified CPU is a no-CBs CPU that does not already have its
2320 * rcuo kthreads, spawn them.
2321 */
2322static void rcu_spawn_all_nocb_kthreads(int cpu)
2323{
2324 struct rcu_state *rsp;
2325
2326 if (rcu_scheduler_fully_active)
2327 for_each_rcu_flavor(rsp)
2328 rcu_spawn_one_nocb_kthread(rsp, cpu);
2329}
2330
2331/*
2332 * Once the scheduler is running, spawn rcuo kthreads for all online
2333 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2334 * non-boot CPUs come online -- if this changes, we will need to add
2335 * some mutual exclusion.
2336 */
2337static void __init rcu_spawn_nocb_kthreads(void)
2338{
2339 int cpu;
2340
2341 for_each_online_cpu(cpu)
2342 rcu_spawn_all_nocb_kthreads(cpu);
2343}
2344
2345/* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2346static int rcu_nocb_leader_stride = -1;
2347module_param(rcu_nocb_leader_stride, int, 0444);
2348
2349/*
2350 * Initialize leader-follower relationships for all no-CBs CPU.
2351 */
2352static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2353{
2354 int cpu;
2355 int ls = rcu_nocb_leader_stride;
2356 int nl = 0; /* Next leader. */
2357 struct rcu_data *rdp;
2358 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2359 struct rcu_data *rdp_prev = NULL;
2360
2361 if (!have_rcu_nocb_mask)
2362 return;
2363 if (ls == -1) {
2364 ls = int_sqrt(nr_cpu_ids);
2365 rcu_nocb_leader_stride = ls;
2366 }
2367
2368 /*
2369 * Each pass through this loop sets up one rcu_data structure and
2370 * spawns one rcu_nocb_kthread().
2371 */
2372 for_each_cpu(cpu, rcu_nocb_mask) {
2373 rdp = per_cpu_ptr(rsp->rda, cpu);
2374 if (rdp->cpu >= nl) {
2375 /* New leader, set up for followers & next leader. */
2376 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2377 rdp->nocb_leader = rdp;
2378 rdp_leader = rdp;
2379 } else {
2380 /* Another follower, link to previous leader. */
2381 rdp->nocb_leader = rdp_leader;
2382 rdp_prev->nocb_next_follower = rdp;
2383 }
2384 rdp_prev = rdp;
2385 }
2386}
2387
2388/* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2389static bool init_nocb_callback_list(struct rcu_data *rdp)
2390{
2391 if (!rcu_is_nocb_cpu(rdp->cpu))
2392 return false;
2393
2394 /* If there are early-boot callbacks, move them to nocb lists. */
2395 if (rdp->nxtlist) {
2396 rdp->nocb_head = rdp->nxtlist;
2397 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2398 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2399 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2400 rdp->nxtlist = NULL;
2401 rdp->qlen = 0;
2402 rdp->qlen_lazy = 0;
2403 }
2404 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2405 return true;
2406}
2407
2408#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2409
2410static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2411{
2412 WARN_ON_ONCE(1); /* Should be dead code. */
2413 return false;
2414}
2415
2416static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2417{
2418}
2419
2420static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2421{
2422}
2423
2424static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2425{
2426 return NULL;
2427}
2428
2429static void rcu_init_one_nocb(struct rcu_node *rnp)
2430{
2431}
2432
2433static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2434 bool lazy, unsigned long flags)
2435{
2436 return false;
2437}
2438
2439static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2440 struct rcu_data *rdp,
2441 unsigned long flags)
2442{
2443 return false;
2444}
2445
2446static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2447{
2448}
2449
2450static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2451{
2452 return false;
2453}
2454
2455static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2456{
2457}
2458
2459static void rcu_spawn_all_nocb_kthreads(int cpu)
2460{
2461}
2462
2463static void __init rcu_spawn_nocb_kthreads(void)
2464{
2465}
2466
2467static bool init_nocb_callback_list(struct rcu_data *rdp)
2468{
2469 return false;
2470}
2471
2472#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2473
2474/*
2475 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2476 * arbitrarily long period of time with the scheduling-clock tick turned
2477 * off. RCU will be paying attention to this CPU because it is in the
2478 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2479 * machine because the scheduling-clock tick has been disabled. Therefore,
2480 * if an adaptive-ticks CPU is failing to respond to the current grace
2481 * period and has not be idle from an RCU perspective, kick it.
2482 */
2483static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2484{
2485#ifdef CONFIG_NO_HZ_FULL
2486 if (tick_nohz_full_cpu(cpu))
2487 smp_send_reschedule(cpu);
2488#endif /* #ifdef CONFIG_NO_HZ_FULL */
2489}
2490
2491
2492#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2493
2494static int full_sysidle_state; /* Current system-idle state. */
2495#define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2496#define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2497#define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2498#define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2499#define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2500
2501/*
2502 * Invoked to note exit from irq or task transition to idle. Note that
2503 * usermode execution does -not- count as idle here! After all, we want
2504 * to detect full-system idle states, not RCU quiescent states and grace
2505 * periods. The caller must have disabled interrupts.
2506 */
2507static void rcu_sysidle_enter(int irq)
2508{
2509 unsigned long j;
2510 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2511
2512 /* If there are no nohz_full= CPUs, no need to track this. */
2513 if (!tick_nohz_full_enabled())
2514 return;
2515
2516 /* Adjust nesting, check for fully idle. */
2517 if (irq) {
2518 rdtp->dynticks_idle_nesting--;
2519 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2520 if (rdtp->dynticks_idle_nesting != 0)
2521 return; /* Still not fully idle. */
2522 } else {
2523 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2524 DYNTICK_TASK_NEST_VALUE) {
2525 rdtp->dynticks_idle_nesting = 0;
2526 } else {
2527 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2528 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2529 return; /* Still not fully idle. */
2530 }
2531 }
2532
2533 /* Record start of fully idle period. */
2534 j = jiffies;
2535 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2536 smp_mb__before_atomic();
2537 atomic_inc(&rdtp->dynticks_idle);
2538 smp_mb__after_atomic();
2539 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2540}
2541
2542/*
2543 * Unconditionally force exit from full system-idle state. This is
2544 * invoked when a normal CPU exits idle, but must be called separately
2545 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2546 * is that the timekeeping CPU is permitted to take scheduling-clock
2547 * interrupts while the system is in system-idle state, and of course
2548 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2549 * interrupt from any other type of interrupt.
2550 */
2551void rcu_sysidle_force_exit(void)
2552{
2553 int oldstate = READ_ONCE(full_sysidle_state);
2554 int newoldstate;
2555
2556 /*
2557 * Each pass through the following loop attempts to exit full
2558 * system-idle state. If contention proves to be a problem,
2559 * a trylock-based contention tree could be used here.
2560 */
2561 while (oldstate > RCU_SYSIDLE_SHORT) {
2562 newoldstate = cmpxchg(&full_sysidle_state,
2563 oldstate, RCU_SYSIDLE_NOT);
2564 if (oldstate == newoldstate &&
2565 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2566 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2567 return; /* We cleared it, done! */
2568 }
2569 oldstate = newoldstate;
2570 }
2571 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2572}
2573
2574/*
2575 * Invoked to note entry to irq or task transition from idle. Note that
2576 * usermode execution does -not- count as idle here! The caller must
2577 * have disabled interrupts.
2578 */
2579static void rcu_sysidle_exit(int irq)
2580{
2581 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2582
2583 /* If there are no nohz_full= CPUs, no need to track this. */
2584 if (!tick_nohz_full_enabled())
2585 return;
2586
2587 /* Adjust nesting, check for already non-idle. */
2588 if (irq) {
2589 rdtp->dynticks_idle_nesting++;
2590 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2591 if (rdtp->dynticks_idle_nesting != 1)
2592 return; /* Already non-idle. */
2593 } else {
2594 /*
2595 * Allow for irq misnesting. Yes, it really is possible
2596 * to enter an irq handler then never leave it, and maybe
2597 * also vice versa. Handle both possibilities.
2598 */
2599 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2600 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2601 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2602 return; /* Already non-idle. */
2603 } else {
2604 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2605 }
2606 }
2607
2608 /* Record end of idle period. */
2609 smp_mb__before_atomic();
2610 atomic_inc(&rdtp->dynticks_idle);
2611 smp_mb__after_atomic();
2612 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2613
2614 /*
2615 * If we are the timekeeping CPU, we are permitted to be non-idle
2616 * during a system-idle state. This must be the case, because
2617 * the timekeeping CPU has to take scheduling-clock interrupts
2618 * during the time that the system is transitioning to full
2619 * system-idle state. This means that the timekeeping CPU must
2620 * invoke rcu_sysidle_force_exit() directly if it does anything
2621 * more than take a scheduling-clock interrupt.
2622 */
2623 if (smp_processor_id() == tick_do_timer_cpu)
2624 return;
2625
2626 /* Update system-idle state: We are clearly no longer fully idle! */
2627 rcu_sysidle_force_exit();
2628}
2629
2630/*
2631 * Check to see if the current CPU is idle. Note that usermode execution
2632 * does not count as idle. The caller must have disabled interrupts,
2633 * and must be running on tick_do_timer_cpu.
2634 */
2635static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2636 unsigned long *maxj)
2637{
2638 int cur;
2639 unsigned long j;
2640 struct rcu_dynticks *rdtp = rdp->dynticks;
2641
2642 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2643 if (!tick_nohz_full_enabled())
2644 return;
2645
2646 /*
2647 * If some other CPU has already reported non-idle, if this is
2648 * not the flavor of RCU that tracks sysidle state, or if this
2649 * is an offline or the timekeeping CPU, nothing to do.
2650 */
2651 if (!*isidle || rdp->rsp != rcu_state_p ||
2652 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2653 return;
2654 /* Verify affinity of current kthread. */
2655 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2656
2657 /* Pick up current idle and NMI-nesting counter and check. */
2658 cur = atomic_read(&rdtp->dynticks_idle);
2659 if (cur & 0x1) {
2660 *isidle = false; /* We are not idle! */
2661 return;
2662 }
2663 smp_mb(); /* Read counters before timestamps. */
2664
2665 /* Pick up timestamps. */
2666 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2667 /* If this CPU entered idle more recently, update maxj timestamp. */
2668 if (ULONG_CMP_LT(*maxj, j))
2669 *maxj = j;
2670}
2671
2672/*
2673 * Is this the flavor of RCU that is handling full-system idle?
2674 */
2675static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2676{
2677 return rsp == rcu_state_p;
2678}
2679
2680/*
2681 * Return a delay in jiffies based on the number of CPUs, rcu_node
2682 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2683 * systems more time to transition to full-idle state in order to
2684 * avoid the cache thrashing that otherwise occur on the state variable.
2685 * Really small systems (less than a couple of tens of CPUs) should
2686 * instead use a single global atomically incremented counter, and later
2687 * versions of this will automatically reconfigure themselves accordingly.
2688 */
2689static unsigned long rcu_sysidle_delay(void)
2690{
2691 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2692 return 0;
2693 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2694}
2695
2696/*
2697 * Advance the full-system-idle state. This is invoked when all of
2698 * the non-timekeeping CPUs are idle.
2699 */
2700static void rcu_sysidle(unsigned long j)
2701{
2702 /* Check the current state. */
2703 switch (READ_ONCE(full_sysidle_state)) {
2704 case RCU_SYSIDLE_NOT:
2705
2706 /* First time all are idle, so note a short idle period. */
2707 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2708 break;
2709
2710 case RCU_SYSIDLE_SHORT:
2711
2712 /*
2713 * Idle for a bit, time to advance to next state?
2714 * cmpxchg failure means race with non-idle, let them win.
2715 */
2716 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2717 (void)cmpxchg(&full_sysidle_state,
2718 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2719 break;
2720
2721 case RCU_SYSIDLE_LONG:
2722
2723 /*
2724 * Do an additional check pass before advancing to full.
2725 * cmpxchg failure means race with non-idle, let them win.
2726 */
2727 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2728 (void)cmpxchg(&full_sysidle_state,
2729 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2730 break;
2731
2732 default:
2733 break;
2734 }
2735}
2736
2737/*
2738 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2739 * back to the beginning.
2740 */
2741static void rcu_sysidle_cancel(void)
2742{
2743 smp_mb();
2744 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2745 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2746}
2747
2748/*
2749 * Update the sysidle state based on the results of a force-quiescent-state
2750 * scan of the CPUs' dyntick-idle state.
2751 */
2752static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2753 unsigned long maxj, bool gpkt)
2754{
2755 if (rsp != rcu_state_p)
2756 return; /* Wrong flavor, ignore. */
2757 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2758 return; /* Running state machine from timekeeping CPU. */
2759 if (isidle)
2760 rcu_sysidle(maxj); /* More idle! */
2761 else
2762 rcu_sysidle_cancel(); /* Idle is over. */
2763}
2764
2765/*
2766 * Wrapper for rcu_sysidle_report() when called from the grace-period
2767 * kthread's context.
2768 */
2769static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2770 unsigned long maxj)
2771{
2772 /* If there are no nohz_full= CPUs, no need to track this. */
2773 if (!tick_nohz_full_enabled())
2774 return;
2775
2776 rcu_sysidle_report(rsp, isidle, maxj, true);
2777}
2778
2779/* Callback and function for forcing an RCU grace period. */
2780struct rcu_sysidle_head {
2781 struct rcu_head rh;
2782 int inuse;
2783};
2784
2785static void rcu_sysidle_cb(struct rcu_head *rhp)
2786{
2787 struct rcu_sysidle_head *rshp;
2788
2789 /*
2790 * The following memory barrier is needed to replace the
2791 * memory barriers that would normally be in the memory
2792 * allocator.
2793 */
2794 smp_mb(); /* grace period precedes setting inuse. */
2795
2796 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2797 WRITE_ONCE(rshp->inuse, 0);
2798}
2799
2800/*
2801 * Check to see if the system is fully idle, other than the timekeeping CPU.
2802 * The caller must have disabled interrupts. This is not intended to be
2803 * called unless tick_nohz_full_enabled().
2804 */
2805bool rcu_sys_is_idle(void)
2806{
2807 static struct rcu_sysidle_head rsh;
2808 int rss = READ_ONCE(full_sysidle_state);
2809
2810 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2811 return false;
2812
2813 /* Handle small-system case by doing a full scan of CPUs. */
2814 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2815 int oldrss = rss - 1;
2816
2817 /*
2818 * One pass to advance to each state up to _FULL.
2819 * Give up if any pass fails to advance the state.
2820 */
2821 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2822 int cpu;
2823 bool isidle = true;
2824 unsigned long maxj = jiffies - ULONG_MAX / 4;
2825 struct rcu_data *rdp;
2826
2827 /* Scan all the CPUs looking for nonidle CPUs. */
2828 for_each_possible_cpu(cpu) {
2829 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2830 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2831 if (!isidle)
2832 break;
2833 }
2834 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2835 oldrss = rss;
2836 rss = READ_ONCE(full_sysidle_state);
2837 }
2838 }
2839
2840 /* If this is the first observation of an idle period, record it. */
2841 if (rss == RCU_SYSIDLE_FULL) {
2842 rss = cmpxchg(&full_sysidle_state,
2843 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2844 return rss == RCU_SYSIDLE_FULL;
2845 }
2846
2847 smp_mb(); /* ensure rss load happens before later caller actions. */
2848
2849 /* If already fully idle, tell the caller (in case of races). */
2850 if (rss == RCU_SYSIDLE_FULL_NOTED)
2851 return true;
2852
2853 /*
2854 * If we aren't there yet, and a grace period is not in flight,
2855 * initiate a grace period. Either way, tell the caller that
2856 * we are not there yet. We use an xchg() rather than an assignment
2857 * to make up for the memory barriers that would otherwise be
2858 * provided by the memory allocator.
2859 */
2860 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2861 !rcu_gp_in_progress(rcu_state_p) &&
2862 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2863 call_rcu(&rsh.rh, rcu_sysidle_cb);
2864 return false;
2865}
2866
2867/*
2868 * Initialize dynticks sysidle state for CPUs coming online.
2869 */
2870static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2871{
2872 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2873}
2874
2875#else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2876
2877static void rcu_sysidle_enter(int irq)
2878{
2879}
2880
2881static void rcu_sysidle_exit(int irq)
2882{
2883}
2884
2885static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2886 unsigned long *maxj)
2887{
2888}
2889
2890static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2891{
2892 return false;
2893}
2894
2895static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2896 unsigned long maxj)
2897{
2898}
2899
2900static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2901{
2902}
2903
2904#endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2905
2906/*
2907 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2908 * grace-period kthread will do force_quiescent_state() processing?
2909 * The idea is to avoid waking up RCU core processing on such a
2910 * CPU unless the grace period has extended for too long.
2911 *
2912 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2913 * CONFIG_RCU_NOCB_CPU CPUs.
2914 */
2915static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2916{
2917#ifdef CONFIG_NO_HZ_FULL
2918 if (tick_nohz_full_cpu(smp_processor_id()) &&
2919 (!rcu_gp_in_progress(rsp) ||
2920 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
2921 return true;
2922#endif /* #ifdef CONFIG_NO_HZ_FULL */
2923 return false;
2924}
2925
2926/*
2927 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2928 * timekeeping CPU.
2929 */
2930static void rcu_bind_gp_kthread(void)
2931{
2932 int __maybe_unused cpu;
2933
2934 if (!tick_nohz_full_enabled())
2935 return;
2936#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2937 cpu = tick_do_timer_cpu;
2938 if (cpu >= 0 && cpu < nr_cpu_ids)
2939 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2940#else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2941 housekeeping_affine(current);
2942#endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2943}
2944
2945/* Record the current task on dyntick-idle entry. */
2946static void rcu_dynticks_task_enter(void)
2947{
2948#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2949 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2950#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2951}
2952
2953/* Record no current task on dyntick-idle exit. */
2954static void rcu_dynticks_task_exit(void)
2955{
2956#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2957 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2958#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2959}