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/* 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
16static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
17{
18 /*
19 * In order to read the offloaded state of an rdp in a safe
20 * and stable way and prevent from its value to be changed
21 * under us, we must either hold the barrier mutex, the cpu
22 * hotplug lock (read or write) or the nocb lock. Local
23 * non-preemptible reads are also safe. NOCB kthreads and
24 * timers have their own means of synchronization against the
25 * offloaded state updaters.
26 */
27 RCU_LOCKDEP_WARN(
28 !(lockdep_is_held(&rcu_state.barrier_mutex) ||
29 (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
30 rcu_lockdep_is_held_nocb(rdp) ||
31 (rdp == this_cpu_ptr(&rcu_data) &&
32 !(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible())) ||
33 rcu_current_is_nocb_kthread(rdp)),
34 "Unsafe read of RCU_NOCB offloaded state"
35 );
36
37 return rcu_segcblist_is_offloaded(&rdp->cblist);
38}
39
40/*
41 * Check the RCU kernel configuration parameters and print informative
42 * messages about anything out of the ordinary.
43 */
44static void __init rcu_bootup_announce_oddness(void)
45{
46 if (IS_ENABLED(CONFIG_RCU_TRACE))
47 pr_info("\tRCU event tracing is enabled.\n");
48 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
49 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
50 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
51 RCU_FANOUT);
52 if (rcu_fanout_exact)
53 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
54 if (IS_ENABLED(CONFIG_PROVE_RCU))
55 pr_info("\tRCU lockdep checking is enabled.\n");
56 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
57 pr_info("\tRCU strict (and thus non-scalable) grace periods are enabled.\n");
58 if (RCU_NUM_LVLS >= 4)
59 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
60 if (RCU_FANOUT_LEAF != 16)
61 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
62 RCU_FANOUT_LEAF);
63 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
64 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
65 rcu_fanout_leaf);
66 if (nr_cpu_ids != NR_CPUS)
67 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
68#ifdef CONFIG_RCU_BOOST
69 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
70 kthread_prio, CONFIG_RCU_BOOST_DELAY);
71#endif
72 if (blimit != DEFAULT_RCU_BLIMIT)
73 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
74 if (qhimark != DEFAULT_RCU_QHIMARK)
75 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
76 if (qlowmark != DEFAULT_RCU_QLOMARK)
77 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
78 if (qovld != DEFAULT_RCU_QOVLD)
79 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
80 if (jiffies_till_first_fqs != ULONG_MAX)
81 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
82 if (jiffies_till_next_fqs != ULONG_MAX)
83 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
84 if (jiffies_till_sched_qs != ULONG_MAX)
85 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
86 if (rcu_kick_kthreads)
87 pr_info("\tKick kthreads if too-long grace period.\n");
88 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
89 pr_info("\tRCU callback double-/use-after-free debug is enabled.\n");
90 if (gp_preinit_delay)
91 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
92 if (gp_init_delay)
93 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
94 if (gp_cleanup_delay)
95 pr_info("\tRCU debug GP cleanup slowdown %d jiffies.\n", gp_cleanup_delay);
96 if (!use_softirq)
97 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
98 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
99 pr_info("\tRCU debug extended QS entry/exit.\n");
100 rcupdate_announce_bootup_oddness();
101}
102
103#ifdef CONFIG_PREEMPT_RCU
104
105static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
106static void rcu_read_unlock_special(struct task_struct *t);
107
108/*
109 * Tell them what RCU they are running.
110 */
111static void __init rcu_bootup_announce(void)
112{
113 pr_info("Preemptible hierarchical RCU implementation.\n");
114 rcu_bootup_announce_oddness();
115}
116
117/* Flags for rcu_preempt_ctxt_queue() decision table. */
118#define RCU_GP_TASKS 0x8
119#define RCU_EXP_TASKS 0x4
120#define RCU_GP_BLKD 0x2
121#define RCU_EXP_BLKD 0x1
122
123/*
124 * Queues a task preempted within an RCU-preempt read-side critical
125 * section into the appropriate location within the ->blkd_tasks list,
126 * depending on the states of any ongoing normal and expedited grace
127 * periods. The ->gp_tasks pointer indicates which element the normal
128 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
129 * indicates which element the expedited grace period is waiting on (again,
130 * NULL if none). If a grace period is waiting on a given element in the
131 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
132 * adding a task to the tail of the list blocks any grace period that is
133 * already waiting on one of the elements. In contrast, adding a task
134 * to the head of the list won't block any grace period that is already
135 * waiting on one of the elements.
136 *
137 * This queuing is imprecise, and can sometimes make an ongoing grace
138 * period wait for a task that is not strictly speaking blocking it.
139 * Given the choice, we needlessly block a normal grace period rather than
140 * blocking an expedited grace period.
141 *
142 * Note that an endless sequence of expedited grace periods still cannot
143 * indefinitely postpone a normal grace period. Eventually, all of the
144 * fixed number of preempted tasks blocking the normal grace period that are
145 * not also blocking the expedited grace period will resume and complete
146 * their RCU read-side critical sections. At that point, the ->gp_tasks
147 * pointer will equal the ->exp_tasks pointer, at which point the end of
148 * the corresponding expedited grace period will also be the end of the
149 * normal grace period.
150 */
151static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
152 __releases(rnp->lock) /* But leaves rrupts disabled. */
153{
154 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
155 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
156 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
157 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
158 struct task_struct *t = current;
159
160 raw_lockdep_assert_held_rcu_node(rnp);
161 WARN_ON_ONCE(rdp->mynode != rnp);
162 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
163 /* RCU better not be waiting on newly onlined CPUs! */
164 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
165 rdp->grpmask);
166
167 /*
168 * Decide where to queue the newly blocked task. In theory,
169 * this could be an if-statement. In practice, when I tried
170 * that, it was quite messy.
171 */
172 switch (blkd_state) {
173 case 0:
174 case RCU_EXP_TASKS:
175 case RCU_EXP_TASKS + RCU_GP_BLKD:
176 case RCU_GP_TASKS:
177 case RCU_GP_TASKS + RCU_EXP_TASKS:
178
179 /*
180 * Blocking neither GP, or first task blocking the normal
181 * GP but not blocking the already-waiting expedited GP.
182 * Queue at the head of the list to avoid unnecessarily
183 * blocking the already-waiting GPs.
184 */
185 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
186 break;
187
188 case RCU_EXP_BLKD:
189 case RCU_GP_BLKD:
190 case RCU_GP_BLKD + RCU_EXP_BLKD:
191 case RCU_GP_TASKS + RCU_EXP_BLKD:
192 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
193 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
194
195 /*
196 * First task arriving that blocks either GP, or first task
197 * arriving that blocks the expedited GP (with the normal
198 * GP already waiting), or a task arriving that blocks
199 * both GPs with both GPs already waiting. Queue at the
200 * tail of the list to avoid any GP waiting on any of the
201 * already queued tasks that are not blocking it.
202 */
203 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
204 break;
205
206 case RCU_EXP_TASKS + RCU_EXP_BLKD:
207 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
208 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
209
210 /*
211 * Second or subsequent task blocking the expedited GP.
212 * The task either does not block the normal GP, or is the
213 * first task blocking the normal GP. Queue just after
214 * the first task blocking the expedited GP.
215 */
216 list_add(&t->rcu_node_entry, rnp->exp_tasks);
217 break;
218
219 case RCU_GP_TASKS + RCU_GP_BLKD:
220 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
221
222 /*
223 * Second or subsequent task blocking the normal GP.
224 * The task does not block the expedited GP. Queue just
225 * after the first task blocking the normal GP.
226 */
227 list_add(&t->rcu_node_entry, rnp->gp_tasks);
228 break;
229
230 default:
231
232 /* Yet another exercise in excessive paranoia. */
233 WARN_ON_ONCE(1);
234 break;
235 }
236
237 /*
238 * We have now queued the task. If it was the first one to
239 * block either grace period, update the ->gp_tasks and/or
240 * ->exp_tasks pointers, respectively, to reference the newly
241 * blocked tasks.
242 */
243 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
244 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
245 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
246 }
247 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
248 WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
249 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
250 !(rnp->qsmask & rdp->grpmask));
251 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
252 !(rnp->expmask & rdp->grpmask));
253 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
254
255 /*
256 * Report the quiescent state for the expedited GP. This expedited
257 * GP should not be able to end until we report, so there should be
258 * no need to check for a subsequent expedited GP. (Though we are
259 * still in a quiescent state in any case.)
260 *
261 * Interrupts are disabled, so ->cpu_no_qs.b.exp cannot change.
262 */
263 if (blkd_state & RCU_EXP_BLKD && rdp->cpu_no_qs.b.exp)
264 rcu_report_exp_rdp(rdp);
265 else
266 WARN_ON_ONCE(rdp->cpu_no_qs.b.exp);
267}
268
269/*
270 * Record a preemptible-RCU quiescent state for the specified CPU.
271 * Note that this does not necessarily mean that the task currently running
272 * on the CPU is in a quiescent state: Instead, it means that the current
273 * grace period need not wait on any RCU read-side critical section that
274 * starts later on this CPU. It also means that if the current task is
275 * in an RCU read-side critical section, it has already added itself to
276 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
277 * current task, there might be any number of other tasks blocked while
278 * in an RCU read-side critical section.
279 *
280 * Unlike non-preemptible-RCU, quiescent state reports for expedited
281 * grace periods are handled separately via deferred quiescent states
282 * and context switch events.
283 *
284 * Callers to this function must disable preemption.
285 */
286static void rcu_qs(void)
287{
288 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
289 if (__this_cpu_read(rcu_data.cpu_no_qs.b.norm)) {
290 trace_rcu_grace_period(TPS("rcu_preempt"),
291 __this_cpu_read(rcu_data.gp_seq),
292 TPS("cpuqs"));
293 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
294 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
295 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
296 }
297}
298
299/*
300 * We have entered the scheduler, and the current task might soon be
301 * context-switched away from. If this task is in an RCU read-side
302 * critical section, we will no longer be able to rely on the CPU to
303 * record that fact, so we enqueue the task on the blkd_tasks list.
304 * The task will dequeue itself when it exits the outermost enclosing
305 * RCU read-side critical section. Therefore, the current grace period
306 * cannot be permitted to complete until the blkd_tasks list entries
307 * predating the current grace period drain, in other words, until
308 * rnp->gp_tasks becomes NULL.
309 *
310 * Caller must disable interrupts.
311 */
312void rcu_note_context_switch(bool preempt)
313{
314 struct task_struct *t = current;
315 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
316 struct rcu_node *rnp;
317
318 trace_rcu_utilization(TPS("Start context switch"));
319 lockdep_assert_irqs_disabled();
320 WARN_ONCE(!preempt && rcu_preempt_depth() > 0, "Voluntary context switch within RCU read-side critical section!");
321 if (rcu_preempt_depth() > 0 &&
322 !t->rcu_read_unlock_special.b.blocked) {
323
324 /* Possibly blocking in an RCU read-side critical section. */
325 rnp = rdp->mynode;
326 raw_spin_lock_rcu_node(rnp);
327 t->rcu_read_unlock_special.b.blocked = true;
328 t->rcu_blocked_node = rnp;
329
330 /*
331 * Verify the CPU's sanity, trace the preemption, and
332 * then queue the task as required based on the states
333 * of any ongoing and expedited grace periods.
334 */
335 WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp));
336 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
337 trace_rcu_preempt_task(rcu_state.name,
338 t->pid,
339 (rnp->qsmask & rdp->grpmask)
340 ? rnp->gp_seq
341 : rcu_seq_snap(&rnp->gp_seq));
342 rcu_preempt_ctxt_queue(rnp, rdp);
343 } else {
344 rcu_preempt_deferred_qs(t);
345 }
346
347 /*
348 * Either we were not in an RCU read-side critical section to
349 * begin with, or we have now recorded that critical section
350 * globally. Either way, we can now note a quiescent state
351 * for this CPU. Again, if we were in an RCU read-side critical
352 * section, and if that critical section was blocking the current
353 * grace period, then the fact that the task has been enqueued
354 * means that we continue to block the current grace period.
355 */
356 rcu_qs();
357 if (rdp->cpu_no_qs.b.exp)
358 rcu_report_exp_rdp(rdp);
359 rcu_tasks_qs(current, preempt);
360 trace_rcu_utilization(TPS("End context switch"));
361}
362EXPORT_SYMBOL_GPL(rcu_note_context_switch);
363
364/*
365 * Check for preempted RCU readers blocking the current grace period
366 * for the specified rcu_node structure. If the caller needs a reliable
367 * answer, it must hold the rcu_node's ->lock.
368 */
369static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
370{
371 return READ_ONCE(rnp->gp_tasks) != NULL;
372}
373
374/* limit value for ->rcu_read_lock_nesting. */
375#define RCU_NEST_PMAX (INT_MAX / 2)
376
377static void rcu_preempt_read_enter(void)
378{
379 WRITE_ONCE(current->rcu_read_lock_nesting, READ_ONCE(current->rcu_read_lock_nesting) + 1);
380}
381
382static int rcu_preempt_read_exit(void)
383{
384 int ret = READ_ONCE(current->rcu_read_lock_nesting) - 1;
385
386 WRITE_ONCE(current->rcu_read_lock_nesting, ret);
387 return ret;
388}
389
390static void rcu_preempt_depth_set(int val)
391{
392 WRITE_ONCE(current->rcu_read_lock_nesting, val);
393}
394
395/*
396 * Preemptible RCU implementation for rcu_read_lock().
397 * Just increment ->rcu_read_lock_nesting, shared state will be updated
398 * if we block.
399 */
400void __rcu_read_lock(void)
401{
402 rcu_preempt_read_enter();
403 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
404 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
405 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
406 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
407 barrier(); /* critical section after entry code. */
408}
409EXPORT_SYMBOL_GPL(__rcu_read_lock);
410
411/*
412 * Preemptible RCU implementation for rcu_read_unlock().
413 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
414 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
415 * invoke rcu_read_unlock_special() to clean up after a context switch
416 * in an RCU read-side critical section and other special cases.
417 */
418void __rcu_read_unlock(void)
419{
420 struct task_struct *t = current;
421
422 barrier(); // critical section before exit code.
423 if (rcu_preempt_read_exit() == 0) {
424 barrier(); // critical-section exit before .s check.
425 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
426 rcu_read_unlock_special(t);
427 }
428 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
429 int rrln = rcu_preempt_depth();
430
431 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
432 }
433}
434EXPORT_SYMBOL_GPL(__rcu_read_unlock);
435
436/*
437 * Advance a ->blkd_tasks-list pointer to the next entry, instead
438 * returning NULL if at the end of the list.
439 */
440static struct list_head *rcu_next_node_entry(struct task_struct *t,
441 struct rcu_node *rnp)
442{
443 struct list_head *np;
444
445 np = t->rcu_node_entry.next;
446 if (np == &rnp->blkd_tasks)
447 np = NULL;
448 return np;
449}
450
451/*
452 * Return true if the specified rcu_node structure has tasks that were
453 * preempted within an RCU read-side critical section.
454 */
455static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
456{
457 return !list_empty(&rnp->blkd_tasks);
458}
459
460/*
461 * Report deferred quiescent states. The deferral time can
462 * be quite short, for example, in the case of the call from
463 * rcu_read_unlock_special().
464 */
465static notrace void
466rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
467{
468 bool empty_exp;
469 bool empty_norm;
470 bool empty_exp_now;
471 struct list_head *np;
472 bool drop_boost_mutex = false;
473 struct rcu_data *rdp;
474 struct rcu_node *rnp;
475 union rcu_special special;
476
477 /*
478 * If RCU core is waiting for this CPU to exit its critical section,
479 * report the fact that it has exited. Because irqs are disabled,
480 * t->rcu_read_unlock_special cannot change.
481 */
482 special = t->rcu_read_unlock_special;
483 rdp = this_cpu_ptr(&rcu_data);
484 if (!special.s && !rdp->cpu_no_qs.b.exp) {
485 local_irq_restore(flags);
486 return;
487 }
488 t->rcu_read_unlock_special.s = 0;
489 if (special.b.need_qs) {
490 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
491 rdp->cpu_no_qs.b.norm = false;
492 rcu_report_qs_rdp(rdp);
493 udelay(rcu_unlock_delay);
494 } else {
495 rcu_qs();
496 }
497 }
498
499 /*
500 * Respond to a request by an expedited grace period for a
501 * quiescent state from this CPU. Note that requests from
502 * tasks are handled when removing the task from the
503 * blocked-tasks list below.
504 */
505 if (rdp->cpu_no_qs.b.exp)
506 rcu_report_exp_rdp(rdp);
507
508 /* Clean up if blocked during RCU read-side critical section. */
509 if (special.b.blocked) {
510
511 /*
512 * Remove this task from the list it blocked on. The task
513 * now remains queued on the rcu_node corresponding to the
514 * CPU it first blocked on, so there is no longer any need
515 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
516 */
517 rnp = t->rcu_blocked_node;
518 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
519 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
520 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
521 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
522 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
523 (!empty_norm || rnp->qsmask));
524 empty_exp = sync_rcu_exp_done(rnp);
525 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
526 np = rcu_next_node_entry(t, rnp);
527 list_del_init(&t->rcu_node_entry);
528 t->rcu_blocked_node = NULL;
529 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
530 rnp->gp_seq, t->pid);
531 if (&t->rcu_node_entry == rnp->gp_tasks)
532 WRITE_ONCE(rnp->gp_tasks, np);
533 if (&t->rcu_node_entry == rnp->exp_tasks)
534 WRITE_ONCE(rnp->exp_tasks, np);
535 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
536 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
537 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx.rtmutex) == t;
538 if (&t->rcu_node_entry == rnp->boost_tasks)
539 WRITE_ONCE(rnp->boost_tasks, np);
540 }
541
542 /*
543 * If this was the last task on the current list, and if
544 * we aren't waiting on any CPUs, report the quiescent state.
545 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
546 * so we must take a snapshot of the expedited state.
547 */
548 empty_exp_now = sync_rcu_exp_done(rnp);
549 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
550 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
551 rnp->gp_seq,
552 0, rnp->qsmask,
553 rnp->level,
554 rnp->grplo,
555 rnp->grphi,
556 !!rnp->gp_tasks);
557 rcu_report_unblock_qs_rnp(rnp, flags);
558 } else {
559 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
560 }
561
562 /*
563 * If this was the last task on the expedited lists,
564 * then we need to report up the rcu_node hierarchy.
565 */
566 if (!empty_exp && empty_exp_now)
567 rcu_report_exp_rnp(rnp, true);
568
569 /* Unboost if we were boosted. */
570 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
571 rt_mutex_futex_unlock(&rnp->boost_mtx.rtmutex);
572 } else {
573 local_irq_restore(flags);
574 }
575}
576
577/*
578 * Is a deferred quiescent-state pending, and are we also not in
579 * an RCU read-side critical section? It is the caller's responsibility
580 * to ensure it is otherwise safe to report any deferred quiescent
581 * states. The reason for this is that it is safe to report a
582 * quiescent state during context switch even though preemption
583 * is disabled. This function cannot be expected to understand these
584 * nuances, so the caller must handle them.
585 */
586static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
587{
588 return (__this_cpu_read(rcu_data.cpu_no_qs.b.exp) ||
589 READ_ONCE(t->rcu_read_unlock_special.s)) &&
590 rcu_preempt_depth() == 0;
591}
592
593/*
594 * Report a deferred quiescent state if needed and safe to do so.
595 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
596 * not being in an RCU read-side critical section. The caller must
597 * evaluate safety in terms of interrupt, softirq, and preemption
598 * disabling.
599 */
600notrace void rcu_preempt_deferred_qs(struct task_struct *t)
601{
602 unsigned long flags;
603
604 if (!rcu_preempt_need_deferred_qs(t))
605 return;
606 local_irq_save(flags);
607 rcu_preempt_deferred_qs_irqrestore(t, flags);
608}
609
610/*
611 * Minimal handler to give the scheduler a chance to re-evaluate.
612 */
613static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
614{
615 struct rcu_data *rdp;
616
617 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
618 rdp->defer_qs_iw_pending = false;
619}
620
621/*
622 * Handle special cases during rcu_read_unlock(), such as needing to
623 * notify RCU core processing or task having blocked during the RCU
624 * read-side critical section.
625 */
626static void rcu_read_unlock_special(struct task_struct *t)
627{
628 unsigned long flags;
629 bool irqs_were_disabled;
630 bool preempt_bh_were_disabled =
631 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
632
633 /* NMI handlers cannot block and cannot safely manipulate state. */
634 if (in_nmi())
635 return;
636
637 local_irq_save(flags);
638 irqs_were_disabled = irqs_disabled_flags(flags);
639 if (preempt_bh_were_disabled || irqs_were_disabled) {
640 bool expboost; // Expedited GP in flight or possible boosting.
641 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
642 struct rcu_node *rnp = rdp->mynode;
643
644 expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
645 (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
646 (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
647 ((rdp->grpmask & READ_ONCE(rnp->qsmask)) || t->rcu_blocked_node)) ||
648 (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
649 t->rcu_blocked_node);
650 // Need to defer quiescent state until everything is enabled.
651 if (use_softirq && (in_hardirq() || (expboost && !irqs_were_disabled))) {
652 // Using softirq, safe to awaken, and either the
653 // wakeup is free or there is either an expedited
654 // GP in flight or a potential need to deboost.
655 raise_softirq_irqoff(RCU_SOFTIRQ);
656 } else {
657 // Enabling BH or preempt does reschedule, so...
658 // Also if no expediting and no possible deboosting,
659 // slow is OK. Plus nohz_full CPUs eventually get
660 // tick enabled.
661 set_tsk_need_resched(current);
662 set_preempt_need_resched();
663 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
664 expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
665 // Get scheduler to re-evaluate and call hooks.
666 // If !IRQ_WORK, FQS scan will eventually IPI.
667 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
668 IS_ENABLED(CONFIG_PREEMPT_RT))
669 rdp->defer_qs_iw = IRQ_WORK_INIT_HARD(
670 rcu_preempt_deferred_qs_handler);
671 else
672 init_irq_work(&rdp->defer_qs_iw,
673 rcu_preempt_deferred_qs_handler);
674 rdp->defer_qs_iw_pending = true;
675 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
676 }
677 }
678 local_irq_restore(flags);
679 return;
680 }
681 rcu_preempt_deferred_qs_irqrestore(t, flags);
682}
683
684/*
685 * Check that the list of blocked tasks for the newly completed grace
686 * period is in fact empty. It is a serious bug to complete a grace
687 * period that still has RCU readers blocked! This function must be
688 * invoked -before- updating this rnp's ->gp_seq.
689 *
690 * Also, if there are blocked tasks on the list, they automatically
691 * block the newly created grace period, so set up ->gp_tasks accordingly.
692 */
693static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
694{
695 struct task_struct *t;
696
697 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
698 raw_lockdep_assert_held_rcu_node(rnp);
699 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
700 dump_blkd_tasks(rnp, 10);
701 if (rcu_preempt_has_tasks(rnp) &&
702 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
703 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
704 t = container_of(rnp->gp_tasks, struct task_struct,
705 rcu_node_entry);
706 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
707 rnp->gp_seq, t->pid);
708 }
709 WARN_ON_ONCE(rnp->qsmask);
710}
711
712/*
713 * Check for a quiescent state from the current CPU, including voluntary
714 * context switches for Tasks RCU. When a task blocks, the task is
715 * recorded in the corresponding CPU's rcu_node structure, which is checked
716 * elsewhere, hence this function need only check for quiescent states
717 * related to the current CPU, not to those related to tasks.
718 */
719static void rcu_flavor_sched_clock_irq(int user)
720{
721 struct task_struct *t = current;
722
723 lockdep_assert_irqs_disabled();
724 if (rcu_preempt_depth() > 0 ||
725 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
726 /* No QS, force context switch if deferred. */
727 if (rcu_preempt_need_deferred_qs(t)) {
728 set_tsk_need_resched(t);
729 set_preempt_need_resched();
730 }
731 } else if (rcu_preempt_need_deferred_qs(t)) {
732 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
733 return;
734 } else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
735 rcu_qs(); /* Report immediate QS. */
736 return;
737 }
738
739 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
740 if (rcu_preempt_depth() > 0 &&
741 __this_cpu_read(rcu_data.core_needs_qs) &&
742 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
743 !t->rcu_read_unlock_special.b.need_qs &&
744 time_after(jiffies, rcu_state.gp_start + HZ))
745 t->rcu_read_unlock_special.b.need_qs = true;
746}
747
748/*
749 * Check for a task exiting while in a preemptible-RCU read-side
750 * critical section, clean up if so. No need to issue warnings, as
751 * debug_check_no_locks_held() already does this if lockdep is enabled.
752 * Besides, if this function does anything other than just immediately
753 * return, there was a bug of some sort. Spewing warnings from this
754 * function is like as not to simply obscure important prior warnings.
755 */
756void exit_rcu(void)
757{
758 struct task_struct *t = current;
759
760 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
761 rcu_preempt_depth_set(1);
762 barrier();
763 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
764 } else if (unlikely(rcu_preempt_depth())) {
765 rcu_preempt_depth_set(1);
766 } else {
767 return;
768 }
769 __rcu_read_unlock();
770 rcu_preempt_deferred_qs(current);
771}
772
773/*
774 * Dump the blocked-tasks state, but limit the list dump to the
775 * specified number of elements.
776 */
777static void
778dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
779{
780 int cpu;
781 int i;
782 struct list_head *lhp;
783 struct rcu_data *rdp;
784 struct rcu_node *rnp1;
785
786 raw_lockdep_assert_held_rcu_node(rnp);
787 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
788 __func__, rnp->grplo, rnp->grphi, rnp->level,
789 (long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
790 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
791 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
792 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
793 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
794 __func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
795 READ_ONCE(rnp->exp_tasks));
796 pr_info("%s: ->blkd_tasks", __func__);
797 i = 0;
798 list_for_each(lhp, &rnp->blkd_tasks) {
799 pr_cont(" %p", lhp);
800 if (++i >= ncheck)
801 break;
802 }
803 pr_cont("\n");
804 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
805 rdp = per_cpu_ptr(&rcu_data, cpu);
806 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
807 cpu, ".o"[rcu_rdp_cpu_online(rdp)],
808 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
809 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
810 }
811}
812
813#else /* #ifdef CONFIG_PREEMPT_RCU */
814
815/*
816 * If strict grace periods are enabled, and if the calling
817 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
818 * report that quiescent state and, if requested, spin for a bit.
819 */
820void rcu_read_unlock_strict(void)
821{
822 struct rcu_data *rdp;
823
824 if (irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
825 return;
826 rdp = this_cpu_ptr(&rcu_data);
827 rdp->cpu_no_qs.b.norm = false;
828 rcu_report_qs_rdp(rdp);
829 udelay(rcu_unlock_delay);
830}
831EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
832
833/*
834 * Tell them what RCU they are running.
835 */
836static void __init rcu_bootup_announce(void)
837{
838 pr_info("Hierarchical RCU implementation.\n");
839 rcu_bootup_announce_oddness();
840}
841
842/*
843 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
844 * how many quiescent states passed, just if there was at least one since
845 * the start of the grace period, this just sets a flag. The caller must
846 * have disabled preemption.
847 */
848static void rcu_qs(void)
849{
850 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
851 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
852 return;
853 trace_rcu_grace_period(TPS("rcu_sched"),
854 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
855 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
856 if (__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
857 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
858}
859
860/*
861 * Register an urgently needed quiescent state. If there is an
862 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
863 * dyntick-idle quiescent state visible to other CPUs, which will in
864 * some cases serve for expedited as well as normal grace periods.
865 * Either way, register a lightweight quiescent state.
866 */
867void rcu_all_qs(void)
868{
869 unsigned long flags;
870
871 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
872 return;
873 preempt_disable(); // For CONFIG_PREEMPT_COUNT=y kernels
874 /* Load rcu_urgent_qs before other flags. */
875 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
876 preempt_enable();
877 return;
878 }
879 this_cpu_write(rcu_data.rcu_urgent_qs, false);
880 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
881 local_irq_save(flags);
882 rcu_momentary_dyntick_idle();
883 local_irq_restore(flags);
884 }
885 rcu_qs();
886 preempt_enable();
887}
888EXPORT_SYMBOL_GPL(rcu_all_qs);
889
890/*
891 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
892 */
893void rcu_note_context_switch(bool preempt)
894{
895 trace_rcu_utilization(TPS("Start context switch"));
896 rcu_qs();
897 /* Load rcu_urgent_qs before other flags. */
898 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
899 goto out;
900 this_cpu_write(rcu_data.rcu_urgent_qs, false);
901 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
902 rcu_momentary_dyntick_idle();
903out:
904 rcu_tasks_qs(current, preempt);
905 trace_rcu_utilization(TPS("End context switch"));
906}
907EXPORT_SYMBOL_GPL(rcu_note_context_switch);
908
909/*
910 * Because preemptible RCU does not exist, there are never any preempted
911 * RCU readers.
912 */
913static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
914{
915 return 0;
916}
917
918/*
919 * Because there is no preemptible RCU, there can be no readers blocked.
920 */
921static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
922{
923 return false;
924}
925
926/*
927 * Because there is no preemptible RCU, there can be no deferred quiescent
928 * states.
929 */
930static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
931{
932 return false;
933}
934
935// Except that we do need to respond to a request by an expedited
936// grace period for a quiescent state from this CPU. Note that in
937// non-preemptible kernels, there can be no context switches within RCU
938// read-side critical sections, which in turn means that the leaf rcu_node
939// structure's blocked-tasks list is always empty. is therefore no need to
940// actually check it. Instead, a quiescent state from this CPU suffices,
941// and this function is only called from such a quiescent state.
942notrace void rcu_preempt_deferred_qs(struct task_struct *t)
943{
944 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
945
946 if (READ_ONCE(rdp->cpu_no_qs.b.exp))
947 rcu_report_exp_rdp(rdp);
948}
949
950/*
951 * Because there is no preemptible RCU, there can be no readers blocked,
952 * so there is no need to check for blocked tasks. So check only for
953 * bogus qsmask values.
954 */
955static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
956{
957 WARN_ON_ONCE(rnp->qsmask);
958}
959
960/*
961 * Check to see if this CPU is in a non-context-switch quiescent state,
962 * namely user mode and idle loop.
963 */
964static void rcu_flavor_sched_clock_irq(int user)
965{
966 if (user || rcu_is_cpu_rrupt_from_idle()) {
967
968 /*
969 * Get here if this CPU took its interrupt from user
970 * mode or from the idle loop, and if this is not a
971 * nested interrupt. In this case, the CPU is in
972 * a quiescent state, so note it.
973 *
974 * No memory barrier is required here because rcu_qs()
975 * references only CPU-local variables that other CPUs
976 * neither access nor modify, at least not while the
977 * corresponding CPU is online.
978 */
979 rcu_qs();
980 }
981}
982
983/*
984 * Because preemptible RCU does not exist, tasks cannot possibly exit
985 * while in preemptible RCU read-side critical sections.
986 */
987void exit_rcu(void)
988{
989}
990
991/*
992 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
993 */
994static void
995dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
996{
997 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
998}
999
1000#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
1001
1002/*
1003 * If boosting, set rcuc kthreads to realtime priority.
1004 */
1005static void rcu_cpu_kthread_setup(unsigned int cpu)
1006{
1007 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1008#ifdef CONFIG_RCU_BOOST
1009 struct sched_param sp;
1010
1011 sp.sched_priority = kthread_prio;
1012 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1013#endif /* #ifdef CONFIG_RCU_BOOST */
1014
1015 WRITE_ONCE(rdp->rcuc_activity, jiffies);
1016}
1017
1018static bool rcu_is_callbacks_nocb_kthread(struct rcu_data *rdp)
1019{
1020#ifdef CONFIG_RCU_NOCB_CPU
1021 return rdp->nocb_cb_kthread == current;
1022#else
1023 return false;
1024#endif
1025}
1026
1027/*
1028 * Is the current CPU running the RCU-callbacks kthread?
1029 * Caller must have preemption disabled.
1030 */
1031static bool rcu_is_callbacks_kthread(struct rcu_data *rdp)
1032{
1033 return rdp->rcu_cpu_kthread_task == current ||
1034 rcu_is_callbacks_nocb_kthread(rdp);
1035}
1036
1037#ifdef CONFIG_RCU_BOOST
1038
1039/*
1040 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1041 * or ->boost_tasks, advancing the pointer to the next task in the
1042 * ->blkd_tasks list.
1043 *
1044 * Note that irqs must be enabled: boosting the task can block.
1045 * Returns 1 if there are more tasks needing to be boosted.
1046 */
1047static int rcu_boost(struct rcu_node *rnp)
1048{
1049 unsigned long flags;
1050 struct task_struct *t;
1051 struct list_head *tb;
1052
1053 if (READ_ONCE(rnp->exp_tasks) == NULL &&
1054 READ_ONCE(rnp->boost_tasks) == NULL)
1055 return 0; /* Nothing left to boost. */
1056
1057 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1058
1059 /*
1060 * Recheck under the lock: all tasks in need of boosting
1061 * might exit their RCU read-side critical sections on their own.
1062 */
1063 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1064 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1065 return 0;
1066 }
1067
1068 /*
1069 * Preferentially boost tasks blocking expedited grace periods.
1070 * This cannot starve the normal grace periods because a second
1071 * expedited grace period must boost all blocked tasks, including
1072 * those blocking the pre-existing normal grace period.
1073 */
1074 if (rnp->exp_tasks != NULL)
1075 tb = rnp->exp_tasks;
1076 else
1077 tb = rnp->boost_tasks;
1078
1079 /*
1080 * We boost task t by manufacturing an rt_mutex that appears to
1081 * be held by task t. We leave a pointer to that rt_mutex where
1082 * task t can find it, and task t will release the mutex when it
1083 * exits its outermost RCU read-side critical section. Then
1084 * simply acquiring this artificial rt_mutex will boost task
1085 * t's priority. (Thanks to tglx for suggesting this approach!)
1086 *
1087 * Note that task t must acquire rnp->lock to remove itself from
1088 * the ->blkd_tasks list, which it will do from exit() if from
1089 * nowhere else. We therefore are guaranteed that task t will
1090 * stay around at least until we drop rnp->lock. Note that
1091 * rnp->lock also resolves races between our priority boosting
1092 * and task t's exiting its outermost RCU read-side critical
1093 * section.
1094 */
1095 t = container_of(tb, struct task_struct, rcu_node_entry);
1096 rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
1097 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1098 /* Lock only for side effect: boosts task t's priority. */
1099 rt_mutex_lock(&rnp->boost_mtx);
1100 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1101 rnp->n_boosts++;
1102
1103 return READ_ONCE(rnp->exp_tasks) != NULL ||
1104 READ_ONCE(rnp->boost_tasks) != NULL;
1105}
1106
1107/*
1108 * Priority-boosting kthread, one per leaf rcu_node.
1109 */
1110static int rcu_boost_kthread(void *arg)
1111{
1112 struct rcu_node *rnp = (struct rcu_node *)arg;
1113 int spincnt = 0;
1114 int more2boost;
1115
1116 trace_rcu_utilization(TPS("Start boost kthread@init"));
1117 for (;;) {
1118 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1119 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1120 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1121 READ_ONCE(rnp->exp_tasks));
1122 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1123 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1124 more2boost = rcu_boost(rnp);
1125 if (more2boost)
1126 spincnt++;
1127 else
1128 spincnt = 0;
1129 if (spincnt > 10) {
1130 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1131 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1132 schedule_timeout_idle(2);
1133 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1134 spincnt = 0;
1135 }
1136 }
1137 /* NOTREACHED */
1138 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1139 return 0;
1140}
1141
1142/*
1143 * Check to see if it is time to start boosting RCU readers that are
1144 * blocking the current grace period, and, if so, tell the per-rcu_node
1145 * kthread to start boosting them. If there is an expedited grace
1146 * period in progress, it is always time to boost.
1147 *
1148 * The caller must hold rnp->lock, which this function releases.
1149 * The ->boost_kthread_task is immortal, so we don't need to worry
1150 * about it going away.
1151 */
1152static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1153 __releases(rnp->lock)
1154{
1155 raw_lockdep_assert_held_rcu_node(rnp);
1156 if (!rnp->boost_kthread_task ||
1157 (!rcu_preempt_blocked_readers_cgp(rnp) && !rnp->exp_tasks)) {
1158 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1159 return;
1160 }
1161 if (rnp->exp_tasks != NULL ||
1162 (rnp->gp_tasks != NULL &&
1163 rnp->boost_tasks == NULL &&
1164 rnp->qsmask == 0 &&
1165 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld ||
1166 IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)))) {
1167 if (rnp->exp_tasks == NULL)
1168 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1169 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1170 rcu_wake_cond(rnp->boost_kthread_task,
1171 READ_ONCE(rnp->boost_kthread_status));
1172 } else {
1173 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1174 }
1175}
1176
1177#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1178
1179/*
1180 * Do priority-boost accounting for the start of a new grace period.
1181 */
1182static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1183{
1184 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1185}
1186
1187/*
1188 * Create an RCU-boost kthread for the specified node if one does not
1189 * already exist. We only create this kthread for preemptible RCU.
1190 */
1191static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1192{
1193 unsigned long flags;
1194 int rnp_index = rnp - rcu_get_root();
1195 struct sched_param sp;
1196 struct task_struct *t;
1197
1198 if (rnp->boost_kthread_task)
1199 return;
1200
1201 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1202 "rcub/%d", rnp_index);
1203 if (WARN_ON_ONCE(IS_ERR(t)))
1204 return;
1205
1206 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1207 rnp->boost_kthread_task = t;
1208 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1209 sp.sched_priority = kthread_prio;
1210 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1211 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1212}
1213
1214static struct task_struct *rcu_boost_task(struct rcu_node *rnp)
1215{
1216 return READ_ONCE(rnp->boost_kthread_task);
1217}
1218
1219#else /* #ifdef CONFIG_RCU_BOOST */
1220
1221static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1222 __releases(rnp->lock)
1223{
1224 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1225}
1226
1227static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1228{
1229}
1230
1231static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1232{
1233}
1234
1235static struct task_struct *rcu_boost_task(struct rcu_node *rnp)
1236{
1237 return NULL;
1238}
1239#endif /* #else #ifdef CONFIG_RCU_BOOST */
1240
1241/*
1242 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
1243 * grace-period kthread will do force_quiescent_state() processing?
1244 * The idea is to avoid waking up RCU core processing on such a
1245 * CPU unless the grace period has extended for too long.
1246 *
1247 * This code relies on the fact that all NO_HZ_FULL CPUs are also
1248 * RCU_NOCB_CPU CPUs.
1249 */
1250static bool rcu_nohz_full_cpu(void)
1251{
1252#ifdef CONFIG_NO_HZ_FULL
1253 if (tick_nohz_full_cpu(smp_processor_id()) &&
1254 (!rcu_gp_in_progress() ||
1255 time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
1256 return true;
1257#endif /* #ifdef CONFIG_NO_HZ_FULL */
1258 return false;
1259}
1260
1261/*
1262 * Bind the RCU grace-period kthreads to the housekeeping CPU.
1263 */
1264static void rcu_bind_gp_kthread(void)
1265{
1266 if (!tick_nohz_full_enabled())
1267 return;
1268 housekeeping_affine(current, HK_TYPE_RCU);
1269}